Skip to contentSkip to conversation

block/rbd: fixes for bdrv_co_block_status

PATCH Superseded Peter Lieven 3 participants 13 messages Jan 10, 2022 – Jan 12, 2022 Block layer

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#6 lore raw

PATCH 2/2 block/rbd: workaround for ceph issue #53784

In reply to Peter Lieven

librbd had a bug until early 2022 that affected all versions of ceph that supported fast-diff. This bug results in reporting of incorrect offsets if the offset parameter to rbd_diff_iterate2 is not object aligned. Work around this bug by rounding down the offset to object boundaries.

---
 block/rbd.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)
block/rbd.c +16 -1
index 5e9dc91d81..260cb9f4b4 100644
@@ -1333,6 +1333,7 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
     int status, r;
     RBDDiffIterateReq req = { .offs = offset };
     uint64_t features, flags;
+    int64_t head;
 
     assert(offset + bytes <= s->image_size);
 
@@ -1360,6 +1361,19 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
         return status;
     }
 
+    /*
+     * librbd had a bug until early 2022 that affected all versions of ceph that
+     * supported fast-diff. This bug results in reporting of incorrect offsets
+     * if the offset parameter to rbd_diff_iterate2 is not object aligned.
+     * Work around this bug by rounding down the offset to object boundaries.
+     *
+     * See: https://tracker.ceph.com/issues/53784
+     */
+    head = offset & (s->object_size - 1);
+    offset -= head;
+    req.offs -= head;
+    bytes += head;
+
     r = rbd_diff_iterate2(s->image, NULL, offset, bytes, true, true,
                           qemu_rbd_diff_iterate_cb, &req);
     if (r < 0 && r != QEMU_RBD_EXIT_DIFF_ITERATE2) {
@@ -1379,7 +1393,8 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
         status = BDRV_BLOCK_ZERO | BDRV_BLOCK_OFFSET_VALID;
     }
 
-    *pnum = req.bytes;
+    assert(req.bytes > head);
+    *pnum = req.bytes - head;
     return status;
 }
 
Signature · 1 line
2.25.1
#0 lore raw

cover block/rbd: fixes for bdrv_co_block_status

Peter Lieven (2):
block/rbd: fix handling of holes in .bdrv_co_block_status block/rbd: workaround for ceph issue #53784

 block/rbd.c | 72 +++++++++++++++++++++++++++++++++++++----------------
 1 file changed, 50 insertions(+), 22 deletions(-)
Signature · 1 line
2.25.1
#1 lore raw

PATCH 1/2 block/rbd: fix handling of holes in .bdrv_co_block_status

the assumption that we can't hit a hole if we do not diff against a snapshot was wrong.

We can see a hole in an image if we diff against base if there exists an older snapshot of the image and we have discarded blocks in the image where the snapshot has data.

Fixes 0347a8fd4c3faaedf119be04c197804be40a384b
---
 block/rbd.c | 55 +++++++++++++++++++++++++++++++++--------------------
 1 file changed, 34 insertions(+), 21 deletions(-)
block/rbd.c +34 -21
index def96292e0..5e9dc91d81 100644
@@ -1279,13 +1279,24 @@ static int qemu_rbd_diff_iterate_cb(uint64_t offs, size_t len,
     RBDDiffIterateReq *req = opaque;
 
     assert(req->offs + req->bytes <= offs);
-    /*
-     * we do not diff against a snapshot so we should never receive a callback
-     * for a hole.
-     */
-    assert(exists);
 
