DynamoRIO/dynamorio
1
0
Fork 0
mirror of https://github.com/DynamoRIO/dynamorio.git synced 2026-10-12 02:02:04 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
master
dynamorio/core/hotpatch.c
T
Bin Wang 9fe0764ea3 i#8021 kernel: Generate obj list for the kernel module by reusing CORE_SRCS/ARCH_SRCS (#8172) 
This PR is a follow up on #8169. Instead of hardcoding the kernel module
object file list, we are reusing existing `CORE_SRCS` and `ARCH_SRCS` to
generate the object file list. Overall, the current process is easier to
maintain and less error prone. It's also easier to replicate in a
different build system. The trade-offs are there are some files we don't
actually need getting pulled in, so the build time is slightly longer
(<10s longer) and the built `.ko` file is slightly larger (~130KB larger
with DEBUG ON, ~20KB larger with DEBUG OFF).

- Use existing `CORE_SRCS` and `ARCH_SRCS` defined in
`core/CMakeLists.txt` for kernel module object file list generation.
- Prevent `emit_utils.c` from being included in `ARCH_SRCS` twice when
`ARCH_NAME` and `ARCH_NAME_SHARED` are the same in
`core/CMakeLists.txt`. This isn't a bug per se: both CMake and kbuild
can handle duplicate files. I just thought this would be cleaner and
less confusing.
- Disable annotations (by `undef ANNOTATIONS`) in `configure.cmake.h`.
- Use `limits_wrapper.h` in `core/hotpatch.c`.
- Remove redundant `lib/dr_annotations.h` in `core/jit_opt.c`.

Issue: #8021
2026-10-07 14:21:33 -07:00

6102 lines
271 KiB
C
Raw Permalink Blame History

/* **********************************************************
* Copyright (c) 2011-2026 Google, Inc. All rights reserved.
* Copyright (c) 2005-2010 VMware, Inc. All rights reserved.
* **********************************************************/
/*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* * Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
*
* * Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* * Neither the name of VMware, Inc. nor the names of its contributors may be
* used to endorse or promote products derived from this software without
* specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL VMWARE, INC. OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
* DAMAGE.
*/
/* Copyright (c) 2005-2007 Determina Corp. */
/*
* hotpatch.c - Hot patching mechanism
*/
#include "globals.h"
#include "fragment.h"
#include "arch.h"
#include "instr.h"
#include "instr_create_shared.h"
#include "decode.h"
#include "instrument.h"
#include "hotpatch.h"
#include "hotpatch_interface.h"
#include "probe_api.h"
#include "moduledb.h" /* macros for nudge; can be moved with nudge to os.c */
#ifndef WINDOWS
#endif
#include "limits_wrapper.h" /* for ULLONG_MAX */
#include "fcache.h" /* for fcache_reset_all_caches_proactively */
#ifdef GBOP
# include "aslr.h"
#endif
#include "perscache.h"
#include "synch.h"
/* Note: HOT_PATCHING_INTERFACE refers to the mechanism for injecting code at
* arbitrary points in the application text. It was formerly known as
* constraint injection. It has nothing to do with the DR mechanism that
* allows for dynamically changing existing instructions, as indicated
* by INSTR_HOT_PATCHABLE in instr.h
*/
#ifdef HOT_PATCHING_INTERFACE /* Around the whole file */
/*----------------------------------------------------------------------------*/
/* Local typed constants. */
/* Local untyped constants. */
/* defined for non-GBOP as well since used for -probe_api */
# define HOTP_ONLY_GBOP_PRECEDENCE 10
/* Limits for vulnerability data provided by the constraint writer. */
enum {
MIN_HOTP_INTERFACE_VERSION = HOTP_INTERFACE_VERSION,
MAX_HOTP_INTERFACE_VERSION = HOTP_INTERFACE_VERSION,
MIN_NUM_VULNERABILITIES = 1,
MAX_NUM_VULNERABILITIES = 10000,
MIN_VULNERABILITY_ID_LENGTH = 1,
MAX_VULNERABILITY_ID_LENGTH = 32,
MIN_POLICY_ID_LENGTH = HOTP_POLICY_ID_LENGTH,
MAX_POLICY_ID_LENGTH = HOTP_POLICY_ID_LENGTH,
MIN_POLICY_VERSION = 1,
MAX_POLICY_VERSION = 10000,
MIN_NUM_SETS = 1,
MAX_NUM_SETS = 10000,
MIN_NUM_MODULES = 1,
MAX_NUM_MODULES = 10000,
/* We don't expect PE files to be more than 1 GB in size. */
MAX_MODULE_SIZE = 1024 * 1024 * 1024,
/* Can have PEs with time stamp as zero, though fairly unlikely; zero
* checksum is more likely, zero file version is somewhat likely and zero
* {image,code} size is extremely unlikely. The max values though are
* unlikely to exist in reality, we use these limits as the don't care
* values for timestamp, checksum, {image,code} size & file version.
*/
MIN_PE_TIMESTAMP = 0,
MAX_PE_TIMESTAMP = UINT_MAX,
PE_TIMESTAMP_IGNORE = UINT_MAX,
PE_TIMESTAMP_UNAVAILABLE = PE_TIMESTAMP_IGNORE - 1,
MIN_PE_CHECKSUM = 0,
MAX_PE_CHECKSUM = UINT_MAX,
PE_CHECKSUM_IGNORE = UINT_MAX,
PE_CHECKSUM_UNAVAILABLE = PE_CHECKSUM_IGNORE - 1,
MIN_PE_IMAGE_SIZE = 0,
MAX_PE_IMAGE_SIZE = UINT_MAX,
PE_IMAGE_SIZE_IGNORE = UINT_MAX,
PE_IMAGE_SIZE_UNAVAILABLE = PE_IMAGE_SIZE_IGNORE - 1,
MIN_PE_CODE_SIZE = 0, /* kbdus.dll has only data in it */
MAX_PE_CODE_SIZE = UINT_MAX,
PE_CODE_SIZE_IGNORE = UINT_MAX,
PE_CODE_SIZE_UNAVAILABLE = PE_CODE_SIZE_IGNORE - 1,
MIN_PE_FILE_VERSION = 0,
MAX_PE_FILE_VERSION = ULLONG_MAX,
PE_FILE_VERSION_IGNORE = IF_WINDOWS_ELSE(MODULE_FILE_VERSION_INVALID, ULLONG_MAX),
PE_FILE_VERSION_UNAVAILABLE = PE_FILE_VERSION_IGNORE - 1,
MIN_NUM_PATCH_POINT_HASHES = 1,
MAX_NUM_PATCH_POINT_HASHES = 10000,
MIN_HASH_START_OFFSET = 1,
MAX_HASH_START_OFFSET = MAX_MODULE_SIZE, /* Can't exceed module size. */
MIN_HASH_LENGTH = 1,
MAX_HASH_LENGTH = MAX_MODULE_SIZE, /* Can't exceed module size. */
MIN_HASH_VALUE = 0,
MAX_HASH_VALUE = UINT_MAX,
MIN_NUM_PATCH_POINTS = MIN_NUM_PATCH_POINT_HASHES,
MAX_NUM_PATCH_POINTS = MAX_NUM_PATCH_POINT_HASHES,
MIN_PATCH_OFFSET = 1,
MAX_PATCH_OFFSET = MAX_MODULE_SIZE, /* Can't exceed module size. */
MIN_PATCH_PRECEDENCE = 1,
MAX_PATCH_PRECEDENCE = 10000,
MIN_DETECTOR_OFFSET = 1,
/* Hot patch dlls shouldn't be anywhere near 10 MB in size;
* this check is just to catch some wrong file being loaded by accident.
* Today a typical hot patch is far less than 1k in size, so to hit 10 MB
* we would need a mininum of 10000 constraints of 1k each - unlikely.
*/
MAX_DETECTOR_OFFSET = 10 * 1024 * 1024,
/* Protectors should exist for all hot patches; even if it does nothing. */
MIN_PROTECTOR_OFFSET = 1,
MAX_PROTECTOR_OFFSET = MAX_DETECTOR_OFFSET,
/* Zero offset either means there is no protector or no control flow
* change is requested by the protector.
*/
MIN_RETURN_ADDR = 0,
/* We don't expect return addresses to be across modules; given that we
* don't expect a module to be more than 1 GB in size, the return address
* offset shouldn't be more than 1 GB too.
*/
MAX_RETURN_ADDR = MAX_PATCH_OFFSET,
MIN_MODE = HOTP_MODE_OFF,
MAX_MODE = HOTP_MODE_PROTECT,
/* case 8051: > 256KB per-process means we should start thinking about
* sharing. 24-Apr-07: sharing is in plan for 4.3 or 4.4; upping to 384k.
* Note: this is used only in debug builds; release builds can handle all
* sizes as long as we don't run out of memory. */
MAX_POLICY_FILE_SIZE = 384 * 1024
};
# define PE_NAME_IGNORE "*" /* Can't have strings in the enum. */
# define PE_NAME_UNAVAILABLE '\0'
/*----------------------------------------------------------------------------*/
/* Local type definitions. */
/* Module signature is used to uniquely describe a module, in our case, a Win32
* PE module.
*/
typedef struct { /* xref case 4688 */
const char *pe_name;
/* Don't-care values for pe_{checksum,timestamp,{image,code}_size,
* file_version} will be their respective MAX values. See enum above.
*/
uint pe_checksum;
uint pe_timestamp;
size_t pe_image_size;
/* Refers to the sum of the unpadded sizes of all executable sections in the
* PE image. The section size used is from get_image_section_unpadded_size()
* which equals VirtualSize (unless that is 0 in which case it equals SizeOfRawData).
*
* As an aside note that VirtualSize usually has no alignment padding while
* SizeOfRawData is typically padded to FileAlignment (the image loader pads
* VirtualSize to SectionAligment), so SizeOfRawData is often larger than
* VirtualSize for fully initialized sections (- this is the opposite of how it is
* in unix/elf, i.e., raw/file size is usually smaller than virtual/mem size because
* the latter does the alignment; also in unix, there are usually two different
* mmaps as opposed to one on windows to load the image). Though xref case 5355,
* what is actually accepted (and generated by some compilers) differs from what is
* typical/legal in pe specifications.
*
* Using _code_ rather than _text_ in the name because text usually refers
* only to the .text section.
*/
size_t pe_code_size;
/* Found in the resource section, some PE file may not have it, in which
* case it will be set to its don't-care value.
*/
uint64 pe_file_version;
} hotp_module_sig_t;
/* A patch point describes what application address to patch and the address of
* the hot patches that will be used for patching. If a hot patch (only a
* protector) intends to change the flow of application's execution, then the
* address to which control should go to after the hot patch is executed is
* also specified. A precedence attribute defines the order (rank) in which a
* particular patch is to be applied if more than one need to be applied at the
* same application offset. All addresses are relative to the base of the
* module.
*/
/* TODO: typedef uint app_rva_t to define offsets; app_pc is actually an address,
* not an offset, so don't use it for defining offsets
*/
/* app_pc is a pointer, not an offset; using it to a compute pointer with
* base address gives a compiler error about adding two pointers. Hence, a new
* type to define module offsets.
*/
typedef struct {
app_rva_t offset; /* offset relative to the base of the module where
* thepatch is to be applied */
/* TODO: clearly split each structure into read only and runtime data
* because things are tending to go out of synch again; can create a
* parallel tree later on.
*/
app_rva_t detector_fn; /* Offset of the detector function from the base
* of the hot patch dll */
app_rva_t protector_fn;
app_rva_t return_addr;
/* NYI (was never needed in practice, though at design time I thought this
* was needed for supporting multiple patches at the same address); lower
* numbers mean high precedence.
*/
uint precedence;
/*------------------------------------------------------------------------*/
/* The following fields are part of runtime policy/vulnerability data, not
* part of vulnerability definitions, i.e., shouldn't be shared across
* processes.
*/
/* TODO: num_injected at the vulnerability level; relevant here? */
/* Buffer to hold the trampoline with which a patch point was hooked in
* order to execute a hot patch in hotp_only mode. Should be NULL for
* regular hot patching, i.e., with fcache.
*/
byte *trampoline;
/* Pointer to the copy of app code that resides inside the trampoline,
* that gets executed at the end of trampoline execution; this is the app
* code that existed at the injection point. Used only by hotp_only.
*/
byte *app_code_copy;
/* Pointer to the cti target inside the trampoline (the one that is used to
* implement AFTER_INTERCEPT_LET_GO_ALT_DYN) that is used to
* change control flow. Used only in hotp_only mode for a patch point
* that requests a control flow change, i.e., has non-zero return_addr.
*/
byte *tramp_exit_tgt;
} hotp_patch_point_t;
/* Experiments showed that the maximum size of a single interception
* trampoline/hook is about 400 to 450 bytes, so 512 should be adequate.
*/
# define HOTP_ONLY_TRAMPOLINE_SIZE 512
# define HOTP_ONLY_NUM_THREADS_AT_INIT -1
/* A patch region size of 5 is used for hotp_only mode. This is done so
* that the same vm_area_t vector (hotp_patch_point_areas) can be used for
* patch point overlap checks and address lookup. Note: 5 is the minimum
* bytes needed to encode/insert a direct jmp with 32-bit displacement, i.e.,
* a hook.
* For hotp in code cache, all patch regions are points, so patch region size
* 1 is used. In this mode it is used only for patch address lookup.
*
* NOTE: Investigate issues when implementing hotp_only for native_exec dlls as
* we would have to have regions with different sizes - might trigger a
* few hotp asserts.
*
* Use -1 as an error catching value if this macro is used without -hot_patching.
*/
# define HOTP_ONLY_PATCH_REGION_SIZE (5)
# define HOTP_CACHE_PATCH_REGION_SIZE (1)
# define HOTP_BAD_PATCH_REGION_SIZE (-1)
# define HOTP_PATCH_REGION_SIZE \
(DYNAMO_OPTION(hot_patching) \
? (DYNAMO_OPTION(hotp_only) ? HOTP_ONLY_PATCH_REGION_SIZE \
: HOTP_CACHE_PATCH_REGION_SIZE) \
: HOTP_BAD_PATCH_REGION_SIZE)
/* This structure is used to define a hash value for a specified region
* around a patch point as decided by the hot patch writer. This hash, which
* is provided by the hot patch writer will be used at run time as part of the
* mechanism to identify a given PE module for injecting hot patches.
*/
typedef struct {
/* Offset, relative to the base of the module, that should be used as the
* starting point of hash computation string; for the module to be patched.
*/
app_rva_t start;
uint len; /* number of bytes to be used for hash computation */
uint hash_value;
} hotp_patch_point_hash_t;
typedef struct {
hotp_module_sig_t sig;
uint num_patch_points;
hotp_patch_point_t *patch_points;
uint num_patch_point_hashes;
hotp_patch_point_hash_t *hashes;
/* Data computed at run time; should be zeroed out at read time */
bool matched; /* True if current module is loaded & matched. */
app_pc base_address;
} hotp_module_t;
typedef struct {
uint num_modules;
hotp_module_t *modules;
} hotp_set_t;
/* Note: status and statistics are kept in a separate structure to allow for
* easy output, either via a file or via read only memory.
* Note: whole struct is runtime data; hence separated out.
*/
typedef struct {
hotp_exec_status_t exec_status;
/* Points to the one in hotp_policy_status_t to avoid duplication. */
hotp_inject_status_t *inject_status;
/* TODO: num_injected at the vulnerability level */
/* TODO: decide on the size of stats (uint or uint64) before finalizing
* interface with jim
*/
uint64 num_detected;
uint64 num_not_detected;
uint64 num_detector_error;
uint64 num_protected;
uint64 num_not_protected;
uint64 num_kill_thread;
uint64 num_kill_process;
uint64 num_raise_exception;
uint64 num_change_control_flow;
uint64 num_protector_error;
uint64 num_aborted;
} hotp_vul_info_t;
/* The types are defined as unique bit flags because it may be possible in
* future that we have a case that is more than one type.
* For example a hot patch with a symbolic offset may be
* HOTP_TYPE_SYMBOLIC_TYPE | HOTP_TYPE_HOT_PATCH, whereas a gbop hook may be
* HOTP_TYPE_SYMBOLIC_TYPE | HOTP_TYPE_GBOP_HOOK.
*/
typedef enum {
/* This represents the patches that fix vulnerabilities, as described by
* the hot patch injection design.
*/
HOTP_TYPE_HOT_PATCH = 0x1,
/* This represents all gbop hook points. This type is different in that it:
* 1. Isn't specified by a config file; well not as of now (XXX?),
* 2. Is specified by gbop_hooks and/or gbop_include_list (TODO: NYI),
* 3. Can't be turned off by modes file; will not as of now (XXX?),
* 4. Can be turned of by gbop_exclude_list (TODO: NYI),
* 5. Uses a symbolic name rather than identifying the PE uniquely,
* 6. Has a generic detector and protector which is part of the core, and,
* 7. Uses the core defaults for events, actions, dumps & forensics
* (TODO: NYI).
*/
HOTP_TYPE_GBOP_HOOK = 0x2,
/* Currently will be exclusive with HOTP_TYPE_{HOT_PATCH,GBOP_HOOK},
* eventually will co-exist.
*/
HOTP_TYPE_PROBE = 0x4
} hotp_type_t;
/* hotp_vul_t defines a vulnerability */
/* the entire expanded structure of hotp_vul_t consists of constant data,
* except for a couple of runtime data; this is so that policies can be
* easily read in from file/memory in a binary format, thus eliminating the
* need to do any data formatting/processing inside the core.
*/
typedef struct {
const char *vul_id;
/* policy_id is of the format XXXX.XXXX so that it can be used to
* generate the corresponding threat_id; so use
* char policy_id[MAX_POLICY_ID_LENGTH + 1];
* to be consistent with hotp_policy_status_t; TODO
* not done now because SET_STR doesn't handle arrays.
*/
const char *policy_id;
uint policy_version;
const char *hotp_dll;
const char *hotp_dll_hash;
hotp_policy_mode_t mode;
uint num_sets;
hotp_set_t *sets;
/* Data computed at run time; should be zeroed out at read time */
hotp_vul_info_t *info;
app_pc hotp_dll_base;
/* TODO: if policy data is going to be shared across multiple processes,
* info (i.e., runtime data) can't be part of this; a parallel runtime
* structure must be created; not a big issue till hot patches reach
* thousands in number
*/
/* XXX: right now this isn't specified by the config file because
* config files are assumed to define only hotpatches. Also, gbop_hooks
* are added to the table by a different routine, so there is no room
* for ambiguity. If we decide to use the config file for all, then
* this type should come from there - that would involve revving up
* the hotp interface, i.e., engine version.
* Note: probe types are provided by client libraries directly via
* dr_register_probes.
*/
hotp_type_t type;
/* The following fields were introduced for probe api. */
/* Unique ID for each probe; must be unique across different clients in the
* same process to avoid one client from controlling another's probes. */
unsigned int id;
} hotp_vul_t;
/* Maintain a list of vulnerability tables so that they can be freed at exit
* time. Nudge for policy reading creates new tables. The old ones should
* be left alone so that races between hot patch execution and table freeing
* are avoided (case 5521). All such tables are freed during DR exit.
* XXX: Release tables using a ref_count in case there are many & memory usage
* is high. It is highly unlikely that a given process will get more
* than a couple of policy read nudges during its lifetime.
* memory usage issue not correctness one, work on it after beta.
*/
typedef struct hotp_vul_tab_t {
hotp_vul_t *vul_tab;
uint num_vuls;
struct hotp_vul_tab_t *next;
} hotp_vul_tab_t;
/* TODO: for now this just has debug information; later on move all hot patch
* related globals into this structure. The debug variable listed below needed
* to be updated during loader activity and that conflicts with our data segment
* protection.
*/
# ifdef DEBUG
typedef struct hotp_globals_t {
/* The variables below help catch removing the same patch twice and
* injecting it twice, which is ok only for loader safety. Technically
* each patch point should have this variable, but given that the loader
* loads/relocates one dll at a time, this should be ok.
*/
bool ldr_safe_hook_removal; /* used only in -hotp_only mode */
bool ldr_safe_hook_injection; /* used only in -hotp_only mode */
} hotp_globals_t;
# endif
/*----------------------------------------------------------------------------*/
/* Macro definitions. */
/* These macros serve two purposes. Firstly they provide a clean interface
* to access the global hotp_vul_table, so that direct use of the global
* variable can be avoided. Secondly they improve readability; given that
* these structures are nested, accessing a member directly would result in
* long lines of code, which aren't very readable.
*/
/* TODO: Derek feels that these macros obfuscate the code rather than making
* them readable, which is opposite to what I thought. Try using local
* variables and if that looks good, remove these macros.
*/
# define VUL(vul_table, i) (vul_table[i])
# define SET(vul_table, v, i) (VUL(vul_table, v).sets[i])
# define MODULE(vul_table, v, s, i) (SET(vul_table, v, s).modules[i])
# define SIG(vul_table, v, s, m) (MODULE(vul_table, v, s, m).sig)
# define PPOINT(vul_table, v, s, m, i) (MODULE(vul_table, v, s, m).patch_points[i])
# define PPOINT_HASH(vul_table, v, s, m, i) (MODULE(vul_table, v, s, m).hashes[i])
# define NUM_GLOBAL_VULS (hotp_num_vuls)
# define GLOBAL_VUL_TABLE (hotp_vul_table)
# define GLOBAL_VUL(i) VUL(GLOBAL_VUL_TABLE, i)
# define GLOBAL_SET(v, i) SET(GLOBAL_VUL_TABLE, v, i)
# define GLOBAL_MODULE(v, s, i) MODULE(GLOBAL_VUL_TABLE, v, s, i)
# define GLOBAL_SIG(v, s, m) SIG(GLOBAL_VUL_TABLE, v, s, m)
# define GLOBAL_PPOINT(v, s, m, i) PPOINT(GLOBAL_VUL_TABLE, v, s, m, i)
# define GLOBAL_HASH(v, s, m, i) PPOINT_HASH(GLOBAL_VUL_TABLE, v, s, m, i)
/* TODO: change it to model ATOMIC_ADD; can't use ATOMIC_ADD directly because
* it wants only uint, not uint64 which is what all vulnerability stats
* are; maybe the easy way is to make the vul stat uint, but don't know
* if that will result in overflows fairly quickly, esp. for long running
* apps. either way, make this increment non racy, the users of this
* macro assume atomic increments.
*/
# define VUL_STAT_INC(x) ((x)++);
# define SET_NUM(var, type, limit_str, input_ptr) \
{ \
char *str = hotp_get_next_str(&(input_ptr)); \
const char *hex_fmt, *dec_fmt; \
type temp; \
\
ASSERT(sizeof(type) == sizeof(uint) || sizeof(type) == sizeof(uint64)); \
hex_fmt = \
(sizeof(type) == sizeof(uint)) ? "0x%x" : "0x" HEX64_FORMAT_STRING; \
dec_fmt = (sizeof(type) == sizeof(uint)) ? "%d" : INT64_FORMAT_STRING; \
\
if (sscanf(str, hex_fmt, &temp) == 1 || sscanf(str, dec_fmt, &temp) == 1) { \
if (temp < (type)(MIN_##limit_str) || temp > (type)(MAX_##limit_str)) \
goto error_reading_policy; /* Range error */ \
(var) = temp; \
} else \
goto error_reading_policy; /* Parse error. */ \
}
/* XXX: range check strs for min & max length; null check already done. */
# define SET_STR_DUP(var, input_ptr) \
{ \
char *str = hotp_get_next_str(&(input_ptr)); \
\
if (str == NULL) \
goto error_reading_policy; \
(var) = dr_strdup(str HEAPACCT(ACCT_HOT_PATCHING)); \
}
# define SET_STR_PTR(var, input_ptr) \
{ \
char *str = hotp_get_next_str(&(input_ptr)); \
\
if (str == NULL) \
goto error_reading_policy; \
(var) = str; \
}
# define SET_STR(var, input_ptr) SET_STR_DUP(var, input_ptr)
# define HOTP_IS_IN_REGION(region_start, region_size, addr) \
(((addr) >= (region_start)) && ((addr) < ((region_start) + (region_size))))
/* This checks addresses. */
# define HOTP_ONLY_IS_IN_TRAMPOLINE(ppoint, addr) \
(((ppoint)->trampoline == NULL || (addr) == NULL) \
? false \
: HOTP_IS_IN_REGION((ppoint)->trampoline, HOTP_ONLY_TRAMPOLINE_SIZE, addr))
/* This checks offsets/RVAs. */
# define HOTP_ONLY_IS_IN_PATCH_REGION(ppoint, addr) \
(((ppoint)->offset <= 0 || (addr) <= 0) \
? false \
: HOTP_IS_IN_REGION((ppoint)->offset, HOTP_PATCH_REGION_SIZE, addr))
/* TODO: PR 225550 - make this a better function so that each probe is
* identified uniquely so as to prevent clients from modifying each others'
* probes - make it a function of the client name, probe def & this counter.
* Note: probe id is generated outside hotp_vul_table_lock because of having
* to load probe/callback dlls without hitting dr hooks, so updates to
* probe_id_counter haver to be atomic.
*/
# define GENERATE_PROBE_ID() (atomic_add_exchange_int((int *)&probe_id_counter, 4))
/*----------------------------------------------------------------------------*/
/* Local function prototypes. */
static void
hotp_change_control_flow(const hotp_context_t *app_reg_ptr, const app_pc target);
static after_intercept_action_t
hotp_gateway(const hotp_vul_t *vul_tab, const uint num_vuls, const uint vul_index,
const uint set_index, const uint module_index, const uint ppoint_index,
hotp_context_t *app_reg_ptr, const bool own_hot_patch_lock);
static void
hotp_free_vul_table(hotp_vul_t *tab, uint num_vuls_alloc);
static hotp_exec_status_t
hotp_execute_patch(hotp_func_t hotp_fn_ptr, hotp_context_t *hotp_cxt,
hotp_policy_mode_t mode, bool dump_excpt_info, bool dump_error_info);
static void
hotp_update_vul_stats(const hotp_exec_status_t exec_status, const uint vul_index);
static void
hotp_event_notify(hotp_exec_status_t exec_status, bool protected,
const hotp_offset_match_t *inject_point, const app_pc bad_addr,
const hotp_context_t *hotp_cxt);
# if defined(DEBUG) && defined(INTERNAL)
static void
hotp_dump_reg_state(const hotp_context_t *reg_state, const app_pc eip,
const uint loglevel);
# endif
static void
hotp_only_inject_patch(const hotp_offset_match_t *ppoint_desc,
const thread_record_t **thread_table, const int num_threads);
static void
hotp_only_remove_patch(dcontext_t *dcontext, const hotp_module_t *module,
hotp_patch_point_t *cur_ppoint);
after_intercept_action_t
hotp_only_gateway(app_state_at_intercept_t *state);
static uint
hotp_compute_hash(app_pc base, hotp_patch_point_hash_t *hash);
# ifdef GBOP
static void
hotp_only_read_gbop_policy_defs(hotp_vul_t *tab, uint *num_vuls);
# endif
/* TODO: add function prototypes for all functions in this file */
/*----------------------------------------------------------------------------*/
/* Local data. */
hotp_policy_status_table_t *hotp_policy_status_table;
/* XXX: create hotp_vul_table_t and put these three into it */
static hotp_vul_t *hotp_vul_table;
static uint hotp_num_vuls;
static hotp_vul_tab_t *hotp_old_vul_tabs;
DECLARE_CXTSWPROT_VAR(static read_write_lock_t hotp_vul_table_lock, { { 0 } });
/* Special heap for hotp_only trampolines; heap is executable. */
static void *hotp_only_tramp_heap;
/* Leak to handle case 9593. This should go if we find a cleaner solution. */
# if defined(DEBUG) && defined(HEAP_ACCOUNTING)
DECLARE_NEVERPROT_VAR(int hotp_only_tramp_bytes_leaked, 0);
# endif
/* This is used to cache hotp_only_tramp_heap for handling leak asserts during
* detach and to track whether or not any hotp_only patch was removed. Case
* 9593 & PR 215520. */
static void *hotp_only_tramp_heap_cache;
/* Trampoline area vector; currently used only to identify if a thread is in
* the middle of hot patch execution during suspension - for multiprocessor
* safe hot patch removal in hotp_only mode.
* Kept on the heap for selfprot (case 7957).
*/
static vm_area_vector_t *hotp_only_tramp_areas;
/* This has all the matched patch points, i.e., patch points that have been
* determined by hotp_process_image() to be ready to be injected. Only that
* function adds or removes from this vector because only that function does
* module matching.
* The custom data stored is a hotp_offset_match_t structure which describes
* the patch point precisely in the GLOBAL_VUL_TABLE.
* For hotp_only this refers to all injected patches because they get injected
* during match/dll load time. For fcache based hot patches, this may or may
* not specify patch injection, but will specify matches. This is because for
* hotp_only matching & injection are done in one shot, whereas they are split
* for fcache based hot patches.
* This vector is not static, it is on the heap because of selfprot; case 8074.
* Uses:
* 1. for hotp_only to solve the overlapping hashes problem (case 7279).
* 2. for offset lookup for both hotp and hotp_only (case 8132).
* 3. NYI - all patch removal & injection; perscache stuff (case 10728).
*/
static vm_area_vector_t *hotp_patch_point_areas;
# ifdef DEBUG
static hotp_globals_t *hotp_globals;
# endif
/* Global counter used to generate unique ids for probes. This is updated
* atomically and isn't guarded by any lock. See GENERATE_PROBE_ID() for
* details.
*/
static unsigned int probe_id_counter;
/*----------------------------------------------------------------------------*/
/* Function definitions. */
/* Don't expose the hot patch lock directly outside this module. */
read_write_lock_t *
hotp_get_lock(void)
{
ASSERT(DYNAMO_OPTION(hot_patching));
return &hotp_vul_table_lock;
}
static inline app_pc
hotp_ppoint_addr(const hotp_module_t *module, const hotp_patch_point_t *ppoint)
{
app_pc ppoint_offset;
ASSERT(module != NULL && ppoint != NULL);
ASSERT(module->base_address != NULL && ppoint->offset != 0);
ppoint_offset = module->base_address + ppoint->offset;
/* The patch point should be inside the code section of a loaded module. */
ASSERT(is_in_code_section(module->base_address, ppoint_offset, NULL, NULL));
return ppoint_offset;
}
static void
hotp_ppoint_areas_add(hotp_offset_match_t *ppoint_desc)
{
hotp_module_t *module;
hotp_patch_point_t *ppoint;
hotp_offset_match_t *copy;
app_pc ppoint_start, ppoint_end;
ASSERT(ppoint_desc != NULL);
ASSERT(GLOBAL_VUL_TABLE != NULL && hotp_patch_point_areas != NULL);
ASSERT_OWN_READWRITE_LOCK(true, &hotp_vul_table_lock);
module = &GLOBAL_MODULE(ppoint_desc->vul_index, ppoint_desc->set_index,
ppoint_desc->module_index);
ppoint = &module->patch_points[ppoint_desc->ppoint_index];
/* Shouldn't be adding to hotp_patch_point_areas if the module hasn't been
* matched.
*/
ASSERT(module->matched);
ppoint_start = hotp_ppoint_addr(module, ppoint);
ppoint_end = ppoint_start + HOTP_PATCH_REGION_SIZE;
/* Each matched (or injected) patch point should be added only
* once and removed only once, so before adding, make sure that it
* is not already in there.
