Skip to contentSkip to conversation

btrfs: output more info for -ENOSPC caused transaction abort and other enhancement

PATCH Superseded Qu Wenruo 5 participants 24 messages Jul 19, 2022 – Aug 25, 2022 Btrfs file system

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Qu Wenruo

author
#0 lore raw

v2 cover btrfs: output more info for -ENOSPC caused transaction abort and other enhancement

[Changelog]
v2:
- Add a new line to show the meaning of the metadata dump. Previous output only includes the reserved bytes and size bytes, but not showing which is which, thus still need to check the code

We have some internal reports of transaction abort with -ENOSPC.

Unfortunately it's really hard to debug, since we really got nothing other than the error message for most cases.

Also we have helpers like __btrfs_dump_space_info(), it needs enospc_debug mount option which only makes sense if the user can reproduce the bug and retry with that mount option.

Considering ENOSPC should not happen for critical paths which failure means transaction abort, we should dump all space info for debug purpose if the transaction arbot is caused by -ENOSPC.

And since we're here, also enhance the output of
__btrfs_dump_space_info().

Now it should be more human readable, the exmaple would look like this:

BTRFS info (device dm-1: state A): dumpping space info:
BTRFS info (device dm-1: state A): space_info DATA has 1175552 free, is not full

 BTRFS info (device dm-1: state A):   total:         8388608
 BTRFS info (device dm-1: state A):   used:          7163904
 BTRFS info (device dm-1: state A):   pinned:        49152
 BTRFS info (device dm-1: state A):   reserved:      0
 BTRFS info (device dm-1: state A):   may_use:       0
 BTRFS info (device dm-1: state A):   read_only:     0
 BTRFS info (device dm-1: state A): space_info META has 263798784 free, is not full
 BTRFS info (device dm-1: state A):   total:         268435456
 BTRFS info (device dm-1: state A):   used:          180224
 BTRFS info (device dm-1: state A):   pinned:        196608
 BTRFS info (device dm-1: state A):   reserved:      0
 BTRFS info (device dm-1: state A):   may_use:       4194304
 BTRFS info (device dm-1: state A):   read_only:     65536
 BTRFS info (device dm-1: state A): space_info SYS has 8372224 free, is not full
 BTRFS info (device dm-1: state A):   total:         8388608
 BTRFS info (device dm-1: state A):   used:          16384
 BTRFS info (device dm-1: state A):   pinned:        0
 BTRFS info (device dm-1: state A):   reserved:      0
 BTRFS info (device dm-1: state A):   may_use:       0
 BTRFS info (device dm-1: state A):   read_only:     0
 BTRFS info (device dm-1: state A): dumpping metadata reservation:
 BTRFS info (device dm-1: state A):   global:          (3670016/3670016)
 BTRFS info (device dm-1: state A):   trans:           (0/0)
 BTRFS info (device dm-1: state A):   chuunk:          (0/0)
 BTRFS info (device dm-1: state A):   delayed_inode:   (0/0)
 BTRFS info (device dm-1: state A):   delayed_refs:    (524288/524288)

BTRFS: error (device dm-1: state A) in cleanup_transaction:1971: errno=-28 No space left BTRFS info (device dm-1: state EA): forced readonly

Qu Wenruo (4):
btrfs: output human readable space info flag
btrfs: make __btrfs_dump_space_info() output better formatted
btrfs: make DUMP_BLOCK_RSV() to have better output
btrfs: dump all space infos if we abort transaction due to ENOSPC

 fs/btrfs/ctree.h      |  6 ++--
 fs/btrfs/space-info.c | 72 +++++++++++++++++++++++++++++++++----------
 fs/btrfs/space-info.h |  2 ++
 fs/btrfs/super.c      |  4 ++-
 4 files changed, 64 insertions(+), 20 deletions(-)
Signature · 1 line
2.37.0
author
#1 lore raw

PATCH v2 1/4 btrfs: output human readable space info flag

For btrfs_space_info, its flags has only 4 possible values:
- BTRFS_BLOCK_GROUP_SYSTEM
- BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA
- BTRFS_BLOCK_GROUP_METADATA
- BTRFS_BLOCK_GROUP_DATA

Thus do debuggers a favor by output a human readable flags in __btrfs_dump_space_info().

Now the summary line of __btrfs_dump_space_info() looks like:

BTRFS info (device dm-1: state A): space_info META has 251494400 free, is not full

---
 fs/btrfs/space-info.c | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)
fs/btrfs/space-info.c +17 -2
index 62d25112310d..36b466525318 100644
@@ -450,14 +450,29 @@ do {									\
 	spin_unlock(&__rsv->lock);					\
 } while (0)
 
+static const char *space_info_flag_to_str(struct btrfs_space_info *space_info)
+{
+	if (space_info->flags == BTRFS_BLOCK_GROUP_SYSTEM)
+		return "SYS";
+
+	/* Handle mixed data+metadata first. */
+	if (space_info->flags == (BTRFS_BLOCK_GROUP_METADATA |
+				  BTRFS_BLOCK_GROUP_DATA))
+		return "DATA+META";
+	if (space_info->flags == BTRFS_BLOCK_GROUP_DATA)
+		return "DATA";
+	return "META";
+}
+
 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 				    struct btrfs_space_info *info)
 {
+	const char *flag_str = space_info_flag_to_str(info);
 	lockdep_assert_held(&info->lock);
 
 	/* The free space could be negative in case of overcommit */
-	btrfs_info(fs_info, "space_info %llu has %lld free, is %sfull",
-		   info->flags,
+	btrfs_info(fs_info, "space_info %s has %lld free, is %sfull",
+		   flag_str,
 		   (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
 		   info->full ? "" : "not ");
 	btrfs_info(fs_info,
Signature · 1 line
2.37.0
#2 lore raw

Re: [PATCH v2 1/4] btrfs: output human readable space info flag

Quoted 30 lines · Qu Wenruo

For btrfs_space_info, its flags has only 4 possible values:
- BTRFS_BLOCK_GROUP_SYSTEM
- BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA
- BTRFS_BLOCK_GROUP_METADATA
- BTRFS_BLOCK_GROUP_DATA

Thus do debuggers a favor by output a human readable flags in __btrfs_dump_space_info().

Now the summary line of __btrfs_dump_space_info() looks like:

BTRFS info (device dm-1: state A): space_info META has 251494400 free, is not full

---
  fs/btrfs/space-info.c | 19 +++++++++++++++++--
  1 file changed, 17 insertions(+), 2 deletions(-)
fs/btrfs/space-info.c +3 -0
index 62d25112310d..36b466525318 100644
@@ -450,14 +450,29 @@ do {									\
  	spin_unlock(&__rsv->lock);					\
  } while (0)
 
+static const char *space_info_flag_to_str(struct btrfs_space_info *space_info)
+{
+	if (space_info->flags == BTRFS_BLOCK_GROUP_SYSTEM)

fs/btrfs/space-info.c do { \ · line 465

Quoted 10 lines
fs/btrfs/space-info.c +10 -0
+		return "SYS";
+
+	/* Handle mixed data+metadata first. */
+	if (space_info->flags == (BTRFS_BLOCK_GROUP_METADATA |
+				  BTRFS_BLOCK_GROUP_DATA))
+		return "DATA+META";
+	if (space_info->flags == BTRFS_BLOCK_GROUP_DATA)
+		return "DATA";
+	return "META";
+}

I think this would be more elegant as a switch statement; and would prefer the default to be "UNKNOWN" just in case there were a flag bitflip. +

Quoted 14 lines
fs/btrfs/space-info.c +3 -2
  static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
  				    struct btrfs_space_info *info)
  {
+	const char *flag_str = space_info_flag_to_str(info);
  	lockdep_assert_held(&info->lock);
 
  	/* The free space could be negative in case of overcommit */
-	btrfs_info(fs_info, "space_info %llu has %lld free, is %sfull",
-		   info->flags,
+	btrfs_info(fs_info, "space_info %s has %lld free, is %sfull",
+		   flag_str,
  		   (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
  		   info->full ? "" : "not ");
  	btrfs_info(fs_info,
author
#3 lore raw

PATCH v2 2/4 btrfs: make __btrfs_dump_space_info() output better formatted

In reply to Qu Wenruo

The format change includes:

- Output each bytes_* in a separate line

- All bytes_* output starts at the same vertical position Do human a favor reading the numbers

- Skip zone specific numbers if zone is not enabled

Now one example of __btrfs_dump_space_info() looks like this for its bytes_* members.