-    if (!req->exists && offs > req->offs) {
+    if (req->exists && offs > req->offs + req->bytes) {
+        /*
+         * we started in an allocated area and jumped over an unallocated area,
+         * req->bytes contains the length of the allocated area before the
+         * unallocated area. stop further processing.
+         */
+        return QEMU_RBD_EXIT_DIFF_ITERATE2;
+    }
+    if (req->exists && !exists) {
+        /*
+         * we started in an allocated area and reached a hole. req->bytes
+         * contains the length of the allocated area before the hole.
+         * stop further processing.
+         */
+        return QEMU_RBD_EXIT_DIFF_ITERATE2;
+    }
+    if (!req->exists && exists && offs > req->offs) {
         /*
          * we started in an unallocated area and hit the first allocated
          * block. req->bytes must be set to the length of the unallocated area
@@ -1295,17 +1306,19 @@ static int qemu_rbd_diff_iterate_cb(uint64_t offs, size_t len,
         return QEMU_RBD_EXIT_DIFF_ITERATE2;
     }
 
-    if (req->exists && offs > req->offs + req->bytes) {
-        /*
-         * we started in an allocated area and jumped over an unallocated area,
-         * req->bytes contains the length of the allocated area before the
-         * unallocated area. stop further processing.
-         */
-        return QEMU_RBD_EXIT_DIFF_ITERATE2;
-    }
+    /*
+     * assert that we caught all cases above and allocation state has not
+     * changed during callbacks.
+     */
+    assert(exists == req->exists || !req->bytes);
+    req->exists = exists;
 
-    req->bytes += len;
-    req->exists = true;
+    /*
+     * assert that we either return an unallocated block or have got callbacks
+     * for all allocated blocks present.
+     */
+    assert(!req->exists || offs == req->offs + req->bytes);
+    req->bytes = offs + len - req->offs;
 
     return 0;
 }
@@ -1354,13 +1367,13 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
     }
     assert(req.bytes <= bytes);
     if (!req.exists) {
-        if (r == 0) {
+        if (r == 0 && !req.bytes) {
             /*
-             * rbd_diff_iterate2 does not invoke callbacks for unallocated
-             * areas. This here catches the case where no callback was
-             * invoked at all (req.bytes == 0).
+             * rbd_diff_iterate2 does not invoke callbacks for unallocated areas
+             * except for the case where an overlay has a hole where the parent
+             * or an older snapshot of the image has not. This here catches the
+             * case where no callback was invoked at all.
              */
-            assert(req.bytes == 0);
             req.bytes = bytes;
         }
         status = BDRV_BLOCK_ZERO | BDRV_BLOCK_OFFSET_VALID;
Signature · 1 line
2.25.1
Tested-by on v1 2/2
#7 lore raw

Re: [PATCH 2/2] block/rbd: workaround for ceph issue #53784

In reply to Peter Lieven
Quoted 45 lines · Peter Lieven

librbd had a bug until early 2022 that affected all versions of ceph that supported fast-diff. This bug results in reporting of incorrect offsets if the offset parameter to rbd_diff_iterate2 is not object aligned. Work around this bug by rounding down the offset to object boundaries.

---
block/rbd.c | 17 ++++++++++++++++-

1 file changed, 16 insertions(+), 1 deletion(-)
block/rbd.c +14 -0
index 5e9dc91d81..260cb9f4b4 100644
@@ -1333,6 +1333,7 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
    int status, r;
    RBDDiffIterateReq req = { .offs = offset };
    uint64_t features, flags;
+    int64_t head;
 
    assert(offset + bytes <= s->image_size);
 
@@ -1360,6 +1361,19 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
        return status;
    }
 
+    /*
+     * librbd had a bug until early 2022 that affected all versions of ceph that
+     * supported fast-diff. This bug results in reporting of incorrect offsets
+     * if the offset parameter to rbd_diff_iterate2 is not object aligned.
+     * Work around this bug by rounding down the offset to object boundaries.
+     *
+     * See: https://tracker.ceph.com/issues/53784
+     */
+    head = offset & (s->object_size - 1);
+    offset -= head;
+    req.offs -= head;
+    bytes += head;
+
    r = rbd_diff_iterate2(s->image, NULL, offset, bytes, true, true,
                          qemu_rbd_diff_iterate_cb, &req);
    if (r < 0 && r != QEMU_RBD_EXIT_DIFF_ITERATE2) {

block/rbd.c static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs, · line 1399

Quoted 8 lines
block/rbd.c +2 -1
@@ -1379,7 +1393,8 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
        status = BDRV_BLOCK_ZERO | BDRV_BLOCK_OFFSET_VALID;
    }
 
-    *pnum = req.bytes;
+    assert(req.bytes > head);
+    *pnum = req.bytes - head;
    return status;
 }

Thanks for the workaround!

