fstests: add mmap page boundary tests
RFC 1/1 fstests: add mmap page boundary tests
mmap() POSIX compliance says we should zero fill data beyond a file size up to page boundary, and issue a SIGBUS if we go beyond. While fsx helps us test zero-fill sometimes, fsstress also let's us sometimes test for SIGBUS however that is based on a random value and its not likley we always test it. Dedicate a specic test for this to make testing for this specific situation and to easily expand on other corner cases.
---
Does enough to get us to use to test this, however I'm aware of a bit more polishing up to do:
* maybe moving mread to common as generic/574 did it first
* sharing round_up_to_page_boundary() as well
generic/574 is special, it was just testing for correctness of integrity if we muck with mmap() however if you don't have verity stuff available obviously you won't end up testing it.
This generalizes mmap() zero-fill and SIGBUS corner case tests.
I've tested so far only 4k and it works well there. For 16k bs on LBS just the SIGBUS issue exists, I'll test smaller block sizes later like 512, 1k, 2k as well. We'll fix triggering the SIBGUS when LBS is used, we'll address that in the next iteration.
Is this a worthy test as a generic test?
common/filter | 6 ++ tests/generic/740 | 231 ++++++++++++++++++++++++++++++++++++++++++ tests/generic/740.out | 2 + 3 files changed, 239 insertions(+) create mode 100755 tests/generic/740 create mode 100644 tests/generic/740.out
common/filter +6 -0
@@ -194,6 +194,12 @@ _filter_xfs_io_unique() common_line_filter | _filter_xfs_io } +_filter_xfs_io_data_unique() +{ + _filter_xfs_io_offset | sed -e 's| |\n|g' | egrep -v "\.|XX|\*" | \ + sort | uniq | tr -d '\n' +} + _filter_xfs_io_units_modified() { UNIT=$1
tests/generic/740 new file +231 -0
231 lines not on this page · raw message
tests/generic/740.out new file +2 -0
@@ -0,0 +1,2 @@ +QA output created by 740 +Silence is golden
Signature · 1 line
2.43.0
Quoted 103 lines · Luis Chamberlain
mmap()POSIX compliance says we should zero fill data beyond a file size up to page boundary, and issue a SIGBUS if we go beyond. While fsx helps us test zero-fill sometimes, fsstress also let's us sometimes test for SIGBUS however that is based on a random value and its not likley we always test it. Dedicate a specic test for this to make testing for this specific situation and to easily expand on other corner cases.Suggested-by Matthew Wilcox <[email protected]>Signed-off-by Luis Chamberlain <[email protected]>---
Does enough to get us to use to test this, however I'm aware of a bit more polishing up to do:
* maybe moving mread to common as generic/574 did it first
* sharinground_up_to_page_boundary()as wellgeneric/574 is special, it was just testing for correctness of integrity if we muck with
mmap()however if you don't have verity stuff available obviously you won't end up testing it.This generalizes
mmap()zero-fill and SIGBUS corner case tests.I've tested so far only 4k and it works well there. For 16k bs on LBS just the SIGBUS issue exists, I'll test smaller block sizes later like 512, 1k, 2k as well. We'll fix triggering the SIBGUS when LBS is used, we'll address that in the next iteration.
Is this a worthy test as a generic test?
