flock/open random error

[Takashi Yano]

On Tue, 21 Oct 2025 17:41:51 -0700
Nahor wrote:
> Hi,
> 
> There is a test in the Fish shell (tests::history::test_history_races)
> that systematically fails when I run it. The test simulates multiple
> processes/threads trying to write to the shell history file at the
> same time.
> In my case, the test freezes/deadlocks with errors like "Bad Addr" and
> "Is Directory".
> When I add a sleep, the freeze/deadlocks disappear but the test
> eventually fails because the fake history is not the right size.
> See https://github.com/fish-shell/fish-shell/issues/11933 for more details.
> 
> I wrote a test case in pure C (attached) that also triggers the issue
> although it's not as systematic (30-50%).
> To compile: gcc main.c -o test.exe
> To run: ./test.exe
> 
> 
> Most failures look like this:
>     ```
>     $ ./test.exe
>     tmp_dir: /tmp/flockc2Hz4c
>     open file error: 21 - Is a directory
>             /tmp/flockc2Hz4c/append_file
>     assertion "file_fd >= 0" failed: file "main.c", line 49, function:
> thread_func
>     Aborted
>     ```
> Occasionally (maybe 10%), it looks like that:
>     ```
>     $ ./test.exe
>     tmp_dir: /tmp/flock5Oly9J
>     lock error: 14 - Bad address
>     assertion "lock_res == 0" failed: file "main.c", line 38,
> function: thread_func
>     Aborted
>     ```
> 
> I believe the freeze/deadlock in the Fish test is because, unlike my
> test, they don't assert/crash, and the next time they access the
> history file, there is a bunch of deadlock in cygwin internals.
> If that helps, this is a partial capture of the stack traces at one such time:
>     ```
>     Thread 9
>     #2  0x00000001800d487f in muto::acquire (this=0x1802c24c0
> <lock_process::locker>, ms=ms@entry=4294967295) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/sync.cc:84
>     #3  0x00000001800dd6e0 in dtable::lock (this=<optimized out>) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/local_includes/dtable.h:77
>     #4  cygheap_fdnew::cygheap_fdnew (this=<synthetic pointer>,
> seed_fd=-1, lockit=true) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/local_includes/cygheap.h:593
>     #5  open (unix_path=0xa0002b3b0
> "[...]/fish-shell/target/fish-test-home", flags=262144) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/syscalls.cc:1576
> 
>     Thread 10
>     #2  0x00000001800d487f in muto::acquire (this=0x1802c24c0
> <lock_process::locker>, ms=ms@entry=4294967295) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/sync.cc:84
>     #3  0x00000001800dd6e0 in dtable::lock (this=<optimized out>) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/local_includes/dtable.h:77
>     #4  cygheap_fdnew::cygheap_fdnew (this=<synthetic pointer>,
> seed_fd=-1, lockit=true) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/local_includes/cygheap.h:593
>     #5  open (unix_path=0xa0002bfe0
> "[...]/fish-shell/target/fish-test-home/race_test_history.FwyAgK",
> flags=264706) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/syscalls.cc:1576
> 
>     Thread 11
>     #2  0x00000001800670bb in inode_t::LOCK (this=0x80000ba20) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/flock.cc:314
>     #3  inode_t::get (dev=1881899537, ino=ino@entry=10977524092162599,
> create_if_missing=create_if_missing@entry=false, lock=lock@entry=true)
> at /d/S/B/src/msys2-runtime/winsup/cygwin/flock.cc:504
>     #4  0x0000000180068eb1 in fhandler_base::del_my_locks
> (this=0x80000b810, from=on_close) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/flock.cc:402
>     #5  0x000000018010d5bf in fhandler_base::close_with_arch
> (this=0x80000b810, flag=flag@entry=-1) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/fhandler/base.cc:1306
>     #6  0x00000001800de36b in __close (fd=5, flag=-1) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/syscalls.cc:1710
>     #7  close (fd=5) at /d/S/B/src/msys2-runtime/winsup/cygwin/syscalls.cc:1722
> 
>     Thread 12
>     #2  0x00000001800d487f in muto::acquire (this=0x1802c24c0
> <lock_process::locker>, ms=ms@entry=4294967295) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/sync.cc:84
>     #3  0x00000001800dd6e0 in dtable::lock (this=<optimized out>) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/local_includes/dtable.h:77
>     #4  cygheap_fdnew::cygheap_fdnew (this=<synthetic pointer>,
> seed_fd=-1, lockit=true) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/local_includes/cygheap.h:593
>     #5  open (unix_path=0x7ff10b488
> "[...]/fish-shell/target/fish-test-home/race_test_history.pZO5DS",
> flags=263169) at
> /d/S/B/src/msys2-runtime/winsup/cygwin/syscalls.cc:1576
>     ```
> The freeze/deadlock can be reproduced in my C code by calling
> "continue" inside the "if (lock_res != 0) {" instead of triggering the
> assert just after.
> 
> 
> I haven't been able to reproduce the missing data in the history file
> so it's unknown if it's an issue in Fish or flock not locking properly
> at times. So far the test passes on Linux and MacOS.

Thanks for the report.

Do you have any evidence that flock() should be MT-safe?
This issue seems to be caused by the fact that flock() in
cygwin is not MT-safe.

If flock is guarded by mutex in your test case, the issue
does not happen.

-- 
Takashi Yano <takashi.yano@nifty.ne.jp>