Skip to contentSkip to conversation

backup: copy_range fixes

PATCH Superseded Vladimir Sementsov-Ogievskiy 2 participants 10 messages Sep 17, 2019 – Sep 20, 2019 Block layer

v12 of 15 Superseded Latest is v15 →

lore
Share & tools
lore
#0 lore raw

v12 cover backup: copy_range fixes

Hi all!

Here are two small fixes. They fixes old commit, so qemu-stable is in CC, but actually, I don't think they are critical.

01: is new
02: is just copied from
[PATCH v11 01/14] block/backup: fix backup_cow_with_offload for last cluster (I only add Fixes: to commit-message)
and this is why I called this v12, to not interfere with previous emails

I'd prefer this to go through Max's block branch, as Max is reviewing my backup-top series, which will refer to these patches and seems simpler to queue them all together.

Vladimir Sementsov-Ogievskiy (2):
block/backup: fix max_transfer handling for copy_range block/backup: fix backup_cow_with_offload for last cluster

 block/backup.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)
Signature · 1 line
2.21.0
#1 lore raw

PATCH v12 1/2 block/backup: fix max_transfer handling for copy_range

Of course, QEMU_ALIGN_UP is a typo, it should be QEMU_ALIGN_DOWN, as we are trying to find aligned size which satisfy both source and target. Also, don't ignore too small max_transfer. In this case seems safer to disable copy_range.

Fixes 9ded4a0114968e
---
 block/backup.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)
block/backup.c +8 -4
index 763f0d7ff6..d8fdbfadfe 100644
@@ -741,12 +741,16 @@ BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
     job->cluster_size = cluster_size;
     job->copy_bitmap = copy_bitmap;
     copy_bitmap = NULL;
-    job->use_copy_range = !compress; /* compression isn't supported for it */
     job->copy_range_size = MIN_NON_ZERO(blk_get_max_transfer(job->common.blk),
                                         blk_get_max_transfer(job->target));
-    job->copy_range_size = MAX(job->cluster_size,
-                               QEMU_ALIGN_UP(job->copy_range_size,
-                                             job->cluster_size));
+    job->copy_range_size = QEMU_ALIGN_DOWN(job->copy_range_size,
+                                           job->cluster_size);
+    /*
+     * Compression is not supported for copy_range. Also, we don't want to
+     * handle small max_transfer for copy_range (which currently don't
+     * handle max_transfer at all).
+     */
+    job->use_copy_range = !compress && job->copy_range_size > 0;
 
     /* Required permissions are already taken with target's blk_new() */
     block_job_add_bdrv(&job->common, "target", target, 0, BLK_PERM_ALL,
Signature · 1 line
2.21.0
#2 lore raw

Re: [Qemu-devel] [PATCH v12 1/2] block/backup: fix max_transfer handling for copy_range

Quoted 27 lines · Vladimir Sementsov-Ogievskiy

Of course, QEMU_ALIGN_UP is a typo, it should be QEMU_ALIGN_DOWN, as we are trying to find aligned size which satisfy both source and target. Also, don't ignore too small max_transfer. In this case seems safer to disable copy_range.

Fixes 9ded4a0114968e
---
  block/backup.c | 12 ++++++++----
  1 file changed, 8 insertions(+), 4 deletions(-)
block/backup.c +1 -4
index 763f0d7ff6..d8fdbfadfe 100644
@@ -741,12 +741,16 @@ BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
      job->cluster_size = cluster_size;
      job->copy_bitmap = copy_bitmap;
      copy_bitmap = NULL;
-    job->use_copy_range = !compress; /* compression isn't supported for it */
      job->copy_range_size = MIN_NON_ZERO(blk_get_max_transfer(job->common.blk),
                                          blk_get_max_transfer(job->target));
-    job->copy_range_size = MAX(job->cluster_size,
-                               QEMU_ALIGN_UP(job->copy_range_size,
-                                             job->cluster_size));
+    job->copy_range_size = QEMU_ALIGN_DOWN(job->copy_range_size,

block/backup.c BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs, · line 756

Quoted 10 lines
block/backup.c +7 -0
+                                           job->cluster_size);
+    /*
+     * Compression is not supported for copy_range. Also, we don't want to
+     * handle small max_transfer for copy_range (which currently don't
+     * handle max_transfer at all).
+     */
+    job->use_copy_range = !compress && job->copy_range_size > 0;
 
      /* Required permissions are already taken with target's blk_new() */
      block_job_add_bdrv(&job->common, "target", target, 0, BLK_PERM_ALL,

I'm clear on the alignment fix, I'm not clear on the comment about max_transfer and how it relates to copy_range_size being non-zero.