common/filter | 6 ++ tests/generic/740 | 231 ++++++++++++++++++++++++++++++++++++++++++ tests/generic/740.out | 2 + 3 files changed, 239 insertions(+) create mode 100755 tests/generic/740 create mode 100644 tests/generic/740.outcommon/filter +6
-0@@ -194,6 +194,12 @@ _filter_xfs_io_unique() common_line_filter | _filter_xfs_io } +_filter_xfs_io_data_unique() +{ + _filter_xfs_io_offset | sed -e 's| |\n|g' | egrep -v "\.|XX|\*" | \ + sort | uniq | tr -d '\n' +} + _filter_xfs_io_units_modified() { UNIT=$1tests/generic/740 new file +43
-0@@ -0,0 +1,231 @@ +#! /bin/bash +# SPDX-License-Identifier: GPL-2.0 +# Copyright (c) Luis Chamberlain. All Rights Reserved. +# +# FS QA Test 740 +# +# As per POSIX NOTES mmap(2) maps multiples of the system page size, but if the +# data mapped is not multiples of the page size the remaining bytes are zeroed +# out when mapped and modifications to that region are not written to the file. +# On Linux when you write data to such partial page after the end of the +# object, the data stays in the page cache even after the file is closed and +# unmapped and even though the data is never written to the file itself, +# subsequent mappings may see the modified content. If you go *beyond* this +# page, you should get a SIGBUS. This test verifies we zero-fill to page +# boundary and ensures we get a SIGBUS if we write to data beyond the system +# page size even if the block size is greater than the system page size. +. ./common/preamble +. ./common/rc +_begin_fstest auto quick + +# Override the default cleanup function. +_cleanup() +{ + cd / + rm -r -f $tmp.* +} + +# Import common functions. +. ./common/filter + +# real QA test starts here +_supported_fs generic +_require_scratch_nocheck +_require_test + +setup_zeroed_file() +{ + local file_len=$1 + local sparse=$2 + + if $sparse; then + $XFS_IO_PROG -f -c "truncate $file_len" $test_file + else
tests/generic/740 line 53
Quoted 10 lines
tests/generic/740 new file +10
-0+ $XFS_IO_PROG -f -c "falloc 0 $file_len" $test_file + fi +} + +round_up_to_page_boundary() +{ + local n=$1 + local page_size=$(_get_page_size) + + echo $(( (n + page_size - 1) & ~(page_size - 1) ))
Does iomap put a large folio into the pagecache that crosses EOF, or does it back down to base page size?
Quoted 9 lines
tests/generic/740 new file +9
-0+} + +mread() +{ + local file=$1 + local map_len=$2 + local offset=$3 + local length=$4 +
tests/generic/740 line 72
Quoted 10 lines
tests/generic/740 new file +10
-0+ # Some callers expect xfs_io to crash with SIGBUS due to the mread, + # causing the shell to print "Bus error" to stderr. To allow this + # message to be redirected, execute xfs_io in a new shell instance. + # However, for this to work reliably, we also need to prevent the new + # shell instance from optimizing out the fork and directly exec'ing + # xfs_io. The easiest way to do that is to append 'true' to the + # commands, so that xfs_io is no longer the last command the shell sees. + bash -c "trap '' SIGBUS; $XFS_IO_PROG -r $file \ + -c 'mmap -r 0 $map_len' \ + -c 'mread $offset $length'; true"
Please hoist the mread() function with generic/574 to common; your copy is out of date with the original.
Quoted 11 lines
tests/generic/740 new file +11
-0+} + +do_mmap_tests() +{ + local block_size=$1 + local file_len=$2 + local offset=$3 + local len=$4 + local use_sparse_file=${5:-false} + local new_filelen=0 + local map_len=0
tests/generic/740 line 93
Quoted 10 lines
tests/generic/740 new file +10
-0+ local csum=0 + local fs_block_size=$(_get_block_size $SCRATCH_MNT) + + echo -en "\n\n==> Testing blocksize $block_size " >> $seqres.full + echo -en "file_len: $file_len offset: $offset " >> $seqres.full + echo -e "len: $len sparse: $use_sparse_file" >> $seqres.full + + if ((fs_block_size != block_size)); then + echo "Block size created ($block_size) doesn't match _get_block_size on mount ($fs_block_size)" + _fail
_fail "Block size created..."
Quoted 19 lines
tests/generic/740 new file +19
-0+ fi + + rm -rf "${SCRATCH_MNT:?}"/* + + # This let's us also test against sparse files + setup_zeroed_file $file_len $use_sparse_file + + # This will overwrite the old data, the file size is the + # delta between offset and len now. + $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 512 $offset $len" \ + $test_file >> $seqres.full + + sync + new_filelen=$(_get_filesize $test_file) + map_len=$(round_up_to_page_boundary $new_filelen) + csum_orig="$(_md5_checksum $test_file)" + + # A couple of mmap() tests: + #
tests/generic/740 line 122
Quoted 10 lines
tests/generic/740 new file +10
-0+ # We are allowed to mmap() up to the boundary of the page size of a + # data object, but there a few rules to follow we must check for: + # + # a) zero-fill test for the data: POSIX says we should zero fill any + # partial page after the end of the object. Verify zero-fill. + # b) do not write this bogus data to disk: on Linux, if we write data + # to a partially filled page, it will stay in the page cache even + # after the file is closed and unmapped even if it never reaches the + # file. Subsequent mappings *may* see the modified content, but it + # also can get other data. Since the data read after the actual
What does "other data" mean?
tests/generic/740 line 124
Quoted 2 lines
tests/generic/740 new file +2
-0+ # object data can vary we just verify the filesize does not change. + # This is not true for tmpfs.
Er... is this file size change a bug?
