* src/tail.c (tail_forever_inotify): Only monitor write()s and
truncate()s to files in --follow=descriptor mode, thus avoiding
the bug where we removed the watch on renamed files.
Also adjust the inotify event processing code that is
now significant only in --follow=name mode.
* tests/tail-2/F-vs-rename.sh: Improve this existing test by running
in both polling and inotify modes.
* tests/tail-2/f-vs-rename.sh: A new test based on the existing one.
* tests/local.mk: Reference the new test.
* NEWS: Mention the bug.
Fixes http://bugs.gnu.org/19760
Each user has a maximum number of inotify watches,
so handle the cases where we exhaust these resources.
* src/tail.c (tail_forever_inotify): Ensure we inotify_rm_watch()
the watch for an inode, when replacing with a new watch for a name.
Return all used inotify resources when reverting to polling.
Revert to polling upon first indication of inotify resource exhaustion.
Revert to polling on any inotify resource exhaustion.
Diagnose resource exhaustion correctly in all cases.
Avoid redundant reinsertion in the hash for unchanged watches
(where only attributes of the file are changed).
* tests/tail-2/retry.sh: Avoid false failure when reverting to polling.
* tests/tail-2/inotify-rotate.sh: Likewise.
* tests/tail-2/symlink.sh: Likewise.
* tests/tail-2/inotify-rotate-resources.sh: New test to check
that we're calling inotify_rm_watch() for replaced files.
* tests/local.mk: Reference the new test.
* NEWS: Mention the bug fix.
* THANKS.in: Thanks for reporting and problem identification.
* tests/tail-2/inotify-rotate.sh (cleanup_fail_): Set fail=1
so that failures are identified. Regression in v8.23-63-g111a2b9
Also use print_ver_ rather than open coding --verbose support.
Also check for more than a single 'b' which seems brittle.
When some program produces unexpected output, that use of
compare-vs-/dev/null will ensure that the surprising output is
printed in the test's output. With "test -s err" only, one
would have to instrument and rerun in order to see the offending
output.
* cfg.mk (sc_prohibit_and_fail_1): Exempt 'compare' from this check.
* tests/dd/misc.sh: Change "tests -s ... || fail=1" to
"compare /dev/null ... && fail=1".
* tests/misc/nice.sh: Likewise.
* tests/rm/read-only.sh: Likewise.
* tests/tail-2/inotify-race.sh: Likewise.
* tests/touch/no-dereference.sh: Likewise.
Suggested by Jim Meyering in
http://lists.gnu.org/archive/html/coreutils/2015-01/msg00042.html
Run "make update-copyright" and then...
* tests/sample-test: Adjust to use the single most recent year.
* tests/du/bind-mount-dir-cycle-v2.sh: Fix case in copyright message,
so that year is updated automatically in future.
Commit v8.23-63-g111a2b9 removed the expensive tag on this test,
as it runs quickly on systems with inotify. However without that
it would take about 8 minutes for the test to complete all iterations.
* tests/tail-2/inotify-rotate.sh: Tag as expensive without inotify.
Also adjust the polling parameters used on systems without inotify
so that the test completes within about 15 seconds.
* tests/tail-2/inotify-rotate.sh: Use retry_delay_
to employ an exponential backoff with a total delay of
up to 25.5s. The 15s delay was seen to trigger a false
failure in http://hydra.nixos.org/build/16546517
Also remove the .1s sleep in each of the 50 iterations
to reduce the running time of the test and thus the
expensive_ tag on this test was removed.
Also ensure that we use the standard exit procedure
upon failure to avoid any erroneous diagnostics due
to persistent files on NFS.
Fix similar problems in head, od, split, tac, and tail.
Reported by George Shuklin in: http://bugs.gnu.org/18621
* NEWS: Document this.
* src/head.c (elseek): Move up.
(elide_tail_bytes_pipe, elide_tail_lines_pipe): New arg
CURRENT_POS. All uses changed.
(elide_tail_bytes_file, elide_tail_lines_file):
New arg ST and remove arg SIZE. All uses changed.
* src/head.c (elide_tail_bytes_file):
* src/od.c (skip): Avoid optimization for /sys files, where
st_size is bogus and st_size == st_blksize.