"small max transfer" -- what happens when it's zero? we're apparently OK with a single cluster, but when it's zero, what happens?

#3 lore raw

Re: [Qemu-devel] [PATCH v12 1/2] block/backup: fix max_transfer handling for copy_range

Quoted 42 lines · John Snow
Quoted 26 lines · Vladimir Sementsov-Ogievskiy

Of course, QEMU_ALIGN_UP is a typo, it should be QEMU_ALIGN_DOWN, as we are trying to find aligned size which satisfy both source and target. Also, don't ignore too small max_transfer. In this case seems safer to disable copy_range.

Fixes 9ded4a0114968e

---
  block/backup.c | 12 ++++++++----

  1 file changed, 8 insertions(+), 4 deletions(-)
block/backup.c +1 -4
index 763f0d7ff6..d8fdbfadfe 100644
@@ -741,12 +741,16 @@ BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
       job->cluster_size = cluster_size;
       job->copy_bitmap = copy_bitmap;
       copy_bitmap = NULL;
-    job->use_copy_range = !compress; /* compression isn't supported for it */
       job->copy_range_size = MIN_NON_ZERO(blk_get_max_transfer(job->common.blk),
                                           blk_get_max_transfer(job->target));
-    job->copy_range_size = MAX(job->cluster_size,
-                               QEMU_ALIGN_UP(job->copy_range_size,
-                                             job->cluster_size));
+    job->copy_range_size = QEMU_ALIGN_DOWN(job->copy_range_size,

block/backup.c BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs, · line 756

Quoted 10 lines
block/backup.c +7 -0
+                                           job->cluster_size);
+    /*
+     * Compression is not supported for copy_range. Also, we don't want to
+     * handle small max_transfer for copy_range (which currently don't
+     * handle max_transfer at all).
+     */
+    job->use_copy_range = !compress && job->copy_range_size > 0;
       /* Required permissions are already taken with target's blk_new() */
       block_job_add_bdrv(&job->common, "target", target, 0, BLK_PERM_ALL,
 

I'm clear on the alignment fix, I'm not clear on the comment about max_transfer and how it relates to copy_range_size being non-zero.

Quoted 1 line

"small max transfer" -- what happens when it's zero? we're apparently OK with a single cluster, but when it's zero, what happens?

if it zero it means that source or target requires max_transfer less than cluster_size. It seems not valid to call copy_range in this case. Still it's OK to use normal read/write, as they handle max_transfer internally in a loop (copy_range doesn't do it).

Signature · 2 lines
Best regards,
Vladimir
Reviewed-by
#4 lore raw

Re: [Qemu-devel] [PATCH v12 1/2] block/backup: fix max_transfer handling for copy_range

Quoted 43 lines · Vladimir Sementsov-Ogievskiy
Quoted 42 lines · John Snow
Quoted 26 lines · Vladimir Sementsov-Ogievskiy

Of course, QEMU_ALIGN_UP is a typo, it should be QEMU_ALIGN_DOWN, as we are trying to find aligned size which satisfy both source and target. Also, don't ignore too small max_transfer. In this case seems safer to disable copy_range.

Fixes 9ded4a0114968e

---
  block/backup.c | 12 ++++++++----

  1 file changed, 8 insertions(+), 4 deletions(-)
block/backup.c +1 -4
index 763f0d7ff6..d8fdbfadfe 100644
@@ -741,12 +741,16 @@ BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
       job->cluster_size = cluster_size;
       job->copy_bitmap = copy_bitmap;
       copy_bitmap = NULL;
-    job->use_copy_range = !compress; /* compression isn't supported for it */
       job->copy_range_size = MIN_NON_ZERO(blk_get_max_transfer(job->common.blk),
                                           blk_get_max_transfer(job->target));
-    job->copy_range_size = MAX(job->cluster_size,
-                               QEMU_ALIGN_UP(job->copy_range_size,
-                                             job->cluster_size));
+    job->copy_range_size = QEMU_ALIGN_DOWN(job->copy_range_size,

block/backup.c BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs, · line 756

Quoted 10 lines
block/backup.c +7 -0
+                                           job->cluster_size);
+    /*
+     * Compression is not supported for copy_range. Also, we don't want to
+     * handle small max_transfer for copy_range (which currently don't
+     * handle max_transfer at all).
+     */
+    job->use_copy_range = !compress && job->copy_range_size > 0;
       /* Required permissions are already taken with target's blk_new() */
       block_job_add_bdrv(&job->common, "target", target, 0, BLK_PERM_ALL,
 

I'm clear on the alignment fix, I'm not clear on the comment about max_transfer and how it relates to copy_range_size being non-zero.

