Summary:
This is the first of a few patches I have to improve the performance of dynamic module loading on Android.
In this first diff I'll describe the context of my main motivation and will then link to it in the other diffs to avoid repeating myself.
## Motivation
I have a few scenarios where opening a specific feature on an Android app takes around 40s when lldb is attached to it. The reason for that is because 40 modules are dynamicly loaded at that point in time and each one of them is taking ~1s.
## The problem
To learn about new modules we have a breakpoint on a linker function that is called twice whenever a module is loaded. One time just before it's loaded (so lldb can check which modules are loaded) and another right after it's loaded (so lldb can check again which ones are loaded and calculate the diference).
It's figuring out which modules are loaded that is taking quite some time. This is currently done by traversing the linked list of loaded shared libraries that the linker maintains in memory. Each item in the linked list requires its own `x` packet sent to the gdb server (this is android so the network also plays a part). In my scenario there are 400+ loaded libraries and even though we read 0x800 worth of bytes at a time we still make ~180 requests that end up taking 150-200ms.
We also do this twice, once before the module is loaded (state = eAdd) and another right after (state = eConsistent) which easly adds up to ~400ms per module.
## A solution
**Implement `xfer:libraries-svr4` in lldb-server:**
I noticed in the code that loads the new modules that it had support for the `xfer:libraries-svr4` packet (added ~4 years ago to support the ds2 debug server) but we didn't support it in lldb-server. This single packet returns an xml list of all the loaded modules by the process. The advantage is that there's no more need to make 180 requests to read the linked list. Additionally this new requests takes around 10ms.
**More efficient usage of the `xfer:libraries-svr4` packet in lldb:**
When `xfer:libraries-svr4` is available the Process class has a `LoadModules` function that requests this packet and then loads or unloads modules based on the current list of loaded modules by the process.
This is the function that is used by the DYLDRendezvous class to get the list of loaded modules before and after the module is loaded. However, this is really not needed since the LoadModules function already loaded or unloaded the modules accordingly. I changed this strategy to call LoadModules only once (after the process has loaded the module).
**Bugs**
I found a few issues in lldb while implementing this and have submitted independent patches for them.
I tried to devide this into multiple logical patches to make it easier to review and discuss.
## Tests
I wanted to put these set of diffs up before having all the tests up and running to start having them reviewed from a techical point of view. I'm also having some trouble making the tests running on linux so I need more time to make that happen.
# This diff
The `xfer` packages follow the same protocol, they are requested with `xfer:<object>:<read|write>:<annex>:<offset,length>` and a return that starts with `l` or `m` depending if the offset and length covers the entire data or not. Before implementing the `xfer:libraries-svr4` I refactored the `xfer:auxv` to generically handle xfer packets so we can easly add new ones.
The overall structure of the function ends up being:
* Parse the packet into its components: object, offset etc.
* Depending on the object do its own logic to generate the data.
* Return the data based on its size, the requested offset and length.
Reviewers: clayborg, xiaobai, labath
Reviewed By: labath
Subscribers: mgorny, krytarowski, lldb-commits
Tags: #lldb
Differential Revision: https://reviews.llvm.org/D62499
llvm-svn: 362982
Summary:
According to [C128] "Virtual functions should specify exactly one
of `virtual`, `override`, or `final`", I've added override where a
virtual function is overriden but the explicit `override` keyword
was missing. Whenever both `virtual` and `override` were specified,
I removed `virtual`. As C.128 puts it:
> [...] writing more than one of these three is both redundant and
> a potential source of errors.
I anticipate a discussion about whether or not to add `override` to
destructors but I went for it because of an example in [ISOCPP1000].
Let me repeat the comment for you here:
Consider this code:
```
struct Base {
virtual ~Base(){}
};
struct SubClass : Base {
~SubClass() {
std::cout << "It works!\n";
}
};
int main() {
std::unique_ptr<Base> ptr = std::make_unique<SubClass>();
}
```
If for some odd reason somebody removes the `virtual` keyword from the
`Base` struct, the code will no longer print `It works!`. So adding
`override` to destructors actively protects us from accidentally
breaking our code at runtime.