BTRFS info (device dm-1: state A): space_info META has 251494400 free, is not full

 BTRFS info (device dm-1: state A):   total:         268435456
 BTRFS info (device dm-1: state A):   used:          376832
 BTRFS info (device dm-1: state A):   pinned:        229376
 BTRFS info (device dm-1: state A):   reserved:      0
 BTRFS info (device dm-1: state A):   may_use:       16269312
 BTRFS info (device dm-1: state A):   read_only:     65536
---
 fs/btrfs/space-info.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)
fs/btrfs/space-info.c +9 -5
index 36b466525318..623fa0488545 100644
@@ -475,11 +475,15 @@ static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 		   flag_str,
 		   (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
 		   info->full ? "" : "not ");
-	btrfs_info(fs_info,
-		"space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
-		info->total_bytes, info->bytes_used, info->bytes_pinned,
-		info->bytes_reserved, info->bytes_may_use,
-		info->bytes_readonly, info->bytes_zone_unusable);
+	btrfs_info(fs_info, "  total:         %llu", info->total_bytes);
+	btrfs_info(fs_info, "  used:          %llu", info->bytes_used);
+	btrfs_info(fs_info, "  pinned:        %llu", info->bytes_pinned);
+	btrfs_info(fs_info, "  reserved:      %llu", info->bytes_reserved);
+	btrfs_info(fs_info, "  may_use:       %llu", info->bytes_may_use);
+	btrfs_info(fs_info, "  read_only:     %llu", info->bytes_readonly);
+	if (btrfs_is_zoned(fs_info))
+		btrfs_info(fs_info,
+			    "  zone_unusable: %llu", info->bytes_zone_unusable);
 
 	DUMP_BLOCK_RSV(fs_info, global_block_rsv);
 	DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
Signature · 1 line
2.37.0
#4 lore raw

Re: [PATCH v2 2/4] btrfs: make __btrfs_dump_space_info() output better formatted

Quoted 38 lines · Qu Wenruo

The format change includes:

- Output each bytes_* in a separate line

- All bytes_* output starts at the same vertical position Do human a favor reading the numbers

- Skip zone specific numbers if zone is not enabled

Now one example of __btrfs_dump_space_info() looks like this for its bytes_* members.

BTRFS info (device dm-1: state A): space_info META has 251494400 free, is not full

  BTRFS info (device dm-1: state A):   total:         268435456
  BTRFS info (device dm-1: state A):   used:          376832
  BTRFS info (device dm-1: state A):   pinned:        229376
  BTRFS info (device dm-1: state A):   reserved:      0
  BTRFS info (device dm-1: state A):   may_use:       16269312
  BTRFS info (device dm-1: state A):   read_only:     65536
---
  fs/btrfs/space-info.c | 14 +++++++++-----
  1 file changed, 9 insertions(+), 5 deletions(-)
fs/btrfs/space-info.c +0 -4
index 36b466525318..623fa0488545 100644
@@ -475,11 +475,15 @@ static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
  		   flag_str,
  		   (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
  		   info->full ? "" : "not ");
-	btrfs_info(fs_info,
-		"space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
-		info->total_bytes, info->bytes_used, info->bytes_pinned,
-		info->bytes_reserved, info->bytes_may_use,

fs/btrfs/space-info.c static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, · line 486

Quoted 10 lines
fs/btrfs/space-info.c +9 -1
-		info->bytes_readonly, info->bytes_zone_unusable);
+	btrfs_info(fs_info, "  total:         %llu", info->total_bytes);
+	btrfs_info(fs_info, "  used:          %llu", info->bytes_used);
+	btrfs_info(fs_info, "  pinned:        %llu", info->bytes_pinned);
+	btrfs_info(fs_info, "  reserved:      %llu", info->bytes_reserved);
+	btrfs_info(fs_info, "  may_use:       %llu", info->bytes_may_use);
+	btrfs_info(fs_info, "  read_only:     %llu", info->bytes_readonly);
+	if (btrfs_is_zoned(fs_info))
+		btrfs_info(fs_info,
+			    "  zone_unusable: %llu", info->bytes_zone_unusable);

I'm (perhaps needlessly) worried about splitting this up into six/seven messages, because of the ratelimiting rolled into btrfs_printk. The ratelimit is 100 messages per 5 * HZ, and it seems like it would be unfortunate if it kicked in during the middle of this dump and prevented later info from being dumped.

Maybe we should add a btrfs_dump_printk() helper that doesn't have a ratelimit built in, for exceptional cases like this where we really, really don't want anything ratelimited?

Quoted 3 lines
fs/btrfs/space-info.c +0 -0
  	DUMP_BLOCK_RSV(fs_info, global_block_rsv);
  	DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
#5 lore raw

Re: [PATCH v2 2/4] btrfs: make __btrfs_dump_space_info() output better formatted

Quoted 31 lines · Sweet Tea Dorminy
Quoted 13 lines · Qu Wenruo
fs/btrfs/space-info.c +0 -4
index 36b466525318..623fa0488545 100644
@@ -475,11 +475,15 @@ static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
  		   flag_str,
  		   (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
  		   info->full ? "" : "not ");
-	btrfs_info(fs_info,
-		"space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
-		info->total_bytes, info->bytes_used, info->bytes_pinned,
-		info->bytes_reserved, info->bytes_may_use,

fs/btrfs/space-info.c static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, · line 486

Quoted 10 lines
fs/btrfs/space-info.c +9 -1
-		info->bytes_readonly, info->bytes_zone_unusable);
+	btrfs_info(fs_info, "  total:         %llu", info->total_bytes);
+	btrfs_info(fs_info, "  used:          %llu", info->bytes_used);
+	btrfs_info(fs_info, "  pinned:        %llu", info->bytes_pinned);
+	btrfs_info(fs_info, "  reserved:      %llu", info->bytes_reserved);
+	btrfs_info(fs_info, "  may_use:       %llu", info->bytes_may_use);
+	btrfs_info(fs_info, "  read_only:     %llu", info->bytes_readonly);
+	if (btrfs_is_zoned(fs_info))
+		btrfs_info(fs_info,
+			    "  zone_unusable: %llu", info->bytes_zone_unusable);

I'm (perhaps needlessly) worried about splitting this up into six/seven messages, because of the ratelimiting rolled into btrfs_printk. The ratelimit is 100 messages per 5 * HZ, and it seems like it would be unfortunate if it kicked in during the middle of this dump and prevented later info from being dumped.

Quoted 3 lines

Maybe we should add a btrfs_dump_printk() helper that doesn't have a ratelimit built in, for exceptional cases like this where we really, really don't want anything ratelimited?

Splitting the message is IMHO wrong thing, there are other subysystems writing to the log so the lines can become scattered or interleaved with the same message from other threads.

We should prefer single line messages if possible, if not only for better grep-ability, pretty printing can be done by any utility that parses the logs.

I did not think about the rate limiting, but you're right that it could be problematic.