Quoted 5 lines
Quoted 1 line

"small max transfer" -- what happens when it's zero? we're apparently OK with a single cluster, but when it's zero, what happens?

if it zero it means that source or target requires max_transfer less than cluster_size. It seems not valid to call copy_range in this case. Still it's OK to use normal read/write, as they handle max_transfer internally in a loop (copy_range doesn't do it).

oh, I'm ... sorry, I just didn't quite understand the comment.

You're just making sure copy_range after all of our checks is non-zero, plain and simple. If max_transfer was *smaller than a job cluster*, we might end up with a copy_range size that's zero, which is obviously... not useful.

So, I might phrase "Also, we don't want to..." as:

"copy_range does not respect max_transfer, so we factor that in here. If it's smaller than the job->cluster_size, we are unable to use copy_range."

Just a suggestion, though, so:

(SHOULD copy_range respect max_transfer? I guess it would be quite different -- it would only count things it had to fall back and actually *transfer*, right? I suppose that because it can have that fallback we need to accommodate it here in backup.c, hence this workaround clamp.)

#5 lore raw

Re: [Qemu-devel] [PATCH v12 1/2] block/backup: fix max_transfer handling for copy_range

Quoted 59 lines · John Snow
Quoted 43 lines · Vladimir Sementsov-Ogievskiy
Quoted 42 lines · John Snow
Quoted 26 lines · Vladimir Sementsov-Ogievskiy

Of course, QEMU_ALIGN_UP is a typo, it should be QEMU_ALIGN_DOWN, as we are trying to find aligned size which satisfy both source and target. Also, don't ignore too small max_transfer. In this case seems safer to disable copy_range.

Fixes 9ded4a0114968e
---
   block/backup.c | 12 ++++++++----
   1 file changed, 8 insertions(+), 4 deletions(-)
block/backup.c +1 -4
index 763f0d7ff6..d8fdbfadfe 100644
@@ -741,12 +741,16 @@ BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
       job->cluster_size = cluster_size;
       job->copy_bitmap = copy_bitmap;
       copy_bitmap = NULL;
-    job->use_copy_range = !compress; /* compression isn't supported for it */
       job->copy_range_size = MIN_NON_ZERO(blk_get_max_transfer(job->common.blk),
                                           blk_get_max_transfer(job->target));
-    job->copy_range_size = MAX(job->cluster_size,
-                               QEMU_ALIGN_UP(job->copy_range_size,
-                                             job->cluster_size));
+    job->copy_range_size = QEMU_ALIGN_DOWN(job->copy_range_size,

block/backup.c BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs, · line 756

Quoted 10 lines
block/backup.c +7 -0
+                                           job->cluster_size);
+    /*
+     * Compression is not supported for copy_range. Also, we don't want to
+     * handle small max_transfer for copy_range (which currently don't
+     * handle max_transfer at all).
+     */
+    job->use_copy_range = !compress && job->copy_range_size > 0;
       /* Required permissions are already taken with target's blk_new() */
       block_job_add_bdrv(&job->common, "target", target, 0, BLK_PERM_ALL,
 

I'm clear on the alignment fix, I'm not clear on the comment about max_transfer and how it relates to copy_range_size being non-zero.

Quoted 5 lines
Quoted 1 line

"small max transfer" -- what happens when it's zero? we're apparently OK with a single cluster, but when it's zero, what happens?

if it zero it means that source or target requires max_transfer less than cluster_size. It seems not valid to call copy_range in this case. Still it's OK to use normal read/write, as they handle max_transfer internally in a loop (copy_range doesn't do it).

oh, I'm ... sorry, I just didn't quite understand the comment.

You're just making sure copy_range after all of our checks is non-zero, plain and simple. If max_transfer was *smaller than a job cluster*, we might end up with a copy_range size that's zero, which is obviously... not useful.

Quoted 4 lines

So, I might phrase "Also, we don't want to..." as:

"copy_range does not respect max_transfer, so we factor that in here. If it's smaller than the job->cluster_size, we are unable to use copy_range."

