This code path has not been used for a while now, as mono_os_(mutex|cond)_destroy return -1 in case of error, and not EBUSY.
This is also the only code path that would check the return value, so we can simply remove the returned value, and assert on the error.
Previously the tests were written so that some failures might occur on the
threadpool in methods that did not catch exceptions. In that case the process
would terminate due to an unhandled exception. So rewrite the tests to catch
exceptions in any code that doesn't run on the main thread, and forward those
exceptions to the main thread.
Also wait for async operations to complete on the main thread, and remove
asserts that are not supposed to be hit.
behaviour when passed times within the lost period at DST start
When passed a time within this period .NET's IsDaylightSavingTime() returns
true while GetUtcOffset() returns the BaseUtcOffset without the DST delta
added. This patch changes the behaviour in Mono to do what .NET does.
* [mono-threads] Remove MonoThreadParm.creation_flags
* [mono-threads] Remove MonoThreadParm.priority and set to MONO_THREAD_PRIORITY_NORMAL by default
* [threads] Wrap use of mono_thread_info_(get|set)_priority
* [threads] Store priority on MonoInternalThread
* [threads] Import mono_thread_info_set_priority
* [mono-threads] Remove MonoThreadParm
[arm64] fixes around icache flushing
This fixes https://bugzilla.xamarin.com/show_bug.cgi?id=39859
The problem in `emit_thunk` and the workarounds for the Cortex A53 errata are unrelated to the bug above, but they don't hurt either.
The basic problem that we see is that the Exynos 8890 SoC (shipped for example in the world edition of the Samsung Galaxy S7) is a big.LITTLE architecture with two different CPUs, four of each. `__clear_cache` of GCC [1], the built-in that we use to flush the instruction cache, correctly reads the cache line sizes for data- and instruction cache from the CPU and then caches that result. However, the two CPUs on the Exynos 8890 have two different cache line sizes for the instruction cache (128 bytes vs. 64 bytes): If we happen to initialize the cache line sizes with the larger one, we only flush every other cache line if the code runs on the second CPU model.
Neither V8 [2] and LLVM [3] do caching of the cache line sizes, but just read it every time from the register. We have a slightly different solution, where we cache _and_ read the value every time and try to determine a global minimum across CPUs, in case we get scheduled to a different CPU by the kernel during the flushing loop.
[1] https://android.googlesource.com/toolchain/gcc/+/master/gcc-4.9/libgcc/config/aarch64/sync-cache.c#33
[2] https://github.com/v8/v8/commit/fec99c689b8587b863df4a5c4793c601772ef663
[3] https://github.com/llvm-mirror/compiler-rt/blob/ff75f2a0260b1940436a483413091c5770427c04/lib/builtins/clear_cache.c#L146
After sweep, depending on the computed allowance, we are freeing empty blocks so that we still keep around enough blocks to fill that allowance (to avoid repeated vmaps).
On both collectors this had the problem that the allowance also accounts for LOS space, which we shouldn't account for in major section heuristics (which means that in LOS intensive benchmarks we only free major sections after the LOS shrinks).
On the concurrent collector, where we have a negative allowance adjustment depending on the duration of the concurrent mark (in order to mimic serial memory usage), we can have very small allowances at the end of a major collection, even 0. This means that we would reserve no empty blocks for later use.
We solve these issues by reserving empty blocks relative to the amount of current alive blocks, which is also the basis of how we compute the allowance in the first place.
This reverts commit 2f88b8eb5a.
There is a subtle problem with turning the monitor thread into a simple
MonoThreadInfo: It creates worker threads via mono_thread_create_internal (),
which allocates the thread object from the GC. No thread can allocate from
the GC unless it's actually attached, which the worker thread is not when
created with just mono_threads_create ().
This handle should have a total of maximum 3 copies:
- in MonoThreadInfo->handle
- in the return value of mono_threads_create_thread
- in MonoInternalThread->handle (if it's a managed thread)
Before, there would always be 2 references (in MonoThreadInfo->handle and another another for MonoInternalThread->handle), even if it would be referenced at one place only (MonoThreadInfo->handle), which would lead to a handle leak.
We also need to duplicate the returned handle, to avoid having a use-after-free of the handle (in case the thread exits before we access that handle)
If you abort a thread that is doing a Monitor:Enter, if the syscall abort lands before
the target thread start waiting it will block in a non aletable way.
To address that, we start by waiting in an alertable way and if interrupted we
decide to continue or not based on whether an abort can start.
Additionally, when alertable, we CAS the thread state to catch early Thread:Abort
calls that set the state by did not managed to init it.
This fixes a rare deadlock on tests/monitor-abort.exe.
* [threads] Allocate StartInfo in create_thread
* [threads] Put start_notify semaphore in StartInfo
* [threads] Factor common attach code in mono_thread_attach_internal
* [threads] Use same registered mechanism as mono-threads
The tests that fail like this actually work just fine on Mono on Windows.
Added the NotDotNet category to the CurrentThread_Domains() test since it
fails on .NET.
* Avoid race condition between domain unload and threadpool.
Previous logic only ensured that domain->threadpool_jobs was 0.
There was a race condition as follows:
* worker jobs are queued into ThreadPoolDomain outstanding_request's
* domain unload starts
* domain->threadpool_jobs is 0 so unload proceeds
* queued jobs are processed from the ThreadPoolDomain outstanding_request's
* the domain can be set/entered by worker threads since
mono_domain_set only fails for domains with state of MONO_APPDOMAIN_UNLOADED,
while our domain is in the MONO_APPDOMAIN_UNLOADING state
* jobs are now being unexpectedly run in domain that is being shutdown
* Fix a hang when unloading the domain.
This change applies the diff from ludovic-henry to correct the hang.
https://gist.github.com/ludovic-henry/6c8a4bc8951091c0ef36df2e52d5e32e
For some reason preamble length was asumed to be 3 when cp 65001
instead of always looking att the preamble length of the encoding.
This resulted in responses cropped by 3 chars on windows.
* When doing async connect on windows, connect returns `WSAEINPROGRESS`
which needs to be returned to managed code for correct handling. However
on Windows, last error gets reset before `WSAGetLastError` is called.
This fix moves all `WSAGetLastError` calls up closer to the socket
operation to make sure correct error code is returned to the managed code.
* Making sure `TcpNoDelay` is set to false for IPv4 by default and
that `DualMode` is default for IPv6.
* Making sure option value is set to 0 before calling getsockopt
because on Windows, a `BOOL` option value (eg TcpNoDelay) is returned
as a one byte value and the other bytes are left untouched which resulted
in TcpNoDelay always seemed to be true.
* Fixed so `TF_REUSE_SOCKET` is set only if `reuse` parameter is set
when disconnecting a socket.
* [threads] Remove use of CREATE_SUSPENDED for domain unload thread
* [threads] Remove use of CREATE_SUSPENDED for managed thread
* [threads] Remove CREATE_SUSPENDED
If the sweeping mechanism has been triggered by a major collection, then the iteration will wait for the general sweeping to finish and also for the sweeping of all the blocks. If this isn't the case (we are doing a major collection) we will iterate over the current state of the heap, some objects being bound to die during the collection, and the callback must account for this.
It doesn't make much sense to explicitly attempt to iterate without trying to sweep blocks that can be swept.
We never enter this case since we never iterate over the major objects without requiring to sweep, while a sweep is running. Even if we would be in this case, we directly access the mark bits which can be cleared at the same time by another thread sweeping the block, so we could miss live objects.
It doesn't make sense to check the remset during a major since majors dont use the cardtable and we clear it at the start. So we will obviously have missing remsets. For the debugging concurrent case, we should use the mod union consistency checks instead.
We would need to deal with scanning forwarded objects and dead objects which is not worth it. Only verify the heap at the start of a collection, as indicated by the name of the flag.
When verifying the heap we need to have the set of pinned objects otherwise we can observe missing remsets if suspended between the store and the write barrier. If we don't have the set of pinned objects, allow missing remsets.
After https://github.com/mono/mono/pull/3452 was merged, the LinkSdkRegressionTest.SecurityDeclaration
test in XI started failing:
```
[FAIL] LinkSdkRegressionTest.SecurityDeclaration : FileIOPermissionAccess
Expected: null
But was: <System.Security.Permissions.FileIOPermissionAccess>
at LinkSdk.LinkSdkRegressionTest.SecurityDeclaration () [0x00033] in /Users/builder/data/lanes/1411/2a96404f/source/xamarin-macios/tests/linker-ios/link sdk/LinkSdkRegressionTest.cs:990
at (wrapper managed-to-native) System.Reflection.MonoMethod:InternalInvoke (System.Reflection.MonoMethod,object,object[],System.Exception&)
at System.Reflection.MonoMethod.Invoke (System.Object obj, System.Reflection.BindingFlags invokeAttr, System.Reflection.Binder binder, System.Object[] parameters, System.Globalization.CultureInfo culture) [0x00038] in /Users/builder/data/lanes/1411/2a96404f/source/xamarin-macios/_ios-build/Library/Frameworks/Xamarin.iOS.framework/Versions/git/src/mono/mcs/class/corlib/System.Reflection/MonoMethod.cs:305
```
Due to the added call in the PR the linker preserved FileIOPermissionAccess enum,
which is apparently something we want to avoid.
Since according to https://github.com/mono/mono/pull/3452#discussion_r76384945 the only point of calling the
FileIOPermission is validating that the path doesn't contain invalid characters (mostly an issue on Windows),
I decided to just copy those checks as it's way cleaner.
Reference source of WindowsIdentity is using `MAXIMUM_ALLOWED`
as desired level by default when getting current token.
Changing to same behaviour to enable duplicate and impersonation
to be made on the token retrieved by GetCurrentToken.
* This improves the marshaling for packed classes and class involved in an inheritance hierarchy.
* Similar changes were made in the Unity fork of mono across a number of commits.
* The tests here have been taken from the repo and converted to work as NUnit tests in Mono.
Refcatored Win32RegistryApi.SetValue() to accept uint/byte/etc values for
DWord registry values.
Added handling of ERROR_CHILD_MUST_BE_VOLATILE errors to GenerateException().
Centralized MarkedForDeletion error handling to GenerateException().
Fixed tests which create volatile subkeys in RegistryKeyTest which used the
same key names but didn't clean up properly.
On Windows the TearDown() cannot delete the temp folder since it contains
loaded assemblies. The next time the test is run it will create a temp folder
inside the previous folder. After a number of runs the temp folder path will
approach the ~260 max path supported on Windows and some of the methods will
start failing.
This patch changes the way the temp folder is created and reduces the depth in
the same way as AssemblyTest was patched to fix a similar issue in
b90617e7b7.
Removed the differing code paths in FixtureSetUp() when running on non-Windows
platforms vs when running Mono on Windows or on .NET. The path previously
taken only on non-Windows platforms has been verified to work as expected on
both Mono on Windows and on .NET.
It failed a couple of times according to the babysitter logs with the following error:
```
MESSAGE:
Expected and actual are both <System.Int32[10]>
Values differ at index [1]
Expected: -1
But was: 0
```
Bump timeouts a little bit to see if this helps.
environment variables
Relying on the HOMEDRIVE/HOMEPATH environment variables is fragile as those
variables don't always have to be set. One such scenario is when running on a
Jenkins slave. When not set g_get_home_dir() in eglib would return NULL which
led to a crash in ves_icall_System_Environment_InternalGetHome().
This patch changes g_get_home_dir() to first try to call
SHGetKnownFolderPath() to get the current user's profile directory (i.e. home
dir). If that fails it falls back to the USERPROFILE environment variable. If
that isn't available the function will fallback to the original behavior and
try with HOMEDRIVE/HOMEPATH.
Creating new sockets between each connect instead of
trying to reuse the client socket. The original tests
was not working in the MS .NET runtime and socket reuse
is not recommended to use unless in very special cases.
On Windows the test used URLs like `file://C:\tmp\foo` which the
Mono.Security.Uri class was incapable of parsing. System.Uri from
referencesource parses such URLs just fine. This patch fixes the Uri class to
be able to parse such URLs. It also changes the Zone class to not throw an
exception if the Zone.CreateFromUrl() method is passed an invalid URL. Instead
the zone is set to NoZone. This seems to be how Zone in .NET behaves.
There are a few places in the runtime that should not be aborted as it would render it unusable.
cctor execution was the initial trigger for this changeset, but then edge triggering of TAE turned out
to expose this issue on other places.
This patch introduces those regions and apply them to the following places:
ves_icall_System_Runtime_Remoting_Messaging_AsyncResult_Invoke: Not delivering the completed CB
screws up any completion logic the user setup up.
mono_unhandled_exception: Protect delivery of the unhandled exception. Can't fail during it or we'll
effectively double fault and abort.
A thread abort must not start if the thread is running a .cctor.
OTOH, if the cctor self aborts, it should start.
Threads now track how many cctors they are running in the cctor_exec_depth variable.
The abort machinery (in async_abort_critical) checks that flag and bail from
starting the abort.
The cctor code (in mono_runtime_class_init_full) ensures that cctor_exec_depth is bumped
around execution of the cctor itself. Additionally, error handling was cleaned up as it was pretty
inconsistent.
This callback allows the runtime to query mini on whether there's a block guard in play.
It allows for the runtime to correctly decide when to start an abort.
Call g_log_set_default_handler in mono_trace_set_log_handler and mono_trace_set_log_handler_internal to ensure that g_ logging is properly routed.
Additionally, change the internal logging callback to take the log message as a single string to simplify things.
The patch of mono/metadata/marshal.c makes sure that
Marshal.StringToHGlobalAnsi() allocates a memory area for the returned C
string which is at least as large as the length of the string object. This is
the behavior on non-Windows platforms. Without this patch the
MarshalTest.StringToHGlobalAnsiWithNullValues() test wrote outside of a memory
area and corrupted the heap which later on caused a heap corruption crash when
running the corlib test suite on Windows.
The patch of mono/metadata/mono-security.c fixes the freeing of the memory
returned by GetNamedSecurityInfoW(). When GetNamedSecurityInfoW() is used to
request a DACL one must also pass an argument for the ppSecurityDescriptor
parameter and this is the one that should be freed later on, not the pointer
to the DACL as the original code used to free. Before this patch
DSACryptoServiceProviderTest.UseMachineKeyStore() crashed because of a heap
corruption.
In defer mode, the debugger thread is started as soon as possible before
any clients are attached, but the debugger is_attached flag was being set
immediately. This would cause the runtime (and the System.Diagnostics.Debugger
class) to wrongly report a client as being attached.
