If the files being allocated during the test do not fit into a single Allocation Group, XFS allocator may disable alignment requirements causing the test to fail even though XFS was working as expected. Fix this by fixing a min AG size, so all files created during the test will fit into a single AG not disabling XFS alignment requirements. Signed-off-by: Carlos Maiolino <cmaiolino@xxxxxxxxxx> --- Hi, I am tagging this patch as a RFC mostly to start a discussion here, regarding this issue found while running generic/223. The generic/223 fails when running it with finobt disabled. Specifically, the last file being fallocated are unaligned. When the finobt is enabled, the allocator does not try to squeeze partial file data into small available extents in AG 0, while it does when finobt is disabled. Here are the bmap of the same file after generic/223 finishes with and without finobt: finobt=0 /mnt/scratch/file-1073745920-falloc: EXT: FILE-OFFSET BLOCK-RANGE AG AG-OFFSET TOTAL FLAGS 0: [0..191]: 320..511 0 (320..511) 192 001011 1: [192..375]: 64..247 0 (64..247) 184 001111 2: [376..1287791]: 678400..1965815 0 (678400..1965815) 1287416 000111 3: [1287792..2097159]: 1966080..2775447 1 (256..809623) 809368 000101 finobt=1 /mnt/scratch/file-1073745920-falloc: EXT: FILE-OFFSET BLOCK-RANGE AG AG-OFFSET TOTAL FLAGS 0: [0..1285831]: 678400..1964231 0 (678400..1964231) 1285832 000111 1: [1285832..2097159]: 1966080..2777407 1 (256..811583) 811328 000101 I still don't know the details about why the allocator takes different decisions depending on finobt being used or not, although I believe it's because the extra space being used in each AG, and the default AG size when running the test, but I'm still reading the code to try to understand this difference. Even though I think there might be room for improvement in the XFS allocator code to avoid this bypass of alignment requirements here, I still think the test should be fixed to avoid forcing the filesystem to drop alignment constraints during file allocation which basically invalidate the test, and that's why I decided to start the discussion with a RFC patch for the test, but sending it to xfs list instead of fstests. Comments? Cheers tests/generic/223 | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/tests/generic/223 b/tests/generic/223 index dfd8c41b..782651e2 100755 --- a/tests/generic/223 +++ b/tests/generic/223 @@ -34,6 +34,13 @@ _require_xfs_io_command "falloc" rm -f $seqres.full +# Ensure we won't trick xfs allocator to disable alignment requirements +if [ "$FSTYP" == "xfs" ]; then + mkfs_opts="-d agsize=2g" +else + mkfs_opts="" +fi + BLOCKSIZE=4096 for SUNIT_K in 8 16 32 64 128; do @@ -41,7 +48,7 @@ for SUNIT_K in 8 16 32 64 128; do let SUNIT_BLOCKS=$SUNIT_BYTES/$BLOCKSIZE echo "=== mkfs with su $SUNIT_BLOCKS blocks x 4 ===" - export MKFS_OPTIONS="" + export MKFS_OPTIONS=$mkfs_opts _scratch_mkfs_geom $SUNIT_BYTES 4 $BLOCKSIZE >> $seqres.full 2>&1 _scratch_mount -- 2.20.1
