Currently we have a bit of a mess related to tids:
- sanitizers re-declare kInvalidTid multiple times
- some call it kUnknownTid
- implicit assumptions that main tid is 0
- asan/memprof claim their tids need to fit into 24 bits,
but this does not seem to be true anymore
- inconsistent use of u32/int to store tids
Introduce kInvalidTid/kMainTid in sanitizer_common
and use them consistently.
Reviewed By: vitalybuka
Differential Revision: https://reviews.llvm.org/D101428
We've got a user report about heap block allocator overflow.
Bump the L1 capacity of all dense slab allocators to maximum
and be careful to not page the whole L1 array in from .bss.
If OS uses huge pages, this still may cause a limited RSS increase
due to boundary huge pages, but avoiding that looks hard.
Reviewed By: vitalybuka
Differential Revision: https://reviews.llvm.org/D101161
Add ThreadClock:: global_acquire_ which is the last time another thread
has done a global acquire of this thread's clock.
It helps to avoid problem described in:
https://github.com/golang/go/issues/39186
See test/tsan/java_finalizer2.cpp for a regression test.
Note the failuire is _extremely_ hard to hit, so if you are trying
to reproduce it, you may want to run something like:
$ go get golang.org/x/tools/cmd/stress
$ stress -p=64 ./a.out
The crux of the problem is roughly as follows.
A number of O(1) optimizations in the clocks algorithm assume proper
transitive cumulative propagation of clock values. The AcquireGlobal
operation may produce an inconsistent non-linearazable view of
thread clocks. Namely, it may acquire a later value from a thread
with a higher ID, but fail to acquire an earlier value from a thread
with a lower ID. If a thread that executed AcquireGlobal then releases
to a sync clock, it will spoil the sync clock with the inconsistent
values. If another thread later releases to the sync clock, the optimized
algorithm may break.
The exact sequence of events that leads to the failure.
- thread 1 executes AcquireGlobal
- thread 1 acquires value 1 for thread 2
- thread 2 increments clock to 2
- thread 2 releases to sync object 1
- thread 3 at time 1
- thread 3 acquires from sync object 1
- thread 1 acquires value 1 for thread 3
- thread 1 releases to sync object 2
- sync object 2 clock has 1 for thread 2 and 1 for thread 3
- thread 3 releases to sync object 2
- thread 3 sees value 1 in the clock for itself
and decides that it has already released to the clock
and did not acquire anything from other threads after that
(the last_acquire_ check in release operation)
- thread 3 does not update the value for thread 2 in the clock from 1 to 2
- thread 4 acquires from sync object 2
- thread 4 detects a false race with thread 2
as it should have been synchronized with thread 2 up to time 2,
but because of the broken clock it is now synchronized only up to time 1
The global_acquire_ value helps to prevent this scenario.
Namely, thread 3 will not trust any own clock values up to global_acquire_
for the purposes of the last_acquire_ optimization.
Reviewed-in: https://reviews.llvm.org/D80474
Reported-by: nvanbenschoten (Nathan VanBenschoten)
realeaseAcquire() is a new function added to TSan in support of the Go data-race detector.
It's semantics is:
void ThreadClock::releaseAcquire(SyncClock *sc) const {
for (int i = 0; i < kMaxThreads; i++) {
tmp = clock[i];
clock[i] = max(clock[i], sc->clock[i]);
sc->clock[i] = tmp;
}
}
For context see: https://go-review.googlesource.com/c/go/+/220419
Reviewed-in: https://reviews.llvm.org/D76322
Author: dfava (Daniel Fava)
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
This 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
1. Add SyncClock::ResetImpl which removes code
duplication between ctor and Reset.
2. Move SyncClock::Resize to SyncClock methods,
currently it's defined between ThreadClock methods.
llvm-svn: 307785
This reverts commit r250823.
Replacing at least some of empty
constructors with "= default" variants is a semantical change which we
don't want. E.g. __tsan::ClockBlock contains a union of large arrays,
and it's critical for correctness and performance that we don't memset()
these arrays in the constructor.
llvm-svn: 251717
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 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
Algorithm description: http://code.google.com/p/thread-sanitizer/wiki/ThreadSanitizerAlgorithm
Status:
The tool is known to work on large real-life applications, but still has quite a few rough edges.
Nothing is guaranteed yet.
The tool works on x86_64 Linux.
Support for 64-bit MacOS 10.7+ is planned for late 2012.
Support for 32-bit OSes is doable, but problematic and not yet planed.
Further commits coming:
- tests
- makefiles
- documentation
- clang driver patch
The code was previously developed at http://code.google.com/p/data-race-test/source/browse/trunk/v2/
by Dmitry Vyukov and Kostya Serebryany with contributions from
Timur Iskhodzhanov, Alexander Potapenko, Alexey Samsonov and Evgeniy Stepanov.
llvm-svn: 156542
Kazuaki Ishizaki