Skip to contentSkip to conversation

btrfs: make ticket wait uninterruptible to address unexpected RO during balance

RFC PATCH Superseded Qu Wenruo 5 participants 12 messages Jul 6, 2020 – Jul 7, 2020 Btrfs file system

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Qu Wenruo

author
#0 lore raw

cover btrfs: make ticket wait uninterruptible to address unexpected RO during balance

There is a report that, unlucky signal timing during balance can cause btrfs to remounted into RO mode.

This is caused by the fact that, most btrfs_start_transaction() or delalloc metadata reserve are interruptible.

That would return -EINTR to a lot of critical code section, and under most case, our way to handle such error is just to abort transaction, without any consideration for -EINTR.

This is never a good idea to allow random Ctrl-C to make btrfs RO, even if the window is pretty small for regular operations.

This patchset will address it in a different direction, since most operations are pretty fast, we don't need that signal check in waiting ticket.

For those long running operations, signal should be checked in their call sites.
E.g. __generic_block_fiemap() calls fatal_signal_pending() to check if it needs to exit, so does btrfs_clone().

We shouldn't check the signal, and just throw a -EINTR for all ticketing system callers, they don't really want to handle that damn -EINTR.

Only long executing operations really need that signal checking, and let them to check, not the infrastructure.

Reason for RFC:
I'm not yet completely sure if uninterruptible ticketing system would cause extra problems.
Any advice on that would be great.

Qu Wenruo (2):
btrfs: relocation: Allow signal to cancel balance
btrfs: space-info: Don't allow signal to interrupt ticket waiting

 fs/btrfs/relocation.c | 3 ++-
 fs/btrfs/space-info.c | 3 ++-
 2 files changed, 4 insertions(+), 2 deletions(-)
Signature · 1 line
2.27.0
author
#1 lore raw

RFC PATCH 1/2 btrfs: relocation: Allow signal to cancel balance

Although btrfs balance can be canceled with "btrfs balance cancel" command, it's still almost muscle memory to press Ctrl-C to cancel a long running btrfs balance.

So allow btrfs balance to check signal to determine if it should exit.

---
 fs/btrfs/relocation.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fs/btrfs/relocation.c +2 -1
index 523d2e5fab8f..2b869fb2e62c 100644
@@ -2656,7 +2656,8 @@ int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
  */
 int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
 {
-	return atomic_read(&fs_info->balance_cancel_req);
+	return atomic_read(&fs_info->balance_cancel_req) ||
+		fatal_signal_pending(current);
 }
 ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
 
Signature · 1 line
2.27.0
Reviewed-by on v1 1/2
#2 lore raw

Re: [PATCH RFC 1/2] btrfs: relocation: Allow signal to cancel balance

Quoted 7 lines · Qu Wenruo

Although btrfs balance can be canceled with "btrfs balance cancel" command, it's still almost muscle memory to press Ctrl-C to cancel a long running btrfs balance.

So allow btrfs balance to check signal to determine if it should exit.

Thanks,

Josef

#3 lore raw

Re: [PATCH RFC 1/2] btrfs: relocation: Allow signal to cancel balance

In reply to Qu Wenruo

Hi,

Quoted 3 lines · Qu Wenruo

Although btrfs balance can be canceled with "btrfs balance cancel" command, it's still almost muscle memory to press Ctrl-C to cancel a long running btrfs balance.

Thanks for investigating all of this.

Has it actually been unsafe (read: undefined behaviour) forever, or only since some recent change?

Or did it just by accident not cause real damage earlier on in the past?

[ 1041.810963] BTRFS info (device xvdb): relocating block group

91423244288 flags metadata

<- ^C made it stop here

[ 1076.189766] BTRFS info (device xvdb): relocating block group
91423244288 flags metadata
[ 1081.300131] BTRFS info (device xvdb): found 6689 extents
[ 1081.311138] BTRFS info (device xvdb): relocating block group
90349502464 flags data
[ 1089.776066] BTRFS info (device xvdb): found 215 extents

The above is with 4.19.118. Now I'm trying this again and looking just a little better at the logging, I see that what I thought (it always stopped after doing 1 block group) is not true. With ^C I can make it stop halfway and then next time it again starts at 91423244288.

