ext4: Change s_last_trim_minblks type to unsigned long

Ext4 has an optimization mechanism for batched disacrd (FITRIM) that should help speed up subsequent calls of FITRIM ioctl by skipping the groups that were previously trimmed. However because the FITRIM allows to set the minimum size of an extent to trim, ext4 stores the last minimum extent size and only avoids trimming the group if it was previously trimmed with minimum extent size equal to, or smaller than the current call.

There is currently no way to bypass the optimization without umount/mount cycle. This becomes a problem when the file system is live migrated to a different storage, because the optimization will prevent possibly useful discard calls to the storage.

Fix it by exporting the s_last_trim_minblks via sysfs interface which will allow us to set the minimum size to the number of blocks larger than subsequent FITRIM call, effectively bypassing the optimization.

By setting the s_last_trim_minblks to ULONG_MAX the optimization will be effectively cleared regardless of the previous state, or file system configuration.

For example:
getconf ULONG_MAX > /sys/fs/ext4/dm-1/last_trim_minblks

Signed-off-by Lukas Czerner <[email protected]>

Reported-by Laurent GUERBY <[email protected]>

---
v2: Remove unnecessary assignment
v3: s_last_trim_minblks is now unsinged long which simplifies this

fs/ext4/sysfs.c | 2 ++

1 file changed, 2 insertions(+)

fs/ext4/sysfs.c +2 -1

index 2314f7446592..95d8a996d2d8 100644
@@ -245,6 +245,7 @@ EXT4_ATTR(last_error_time, 0444, last_error_time);
 EXT4_ATTR(journal_task, 0444, journal_task);
 EXT4_RW_ATTR_SBI_UI(mb_prefetch, s_mb_prefetch);
 EXT4_RW_ATTR_SBI_UI(mb_prefetch_limit, s_mb_prefetch_limit);
+EXT4_RW_ATTR_SBI_UL(last_trim_minblks, s_last_trim_minblks);
 
 static unsigned int old_bump_val = 128;
 EXT4_ATTR_PTR(max_writeback_mb_bump, 0444, pointer_ui, &old_bump_val);

@@ -295,6 +296,7 @@ static struct attribute *ext4_attrs[] = {
 #endif
 	ATTR_LIST(mb_prefetch),
 	ATTR_LIST(mb_prefetch_limit),
+	ATTR_LIST(last_trim_minblks),
 	NULL,
 };
 ATTRIBUTE_GROUPS(ext4);
--

2.31.1

Ext4 has an optimization mechanism for batched disacrd (FITRIM) that should help speed up subsequent calls of FITRIM ioctl by skipping the groups that were previously trimmed. However because the FITRIM allows to set the minimum size of an extent to trim, ext4 stores the last minimum extent size and only avoids trimming the group if it was previously trimmed with minimum extent size equal to, or smaller than the current call.

There is currently no way to bypass the optimization without umount/mount cycle. This becomes a problem when the file system is live migrated to a different storage, because the optimization will prevent possibly useful discard calls to the storage.

Fix it by exporting the s_last_trim_minblks via sysfs interface which will allow us to set the minimum size to the number of blocks larger than subsequent FITRIM call, effectively bypassing the optimization.

By setting the s_last_trim_minblks to ULONG_MAX the optimization will be effectively cleared regardless of the previous state, or file system configuration.

For example:
getconf ULONG_MAX > /sys/fs/ext4/dm-1/last_trim_minblks

Signed-off-by Lukas Czerner <[email protected]>

Reported-by Laurent GUERBY <[email protected]>

---
v2: Remove unnecessary assignment
v3: s_last_trim_minblks is now unsinged long which simplifies this

 fs/ext4/sysfs.c | 2 ++
 1 file changed, 2 insertions(+)

fs/ext4/sysfs.c +2 -0

index 2314f7446592..95d8a996d2d8 100644
@@ -245,6 +245,7 @@ EXT4_ATTR(last_error_time, 0444, last_error_time);
 EXT4_ATTR(journal_task, 0444, journal_task);
 EXT4_RW_ATTR_SBI_UI(mb_prefetch, s_mb_prefetch);
 EXT4_RW_ATTR_SBI_UI(mb_prefetch_limit, s_mb_prefetch_limit);
+EXT4_RW_ATTR_SBI_UL(last_trim_minblks, s_last_trim_minblks);
 
 static unsigned int old_bump_val = 128;
 EXT4_ATTR_PTR(max_writeback_mb_bump, 0444, pointer_ui, &old_bump_val);

