libext2fs: optimize ext2fs_convert_subcluster_bitmap()
PATCH v3 1/5 libext2fs: optimize ext2fs_convert_subcluster_bitmap()
For a bigalloc filesystem, converting the block bitmap from blocks to chunks in ext2fs_convert_subcluster_bitmap() can take a long time when the device is huge, because we test the bitmap bit-by-bit using ext2fs_test_block_bitmap2().
Use ext2fs_find_first_set_block_bitmap2() which is more efficient for mke2fs when the fs is mostly empty.
e2fsck can also benefit from this during pass1 block scanning.
Time taken for "mke2fs -O bigalloc,extent -C 131072 -b 4096" on a 1PB device:
without patch:
real 27m49.457s user 21m36.474s sys 6m9.514s
with patch:
real 6m31.908s
user 0m1.806s
sys 6m29.697s
--- lib/ext2fs/gen_bitmap64.c | 20 +++++++------------- 1 file changed, 7 insertions(+), 13 deletions(-)
lib/ext2fs/gen_bitmap64.c +7 -13
@@ -799,8 +799,7 @@ errcode_t ext2fs_convert_subcluster_bitmap(ext2_filsys fs,ext2fs_generic_bitmap_64 bmap, cmap;ext2fs_block_bitmap gen_bmap = *bitmap, gen_cmap;errcode_t retval; -blk64_t i, b_end, c_end; -int n, ratio; +blk64_t i, next, b_end, c_end; bmap = (ext2fs_generic_bitmap_64 ) gen_bmap;if (fs->cluster_ratio_bits ==ext2fs_get_bitmap_granularity (gen_bmap))
@@ -817,18 +816,13 @@ errcode_t ext2fs_convert_subcluster_bitmap(ext2_filsys fs, bmap->end = bmap->real_end; c_end = cmap->end; cmap->end = cmap->real_end; - n =0 ; - ratio =1 << fs->cluster_ratio_bits;while (i < bmap->real_end) { -if (ext2fs_test_block_bitmap2 (gen_bmap, i)) { -ext2fs_mark_block_bitmap2 (gen_cmap, i); - i += ratio - n; - n =0 ; -continue ; - } - i++; n++; -if (n >= ratio) - n =0 ; + retval =ext2fs_find_first_set_block_bitmap2 (gen_bmap, + i, bmap->real_end, &next); +if (retval) +break ; +ext2fs_mark_block_bitmap2 (gen_cmap, next); + i =EXT2FS_C2B (fs,EXT2FS_B2C (fs, next) +1 ); } bmap->end = b_end; cmap->end = c_end;
Signature · 1 line
2.24.0
PATCH v3 2/5 mke2fs: fix setting bad blocks in the block bitmap
We mark the bad blocks as used on fs->block_map before allocating group tables.
Don't translate the block number to cluster number when doing this, the fs->block_map is still a block-granularity allocation map, it will be coverted later by ext2fs_convert_subcluster_bitmap().
--- misc/mke2fs.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
misc/mke2fs.c +1 -1
@@ -344,7 +344,7 @@ _("Warning: the backup superblock/group descriptors at block %u contain\n"exit (1 ); }while (ext2fs_badblocks_list_iterate (bb_iter, &blk)) -ext2fs_mark_block_bitmap2 (fs->block_map,EXT2FS_B2C (fs, blk)); +ext2fs_mark_block_bitmap2 (fs->block_map, blk);ext2fs_badblocks_list_iterate_end (bb_iter); }
Signature · 1 line
2.24.0
Theodore Y. Ts'o
maintainerRe: [PATCH v3 2/5] mke2fs: fix setting bad blocks in the block bitmap
Quoted 8 lines · Li Dongyang
We mark the bad blocks as used on
fs->block_mapbefore allocating group tables.
Don't translate the block number to cluster number when doing this, thefs->block_mapis still a block-granularity allocation map, it will be coverted later byext2fs_convert_subcluster_bitmap().Signed-off-by Li Dongyang <[email protected]>
Applied on the maint branch, thanks. (The other patches need to go on the e2fsprogs master/next branch, and the last is a kernel patch.)
