ext4: Improve scalability of ext4 orphan file handling

Even though the length of the critical section when adding / removing
orphaned inodes was significantly reduced by using orphan file, the
contention of lock protecting orphan file still appears high in profiles
for truncate / unlink intensive workloads with high number of threads.

This patch makes handling of orphan file completely lockless. Also to
reduce conflicts between CPUs different CPUs start searching for empty
slot in orphan file in different blocks.

Performance comparison of locked orphan file handling, lockless orphan
file handling, and completely disabled orphan inode handling
from 80 CPU Xeon Server with 526 GB of RAM, filesystem located on
SAS SSD disk, average of 5 runs:

stress-orphan (microbenchmark truncating files byte-by-byte from N
processes in parallel)

Threads Time            Time            Time
        Orphan locked   Orphan lockless No orphan
  1       0.945600       0.939400        0.891200
  2       1.331800       1.246600        1.174400
  4       1.995000       1.780600        1.713200
  8       6.424200       4.900000        4.106000
 16      14.937600       8.516400        8.138000
 32      33.038200      24.565600       24.002200
 64      60.823600      39.844600       38.440200
128     122.941400      70.950400       69.315000

So we can see that with lockless orphan file handling, addition /
deletion of orphaned inodes got almost completely out of picture even
for a microbenchmark stressing it.

For reaim creat_clo workload on ramdisk there are also noticeable gains
(average of 5 runs):

Clients         Vanilla (ops/s)        Patched (ops/s)
creat_clo-1     14705.88 (   0.00%)    14354.07 *  -2.39%*
creat_clo-3     27108.43 (   0.00%)    28301.89 (   4.40%)
creat_clo-5     37406.48 (   0.00%)    45180.73 *  20.78%*
creat_clo-7     41338.58 (   0.00%)    54687.50 *  32.29%*
creat_clo-9     45226.13 (   0.00%)    62937.07 *  39.16%*
creat_clo-11    44000.00 (   0.00%)    65088.76 *  47.93%*
creat_clo-13    36516.85 (   0.00%)    68661.97 *  88.03%*
creat_clo-15    30864.20 (   0.00%)    69551.78 * 125.35%*
creat_clo-17    27478.45 (   0.00%)    67729.08 * 146.48%*
creat_clo-19    25000.00 (   0.00%)    61621.62 * 146.49%*
creat_clo-21    18772.35 (   0.00%)    63829.79 * 240.02%*
creat_clo-23    16698.94 (   0.00%)    61938.96 * 270.92%*
creat_clo-25    14973.05 (   0.00%)    56947.61 * 280.33%*
creat_clo-27    16436.69 (   0.00%)    65008.03 * 295.51%*
creat_clo-29    13949.01 (   0.00%)    69047.62 * 395.00%*
creat_clo-31    14283.52 (   0.00%)    67982.45 * 375.95%*

Reviewed-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Jan Kara <jack@suse.cz>

fs/ext4/ext4.h

@@ -1489,15 +1489,14 @@ static inline int ext4_inodes_per_orphan_block(struct super_block *sb)
 }
 
 struct ext4_orphan_block {
-	int ob_free_entries;	/* Number of free orphan entries in block */
+	atomic_t ob_free_entries;	/* Number of free orphan entries in block */
 	struct buffer_head *ob_bh;	/* Buffer for orphan block */
 };
 
 /*
  * Info about orphan file.
  */
 struct ext4_orphan_info {
-	spinlock_t of_lock;
 	int of_blocks;			/* Number of orphan blocks in a file */
 	__u32 of_csum_seed;		/* Checksum seed for orphan file */
 	struct ext4_orphan_block *of_binfo;	/* Array with info about orphan

fs/ext4/orphan.c

@@ -10,16 +10,31 @@
 
 static int ext4_orphan_file_add(handle_t *handle, struct inode *inode)
 {
-	int i, j;
+	int i, j, start;
 	struct ext4_orphan_info *oi = &EXT4_SB(inode->i_sb)->s_orphan_info;
 	int ret = 0;
+	bool found = false;
 	__le32 *bdata;
 	int inodes_per_ob = ext4_inodes_per_orphan_block(inode->i_sb);
+	int looped = 0;
 
