fix: keep mtime when utimens only updates atime (UTIME_OMIT)

[Yike-Ye]

Problem

On an sshfs mount, updating only a file's access time also changes its modification time on the server. On macOS this is very visible: pressing Space in Finder to Quick Look a file immediately changes its "Date Modified" (macfuse/macfuse#1157). The same bug was reported for FreeBSD (and macOS) in #282.

Minimal reproducer:

$ touch -t 202001010000 f      # atime = mtime = 2020-01-01
$ touch -a f                   # update atime only
$ stat -f 'atime=%Sa  mtime=%Sm' f

On a local file system mtime stays at 2020-01-01; on sshfs it becomes the current time.

Root cause

When only one of the two times is being changed, libfuse's high-level fuse_lib_setattr() passes the other one as { .tv_sec = 0, .tv_nsec = UTIME_OMIT } (and "now" as { .tv_sec = 0, .tv_nsec = UTIME_NOW }).

f045211 ("Make utimens(NULL) work correctly", #130, fixing #112) made sshfs_utimens() treat tv_sec == 0 as "now". That fixed UTIME_NOW, but UTIME_OMIT also has tv_sec == 0, so an omitted time is replaced with the current time. Because SSH_FILEXFER_ATTR_ACMODTIME always sets both atime and mtime, the server's mtime is overwritten.

Tracing the arguments (temporary debug print in sshfs_utimens(), macOS 27 + macFUSE 5.3.3) for utimensat(AT_FDCWD, f, {{0, UTIME_NOW}, {0, UTIME_OMIT}}, 0):

in atime={1789113909,180150000} mtime={0,-2} (UTIME_OMIT=-2) -> SETSTAT atime=1789113909 mtime=1789113909

@bfleischer also noted in macfuse/macfuse#1157 that the LoopbackFS reference file system does not show this, which points at sshfs.

Fix

Decide by tv_nsec instead of guessing from tv_sec:

• UTIME_OMIT: send the time's current value, read with sshfs_getattr() (only when a time is actually omitted)
• UTIME_NOW: send the current time (keeps the utimensat(NULL) not handled correctly #112 fix)
• both omitted: send nothing
• otherwise: send the given tv_sec

SFTP v3 cannot set just one of the two times, so reading the current value first is unavoidable; a concurrent change to the omitted time between the read and the SETSTAT would be lost.

Testing

• New tst_utimens_omit in test/test_sshfs.py: set both times 1000 s in the past, run touch -a, assert that mtime is unchanged. touch -a passes UTIME_OMIT for mtime via utimensat().
• The pytest suite skips on macOS (it needs fusermount), so I called tst_utimens, tst_utimens_now, tst_utimens_omit, tst_create and tst_open_write directly against a localhost: mount using the same cache options as test_sshfs(): with this patch 5/5 pass; current master fails only tst_utimens_omit.
• Against a remote Linux/OpenSSH server: an atime-only update via utimensat() and via setattrlist(ATTR_CMN_ACCTIME) now keeps mtime; touch, touch -m -t, touch -t and creating a new file behave as before (no 1970 timestamps).
• Built with clang and the strict-warnings CI flags: no new warnings. I have not run the Linux CI jobs locally.

Fixes #282
Refs: #112, #130, macfuse/macfuse#1157

🤖 Generated with Claude Code