Fix possible deadlock for jemalloc with enabled profiler

ClickHouse uses JEMALLOC_PROF_LIBGCC - _Unwind_Backtrace() for jemalloc,
which calls _Unwind_Backtrace() during bootstrap of jemalloc, and if,
dlsym() has allocations it will lead to deadlock:

    0  __lll_lock_wait (futex=futex@entry=0x16516410 <init_lock+64>, private=0) at lowlevellock.c:52
    1  0x00007ffff7f9d0a3 in __GI___pthread_mutex_lock (mutex=0x16516410 <init_lock+64>) at ../nptl/pthread_mutex_lock.c:80
    2  0x000000000ba6ec23 in pthread_mutex_lock (arg=arg@entry=0x16516410 <init_lock+64>) at ./build/RelWithDebInfo/./src/Common/ThreadFuzzer.cpp:445
    3  0x0000000014215fd4 in malloc_mutex_lock_final (mutex=0x165163d0 <init_lock>) at ./contrib/jemalloc/include/jemalloc/internal/mutex.h:151
    4  malloc_mutex_lock_slow (mutex=0x165163d0 <init_lock>) at ./build/RelWithDebInfo/./contrib/jemalloc/src/mutex.c:90
    5  0x00000000141defc6 in malloc_mutex_lock (tsdn=0x0, mutex=<optimized out>) at ./contrib/jemalloc/include/jemalloc/internal/mutex.h:217
    6  malloc_init_hard () at ./build/RelWithDebInfo/./contrib/jemalloc/src/jemalloc.c:2118
    7  0x00000000141d05db in malloc_init () at ./build/RelWithDebInfo/./contrib/jemalloc/src/jemalloc.c:298
    8  imalloc_init_check (sopts=<optimized out>, dopts=<optimized out>) at ./build/RelWithDebInfo/./contrib/jemalloc/src/jemalloc.c:2658
    9  imalloc (sopts=<optimized out>, dopts=<optimized out>) at ./build/RelWithDebInfo/./contrib/jemalloc/src/jemalloc.c:2689
    10 calloc (num=num@entry=1, size=size@entry=32) at ./build/RelWithDebInfo/./contrib/jemalloc/src/jemalloc.c:2852
    11 0x00007ffff7c42c05 in _dlerror_run (operate=operate@entry=0x7ffff7c42490 <dlsym_doit>, args=args@entry=0x7fffffffda40) at dlerror.c:148
    12 0x00007ffff7c42525 in __dlsym (handle=<optimized out>, name=0xe7f568 "dl_iterate_phdr") at dlsym.c:70
    13 0x000000001408e1a0 in (anonymous namespace)::getOriginalDLIteratePHDR () at ./build/RelWithDebInfo/./base/base/phdr_cache.cpp:44
    14 dl_iterate_phdr (callback=0x16287a60 <libunwind::findUnwindSectionsByPhdr(dl_phdr_info*, unsigned long, void*)>, data=0x7fffffffdb58) at ./build/RelWithDebInfo/./base/base/phdr_cache.cpp:64
    15 0x0000000016288468 in libunwind::LocalAddressSpace::findUnwindSections (this=<optimized out>, targetAddr=371765377, info=...) at ./contrib/libunwind/src/AddressSpace.hpp:605
    16 libunwind::UnwindCursor<libunwind::LocalAddressSpace, libunwind::Registers_x86_64>::setInfoBasedOnIPRegister (this=this@entry=0x7fffffffde70, isReturnAddress=false) at ./contrib/libunwind/src/UnwindCursor.hpp:2553
    17 0x0000000016287617 in __unw_init_local (cursor=cursor@entry=0x7fffffffde70, context=context@entry=0x7fffffffddc8) at ./build/RelWithDebInfo/./contrib/libunwind/src/libunwind.cpp:91
    18 0x000000001628b094 in _Unwind_Backtrace (callback=0x14226f40 <prof_unwind_init_callback>, ref=0x0) at ./build/RelWithDebInfo/./contrib/libunwind/src/UnwindLevel1-gcc-ext.c:106
    19 0x000000001421a7e0 in prof_boot2 (tsd=tsd@entry=0x7ffff7c357c0, base=<optimized out>) at ./build/RelWithDebInfo/./contrib/jemalloc/src/prof.c:693
    20 0x00000000141df3e7 in malloc_init_hard () at ./build/RelWithDebInfo/./contrib/jemalloc/src/jemalloc.c:2151
    21 0x00000000141d96f1 in malloc_init () at ./build/RelWithDebInfo/./contrib/jemalloc/src/jemalloc.c:298
    22 nallocx (size=32, flags=0) at ./build/RelWithDebInfo/./contrib/jemalloc/src/jemalloc.c:4011
    23 0x000000000b9fdd9b in _ZN6Memory23getActualAllocationSizeITpTkNSt3__17same_asISt11align_val_tEEJEQsr2DBE16OptionalArgumentIDpT_EEEmmS5_ (size=32) at ./src/Common/memory.h:197
    24 _ZN6Memory11trackMemoryITpTkNSt3__17same_asISt11align_val_tEEJEQsr2DBE16OptionalArgumentIDpT_EEEmmR15AllocationTraceS5_ (size=32, trace=...) at ./src/Common/memory.h:208
    25 operator new (size=size@entry=32) at ./build/RelWithDebInfo/./src/Common/new_delete.cpp:53
    26 0x000000001540e0a6 in google::protobuf::internal::ShutdownData::get () at ./build/RelWithDebInfo/./contrib/google-protobuf/src/google/protobuf/message_lite.cc:679
    27 google::protobuf::internal::OnShutdownRun (f=0x15407420 <google::protobuf::internal::DestroyString(void const*)>, arg=0x167a6368 <google::protobuf::internal::fixed_address_empty_string>) at ./build/RelWithDebInfo/./contrib/google-protobuf/src/google/protobuf/message_lite.cc:697
    28 0x0000000015407560 in global constructors keyed to 000101 () at ./contrib/google-protobuf/src/google/protobuf/generated_message_util.h:180
    29 0x000000001629101d in __libc_csu_init ()
    30 0x00007ffff7dba010 in __libc_start_main (main=0x66e3b00 <main(int, char**)>, argc=4, argv=0x7fffffffe258, init=0x16290fd0 <__libc_csu_init>, fini=<optimized out>, rtld_fini=<optimized out>, stack_end=0x7fffffffe248) at ../csu/libc-start.c:264
    31 0x00000000066e302e in _start ()

