utimensat doesn't exist, or is in glibc but kernel 2.6.18 fails with ENOSYS
kernel 2.6.22 and earlier rejects AT_SYMLINK_NOFOLLOW
kernel 2.6.25 and earlier reject UTIME_NOW/UTIME_OMIT with non-zero tv_sec
- kernel 2.6.32 and earlier fail to bump ctime if mtime is UTIME_OMIT
+ kernel 2.6.32 used with xfs or ntfs-3g fail to honor UTIME_OMIT
utimensat completely works
For each cache variable: 0 = unknown, 1 = yes, -1 = no. */
static int utimensat_works_really;
static int lutimensat_works_really;
-static int utimensat_ctime_really;
-
-/* Determine whether the kernel has a ctime bug. ST1 and ST2
- correspond to stat data before and after a successful time change.
- TIMES contains the timestamps that were used during the time change
- (mtime will be UTIME_OMIT). Update the cache variable if there is
- conclusive evidence of the kernel working or being buggy. Return
- true if TIMES has been updated and another kernel call is needed,
- whether or not the kernel is known to have the bug. */
-static bool
-detect_ctime_bug (struct stat *st1, struct stat *st2, struct timespec times[2])
-{
- struct timespec now;
- if (st1->st_ctime != st2->st_ctime
- || get_stat_ctime_ns (st1) != get_stat_ctime_ns (st2))
- {
- utimensat_ctime_really = 1;
- return false;
- }
- /* The results are inconclusive if the ctime in st1 is within a file
- system quantization window of now. For FAT, this is 2 seconds,
- for systems with sub-second resolution, a typical resolution is
- 10 milliseconds; to be safe we declare an inconsistent result if
- ctime is within a 20 millisecond window. Avoid an extra gettime
- call if atime makes sense. It is unlikely that the original
- ctime is later than now, but rather than deal with the overflow,
- we treat that as consistent evidence of the bug. */
- if (times[0].tv_nsec == UTIME_NOW)
- now = get_stat_atime (st2);
- else
- gettime (&now);
- if (now.tv_sec < st2->st_ctime
- || 2 < now.tv_sec - st2->st_ctime
- || (get_stat_ctime_ns (st2)
- && now.tv_sec - st2->st_ctime < 2
- && (20000000 < (1000000000 * (now.tv_sec - st2->st_ctime)
- + now.tv_nsec - get_stat_ctime_ns (st2)))))
- utimensat_ctime_really = -1;
- times[1] = get_stat_mtime (st2);
- return true;
-}
#endif /* HAVE_UTIMENSAT || HAVE_FUTIMENS */
/* Validate the requested timestamps. Return 0 if the resulting
timespec can be used for utimensat (after possibly modifying it to
work around bugs in utimensat). Return a positive value if the
timespec needs further adjustment based on stat results: 1 if any
- adjustment is needed for utimes, and 2 if mtime was UTIME_OMIT and
- an adjustment is needed for utimensat. Return -1, with errno set
- to EINVAL, if timespec is out of range. */
+ adjustment is needed for utimes, and 2 if any adjustment is needed
+ for Linux utimensat. Return -1, with errno set to EINVAL, if
+ timespec is out of range. */
static int
validate_timespec (struct timespec timespec[2])
{
int result = 0;
+ int utime_omit_count = 0;
assert (timespec);
if ((timespec[0].tv_nsec != UTIME_NOW
&& timespec[0].tv_nsec != UTIME_OMIT
{
timespec[0].tv_sec = 0;
result = 1;
+ if (timespec[0].tv_nsec == UTIME_OMIT)
+ utime_omit_count++;
}
- if (timespec[1].tv_nsec == UTIME_NOW)
+ if (timespec[1].tv_nsec == UTIME_NOW
+ || timespec[1].tv_nsec == UTIME_OMIT)
{
timespec[1].tv_sec = 0;
result = 1;
+ if (timespec[1].tv_nsec == UTIME_OMIT)
