* tests/nap.h (nap_get_stat): Renamed from get_stat. Remove argument fd;
use nap_fd instead. On native Windows, close and reopen nap_fd.
(nap_works): Don't compare the ctimes, because on native Windows, these
are the creation times.
(nap): Update.
* build-aux/gendocs.sh (version):
* doc/gendocs_template:
* doc/gendocs_template_min:
* doc/gnulib.texi:
* lib/version-etc.c (COPYRIGHT_YEAR):
Update copyright dates by hand in templates and the like.
* all files: Run 'make update-copyright'.
* build-aux/gendocs.sh (version):
* doc/gendocs_template:
* doc/gendocs_template_min:
* doc/gnulib.texi:
* lib/version-etc.c (COPYRIGHT_YEAR):
Update copyright dates by hand in templates and the like.
* all files: Run 'make update-copyright'.
The recent change in nap.h (5191133e) decreased the probability of lost
races to about a third, however such problems could still be observed
in virtual machines and openSUSE's OBS.
Instead of calulating the nap() time once and using it (together with
a small correction multiplier), avoid the race alltogether by verifying
on a reference file whether a timestamp difference has happened.
Before, nap() detected the needed time once empirically and then used
that delay (together with a small correction multiplier) in further
calls. This problem has been reported and discussed several times,
including guesses about possible kernel issues:
https://lists.gnu.org/archive/html/bug-gnulib/2013-04/msg00071.htmlhttp://lists.gnu.org/archive/html/coreutils/2012-03/msg00088.htmlhttps://lists.gnu.org/archive/html/bug-gnulib/2011-11/msg00226.htmlhttp://bugs.gnu.org/12820https://lists.gnu.org/archive/html/bug-gnulib/2010-11/msg00113.htmlhttps://lists.gnu.org/archive/html/bug-gnulib/2009-11/msg00007.html
Now, nap() avoids the race alltogether by verifying on a reference
file whether a timestamp difference has happened.
* tests/nap.h (nap_fd): Define file descriptor variable for the
witness file.
(nap_works): Change return value to bool. Change passing
the old file's status by value instead of by reference as this function
does no longer update that timestamp; rename the function argument from
st to old_st. Remove the local variables cdiff and mdiff because that
function now returns true/false instead of the precise delay.
(guess_delay): Remove function.
(clear_tmp_file): Add new function to close and unlink the witness file.
(nap): Instead of re-using the delay which has been calculated during
the first call, avoid the race by actually verifying that a timestamp
difference can be observed on the current file system. Use an adaptive
approach for the delay to minimize execution time. Assert that the
maximum delay is <= ~2 seconds, more precisely sum(2^n) from 0 to 30
= 2^31 - 1 = 2.1s.
Use atexit to call clear_tmp_file when the process terminates.
Problem reported by Bernhard Voelker in
<http://lists.gnu.org/archive/html/bug-gnulib/2013-04/msg00071.html>.
I don't know whether this patch fixes that race condition, but it
fixes *some* race conditions, so it should be a win.
* modules/chown-tests (Depends-on):
* modules/fchownat-tests (Depends-on):
* modules/fdutimensat-tests (Depends-on):
* modules/futimens-tests (Depends-on):
* modules/lchown-tests (Depends-on):
* modules/stat-time-tests (Depends-on):
* modules/utimens-tests (Depends-on):
* modules/utimensat-tests (Depends-on):
Depend on nanosleep, not usleep.
* modules/chown-tests (test_chown_LDADD):
* modules/lchown-tests (test_lchown_LDADD):
* modules/stat-time-tests (test_stat_time_LDADD):
New macro.
* modules/fchownat-tests (test_fchownat_LDADD):
* modules/fdutimensat-tests (test_fdutimensat_LDADD):
* modules/futimens-tests (test_futimens_LDADD):
* modules/utimens-tests (test_utimens_LDADD):
* modules/utimensat-tests (test_utimensat_LDADD):
Add $(LIB_NANOSLEEP).
* modules/stat-time-tests (Files): Add tests/nap.h.
* tests/nap.h: Include <limits.h>, for INT_MAX.
(lt_mtime): Remove.
(diff_timespec): New function.
(get_stat): Rename from get_mtime. All callers changed.
(nap_works): Determine the needed delay by inspecting the
file system's timestamp jumps; this should be more reliable.
Look at both mtime and ctime, and take the maximum of the two jumps.
(nap_works, guess_delay):
Return a nanosecond cound, not a microsecond count.
All callers changed.
(nap_works, nap): Use nanosleep, not usleep. Check for nanosleep
failure.
(nap): Multiply the guess by 1.125, to accommodate the case where
the file system's clock is a bit slower than nanosleep's clock.
* tests/test-stat-time.c (BASE): New macro.
Include nap.h.
(nap): Remove; nap.h now defines this. This removes a duplicate
implementation of 'nap'.
* tests/nap.h (lt_mtime, get_mtime, nap_works, guess_delay):
New functions.
(nap): Use them, to do a better job of guessing the delay.
On Fedora 17 with ext4 atop md atop hard disks, this made
test-utimens run 10x faster, because the test napped for
1 ms at a time rather than 20 ms. Reported by Stefano Lattarini in
<http://bugs.gnu.org/12820#11>.
Fixes a spurious failure on ext3, with one-second resolution,
now that ctime effects are being tested for inequality.
* tests/test-chown.h (nap): Factor...
* tests/nap.h: ...into new file.
* tests/test-lchown.h (nap): Avoid duplication.
* tests/test-utimens-common.h (nap): Use shared implementation,
necessary on file systems with 1-second resolution.
* modules/chown-tests (Files): Include new file.
* modules/fdutimensat-tests (Files): Likewise.
* modules/futimens-tests (Files): Likewise.
* modules/lchown-tests (Files): Likewise.
* modules/openat-tests (Files): Likewise.
* modules/utimens-tests (Files): Likewise.
* modules/utimensat-tests (Files): Likewise.
Signed-off-by: Eric Blake <ebb9@byu.net>