Originally fixed in the Unity fork here: https://github.com/Unity-Technologies/mono/commit/797a210e0f2416d779c448bb9e03bb16933e261f
In defer mode, the debugger thread is started as soon as possible before
any clients are attached, but the debugger is_attached flag was being set
immediately. This would cause the runtime (and the System.Diagnostics.Debugger
class) to wrongly report a client as being attached.
Originally fixed in the Unity fork here: https://github.com/Unity-Technologies/mono/commit/797a210e0f2416d779c448bb9e03bb16933e261f
All ia64 compilers have intrinsics that we use. Also, this ia64 code was very
out of date (missing many functions we use now) so it wouldn't compile anyway.
When encoding a UTF-16 C# string into bytes using "latin-1"/"iso-8859-1" encoding we were
not properly taking surrogate pairs into account. MSDN[1] provides a good introduction about the topic.
For example, this string would fail to encode and throw an IndexOutOfRangeException: "\ud83d\ude0a "
However, this string would not: "\ud83d\ude0a" (note missing space at the end). More on that later.
In https://bugzilla.xamarin.com/show_bug.cgi?id=42887 it was noted that encoding the string
didn't cause the exception on Mono 2.10.8.1 and 3.2.8. Indeed, it doesn't happen until Mono 4.0
where we switched most of the string code to referencesources.
While looking at the output of `Encoding.GetEncoding("latin1").GetBytes("\ud83d\ude0a ")` on the
"working" Mono verions I noticed that they essentially returned wrong results, e.g.:
```
csharp> Encoding.GetEncoding("latin1").GetBytes("\ud83d\ude0a ")
{ 63, 32, 0 }
```
This is wrong, because a) there's only one 63 (i.e. the '?' question mark char which is used as
a replacement char whenever the target encoding doesn't support a specific char,
and b) there is a trailing '\0' which should not be there.
Checking on .NET confirms the expected result: { 63, 63, 32 } i.e. '?? '
Back to the IndexOutOfRangeException in Mono 4.0+: digging into Latin1Encoding.InternalGetBytes() it
turns out that when we detect a surrogate pair we were incrementing charIndex, but we didn't decrement count.
This means that after the fallback characters (the two '?') were added we would still try to encode one more
char in the loop even though we're already at the end of the string, resulting in the exception.
For the "\ud83d\ude0a" example we would not hit the `count > 1` code path which made it work in this case.
Adding the decrement fixes the bug and makes the code consistent with MonoEncoding.cs[2].
The reason why it "worked" in earlier Mono versions was that we had a very simple EncoderReplacementFallbackBuffer
implementation[3] which only output a single '?' for the surrogate pair:
```
csharp> var buf = Encoding.GetEncoding("latin1").EncoderFallback.CreateFallbackBuffer()
csharp> buf.Fallback((char)0xD800, (char)0xDC00, 0)
true
csharp> buf.GetNextChar()
'?'
csharp> buf.GetNextChar()
'\x0
csharp> buf.Fallback((char)0xD800, (char)0xDC00, 0)
true
csharp> buf.GetNextChar()
'?'
csharp> buf.GetNextChar()
'?'
csharp> buf.GetNextChar()
'\x0'
```
This means that for the "\ud83d\ude0a " example from above we would only add one '?' character
for the surrogate pair to the result buffer, then the space and then we'd be done, leaving the
uninitialized array value 0 in the buffer on index 2.
I added tests for Latin1Encoding, mostly copied from ASCIIEncoding with additions for these
specific cases.
Fixes https://bugzilla.xamarin.com/show_bug.cgi?id=42887
[1] https://msdn.microsoft.com/en-us/library/ms404377(v=vs.110).aspx
[2] https://github.com/mono/mono/blob/5b69e839e12b3d3512f310aa5a0e873346d7e50e/mcs/class/I18N/Common/MonoEncoding.cs#L62-L66
[3] https://github.com/mono/mono/blob/mono-3.12.0-branch/mcs/class/corlib/System.Text/EncoderReplacementFallbackBuffer.cs
Previously, we would try to satisfy the heapshot request every single time
process_requests () was invoked, but since heap_walk () would never unset
heapshot_requested, we'd end up doing GCs over and over, causing massive
overhead.
All events now come with a time field right after the event byte, and it is
encoded in the same way for all events (based on LogBuffer.time_base).
This allows readers to queue up events until a sync point event, then sort all
of those events by time, and process them in the order they were actually
written in.
Maintaining it for every format version is getting unwieldy and most readers
likely won't care about ancient format versions. Further, we haven't really
stabilized the format yet, so documenting every iteration of it doesn't make
a whole lot of sense. Finally, if somebody does care about older versions,
they can simply look in decode.c to see how mprof-report deals with them.
* Methods in backtraces are now properly encoded based on the method_base
rather than the ptr_base.
* The unused IL and native offset fields in TYPE_SAMPLE_HIT's backtrace frame
format have been removed.
* The unused flags fields in the backtrace format and in metadata events have
been removed.
* Changed a number of fields to be encoded as a single byte instead of LEB128.
These events basically establish points in the event stream where all previous
events can be processed and a complete view of the profiling state can be
formed. For instance, if an allocation event for object A of type T was seen
before a sync point, then it's guaranteed that the class load event for T also
came before that sync point (although it may have arrived after the allocation
event for A).
Other than timely arrival, there was never really a need to force buffers to be
flushed at these arbitrary points. Now that we flush all buffers periodically,
that reason is no longer relevant.
Heapshots have nothing to do with call depth, so they should be processed in
any case.
Also flush the local buffer in all enter/leave callbacks if needed.
This ensures that even for low-activity threads, it won't take ages before the
events from it arrive. We do this flush every 1sec (as with sampling of perf
counters).
The problem is this: Given an allocation of object O on address A1 in thread T1
and a subsequent GC move of O from A1 to A2 in T2, if T2's event buffer is
flushed before T1's buffer, the reader of the log file will a seemingly
nonexistent object being moved, and then see the actual allocation of that
object later. This makes it extremely difficult to keep track of the managed
heap.
The way we solve this is by flushing all thread-local buffers when the GC stops
the world. This way, we know that all prior allocations in the program have
been committed to the log file before any moves that are about to follow as
part of the collection. Similarly, we flush the GC thread's buffer before
restarting the world to ensure that all moves are committed before allocations
that may happen later (which could use some of the addresses freed up for use
due to the moves).
This is reasonably easy now that we have a global LLS of threads which are
attached to the profiler. However, we have to use some kind of lock to ensure
that buffers aren't being written to while we're trying to flush them, or that
a suspended thread was in the middle of modifying its buffer. For this, we use
a reader/writer spin lock of sorts, somewhat similar to the lock used by the
SGen binary protocol. When threads are modifying their own buffer (whether it
be writing to it or flushing it), they take the reader lock. This means that
many different threads can be operating on their own buffers without any lock
contention (which would kinda defeat the purpose of the thread-local buffers).
Now, when a thread -- such as the thread that initiated a GC -- wants to access
the buffers of other threads, it takes the writer lock. In doing so, all other
threads are prevented from modifying their own buffers. This gives the thread
holding the writer lock an opportunity to flush the buffers of other threads.
* Actual thread-local state is now stored in a MonoProfilerThread struct which
is stored in a single TLS slot per thread. This avoids storing things like
call_depth in LogBuffer. The current LogBuffer of a thread is also stored in
the MonoProfilerThread.
* All MonoProfilerThread structs are stored in a global linked list set. This
LLS is used to manually clean up the thread-local state of threads that, for
whatever reason, don't have the thread_end () callback invoked.
* Lifetime management of thread-local buffers is now much more straightforward.
* The thread-local list of referenced methods is now only allocated if needed.
* Fixed a MethodInfo memory leak in handle_writer_queue_entry ().
* Fixed some class load events that could be lost in the writer thread when
fetching the full name of a method.
* TLS slots are now actually freed on shutdown.
Make GC bridge perform well in the "double fan" scenario
The GC bridge currently performs VERY poorly when the object graph has a "double fan" shape. If M Java objects point to 1 C# object which points to N Java objects, the graph exported to Monodroid will currently drop the C# object node and replace it with (M\*N) edges. It is easy to write code where (M\*N) grows ludicrously big.
In testing on device, this patch solves that problem while leaving other cases unaffected. In testing on device, I found:
* All performance tests except double fan: No discernible performance difference after patch
* Double fan test with M=N=1000: Before the patch this took between three and seven seconds. After the patch this took between 0.2 and 0.3 seconds.
* Double fan test with M=N=4000: Before the patch this took anywhere from one to two minutes. After the patch, this took about 1.2 seconds.
* Double fan test with M=N=6000: Before the patch, this took so long I was not able to get it to ever complete. After the patch, this took about 1.3 seconds.
* Double fan test with M=N=20000: I did not attempt this test pre-patch. After the patch, it took about 5.6 seconds.
The commit messages describe the changes in some detail but the short version is:
* Add a mechanism for exporting non-bridged SCCs to the bridge client
* Turn that mechanism on whenever the product fanin*fanout grows too large
* Make the merge cache more aggressive
To function, this patch requires a corresponding change to monodroid. See https://github.com/xamarin/xamarin-android/pull/154
When running msbuild on mono against an architecture-specific
project that also defines satellite assemblies, the /platform
switch will be passed to al.exe. Mono's implementation of al.exe
does not support this switch, which results in a failed build.
This commit adds the /platform switch, which also required
changing the implementation from System.Reflection to
IKVM.Reflection because Mono's implementation of AssemblyBuilder
does not support the necessary options.
While testing this change, I also noticed a few attributes that
the MS version of al.exe will copy from "template" assemblies,
so I added support for those, as well.
SetUp() already sets up a temp folder for the test methods and TearDown()
handles cleaning up and taking into account that on Windows, loaded assemblies
cannot be deleted. However, some of the tests created temp files outside of
this folder and had their own cleanup code which didn't expect temp files to
be undeletable on Windows.
This patch removes any custom temp file creation/cleanup from individual test
methods. It also slightly changes the way SetUp() created the temp folder for
each test, now the paths will be shallower.
When AppDomain.Unload() is called a new thread is created and attached. When
this thread is done the thread handle is immediately closed using
CloseHandle() but the same handle is also passed to CloseHandle() later on by
the InternalThread destructor. On the second CloseHandle() call Windows will
abort the process with an "Invalid handle" error.
This patch removes the CloseHandle() calls in mono_domain_try_unload() which
is called by AppDomain.Unload(). The patch also wires up some more functions
in mono/utils/mono-threads-windows.c to set the MonoThreadInfo.handle field to
NULL when a thread has exited.
* Get rid of the arch-specific headers.
* Do verbose printing in one place.
* Simplify the cross compilation case to avoid repetition.
* Define all variables with a macro in a single header.
* Print detected hwcap info via g_print () so it's visible on mobile.
* Expose plain bool variables in the s390x implementation.
It doesn't work on nunitlite e.g. on XA:
```
src/corlib/Test/System.Runtime.InteropServices/MarshalTest.cs(268,11): error CS0117: `NUnit.Framework.Assert' does not contain a definition for `Less'
```
Previously every time we add a new runtime setting for the bridge
processor we have added a new callback for it. Because this is getting
unweildy, and also because some of these settings might have been
broken by recent init order changes, I created a new set_config
callback and merged the existing set_dump_prefix and enable_processing
with it.
hash_perturb was being mixed in in the wrong place. Fix that, also
make the hash function take some input from high bits on 64-bit
systems, also make the types around the hash more reasonable.
The names were fixed only for the subset of methods that are included in netstandard,
and in cases where netstandard differs from Desktop .NET we match with .NET.
* [runtime] Remove most of the contents of MonoDynamicGenericClass, the elements of the 'fields' array are not used anywhere, making all the other fields unneccesary too.
* [runtime] Remove the MonoDynamicGenericClass structure, merge it into MonoGenericClass. Add some comments.
On Windows we should check the path for invalid characters. Also expand
relative paths set using file:// URLs like .NET does. This patch also removes
the code which prepends "//" on Windows when a file:// URL is given as .NET
doesn't seem to do that in more recent versions.
When having a float variable in memory, we should always access its contents through the fload_r4/fstore_r4 funcs, which do the conversion between float and double and different memsize vs var size. It is common case, to have a temp float, have its address loaded with ldaddr and later do a load_r4_membase on that address to access the contents. If alias analyis lowers this load_r4_membase to an op_fmove directly on the register pair, this would in turn be decomposed into 2 simple int op_moves on soft float, which is wrong if these op_moves are converted to memory accesses.
Even though we register emulations for these opcodes, the native function doesn't not map to the opcode logic, being used instead as part of the opcode decomposition along with other instructions. Make sure we don't try to emulate these opcodes through the general emulation pass.
Add missing source file, sgen-os-coop.c to project.
Add source files and headers into correct filters (Header/Source Files).
Eliminated Visual Studio linker warning LNK4221 for libmonoruntime project generated
by empty source files under certain configurations, boehm-gc.c, lock-tracer.c,
null-gc.c and sgen-os-coop.c
If we happen to have the debugger or the log profiler enabled, we would run into a use before initialization case for the w32handle. The debugger or profiler would call `mono_gc_base_init` which would attach the current thread, and that would crash in `mono_w32handle_new` at w32handle.c:437 with `scan_mutex` being an invalid value (EINVAL).
FileNotFoundException on Mono on Windows and on .NET when target doesn't exist
FileInfoTest.MoveTo_DestFileName_DirectoryDoesNotExist() expects
FileNotFoundException to be thrown but on Windows a DirectoryNotFoundException
is thrown instead.
This patch changes MoveFile() in mono/io-layer/io.c to set the last error to
ERROR_PATH_NOT_FOUND rather than ERROR_FILE_NOT_FOUND when the destination
path doesn't exist. This results in a DirectoryNotFoundException being thrown
in managed code rather than FileNotFoundException. This is the behaviour
observed in .NET.
Also updated FileInfoTest.MoveTo_DestFileName_DirectoryDoesNotExist() to check
for DirectoryNotFoundException rather than FileNotFoundException.
The methods are already available in our implementations of those types so we can just forward the calls
to the "real" types instead of throwing NotImplementedException.
I audited all code in mcs/class/Facades and the only places where we still throw NotImplementedException
are ones where we really need to add a custom implementation.
Fixes https://bugzilla.xamarin.com/show_bug.cgi?id=43695
It's unlikely the ServiceController APIs will be supported on mobile since they don't really make sense there,
they were initially designed for controlling Windows Services.