+ utime_omit_count++;
}
- else if (timespec[1].tv_nsec == UTIME_OMIT)
- {
- timespec[1].tv_sec = 0;
- result = 2;
- }
- return result;
+ return result + (utime_omit_count == 1);
}
/* Normalize any UTIME_NOW or UTIME_OMIT values in *TS, using stat
if (0 <= utimensat_works_really)
{
int result;
- struct stat st1;
- struct stat st2;
- /* Linux kernel 2.6.32 has a bug where it fails to bump ctime if
- UTIME_OMIT was used for mtime. It costs time to do an extra
- [f]stat up front, so we cache whether the function works. */
- if (utimensat_ctime_really <= 0 && adjustment_needed == 2)
+# if __linux__
+ struct stat st;
+ /* As recently as Linux kernel 2.6.32 (Dec 2009), several file
+ systems (xfs, ntfs-3g) have bugs with a single UTIME_OMIT,
+ but work if both times are either explicitly specified or
+ UTIME_NOW. Work around it with a preparatory [f]stat prior
+ to calling futimens/utimensat; fortunately, there is not much
+ timing impact due to the extra syscall even on file systems
+ where UTIME_OMIT would have worked. FIXME: Simplify this in
+ 2012, when file system bugs are no longer common. */
+ if (adjustment_needed == 2)
{
- if (fd < 0 ? stat (file, &st1) : fstat (fd, &st1))
+ if (fd < 0 ? stat (file, &st) : fstat (fd, &st))
return -1;
if (ts[0].tv_nsec == UTIME_OMIT)
- return 0;
- if (utimensat_ctime_really < 0)
- ts[1] = get_stat_mtime (&st1);
+ ts[0] = get_stat_atime (&st);
+ else if (ts[1].tv_nsec == UTIME_OMIT)
+ ts[1] = get_stat_mtime (&st);
}
+# endif /* __linux__ */
# if HAVE_UTIMENSAT
if (fd < 0)
{
if (result == 0 || errno != ENOSYS)
{
utimensat_works_really = 1;
- if (result == 0 && utimensat_ctime_really == 0
- && adjustment_needed == 2)
- {
- /* Perform a followup stat to see if the kernel has
- a ctime bug. */
- if (stat (file, &st2))
- return -1;
- if (detect_ctime_bug (&st1, &st2, ts))
- result = utimensat (AT_FDCWD, file, ts, 0);
- }
return result;
}
}
if (result == 0 || errno != ENOSYS)
{
utimensat_works_really = 1;
- /* Work around the same bug as above. */
- if (result == 0 && utimensat_ctime_really == 0
- && adjustment_needed == 2)
- {
- if (fstat (fd, &st2))
- return -1;
- if (detect_ctime_bug (&st1, &st2, ts))
- result = futimens (fd, ts);
- }
return result;
}
}
if (0 <= lutimensat_works_really)
{
int result;
- struct stat st1;
- struct stat st2;
- /* Linux kernel 2.6.32 has a bug where it fails to bump ctime if
- UTIME_OMIT was used for mtime. It costs time to do an extra
- lstat up front, so we cache whether the function works. */
- if (utimensat_ctime_really <= 0 && adjustment_needed == 2)
+# if __linux__
+ struct stat st;
+ /* As recently as Linux kernel 2.6.32 (Dec 2009), several file
+ systems (xfs, ntfs-3g) have bugs with a single UTIME_OMIT,
+ but work if both times are either explicitly specified or
+ UTIME_NOW. Work around it with a preparatory lstat prior to
+ calling utimensat; fortunately, there is not much timing
+ impact due to the extra syscall even on file systems where
+ UTIME_OMIT would have worked. FIXME: Simplify this in 2012,
+ when file system bugs are no longer common. */
+ if (adjustment_needed == 2)
{
- if (lstat (file, &st1))
+ if (lstat (file, &st))
return -1;
if (ts[0].tv_nsec == UTIME_OMIT)
- return 0;
- if (utimensat_ctime_really < 0)
- ts[1] = get_stat_mtime (&st1);
+ ts[0] = get_stat_atime (&st);
+ else if (ts[1].tv_nsec == UTIME_OMIT)
+ ts[1] = get_stat_mtime (&st);