*/
ASSERT(!vmvector_overlap(hotp_patch_point_areas, ppoint_start, ppoint_end));
copy = HEAP_TYPE_ALLOC(GLOBAL_DCONTEXT, hotp_offset_match_t, ACCT_HOT_PATCHING,
PROTECTED);
*copy = *ppoint_desc;
vmvector_add(hotp_patch_point_areas, ppoint_start, ppoint_end, (void *)copy);
}
static void
hotp_ppoint_areas_remove(app_pc pc)
{
hotp_offset_match_t *ppoint_desc;
DEBUG_DECLARE(bool ok;)
ASSERT(pc != NULL);
ASSERT(GLOBAL_VUL_TABLE != NULL && hotp_patch_point_areas != NULL);
ASSERT_OWN_READWRITE_LOCK(true, &hotp_vul_table_lock);
ppoint_desc = (hotp_offset_match_t *)vmvector_lookup(hotp_patch_point_areas, pc);
DOCHECK(1, {
hotp_module_t *module;
hotp_patch_point_t *ppoint;
/* Shouldn't be trying to remove something that wasn't added. */
ASSERT(ppoint_desc != NULL);
/* Verify that the ppoint_desc in the vmvector corresponds to pc. */
module = &GLOBAL_MODULE(ppoint_desc->vul_index, ppoint_desc->set_index,
ppoint_desc->module_index);
ppoint = &module->patch_points[ppoint_desc->ppoint_index];
ASSERT(pc == hotp_ppoint_addr(module, ppoint));
});
HEAP_TYPE_FREE(GLOBAL_DCONTEXT, ppoint_desc, hotp_offset_match_t, ACCT_HOT_PATCHING,
PROTECTED);
DEBUG_DECLARE(ok =)
vmvector_remove(hotp_patch_point_areas, pc, pc + HOTP_PATCH_REGION_SIZE);
ASSERT(ok);
}
static void
hotp_ppoint_areas_release(void)
{
app_pc vm_start, vm_end;
hotp_offset_match_t *ppoint_desc;
vmvector_iterator_t iterator;
ASSERT_OWN_READWRITE_LOCK(true, &hotp_vul_table_lock);
/* Release all patch point descriptors. */
vmvector_iterator_start(hotp_patch_point_areas, &iterator);
while (vmvector_iterator_hasnext(&iterator)) {
ppoint_desc = vmvector_iterator_next(&iterator, &vm_start, &vm_end);
ASSERT(ppoint_desc != NULL);
HEAP_TYPE_FREE(GLOBAL_DCONTEXT, ppoint_desc, hotp_offset_match_t,
ACCT_HOT_PATCHING, PROTECTED);
}
vmvector_iterator_stop(&iterator);
/* Remove all vm_areas in the vmvector. */
vmvector_remove(hotp_patch_point_areas, UNIVERSAL_REGION_BASE, UNIVERSAL_REGION_END);
ASSERT(vmvector_empty(hotp_patch_point_areas));
}
/* Used to read in vulnerability definitions from file */
static char *
hotp_get_next_str(char **start)
{
char *end = *start, *temp = *start;
bool dos_line_terminator = false;
if (start == NULL || *start == NULL)
return NULL;
while (*end != '\n' && *end != '\r' && *end != '\0')
end++;
if (*end != '\0') {
if (*end == '\r') {
if (*(end + 1) == '\n')
dos_line_terminator = true;
else
SYSLOG_INTERNAL_WARNING("Mac OS style line separator!");
}
*end++ = '\0';
if (dos_line_terminator)
end++;
}
*start = end;
return temp;
}
/* Used to read either the policy file or the modes file. */
enum { POLICY_FILE = 1, MODES_FILE };
static char *
hotp_read_data_file(uint type, size_t *buf_len /* OUT */)
{
int retval;
char file[MAXIMUM_PATH];
ASSERT(type == POLICY_FILE || type == MODES_FILE);
ASSERT(buf_len != NULL);
*buf_len = 0;
retval = d_r_get_parameter(type == POLICY_FILE
? PARAM_STR(DYNAMORIO_VAR_HOT_PATCH_POLICIES)
: PARAM_STR(DYNAMORIO_VAR_HOT_PATCH_MODES),
file, BUFFER_SIZE_ELEMENTS(file));
if (IS_GET_PARAMETER_FAILURE(retval)) {
SYSLOG_INTERNAL_WARNING("Can't find %s definition directory name.",
(type == POLICY_FILE) ? "policy" : "mode");
return NULL;
}
/* The {defs,modes} file is
* $DYNAMORIO_HOT_PATCH_POLICIES/<engine>/HOTP_{POLICIES,MODES}_FILENAME
*/
CAT_AND_TERMINATE(file, "\\" STRINGIFY(HOTP_INTERFACE_VERSION) "\\");
CAT_AND_TERMINATE(file,
type == POLICY_FILE ? HOTP_POLICIES_FILENAME : HOTP_MODES_FILENAME);
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "Hot patch %s definition file: %s\n",
(type == POLICY_FILE) ? "policy" : "mode", file);
return read_entire_file(file, buf_len HEAPACCT(ACCT_HOT_PATCHING));
}
/* On a successful read, this should return a valid pointer to a vulnerability
* table and modify the size argument passed to it. If it fails, it should
* dump a log event, return NULL & not modify the size.
*
* The caller should release the old table & make the return value the new
* table; the reason for doing this table swap outside this function is to
* allow (future work) identification of vulnerabilities that have actually
* changed; from this set of changed vulnerabilities, identify those that have
* been injected and flush only those (an optimization issue).
*
* Policy file format: (indentations don't appear in the actual file, they exist
* here to illustrate the format & to show where multiple data can occur; also
* format is close to binary as it is now)
* All integers/hex_numbers are 32-bits unless explicitly stated otherwise.
<engine_version-str>
<num_vulnerabilities-decimal_integer>
<vulnerability_id-str>
<policy_id-str>
<version-decimal_integer>
<hotpatch_dll-str>
<hotpatch_dll_hash-str>
<num_sets-decimal_integer>
<num_modules-decimal_integer>
<pe_name-str>
<pe_timestamp-hex_number>
<pe_checksum-hex_number>
<pe_image_size-hex_number>
<pe_code_size-hex_number>
<pe_file_version-hex_number-64_bits>
<num_hashes-decimal_integer>
<start-hex_number>
<length-hex_number>
<hash-decimal_integer>
<num_patch_points-decimal_integer>
<offset-hex_number>
<precedence-decimal_integer>
<detector_offset-hex_number>
<protector_offset-hex_number>
<return_addr-hex_number>
* TODO: all unused fields, i.e., runtime fields in the data structures should
* be set to NULL/0 to avoid any assumption violations down stream.
* TODO: after reading in the vulnerability data, that region should be write
* protected
*/
static hotp_vul_t *
hotp_read_policy_defs(uint *num_vuls_read)
{
hotp_vul_t *tab = NULL;
uint hotp_interface_version;
uint vul = 0, set, module, hash, ppoint;
uint num_vuls = 0, num_vuls_alloc = 0;
char *buf = NULL; /* TODO: for now only; will go after file mapping */
size_t buf_len = 0;
char *start = NULL;
DEBUG_DECLARE(bool started_parsing = false;)
/* Read the config file only if -liveshields is turned on. If it isn't
* turned on, read gbop hooks if -gbop is specified.
*/
if (!DYNAMO_OPTION(liveshields)) {
# ifdef GBOP
if (DYNAMO_OPTION(gbop))
goto read_gbop_only;
# endif
return NULL;
}
buf = hotp_read_data_file(POLICY_FILE, &buf_len);
if (buf == NULL) {
ASSERT(buf_len == 0);
goto error_reading_policy;
} else {
ASSERT(buf_len > 0);
ASSERT_CURIOSITY(buf_len < MAX_POLICY_FILE_SIZE);
}
start = buf;
DEBUG_DECLARE(started_parsing = true;)
SET_NUM(hotp_interface_version, uint, HOTP_INTERFACE_VERSION, start);
SET_NUM(num_vuls, uint, NUM_VULNERABILITIES, start);
# ifdef GBOP
if (DYNAMO_OPTION(gbop))
num_vuls_alloc = gbop_get_num_hooks();
# endif
num_vuls_alloc += num_vuls;
ASSERT(num_vuls_alloc > 0 && num_vuls_alloc <= MAX_NUM_VULNERABILITIES);
/* Zero out all dynamically allocated hotpatch table structures to avoid
* leaks when there is a parse error. See case 8272, 9045.
*/
tab = HEAP_ARRAY_ALLOC_MEMSET(GLOBAL_DCONTEXT, hotp_vul_t, num_vuls_alloc,
ACCT_HOT_PATCHING, PROTECTED, 0);
for (vul = 0; vul < num_vuls; vul++) {
/* XXX: bounds checking; length should be > 2 && < 32; not null */
SET_STR(VUL(tab, vul).vul_id, start);
SET_STR(VUL(tab, vul).policy_id, start);
SET_NUM(VUL(tab, vul).policy_version, uint, POLICY_VERSION, start);
/* XXX: strdup strings because the buffer/mapped file will be deleted
* after processing; don't use strdup though!
* works right now till the next time I read in a policy file
* into buf[]; if that read fails the old data will be corrupt!
* remember, if not strdup'ed, all strings are in writable memory
*/
SET_STR(VUL(tab, vul).hotp_dll, start);
SET_STR(VUL(tab, vul).hotp_dll_hash, start);
SET_NUM(VUL(tab, vul).num_sets, uint, NUM_SETS, start);
/* Initialize all runtime values in the structure. */
VUL(tab, vul).mode = HOTP_MODE_OFF; /* Fix for case 5326. */
VUL(tab, vul).type = HOTP_TYPE_HOT_PATCH;
VUL(tab, vul).sets =
HEAP_ARRAY_ALLOC_MEMSET(GLOBAL_DCONTEXT, hotp_set_t, VUL(tab, vul).num_sets,
ACCT_HOT_PATCHING, PROTECTED, 0);
VUL(tab, vul).info = HEAP_ARRAY_ALLOC_MEMSET(GLOBAL_DCONTEXT, hotp_vul_info_t, 1,
ACCT_HOT_PATCHING, PROTECTED, 0);
for (set = 0; set < VUL(tab, vul).num_sets; set++) {
SET_NUM(SET(tab, vul, set).num_modules, uint, NUM_MODULES, start);
SET(tab, vul, set).modules = HEAP_ARRAY_ALLOC_MEMSET(
GLOBAL_DCONTEXT, hotp_module_t, SET(tab, vul, set).num_modules,
ACCT_HOT_PATCHING, PROTECTED, 0);
for (module = 0; module < SET(tab, vul, set).num_modules; module++) {
SET_STR(SIG(tab, vul, set, module).pe_name, start);
SET_NUM(SIG(tab, vul, set, module).pe_timestamp, uint, PE_TIMESTAMP,
start);
SET_NUM(SIG(tab, vul, set, module).pe_checksum, uint, PE_CHECKSUM, start);
SET_NUM(SIG(tab, vul, set, module).pe_image_size, uint, PE_IMAGE_SIZE,
start);
SET_NUM(SIG(tab, vul, set, module).pe_code_size, uint, PE_CODE_SIZE,
start);
SET_NUM(SIG(tab, vul, set, module).pe_file_version, uint64,
PE_FILE_VERSION, start);
/* Initialize all runtime values in the structure. */
MODULE(tab, vul, set, module).matched = false;
MODULE(tab, vul, set, module).base_address = NULL;
SET_NUM(MODULE(tab, vul, set, module).num_patch_point_hashes, uint,
NUM_PATCH_POINT_HASHES, start);
MODULE(tab, vul, set, module).hashes = HEAP_ARRAY_ALLOC_MEMSET(
GLOBAL_DCONTEXT, hotp_patch_point_hash_t,
MODULE(tab, vul, set, module).num_patch_point_hashes,
ACCT_HOT_PATCHING, PROTECTED, 0);
for (hash = 0;
hash < MODULE(tab, vul, set, module).num_patch_point_hashes;
hash++) {
SET_NUM(PPOINT_HASH(tab, vul, set, module, hash).start, app_rva_t,
HASH_START_OFFSET, start);
SET_NUM(PPOINT_HASH(tab, vul, set, module, hash).len, uint,
HASH_LENGTH, start);
SET_NUM(PPOINT_HASH(tab, vul, set, module, hash).hash_value, uint,
HASH_VALUE, start);
}
SET_NUM(MODULE(tab, vul, set, module).num_patch_points, uint,
NUM_PATCH_POINTS, start);
MODULE(tab, vul, set, module).patch_points = HEAP_ARRAY_ALLOC_MEMSET(
GLOBAL_DCONTEXT, hotp_patch_point_t,
MODULE(tab, vul, set, module).num_patch_points, ACCT_HOT_PATCHING,
PROTECTED, 0);
for (ppoint = 0; ppoint < MODULE(tab, vul, set, module).num_patch_points;
ppoint++) {
SET_NUM(PPOINT(tab, vul, set, module, ppoint).offset, app_rva_t,
PATCH_OFFSET, start);
SET_NUM(PPOINT(tab, vul, set, module, ppoint).precedence, uint,
PATCH_PRECEDENCE, start);
SET_NUM(PPOINT(tab, vul, set, module, ppoint).detector_fn, app_rva_t,
DETECTOR_OFFSET, start);
/* Both protector and return address can be NULL */
SET_NUM(PPOINT(tab, vul, set, module, ppoint).protector_fn, app_rva_t,
PROTECTOR_OFFSET, start);
SET_NUM(PPOINT(tab, vul, set, module, ppoint).return_addr, app_rva_t,
RETURN_ADDR, start);
PPOINT(tab, vul, set, module, ppoint).trampoline = NULL;
PPOINT(tab, vul, set, module, ppoint).app_code_copy = NULL;
PPOINT(tab, vul, set, module, ppoint).tramp_exit_tgt = NULL;
}
}
}
}
# ifdef GBOP
if (DYNAMO_OPTION(gbop)) {
hotp_only_read_gbop_policy_defs(tab, &num_vuls /* IN OUT arg */);
ASSERT(num_vuls_alloc == num_vuls);
}
# endif
*num_vuls_read = num_vuls;
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "read %d vulnerability definitions\n", num_vuls);
heap_free(GLOBAL_DCONTEXT, buf, buf_len HEAPACCT(ACCT_HOT_PATCHING));
return tab;
error_reading_policy:
/* TODO: log error, free stuff, set tab to null, leave size intact & exit
* for now just assert to make sure that bugs don't escape.
*/
/* TODO: provide line #, not offset; alex couldn't use the offset */
SYSLOG_INTERNAL_WARNING("Error reading or parsing hot patch definitions");
/* Need this curiosity to make qa notice; the warning is handy for
* development testing only. No hot patching on Linux, so don't assert.
* XXX: Convert to assert after case 9066 has been fixed & tested.
* Note: Warn for missing file, but assert for parsing error; latter is
* bug, former may just be a hotpatch-less installation - mostly coredev.
*/
IF_WINDOWS(
ASSERT_CURIOSITY(!started_parsing && "Error parsing hot patch definitions");)
*num_vuls_read = 0;
if (tab != NULL) {
ASSERT(num_vuls_alloc > 0);
/* If gbop is on, then during a parse error num_vuls (parsed) must be
* less than num_vuls_alloc because if table as been allocated space
* has been allocated for gbop entries as well which wouldn't have been
* read on a parse error. It is read after this point; see below.
*/
IF_GBOP(ASSERT(!DYNAMO_OPTION(gbop) || num_vuls < num_vuls_alloc);)
/* On error free the whole table, not just what was read; case 9044. */
hotp_free_vul_table(tab, num_vuls_alloc);
tab = NULL;
}
/* buf can be allocated even if vulnerability table hasn't been allocated.
* See case 8332.
*/
if (buf != NULL) {
ASSERT(buf_len > 0);
heap_free(GLOBAL_DCONTEXT, buf, buf_len HEAPACCT(ACCT_HOT_PATCHING));
LOG(GLOBAL, LOG_HOT_PATCHING, 1,
"error reading vulnerability file at offset " SZFMT "\n",
(ptr_uint_t)(start - buf));
}
# ifdef GBOP
/* Even if we couldn't read the hot patch policies, we should still
* allocate a new table and read in the gbop hooks.
*/
read_gbop_only:
if (DYNAMO_OPTION(gbop)) {
num_vuls_alloc = gbop_get_num_hooks();
ASSERT(num_vuls_alloc > 0 && num_vuls_alloc <= MAX_NUM_VULNERABILITIES);
num_vuls = 0;
tab = HEAP_ARRAY_ALLOC(GLOBAL_DCONTEXT, hotp_vul_t, num_vuls_alloc,
ACCT_HOT_PATCHING, PROTECTED);
hotp_only_read_gbop_policy_defs(tab, &num_vuls /* IN OUT arg */);
ASSERT(num_vuls_alloc == num_vuls);
*num_vuls_read = num_vuls;
}
# endif
return tab;
}
/* TODO: An efficiency issue: don't load all hot patch dlls unless the mode
* for at least one corresponding policy is detect or protect; this will
* avoid loading all hot patch dlls whether they are used or not. Note:
* this is still eager loading as per the design.
*/
static void
hotp_load_hotp_dlls(hotp_vul_t *vul_tab, uint num_vuls)
{
uint vul;
int retval;
/* TODO: these arrays are large so make them static with a lock to avoid
* a potential runtime stack overflow.
*/
char hotp_dll_path[MAXIMUM_PATH];
char hotp_dll_cache[MAXIMUM_PATH];
/* Only liveshields need to know DYNAMORIO_HOME, probes give full paths. */
if (DYNAMO_OPTION(liveshields)) {
/* If null or non-existent hotp_dll_cache directory raise error log,
* disable all associated vuls? We are going to assert/log if we can't
* find the dll (below) anyway.
*/
retval = d_r_get_parameter(PARAM_STR(DYNAMORIO_VAR_HOME), hotp_dll_cache,
BUFFER_SIZE_ELEMENTS(hotp_dll_cache));
if (IS_GET_PARAMETER_FAILURE(retval)) {
SYSLOG_INTERNAL_WARNING("Can't read %s. Hot patch dll loading "
"failed; hot patching won't work.",
DYNAMORIO_VAR_HOME);
return;
}
} else {
ASSERT(DYNAMO_OPTION(probe_api));
}
/* Compute dll cache path, i.e., $DYNAMORIO_HOME/lib/hotp/<engine>/ */
NULL_TERMINATE_BUFFER(hotp_dll_cache);
CAT_AND_TERMINATE(hotp_dll_cache, HOT_PATCHING_DLL_CACHE_PATH);
CAT_AND_TERMINATE(hotp_dll_cache, STRINGIFY(HOTP_INTERFACE_VERSION) "\\");
for (vul = 0; vul < num_vuls; vul++) {
/* Hot patch dlls exist only for the type hot_patch and probe, not for
* gbop hooks; well, not at least for now.
*/
if (!TESTANY(HOTP_TYPE_HOT_PATCH | HOTP_TYPE_PROBE, VUL(vul_tab, vul).type)) {
ASSERT(TESTALL(HOTP_TYPE_GBOP_HOOK, VUL(vul_tab, vul).type));
/* TODO: also assert that the base is dynamorio.dll & remediator
* offsets are what they should be - use a DODEBUG
*/
continue;
}
if (VUL(vul_tab, vul).hotp_dll_base == NULL) { /* Not loaded yet. */
ASSERT(
TESTANY(HOTP_TYPE_HOT_PATCH | HOTP_TYPE_PROBE, VUL(vul_tab, vul).type));
ASSERT(VUL(vul_tab, vul).hotp_dll != NULL);
/* Liveshields give just the base name which is used to compute
* full path, i.e., DYNAMORIO_HOME/lib/hotp/hotp_dll. */
if (TESTANY(HOTP_TYPE_HOT_PATCH, VUL(vul_tab, vul).type)) {
strncpy(hotp_dll_path, hotp_dll_cache,
BUFFER_SIZE_ELEMENTS(hotp_dll_path) - 1);
NULL_TERMINATE_BUFFER(hotp_dll_path);
/* Hot patch dll names should just be base names; with no / or \. */
ASSERT(strchr(VUL(vul_tab, vul).hotp_dll, '\\') == NULL &&
strchr(VUL(vul_tab, vul).hotp_dll, '/') == NULL);
strncat(hotp_dll_path, VUL(vul_tab, vul).hotp_dll,
BUFFER_SIZE_ELEMENTS(hotp_dll_path) - strlen(hotp_dll_path) - 1);
} else {
/* Probe api calls provide full path to hotp dlls. */
strncpy(hotp_dll_path, VUL(vul_tab, vul).hotp_dll,
BUFFER_SIZE_ELEMENTS(hotp_dll_path) - 1);
}
NULL_TERMINATE_BUFFER(hotp_dll_path);
ASSERT(strlen(hotp_dll_path) < BUFFER_SIZE_ELEMENTS(hotp_dll_path));
/* TODO: check if file exists; if not log, turn off associated
* vulnerabilities & bail out; need to think through the
* error exit mechanism while reading polcy-{defs,modes}.
*/
/* XXX: currently our loadlibrary hits our own syscall_while_native
* hook and goes to d_r_dispatch, which expects protected data sections.
* Once we have our own loader we can remove this.
*/
VUL(vul_tab, vul).hotp_dll_base =
load_shared_library(hotp_dll_path, false /*!reachable*/);
/* TODO: if module base is null, raise a log event, mark vul as not
* usable (probably a new status) and move on; for now just
* assert.
*/
/* TODO: assert that all detector_fn & protector_fn offsets
* associated with this hotp_dll actually lie within its
* text space.
*/
if (VUL(vul_tab, vul).hotp_dll_base == NULL) {
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "unable to load hotp_dll: %s\n",
hotp_dll_path);
ASSERT(VUL(vul_tab, vul).hotp_dll_base != NULL);
}
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "loaded hotp_dll: %s at " PFX "\n",
hotp_dll_path, VUL(vul_tab, vul).hotp_dll_base);
/* TODO: this one must be done asap; add the hot patch dll's text
* either to a new vm_area_vector_t or executable_vm_areas;
* check with the team first. case 5381
* add_executable_vm_area(hotp_dll_text_base, size_of_text,
* VM_UNMOD_IMAGE, false
_IF_DEBUG("hot patch dll loading"));
*/
/* TODO: assert that hotp_dll's dllmain is null to prevent control
* flow from going there during the thread creation due to
* nudge; but how?
*/
}
}
}
/* TODO: need a lot more LOG, ASSERT and SYSLOG statements */
/*----------------------------------------------------------------------------*/
/* TODO: for now just read from a flat file; change it in next phase to
* file/shmem depending upon what we decide; same goes for binary vs.
* text format; either way, the format of communication has to be defined
* so that nodemanager & core know what to write & read - key items
* include number of mode changes transmitted & the structure of each.
*
* Mode file format:
* <num_mode_update_entries>
* <policy_id-str>:<mode-decimal_integer>
* ...
* mode 0 - off, 1 - detect, 2 - protect;
*
* TODO: eventually, modes will be strings (good idea?, not binary; might be
* better to leave it as it is today.
*/
static void
hotp_read_policy_modes(hotp_policy_mode_t **old_modes)
{
/* TODO: for the next phase decide whether to use registry key or option
* string; for use a registry key.
*/
uint mode = 0, vul, num_mode_update_entries, i;
char *buf = NULL;
size_t buf_len = 0;
char *start = NULL;
/* Old modes are needed only by regular hotp for flushing patches;
* hotp_only shouldn't use them.
*/
ASSERT(!DYNAMO_OPTION(hotp_only) || old_modes == NULL);
if (old_modes != NULL) /* init to NULL because there are error exits */
*old_modes = NULL;
/* Can be called only during hotp_init() or during a nudge. */
ASSERT_OWN_WRITE_LOCK(true, &hotp_vul_table_lock);
/* This function shouldn't be called before policies are read.
* Sometimes, the node manager can nudge for a mode read without specifying
* policies first! This may happen during startup. Case 5448.
*/
if (GLOBAL_VUL_TABLE == NULL) {
LOG(GLOBAL, LOG_HOT_PATCHING, 1,
"Modes can't be set without "
"policy definitions. Probably caused due to a nudge by the node "
"manager to read modes when there were no policies.");
return;
}
buf = hotp_read_data_file(MODES_FILE, &buf_len);
if (buf == NULL) {
ASSERT(buf_len == 0);
return;
}
ASSERT(buf_len > 0);
/* Allocate space to save the old modes if they were requested for. */
if (old_modes != NULL) {
*old_modes = HEAP_ARRAY_ALLOC(GLOBAL_DCONTEXT, hotp_policy_mode_t,
NUM_GLOBAL_VULS, ACCT_HOT_PATCHING, PROTECTED);
ASSERT(*old_modes != NULL); /* make sure that space was allocated */
}
/* Turn off all vulnerabilities before reading modes. Only those for which
* a mode is specified should be on. Fix for case 5565. As the write lock
* is held, there is no danger of any lookup providing a no-match when there
* is one.
*/
for (vul = 0; vul < NUM_GLOBAL_VULS; vul++) {
if (old_modes != NULL)
(*old_modes)[vul] = GLOBAL_VUL(vul).mode;
/* Only hot patch types can be turned off by mode files. Other types
* like gbop hooks can't be.
*/
if (TESTALL(HOTP_TYPE_HOT_PATCH, GLOBAL_VUL(vul).type))
GLOBAL_VUL(vul).mode = HOTP_MODE_OFF;
}
start = buf;
SET_NUM(num_mode_update_entries, uint, NUM_VULNERABILITIES, start);
/* TODO: what if num_mode_update_entries is more than the entries in the
* file?
*/
for (i = 0; i < num_mode_update_entries; i++) {
bool matched = false;
char *split, *policy_id;
SET_STR_PTR(policy_id, start);
split = strchr(policy_id, ':');
if (split == NULL)
goto error_reading_policy;
*split++ = '\0'; /* TODO: during file mapping, this won't work */
SET_NUM(mode, uint, MODE, split);
/* Must set mode for all vulnerabilities with a matching policy_id, not
* just the first one.
*/
for (vul = 0; vul < NUM_GLOBAL_VULS; vul++) {
if (strncmp(GLOBAL_VUL(vul).policy_id, policy_id, MAX_POLICY_ID_LENGTH) ==
0) {
GLOBAL_VUL(vul).mode = mode;
matched = true;
}
}
/* If during mode update policy_id from a mode file doesn't have
* a corresponding vul_t, log a warning. When the node manager is
* starting up, modes file can be inconsistent, so this may happen
* (cases 5500 & 5526). However this could be a bug somewhere in the
* pipe line (EV, nm, policy package, etc) too.
*/
if (!matched) {
SYSLOG_INTERNAL_WARNING("While reading modes, found a mode "
"definition for a policy (%s) that didn't exist",
policy_id);
}
}
/* TODO: make the macro take this as an argument or find a neutral name */
error_reading_policy:
ASSERT(buf != NULL);
heap_free(GLOBAL_DCONTEXT, buf, buf_len HEAPACCT(ACCT_HOT_PATCHING));
return;
}
static void
hotp_set_policy_status(const uint vul_index, const hotp_inject_status_t status)
{
uint crc_buf_size;
ASSERT_OWN_WRITE_LOCK(true, &hotp_vul_table_lock);
ASSERT(hotp_policy_status_table != NULL);
ASSERT(status == HOTP_INJECT_NO_MATCH || status == HOTP_INJECT_PENDING ||
status == HOTP_INJECT_IN_PROGRESS || status == HOTP_INJECT_DETECT ||
status == HOTP_INJECT_PROTECT || status == HOTP_INJECT_ERROR);
/* Given that no other thread, app or nudge, will change this without the
* hot patch lock, this can be done without an atomic write.
*/
ASSERT(GLOBAL_VUL(vul_index).info->inject_status != NULL);
*(GLOBAL_VUL(vul_index).info->inject_status) = status;
/* Compute CRC after this status update and put it in the
* policy status table so that the node manager is protected from
* reading invalid status due to policy status table being reset/
* relllocated due to hotp_init or nudge or detach taking place.
*
* Note: The CRC write to the table doesn't need to be atomic too. Also,
* the CRC value is for all bytes of the policy status table except
* the CRC itself. Otherwise we would have to do the CRC computation
* twice; wastefully expensive.
*/
crc_buf_size = hotp_policy_status_table->size - sizeof(hotp_policy_status_table->crc);
hotp_policy_status_table->crc =
d_r_crc32((char *)&hotp_policy_status_table->size, crc_buf_size);
}
/* The status of hot patches is directly read by the node manager from the
* memory address specified in the drmarker; no nudge is needed. While the
* table is being created, the drmarker pointer will be null and set only
* after the table is fully initialized. Also, updates to the table entries
* are made with the hot patch lock, as with creation. The only way the node
* manager can get invalid data is after it reads drmarker, this routine
* releases the old policy status table before the node manager can read it.
* That is guarded by the table CRC, which is likely to be wrong. If drmarker
* points to memory released to the os or NULL, node manager will get a memory
* read error and it should be able to reattempt within which the new table will
* be ready.
*
* Format of policy status table in memory:
* <CRC32-uint> - CRC of size_in_bytes - sizeof(CRC32, i.e, uint).
* <size_in_bytes-uint>
* <num_policy_entries-uint>
* <hotp_policy_status_t>*
*/
static void
hotp_init_policy_status_table(void)
{
uint i, num_policies = NUM_GLOBAL_VULS, size_in_bytes, crc_buf_size;
hotp_policy_status_table_t *temp;
/* Can be called only during hotp_init() or during a nudge. */
ASSERT_OWN_WRITE_LOCK(true, &hotp_vul_table_lock);
ASSERT(!DATASEC_PROTECTED(DATASEC_RARELY_PROT));
/* This function shouldn't be called before policies and/or modes are read.
* Sometimes, the node manager can nudge for a mode read without specifying
* policies first! This may happen during startup. Case 5448.
*/
if (GLOBAL_VUL_TABLE == NULL) {
LOG(GLOBAL, LOG_HOT_PATCHING, 1,
"Policy status table can't be created "
"without policy definitions. Probably caused due to a nudge by the "
"node manager to read modes when there were no policies. Or "
"because all probes registered using the probe api were invalid.");
return;
}
/* This function is called each time a new policies and/or modes are read
* in. Each such time all existing injected hot patches are removed, so
* the policy status table associated with the old global vulnerability
* table must be released or resized to fit only the new set of
* hot patches turned on. The former is simpler to do.
*
* Note: if the optimization of flushing only those policies that
* have changed is implemented, which is not the case today, then just
* releasing policy status table will result in incorrect inject status.
* It should be released after the new table is created and filled with
* old values.
*/
if (hotp_policy_status_table != NULL) {
temp = hotp_policy_status_table;
hotp_policy_status_table = NULL;
/* If dr_marker_t isn't initialized, this won't be set. In that case,
* the dr_marker_t initialization code will set up the policy status table.
* This can happen at init time because hotp_init() is called before
* callback_interception_init().
*/
set_drmarker_hotp_policy_status_table(NULL);
heap_free(GLOBAL_DCONTEXT, temp, temp->size HEAPACCT(ACCT_HOT_PATCHING));
}
/* Right now, the status table contains as many elements as
* vulnerabilities. The original idea was to have only policies which
* are turned on in the table. This caused failures in the core
* because we need to maintain status internally for vulnerabilities
* that are turned off too. Case 5326.
*/
size_in_bytes =
sizeof(hotp_policy_status_table_t) + sizeof(hotp_policy_status_t) * num_policies;
temp = heap_alloc(GLOBAL_DCONTEXT, size_in_bytes HEAPACCT(ACCT_HOT_PATCHING));
temp->size = size_in_bytes;
temp->policy_status_array =
(hotp_policy_status_t *)((char *)temp + sizeof(hotp_policy_status_table_t));
/* Init status buffer elements & set up global vul table pointers */
/* TODO: two vulnerabilities can belong to the same policy; need to check
* for that and avoid duplication in the table; not needed now
* because we don't have such policies yet.
*/
for (i = 0; i < NUM_GLOBAL_VULS; i++) {
strncpy(temp->policy_status_array[i].policy_id, GLOBAL_VUL(i).policy_id,
MAX_POLICY_ID_LENGTH);
NULL_TERMINATE_BUFFER(temp->policy_status_array[i].policy_id);
temp->policy_status_array[i].inject_status = HOTP_INJECT_NO_MATCH;
/* Fix for case 5484, where the node manager wasn't able to tell if an
* inject status was for a policy that was turned on or off.
*/
temp->policy_status_array[i].mode = GLOBAL_VUL(i).mode;
/* The inject status in global vulnerability table should point
* to the corresponding element in this table.
*/
GLOBAL_VUL(i).info->inject_status = &temp->policy_status_array[i].inject_status;
}
temp->num_policies = i;
/* Set the table CRC now that the table has been initialized. */
crc_buf_size = temp->size - sizeof(temp->crc);
temp->crc = d_r_crc32((char *)&temp->size, crc_buf_size);
/* Make the policy status table live. If the dr_marker_t isn't initialized
* this won't be set. In that case, the dr_marker_t initialization code will
* set up the policy status table; happens during startup/initialization.
*/
hotp_policy_status_table = temp;
set_drmarker_hotp_policy_status_table((void *)temp);
}
/* Frees all the dynamically allocated members of vul (strings, info, sets,
* modules and patch points). NOTE: It doesn't free the vul itself.