#6 lore raw

Re: [PATCH v2 2/4] btrfs: make __btrfs_dump_space_info() output better formatted

Quoted 32 lines · David Sterba
Quoted 31 lines · Sweet Tea Dorminy
Quoted 13 lines · Qu Wenruo
fs/btrfs/space-info.c +0 -4
index 36b466525318..623fa0488545 100644
@@ -475,11 +475,15 @@ static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
   		   flag_str,
   		   (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
   		   info->full ? "" : "not ");
-	btrfs_info(fs_info,
-		"space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
-		info->total_bytes, info->bytes_used, info->bytes_pinned,
-		info->bytes_reserved, info->bytes_may_use,

fs/btrfs/space-info.c static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, · line 486

Quoted 10 lines
fs/btrfs/space-info.c +9 -1
-		info->bytes_readonly, info->bytes_zone_unusable);
+	btrfs_info(fs_info, "  total:         %llu", info->total_bytes);
+	btrfs_info(fs_info, "  used:          %llu", info->bytes_used);
+	btrfs_info(fs_info, "  pinned:        %llu", info->bytes_pinned);
+	btrfs_info(fs_info, "  reserved:      %llu", info->bytes_reserved);
+	btrfs_info(fs_info, "  may_use:       %llu", info->bytes_may_use);
+	btrfs_info(fs_info, "  read_only:     %llu", info->bytes_readonly);
+	if (btrfs_is_zoned(fs_info))
+		btrfs_info(fs_info,
+			    "  zone_unusable: %llu", info->bytes_zone_unusable);

I'm (perhaps needlessly) worried about splitting this up into six/seven messages, because of the ratelimiting rolled into btrfs_printk. The ratelimit is 100 messages per 5 * HZ, and it seems like it would be unfortunate if it kicked in during the middle of this dump and prevented later info from being dumped.

Quoted 7 lines
Quoted 3 lines

Maybe we should add a btrfs_dump_printk() helper that doesn't have a ratelimit built in, for exceptional cases like this where we really, really don't want anything ratelimited?

Splitting the message is IMHO wrong thing, there are other subysystems writing to the log so the lines can become scattered or interleaved with the same message from other threads.

But that one line output is really hard to read for human beings.

Or do you mean that, as long as it's debug info, we should not care about readability at all?

Thanks,
Qu

Quoted 7 lines

We should prefer single line messages if possible, if not only for better grep-ability, pretty printing can be done by any utility that parses the logs.

I did not think about the rate limiting, but you're right that it could be problematic.

#7 lore raw

Re: [PATCH v2 2/4] btrfs: make __btrfs_dump_space_info() output better formatted

Quoted 17 lines · Qu Wenruo
Quoted 17 lines
Quoted 16 lines

fs/btrfs/space-info.c static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, · line 486

Quoted 6 lines
fs/btrfs/space-info.c +6 -0
+	btrfs_info(fs_info, "  reserved:      %llu", info->bytes_reserved);
+	btrfs_info(fs_info, "  may_use:       %llu", info->bytes_may_use);
+	btrfs_info(fs_info, "  read_only:     %llu", info->bytes_readonly);
+	if (btrfs_is_zoned(fs_info))
+		btrfs_info(fs_info,
+			    "  zone_unusable: %llu", info->bytes_zone_unusable);

I'm (perhaps needlessly) worried about splitting this up into six/seven messages, because of the ratelimiting rolled into btrfs_printk. The ratelimit is 100 messages per 5 * HZ, and it seems like it would be unfortunate if it kicked in during the middle of this dump and prevented later info from being dumped.

Maybe we should add a btrfs_dump_printk() helper that doesn't have a ratelimit built in, for exceptional cases like this where we really, really don't want anything ratelimited?

Quoted 8 lines
Quoted 3 lines

Splitting the message is IMHO wrong thing, there are other subysystems writing to the log so the lines can become scattered or interleaved with the same message from other threads.

But that one line output is really hard to read for human beings.

Or do you mean that, as long as it's debug info, we should not care about readability at all?

Yes we shold care about readability but kernel printk output lines can be interleaved, single line is much easier to grep for and all the values are from one event. The format where it's a series of "key=value" is common and I think we're used to it from tracepoints too. There are lines that do not put "=" between keys and values we could unify that eventually.

#8 lore raw

Re: [PATCH v2 2/4] btrfs: make __btrfs_dump_space_info() output better formatted

Quoted 27 lines · David Sterba
Quoted 17 lines · Qu Wenruo
Quoted 17 lines
Quoted 16 lines

fs/btrfs/space-info.c static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, · line 486

Quoted 6 lines
fs/btrfs/space-info.c +6 -0
+	btrfs_info(fs_info, "  reserved:      %llu", info->bytes_reserved);
+	btrfs_info(fs_info, "  may_use:       %llu", info->bytes_may_use);
+	btrfs_info(fs_info, "  read_only:     %llu", info->bytes_readonly);
+	if (btrfs_is_zoned(fs_info))
+		btrfs_info(fs_info,
+			    "  zone_unusable: %llu", info->bytes_zone_unusable);

I'm (perhaps needlessly) worried about splitting this up into six/seven messages, because of the ratelimiting rolled into btrfs_printk. The ratelimit is 100 messages per 5 * HZ, and it seems like it would be unfortunate if it kicked in during the middle of this dump and prevented later info from being dumped.

Maybe we should add a btrfs_dump_printk() helper that doesn't have a ratelimit built in, for exceptional cases like this where we really, really don't want anything ratelimited?

Quoted 8 lines
Quoted 3 lines

Splitting the message is IMHO wrong thing, there are other subysystems writing to the log so the lines can become scattered or interleaved with the same message from other threads.

But that one line output is really hard to read for human beings.

Or do you mean that, as long as it's debug info, we should not care about readability at all?

Quoted 6 lines

Yes we shold care about readability but kernel printk output lines can be interleaved, single line is much easier to grep for and all the values are from one event. The format where it's a series of "key=value" is common and I think we're used to it from tracepoints too. There are lines that do not put "=" between keys and values we could unify that eventually.

Agreed that a long line is OK, and preferable to full on splitting.

What about making some btrfs printing macros that use KERN_CONT? I think that would do what Qu wants without splitting the lines or being bad for ratelimiting.

#9 lore raw

Re: [PATCH v2 2/4] btrfs: make __btrfs_dump_space_info() output better formatted

Quoted 7 lines · Boris Burkov
Quoted 6 lines

Yes we shold care about readability but kernel printk output lines can be interleaved, single line is much easier to grep for and all the values are from one event. The format where it's a series of "key=value" is common and I think we're used to it from tracepoints too. There are lines that do not put "=" between keys and values we could unify that eventually.

Quoted 5 lines

Agreed that a long line is OK, and preferable to full on splitting.

What about making some btrfs printing macros that use KERN_CONT? I think that would do what Qu wants without splitting the lines or being bad for ratelimiting.

IIRC I've read some discussions about KERN_CONT suggesting not to use it, I'll ask what's the status.

#10 lore raw

Re: [PATCH v2 2/4] btrfs: make __btrfs_dump_space_info() output better formatted

Quoted 8 lines · David Sterba
Quoted 7 lines · Boris Burkov
Quoted 6 lines

Yes we shold care about readability but kernel printk output lines can be interleaved, single line is much easier to grep for and all the values are from one event. The format where it's a series of "key=value" is common and I think we're used to it from tracepoints too. There are lines that do not put "=" between keys and values we could unify that eventually.

Quoted 8 lines
Quoted 5 lines

Agreed that a long line is OK, and preferable to full on splitting.

What about making some btrfs printing macros that use KERN_CONT? I think that would do what Qu wants without splitting the lines or being bad for ratelimiting.

IIRC I've read some discussions about KERN_CONT suggesting not to use it, I'll ask what's the status.

I just saw a comment at its definition that reads:

/*
* Annotation for a "continued" line of log printout (only done after a
* line that had no enclosing \n). Only to be used by core/arch code
* during early bootup (a continued line is not SMP-safe otherwise). */
#define KERN_CONT KERN_SOH "c"

So that's not an encouraging sign. OTOH, I found some code in ext4/super.c that prints its errors with KERN_CONT here:
'ext4: super.c: Update logging style using KERN_CONT'

#11 lore raw

Re: [PATCH v2 2/4] btrfs: make __btrfs_dump_space_info() output better formatted

Quoted 27 lines · Boris Burkov
Quoted 8 lines · David Sterba
Quoted 7 lines · Boris Burkov
Quoted 6 lines

Yes we shold care about readability but kernel printk output lines can be interleaved, single line is much easier to grep for and all the values are from one event. The format where it's a series of "key=value" is common and I think we're used to it from tracepoints too. There are lines that do not put "=" between keys and values we could unify that eventually.

