On Jun 2, 2016, at 6:01 AM, Lin Feng wrote: > > Descriptions: > ext4 block allocation core stack: > ext4_mb_new_blocks > ext4_mb_normalize_request > ext4_mb_regular_allocator > ext4_mb_find_by_goal > mb_find_extent(e4b, ac->ac_g_ex.fe_start, ac->ac_g_ex.fe_len, &ex); > > The start block searching hint for merging(use EXT4_MB_HINT_TRY_GOAL flag) > set in ext4_mb_normalize_request is stored in ac_f_ex, while in > EXT4_MB_HINT_TRY_GOAL path which falls in ext4_mb_find_by_goal always use > ac_g_ex as a hint and the hint set in ext4_mb_normalize_request is never > use. > > We could hit this bug by writing a sparse file from backward mode and the > file may get fragments even if the physical blocks in the hole is free, > which is expected to be merged into a single extent. This looks reasonable. Do you have any kind of test case that shows the effect of the change (e.g. fragmentation counts per file before/after)? Reviewed-by: Andreas Dilger > Signed-off-by: Lin Feng > --- > fs/ext4/mballoc.c | 8 ++++---- > 1 file changed, 4 insertions(+), 4 deletions(-) > > diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c > index c1ab3ec..e31fc63 100644 > --- a/fs/ext4/mballoc.c > +++ b/fs/ext4/mballoc.c > @@ -3198,15 +3198,15 @@ ext4_mb_normalize_request(struct ext4_allocation_context *ac, > if (ar->pright && (ar->lright == (start + size))) { > /* merge to the right */ > ext4_get_group_no_and_offset(ac->ac_sb, ar->pright - size, > - &ac->ac_f_ex.fe_group, > - &ac->ac_f_ex.fe_start); > + &ac->ac_g_ex.fe_group, > + &ac->ac_g_ex.fe_start); > ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; > } > if (ar->pleft && (ar->lleft + 1 == start)) { > /* merge to the left */ > ext4_get_group_no_and_offset(ac->ac_sb, ar->pleft + 1, > - &ac->ac_f_ex.fe_group, > - &ac->ac_f_ex.fe_start); > + &ac->ac_g_ex.fe_group, > + &ac->ac_g_ex.fe_start); > ac->ac_flags |= EXT4_MB_HINT_TRY_GOAL; > } > > -- > 1.9.3 > > Cheers, Andreas