*/
static void
hotp_free_one_vul(hotp_vul_t *vul)
{
uint set_idx, module_idx, ppoint_idx;
/* If this routine is called with a NULL for argument then there is a bug
* somewhere.*/
ASSERT(vul != NULL);
if (vul == NULL)
return;
if (vul->vul_id != NULL)
dr_strfree(vul->vul_id HEAPACCT(ACCT_HOT_PATCHING));
if (vul->policy_id != NULL)
dr_strfree(vul->policy_id HEAPACCT(ACCT_HOT_PATCHING));
if (vul->hotp_dll != NULL)
dr_strfree(vul->hotp_dll HEAPACCT(ACCT_HOT_PATCHING));
if (vul->hotp_dll_hash != NULL)
dr_strfree(vul->hotp_dll_hash HEAPACCT(ACCT_HOT_PATCHING));
if (vul->info != NULL) {
HEAP_TYPE_FREE(GLOBAL_DCONTEXT, vul->info, hotp_vul_info_t, ACCT_HOT_PATCHING,
PROTECTED);
}
if (vul->sets == NULL)
return;
/* If a set's array isn't null, then the number of sets can't be zero. */
ASSERT(vul->num_sets > 0);
for (set_idx = 0; set_idx < vul->num_sets; set_idx++) {
hotp_set_t *set = &vul->sets[set_idx];
if (set->modules == NULL)
continue;
/* If a modules array isn't null, then the number of modules can't
* be zero.
*/
ASSERT(set->num_modules > 0);
for (module_idx = 0; module_idx < set->num_modules; module_idx++) {
hotp_module_t *module = &set->modules[module_idx];
if (module->sig.pe_name != NULL) {
dr_strfree(module->sig.pe_name HEAPACCT(ACCT_HOT_PATCHING));
}
if (module->hashes != NULL) {
ASSERT(module->num_patch_point_hashes > 0);
HEAP_ARRAY_FREE(GLOBAL_DCONTEXT, module->hashes, hotp_patch_point_hash_t,
module->num_patch_point_hashes, ACCT_HOT_PATCHING,
PROTECTED);
}
if (module->patch_points != NULL) {
ASSERT(module->num_patch_points > 0);
for (ppoint_idx = 0; ppoint_idx < module->num_patch_points;
ppoint_idx++) {
hotp_patch_point_t *ppoint = &module->patch_points[ppoint_idx];
if (ppoint->trampoline != NULL) {
ASSERT(DYNAMO_OPTION(hotp_only));
ASSERT(ppoint->app_code_copy != NULL);
special_heap_free(hotp_only_tramp_heap,
(void *)ppoint->trampoline);
}
}
HEAP_ARRAY_FREE(GLOBAL_DCONTEXT, module->patch_points, hotp_patch_point_t,
module->num_patch_points, ACCT_HOT_PATCHING, PROTECTED);
}
}
HEAP_ARRAY_FREE(GLOBAL_DCONTEXT, set->modules, hotp_module_t, set->num_modules,
ACCT_HOT_PATCHING, PROTECTED);
}
HEAP_ARRAY_FREE(GLOBAL_DCONTEXT, vul->sets, hotp_set_t, vul->num_sets,
ACCT_HOT_PATCHING, PROTECTED);
}
/* Release all memory used by the hot patch vulnerability table, tab.
* num_vuls_alloc is number of vulnerability defs. the table has space for.
* The table may not always contain num_vuls_alloc policy defs. Where there
* is an error during policy defs file parsing they can be fewer in number with
* the last one (one where the error happened) being partial. Cases 8272, 9045.
*/
static void
hotp_free_vul_table(hotp_vul_t *tab, uint num_vuls_alloc)
{
uint vul_idx;
if (tab == NULL) {
ASSERT(num_vuls_alloc == 0);
return;
}
/* If the table isn't NULL, the number of vulnerabilities can't be zero. */
ASSERT(num_vuls_alloc > 0);
for (vul_idx = 0; vul_idx < num_vuls_alloc; vul_idx++) {
hotp_free_one_vul(&tab[vul_idx]);
}
HEAP_ARRAY_FREE(GLOBAL_DCONTEXT, tab, hotp_vul_t, num_vuls_alloc, ACCT_HOT_PATCHING,
PROTECTED);
}
/* This routine flushes all fragments in fcache that have been injected with a
* hot patch, i.e., restoring an app text to its pre-hot-patch state.
*
* Note: hot patch removal is not optimized, i.e., changes to existing
* policy definitions, modes or actual injection status aren't used to limit
* flushing. Not a performance issue for now.
* TODO: flush only those vulnerabilities that have actually changed, not
* every thing that is active or has been injected.
* TODO: make this use loaded_module_areas & get rid off the 4-level nested
* loops.
*/
static void
hotp_remove_patches_from_module(const hotp_vul_t *vul_tab, const uint num_vuls,
const bool hotp_only, const app_pc mod_base,
const hotp_policy_mode_t *old_modes)
{
uint vul_idx, set_idx, module_idx, ppoint_idx;
hotp_module_t *module;
hotp_patch_point_t *ppoint;
dcontext_t *dcontext = get_thread_private_dcontext();
ASSERT(dcontext != NULL && dcontext != GLOBAL_DCONTEXT);
/* For hotp_only patch removal, we should be running in hotp_only mode. */
ASSERT(!hotp_only || DYNAMO_OPTION(hotp_only));
/* Old vulnerability modes shouldn't be used with hotp_only. */
ASSERT(!DYNAMO_OPTION(hotp_only) || old_modes == NULL);
/* Alternate modes shouldn't be used during module specific removal also. */
ASSERT(mod_base == NULL || old_modes == NULL);
/* Though trying to flush a NULL vul table is a bug, this can happen
* because the node manager can nudge the core to read modes when it hasn't
* provided the policies! See case 5448. Hence just a warning & no assert.
*/
if (vul_tab == NULL) {
LOG(GLOBAL, LOG_HOT_PATCHING, 1,
"Hot patch flushing has been invoked "
"with a NULL table");
return;
}
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "flushing as a result of nudge\n");
for (vul_idx = 0; vul_idx < num_vuls; vul_idx++) {
bool set_processed = false;
const hotp_vul_t *vul = &VUL(vul_tab, vul_idx);
/* Nothing to remove or flush if the mode is off, i.e., nothing would
* have been injected.
* Note: Both vul's current mode & its old mode should be off to skip
* removal; even if one is not, then that vulnerability's patches need
* to be removed. In otherwords, if patch was previously on (injected)
* or is now on (to be injected), corresponding bbs must be flushed;
* this is for regular hotp, not for hotp_only which has no flushing.
*/
if (vul->mode == HOTP_MODE_OFF) {
if (old_modes == NULL) {
/* If there is no old_mode, skip right here. */
continue;
} else if (old_modes[vul_idx] == HOTP_MODE_OFF) {
/* If old_mode exists, that must be off too in order to skip. */
continue;
}
}
ASSERT(vul->mode == HOTP_MODE_DETECT || vul->mode == HOTP_MODE_PROTECT ||
(old_modes != NULL &&
(old_modes[vul_idx] == HOTP_MODE_DETECT ||
old_modes[vul_idx] == HOTP_MODE_PROTECT)));
for (set_idx = 0; set_idx < VUL(vul_tab, vul_idx).num_sets; set_idx++) {
/* Only the first matching set should be used; case 10248. */
if (set_processed)
break;
for (module_idx = 0; module_idx < SET(vul_tab, vul_idx, set_idx).num_modules;
module_idx++) {
module = &MODULE(vul_tab, vul_idx, set_idx, module_idx);
if (module->matched) {
/* If a specific module is mentioned remove patches from
* just that.
*/
if (mod_base != NULL && mod_base != module->base_address)
continue;
set_processed = true;
/* Otherwise, flush all patch points in any module that
* matches. Nothing to flush in unmatched modules.
*/
for (ppoint_idx = 0; ppoint_idx < module->num_patch_points;
ppoint_idx++) {
ppoint = &module->patch_points[ppoint_idx];
if (hotp_only) {
/* For a hotp_only patch, we can only remove that
* which has been injected, unlike the hotp mode
* where we might just be flushing out uninjected
* fragments or don't know which particular patch
* point has been injected (in hotp_only mode all
* of them should be injected if one is injected).
*/
if (ppoint->trampoline != NULL)
hotp_only_remove_patch(dcontext, module, ppoint);
else {
/* If module is matched and mode is on, then
* hotp_only patch targeting the current
* ppoint must be injected unless it has
* been removed to handle loader-safety issues.
*/
ASSERT((ppoint->trampoline == NULL ||
hotp_globals->ldr_safe_hook_removal) &&
"hotp_only - double patch removal");
}
} else {
app_pc flush_addr = hotp_ppoint_addr(module, ppoint);
ASSERT_OWN_NO_LOCKS();
LOG(GLOBAL, LOG_HOT_PATCHING, 4,
"flushing " PFX " due to a nudge\n", flush_addr);
flush_fragments_in_region_start(
dcontext, flush_addr, 1, false /* no lock */,
false /* keep futures */, false /*exec still valid*/,
false /*don't force synchall*/
_IF_DGCDIAG(NULL));
flush_fragments_in_region_finish(dcontext, false);
/* TODO: ASSERT (flushed fragments have really been)
* flushed but how, using a vm_areas_overlap()
* or fragment_lookup() check?
*/
}
}
}
}
}
}
}
/* TODO: make this use hotp_patch_point_areas & get rid off the 4-level nested
* loops which is used in hotp_remove_patches_from_module.
*/
static void
hotp_remove_hot_patches(const hotp_vul_t *vul_tab, const uint num_vuls,
const bool hotp_only, const hotp_policy_mode_t *old_modes)
{
/* Old vulnerability modes shouldn't be used with hotp_only. */
ASSERT(!DYNAMO_OPTION(hotp_only) || old_modes == NULL);
hotp_remove_patches_from_module(vul_tab, num_vuls, hotp_only, NULL, old_modes);
}
/* TODO: vlad wanted the ability to ignore some attributes during checking;
* this is not for constraints, but if he wants an ad-hoc patch to fix
* something other than a vulnerability, say, broken code that is not
* a vulnerability; for hot patches/constraints all attributes must be
* checked, no ignoring stuff.
*/
static bool
hotp_module_match(const hotp_module_t *module, const app_pc base, const uint checksum,
const uint timestamp, const size_t image_size, const size_t code_size,
const uint64 file_version, const char *name, hotp_type_t type)
{
uint hash_index, computed_hash;
hotp_patch_point_hash_t *hash;
bool matched;
ASSERT(module != NULL && base != NULL);
ASSERT(TESTANY(HOTP_TYPE_HOT_PATCH | HOTP_TYPE_GBOP_HOOK | HOTP_TYPE_PROBE, type));
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "Matching module base " PFX " %s\n", base, name);
/* For library offset or export function based patch points, the probe will
* define a library by name (if needed we expand it to include the
* liveshield type matching, but the client can do it outside) */
/* gbop type patches provide a symbolic name to hook, so there is nothing
* to match it with other than the pe name.
*/
if (TESTALL(HOTP_TYPE_PROBE, type) IF_GBOP(|| (TESTALL(HOTP_TYPE_GBOP_HOOK, type)))) {
ASSERT(module->sig.pe_checksum == 0 && module->sig.pe_timestamp == 0 &&
module->sig.pe_image_size == 0 && module->sig.pe_code_size == 0 &&
module->sig.pe_file_version == 0 && module->num_patch_points == 1 &&
module->patch_points != NULL && module->num_patch_point_hashes == 0 &&
module->hashes == NULL);
if (name == NULL) {
/* if the only check is the module name, then a NULL name means
* the module wasn't matched; otherwise this check would be bogus.
*/
return false;
} else if (IF_UNIX_ELSE(strncmp, strncasecmp)(module->sig.pe_name, name,
MAXIMUM_PATH) == 0) {
/* XXX: strcmp() is faster than the ignore case version,
* but we shouldn't rely on the PE name case to be the
* same in all versions of Windows.
*/
return true;
} else {
return false;
}
}
/* These checks are for hot patch types, i.e., ones that have offset rvas
* specified for each known version.
* First stage check: PE timestamp, PE checksum, PE code_size, PE file
* version & PE name, i.e., signature match.
*
* XXX: Today error handling of PE parsing is not done by the core, so
* unavailability of an attribute isn't recorded. Thus IGNORE and
* UNAVAILABLE are treated the same for module matching. When the core can
* handle it the UNAVAILABLE part should be removed from the checks, and
* checks for unavailability should be done. Case 9215 tracks the core not
* handling PE parsing for malformed files and their impact on hot patching.
*/
ASSERT(TESTALL(HOTP_TYPE_HOT_PATCH, type));
matched = module->sig.pe_timestamp == timestamp ||
module->sig.pe_timestamp == PE_TIMESTAMP_IGNORE ||
module->sig.pe_timestamp == PE_TIMESTAMP_UNAVAILABLE;
matched = matched &&
(module->sig.pe_checksum == checksum ||
module->sig.pe_checksum == PE_CHECKSUM_IGNORE ||
module->sig.pe_checksum == PE_CHECKSUM_UNAVAILABLE);
matched = matched &&
(module->sig.pe_image_size == image_size ||
module->sig.pe_image_size == PE_IMAGE_SIZE_IGNORE ||
module->sig.pe_image_size == PE_IMAGE_SIZE_UNAVAILABLE);
matched = matched &&
(module->sig.pe_code_size == code_size ||
module->sig.pe_code_size == PE_CODE_SIZE_IGNORE ||
module->sig.pe_code_size == PE_CODE_SIZE_UNAVAILABLE);
matched = matched &&
(module->sig.pe_file_version == file_version ||
module->sig.pe_file_version == PE_FILE_VERSION_IGNORE ||
module->sig.pe_file_version == PE_FILE_VERSION_UNAVAILABLE);
matched = matched &&
((strncmp(module->sig.pe_name, PE_NAME_IGNORE, sizeof(PE_NAME_IGNORE)) == 0) ||
(name == NULL && /* no name case */
module->sig.pe_name[0] == PE_NAME_UNAVAILABLE) ||
(name != NULL && (strncmp(module->sig.pe_name, name, MAXIMUM_PATH) == 0)));
if (matched) {
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "Module signature check passed\n");
/* First stage check was true, now let us do the second stage check,
* i.e., check the hashes of patch points in the module.
*/
ASSERT(module->num_patch_point_hashes > 0 && module->hashes != NULL);
for (hash_index = 0; hash_index < module->num_patch_point_hashes; hash_index++) {
hash = &module->hashes[hash_index];
computed_hash = hotp_compute_hash(base, hash);
if (computed_hash != hash->hash_value)
return false;
}
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "Patch point hash check passed\n");
return true;
}
return false;
}
/* Used to compute the hash of a patch point hash region. In hotp_only mode,
* if there is an overlap between a hash region and a patch region, the
* image bytes, stored at the top of the trampoline, are used to create copy
* of the image on which d_r_crc32 is computed. In regular hotp mode, d_r_crc32 is
* called directly.
*/
static uint
hotp_compute_hash(app_pc base, hotp_patch_point_hash_t *hash)
{
uint crc, copy_size;
char *hash_buf, *copy;
app_pc hash_start, hash_end, vm_start, vm_end, src, dst, trampoline;
vmvector_iterator_t iterator;
hotp_offset_match_t *ppoint_desc;
ASSERT(base != NULL && hash != NULL);
ASSERT(hash->start > 0 && hash->len > 0);
hash_start = base + hash->start;
hash_end = hash_start + hash->len;
/* If the hash region overlaps with any patch point region, then use the
* original image bytes to compute the d_r_crc32. Valid for hotp_only because
* in hotp mode, i.e., with a code cache, we don't modify the original code.
*/
if (DYNAMO_OPTION(hotp_only) &&
vmvector_overlap(hotp_patch_point_areas, hash_start, hash_end)) {
/* Make sure that the patch region size for hotp_only is correct. */
ASSERT(HOTP_PATCH_REGION_SIZE == HOTP_ONLY_PATCH_REGION_SIZE);
/* Allocate a buffer & copy the image bytes represented by the hash.
* This will include bytes modified by a prior hotp_only patch.
* Note: an extra 2 x HOTP_PATCH_REGION_SIZE is allocated to be used
* as overflow buffers at the front & back of the copy; makes handling
* the overlap scenarios (4 different ones) easy.
*/
copy_size = hash->len + (2 * HOTP_PATCH_REGION_SIZE);
copy = HEAP_ARRAY_ALLOC(GLOBAL_DCONTEXT, char, copy_size, ACCT_HOT_PATCHING,
PROTECTED);
hash_buf = copy + HOTP_PATCH_REGION_SIZE;
memcpy(hash_buf, hash_start, hash->len);
/* Now, for each vmarea that overlaps, copy the original image bytes
* into the buffer.
* XXX: we do a linear walk as opposed to walking over only those
* regions that overlap, ineffcient; see case 8211 about a new
* vmvector iterator that walks over only overlapping regions.
*/
vmvector_iterator_start(hotp_patch_point_areas, &iterator);
while (vmvector_iterator_hasnext(&iterator)) {
ppoint_desc = vmvector_iterator_next(&iterator, &vm_start, &vm_end);
trampoline =
GLOBAL_PPOINT(ppoint_desc->vul_index, ppoint_desc->set_index,
ppoint_desc->module_index, ppoint_desc->ppoint_index)
.trampoline;
/* If the patch isn't injected, overlap doesn't matter because
* the image hasn't been changed. Overlap with an uninjected patch
* region can only happen when loader safety is in progress during
* which a patch point is removed (only from the image, not
* hotp_patch_point_areas) and it is re-injected; the re-injection
* of the patch point will overlap with itself. See case 8222.
*/
if (trampoline == NULL) {
/* If hash belongs ppoint_desc, i.e., overlaps with self, then
* base and module's base must match.
*/
ASSERT(base ==
GLOBAL_MODULE(ppoint_desc->vul_index, ppoint_desc->set_index,
ppoint_desc->module_index)
.base_address);
continue;
}
/* If the trampoline exists, it better be a valid one, i.e., the
* patch corresponding to this vmarea must be injected.
*/
ASSERT(vmvector_overlap(hotp_only_tramp_areas, trampoline,
trampoline + HOTP_ONLY_TRAMPOLINE_SIZE));
/* The size of each vmarea in hotp_patch_point_areas must be
* equal to that of the patch region.
*/
ASSERT(vm_end - vm_start == HOTP_PATCH_REGION_SIZE);
/* The module corresponding to this vm area (patch point) should
* have been matched by a vul. def. (in hotp_process_image).
*/
ASSERT(GLOBAL_MODULE(ppoint_desc->vul_index, ppoint_desc->set_index,
ppoint_desc->module_index)
.matched);
/* There are a few scenarios for a hash & patch point to overlap,
* vmarea fully within the hash area, vice versa, partial below,
* partial above, and exact on either side or both
* Using an extra buffer the size of a patch region at the front
* and back allows all the scenarios to be handled with a single
* equation - eliminates messy code; worth allocating 10 bytes
* extra.
* Note: the extra buffer can be 1 byte shorter on either side, but
* leaving it at patch point region size, just to be safe.
*/
if (vm_start < hash_end && vm_end > hash_start) {
src = trampoline;
dst = (app_pc)hash_buf + (vm_start - hash_start);
/* Just make sure that we don't trash anything when copying the
* original image over the bytes in hash_buf.
*/
ASSERT((dst >= (app_pc)copy) &&
((dst + HOTP_PATCH_REGION_SIZE) <= ((app_pc)copy + copy_size)));
/* If the hash overlaps with a patch point region, then the
* current image & the copy should be different, i.e., a patch
* must exist at that point.
*/
ASSERT(memcmp(dst, src, HOTP_PATCH_REGION_SIZE) != 0);
/* CAUTION: this memcpy assumes the location & size of
* app code copy in the trampoline, i.e., the first 5 bytes of
* trampoline contain the original app code; so any changes
* should be kept in sync.
*/
memcpy(dst, src, HOTP_PATCH_REGION_SIZE);
}
/* XXX: if the iterator guaranteed order, we can break out after
* the first non-match - optimization.
*/
}
vmvector_iterator_stop(&iterator);
crc = d_r_crc32(hash_buf, hash->len);
HEAP_ARRAY_FREE(GLOBAL_DCONTEXT, copy, char, copy_size, ACCT_HOT_PATCHING,
PROTECTED);
} else {
/* No overlap; image is unmodified, so image's d_r_crc32 should be valid. */
crc = d_r_crc32((char *)hash_start, hash->len);
}
return crc;
}
/* TODO: this function should be used for startup & nudge dll list walking,
* dll loading and unloading.
* TODO: assert somehow that every time this function is invoked there must
* be a flush preceding or succeeding it, except at startup
* TODO: os specific routine; move to win32/
* TODO: this function is called when vm_areas_init() is invoked, but
* hotp_init() is called after vm_areas_init()! bogus - check
* other start up scenarios like retakeover to see if policy reading
* & activation get out of order; this is the same issue that vlad
* pointed out: make sure that process_image() is called after
* hotp_init()
* TODO: process_{image,mmap}() should never be called on hot patch dlls
* because dr is loading them; assert for this somewhere to prevent
* assumption violation bugs.
*/
static void
hotp_process_image_helper(const app_pc base, const bool loaded,
const bool own_hot_patch_lock, const bool just_check,
bool *needs_processing, const thread_record_t **thread_table,
const int num_threads, const bool ldr_safety,
vm_area_vector_t *toflush);
void
hotp_process_image(const app_pc base, const bool loaded, const bool own_hot_patch_lock,
const bool just_check, bool *needs_processing,
const thread_record_t **thread_table, const int num_threads)
{
hotp_process_image_helper(base, loaded, own_hot_patch_lock, just_check,
needs_processing, thread_table, num_threads, false, NULL);
}
/* Helper routine for seeing if point is in hotp_ppoint_vec */
bool
hotp_ppoint_on_list(app_rva_t ppoint, app_rva_t *hotp_ppoint_vec,
uint hotp_ppoint_vec_num)
{
bool on_list = false;
uint i;
/* We assume there are at most a handful of these so we don't sort.
* If we add GBOP hooks we may want to do that. */
# ifdef GBOP
ASSERT(DYNAMO_OPTION(gbop) == 0);
# endif
ASSERT(ppoint != 0);
ASSERT(hotp_ppoint_vec != NULL && hotp_ppoint_vec_num > 0);
if (hotp_ppoint_vec == NULL)
return false;
for (i = 0; i < hotp_ppoint_vec_num; i++) {
if (ppoint == hotp_ppoint_vec[i]) {
on_list = true;
break;
}
}
return on_list;
}
/* Returns true if there is a persistent cache in [base,base+image_size) that
* may contain code for any of the patch points of module
*/
static bool
hotp_perscache_overlap(uint vul, uint set, uint module, app_pc base, size_t image_size)
{
vmvector_iterator_t vmvi;
coarse_info_t *info;
uint pp;
bool flush_perscache = false;
ASSERT(DYNAMO_OPTION(use_persisted_hotp));
ASSERT(!DYNAMO_OPTION(hotp_only));
vm_area_coarse_iter_start(&vmvi, base);
/* We have a lot of nested linear walks here, esp. when called from
* hotp_process_image_helper inside nested loops, but typically the coarse
* iterator involves one binary search and only one match, and
* hotp_ppoint_on_list and the pp for loop here only a few entries each;
* so this routine shouldn't be a perf bottleneck by itself.
*/
while (!flush_perscache && vm_area_coarse_iter_hasnext(&vmvi, base + image_size)) {
info = vm_area_coarse_iter_next(&vmvi, base + image_size);
ASSERT(info != NULL);
if (info == NULL) /* be paranoid */
continue;
if (info->hotp_ppoint_vec == NULL)
flush_perscache = true;
else {
ASSERT(info->persisted);
for (pp = 0; pp < GLOBAL_MODULE(vul, set, module).num_patch_points; pp++) {
if (!hotp_ppoint_on_list(GLOBAL_PPOINT(vul, set, module, pp).offset,
info->hotp_ppoint_vec,
info->hotp_ppoint_vec_num)) {
flush_perscache = true;
break;
}
}
}
/* Should be able to ignore 2ndary unit */
ASSERT(info->non_frozen == NULL || info->non_frozen->hotp_ppoint_vec == NULL);
}
vm_area_coarse_iter_stop(&vmvi);
return flush_perscache;
}
/* This helper exists mainly to handle the loader safety case for adding
* ppoint areas. vm_areas should be added to ppoint_areas only during module
* load/unload (including the initial stack walk) and during policy read
* nudge, not during a reinjection during loader safety.
* The same holds good for removal, but today that isn't an
* issue because loader safety uses hotp_remove_patches_from_module() to do
* it, which doesn't modify ppoint areas.
* XXX: once hotp_inject_patches_into_module() is implemented
* based on loaded_module_areas and used in hotp_only_mem_prot_change() instead
* of hotp_process_image_helper, this can go.
*/
static void
hotp_process_image_helper(const app_pc base, const bool loaded,
const bool own_hot_patch_lock, const bool just_check,
bool *needs_processing, const thread_record_t **thread_table,
const int num_threads_arg, const bool ldr_safety,
vm_area_vector_t *toflush)
{
uint vul_idx, set_idx, module_idx, ppoint_idx;
hotp_module_t *module;
dcontext_t *dcontext = get_thread_private_dcontext();
uint checksum, timestamp;
size_t image_size = 0, code_size;
uint64 file_version;
module_names_t *names = NULL;
const char *name = NULL, *pe_name = NULL, *mod_name = NULL;
int num_threads = num_threads_arg;
bool any_matched = false;
bool flush_perscache = false;
bool perscache_range_overlap = false;
ASSERT(base != NULL); /* Is it a valid dll in loaded memory? */
LOG(GLOBAL, LOG_HOT_PATCHING, 2, "hotp_process_image " PFX " %s w/ %d vuls\n", base,
loaded ? "load" : "unload", NUM_GLOBAL_VULS);
ASSERT(dcontext != GLOBAL_DCONTEXT);
/* note that during startup processing due to
* find_executable_vm_areas() dcontext can in fact be NULL
*/
if (dcontext != NULL && dcontext->nudge_thread) /* Fix for case 5367. */
return;
# ifdef WINDOWS
if (num_threads == 0 && !just_check && DYNAMO_OPTION(hotp_only)) {
/* XXX PR 225578: dr_register_probes passes 0 for the thread count
* b/c post-init probes are NYI: but to enable at-your-own risk probes
* relaxing the assert
*/
ASSERT_CURIOSITY_ONCE(!dynamo_initialized &&
"post-init probes at your own risk: PR 225578!");
num_threads = HOTP_ONLY_NUM_THREADS_AT_INIT;
/* For hotp_only, all threads should be suspended before patch injection.
* However, at this point in startup, callback hooks aren't in place and
* we don't know if any other thread is running around that the core
* doesn't know about. This would be rare and with early injection, rarer.
* However, if that thread is executing in a region being patched we can
* fail spectacularly. Curiosity in the meanwhile.
* Also, to be on the safe side grab the synchronization locks.
*/
ASSERT_CURIOSITY(check_sole_thread());
ASSERT(!own_hot_patch_lock); /* can't get hotp lock before sync locks */
d_r_mutex_lock(&all_threads_synch_lock);
d_r_mutex_lock(&thread_initexit_lock);
}
# endif
if (!own_hot_patch_lock)
d_r_write_lock(&hotp_vul_table_lock);
ASSERT_OWN_READWRITE_LOCK(true, &hotp_vul_table_lock);
/* Caller doesn't want to process the image, but to know if it matches. */
if (just_check) {
/* Only hotp_only needs this; not regular hot patching. */
ASSERT(DYNAMO_OPTION(hotp_only));
ASSERT(needs_processing != NULL);
*needs_processing = false; /* will be set it to true, if needed */
} else
ASSERT(needs_processing == NULL);
/* get module information from PE once (case 7990) */
/* XXX: once all pe information is available in loaded_module_areas, use
* that here
* XXX: file_version is obtained by walking the resouce section which is expensive;
* the same is true for code_size to some extent, i.e., expensive but not
* that much. So we may be better off by computing them in separate routines
* predicated by the first check - and put all these into hotp_get_module_sig()
*/
os_get_module_info_lock();
if (!os_get_module_info_all_names(base, &checksum, &timestamp, &image_size, &names,
&code_size, &file_version)) {
/* XXX: case 9778 - module info is now obtained from
* loaded_module_areas vector, which doesn't seem to have hotp dll
* info, so we hit this. As a first step this is converted to a log
* to make tests work; will have to read it from pe directly (using
* try/except) if it isn't a hotp dll - if that doesn't work then be
* curious. Also, need to find out if it was triggered only for hotp
* dlls. */
LOG(GLOBAL, LOG_HOT_PATCHING, 2, "unreadable PE base (" PFX ")?\n", base);
os_get_module_info_unlock();
goto hotp_process_image_exit;
} else {
/* Make our own copy of both the pe name and the general module name.
* This is because pe name can be null for executables, which is fine
* for liveshields, but not for gbop or probe api - they just specify a
* module name, so we have to use any name that is available. Note: as
* of today, gbop hasn't been done on executables, which is why it
* worked - it is broken for hooks in exes - a XXX, but gbop is going
* away anyway. */
pe_name = dr_strdup(names->module_name HEAPACCT(ACCT_HOT_PATCHING));
mod_name = dr_strdup(GET_MODULE_NAME(names) HEAPACCT(ACCT_HOT_PATCHING));
os_get_module_info_unlock();
/* These values can't be read in from a module, they are used by the
* patch writer to hint to the core to ignore the corresponding checks.
*/
ASSERT_CURIOSITY(timestamp != PE_TIMESTAMP_IGNORE &&
checksum != PE_CHECKSUM_IGNORE &&
image_size != PE_IMAGE_SIZE_IGNORE);
}
# ifdef WINDOWS
DOCHECK(1, {
if (TESTANY(ASLR_DLL, DYNAMO_OPTION(aslr)) &&
TESTANY(ASLR_SHARED_CONTENTS, DYNAMO_OPTION(aslr_cache))) {
/* case 8507 - the timestamp and possibly checksum of the current mapping,
possibly ASLRed, may not be the same as the application DLL */
uint pe_timestamp;
uint pe_checksum;
bool ok = os_get_module_info(base, &pe_checksum, &pe_timestamp, NULL, NULL,
NULL, NULL);
ASSERT_CURIOSITY(timestamp != 0);
/* Note that if we don't find the DLL in the module list,
* we'll keep using the previously found checksum and
* timestamp. Although normally all DLLs are expected to be
* listed, currently that is done only with ASLR_TRACK_AREAS.
*/
/* case 5381: we don't ASSERT(ok) b/c hotpatch DLLs aren't listed in our
* own module areas, so we don't always find all modules */
/* with the current scheme the checksum is still the original DLLs checksum
* though it won't check, and the timestamp is bumped by one second
*/
ASSERT(!ok || pe_checksum == checksum);
ASSERT_CURIOSITY(!ok || pe_timestamp == timestamp ||
pe_timestamp == timestamp + 1);
}
});
# endif /* WINDOWS */
if (!DYNAMO_OPTION(hotp_only)) {
perscache_range_overlap =
executable_vm_area_persisted_overlap(base, base + image_size);
}
/* TODO: assert that 'base' is the module's base address,
* get_dll_short_name() expects this; will be used for sig check
* use the fn() that gets only what is in the PE
* XXX: eliminate this n^4 loop for each module {load,unload}; case 10683
*/
for (vul_idx = 0; vul_idx < NUM_GLOBAL_VULS; vul_idx++) {
bool set_matched = false;
if (TESTALL(HOTP_TYPE_PROBE, GLOBAL_VUL(vul_idx).type)
IF_GBOP(|| (TESTALL(HOTP_TYPE_GBOP_HOOK, GLOBAL_VUL(vul_idx).type)))) {
/* XXX PR 533522: state in the docs/comments which name is
* used where! pe_name vs mod_name
*/
name = mod_name;
} else {
ASSERT(TESTALL(HOTP_TYPE_HOT_PATCH, GLOBAL_VUL(vul_idx).type));
/* XXX PR 533522: state in the docs/comments which name is
* used where! pe_name vs mod_name
*/
name = pe_name;
}
for (set_idx = 0; set_idx < GLOBAL_VUL(vul_idx).num_sets; set_idx++) {
/* Case 10248 - multiple sets can match, but only the first such set
* should be used, the rest discarded. In the old model only one
* set matched, but it was changed to let the patch writer to
* progressively relax the matching criteria. */
if (set_matched)
break;
for (module_idx = 0; module_idx < GLOBAL_SET(vul_idx, set_idx).num_modules;
module_idx++) {
module = &GLOBAL_MODULE(vul_idx, set_idx, module_idx);
/* When unloading a matched dll in hotp_only mode, all injected
* patches must be removed before proceeding any further.