If the declaring type of a nested type marked by an xml descriptor
is not otherwise marked, the members of the nested type will not
be preserved, because the linker will not process the nested type
(since the parent type isn't marked, MarkStep:InitializeAssembly
will skip it, and thus skip akk its nested types as well).
https://bugzilla.xamarin.com/show_bug.cgi?id=43658
If sub == 0, i.e. there are no repeated separators, this method just returns
the input string. On Windows, where DirectorySeparatorChar (\) is different
from AltDirectorySeparatorChar (/), this means that the call
Path.GetDirectoryName ("foo/bar/dingus")
won't return "foo\bar" as .NET would but rather returns "foo/bar".
This patch makes sure the '/' -> '\' conversion is run when
AltDirectorySeparatorChar characters have been seen in the input string, even
if sub == 0.
PathTest.GetDirectoryName() triggered this bug.
When a FileStream is opened with the FileOptions.Asynchronous flag set the
code in file-io.c will pass FILE_FLAG_OVERLAPPED to the CreateFile() function.
On Windows this means the file will be in asynchronous mode and when
ReadFile()/WriteFile() are called an OVERLAPPED struct has to be passed. Mono
doesn't do this which means that when reading/writing asynchronous FileStreams
Windows returns an ERROR_INVALID_PARAMETER and an exception is raised in
managed code.
This patch simply removes the code which passes the FILE_FLAG_OVERLAPPED to
CreateFile(). This will work since asynchronous I/O is handled in managed code
anyway, using threads. On the OS level the I/O is still synchronous. The patch
won't affect other platforms since the implementation of
ReadFile()/WriteFile()/etc in mono/io-layer/, which is used on non-Windows, is
always synchronous and ignores the FILE_FLAG_OVERLAPPED.
This bug was triggered by the
MonoTests.System.IO.BinaryWriterTest.AsynchronousModeWrites test.
This test assumes that `new DirectoryInfo (@"c:").Name == "C:\"` when run on
Windows but the actual value depends on the current directory. If current dir
is C:\temp the Name property will return "temp". Similar problem in the
FullName() and Parent() tests.
This patch sets the current directory to C:\ explicitly before running the
failing tests on Windows. The current dir is reset in TearDown().
In an effort to cut down on the number of different scripts we have that try to output NUnit xml.
It also fixes a bug in emitnunit.pl where it would not output anything when one of the tests crashes
the process, because then the final "Overall results: tests: 22425, 100% pass, opt combinations: 25"
line isn't printed and emitnunit.pl just exits. This happens on Alpine.
DoubleTest.Parse() expects "3 5" to be unparseable in the current culture and
expects an exception to be thrown. In sv-SE however "3 5" parses to 35 and
this test fails.
This patch sets the current culture to en-US when the DoubleTest.Parse() test
is run.
When running DirectoryInfo.EnumerateFileSystemInfos() on an empty directory
MonoIO.FindFirst() will return null and the method returns without closing the
handle which is opened by FindFirst(). This patch makes sure the finally block
which closes the handle is run even when FindFirst() returns null.
Also added a call to FindClose() and some more cleanup in the native
implementation of MonoIO.FindFirst() when an allocation fails.
This bug caused an exception in DirectyoryInfoTest.TearDown() and several
other TearDown() methods in the MonoTests.System.IO tests on Windows due to
RemoveDirectory() failing on an empty directory which still had an open handle
associated with it.
The targets depended on a feature that doesn't exist in
xbuild (https://bugzilla.xamarin.com/show_bug.cgi?id=43645)
and as a consequence would fail to add the generated file to
the build when the GenerateTargetFrameworkMonikerAttribute
target was up to date.
With the .NET 4.6.2 merge in https://github.com/mono/mono/commit/b6a30059db8319348ebb454cc303633dfabd6a29#diff-799d44d25450b5b84906cbe7095708a0L86
the RSAPKCS1SignatureDescription class switched from using SHA1CryptoServiceProvider as the DigestAlgorithm to SHA1Cng.
This uncovered an issue when running SignatureDescriptionTest in FullAOT, as we were missing an entry for SHA1Cng
in CryptoConfig.fullaot.cs and so we didn't find the algorithm.
Additionally, 4.6.2 added new variants of the SignatureDescription class using SHA256, SHA384 and 512 which
had to be added alongside the corresponding digest/signature URLs defined by https://msdn.microsoft.com/en-us/library/windows/desktop/dd979768(v=vs.85).aspx
Note that on mobile/FullAOT we're falling back to the *Managed variants of the SHA algorithms as the *Cng/*CryptoServiceProvider variants
aren't exposed there (except for SHA1CryptoServiceProvider since it was in the framework from the beginning) so we need to special case these in the tests.
They were all using the same .cs source files so just including the generic monotouch .sources
makes it easier to reason about and reduces the risk of forgetting to update one of the lists.
SetUp() calls uname() in libc to check whether the test is running on OS X. On
Windows this fails since libc isn't available.
This patch removes the call to uname() completely since the result wasn't used
by the test anyway.
* Fixing linker warning when building libmono DLL.
After splitting mono runtime into a static library used when building libmono
DLL, we got a linker warning since there were no object files left in the
libmono project. There was also a sematic issue with the current organization
of DllMain since it is located in driver.c that will be compiled into the static
library and then consumed by the linker building the DLL. In cases where the
static library was consumed directly it would still include DllMain entry point.
By splitting the DllMain method implementation into a separate specific windows file
we can both resolve the linker warning and remove the DllMain method implementation from the static
library and only include it when building the libmono DLL.
* Visual Studio projects support to optimally link mono applications using static libmono library.
Added a new property, MONO_USE_STATIC_LIBMONO that can be used to link some
mono applications, for example mono(-sgen).exe, towards static version of libmono.
This gives us the option to build a binary without dependencies on other runtime binaries
minimizing what’s needs to be deployed in order to run mono. This in combination with static linked c-runtime
(another option available through a VS property) will produce a mono runtime binary without
addition dependencies except .NET assemblies deployed by mono runtime users.
The default is still using the dynamic version of libmono as before.
* Visual Studio project to test/debug testdriver tests from within Visual Studio.
Added an additional Visual Studio project preconfigured to run testdriver related tests
directly from within Visual Studio.
Commit also includes an additional configuration script that can be used test projects in order
to setup a mono config file. The testdriver project uses this to create a unique configuration
files pointing to the build libtest.dll for running configuration. The configuration file will then
be used when the tests are executed to make sure the correct libtest.dll for current build configuration
gets used when running test different tests from within the Visual Studio debugger.
The project comes with a number of user macros as well that can be used in order to tailor how the tests are run.
Most of the values are preconfigured and shouldn't need to be changed in order to use
current build. In order to change the test that is executed, use the following user macro
defined in mono-testdriver-test property sheet added to the project:
MONO_TESTDRIVER_RUN_TARGET
and point it to the assembly hosting the tests to execute from within Visual Studio.
* Visual Studio project to test/debug nunit tests from within Visual Studio.
Added an additional Visual Studio project preconfigured to run nunit related tests
directly from within Visual Studio.
The project comes with a number of user macros as well that can be used in order to tailor how the tests are run.
Most of the values are preconfigured and shouldn't need to be changed in order to use
current build. In order to change the test that is executed, use the following user macro
defined in mono-nunit-test property sheet added to the project:
MONO_NUNIT_RUN_TARGET
and point it to the assembly hosting the tests to execute from within Visual Studio.
There is also an additional user macro defined:
MONO_NUNIT_FIXTURE
This can be used to limit the number of nunit test run within a test suite.
The merge_cache previously, if it encountered an SCC ("color") with
many outgoing links, would not attempt to merge that SCC with existing
SCCs and instead would just create a new color. Because unbridged SCCs
are now reported to the bridge client, the cost of creating extra
colors has increased. To mitigate this the merge cache now attempts to
merge objects of all sizes, and has been changed to be able to do this
performantly.
The biggest change is that in the case of a hash collision, SCCs will
sometimes be merged without fully checking if they are exactly equal.
This is considered okay because false merges will be infrequent and
nothing worse will happen than short-term failure to reclaim memory.
The bridge-pathologies double fan test has been amended to verify
large-object merging works.
Augments each tarjan-bridge SCC ("color") with a count of how many
SCCs are pointing to it. The mechanism for exporting non-bridged SCCs
uses this count to only export those SCCs where the "double fan"
scenario appears to be occurring. In testing this eliminates the
performance overhead imposed by the previous commit.
Also renamed/moved some variable declarations because namespace bugs
were creeping in.
Introduces a mechanism for non-bridged SCCs which lay between bridged
SCCs in the SCC forest to be reported to the bridge client. In this
version, such non-bridged SCCs are always reported. The result is that
most GCs get a little slower but GCs in which the "double fan" shape
appears (many bridged objects link to one C# object which links to
many bridged objects) become massively faster.
Because before clients were allowed to assume exported SCCs were
always bridged, SGEN_BRIDGE_VERSION has been incremented.
Windows
The error (ERROR_DIRECTORY) wasn't recognized in MonoIO so the error message
became "Win32 IO returned 267. Path: C:\foo\bar" which includes the path. The
test tests that the path isn't include in the message in the thrown
IOException.
Previously we were wrongly trying to achieve this by doing a load-acq followed by a store-rel. Assuming we have memory access on address A (before) and C (after) while the atomic operation happens on address B, the order of the instructions would look like this :
[A], [B]load, load-acq, store-rel, [B]store, [C]
In this case a reordering of memory accesses like the following is possible, which clearly breaks the desired constraints :
[B]load, load-acq, [C], [A], store-rel, [B]store
In order to provide the sequential memory constraint we could emit the barriers like this instead :
[A], membar, [B]load, [B]store, membar, [C]
We can additionally save an instruction and emit the barriers like this instead (this is also what gcc does) :
[A], [B]load, store-rel, [B]store, membar, [C]
In this case we only need to worry about the relation between memory accesses on A and the loading of B. We need to consider what happens if B is loaded before the access on A is fulfilled. This only matters if we load an older value for B (rather than the one visible by other threads) before A. In that case the erroneous store will fail due to using exclusive stores. At the retry the value read from B will be the right one.
Loads with acquire semantics guarantee that following memory accesses are not reordered with the load. Add full membar before it to ensure full sequencing.
Stores with release semantics guarantee that previous memory accesses are not reordered with the store. Add a full membar after it to ensure full sequencing.
NET_2_1 is also defined whenever we define MOBILE, so they have the same effect. Silverlight was 2.1 but it's been dead for a while now, and we do not even support it.
On Win64 Mono doesn't (yet) implement fast JIT TLS (mono_get_jit_tls_offset()
returns -1). This means that the code in
mono_arch_create_handler_block_trampoline() takes the code path in the "else"
which is the same code path taken when AOT compiling. The code generated in
this code path manipulates the stack pointer to ensure it's aligned properly
according to the ABI (RSP-8 is 16-byte aligned on function entry) for the AOT
case where the mono_handler_block_guard_trampoline() C function is called
directly from the handler_block_trampoline.
This patch unifies the JIT and AOT generated code to be as similar as possible
and gets rid of the fast path. There's no need for this code to be fast. The
handler_block_trampoline will after this patch call
mono_handler_block_guard_trampoline() directly, both in JIT and in AOT mode,
and then jump back to the saved return address itself rather than let the
MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD generic trampoline do this.
The tests mono/tests/finally_block_ending_in_dead_bb.exe and
mono/tests/block_guard_restore_aligment_on_exit.exe trigger this bug.
Split up reflection.c into three separate files:
1. reflection-query.c: This contains System.Type and the interrogative System.Reflection APIs.
2. reflection.c: This contains the System.Reflection.Emit APIs
3. reflection-custom-attrs.c: All about dealing with custom attributes.
Simplify code by always accessing tlab information and inside_critial_region through the SgenThreadInfo pointer. This also fixes potential failure to detect that we are in a critical region within the managed allocator.
MONO_THREAD_VAR_OFFSET definitions don't work when loading mono as a dynamic library. This means that we need to be conservative and don't aot code that contains these tls chunks.
In the future, we should remove HAVE_KW_THREAD altogether and use only pthread since it simplifies the code alot.
Added a couple of new/updated test/debug Visual Studio utility projects.
These projects can be used to quicker get up to speed when there is a need
to run/test/debug mono runtime tests inside Visual Studio debugger.
Added utility projects can be used to run mini regression tests using JIT as
well as compile/run full AOT:ed versions of mini regression tests or other assemblies.
The different .prop files are also organized so they can be reused when adding additional
run/test/debug projects going forward.
The code in mono_arch_create_handler_block_trampoline() in tramp-amd64.c emits
a relative CALL which calls the mono_amd64_handler_block_trampoline_helper()
functions via:
amd64_call_code (code, mono_amd64_handler_block_trampoline_helper)
This macro however only works properly for offsets less than 32-bits. For
64-bit offsets it truncates the offset to a 32-bit number. On Win64 the offset
in this case is always a number larger than 2^32. The emitted CALL will jump
to an illegal address causing a crash.
This patch changes the emitted code sequence to load the address of
mono_amd64_handler_block_trampoline_helper() to RAX and then calls the address
in RAX.
The test mono/tests/finally_block_ending_in_dead_bb.exe triggered this bug.
After this patch the test still crashes but in a new way related to another
bug.
* [corlib] Regression test for non-visible custom attributes.
See https://bugzilla.xamarin.com/show_bug.cgi?id=43291
* [reflection] Don't crash when building dynamic custom attributes on dynamic types (Fixes#43291)
1. Don't dereference a NULL custom attribute ctor, throw a type load
exception.
2. When building MonoCustomAttrInfo from an array of
MonoReflectinoCustomAttr*, compute the number of
non-visible (non-public) attributes correctly, and make sure to
iterate over all the attributes when populating the result array.
Fixes [#43291](https://bugzilla.xamarin.com/show_bug.cgi?id=43291)
* [reflection] Marginally better TLE for custom attrs
At least include a message about what went wrong when trying to
construct a custom attribute from a type that isn't finished yet.
I accidentally added this in 3c3ef17376 because I missed that on
full .NET this class inherits from NameValueCollection which provides `this[string name]` already
(the diff tool showed it as missing because it didn't take this into account).
The test fails due an ArgumentException being thrown when the catch handler in
the test's Run() method tries to write to the console. The ArgumentException
is thrown in EncodingTable.GetCodePageFromName(string) since the encoding name
passed to it is null. Normally this exception would be caught and the default
encoding would fall back to UTF8. However, since the thread is aborted the
ThreadAbortException will be rethrown and terminates the test with an uncaught
exception. I have verified that the same behavior is seen on OS X if I mock up
the default encoding code to throw the same ArgumentException.
This patch adds a call to Thread.ResetAbort() just before the call to
Console.WriteLine() in the catch handler. This removes the abort status of the
thread and avoids the rethrow of the ThreadAbortException.
We need a valid MonoThreadInfo in the TLS key while we run
unregister_thread because we may need to block the thread when it tries
to take the sgen GC mutex.