Related question: should, additionally, the btrfs balance in progs also be changed to catch the SIGINT while it's doing the balance ioctl, to prevent the signal from leaking to the kernel space? I mean, kernels with the bug are already 'out there' now...

Thanks

Quoted 22 lines

So allow btrfs balance to check signal to determine if it should exit.

---
 fs/btrfs/relocation.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fs/btrfs/relocation.c +2 -1
index 523d2e5fab8f..2b869fb2e62c 100644
@@ -2656,7 +2656,8 @@ int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
  */
 int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
 {
-	return atomic_read(&fs_info->balance_cancel_req);
+	return atomic_read(&fs_info->balance_cancel_req) ||
+		fatal_signal_pending(current);
 }
 ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
 
#4 lore raw

Re: [PATCH RFC 1/2] btrfs: relocation: Allow signal to cancel balance

Quoted 11 lines · Hans van Kranenburg

Hi,

Quoted 3 lines · Qu Wenruo

Although btrfs balance can be canceled with "btrfs balance cancel" command, it's still almost muscle memory to press Ctrl-C to cancel a long running btrfs balance.

Thanks for investigating all of this.

Has it actually been unsafe (read: undefined behaviour) forever, or only since some recent change?

Forever.

That -EINTR from metadata reservation path is there for a long long time.

Quoted 20 lines

Or did it just by accident not cause real damage earlier on in the past?

[ 1041.810963] BTRFS info (device xvdb): relocating block group

91423244288 flags metadata

<- ^C made it stop here

[ 1076.189766] BTRFS info (device xvdb): relocating block group
91423244288 flags metadata
[ 1081.300131] BTRFS info (device xvdb): found 6689 extents
[ 1081.311138] BTRFS info (device xvdb): relocating block group
90349502464 flags data
[ 1089.776066] BTRFS info (device xvdb): found 215 extents

The above is with 4.19.118. Now I'm trying this again and looking just a little better at the logging, I see that what I thought (it always stopped after doing 1 block group) is not true. With ^C I can make it stop halfway and then next time it again starts at 91423244288.

Quoted 4 lines

Related question: should, additionally, the btrfs balance in progs also be changed to catch the SIGINT while it's doing the balance ioctl, to prevent the signal from leaking to the kernel space? I mean, kernels with the bug are already 'out there' now...

It has no way to catch signal while trapped into kernel space.

My previous assumption of the whole ioctl thing is still right, when we're in kernel space, we can't catch any signal.

It's just the metadata reservation code manually checking the pending signal and return -EINTR.

So even if we make btrfs-progs to catch that signal, it won't work. And even if it seems to work, it's because in btrfs module we're checking signal explicitly.

Thanks,
Qu

Quoted 26 lines

Thanks

Quoted 22 lines

So allow btrfs balance to check signal to determine if it should exit.

---
 fs/btrfs/relocation.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fs/btrfs/relocation.c +2 -1
index 523d2e5fab8f..2b869fb2e62c 100644
@@ -2656,7 +2656,8 @@ int setup_extent_mapping(struct inode *inode, u64 start, u64 end,
  */
 int btrfs_should_cancel_balance(struct btrfs_fs_info *fs_info)
 {
-	return atomic_read(&fs_info->balance_cancel_req);
+	return atomic_read(&fs_info->balance_cancel_req) ||
+		fatal_signal_pending(current);
 }
 ALLOW_ERROR_INJECTION(btrfs_should_cancel_balance, TRUE);
 
author
#5 lore raw

RFC PATCH 2/2 btrfs: space-info: Don't allow signal to interrupt ticket waiting

In reply to Qu Wenruo

[BUG]
When balance receive a fatal signal, it can make the fs to read-only mode if the timing is unlucky enough:

BTRFS info (device xvdb): balance: start -d -m -s
BTRFS info (device xvdb): relocating block group 73001861120 flags metadata BTRFS info (device xvdb): found 12236 extents, stage: move data extents BTRFS info (device xvdb): relocating block group 71928119296 flags data BTRFS info (device xvdb): found 3 extents, stage: move data extents BTRFS info (device xvdb): found 3 extents, stage: update data pointers BTRFS info (device xvdb): relocating block group 60922265600 flags metadata
BTRFS: error (device xvdb) in btrfs_drop_snapshot:5505: errno=-4 unknown BTRFS info (device xvdb): forced readonly
BTRFS info (device xvdb): balance: ended with status: -4

[CAUSE]
This is caused by the fact that btrfs ticketing space system can be interrupted, and cause all kind of -EINTR returned to various critical section, where we never thought of -EINTR at all.

Even for things like btrfs_start_transaction() can be affected by signal:
btrfs_start_transaction()
|- start_transaction(flush = FLUSH_ALL)

    |- btrfs_block_rsv_add()
       |- btrfs_reserve_metadata_bytes()
          |- __reserve_metadata_bytes()
             |- handle_reserve_ticket()
                |- wait_reserve_ticket()
                   |- prepare_to_wait_event(TASK_KILLABLE)
                   |- ticket->error = -EINTR;

And all related callers get -EINTR error.

In fact, there are really very limited call sites can really handle that -EINTR properly, above btrfs_drop_snapshot() is one case.

[FIX]
Things like metadata allocation is really a critical section for btrfs, we don't really want it to be that killable by some impatient users.

In fact, for really long duration calls, it should have their own checks on signal, like balance, reflink, generic fiemap calls.

So this patch will make ticket waiting uninterruptible, relying on each long duration calls to handle their signals more properly.

---
 fs/btrfs/space-info.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fs/btrfs/space-info.c +2 -1
index c7bd3fdd7792..c5cfc759b804 100644
@@ -1099,7 +1099,8 @@ static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
 
 	spin_lock(&space_info->lock);
 	while (ticket->bytes > 0 && ticket->error == 0) {
-		ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
+		ret = prepare_to_wait_event(&ticket->wait, &wait,
+					    TASK_UNINTERRUPTIBLE);
 		if (ret) {
 			/*
 			 * Delete us from the list. After we unlock the space
Signature · 1 line
2.27.0
#6 lore raw

Re: [PATCH RFC 2/2] btrfs: space-info: Don't allow signal to interrupt ticket waiting

Quoted 41 lines · Qu Wenruo

[BUG]
When balance receive a fatal signal, it can make the fs to read-only mode if the timing is unlucky enough:

BTRFS info (device xvdb): balance: start -d -m -s
BTRFS info (device xvdb): relocating block group 73001861120 flags metadata BTRFS info (device xvdb): found 12236 extents, stage: move data extents BTRFS info (device xvdb): relocating block group 71928119296 flags data BTRFS info (device xvdb): found 3 extents, stage: move data extents BTRFS info (device xvdb): found 3 extents, stage: update data pointers BTRFS info (device xvdb): relocating block group 60922265600 flags metadata
BTRFS: error (device xvdb) in btrfs_drop_snapshot:5505: errno=-4 unknown BTRFS info (device xvdb): forced readonly
BTRFS info (device xvdb): balance: ended with status: -4

[CAUSE]
This is caused by the fact that btrfs ticketing space system can be interrupted, and cause all kind of -EINTR returned to various critical section, where we never thought of -EINTR at all.

Even for things like btrfs_start_transaction() can be affected by signal:
btrfs_start_transaction()
|- start_transaction(flush = FLUSH_ALL)

     |- btrfs_block_rsv_add()
        |- btrfs_reserve_metadata_bytes()
           |- __reserve_metadata_bytes()
              |- handle_reserve_ticket()
                 |- wait_reserve_ticket()
                    |- prepare_to_wait_event(TASK_KILLABLE)
                    |- ticket->error = -EINTR;

And all related callers get -EINTR error.