Quoted 8 lines
Quoted 5 lines

Agreed that a long line is OK, and preferable to full on splitting.

What about making some btrfs printing macros that use KERN_CONT? I think that would do what Qu wants without splitting the lines or being bad for ratelimiting.

IIRC I've read some discussions about KERN_CONT suggesting not to use it, I'll ask what's the status.

I just saw a comment at its definition that reads:

/*
* Annotation for a "continued" line of log printout (only done after a
* line that had no enclosing \n). Only to be used by core/arch code
* during early bootup (a continued line is not SMP-safe otherwise). */
#define KERN_CONT KERN_SOH "c"

Quoted 3 lines

So that's not an encouraging sign. OTOH, I found some code in ext4/super.c that prints its errors with KERN_CONT here:
'ext4: super.c: Update logging style using KERN_CONT

Some other log message from somewhere else could be emitted to the printk ringbuffer between the original and the continued message. In such a case, the continued message instead gets treated as its own message of loglevel default. (kernel/printk/printk.c:2173ish) Using KERN_CONT seems like it has a lot of potential for confusion, especially if the default message level has been changed to be different from the original messages' level.

#12 lore raw

Re: [PATCH v2 2/4] btrfs: make __btrfs_dump_space_info() output better formatted

Quoted 37 lines · Sweet Tea Dorminy
Quoted 30 lines · Boris Burkov
Quoted 11 lines · David Sterba
Quoted 10 lines · Boris Burkov
Quoted 7 lines

Yes we shold care about readability but kernel printk output lines can be interleaved, single line is much easier to grep for and all the values are from one event. The format where it's a series of "key=value"
is common and I think we're used to it from tracepoints too. There are lines that do not put "=" between keys and values we could unify that eventually.

Agreed that a long line is OK, and preferable to full on splitting.

Quoted 8 lines
Quoted 5 lines

What about making some btrfs printing macros that use KERN_CONT? I think
that would do what Qu wants without splitting the lines or being bad for
ratelimiting.

IIRC I've read some discussions about KERN_CONT suggesting not to use it, I'll ask what's the status.

I just saw a comment at its definition that reads:

/*
  * Annotation for a "continued" line of log printout (only done after a
  * line that had no enclosing \n). Only to be used by core/arch code
  * during early bootup (a continued line is not SMP-safe otherwise). */
#define KERN_CONT       KERN_SOH "c"

Quoted 3 lines

So that's not an encouraging sign. OTOH, I found some code in ext4/super.c that prints its errors with KERN_CONT here:
'ext4: super.c: Update logging style using KERN_CONT

Quoted 7 lines

Some other log message from somewhere else could be emitted to the printk ringbuffer between the original and the continued message. In such a case, the continued message instead gets treated as its own message of loglevel default. (kernel/printk/printk.c:2173ish) Using KERN_CONT seems like it has a lot of potential for confusion, especially if the default message level has been changed to be different from the original messages' level.

Thanks for all the discussion, it looks like the current long single line is the way to go (in fact, the space info dumping itself is still two lines, and we may want to fix it).

Although it's not that human readable, the racy nature of message output is indeed a concern.
This also means the old DUMP_BLOCK_RSV() function calls are not safe either.

But on the other hand, what if we output one line with multiple '\n'? Would it keep things readable while still count as one single line?

If so, I'd prefer multiple line split.

Thanks,
Qu

#13 lore raw

Re: [PATCH v2 2/4] btrfs: make __btrfs_dump_space_info() output better formatted

Quoted 52 lines · Qu Wenruo
Quoted 37 lines · Sweet Tea Dorminy
Quoted 30 lines · Boris Burkov
Quoted 11 lines · David Sterba
Quoted 10 lines · Boris Burkov
Quoted 7 lines

Yes we shold care about readability but kernel printk output lines can be interleaved, single line is much easier to grep for and all the values are from one event. The format where it's a series of "key=value"
is common and I think we're used to it from tracepoints too. There are lines that do not put "=" between keys and values we could unify that eventually.

Agreed that a long line is OK, and preferable to full on splitting.

Quoted 8 lines
Quoted 5 lines

What about making some btrfs printing macros that use KERN_CONT? I think
that would do what Qu wants without splitting the lines or being bad for
ratelimiting.

IIRC I've read some discussions about KERN_CONT suggesting not to use it, I'll ask what's the status.

I just saw a comment at its definition that reads:

/*
  * Annotation for a "continued" line of log printout (only done after a
  * line that had no enclosing \n). Only to be used by core/arch code
  * during early bootup (a continued line is not SMP-safe otherwise). */
#define KERN_CONT       KERN_SOH "c"

Quoted 3 lines

So that's not an encouraging sign. OTOH, I found some code in ext4/super.c that prints its errors with KERN_CONT here:
'ext4: super.c: Update logging style using KERN_CONT

Quoted 7 lines

Some other log message from somewhere else could be emitted to the printk ringbuffer between the original and the continued message. In such a case, the continued message instead gets treated as its own message of loglevel default. (kernel/printk/printk.c:2173ish) Using KERN_CONT seems like it has a lot of potential for confusion, especially if the default message level has been changed to be different from the original messages' level.

Thanks for all the discussion, it looks like the current long single line is the way to go (in fact, the space info dumping itself is still two lines, and we may want to fix it).

Quoted 7 lines

Although it's not that human readable, the racy nature of message output is indeed a concern.
This also means the old DUMP_BLOCK_RSV() function calls are not safe either.

But on the other hand, what if we output one line with multiple '\n'? Would it keep things readable while still count as one single line?

Yes, it would count as only one message, but it would break scripts that assume dmesg etc outputs a timestamp for every line; I think many kernel log scrapers make that assumption.

sample dmesg -T

[Wed Jul 27 10:59:54 2022] test three
                           line message
[Wed Jul 27 11:03:50 2022] next message.

sample journalctl:

Jul 27 10:59:54 vmcentral unknown: test three
                           line message
Jul 27 11:03:50 vmcentral unknown: next message.
author
#14 lore raw

PATCH v2 3/4 btrfs: make DUMP_BLOCK_RSV() to have better output

In reply to Qu Wenruo

This involves the following change:

- Remove "_block_rsv" suffix

- Better alignment for the numbers

- Better naming distinguish for delayed refs and delayed inode

- Skip "size" and "reserved" output
Just output the numbers directly

Before:

BTRFS info (device dm-1: state A): global_block_rsv: size 3670016 reserved 3653632 BTRFS info (device dm-1: state A): trans_block_rsv: size 0 reserved 0 BTRFS info (device dm-1: state A): chunk_block_rsv: size 0 reserved 0 BTRFS info (device dm-1: state A): delayed_block_rsv: size 0 reserved 0 BTRFS info (device dm-1: state A): delayed_refs_rsv: size 524288 reserved 360448

After:

BTRFS info (device dm-1: state A): dumping metadata reservation: (reserved/size)

 BTRFS info (device dm-1: state A):   global:          (3670016/3670016)
 BTRFS info (device dm-1: state A):   trans:           (0/0)
 BTRFS info (device dm-1: state A):   chunk:           (0/0)
 BTRFS info (device dm-1: state A):   delayed_inode:   (0/0)
 BTRFS info (device dm-1: state A):   delayed_refs:    (524288/524288)

Also, this patch will separate the metadata reservation dumping into a separate function for later usage.