[C128]: https://github.com/isocpp/CppCoreGuidelines/blob/master/CppCoreGuidelines.md#c128-virtual-functions-should-specify-exactly-one-of-virtual-override-or-final
[ISOCPP1000]: https://github.com/isocpp/CppCoreGuidelines/issues/1000#issuecomment-476951555
Reviewers: teemperor, JDevlieghere, davide, shafik
Reviewed By: teemperor
Subscribers: kwk, arphaman, kadircet, lldb-commits
Tags: #lldb
Differential Revision: https://reviews.llvm.org/D61440
llvm-svn: 359868
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
This patch removes the logic for resolving paths out of FileSpec and
updates call sites to rely on the FileSystem class instead.
Differential revision: https://reviews.llvm.org/D53915
llvm-svn: 345890
Summary:
During the previous attempt to generalize the UUID class, it was
suggested that we represent invalid UUIDs as length zero (previously, we
used an all-zero UUID for that). This meant that some valid build-ids
could not be represented (it's possible however unlikely that a checksum of
some file would be zero) and complicated adding support for variable
length build-ids (should a 16-byte empty UUID compare equal to a 20-byte
empty UUID?).
This patch resolves these issues by introducing a canonical
representation for an invalid UUID. The slight complication here is that
some clients (MachO) actually use the all-zero notation to mean "no UUID
has been set". To keep this use case working (while making it very
explicit about which construction semantices are wanted), replaced the
UUID constructors and the SetBytes functions with named factory methods.
- "fromData" creates a UUID from the given data, and it treats all bytes
equally.
- "fromOptionalData" first checks the data contents - if all bytes are
zero, it treats this as an invalid/empty UUID.
Reviewers: clayborg, sas, lemo, davide, espindola
Subscribers: emaste, lldb-commits, arichardson
Differential Revision: https://reviews.llvm.org/D48479
llvm-svn: 335612
Summary:
The llvm version of the enum has the same enumerators, with stlightly
different names, so this is mostly just a search&replace exercise. One
concrete benefit of this is that we can remove the function for
converting between the two enums.
To avoid typing llvm::sys::path::Style::windows everywhere I import the
enum into the FileSpec class, so it can be referenced as
FileSpec::Style::windows.
Reviewers: zturner, clayborg
Subscribers: lldb-commits
Differential Revision: https://reviews.llvm.org/D46753
llvm-svn: 332247
In case we are building with xml enabled, the GetMemoryRegionInfo
function will send extra packets to query te extended memory map, which
the tests were not expecting.
Add an expectation for this to the test. Right now, it's just a basic
one which pretends we don't support the extension, however, it would be
also interesting the add a test which verifies the extension-enabled
case.
I also noticed that the test does a pretty lousy job of validating the
returned memory region info, so I add a couple of extra assertions to
improve that.
llvm-svn: 331374
Summary:
We've had a mismatch in the checksum computation between the sender and
receiver. The sender computed the payload checksum using the wire
encoding of the packet, while the receiver did this after expanding
un-escaping and expanding run-length-encoded sequences. This resulted in
communication breakdown if packets using these feature were sent in the
ack mode.
Normally, this did not cause any issues since the only packet we send in
the ack-mode is the QStartNoAckMode packet, but I ran into this when
debugging the lldb-server tests which (for better or worse) don't use
this mode.
According to the gdb-remote documentation "The two-digit checksum is computed as
the modulo 256 sum of all characters between the leading ‘$’ and the
trailing ‘#’", it seems that our sender is doing the right thing here.
Therefore, I fix the receiver the match the sender behavior and add a
test.
With this bug fixed, we can see that lldb-server is sending a stop-reply
after receiving the "k" in the same way as debugserver does (but we
weren't detecting this because at that point the connection was dead
already). I fix that expectation as well.