* Otherwise hotp_module_match() will fail in the id hash
* computation part due to a changed image, due to injection.
*/
if (base == module->base_address && !loaded) {
if (just_check) { /* caller doesn't want processing */
*needs_processing = true;
goto hotp_process_image_exit;
}
/* For hotp_only if a module matches all patch points
* in it must be removed in one shot; just as they are
* injected in one shot.
*/
if (GLOBAL_VUL(vul_idx).mode == HOTP_MODE_DETECT ||
GLOBAL_VUL(vul_idx).mode == HOTP_MODE_PROTECT) {
for (ppoint_idx = 0; ppoint_idx < module->num_patch_points;
ppoint_idx++) {
hotp_patch_point_t *ppoint;
ppoint = &module->patch_points[ppoint_idx];
if (DYNAMO_OPTION(hotp_only)) {
if (ppoint->trampoline != NULL) {
hotp_only_remove_patch(dcontext, module, ppoint);
} else {
/* If module is matched & mode is on, then the
* patch must be injected unless it has been
* removed to handle loader-safety issues.
*/
ASSERT(hotp_globals->ldr_safe_hook_removal &&
"hotp_only - double patch removal");
}
}
/* xref case 10736.
* For hotp_only, module load & inject, and
* similarly, module unload and remove are done
* together, so hot_patch_point_areas won't be out
* of synch. However, for hotp with fcache, a
* module unload can remove the patches from
* hotp_patch_point_areas before flushing them.
* This can prevent the flush from happening if
* hotp_patch_point_areas is used for it (which
* isn't done today; case 10728). It can also
* result in voiding a patch injection for a new bb
* in that module, i.e., module can be without a
* patch for a brief period till it is unloaded.
*/
hotp_ppoint_areas_remove(hotp_ppoint_addr(module, ppoint));
}
}
/* Once hotp_only patches are removed, the module must
* match at this point.
* TODO: multiple vulnerabilities targeting the same module
* & whose hashes overlap, won't be {inject,remove}d
* because the image gets modified with the injection
* of the first one and the hash check for the second
* one will fail.
*/
ASSERT_CURIOSITY(hotp_module_match(
module, base, checksum, timestamp, image_size, code_size,
file_version, name, GLOBAL_VUL(vul_idx).type));
}
/* XXX: there's no reason to compute whether an OFF patch
* matches; just wasted cycles, as we come back here on
* any path that later turns the patch on, and no external
* stats rely on knowing whether an off patch matches.
*/
if (hotp_module_match(module, base, checksum, timestamp, image_size,
code_size, file_version, name,
GLOBAL_VUL(vul_idx).type)) {
set_matched = true;
if (just_check) { /* caller doesn't want processing */
*needs_processing = true;
goto hotp_process_image_exit;
}
if (loaded) { /* loading dll */
bool patch_enabled =
(GLOBAL_VUL(vul_idx).mode == HOTP_MODE_DETECT ||
GLOBAL_VUL(vul_idx).mode == HOTP_MODE_PROTECT);
LOG(GLOBAL, LOG_HOT_PATCHING, 1,
"activating vulnerability %s while loading %s\n",
GLOBAL_VUL(vul_idx).vul_id, module->sig.pe_name);
any_matched = true;
/* Case 9970: See if we need to flush any
* perscaches in the region. Once we decide to flush
* we're going to flush everything. We avoid the later
* flush on a nudge in vm_area_allsynch_flush_fragments().
* We currently come here for OFF patches, so we explicitly
* check for that before flushing.
*/
if (patch_enabled && perscache_range_overlap &&
!flush_perscache && DYNAMO_OPTION(use_persisted_hotp)) {
flush_perscache = hotp_perscache_overlap(
vul_idx, set_idx, module_idx, base, image_size);
}
/* TODO: check if all modules in the current
* vulnerability are active; if so activate the
* policy
* also, add patch points to lookup structures
* only if entire vulnerability is active;
* needed to enforce atomicity of patch injection
*/
/* the base is used to find the runtime address of
* patch offset in the current lookup routine; till a
* offset lookup hash is constructed the base address
* is needed because the offset in the patchpoint
* structure is read only data that should be fixed to
* point to the runtime address. even then, the flush
* routine would need to know which offset, i.e.,
* runtime offset, to flush; so this base_address is
* needed or a runtime data field must be created.
*/
module->base_address = base;
module->matched = true;
hotp_set_policy_status(vul_idx, HOTP_INJECT_PENDING);
/* gbop type hooks don't have patch offsets defined,
* as they use function names; must set them otherwise
* patching will blow up.
*/
if (TESTALL(HOTP_TYPE_GBOP_HOOK, GLOBAL_VUL(vul_idx).type)) {
/* XXX: assert on all patch point fields being 0,
* except precedence.
* also, ASSERT on func_addr & func_name;
*/
app_pc func_addr;
app_rva_t offset;
/* gbop is only in -client mode, i.e., hotp_only */
ASSERT(DYNAMO_OPTION(hotp_only));
func_addr = (app_pc)d_r_get_proc_address(
(module_handle_t)base, GLOBAL_VUL(vul_idx).vul_id);
if (func_addr != NULL) { /* fix for case 7969 */
ASSERT(func_addr > base);
offset = func_addr - base;
module->patch_points[0].offset = offset;
} else {
/* Some windows versions won't have some gbop
* hook funcs or get_proc_address might just
* fail; either way just ignore such hooks.
* TODO: think about this - design issue.
*/
module->base_address = NULL;
module->matched = false;
continue;
}
}
/* For hotp_only if a module matches all patch points
* in it must be injected in one shot.
*/
if (patch_enabled) {
hotp_offset_match_t ppoint_desc;
ppoint_desc.vul_index = vul_idx;
ppoint_desc.set_index = set_idx;
ppoint_desc.module_index = module_idx;
for (ppoint_idx = 0; ppoint_idx < module->num_patch_points;
ppoint_idx++) {
ppoint_desc.ppoint_index = ppoint_idx;
/* ldr_safety can happen only for hotp_only. */
ASSERT(DYNAMO_OPTION(hotp_only) || !ldr_safety);
/* Don't re-add a patch point to the vector
* during patch injection while handling
* loader safe injection.
*/
if (!ldr_safety)
hotp_ppoint_areas_add(&ppoint_desc);
if (DYNAMO_OPTION(hotp_only)) {
hotp_only_inject_patch(&ppoint_desc, thread_table,
num_threads);
}
}
}
} else { /* unloading dll */
/* TODO: same issues as in the 'if' block above, but
* reverse.
*/
module->base_address = NULL;
module->matched = false;
hotp_set_policy_status(vul_idx, HOTP_INJECT_NO_MATCH);
}
}
}
}
}
if (!DYNAMO_OPTION(use_persisted_hotp)) /* else we check in loop above */
flush_perscache = any_matched && perscache_range_overlap;
if (flush_perscache) {
ASSERT(any_matched && perscache_range_overlap);
ASSERT(!DYNAMO_OPTION(hotp_only));
/* During startup we process hotp before we add exec areas, so we
* should only get a match in a later nudge, when we pass in toflush.
*/
ASSERT(dynamo_initialized);
ASSERT(toflush != NULL);
# ifdef WINDOWS
ASSERT(dcontext->nudge_target != NULL);
# else
ASSERT_NOT_REACHED(); /* No nudge on Linux, should only be here for nudge. */
# endif
if (toflush != NULL) { /* be paranoid (we fail otherwise though) */
LOG(GLOBAL, LOG_HOT_PATCHING, 2,
"Hotp for " PFX "-" PFX " %s overlaps perscache, flushing\n", base,
base + image_size, name);
/* As we hold the hotp_vul_table_lock we cannot flush here;
* instead we add to a pending-flush vmvector.
*/
vmvector_add(toflush, base, base + image_size, NULL);
STATS_INC(hotp_persist_flush);
/* XXX: we could eliminate this and rely on our later flush of the
* patch area, as we're only coming here for nudges; we technically
* only need an explicit check when loading a perscache, as long as
* hotp defs are set up first.
*/
}
}
hotp_process_image_exit:
if (pe_name != NULL)
dr_strfree(pe_name HEAPACCT(ACCT_HOT_PATCHING));
if (mod_name != NULL)
dr_strfree(mod_name HEAPACCT(ACCT_HOT_PATCHING));
/* Don't unlock in case the lock was already obtained before reaching this
* function. Only in that case lock_acquired will be false.
*/
/* TODO: or does this go after flush? */
if (!own_hot_patch_lock)
d_r_write_unlock(&hotp_vul_table_lock);
/* TODO: also there are some race conditions with nudging & policy lookup/
* injection; sort those out; flushing before or after reading the
* policy plays a role too.
*/
# ifdef WINDOWS
if (num_threads == HOTP_ONLY_NUM_THREADS_AT_INIT) {
ASSERT(DYNAMO_OPTION(hotp_only));
ASSERT(!just_check);
ASSERT_CURIOSITY(check_sole_thread());
d_r_mutex_unlock(&thread_initexit_lock);
d_r_mutex_unlock(&all_threads_synch_lock);
}
# endif
}
/* If vec==NULL, returns the number of patch points for the
* matched vuls in [start,end).
* Else, stores in vec the offsets for all the matched patch points in [start,end).
* Returns -1 if vec!=NULL and vec_num is too small (still fills it up).
* For now this routine assumes that [start,end) is contained in a single module.
* The caller must own the hotp_vul_table_lock (as a read lock).
*/
static int
hotp_patch_point_persist_helper(const app_pc start, const app_pc end, app_rva_t *vec,
uint vec_num)
{
uint num_ppoints = 0;
uint vul, set, module, pp;
/* XXX: check [start,end) instead of module */
app_pc modbase = get_module_base(start);
ASSERT(modbase == get_module_base(end));
ASSERT(start != NULL); /* only support single module for now */
ASSERT_OWN_READ_LOCK(true, &hotp_vul_table_lock);
if (GLOBAL_VUL_TABLE == NULL)
return 0;
/* XXX: once hotp_patch_point_areas is not just hotp_only, use it here */
for (vul = 0; vul < NUM_GLOBAL_VULS; vul++) {
bool set_processed = false;
/* Ignore if off or dll wasn't loaded */
if (GLOBAL_VUL(vul).mode == HOTP_MODE_OFF ||
GLOBAL_VUL(vul).hotp_dll_base == NULL)
continue;
for (set = 0; set < GLOBAL_VUL(vul).num_sets; set++) {
/* Only the first matching set should be used; case 10248. */
if (set_processed)
break;
for (module = 0; module < GLOBAL_SET(vul, set).num_modules; module++) {
if (!GLOBAL_MODULE(vul, set, module).matched ||
modbase != GLOBAL_MODULE(vul, set, module).base_address)
continue;
set_processed = true;
if (vec == NULL) {
num_ppoints += GLOBAL_MODULE(vul, set, module).num_patch_points;
} else {
for (pp = 0; pp < GLOBAL_MODULE(vul, set, module).num_patch_points;
pp++) {
if (num_ppoints >= vec_num) {
/* It's ok to get here, just currently no callers do */
ASSERT_NOT_TESTED();
return -1;
}
vec[num_ppoints++] = GLOBAL_PPOINT(vul, set, module, pp).offset;
}
}
}
}
}
return num_ppoints;
}
/* Returns the number of patch points for the matched vuls in [start,end).
* For now this routine assumes that [start,end) is contained in a single module.
* The caller must own the hotp_vul_table_lock (as a read lock).
*/
int
hotp_num_matched_patch_points(const app_pc start, const app_pc end)
{
return hotp_patch_point_persist_helper(start, end, NULL, 0);
}
/* Stores in vec the offsets for all the matched patch points in [start,end).
* Returns -1 if vec_num is too small (still fills it up).
* For now this routine assumes that [start,end) is contained in a single module.
* The caller must own the hotp_vul_table_lock (as a read lock).
*/
int
hotp_get_matched_patch_points(const app_pc start, const app_pc end, app_rva_t *vec,
uint vec_num)
{
return hotp_patch_point_persist_helper(start, end, vec, vec_num);
}
/* Checks whether any matched patch point in [start, end) is not listed on
* hotp_ppoint_vec. If hotp_ppoint_vec is NULL just checks whether any patch
* point is matched in the region. For now this routine assumes that
* [start,end) is contained in a single module.
*/
bool
hotp_point_not_on_list(const app_pc start, const app_pc end, bool own_hot_patch_lock,
app_rva_t *hotp_ppoint_vec, uint hotp_ppoint_vec_num)
{
/* We could use hotp_process_image_helper()'s just_check but would have
* to add hotp_ppoint_vec arg; plus we don't care about module matching.
*/
bool not_on_list = false;
uint vul, set, module, pp;
/* XXX: check [start,end) instead of module */
app_pc modbase = get_module_base(start);
DEBUG_DECLARE(bool matched = false;)
ASSERT(modbase == get_module_base(end));
if (!own_hot_patch_lock)
d_r_read_lock(&hotp_vul_table_lock);
ASSERT_OWN_READWRITE_LOCK(true, &hotp_vul_table_lock);
if (GLOBAL_VUL_TABLE == NULL)
goto hotp_policy_list_exit;
/* XXX: I would make an iterator to share w/ patch_point_persist_helper but
* this many-nested loop lookup should go away in general ASAP and be
* replaced w/ hotp_patch_point_areas which is currently only hotp_only.
*/
for (vul = 0; vul < NUM_GLOBAL_VULS; vul++) {
bool set_processed = false;
/* Ignore if off or dll wasn't loaded */
if (GLOBAL_VUL(vul).mode == HOTP_MODE_OFF ||
GLOBAL_VUL(vul).hotp_dll_base == NULL)
continue;
for (set = 0; set < GLOBAL_VUL(vul).num_sets; set++) {
/* Only the first matching set should be used; case 10248. */
if (set_processed)
break;
for (module = 0; module < GLOBAL_SET(vul, set).num_modules; module++) {
if (!GLOBAL_MODULE(vul, set, module).matched ||
modbase != GLOBAL_MODULE(vul, set, module).base_address)
continue;
/* We have a match; only ok if on the list */
DODEBUG({ matched = true; });
set_processed = true;
ASSERT(!not_on_list); /* should have exited if not on list */
not_on_list = true;
if (hotp_ppoint_vec != NULL && DYNAMO_OPTION(use_persisted_hotp)) {
for (pp = 0; pp < GLOBAL_MODULE(vul, set, module).num_patch_points;
pp++) {
if (!hotp_ppoint_on_list(
GLOBAL_PPOINT(vul, set, module, pp).offset,
hotp_ppoint_vec, hotp_ppoint_vec_num))
goto hotp_policy_list_exit;
}
not_on_list = false;
} else
goto hotp_policy_list_exit;
}
}
}
hotp_policy_list_exit:
if (!own_hot_patch_lock)
d_r_read_unlock(&hotp_vul_table_lock);
DOSTATS({
if (matched && !not_on_list) {
ASSERT(hotp_ppoint_vec != NULL && DYNAMO_OPTION(use_persisted_hotp));
STATS_INC(perscache_hotp_conflict_avoided);
}
});
return not_on_list;
}
/* TODO: change this to walk the new PE list (not for now though); needed only
* during nudge; start up walk is already done by the core, piggyback
* on that and call hotp_process_image() there; basically, get rid
* of the need to walk the loader list
* Note: for -probe_api, we walk the module list at start up
* because client init is done after vmareas_init, i.e., after
* scanning for modules in memory and processing them.
*
*/
static void
hotp_walk_loader_list(thread_record_t **thread_table, const int num_threads,
vm_area_vector_t *toflush, bool probe_init)
{
/* This routine will go away; till then need to compile on linux. Not walking
* the module list on linux means that no vulnerability will get activated
* for injection; that is ok as we aren't trying to build a linux version now.
*/
# ifdef WINDOWS
/* TODO: this routine uses PEB, etc, this has to be os specific */
PEB *peb = get_own_peb();
PEB_LDR_DATA *ldr = peb->LoaderData;
LIST_ENTRY *e, *start;
LDR_MODULE *mod;
/* For hotp_only, thread_table can be valid, i.e., all known threads may be
* suspended. Even if they are not, all synch locks will be held, so that
* module processing can happen without races. Check for that.
* Note: for -probe_api, this routine can be called during dr init time,
* i.e., synch locks won't be held, so we shouldn't assert.
*/
if (!probe_init) {
ASSERT_OWN_MUTEX(DYNAMO_OPTION(hotp_only), &all_threads_synch_lock);
ASSERT_OWN_MUTEX(DYNAMO_OPTION(hotp_only), &thread_initexit_lock);
}
/* Flushing of pcaches conflicting with hot patches is handled for dll
* loads by the pcache loads. Conflicts at hotp_init time can't happen as
* pcaches won't be loaded then (they are loaded in vm_areas_init which
* comes afterwards). However for nudging and client init
* (dr_register_probes) this is needed because pcaches can be loaded by
* then. Note even though client init happens during startup, it happens
* after vm_areas_init, hence pcaches can be loaded. PR 226578 tracks
* implementing pcache flushes for probe api - till then this assert is
* relaxed.
*/
ASSERT(toflush != NULL || DYNAMO_OPTION(hotp_only) ||
(DYNAMO_OPTION(probe_api) && !DYNAMO_OPTION(use_persisted)));
start = &ldr->InLoadOrderModuleList;
for (e = start->Flink; e != start; e = e->Flink) {
mod = (LDR_MODULE *)e;
/* TODO: ASSERT that the module is loaded? */
hotp_process_image_helper(mod->BaseAddress, true, probe_init ? false : true,
false, NULL, thread_table, num_threads, false /*!ldr*/,
toflush);
/* TODO: make hotp_process_image() emit different log messages
* depending upon which path it is invoked from.
*/
}
# endif /* WINDOWS */
}
void
hotp_init(void)
{
ASSIGN_INIT_READWRITE_LOCK_FREE(hotp_vul_table_lock, hotp_vul_table_lock);
/* Assuming no locks are held on while initializing hot patching. */
ASSERT_OWN_NO_LOCKS();
ASSERT(DYNAMO_OPTION(hot_patching));
# ifdef GBOP
/* gbop can't be turned on without hotp_only. */
ASSERT(DYNAMO_OPTION(hotp_only) || !DYNAMO_OPTION(gbop));
# endif
if (DYNAMO_OPTION(hotp_only)) {
VMVECTOR_ALLOC_VECTOR(hotp_only_tramp_areas, GLOBAL_DCONTEXT,
VECTOR_SHARED | VECTOR_NEVER_MERGE,
hotp_only_tramp_areas_lock);
}
d_r_write_lock(&hotp_vul_table_lock);
# ifdef DEBUG
hotp_globals =
HEAP_TYPE_ALLOC(GLOBAL_DCONTEXT, hotp_globals_t, ACCT_HOT_PATCHING, PROTECTED);
hotp_globals->ldr_safe_hook_removal = false;
hotp_globals->ldr_safe_hook_injection = false;
# endif
/* Currently hotp_patch_point_areas is used for hotp_only to do module
* matching (case 7279) and offset lookup (case 8132), and offset lookup
* only for hotp with fcache (case 10075). Later on it will be
* used for patch injection, removal, perscache stuff, etc; case 10728.
*/
VMVECTOR_ALLOC_VECTOR(hotp_patch_point_areas, GLOBAL_DCONTEXT,
VECTOR_SHARED | VECTOR_NEVER_MERGE,
hotp_patch_point_areas_lock);
/* hotp_only trampolines should be allocated on a special heap that allows
* code to be executed in it.
*/
if (DYNAMO_OPTION(hotp_only)) {
hotp_only_tramp_heap =
special_heap_init(HOTP_ONLY_TRAMPOLINE_SIZE, true, /* yes, use a lock */
true, /* make it executable */
true /* it is persistent */);
}
ASSERT(GLOBAL_VUL_TABLE == NULL && NUM_GLOBAL_VULS == 0);
GLOBAL_VUL_TABLE = hotp_read_policy_defs(&NUM_GLOBAL_VULS);
if (GLOBAL_VUL_TABLE != NULL) {
hotp_load_hotp_dlls(GLOBAL_VUL_TABLE, NUM_GLOBAL_VULS);
hotp_read_policy_modes(NULL);
/* Policy status table must be initialized after the global
* vulnerability table and modes are read, but before module list
* is iterated over.
*/
hotp_init_policy_status_table();
/* We don't need to call hotp_walk_loader_list() here as
* find_executable_vm_areas() will call hotp_process_image() for us.
*/
} else {
LOG(GLOBAL, LOG_HOT_PATCHING, 2, "No hot patch definitions to read\n");
}
/* Release locks. */
d_r_write_unlock(&hotp_vul_table_lock);
/* Can't hold any locks at the end of hot patch initializations. */
ASSERT_OWN_NO_LOCKS();
}
/* thread-shared initialization that should be repeated after a reset */
void
hotp_reset_init(void)
{
/* nothing to do */
}
/* Free all thread-shared state not critical to forward progress;
* hotp_reset_init() will be called before continuing.
*/
void
hotp_reset_free(void)
{
/* Free old tables. Hot patch code must ensure that no old table
* pointer is kept across a synch-all point, which is also a reset
* point (case 7760 & 8921).
*/
hotp_vul_tab_t *current_tab, *temp_tab;
if (!DYNAMO_OPTION(hot_patching))
return;
d_r_write_lock(&hotp_vul_table_lock);
temp_tab = hotp_old_vul_tabs;
while (temp_tab != NULL) {
current_tab = temp_tab;
temp_tab = temp_tab->next;
hotp_free_vul_table(current_tab->vul_tab, current_tab->num_vuls);
heap_free(GLOBAL_DCONTEXT, current_tab,
sizeof(hotp_vul_tab_t) HEAPACCT(ACCT_HOT_PATCHING));
}
hotp_old_vul_tabs = NULL;
d_r_write_unlock(&hotp_vul_table_lock);
}
/* Free up all allocated memory and delete hot patching lock. */
void
hotp_exit(void)
{
/* This assert will ensure that there is only one thread in the process
* during exit. Grab the hot patch lock all the same because a nudge
* can come in at this point; freeing things without the lock is bad.
*/
ASSERT(dynamo_exited);
ASSERT(DYNAMO_OPTION(hot_patching));
d_r_write_lock(&hotp_vul_table_lock);
/* Release the hot patch policy status table if allocated. This table
* may not be allocated till the end if there were no hot patch definitions
* but -hot_patching was turned on.
*/
if (hotp_policy_status_table != NULL) {
heap_free(GLOBAL_DCONTEXT, hotp_policy_status_table,
hotp_policy_status_table->size HEAPACCT(ACCT_HOT_PATCHING));
hotp_policy_status_table = NULL;
}
/* Release the patch point areas vector before the table. */
hotp_ppoint_areas_release();
vmvector_delete_vector(GLOBAL_DCONTEXT, hotp_patch_point_areas);
hotp_patch_point_areas = NULL;
/* Release the global vulnerability table and old tables if any. */
hotp_free_vul_table(GLOBAL_VUL_TABLE, NUM_GLOBAL_VULS);
/* case 8118: set to NULL since referenced in hotp_print_diagnostics() */
GLOBAL_VUL_TABLE = NULL;
# ifdef DEBUG
HEAP_TYPE_FREE(GLOBAL_DCONTEXT, hotp_globals, hotp_globals_t, ACCT_HOT_PATCHING,
PROTECTED);
# endif
d_r_write_unlock(&hotp_vul_table_lock);
hotp_reset_free();
if (DYNAMO_OPTION(hotp_only)) {
# ifdef WINDOWS
/* Don't free the heap upon detach - app may have hooked with our
* trampoline code; case 9593. Make this memory efficient, i.e., delete
* the heap if no collisions were detected; part of bookkeepping needed
* to not leak all removed hotp trampolines, but only those that have a
* potential collision; a minor TODO - - would save a max of 50kb.
* Note: heap lock should be deleted even if heap isn't! */
/* If hotp_only_tramp_heap_cache is NULL, it means that no patches
* were removed (either because they weren't injected or just not
* removed). This means we don't have to leak the trampolines even
* for detach (PR 215520). */
if (!doing_detach || hotp_only_tramp_heap_cache == NULL)
special_heap_exit(hotp_only_tramp_heap);
# ifdef DEBUG
else
special_heap_delete_lock(hotp_only_tramp_heap);
# endif
# else
special_heap_exit(hotp_only_tramp_heap);
# endif
hotp_only_tramp_heap = NULL;
vmvector_delete_vector(GLOBAL_DCONTEXT, hotp_only_tramp_areas);
hotp_only_tramp_areas = NULL;
}
DELETE_READWRITE_LOCK(hotp_vul_table_lock);
}
/* Hot patch policy update action handler */
bool
nudge_action_read_policies(void)
{
hotp_vul_t *old_vul_table = NULL, *new_vul_table = NULL;
uint num_old_vuls = 0, num_new_vuls;
int num_threads = 0;
thread_record_t **thread_table = NULL;
STATS_INC(hotp_num_policy_nudge);
/* Fix for case 6090; TODO: remove when -hotp_policy_size is removed */
synchronize_dynamic_options();
new_vul_table = hotp_read_policy_defs(&num_new_vuls);
if (new_vul_table != NULL) {
bool old_value;
hotp_vul_tab_t *temp;
dr_where_am_i_t wherewasi;
dcontext_t *dcontext = get_thread_private_dcontext();
vm_area_vector_t toflush; /* never initialized for hotp_only */
/* If dynamo_exited was false till the check in this routine, then
* this thread would have been intercepted by the core, i.e., it
* would have got a dcontext. The assert is to catch
* bugs; the if is to make sure that the release build doesn't
* crash in case this happens.
*/
ASSERT(dcontext != NULL && dcontext != GLOBAL_DCONTEXT);
if (dcontext == NULL) {
return false; /* skip further processing */
}
/* When the nudge thread starts up, the core takes control and
* lets it go once it is identified as nudge. However,
* under_dynamo_control is still true because we come here from the cache.
* We need to set under_dynamo_control to false during hot patch dll loading,
* otherwise the core will take over again at the dll loading interception point.
* Once hot patch dlls are loaded we restore under_dynamo_control in case it's
* relied on elsewhere. Note - this isn't needed for
* loading hot patch dlls at startup because thread init comes
* after hotp_init(), so under_dynamo_control isn't set. Only
* hot patch dll loading during nudge needs this.
* TODO: under_dynamo_control needs cleanup - see case 529, 5183.
*/
old_value = dcontext->thread_record->under_dynamo_control;
dcontext->thread_record->under_dynamo_control = false;
/* Fix for case 5367. TODO: undo fix after writing own loader. */
wherewasi = dcontext->whereami;
dcontext->whereami = DR_WHERE_APP; /* DR_WHERE_APP? more like DR_WHERE_DR */
dcontext->nudge_thread = true;
/* Fix for case 5376. There can be a deadlock if a nudge happened
* to result in hot patch dlls being loaded when at the same time
* an app dll was being loaded; hotp_vul_table_lock & LoaderLock
* would create a deadlock. So while loading the hot patch dlls
* the hotp_vul_table_lock shouldn't be held.
* To avoid this the table is read, stored in a temporary variable
* and hot patch dlls are loaded using that temp. table - all this
* is now done without the hotp_vul_table_lock. Then the vul table
* lock is grabbed (see below) and the global table is setup.
*
* XXX: The longer term solution is to have our own loader to
* load hot patch dlls.
*/
hotp_load_hotp_dlls(new_vul_table, num_new_vuls);
/* Must be set to false, otherwise the subsequent module list
* walking will be useless, i.e., won't be able to identify
* modules for hot patching because hotp_process_image() won't work.
*/
dcontext->nudge_thread = false;
/* If whereami changed, that means, the probably was a callback,
* which can lead to other bugs. So, let us make sure it doesn't.
*/
ASSERT(dcontext->whereami == DR_WHERE_APP);
dcontext->whereami = wherewasi;
dcontext->thread_record->under_dynamo_control = old_value;
/* Suspend all threads (for hotp_only) and grab locks. */
if (DYNAMO_OPTION(hotp_only)) {
# ifdef WINDOWS
DEBUG_DECLARE(bool ok =)
synch_with_all_threads(THREAD_SYNCH_SUSPENDED, &thread_table,
/* Case 6821: other synch-all-thread uses that
* only care about threads carrying fcache
* state can ignore us
*/
&num_threads, THREAD_SYNCH_NO_LOCKS_NO_XFER,
/* if we fail to suspend a thread (e.g., privilege
* problems) ignore it. XXX: retry instead? */
THREAD_SYNCH_SUSPEND_FAILURE_IGNORE);
ASSERT(ok);
# endif
}
/* CAUTION: Setting up the global table, reading modes, setting up
* policy status table and module list walking MUST all be
* done in that order with the table lock held as all of them
* update the global table.
*/
d_r_write_lock(&hotp_vul_table_lock);
/* For hotp_only, all patches have to be removed before doing
* anything with new vulnerability data, and nothing after that,
* which is unlike hotp, where removal has to be done before & after.
*/
if (DYNAMO_OPTION(hotp_only)) {
hotp_remove_hot_patches(GLOBAL_VUL_TABLE, NUM_GLOBAL_VULS, true, NULL);
}
/* Save the old table for flushing & launch the new table. */
old_vul_table = GLOBAL_VUL_TABLE;
num_old_vuls = NUM_GLOBAL_VULS;
hotp_ppoint_areas_release(); /* throw out the old patch points */
SELF_UNPROTECT_DATASEC(DATASEC_RARELY_PROT);
GLOBAL_VUL_TABLE = new_vul_table;
NUM_GLOBAL_VULS = num_new_vuls;
SELF_PROTECT_DATASEC(DATASEC_RARELY_PROT);
hotp_read_policy_modes(NULL);
/* Policy status table must be initialized after the global
* vulnerability table and modes are read, but before module list
* is iterated over.
*/
SELF_UNPROTECT_DATASEC(DATASEC_RARELY_PROT);
hotp_init_policy_status_table();
if (!DYNAMO_OPTION(hotp_only))
vmvector_init_vector(&toflush, 0); /* no lock init needed since not used */
hotp_walk_loader_list(thread_table, num_threads,
DYNAMO_OPTION(hotp_only) ? NULL : &toflush,
false /* !probe_init */);
SELF_PROTECT_DATASEC(DATASEC_RARELY_PROT);
/* Release all locks. */
d_r_write_unlock(&hotp_vul_table_lock);
# ifdef WINDOWS
if (DYNAMO_OPTION(hotp_only)) {
end_synch_with_all_threads(thread_table, num_threads, true /*resume*/);
}
# endif
/* If a new vulnerability table was created, then flush the bbs
* with hot patches from the old table and then free that table.
* Note, the old table has to be freed outside the scope of the
* hotp_vul_table_lock because bbs corresponding to that table
* can't be flushed inside it. See flushing comments below.
*/
ASSERT(old_vul_table != GLOBAL_VUL_TABLE);
if (!DYNAMO_OPTION(hotp_only)) {
if (!vmvector_empty(&toflush)) {
ASSERT(DYNAMO_OPTION(coarse_units) && DYNAMO_OPTION(use_persisted));
/* case 9970: we must flush the perscache and ibl tables.
* XXX optimization: don't flush the fine-grained fragments
* or non-persisted unit(s) (there can be multiple).
*/
flush_vmvector_regions(get_thread_private_dcontext(), &toflush,
false /*keep futures*/,
false /*exec still valid*/);
}
/* XXX: don't need to flush non-persisted coarse units since
* patch points are fine-grained: would have to widen
* flush interface. Note that we do avoid flushing perscaches
* that do not contain the old patch points.
*/
hotp_remove_hot_patches(old_vul_table, num_old_vuls, false, NULL);
vmvector_reset_vector(GLOBAL_DCONTEXT, &toflush);
} /* else toflush is uninitialized */
/* Freeing the old vulnerability table immediately causes a race
* with hot patch execution (see case 5521), so it is put on a free
* list and freed at a reset or dr exit. hotp_vul_table_lock must be held here;
* though this list is a new structure, a new lock is unnecessary.
* Also, don't chain empty tables; a NULL table can occur when
* no hot patches are loaded during startup, but are nudged in.
*
* Case 8921: We can't add to the old list prior to removing
* hot patches since the synch-all for -coarse_unit or
* -hotp_only flushing is a reset point and the table can then
* be freed underneath us. Thus we pay the cost of re-acquiring the lock.
