| Message ID | 20190603064009.9891-1-wqu@suse.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v2] fstests: generic/260: Make it handle btrfs more gracefully | expand |
On 3.06.19 г. 9:40 ч., Qu Wenruo wrote: > If a filesystem doesn't map its logical address space (normally the > bytenr/blocknr returned by fiemap) directly to its devices(s), the > following assumptions used in the test case is no longer true: > - trim range start beyond the end of fs should fail > - trim range start beyond the end of fs with len set should fail > > Under the following example, even with just one device, btrfs can still > trim the fs correctly while breaking above assumption: > > 0 1G 1.25G > |---------------|///////////////|-----------------| <- btrfs logical > | address space > ------------ mapped as SINGLE > | > 0 V 256M > |///////////////| <- device address space > > Thus trim range start=1G len=256M will cause btrfs to trim the 256M > block group, thus return correct result. > > Furthermore, there is no cleared defined behavior for whether a fs should > trim the unmapped space. (only for indirectly mapped fs) > > Btrfs currently will always trim the unmapped space, but the behavior > can change as large trim can be very expensive. > > Despite the change to skip certain tests for btrfs, still run the > following tests for btrfs: > - trim start=U64_MAX with lenght set > This will expose a bug that btrfs doesn't check overflow of the range. > This bug will be fixed soon. > > - trim beyond the end of the fs > This will expose a bug where btrfs could send trim command beyond the > end of its device. > This bug is a regression, can be fixed by reverting c2d1b3aae336 ("btrfs: > Honour FITRIM range constraints during free space trim") > > With proper fixes for btrfs, this test case should pass on btrfs, ext4, > xfs. > > Signed-off-by: Qu Wenruo <wqu@suse.com> > --- > changelog: > v2: > - Return 0/1 instead of echo "1"/"0" for _is_fs_directly_mapped > Although it may be a little confusing, but make > "if _is_fs_directly_mapped; then" much cleaner. > - Comment change. > --- Nope, the output is rather unhelpful. Current misc-next of btrfs fails and the output is: [+] Start = 2^64-1 and len is set (should fail) [+] Trim an empty fs 13554941952 trimed [+] Try to trim beyond the end of the fs [+] Try to trim the fs with large enough len 15727198208 trimed generic/260 [failed, exit status 1] There is no 260.out file which is supposed to contain some of the error strings which in turn makes the test tedious to debug... <snip>
On 2019/6/5 下午7:16, Nikolay Borisov wrote: > > > On 3.06.19 г. 9:40 ч., Qu Wenruo wrote: >> If a filesystem doesn't map its logical address space (normally the >> bytenr/blocknr returned by fiemap) directly to its devices(s), the >> following assumptions used in the test case is no longer true: >> - trim range start beyond the end of fs should fail >> - trim range start beyond the end of fs with len set should fail >> >> Under the following example, even with just one device, btrfs can still >> trim the fs correctly while breaking above assumption: >> >> 0 1G 1.25G >> |---------------|///////////////|-----------------| <- btrfs logical >> | address space >> ------------ mapped as SINGLE >> | >> 0 V 256M >> |///////////////| <- device address space >> >> Thus trim range start=1G len=256M will cause btrfs to trim the 256M >> block group, thus return correct result. >> >> Furthermore, there is no cleared defined behavior for whether a fs should >> trim the unmapped space. (only for indirectly mapped fs) >> >> Btrfs currently will always trim the unmapped space, but the behavior >> can change as large trim can be very expensive. >> >> Despite the change to skip certain tests for btrfs, still run the >> following tests for btrfs: >> - trim start=U64_MAX with lenght set >> This will