Fixes [#43099](https://bugzilla.xamarin.com/show_bug.cgi?id=43099)
The mono_threads_platform_get_priority()/mono_threads_platform_set_priority()
functions in mono-threads-windows.c assume the "handle" member has been set in
the MonoThreadInfo struct passed to it but it's always NULL. This patch
implements mono_threads_suspend_register() in the same file which sets the
"handle" member at thread registration time, like it's done in
mono-threads-posix.c.
This patch also makes sure the priority of a new thread is set as expected
when the Thread.Priority property has been changed prior to the thread being
started.
Also added a test method to mono/tests/priority.cs which tests that the
priority of the main thread is Normal by default and that it can be changed as
expected.
The changes in this patch fixes the failures in mono/tests/priority.exe on
Windows i386 and x64.
typeof (int).GetType ().GetType ().Name is now RuntimeType since https://github.com/mono/mono/pull/3045.
We did not notice this since most of our bots don't run with en-US locale set which this test requires.
We don't use svn anymore since a very long time and some diff implementations (e.g. in Alpine) don't support --exclude.
Removing the argument is an easy fix.
Previous implementation laid out the BSTR structure
in directly memory in C# code. This layout was not correct
for windows.
Furthermore memory was allocated using AllocCoTaskMem but later freed
using SysFreeString which on windows are different memory areas.
Third issue is that SecureString asumes big-endian when copying the chars
to its internal byte buffer. So on little-endian environments,
a swap of bytes is needed before the buffer is sent to native side
for BSTR conversion. This should probably be handled internally in
SecureString instead.
In the past there were wars between constified and unconstified
second parameter.. NetBSD went with constifying it. POSIX went without
const and most other OSes standarized. NetBSD is still having
incompatible function prototype.
LIBRARY
Standard C Library (libc, -lc)
SYNOPSIS
#include <iconv.h>
size_t
iconv(iconv_t cd, const char ** restrict src, size_t * restrict srcleft,
char ** restrict dst, size_t * restrict dstleft);
iteritems() was removed in py3, we can just use items() instead since we're iterating
over the list anyway so we don't need the lazy evaluation behavior of iteritems().
TODO: there's a gratuitous call to mono_lookup_internal_call_full().
Should really cache the uses_handle info along with the addr since the
lookup by name is rather involved.
Since UNICODE is defined when Mono is compiled under Windows the
GetLocaleInfo() macro will expand to GetLocaleInfoW() which returns the
current locale identifier in g_win32_getlocale() in 16-bit Unicode.
g_win32_getlocale() however expects an 8-bit encoding. By using
GetLocaleInfoA() explicitly we'll get the identifier back as an ordinary 8-bit
C string.
This patch fixes the mono/tests/threadpool-exceptions5.exe test.
MVID and AOTID metadata is no longer displayed at the end of the stack
trace.
Now, every stack frame contains a MVID and AOTID inlined.
An IL offset stackframe looks like:
at StackTraceDumper.Main () [0x00000] in <a3ec873a838c42c29f4eba7cab7d1dfb>:0
A native offset stack frame looks like:
at StackTraceDumper.<Main>m__0 () <0x100e15d50 + 0x00029> in <a3ec873a838c42c29f4eba7cab7d1dfb#b37a9d11247b811ffffdaff8090d8a59>:0
The crash would come from a use-after-free of stack memory: in coop_cond_timedwait_alertable, we would allocate the struct we pass to the mono_thread_info_install_interrupt function on the stack. But there is no guarantee as to when break_coop_alertable_wait is run. The fix is to allocate the data passed to mono_thread_info_install_interrupt on the heap, and free it in the callback, or if it hasn't been interrupted.
the stack.
Windows uses a guard page before the committed stack memory pages to detect
when the stack needs to be grown. If we suspend a thread just after a function
prolog has decremented the stack pointer to point into the guard page but
before the thread has been able to read or write to that page, starting the
stack scan at the thread's ESP/RSP will raise a STATUS_GUARD_PAGE_VIOLATION and
the process will crash.
This patch uses VirtualQuery() to determine whether the thread's stack_start
value points into the guard page. If it does the stack scan will instead be
started at the following, non-guard page.
Windows x64 only has 1 MB of stack per thread which is less than other x86
64-bit OSes. Using 10000 for "depth" in this test causes a stack overflow on
Windows x64. This patch changes "depth" to 5000 only on Windows x64. 5000 will
fit within the default stack size.
This is the recent behavior in .NET/MSBuild as well [1]:
> Starting in Visual Studio 2013, the MSBuild Toolset version is the same
> as the Visual Studio version number. MSBuild defaults to this Toolset
> within Visual Studio and on the command line, regardless of the
> toolset version specified in project file.
With this change xbuild will always use its own version instead
of the one specified in the project, unlessa ToolsVersion is forced
via the /tv switch on the command line.
Fixes https://bugzilla.xamarin.com/show_bug.cgi?id=42938
[1] https://msdn.microsoft.com/en-us/library/bb383796.aspx#Anchor_1
For test harnesses which don't support the full babysitter protocol,
the test harness can now report back paths to NUnit-format XML files
containing test results, or the paths to those XML files can be given
by whoever invoked the babysitter script. Failures listed in the XMLs
are included in the babysitter log.
test-runner.cs has been modified to report its XMLs, so we should now
see mono/tests failures in the babysitter log.
The Process.Exited event would be triggered too early, leading to crash in code using this event to access the process exit code. This is because in Process.ExitCode, we first check that the process has really exited.
ABI states that "All memory beyond the current address of RSP is considered
volatile".
On Windows the stack grows dynamically as the thread needs more stack space.
Windows places a guard page before the page at the top of stack and when this
guard page is accessed by the thread Windows will add a new page to the stack
and resume the thread.
Before this fix SGen on Win64 subtracted 128 from the stack pointer of a
suspended thread and started the stack scan from this address. In the case
when the stack pointer would be less than 128 from the start of the memory
page SGen would attempt to read from the previous page. If this happens to be
the guard page a STATUS_GUARD_PAGE_VIOLATION exception is raised and the
process crashes. The sgen-weakref-stress.exe triggered this bug on almost
every invocation.
We split mono-threads-{posix,windows,mach}.c functions in 2 categories:
- platform: depends on the host the runtime is compiled on (possible values: posix or windows)
- suspend: depends on the mono-threads suspend backend in use (possible values: posix, windows or mach)
This involves a few steps.
First define MonoAssemblyName in RuntimeStructs.cs
Then add some helped to decode blob arrays and handle hex digits.
Finally, move the actual implementation from native to managed.
And call it consistently in thread start functions when the managed
thread start function returns with an exception.
(If coop is not enabled, the exception handler unwinds the thread and
when there's no handler invokes the unhandled exception hook via
jit_tls->abort_func, so this is just for the coop case.)
Fixes mono/tests/unhandled-exception-1.cs for coop and bitcode.
Allows safe allocation of native utf8 char* strings from managed
strings.
using (var h = new SafeStringMarshal (str)) {
/* use h.Value IntPtr */
} // g_free'd by h.Dispose ()
On musl __GLIBC__ is not defined, so the conditional logic will
not produce correct result. Add a specific case to handle when
__GLIBC__ is not defined.
It showed the following error before:
```
In file included from wapi.h:23:0,
from wapi.c:2:
../../mono/io-layer/processes.h:218:31: error: unknown type name 'pid_t'
gchar* wapi_process_get_path (pid_t pid);
^
Makefile:603: recipe for target 'wapi.lo' failed
make[3]: *** [wapi.lo] Error 1
```
This is necessary on weird devices where the hwcap code might fail to detect
the ARM version it's running on (usually caused by differences in /proc/cpuinfo
output).
Some Android devices for some reason have a 'model name' line instead of a
'Processor' line. On such devices, we would fail to detect the architecture
version at all, and end up thinking that it's ARMv4 (even if the runtime
was compiled for ARMv7).
When we shutdown we set suspend_finalizers flag in metadata so we no longer run any finalizers. After recent changes, we avoid traversing gc hashes for finalizer enqueuing when this flag is set. The counterpart, popping the queues empty (even without running the finalizers) can still take some time, especially since we lock for each entry. Exit early from this loop if we no longer need to run finalizers.
Achieve this by setting a new flag inside sgen, instead of passing the address of suspend_finalizers all the time to the sgen api.
By asserting on error, we ensure that we return from these functions only when successfull, or when it makes sense to return an error (for example, timedwait return -1 only on timeout or interrupt).
By asserting on error, we ensure that we return from these functions only when successfull, or when it makes sense to return an error (for example, timedwait return -1 only on timeout).
The following scenario might happen:
- thread A: returns from tpdomain D1, then parks
- thread B: returns from tpdomain D1, D1 is unloading so it's removed with `domain_remove`
- thread A: unparks, calls `domain_get_next` with tpdomain D1
- domain_get_next asserts as it can't find tpdomain D1 in the list of domains
To fix that, we simply set the previous_tpdomain to NULL when unparking.
This was from when we did a nursery collection as part of the
finishing collection pause. Now that we don't do that anymore, we
underaccount major collection times.
The store/store fence incurs quite a high cost, so we only want it to
run when absolutely necessary. The most efficient way is to
separately compile nursery object operations for when the concurrent
collector is running, so there has to be no dynamic check.
In the future we might want to try batching updates of references to
moved objects so we have to do only one fence per batch.
The missing remset check uses the binary protocol to log missing
remsets it found. Debugging is easier if we don't have to manually
flush the binary protocol.
If a thread is interrupted after a reference write but before the corresponding
write barrier, the remset will be missing, but the object will be pinned.
After the latest code changes, we may land here with `errors == SslPolicyErrors.None`,
so we need to make sure that we actually set it to `RemoteCertificateChainErrors` if
the system validator returns an error.
(cherry picked from commit b4304354b1)
Alternative implementation to 8f092147bb.
In CI we explicitly pass EMIT_NUNIT=1 and check-seq-points to generate the nunit xml and run the tests,
locally those are omitted. This avoids needing to sniff for Jenkins/Wrench variables.
Same for a9eec21d62, but we already pass V=1 in Jenkins.
When doing the first phase of sweeping we tag the block pointer within the block list with a CAS, after which we own the block and we are free to update the state and do sweep computations on the block. When we finish this phase we directly write the untagged block pointer back in the list, making the block available to inspect by other threads. We need a write barrier before writting the untagged pointer back because, when another thread might acquire the block, it might encounter an invalid block state.
I noticed this while looking at why Mono.CSharp is registered as "crashed" in the the new
fullaot PR job on Jenkins. It turns out all tests were excluded in mobile_static which means
no NUnit xml was produced and that is exactly what the script then interprets as a crash.
I saw that ASTVisitorTest.cs was not included in the net_4_x tests as well, likely inadvertently.
It got bitrotten over the years so I had to fix it up a bit, now it compiles/runs on mobile_static as well.
Since acceptance-tests/Makefile (and llvm/Makefile) include scripts/submodules/versions.mk which evals versions.py which looks
for a SUBMODULES.json file in the current working directory we got confusing error messages if the build directory is not the src dir:
```
Traceback (most recent call last):
File "../../mono/scripts/submodules/versions.py", line 22, in <module>
submodules = json.load(open(CONFIG_FILE))
IOError: [Errno 2] No such file or directory: 'SUBMODULES.json'
```
Instead we now pass the path to SUBMODULES.json as a variable to versions.py and set it based on $(top_srcdir) so it works
even for out-of-tree builds.
In order for our exception handling to work on linux, we need the
unwind-dw2-fde-dip implementation in libgcc to work with our
.o file. This requires that `--eh-frame-hdr` be passed by the linker.
By using clang++ as the linker, we will get this and other flags
necessary for correct interop with c++ libraries.
Fixes issues where ResponseStream.Close would throw an exception when
peer had already closed the connection and ResponseStream.Close would
still try to communicating to peer that it was closing.
The issue was brought to our attention by failures on:
- MonoTests.System.Net.WebRequestTest.TestReceiveCancelation
- MonoTests.System.Net.HttpWebRequestTest.TestLargeDataReading
A common pattern that we use across the test suite is to create objects on a new thread, join the thread, collect then expect certain objects to be dead/finalized since we assumed that thread to be no longer a source of pinning. This is not always the case since, after joining, the thread might still be alive and running unmanaged code related to the thread shutdown, which can reuse the stack that previously contained the managed references.
The workaround is to still allocate objects far down the stack, even if we are on a separate thread.
At the moment the helper can be used in tests which run under the TestDriver (since it's included in TestDriver.dll) and can also be included as a helper source for test bcl classes that need it.
This reverts commit cf94a6fcd4.
It seems this causes something that prevents the chroot from unmounting in the Linux builds on Jenkins,
resulting in an endless "can't unmount" loop and stuck builds. Reverting to see if this fixes it.
For objects that die at the same time, normal finalizers should be run before critical finalizers. When we were unloading a domain we were running finalizers for all objects in that domain without checking whether the finalizer is critical or not. Now we enqueue the objects first in the finalization queues so we can run all the normal finalizers before the critical ones.
Fix for the following failing pinvoke tests in runtime tests suite on win x64:
* pinvoke11.exe
* winx64structs.exe
* pinvoke_ppcs.exe
* pinvoke_ppcc.exe
- Microsoft.WebApplication.targets was inadvertently removed.
- The new Microsoft.NuGet.Build.Tasks assembly sources were not getting dist'ed. Adding to DIST_ONLY_SUBDIRS fixes this for now,
though we can maybe add it to SUBDIRS instead (but neither me nor Marek know for what exactly that's used)
ILASM.exe has features which a mobile_static runtime will not support.
It is invoked with an external mono when used in the runtime.
We skip it here because otherwise it will fail to verify.
Wrapper stack frames don't have MVID or IL associated.
ex: (wrapper managed-to-native) System.Reflection.MonoMethod:InternalInvoke (System.Reflection.MonoMethod,object,object[],System.Exception&)
This fixes an inconsistency where xbuild would add System.Core
references to all builds, and MSBuild would only add them for
.NETFramework.
This change could break user XA/XI/XM projects on Mac if the user has
manually removed the explicit System.Core reference from the project,
however such projects would already be broken on Windows.
This is required for interoperating with project.json-based NuGet package configuration files
(not to be confused with DNX or AspNet Core which use it as a project file too) from MSBuild/xbuild.
The task hooks into the build process, resolves assembly references from the nuget package cache
and injects them into the compilation project so the csproj doesn't need to be modified when
adding new references/packages.
It is the first step on our road to PCL5/NuGet3.
Also add Mono.RuntimePropertyHandle struct representing a native MonoProperty pointer.
