lkml.org 
[lkml]   [2011]   [Jun]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] xfstests 255: add a seek_data/seek_hole tester
On 06/29/2011 02:53 AM, Dave Chinner wrote:
> On Tue, Jun 28, 2011 at 11:33:19AM -0400, Josef Bacik wrote:
>> This is a test to make sure seek_data/seek_hole is acting like it does on
>> Solaris. It will check to see if the fs supports finding a hole or not and will
>> adjust as necessary.
>
> So I just looked at this with an eye to validating an XFS
> implementation, and I came up with this list of stuff that the test
> does not cover that I'd need to test in some way:
>
> - files with clean unwritten extents. Are they a hole or
> data? What's SEEK_DATA supposed to return on layout like
> hole-unwritten-data? i.e. needs to add fallocate to the
> picture...
>
> - files with dirty unwritten extents (i.e. dirty in memory,
> not on disk). They are most definitely data, and most
> filesystems will need a separate lookup path to detect
> dirty unwritten ranges because the state is kept
> separately (page cache vs extent cache). Plenty of scope
> for filesystem specific bugs here so needs a roubust test.
>
> - cold cache behaviour - all dirty data ranges the test
> creates are hot in cache and not even forced to disk, so
> it is not testing the no-page-cache-over-the-data-range
> case. i.e. it tests delalloc state tracking but not
> data-extent-already exists lookups during a seek.
>
> - assumes that allocation size is the block size and that
> holes follows block size alignment. We already know that
> ext4 does not follow that rule when doing small sparse
> writes close together in a file, and XFS is also known to
> fill holes when doing sparse writes past EOF.
>
> - only tests single block data extents ѕo doesn't cover
> corner cases like skipping over multiple fragmented data
> extents to the next hole.
>

Yeah I intentionally left out preallocated stuff because these are going
to be implementation specific, so I was going to leave that for a later
exercise when people actually start doing proper implementations.

> Some more comments in line....
>
>> +_cleanup()
>> +{
>> + rm -f $tmp.*
>> +}
>> +
>> +trap "_cleanup ; exit \$status" 0 1 2 3 15
>> +
>> +# get standard environment, filters and checks
>> +. ./common.rc
>> +. ./common.filter
>> +
>> +# real QA test starts here
>> +_supported_fs generic
>> +_supported_os Linux
>> +
>> +testfile=$TEST_DIR/seek_test.$$
>> +logfile=$TEST_DIR/seek_test.$$.log
>
> The log file is usually named $seq.full, and doesn't get placed in
> the filesystem being tested. It gets saved in the xfstests directory
> along side $seq.out.bad for analysis whenteh test fails...
>

I only want it to see if SEEK_HOLE fails so I can say it didn't run. I
followed the same example as the fiemap test that Eric wrote.

>> +[ -x $here/src/seek-tester ] || _notrun "seek-tester not built"
>> +
>> +_cleanup()
>> +{
>> + rm -f $testfile
>> + rm -f $logfile
>> +}
>> +trap "_cleanup; exit \$status" 0 1 2 3 15
>> +
>> +echo "Silence is golden"
>> +$here/src/seek-tester -q $testfile 2>&1 | tee -a $logfile
>
> Personally I'd prefer the test to be a bit noisy about what it is
> running, especially when there are so many subtests the single
> invocation is running. It makes no difference to the run time ofthe
> test, or the output when something fails, but it at least allows you
> to run the test manually and see what it is doing easily...
>

Right, the problem with this test is it will run differently depending
on the implementation. I agree, I really like the noisy output tests,
but unfortunately if I run this test on ext4 where it currently treats
the entire file as data, and then run it on btrfs where it is smarter
and actually recognizes holes, we end up with two different outputs that
are both correct.

>> +
>> +if grep -q "SEEK_HOLE is not supported" $logfile; then
>> + _notrun "SEEK_HOLE/SEEK_DATA not supported by this kernel"
>> +fi
>> +
>> +rm -f $logfile
>> +rm -f $testfile
>> +
>> +status=0 ; exit
>> diff --git a/255.out b/255.out
>> new file mode 100644
>> index 0000000..7eefb82
>> --- /dev/null
>> +++ b/255.out
>> @@ -0,0 +1,2 @@
>> +QA output created by 255
>> +Silence is golden
>> diff --git a/group b/group
>> index 1f86075..c045e70 100644
>> --- a/group
>> +++ b/group
>> @@ -368,3 +368,4 @@ deprecated
>> 252 auto quick prealloc
>> 253 auto quick
>> 254 auto quick
>> +255 auto quick
>
> I'd suggest that rw and prealloc (once unwritten extent
> testing is added) groups should also be defined for this test.
>
> Otherwise, the test code looks ok if a bit over-engineered....
>
>> +struct testrec {
>> + int test_num;
>> + int (*test_func)(int fd, int testnum);
>> + char *test_desc;
>> +};
>> +
>> +struct testrec seek_tests[] = {
>> + { 1, test01, "Test basic support" },
>> + { 2, test02, "Test an empty file" },
>> + { 3, test03, "Test a full file" },
>> + { 4, test04, "Test file hole at beg, data at end" },
>> + { 5, test05, "Test file data at beg, hole at end" },
>> + { 6, test06, "Test file hole data hole data" },
>
> So, to take from the hole punch test matrix, it covers a bunch more
> file state transitions and cases that are just as relevant to
> SEEK_HOLE/SEEK_DATA. Those cases are:
>
> # 1. into a hole
> # 2. into allocated space
> # 3. into unwritten space
> # 4. hole -> data
> # 5. hole -> unwritten
> # 6. data -> hole
> # 7. data -> unwritten
> # 8. unwritten -> hole
> # 9. unwritten -> data
> # 10. hole -> data -> hole
> # 11. data -> hole -> data
> # 12. unwritten -> data -> unwritten
> # 13. data -> unwritten -> data
> # 14. data -> hole @ EOF
> # 15. data -> hole @ 0
> # 16. data -> cache cold ->hole
> # 17. data -> hole in single block file
>
> I thikn we also need to cover most of these same cases, right? And
> SEEK_HOLE/SEEK data also need to explicitly separate the unwritten
> tests into "clean unwritten" and "dirty unwritten" and cover the
> transitions between regions of those states as well, right?
>

Yeah you are right, but again doing preallocated stuff is tricky, but I
can expand it now if that's what we want. Thanks,

Josef
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2011-06-29 15:23    [W:0.143 / U:0.024 seconds]
©2003-2014 Jasper Spaans. Advertise on this site