ext4: add prefetching for block allocation bitmaps

This should significantly improve bitmap loading, especially for flex
groups as it tries to load all bitmaps within a flex.group instead of
one by one synchronously.

Prefetching is done in 8 * flex_bg groups, so it should be 8
read-ahead reads for a single allocating thread. At the end of
allocation the thread waits for read-ahead completion and initializes
buddy information so that read-aheads are not lost in case of memory
pressure.

At cr=0 the number of prefetching IOs is limited per allocation
context to prevent a situation when mballoc loads thousands of bitmaps
looking for a perfect group and ignoring groups with good chunks.

Together with the patch "ext4: limit scanning of uninitialized groups"
the mount time (which includes few tiny allocations) of a 1PB
filesystem is reduced significantly:

               0% full    50%-full unpatched    patched
  mount time       33s                9279s       563s

[ Restructured by tytso; removed the state flags in the allocation
  context, so it can be used to lazily prefetch the allocation bitmaps
  immediately after the file system is mounted.  Skip prefetching
  block groups which are uninitialized.  Finally pass in the
  REQ_RAHEAD flag to the block layer while prefetching. ]

Signed-off-by: Alex Zhuravlev <bzzz@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

1 files changed, 131 insertions, 2 deletions

diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
index 6dc2c6c535ef..c350ae9cb0d9 100644
--- a/fs/ext4/mballoc.c
+++ b/fs/ext4/mballoc.c
@@ -922,7 +922,7 @@ static int ext4_mb_init_cache(struct page *page, char *incore, gfp_t gfp)
 			bh[i] = NULL;
 			continue;
 		}
-		bh[i] = ext4_read_block_bitmap_nowait(sb, group);
+		bh[i] = ext4_read_block_bitmap_nowait(sb, group, false);
 		if (IS_ERR(bh[i])) {
 			err = PTR_ERR(bh[i]);
 			bh[i] = NULL;
@@ -2209,12 +2209,93 @@ out:
 	return ret;
 }
 
+/*
+ * Start prefetching @nr block bitmaps starting at @group.
+ * Return the next group which needs to be prefetched.
+ */
+static ext4_group_t
+ext4_mb_prefetch(struct super_block *sb, ext4_group_t group,
+		 unsigned int nr, int *cnt)
+{
+	ext4_group_t ngroups = ext4_get_groups_count(sb);
+	struct buffer_head *bh;
+	struct blk_plug plug;
+
+	blk_start_plug(&plug);
+	while (nr-- > 0) {
+		struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
+								  NULL);
+		struct ext4_group_info *grp = ext4_get_group_info(sb, group);
+
+		/*
+		 * Prefetch block groups with free blocks; but don't
+		 * bother if it is marked uninitialized on disk, since
+		 * it won't require I/O to read.  Also only try to
+		 * prefetch once, so we avoid getblk() call, which can
+		 * be expensive.
+		 */
+		if (!EXT4_MB_GRP_TEST_AND_SET_READ(grp) &&
+		    EXT4_MB_GRP_NEED_INIT(grp) &&
+		    ext4_free_group_clusters(sb, gdp) > 0 &&
+		    !(ext4_has_group_desc_csum(sb) &&
+		      (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
+			bh = ext4_read_block_bitmap_nowait(sb, group, true);
+			if (bh && !IS_ERR(bh)) {
+				if (!buffer_uptodate(bh) && cnt)
+					(*cnt)++;
+				brelse(bh);
+			}
+		}
+		if (++group >= ngroups)
+			group = 0;
+	}
+	blk_finish_plug(&plug);
+	return group;
+}
+
+/*
+ * Prefetching reads the block bitmap into the buffer cache; but we
+ * need to make sure that the buddy bitmap in the page cache has been
+ * initialized.  Note that ext4_mb_init_group() will block if the I/O
+ * is not yet completed, or indeed if it was not initiated by
+ * ext4_mb_prefetch did not start the I/O.
+ *
+ * TODO: We should actually kick off the buddy bitmap setup in a work
+ * queue when the buffer I/O is completed, so that we don't block
+ * waiting for the block allocation bitmap read to finish when
+ * ext4_mb_prefetch_fini is called from ext4_mb_regular_allocator().
+ */
+static void
+ext4_mb_prefetch_fini(struct super_block *sb, ext4_group_t group,
+		      unsigned int nr)
+{
+	while (nr-- > 0) {
+		struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group,
+								  NULL);
+		struct ext4_group_info *grp = ext4_get_group_info(sb, group);
+
+		if (!group)
+			group = ext4_get_groups_count(sb);
+		group--;
+		grp = ext4_get_group_info(sb, group);
+
+		if (EXT4_MB_GRP_NEED_INIT(grp) &&
+		    ext4_free_group_clusters(sb, gdp) > 0 &&
+		    !(ext4_has_group_desc_csum(sb) &&
+		      (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)))) {
+			if (ext4_mb_init_group(sb, group, GFP_NOFS))
+				break;
+		}
+	}
+}
+
 static noinline_for_stack int
 ext4_mb_regular_allocator(struct ext4_allocation_context *ac)
 {
-	ext4_group_t ngroups, group, i;
+	ext4_group_t prefetch_grp = 0, ngroups, group, i;
 	int cr = -1;
 	int err = 0, first_err = 0;
+	unsigned int nr = 0, prefetch_ios = 0;
 	struct ext4_sb_info *sbi;
 	struct super_block *sb;
 	struct ext4_buddy e4b;
@@ -2282,6 +2363,7 @@ repeat:
 		 * from the goal value specified
 		 */
 		group = ac->ac_g_ex.fe_group;
+		prefetch_grp = group;
 