In fact, there are really very limited call sites can really handle that -EINTR properly, above btrfs_drop_snapshot() is one case.

[FIX]
Things like metadata allocation is really a critical section for btrfs, we don't really want it to be that killable by some impatient users.

Quoted 6 lines

In fact, for really long duration calls, it should have their own checks on signal, like balance, reflink, generic fiemap calls.

So this patch will make ticket waiting uninterruptible, relying on each long duration calls to handle their signals more properly.

Nope, everybody that calls start_transaction() should be able to handle -EINTR, so we need to find those callsites and fix them, not make it so we hang the box because we don't like fixing error paths. Thanks,

Josef

author
#7 lore raw

Re: [PATCH RFC 2/2] btrfs: space-info: Don't allow signal to interrupt ticket waiting

Quoted 50 lines · Josef Bacik
Quoted 49 lines · Qu Wenruo

[BUG]
When balance receive a fatal signal, it can make the fs to read-only mode if the timing is unlucky enough:

   BTRFS info (device xvdb): balance: start -d -m -s
   BTRFS info (device xvdb): relocating block group 73001861120 flags metadata
   BTRFS info (device xvdb): found 12236 extents, stage: move data extents
   BTRFS info (device xvdb): relocating block group 71928119296 flags data
   BTRFS info (device xvdb): found 3 extents, stage: move data extents BTRFS info (device xvdb): found 3 extents, stage: update data pointers BTRFS info (device xvdb): relocating block group 60922265600 flags metadata
   BTRFS: error (device xvdb) in btrfs_drop_snapshot:5505: errno=-4 unknown
   BTRFS info (device xvdb): forced readonly
   BTRFS info (device xvdb): balance: ended with status: -4

[CAUSE]
This is caused by the fact that btrfs ticketing space system can be interrupted, and cause all kind of -EINTR returned to various critical section, where we never thought of -EINTR at all.

Even for things like btrfs_start_transaction() can be affected by signal:
  btrfs_start_transaction()
  |- start_transaction(flush = FLUSH_ALL)
     |- btrfs_block_rsv_add()
        |- btrfs_reserve_metadata_bytes()
           |- __reserve_metadata_bytes()
              |- handle_reserve_ticket()
                 |- wait_reserve_ticket()
                    |- prepare_to_wait_event(TASK_KILLABLE) |- ticket->error = -EINTR;

And all related callers get -EINTR error.

In fact, there are really very limited call sites can really handle that -EINTR properly, above btrfs_drop_snapshot() is one case.

[FIX]
Things like metadata allocation is really a critical section for btrfs, we don't really want it to be that killable by some impatient users.

In fact, for really long duration calls, it should have their own checks on signal, like balance, reflink, generic fiemap calls.

Quoted 7 lines
Quoted 3 lines

So this patch will make ticket waiting uninterruptible, relying on each long duration calls to handle their signals more properly.

Nope, everybody that calls start_transaction() should be able to handle -EINTR, so we need to find those callsites and fix them, not make it so we hang the box because we don't like fixing error paths.  Thanks,

Then we also need to handle btrfs_delalloc_reserve_metadata(), btrfs_block_rsv_refill(), btrfs_use_block_rsv(), btrfs_block_rsv_add().

Are you really willing to go to that rabbit hole?

To me, there are only limited call sites would benefit from signal checking, while we need to handle tons of unnecessary possible -EINTR errors just for no obvious benefits for other calls sites?

That doesn't sound sane to me.