Also add ves_icall_System_Reflection_PropertyInfo_internal_from_handle_type icall
that wraps mono_property_get_object_checked.
Add new Mono.RuntimeEventHandle struct wrapping a pointer to a native MonoEvent
Add new icall
ves_icall_System_Reflection_EventInfo_internal_from_handle_type that's
mostly a wrapper around mono_event_get_object_checked.
This happens because current implementation is doing a direct call to Win32
GetLastError in order to get the value before it get clobbered by the icall
setting the last error value in mono. When running AOT all constructs that tries
to call a method using a direct address will be dropped since that will not work
in full AOT. This results in incomplete compiled libraries causing runtime exceptions
in full AOT.
The fix on Windows x64 is to introduce a new OP code, GET_LAST_ERROR that will
retrieve the last error without clobbering it in the process. This is done by
emitting code reading the value directly from the TEB (thread environment block)
and will be lowered to just a move from the segment register holding the TEB data
structure on Windows x64. This is how GetLastError is doing this as well, but since
we won’t be able to call the method in a “safe” way in full AOT, using the segment
register directly will solve the issue. We use the segment register for accessing TLS,
so we already relying on the TEB structure.
NOTE, this will only be applied to Windows x64, full AOT for now.
be detected at runtime. Since the process crash, lot of valuable information was never
written out to the console (like the managed stack trace) needed to effectively resolve
the issue.
Current implementation used different value type ABI’s depending on if call
was a managed->native, native->managed or managed->managed call. The implementation
was also overcomplicated since it was a tweaked copy from System V x64 value type
ABI with scenarios and conditions that wouldn’t apply to the Windows x64 value
type ABI. Having a uniform value type ABI when possible regardless of managed->native,
native->managed or managed->managed calls simplifies scenarios and reduce complexity.
While all the above are legitimate reasons the primary driving factor behind
the rewrite was a different issue. There are .NET wrapper libraries around DirectX API
used by game engines and other libraries that uses a pattern using a call into a stub
managed method that gets patched directly in the MSIL to a calli that will end up in a
native DirectX vtable method. In current ABI we can’t detect scenario like this, so arguments
are passed as if the call was a managed->managed call causing issues for value types. Since DirectX
uses value types a lot (vectors, matrixes etc.), this was a blocker getting these technologies
running on Windows using DirectX at all. Another approach could be to change the behavior of the
library, but this would require a major redesign of the library and since several of these libraries
are mature, that wasn’t a realistic option, especially since MS .NET runtime handles the libraries as is
without modifications.
The rewrite makes sure Windows x64 value type ABI follow the platform value type ABI in all scenarios.
Also:
1. MethodBase.GetMethodFromHandleInternalType_native icall
2. reimplemented the other MethodBase GetMethodFromHandleXXX methods
in terms of this new icall.
3. MonoError-ize mono_class_get_methods_by_name
The mutex on pending_bridge_callbacks was overkill because the init
model for the mono runtime already enforces the callbacks being set on
the init thread. Added a couple comments to formalize this.
If we have an invalid SP in the suspended thread, we don't set stack_start in the thread info, expecting it to be 0 in order to signal a resume and suspend retry to the suspender thread. If the thread was already restarted before (from a different reason like IP in critical section), it would have a nonzero stack_start from the previous suspend, leading to an assertion in the previously described case.
This is also what we do on mach.
On Windows, xbuild can fail with the somewhat obscure error message: "error : Error building target GetReferenceAssemblyPaths: UNC paths should be of the form \\server\share." This occurs because a call to Path.GetFullPath(@"\") returns the string "\" on Windows, when it should return a root path that includes the drive letter, like "C:\".
The code in xbuild tries to scan directories for files. Specifically, it looks for "\.*dll". If any .dll files exist in the root directory, each will be returned with the path "/\foo.dll". This path is used to create a FileInfo, and the exception about a UNC path occurs.
So if the path is "\" or "/" Path.GetFullPath should return the correct root path, including the drive letter, on Windows.
Here is the full call stack from xbuild:
Error building target GetReferenceAssemblyPaths: System.ArgumentException: UNC paths should be of the form \\server\share.
at System.IO.Path.InsecureGetFullPath (System.String path) <0x25343f0 + 0x003fb> in <filename unknown>:0
at System.IO.Path.GetFullPath (System.String path) <0x2534378 + 0x00013> in <filename unknown>:0
at System.IO.FileInfo..ctor (System.String fileName) <0x25fae20 + 0x00033> in <filename unknown>:0
at (wrapper remoting-invoke-with-check) System.IO.FileInfo:.ctor (string)
at System.IO.DirectoryInfo.GetFiles (System.String searchPattern) <0x25fa7c0 + 0x00087> in <filename unknown>:0
at (wrapper remoting-invoke-with-check) System.IO.DirectoryInfo:GetFiles (string)
input, Int32 ptr, System.IO.DirectoryInfo directory, Boolean recursive) <0x25fa2b0 + 0x0016b> in <filename unknown>:0
input, Int32 ptr, System.IO.DirectoryInfo directory) <0x25fa270 + 0x0002f> in <filename unknown>:0
at Microsoft.Build.BuildEngine.DirectoryScanner.ProcessInclude (ITaskItem include_item, System.Collections.Generic.Dictionary`2 excludedItems, System.Collections.Generic.List`1 includedItems) <0x25d72d0 + 0x002ab> in <filename unknown>:0
at Microsoft.Build.BuildEngine.DirectoryScanner.Scan () <0x25d70e0 + 0x000ff> in <filename unknown>:0
at Microsoft.Build.BuildEngine.BuildItem.Evaluate (Microsoft.Build.BuildEngine.Project project, Boolean evaluatedTo) <0x25d5a50 + 0x005d3> in <filename unknown>:0
at Microsoft.Build.BuildEngine.BuildTaskItem.Execute () <0x25e5ad8 + 0x0002b> in <filename unknown>:0
at Microsoft.Build.BuildEngine.TaskBatchingImpl.Execute (IBuildTask buildTask, TaskExecutionMode taskExecutionMode) <0x25e3698 + 0x0004a> in <filename unknown>:0
at Microsoft.Build. BuildEngine.TaskBatchingImpl.Run (IBuildTask buildTask, TaskExecutionMode taskExecutionMode, System.Boolean& executeOnErrors) <0x25e50e0 + 0x0004f> in <filename unknown>:0
at Microsoft.Build.BuildEngine.TaskBatchingImpl.Build (IBuildTask buildTask, TaskExecutionMode taskExecutionMode, System.Boolean& executeOnErrors) <0x25e21b0 + 0x00097> in <filename unknown>:0
at Microsoft.Build.BuildEngine.TargetBatchingImpl.RunTargetWithBucket (System.Collections.Generic.Dictionary`2 bucket, Microsoft.Build.BuildEngine.Target target, System.Boolean& executeOnErrors) <0x25e14b8 + 0x00213> in <filename unknown>:0
at Microsoft.Build.BuildEngine.TargetBatchingImpl.Run (Microsoft.Build.BuildEngine.Target target, System.Boolean& executeOnErrors) <0x25e1360 + 0x000f7> in <filename unknown>:0
at Microsoft.Build.BuildEngine.TargetBatchingImpl.Build (Microsoft.Build.BuildEngine.Target target, System.Boolean& executeOnErrors) <0x25dfee0 + 0x00053> in <filename unknown>:0
at Microsoft.Build.BuildEngine.Target.DoBuild (System.Boolean& executeOnErrors) <0x25dfcf0 + 0x0004f> in <filename unknown>:0
Enable compact seq points by default so that the stack trace have IL
offsets instead os native ones.
This changes will allow any stacktrace generated from now on to be
symbolicated using mono-symbolicate.
Deprecates options gen-compact-seq-points.
This is cleanup to the previous emergency fix for bug 42469. The goal of that fix was to defer bridge processor initialization until after sgen init, but it introduced a requirement mono_gc_register_bridge_callback be called before sgen init. This breaks Xamarin Designer, which sets up callbacks late.
This patch introduces an sgen_init_bridge () which is called both at startup and after callbacks are set, and performs the initialization once all pieces are in place. Additional concurrency and style issues are also addressed.
* [corlib] Rewrite RuntimeType.GetGenericParameterConstraints
using managed code for managed object manipulation by surfacing the
native MonoGenericParamInfo struct to managed code as Mono.RuntimeStructs.GenericParamInfo
* [corlib] RuntimeType::GetGenericParameterAttributes can be managed code now.
* [corlib] Fix x86 breakage
RuntimeTypeHandle.GetGenericParameterInfo returns a pointer
Not a struct that's the same size as a pointer. On the x86 Linux
calling convention they behave differently.
mono_gc_register_bridge_callbacks is a function which the android implementation calls before initializing the runtime. One of the things this function previously did was initialize the default bridge processor if none has been initialized yet. However, because the function gets called before initializing the runtime, a bridge processor was never initialized. This lead to two bad effects:
- If a bridge processor was ever manually specified, the bridge processor would get initialized twice
- Tarjan crashes if initialized before sgen, so when Tarjan became the default bridge processor Android always crashed on launch
The default bridge processor is now initialized at the end of sgen_gc_init.
MVID and AOTID stacktrace metadata no longer displays guids with '-'.
Metadata is now also using lowercase instead of uppercase.
Changed mono-symbolicate and mono --aot to use the minimalistic GUIDs.
Replaced static constructor with delayed initialization.
Fixed possible undeterministic order of dumped MVIDs.
Used AppendFormat when possible.
Added comment to StackTrace.AddMetadataHandler explaining that it can be
used with reflection by outside code, to avoid removing it or changing
its signature.
Added StringComparer.Ordinal to dictionary.
Extracted Symbolicate.Process stacktrace dump logic into DumpStackTrace
method. Which can also be used to dump the last stack trace after the
process loop.
mono-symbolicate can now be called:
symbolicate store-symbols <msym dir> [<dir>]+
This will store in <msym dir> all the managed symbols found in the
provided directories.
mono-symbolicate now parses stack traces with metadata at their end.
All the stack trace and metada are read, so that symbolicate can use
MVID and AOTID to retrieve the symbols from the msym directory..
The MVID metadata handler displays multiple MVID(s) followed by the
comma separated list of StackFrame lines whose method belongs to the
module with the displayed MVID.
Added a reflectable static method that can be used to customize data
displayed after a stack trace.
Added metadata handler that display the AOTID of the executing assembly.
AOTID is a GUID generated by the AOT compiler and stored along with the
native code.
The AOTID uniquely identifies the generated native assembly.
The AOTID is required when we want to map native offsets to IL offsets
and the sequence points mapping those values need to be determined.
Previously handles were pointer-sized structs. Unfortunately not all
calling conventions treat pointer-sized structs as pointers, so if we
wanted to interact with handles from wrappers, we would have to do
something complicated.
With the new design, `TYPED_HANDLE_DECL(MonoFoo)` will expand to:
```c
typedef struct {
MonoFoo *__obj;
} MonoFooHandlePayload;
typedef MonoFooHandlePayload *MonoFooHandle;
```
- A single dll can be mapped to different targets based on function name.
- The dllentry target can be omitted if it is the same as the source.
This change is released under the MIT license.
Exceptions.exe regression test started to fail on win x64 release builds.
Turns out that the Visual Studio C compiler picked a different optimization
for release builds (due to a change of library type that uses a different default optimization level)
using RSI register in native code calling into JIT:ed code. If the JIT:ed code used reflection
and exceptions, that triggers InternalInvoke and lmf_filter, the code only
stored System V x64 non-volatile registers in the unwind information. On Windows,
RSI is also a non-volatile register (RDI as well) that needs to be preserved and
since the code didn’t include those in the unwind information, the correct RSI value
was not restored during the unwind. Once the control was handled back to the native
code it dereferenced RSI (that was set to 0) and crashed.
The fix is to make sure we store all non-volatile registers on x64 Windows
when we save an lmf so the unwind code can restore them when unwinding through
the lmf.
* Making sure ownerThread is not nulled unless lockCount is zero.
* Fixed so Locked property checks lockCount when determining if locked
instead of _locked field which was never set.
* Added test for setting Locked property.
* Added test capturing the original issue.
I saw this failing occasionally on iOS on Wrench:
```
[FAIL] TimerTest.AutoResetEventFalseStopsFiringElapsed : #2 wrong elapsedCount
Expected: 1
But was: 0
at MonoTests.System.Timers.TimerTest.AutoResetEventFalseStopsFiringElapsed () [0x00091] in /Users/builder/data/lanes/1411/b07703ed/source/xamarin-macios/external/mono/mcs/class/System/Test/System.Timers/TimerTest.cs:327
at (wrapper managed-to-native) System.Reflection.MonoMethod:InternalInvoke (System.Reflection.MonoMethod,object,object[],System.Exception&)
at System.Reflection.MonoMethod.Invoke (System.Object obj, System.Reflection.BindingFlags invokeAttr, System.Reflection.Binder binder, System.Object[] parameters, System.Globalization.CultureInfo culture) [0x00038] in /Users/builder/data/lanes/1411/b07703ed/source/xamarin-macios/_ios-build/Library/Frameworks/Xamarin.iOS.framework/Versions/git/src/mono/mcs/class/corlib/System.Reflection/MonoMethod.cs:309
```
* Fix for non GCC win x86 calling convention when passing empty structs over pinvoke.
Port of https://github.com/mono/mono/pull/3085 to Winwdows x86 fixing failures
in pinvoke2 tests:
test_0_marshal_empty_struct
* Fix for incorrect implementation of small struct (single float/double) on Windows x86.
Failure in pinvoke3 test_0_marshal_small_struct_delegate9 is caused by incorrect
implemented small struct ABI on Windows x86.
Windows x86 ABI for returning structs of size 4 or 8 bytes (regardless of type)
dictates that values are passed in EDX:EAX register pairs,
https://msdn.microsoft.com/en-us/library/984x0h58.aspx.
This is different compared to for example float or double return types
(not in struct) that will be returned in ST(0),
https://msdn.microsoft.com/en-us/library/ha59cbfz.aspx.
Fix disables the alternative to do special handling of structs with 1 float or double
and make sure same logic is always applied to all small structs regardless type
as dictated by Windows x86 ABI.
* mono_test_marshal_delegate_ref_delegate uses incorrect calling convention.
mono_test_marshal_delegate_ref_delegate called by pinvoke3::test_55_marshal_delegate_ref_delegate
uses incorrect calling convention. Mono assumes delegates to be stdcall but
current definitions will be cdecl, this will cause an incorrect stack pointer
on architectures where cdecl and stdcall are different (Windows x86). Test fails
with "corrupt" stack pointer on Windows x86.