Don't report error at EOF when not optimizing.
* src/head.c, src/od.c, src/tail.c: Include "stat-size.h".
* src/split.c (input_file_size): New function.
(bytes_split, lines_chunk_split, bytes_chunk_extract): New arg
INITIAL_READ. All uses changed. Use it to double-check st_size.
* src/tac.c (tac_seekable): New arg FILE_POS. All uses changed.
(copy_to_temp): Return size of temp file. All uses changed.
* src/tac.c (tac_seekable):
* src/tail.c (tail_bytes):
* src/wc.c (wc):
Don't trust st_size; double-check by reading.
* src/wc.c (wc): New arg CURRENT_POS. All uses changed.
* tests/local.mk (all_tests): Add tests/misc/wc-proc.sh,
tests/misc/od-j.sh, tests/tail-2/tail-c.sh.
* tests/misc/head-c.sh:
* tests/misc/tac-2-nonseekable.sh:
* tests/split/b-chunk.sh:
Add tests for problems with /proc and /sys files.
* tests/misc/od-j.sh, tests/misc/wc-proc.sh, tests/tail-2/tail-c.sh:
New files.
On most shells `:>file || framework_failure_` will not evaluate
the framework_failure_ even if there was an error writing the file.
shells which do evaluate the failure are ksh 93u+ and bash 4.2,
while shells wich don't include bash 4.3, solaris, freebsd, dash.
Furthermore this construct is problematic on Solaris 10 sh,
which will try to optimize away a `:' command in a loop
after the first iteration, even if it is redirected.
* tests/cp/link-deref.sh: Remove the leading colon on redirections.
* tests/cp/reflink-perm.sh: Likewise.
* tests/id/zero.sh: Likewise.
* tests/install/install-C.sh: Likewise.
* tests/misc/env.sh: Likewise.
* tests/misc/md5sum-bsd.sh: Likewise.
* tests/misc/runcon-no-reorder.sh: Likewise.
* tests/mv/partition-perm.sh: Likewise.
* tests/rm/r-root.sh: Likewise.
* tests/split/l-chunk.sh: Likewise.
* tests/split/line-bytes.sh: Likewise.
* tests/tail-2/inotify-rotate.sh: Likewise.
* tests/tail-2/retry.sh: Likewise.
* tests/tail-2/symlink.sh: Likewise.
* tests/tail-2/wait.sh: Likewise.
* tests/touch/read-only.sh: Likewise.
+ cfg.mk (sc_prohibit_colon_redirection): A new syntax check
to avoid further instances of this creeping in.
Run "make update-copyright", but then also run this,
perl -pi -e 's/2\d\d\d-//' tests/sample-test
to make that one script use the single most recent year number.
Inline functions are awkward to breakpoint as mentioned at:
https://sourceware.org/bugzilla/show_bug.cgi?id=10738
The normal case here was for the breakpoint on the inline function
to fail, resulting in a 10s delay before skipping the test.
However on GCC 4.7.2 on ppc64 at least it was seen that
the test failed erroneously due to the breakpoint being successfully
set on an "out of line" function, but an inline function was
actually being called.
* tests/tail-2/inotify-race.sh: Switch to a line based breakpoint,
rather than a symbol base one, which avoids issues with breakpoints
on inline functions. Also skip_ on the initial breakpoint check
in case the breakpoint is not traversed which would be the case
on remote file systems for example.
* src/tail.c: With inotify, when a file is initially absent,
we fstat(-1) for that file spec, thus recording an errnum of EBADF,
which caused the "has become accessible" diagnostic to be issued,
when the file first appears. Instead we avoid the fstat(-1) and
thus emit the more natural and consistent "has appeared" diagnostic.
* tests/tail-2/retry.sh: Use the new diagnostic which also causes
this test to pass on systems without inotify.
Previous behavior failed to read contents of a (re)appearing file,
when symlinked by tail's watched file. Also we now diagnose other
edge cases when running in inotify mode, where an initially
missing or regular file changes to a symlink.
* src/tail.c (main): If any arg is a symlink, use polling mode.
(recheck): Diagnose the edge case where a symlink appears during
inotify processing.
* tests/tail-2/symlink.sh: Test the fix. Mention the edge cases.