-	spin_lock(&oi->of_lock);
-	for (i = 0; i < oi->of_blocks && !oi->of_binfo[i].ob_free_entries; i++);
-	if (i == oi->of_blocks) {
-		spin_unlock(&oi->of_lock);
+	/*
+	 * Find block with free orphan entry. Use CPU number for a naive hash
+	 * for a search start in the orphan file
+	 */
+	start = raw_smp_processor_id()*13 % oi->of_blocks;
+	i = start;
+	do {
+		if (atomic_dec_if_positive(&oi->of_binfo[i].ob_free_entries)
+		    >= 0) {
+			found = true;
+			break;
+		}
+		if (++i >= oi->of_blocks)
+			i = 0;
+	} while (i != start);
+
+	if (!found) {
 		/*
 		 * For now we don't grow or shrink orphan file. We just use
 		 * whatever was allocated at mke2fs time. The additional
@@ -28,28 +43,40 @@ static int ext4_orphan_file_add(handle_t *handle, struct inode *inode)
 		 */
 		return -ENOSPC;
 	}
-	oi->of_binfo[i].ob_free_entries--;
-	spin_unlock(&oi->of_lock);
 
-	/*
-	 * Get access to orphan block. We have dropped of_lock but since we
-	 * have decremented number of free entries we are guaranteed free entry
-	 * in our block.
-	 */
 	ret = ext4_journal_get_write_access(handle, inode->i_sb,
 				oi->of_binfo[i].ob_bh, EXT4_JTR_ORPHAN_FILE);
 	if (ret)
 		return ret;
 
 	bdata = (__le32 *)(oi->of_binfo[i].ob_bh->b_data);
-	spin_lock(&oi->of_lock);
 	/* Find empty slot in a block */
-	for (j = 0; j < inodes_per_ob && bdata[j]; j++);
-	BUG_ON(j == inodes_per_ob);
-	bdata[j] = cpu_to_le32(inode->i_ino);
+	j = 0;
+	do {
+		if (looped) {
+			/*
+			 * Did we walk through the block several times without
+			 * finding free entry? It is theoretically possible
+			 * if entries get constantly allocated and freed or
+			 * if the block is corrupted. Avoid indefinite looping
+			 * and bail. We'll use orphan list instead.
+			 */
+			if (looped > 3) {
+				atomic_inc(&oi->of_binfo[i].ob_free_entries);
+				return -ENOSPC;
+			}
+			cond_resched();
+		}
+		while (bdata[j]) {
+			if (++j >= inodes_per_ob) {
+				j = 0;
+				looped++;
+			}
+		}
+	} while (cmpxchg(&bdata[j], 0, cpu_to_le32(inode->i_ino)) != 0);
+
 	EXT4_I(inode)->i_orphan_idx = i * inodes_per_ob + j;
 	ext4_set_inode_state(inode, EXT4_STATE_ORPHAN_FILE);
-	spin_unlock(&oi->of_lock);
 
 	return ext4_handle_dirty_metadata(handle, NULL, oi->of_binfo[i].ob_bh);
 }
@@ -180,10 +207,8 @@ static int ext4_orphan_file_del(handle_t *handle, struct inode *inode)
 		goto out;
 
 	bdata = (__le32 *)(oi->of_binfo[blk].ob_bh->b_data);
-	spin_lock(&oi->of_lock);
 	bdata[off] = 0;
-	oi->of_binfo[blk].ob_free_entries++;
-	spin_unlock(&oi->of_lock);
+	atomic_inc(&oi->of_binfo[blk].ob_free_entries);
 	ret = ext4_handle_dirty_metadata(handle, NULL, oi->of_binfo[blk].ob_bh);
 out:
 	ext4_clear_inode_state(inode, EXT4_STATE_ORPHAN_FILE);
@@ -552,8 +577,6 @@ int ext4_init_orphan_info(struct super_block *sb)
 	struct ext4_orphan_block_tail *ot;
 	ino_t orphan_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_orphan_file_inum);
 
-	spin_lock_init(&oi->of_lock);
-
 	if (!ext4_has_feature_orphan_file(sb))
 		return 0;
 
@@ -597,7 +620,7 @@ int ext4_init_orphan_info(struct super_block *sb)
 		for (j = 0; j < inodes_per_ob; j++)
 			if (bdata[j] == 0)
 				free++;
-		oi->of_binfo[i].ob_free_entries = free;
+		atomic_set(&oi->of_binfo[i].ob_free_entries, free);
 	}
 	iput(inode);
 	return 0;
@@ -619,7 +642,8 @@ int ext4_orphan_file_empty(struct super_block *sb)
 	if (!ext4_has_feature_orphan_file(sb))
 		return 1;
 	for (i = 0; i < oi->of_blocks; i++)
-		if (oi->of_binfo[i].ob_free_entries != inodes_per_ob)
+		if (atomic_read(&oi->of_binfo[i].ob_free_entries) !=
+		    inodes_per_ob)
 			return 0;
 	return 1;
 }