* Smaller adjustment based on feedback.
* Changed boolean struct member to bit field (on both x86 and x64)
* Changed compare code style from 0 == x to x == 0 to comply with coding standard.
libmono will use static library when building its dynamic library. This gives
the option to build binaries that could link the static version of mono runtime
or use the dynamic version of the same library. The default build binaries still
link against the dynamic version of the runtime.
The conflict happens when a type is declared like this:
.class value private auto ansi serializable sealed foo extends [mscorlib]System.Enum
The value keyword would make it a value type and ignore the parent declaration.
.net ilasm favors the later, which makes more sense.
I addressed the issue mentioned in the comments of mono_method_can_access_method:
* FIXME:
* with generics calls to explicit interface implementations can be expressed
* directly: the method is private, but we must allow it. This may be opening
* a hole or the generics code should handle this differently.
* Maybe just ensure the interface type is public.
The way to do it is to check access between gmds
Covers #41564
When multiple methods shared soap header a duplicate key exception would
be thrown.
This issue only occurs on XAMMAC and MOBILE where the non reference
source serializer is still used.
.NET supports this, from MSDN: "If an entry with the specified path and name already exists in the archive, a second entry is created with the same path and name.".
Fixes https://bugzilla.xamarin.com/show_bug.cgi?id=42219.
* [remoting] Convert mono_load_remote_field_new_checked to managed code.
* [corlib] Expose runtime struct MonoRemoteClass and MonoClass to managed
Also expose mono_class_get_type to managed code.
* [corlib] Add Mono.RuntimeClassHandle to encapsulate MonoClass*
and Mono.RuntimeRemoteClassHandle to encapsulate MonoRemoteClass*
* [runtime] mono_marshal_get_ldfld_remote_wrapper invoke the managed TransparentProxy.LoadRemoteFieldNew method directly
Don't bother going through a native call and back out to managed.
* [runtime] Get rid of MONO_WRAPPER_LDFLD_REMOTE
When the MONO_WRAPPER_LDFLD is operating on a proxy, it always invokes
the TransparentProxy::LoadRemoteFieldNew managed method
* [runtime] Remove dead mono_load_remote_field_new_icall
* [mini] Bump AOT file format version
Because we removed the LDFLD_REMOTE wrapper type so all the
MonoWrapperType enum values shifted.
[utils] Fix retiring of live block for the lock free allocator
Freeing a block descriptor (along with the block) can happen either when freeing a pointer in the block (we do this in order to prevent situations where we need to wait for the next alloc in order to free a block), either when we acquire an empty block while allocating. Either way, in order to free a block descriptor, we need to acquire it first.
When freeing a slot in a block, we are checking if the block is now empty in which case we acquire the block descriptor in order to free it (along with the block). While doing so, we were failing to account for the case where, by the time we make the block descriptor empty and the time we acquire it, another thread would try to allocate from the same descriptor, see that it's empty, free it, allocate a new descriptor (which would happen to have the same address as the retired one) to fulfill the allocation and then set it as the active descriptor for the allocator. When the first thread finally acquires the descriptor, we need to double check that it is still empty since it might be logically a different descriptor than the one we actually freed from. In the unlikely case it's not empty, we need to put it back in the allocator's data structures.
Fixes#41616
ConnectAsyncCallback now throws an exception when called a second time.
ConnectAsync reuses SocketAsyncEventArgs in BeginConnect which
already updates its state and calls ConnectAsyncCallback.
The changes make sure that ConnectAsyncCallback no longer called on
ConnectAsync catch block, while the state is still updated with the
exception.
Covers #41616
Our console runner is currently swallowing this test unhandled
exceptions.
To make it test show the failure I had to set legacyUnhandledExceptionPolicy
to 0 in mcs/class/lib/net_4_x/nunit-console.exe.config
Then running we can observe the crash by running:
make -C mcs/class/System check TESTNAME=System.Net.Sockets.SocketTest.ConnectAsyncUnhandledEx
Other tests appear to crash with legacyUEP set to 0, while the runner
shows no errors.
When we cleanup the gc at shutdown, we no longer care about running the remaining finalizers so we set the suspend_finalizers flag. This flag skipped running the finalizers but we would still traverse and modify the sgen finalizer hashes (on an arm box this could take even longer than 1s and lead to finalizer_thread_exited assertion failure since it's over the 100ms limit).
Avoid traversing sgen finalization hashes as soon as we see that we are shutting down.
libmonosgen-2.0.dll and breaks the build of monograph.exe which links against
it.
Also adds -O2 to CFLAGS on all platforms to align with how releases are built.
This should make it easier to programmatically check for errors, since it will remove the need to test the output string, which is an unreliable way of doing things.
Freeing a block descriptor (along with the block) can happen either when freeing a pointer in the block (we do this in order to prevent situations where we need to wait for the next alloc in order to free a block), either when we acquire an empty block while allocating. Either way, in order to free a block descriptor, we need to acquire it first.
When freeing a slot in a block, we are checking if the block is now empty in which case we acquire the block descriptor in order to free it (along with the block). While doing so, we were failing to account for the case where, by the time we make the block descriptor empty and the time we acquire it, another thread would try to allocate from the same descriptor, see that it's empty, free it, allocate a new descriptor (which would happen to have the same address as the retired one) to fulfill the allocation and then set it as the active descriptor for the allocator. When the first thread finally acquires the descriptor, we need to double check that it is still empty since it might be logically a different descriptor than the one we actually freed from. In the unlikely case it's not empty, we need to put it back in the allocator's data structures.
As described in https://bugzilla.xamarin.com/show_bug.cgi?id=34883 at
least Samsung SG3 doesn't allow access to the local network
interface (in fact, even ifconfig doesn't work on the device). Therefore
the test needs to be disabled on those devices. This is the first part
of the fix, the second one will be implemented in Xamarin.Android test
suite which will provide the AndroidShouldPingWork method which will
check the device id against list of those that do not work for this
test.
Part #1 of fix for https://bugzilla.xamarin.com/show_bug.cgi?id=34883
MXR is not coming back (see https://bugzilla.mozilla.org/show_bug.cgi?id=1279952 and https://github.com/SuaveIO/mono-script-buildpack/issues/9#issuecomment-226749314)
and this broke everyone who still relied on mozroots. Fix by using the version from the hg repository instead.
Starting from Mono 3.12 we've been shipping cert-sync to import CA certificates from the local computer trust store to Mono's trust store,
we strongly encourage everyone to move to that instead. Added a deprecation warning to mozroots to further strengthen this message.
Note that when fetching the CA list over the new https URL we need to skip cert validation errors, as the common use case for mozroots is when
no certs are present on the system (bootstrapping). This enables MITM but we don't really have any other options. Another reason why cert-sync is superior.
In some cases the "git log -1 <remote>" check fails even though the remote is configured
and an unnecessary fetch is done. Adding the branch to the check fixes this.
* An enhancement to the thread priority work added earlier. This will support starting a thread with a specified priority. Currently, the thread will start with the priority of the creator. The semantics of threads allow you to specify the priority of a thread before it has started. Therefore, we take note of this value at thread creation time.
* Missed committing this file the first time
* Correct typo
* Fix creation_flags reference
* MonoThreadParm really should not be in objects-internal
* Place priority in MonoInternalThread where it can be accessed by GetThreadPriority(). This enablses the priority to be set/retrieved before the thread is started. It also ensures that new threads default to a priority of THREAD_PRIORITY_NORMAL rather than inheriting the parent thread's priority.
* Revert changes to MonoInternalThread/Thread.cs. Properly use the handle for thread lookup. Initialize priority at thread creation time. Use stored value of priority if thread has not been started.
* Add code to handle retrieving the thread priority from a started thread that had been set before thread had started. This is necessary as before the thread is started the priority value can be manipulated. This value is stored in the MonoThread structure. Before the thread is started the value can be interrogated so we return this value rather than invoking the GetThreadPriority() API. When a thread is started this value is used to set the threads priority (if scheduling policy is SCHED_FIFO or SCHED_RR). For POSIX threads, after the thread is started we now keep the priority in the _Wapi_thread structure which is located via the thread handle. This structure doesn't exist until after the thread is started so we can't store the value there but after we can cache the priority value there and return it for non-SCHED_[RR|FIFO] threads. This way the expected value will always be set and returned by ThreadPriority property (currently for these threads it will simply return Normal). A test case has been added to the mono/tests directory.
* Use a copy of creation_flags so local modifications aren't propagated
* Use pthread_setschedparam as OSX doesn't support pthread_setschedprio
This fixes the bug where form authentication tickets do not expire until 50 years later.
I bumped into this bug after upgrading mono from 4.2.2 to 4.4.0
Firstly, the authentication ticket's expiry must always equal to the timeout attribute on the <forms/> element. It doesn't matter whether persistent or session cookies are used.
Secondly, if the cookie is persistent, then it's expiry should be set to the same as that of the authentication ticket.
Reference: https://blogs.msdn.microsoft.com/dansellers/2006/02/15/change-to-asp-net-2-0-forms-authentication-persistent-cookies/
This should fix some problem with Nuget archives created by Mono.
No test case for this one because it would involve writing a mini-ZIP file parser to make sure the value is written correctly (or hardcore the value offset in the generated file, which could lead to a very flakey test).
Fixes https://bugzilla.xamarin.com/show_bug.cgi?id=41775.
Updated Visual Studio build uses a different default build location of modules
compared to previous version. The tests config file includes a relative
path to libtest used for by all pinvoke tests. Since the module is now
placed at a different location the load failed during the test.
naming for intermediate pdb files for exe/dll's.
All artifacts will now have the same name of their final pdb file as
the library/dll/exe target file name.
The problem with the naming turns out to be a problem in Visual Studio
when a exe/dll use the same name of its intermediate pdb files and final
pdb file. In that case the final pdb file produced by the linker will be
incorrect and not usable by the debugger. This was the reason for the initial
change, giving intermediate pdb files a different name. Moving back to the
default naming for intermediate pdb files for exe/dll's will give the same
effect (a different name) and since libraries are not affected by this
problem, libraries can use the target name for their pdb files as well.
Also added support to not always replace config.h/version.h since that triggers
unnecessary rebuilds during development. New script will check if there is
an actual difference in the config.h/version.h file before doing the copy.
* Added regression test running mono/mini --regression [regression_tests]
* Added regression test running mono/mini --aot [regression_tests]
* Added default debugging options to run regression tests from within visual studio debugger.
* On new build system, setting up include folder if not availalble.
All changes applies to debug/release 32/64 bit.
* Added support for user macros in order to set build/install/sgen.
* Added support for more arguments to bat files + support for quoted paths.
* Added support for verbose/quite output in bat files.
* Moved libmono xcopy post build step into separate bat file.
* Cleaned some sgen configs still left in vcxprojs.
* Enabled build target suffix in order to lable a build (primarly for sgen).
* Enabled separate build tree.
* Dropped all sgen configurations, replaced by properties and target suffix.
* Fixed pdb naming of intermediate pdb files.
* Upgraded all projects to new vs2015 tool chain.
* Used mono.props through property manager instead of import.
* Fixed dependencies to be more explicitly in additional dependecies.
* Fixed error in mono.props, switched $ to % when referencing internal variables.
* Added property variables for includes and libraries, defined in mono.props.
The reason why mkbundle is not working is that for some reason
this constant expression is producing different code than the
-24 that it is supposed to.
What is odd is that the return value from the lseek system
call is the same, but yet, read fails with an end-of-file with
this value, and otherwise it passes.
With this call:
lseek (fd, -(16+sizeof(uint64_t)), SEEK_END)
strace displays:
lseek(0x3, 0xFFFFFFE8, 0x2) = 19375512 0
With this call:
lseek (fd, -24, SEEK_END)
strace displays:
lseek(0x3, 0xFFFFFFFFFFFFFFE8, 0x2) = 19375512 0
Now, while both results from lseek are the same, a subsequent read
with 24 bytes to read returns zero on the former case, and 24 on
the latter case.
When building a mobile_static build, we need to pass a lot of flags
which disable features. These make the net_4_x profile not build
correctly when using the official jenkins script. This helps.
on Windows. Updates the tests-config.in file with Windows specific <dllmap>
elements that map 'libtest' to the MSVC built libtest.dll for the current
architecture. test-runner.cs has been modified to accept a --config option to
be able to use a specific config file. Makefile.am now configures test-
runner.exe to use the tests-config config when running the tests.
Parsing ```default-allowance-ratio``` in ```MONO_GC_PARAMS``` is completely broken and accepts only one value - 1.0, while it should accept range of 1.0-10.0. We can clearly notice that when setting ```default-allowance-ratio``` to e.g. ```2.0``` and trying to run any program through Mono runtime:
```
Warning: In environment variable `MONO_GC_PARAMS': `default-allowance-ratio` must be between 1.00 - 1.00. - Using default value.
```
Therefore, correct this for using proper 1.0-10.0 range.
The do_try_exec_main function is always called in contexts where any
exception from doing the invoke is going to be stored in the exc outarg.
That means any MonoError will be handled locally by do_try_exec_main.
So we don't need a MonoError argument, we just need an inner_error.
We're seeing it randomly fail on Wrench with durations of 1550 - 1800ms.
Talked to Zoltan, this might be normal so bumping the timeout to reduce the failure rate.
The llvmonly backend will use `mono_aot_init_gshared_method_this`
and other initializers to initialize methods. These will transitively call
`mono_icall_get_wrapper_full` when icall wrappers are needed.
The init functions will set the
initialized flag after initialization. They set it in a non-atomic manner. This means that
in a multithreaded environment, it's possible to call the functions
twice.
When `mono_icall_get_wrapper_full` is called without this patch, the
pointers to wrappers it gets from `mono_marshal_get_icall_wrapper`
will not have the same addresses. They will hash to the same value
though.
This will lead to them not being found or inserted by
`find_aot_method_in_amodule`.
The fact that we already have a value with this hash means that
the new wrapper doesn't get inserted. The fact that the pointers differ
mean that we don't say that we've found our desired wrapper.
This leads to an assertion in `mono_icall_get_wrapper_full`
crashing the runtime.
Since we intend to have the semantics of a wrapper for each icall,
this change uses a cache to ensure uniqueness. The observed crashes
do not happen when this change is applied.
MONO_EXECUTABLE env variable to test-runner.exe using the --runtime command
line parameter. Defaults to 'mono' which is also the defult in test- runner.cs
if no --runtime argument is given.
* [runtime] Move 'pedump' to tools/pedump, build tools by default.
* Add missing file.
* [runtime] Mark mono_opcodes[] with MONO_API.
* [runtime] Build monograph when sgen is supported instead of boehm.
* [runtime] Link monograph against the static runtime libraries.