* This can also end up with tables on the old list in a different order
* than their nudges, but that's not a problem.
* XXX case 8921: -hotp_only should free the table up front
* XXX: we should synch-all once, up front, and then avoid this ugliness
* as well as multiple flush synchs.
* XXX: hotp could indirect the table like hotp_only to allow
* earlier freeing.
*/
/* XXX: don't add to old table list if in hotp_only mode, there is no
* eed because there is a lookup before execution and there is no lazy
* flush going on.
*/
if (old_vul_table != NULL) {
temp = HEAP_TYPE_ALLOC(GLOBAL_DCONTEXT, hotp_vul_tab_t, ACCT_HOT_PATCHING,
PROTECTED);
temp->vul_tab = old_vul_table;
temp->num_vuls = num_old_vuls;
d_r_write_lock(&hotp_vul_table_lock);
SELF_UNPROTECT_DATASEC(DATASEC_RARELY_PROT);
temp->next = hotp_old_vul_tabs;
hotp_old_vul_tabs = temp;
SELF_PROTECT_DATASEC(DATASEC_RARELY_PROT);
d_r_write_unlock(&hotp_vul_table_lock);
}
} else {
/* Note, if the new table wasn't read in successfully, then the old
* table isn't touched, i.e., status quo is maintained.
*/
LOG(GLOBAL, LOG_HOT_PATCHING, 2, "No hot patch policies to read\n");
}
return true;
}
# ifdef WINDOWS /* no nudging yet on Linux */
/* This routine handles hot patch nudges. */
void
hotp_nudge_handler(uint nudge_action_mask)
{
/* Note, multiple nudges will be synchronized by the hotp_vul_table_lock.
* It is irrelevant if nudge threads change order between reading and
* flushing. */
ASSERT(DYNAMO_OPTION(liveshields) && DYNAMO_OPTION(hot_patching));
ASSERT(nudge_action_mask != 0); /* else shouldn't be called */
if (TESTANY(NUDGE_GENERIC(lstats), nudge_action_mask)) {
SYSLOG_INTERNAL_WARNING("Stat dumping for hot patches not done yet.");
}
if (TESTANY(NUDGE_GENERIC(policy), nudge_action_mask)) {
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "\n nudged to read in policy\n");
nudge_action_read_policies();
}
if (TESTANY(NUDGE_GENERIC(mode), nudge_action_mask)) {
thread_record_t **thread_table = NULL;
int num_threads = 0;
hotp_policy_mode_t *old_modes = NULL;
vm_area_vector_t toflush; /* never initialized for hotp_only */
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "\n nudged to read in policy\n");
STATS_INC(hotp_num_mode_nudge);
/* If -liveshields isn't on, then modes nudges are irrelevant. */
if (!DYNAMO_OPTION(liveshields))
return;
/* Suspend all threads (for hotp_only) and grab locks. */
if (DYNAMO_OPTION(hotp_only)) {
DEBUG_DECLARE(bool ok =)
synch_with_all_threads(THREAD_SYNCH_SUSPENDED, &thread_table,
/* Case 6821: other synch-all-thread uses that
* only care about threads carrying fcache
* state can ignore us
*/
&num_threads, THREAD_SYNCH_NO_LOCKS_NO_XFER,
/* if we fail to suspend a thread (e.g., privilege
* problems) ignore it. XXX: retry instead? */
THREAD_SYNCH_SUSPEND_FAILURE_IGNORE);
ASSERT(ok);
}
d_r_write_lock(&hotp_vul_table_lock);
/* For hotp_only, all patches have to be removed before doing anything
* with new mode data; loader list walking will inject new ones.
*/
if (DYNAMO_OPTION(hotp_only)) {
hotp_remove_hot_patches(GLOBAL_VUL_TABLE, NUM_GLOBAL_VULS, true, NULL);
}
hotp_ppoint_areas_release(); /* throw out the old patch points */
/* Old modes are for regular hot patching, not for hotp_only. */
hotp_read_policy_modes(DYNAMO_OPTION(hotp_only) ? NULL : &old_modes);
/* Policy status table must be initialized after the global
* vulnerability table and modes are read, but before module list
* is iterated over.
*/
SELF_UNPROTECT_DATASEC(DATASEC_RARELY_PROT);
hotp_init_policy_status_table();
SELF_PROTECT_DATASEC(DATASEC_RARELY_PROT);
if (!DYNAMO_OPTION(hotp_only))
vmvector_init_vector(&toflush, 0); /* no lock init needed since not used */
hotp_walk_loader_list(thread_table, num_threads,
DYNAMO_OPTION(hotp_only) ? NULL : &toflush,
false /* !probe_init */);
/* Release all locks. */
d_r_write_unlock(&hotp_vul_table_lock);
if (DYNAMO_OPTION(hotp_only)) {
end_synch_with_all_threads(thread_table, num_threads, true /*resume*/);
}
/* If modes did change, then we need to flush out patches that were
* injected because their old modes were on (detect or protect).
* Fix for case 6619; resulted in using old_modes for patch removal.
* Note: Just like policy reading, flushing has to be done outside the
* scope of the hotp_vul_table_lock & ONLY after reading the new modes.
*/
if (!DYNAMO_OPTION(hotp_only)) {
if (!vmvector_empty(&toflush)) {
/* case 9970: we must flush the perscache and ibl tables.
* XXX optimization: don't flush the fine-grained fragments
* or non-persisted unit(s) (there can be multiple).
*/
flush_vmvector_regions(get_thread_private_dcontext(), &toflush,
false /*keep futures*/,
false /*exec still valid*/);
}
hotp_remove_hot_patches(GLOBAL_VUL_TABLE, NUM_GLOBAL_VULS, false, old_modes);
if (old_modes != NULL) {
HEAP_ARRAY_FREE(GLOBAL_DCONTEXT, old_modes, hotp_policy_mode_t,
NUM_GLOBAL_VULS, ACCT_HOT_PATCHING, PROTECTED);
}
vmvector_reset_vector(GLOBAL_DCONTEXT, &toflush);
} /* else toflush is uninitialized */
}
/* Flushing injected bbs must be outside the scope of hotp_vul_table_lock.
* Otherwise, flushing will deadlock. See case 5415.
* Though it happens so, it is safe. The side effect of this is that
* bbs with hot patches that have been turned off would still be active
* till the flush below, which is ok as they were already active. Similarly
* hot patches that have been turned on will not work until the flush
* happens.
*
* There are two flushes for hotp mode per nudge (policy or mode read) and
* one for hotp_only mode. For hotp, the first flush is to clean out bbs
* with old/injected patches and is done above (nudge_action_read_policies -
* in the case of policy nudge).
*
* The second flush is to remove bbs corresponding to new policies/modes,
* i.e., bbs that were already translated but weren't injected based on
* any old policies/modes, but are by new ones. This is applicable to
* both policy & mode reading.
* Note that for case 9995 we avoided flushing perscaches that do not contain
* the new patch points at match time, and we avoid flushing here with
* checks in vm_area_allsynch_flush_fragments.
*/
if (TESTANY(NUDGE_GENERIC(mode), nudge_action_mask) ||
TESTANY(NUDGE_GENERIC(policy), nudge_action_mask)) {
if (!DYNAMO_OPTION(hotp_only))
hotp_remove_hot_patches(GLOBAL_VUL_TABLE, NUM_GLOBAL_VULS, false, NULL);
}
}
# endif /* WINDOWS */
/* This is a faster lookup of hotp_vul_table, see case 8132.
* XXX: try to see if this can be merged with hotp_lookup_patch_addr.
*/
static bool
hotp_only_lookup_patch_addr(const app_pc pc, hotp_offset_match_t *match)
{
hotp_offset_match_t *ppoint_desc;
ASSERT(pc != NULL && match != NULL);
ASSERT(DYNAMO_OPTION(hotp_only));
/* This is always initialized at startup, so can't be NULL at this point. */
ASSERT(hotp_patch_point_areas != NULL);
/* Table read & injection as done together, if a module matches; even if it
* doesn't no patching will take place when table is NULL. Similarly, no
* patch is left when table is emptied/cleared for update. Thus, hotp
* won't execute if the global table is null, which is where this lookup is
* done. */
ASSERT(GLOBAL_VUL_TABLE != NULL);
ASSERT_OWN_READ_LOCK(true, &hotp_vul_table_lock);
ppoint_desc = (hotp_offset_match_t *)vmvector_lookup(hotp_patch_point_areas, pc);
if (ppoint_desc == NULL) {
/* Custom data for this vector can't be NULL, so NULL means failure. */
return false;
} else {
*match = *ppoint_desc;
return true;
}
}
/* TODO: need to use the concept of a policy activation in addition with pc
* to ensure that the library of the patch point is actually loaded!
* i think this is best if done at the time of adding/removing patch
* points to lookup structures
* TODO: start using the source (i.e., pc's) dll name to verify
* patch point/policy with the policy's dll; similar issue as above.
* TODO: take an argument for lock; in the bb stage call it with no lock; in
* the injecting stage call it with lock;
* TODO: split up lookup into two, a vm area lookup in the outer decode loop in
* build_bb_ilist() & a pc lookup inside hotp_inject(); the former
* will be racy and will serve as a first level check; the latter is to
* be used only for injection purposes and won't be racy (because it will
* be called within the scope of GLOBAL_VUL_TABLE or pc hash lock),
* and won't be visible outside the hotpatch module;
* see is_executable_address() for sample.
* may need new locks for lookup data structures.
*
* pc lookup should match only if pc matches, dll matches, mode is not
* off and all dlls are available (i.e., vulnerability is active).
*/
static bool
hotp_lookup_patch_addr(const app_pc pc, hotp_offset_match_t *match,
bool own_hot_patch_lock)
{
bool res = false;
hotp_offset_match_t *ppoint_desc = NULL;
ASSERT(pc != NULL && match != NULL);
if (pc == NULL) /* Defensively exit. */
return false;
/* This is called only during patch clean call injection into fcache, hence
* not applicable to hotp_only.
*/
ASSERT(!DYNAMO_OPTION(hotp_only));
/* There is a remote possibility that GLOBAL_VUL_TABLE can become NULL
* between the time the hotp lookup in bb building succeeded and the time
* actual patch injection takes place. This can be caused by a nudge
* with an empty or faulty policy config file. So, can't assert on
* GLOBAL_VUL_TABLE not being NULL.
*/
if (GLOBAL_VUL_TABLE == NULL) /* Nothing to lookup. */
return res;
/* This is always initialized at startup, so can't be NULL at this point. */
ASSERT(hotp_patch_point_areas != NULL);
if (!own_hot_patch_lock) /* Fix for case 5323. */
d_r_read_lock(&hotp_vul_table_lock);
/* Can come here with either the read lock (during instruction matching) or
* with the write lock (during injection).
*/
ASSERT_OWN_READWRITE_LOCK(true, &hotp_vul_table_lock);
ppoint_desc = (hotp_offset_match_t *)vmvector_lookup(hotp_patch_point_areas, pc);
if (ppoint_desc != NULL) {
/* If the hot patch dll for this vulnerability wasn't loaded for any
* reason, don't even bother with pc matching; we can't execute the
* corresponding patch as it hasn't been loaded. Fix for case 6032.
* TODO: when splitting up the pc lookup, this should be taken care
* of too.
* Assert as it is a LiveShield product bug, not dr bug; but handle it.
*/
ASSERT(GLOBAL_VUL(ppoint_desc->vul_index).hotp_dll_base != NULL &&
"hot patch dll loaded");
if (GLOBAL_VUL(ppoint_desc->vul_index).hotp_dll_base == NULL)
goto hotp_lookup_patch_addr_exit; /* lookup failed */
/* TODO: check if vul. is ready, i.e., all modules match */
ASSERT(GLOBAL_MODULE(ppoint_desc->vul_index, ppoint_desc->set_index,
ppoint_desc->module_index)
.matched);
ASSERT(GLOBAL_VUL(ppoint_desc->vul_index).mode == HOTP_MODE_DETECT ||
GLOBAL_VUL(ppoint_desc->vul_index).mode == HOTP_MODE_PROTECT);
/* TODO: assert that the indices are within limits */
/* TODO: vulnerability is returned without a lock for it,
* definitely a problem because it can be updated while being used.
* TODO: Also, need to figure out a way to return multiple matches.
*/
res = true; /* vmvector lookup succeeded */
if (match != NULL) /* match can't be NULL, but be cautious. */
*match = *ppoint_desc;
LOG(GLOBAL, LOG_HOT_PATCHING, 1,
"lookup for " PFX " succeeded with vulnerability #%s\n", pc,
GLOBAL_VUL(ppoint_desc->vul_index).vul_id);
}
hotp_lookup_patch_addr_exit:
if (!own_hot_patch_lock)
d_r_read_unlock(&hotp_vul_table_lock);
return res;
}
/* Returns true if the region passed in should be patched and the module is
* ready, i.e., loaded & matched.
* Note: start and end define a region that is looked up a vmvector,
* hotp_patch_point_ares. Though our vmvector can accept end being NULL,
* signifying no upper ceiling, it doesn't make sense of hot patch lookup - at
* best it signifies an error somewhere. So a NULL for end will be treated as a
* lookup failure.
*/
bool
hotp_does_region_need_patch(const app_pc start, const app_pc end, bool own_hot_patch_lock)
{
bool res = false;
ASSERT(start != NULL && end != NULL);
if (start == NULL || end == NULL)
return false;
/* This is called only for finding out if a bb needs a hot patch, so can't
* be used for hotp_only.
*/
ASSERT(!DYNAMO_OPTION(hotp_only));
/* Called during bb building even when there is no hot patch info available. */
if (GLOBAL_VUL_TABLE == NULL)
return false;
/* This is always initialized at startup, so can't be NULL at this point. */
ASSERT(hotp_patch_point_areas != NULL);
if (!own_hot_patch_lock) /* Fix for case 5323. */
d_r_read_lock(&hotp_vul_table_lock);
/* Caller must come in with lock - that is the use today. However, this
* doesn't need the caller to hold the hotp_vul_table_lock; can do so by
* itself. Imposed by fix for case 8780 - excessive holding of hotp lock.
* need to find a better solution (XXX).
*/
ASSERT_OWN_READWRITE_LOCK(true, &hotp_vul_table_lock);
res = vmvector_overlap(hotp_patch_point_areas, start, end);
if (!own_hot_patch_lock)
d_r_read_unlock(&hotp_vul_table_lock);
return res;
}
/* for the given ilist, it will insert the call to hot patch gateway before
* instruction 'where' such that the hot patch corresponding to the given
* 'policy' will be invoked with the mode specified by the policy.
*
* The disassembly of what is injected at each patch point is shown below.
* Note: the disassembly below may change based on the new design.
*
Clean call preparation
Note: this clean call prep shows accessing dcontext directly, i.e., as
in thread private case. In the shared fragments case dcontext will
first be loaded from the TLS. See prepare_for_clean_call() for details.
mov %esp -> dcontext.mcontext.mcontext.xsp
mov dcontext.dstack -> %esp
pushf
pusha ; save app reg. state
push $0x00000000 %esp -> %esp (%esp)
popf %esp (%esp) -> %esp
addr16 mov %fs:0x34 -> %eax ; last error value
push %eax %esp -> %esp (%esp)
Save app state & make call to hotp_gateway
if (SHARED_FRAGMENTS_ENABLED()) {
mov %fs:TLS_DCONTEXT_SLOT -> %eax
mov %eax(DSTACK_OFFSET) -> %eax
} else {
mov dcontext.dstack -> %eax ; locate app reg state on stack
}
sub HOTP_CONTEXT_OFFSET_ON_DSTACK,%eax -> %eax
pusha was done on dr stack, so esp is dr's; get and spill the app's esp
if (SHARED_FRAGMENTS_ENABLED()) {
mov %fs:TLS_DCONTEXT_SLOT -> %ecx
mov %ecx(XSP_OFFSET) -> %ecx
} else {
mov dcontext.mcontext.mcontext.xsp -> %ecx
}
mov %ecx -> [%eax + 0xc] ; 0xc == offsetof(hotp_context_t, xsp)
Note: Don't send func_ptr; security hazard; use indices into hotp_vul_table -
one for vul, set, mod & ppt; this way the gateway can pick out the exact
hotpatch offset from the table which is in read only memory; this also
avoids the need to maintain a hash for hot patch offsets which can be
looked up by hotp_gateway() before doing the hot patch call.
push false; don't have the hotp_vul_table_lock
push $eax ; app_reg_ptr
push ppoint_index
push module_index
push set_index
push vul_index
push num_vuls
push vul_table_ptr
call hotp_gateway()
The hot patch could have changed esp we sent to it via app_reg_ptr. As we
restore esp from dcontext, save app_reg_ptr->xsp in the dcontext.
Fix for case 5594.
app_reg_ptr = dstack - HOTP_CONTEXT_OFFSET_ON_DSTACK
app_esp_p = app_reg_ptr + offsetof(hotp_context_t, xsp)
See clean call above.
if (SHARED_FRAGMENTS_ENABLED()) {
mov %fs:TLS_DCONTEXT_SLOT -> %eax
mov %eax(DSTACK_OFFSET) -> %eax
} else {
mov dcontext.dstack -> %eax
}
mov (%eax-$0x14) -> %eax ; eax = [app_esp_p]
if (SHARED_FRAGMENTS_ENABLED()) {
mov %fs:TLS_DCONTEXT_SLOT -> %ecx
mov $eax -> %ecx(XSP_OFFSET)
} else {
mov %eax -> dcontext.mcontext.mcontext.xsp
}
Clean call cleanup
add $0x1c %esp -> %esp ; pop off the 7 args to hotp_gateway()
pop %esp (%esp) -> %eax %esp
addr16 mov %eax -> %fs:0x34 ; last error value
popa ; restore app reg. state
popf
mov dcontext.mcontext.mcontext.xsp -> %esp
* CAUTION: Any change to this function will affect hotp_change_control_flow().
* What is stored in the app/dr stack by the code generated by this
* routine is used and modified by hotp_change_control_flow().
*/
/* TODO: PR 226888 - make hotp bbs shared - they are enabled to be shared, but
* actually aren't shared yet.
*/
static int
hotp_inject_gateway_call(dcontext_t *dcontext, instrlist_t *ilist, instr_t *where,
const hotp_offset_match_t *match)
{
/* TODO: use a separate stack later on; don't pollute the dr stack;
* for now use dr stack for executing the hot patch code.
*/
/* NOTE: app_reg pointer computation assumes certain behavior from
* dr_prepare_for_call, i.e., first thing is all app registers are
* pushed on to DR stack; eax is scratch at this point.
* TODO: Add asserts here for these.
*/
# define HOTP_CONTEXT_OFFSET_ON_DSTACK sizeof(hotp_context_t)
/* Loads contents of dcontext at offset to reg. For shared fragments it is
* loaded via dc_reg; load dc into dc_reg if it isn't available (!have_dc).
*/
# define GET_FROM_DC_OFFS_TO_REG(offset, reg, have_dc, dc_reg) \
do { \
if (SHARED_FRAGMENTS_ENABLED()) { \
if (!have_dc) \
insert_get_mcontext_base(dcontext, ilist, where, dc_reg); \
MINSERT(ilist, where, \
instr_create_restore_from_dc_via_reg(dcontext, dc_reg, reg, \
offset)); \
} else { \
MINSERT(ilist, where, \
instr_create_restore_from_dcontext(dcontext, reg, offset)); \
} \
} while (0)
/* Using client api to avoid duplicating code. */
/* XXX PR 226036: set hotp_context_t pc field? left as 0 by dr_prepare_for_call */
dr_prepare_for_call(dcontext, ilist, where);
/* DSTACK_OFFSET isn't within the upcontext so if it's separate our use of
* insert_get_mcontext_base() above is incorrect. */
ASSERT_NOT_IMPLEMENTED(!TESTANY(SELFPROT_DCONTEXT, dynamo_options.protect_mask));
/* We push eax as a parameter to the call */
GET_FROM_DC_OFFS_TO_REG(DSTACK_OFFSET, REG_XAX, false /* !have_dc */, REG_XBX);
/* app reg ptr is put in eax */
MINSERT(ilist, where,
INSTR_CREATE_sub(dcontext, opnd_create_reg(REG_XAX),
OPND_CREATE_INT8(HOTP_CONTEXT_OFFSET_ON_DSTACK)));
/* Get the app esp stored in dcontext.mcontext & spill it in the right
* location for the hot patch code.
*/
GET_FROM_DC_OFFS_TO_REG(XSP_OFFSET, REG_XCX, true /* have_dc */, REG_XBX);
MINSERT(ilist, where,
INSTR_CREATE_mov_st(dcontext,
OPND_CREATE_MEM32(REG_XAX, offsetof(hotp_context_t, xsp)),
opnd_create_reg(REG_XCX)));
dr_insert_call(
dcontext, ilist, where, (app_pc)&hotp_gateway, 8,
OPND_CREATE_INTPTR(GLOBAL_VUL_TABLE), OPND_CREATE_INT32(NUM_GLOBAL_VULS),
OPND_CREATE_INT32(match->vul_index), OPND_CREATE_INT32(match->set_index),
OPND_CREATE_INT32(match->module_index), OPND_CREATE_INT32(match->ppoint_index),
/* app reg ptr put in eax above */
opnd_create_reg(REG_XAX), OPND_CREATE_INT32(false));
/* TODO: also, for multiple patch points for one offset, gateway will have
* to take variable arguments, i.e., one set per patch.
*/
/* Copy app esp from context passed to hot patch into mcontext to set up
* for restore. Fix for case 5594.
*/
GET_FROM_DC_OFFS_TO_REG(DSTACK_OFFSET, REG_XAX, false /* !have_dc */, REG_XBX);
IF_X64(ASSERT(CHECK_TRUNCATE_TYPE_int(offsetof(hotp_context_t, xsp))));
IF_X64(ASSERT(CHECK_TRUNCATE_TYPE_int(HOTP_CONTEXT_OFFSET_ON_DSTACK)));
MINSERT(
ilist, where,
INSTR_CREATE_mov_ld(dcontext, opnd_create_reg(REG_XAX),
OPND_CREATE_MEM32(REG_XAX,
(int)offsetof(hotp_context_t, xsp) -
(int)HOTP_CONTEXT_OFFSET_ON_DSTACK)));
if (SHARED_FRAGMENTS_ENABLED()) {
MINSERT(ilist, where,
instr_create_save_to_dc_via_reg(dcontext, REG_XBX, REG_XAX, XSP_OFFSET));
} else {
MINSERT(ilist, where,
instr_create_save_to_dcontext(dcontext, REG_XAX, XSP_OFFSET));
}
dr_cleanup_after_call(dcontext, ilist, where, 0);
return 1; /* TODO: why return anything here? */
}
/* If the given ilist has instructions that are targeted by any vulnerabilities,
* this routine will identify those policies and insert code into the basic block to
* call the hot patch code corresponding to the matching vulnerabilities.
*
* Note: Expand the ilist corresponding to the bb only if a hot patch needs to
* be injected into it; taken care of by the boolean that predicates the
* call to this function.
*/
bool
hotp_inject(dcontext_t *dcontext, instrlist_t *ilist)
{
bool injected_hot_patch = false;
instr_t *instr, *next;
hotp_offset_match_t match = { -1, -1, -1, -1 };
app_pc translation_target = NULL; /* Fix for case 5981. */
bool caller_owns_hotp_lock = self_owns_write_lock(hotp_get_lock());
/* This routine is for injecting hot patches into an ilist, i.e., into the
* fcache. Shouldn't be here for -hotp_only which patches the image.
*/
ASSERT(!DYNAMO_OPTION(hotp_only));
if (!caller_owns_hotp_lock)
d_r_write_lock(&hotp_vul_table_lock); /* Fix for case 5323. */
/* Expand the ilist corresponding to the basic block and for each
* instruction in the ilist, check if one or more injections to
* the gateway should be made and then do so.
*/
instr = instrlist_first_expanded(dcontext, ilist);
while (instr != NULL) {
next = instr_get_next_expanded(dcontext, ilist, instr);
/* TODO: must have way to ensure that all offsets matched for this
* basic block are patched (not missed) and correctly too. but how?
*/
/* TODO: hotp_lookup_patch_addr(), i.e., the second/internal lookup should
* be able to return multiple matching vulnerabilities/ppoints -
* need a new data structure for it.
*/
/* TODO: for now this is just one vul, so no loop is used inside
* this if; must change to handle multiple matching policies, i.e,
* multiple injections; that should handle precedences if offsets are
* the same.
*/
if (hotp_lookup_patch_addr(instr_get_raw_bits(instr), &match,
true /* own hotp_vul_table_lock */)) {
/* The mode better be either protect or detect at this point! */
hotp_policy_mode_t mode = GLOBAL_VUL(match.vul_index).mode;
ASSERT(mode == HOTP_MODE_DETECT || mode == HOTP_MODE_PROTECT);
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "injecting into %s at " PIFX "\n",
GLOBAL_MODULE(match.vul_index, match.set_index, match.module_index)
.sig.pe_name,
GLOBAL_PPOINT(match.vul_index, match.set_index, match.module_index,
match.ppoint_index)
.offset);
/* TODO: assert somewhere that a given vul can't patch the same
* offset twice in a given module. guess this can be done at vul
* table creation time, i.e., during startup or nudge from nodemgr.
*/
/* The translation target for the inserted instructions is set to
* the address of the instruction preceding the one to be patched.
* Otherwise if an app exception happens in this bb
* recreate_app_state_from_ilist() would fail.
* Note: the only problem is if the first instruction in a bb is
* the patchee; in that case we use that address itself
* though the exception handler will complain about not
* being able to create app state. However, it will get the
* right state, so we are fine in release builds. XXX.
* Part of fix for case 5981.
*/
if (translation_target == NULL)
translation_target = instr_get_raw_bits(instr);
instrlist_set_translation_target(ilist, translation_target);
instrlist_set_our_mangling(ilist, true); /* PR 267260 */
hotp_inject_gateway_call(dcontext, ilist, instr, &match);
instrlist_set_translation_target(ilist, NULL);
instrlist_set_our_mangling(ilist, false); /* PR 267260 */
STATS_INC(hotp_num_inject);
injected_hot_patch = true;
if (mode == HOTP_MODE_DETECT)
hotp_set_policy_status(match.vul_index, HOTP_INJECT_DETECT);
else
hotp_set_policy_status(match.vul_index, HOTP_INJECT_PROTECT);
}
translation_target = instr_get_raw_bits(instr);
instr = next;
}
if (!caller_owns_hotp_lock)
d_r_write_unlock(&hotp_vul_table_lock);
return injected_hot_patch;
}
/* For hotp_only, a patch region shouldn't contain any jmp, call, ret or int
* instructions that start and end within it; it is ok if a jmp, a call,
* a ret or a int spans the entire patch region or beyond it. This is to
* ensure that no control flow can come into the middle of a patch region.
* Those valid calls/jmps that can exist in the patch region should only target
* some image address that belongs to the app, not stack or heap ==>
* ==> XXX case 7657: need to relax that to allow 3rd party hookers (and, app
* itself could be targeting heap).
* Also, the patch region shouldn't be already hooked by the core's hooks, i.e.,
* non hotp_only core hooks.
* TODO: strengthen this function; today it checks for what is not allowed and
* allows all else; make it check for what is allow too, i.e., be precise
* because assumptions can break with instruction extensions.
*/
static bool
hotp_only_patch_region_valid(const app_pc addr_to_hook)
{
instr_t *inst;
app_pc pc = addr_to_hook;
app_pc start_pc = addr_to_hook;
dcontext_t *dcontext = get_thread_private_dcontext();
bool res = true;
/* As of today hot patches can only target the .text section in a module. */
DODEBUG({
if (!is_in_code_section(get_module_base(addr_to_hook), addr_to_hook, NULL, NULL))
return false;
});
/* Happens during hotp_init(); thread init happens afterwards so dcontext
* isn't set up.
*/
if (dcontext == NULL)
dcontext = GLOBAL_DCONTEXT;
inst = instr_create(dcontext);
while (pc < addr_to_hook + HOTP_PATCH_REGION_SIZE) {
instr_reset(dcontext, inst);
pc = decode(dcontext, pc, inst);
if (instr_is_cti(inst) || instr_is_interrupt(inst)) {
/* cti is in patch region followed by other instructions in it.
* Shouldn't patch this as control can come into the middle of the
* patch region.
*/
if ((start_pc + instr_length(dcontext, inst)) <
(addr_to_hook + HOTP_PATCH_REGION_SIZE)) {
LOG(GLOBAL, LOG_HOT_PATCHING, 2,
"invalid hotp_only patch point"
" at " PFX "; there is cti inside it\n",
start_pc);
res = false;
goto hotp_only_patch_region_valid_exit;
} else {
/* cti is in the patch region & spans till or beyond the end
* of the patch region, i.e., this region is valid.
*/
if (instr_is_call(inst)) {
/* XXX: Mangling calls in patch regions hasn't been done
* yet. See case 6839.
*/
LOG(GLOBAL, LOG_HOT_PATCHING, 1,
"Warning: not mangling "
"valid call in hotp_only patch region; not supported "
"yet, see case 6839.");
}
if (instr_is_call_direct(inst) || instr_is_ubr(inst) ||
instr_is_cbr(inst)) {
app_pc target = instr_get_branch_target_pc(inst);
/* XXX: core doesn't handle far ctis today, see case 6962;
* when far ctis are handled, this assert can go.
*/
ASSERT(!instr_is_far_abs_cti(inst));
# ifdef WINDOWS
/* Does it overlap with any of the core's hooks?
* Note: native_exec_syscalls don't use the landing pad as
* of now, so we still have to look at the
* interception_buffer. Also, the vmvector_overlap may
* trigger for hotp_only hooks too. Once native_exec
* hooking uses landing pads change this so that the target
* of the landing pad is checked to see if it is in the
* interception buffer. Not a big deal as both result in
* the hooking being aborted - just the log message
* changes.
*/
if (is_part_of_interception(target)) {
LOG(GLOBAL, LOG_HOT_PATCHING, 2,
"invalid hotp_only "
"patch point at " PFX "; it collides with a core hook\n",
start_pc);
res = false;
goto hotp_only_patch_region_valid_exit;
}
# endif /* WINDOWS */
/* Overlaps with any injected hot patch? Case 7998.
* This is not infrequent, some dlls like urlmon or rpcrt4
* have a .orpc section which results in an unmatched page
* protection change, like rw-, r-x, r-x; the last one
* results in double injection, which should be ignored.
* See case 9588 and 9906 where this causes a crash.
* Note: as all hotp_only hooks go through landing pads we
* don't have to check hotp_only_tramp_areas.
*/
if (vmvector_overlap(landing_pad_areas, target, target + 1)) {
# ifdef WINDOWS /* WINDOWS_VERSION_2003 doesn't exist on linux. */
DODEBUG({
const char *reason = "unknown";
if (hotp_globals->ldr_safe_hook_injection) {
reason = "due to loader safety";
} else if (get_os_version() >= WINDOWS_VERSION_2003) {
/* On 2k3 loader lock isn't held during dll
* loading before executing image entry, so we
* can't tell for sure. */
/* XXX case 10636: what about vista? */
reason = "2003; may be due to loader safety";
} else {
ASSERT_NOT_REACHED(); /* Unknown reason. */
}
LOG(GLOBAL, LOG_HOT_PATCHING, 2,
"Blocking double "
"injection at " PFX " in module at " PFX " - %s\n",
start_pc, get_module_base(start_pc), reason);
});
# endif
res = false;
goto hotp_only_patch_region_valid_exit;
}
/* This check concludes it is a 3rd party hook if
* target is not in current image; target may be in another
* image, mapped read-only file, or heap. The first two
* may not be 3rd party hook conflicts (rare). For now, we
* conservatively conclude these to be hook conflicts.
* Note: Whether a hook targets image or heap has no
* bearing on how easily we can interop with it.
* XXX: track patch point from mmap to point of hooking
* to see if it is hooked before concluding hook conflict;
* case 10433.