Reviewers: clayborg, jasonmolenda
Subscribers: mgorny, lldb-commits
Differential Revision: https://reviews.llvm.org/D44922
llvm-svn: 328693
Summary:
Gdb servers like openocd may send many $O reply packets for the client to output during a qRcmd command sequence. Currently, lldb interprets the first O packet as an unexpected response. Besides generating no output, this causes lldb to get out of sync with future commands because it continues reading O packets from the first command as response to subsequent commands.
This patch handles any O packets during an qRcmd, treating the first non-O packet as the true response.
Preliminary discussion at http://lists.llvm.org/pipermail/lldb-dev/2018-January/013078.html
Reviewers: clayborg
Reviewed By: clayborg
Subscribers: labath, lldb-commits
Differential Revision: https://reviews.llvm.org/D41745
Patch by Owen Shaw <llvm@owenpshaw.net>
llvm-svn: 322190
The recent UUID cleanups exposed a bug in the parsing code for the
jModulesInfo response, which was passing wrong value for the second
argument to UUID::SetFromStringRef (it passed the length of the string,
whereas the correct value should be the number of decoded bytes we
expect to receive).
This was not picked up by tests, because they test with 16-byte uuids,
for which the function happens to do the right thing even if the length
does not match (if the length does not match, the function does not
update m_num_uuid_bytes member, but that member is already 16 to begin
with).
I fix that and add a test with 20-byte uuid to catch if this regresses.
I have also added more safeguards into the parsing code to fail if we
cannot parse the entire uuid field we recieve. While testing the latter
part, I noticed that the "negative" jModulesInfo tests were succeeding
because we were sending malformed json (and not because the json
contents was invalid), so I make those tests a bit more robuts as well.
llvm-svn: 320985
Summary:
The classes have no dependencies, and they are used both by lldb and
lldb-server, so it makes sense for them to live in the lowest layers.
Reviewers: zturner, jingham
Subscribers: emaste, mgorny, lldb-commits
Differential Revision: https://reviews.llvm.org/D34746
llvm-svn: 306682
Summary:
It had a dependency on StringConvert and file reading code, which is not
in Utility. I've replaced that code by equivalent llvm operations.
I've added a unit test to demonstrate that parsing a file still works.
Reviewers: zturner, jingham
Subscribers: kubamracek, mgorny, lldb-commits
Differential Revision: https://reviews.llvm.org/D34625
llvm-svn: 306394
Instead of every test creating a client-server combo, do that in the
SetUp method of the test fixture. This also means that we can rely on
gtest to not run the test if the SetUp method fails and delete the
if(HasFailure) calls.
llvm-svn: 306013
Summary:
The changes consist of new packets for trace manipulation and
trace collection. The new packets are also documented. The packets
are capable of providing custom trace specific parameters to start
tracing and also retrieve such configuration from the server.
Reviewers: clayborg, lldb-commits, tberghammer, labath, zturner
Reviewed By: clayborg, labath
Subscribers: krytarowski, lldb-commits
Differential Revision: https://reviews.llvm.org/D32585
llvm-svn: 303972
This renames the LLDB error class to Status, as discussed
on the lldb-dev mailing list.
A change of this magnitude cannot easily be done without
find and replace, but that has potential to catch unwanted
occurrences of common strings such as "Error". Every effort
was made to find all the obvious things such as the word "Error"
appearing in a string, etc, but it's possible there are still
some lingering occurences left around. Hopefully nothing too
serious.
llvm-svn: 302872
This support was landed in r300579, and reverted in r300669 due to failures on the bots.
The failures were caused by sockets not being properly closed, and this updated version of the patches should resolve that.
Summary from the original change:
This patch adds IPv6 support to LLDB/Host's TCP socket implementation. Supporting IPv6 involved a few significant changes to the implementation of the socket layers, and I have performed some significant code cleanup along the way.