expose a bug that btrfs doesn't check overflow of the range. >> This bug will be fixed soon. >> >> - trim beyond the end of the fs >> This will expose a bug where btrfs could send trim command beyond the >> end of its device. >> This bug is a regression, can be fixed by reverting c2d1b3aae336 ("btrfs: >> Honour FITRIM range constraints during free space trim") >> >> With proper fixes for btrfs, this test case should pass on btrfs, ext4, >> xfs. >> >> Signed-off-by: Qu Wenruo <wqu@suse.com> >> --- >> changelog: >> v2: >> - Return 0/1 instead of echo "1"/"0" for _is_fs_directly_mapped >> Although it may be a little confusing, but make >> "if _is_fs_directly_mapped; then" much cleaner. >> - Comment change. >> --- > > Nope, the output is rather unhelpful. Current misc-next of btrfs fails > and the output is: This is not the output. This is seqres.full. For output, using 5.2-rc2, btrfs would fail like: [+] Optional trim range test (fs dependent) [+] Default length (should succeed) [+] Default length with start set (should succeed) [+] Length beyond the end of fs (should succeed) [+] Length beyond the end of fs with start set (should succeed) Unexpected error happened during trim Test done Which is already good enough to show what's wrong. > > [+] Start = 2^64-1 and len is set (should fail) > > [+] Trim an empty fs > > 13554941952 trimed > > [+] Try to trim beyond the end of the fs > > [+] Try to trim the fs with large enough len > > 15727198208 trimed For this full, try it on 5.2-rc2, then you would understand why it's here: [+] Start = 2^64-1 and len is set (should fail) << It doesn't fail [+] Trim an empty fs 0 trimed << It trimmed 0 bytes, isn't it already a problem? [+] Try to trim beyond the end of the fs fstrim: /mnt/scratch: FITRIM ioctl failed: Input/output error << Beyond device end bug [+] Try to trim the fs with large enough len 5367267328 trimed << The only good result here. > > generic/260 [failed, exit status 1] > > > There is no 260.out file which is supposed to contain some of the error > strings which in turn makes the test tedious to debug... I'm afraid you're checking the wrong file. Thanks, Qu > > > <snip> >
On 5.06.19 г. 14:53 ч., Qu Wenruo wrote: > > > On 2019/6/5 下午7:16, Nikolay Borisov wrote: >> >> >> On 3.06.19 г. 9:40 ч., Qu Wenruo wrote: >>> If a filesystem doesn't map its logical address space (normally the >>> bytenr/blocknr returned by fiemap) directly to its devices(s), the >>> following assumptions used in the test case is no longer true: >>> - trim range start beyond the end of fs should fail >>> - trim range start beyond the end of fs with len set should fail >>> >>> Under the following example, even with just one device, btrfs can still >>> trim the fs correctly while breaking above assumption: >>> >>> 0 1G 1.25G >>> |---------------|///////////////|-----------------| <- btrfs logical >>> | address space >>> ------------ mapped as SINGLE >>> | >>> 0 V 256M >>> |///////////////| <- device address space >>> >>> Thus trim range start=1G len=256M will cause btrfs to trim the 256M >>> block group, thus return correct result. >>> >>> Furthermore, there is no cleared defined behavior for whether a fs should >>> trim the unmapped space. (only for indirectly mapped fs) >>> >>> Btrfs currently will always trim the unmapped space, but the behavior >>> can change as large trim can be very expensive. >>> >>> Despite the change to skip certain tests for btrfs, still run the >>> following tests for btrfs: >>> - trim start=U64_MAX with lenght set >>> This will expose a bug that btrfs doesn't check overflow of the range. >>> This bug will be fixed soon. >>> >>> - trim beyond the end of the fs >>> This will expose a bug where btrfs could send trim command beyond the >>> end of its device. >>> This bug is a regression, can be fixed by reverting c2d1b3aae336 ("btrfs: >>> Honour FITRIM range constraints during free space trim") >>> >>> With proper fixes for btrfs, this test case should pass on btrfs, ext4, >>> xfs. >>> >>> Signed-off-by: Qu Wenruo <wqu@suse.com> >>> --- >>> changelog: >>> v2: >>> - Return 0/1 instead of echo "1"/"0" for _is_fs_directly_mapped >>> Although it may be a little confusing, but make >>> "if _is_fs_directly_mapped; then" much cleaner. >>> - Comment change. >>> --- >> >> Nope, the output is rather unhelpful. Current misc-next of btrfs fails >> and the output is: > > This is not the output. This is seqres.full. > > For output, using 5.2-rc2, btrfs would fail like: > [+] Optional trim range test (fs dependent) > [+] Default length (should succeed) > [+] Default length with start set (should succeed) > [+] Length beyond the end of fs (should succeed) > [+] Length beyond the end of fs with start set (should succeed) > Unexpected error happened during trim > Test done > > Which is already good enough to show what's wrong. > >> >> [+] Start = 2^64-1 and len is set (should fail) >> >> [+] Trim an empty fs >> >> 13554941952 trimed >> >> [+] Try to trim beyond the end of the fs >> >> [+] Try to trim the fs with large enough len >> >> 15727198208 trimed > > For this full, try it on 5.2-rc2, then you would understand why it's here: > > [+] Start = 2^64-1 and len is set (should fail) << It doesn't fail > [+] Trim an empty fs > 0 trimed << It trimmed 0 bytes, isn't it already a problem? > [+] Try to trim beyond the end of the fs > fstrim: /mnt/scratch: FITRIM ioctl failed: Input/output error << Beyond > device end bug > [+] Try to trim the fs with large enough len > 5367267328 trimed << The only good result here. > >> >> generic/260 [failed, exit status 1] >> >> >> There is no 260.out file which is supposed to contain some of the error >> strings which in turn makes the test tedious to debug... > > I'm afraid you're checking the wrong file. I don't have an .out file produced! > > Thanks, > Qu > >> >> >> <snip> >> >
On 2019/6/5 下午7:54, Nikolay Borisov wrote: > > > On 5.06.19 г. 14:53 ч., Qu Wenruo wrote: >> >> >> On 2019/6/5 下午7:16, Nikolay Borisov wrote: >>> >>> >>> On 3.06.19 г. 9:40 ч., Qu Wenruo wrote: >>>> If a filesystem doesn't map its logical address space (normally the >>>> bytenr/blocknr returned by fiemap) directly to its devices(s), the >>>> following assumptions used in the test case is no longer true: >>>> - trim range start beyond the end of fs should fail >>>> - trim range start beyond the end of fs with len set should fail >>>> >>>> Under the following example, even with just one device, btrfs can still >>>> trim the fs correctly while breaking above assumption: >>>> >>>> 0 1G 1.25G >>>> |---------------|///////////////|-----------------| <- btrfs logical >>>> | address space >>>> ------------ mapped as SINGLE >>>> | >>>> 0 V 256M >>>> |///////////////| <- device address space >>>> >>>> Thus trim range start=1G len=256M will cause btrfs to trim the 256M >>>> block group, thus return correct result. >>>> >>>> Furthermore, there is no cleared defined behavior for whether a fs should >>>> trim the unmapped space. (only for indirectly mapped fs) >>>> >>>> Btrfs currently will always trim the unmapped space, but the behavior >>>> can change as large trim can be very expensive. >>>> >>>> Despite the change to skip certain tests for btrfs, still run the >>>> following tests for btrfs: >>>> - trim start=U64_MAX with lenght set >>>> This will expose a bug that btrfs doesn't check overflow of the range. >>>> This bug will be fixed soon. >>>> >>>> - trim beyond the end of the fs >>>> This will expose a bug where btrfs could send trim command beyond the >>>> end of its device. >>>> This bug is a regression, can be fixed by reverting c2d1b3aae336 ("btrfs: >>>> Honour FITRIM range constraints during free space