*/
DODEBUG({
if (!is_in_any_section(get_module_base(start_pc), target, NULL,
NULL)) {
LOG(GLOBAL, LOG_HOT_PATCHING, 2,
"cti in patch region " PFX "; cti target " PFX
" isn't inside image "
"- potential 3rd-party hooker",
start_pc, target);
SYSLOG_INTERNAL_WARNING("Potential 3rd party hook "
"conflict at " PFX,
start_pc);
}
});
/* No app jump should be targeting the core. */
if (is_in_dynamo_dll(target)) {
LOG(GLOBAL, LOG_HOT_PATCHING, 2,
"invalid hotp_only "
"patch point at " PFX "; cti targets dynamorio.dll!\n",
start_pc);
ASSERT_NOT_REACHED();
res = false;
goto hotp_only_patch_region_valid_exit;
}
SYSLOG_INTERNAL_WARNING_ONCE("cti found at hotp point, will chain");
LOG(GLOBAL, LOG_HOT_PATCHING, 2,
"found chainable cti at "
"patch point at " PFX "\n",
start_pc);
}
}
}
start_pc = pc;
}
hotp_only_patch_region_valid_exit:
instr_destroy(dcontext, inst);
return res;
}
static void
patch_cti_tgt(byte *tgt_loc, byte *new_val, bool hot_patch)
{
# ifdef X64
ATOMIC_8BYTE_WRITE(tgt_loc, (int64)new_val, hot_patch);
# else
insert_relative_target(tgt_loc, new_val, hot_patch);
# endif
}
/* Injects one hotp_only patch, i.e., inserts trampoline to execute a hot patch.
*
* XXX: multi-thread safe injection hasn't been implemented; when that is
* implemented this routine will have to assert that all threads in this
* process have stopped. see case 6662.
* Note: injections are done per module, not for the whole policy table,
* so there might be performance issues with stopping and resuming
* all threads for each module to be patched.
*
* XXX: injection currently doesn't check if loader is finished with a module
* before injecting; needs to be done. also, while injecting the loader
* shouldn't be allowed to modify the module. see case 6662.
*
* XXX: patch removal hasn't been implemented yet for hotp_only; when doing so
* trampoline code must be released, hook removed & image processed to
* set it to unmatched. see case 6663.
*/
static void
hotp_only_inject_patch(const hotp_offset_match_t *ppoint_desc,
const thread_record_t **thread_table, const int num_threads)
{
hotp_vul_t *vul;
hotp_set_t *set;
hotp_module_t *module;
hotp_patch_point_t *ppoint, *cur_ppoint;
uint ppoint_idx;
app_pc addr_to_hook, cflow_target;
byte *end;
bool patched = false;
ASSERT(DYNAMO_OPTION(hotp_only));
/* At startup there should be no other thread than this, so thread_table
* won't be valid.
*/
if (num_threads != HOTP_ONLY_NUM_THREADS_AT_INIT) {
ASSERT(ppoint_desc != NULL && thread_table != NULL);
} else {
ASSERT(ppoint_desc != NULL && thread_table == NULL);
}
/* Check if it is safe to patch, i.e., no known threads should be running
* around (of course for the unknown thread this won't help; see
* hotp_init() for the comment about that corner case).
*/
ASSERT_OWN_MUTEX(true, &all_threads_synch_lock);
ASSERT_OWN_MUTEX(true, &thread_initexit_lock);
ASSERT_OWN_READWRITE_LOCK(true, &hotp_vul_table_lock);
vul = &GLOBAL_VUL(ppoint_desc->vul_index);
set = &vul->sets[ppoint_desc->set_index];
module = &set->modules[ppoint_desc->module_index];
cur_ppoint = &module->patch_points[ppoint_desc->ppoint_index];
addr_to_hook = hotp_ppoint_addr(module, cur_ppoint);
/* Can't inject a hot patch if its container dll isn't loaded. This can
* happen if there is a bug in the policy def file or while core was loading
* the dll.
*/
if (vul->hotp_dll_base == NULL) {
SYSLOG_INTERNAL_WARNING("Hot patch dll (%s) hasn't been loaded; "
"aborting hotp_only injection",
vul->hotp_dll);
ASSERT(false);
return;
}
/* If addr_to_hook doesn't conform to the patch region definition, then
* don't inject the patch.
*/
if (!hotp_only_patch_region_valid(addr_to_hook)) {
STATS_INC(hotp_only_aborted_injects);
return;
}
if (cur_ppoint->trampoline != NULL) {
/* XXX case 9148/7657: we can have hookers who chain off our old
* trampoline via a +rwx prot change followed by a +rx change that
* triggers us adding new hooks without removing the old. We go ahead
* and leave that bug in and live with the leak for now since it works
* out better in terms of chaining (o/w we will re-use old trampoline
* buffers that are pointed to by the hooker's chaining, causing
* infinite recursion, incorrect API calls, or worse). Our hook code is
* then called twice, but this can only happen for GBOP (o/w the hash
* wouldn't match) which can handle duplicate checks. We need a
* comprehensive hooker + loader compatibility policy that minimizes
* these types of problems (case 7657).
*/
SYSLOG_INTERNAL_WARNING("patch point " PFX " in module %s being re-patched; "
"old patch leaked",
addr_to_hook, module->sig.pe_name);
ASSERT(cur_ppoint->app_code_copy != NULL);
} else {
ASSERT(cur_ppoint->app_code_copy == NULL);
ASSERT(cur_ppoint->tramp_exit_tgt == NULL);
}
/* Shouldn't be injecting anything that is isn't turned on. */
ASSERT(vul->mode == HOTP_MODE_DETECT || vul->mode == HOTP_MODE_PROTECT);
/* Make sure that patch region size isn't messed up. */
ASSERT(HOTP_PATCH_REGION_SIZE == HOTP_ONLY_PATCH_REGION_SIZE);
cur_ppoint->trampoline = (byte *)special_heap_alloc(hotp_only_tramp_heap);
/* The patch region has been checked for validity by now, so if there are
* other hooks in there smash them. Also, control flow change is
* implemented using AFTER_INTERCEPT_DYNAMIC_DECISION hooking model and
* using AFTER_INTERCEPT_LET_GO_ALT_DYN; the only difference being that the
* alternate target is not provided at hook time because it is unknown till
* hooking is completed. The alternate target is provided after hooking;
* see below in the hook conflict resolution code.
*/
end = hook_text(cur_ppoint->trampoline, addr_to_hook, hotp_only_gateway,
(void *)addr_to_hook,
cur_ppoint->return_addr != 0 ? AFTER_INTERCEPT_DYNAMIC_DECISION
: AFTER_INTERCEPT_LET_GO,
false, /* don't abort if hooked, smash it */
true, /* ignore ctis; they have been checked for already */
&cur_ppoint->app_code_copy,
cur_ppoint->return_addr != 0 ? &cur_ppoint->tramp_exit_tgt : NULL);
/* Did we hook it successfully? */
ASSERT(*addr_to_hook == JMP_REL32_OPCODE);
/* Trampoline code shouldn't overflow the trampoline buffer here. By now
* the damage is already done. In a debug build it is ok, but in a release
* build? XXX: need to make intercept_call() take a buffer length.
*/
ASSERT((end - cur_ppoint->trampoline) <= HOTP_ONLY_TRAMPOLINE_SIZE);
/* The copy of the hooked app code should be within the trampoline. */
ASSERT(HOTP_ONLY_IS_IN_TRAMPOLINE(cur_ppoint, cur_ppoint->app_code_copy));
/* If the current hot patch has a control flow change address then the
* cti that does the control flow change should be inside the trampoline.
*/
ASSERT(cur_ppoint->return_addr == 0 ||
HOTP_ONLY_IS_IN_TRAMPOLINE(cur_ppoint, cur_ppoint->tramp_exit_tgt));
/* Now that the trampoline has been created to our satisfaction, add it to
* the trampoline vector. Note, all thread synch locks & hot patch locks
* must be held before adding anything to the vector.
*/
vmvector_add(hotp_only_tramp_areas, cur_ppoint->trampoline,
cur_ppoint->trampoline + HOTP_ONLY_TRAMPOLINE_SIZE, (void *)cur_ppoint);
if (cur_ppoint->return_addr != 0) {
/* A hot patch can't change control flow to go to the point where it
* is injected; would lead to an infinite loop.
*/
ASSERT(cur_ppoint->return_addr != cur_ppoint->offset);
/* Go through all the patch points in this module, including the
* current one, to see if the current patch point's
* control-flow-change-target is in the middle of any patch region
* that has been hooked by the core; this is to make sure that we
* end up jumping to the copy of the app code in the trampoline as
* opposed to jumping to the hook itself!
*/
for (ppoint_idx = 0; ppoint_idx < module->num_patch_points; ppoint_idx++) {
ppoint = &module->patch_points[ppoint_idx];
/* If a ppoint hasn't been patched yet, don't do try to a resolve
* control flow change conflict targeting it! If a ppoint has
* been patched, is cur_ppoint inside it? If so, resolve conflict.
*/
if (ppoint->trampoline != NULL &&
HOTP_ONLY_IS_IN_PATCH_REGION(ppoint, cur_ppoint->return_addr)) {
/* If ppoint has been injected, then its app_code_copy must
* point to the copy of the app code that was overwritten by
* the hook.
*/
ASSERT(HOTP_ONLY_IS_IN_TRAMPOLINE(ppoint, ppoint->app_code_copy));
/* Without multiple patch points at the same offset, a control
* flow change target can collide with only one patch region.
*/
ASSERT(!patched);
/* Control flow transfer is going to the middle of another
* hot patch's patch region; one which has been injected.
* So fix the cur_ppoint trampoline's exit cti to target the
* app code copy stored in the target hot patch's trampoline as
* opposed to actual image.
*/
cflow_target =
ppoint->app_code_copy + (cur_ppoint->return_addr - ppoint->offset);
ASSERT(HOTP_ONLY_IS_IN_TRAMPOLINE(ppoint, cflow_target));
patch_cti_tgt(cur_ppoint->tramp_exit_tgt, cflow_target, false);
patched = true;
# ifndef DEBUG
/* Cycle through all patches even if patched for debug builds;
* it helps to catch multiple ppoints in the same offset. In
* release builds, this is an inefficiency, so just break.
*/
break;
# endif
STATS_INC(hotp_only_cflow_collision);
}
}
/* Control flow change is to a point inside the module which isn't a
* patch point.
*/
if (!patched) {
cflow_target = module->base_address + cur_ppoint->return_addr;
patch_cti_tgt(cur_ppoint->tramp_exit_tgt, cflow_target, false);
}
}
/* Now, check in the current module, if any other injected patch point's
* control-flow-change target is the current patch point's region; if so
* make it jump to the app_code_copy in the trampoline buffer of the
* current ppoint.
*/
for (ppoint_idx = 0; ppoint_idx < module->num_patch_points; ppoint_idx++) {
ppoint = &module->patch_points[ppoint_idx];
/* No point in checking the current patch point with itself; of course
* there will be a collision.
*/
if (ppoint != cur_ppoint) {
/* If ppoint hasn't been injected, nothing to do. If it has been &
* its return_addr collides with cur_ppoint's patch region,
* then resolve conflict, i.e., change control flow to the copy
* of app code inside cur_ppoint's trampoline.
*/
if (ppoint->trampoline != NULL &&
HOTP_ONLY_IS_IN_PATCH_REGION(cur_ppoint, ppoint->return_addr)) {
ASSERT(HOTP_ONLY_IS_IN_TRAMPOLINE(ppoint, ppoint->app_code_copy));
ASSERT(HOTP_ONLY_IS_IN_TRAMPOLINE(ppoint, ppoint->tramp_exit_tgt));
cflow_target = cur_ppoint->app_code_copy +
(ppoint->return_addr - cur_ppoint->offset);
ASSERT(HOTP_ONLY_IS_IN_TRAMPOLINE(cur_ppoint, cflow_target));
patch_cti_tgt(ppoint->tramp_exit_tgt, cflow_target, false);
STATS_INC(hotp_only_cflow_collision);
}
}
}
# ifdef WINDOWS
/* If any suspended app thread is in the middle of the current patch point
* then it needs to be relocated, i.e., its eip needs to be changed to point
* to the correct offset in the app_code_copy in the trampoline.
*/
if (num_threads != HOTP_ONLY_NUM_THREADS_AT_INIT) {
int i;
bool res;
app_pc eip;
CONTEXT cxt;
thread_id_t my_tid = d_r_get_thread_id();
for (i = 0; i < num_threads; i++) {
/* Skip the current thread; nudge thread's Eip isn't relevant. */
if (my_tid == thread_table[i]->id)
continue;
/* App thread can't be in the core holding a lock when suspended. */
ASSERT(thread_owns_no_locks(thread_table[i]->dcontext));
cxt.ContextFlags = CONTEXT_FULL; /* PR 264138: don't need xmm regs */
res = thread_get_context((thread_record_t *)thread_table[i], &cxt);
ASSERT(res);
eip = (app_pc)cxt.CXT_XIP;
/* 3 conditions have to be met to relocate an app thread during
* hotp_only patching.
* 1. thread's eip should be greater than the module base of the
* current ppoint; if not, negative offsets will result which
* can cause wrap arounds in the HOTP_ONLY_IS_IN_PATCH_REGION
* check which uses app_rva_t (size_t).
* 2. if eip is at the start of the patch region, don't relocate
* it; just let it go to the trampoline. Fixes a security
* issue: a live process which is blocked on a system call can
* be patched right after the syscall so that a vulnerability
* in the results can be caught; if relocated, the first time,
* the hotpatch won't execute, just the app code copy, thereby
* letting the attack slip. Rare & theoritical (because we
* don't allow returns inside the ppoint & because it is hard
* the attack has to be timed to be after the patch but
* before it is executed) hole.
* 3. eip should be inside the patch region defined by cur_ppoint.
*/
if (eip > module->base_address && eip != addr_to_hook &&
HOTP_ONLY_IS_IN_PATCH_REGION(cur_ppoint,
(app_rva_t)(eip - module->base_address))) {
/* XXX: this is one place that may need work if we mangle
* cti_short in the patch region; see case 6839.
*/
cxt.CXT_XIP =
(ptr_uint_t)(cur_ppoint->app_code_copy +
(eip - (module->base_address + cur_ppoint->offset)));
res = thread_set_context((thread_record_t *)thread_table[i], &cxt);
ASSERT(res);
}
}
}
# endif /* WINDOWS */
STATS_INC(hotp_only_num_inject);
if (vul->mode == HOTP_MODE_DETECT)
hotp_set_policy_status(ppoint_desc->vul_index, HOTP_INJECT_DETECT);
else
hotp_set_policy_status(ppoint_desc->vul_index, HOTP_INJECT_PROTECT);
}
/* Does mp safe removal of one hotp_only patch. At the point of suspension,
* each thread shouldn't be in all of the following: dr, hotp_dll and dr_stack.
*/
static void
hotp_only_remove_patch(dcontext_t *dcontext, const hotp_module_t *module,
hotp_patch_point_t *cur_ppoint)
{
bool res;
app_pc addr_to_unhook;
DEBUG_DECLARE(char ppoint_content[HOTP_ONLY_PATCH_REGION_SIZE]);
ASSERT(DYNAMO_OPTION(hotp_only));
/* Are we at a mp-safe spot to remove the patches? */
ASSERT_OWN_MUTEX(true, &all_threads_synch_lock);
ASSERT_OWN_MUTEX(true, &thread_initexit_lock);
ASSERT_OWN_READWRITE_LOCK(true, &hotp_vul_table_lock);
addr_to_unhook = hotp_ppoint_addr(module, cur_ppoint);
/* Is there a hook at this place? */
ASSERT(*addr_to_unhook == JMP_REL32_OPCODE);
/* Is there a valid trampoline? */
ASSERT(cur_ppoint->trampoline != NULL);
ASSERT(cur_ppoint->app_code_copy != NULL);
/* Save the 5 original app code bytes by getting it from the trampoline
* (today we store those at the start of the trampoline). Check that
* those bytes match after unhooking.
*/
ASSERT(HOTP_PATCH_REGION_SIZE == HOTP_ONLY_PATCH_REGION_SIZE);
DODEBUG({ memcpy(ppoint_content, cur_ppoint->trampoline, HOTP_PATCH_REGION_SIZE); });
unhook_text(cur_ppoint->trampoline, addr_to_unhook);
ASSERT(!memcmp(ppoint_content, addr_to_unhook, HOTP_PATCH_REGION_SIZE));
/* Don't release the trampoline, just leak it, i.e., don't call
* special_heap_free. This is how we handle the interop and detach
* problems created by hotp & 3rd-party hooks colliding. Not elegant or
* memory efficient, but will handle the cases of the 3rd party reading our
* hook before hooking and/or leaving the page marked rwx. See cases
* 9906, 9588, 9593, 9148 & 9157.
* XXX: have a better mechanism to resolve hook conflict issues; currently
* only a minimalist solution is in place; case 7657, case 10433.
* XXX: make leaking selective, i.e., don't leak all trampoline, leak
* only the ones that collide with 3rd party hooks - need to do
* some bookkeepping; not a big issue, but about 20k to 50k can be
* lost for each process otherwise, case 10433.
*/
/* The ifdef mess in the next few lines is to handle the leak for case 9593. */
# ifdef DEBUG
# ifdef HEAP_ACCOUNTING
hotp_only_tramp_bytes_leaked += HOTP_ONLY_TRAMPOLINE_SIZE;
# endif
# endif
/* Tramp heap is freed before memory leak is checked, so cache the value.
* hotp_only_tramp_heap_cache also tracks if there was patch removal. */
if (hotp_only_tramp_heap_cache == NULL) {
SELF_UNPROTECT_DATASEC(DATASEC_RARELY_PROT);
hotp_only_tramp_heap_cache = hotp_only_tramp_heap;
SELF_PROTECT_DATASEC(DATASEC_RARELY_PROT);
}
/* Note: as we aren't freeing the trampoline, we have to nop it else we can
* have bad consequences like a conflicting 3rd party hook jumping to our
* trampoline after we detach! See case 9593.
* This is done by bypassing the whole trampoline and jumping to the part
* that executes the original app code and returns to the address after
* the hook point. */
insert_jmp_at_tramp_entry(dcontext, cur_ppoint->trampoline,
cur_ppoint->app_code_copy);
/* Note, all thread synch locks & hot patch locks must be held before
* removing anything from the vector.
*/
res = vmvector_remove(hotp_only_tramp_areas, cur_ppoint->trampoline,
cur_ppoint->trampoline + HOTP_ONLY_TRAMPOLINE_SIZE);
ASSERT(res);
/* Today for hotp_only all patches in a module are applied and removed in
* one shot, and control flow change doesn't go across modules, so there
* is no need to patch no tramp_exit_tgt (to make sure that control flow
* change requested is not affected) as a result of patch removal
* (remember that all threads are suspended at outside of any hot patches
* during the patch removal process). If in future we allow control flow
* change to go across modules, then we will need to go through all
* modules & their patch points to fix the tramp_exit_tgt.
*/
cur_ppoint->trampoline = NULL;
cur_ppoint->tramp_exit_tgt = NULL;
cur_ppoint->app_code_copy = NULL;
}
/* Returns true if the eip is inside any hotp_only trampoline. */
bool
hotp_only_in_tramp(const app_pc pc)
{
/* Only after successfully stopping all threads will hotp_only_tramp_areas
* vector will be written to. This means that during synching when each
* thread is suspended, where this function is called, there should be no
* one updating the hotp_only_tramp_areas vector.
*/
if (DYNAMO_OPTION(hotp_only)) {
ASSERT(!WRITE_LOCK_HELD(&hotp_only_tramp_areas->lock));
return vmvector_overlap(hotp_only_tramp_areas, pc, pc + 1);
} else
return false; /* check is moot if there is no trampoline */
}
/* This routine is used to remove hotp_only patches on a detach.
* Note: Though hotp_exit gets called by detach, the removal of patches can't
* be done there because the synch locks and thread data structures won't be
* available at that point. Hence the patch removal has to be done earlier
* inside detach.
*/
void
hotp_only_detach_helper(void)
{
/* Can't be removing hotp_only patches when hotp_only mode isn't on. Though
* we assert, it is safe to doing nothing in release builds and just return.
*/
ASSERT(DYNAMO_OPTION(hotp_only));
if (!DYNAMO_OPTION(hotp_only))
return;
/* Thread synch locks must be held before removing. */
ASSERT_OWN_MUTEX(true, &all_threads_synch_lock);
ASSERT_OWN_MUTEX(true, &thread_initexit_lock);
d_r_write_lock(&hotp_vul_table_lock);
hotp_remove_hot_patches(GLOBAL_VUL_TABLE, NUM_GLOBAL_VULS, true, NULL);
d_r_write_unlock(&hotp_vul_table_lock);
}
/* This function is used to handle loader safe injection for hotp_only mode.
* This is done by removing patches in regions the loader wants to write to
* and reinjecting them afterwards; done by monitoring memory protection
* changes made.
*
* XXX case 9148: this causes problems with hookers who read prior to marking
* +w and thus chain with our old trampoline that we are about to remove here!
* That's why we don't call here for +rwx changes, where we live with a leak on
* the +rx change (case 9148), which is better than hookers who mark +rw and can
* end up with infinite recursions or wrong API calls. We need a better
* approach to handling both loader and hooker interop (case 7657).
*
* Note: all patches in a module come out, not just the page in question because
* if a non-loader-agent changes the image and messes up our hash checks, then
* we wouldn't be able to reinsert any into that page, leaving a multiple ppoint
* policy in an inconsistent state or a ppoint in protect mode unprotected.
* Another reason for pulling out all is module atomicity; set atomicity will
* involve removing patches from other modules too!
*
* Note: we don't handle with some one trying to change memory protection
* across two modules in with a single syscall; don't think it is allowed.
*/
void
hotp_only_mem_prot_change(const app_pc start, const size_t size, const bool remove,
const bool inject)
{
app_pc base;
bool needs_processing = false;
int num_threads = 0;
thread_record_t **thread_table = NULL;
# ifdef WINDOWS
DEBUG_DECLARE(bool ok;)
# endif
/* For hotp_only, for regular mode, vmarea tracking will flush the
* necessary fragments.
*/
ASSERT(DYNAMO_OPTION(hotp_only));
ASSERT(start != NULL && size > 0);
ASSERT((remove == true || remove == false) && (inject == true || inject == false));
ASSERT(remove != inject); /* One and only one must be true. */
if (remove == inject) /* Defensively just ignore. */
return;
base = get_module_base(start);
/* If base doesn't belong to any module; ignore. We don't hot patch DGC. */
if (base == NULL)
return;
/* The end of the region better be in the image! */
ASSERT(base == get_module_base(base + size));
# ifdef WINDOWS
DODEBUG({
if (get_loader_lock_owner() != d_r_get_thread_id()) {
LOG(GLOBAL, LOG_HOT_PATCHING, 1,
"Warning: Loader lock not held "
"during image memory protection change; possible incompatible "
"hooker or w2k3 loader.");
}
});
# endif
/* Inefficient check to see if this module has been matched for hot
* patching. hotp_process_image() is needed only when loading or
* unloading a dll, not here, which is post module loading.
* XXX: Use vmvector_overlap check on loaded_module_areas after
* integrating it with hotp. Optimization.
*/
hotp_process_image(base, inject ? true : false, false, true, &needs_processing, NULL,
0);
if (!needs_processing) { /* Ignore if it isn't a vulnerable module. */
LOG(THREAD_GET, LOG_HOT_PATCHING, 2,
"hotp_only_mem_prot_change: no work to be done for base " PFX "\n", base);
return;
}
# ifdef WINDOWS
/* Ok, let's suspend all threads and do the injection/removal. */
DEBUG_DECLARE(ok =)
synch_with_all_threads(THREAD_SYNCH_SUSPENDED, &thread_table, &num_threads,
/* Case 6821: other synch-all-thread uses that
* only care about threads carrying fcache
* state can ignore us
*/
THREAD_SYNCH_NO_LOCKS_NO_XFER,
/* if we fail to suspend a thread (e.g., privilege
* problems) ignore it. XXX: retry instead? */
THREAD_SYNCH_SUSPEND_FAILURE_IGNORE);
ASSERT(ok);
# endif
d_r_write_lock(&hotp_vul_table_lock);
/* Using hotp_process_image to inject is inefficient because it goes
* through the whole vul table.
* XXX: Optimization: write hotp_only_inject_patches() which should use
* hotp_patch_point_areas; use that to do the injection here.
*/
if (inject) {
LOG(THREAD_GET, LOG_HOT_PATCHING, 1,
"hotp_only_mem_prot_change: injecting for base " PFX "\n", base);
DODEBUG(hotp_globals->ldr_safe_hook_injection = true;); /* Case 7998. */
DODEBUG(hotp_globals->ldr_safe_hook_removal = false;); /* Case 7832. */
hotp_process_image_helper(base, true, true, false, NULL,
(const thread_record_t **)thread_table, num_threads,
true, NULL);
DODEBUG(hotp_globals->ldr_safe_hook_injection = false;);
/* Similarly, hotp_remove_patches_from_module() is inefficient too.
* XXX: using loaded_module_areas in that routine.
*/
} else if (remove) {
LOG(THREAD_GET, LOG_HOT_PATCHING, 1,
"hotp_only_mem_prot_change: removing for base " PFX "\n", base);
hotp_remove_patches_from_module(GLOBAL_VUL_TABLE, NUM_GLOBAL_VULS, true, base,
NULL);
/* Used to detect double removal while handling loader-safety. */
DODEBUG(hotp_globals->ldr_safe_hook_removal = true;); /* Case 7832. */
}
d_r_write_unlock(&hotp_vul_table_lock);
# ifdef WINDOWS
end_synch_with_all_threads(thread_table, num_threads, true /*resume*/);
# endif
}
/* This is the routine that will serve as the entry point into the core for
* executing hot patches in the -hotp_only mode.
* XXX: for now, dr stack is used to execute the hot patch; later on a
* separate stack should be used.
*/
after_intercept_action_t
hotp_only_gateway(app_state_at_intercept_t *state)
{
hotp_offset_match_t match;
hotp_context_t cxt;
app_pc hook_addr = (app_pc)state->callee_arg;
after_intercept_action_t res = AFTER_INTERCEPT_LET_GO;
d_r_read_lock(&hotp_vul_table_lock);
ASSERT(DYNAMO_OPTION(hotp_only));
/* Callee_arg contains the application eip to be patched. It better be
* inside a code region.
*/
ASSERT(is_in_code_section(get_module_base(hook_addr), hook_addr, NULL, NULL));
/* Note: for -hotp_only vulnerability table access during hot patch
* execution is indirect, i.e., we do a lookup. For hot patches in the
* code cache, this information is embedded in the injected code.
*/
if (hotp_only_lookup_patch_addr(hook_addr, &match)) {
cxt = state->mc;
res = hotp_gateway(GLOBAL_VUL_TABLE, NUM_GLOBAL_VULS, match.vul_index,
match.set_index, match.module_index, match.ppoint_index, &cxt,
true /* have lock */);
/* The hot patch could have modified app state as part of the fix, so
* copy it back.
*/
state->mc = cxt;
} else {
/* If we reached here, there was a hot patch that was injected that no
* longer matches, i.e., there is no matching definition. Could be
* because the mode was changed, the module got unloaded or new defs
* came in, etc. With mp-safe hotp_only patch injection, vulnerability/
* policy data changes are preceded by removal of all injected patches;
* and patch removal guarantees that no patch will be executing.
* This means that offset lookup should always succeed in this routine.
*/
ASSERT_NOT_REACHED();
}
d_r_read_unlock(&hotp_vul_table_lock);
return res;
}
/* TODO: for multiple patch points, need to pass the number of patch points;
* preferably as the first argument.
* TODO: this routine calls dr routines, i.e., switches to dr from the fcache.
* the switching involves protections changes (ENTER_DR_HOOK); however,
* the clean call mechanism used to reach here doesn't call the hook!
* also, there are assumptions in dr about locks being held across fcache
* - Derek raised these issues as some that came up during client
* interface design; he also raised some interesing points about
* generating control flow change code rather than doing it in C.
* - Derek also mentioned if the gateway was called from C code within
* dr, then things should be fine; in other words, bail out of the
* code cache for bb that need hot patching and execute the gateway
* from within dr - this is the model we will be switching to in
* the immediate future.
*/
static after_intercept_action_t
hotp_gateway(const hotp_vul_t *vul_tab, const uint num_vuls, const uint vul_index,
const uint set_index, const uint module_index, const uint ppoint_index,
hotp_context_t *app_reg_ptr, const bool own_hot_patch_lock)
{
/* XXX: racy access here; getting lock may be expensive; even if that is
* ok, must make sure that no one will come in here and wait on the
* lock while a hot patch flush happens due to a nudge - deadlock.
* also, before executing each hot patch, it must be verfied
* that it still is valid because it could have been changed by a
* nudge; see case 5052.
* looks like the whole hot patch execution should be covered by a
* lock - same hotp_vul_table lock or a new one? new one I think.
* Derek: let flush worry about invalidation; just grab locks for
* table lookup or stats update.
* see case 5521.
*/
hotp_policy_mode_t mode;
hotp_type_t hotp_type;
app_rva_t detector_offset, protector_offset;
hotp_func_t detector_fn, protector_fn = NULL;
hotp_exec_status_t exec_status, temp;
hotp_offset_match_t ppoint = { vul_index, set_index, module_index, ppoint_index };
bool dump_excpt_info, dump_error_info;
after_intercept_action_t res = AFTER_INTERCEPT_LET_GO;
/* XXX: till hotp interface is expanded to send arguments to detectors
* and protectors, this spill is the simplest way to send/receive args.
* xref case 6804.
*/
reg_t gbop_eax_spill = 0, gbop_edx_spill = 0;
app_pc gbop_bad_addr = NULL;
app_pc ppoint_addr; /* Fix for case 6054. Exposed for gbop. */
DOCHECK(1, {
dcontext_t *dcontext = get_thread_private_dcontext();
ASSERT_CURIOSITY(dcontext != NULL && dcontext != GLOBAL_DCONTEXT &&
"unknown thread");
/* App esp should neither be on dr stack nor on d_r_initstack; see case 7058.
* TODO: when the hot patch interface expands to having eip, assert that
* the eip isn't inside dr.
*/
ASSERT(!is_on_dstack(dcontext, (byte *)app_reg_ptr->xsp) &&
!is_on_initstack((byte *)app_reg_ptr->xsp));
});
ppoint_addr = hotp_ppoint_addr(
&MODULE(vul_tab, vul_index, set_index, module_index),
&PPOINT(vul_tab, vul_index, set_index, module_index, ppoint_index));
/* If we change this to be invoked from d_r_dispatch, should remove this.
* Note that we assume that hotp_only, which is invoked from
* interception code that has its own enter hook embedded, will
* not call any of these hooks -- else we do a double-enter here and
* the exit via hotp_change_control_flow() results in unprotected .data!
*/
ENTERING_DR();
if (!own_hot_patch_lock) {
/* Note: for regular hot patches (!hotp_only) vulnerability table
* access during execution isn't via a lookup and all the old tables
* are alive, so we don't need to grab the lock here; if we do an
* indirect access then we need it. It is left in there for safety.
*/
d_r_read_lock(&hotp_vul_table_lock); /* Part of fix for case 5521. */
} else
ASSERT_OWN_READ_LOCK(true, &hotp_vul_table_lock);
/* Check the validity of the input indices before using them. The injection
* routine should be generating code to send the right values. These
* asserts will trigger if either the injected code is messed up or a nudge
* resulted in a vulnerability change that didn't have a corresponding
* flush of injected bbs/traces. One other possibility is that while in
* this function the vulnerability table changed due to a nudge - this can't
* happen because a nudge would result in a flush, which would wait for all
* threads to come out of the cache, thus out of this function before
* modifying the table.
*/
ASSERT(vul_index >= 0 && vul_index < num_vuls);
ASSERT(set_index >= 0 && set_index < VUL(vul_tab, vul_index).num_sets);
ASSERT(module_index >= 0 &&
module_index < SET(vul_tab, vul_index, set_index).num_modules);
ASSERT(ppoint_index >= 0 &&
ppoint_index <
MODULE(vul_tab, vul_index, set_index, module_index).num_patch_points);
mode = VUL(vul_tab, vul_index).mode; /* Racy; see assert comments above. */
hotp_type = VUL(vul_tab, vul_index).type;
/* For hotp_only mode control can't reach here if the mode is off because
* in order to change modes all patches are removed first. However, for
* regular hot patching, patch removal (flushing) is done after mode change
* and outside the scope of the hotp_vul_table_lock, so control can be in
* the gateway with the mode set to off, but only for one execution per
* thread because the fragment has been unlinked by the flush and
* scheduled for deletion, so there is no entry to it.