This patch changes the Socket constructors for all types of sockets to not create sockets until first use. This is required for IPv6 support because the socket type will vary based on the address you are connecting to. This also has the benefit of removing code that could have errors from the Socket subclass constructors (which seems like a win to me).
The patch also slightly changes the API and behaviors of the Listen/Accept pattern. Previously both Listen and Accept calls took an address specified as a string. Now only listen does. This change was made because the Listen call can result in opening more than one socket. In order to support listening for both IPv4 and IPv6 connections we need to open one AF_INET socket and one AF_INET6 socket. During the listen call we construct a map of file descriptors to addrin structures which represent the allowable incoming connection address. This map removes the need for taking an address into the Accept call.
This does have a change in functionality. Previously you could Listen for connections based on one address, and Accept connections from a different address. This is no longer supported. I could not find anywhere in LLDB where we actually used the APIs in that way. The new API does still support AnyAddr for allowing incoming connections from any address.
The Listen implementation is implemented using kqueue on FreeBSD and Darwin, WSAPoll on Windows and poll(2) everywhere else.
https://reviews.llvm.org/D31823
llvm-svn: 301492
The break the linux bots (and probably any other machine which would
run the test suite in a massively parallel way). The problem is that it
can happen that we only successfully create an IPv6 listening socket
(because the relevant IPv4 port is used by another process) and then the
connecting side attempts to connect to the IPv4 port and fails.
It's not very obvious how to fix this problem, so I am reverting this
until we come up with a solution.
llvm-svn: 300669
Summary:
This patch adds IPv6 support to LLDB/Host's TCP socket implementation. Supporting IPv6 involved a few significant changes to the implementation of the socket layers, and I have performed some significant code cleanup along the way.
This patch changes the Socket constructors for all types of sockets to not create sockets until first use. This is required for IPv6 support because the socket type will vary based on the address you are connecting to. This also has the benefit of removing code that could have errors from the Socket subclass constructors (which seems like a win to me).
The patch also slightly changes the API and behaviors of the Listen/Accept pattern. Previously both Listen and Accept calls took an address specified as a string. Now only listen does. This change was made because the Listen call can result in opening more than one socket. In order to support listening for both IPv4 and IPv6 connections we need to open one AF_INET socket and one AF_INET6 socket. During the listen call we construct a map of file descriptors to addrin structures which represent the allowable incoming connection address. This map removes the need for taking an address into the Accept call.
This does have a change in functionality. Previously you could Listen for connections based on one address, and Accept connections from a different address. This is no longer supported. I could not find anywhere in LLDB where we actually used the APIs in that way. The new API does still support AnyAddr for allowing incoming connections from any address.
The Listen implementation is implemented using kqueue on FreeBSD and Darwin, WSAPoll on Windows and poll(2) everywhere else.
Reviewers: zturner, clayborg
Subscribers: jasonmolenda, labath, lldb-commits, emaste
Differential Revision: https://reviews.llvm.org/D31823
llvm-svn: 300579
Summary:
This aims to verify the validity of the response from the debugging
server in GDBRemoteCommunicationClient::GetMemoryRegionInfo. I was
working with ds2 (https://github.com/facebook/ds2) and encountered a bug
that caused the server's response to have a 'size' value of 0, which
caused lldb to behave incorrectly.
Reviewers: k8stone, labath, clayborg
Reviewed By: labath, clayborg
Subscribers: clayborg, sas, lldb-commits
Differential Revision: https://reviews.llvm.org/D31485
Change by Alex Langford <apl@fb.com>
llvm-svn: 299239
If QPassSignals packaet is supported by lldb-server, lldb-client will
utilize it and ask the server to ignore signals that don't require stops
or notifications.
Such signals will be immediately re-injected into inferior to continue
normal execution.