trim") >>>> >>>> With proper fixes for btrfs, this test case should pass on btrfs, ext4, >>>> xfs. >>>> >>>> Signed-off-by: Qu Wenruo <wqu@suse.com> >>>> --- >>>> changelog: >>>> v2: >>>> - Return 0/1 instead of echo "1"/"0" for _is_fs_directly_mapped >>>> Although it may be a little confusing, but make >>>> "if _is_fs_directly_mapped; then" much cleaner. >>>> - Comment change. >>>> --- >>> >>> Nope, the output is rather unhelpful. Current misc-next of btrfs fails >>> and the output is: >> >> This is not the output. This is seqres.full. >> >> For output, using 5.2-rc2, btrfs would fail like: >> [+] Optional trim range test (fs dependent) >> [+] Default length (should succeed) >> [+] Default length with start set (should succeed) >> [+] Length beyond the end of fs (should succeed) >> [+] Length beyond the end of fs with start set (should succeed) >> Unexpected error happened during trim >> Test done >> >> Which is already good enough to show what's wrong. >> >>> >>> [+] Start = 2^64-1 and len is set (should fail) >>> >>> [+] Trim an empty fs >>> >>> 13554941952 trimed >>> >>> [+] Try to trim beyond the end of the fs >>> >>> [+] Try to trim the fs with large enough len >>> >>> 15727198208 trimed >> >> For this full, try it on 5.2-rc2, then you would understand why it's here: >> >> [+] Start = 2^64-1 and len is set (should fail) << It doesn't fail >> [+] Trim an empty fs >> 0 trimed << It trimmed 0 bytes, isn't it already a problem? >> [+] Try to trim beyond the end of the fs >> fstrim: /mnt/scratch: FITRIM ioctl failed: Input/output error << Beyond >> device end bug >> [+] Try to trim the fs with large enough len >> 5367267328 trimed << The only good result here. >> >>> >>> generic/260 [failed, exit status 1] >>> >>> >>> There is no 260.out file which is supposed to contain some of the error >>> strings which in turn makes the test tedious to debug... >> >> I'm afraid you're checking the wrong file. > > I don't have an .out file produced! Then it's because the output matches. Only status is wrong. For latest misc branch, the only test it will fail should be that start = -1 with len set one. Since that subtest is inside the optional fs dependent test, thus it doesn't output into stdio but completely rely on seqres.full. In that case, the seqres.full output should be enough to show the problem. > [+] Start = 2^64-1 and len is set (should fail) Should fail but no error message? Isn't that the reason why it fails? Thanks, Qu > > >> >> Thanks, >> Qu >> >>> >>> >>> <snip> >>> >>
On Mon, Jun 03, 2019 at 02:40:09PM +0800, Qu Wenruo wrote: > If a filesystem doesn't map its logical address space (normally the > bytenr/blocknr returned by fiemap) directly to its devices(s), the > following assumptions used in the test case is no longer true: > - trim range start beyond the end of fs should fail > - trim range start beyond the end of fs with len set should fail All these assumptions seem generally bogus. I don't think we can just run this as a generic test.
diff --git a/common/rc b/common/rc index 17b89d5d..117dcec2 100644 --- a/common/rc +++ b/common/rc @@ -4005,6 +4005,21 @@ _require_fibmap() rm -f $file } +# Check if the logical address (returned by fiemap) can be larger than the +# block device. +# Currently only btrfs has such behavior even for single device usage. +# Return 0 if it's directly mapped, return 1 if not. +# This return value looks confusing but allows us to use +# "if _is_fs_direct_mapped; then" directly +_is_fs_direct_mapped() +{ + if [ "$FSTYP" == "btrfs" ]; then + return 1 + else + return 0 + fi +} + _try_wipe_scratch_devs() { test -x "$WIPEFS_PROG" || return 0 diff --git a/tests/generic/260 b/tests/generic/260 index 9e652dee..77ace622 100755 --- a/tests/generic/260 +++ b/tests/generic/260 @@ -27,40 +27,50 @@ _supported_fs generic _supported_os Linux _require_math +rm -f $seqres.full + _require_scratch _scratch_mkfs >/dev/null 2>&1 _scratch_mount _require_batched_discard $SCRATCH_MNT + fssize=$($DF_PROG -k | grep "$SCRATCH_MNT" | grep "$SCRATCH_DEV" | awk '{print $3}') beyond_eofs=$(_math "$fssize*2048") max_64bit=$(_math "2^64 - 1") -# All these tests should return EINVAL -# since the start is beyond the end of -# the file system - -echo "[+] Start beyond the end of fs (should fail)" -out=$($FSTRIM_PROG -o $beyond_eofs $SCRATCH_MNT 2>&1) -[ $? -eq 0 ] && status=1 -echo $out | _filter_scratch - -echo "[+] Start beyond the end of fs with len set (should fail)" -out=$($FSTRIM_PROG -o $beyond_eofs -l1M $SCRATCH_MNT 2>&1) -[ $? -eq 0 ] && status=1 -echo $out | _filter_scratch - -echo "[+] Start = 2^64-1 (should fail)" -out=$($FSTRIM_PROG -o $max_64bit $SCRATCH_MNT 2>&1) -[ $? -eq 0 ] && status=1 -echo $out | _filter_scratch +# For filesystem with direct mapping, all these tests should return EINVAL +# since the start is beyond the end of the file system +# +# Skip these tests if the filesystem has its own address space mapping, +# as it's implementation dependent. +# E.g btrfs can map its physical address of (devid=1, physical=1M, len=1M) +# to its logical address (logical=1G, len=1M). Making trim beyond device +# boundary to success. + +echo "[+] Optional trim range test (fs dependent)" +if _is_fs_direct_mapped; then + echo "[+] Start beyond the end of fs (should fail)" >> $seqres.full + $FSTRIM_PROG -o $beyond_eofs $SCRATCH_MNT >> $seqres.full 2>&1 + [ $? -eq 0 ] && status=1 + + echo "[+] Start beyond the end of fs with len set (should fail)" >> $seqres.full + $FSTRIM_PROG -o $beyond_eofs -l1M $SCRATCH_MNT >> $seqres.full 2>&1 + [ $? -eq 0 ] && status=1 + + # indirectly mapped fs may use this special value to trim their + # unmapped space, so don't do this for indirectly mapped fs. + echo "[+] Start = 2^64-1 (should fail)" >> $seqres.full + $FSTRIM_PROG -o $max_64bit $SCRATCH_MNT 2>&1 >> $seqres.full 2>&1 + [ $? -eq 0 ] && status=1 +fi -echo "[+] Start = 2^64-1 and len is set (should fail)" -out=$($FSTRIM_PROG -o $max_64bit -l1M $SCRATCH_MNT 2>&1) +# This should fail due to overflow no matter how the fs is implemented +echo "[+] Start = 2^64-1 and len is set (should fail)" >> $seqres.full +$FSTRIM_PROG -o $max_64bit -l1M $SCRATCH_MNT >> $seqres.full 2>&1 [ $? -eq 0 ] && status=1 -echo $out | _filter_scratch _scratch_unmount _scratch_mkfs >/dev/null 2>&1 @@ -86,10 +96,12 @@ _scratch_unmount _scratch_mkfs >/dev/null 2>&1 _scratch_mount +echo "[+] Trim an empty fs" >> $seqres.full # This is a bit fuzzy, but since the file system is fresh # there should be at least (fssize/2) free space to trim. # This is supposed to catch wrong FITRIM argument handling bytes=$($FSTRIM_PROG -v -o10M $SCRATCH_MNT | _filter_fstrim) +echo "$bytes trimed" >> $seqres.full if [ $bytes -gt $(_math "$fssize*1024") ]; then status=1 @@ -97,11 +109,9 @@ if [ $bytes -gt $(_math "$fssize*1024") ]; then "however the file system is $(_math "$fssize*1024") bytes long." fi -# Btrfs is special and this test does not apply to it -# It is because btrfs does not have not-yet-used parts of the device -# mapped and since we got here right after the mkfs, there is not -# enough free extents in the root tree. -if [ $bytes -le $(_math "$fssize*512") ] && [ $FSTYP != "btrfs" ]; then +# Indirect mapped fs can have their own decision on whether to trim +# unmapped blocks. +if [ $bytes -le $(_math "$fssize*512") ] && _is_fs_direct_mapped; then status=1 echo "After the full fs discard $bytes bytes were discarded"\ "however the file system is $(_math "$fssize*1024") bytes long." @@ -141,14 +151,23 @@ esac _scratch_unmount _scratch_mkfs >/dev/null 2>&1 _scratch_mount + +echo "[+] Try to trim beyond the end of the fs" >> $seqres.full # It should fail since $start is beyond the end of file system -$FSTRIM_PROG -o$start -l10M $SCRATCH_MNT &> /dev/null -if [ $? -eq 0 ]; then +$FSTRIM_PROG -o$start -l10M $SCRATCH_MNT >> $seqres.full 2>&1 +ret=$? +if [ $ret -eq 0 ] && _is_fs_direct_mapped; then status=1 echo "It seems that fs logic handling start"\ "argument overflows" fi +# For indirectly mapped fs, it shouldn't fail. +if [ $ret -ne 0 ] && ! _is_fs_direct_mapped; then + status=1 + echo "Unexpected error happened during trim" +fi + _scratch_unmount _scratch_mkfs >/dev/null 2>&1 _scratch_mount @@ -160,8 +179,10 @@ _scratch_mount # It is because btrfs does not have not-yet-used parts of the device # mapped and since we got here right after the mkfs, there is not # enough free extents in the root tree. +echo "[+] Try to trim the fs with large enough len" >> $seqres.full bytes=$($FSTRIM_PROG -v -l$len $SCRATCH_MNT | _filter_fstrim) -if [ $bytes -le $(_math "$fssize*512") ] && [ $FSTYP != "btrfs" ]; then +echo "$bytes trimed" >> $seqres.full +if [ $bytes -le $(_math "$fssize*512") ] && _is_fs_direct_mapped; then status=1 echo "It seems that fs logic handling len argument overflows" fi diff --git a/tests/generic/260.out b/tests/generic/260.out index a16c4f74..f4ee2f72 100644 --- a/tests/generic/260.out +++ b/tests/generic/260.out @@ -1,12 +1,5 @@ QA output created by 260 -[+] Start beyond the end of fs (should fail) -fstrim: SCRATCH_MNT: FITRIM ioctl failed: Invalid argument -[+] Start beyond the end of fs with len set (should fail) -fstrim: SCRATCH_MNT: FITRIM ioctl failed: Invalid argument -[+] Start = 2^64-1 (should fail) -fstrim: SCRATCH_MNT: FITRIM ioctl failed: Invalid argument -[+] Start = 2^64-1 and len is set (should fail) -fstrim: SCRATCH_MNT: FITRIM ioctl failed: Invalid argument +[+] Optional trim range test (fs dependent) [+] Default length (should succeed) [+] Default length with start set (should succeed) [+] Length beyond the end of fs (should succeed)
If a filesystem doesn't map its logical address space (normally the bytenr/blocknr returned by fiemap) directly to its devices(s), the following assumptions used in the test case is no longer true: - trim range start beyond the end of fs should fail - trim range start beyond the end of fs with len set should fail Under the following example, even with just one device, btrfs can still trim the fs correctly while breaking above assumption: 0 1G 1.25G |---------------|///////////////|-----------------| <- btrfs logical | address space ------------ mapped as SINGLE | 0 V 256M |///////////////| <- device address space Thus trim range start=1G len=256M will cause btrfs to trim the 256M block group, thus return correct result. Furthermore, there is no cleared defined behavior for whether a fs should trim the unmapped space. (only for indirectly mapped fs) Btrfs currently will always trim the unmapped space, but the behavior can change as large trim can be very expensive. Despite the change to skip certain tests for btrfs, still run the following tests for btrfs: - trim start=U64_MAX with lenght set This will expose a bug that btrfs doesn't check overflow of the range. This bug will be fixed soon. - trim beyond the end of the fs This will expose a bug where btrfs could send trim command beyond the end of its device. This bug is a regression, can be fixed by reverting c2d1b3aae336 ("btrfs: Honour FITRIM range constraints during free space trim") With proper fixes for btrfs, this test case should pass on btrfs, ext4, xfs. Signed-off-by: Qu Wenruo <wqu@suse.com> --- changelog: v2: - Return 0/1 instead of echo "1"/"0" for _is_fs_directly_mapped Although it may be a little confusing, but make "if _is_fs_directly_mapped; then" much cleaner. - Comment change. --- common/rc | 15 ++++++++ tests/generic/260 | 79 +++++++++++++++++++++++++++---------------- tests/generic/260.out | 9 +---- 3 files changed, 66 insertions(+), 37 deletions(-)