Thanks,
Qu

Quoted 3 lines

Josef

#8 lore raw

Re: [PATCH RFC 2/2] btrfs: space-info: Don't allow signal to interrupt ticket waiting

Quoted 57 lines · Qu Wenruo
Quoted 54 lines · Josef Bacik
Quoted 52 lines · Qu Wenruo

[BUG]
When balance receive a fatal signal, it can make the fs to read-only mode if the timing is unlucky enough:

   BTRFS info (device xvdb): balance: start -d -m -s
   BTRFS info (device xvdb): relocating block group 73001861120 flags
metadata
   BTRFS info (device xvdb): found 12236 extents, stage: move data
extents
   BTRFS info (device xvdb): relocating block group 71928119296 flags
data
   BTRFS info (device xvdb): found 3 extents, stage: move data extents BTRFS info (device xvdb): found 3 extents, stage: update data pointers BTRFS info (device xvdb): relocating block group 60922265600 flags
metadata
   BTRFS: error (device xvdb) in btrfs_drop_snapshot:5505: errno=-4
unknown
   BTRFS info (device xvdb): forced readonly
   BTRFS info (device xvdb): balance: ended with status: -4

[CAUSE]
This is caused by the fact that btrfs ticketing space system can be interrupted, and cause all kind of -EINTR returned to various critical section, where we never thought of -EINTR at all.

Even for things like btrfs_start_transaction() can be affected by signal:
  btrfs_start_transaction()
  |- start_transaction(flush = FLUSH_ALL)
     |- btrfs_block_rsv_add()
        |- btrfs_reserve_metadata_bytes()
           |- __reserve_metadata_bytes()
              |- handle_reserve_ticket()
                 |- wait_reserve_ticket()
                    |- prepare_to_wait_event(TASK_KILLABLE) |- ticket->error = -EINTR;

And all related callers get -EINTR error.

In fact, there are really very limited call sites can really handle that -EINTR properly, above btrfs_drop_snapshot() is one case.

[FIX]
Things like metadata allocation is really a critical section for btrfs, we don't really want it to be that killable by some impatient users.

In fact, for really long duration calls, it should have their own checks on signal, like balance, reflink, generic fiemap calls.

So this patch will make ticket waiting uninterruptible, relying on each long duration calls to handle their signals more properly.

Quoted 7 lines
Quoted 3 lines

Nope, everybody that calls start_transaction() should be able to handle -EINTR, so we need to find those callsites and fix them, not make it so we hang the box because we don't like fixing error paths.  Thanks,

Then we also need to handle btrfs_delalloc_reserve_metadata(), btrfs_block_rsv_refill(), btrfs_use_block_rsv(), btrfs_block_rsv_add().

This only needs to be handled for FLUSH_ALL and FLUSH_STEAL. Anybody doing btrfs_start_transaction() should be able to fail with -EINTR, because they should be close to the syscall path. Balance needs to be fixed to deal with it, and I assume there might be a few other places, but by-in-large none of these places should flip read only. Thanks,

Josef

#9 lore raw

Re: [PATCH RFC 2/2] btrfs: space-info: Don't allow signal to interrupt ticket waiting

Quoted 69 lines · Josef Bacik
Quoted 60 lines · Qu Wenruo
Quoted 57 lines · Josef Bacik
Quoted 55 lines · Qu Wenruo

[BUG]
When balance receive a fatal signal, it can make the fs to read-only mode if the timing is unlucky enough:

    BTRFS info (device xvdb): balance: start -d -m -s
    BTRFS info (device xvdb): relocating block group 73001861120 flags metadata
    BTRFS info (device xvdb): found 12236 extents, stage: move data extents
    BTRFS info (device xvdb): relocating block group 71928119296 flags data
    BTRFS info (device xvdb): found 3 extents, stage: move data extents BTRFS info (device xvdb): found 3 extents, stage: update data pointers
    BTRFS info (device xvdb): relocating block group 60922265600 flags metadata
    BTRFS: error (device xvdb) in btrfs_drop_snapshot:5505: errno=-4 unknown
    BTRFS info (device xvdb): forced readonly
    BTRFS info (device xvdb): balance: ended with status: -4

[CAUSE]
This is caused by the fact that btrfs ticketing space system can be interrupted, and cause all kind of -EINTR returned to various critical section, where we never thought of -EINTR at all.