- Ted
PATCH v3 3/5 ext2fs: rename "s_overhead_blocks" to "s_overhead_clusters"
In reply to Li DongyangRename s_overhead_blocks field from struct ext2_super_block to make it consistent with the kernel counterpart.
--- debugfs/set_fields.c | 2 +- lib/e2p/ls.c | 6 +++--- lib/ext2fs/ext2_fs.h | 2 +- lib/ext2fs/swapfs.c | 2 +- lib/ext2fs/tst_super_size.c | 2 +- 5 files changed, 7 insertions(+), 7 deletions(-)
debugfs/set_fields.c +1 -1
@@ -160,7 +160,7 @@ static struct field_set_info super_fields[] = { { "usr_quota_inum", &set_sb.s_usr_quota_inum, NULL, 4, parse_uint }, { "grp_quota_inum", &set_sb.s_grp_quota_inum, NULL, 4, parse_uint }, { "prj_quota_inum", &set_sb.s_prj_quota_inum, NULL, 4, parse_uint }, - { "overhead_blocks", &set_sb.s_overhead_blocks, NULL, 4, parse_uint }, + { "overhead_clusters", &set_sb.s_overhead_clusters, NULL, 4, parse_uint }, { "backup_bgs", &set_sb.s_backup_bgs[0], NULL, 4, parse_uint, FLAG_ARRAY, 2 }, { "checksum", &set_sb.s_checksum, NULL, 4, parse_uint },
lib/e2p/ls.c +3 -3
@@ -272,9 +272,9 @@ void list_super2(struct ext2_super_block * sb, FILE *f)fprintf (f,"Inode count: %u\n" , sb->s_inodes_count);fprintf (f,"Block count: %llu\n" ,e2p_blocks_count (sb));fprintf (f,"Reserved block count: %llu\n" ,e2p_r_blocks_count (sb)); -if (sb->s_overhead_blocks) -fprintf (f,"Overhead blocks: %u\n" , - sb->s_overhead_blocks); +if (sb->s_overhead_clusters) +fprintf (f,"Overhead clusters: %u\n" , + sb->s_overhead_clusters);fprintf (f,"Free blocks: %llu\n" ,e2p_free_blocks_count (sb));fprintf (f,"Free inodes: %u\n" , sb->s_free_inodes_count);fprintf (f,"First block: %u\n" , sb->s_first_data_block);
lib/ext2fs/ext2_fs.h +1 -1
@@ -742,7 +742,7 @@ struct ext2_super_block {/*200*/ __u8 s_mount_opts[64 ];/*240*/ __u32 s_usr_quota_inum;/* inode number of user quota file */ __u32 s_grp_quota_inum;/* inode number of group quota file */ -__u32 s_overhead_blocks;/* overhead blocks/clusters in fs */ +__u32 s_overhead_clusters;/* overhead blocks/clusters in fs */ /*24c*/ __u32 s_backup_bgs[2 ];/* If sparse_super2 enabled */ /*254*/ __u8 s_encrypt_algos[4 ];/* Encryption algorithms in use */ /*258*/ __u8 s_encrypt_pw_salt[16 ];/* Salt used for string2key algorithm */
lib/ext2fs/swapfs.c +1 -1
@@ -121,7 +121,7 @@ void ext2fs_swap_super(struct ext2_super_block * sb)/* sb->s_mount_opts is __u8 and does not need swabbing */ sb->s_usr_quota_inum =ext2fs_swab32 (sb->s_usr_quota_inum); sb->s_grp_quota_inum =ext2fs_swab32 (sb->s_grp_quota_inum); - sb->s_overhead_blocks =ext2fs_swab32 (sb->s_overhead_blocks); + sb->s_overhead_clusters =ext2fs_swab32 (sb->s_overhead_clusters); sb->s_backup_bgs[0 ] =ext2fs_swab32 (sb->s_backup_bgs[0 ]); sb->s_backup_bgs[1 ] =ext2fs_swab32 (sb->s_backup_bgs[1 ]);/* sb->s_encrypt_algos is __u8 and does not need swabbing */
lib/ext2fs/tst_super_size.c +1 -1
@@ -135,7 +135,7 @@ int main(int argc, char **argv)check_field (s_mount_opts,64 );check_field (s_usr_quota_inum,4 );check_field (s_grp_quota_inum,4 ); -check_field (s_overhead_blocks,4 ); +check_field (s_overhead_clusters,4 );check_field (s_backup_bgs,8 );check_field (s_encrypt_algos,4 );check_field (s_encrypt_pw_salt,16 );
Signature · 1 line
2.24.0
Re: [PATCH v3 3/5] ext2fs: rename "s_overhead_blocks" to "s_overhead_clusters"
Quoted 5 lines · Li Dongyang
Rename
s_overhead_blocksfield from structext2_super_blockto make it consistent with the kernel counterpart.Signed-off-by Li Dongyang <[email protected]>Reviewed-by Andreas Dilger <[email protected]>
Applied, thanks.