 		for (i = 0; i < ngroups; group++, i++) {
 			int ret = 0;
@@ -2293,6 +2375,29 @@ repeat:
 			if (group >= ngroups)
 				group = 0;
 
+			/*
+			 * Batch reads of the block allocation bitmaps
+			 * to get multiple READs in flight; limit
+			 * prefetching at cr=0/1, otherwise mballoc can
+			 * spend a lot of time loading imperfect groups
+			 */
+			if ((prefetch_grp == group) &&
+			    (cr > 1 ||
+			     prefetch_ios < sbi->s_mb_prefetch_limit)) {
+				unsigned int curr_ios = prefetch_ios;
+
+				nr = sbi->s_mb_prefetch;
+				if (ext4_has_feature_flex_bg(sb)) {
+					nr = (group / sbi->s_mb_prefetch) *
+						sbi->s_mb_prefetch;
+					nr = nr + sbi->s_mb_prefetch - group;
+				}
+				prefetch_grp = ext4_mb_prefetch(sb, group,
+							nr, &prefetch_ios);
+				if (prefetch_ios == curr_ios)
+					nr = 0;
+			}
+
 			/* This now checks without needing the buddy page */
 			ret = ext4_mb_good_group_nolock(ac, group, cr);
 			if (ret <= 0) {
@@ -2367,6 +2472,10 @@ out:
 	mb_debug(sb, "Best len %d, origin len %d, ac_status %u, ac_flags 0x%x, cr %d ret %d\n",
 		 ac->ac_b_ex.fe_len, ac->ac_o_ex.fe_len, ac->ac_status,
 		 ac->ac_flags, cr, err);
+
+	if (nr)
+		ext4_mb_prefetch_fini(sb, prefetch_grp, nr);
+
 	return err;
 }
 
@@ -2613,6 +2722,26 @@ static int ext4_mb_init_backend(struct super_block *sb)
 			goto err_freebuddy;
 	}
 
+	if (ext4_has_feature_flex_bg(sb)) {
+		/* a single flex group is supposed to be read by a single IO */
+		sbi->s_mb_prefetch = 1 << sbi->s_es->s_log_groups_per_flex;
+		sbi->s_mb_prefetch *= 8; /* 8 prefetch IOs in flight at most */
+	} else {
+		sbi->s_mb_prefetch = 32;
+	}
+	if (sbi->s_mb_prefetch > ext4_get_groups_count(sb))
+		sbi->s_mb_prefetch = ext4_get_groups_count(sb);
+	/* now many real IOs to prefetch within a single allocation at cr=0
+	 * given cr=0 is an CPU-related optimization we shouldn't try to
+	 * load too many groups, at some point we should start to use what
+	 * we've got in memory.
+	 * with an average random access time 5ms, it'd take a second to get
+	 * 200 groups (* N with flex_bg), so let's make this limit 4
+	 */
+	sbi->s_mb_prefetch_limit = sbi->s_mb_prefetch * 4;
+	if (sbi->s_mb_prefetch_limit > ext4_get_groups_count(sb))
+		sbi->s_mb_prefetch_limit = ext4_get_groups_count(sb);
+
 	return 0;
 
 err_freebuddy: