Fixes:
1. Setting the number of entries in a thread's clock to max between
the thread and the SyncClock the thread is acquiring from
2. Setting last_acquire_
Unit- and stress-test for releaseStoreAcquire added to
tests/unit/tsan_clock_test.cpp
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
Summary:
Following up on and complementing D44404 and other sanitizer allocators.
Currently many allocator specific errors (OOM, for example) are reported as
a text message and CHECK(0) termination, no stack, no details, not too
helpful nor informative. To improve the situation, detailed and structured
common errors were defined and reported under the appropriate conditions.
Common tests were generalized a bit to cover a slightly different TSan
stack reporting format, extended to verify errno value and returned
pointer value check is now explicit to facilitate debugging.
Reviewers: dvyukov
Subscribers: srhines, kubamracek, delcypher, #sanitizers, llvm-commits
Differential Revision: https://reviews.llvm.org/D48087
llvm-svn: 334975
Summary:
The low-fat STL-like vector container will be reused in MSan.
It is needed to implement an atexit(3) interceptor on NetBSD/amd64 in MSan.
Sponsored by <The NetBSD Foundation>
Reviewers: joerg, dvyukov, eugenis, vitalybuka, kcc
Reviewed By: dvyukov
Subscribers: kubamracek, mgorny, llvm-commits, #sanitizers
Tags: #sanitizers
Differential Revision: https://reviews.llvm.org/D40726
llvm-svn: 319650
Summary:
`CheckForPvallocOverflow` was introduced with D35818 to detect when pvalloc
would wrap when rounding up to the next multiple of the page size.
Add this check to TSan's pvalloc implementation.
Reviewers: alekseyshl
Reviewed By: alekseyshl
Subscribers: llvm-commits, kubamracek
Differential Revision: https://reviews.llvm.org/D36245
llvm-svn: 309897
Summary:
Set proper errno code on allocation failures and change realloc, pvalloc,
aligned_alloc, memalign and posix_memalign implementation to satisfy
their man-specified requirements.
Modify allocator API implementation to bring it closer to other
sanitizers allocators.
Reviewers: dvyukov
Subscribers: llvm-commits, kubamracek
Differential Revision: https://reviews.llvm.org/D35690
llvm-svn: 308929
This change implements 2 optimizations of sync clocks that reduce memory consumption:
Use previously unused first level block space to store clock elements.
Currently a clock for 100 threads consumes 3 512-byte blocks:
2 64-bit second level blocks to store clock elements
+1 32-bit first level block to store indices to second level blocks
Only 8 bytes of the first level block are actually used.
With this change such clock consumes only 2 blocks.
Share similar clocks differing only by a single clock entry for the current thread.
When a thread does several release operations on fresh sync objects without intervening
acquire operations in between (e.g. initialization of several fields in ctor),
the resulting clocks differ only by a single entry for the current thread.
This change reuses a single clock for such release operations. The current thread time
(which is different for different clocks) is stored in dirty entries.
We are experiencing issues with a large program that eats all 64M clock blocks
(32GB of non-flushable memory) and crashes with dense allocator overflow.
Max number of threads in the program is ~170 which is currently quite unfortunate
(consume 4 blocks per clock). Currently it crashes after consuming 60+ GB of memory.
The first optimization brings clock block consumption down to ~40M and
allows the program to work. The second optimization further reduces block consumption
to "modest" 16M blocks (~8GB of RAM) and reduces overall RAM consumption to ~30GB.
Measurements on another real world C++ RPC benchmark show RSS reduction
from 3.491G to 3.186G and a modest speedup of ~5%.
Go parallel client/server HTTP benchmark:
https://github.com/golang/benchmarks/blob/master/http/http.go
shows RSS reduction from 320MB to 240MB and a few percent speedup.
Reviewed in https://reviews.llvm.org/D35323
llvm-svn: 308018
Current interface assumes that Go calls ProcWire/ProcUnwire
to establish the association between thread and proc.
With the wisdom of hindsight, this interface does not work
very well. I had to sprinkle Go scheduler with wire/unwire
calls, and any mistake leads to hard to debug crashes.
This is not something one wants to maintian.
Fortunately, there is a simpler solution. We can ask Go
runtime as to what is the current Processor, and that
question is very easy to answer on Go side.
Switch to such interface.
llvm-svn: 267703
This is reincarnation of http://reviews.llvm.org/D17648 with the bug fix pointed out by Adhemerval (zatrazz).
Currently ThreadState holds both logical state (required for race-detection algorithm, user-visible)
and physical state (various caches, most notably malloc cache). Move physical state in a new
Process entity. Besides just being the right thing from abstraction point of view, this solves several
problems:
Cache everything on P level in Go. Currently we cache on a mix of goroutine and OS thread levels.
This unnecessary increases memory consumption.