Quoted 90 lines
tests/generic/740 new file +90
-0+ if [[ $map_len -gt $new_filelen ]]; then + zero_filled_data_len=$((map_len - new_filelen)) + _scratch_cycle_mount + expected_zero_data="00" + zero_filled_data=$($XFS_IO_PROG -r $test_file \ + -c "mmap -r 0 $map_len" \ + -c "mread -v $new_filelen $zero_filled_data_len" \ + -c "munmap" | \ + _filter_xfs_io_data_unique) + if [[ "$zero_filled_data" != "$expected_zero_data" ]]; then + echo "Expected data: $expected_zero_data" + echo " Actual data: $zero_filled_data" + echo "Zero-fill broken see mmap() requirements" + _fail + fi + + if [[ "$FSTYP" != "tmpfs" ]]; then + _scratch_cycle_mount + $XFS_IO_PROG $test_file \ + -c "mmap -w 0 $map_len" \ + -c "mwrite $new_filelen $zero_filled_data_len" \ + -c "munmap" + sync + csum_post="$(_md5_checksum $test_file)" + if [[ "$csum_orig" != "$csum_post" ]]; then + echo "Expected csum: $csum_orig" + echo " Actual csum: $csum_post" + _fail + fi + + local filelen_test=$(_get_filesize $test_file) + if [[ "$filelen_test" != "$new_filelen" ]]; then + echo "Expected file length: $new_filelen" + echo " Actual file length: $filelen_test" + _fail + fi + fi + fi + + # Now lets ensure we get SIGBUS when we go beyond the page boundary + if [[ "$FSTYP" != "tmpfs" ]]; then + _scratch_cycle_mount + new_filelen=$(_get_filesize $test_file) + map_len=$(round_up_to_page_boundary $new_filelen) + csum_orig="$(_md5_checksum $test_file)" + mread $test_file $map_len 0 $map_len >> $seqres.full 2>$tmp.err + if grep -q 'Bus error' $tmp.err; then + echo "Not expecting SIGBUS when reading up to page boundary" + cat $tmp.err + _fail + fi + + # This should just work + mread $test_file $map_len 0 $map_len >> $seqres.full 2>$tmp.err + if [[ $? -ne 0 ]]; then + _fail + fi + + # If we mmap() on the boundary but try to read beyond it just + # fails, we don't get a SIGBUS + $XFS_IO_PROG -r $test_file \ + -c "mmap -r 0 $map_len" \ + -c "mread 0 $((map_len + 10))" >> $seqres.full 2>$tmp.err + local mread_err=$? + if [[ $mread_err -eq 0 ]]; then + echo "mmap() to page boundary works as expected but reading beyond should fail" + echo "err: $?" + _fail + fi + + # Now let's go beyond the allowed mmap() page boundary + mread $test_file $((map_len + 10)) 0 $((map_len + 10)) >> $seqres.full 2>$tmp.err + if ! grep -q 'Bus error' $tmp.err; then + echo "Expected SIGBUS when mmap() reading beyond page boundary" + _fail + fi + local filelen_test=$(_get_filesize $test_file) + if [[ "$filelen_test" != "$new_filelen" ]]; then + echo "Expected file length: $new_filelen" + echo " Actual file length: $filelen_test" + _fail + fi + fi +} + +test_block_size() +{ + local block_size=$1 + + do_mmap_tests $block_size 512 3 5
tests/generic/740 line 224
Quoted 10 lines
tests/generic/740 new file +10
-0+ do_mmap_tests $block_size 16k 0 $((16384+3)) + do_mmap_tests $block_size 16k $((16384-10)) $((16384+20)) + do_mmap_tests $block_size 64k 0 $((65536+3)) + do_mmap_tests $block_size 4k 4090 30 true +} + +_scratch_mkfs >> $seqres.full 2>&1 || _fail "mkfs failed" +_scratch_mount +test_file=$SCRATCH_MNT/file +block_size=$(_get_block_size "$SCRATCH_MNT")
_get_file_block_size ?
tests/generic/740 line 227
Quoted 3 lines
tests/generic/740 new file +3
-0+test_block_size $block_size +_scratch_unmount +_check_scratch_fs
This work is done for you automatically if you simply exit the test with the scratch fs still mounted.