Differential Revision: https://reviews.llvm.org/D30520
llvm-svn: 297231
Prior to MSVC 2015 we had to manually include this header any
time we were going to include <thread> or <future> due to a
bug in MSVC's STL implementation. This has been fixed in MSVC
for some time now, and we require VS 2015 minimum, so we can
remove this across all subprojects.
llvm-svn: 296906
Summary:
Communication classes use the Timeout<> class to specify the timeout. Listener
class was converted to chrono some time ago, but it used a different meaning for
a timeout of zero (Listener: infinite wait, Communication: no wait). Instead,
Listener provided separate functions which performed a non-blocking event read.
This converts the Listener class to the new Timeout class, to improve
consistency. It also allows us to get merge the different GetNextEvent*** and
WaitForEvent*** functions into one. No functional change intended.
Reviewers: jingham, clayborg, zturner
Subscribers: lldb-commits
Differential Revision: https://reviews.llvm.org/D27136
llvm-svn: 288238
the chrono library there uses long long as the underlying chrono type, but
defines int64_t as long (or the other way around, I am not sure). In any case,
this caused the implicit conversion to not trigger. This should address that.
Also fix up the relevant unit test.
llvm-svn: 287867
The mock server was listening for only one packet (I forgot to put a loop around
it), which caused the client to stall in debug builds, as the timeout there is
1000 seconds. In case of a release builds the test would just silently succeed as
the tested function does not check or report errors (which should be fixed).
This fixes the test by adding the server loop. Since the test was taking quite a
long time now (8s), I have added a parameter to control the amount of data sent
(default 4MB), and call it with a smaller value in the test, to make the test run
faster.
llvm-svn: 285992
Summary:
Most of the changes are very straight-forward, the only tricky part was the
"packet speed-test" function, which is very time-heavy. As the function was
completely untested, I added a quick unit smoke test for it.
Reviewers: clayborg, zturner
Subscribers: lldb-commits
Differential Revision: https://reviews.llvm.org/D25391
llvm-svn: 285602
This change does the following:
* Changes the signature for the continuation delegate method that handles
async structured data from accepting an already-parsed structured data
element to taking just the packet contents.
* Moves the conversion of the JSON-async: packet contents from
GDBRemoteClientBase to the continuation delegate method.
* Adds a new unit test for verifying that the $JSON-asyc: packets get
decoded and that the decoded packets get forwarded on to the delegate
for further processing. Thanks to Pavel for making that whole section of
code easily unit testable!
* Tightens up the packet verification on reception of a $JSON-async:
packet contents. The code prior to this change is susceptible to a
segfault if a packet is carefully crafted that starts with $J but
has a total length shorter than the length of "$JSON-async:".
Reviewers: labath, clayborg, zturner
Differential Revision: https://reviews.llvm.org/D23884
llvm-svn: 281121
The behaviour of FileSpec differed between host OS versions. Hardcode the path
syntax to posix, as we don't care about that in this test.
llvm-svn: 281025
Summary:
This adds the jModulesInfo packet, which is the equivalent of qModulesInfo, but it enables us to
query multiple modules at once. This makes a significant speed improvement in case the
application has many (over a hundred) modules, and the communication link has a non-negligible
latency. This functionality is accessed by ProcessGdbRemote::PrefetchModuleSpecs(), which does
the caching. GetModuleSpecs() is modified to first consult the cache before asking the remote
stub. PrefetchModuleSpecs is currently only called from POSIX-DYLD dynamic loader plugin, after
it reads the list of modules from the inferior memory, but other uses are possible.
This decreases the attach time to an android application by about 40%.