Even for things like btrfs_start_transaction() can be affected by signal:
   btrfs_start_transaction()
   |- start_transaction(flush = FLUSH_ALL)
      |- btrfs_block_rsv_add()
         |- btrfs_reserve_metadata_bytes()
            |- __reserve_metadata_bytes()
               |- handle_reserve_ticket()
                  |- wait_reserve_ticket()
                     |- prepare_to_wait_event(TASK_KILLABLE) |- ticket->error = -EINTR;

And all related callers get -EINTR error.

In fact, there are really very limited call sites can really handle that
-EINTR properly, above btrfs_drop_snapshot() is one case.

[FIX]
Things like metadata allocation is really a critical section for btrfs, we don't really want it to be that killable by some impatient users.

In fact, for really long duration calls, it should have their own checks
on signal, like balance, reflink, generic fiemap calls.

So this patch will make ticket waiting uninterruptible, relying on each long duration calls to handle their signals more properly.

Quoted 7 lines
Quoted 3 lines

Nope, everybody that calls start_transaction() should be able to handle -EINTR, so we need to find those callsites and fix them, not make it so we hang the box because we don't like fixing error paths.  Thanks,

Then we also need to handle btrfs_delalloc_reserve_metadata(), btrfs_block_rsv_refill(), btrfs_use_block_rsv(), btrfs_block_rsv_add().

Quoted 5 lines

This only needs to be handled for FLUSH_ALL and FLUSH_STEAL.  Anybody doing btrfs_start_transaction() should be able to fail with -EINTR, because they should be close to the syscall path.  Balance needs to be fixed to deal with it, and I assume there might be a few other places, but by-in-large none of these places should flip read only.  Thanks,

There are already ones existing, for btrfs_drop_snapshot().

This is mostly caused by btrfs_delalloc_reserve_metadata(), which always use FLUSH_ALL unless there is a running trans or its space cache inode.

But still, for a lot of relocation code, we don't really want to bother the EINTR and just want uninterruptible at least for now.

Any idea for that?
Or just rework how we handle errors in a lot of places?

It doesn't look correct for such a low level mechanism to return -EINTR while most of callers doesn't really want to bother.

Thanks,
Qu

Quoted 2 lines

Josef

#10 lore raw

Re: [PATCH RFC 2/2] btrfs: space-info: Don't allow signal to interrupt ticket waiting

Quoted 83 lines · Qu Wenruo
Quoted 69 lines · Josef Bacik
Quoted 60 lines · Qu Wenruo
Quoted 57 lines · Josef Bacik
Quoted 55 lines · Qu Wenruo

[BUG]
When balance receive a fatal signal, it can make the fs to read-only mode if the timing is unlucky enough:

    BTRFS info (device xvdb): balance: start -d -m -s
    BTRFS info (device xvdb): relocating block group 73001861120 flags
metadata
    BTRFS info (device xvdb): found 12236 extents, stage: move data
extents
    BTRFS info (device xvdb): relocating block group 71928119296 flags
data
    BTRFS info (device xvdb): found 3 extents, stage: move data extents BTRFS info (device xvdb): found 3 extents, stage: update data
pointers
    BTRFS info (device xvdb): relocating block group 60922265600 flags
metadata
    BTRFS: error (device xvdb) in btrfs_drop_snapshot:5505: errno=-4
unknown
    BTRFS info (device xvdb): forced readonly
    BTRFS info (device xvdb): balance: ended with status: -4

[CAUSE]
This is caused by the fact that btrfs ticketing space system can be interrupted, and cause all kind of -EINTR returned to various critical section, where we never thought of -EINTR at all.