--D
Quoted 16 lines
tests/generic/740 new file +4
-0+ +echo "Silence is golden" +status=0 +exittests/generic/740.out new file +2
-0@@ -0,0 +1,2 @@ +QA output created by 740 +Silence is goldenSignature · 1 line
2.43.0
Quoted 42 lines · Luis Chamberlain
mmap()POSIX compliance says we should zero fill data beyond a file size up to page boundary, and issue a SIGBUS if we go beyond. While fsx helps us test zero-fill sometimes, fsstress also let's us sometimes test for SIGBUS however that is based on a random value and its not likley we always test it. Dedicate a specic test for this to make testing for this specific situation and to easily expand on other corner cases.Suggested-by Matthew Wilcox <[email protected]>Signed-off-by Luis Chamberlain <[email protected]>---
Does enough to get us to use to test this, however I'm aware of a bit more polishing up to do:
* maybe moving mread to common as generic/574 did it first
* sharinground_up_to_page_boundary()as wellgeneric/574 is special, it was just testing for correctness of integrity if we muck with
mmap()however if you don't have verity stuff available obviously you won't end up testing it.This generalizes
mmap()zero-fill and SIGBUS corner case tests.I've tested so far only 4k and it works well there. For 16k bs on LBS just the SIGBUS issue exists, I'll test smaller block sizes later like 512, 1k, 2k as well. We'll fix triggering the SIBGUS when LBS is used, we'll address that in the next iteration.
Is this a worthy test as a generic test?
common/filter | 6 ++ tests/generic/740 | 231 ++++++++++++++++++++++++++++++++++++++++++ tests/generic/740.out | 2 + 3 files changed, 239 insertions(+) create mode 100755 tests/generic/740 create mode 100644 tests/generic/740.outcommon/filter +0
-0
common/filter _filter_xfs_io_unique() · line 199
Quoted 6 lines
common/filter +3
-0@@ -194,6 +194,12 @@ _filter_xfs_io_unique() common_line_filter | _filter_xfs_io } +_filter_xfs_io_data_unique() +{ + _filter_xfs_io_offset | sed -e 's| |\n|g' | egrep -v "\.|XX|\*" | \
The egrep is deprecated, please use 'grep -E'.
common/filter _filter_xfs_io_unique() · line 200
Quoted 1 line
common/filter +1
-0+ sort | uniq | tr -d '\n'
Isn't "sort | uniq" equal to "sort -u" ?
common/filter _filter_xfs_io_unique() · line 201
Quoted 1 line
common/filter +1
-0+}
Do we need this filter to be a common helper? Will it be used widely? If not, this can be a local function of below single test case.
Quoted 26 lines
common/filter +1
-0+ _filter_xfs_io_units_modified() { UNIT=$1tests/generic/740 new file +16
-0@@ -0,0 +1,231 @@ +#! /bin/bash +# SPDX-License-Identifier: GPL-2.0 +# Copyright (c) Luis Chamberlain. All Rights Reserved. +# +# FS QA Test 740 +# +# As per POSIX NOTES mmap(2) maps multiples of the system page size, but if the +# data mapped is not multiples of the page size the remaining bytes are zeroed +# out when mapped and modifications to that region are not written to the file. +# On Linux when you write data to such partial page after the end of the +# object, the data stays in the page cache even after the file is closed and +# unmapped and even though the data is never written to the file itself, +# subsequent mappings may see the modified content. If you go *beyond* this +# page, you should get a SIGBUS. This test verifies we zero-fill to page +# boundary and ensures we get a SIGBUS if we write to data beyond the system +# page size even if the block size is greater than the system page size.
tests/generic/740 line 26
Quoted 10 lines
tests/generic/740 new file +10
-0+. ./common/preamble +. ./common/rc +_begin_fstest auto quick + +# Override the default cleanup function. +_cleanup() +{ + cd / + rm -r -f $tmp.* +}
This _cleanup is same with the default one, don't need this override.