- Ted
PATCH v3 4/5 mke2fs: set overhead in super block
In reply to Li DongyangIf overhead is not recorded in the super block, it is caculated during mount in kernel, for bigalloc file systems the it takes O(groups**2) in time.
For a 1PB deivce with 32K cluste size it takes ~12 mins to mount, with most of the time spent on figuring out overhead.
While we can not improve the overhead algorithm in kernel due to the nature of bigalloc, we can work out the overhead during mke2fs and set it in the super block, avoiding calculating it every time when it mounts.
Overhead is s_first_data_block plus internal journal blocks plus the block and inode bitmaps, inode table, super block backups and group descriptor blocks for every group. This patch introduces ext2fs_count_used_clusters(), which calculates the clusters used in the block bitmap for the given range.
When bad blocks are involved, it gets tricky because the blocks counted as overhead and the bad blocks can end up in the same allocation cluster. In this case we will unmark the bad blocks from the block bitmap, convert to cluster bitmap and get the overhead, then mark the bad blocks back in the cluster bitmap.
Reset the overhead to zero when resizing, we can not simplly count the used blocks as overhead like we do when mke2fs. The overhead can be calculated by kernel side during mount.
--- lib/ext2fs/ext2fs.h | 2 ++ lib/ext2fs/gen_bitmap64.c | 35 +++++++++++++++++++++++++++++ misc/mke2fs.c | 47 +++++++++++++++++++++++++++++++++++++++ resize/resize2fs.c | 1 + 4 files changed, 85 insertions(+)
lib/ext2fs/ext2fs.h +2 -0
@@ -1442,6 +1442,8 @@ errcode_t ext2fs_set_generic_bmap_range(ext2fs_generic_bitmap bmap, void *in);errcode_t ext2fs_convert_subcluster_bitmap (ext2_filsys fs,ext2fs_block_bitmap *bitmap); +errcode_t ext2fs_count_used_clusters (ext2_filsys fs,blk64_t start, +blk64_t end,blk64_t *out);/* get_num_dirs.c */ extern errcode_t ext2fs_get_num_dirs (ext2_filsys fs,ext2_ino_t *ret_num_dirs);
lib/ext2fs/gen_bitmap64.c +35 -0
@@ -940,3 +940,38 @@ errcode_t ext2fs_find_first_set_generic_bmap(ext2fs_generic_bitmap bitmap,return ENOENT ; } + +errcode_t ext2fs_count_used_clusters (ext2_filsys fs,blk64_t start, +blk64_t end,blk64_t *out) +{ +blk64_t next; +blk64_t tot_set =0 ; +errcode_t retval; + +while (start < end) { + retval =ext2fs_find_first_set_block_bitmap2 (fs->block_map, + start, end, &next); +if (retval) { +if (retval ==ENOENT ) + retval =0 ; +break ; + } + start = next; + + retval =ext2fs_find_first_zero_block_bitmap2 (fs->block_map, + start, end, &next); +if (retval ==0 ) { + tot_set += next - start; + start = next +1 ; + }else if (retval ==ENOENT ) { + retval =0 ; + tot_set += end - start +1 ; +break ; + }else +break ; + } + +if (!retval) + *out =EXT2FS_NUM_B2C (fs, tot_set); +return retval; +}
misc/mke2fs.c +47 -0
83 lines not on this page · raw message
resize/resize2fs.c +1 -0
@@ -703,6 +703,7 @@ errcode_t adjust_fs_info(ext2_filsys fs, ext2_filsys old_fs,double percent;ext2fs_blocks_count_set (fs->super, new_size); + fs->super->s_overhead_clusters =0 ; retry: fs->group_desc_count = ext2fs_div64_ceil(ext2fs_blocks_count(fs->super) -
Signature · 1 line
2.24.0
Re: [PATCH v3 4/5] mke2fs: set overhead in super block
Quoted 23 lines · Li Dongyang
If overhead is not recorded in the super block, it is caculated during mount in kernel, for bigalloc file systems the it takes
O(groups**2)in time.