We actually able to: just using a loop and calling copy_range several times. May be just:

copy_range does not respect max_transfer, so we factor that in here. If it's smaller than the job->cluster_size, we do not use copy_range.

Quoted 11 lines

Just a suggestion, though, so:

(SHOULD copy_range respect max_transfer? I guess it would be quite different -- it would only count things it had to fall back and actually *transfer*, right? I suppose that because it can have that fallback we need to accommodate it here in backup.c, hence this workaround clamp.)

Signature · 2 lines
Best regards,
Vladimir
#6 lore raw

PATCH v12 2/2 block/backup: fix backup_cow_with_offload for last cluster

In reply to Vladimir Sementsov-Ogievskiy

We shouldn't try to copy bytes beyond EOF. Fix it.

---
 block/backup.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
block/backup.c +1 -1
index d8fdbfadfe..89f7f89200 100644
@@ -161,7 +161,7 @@ static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job,
 
     assert(QEMU_IS_ALIGNED(job->copy_range_size, job->cluster_size));
     assert(QEMU_IS_ALIGNED(start, job->cluster_size));
-    nbytes = MIN(job->copy_range_size, end - start);
+    nbytes = MIN(job->copy_range_size, MIN(end, job->len) - start);
     nr_clusters = DIV_ROUND_UP(nbytes, job->cluster_size);
     bdrv_reset_dirty_bitmap(job->copy_bitmap, start,
                             job->cluster_size * nr_clusters);
Signature · 1 line
2.21.0
#7 lore raw

Re: [Qemu-devel] [PATCH v12 2/2] block/backup: fix backup_cow_with_offload for last cluster

Quoted 14 lines · Vladimir Sementsov-Ogievskiy

We shouldn't try to copy bytes beyond EOF. Fix it.

---
  block/backup.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
block/backup.c +0 -0
index d8fdbfadfe..89f7f89200 100644

block/backup.c static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job, · line 164

Quoted 5 lines
block/backup.c +1 -1
@@ -161,7 +161,7 @@ static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job,
 
      assert(QEMU_IS_ALIGNED(job->copy_range_size, job->cluster_size));
      assert(QEMU_IS_ALIGNED(start, job->cluster_size));
-    nbytes = MIN(job->copy_range_size, end - start);
+    nbytes = MIN(job->copy_range_size, MIN(end, job->len) - start);

I'm a little confused. I think the patch as written is correct, but I don't know what problem it solves.

If we're going to allow the caller to pass in an end that's beyond EOF, does that mean we are *requiring* the caller to pass in a value that's aligned?

We should probably assert what kind of a value we're accepted here, right? We do for start, but should we for 'end' as well?

Then ...

block/backup.c static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job, · line 165

Quoted 1 line
block/backup.c +0 -0
      nr_clusters = DIV_ROUND_UP(nbytes, job->cluster_size);

Don't we just round this right back up immediately anyway? Does that mean we have callers that are passing in an 'end' that's more than 1 job-cluster beyond EOF? That seems like something that should be fixed in the caller, surely?

Quoted 3 lines
block/backup.c +0 -0
      bdrv_reset_dirty_bitmap(job->copy_bitmap, start,
                              job->cluster_size * nr_clusters);
#8 lore raw

Re: [Qemu-devel] [PATCH v12 2/2] block/backup: fix backup_cow_with_offload for last cluster

Quoted 12 lines · John Snow
Quoted 9 lines · Vladimir Sementsov-Ogievskiy

We shouldn't try to copy bytes beyond EOF. Fix it.

---
  block/backup.c | 2 +-

  1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 11 lines

block/backup.c static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job, · line 163

Quoted 9 lines
block/backup.c +1 -1
index d8fdbfadfe..89f7f89200 100644
@@ -161,7 +161,7 @@ static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job,
       assert(QEMU_IS_ALIGNED(job->copy_range_size, job->cluster_size));
       assert(QEMU_IS_ALIGNED(start, job->cluster_size));
-    nbytes = MIN(job->copy_range_size, end - start);
+    nbytes = MIN(job->copy_range_size, MIN(end, job->len) - start);

I'm a little confused. I think the patch as written is correct, but I don't know what problem it solves.

last cluster may exceed EOF. And backup_do_cow (who calls backup_cow_with_offload) rounds all to clusters. It's not bad, but we need to crop nbytes before calling actual io functions. backup_cow_with_bounce_buffer does the same thing.