Properly handle free operations in Go. Frees are issue by GC which don't have goroutine context.
As the result we could not do anything more than just clearing shadow. For example, we leaked
sync objects and heap block descriptors.
This will allow to get rid of libc malloc in Go (now we have Processor context for internal allocator cache).
This in turn will allow to get rid of dependency on libc entirely.
Potentially we can make Processor per-CPU in C++ mode instead of per-thread, which will
reduce resource consumption.
The distinction between Thread and Processor is currently used only by Go, C++ creates Processor per OS thread,
which is equivalent to the current scheme.
llvm-svn: 267678
On OS X 10.11+, we have "automatic interceptors", so we don't need to use DYLD_INSERT_LIBRARIES when launching instrumented programs. However, non-instrumented programs that load TSan late (e.g. via dlopen) are currently broken, as TSan will still try to initialize, but the program will crash/hang at random places (because the interceptors don't work). This patch adds an explicit check that interceptors are working, and if not, it aborts and prints out an error message suggesting to explicitly use DYLD_INSERT_LIBRARIES.
TSan unit tests run with a statically linked runtime, where interceptors don't work. To avoid aborting the process in this case, the patch replaces `DisableReexec()` with a weak `ReexecDisabled()` function which is defined to return true in unit tests.
Differential Revision: http://reviews.llvm.org/D18212
llvm-svn: 263695
Currently ThreadState holds both logical state (required for race-detection algorithm, user-visible)
and physical state (various caches, most notably malloc cache). Move physical state in a new
Process entity. Besides just being the right thing from abstraction point of view, this solves several
problems:
1. Cache everything on P level in Go. Currently we cache on a mix of goroutine and OS thread levels.
This unnecessary increases memory consumption.
2. Properly handle free operations in Go. Frees are issue by GC which don't have goroutine context.
As the result we could not do anything more than just clearing shadow. For example, we leaked
sync objects and heap block descriptors.
3. This will allow to get rid of libc malloc in Go (now we have Processor context for internal allocator cache).
This in turn will allow to get rid of dependency on libc entirely.
4. Potentially we can make Processor per-CPU in C++ mode instead of per-thread, which will
reduce resource consumption.
The distinction between Thread and Processor is currently used only by Go, C++ creates Processor per OS thread,
which is equivalent to the current scheme.
llvm-svn: 262037
On OS X, interceptors don't work in unit tests, so calloc() calls the system allocator. We need to use user_calloc() instead.
Differential Revision: http://reviews.llvm.org/D14918
llvm-svn: 253979
Summary:
Merge "exitcode" flag from ASan, LSan, TSan and "exit_code" from MSan
into one entity. Additionally, make sure sanitizer_common now uses the
value of common_flags()->exitcode when dying on error, so that this
flag will automatically work for other sanitizers (UBSan and DFSan) as
well.
User-visible changes:
* "exit_code" MSan runtime flag is now deprecated. If explicitly
specified, this flag will take precedence over "exitcode".
The users are encouraged to migrate to the new version.
* __asan_set_error_exit_code() and __msan_set_exit_code() functions
are removed. With few exceptions, we don't support changing runtime
flags during program execution - we can't make them thread-safe.
The users should use __sanitizer_set_death_callback()
that would call _exit() with proper exit code instead.
* Plugin tools (LSan and UBSan) now inherit the exit code of the parent
tool. In particular, this means that ASan would now crash the program
with exit code "1" instead of "23" if it detects leaks.
Reviewers: kcc, eugenis
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D12120
llvm-svn: 245734
Summary:
This change removes `__tsan::StackTrace` class. There are
now three alternatives:
# Lightweight `__sanitizer::StackTrace`, which doesn't own a buffer
of PCs. It is used in functions that need stack traces in read-only
mode, and helps to prevent unnecessary allocations/copies (e.g.
for StackTraces fetched from StackDepot).
# `__sanitizer::BufferedStackTrace`, which stores buffer of PCs in
a constant array. It is used in TraceHeader (non-Go version)
# `__tsan::VarSizeStackTrace`, which owns buffer of PCs, dynamically
allocated via TSan internal allocator.
Test Plan: compiler-rt test suite
Reviewers: dvyukov, kcc
Reviewed By: kcc
Subscribers: llvm-commits, kcc
Differential Revision: http://reviews.llvm.org/D6004
llvm-svn: 221194
Vector clocks is the most actively allocated object in tsan runtime.