And this is indeed what happens for glibc prio 2.34, since only in [2]
the allocation had been removed

  [2]: https://sourceware.org/git/?p=glibc.git;a=commit;h=fada9018199c21c469ff0e731ef75c6020074ac9

But, jemalloc provides also JEMALLOC_PROF_LIBUNWIND, yes it is not llvm
libunwind, but gnu one, but ClickHouse version of llvm libunwind
provides unw_backtrace(), so we can use it, and it will not be called
during jemalloc bootstrap, so deadlock should not be possible.

Funny thing that for this deadlock the workaround is to enable
background_thread, because a thread created for it will initialize TLS for
__dlerror().

Signed-off-by: Azat Khuzhin <a.khuzhin@semrush.com>
This commit is contained in:
Azat Khuzhin 2024-07-10 15:58:42 +02:00
parent d2ad1cef33
commit e6e8b065bc

View File

@ -179,12 +179,19 @@ endif ()
target_compile_definitions(_jemalloc PRIVATE -DJEMALLOC_PROF=1)
# jemalloc provides support for two different libunwind flavors: the original HP libunwind and the one coming with gcc / g++ / libstdc++.
# The latter is identified by `JEMALLOC_PROF_LIBGCC` and uses `_Unwind_Backtrace` method instead of `unw_backtrace`.
# At the time ClickHouse uses LLVM libunwind which follows libgcc's way of backtracking.
# jemalloc provides support two unwind flavors:
# - JEMALLOC_PROF_LIBUNWIND - unw_backtrace() - gnu libunwind (compatible with llvm libunwind)
# - JEMALLOC_PROF_LIBGCC - _Unwind_Backtrace() - the original HP libunwind and the one coming with gcc / g++ / libstdc++.
#
# ClickHouse has to provide `unw_backtrace` method by the means of [commit 8e2b31e](https://github.com/ClickHouse/libunwind/commit/8e2b31e766dd502f6df74909e04a7dbdf5182eb1).
target_compile_definitions (_jemalloc PRIVATE -DJEMALLOC_PROF_LIBGCC=1)
# But for JEMALLOC_PROF_LIBGCC it also calls _Unwind_Backtrace() during
# bootstraping of jemalloc, which may lead to deadlock, if the dlsym will do
# allocations somewhere (like glibc does prio 2.34, see [1]).
#
# [1]: https://sourceware.org/git/?p=glibc.git;a=commit;h=fada9018199c21c469ff0e731ef75c6020074ac9
#
# And since ClickHouse unwind already supports unw_backtrace() we can safely
# switch to it to avoid this deadlock.
target_compile_definitions (_jemalloc PRIVATE -DJEMALLOC_PROF_LIBUNWIND=1)
target_link_libraries (_jemalloc PRIVATE unwind)
# for RTLD_NEXT