Skip to contentSkip to conversation

zonefs: do not use append if device does not support it

PATCH In review Andreas Hindborg 4 participants 10 messages Jun 26, 2023 – Jun 27, 2023

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 zonefs: do not use append if device does not support it

From: "Andreas Hindborg (Samsung)" <[email protected]>

Zonefs will try to use zonefs_file_dio_append() for direct sync writes even if device max_zone_append_sectors is zero. This will cause the IO to fail as the io vector is truncated to zero. It also causes a call to invalidate_inode_pages2_range() with end set to UINT_MAX, which is probably not intentional. Thus, do not use append when device does not support it.

---
 fs/zonefs/file.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
fs/zonefs/file.c +3 -1
index 132f01d3461f..c97fe2aa20b0 100644
@@ -536,9 +536,11 @@ static ssize_t zonefs_write_checks(struct kiocb *iocb, struct iov_iter *from)
 static ssize_t zonefs_file_dio_write(struct kiocb *iocb, struct iov_iter *from)
 {
 	struct inode *inode = file_inode(iocb->ki_filp);
+	struct block_device *bdev = inode->i_sb->s_bdev;
 	struct zonefs_inode_info *zi = ZONEFS_I(inode);
 	struct zonefs_zone *z = zonefs_inode_zone(inode);
 	struct super_block *sb = inode->i_sb;
+	unsigned int max_append = bdev_max_zone_append_sectors(bdev);
 	bool sync = is_sync_kiocb(iocb);
 	bool append = false;
 	ssize_t ret, count;
@@ -581,7 +583,7 @@ static ssize_t zonefs_file_dio_write(struct kiocb *iocb, struct iov_iter *from)
 		append = sync;
 	}
 
-	if (append) {
+	if (append && max_append) {
 		ret = zonefs_file_dio_append(iocb, from);
 	} else {
 		/*

base-commit: 45a3e24f65e90a047bef86f927ebdc4c710edaa1

Signature · 1 line
2.41.0
#1 lore raw
Quoted 8 lines · Andreas Hindborg

From: "Andreas Hindborg (Samsung)" <[email protected]>

Zonefs will try to use zonefs_file_dio_append() for direct sync writes even if device max_zone_append_sectors is zero. This will cause the IO to fail as the io vector is truncated to zero. It also causes a call to invalidate_inode_pages2_range() with end set to UINT_MAX, which is probably not intentional. Thus, do not use append when device does not support it.

I'm sorry but I think it has been stated often enough that for Linux Zone Append is a mandatory feature for a Zoned Block Device. Therefore this path is essentially dead code as max_zone_append_sectors will always be greater than zero.

So this is a clear NAK from my side.

#2 lore raw
Quoted 10 lines · Johannes Thumshirn
Quoted 8 lines · Andreas Hindborg

From: "Andreas Hindborg (Samsung)" <[email protected]>

Zonefs will try to use zonefs_file_dio_append() for direct sync writes even if device max_zone_append_sectors is zero. This will cause the IO to fail as the io vector is truncated to zero. It also causes a call to invalidate_inode_pages2_range() with end set to UINT_MAX, which is probably not intentional. Thus, do not use append when device does not support it.

Quoted 5 lines

I'm sorry but I think it has been stated often enough that for Linux Zone Append is a mandatory feature for a Zoned Block Device. Therefore this path is essentially dead code as max_zone_append_sectors will always be greater than zero.

So this is a clear NAK from my side.

OK, thanks for clarifying 👍 I came across this bugging out while playing around with zone append for ublk. The code makes sense if the stack expects append to always be present.

I didn't follow the discussion, could you reiterate why the policy is that zoned devices _must_ support append?

Best regards,
Andreas

maintainer
#3 lore raw
Quoted 17 lines · Andreas Hindborg (Samsung)
Quoted 14 lines · Johannes Thumshirn
Quoted 8 lines · Andreas Hindborg

From: "Andreas Hindborg (Samsung)" <[email protected]>

Zonefs will try to use zonefs_file_dio_append() for direct sync writes even if device max_zone_append_sectors is zero. This will cause the IO to fail as the io vector is truncated to zero. It also causes a call to invalidate_inode_pages2_range() with end set to UINT_MAX, which is probably not intentional. Thus, do not use append when device does not support it.

I'm sorry but I think it has been stated often enough that for Linux Zone Append is a mandatory feature for a Zoned Block Device. Therefore this path is essentially dead code as max_zone_append_sectors will always be greater than zero.

Quoted 8 lines
Quoted 1 line

So this is a clear NAK from my side.

OK, thanks for clarifying 👍 I came across this bugging out while playing around with zone append for ublk. The code makes sense if the stack expects append to always be present.

I didn't follow the discussion, could you reiterate why the policy is that zoned devices _must_ support append?

To avoid support fragmentation and for performance. btrfs zoned block device support requires zone append and using that command makes writes much faster as we do not have to go through zone locking.
Note that for zonefs, I plan to add async zone append support as well, linked with O_APPEND use to further improve write performance with ZNS drives.

Quoted 4 lines

Best regards,
Andreas

Signature · 2 lines
Damien Le Moal
Western Digital Research
#4 lore raw
Quoted 24 lines · Damien Le Moal
Quoted 23 lines · Andreas Hindborg (Samsung)
Quoted 10 lines · Johannes Thumshirn
Quoted 8 lines · Andreas Hindborg

From: "Andreas Hindborg (Samsung)" <[email protected]>

Zonefs will try to use zonefs_file_dio_append() for direct sync writes even if device max_zone_append_sectors is zero. This will cause the IO to fail as the io vector is truncated to zero. It also causes a call to invalidate_inode_pages2_range() with end set to UINT_MAX, which is probably not intentional. Thus, do not use append when device does not support it.

Quoted 5 lines

I'm sorry but I think it has been stated often enough that for Linux Zone Append is a mandatory feature for a Zoned Block Device. Therefore this path is essentially dead code as max_zone_append_sectors will always be greater than zero.

So this is a clear NAK from my side.

OK, thanks for clarifying 👍 I came across this bugging out while playing around with zone append for ublk. The code makes sense if the stack expects append to always be present.

Quoted 9 lines
Quoted 2 lines

I didn't follow the discussion, could you reiterate why the policy is that zoned devices _must_ support append?

To avoid support fragmentation and for performance. btrfs zoned block device support requires zone append and using that command makes writes much faster as we do not have to go through zone locking.
Note that for zonefs, I plan to add async zone append support as well, linked with O_APPEND use to further improve write performance with ZNS drives.

Thanks for clarifying, Damien 👍

BR Andreas

#5 lore raw
In reply to Andreas Hindborg
Quoted 7 lines · Andreas Hindborg

From: "Andreas Hindborg (Samsung)" <[email protected]>

Zonefs will try to use zonefs_file_dio_append() for direct sync writes even if device max_zone_append_sectors is zero. This will cause the IO to fail as the io vector is truncated to zero. It also causes a call to invalidate_inode_pages2_range() with end set to UINT_MAX, which is probably not intentional. Thus, do not use append when device does not support it.

How do you even manage to hit this code? Zone Append is a mandatory feature and driver need to check it is available.

maintainer
#6 lore raw
Quoted 11 lines · Christoph Hellwig
Quoted 7 lines · Andreas Hindborg

From: "Andreas Hindborg (Samsung)" <[email protected]>

Zonefs will try to use zonefs_file_dio_append() for direct sync writes even if device max_zone_append_sectors is zero. This will cause the IO to fail as the io vector is truncated to zero. It also causes a call to invalidate_inode_pages2_range() with end set to UINT_MAX, which is probably not intentional. Thus, do not use append when device does not support it.

How do you even manage to hit this code? Zone Append is a mandatory feature and driver need to check it is available.

ublk driver probably is missing that check ? I have not looked at the code for zone support.

But thinking of it, we probably would be better off having a generic check for "q->limits.max_zone_append_sectors != 0" in blk_revalidate_disk_zones().

Signature · 2 lines
Damien Le Moal
Western Digital Research
#7 lore raw
Quoted 2 lines · Damien Le Moal

But thinking of it, we probably would be better off having a generic check for "q->limits.max_zone_append_sectors != 0" in blk_revalidate_disk_zones().

Agreed.

maintainer
#8 lore raw
Quoted 5 lines · Christoph Hellwig
Quoted 2 lines · Damien Le Moal

But thinking of it, we probably would be better off having a generic check for "q->limits.max_zone_append_sectors != 0" in blk_revalidate_disk_zones().

Agreed.

I'll send something.

Signature · 2 lines
Damien Le Moal
Western Digital Research
#9 lore raw
In reply to Damien Le Moal
Quoted 10 lines · Damien Le Moal
Quoted 9 lines · Christoph Hellwig
Quoted 7 lines · Andreas Hindborg

From: "Andreas Hindborg (Samsung)" <[email protected]>

Zonefs will try to use zonefs_file_dio_append() for direct sync writes even if device max_zone_append_sectors is zero. This will cause the IO to fail as the io vector is truncated to zero. It also causes a call to invalidate_inode_pages2_range() with end set to UINT_MAX, which is probably not intentional. Thus, do not use append when device does not support it.

Quoted 8 lines
Quoted 2 lines

How do you even manage to hit this code? Zone Append is a mandatory feature and driver need to check it is available.

ublk driver probably is missing that check ? I have not looked at the code for zone support.

But thinking of it, we probably would be better off having a generic check for "q->limits.max_zone_append_sectors != 0" in blk_revalidate_disk_zones().

I was playing with ublk zone support. It seems I made it buggy by allowing zone append size to go to zero.

Adding the check would be a nice help to people like me that will implement whatever in their driver :)

Best regards
Andreas