For a 1PB deivce with 32K cluste size it takes ~12 mins to mount, with most of the time spent on figuring out overhead.While we can not improve the overhead algorithm in kernel due to the nature of bigalloc, we can work out the overhead during mke2fs and set it in the super block, avoiding calculating it every time when it mounts.
Overhead is
s_first_data_blockplus internal journal blocks plus the block and inode bitmaps, inode table, super block backups and group descriptor blocks for every group. This patch introducesext2fs_count_used_clusters(), which calculates the clusters used in the block bitmap for the given range.When bad blocks are involved, it gets tricky because the blocks counted as overhead and the bad blocks can end up in the same allocation cluster. In this case we will unmark the bad blocks from the block bitmap, convert to cluster bitmap and get the overhead, then mark the bad blocks back in the cluster bitmap.
Quoted 5 lines
Reset the overhead to zero when resizing, we can not simplly count the used blocks as overhead like we do when mke2fs. The overhead can be calculated by kernel side during mount.
Signed-off-by Li Dongyang <[email protected]>
Applied, but I had to fix up a number of spelling errors in the commit description, and fix up a large number of test failures caused by this change.
- Ted
PATCH v3 5/5 ext4: record overhead in super block
In reply to Li DongyangStore the overhead in super block so we don't have to calculate again during next mount.
It also gets updated after online resizing.
--- fs/ext4/resize.c | 1 + fs/ext4/super.c | 4 ++++ 2 files changed, 5 insertions(+)
fs/ext4/resize.c +1 -0
@@ -1431,6 +1431,7 @@ static void ext4_update_super(struct super_block *sb,* Update the fs overhead information */ ext4_calculate_overhead (sb); + es->s_overhead_clusters =cpu_to_le32 (sbi->s_overhead);if (test_opt (sb,DEBUG )) printk(KERN_DEBUG "EXT4-fs: added group %u:"
fs/ext4/super.c +4 -0
@@ -4467,6 +4467,10 @@ static int ext4_fill_super(struct super_block *sb, void *data, int silent) err =ext4_calculate_overhead (sb);if (err) goto failed_mount_wq; +if (!sb_rdonly (sb)) { + es->s_overhead_clusters =cpu_to_le32 (sbi->s_overhead); +ext4_commit_super (sb,1 ); + } }/*
Signature · 1 line
2.24.0
Quoted 17 lines · Li Dongyang
For a bigalloc filesystem, converting the block bitmap from blocks to chunks in
ext2fs_convert_subcluster_bitmap()can take a long time when the device is huge, because we test the bitmap bit-by-bit usingext2fs_test_block_bitmap2().
Useext2fs_find_first_set_block_bitmap2()which is more efficient for mke2fs when the fs is mostly empty.e2fsck can also benefit from this during pass1 block scanning.
Time taken for "mke2fs -O bigalloc,extent -C 131072 -b 4096" on a 1PB device:
without patch:
real 27m49.457s user 21m36.474s sys 6m9.514s
Quoted 7 lines
with patch:
real 6m31.908s
user 0m1.806s
sys 6m29.697sSigned-off-by Li Dongyang <[email protected]>Reviewed-by Andreas Dilger <[email protected]>
Applied, thanks.
- Ted