* tests/local.mk: Reference the new test.
* NEWS: Mention the fix.
Reported by: Ondrej Oprala
Some newer test scripts - partially ones from me - are not executable.
It does not seem to be a problem, but for consistency and to avoid
future problems on unusual platforms or shells change the permissions
by adding the executable bit.
* cfg.mk (sc_tests_executable): Add new syntax-check rule to ensure
that all test scripts are executable.
* tests/df/df-output.sh: Change file mode from 644 to 755.
* tests/du/threshold.sh: Likewise.
* tests/factor/run.sh: Likewise.
* tests/init.sh: Likewise.
* tests/misc/csplit-suppress-matched.pl: Likewise.
* tests/misc/numfmt.pl: Likewise.
* tests/tail-2/retry.sh: Likewise.
Prompted by the continuous integration build failure at:
http://hydra.nixos.org/build/5221053
* tests/tail-2/retry.sh: Ensure the 'out' file is truncated,
as it's used to arbitrate the run order of commands.
Relying on the truncation in the background tail command
is racy because the truncation can occur after the fork
of the background shell and thus wait4lines would not wait
for output to occur in 'out', which would mean that the
'missing' file could be populated by the time tail(1)
gets to process it.
As a side effect of the previous commit which fixes 'tail -f --retry'
to wait for a file to appear, tail would not exit when the last file
appears untailable and gives up on this file.
This can happen, for example, when the argument file name appears
as directory. Tail sets the 'ignore' flag of this file to true,
but instead of exiting the program, tail would continue the loop.
* src/tail.c (any_live_files): Change the function to return true
if any of the files is still tailable or if tail should continue to
try to check again.
(tail_forever): Change the condition to break the loop in the
"no files remaining" case, because now any_live_files() will care
about it, as mentioned above.
(parse_options): When --retry is used without any follow mode,
then reset reopen_inaccessible_files to false.
* tests/tail-2/retry.sh: Add test case.
The --retry option is indeed useful for both following modes
by name and by file descriptor. The difference is that in the
latter case, it is effective only during the initial open.
As a regression of the implementation of the inotify support,
tail -f --retry would immediately exit if the given file is
inaccessible.
* src/tail.c (usage): Change the description of the --retry option:
remove the note that this option would mainly be useful when
following by name.
(main): Change diagnosing dubios uses of --retry option:
when the --retry option is used without following, then issue
a warning that this option is ignored; when it is used together
with --follow=descriptor, then issue a warning that it is only
effective for the initial open.
Disable inotify also in the case when the initial open in tail_file()
failed (which is the actual bug fix).
* init.cfg (retry_delay_): Pass excess arguments to the test function.
* tests/tail-2/retry.sh: Add new tests.
* tests/local.mk (all_tests): Mention it.
* doc/coreutils.texi (tail invocation): Enhance the documentation
of the --retry option. Clarify the difference in tail's behavior
regarding the --retry option when combined with the following modes
name versus descriptor.
* NEWS (Bug fixes): Mention the fix.
Reported by Noel Morrison in:
http://lists.gnu.org/archive/html/coreutils/2013-04/msg00003.html
* src/tail.c (main): If -n0 or -c0 were specified without -f,
then no data would ever be output, so exit without reading input.
* tests/tail-2/tail-n0f.sh: Augment the related test with this case.
* tests/tail-2/inotify-rotate.sh: Avoid a subshell with bash,
which in turn causes the `kill` to be ineffective to the tail
processes (as the SIGTERM is sent to the subshell which doesn't
propagate the signal on to its children). On NFS the test
cleanup will then fail as there will be .nfs files maintained
in the directory for the files still opened by the tail processes.
Reported by Bernhard Voelker.
Run "make update-copyright", but then also run this,
perl -pi -e 's/2\d\d\d-//' tests/sample-test
to make that one script use the single most recent year number.
* Makefile.am (SUBDIRS): Remove 'tests'.
(include): The '$(top_srcdir)/tests/local.mk' file.
(check-root): Remove this convenience target, it's no longer needed
now that the "real" check-root target once in 'tests/Makefile' will
land in the top-level makefile.
* configure.ac (AC_CONFIG_FILES): Remove 'tests/Makefile'.