---
 fs/btrfs/space-info.c | 25 +++++++++++++++----------
 1 file changed, 15 insertions(+), 10 deletions(-)
fs/btrfs/space-info.c +15 -10
index 623fa0488545..81457049816e 100644
@@ -441,12 +441,12 @@ void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
 	}
 }
 
-#define DUMP_BLOCK_RSV(fs_info, rsv_name)				\
+#define DUMP_BLOCK_RSV(fs_info, member, name)				\
 do {									\
-	struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;		\
+	struct btrfs_block_rsv *__rsv = &(fs_info)->member;		\
 	spin_lock(&__rsv->lock);					\
-	btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",	\
-		   __rsv->size, __rsv->reserved);			\
+	btrfs_info(fs_info, "  %-16s (%llu/%llu)",			\
+		   name, __rsv->reserved, __rsv->size);			\
 	spin_unlock(&__rsv->lock);					\
 } while (0)
 
@@ -484,13 +484,16 @@ static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 	if (btrfs_is_zoned(fs_info))
 		btrfs_info(fs_info,
 			    "  zone_unusable: %llu", info->bytes_zone_unusable);
+}
 
-	DUMP_BLOCK_RSV(fs_info, global_block_rsv);
-	DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
-	DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
-	DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
-	DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
-
+static void dump_metadata_rsv(struct btrfs_fs_info *fs_info)
+{
+	btrfs_info(fs_info, "dumping metadata reservation: (reserved/size)");
+	DUMP_BLOCK_RSV(fs_info, global_block_rsv, "global:");
+	DUMP_BLOCK_RSV(fs_info, trans_block_rsv, "trans:");
+	DUMP_BLOCK_RSV(fs_info, chunk_block_rsv, "chunk:");
+	DUMP_BLOCK_RSV(fs_info, delayed_block_rsv, "delayed_inode:");
+	DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv, "delayed_refs:");
 }
 
 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
@@ -502,6 +505,7 @@ void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 
 	spin_lock(&info->lock);
 	__btrfs_dump_space_info(fs_info, info);
+	dump_metadata_rsv(fs_info);
 	spin_unlock(&info->lock);
 
 	if (!dump_block_groups)
@@ -957,6 +961,7 @@ static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
 	if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
 		btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
 		__btrfs_dump_space_info(fs_info, space_info);
+		dump_metadata_rsv(fs_info);
 	}
 
 	while (!list_empty(&space_info->tickets) &&
Signature · 1 line
2.37.0
author
#15 lore raw

PATCH v2 4/4 btrfs: dump all space infos if we abort transaction due to ENOSPC

In reply to Qu Wenruo

We have hit some transaction abort due to -ENOSPC internally.

Normally we should always reserve enough space for metadata for every transaction, thus hitting -ENOSPC should really indicate some cases we didn't expect.

But unfrotunately current error reporting will only give a kernel wanring and backtrace, not really helpful to debug what's causing the problem.

And debug_enospc can only help when user can reproduce the problem, but under most cases, such transaction abort by -ENOSPC is really hard to reproduce.

So this patch will dump all space infos (data, metadata, system) when we abort the first transaction with -ENOSPC.

This should at least provide some clue to us.

The example of a dump would look like this:

 ------------[ cut here ]------------
 <skip stack dump>
 ---[ end trace 0000000000000000 ]---
 BTRFS info (device dm-4: state A): dumpping space info:
 BTRFS info (device dm-4: state A): space_info DATA has 8388608 free, is not full
 BTRFS info (device dm-4: state A):   total:         8388608
 BTRFS info (device dm-4: state A):   used:          0
 BTRFS info (device dm-4: state A):   pinned:        0
 BTRFS info (device dm-4: state A):   reserved:      0
 BTRFS info (device dm-4: state A):   may_use:       0
 BTRFS info (device dm-4: state A):   read_only:     0
 BTRFS info (device dm-4: state A): space_info META has 263979008 free, is not full
 BTRFS info (device dm-4: state A):   total:         268435456
 BTRFS info (device dm-4: state A):   used:          131072
 BTRFS info (device dm-4: state A):   pinned:        65536
 BTRFS info (device dm-4: state A):   reserved:      0
 BTRFS info (device dm-4: state A):   may_use:       4194304
 BTRFS info (device dm-4: state A):   read_only:     65536
 BTRFS info (device dm-4: state A): space_info SYS has 8372224 free, is not full
 BTRFS info (device dm-4: state A):   total:         8388608
 BTRFS info (device dm-4: state A):   used:          16384
 BTRFS info (device dm-4: state A):   pinned:        0
 BTRFS info (device dm-4: state A):   reserved:      0
 BTRFS info (device dm-4: state A):   may_use:       0
 BTRFS info (device dm-4: state A):   read_only:     0
 BTRFS info (device dm-4: state A): dumping metadata reservation: (reserved/size)
 BTRFS info (device dm-4: state A):   global:          (3670016/3670016)
 BTRFS info (device dm-4: state A):   trans:           (0/0)
 BTRFS info (device dm-4: state A):   chunk:           (0/0)
 BTRFS info (device dm-4: state A):   delayed_inode:   (0/0)
 BTRFS info (device dm-4: state A):   delayed_refs:    (524288/524288)
 BTRFS: error (device dm-1: state A) in cleanup_transaction:1971: errno=-28 No space left
 BTRFS info (device dm-1: state EA): forced readonly
---
 fs/btrfs/ctree.h      |  6 ++++--
 fs/btrfs/space-info.c | 14 ++++++++++++++
 fs/btrfs/space-info.h |  2 ++
 fs/btrfs/super.c      |  4 +++-
 4 files changed, 23 insertions(+), 3 deletions(-)
fs/btrfs/ctree.h +4 -2
index 4e2569f84aab..3d6fd7f6b339 100644
@@ -3739,7 +3739,7 @@ const char * __attribute_const__ btrfs_decode_error(int errno);
 __cold
 void __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
 			       const char *function,
-			       unsigned int line, int errno);
+			       unsigned int line, int errno, bool first_hit);
 
 /*
  * Call btrfs_abort_transaction as early as possible when an error condition is
@@ -3747,9 +3747,11 @@ void __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
  */
 #define btrfs_abort_transaction(trans, errno)		\
 do {								\
+	bool first = false;					\
 	/* Report first abort since mount */			\
 	if (!test_and_set_bit(BTRFS_FS_STATE_TRANS_ABORTED,	\
 			&((trans)->fs_info->fs_state))) {	\
+		first = true;					\
 		if ((errno) != -EIO && (errno) != -EROFS) {		\
 			WARN(1, KERN_DEBUG				\
 			"BTRFS: Transaction aborted (error %d)\n",	\
@@ -3761,7 +3763,7 @@ do {								\
 		}						\
 	}							\
 	__btrfs_abort_transaction((trans), __func__,		\
-				  __LINE__, (errno));		\
+				  __LINE__, (errno), first);	\
 } while (0)
 
 #ifdef CONFIG_PRINTK_INDEX
fs/btrfs/space-info.c +14 -0
index 81457049816e..af2b3f5ef2b0 100644
@@ -1717,3 +1717,17 @@ int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
 	}
 	return ret;
 }
+
+/* Dump all the space infos when we abort a transaction due to ENOSPC. */
+__cold void btrfs_dump_fs_space_info(struct btrfs_fs_info *fs_info)
+{
+	struct btrfs_space_info *space_info;
+
+	btrfs_info(fs_info, "dumping space info:");
+	list_for_each_entry(space_info, &fs_info->space_info, list) {
+		spin_lock(&space_info->lock);
+		__btrfs_dump_space_info(fs_info, space_info);
+		spin_unlock(&space_info->lock);
+	}
+	dump_metadata_rsv(fs_info);
+}
fs/btrfs/space-info.h +2 -0
index e7de24a529cf..01287a7a22a4 100644
@@ -157,4 +157,6 @@ static inline void btrfs_space_info_free_bytes_may_use(
 }
 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
 			     enum btrfs_reserve_flush_enum flush);
+void btrfs_dump_fs_space_info(struct btrfs_fs_info *fs_info);
+
 #endif /* BTRFS_SPACE_INFO_H */
fs/btrfs/super.c +3 -1
index 4c7089b1681b..f6bc8aa00f44 100644
@@ -346,12 +346,14 @@ void __cold btrfs_err_32bit_limit(struct btrfs_fs_info *fs_info)
 __cold
 void __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
 			       const char *function,
-			       unsigned int line, int errno)
+			       unsigned int line, int errno, bool first_hit)
 {
 	struct btrfs_fs_info *fs_info = trans->fs_info;
 
 	WRITE_ONCE(trans->aborted, errno);
 	WRITE_ONCE(trans->transaction->aborted, errno);
+	if (first_hit && errno == -ENOSPC)
+		btrfs_dump_fs_space_info(fs_info);
 	/* Wake up anybody who may be waiting on this transaction */
 	wake_up(&fs_info->transaction_wait);
 	wake_up(&fs_info->transaction_blocked_wait);
Signature · 1 line
2.37.0
#16 lore raw

Re: [PATCH v2 4/4] btrfs: dump all space infos if we abort transaction due to ENOSPC

Quoted 17 lines · Qu Wenruo

We have hit some transaction abort due to -ENOSPC internally.