*/
DODEBUG({
if (mode == HOTP_MODE_OFF) {
ASSERT(!DYNAMO_OPTION(hotp_only));
STATS_INC(hotp_exec_mode_off);
} else {
ASSERT(mode == HOTP_MODE_DETECT || mode == HOTP_MODE_PROTECT);
}
});
/* The hot patch dll specified by the vulnerability better be loaded by
* this point. Unloaded hot patch dlls will result in the vulnerability
* being deactivated, so we should never reach this point for such
* vulnerabilities.
*/
ASSERT(VUL(vul_tab, vul_index).hotp_dll_base != NULL);
detector_offset =
PPOINT(vul_tab, vul_index, set_index, module_index, ppoint_index).detector_fn;
/* TODO: make the assertion range tigher by using the actual size of the
* text section of the hot patch dll.
*/
ASSERT((detector_offset >= MIN_DETECTOR_OFFSET &&
detector_offset <= MAX_DETECTOR_OFFSET) ||
TESTALL(HOTP_TYPE_PROBE, hotp_type)); /* no detector for probes */
protector_offset =
PPOINT(vul_tab, vul_index, set_index, module_index, ppoint_index).protector_fn;
ASSERT(protector_offset >= MIN_PROTECTOR_OFFSET &&
protector_offset <= MAX_PROTECTOR_OFFSET);
/* Compute the hot patch function addresses with the hot patch dll base. */
detector_fn = (hotp_func_t)((ptr_uint_t)VUL(vul_tab, vul_index).hotp_dll_base +
detector_offset);
protector_fn = (hotp_func_t)((ptr_uint_t)VUL(vul_tab, vul_index).hotp_dll_base +
protector_offset);
ASSERT(detector_fn != protector_fn); /* can't be the same code! */
LOG(GLOBAL, LOG_HOT_PATCHING, 2, "Invoking detector for vulnerability %s\n",
VUL(vul_tab, vul_index).vul_id);
LOG(GLOBAL, LOG_HOT_PATCHING, 4, "Register state sent to detector\n");
DOLOG(4, LOG_HOT_PATCHING, { hotp_dump_reg_state(app_reg_ptr, ppoint_addr, 4); });
/* gbop hooks need to know current pc & will return the bad return address
* if it is faulty. As hotp_context_t doesn't have eip as of now, we pass
* it via edx. eax is used to get the return value.
* TODO: make this a function & move it to the gbop section
* XXX PR 226036: hotp_context_t does have eip now, use it!
*/
if (TESTALL(HOTP_TYPE_GBOP_HOOK, hotp_type)) {
# ifdef GBOP
ASSERT(DYNAMO_OPTION(gbop) && DYNAMO_OPTION(hotp_only));
# endif
gbop_eax_spill = APP_XAX(app_reg_ptr);
gbop_edx_spill = APP_XDX(app_reg_ptr);
APP_XDX(app_reg_ptr) = (reg_t)ppoint_addr;
} else {
/* A hot patch can be only one type. */
ASSERT(TESTANY(HOTP_TYPE_HOT_PATCH, hotp_type) ^
TESTANY(HOTP_TYPE_PROBE, hotp_type));
}
/* Under the current design, a detector will always be called; a protector
* will be called only if the mode says so.
*
* Forensics dumped for hot patch exceptions and errors are done so once
* for each vulnerability; otherwise we could flood the machine. Events
* are logged every time; for errors, this is predicated by the patch
* returning HOTP_EXEC_LOG_EVENT. Cores are dumped only if the
* dumpcore mask is set; for exceptions, it is done every time and
* for errors it is done once, if the mask is set (because for exceptions
* it hard to convey the "once only" information to the exception handler).
* We use num_{aborted,detector_error,protector_error} as booleans to
* control this.
* TODO: area to revisit when we work on information throttling.
*/
dump_excpt_info = (VUL(vul_tab, vul_index).info->num_aborted == 0);
dump_error_info = (VUL(vul_tab, vul_index).info->num_detector_error == 0);
if (TESTALL(HOTP_TYPE_PROBE, hotp_type)) {
/* No detectors for probes. This status means execute the protector. */
exec_status = HOTP_EXEC_EXPLOIT_DETECTED;
} else {
/* A hot patch can be only one type. */
ASSERT(TESTANY(HOTP_TYPE_HOT_PATCH, hotp_type) ^
TESTANY(HOTP_TYPE_GBOP_HOOK, hotp_type));
exec_status = hotp_execute_patch(detector_fn, app_reg_ptr, mode, dump_excpt_info,
dump_error_info);
LOG(GLOBAL, LOG_HOT_PATCHING, 3, "Detector finished for vulnerability %s\n",
VUL(vul_tab, vul_index).vul_id);
STATS_INC(hotp_num_det_exec);
hotp_update_vul_stats(exec_status, vul_index);
}
temp = exec_status & ~HOTP_EXEC_LOG_EVENT;
ASSERT(temp == HOTP_EXEC_EXPLOIT_DETECTED || temp == HOTP_EXEC_EXPLOIT_NOT_DETECTED ||
temp == HOTP_EXEC_DETECTOR_ERROR || temp == HOTP_EXEC_ABORTED);
/* Restore eax & edx spilled for executing gbop remediators.
* TODO: make this a function & move it to the gbop section
*/
if (TESTALL(HOTP_TYPE_GBOP_HOOK, hotp_type)) {
# ifdef GBOP
ASSERT(DYNAMO_OPTION(gbop) && DYNAMO_OPTION(hotp_only));
# endif
if (temp == HOTP_EXEC_EXPLOIT_DETECTED)
gbop_bad_addr = (app_pc)APP_XAX(app_reg_ptr);
APP_XAX(app_reg_ptr) = gbop_eax_spill;
APP_XDX(app_reg_ptr) = gbop_edx_spill;
}
if (temp == HOTP_EXEC_ABORTED || temp == HOTP_EXEC_DETECTOR_ERROR)
goto hotp_gateway_ret;
/* If the patch asked for violation notification, do so only if its mode
* is set to detect. For protect mode, the protector will report the
* violation if asked, so don't worry about it. The exception here is for
* gbop hooks, which currently always run in protect mode, which need to
* honor -detect_mode, in which case we report the violation right here.
* Note: In this case, the gbop protector won't get executed even though
* its mode is set to protect. See below.
*/
if (TESTANY(exec_status, HOTP_EXEC_LOG_EVENT) &&
((mode == HOTP_MODE_DETECT) ||
(TESTALL(HOTP_TYPE_GBOP_HOOK, hotp_type)
# ifdef PROGRAM_SHEPHERDING
&& DYNAMO_OPTION(detect_mode)
# endif
))) {
hotp_event_notify(exec_status, false, &ppoint, gbop_bad_addr, app_reg_ptr);
}
/* The protector should be invoked only if an exploit was detected and the
* mode was set to protect.
* In the case of gbop hooks, -detect_mode shouldn't invoke the protector.
* Note: Unlike hot patches, gbop hooks must conform to core reporting and
* remediation options. As of today hot patch actions are specified by the
* patch {writer}. There are plans to have an override, case 8095.
*/
if (TESTALL(HOTP_TYPE_GBOP_HOOK, hotp_type)
# ifdef PROGRAM_SHEPHERDING
&& DYNAMO_OPTION(detect_mode)
# endif
)
goto hotp_gateway_ret;
if (mode == HOTP_MODE_PROTECT && temp == HOTP_EXEC_EXPLOIT_DETECTED) {
LOG(GLOBAL, LOG_HOT_PATCHING, 2, "Invoking protector for vulnerability %s\n",
VUL(vul_tab, vul_index).vul_id);
LOG(GLOBAL, LOG_HOT_PATCHING, 6, "Register state sent to protector\n");
DOLOG(6, LOG_HOT_PATCHING, { hotp_dump_reg_state(app_reg_ptr, ppoint_addr, 6); });
/* See detector execution comments above for details about hot patch
* error & exception handling.
* TODO: area to revisit when we work on information throttling.
*/
dump_error_info = (VUL(vul_tab, vul_index).info->num_protector_error == 0);
exec_status = hotp_execute_patch(protector_fn, app_reg_ptr, mode, dump_excpt_info,
dump_error_info);
/* TODO: probes have no return codes defined. PR 229881. */
if (TESTALL(HOTP_TYPE_PROBE, hotp_type)) {
exec_status = HOTP_EXEC_EXPLOIT_PROTECTED;
}
temp = exec_status & ~HOTP_EXEC_LOG_EVENT;
ASSERT(temp == HOTP_EXEC_EXPLOIT_PROTECTED ||
temp == HOTP_EXEC_EXPLOIT_NOT_PROTECTED ||
temp == HOTP_EXEC_EXPLOIT_KILL_THREAD ||
temp == HOTP_EXEC_EXPLOIT_KILL_PROCESS ||
temp == HOTP_EXEC_EXPLOIT_RAISE_EXCEPTION ||
temp == HOTP_EXEC_CHANGE_CONTROL_FLOW ||
temp == HOTP_EXEC_PROTECTOR_ERROR || temp == HOTP_EXEC_ABORTED);
LOG(GLOBAL, LOG_HOT_PATCHING, 4, "Register state after protector\n");
DOLOG(4, LOG_HOT_PATCHING, { hotp_dump_reg_state(app_reg_ptr, ppoint_addr, 4); });
LOG(GLOBAL, LOG_HOT_PATCHING, 3, "Protector finished for vulnerability %s\n",
VUL(vul_tab, vul_index).vul_id);
STATS_INC(hotp_num_prot_exec);
hotp_update_vul_stats(exec_status, vul_index);
if (temp == HOTP_EXEC_ABORTED || temp == HOTP_EXEC_PROTECTOR_ERROR)
goto hotp_gateway_ret;
/* Which one comes first, esp with kill/raise & cflow change? */
/* Raise an event only if requested by the protector. */
if (TESTANY(exec_status, HOTP_EXEC_LOG_EVENT)) {
hotp_event_notify(exec_status, true, &ppoint, gbop_bad_addr, app_reg_ptr);
}
if (TESTANY(exec_status, HOTP_EXEC_CHANGE_CONTROL_FLOW)) {
app_rva_t return_rva =
PPOINT(vul_tab, vul_index, set_index, module_index, ppoint_index)
.return_addr;
app_pc module_base =
MODULE(vul_tab, vul_index, set_index, module_index).base_address;
/* gbop hooks shouldn't be changing control flow. */
ASSERT(!TESTALL(HOTP_TYPE_GBOP_HOOK, hotp_type));
/* hotp_only control flow change is implemented using the alt exit
* feature in our intercept_call() mechanism.
*/
if (DYNAMO_OPTION(hotp_only)) {
res = AFTER_INTERCEPT_LET_GO_ALT_DYN;
goto hotp_gateway_ret;
}
/* If control flow change is requested by a protector, then the
* offset to which the control should go to shouldn't be zero
* and the dll should be in memory!
*/
ASSERT(return_rva != 0 && module_base != NULL);
LOG(GLOBAL, LOG_HOT_PATCHING, 1,
"Control flow change requested by "
"vulnerability %s\n",
VUL(vul_tab, vul_index).vul_id);
/* Release the lock because control flow change won't return. */
d_r_read_unlock(&hotp_vul_table_lock); /* Part of fix for case 5521. */
hotp_change_control_flow(app_reg_ptr, module_base + return_rva);
ASSERT_NOT_REACHED();
}
}
hotp_gateway_ret:
if (!own_hot_patch_lock)
d_r_read_unlock(&hotp_vul_table_lock); /* Part of fix for case 5521. */
/* if we change this to be invoked from d_r_dispatch, should remove this */
EXITING_DR();
return res;
}
/* This routine will execute the given hot patch (either detector or protector)
* and return the appropriate execution status. If the hot patch causes an
* exception, it will be terminated and the exception handler will automatically
* return to this function and this function will return with status
* HOTP_EXEC_ABORTED.
* If a hot patch exception occurs
* - it dumps a forenscis file if asked for (using dump_excpt_info)
* - the exception handler dumps a core if the mask is set and logs an event
* If a hot patch returns HOTP_{DETECTOR,PROTECTOR}_ERROR,
* - it dumps a forenscis file if asked for (using dump_error_info)
* - it dumps a core if asked for and the mask is set
* - it logs an event if the patch returns HOTP_EXEC_LOG_EVENT too.
* Hot patch exceptions and errors are treated similarly because they both
* point to a faulty hot patch. The only differences are in the string of the
* event logged and the cause-string of the forensics file.
*
* Note: this routine uses a shadow app reg state to recover from a hot patch
* exception cleanly.
* XXX: using setjmp & longjmp can cause problems if the compiler decides
* to reuse unused args/locals; should probably use volatile for those.
*/
static hotp_exec_status_t
hotp_execute_patch(hotp_func_t hotp_fn_ptr, hotp_context_t *hotp_cxt,
hotp_policy_mode_t mode, bool dump_excpt_info, bool dump_error_info)
{
hotp_exec_status_t exec_status, exec_status_only;
hotp_context_t local_cxt;
dcontext_t *dcontext = get_thread_private_dcontext();
dr_where_am_i_t wherewasi;
ASSERT(dcontext != NULL && dcontext != GLOBAL_DCONTEXT);
if (dcontext == NULL || /* case 9385: unknown thread */
dcontext == GLOBAL_DCONTEXT /* just bug */) {
SYSLOG_INTERNAL_WARNING("hotp_execute_patch: unknown thread");
return HOTP_EXEC_ABORTED;
}
/* For hot patching with fcache, today, hot patches are executed only from
* within the fcache. For hotp_only, there is no fcache; hot patches are
* executed directly when they are DR_WHERE_APP.
Question for reviewer: for hotp_only, when control comes to the gateway, should
whereami be changed to something other than DR_WHERE_APP because we are technically
in the core now; if so, would it be DR_WHERE_TRAMPOLINE?
*/
ASSERT(dcontext->whereami == DR_WHERE_FCACHE ||
(dcontext->whereami == DR_WHERE_APP && DYNAMO_OPTION(hotp_only)));
wherewasi = dcontext->whereami;
/* In case the hot patch causes an exception, the context may be in an
* inconsistent state. To prevent that make a copy of the app's context
* and pass the copy to the hot patch.
* Note: nothing is done for partial memory writes. TODO: how to fix this?
*/
memcpy(&local_cxt, hotp_cxt, sizeof(hotp_context_t));
dcontext->whereami = DR_WHERE_HOTPATCH;
if (DR_SETJMP(&dcontext->hotp_excpt_state) == 0) { /* TRY block */
exec_status = (*hotp_fn_ptr)(&local_cxt);
exec_status_only = exec_status & ~HOTP_EXEC_LOG_EVENT;
/* Successful execution can't return exception code. */
ASSERT(exec_status_only != HOTP_EXEC_ABORTED);
if (mode == HOTP_MODE_DETECT) {
/* The detector shouldn't have modified register state.
* Note: currently there is no way to find out if the memory state
* was modified.
*/
ASSERT(memcmp(hotp_cxt, &local_cxt, sizeof(hotp_context_t)) == 0);
} else if (mode == HOTP_MODE_PROTECT) {
/* Copy back local context which may have been modified by the
* protector back to the context passed in, i.e., apply the
* changes enforced by the hot patch. Note: this is applicable
* only for registers not memory
*/
memcpy(hotp_cxt, &local_cxt, sizeof(hotp_context_t));
}
if ((TESTANY(HOTP_EXEC_DETECTOR_ERROR, exec_status_only) ||
(TESTANY(HOTP_EXEC_PROTECTOR_ERROR, exec_status_only)))) {
const char *msg = TESTANY(HOTP_EXEC_DETECTOR_ERROR, exec_status_only)
? "Hot patch detector error"
: "Hot patch protector error";
if (dump_error_info) {
if (TESTANY(DUMPCORE_HOTP_FAILURE, DYNAMO_OPTION(dumpcore_mask)))
os_dump_core(msg);
}
if (TESTANY(HOTP_EXEC_LOG_EVENT, exec_status)) {
SYSLOG_CUSTOM_NOTIFY(SYSLOG_ERROR, MSG_HOT_PATCH_FAILURE, 3,
"Hot patch error, continuing.",
get_application_name(), get_application_pid(),
"<none>");
}
# ifdef PROGRAM_SHEPHERDING
if (dump_error_info) {
report_diagnostics(msg, NULL, HOT_PATCH_FAILURE);
}
# endif
}
} else { /* EXCEPT block */
/* Hot patch crashed! */
if (dump_excpt_info) {
/* Usually logging the event, dumping core and forensics are done
* together. In this case the first two are done in the exception
* handler because that is where the exception specific information
* is available. Forensics are dumped here because this is where
* the failing vulnerability's information is available. Trying to
* do all in one place would require too many pieces of information
* being passed around.
*/
/* TODO: no hot patch exception specific information is dumped
* in the forensics files today; need to do so.
*/
/* TODO: title should say detector or protector exception. */
# ifdef PROGRAM_SHEPHERDING
report_diagnostics("Hot patch exception", NULL, HOT_PATCH_FAILURE);
# endif
}
exec_status = HOTP_EXEC_ABORTED;
}
dcontext->whereami = wherewasi;
/* Reset hotp_excpt_state to unused. This will be used in
* create_callback_dcontext() to catch potential callbacks, which might
* lead to nested hot patch exceptions, that result due to system calls
* mades from a hot patch. Hot patches shouldn't make system calls.
*/
DODEBUG(memset(&dcontext->hotp_excpt_state, -1, sizeof(dr_jmp_buf_t)););
return exec_status;
}
/* This routine plugs the hot patch violation event into the core's existing
* reporting mechanism. A new threat id (.H) will be generated for hot patch
* violations and the event log will mention whether the violation was detected
* or protected. -report_max will apply these violations. However,
* -detect_mode and other termination options like -kill_thread, etc. won't be.
*
* TODO: currently, -kill_thread and such will apply to hot patches. They
* must be decoupled. Not done currently because a clean way of doing it is
* out of the scope of blowfish beta.
*/
static void
hotp_event_notify(hotp_exec_status_t exec_status, bool protected,
const hotp_offset_match_t *inject_point, const app_pc bad_addr,
const hotp_context_t *hotp_cxt)
{
# ifdef PROGRAM_SHEPHERDING
const char *threat_id = NULL;
hotp_type_t hotp_type = GLOBAL_VUL(inject_point->vul_index).type;
action_type_t action;
app_pc faulting_addr = NULL, inject_addr, old_next_tag;
security_violation_t violation_type = INVALID_VIOLATION, res;
dcontext_t *dcontext = get_thread_private_dcontext();
fragment_t src_frag = { 0 }, *old_last_frag;
priv_mcontext_t old_mc;
ASSERT(dcontext != NULL && dcontext != GLOBAL_DCONTEXT);
ASSERT(inject_point != NULL && hotp_cxt != NULL);
/* action mapping */
switch (exec_status & ~HOTP_EXEC_LOG_EVENT) {
case HOTP_EXEC_EXPLOIT_KILL_THREAD: action = ACTION_TERMINATE_THREAD; break;
case HOTP_EXEC_EXPLOIT_KILL_PROCESS: action = ACTION_TERMINATE_PROCESS; break;
case HOTP_EXEC_EXPLOIT_RAISE_EXCEPTION: action = ACTION_THROW_EXCEPTION; break;
default: action = ACTION_CONTINUE;
};
inject_addr = GLOBAL_MODULE(inject_point->vul_index, inject_point->set_index,
inject_point->module_index)
.base_address +
GLOBAL_PPOINT(inject_point->vul_index, inject_point->set_index,
inject_point->module_index, inject_point->ppoint_index)
.offset;
/* Determine the faulting address, violation type and threat id. */
if (TESTALL(HOTP_TYPE_GBOP_HOOK, hotp_type)) { /* gbop hook type */
# ifdef GBOP
/* XXX: share reporting code with gbop_validate_and_act() - have
* one reporting interface for gbop. Case 8096. Changes here or
* in gbop_validate_and_act() should be kept in sync.
*/
/* Even though many gbop hooks are implemented with hotp_only interface
* gbop violations are treated separately.
*/
ASSERT(DYNAMO_OPTION(hotp_only)); /* no gbopping in code cache mode */
/* NOTE: For gbop, the source is actually the hook address and the
* target is the failing address, not vice versa.
*/
faulting_addr = bad_addr;
violation_type = GBOP_SOURCE_VIOLATION;
/* gbop remediations are decided in security_violation(), so set it to
* the expected default here otherwise security_violation would assert.
*/
action = ACTION_TERMINATE_PROCESS;
# endif
} else { /* hot patch type */
ASSERT(TESTALL(HOTP_TYPE_HOT_PATCH, hotp_type));
ASSERT(bad_addr == NULL);
faulting_addr = inject_addr;
if (protected) {
violation_type = HOT_PATCH_PROTECTOR_VIOLATION;
} else {
violation_type = HOT_PATCH_DETECTOR_VIOLATION;
}
threat_id = GLOBAL_VUL(inject_point->vul_index).policy_id;
ASSERT(threat_id != NULL);
}
/* Save the last fragment, next tag & register state, set them up with the
* right values, report and restore afterwards.
* Note: for hotp_only, src & tgt are the same; for gbop see comment above.
*/
hotp_spill_before_notify(dcontext, &old_last_frag, &src_frag, inject_addr,
&old_next_tag, faulting_addr, &old_mc, (void *)hotp_cxt,
CXT_TYPE_HOT_PATCH);
res = security_violation_internal(dcontext, faulting_addr, violation_type,
OPTION_REPORT | OPTION_BLOCK, threat_id, action,
&hotp_vul_table_lock);
/* Some sanity checks before we go on our merry way. BTW, couldn't use
* DODEBUG because gcc complained about using the #ifdef GBOP inside a
* macro; not an issue; the compiler should optimize out the whole if-block.
*/
if (res == ALLOWING_BAD) {
/* Threat exemptions are only for gbop hooks, they don't make sense
* for hot patches - if you don't want a hot patch's event, just
* turn it off.
*/
ASSERT(TESTALL(HOTP_TYPE_GBOP_HOOK, hotp_type));
ASSERT(!TESTALL(HOTP_TYPE_HOT_PATCH, hotp_type));
} else if (res == HOT_PATCH_DETECTOR_VIOLATION ||
res == HOT_PATCH_PROTECTOR_VIOLATION) {
/* Can return only to continue. */
ASSERT(action == ACTION_CONTINUE);
} else {
# ifdef GBOP
ASSERT(res == GBOP_SOURCE_VIOLATION);
ASSERT(action == ACTION_CONTINUE || DYNAMO_OPTION(detect_mode));
# endif
}
hotp_restore_after_notify(dcontext, old_last_frag, old_next_tag, &old_mc);
/* Can't leave this function without holding the hotp lock! */
ASSERT_OWN_READ_LOCK(true, &hotp_vul_table_lock);
# endif
}
/* This is a hack to make hotp use our existing security violation reporting
* mechanism, which relies on fragments & tags to report violations & generate
* forensics. Case 8079 talks about cleaning up the reporting interface.
*/
void
hotp_spill_before_notify(dcontext_t *dcontext, fragment_t **frag_spill /* OUT */,
fragment_t *new_frag, const app_pc new_frag_tag,
app_pc *new_tag_spill /* OUT */, const app_pc new_next_tag,
priv_mcontext_t *cxt_spill /* OUT */, const void *new_cxt,
cxt_type_t cxt_type)
{
priv_mcontext_t *mc;
ASSERT(dcontext != NULL && dcontext != GLOBAL_DCONTEXT);
ASSERT(frag_spill != NULL && new_frag != NULL && new_frag_tag != NULL);
ASSERT(new_tag_spill != NULL && new_next_tag != NULL);
ASSERT(cxt_spill != NULL && new_cxt != NULL);
ASSERT(cxt_type == CXT_TYPE_HOT_PATCH || cxt_type == CXT_TYPE_CORE_HOOK);
*frag_spill = dcontext->last_fragment;
*new_tag_spill = dcontext->next_tag;
new_frag->tag = new_frag_tag;
dcontext->last_fragment = (fragment_t *)new_frag;
dcontext->next_tag = new_next_tag;
/* For hotp_only the last_fragment should be linkstub_empty_fragment,
* which is static in link.c
*
* next_tag can be set to BACK_TO_NATIVE_AFTER_SYSCALL, so can't
* easily assert on that.
*/
ASSERT(!DYNAMO_OPTION(hotp_only) ||
((*frag_spill)->tag == NULL && (*frag_spill)->flags == FRAG_FAKE));
/* Saving & swapping contexts - this is needed to produce the
* correct machine context for forensics; there can be two types, viz.,
* hotp_context_t if called from hotp_event_notify() and
* app_state_at_intercept_t if called from gbop_validate_and_act().
*/
mc = get_mcontext(dcontext);
ASSERT(mc != NULL);
memcpy(cxt_spill, mc, sizeof(*cxt_spill));
if (cxt_type == CXT_TYPE_HOT_PATCH) {
hotp_context_t *new = (hotp_context_t *)new_cxt;
*mc = *new;
/* XXX PR 226036: use hotp_context_t.xip */
mc->pc = NULL; /* pc reported in source, so NULL here is ok. */
} else if (cxt_type == CXT_TYPE_CORE_HOOK) {
app_state_at_intercept_t *new = (app_state_at_intercept_t *)new_cxt;
*mc = new->mc;
/* XXX PR 226036: use hotp_context_t.xip */
mc->pc = NULL; /* pc reported in source, so NULL here is ok. */
} else
ASSERT_NOT_REACHED();
}
/* Restore dcontext last_fragment & next_tag after reporting the violation. */
/* XXX: old_cxt is unused; see case 8099 about dumping context. */
void
hotp_restore_after_notify(dcontext_t *dcontext, const fragment_t *old_frag,
const app_pc old_next_tag, const priv_mcontext_t *old_cxt)
{
priv_mcontext_t *mc;
ASSERT(dcontext != NULL && dcontext != GLOBAL_DCONTEXT);
ASSERT(old_cxt != NULL);
dcontext->last_fragment = (fragment_t *)old_frag;
dcontext->next_tag = old_next_tag;
mc = get_mcontext(dcontext);
ASSERT(mc != NULL);
memcpy(mc, old_cxt, sizeof(*old_cxt));
}
# if defined(DEBUG) && defined(INTERNAL)
/* XXX PR 226036: eip is now part of hotp_context_t */
static void
hotp_dump_reg_state(const hotp_context_t *reg_state, const app_pc eip,
const uint loglevel)
{
LOG(GLOBAL, LOG_HOT_PATCHING, loglevel, "eax: " PFX "\n", APP_XAX(reg_state));
LOG(GLOBAL, LOG_HOT_PATCHING, loglevel, "ecx: " PFX "\n", APP_XCX(reg_state));
LOG(GLOBAL, LOG_HOT_PATCHING, loglevel, "edx: " PFX "\n", APP_XDX(reg_state));
LOG(GLOBAL, LOG_HOT_PATCHING, loglevel, "ebx: " PFX "\n", APP_XBX(reg_state));
LOG(GLOBAL, LOG_HOT_PATCHING, loglevel, "esp: " PFX "\n", APP_XSP(reg_state));
LOG(GLOBAL, LOG_HOT_PATCHING, loglevel, "ebp: " PFX "\n", APP_XBP(reg_state));
LOG(GLOBAL, LOG_HOT_PATCHING, loglevel, "esi: " PFX "\n", APP_XSI(reg_state));
LOG(GLOBAL, LOG_HOT_PATCHING, loglevel, "edi: " PFX "\n", APP_XDI(reg_state));
LOG(GLOBAL, LOG_HOT_PATCHING, loglevel, "eip: " PFX "\n", eip);
}
# endif
static void
hotp_update_vul_stats(const hotp_exec_status_t exec_status, const uint vul_index)
{
hotp_exec_status_t temp = exec_status & ~HOTP_EXEC_LOG_EVENT;
ASSERT(temp == HOTP_EXEC_EXPLOIT_DETECTED || temp == HOTP_EXEC_EXPLOIT_NOT_DETECTED ||
temp == HOTP_EXEC_DETECTOR_ERROR || temp == HOTP_EXEC_EXPLOIT_PROTECTED ||
temp == HOTP_EXEC_EXPLOIT_NOT_PROTECTED ||
temp == HOTP_EXEC_EXPLOIT_KILL_THREAD ||
temp == HOTP_EXEC_EXPLOIT_KILL_PROCESS ||
temp == HOTP_EXEC_EXPLOIT_RAISE_EXCEPTION ||
temp == HOTP_EXEC_CHANGE_CONTROL_FLOW || temp == HOTP_EXEC_PROTECTOR_ERROR ||
temp == HOTP_EXEC_ABORTED);
/* XXX: Grabbing the hot patch lock here to update stats will deadlock
* if a nudge is waiting for this thread to get out. If a lock
* isn't grabbed, then the stats may be slightly inaccurate if 2
* threads update the same stat for a given vulnerability at
* the same time; odds are low and inaccurate stats aren't a problem.
* We aren't trying to provide accurate stats; besides stats for a
* vulnerability for all process in all nodes using it is vague
* data anyway.
* if VUL_STAT_INC becomes atomic, we won't need a lock here.
* XXX: Vlad suggested creating a stats lock; good idea;
*/
switch (temp) {
case HOTP_EXEC_EXPLOIT_DETECTED: {
VUL_STAT_INC(GLOBAL_VUL(vul_index).info->num_detected);
break;
}
case HOTP_EXEC_EXPLOIT_NOT_DETECTED: {
VUL_STAT_INC(GLOBAL_VUL(vul_index).info->num_not_detected);
break;
}
case HOTP_EXEC_DETECTOR_ERROR: {
VUL_STAT_INC(GLOBAL_VUL(vul_index).info->num_detector_error);
break;
}
case HOTP_EXEC_EXPLOIT_PROTECTED: {
VUL_STAT_INC(GLOBAL_VUL(vul_index).info->num_protected);
break;
}
case HOTP_EXEC_EXPLOIT_NOT_PROTECTED: {
VUL_STAT_INC(GLOBAL_VUL(vul_index).info->num_not_protected);
break;
}
case HOTP_EXEC_EXPLOIT_KILL_THREAD: {
VUL_STAT_INC(GLOBAL_VUL(vul_index).info->num_kill_thread);
break;
}
case HOTP_EXEC_EXPLOIT_KILL_PROCESS: {
VUL_STAT_INC(GLOBAL_VUL(vul_index).info->num_kill_process);
break;
}
case HOTP_EXEC_EXPLOIT_RAISE_EXCEPTION: {
VUL_STAT_INC(GLOBAL_VUL(vul_index).info->num_raise_exception);
break;
}
case HOTP_EXEC_CHANGE_CONTROL_FLOW: {
VUL_STAT_INC(GLOBAL_VUL(vul_index).info->num_change_control_flow);
break;
}
case HOTP_EXEC_PROTECTOR_ERROR: {
VUL_STAT_INC(GLOBAL_VUL(vul_index).info->num_protector_error);
break;
}
case HOTP_EXEC_ABORTED: {
VUL_STAT_INC(GLOBAL_VUL(vul_index).info->num_aborted);
break;
}
default: ASSERT_NOT_REACHED();
}
}
/* Note: 1. This function will not return, unless there is an error.
* 2. The number of patch points must be passed to this function to handle
* control flow changes with multiple patches at the same offset.
*
* CAUTION: Any change to the code generated by hotp_inject_gateway_call() (and,
* thus, prepare_for_clean_call()), will affect how the app. state
* is spilled on the dr stack. This function uses that app. state,
* hence, relies on that order being constant.
* TODO: How to link an assert to these two, so that any change is
* caught immediately?
*
* TODO: show stack diagrams otherwise it is going to be messy
*/
/* These constants refer to the offset of eflags and errno that are saved on
* the stack as part of the clean call. The offsets are relative to the
* location of the pushed register state, i.e., esp after pusha in the clean
* call sequence. Any change to prepare_for_clean_call() will affect this.
*/
# define CLEAN_CALL_XFLAGS_OFFSET 1
# define CLEAN_CALL_ERRNO_OFFSET 2
static void
hotp_change_control_flow(const hotp_context_t *app_reg_ptr, const app_pc target)
{
dcontext_t *dcontext;
priv_mcontext_t mc = *app_reg_ptr;
/* TODO: Eventually, must assert that target is in some module. */
ASSERT(app_reg_ptr != NULL && target != NULL);
dcontext = get_thread_private_dcontext();
ASSERT(dcontext != NULL && dcontext != GLOBAL_DCONTEXT);
dcontext->next_tag = target; /* Set up actual control flow change. */
dcontext->whereami = DR_WHERE_FCACHE;
/* XXX: should determine the actual fragment exiting from */
set_last_exit(dcontext, (linkstub_t *)get_hot_patch_linkstub());
STATS_INC(hotp_num_cflow_change);
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "Changing control flow to " PFX "\n", target);
transfer_to_dispatch(dcontext, &mc, true /*full_DR_state*/);
ASSERT_NOT_REACHED();
}
/* Prints hotpatch vulnerability table information to forensics file in xml
* format.