* tests/Makefile.am: Rename ...
* tests/local.mk: ... like this, with a lot of adjustments.
* tests/init.cfg: Move ...
* init.cfg: ... here. This is necessary, for a limitation of the
gnulib-provided 'tests/init.sh', which unconditionally look for
'init.cfg' in the $(srcdir) directory.
* tests/*/*.sh: Adjust: expect init.sh to be in '$srcdir/tests',
not in '$srcdir', and extend $PATH with './src', not with '../src'.
* tests/Coreutils.pm: Adjust similarly.
* tests/pr/pr-tests.pl ($pfx): Likewise.
Not only this shrinks the size of the generated Makefile (from > 6300
lines to ~3000), but will allow further simplifications in future
changes.
* tests/Makefile.am (TEST_EXTENSIONS): Add '.sh' and '.pl'.
(PL_LOG_COMPILER, SH_LOG_COMPILER): New, still defined simply to
$(LOG_COMPILER) for the time being.
(TESTS, root_tests): Adjust as described.
* All tests: Rename as described.
Exempt init.sh because it runs before we're assured to have a
shell that groks $(...). Exempt *.mk because "$" would have to
be doubled, and besides, any `...` expression in a .mk file is
almost certainly evaluated before init.sh is run. Finally, also
exempt the perl-based tests, because perl's `...` cannot be
converted to $(...). Do that by running this command:
git grep -l '`.*`' tests \
| grep -Ev 'init\.sh|\.mk$' | xargs grep -Lw perl \
| xargs perl -pi -e 's/`(.*?)`/\$($1)/g'
One minor fix-up change was required after that, due to how
quoting differs:
diff --git a/tests/chmod/equals b/tests/chmod/equals
- expected_perms=$(eval 'echo \$expected_'$dest)
+ expected_perms=$(eval 'echo $expected_'$dest)
Another was to make these required quoting adjustments:
diff --git a/tests/misc/stty b/tests/misc/stty
...
- rev=$(eval echo "\\\$REV_$opt")
+ rev=$(eval echo "\$REV_$opt")
...
- rev1=$(eval echo "\\\$REV_$opt1")
- rev2=$(eval echo "\\\$REV_$opt2")
+ rev1=$(eval echo "\$REV_$opt1")
+ rev2=$(eval echo "\$REV_$opt2")
Also, transform two files that were needlessly excluded above:
(both use perl, but are mostly bourne shell)
perl -pi -e 's/`(.*?)`/\$($1)/g' \
tests/du/long-from-unreadable tests/init.cfg
Exempt lines with '$' or '=', since those are prone to improper
conversion. Run this:
git grep -l "\`[^']*'" tests \
|xargs perl -pi -e '/[=\$]/ and next;s/\`([^'\''"]*?'\'')/'\''$1/g'
* tests/tail-2/pipe-f2: Don't always wait 10 seconds.
Before, this test would always wait 10 seconds.
Now, it stops early when it detects that tail -f has written output.
BTW, the race condition that prompted changing the timeout from 1 second
to 10 was that tail -f could be killed by the timeout before producing
any output.
These changes were mostly mechanical, made by running the
following command:
git grep -lw framework_failure | grep -v ChangeLog \
| xargs perl -pi -e 's/\b(framework_failure)\b/${1}_/'
and then editing init.cfg and `tests/cp/cp-a-selinux' by hand.
* tests/init.cfg (framework_failure): Remove, `framework_failure_'
from init.sh should be used instead in the tests.
Remove now-obsolete "FIXME" comment.
(is_local_dir_, require_strace_, require_membership_in_two_groups_,
require_sparse_support_, skip_if_mcstransd_is_running_,
mkfifo_or_skip_) Use `framework_failure_', not `framework_failure'.
* Many test scripts: Likewise.
* tests/init.cfg (skip_test_): Remove function.
Use skip_ in place of skip_test_ everywhere else.
* cfg.mk (sc_prohibit_skip_): Remove rule.
* tests/**: Use skip_, not skip_test_, everywhere.
* src/tail.c (tail_forever_inotify): Just as without inotify,
tail --follow=name now terminates when the last tailed-by-name file
is unlinked or moved aside. This bug was introduced on 2009-06-15
via commit ae494d4b, "tail: use inotify if it is available".