Reviewers: clayborg
Subscribers: tberghammer, lldb-commits, danalbert
Differential Revision: https://reviews.llvm.org/D24236
llvm-svn: 280919
After the reformat, the unittests do not compile due to missing due to redefinition errors
between PosixApi.h and ucrt/direct.h. This is a bit of a shot in the dark, as I have not tested
it on windows, but I am restoring the original include order, so it should hopefully fix it.
llvm-svn: 280793
*** to conform to clang-format’s LLVM style. This kind of mass change has
*** two obvious implications:
Firstly, merging this particular commit into a downstream fork may be a huge
effort. Alternatively, it may be worth merging all changes up to this commit,
performing the same reformatting operation locally, and then discarding the
merge for this particular commit. The commands used to accomplish this
reformatting were as follows (with current working directory as the root of
the repository):
find . \( -iname "*.c" -or -iname "*.cpp" -or -iname "*.h" -or -iname "*.mm" \) -exec clang-format -i {} +
find . -iname "*.py" -exec autopep8 --in-place --aggressive --aggressive {} + ;
The version of clang-format used was 3.9.0, and autopep8 was 1.2.4.
Secondly, “blame” style tools will generally point to this commit instead of
a meaningful prior commit. There are alternatives available that will attempt
to look through this change and find the appropriate prior commit. YMMV.
llvm-svn: 280751
This started as an effort to change StringExtractor to store a
StringRef internally instead of a std::string. I got that working
locally with just 1 test failure which I was unable to figure out the
cause of. But it was also a massive changelist due to a trickle
down effect of changes.
So I'm starting over, using what I learned from the first time to
tackle smaller, more isolated changes hopefully leading up to
a full conversion by the end.
At first the changes (such as in this CL) will seem mostly
a matter of preference and pointless otherwise. However, there
are some places in my larger CL where using StringRef turned 20+
lines of code into 2, drastically simplifying logic. Hopefully
once these go in they will illustrate some of the benefits of
thinking in terms of StringRef.
llvm-svn: 279917
Summary:
This is a preparatory commit for D22914, where I'd like to replace this mutex by an R/W lock
(which is also not recursive). This required a couple of changes:
- The only caller of Read/WriteRegister, GDBRemoteRegisterContext class, was already acquiring
the mutex, so these functions do not need to. All functions which now do not take a lock, take
an lock argument instead, to remind the caller of this fact.
- GetThreadSuffixSupported() was being called from locked and unlocked contexts (including
contexts where the process was running, and the call would fail if it did not have the result
cached). I have split this into two functions, one which computes the thread suffix support and
caches it (this one always takes the lock), and another, which returns the cached value (and
never needs to take the lock). This feels quite natural as ProcessGdbRemote was already
pre-caching this value at the start.
Reviewers: clayborg
Subscribers: lldb-commits
Differential Revision: https://reviews.llvm.org/D23802
llvm-svn: 279725
Summary:
The tricky part here was that the exisiting implementation of WriteAllRegisters was expecting
hex-encoded data (as that was what the first implementation I replaced was using, but here we had
binary data to begin with. I thought the read/write register functions would be more useful if
they handled the hex-encoding themselves (all the other client functions provide the responses in
a more-or-less digested form). The read functions return a DataBuffer, so they can allocate as
much memory as they need to, while the write functions functions take an llvm::ArrayRef, as that
can be constructed from pretty much anything.
Reviewers: clayborg
Subscribers: lldb-commits
Differential Revision: https://reviews.llvm.org/D23659
llvm-svn: 279232
This change adds the Process/gdb-remote gtests to the Xcode
build. It also adds a virtual method impl to the continuation
delegate that I added with the StructuredDataPlugin change.
llvm-svn: 279203
Summary:
Before this, each function had a copy of the code which handled appending of the thread suffix to
the packet (or using $Hg instead). I have moved that code into a single function and made
everyone else use that. The function takes the partial packet as a StreamString rvalue reference,
to avoid a copy and to remind the users that the packet will have undeterminate contents after
the call.
This also fixes the incorrect formatting of the QRestoreRegisterState packet in case thread
suffix is not supported.
Reviewers: clayborg
Subscribers: lldb-commits
Differential Revision: https://reviews.llvm.org/D23604
llvm-svn: 279040
Apparently clang will happily capture a const variable in a lambda without it being specified in
the capture clause. MSVC does not like that.
llvm-svn: 278925