}
+# endif /* __linux__ */
result = utimensat (AT_FDCWD, file, ts, AT_SYMLINK_NOFOLLOW);
# ifdef __linux__
/* Work around a kernel bug:
{
utimensat_works_really = 1;
lutimensat_works_really = 1;
- if (result == 0 && utimensat_ctime_really == 0
- && adjustment_needed == 2)
- {
- /* Perform a followup stat to see if the kernel has a
- ctime bug. */
- if (lstat (file, &st2))
- return -1;
- if (detect_ctime_bug (&st1, &st2, ts))
- result = utimensat (AT_FDCWD, file, ts, AT_SYMLINK_NOFOLLOW);
- }
return result;
}
}
{
/* See comments in utimens.c for details. */
static int utimensat_works_really; /* 0 = unknown, 1 = yes, -1 = no. */
- static int utimensat_ctime_really; /* 0 = unknown, 1 = yes, -1 = no. */
if (0 <= utimensat_works_really)
{
int result;
- struct stat st1;
- struct stat st2;
+# ifdef __linux__
+ struct stat st;
struct timespec ts[2];
- /* Linux kernel 2.6.32 has a bug where mtime of UTIME_OMIT fails
- to change ctime. */
- if (utimensat_ctime_really <= 0 && times
- && times[0].tv_nsec != UTIME_OMIT && times[1].tv_nsec == UTIME_OMIT)
+ /* As recently as Linux kernel 2.6.32 (Dec 2009), several file
+ systems (xfs, ntfs-3g) have bugs with a single UTIME_OMIT,
+ but work if both times are either explicitly specified or
+ UTIME_NOW. Work around it with a preparatory [l]stat prior
+ to calling utimensat; fortunately, there is not much timing
+ impact due to the extra syscall even on file systems where
+ UTIME_OMIT would have worked. FIXME: Simplify this in 2012,
+ when file system bugs are no longer common. */
+ if (times && (times[0].tv_nsec == UTIME_OMIT
+ || times[1].tv_nsec == UTIME_OMIT))
{
- if (fstatat (fd, file, &st1, flag))
+ if (fstatat (fd, file, &st, flag))
return -1;
- if (utimensat_ctime_really < 0)
- {
- ts[0] = times[0];
- ts[1] = get_stat_mtime (&st1);
- times = ts;
- }
+ if (times[0].tv_nsec == UTIME_OMIT && times[1].tv_nsec == UTIME_OMIT)
+ return 0;
+ if (times[0].tv_nsec == UTIME_OMIT)
+ ts[0] = get_stat_atime (&st);
+ else
+ ts[0] = times[0];
+ if (times[1].tv_nsec == UTIME_OMIT)
+ ts[1] = get_stat_mtime (&st);
+ else
+ ts[1] = times[1];
+ times = ts;
}
+# endif /* __linux__ */
result = utimensat (fd, file, times, flag);
/* Linux kernel 2.6.25 has a bug where it returns EINVAL for
UTIME_NOW or UTIME_OMIT with non-zero tv_sec, which
if (result == 0 || (errno != ENOSYS && errno != EINVAL))
{
utimensat_works_really = 1;
- if (result == 0 && utimensat_ctime_really == 0 && times
- && times[0].tv_nsec != UTIME_OMIT
- && times[1].tv_nsec == UTIME_OMIT)
- {
- /* Perform a followup [l]stat. See detect_ctime_bug in
- utimens.c for more details. */
- struct timespec now;
- if (fstatat (fd, file, &st2, flag))
- return -1;
- if (st1.st_ctime != st2.st_ctime
- || get_stat_ctime_ns (&st1) != get_stat_ctime_ns (&st2))
- {
- utimensat_ctime_really = 1;
- return result;
- }
- if (times[0].tv_nsec == UTIME_NOW)
- now = get_stat_atime (&st2);
- else
- gettime (&now);
- if (now.tv_sec < st2.st_ctime
- || 2 < now.tv_sec - st2.st_ctime
- || (get_stat_ctime_ns (&st2)
- && now.tv_sec - st2.st_ctime < 2
- && (20000000 < (1000000000 * (now.tv_sec - st2.st_ctime)
- + now.tv_nsec
- - get_stat_ctime_ns (&st2)))))
- utimensat_ctime_really = -1;
- ts[0] = times[0];
- ts[1] = get_stat_mtime (&st2);
- result = utimensat (fd, file, ts, flag);
- }
return result;
}
}
projects / pspp / commitdiff