Quoted 6 lines
tests/generic/740 new file +6
-0+ +# Import common functions. +. ./common/filter + +# real QA test starts here +_supported_fs generic
tests/generic/740 line 42
Quoted 10 lines
tests/generic/740 new file +10
-0+_require_scratch_nocheck +_require_test + +setup_zeroed_file() +{ + local file_len=$1 + local sparse=$2 + + if $sparse; then + $XFS_IO_PROG -f -c "truncate $file_len" $test_file
_require_xfs_io_command "truncate"
tests/generic/740 line 44
Quoted 2 lines
tests/generic/740 new file +2
-0+ else + $XFS_IO_PROG -f -c "falloc 0 $file_len" $test_file
_require_xfs_io_command "falloc"
And with this, better to add this case into "prealloc" group
Quoted 31 lines
tests/generic/740 new file +31
-0+ fi +} + +round_up_to_page_boundary() +{ + local n=$1 + local page_size=$(_get_page_size) + + echo $(( (n + page_size - 1) & ~(page_size - 1) )) +} + +mread() +{ + local file=$1 + local map_len=$2 + local offset=$3 + local length=$4 + + # Some callers expect xfs_io to crash with SIGBUS due to the mread, + # causing the shell to print "Bus error" to stderr. To allow this + # message to be redirected, execute xfs_io in a new shell instance. + # However, for this to work reliably, we also need to prevent the new + # shell instance from optimizing out the fork and directly exec'ing + # xfs_io. The easiest way to do that is to append 'true' to the + # commands, so that xfs_io is no longer the last command the shell sees. + bash -c "trap '' SIGBUS; $XFS_IO_PROG -r $file \ + -c 'mmap -r 0 $map_len' \ + -c 'mread $offset $length'; true" +} + +do_mmap_tests()
tests/generic/740 line 85
Quoted 10 lines
tests/generic/740 new file +10
-0+{ + local block_size=$1 + local file_len=$2 + local offset=$3 + local len=$4 + local use_sparse_file=${5:-false} + local new_filelen=0 + local map_len=0 + local csum=0 + local fs_block_size=$(_get_block_size $SCRATCH_MNT)
Generally the _get_file_block_size is recommended, rather than _get_block_size.
tests/generic/740 line 93
Quoted 8 lines
tests/generic/740 new file +8
-0+ + echo -en "\n\n==> Testing blocksize $block_size " >> $seqres.full + echo -en "file_len: $file_len offset: $offset " >> $seqres.full + echo -e "len: $len sparse: $use_sparse_file" >> $seqres.full + + if ((fs_block_size != block_size)); then + echo "Block size created ($block_size) doesn't match _get_block_size on mount ($fs_block_size)" + _fail
_fail "Block size created ($block_size) doesn't match _get_block_size on mount ($fs_block_size)" ?
Generally we don't use an empty _fail, same as all "_fail" call in this patch.
Quoted 124 lines
tests/generic/740 new file +124
-0+ fi + + rm -rf "${SCRATCH_MNT:?}"/* + + # This let's us also test against sparse files + setup_zeroed_file $file_len $use_sparse_file + + # This will overwrite the old data, the file size is the + # delta between offset and len now. + $XFS_IO_PROG -f -c "pwrite -S 0xaa -b 512 $offset $len" \ + $test_file >> $seqres.full + + sync + new_filelen=$(_get_filesize $test_file) + map_len=$(round_up_to_page_boundary $new_filelen) + csum_orig="$(_md5_checksum $test_file)" + + # A couple of mmap() tests: + # + # We are allowed to mmap() up to the boundary of the page size of a + # data object, but there a few rules to follow we must check for: + # + # a) zero-fill test for the data: POSIX says we should zero fill any + # partial page after the end of the object. Verify zero-fill. + # b) do not write this bogus data to disk: on Linux, if we write data + # to a partially filled page, it will stay in the page cache even + # after the file is closed and unmapped even if it never reaches the + # file. Subsequent mappings *may* see the modified content, but it + # also can get other data. Since the data read after the actual + # object data can vary we just verify the filesize does not change. + # This is not true for tmpfs. + if [[ $map_len -gt $new_filelen ]]; then + zero_filled_data_len=$((map_len - new_filelen)) + _scratch_cycle_mount + expected_zero_data="00" + zero_filled_data=$($XFS_IO_PROG -r $test_file \ + -c "mmap -r 0 $map_len" \ + -c "mread -v $new_filelen $zero_filled_data_len" \ + -c "munmap" | \ + _filter_xfs_io_data_unique) + if [[ "$zero_filled_data" != "$expected_zero_data" ]]; then + echo "Expected data: $expected_zero_data" + echo " Actual data: $zero_filled_data" + echo "Zero-fill broken see mmap() requirements" + _fail + fi + + if [[ "$FSTYP" != "tmpfs" ]]; then + _scratch_cycle_mount + $XFS_IO_PROG $test_file \ + -c "mmap -w 0 $map_len" \ + -c "mwrite $new_filelen $zero_filled_data_len" \ + -c "munmap" + sync + csum_post="$(_md5_checksum $test_file)" + if [[ "$csum_orig" != "$csum_post" ]]; then + echo "Expected csum: $csum_orig" + echo " Actual csum: $csum_post" + _fail + fi + + local filelen_test=$(_get_filesize $test_file) + if [[ "$filelen_test" != "$new_filelen" ]]; then + echo "Expected file length: $new_filelen" + echo " Actual file length: $filelen_test" + _fail + fi + fi + fi + + # Now lets ensure we get SIGBUS when we go beyond the page boundary + if [[ "$FSTYP" != "tmpfs" ]]; then + _scratch_cycle_mount + new_filelen=$(_get_filesize $test_file) + map_len=$(round_up_to_page_boundary $new_filelen) + csum_orig="$(_md5_checksum $test_file)" + mread $test_file $map_len 0 $map_len >> $seqres.full 2>$tmp.err + if grep -q 'Bus error' $tmp.err; then + echo "Not expecting SIGBUS when reading up to page boundary" + cat $tmp.err + _fail + fi + + # This should just work + mread $test_file $map_len 0 $map_len >> $seqres.full 2>$tmp.err + if [[ $? -ne 0 ]]; then + _fail + fi + + # If we mmap() on the boundary but try to read beyond it just + # fails, we don't get a SIGBUS + $XFS_IO_PROG -r $test_file \ + -c "mmap -r 0 $map_len" \ + -c "mread 0 $((map_len + 10))" >> $seqres.full 2>$tmp.err + local mread_err=$? + if [[ $mread_err -eq 0 ]]; then + echo "mmap() to page boundary works as expected but reading beyond should fail" + echo "err: $?" + _fail + fi + + # Now let's go beyond the allowed mmap() page boundary + mread $test_file $((map_len + 10)) 0 $((map_len + 10)) >> $seqres.full 2>$tmp.err + if ! grep -q 'Bus error' $tmp.err; then + echo "Expected SIGBUS when mmap() reading beyond page boundary" + _fail + fi + local filelen_test=$(_get_filesize $test_file) + if [[ "$filelen_test" != "$new_filelen" ]]; then + echo "Expected file length: $new_filelen" + echo " Actual file length: $filelen_test" + _fail + fi + fi +} + +test_block_size() +{ + local block_size=$1 + + do_mmap_tests $block_size 512 3 5 + do_mmap_tests $block_size 16k 0 $((16384+3)) + do_mmap_tests $block_size 16k $((16384-10)) $((16384+20)) + do_mmap_tests $block_size 64k 0 $((65536+3))
tests/generic/740 line 227
Quoted 10 lines
tests/generic/740 new file +10
-0+ do_mmap_tests $block_size 4k 4090 30 true +} + +_scratch_mkfs >> $seqres.full 2>&1 || _fail "mkfs failed" +_scratch_mount +test_file=$SCRATCH_MNT/file +block_size=$(_get_block_size "$SCRATCH_MNT") +test_block_size $block_size +_scratch_unmount +_check_scratch_fs
The _scratch_unmount && _check_scratch_fs aren't needed, they'll be done after this test done.
Quoted 16 lines
tests/generic/740 new file +4
-0+ +echo "Silence is golden" +status=0 +exittests/generic/740.out new file +2
-0@@ -0,0 +1,2 @@ +QA output created by 740 +Silence is goldenSignature · 1 line
2.43.0
Quoted 13 lines
Quoted 13 lines
tests/generic/740 new file +13
-0+setup_zeroed_file() +{ + local file_len=$1 + local sparse=$2 + + if $sparse; then + $XFS_IO_PROG -f -c "truncate $file_len" $test_file + else + $XFS_IO_PROG -f -c "falloc 0 $file_len" $test_file + fi +} + +round_up_to_page_boundary()
Quoted 8 lines
tests/generic/740 line 53
Quoted 5 lines
tests/generic/740 new file +5
-0+{ + local n=$1 + local page_size=$(_get_page_size) + + echo $(( (n + page_size - 1) & ~(page_size - 1) ))Does iomap put a large folio into the pagecache that crosses EOF, or does it back down to base page size?
No, we back it down to the base page size if the higher order folio crosses EOF. But this changes when we have the LBS support as we need to guarantee the base folio order that will be based on the filesystem BS.