Normally we should always reserve enough space for metadata for every transaction, thus hitting -ENOSPC should really indicate some cases we didn't expect.

But unfrotunately current error reporting will only give a kernel wanring and backtrace, not really helpful to debug what's causing the problem.

And debug_enospc can only help when user can reproduce the problem, but under most cases, such transaction abort by -ENOSPC is really hard to reproduce.

So this patch will dump all space infos (data, metadata, system) when we abort the first transaction with -ENOSPC.

Quoted 8 lines

This should at least provide some clue to us.

The example of a dump would look like this:

  ------------[ cut here ]------------
  <skip stack dump>
  ---[ end trace 0000000000000000 ]---
  BTRFS info (device dm-4: state A): dumpping space info >   BTRFS info (device dm-4: state A): space_info DATA has 8388608 

free, is not full

Quoted 84 lines
  BTRFS info (device dm-4: state A):   total:         8388608
  BTRFS info (device dm-4: state A):   used:          0
  BTRFS info (device dm-4: state A):   pinned:        0
  BTRFS info (device dm-4: state A):   reserved:      0
  BTRFS info (device dm-4: state A):   may_use:       0
  BTRFS info (device dm-4: state A):   read_only:     0
  BTRFS info (device dm-4: state A): space_info META has 263979008 free, is not full
  BTRFS info (device dm-4: state A):   total:         268435456
  BTRFS info (device dm-4: state A):   used:          131072
  BTRFS info (device dm-4: state A):   pinned:        65536
  BTRFS info (device dm-4: state A):   reserved:      0
  BTRFS info (device dm-4: state A):   may_use:       4194304
  BTRFS info (device dm-4: state A):   read_only:     65536
  BTRFS info (device dm-4: state A): space_info SYS has 8372224 free, is not full
  BTRFS info (device dm-4: state A):   total:         8388608
  BTRFS info (device dm-4: state A):   used:          16384
  BTRFS info (device dm-4: state A):   pinned:        0
  BTRFS info (device dm-4: state A):   reserved:      0
  BTRFS info (device dm-4: state A):   may_use:       0
  BTRFS info (device dm-4: state A):   read_only:     0
  BTRFS info (device dm-4: state A): dumping metadata reservation: (reserved/size)
  BTRFS info (device dm-4: state A):   global:          (3670016/3670016)
  BTRFS info (device dm-4: state A):   trans:           (0/0)
  BTRFS info (device dm-4: state A):   chunk:           (0/0)
  BTRFS info (device dm-4: state A):   delayed_inode:   (0/0)
  BTRFS info (device dm-4: state A):   delayed_refs:    (524288/524288)

BTRFS: error (device dm-1: state A) in cleanup_transaction:1971: errno=-28 No space left BTRFS info (device dm-1: state EA): forced readonly

---
  fs/btrfs/ctree.h      |  6 ++++--
  fs/btrfs/space-info.c | 14 ++++++++++++++
  fs/btrfs/space-info.h |  2 ++
  fs/btrfs/super.c      |  4 +++-
  4 files changed, 23 insertions(+), 3 deletions(-)
fs/btrfs/ctree.h +4 -2
index 4e2569f84aab..3d6fd7f6b339 100644
@@ -3739,7 +3739,7 @@ const char * __attribute_const__ btrfs_decode_error(int errno);
  __cold
  void __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
  			       const char *function,
-			       unsigned int line, int errno);
+			       unsigned int line, int errno, bool first_hit);
 
  /*
   * Call btrfs_abort_transaction as early as possible when an error condition is
@@ -3747,9 +3747,11 @@ void __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
   */
  #define btrfs_abort_transaction(trans, errno)		\
  do {								\
+	bool first = false;					\
  	/* Report first abort since mount */			\
  	if (!test_and_set_bit(BTRFS_FS_STATE_TRANS_ABORTED,	\
  			&((trans)->fs_info->fs_state))) {	\
+		first = true;					\
  		if ((errno) != -EIO && (errno) != -EROFS) {		\
  			WARN(1, KERN_DEBUG				\
  			"BTRFS: Transaction aborted (error %d)\n",	\
@@ -3761,7 +3763,7 @@ do {								\
  		}						\
  	}							\
  	__btrfs_abort_transaction((trans), __func__,		\
-				  __LINE__, (errno));		\
+				  __LINE__, (errno), first);	\
  } while (0)
 
  #ifdef CONFIG_PRINTK_INDEX
fs/btrfs/space-info.c +4 -0
index 81457049816e..af2b3f5ef2b0 100644
@@ -1717,3 +1717,17 @@ int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
  	}
  	return ret;
  }
+
+/* Dump all the space infos when we abort a transaction due to ENOSPC. */
+__cold void btrfs_dump_fs_space_info(struct btrfs_fs_info *fs_info)
+{

fs/btrfs/space-info.c int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes, · line 1733

Quoted 10 lines
fs/btrfs/space-info.c +10 -0
+	struct btrfs_space_info *space_info;
+
+	btrfs_info(fs_info, "dumping space info:");
+	list_for_each_entry(space_info, &fs_info->space_info, list) {
+		spin_lock(&space_info->lock);
+		__btrfs_dump_space_info(fs_info, space_info);
+		spin_unlock(&space_info->lock);
+	}
+	dump_metadata_rsv(fs_info);
+}

This function looks similar to btrfs_dump_space_info(), and the name and callsite doesn't help distinguish it very much to me. It seems potentially useful to print all the space_infos when one space_info encounters a problem, and it seems potentially useful to print the block group infos when we're dumping all the space infos already, so maybe the two functions could be combined.

Maybe you could adjust btrfs_dump_space_info() to print all the space infos, starting with the one passed in (potentially NULL), and call it instead of this new function?

Quoted 21 lines
fs/btrfs/space-info.h +2 -0
index e7de24a529cf..01287a7a22a4 100644
@@ -157,4 +157,6 @@ static inline void btrfs_space_info_free_bytes_may_use(
  }
  int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
  			     enum btrfs_reserve_flush_enum flush);
+void btrfs_dump_fs_space_info(struct btrfs_fs_info *fs_info);
+
  #endif /* BTRFS_SPACE_INFO_H */
fs/btrfs/super.c +1 -1
index 4c7089b1681b..f6bc8aa00f44 100644
@@ -346,12 +346,14 @@ void __cold btrfs_err_32bit_limit(struct btrfs_fs_info *fs_info)
  __cold
  void __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
  			       const char *function,
-			       unsigned int line, int errno)
+			       unsigned int line, int errno, bool first_hit)

fs/btrfs/super.c void __cold btrfs_err_32bit_limit(struct btrfs_fs_info *fs_info) · line 359

Quoted 10 lines
fs/btrfs/super.c +2 -0
  {
  	struct btrfs_fs_info *fs_info = trans->fs_info;
 
  	WRITE_ONCE(trans->aborted, errno);
  	WRITE_ONCE(trans->transaction->aborted, errno);
+	if (first_hit && errno == -ENOSPC)
+		btrfs_dump_fs_space_info(fs_info);
  	/* Wake up anybody who may be waiting on this transaction */
  	wake_up(&fs_info->transaction_wait);
  	wake_up(&fs_info->transaction_blocked_wait);

DO_ONCE_LITE(btrfs_dump_fs_space_info, fs_info) from <linux/once_lite.h> seems like a more lightweight way to dump the space infos once upon first transaction abort. Then you don't have to plumb through the 'first_hit' parameter from btrfs_abort_transaction(), and this change becomes even more minimal than it already is.