* [runtime] Use the same SUBDIRS/DIST_SUBDIRS for the CROSS_COMPILING case.
We're seeing the following exception on devices in CI:
```
[FAIL] HttpListenerRequestTest.HttpRequestIsLocal : System.Net.Sockets.SocketException : Could not resolve host 'XQA-AppleTV-1'
at System.Net.Dns.Error_11001 (System.String hostName) [0x00000] in /Users/builder/data/lanes/1381/d8287824/source/xamarin-macios/_ios-build/Library/Frameworks/Xamarin.iOS.framework/Versions/git/src/mono/mcs/class/System/System.Net/Dns.cs:292
at System.Net.Dns.GetHostByName (System.String hostName) [0x00024] in /Users/builder/data/lanes/1381/d8287824/source/xamarin-macios/_ios-build/Library/Frameworks/Xamarin.iOS.framework/Versions/git/src/mono/mcs/class/System/System.Net/Dns.cs:419
at System.Net.Dns.GetHostEntry (System.String hostNameOrAddress) [0x00061] in /Users/builder/data/lanes/1381/d8287824/source/xamarin-macios/_ios-build/Library/Frameworks/Xamarin.iOS.framework/Versions/git/src/mono/mcs/class/System/System.Net/Dns.cs:380
at System.Net.Dns.GetHostAddresses (System.String hostNameOrAddress) [0x00065] in /Users/builder/data/lanes/1381/d8287824/source/xamarin-macios/_ios-build/Library/Frameworks/Xamarin.iOS.framework/Versions/git/src/mono/mcs/class/System/System.Net/Dns.cs:406
at MonoTests.System.Net.HttpListenerRequestTest.HttpRequestIsLocal () [0x00006] in /Users/xamarinqa/QABot/data/lanes/2441/d8287824/source/xamarin-macios/external/mono/mcs/class/System/Test/System.Net/HttpListenerRequestTest.cs:197
at (wrapper managed-to-native) System.Reflection.MonoMethod:InternalInvoke (System.Reflection.MonoMethod,object,object[],System.Exception&)
at System.Reflection.MonoMethod.Invoke (System.Object obj, System.Reflection.BindingFlags invokeAttr, System.Reflection.Binder binder, System.Object[] parameters, System.Globalization.CultureInfo culture) [0x00038] in /Users/builder/data/lanes/1381/d8287824/source/xamarin-macios/_ios-build/Library/Frameworks/Xamarin.iOS.framework/Versions/git/src/mono/mcs/class/corlib/System.Reflection/MonoMethod.cs:309
```
From looking at the code, resolving the hostname to IP addresses via DNS fails so the test doesn't even run.
We don't care where the local IPs come from so we can use an alternative way of getting them - via querying the network interfaces.
Fixes https://bugzilla.xamarin.com/show_bug.cgi?id=41671
PS: fixed a hardcoded port while I was touching the code
[mini] lower typechecks later
lower typecheck instructions later in the compilation, so we've a chance to do high-level optimization after `method-to-ir`
Rewriite `MonoMethodMessage::InitMessage` in managed code
Change icall
`System.Runtime.Remoting.Messaging/MonoMethodMessage::InitMessage` to
managed code.
Also update `mono_message_init()` to invoke `InitMessage`.
Using mono_marshal_free n mono_string_builder_to_utf8
`mono_string_builder_to_utf16` is allocating using `mono_marshal_alloc` which, on windows, allocates memory using `CoTaskMemAlloc`. This buffer was then released in `mono_string_builder_to_utf8` using `g_free` which resulted in a crash since that's a different memory on windows. The fix in this PR makes sure `mono_marshal_free` is called instead which has the correct behaviour and results in a `CoTaskMemFree` on windows.
[System] Make ServicePointTest.DnsRefreshTimeout test more reliable
We are still seeing quite rare failures of this test:
```
HostEntry should result in the same IPHostEntry object.
Expected: same as <System.Net.IPHostEntry>
But was: <System.Net.IPHostEntry>
+++++++++++++++++++
STACK TRACE:
at MonoTests.System.Net.ServicePointTest.DnsRefreshTimeout () [0x00045] in /Users/builder/jenkins/workspace/test-mono-mainline/label/osx-i386/mcs/class/System/Test/System.Net/ServicePointTest.cs:228
at (wrapper managed-to-native) System.Reflection.MonoMethod:InternalInvoke (System.Reflection.MonoMethod,object,object[],System.Exception&)
at System.Reflection.MonoMethod.Invoke (System.Object obj, System.Reflection.BindingFlags invokeAttr, System.Reflection.Binder binder, System.Object[] parameters, System.Globalization.CultureInfo culture) [0x00038] in /Users/builder/jenkins/workspace/test-mono-mainline/label/osx-i386/mcs/class/corlib/System.Reflection/MonoMethod.cs:309
```
Since the test is only checking if expiring of the cached DNS lookup works, we don't really need to use an external domain name for this as "localhost" works just fine as well and typically doesn't trigger a DNS query which makes the test more reliable against network issues.
@monojenkins merge
We are still seeing quite rare failures of this test:
```
HostEntry should result in the same IPHostEntry object.
Expected: same as <System.Net.IPHostEntry>
But was: <System.Net.IPHostEntry>
+++++++++++++++++++
STACK TRACE:
at MonoTests.System.Net.ServicePointTest.DnsRefreshTimeout () [0x00045] in /Users/builder/jenkins/workspace/test-mono-mainline/label/osx-i386/mcs/class/System/Test/System.Net/ServicePointTest.cs:228
at (wrapper managed-to-native) System.Reflection.MonoMethod:InternalInvoke (System.Reflection.MonoMethod,object,object[],System.Exception&)
at System.Reflection.MonoMethod.Invoke (System.Object obj, System.Reflection.BindingFlags invokeAttr, System.Reflection.Binder binder, System.Object[] parameters, System.Globalization.CultureInfo culture) [0x00038] in /Users/builder/jenkins/workspace/test-mono-mainline/label/osx-i386/mcs/class/corlib/System.Reflection/MonoMethod.cs:309
```
Since the test is only checking if expiring of the cached DNS lookup works, we don't really need to use an external domain name for this
as "localhost" works just fine as well and typically doesn't trigger a DNS query which makes the test more reliable against network issues.
We were not exiting the critical section on this code path. This was leading to hangs at stw since the native code would take locks when compiling the code.
We were seeing intermittent failures on devices like:
```
[FAIL] TimerTest.EnabledInElapsed : #1 loss of events
Expected: True
But was: False
at MonoTests.System.Timers.TimerTest.EnabledInElapsed () <0x33b168 + 0x0014f> in <filename unknown>:0
at (wrapper managed-to-native) System.Reflection.MonoMethod:InternalInvoke (System.Reflection.MonoMethod,object,object[],System.Exception&)
at System.Reflection.MonoMethod.Invoke (System.Object , BindingFlags , System.Reflection.Binder , System.Object[] , System.Globalization.CultureInfo ) <0x66d400 + 0x000b7> in <filename unknown>:0
```
Since OS scheduling and GCs can influence the exact timing of the timer intervals sleeping for 200ms is likely just too short.
Instead of bumping the timeout, I decided to rewrite the test to not rely on Thread.Sleep ().
I also added another test which verifies setting AutoResetEvent=false stops the timer event from firing after the first one,
which was implicitly tested before by the other test.
Fixes https://bugzilla.xamarin.com/show_bug.cgi?id=41530
Without doing this, the check would consider a valid option name like
IpTimeToLive to be the same as ReuseAddress, as they share the same enum
value.
Fixes SocketTest.TtlChange and SocketTest.UdpMulticasTimeToLive on Android and
other platforms that don't support ReuseAddress.
* Move the versions.mk infrastructure to scripts/submodules so it can be used elsewhere too.
* Add a conditional submodule for llvm.
* Fix commit messages emitted by the bump- targets.
* Use a anonymous git url for llvm.
Before closing a socket we try to interrupt and wait for all socket
pending system calls to return. For that we set the socket as non
blocking which makes the system calls to return with EAGAIN\EWOULDBLOCK.
This issue was caught by System.Net.Sockets.SocketTest.SendAsyncFile
that was intermittently failing, we were only able to reproduce it while
running multiple test suites in parallel. We now can say that it only failed when
the the socket was closed and the sendfile was waiting for the OS to
allocate enough resources to send the file.
This was fixed by changing wapi_sendfile to return on EAGAIN.
Stack traces of blocked threads are now displayed on the exception.
The threads stacktraces are only stored when MONO_TESTS_IN_PROGRESS env
var is set to yes.
When we allocate the complex descriptor for a new type we might have the concurrent marker scan objects. While the concurrent marker scans the bitmap for a complex object, which resides in the complex_descriptor array, another thread might allocate a new complex descriptor, thus reallocate the complex_descriptor array, which would lead to corruption of the bitmap used by the concurrent marker.
We reuse the SgenArrayList which never reallocates, using instead a list of exponentially increasing arrays.
This is the initial code drop of a new handles implementation.
It differs from the previous one is a few ways:
- It keeps a separate stack of handles which is shared by all "arenas".
This eliminates malloc in the majority of the cases without the need for any tuning
- Custom gc scanning code for the handles stack.
The original implementation registered the chunks as a GC roots. Which is a terribly
slow operation as it takes the GC lock.
Instead of allocating a new handle arena for each icall, we compute a stack mark,
which is the position in the handles stack at icall entry. On exit, we simply restore that
position, which is just a couple of stores.
Since the stack is implemented using arraylets, addressed are stable in the face of
expansion.
Only the typed handles function/macros are in as having both typed and untyped handles
adds extra complexity for no clear value.
The macros only handle what was needed to port a sizeable chunk of locales.c.
This commit just sets the stage for the feature, there's quite some work left
before we can start using it, as it can be seen in the big TODO in handle.c
We were seeing an issue where attempts to create the MonoClass for an
exception in a no-exec context was causing failure. No-exec will forbid
the constructor call. Our solution was to realize that there was no
need to allocate an exception, as MonoError could already encode the
errors.
Fixes a bug in exiting modal runloop that caused not really exiting it
(not calling XplatUI.SetModal(..., false).
It caused issues with menu, for example, if you first ran a modal
dialog (a wizard), and then you started the application.
This would lead to a crash when unregistering the thread from the currnt MonoThreadInfo* destructor, as `mono_thread_info_current_unchecked` would return NULL. This would trigger an infinite loop in `mono_threads_enter_gc_safe_region_unbalanced_with_info`.
unregister_thread is the destructor for the current MonoThreadInfo* TLS key, meaning that when in this function, calling mono_thread_info_current_unchecked will always return NULL.
On Unix, errno value might change between the native call and the call to mono_marshal_set_last_error, if we exit the gc safe region. We need to ensure that it's done right after the native call.
[gc] Register a critical region when executing managed allocators.
Previously, if we stopped a thread and it had a stack looking like this:
...
profiler_signal_handler
<signal handler called>
managed_allocator
...
We would fail to identify the fact that we've just interrupted a thread that's
in a managed allocator (uninterruptible code) because we only look at the very
latest instruction pointer of the thread, which will be pointing into
profiler_signal_handler or some other function called by it. So we would
happily continue along with the STW process and proceed to doing the actual GC,
where we would see a broken heap.
We could solve this by unwinding the stack and checking all frames, but that's
complicated and error-prone. Instead, register a critical region while the
managed allocator runs. This way, if we don't identify it by instruction pointer, we
will identify it by the fact that the thread is in a critical region.
This fixes a regression introduced in 123aaa5db0.
We would sometimes see an ObjectDisposedException when calling AcceptAsync ()
because that patch removed the try-catch around it.
Xamarin.Android internal test suite takes advantage of the BCL tests
in Mono, so far their inclusion was a manual process and this commit
fixes it. A set of source inclusion and exclusion lists is added
and a handful of code fixes to make them compile on Xamarin.Android
[System] Fixes socket async exception.
SendAsyncFile uses StartSocketServer to start a server and at the same
time begins to accept a client socket. The issue is that when
the accept callback is called SendAsyncFile already closed the socket.
Fixed race condition by moving StartSocketServer and ReceiveCallback
logic into SendAsyncFile, and by using a new MRE to ensure that the
client receives all the data before the server is closed.
Fixes [#41264](https://bugzilla.xamarin.com/show_bug.cgi?id=41264)
SendAsyncFile uses StartSocketServer to start a server and at the same
time begins to accept a client socket. The issue is that when
the accept callback is called SendAsyncFile already closed the socket.
Fixed race condition by moving StartSocketServer and ReceiveCallback
logic into SendAsyncFile, and by using a new MRE to ensure that the
client receives all the data before the server is closed.
Fixes#41264
Standard C and C++ doesn't allow empty structs, empty structs will always have a
size of 1 byte. GCC have an extension to allow empty structs,
https://gcc.gnu.org/onlinedocs/gcc/Empty-Structures.html. This cause a
little dilemma since runtime build using none GCC compiler will not be
compatible with GCC build C libraries and the other way around. On platforms
where empty structs has size of 1 byte it must be represented in call and cannot
be dropped.
Libraries build using for example MSVC failed pinvoke call when empty
structs where used in pinvoke2.cs tests. The problem was that current
implementation dropped value types of size 0, eliminating them from the
call. Since MSVC implements C/C++ standard regarding empty structs (no extension used),
the function expected the parameter to be present, even if was not used.
This became a problem if the empty struct ended up in one of the 4 first
parameters (backed by registers) since if excluded, wrong parameter gets
into wrong registers (the same will apply to stack slots as well). For example,
a C signature like this, (int a, EmptyStruct es, int b) will get lowered to the
following on MSVC win x64, a -> rcx, es -> rdx and b -> r8, but since caller
eliminated es from the call, the register assignment from the caller was
a -> rcx and b -> rdx causing a violation of the calling convention and
error in callee.
The fix is to make sure that empty structs are still represented on none GCC
compilers on windows x64 (MSVC), since they must be represented in the call
to pinvoked methods. Fix will also use a const value of 0 to represent the value
passed as the struct instead of loading a value from stack into register/stack slot,
this is inline with the value the .NET runtime passes to pinvoked method. The reason
why we use a 0 instead of loading from stack is because a struct of size 0 is not explicitly
represented on stack and will have garbage data if loaded. Since it’s an empty struct value
type there is no need to load anything, just represent it as 0 in the register/stack slot passed
to caller.
There is also a case when returning an empty struct from a pinvoked method. Result will be returned
as 0 in RAX by MSVC compiled libraries (or whatever value stored in the 1 byte representation).
Since we have similar issues as above with empty structs of size 0 not explicitly represented on the
stack, the store will be eliminated and not written back to stack by caller since an empty struct doesn’t
represent any unique data.