Even for things like btrfs_start_transaction() can be affected by signal:
   btrfs_start_transaction()
   |- start_transaction(flush = FLUSH_ALL)
      |- btrfs_block_rsv_add()
         |- btrfs_reserve_metadata_bytes()
            |- __reserve_metadata_bytes()
               |- handle_reserve_ticket()
                  |- wait_reserve_ticket()
                     |- prepare_to_wait_event(TASK_KILLABLE) |- ticket->error = -EINTR;

And all related callers get -EINTR error.

In fact, there are really very limited call sites can really handle that
-EINTR properly, above btrfs_drop_snapshot() is one case.

[FIX]
Things like metadata allocation is really a critical section for btrfs, we don't really want it to be that killable by some impatient users.

In fact, for really long duration calls, it should have their own checks
on signal, like balance, reflink, generic fiemap calls.

So this patch will make ticket waiting uninterruptible, relying on each long duration calls to handle their signals more properly.

Quoted 7 lines
Quoted 3 lines

Nope, everybody that calls start_transaction() should be able to handle -EINTR, so we need to find those callsites and fix them, not make it so we hang the box because we don't like fixing error paths.  Thanks,

Then we also need to handle btrfs_delalloc_reserve_metadata(), btrfs_block_rsv_refill(), btrfs_use_block_rsv(), btrfs_block_rsv_add().

Quoted 5 lines

This only needs to be handled for FLUSH_ALL and FLUSH_STEAL.  Anybody doing btrfs_start_transaction() should be able to fail with -EINTR, because they should be close to the syscall path.  Balance needs to be fixed to deal with it, and I assume there might be a few other places, but by-in-large none of these places should flip read only.  Thanks,

There are already ones existing, for btrfs_drop_snapshot().

This is mostly caused by btrfs_delalloc_reserve_metadata(), which always use FLUSH_ALL unless there is a running trans or its space cache inode.

Quoted 9 lines

But still, for a lot of relocation code, we don't really want to bother the EINTR and just want uninterruptible at least for now.

Any idea for that?
Or just rework how we handle errors in a lot of places?

It doesn't look correct for such a low level mechanism to return -EINTR while most of callers doesn't really want to bother.

That's the point, most callers shouldn't have to, because most callers shouldn't be far enough into their operations that -EINTR causes problems.

We should probably just change btrfs_drop_snapshot to use join, and then have it do any space reservation it needs outside of the trans handle. The other option is a FLUSH_ALL_UNKILLABLE. Thanks,

Josef

#11 lore raw

Re: [PATCH RFC 2/2] btrfs: space-info: Don't allow signal to interrupt ticket waiting

Quoted 18 lines · Josef Bacik
Quoted 17 lines · Qu Wenruo
Quoted 5 lines

This only needs to be handled for FLUSH_ALL and FLUSH_STEAL.  Anybody doing btrfs_start_transaction() should be able to fail with -EINTR, because they should be close to the syscall path.  Balance needs to be fixed to deal with it, and I assume there might be a few other places, but by-in-large none of these places should flip read only.  Thanks,

There are already ones existing, for btrfs_drop_snapshot().

This is mostly caused by btrfs_delalloc_reserve_metadata(), which always use FLUSH_ALL unless there is a running trans or its space cache inode.

But still, for a lot of relocation code, we don't really want to bother the EINTR and just want uninterruptible at least for now.

Any idea for that?
Or just rework how we handle errors in a lot of places?

Quoted 6 lines
Quoted 3 lines

It doesn't look correct for such a low level mechanism to return -EINTR while most of callers doesn't really want to bother.

That's the point, most callers shouldn't have to, because most callers shouldn't be far enough into their operations that -EINTR causes problems.

Agreed, that's a sane pattern to follow so we should convert the remaining cases, perhaps also auditing all existing btrfs_start_transaction calls but after a quick look I haven't spotted anything else than the reloc and drop snapshot.

Quoted 3 lines

We should probably just change btrfs_drop_snapshot to use join, and then have it do any space reservation it needs outside of the trans handle. The other option is a FLUSH_ALL_UNKILLABLE. Thanks,

I'd rather not push the killable semantics to the flushing state machine and let it up to the caller context to decide.