Current internal allocator is not scalable enough to handle allocation
of clocks in scalable way (too small caches). This changes transforms
clocks to 2-level array with 512-byte blocks. Since all blocks are of
the same size, it's possible to cache them more efficiently in per-thread caches.
llvm-svn: 214912
The bug happens in the following case:
Mutex is located at heap block beginning,
when we call MutexDestroy, s->next is set to 0,
so free can't find the MBlock related to the block.
llvm-svn: 212531
Introduce new public header <sanitizer/allocator_interface.h> and a set
of functions __sanitizer_get_ownership(), __sanitizer_malloc_hook() etc.
that will eventually replace their tool-specific equivalents
(__asan_get_ownership(), __msan_get_ownership() etc.). Tool-specific
functions are now deprecated and implemented as stubs redirecting
to __sanitizer_ versions (which are implemented differently in each tool).
Replace all uses of __xsan_ versions with __sanitizer_ versions in unit
and lit tests.
llvm-svn: 212469
The former used to crash with a null deref if it was given a not owned pointer,
while the latter returned 0. Now they both return 0. This is still not the best possible
behavior: it is better to print an error report with a stack trace, pointing
to the error in user code, as we do in ASan.
llvm-svn: 212112
The optimization is two-fold:
First, the algorithm now uses SSE instructions to
handle all 4 shadow slots at once. This makes processing
faster.
Second, if shadow contains the same access, we do not
store the event into trace. This increases effective
trace size, that is, tsan can remember up to 10x more
previous memory accesses.
Perofrmance impact:
Before:
[ OK ] DISABLED_BENCH.Mop8Read (2461 ms)
[ OK ] DISABLED_BENCH.Mop8Write (1836 ms)
After:
[ OK ] DISABLED_BENCH.Mop8Read (1204 ms)
[ OK ] DISABLED_BENCH.Mop8Write (976 ms)
But this measures only fast-path.
On large real applications the speedup is ~20%.
Trace size impact:
On app1:
Memory accesses : 1163265870
Including same : 791312905 (68%)
on app2:
Memory accesses : 166875345
Including same : 150449689 (90%)
90% of filtered events means that trace size is effectively 10x larger.
llvm-svn: 209897
The new storage (MetaMap) is based on direct shadow (instead of a hashmap + per-block lists).
This solves a number of problems:
- eliminates quadratic behaviour in SyncTab::GetAndLock (https://code.google.com/p/thread-sanitizer/issues/detail?id=26)
- eliminates contention in SyncTab
- eliminates contention in internal allocator during allocation of sync objects
- removes a bunch of ad-hoc code in java interface
- reduces java shadow from 2x to 1/2x
- allows to memorize heap block meta info for Java and Go
- allows to cleanup sync object meta info for Go
- which in turn enabled deadlock detector for Go
llvm-svn: 209810
Make vector clock operations O(1) for several important classes of use cases.
See comments for details.
Below are stats from a large server app, 77% of all clock operations are handled as O(1).
Clock acquire : 25983645
empty clock : 6288080
fast from release-store : 14917504
contains my tid : 4515743
repeated (fast) : 2141428
full (slow) : 2636633
acquired something : 1426863
Clock release : 2544216
resize : 6241
fast1 : 197693
fast2 : 1016293
fast3 : 2007
full (slow) : 1797488
was acquired : 709227
clear tail : 1
last overflow : 0
Clock release store : 3446946
resize : 200516
fast : 469265
slow : 2977681
clear tail : 0
Clock acquire-release : 820028
llvm-svn: 204656
Make behavior introduced in r202820 conditional (under legacy_pthread_cond flag).
The new issue that we've hit with the satellite pthread_cond_t struct is
that pthread_condattr_getpshared does not work (satellite data is not shared between processes).
The idea is that most processes do not use pthread 2.2.5.
The rare ones that use (2.2.5 is dated by 2002) must specify legacy_pthread_cond=1
on their own risk.
llvm-svn: 204032
Currently correct programs can deadlock after fork, because atomic operations and async-signal-safe calls are not async-signal-safe under tsan.
With this change:
- if a single-threaded program forks, the child continues running with verification enabled (the tsan background thread is recreated as well)
- if a multi-threaded program forks, then the child runs with verification disabled (memory accesses, atomic operations and interceptors are disabled); it's expected that it will exec soon anyway
- if the child tries to create more threads after multi-threaded fork, the program aborts with error message
- die_after_fork flag is added that allows to continue running, but all bets are off
http://llvm-reviews.chandlerc.com/D2614
llvm-svn: 199993
This is intended to address the following problem.
Episodically we see CHECK-failures when recursive interceptors call back into user code. Effectively we are not "in_rtl" at this point, but it's very complicated and fragile to properly maintain in_rtl property. Instead get rid of it. It was used mostly for sanity CHECKs, which basically never uncover real problems.
Instead introduce ignore_interceptors flag, which is used in very few narrow places to disable recursive interceptors (e.g. during runtime initialization).
llvm-svn: 197979
This allows to increase max shadow stack size to 64K,
and reliably catch shadow stack overflows instead of silently
corrupting memory.
llvm-svn: 192797
Daniel S Fava