Quoted 2 lines

If we're going to allow the caller to pass in an end that's beyond EOF, does that mean we are *requiring* the caller to pass in a value that's aligned?

Actually yes, as we are resetting dirty bitmap.

Quoted 2 lines

We should probably assert what kind of a value we're accepted here, right? We do for start, but should we for 'end' as well?

Yes assertion may be added.

Quoted 6 lines

Then ...

block/backup.c static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job, · line 165

Quoted 1 line
block/backup.c +0 -0
       nr_clusters = DIV_ROUND_UP(nbytes, job->cluster_size);

Don't we just round this right back up immediately anyway? Does that mean we have callers that are passing in an 'end' that's more than 1 job-cluster beyond EOF? That seems like something that should be fixed in the caller, surely?

nr_clusters are used to set/reset dirty bitmap. It's OK. Still, for last cluster we can drop it and use nbytes directly. No there should not be such callers. nbytes is used to call blk_co_copy_range, and must be cropped to not exceed EOF.

Also, backup_cow_with_bounce_buffer behave in similar way: it crops nbytes.

Of course, there is a place for good refactoring, but I think not in this patch, it's a small bug fix.

Quoted 5 lines
Quoted 4 lines
block/backup.c +0 -0
       bdrv_reset_dirty_bitmap(job->copy_bitmap, start,
                               job->cluster_size * nr_clusters);
 
Signature · 2 lines
Best regards,
Vladimir
Reviewed-by
#9 lore raw

Re: [Qemu-devel] [PATCH v12 2/2] block/backup: fix backup_cow_with_offload for last cluster

Quoted 39 lines · Vladimir Sementsov-Ogievskiy
Quoted 12 lines · John Snow
Quoted 9 lines · Vladimir Sementsov-Ogievskiy

We shouldn't try to copy bytes beyond EOF. Fix it.

---
  block/backup.c | 2 +-

  1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 11 lines

block/backup.c static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job, · line 163

Quoted 9 lines
block/backup.c +1 -1
index d8fdbfadfe..89f7f89200 100644
@@ -161,7 +161,7 @@ static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job,
       assert(QEMU_IS_ALIGNED(job->copy_range_size, job->cluster_size));
       assert(QEMU_IS_ALIGNED(start, job->cluster_size));
-    nbytes = MIN(job->copy_range_size, end - start);
+    nbytes = MIN(job->copy_range_size, MIN(end, job->len) - start);

I'm a little confused. I think the patch as written is correct, but I don't know what problem it solves.

last cluster may exceed EOF. And backup_do_cow (who calls backup_cow_with_offload) rounds all to clusters. It's not bad, but we need to crop nbytes before calling actual io functions. backup_cow_with_bounce_buffer does the same thing.

Quoted 2 lines

If we're going to allow the caller to pass in an end that's beyond EOF, does that mean we are *requiring* the caller to pass in a value that's aligned?

Actually yes, as we are resetting dirty bitmap.

Quoted 2 lines

We should probably assert what kind of a value we're accepted here, right? We do for start, but should we for 'end' as well?

Yes assertion may be added.

Quoted 1 line
Quoted 9 lines
Quoted 5 lines

Then ...

block/backup.c static int coroutine_fn backup_cow_with_offload(BackupBlockJob *job, · line 165

Quoted 1 line
block/backup.c +0 -0
       nr_clusters = DIV_ROUND_UP(nbytes, job->cluster_size);

Don't we just round this right back up immediately anyway? Does that mean we have callers that are passing in an 'end' that's more than 1 job-cluster beyond EOF? That seems like something that should be fixed in the caller, surely?

nr_clusters are used to set/reset dirty bitmap. It's OK. Still, for last cluster we can drop it and use nbytes directly. No there should not be such callers. nbytes is used to call blk_co_copy_range, and must be cropped to not exceed EOF.

Ah, right, right ... I *was* confused. We don't use nr_clusters for the IO itself, just the bitmap. So we effectively re-calculate aligned and unaligned values for use in different places.

Quoted 11 lines

Also, backup_cow_with_bounce_buffer behave in similar way: it crops nbytes.

Of course, there is a place for good refactoring, but I think not in this patch, it's a small bug fix.

Quoted 5 lines
Quoted 4 lines
block/backup.c +0 -0
       bdrv_reset_dirty_bitmap(job->copy_bitmap, start,
                               job->cluster_size * nr_clusters);
 

We should make the interface here a little more clear I think, but what you wrote is correct.