Syscalls stat() and faccessat() sometimes fail with errno EINTR when accessing SMB share through VPN

[Tobias Bading <tbading@web.de>]

Hi.

Everyone doing alright hacking/working from home?

Yesterday I encountered a curious problem while using (a self-built)
Emacs 26.3.50 on my GNU/Linux machine at home:

I've set up the automounter to mount SMB shares of Windows servers in
the office through the company's VPN. This works fine for e.g. "ls -lR
/smb/server/share/dir" in a shell, except for bad performance. The
problems start when I try to work with the same directory from within
Emacs with dired-mode, i.e. a simple C-x C-f /smb/server/share/dir.
Quite regularly I get errors like "dired-get-file-for-visit: File no
longer exists; type ‘g’ to update Dired buffer", although nothing has
changed in the directory.

So I put Emacs under a microscope with

strace -f -e trace=%file -tt emacs 2>&1 | grep --line-buffered -A2 /smb/
 >emacs.log

which revealed errors like

faccessat(AT_FDCWD, "/smb/server/share/dir", F_OK) = -1 EINTR
(Interrupted system call)
--- SIGIO {si_signo=SIGIO, si_code=SI_KERNEL} ---
[...]
stat("/smb/server/share/dir", 0x7fffe49383b0) = -1 EINTR (Interrupted
system call)
--- SIGIO {si_signo=SIGIO, si_code=SI_KERNEL} ---
--- SIGIO {si_signo=SIGIO, si_code=SI_KERNEL} ---

(The timestamps of the SIGIO lines suggest that these signals have
nothing to do with the EINTR errors reported beforehand, the timestamps
are often over 0.1 seconds apart.)

So far I've only seen stat() and faccessat() failing with EINTR. The
funny thing is, the man pages of those two system calls don't mention
EINTR at all. man signal(7) also doesn't mention these functions in the
paragraph about SA_RESTART. Anyway, I've checked the source code of my
Emacs 26.3.50 build and found emacs_sigaction_flags() in src/sysdep.c,
which does return 0 (as intended by the dev(s) who wrote the code). I've
changed the implementation to "return SA_RESTART;", but that had no effect.

To make sure I didn't mess up my own Emacs 26 git branch somehow, I did
a quick test with the current HEAD of origin/master and "src/emacs -Q",
which seems to have the same problem, revealed by error messages like
"apply: Setting current directory: Interrupted system call,
/smb/server/share/dir/".

I'm stumped. A (shell-command "ls -lAFNR /smb/server/share/big-dir/")
works fine, as does a "cp -a" of that directory. But when the Emacs
process itself calls stat() or faccessat(), things go sideways? Why?
What am I missing? Are stat() and faccessat() even allowed to fail with
EINTR? Is this a kernel bug, maybe somewhere in the CIFS client
implementation? But an strace of "ls -lAFNR /smb/server/share/big-dir/"
shows not a single EINTR! So why would only Emacs be affected?

Please enlighten me... ;)

Tobias