#17 lore raw

Re: [PATCH v2 4/4] btrfs: dump all space infos if we abort transaction due to ENOSPC

Quoted 127 lines · Sweet Tea Dorminy
Quoted 19 lines · Qu Wenruo

We have hit some transaction abort due to -ENOSPC internally.

Normally we should always reserve enough space for metadata for every transaction, thus hitting -ENOSPC should really indicate some cases we didn't expect.

But unfrotunately current error reporting will only give a kernel wanring and backtrace, not really helpful to debug what's causing the problem.

And debug_enospc can only help when user can reproduce the problem, but under most cases, such transaction abort by -ENOSPC is really hard to reproduce.

So this patch will dump all space infos (data, metadata, system) when we abort the first transaction with -ENOSPC.

This should at least provide some clue to us.

Quoted 7 lines

The example of a dump would look like this:

  ------------[ cut here ]------------
  <skip stack dump>
  ---[ end trace 0000000000000000 ]---

  BTRFS info (device dm-4: state A): dumpping space info >   BTRFS info (device dm-4: state A): space_info DATA has 8388608

free, is not full

Quoted 97 lines

  BTRFS info (device dm-4: state A):   total:         8388608 BTRFS info (device dm-4: state A):   used:          0
  BTRFS info (device dm-4: state A):   pinned:        0
  BTRFS info (device dm-4: state A):   reserved:      0
  BTRFS info (device dm-4: state A):   may_use:       0
  BTRFS info (device dm-4: state A):   read_only:     0
  BTRFS info (device dm-4: state A): space_info META has 263979008 free, is not full
  BTRFS info (device dm-4: state A):   total:         268435456 BTRFS info (device dm-4: state A):   used:          131072 BTRFS info (device dm-4: state A):   pinned:        65536 BTRFS info (device dm-4: state A):   reserved:      0
  BTRFS info (device dm-4: state A):   may_use:       4194304 BTRFS info (device dm-4: state A):   read_only:     65536 BTRFS info (device dm-4: state A): space_info SYS has 8372224 free, is not full
  BTRFS info (device dm-4: state A):   total:         8388608 BTRFS info (device dm-4: state A):   used:          16384 BTRFS info (device dm-4: state A):   pinned:        0
  BTRFS info (device dm-4: state A):   reserved:      0
  BTRFS info (device dm-4: state A):   may_use:       0
  BTRFS info (device dm-4: state A):   read_only:     0
  BTRFS info (device dm-4: state A): dumping metadata reservation:
(reserved/size)
  BTRFS info (device dm-4: state A):   global:          (3670016/3670016) BTRFS info (device dm-4: state A):   trans:           (0/0) BTRFS info (device dm-4: state A):   chunk:           (0/0) BTRFS info (device dm-4: state A):   delayed_inode:   (0/0) BTRFS info (device dm-4: state A):   delayed_refs:    (524288/524288)
  BTRFS: error (device dm-1: state A) in cleanup_transaction:1971:
errno=-28 No space left
  BTRFS info (device dm-1: state EA): forced readonly

---
  fs/btrfs/ctree.h      |  6 ++++--
  fs/btrfs/space-info.c | 14 ++++++++++++++
  fs/btrfs/space-info.h |  2 ++
  fs/btrfs/super.c      |  4 +++-

  4 files changed, 23 insertions(+), 3 deletions(-)
fs/btrfs/ctree.h +0 -0
index 4e2569f84aab..3d6fd7f6b339 100644
@@ -3739,7 +3739,7 @@ const char * __attribute_const__
 btrfs_decode_error(int errno);

  __cold
  void __btrfs_abort_transaction(struct btrfs_trans_handle *trans, const char *function,
-                   unsigned int line, int errno);
+                   unsigned int line, int errno, bool first_hit); /*
   * Call btrfs_abort_transaction as early as possible when an error condition is

fs/btrfs/ctree.h +2 -0
@@ -3747,9 +3747,11 @@ void __btrfs_abort_transaction(struct
 btrfs_trans_handle *trans,
    */
   #define btrfs_abort_transaction(trans, errno)        \
   do {                                \
+    bool first = false;                    \
       /* Report first abort since mount */            \
       if (!test_and_set_bit(BTRFS_FS_STATE_TRANS_ABORTED,    \
               &((trans)->fs_info->fs_state))) {    \
+        first = true;                    \
           if ((errno) != -EIO && (errno) != -EROFS) {        \
               WARN(1, KERN_DEBUG                \

              "BTRFS: Transaction aborted (error %d)\n",    \

fs/btrfs/ctree.h +1 -1
@@ -3761,7 +3763,7 @@ do {                                \
           }                        \
       }                            \
       __btrfs_abort_transaction((trans), __func__,        \
-                  __LINE__, (errno));        \
+                  __LINE__, (errno), first);    \
   } while (0)
   #ifdef CONFIG_PRINTK_INDEX
fs/btrfs/space-info.c +0 -0
index 81457049816e..af2b3f5ef2b0 100644
@@ -1717,3 +1717,17 @@ int btrfs_reserve_data_bytes(struct
 btrfs_fs_info *fs_info, u64 bytes,
       }
       return ret;

  }

fs/btrfs/space-info.c +2 -0
+
+/* Dump all the space infos when we abort a transaction due to

ENOSPC. */

fs/btrfs/space-info.c +8 -0
+__cold void btrfs_dump_fs_space_info(struct btrfs_fs_info *fs_info)
+{
+    struct btrfs_space_info *space_info;
+
+    btrfs_info(fs_info, "dumping space info:");
+    list_for_each_entry(space_info, &fs_info->space_info, list) {
+        spin_lock(&space_info->lock);
+        __btrfs_dump_space_info(fs_info, space_info);
Quoted 8 lines
Quoted 4 lines

+        spin_unlock(&space_info->lock);
+    }
+    dump_metadata_rsv(fs_info);
+}

This function looks similar to btrfs_dump_space_info(), and the name and callsite doesn't help distinguish it very much to me. It seems potentially useful to print all the space_infos when one space_info encounters a problem,

This is fine for trans abort dump, but may not be a good idea for enospc_debug output.

enospc_debug can be triggered way more frequent than trans abort, and the extra info of unrelated space info may just be noise.

Quoted 3 lines

and it seems potentially useful to print the block
group infos when we're dumping all the space infos already, so maybe the two functions could be combined.

You mean block group infos? That can be very large for large fses. Thus it's avoided for most call sites other than btrfs_reserve_extent().

Quoted 34 lines

Maybe you could adjust btrfs_dump_space_info() to print all the space infos, starting with the one passed in (potentially NULL), and call it instead of this new function?

Quoted 29 lines
fs/btrfs/space-info.h +0 -0
index e7de24a529cf..01287a7a22a4 100644
@@ -157,4 +157,6 @@ static inline void

btrfs_space_info_free_bytes_may_use(
  }
  int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes, enum btrfs_reserve_flush_enum flush);

fs/btrfs/space-info.h +2 -0
+void btrfs_dump_fs_space_info(struct btrfs_fs_info *fs_info);
+

  #endif /* BTRFS_SPACE_INFO_H */

fs/btrfs/super.c +0 -0
index 4c7089b1681b..f6bc8aa00f44 100644
@@ -346,12 +346,14 @@ void __cold btrfs_err_32bit_limit(struct
 btrfs_fs_info *fs_info)

  __cold
  void __btrfs_abort_transaction(struct btrfs_trans_handle *trans, const char *function,
-                   unsigned int line, int errno)
+                   unsigned int line, int errno, bool first_hit) {
      struct btrfs_fs_info *fs_info = trans->fs_info; WRITE_ONCE(trans->aborted, errno); WRITE_ONCE(trans->transaction->aborted, errno);
+    if (first_hit && errno == -ENOSPC)
+        btrfs_dump_fs_space_info(fs_info);

Quoted 8 lines
Quoted 3 lines

      /* Wake up anybody who may be waiting on this transaction */ wake_up(&fs_info->transaction_wait); wake_up(&fs_info->transaction_blocked_wait);

DO_ONCE_LITE(btrfs_dump_fs_space_info, fs_info) from <linux/once_lite.h> seems like a more lightweight way to dump the space infos once upon first transaction abort. Then you don't have to plumb through the 'first_hit' parameter from btrfs_abort_transaction(), and this change becomes even more minimal than it already is.