Quoted 28 lines · Luis Chamberlain
mmap()POSIX compliance says we should zero fill data beyond a file size up to page boundary, and issue a SIGBUS if we go beyond. While fsx helps us test zero-fill sometimes, fsstress also let's us sometimes test for SIGBUS however that is based on a random value and its not likley we always test it. Dedicate a specic test for this to make testing for this specific situation and to easily expand on other corner cases.Suggested-by Matthew Wilcox <[email protected]>Signed-off-by Luis Chamberlain <[email protected]>---
Does enough to get us to use to test this, however I'm aware of a bit more polishing up to do:
* maybe moving mread to common as generic/574 did it first
* sharinground_up_to_page_boundary()as wellgeneric/574 is special, it was just testing for correctness of integrity if we muck with
mmap()however if you don't have verity stuff available obviously you won't end up testing it.This generalizes
mmap()zero-fill and SIGBUS corner case tests.I've tested so far only 4k and it works well there. For 16k bs on LBS just the SIGBUS issue exists, I'll test smaller block sizes later like 512, 1k, 2k as well. We'll fix triggering the SIBGUS when LBS is used, we'll address that in the next iteration.
Quoted 9 lines
Is this a worthy test as a generic test?
common/filter | 6 ++ tests/generic/740 | 231 ++++++++++++++++++++++++++++++++++++++++++ tests/generic/740.out | 2 + 3 files changed, 239 insertions(+) create mode 100755 tests/generic/740 create mode 100644 tests/generic/740.out
We could also extend the test to include mmap writes.
I took the lazy approach here but I think we can make the code look prettier by adding some more functions that can do both mmap read and writes.
tests/generic/740 +42 -0
@@ -53,6 +53,25 @@ round_up_to_page_boundary() echo $(( (n + page_size - 1) & ~(page_size - 1) )) } +mwrite() +{ + local file=$1 + local map_len=$2 + local offset=$3 + local length=$4 + + # Some callers expect xfs_io to crash with SIGBUS due to the mread, + # causing the shell to print "Bus error" to stderr. To allow this + # message to be redirected, execute xfs_io in a new shell instance. + # However, for this to work reliably, we also need to prevent the new + # shell instance from optimizing out the fork and directly exec'ing + # xfs_io. The easiest way to do that is to append 'true' to the + # commands, so that xfs_io is no longer the last command the shell sees. + bash -c "trap '' SIGBUS; $XFS_IO_PROG $file \ + -c 'mmap -w 0 $map_len' \ + -c 'mwrite $offset $length'; true" +} + mread() { local file=$1
@@ -180,6 +199,12 @@ do_mmap_tests() _fail fi + # This should just work + mwrite $test_file $map_len 0 $map_len >> $seqres.full 2>$tmp.err + if [[ $? -ne 0 ]]; then + _fail + fi + # If we mmap() on the boundary but try to read beyond it just # fails, we don't get a SIGBUS $XFS_IO_PROG -r $test_file \
@@ -192,12 +217,29 @@ do_mmap_tests() _fail fi + $XFS_IO_PROG -w $test_file \ + -c "mmap -w 0 $map_len" \ + -c "mwrite 0 $((map_len + 10))" >> $seqres.full 2>$tmp.err + local mwrite_err=$? + if [[ $mwrite_err -eq 0 ]]; then + echo "mmap() to page boundary works as expected but writing beyond should fail" + echo "err: $?" + _fail + fi + # Now let's go beyond the allowed mmap() page boundary mread $test_file $((map_len + 10)) 0 $((map_len + 10)) >> $seqres.full 2>$tmp.err if ! grep -q 'Bus error' $tmp.err; then echo "Expected SIGBUS when mmap() reading beyond page boundary" _fail fi + + mwrite $test_file $((map_len + 10)) 0 $((map_len + 10)) >> $seqres.full 2>$tmp.err + if ! grep -q 'Bus error' $tmp.err; then + echo "Expected SIGBUS when mmap() writing beyond page boundary" + _fail + fi + local filelen_test=$(_get_filesize $test_file) if [[ "$filelen_test" != "$new_filelen" ]]; then echo "Expected file length: $new_filelen"
--
Pankaj
Quoted 9 lines · Darrick J. Wong
tests/generic/740 line 53
Quoted 6 lines · Luis Chamberlain
tests/generic/740 new file +6
-0+round_up_to_page_boundary() +{ + local n=$1 + local page_size=$(_get_page_size) + + echo $(( (n + page_size - 1) & ~(page_size - 1) ))Does iomap put a large folio into the pagecache that crosses EOF
When minorder is used care had to be taken to ensure the EOF is properly respected. Refer to the patch titled "filemap: cap PTE range to be created to allowed zero fill in folio_map_range()", with that, we fix that case to respect the EOF. But since minorder is used the folio is there.