Reported by Tim Underwood in
http://thread.gmane.org/gmane.comp.gnu.coreutils.bugs/22286
* NEWS (Bug fixes): Mention this.
* tests/tail-2/follow-name: Test for this.
* tests/Makefile.am (TESTS): Add it.
Otherwise, this would fail (albeit rarely) on a "make -j24 check" run.
* tests/tail-2/pipe-f2: Increase timeout from 1 second to 10,
to avoid false positive failure.
* tests/tail-2/inotify-rotate: Wait 50% longer for grep to succeed.
Without this change, this test would fail consistently when using
"make -j25 check" with F15 in a virtio- and spinning-rust-backed
virtual machine.
Initially, I did this,
git grep -l srcdir/test-lib.sh|xargs perl -p0i -e '~180-byte script'
but that line would have been much longer than the maximum permitted
by coreutils' commit hook, and wasn't readable besides, so here's a
more readable version:
lhs=$(printf '%s\\n' \
'if test "$VERBOSE" = yes; then' \
' set -x' \
' touch --version' \
'fi' \
'' \
'. $srcdir/test-lib.sh' \
| sed 's/\$/\\\$/g;s/touch/(\\w+)/')
rhs=$(printf '%s\\n' \
'. "${srcdir=.}/init.sh"; path_prepend_ ../src' \
'test "$VERBOSE" = yes && FIXME --version' \
| sed 's/\$/\\\$/g;s/FIXME/\$1/')
git grep -l srcdir/test-lib.sh|xargs perl -p0i -e "s,$lhs,$rhs,"
* src/tail.c (tail_forever_inotify): Handle the case where
tail --follow=name with inotify, is not able to add a watch on
a specified directory. This may happen due to inotify resource
limits or if the directory is currently missing or inaccessible.
In all these cases, revert to polling which will try to reopen
the file later. Note inotify returns ENOSPC when it runs out
of resources, and instead we report a particular error message,
lest users think one of their file systems is full.
(main): Document another caveat with using inotify, where we
currently don't recheck directories recreated after the
initial watch is setup.
* tests/tail-2/F-vs-rename: Fix the endless loop triggered by
the above issue.
* tests/tail-2/inotify-hash-abuse: Likewise.
* tests/tail-2/wait: Don't fail in the resource exhaustion case.
* tests/tail-2/F-vs-missing: A new test for this failure mode
which was until now just triggered on older buggy linux kernels
which returned ENOSPC constantly from inotify_add_watch().
* NEWS: Mention the fix.
* tests/init.cfg: Introduce a retry_delay_() function to
repeatedly call a test function that requires a delay.
This delay can now be shorter for the common case on fast
systems, but will double until a configurable limit it reached
before failing on slower systems.
* tests/dd/reblock: Use retry_delay_.
* tests/misc/cat-buf: Likewise.
* tests/misc/stdbuf: Likewise.
* tests/tail-2/F-vs-rename: Likewise.
* tests/tail-2/flush-initial: Likewise.
* tests/tail-2/tail-n0f: Likewise.
* tests/tail-2/wait: Likewise.
* test/dd/misc: Comment that delay is needed to trigger failure.
* tests/tail-2/inotify-rotate: Switch to new init.sh-based framework.
(grep_timeout): New function.
Use it in place of open-coded loops that might infloop.
This was prompted by my encountering an inexplicable, and so far
unreproducible, infloop in the code that was waiting for "b" to
appear in "out".
* tests/tail-2/inotify-hash-abuse2: Explicitly kill the process
by using cleanup_() rather than using a timeout which may trigger
a failure on very slow systems (< 20 iterations of the loop per second).
Prompted by the continuous integration build failure at:
http://hydra.nixos.org/build/277485
* tests/misc/timeout: Set all expected timeouts to 1s and all
unexpected timeouts to 10s. In this way, tests normally proceed
quickly but may delay up to 10s before reporting failures.
* tests/ls/infloop: Likewise.
* tests/tail-2/pid: Likewise.
* tests/tail-2/pipe-f: Likewise.
* tests/tail-2/wait: Likewise.
* tests/dd/skip-seek-past-dev: Likewise.