*/
void
hotp_print_diagnostics(file_t diagnostics_file)
{
uint vul, set, module, pp, hash;
if (GLOBAL_VUL_TABLE == NULL) {
print_file(diagnostics_file,
"<hotpatching-information>\n"
"Hotpatch vulnerability table is NULL\n"
"</hotpatching-information>\n");
return;
}
print_file(diagnostics_file,
"<hotpatching-information>\n"
"<vulnerability-table num-vulnerabilities=\"%u\">\n",
NUM_GLOBAL_VULS);
for (vul = 0; vul < NUM_GLOBAL_VULS; vul++) {
hotp_inject_status_t inject_status;
print_file(diagnostics_file, " <vulnerability id=\"%s\" num-sets=\"%u\">\n",
GLOBAL_VUL(vul).vul_id, GLOBAL_VUL(vul).num_sets);
print_file(diagnostics_file,
" <policy id=\"%s\" mode=\"%d\" version=\"%u\"/>\n",
GLOBAL_VUL(vul).policy_id, GLOBAL_VUL(vul).mode,
GLOBAL_VUL(vul).policy_version);
/* For vulnerabilities that haven't been used, print only the
* {vul,policy}_id and policy_version; helps to prevent clustter in the
* forensics file. Using the ids & version we can get the hot patch
* definition from our packages/code in house, so not dumping them
* doesn't hamper diagnosis. Case 8549.
*/
inject_status = *(GLOBAL_VUL(vul).info->inject_status);
if (inject_status == HOTP_INJECT_NO_MATCH || inject_status == HOTP_INJECT_OFF) {
print_file(diagnostics_file, " </vulnerability>\n");
continue;
}
print_file(diagnostics_file,
" <hotpatch-dll name=\"%s\" base=\"" PFX "\" hash=\"%s\"/>\n",
GLOBAL_VUL(vul).hotp_dll, GLOBAL_VUL(vul).hotp_dll_base,
GLOBAL_VUL(vul).hotp_dll_hash);
for (set = 0; set < GLOBAL_VUL(vul).num_sets; set++) {
bool print_sets = false;
for (module = 0; module < GLOBAL_SET(vul, set).num_modules; module++) {
/* If a module isn't matched, then it hasn't been used, don't
* dump it; case 8549.
*/
if (!GLOBAL_MODULE(vul, set, module).matched)
continue;
if (!print_sets) { /* Print set title if needed, case 8549. */
print_file(diagnostics_file, " <set num-modules=\"%u\">\n",
GLOBAL_SET(vul, set).num_modules);
print_sets = true;
}
print_file(diagnostics_file,
" <module pe_name=\"%s\" pe_checksum=\"0x%x\" "
"pe_timestamp=\"0x%x\" pe_image_size=\"" PIFX "\" "
"pe_code_size=\"" PIFX "\" "
"pe_file_version=\"0x" HEX64_FORMAT_STRING "\" "
"num-hashes=\"%u\" num-patch-points=\"%u\">\n",
GLOBAL_SIG(vul, set, module).pe_name,
GLOBAL_SIG(vul, set, module).pe_checksum,
GLOBAL_SIG(vul, set, module).pe_timestamp,
GLOBAL_SIG(vul, set, module).pe_image_size,
GLOBAL_SIG(vul, set, module).pe_code_size,
GLOBAL_SIG(vul, set, module).pe_file_version,
GLOBAL_MODULE(vul, set, module).num_patch_point_hashes,
GLOBAL_MODULE(vul, set, module).num_patch_points);
for (hash = 0;
hash < GLOBAL_MODULE(vul, set, module).num_patch_point_hashes;
hash++) {
print_file(diagnostics_file,
" <hash start=\"" PFX "\" length=\"0x%x\" "
"hash=\"%u\"/>\n",
GLOBAL_HASH(vul, set, module, hash).start,
GLOBAL_HASH(vul, set, module, hash).len,
GLOBAL_HASH(vul, set, module, hash).hash_value);
}
for (pp = 0; pp < GLOBAL_MODULE(vul, set, module).num_patch_points;
pp++) {
print_file(diagnostics_file,
" <hotpatch precedence=\"%u\" offset=\"" PIFX
"\">\n",
GLOBAL_PPOINT(vul, set, module, pp).precedence,
GLOBAL_PPOINT(vul, set, module, pp).offset);
print_file(diagnostics_file,
" <function type=\"detector\" "
"offset=\"" PIFX "\"/>\n",
GLOBAL_PPOINT(vul, set, module, pp).detector_fn);
print_file(diagnostics_file,
" <function type=\"protector\" "
"offset=\"" PIFX "\" return=\"" PFX "\"/>\n",
GLOBAL_PPOINT(vul, set, module, pp).protector_fn,
GLOBAL_PPOINT(vul, set, module, pp).return_addr);
print_file(diagnostics_file, " </hotpatch>\n");
}
print_file(diagnostics_file, " </module>\n");
}
if (print_sets) { /* xref case 8549 */
print_file(diagnostics_file, " </set>\n");
}
}
print_file(
diagnostics_file,
" <stats "
"num-detected=\"" UINT64_FORMAT_STRING "\" "
"num-not-detected=\"" UINT64_FORMAT_STRING "\" "
"num-detector-error=\"" UINT64_FORMAT_STRING "\" "
"num-protected=\"" UINT64_FORMAT_STRING "\" "
"num-not-protected=\"" UINT64_FORMAT_STRING "\" "
"num-kill-thread=\"" UINT64_FORMAT_STRING "\" "
"num-kill-process=\"" UINT64_FORMAT_STRING "\" "
"num-raise-exception=\"" UINT64_FORMAT_STRING "\" "
"num-change-control-flow=\"" UINT64_FORMAT_STRING "\" "
"num-protector-error=\"" UINT64_FORMAT_STRING "\" "
"num-aborted=\"" UINT64_FORMAT_STRING "\">\n",
GLOBAL_VUL(vul).info->num_detected, GLOBAL_VUL(vul).info->num_not_detected,
GLOBAL_VUL(vul).info->num_detector_error, GLOBAL_VUL(vul).info->num_protected,
GLOBAL_VUL(vul).info->num_not_protected,
GLOBAL_VUL(vul).info->num_kill_thread, GLOBAL_VUL(vul).info->num_kill_process,
GLOBAL_VUL(vul).info->num_raise_exception,
GLOBAL_VUL(vul).info->num_change_control_flow,
GLOBAL_VUL(vul).info->num_protector_error, GLOBAL_VUL(vul).info->num_aborted);
print_file(diagnostics_file,
" <status type=\"execution\">%d</status>\n"
" <status type=\"injection\">%d</status>\n",
GLOBAL_VUL(vul).info->exec_status,
*(GLOBAL_VUL(vul).info->inject_status));
print_file(diagnostics_file,
" </stats>\n"
" </vulnerability>\n");
}
print_file(diagnostics_file,
"</vulnerability-table>\n"
"</hotpatching-information>\n");
}
# if defined(DEBUG) && defined(DEBUG_MEMORY)
/* Part of bug fix for case 9593 which required leaking trampolines. */
bool
hotp_only_contains_leaked_trampoline(byte *pc, size_t size)
{
if (!DYNAMO_OPTION(hotp_only) IF_WINDOWS(|| !doing_detach))
return false;
/* Today memory debug checks for special heap units only do heap accouting,
* but not memcmp, both of which are done for regular heaps. Special heaps
* are where the leaked trampolines are located. If we do implement that
* check then this code in #if 0 would be needed. Case 10434. */
# if 0
for (i = 0; i < hotp_only_num_tramps_leaked; i++) {
if (hotp_only_tramps_leaked[i] >= pc &&
hotp_only_tramps_leaked[i] < (pc + size)) {
/* Make sure we don't have trampolines across heap units! */
ASSERT((hotp_only_tramps_leaked[i] + HOTP_ONLY_TRAMPOLINE_SIZE) <=
(pc + size));
return true;
}
}
# endif
/* The actual special_units_t structure (pointed to by
* hotp_only_tramp_heap) is also leaked.
* Note: hotp_only_tramp_heap_cache can be NULL if no hotp_only type
* patches were ever removed either because they were never injected or
* just weren't removed.
*/
if ((byte *)hotp_only_tramp_heap_cache >= pc &&
(byte *)hotp_only_tramp_heap_cache < (pc + size))
return true;
return false;
}
# endif
/*----------------------------------------------------------------------------*/
# ifdef GBOP
/* This section contains most of the functionality needed to treat gbop hooks
* as hotp_only patches, thus giving gbop hooks access to all hotp_only patch
* functionality. See case 7949 & 7127.
*/
/* Note: Both the gbop detector and protector request for log events. However,
* the detector events are reported only in -detect_mode and protector ones
* in !-detect_mode.
* Note: The app eax & edx are spilled by the gateway and used as scratch;
* eax to get the faulting address & edx to send in the current pc (which is
* also set by the gateway). xref case 6804 about hotp interface expansion.
*/
static hotp_exec_status_t
hotp_only_gbop_detector(hotp_context_t *cxt)
{
if (gbop_check_valid_caller((app_pc)APP_XBP(cxt), (app_pc)APP_XSP(cxt),
(app_pc)APP_XDX(cxt), (app_pc *)&APP_XAX(cxt))) {
return HOTP_EXEC_EXPLOIT_NOT_DETECTED;
} else {
/* Ask for event; needed to log event if -detect_mode is specified. */
return HOTP_EXEC_EXPLOIT_DETECTED | HOTP_EXEC_LOG_EVENT;
}
}
static hotp_exec_status_t
hotp_only_gbop_protector(hotp_context_t *cxt)
{
# ifdef PROGRAM_SHEPHERDING
ASSERT(!DYNAMO_OPTION(detect_mode)); /* No protection in detect_mode. */
# endif
/* Just log the event; the remediation action for gbop is determined by
* security_violation() using core options like -kill_thread.
*/
return HOTP_EXEC_EXPLOIT_PROTECTED | HOTP_EXEC_LOG_EVENT;
}
/* Note: num_vuls is an IN OUT argument; is specifies the current table size and
* is updated to the new size after reading gbop hooks. The IN value is used
* as the append index into the table.
*/
static void
hotp_only_read_gbop_policy_defs(hotp_vul_t *tab, uint *num_vuls)
{
uint vul_idx;
hotp_vul_t *vul;
hotp_set_t *set;
hotp_module_t *module;
hotp_patch_point_t *patch_point;
uint gbop_num_hooks = gbop_get_num_hooks();
app_pc dr_base = NULL;
ASSERT(tab != NULL && num_vuls != NULL);
ASSERT(gbop_num_hooks > 0);
/* No gbopping for regular hotp, at least not until hotp_only and regular
* hotp coexist, i.e., hotp_only for native_exec dlls (case 6892)
*/
ASSERT(DYNAMO_OPTION(hotp_only) && DYNAMO_OPTION(gbop));
dr_base = get_module_base((app_pc)hotp_only_read_gbop_policy_defs);
ASSERT(dr_base != NULL);
for (vul_idx = *num_vuls; vul_idx < (*num_vuls + gbop_num_hooks); vul_idx++) {
uint gbop_hook_idx = vul_idx - *num_vuls;
const gbop_hook_desc_t *gbop_hook = gbop_get_hook(gbop_hook_idx);
ASSERT(gbop_hook != NULL);
vul = &tab[vul_idx];
vul->vul_id = dr_strdup(gbop_hook->func_name HEAPACCT(ACCT_HOT_PATCHING));
/* XXX: construct this from a combination of {mod,func}_name, or
* func_name and bad_ret_address. For now, just hard code it.
* strdup because hotp_free() thinks this is allocated.
*/
vul->policy_id = dr_strdup("GBOP.VIOL" HEAPACCT(ACCT_HOT_PATCHING));
/* XXX: should this be used to track changes to the gbop detector
* and protector? does it matter, after all these patches will only
* go as part of the core?
*/
vul->policy_version = 1;
vul->hotp_dll = NULL;
vul->hotp_dll_hash = NULL;
/* There is no notion of only detecting and doing nothing for gbop, so
* the mode is always protect.
*/
/* gbop_execlude_filter handles any os specific gbop set removals, xref 9772 */
if (gbop_exclude_filter(gbop_hook)) {
vul->mode = HOTP_MODE_OFF;
LOG(GLOBAL, LOG_HOT_PATCHING, 1, "Excluding %s!%s\n", gbop_hook->mod_name,
gbop_hook->func_name);
} else {
vul->mode = HOTP_MODE_PROTECT;
}
vul->num_sets = 1;
set = HEAP_ARRAY_ALLOC(GLOBAL_DCONTEXT, hotp_set_t, 1, ACCT_HOT_PATCHING,
PROTECTED);
vul->sets = set;
vul->info = HEAP_TYPE_ALLOC(GLOBAL_DCONTEXT, hotp_vul_info_t, ACCT_HOT_PATCHING,
PROTECTED);
memset(vul->info, 0, sizeof(hotp_vul_info_t)); /* Initialize stats. */
vul->hotp_dll_base = dr_base;
vul->type = HOTP_TYPE_GBOP_HOOK;
set->num_modules = 1;
module = HEAP_ARRAY_ALLOC(GLOBAL_DCONTEXT, hotp_module_t, 1, ACCT_HOT_PATCHING,
PROTECTED);
set->modules = module;
module->sig.pe_name = dr_strdup(gbop_hook->mod_name HEAPACCT(ACCT_HOT_PATCHING));
module->sig.pe_checksum = 0;
module->sig.pe_timestamp = 0;
module->sig.pe_image_size = 0;
module->sig.pe_code_size = 0;
module->sig.pe_file_version = 0;
module->num_patch_points = 1;
patch_point = HEAP_ARRAY_ALLOC(GLOBAL_DCONTEXT, hotp_patch_point_t, 1,
ACCT_HOT_PATCHING, PROTECTED);
module->patch_points = patch_point;
module->num_patch_point_hashes = 0;
module->hashes = NULL;
module->matched = false;
module->base_address = NULL;
/* The actual patch offset will be computed if the module matches.
* vul_id holds the function name, which will be used to compute the
* offset. See hotp_process_image().
* XXX: we could use a union for offset to hold offset or func_name;
* that would be elegant, but would require changing too many things
* in hotp - not a good idea, not at least for the first implementation
*/
patch_point->offset = 0;
patch_point->detector_fn = (app_pc)hotp_only_gbop_detector - dr_base;
patch_point->protector_fn = (app_pc)hotp_only_gbop_protector - dr_base;
/* Longer term issue: do we want to have the notion of changing
* control flow for gbop hooks?
*/
patch_point->return_addr = 0;
/* Precedence hasn't been implemented yet, however, if it had been,
* then we don't want gbop hooks to interfere with other patches.
*/
patch_point->precedence = HOTP_ONLY_GBOP_PRECEDENCE;
patch_point->trampoline = NULL;
patch_point->app_code_copy = NULL;
patch_point->tramp_exit_tgt = NULL;
}
*num_vuls += gbop_num_hooks;
ASSERT(*num_vuls == vul_idx);
}
# endif /* GBOP */
/* Both dr_{insert,update}_probes() will be replaced by dr_register_probes() -
* PR 225547. The user will call the same routine to insert or update probes.
* Subsequent calls will result in old probes being removed and new ones
* inserted. By manipulating the input array the user can do inserts (adding
* new defs. to the array), updates (modifying existing defs) or removes (just
* removing unwanted defs from the array).
* Depending upon the context (init or other place) an internal nudge will be
* created.
* NOTE: for beta, there is no update, i.e., this routine can be called only
* once.
* NOTE: The input is an array of probes because allowing the user to do
* individual probe registration will result in a nudge for each one, which is
* very expensive. Also of note is that it isn't uncommon for API to request
* arrays; WIN32 native API does it many places.
* TODO: change hotp vul table to be a list - better for clients, esp. multiple
* ones; also good if probes are used with LS - PR 225673.
*/
void
dr_register_probes(dr_probe_desc_t *probes, unsigned int num_probes)
{
unsigned int i, valid_probes;
hotp_vul_t *tab, *vul;
hotp_set_t *set;
hotp_module_t *module;
hotp_patch_point_t *ppoint;
static bool probes_initialized = false;
/* For now, probes are supported iff probe api is explicitly turned on.
* Also, liveshields shouldn't be on when probe api is on.
*/
CLIENT_ASSERT(DYNAMO_OPTION(hot_patching) && DYNAMO_OPTION(probe_api) &&
!DYNAMO_OPTION(liveshields),
"To use Probe API, "
"-hot_patching, -probe_api and -no_liveshields options "
"should be used.");
if (!DYNAMO_OPTION(probe_api))
return; /* be safe */
/* Hot patching subsystem should be initialized by now. */
ASSERT(hotp_patch_point_areas != NULL);
ASSERT(!DYNAMO_OPTION(hotp_only) || hotp_only_tramp_areas != NULL);
if (num_probes < MIN_NUM_VULNERABILITIES || num_probes > MAX_NUM_VULNERABILITIES ||
probes == NULL) {
/* XXX PR 533384: return a status code! */
return;
}
/* For beta probe registration can be done only once. However, multiple
* calls to this routine should be allowed during
* 1. dr init time - which doesn't need a nudge but needs remove & reinsert
* - PR 225580
* 2. any other point in dr - which requires an internal nudge - PR 225578
* what about DR event callbacks like module load/unload?
* Once both these are implemented the probes_initialized bool can go.
*
* Note: as we don't have multiple clients and at startup as we are single
* threaded here, there is no need for a lock for this temp. bool.
*/
if (probes_initialized) {
/* XXX PR 533384: return a status code!
* actually I'm having this continue for at-your-own-risk probes
*/
ASSERT_CURIOSITY_ONCE(false &&
"register probes >1x at your own risk: PR 225580!");
} else {
SELF_UNPROTECT_DATASEC(DATASEC_RARELY_PROT);
probes_initialized = true;
SELF_PROTECT_DATASEC(DATASEC_RARELY_PROT);
}
/* Zero out all dynamically allocated hotpatch table structures to avoid
* leaks when there is a parse error. See PR 212707, 213480.
*/
tab = HEAP_ARRAY_ALLOC_MEMSET(GLOBAL_DCONTEXT, hotp_vul_t, num_probes,
ACCT_HOT_PATCHING, PROTECTED, 0);
for (i = 0, valid_probes = 0; i < num_probes; i++) {
char *temp;
vul = &tab[valid_probes];
/* memset vul to 0 here because parse errors can leave freed pointers */
memset(vul, 0, sizeof(hotp_vul_t));
/* TODO: remove this once support is added for exported functions
* (PR 225654) & raw addresses (PR 225658); for now just prevent
* needless user errors.
*/
if (probes[i].insert_loc.type != DR_PROBE_ADDR_LIB_OFFS ||
probes[i].callback_func.type != DR_PROBE_ADDR_LIB_OFFS) {
probes[i].status = DR_PROBE_STATUS_UNSUPPORTED;
goto dr_probe_parse_error;
}
/* TODO: validate probe def PR 225663 */
vul->vul_id = dr_strdup(probes[i].name HEAPACCT(ACCT_HOT_PATCHING));
/* For probe api policy_id isn't needed, but it can't be set it to NULL
* because policy status table (used by drview) set up will crash. */
temp = HEAP_ARRAY_ALLOC(GLOBAL_DCONTEXT, char, MAX_POLICY_ID_LENGTH + 1,
ACCT_HOT_PATCHING, PROTECTED);
strncpy(temp, probes[i].name, MAX_POLICY_ID_LENGTH);
temp[MAX_POLICY_ID_LENGTH] = '\0';
/* TODO: validate probe def PR 225663 */
vul->policy_id = temp;
/* Note: if there is a need (highly doubt it) we can expand the probe
* api to support versioning; for now just set it to 1. */
vul->policy_version = 1;
vul->hotp_dll = NULL;
if (probes[i].callback_func.type == DR_PROBE_ADDR_LIB_OFFS) {
if (probes[i].callback_func.lib_offs.library != NULL) {
/* TODO: validate probe def PR 225663 */
vul->hotp_dll = dr_strdup(
probes[i].callback_func.lib_offs.library HEAPACCT(ACCT_HOT_PATCHING));
} else {
probes[i].status = DR_PROBE_STATUS_INVALID_LIB;
goto dr_probe_parse_error;
}
} else if (probes[i].callback_func.type == DR_PROBE_ADDR_EXP_FUNC) {
/* TODO: NYI - support for exported functions (PR 225654) */
probes[i].status = DR_PROBE_STATUS_UNSUPPORTED;
goto dr_probe_parse_error;
} else if (probes[i].callback_func.type != DR_PROBE_ADDR_VIRTUAL) {
/* TODO: NYI - support for virtual addresses (PR 225658) */
probes[i].status = DR_PROBE_STATUS_UNSUPPORTED;
goto dr_probe_parse_error;
}
vul->mode = HOTP_MODE_PROTECT;
vul->num_sets = 1;
set = HEAP_TYPE_ALLOC(GLOBAL_DCONTEXT, hotp_set_t, ACCT_HOT_PATCHING, PROTECTED);
vul->sets = set;
vul->info = HEAP_ARRAY_ALLOC_MEMSET(GLOBAL_DCONTEXT, hotp_vul_info_t, 1,
ACCT_HOT_PATCHING, PROTECTED, 0);
/* Note: if probe is inside client dll, then client dll SHOULD be in
* our module_areas - how to assert on this?
* Update: I found that neither the client dll nor any dll loaded
* during client init is in our loaded_module_areas because these dlls
* are loaded after the vm scan in vm_areas_init() but before dr hooks
* are inserted - I got pop ups in os_get_module_info() because of
* this. PR 225670.
*/
vul->hotp_dll_base = NULL;
vul->type = HOTP_TYPE_PROBE;
set->num_modules = 1;
module = HEAP_ARRAY_ALLOC_MEMSET(GLOBAL_DCONTEXT, hotp_module_t, 1,
ACCT_HOT_PATCHING, PROTECTED, 0);
set->modules = module;
if (probes[i].insert_loc.type == DR_PROBE_ADDR_LIB_OFFS) {
if (probes[i].insert_loc.lib_offs.library != NULL) {
module->sig.pe_name = dr_strdup(
probes[i].insert_loc.lib_offs.library HEAPACCT(ACCT_HOT_PATCHING));
} else {
probes[i].status = DR_PROBE_STATUS_INVALID_LIB;
goto dr_probe_parse_error;
}
} else if (probes[i].insert_loc.type == DR_PROBE_ADDR_EXP_FUNC) {
/* TODO: NYI - support for exported functions (PR 225654) */
probes[i].status = DR_PROBE_STATUS_UNSUPPORTED;
} else if (probes[i].insert_loc.type != DR_PROBE_ADDR_VIRTUAL) {
/* TODO: NYI - support for virtual addresses (PR 225658) */
probes[i].status = DR_PROBE_STATUS_UNSUPPORTED;
goto dr_probe_parse_error;
}
ppoint = HEAP_ARRAY_ALLOC_MEMSET(GLOBAL_DCONTEXT, hotp_patch_point_t, 1,
ACCT_HOT_PATCHING, PROTECTED, 0);
module->num_patch_points = 1;
module->patch_points = ppoint;
/* The actual patch address will be computed if the module matches.
* vul_id holds the function name, which will be used to compute the
* offset. See hotp_process_image().
*/
/* TODO: validate probe & callback addr here if possible; PR 225663 */
ppoint->offset = probes[i].insert_loc.lib_offs.offset;
ppoint->detector_fn = 0; /* No detector for probes. */
ppoint->protector_fn = probes[i].callback_func.lib_offs.offset;
/* Precedence hasn't been implemented yet, however, if it had been,
* then we don't want gbop hooks to interfere with client probes.
*/
# define HOTP_PROBE_PRECEDENCE (HOTP_ONLY_GBOP_PRECEDENCE - 1)
ppoint->precedence = HOTP_PROBE_PRECEDENCE;
/* id generation should be the last step because parsing of a
* probe can be aborted before that and we don't want an id being
* returned for a probe that is rejected.
*/
SELF_UNPROTECT_DATASEC(DATASEC_RARELY_PROT);
vul->id = GENERATE_PROBE_ID();
SELF_PROTECT_DATASEC(DATASEC_RARELY_PROT);
/* If we parsed a probe definition to this point then it is valid. */
valid_probes++;
continue;
dr_probe_parse_error:
/* Invalid probes are not kept inside dr, but discarded, so a 0 id
* should be returned for them
*/
probes[i].id = 0;
hotp_free_one_vul(vul);
}
/* If there were some invalid probes then free extra memory in the initial
* table allocation.
*/
if (valid_probes < num_probes) {
hotp_vul_t *old_tab = tab;
if (valid_probes > 0) {
tab = HEAP_ARRAY_ALLOC(GLOBAL_DCONTEXT, hotp_vul_t, valid_probes,
ACCT_HOT_PATCHING, PROTECTED);
memcpy(tab, old_tab, sizeof(hotp_vul_t) * valid_probes);
} else {
tab = NULL;
}
HEAP_ARRAY_FREE(GLOBAL_DCONTEXT, old_tab, hotp_vul_t, num_probes,
ACCT_HOT_PATCHING, PROTECTED);
}
hotp_load_hotp_dlls(tab, valid_probes);
/* Can't load dlls with hotp lock held - can deadlock if app is loading a
* dll too (see hotp nudge for details). We solve this by setting up the
* hotp table in a temp var, doing the load on it and then grabbing the
* hotp lock and setting the global hotp table. If we have our own loader
* (PR 209430) we won't need to do this.
*/
d_r_write_lock(&hotp_vul_table_lock);
SELF_UNPROTECT_DATASEC(DATASEC_RARELY_PROT);
GLOBAL_VUL_TABLE = tab;
NUM_GLOBAL_VULS = valid_probes;
if (GLOBAL_VUL_TABLE != NULL) {
ASSERT(NUM_GLOBAL_VULS > 0);
/* Policy status table must be initialized after the global
* vulnerability table is setup, but before module list is iterated
* over because it uses the former and the latter will set status.
*/
hotp_init_policy_status_table();
}
SELF_PROTECT_DATASEC(DATASEC_RARELY_PROT);
d_r_write_unlock(&hotp_vul_table_lock);
/* Unlike hotp_init(), client init happens after vmareas_init(), i.e.,
* after module processing, so we have to walk the module list again. It
* is ok to do the walk without the hotp lock because
* 1. that is what is done between hotp_init() and vmareas_init() as no
* change can happen to the hotp table at that time (nudges
* are nop'ed during dr init).
* 2. dr_register_probes()'s execution at init time can't overlap with
* another dr_register_probes() because
* a. a second instance can't be called in dr_client_main() before the first
* one returns.
* b. if a second one is called via a custom nudge or from a callback
* it is nop'ed by this routine (at least for this release - for
* next release will have to figure the callback part); btw, nudges
* during dr init are nop'ed anyway.
* 3. dr_register_probes()'s execution at init time also can't overlap
* with a liveshield nudge (even if we support them both
* simultaneously) because during dr init all nudges are nop'ed.
*
* NOTE: for probe/hot patch related nudges after dr init (whether custom
* nudge, liveshield nudge or internal nudge triggered by calling
* dr_register_probes() after init), loader walking has to be done with the
* hotp lock held otherwise two nudges can mess up each other (one common
* problem would be double injection/removal for hotp_only). This is done
* in nudge_action_read_policies() and hotp_nudge_handler() for
* liveshields.
*
* For probes, nudge (custom or internally triggered) isn't supported today
* - a TODO. When we do that this routine can't be shared as is for both
* probe registration at init and probe registration after init.
*/
if (GLOBAL_VUL_TABLE != NULL) {
/* TODO: opt: if it is safe move client_init() between hotp_init() &
* vm_areas_init() then this loader-list-walk can be eliminated.
* UPDATE: no it can't b/c this can be called post-dr_client_main()!
*/
/* XXX: to support calling post-dr_client_main() the actual num_threads needs
* to be passed (and should do a synchall): does PR 225578 cover this?
*/
hotp_walk_loader_list(NULL, 0, NULL, true /* probe_init */);
}
}
/* TODO: currently no status is set, probe status & LS status codes needed to
* be merged, status code groups have to be defined (invalid, wating to be
* injected, etc.) so nothing is returned. PR 225548.
*/
int
dr_get_probe_status(unsigned int id, dr_probe_status_t *status)
{
unsigned int i;
bool res = false;
/* For now, probes are supported iff probe api is explicitly turned on.
* Also, liveshields shouldn't be on when probe api is on.
*/
CLIENT_ASSERT(DYNAMO_OPTION(hot_patching) && DYNAMO_OPTION(probe_api) &&
!DYNAMO_OPTION(liveshields),
"To use Probe API, "
"-hot_patching, -probe_api and -no_liveshields options "
"should be used.");
if (!DYNAMO_OPTION(probe_api))
return res; /* be safe */
if (status == NULL)
return res;
*status = DR_PROBE_STATUS_INVALID_ID;
d_r_read_lock(&hotp_vul_table_lock);
for (i = 0; i < NUM_GLOBAL_VULS; i++) {
if (id == GLOBAL_VUL(i).id) {
*status = *(GLOBAL_VUL(i).info->inject_status);
res = true;
break;
}
}
d_r_read_unlock(&hotp_vul_table_lock);
return res;
}
#endif /* HOT_PATCHING_INTERFACE */
/* Got hotp_read_policy_defs() working, so this can be used for testing now. */
#ifdef HOT_PATCHING_INTERFACE_UNIT_TESTS
static void
hotp_read_policies(void)
{
dcontext_t *dcontext = get_thread_private_dcontext();
/* TODO: once this function starts reading from file/memory, the data below
* can be used as part of unit-hotpatch.c
*/
static hotp_patch_point_t pp1 = { (app_rva_t)0x673e, (app_rva_t)0x1010,
(app_rva_t)0x1010, (app_rva_t)0x6741, 0 };
static hotp_module_t mod1 = {
{ "ci_loop_test.exe", 0, 0x4241c037, 0 }, 1, &pp1, false, NULL
};
static hotp_set_t set1 = { 1, &mod1 };
static hotp_vul_info_t info1 = { 0 }, info2 = { 0 };
static hotp_patch_point_t pp2 = { (app_rva_t)0x440f, (app_rva_t)&hotp_nimda,
(app_rva_t)&hotp_nimda, (app_rva_t)0, 0 };
static hotp_module_t mod2 = {
{ "iisrtl.dll", 0x20190, 0x384399bc, 0x21000 }, 1, &pp2, false, NULL
};
static hotp_set_t set2 = { 1, &mod2 };
NUM_GLOBAL_VULS = 2;
hotp_vul_table = (hotp_vul_t *)heap_alloc(
GLOBAL_DCONTEXT,
sizeof(hotp_vul_t) * NUM_GLOBAL_VULS HEAPACCT(ACCT_HOT_PATCHING));
/* ci_loop_text.exe vulnerability */
GLOBAL_VUL(0).vul_id = "ci_loop_test-vul";
GLOBAL_VUL(0).policy_id = "ci_loop_test-policy";
GLOBAL_VUL(0).hotp_dll = "c:\\cygwin\\home\\bharath\\ci\\hotp_2.5.dll";
GLOBAL_VUL(0).hotp_dll_hash = NULL;
GLOBAL_VUL(0).hotp_dll_base = NULL; /* Runtime data! just for now */
GLOBAL_VUL(0).mode = HOTP_MODE_OFF;
GLOBAL_VUL(0).num_sets = 1;
GLOBAL_VUL(0).sets = &set1;
GLOBAL_VUL(0).info = &info1;
/* nimda vulnerability */
GLOBAL_VUL(1).vul_id = "nimda-vul";
GLOBAL_VUL(1).policy_id = "nimda-policy";
GLOBAL_VUL(1).hotp_dll = "hotp_2_5.dll";
GLOBAL_VUL(1).hotp_dll_hash = NULL;
GLOBAL_VUL(0).hotp_dll_base = NULL; /* Runtime data! just for now */
GLOBAL_VUL(1).mode = HOTP_MODE_OFF;
GLOBAL_VUL(1).num_sets = 1;
GLOBAL_VUL(1).sets = &set2;
GLOBAL_VUL(1).info = &info2;
}
#endif /* HOT_PATCHING_INTERFACE_UNIT_TESTS */
Reference in New Issue View Git Blame Copy Permalink