Quoted 12 lines
Quoted 12 lines
tests/generic/740 new file +12
-0+mread() +{ + local file=$1 + local map_len=$2 + local offset=$3 + local length=$4 + + # Some callers expect xfs_io to crash with SIGBUS due to the mread, + # causing the shell to print "Bus error" to stderr. To allow this + # message to be redirected, execute xfs_io in a new shell instance. + # However, for this to work reliably, we also need to prevent the new + # shell instance from optimizing out the fork and directly exec'ing
Quoted 8 lines
tests/generic/740 line 72
Quoted 5 lines
tests/generic/740 new file +5
-0+ # xfs_io. The easiest way to do that is to append 'true' to the + # commands, so that xfs_io is no longer the last command the shell sees. + bash -c "trap '' SIGBUS; $XFS_IO_PROG -r $file \ + -c 'mmap -r 0 $map_len' \ + -c 'mread $offset $length'; true"Please hoist the
mread()function with generic/574 to common; your copy is out of date with the original.
Sure thing! Shall we add _mwrite to common/rc too while at it or do we wait to get a user for that?
Quoted 6 lines
Quoted 6 lines
tests/generic/740 new file +6
-0+ # A couple of mmap() tests: + # + # We are allowed to mmap() up to the boundary of the page size of a + # data object, but there a few rules to follow we must check for: + # + # a) zero-fill test for the data: POSIX says we should zero fill any
Quoted 8 lines
tests/generic/740 line 122
Quoted 6 lines
tests/generic/740 new file +6
-0+ # partial page after the end of the object. Verify zero-fill. + # b) do not write this bogus data to disk: on Linux, if we write data + # to a partially filled page, it will stay in the page cache even + # after the file is closed and unmapped even if it never reaches the + # file. Subsequent mappings *may* see the modified content, but it + # also can get other data. Since the data read after the actualWhat does "other data" mean?
That depends on the filesystem implementation, it just means we don't provide a consistent behaviour or enforce a strategy for all filesystems.
Quoted 4 lines
tests/generic/740 line 124
Quoted 2 lines
tests/generic/740 new file +2
-0+ # object data can vary we just verify the filesize does not change. + # This is not true for tmpfs.Er... is this file size change a bug?
There is no filesize bug, the comment about tmpfs always ensuring seeing the actual data since, well, there its kind of write-through. Since we share the same filemap_map_pages() I'd expect the rest should behave the same with tmpfs, but since I didn't test that the test skips it for now.
We'll test it, with all the patch "filemap: cap PTE range to be created to allowed zero fill in folio_map_range()" on tmpfs, and see if we can just enable this test there too. Might as well as we're driving by and sprinkling large folios there too.
Luis
Quoted 15 lines · Zorro Lang
common/filter _filter_xfs_io_unique() · line 199
Quoted 11 lines · Luis Chamberlain
common/filter +3
-0@@ -194,6 +194,12 @@ _filter_xfs_io_unique() common_line_filter | _filter_xfs_io } +_filter_xfs_io_data_unique() +{ + _filter_xfs_io_offset | sed -e 's| |\n|g' | egrep -v "\.|XX|\*" | \The egrep is deprecated, please use 'grep -E'.
Quoted 3 lines
common/filter _filter_xfs_io_unique() · line 200
Quoted 1 line
common/filter +1
-0+ sort | uniq | tr -d '\n'Isn't "sort | uniq" equal to "sort -u" ?
I'll try both suggestions.
Quoted 2 lines
Do we need this filter to be a common helper? Will it be used widely? If not, this can be a local function of below single test case.
I suspect it will be but its OK I can stuff it alone, same as for _mwrite() too then.
Luis
Quoted 11 lines · Luis Chamberlain
Quoted 4 lines · Darrick J. Wong
tests/generic/740 line 124
Quoted 1 line · Luis Chamberlain
tests/generic/740 new file +1
-0+ # This is not true for tmpfs.Er... is this file size change a bug?
There is no filesize bug, the comment about tmpfs always ensuring seeing the actual data since, well, there its kind of write-through. Since we share the same
filemap_map_pages()I'd expect the rest should behave the same with tmpfs, but since I didn't test that the test skips it for now.
Quoted 4 lines
We'll test it, with all the patch "filemap: cap PTE range to be created to allowed zero fill in
folio_map_range()" on tmpfs, and see if we can just enable this test there too. Might as well as we're driving by and sprinkling large folios there too.
Turns out our patch "filemap: cap PTE range to be created to allowed zero fill in folio_map_range()" ends up fixing this on tmpfs for huge pages. So consider this now tested on tmpfs as well, I'll adjust the test for that.
Luis