Calling convention for managed->managed code is kept intact since it is working as expected
handling the case with empty structs as input and return values.
With fix all pinvoke2.cs tests now pass running on Windows x64 JIT.
This appears to serve no purpose as tlab_next is a thread-local variable. It's
only ever accessed by other threads in order to write to it, not read it. We
also don't do this memory barrier in the native allocation functions.
Previously, if we stopped a thread and it had a stack looking like this:
...
profiler_signal_handler
<signal handler called>
managed_allocator
...
We would fail to identify the fact that we've just interrupted a thread that's
in a managed allocator (uninterruptible code) because we only look at the very
latest instruction pointer of the thread, which will be pointing into
profiler_signal_handler or some other function called by it. So we would
happily continue along with the STW process and proceed to doing the actual GC,
where we would see a broken heap.
We could solve this by unwinding the stack and checking all frames, but that's
complicated and error-prone. Instead, register a critical region while the
managed allocator. This way, if we don't identify it by instruction pointer, we
will identify it by the fact that the thread is in a critical region.
Made breakpoint uniqueness test switch to a hash after 8th item added,
consolidated the code that handles the uniqueness test and makes the
breakpoint in one function, fixed style issues.
* Add TODO list
* Use -useSourcePath when building resources, so we do not need to add the explicit extra path
* More updates, now battling resource generation across the board
* Use newlines in the Unix commands to avoid generating things like cs-parser.cs^M
The tests added during the initial development of the -multiassembly argument failed to test the presence of extension methods;
the result was a crash that occurred when there were extension methods present in multiple assemblies.
This patch addresses the issue by putting a guard clause around the code that adds the extension method to the index.
Tests have been modified accordingly to verify this bug and include extension methods
Additionally, the -multiassembly parameter was removed entirely. This behavior is now triggered by the presence of -apistyle,
which is where support for "multiple assemblies" was meant to work anyways. To review, this is to support cross-platform
development, where you have multiple -- *different* -- assemblies for each platform. These assemblies will share a certain
cross-section of APIs that will be *exactly* the same in each assembly (as opposed to multiple versions of literally the same assembly).
As a result, every member will now contain an individual `AssemblyInfo` tag that also contains the name of the assembly.
The date/time struct in zip.h defines the month value as
uInt tm_mon; /* months since January - [0,11] */
.NET's DateTime object in turn stores months from 1-12. We didn't
take this into account when passing the values from managed->native
so the month ended up being off by one.
We didn't use this codepath before https://github.com/mono/mono/pull/3036
that's why it didn't show up until now.
The test assembly used *.dll.excludes before while the library uses *.dll.exclude.sources.
Make this consistent to reduce confusion.
Also remove corlib_cmp.dll.excludes which is from a profile that no longer exists.
We were seeing timeouts in the runtime step if two builds happen to be scheduled
at the exact same time on the ARM workers on Jenkins, causing increased load and
longer test times. They aborted the runtime step right before it'd finish.
Bump the timeout a bit to avoid this.
When we scan a block we need to have it checked for sweeping. Instead of waiting for all the blocks to be checked for sweeping, we traverse all the blocks (even though they might not have live objects), and do the sweeping check only for blocks that have dirty cards associated with them, without waiting for sweeping to end.
Allocating a major block needs addition to the allocated_blocks array, operation that needs sweep to be finished because sweep moves elements and changes the size when compacting the array. Do this operation in the sweep start instead, when we are not racing.
Support for GSHAREDVT added for Windows x64. Changes needed
(ones gsharedvt concept was understood, crashes debugged and adjustments located)
where quite small. Primarily around shadow stack used by windows + incorrect handling
of none volatile registers on windows when going over the gsharedvt trampolines.
All GSHAREDVT tests, gshared.exe and aot.tests.exe pass in full AOT builds with above adjustments.
Support for DYNCALL added/enabled for Windows x64. Current AMD64 support was working
for basic scenarios but not enabled on Windows. Identified a problem on AMD64
when structs are passed as input parameters or returned by methods for DYNCALL scenarios
in full AOT:ed builds. This scenario doesn't seem to work for any AMD64 platform for
full AOT and the tests in aot-test.exe handling these scenarios are not run on AMD64.
Adjusted categories on test in aot-tests.exe to reflect what’s really needed by the tests
to run correctly. Also added a new category to some DYNCALL tests that uses struct support
currently not implemented on AMD64, !FULLAOT-AMD64, so at least some of the DYNCALL tests
can be run on AMD64 platforms in full AOT:ed builds.
Testing with the debugger showed multiple failures around recursive
functions:
1. Step-out breakpoints can trigger on inner frames.
Solution: Extend existing inner-frame test from step-over to step out
also.
2. "Same source line, continuing" test triggers even on wrong frame
(affects both step over and step out)
Solution: Check frame as part of this test
3. When step breakpoints pick next sequence points, they can register
multiple breakpoints at the same place if a recursive call happens to
exit there. (Possibly harmless but may have resulted in duplicate
debugger events being generated in one test.)
Solution: Check for duplicates in step-request breakpoint list before
adding new breakpoints
4. Mono.Debugging.Soft in the debugger-libs repo sometimes resumes
after breakpoints if it decides Mono.Debugger.Soft has not stepped
far enough, but the check to trigger this was not recursion-aware.
Solution: This one is not addressed in this commit, but a fix was
checked in to debugger-libs. Any code using this commit therefore
needs to be run in conjunction with debugger-libs dfbf6d8b13.
Tests were also added for issues 1-3 above.
The NetBSD version of mremap takes additional argument "newp" to hint
new position of a mapped range.
void *
mremap(void *oldp, size_t oldsize, void *newp, size_t newsize,
int flags);
-- NetBSD's mremap(2)
While Linux by default sets old_address and fails if the resize
operation cannot be finished for the original address.
void *mremap(void *old_address, size_t old_size,
size_t new_size, int flags, ... /* void *new_address */);
-- Linux's mremap(2)
NetBSD offers inversed logic with the MAP_FIXED flag to Linux's MREMAP_MAYMOVE:
MAP_FIXED newp is tried and mremap() fails if that address can't
be used as new base address for the range. Otherwise,
oldp and newp are used as hints for the position,
factoring in the given alignment.
-- NetBSD's mremap(2)
MREMAP_MAYMOVE
By default, if there is not sufficient space to expand a mapping
at its current location, then mremap() fails. If this flag is
specified, then the kernel is permitted to relocate the
mapping to a new virtual address, if necessary. If the mapping
is relocated, then absolute pointers into the old mapping
location become invalid (offsets relative to the starting address
of the mapping should be employed).
-- Linux's mremap(2)
The thing that bothers is calling mremap(2) syntax from Linux "posix", but we
cannot do anything with it, for now just simulate the Linux behavior on NetBSD.
- Add logfile and syslog capability. For Windows, the latter is simulated by writing to a file. Adding Eventlog support is for a future effort. The choice of logging mode is made via MONO_LOG_DEST environment variable.
- Add timestamps and pid to the trace messages to make them more useful
- Add timestamps to SGEN_LOG
- Related to tracing: when tracing and using something like Valgrind the system may be slowed so much that the sleep abort limit may be exceeded for threads. This fix adds an environment variable lookup of "MONO_SLEEP_ABORT_LIMIT" which is a value greater than 40, which defines the number of milliseconds to wait for threads to complete pending operations. If not set, the valus is 200.
2016-03-08 14:54:11 -05:00
2140 changed files with 76020 additions and 41241 deletions
# Set the build profiles and options before things which use them
#
AC_ARG_WITH(profile4_x, [ --with-profile4=yes,no If you want to install the 4.x FX (defaults to yes)], [], [with_profile4_x=default])
AC_ARG_WITH(monodroid, [ --with-monodroid=yes,no If you want to build the MonoDroid assemblies (defaults to no)], [], [with_monodroid=default])
AC_ARG_WITH(monotouch, [ --with-monotouch=yes,no If you want to build the Xamarin.iOS assemblies (defaults to no)], [], [with_monotouch=default])
AC_ARG_WITH(monotouch_watch, [ --with-monotouch_watch=yes,no If you want to build the Xamarin.WatchOS assemblies (defaults to no)],[], [with_monotouch_watch=default])
AC_ARG_WITH(monotouch_tv, [ --with-monotouch_tv=yes,no If you want to build the Xamarin.TVOS assemblies (defaults to no)], [], [with_monotouch_tv=default])
AC_ARG_WITH(bitcode, [ --with-bitcode=yes,no If bitcode is enabled (defaults to no)], [], [with_bitcode=default])
AC_ARG_WITH(xammac, [ --with-xammac=yes,no If you want to build the Xamarin.Mac assemblies (defaults to no)], [], [with_xammac=default])
AC_ARG_WITH(mobile_static, [ --with-mobile_static=yes,no If you want to build the mobile_static assemblies (defaults to no)], [], [with_mobile_static=default])
AC_ARG_WITH(runtime_preset, [ --with-runtime_preset=net_4_x,all,mobile_static,bitcode_mobile_static Which default profile to build (defaults to net_4_x)], [], [with_runtime_preset=net_4_x])
AC_ARG_WITH(large-heap, [ --with-large-heap=yes,no Enable support for GC heaps larger than 3GB (defaults to no)], [large_heap=$withval], [large_heap=no])
AC_ARG_WITH(large-heap, [ --with-large-heap=yes,no Enable support for GC heaps larger than 3GB (defaults to no)], [large_heap=$withval], [large_heap=no])
if test "x$large_heap" = "xyes"; then
CPPFLAGS="$CPPFLAGS -DLARGE_CONFIG"
fi
AC_SUBST(LIBGC_CPPFLAGS)
AC_SUBST(LIBGC_LIBS)
AC_SUBST(LIBGC_STATIC_LIBS)
AC_SUBST(libgc_dir)
AC_SUBST(BOEHM_DEFINES)
dnl
dnl End of Boehm GC Configuration
dnl
@@ -2572,13 +2788,17 @@ dnl **************
dnl *** LLVM ***
dnl **************
AC_ARG_ENABLE(llvm,[ --enable-llvm Enable the LLVM back-end], enable_llvm=$enableval, enable_llvm=no)
AC_ARG_ENABLE(llvm,[ --enable-llvm Enable the LLVM back-end], enable_llvm=$enableval, enable_llvm=default)
AC_ARG_ENABLE(llvm-version-check,[ --enable-llvm-version-check Check that the LLVM matches the version expected by mono], enable_llvm_version_check=$enableval, enable_llvm_version_check=no)
AC_ARG_ENABLE(llvm-runtime,[ --enable-llvm-runtime Enable runtime support for llvmonly code], enable_llvm_runtime=$enableval, enable_llvm_runtime=no)
AC_ARG_WITH(llvm, [ --with-llvm=<llvm prefix> Enable the LLVM back-end], enable_llvm=yes,)
if test "x$enable_llvm" = "xdefault"; then
enable_llvm=$enable_llvm_default
fi
if test "x$enable_llvm" = "xyes"; then
if test "x$with_llvm" != "x"; then
LLVM_CONFIG=$with_llvm/bin/llvm-config
@@ -2636,13 +2856,12 @@ if test "x$enable_llvm" = "xyes"; then
if test x$with_sgen_default_concurrent != xno; then
AC_DEFINE(HAVE_CONC_GC_AS_DEFAULT, 1, [Defaults to concurrent GC])
conc_gc_msg=" (concurrent by default)"
fi
if test "x$gc_msg" = "x"; then
gc_msg="sgen$conc_gc_msg"
else
gc_msg="sgen$conc_gc_msg and $gc_msg"
fi
fi
AC_SUBST(SGEN_DEFINES)
AM_CONDITIONAL(SUPPORT_SGEN, test x$buildsgen = xyes)
@@ -3547,15 +3778,6 @@ case "x$libgc" in
;;
esac
AC_ARG_WITH(profile4_x, [ --with-profile4=yes,no If you want to install the 4.6 FX (defaults to yes)], [], [with_profile4_x=yes])
AC_ARG_WITH(monodroid, [ --with-monodroid=yes,no If you want to build the MonoDroid assemblies (defaults to no)], [], [with_monodroid=no])
AC_ARG_WITH(monotouch, [ --with-monotouch=yes,no If you want to build the Xamarin.iOS assemblies (defaults to no)], [], [with_monotouch=no])
AC_ARG_WITH(monotouch_watch, [ --with-monotouch_watch=yes,no If you want to build the Xamarin.WatchOS assemblies (defaults to no)],[], [with_monotouch_watch=no])
AC_ARG_WITH(monotouch_tv, [ --with-monotouch_tv=yes,no If you want to build the Xamarin.TVOS assemblies (defaults to no)], [], [with_monotouch_tv=no])
AC_ARG_WITH(bitcode, [ --with-bitcode=yes,no If bitcode is enabled (defaults to no)], [], [with_bitcode=no])
AC_ARG_WITH(xammac, [ --with-xammac=yes,no If you want to build the Xamarin.Mac assemblies (defaults to no)], [], [with_xammac=no])
AC_ARG_WITH(mobile_static, [ --with-mobile_static=yes,no If you want to build the mobile_static assemblies (defaults to no)], [], [with_mobile_static=no])
MALLOC_MEMPOOLS=no
AC_ARG_WITH(malloc_mempools,[ --with-malloc-mempools=yes,no Use malloc for each single mempool allocation (only for runtime debugging, defaults to no)],[
if test x$with_malloc_mempools = xyes; then
@@ -3565,15 +3787,18 @@ AC_ARG_WITH(malloc_mempools,[ --with-malloc-mempools=yes,no Use malloc for eac
])
DISABLE_MCS_DOCS=no
DISABLE_MCS_DOCS=default
AC_ARG_WITH(mcs_docs,[ --with-mcs-docs=yes,no If you want to build the documentation under mcs (defaults to yes)],[
if test x$with_mcs_docs != xyes; then
DISABLE_MCS_DOCS=yes
fi
])
if test x$with_profile4_x != xyes; then
if test -n "$INSTALL_4_x_TRUE"; then :
DISABLE_MCS_DOCS=yes
fi
if test "x$DISABLE_MCS_DOCS" = "xdefault"; then
DISABLE_MCS_DOCS=$DISABLE_MCS_DOCS_default
fi
AC_ARG_WITH(lazy_gc_thread_creation, [ --with-lazy-gc-thread-creation=yes|no Enable lazy runtime thread creation, embedding host must do it explicitly (defaults to no)],[
if test x$with_lazy_gc_thread_creation != xno ; then
Loaded 100 of 2140 files, more files were not shown because too many files have changed in this diff.
Show more
Reference in new issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.