Memory corruption sometimes.

[vext01]

With a release build of yklua you can shake various memory corruption crashes out when running the test suite.

I've only observed this with a release build and with YKD_SERIALISE_COMPILATION=0 together.

Example:

$ YKD_SERIALISE_COMPILATION=0 ../src/lua -e"_U=true" all.lua
...
free(): invalid size
[Thread 0x7fffe67fc6c0 (LWP 2471092) exited]

Thread 1 "lua" received signal SIGABRT, Aborted.
__pthread_kill_implementation (threadid=<optimized out>, signo=signo@entry=6, no_tid=no_tid@entry=0) at ./nptl/pthread_kill.c:44
44      ./nptl/pthread_kill.c: No such file or directory.
(gdb) bt
#0  __pthread_kill_implementation (threadid=<optimized out>, signo=signo@entry=6, no_tid=no_tid@entry=0) at ./nptl/pthread_kill.c:44
#1  0x00007ffff7c1ee8f in __pthread_kill_internal (signo=6, threadid=<optimized out>) at ./nptl/pthread_kill.c:78
#2  0x00007ffff7bcffb2 in __GI_raise (sig=sig@entry=6) at ../sysdeps/posix/raise.c:26
#3  0x00007ffff7bba472 in __GI_abort () at ./stdlib/abort.c:79
#4  0x00007ffff7c13430 in __libc_message (action=action@entry=do_abort, fmt=fmt@entry=0x7ffff7d2d459 "%s\n") at ../sysdeps/posix/libc_fatal.c:155
#5  0x00007ffff7c287aa in malloc_printerr (str=str@entry=0x7ffff7d2b0c9 "free(): invalid size") at ./malloc/malloc.c:5660
#6  0x00007ffff7c2a544 in _int_free (av=<optimized out>, p=0xe212c0, have_lock=have_lock@entry=0) at ./malloc/malloc.c:4439
#7  0x00007ffff7c2ce8f in __GI___libc_free (mem=<optimized out>) at ./malloc/malloc.c:3385
#8  0x00007ffff7efb106 in __replace_stack () from /home/vext01/research/yk/bin/../target/release/deps/libykcapi.so
#9  0x00007ffff7a7dcf8 in ?? ()
#10 0x00007ffff7a7dd50 in ?? ()
#11 0x00007ffff7a7dd70 in ?? ()
#12 0x00007ffff7a7dc6c in ?? ()
#13 0x0000000000000000 in ?? ()

Discussed briefly with @ptersilie and this particular trace looks like free(3) crashing when releasing scratch memory used during stack reconstruction.

[ltratt]

@vext01 When #1318 merges, can you retry this please? There is a fix (in ykllvm) which might fix this problem.

[vext01]

Using yk master as of about 10 minutes ago (65e45bea3762a), I was able to get similar crashes in release mode.

The trace looks pretty much the same:

#0  __pthread_kill_implementation (threadid=<optimized out>, signo=signo@entry=6, no_tid=no_tid@entry=0) at ./nptl/pthread_kill.c:44
#1  0x00007ffff7c1de8f in __pthread_kill_internal (signo=6, threadid=<optimized out>) at ./nptl/pthread_kill.c:78
#2  0x00007ffff7bcefb2 in __GI_raise (sig=sig@entry=6) at ../sysdeps/posix/raise.c:26
#3  0x00007ffff7bb9472 in __GI_abort () at ./stdlib/abort.c:79
#4  0x00007ffff7c12430 in __libc_message (action=action@entry=do_abort, fmt=fmt@entry=0x7ffff7d2c459 "%s\n") at ../sysdeps/posix/libc_fatal.c:155
#5  0x00007ffff7c277aa in malloc_printerr (str=str@entry=0x7ffff7d2ebb8 "munmap_chunk(): invalid pointer") at ./malloc/malloc.c:5660
#6  0x00007ffff7c2796c in munmap_chunk (p=<optimized out>) at ./malloc/malloc.c:3054
#7  0x00007ffff7c2bed8 in __GI___libc_free (mem=<optimized out>) at ./malloc/malloc.c:3375
#8  0x00007ffff7f19d76 in __replace_stack () from /home/vext01/research/yk/bin/../target/release/deps/libykcapi.so
#9  0x00007ffff7a7ccf8 in ?? ()
#10 0x00007ffff7a7cd50 in ?? ()
#11 0x00007ffff7a7cd70 in ?? ()
#12 0x00007ffff7a7cc6c in ?? ()
#13 0x00007ffff7a7c4b0 in ?? ()
#14 0x0000000000000000 in ?? ()

[ltratt]

I can't replicate this. I suggest trying (in order):

1. make clean and make your yklua. If it was built with an old ykllvm, all bets are off.
2. A totally fresh yk + ykllvm clone.

My guess is that (1) is the most likely culprit, but I'm not really sure.

[vext01]

I always make clean, but today I'll try to make a reproducible script showing the problem.

[vext01]

I was unable to reproduce this using the following script:

#!/bin/sh

set -xeu

unset YKB_YKLLVM_BIN_DIR

if [ ! -e yk ]; then
	git clone --recurse-submodules https://github.com/ykjit/yk
fi

if [ ! -e yklua ]; then
	git clone https://github.com/ykjit/yklua
fi

if [ ! -e try_repeat ]; then
	git clone https://github.com/ltratt/try_repeat
fi

cd yk
cargo test --release
export PATH=$(pwd)/bin:${PATH}
cd ..

cd yklua
export YK_BUILD_TYPE=release

make clean
make -j $(nproc)
cd tests
YKD_SERIALISE_COMPILATION=0 ../../try_repeat/try_repeat 100 \
	../src/lua -e"_U=true" all.lua
YKD_SERIALISE_COMPILATION=1 ../../try_repeat/try_repeat 100 \
	../src/lua -e"_U=true" all.lua
cd ..

# Repeat using YKB_YKLLVM_BIN_DIR in case it has to do with that.
make clean
export YKB_YKLLVM_BIN_DIR=$(pwd)/../yk/target/release/ykllvm/bin
make -j $(nproc)
cd tests
YKD_SERIALISE_COMPILATION=0 ../../try_repeat/try_repeat 100 \
	../src/lua -e"_U=true" all.lua
YKD_SERIALISE_COMPILATION=1 ../../try_repeat/try_repeat 100 \
	../src/lua -e"_U=true" all.lua

So I can only assume it's something up the local setup I was using, or a recent commit fixed it.

Will close.

[vext01]

We just saw some pointer corruption in a CI job. This time it was while running the awfy benchmarks. Reopening.

ykjit/yk-benchmarks#3 (comment)

[ltratt]

I think try_repeat and gdb might be in order?