@@ -295,6 +296,7 @@ static struct attribute *ext4_attrs[] = {
 #endif
 	ATTR_LIST(mb_prefetch),
 	ATTR_LIST(mb_prefetch_limit),
+	ATTR_LIST(last_trim_minblks),
 	NULL,
 };
 ATTRIBUTE_GROUPS(ext4);

Signature · 1 line

2.31.1

Ext4 has an optimization mechanism for batched disacrd (FITRIM) that should help speed up subsequent calls of FITRIM ioctl by skipping the groups that were previously trimmed. However because the FITRIM allows to set the minimum size of an extent to trim, ext4 stores the last minimum extent size and only avoids trimming the group if it was previously trimmed with minimum extent size equal to, or smaller than the current call.

[...]

There is no good reason for the s_last_trim_minblks to be atomic. There is no data integrity needed and there is no real danger in setting and reading it in a racy manner. Change it to be unsigned long, the same type as s_clusters_per_group which is the maximum that's allowed.

Signed-off-by Lukas Czerner <[email protected]>

Suggested-by Andreas Dilger <[email protected]>

---
fs/ext4/ext4.h | 2 +-
fs/ext4/mballoc.c | 4 ++--

2 files changed, 3 insertions(+), 3 deletions(-)

fs/ext4/ext4.h +1 -1

index 3825195539d7..92a155401f61 100644
@@ -1660,7 +1660,7 @@ struct ext4_sb_info {
 	struct task_struct *s_mmp_tsk;
 
 	/* record the last minlen when FITRIM is called. */
-	atomic_t s_last_trim_minblks;
+	unsigned long s_last_trim_minblks;
 
 	/* Reference to checksum algorithm driver via cryptoapi */
 	struct crypto_shash *s_chksum_driver;

fs/ext4/mballoc.c +2 -2

index 72bfac2d6dce..eda550ec3956 100644
@@ -6374,7 +6374,7 @@ ext4_trim_all_free(struct super_block *sb, ext4_group_t group,
 	ext4_lock_group(sb, group);
 
 	if (!EXT4_MB_GRP_WAS_TRIMMED(e4b.bd_info) ||
-	    minblocks < atomic_read(&EXT4_SB(sb)->s_last_trim_minblks)) {
+	    minblocks < EXT4_SB(sb)->s_last_trim_minblks) {
 		ret = ext4_try_to_trim_range(sb, &e4b, start, max, minblocks);
 		if (ret >= 0)
 			EXT4_MB_GRP_SET_TRIMMED(e4b.bd_info);

@@ -6475,7 +6475,7 @@ int ext4_trim_fs(struct super_block *sb, struct fstrim_range *range)
 	}
 
 	if (!ret)
-		atomic_set(&EXT4_SB(sb)->s_last_trim_minblks, minlen);
+		EXT4_SB(sb)->s_last_trim_minblks = minlen;
 

out:

	range->len = EXT4_C2B(EXT4_SB(sb), trimmed) << sb->s_blocksize_bits;

--
2.31.1

There is no good reason for the s_last_trim_minblks to be atomic. There is no data integrity needed and there is no real danger in setting and reading it in a racy manner. Change it to be unsigned long, the same type as s_clusters_per_group which is the maximum that's allowed.

Signed-off-by Lukas Czerner <[email protected]>

Suggested-by Andreas Dilger <[email protected]>

Reviewed-by Andreas Dilger <[email protected]>