Sounds pretty awesome!

Thanks,
Qu

#18 lore raw

Re: [PATCH v2 4/4] btrfs: dump all space infos if we abort transaction due to ENOSPC

Quoted 10 lines
Quoted 6 lines
Quoted 2 lines

+}

This function looks similar to btrfs_dump_space_info(), and the name and callsite doesn't help distinguish it very much to me. It seems potentially useful to print all the space_infos when one space_info encounters a problem,

This is fine for trans abort dump, but may not be a good idea for enospc_debug output.

Quoted 8 lines

enospc_debug can be triggered way more frequent than trans abort, and the extra info of unrelated space info may just be noise.

Quoted 3 lines

and it seems potentially useful to print the block
group infos when we're dumping all the space infos already, so maybe the two functions could be combined.

You mean block group infos? That can be very large for large fses. Thus it's avoided for most call sites other than btrfs_reserve_extent().

I appreciate the explanation, I now agree it makes sense to keep the dump-all function separate.
Maybe you could move the new dump-all function up beside the dump-one function, and name it btrfs_dump_all_space_infos()? I still find the name btrfs_dump_fs_space_info() a little hard to distinguish from the others, and prefer 'all' being in the function name, but it's only a mild preference.

author
#19 lore raw

Re: [PATCH v2 4/4] btrfs: dump all space infos if we abort transaction due to ENOSPC

Quoted 12 lines · Sweet Tea Dorminy
Quoted 12 lines
Quoted 6 lines
Quoted 2 lines

+}

This function looks similar to btrfs_dump_space_info(), and the name and callsite doesn't help distinguish it very much to me. It seems potentially useful to print all the space_infos when one space_info encounters a problem,

This is fine for trans abort dump, but may not be a good idea for enospc_debug output.

enospc_debug can be triggered way more frequent than trans abort, and the extra info of unrelated space info may just be noise.

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

and it seems potentially useful to print the block
group infos when we're dumping all the space infos already, so maybe the two functions could be combined.

You mean block group infos? That can be very large for large fses. Thus it's avoided for most call sites other than btrfs_reserve_extent().

I appreciate the explanation, I now agree it makes sense to keep the dump-all function separate.

Although I'm considering extra dumping, not the level of each block groups, but each device (including used, unallocated, and the max hole size).

That info can be important for metadata over-reserve cases, but that can be done later.

Quoted 1 line

Maybe you could move the new dump-all function up beside the dump-one function, and name it btrfs_dump_all_space_infos()?

That indeed looks like a much better name, would go that solution, after determining if the split lines are to stay or not.

Thanks,
Qu

Quoted 1 line

I still find the name btrfs_dump_fs_space_info() a little hard to distinguish from the others, and prefer 'all' being in the function name, but it's only a mild preference.

#20 lore raw

Re: [PATCH v2 4/4] btrfs: dump all space infos if we abort transaction due to ENOSPC

In reply to Qu Wenruo
Quoted 22 lines · Qu Wenruo
Quoted 13 lines
Quoted 13 lines
(quoted diff) +0 -0
@@ -346,12 +346,14 @@ void __cold btrfs_err_32bit_limit(struct
 btrfs_fs_info *fs_info)

  __cold
  void __btrfs_abort_transaction(struct btrfs_trans_handle *trans, const char *function,
-                   unsigned int line, int errno)
+                   unsigned int line, int errno, bool first_hit) {
      struct btrfs_fs_info *fs_info = trans->fs_info; WRITE_ONCE(trans->aborted, errno); WRITE_ONCE(trans->transaction->aborted, errno);
+    if (first_hit && errno == -ENOSPC)
+        btrfs_dump_fs_space_info(fs_info);

Quoted 8 lines
Quoted 3 lines

      /* Wake up anybody who may be waiting on this transaction */ wake_up(&fs_info->transaction_wait); wake_up(&fs_info->transaction_blocked_wait);

DO_ONCE_LITE(btrfs_dump_fs_space_info, fs_info) from <linux/once_lite.h> seems like a more lightweight way to dump the space infos once upon first transaction abort. Then you don't have to plumb through the 'first_hit' parameter from btrfs_abort_transaction(), and this change becomes even more minimal than it already is.

Quoted 1 line

Sounds pretty awesome!

But DO_ONCE_LITE stores the status in one static variable, this cant' be used because we want to track the status per filesystem, and also per mount. Ie. repeated ENOSPC after umount/mount cycle won't be reported, also another filesystem hitting abort due to ENOSPC.

The first_hit logic as you have it now that passes the first transaction abort to the report is correct.

#21 lore raw

Re: [PATCH v2 4/4] btrfs: dump all space infos if we abort transaction due to ENOSPC

Quoted 23 lines · David Sterba
Quoted 22 lines · Qu Wenruo
Quoted 21 lines
Quoted 6 lines
fs/btrfs/super.c +0 -1
@@ -346,12 +346,14 @@ void __cold btrfs_err_32bit_limit(struct
 btrfs_fs_info *fs_info)
   __cold
   void __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
                      const char *function,
-                   unsigned int line, int errno)

fs/btrfs/super.c void __cold btrfs_err_32bit_limit(struct · line 359

Quoted 10 lines
fs/btrfs/super.c +3 -0
+                   unsigned int line, int errno, bool first_hit)
   {
       struct btrfs_fs_info *fs_info = trans->fs_info;
       WRITE_ONCE(trans->aborted, errno);
       WRITE_ONCE(trans->transaction->aborted, errno);
+    if (first_hit && errno == -ENOSPC)
+        btrfs_dump_fs_space_info(fs_info);
       /* Wake up anybody who may be waiting on this transaction */
       wake_up(&fs_info->transaction_wait);
       wake_up(&fs_info->transaction_blocked_wait);

DO_ONCE_LITE(btrfs_dump_fs_space_info, fs_info) from <linux/once_lite.h> seems like a more lightweight way to dump the space infos once upon first transaction abort. Then you don't have to plumb through the 'first_hit' parameter from btrfs_abort_transaction(), and this change becomes even more minimal than it already is.

Quoted 6 lines
Quoted 1 line

Sounds pretty awesome!

But DO_ONCE_LITE stores the status in one static variable, this cant' be used because we want to track the status per filesystem, and also per mount. Ie. repeated ENOSPC after umount/mount cycle won't be reported, also another filesystem hitting abort due to ENOSPC.

That's a great point, I agree it would be better to print the first abort per filesystem/mount instead of only once per reboot now that it's noted.

#22 lore raw
In reply to Qu Wenruo
Quoted 5 lines · Qu Wenruo

[Changelog]
v2:
- Add a new line to show the meaning of the metadata dump. Previous output only includes the reserved bytes and size bytes, but not showing which is which, thus still need to check the code

There are some nontrivial pending changes for this patchset, so I can't merge it to for-next. Please send v3.

#23 lore raw
Quoted 9 lines · David Sterba
Quoted 5 lines · Qu Wenruo

[Changelog]
v2:
- Add a new line to show the meaning of the metadata dump. Previous output only includes the reserved bytes and size bytes, but not showing which is which, thus still need to check the code

There are some nontrivial pending changes for this patchset, so I can't merge it to for-next. Please send v3.

Thanks for reminding me of this series, I'll update them to address the comments on the multi-line output format (aka, drop all the re-format patches) and refresh them soon.

Thanks,
Qu