I just tested this patch for the issue reported in this BZ [1] and the test now works correctly!

[1] https://bugzilla.redhat.com/show_bug.cgi?id=2034791

#8 lore raw

Re: [PATCH 2/2] block/rbd: workaround for ceph issue #53784

Am 10.01.22 um 15:18 schrieb Stefano Garzarella:

Quoted 57 lines · Peter Lieven
Quoted 45 lines

librbd had a bug until early 2022 that affected all versions of ceph that supported fast-diff. This bug results in reporting of incorrect offsets if the offset parameter to rbd_diff_iterate2 is not object aligned. Work around this bug by rounding down the offset to object boundaries.

---
block/rbd.c | 17 ++++++++++++++++-

1 file changed, 16 insertions(+), 1 deletion(-)
block/rbd.c +14 -0
index 5e9dc91d81..260cb9f4b4 100644
@@ -1333,6 +1333,7 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
     int status, r;
     RBDDiffIterateReq req = { .offs = offset };
     uint64_t features, flags;
+    int64_t head;
 
     assert(offset + bytes <= s->image_size);
 
@@ -1360,6 +1361,19 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
         return status;
     }
 
+    /*
+     * librbd had a bug until early 2022 that affected all versions of ceph that
+     * supported fast-diff. This bug results in reporting of incorrect offsets
+     * if the offset parameter to rbd_diff_iterate2 is not object aligned.
+     * Work around this bug by rounding down the offset to object boundaries.
+     *
+     * See: https://tracker.ceph.com/issues/53784
+     */
+    head = offset & (s->object_size - 1);
+    offset -= head;
+    req.offs -= head;
+    bytes += head;
+
     r = rbd_diff_iterate2(s->image, NULL, offset, bytes, true, true,
                           qemu_rbd_diff_iterate_cb, &req);
     if (r < 0 && r != QEMU_RBD_EXIT_DIFF_ITERATE2) {

block/rbd.c static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs, · line 1399

Quoted 8 lines
block/rbd.c +2 -1
@@ -1379,7 +1393,8 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
         status = BDRV_BLOCK_ZERO | BDRV_BLOCK_OFFSET_VALID;
     }
 
-    *pnum = req.bytes;
+    assert(req.bytes > head);
+    *pnum = req.bytes - head;
     return status;
 }

Thanks for the workaround!

Quoted 6 lines

I just tested this patch for the issue reported in this BZ [1] and the test now works correctly!

[1] https://bugzilla.redhat.com/show_bug.cgi?id=2034791

Hi Stefano,

thanks for the feedback. Please note that you also need the other patch or you will sooner or later run into another assertion as soon as rbd snapshots are involved.

Regarding the workaround I need confirmation from Ilya that it covers all cases. I do not know if it works if striping or EC is configured on the pool.

Best,

Peter

#9 lore raw

Re: [PATCH 2/2] block/rbd: workaround for ceph issue #53784

Hi Peter,

Quoted 4 lines · Peter Lieven

Hi Stefano,

thanks for the feedback. Please note that you also need the other patch or you will sooner or later run into another assertion as soon as rbd snapshots are involved.

Yep, I tested with the entire series applied.
Anyway, thanks for clarifying that.

Quoted 2 lines

Regarding the workaround I need confirmation from Ilya that it covers all cases. I do not know if it works if striping or EC is configured on the pool.

Sure :-)

Thanks,
Stefano

#2 lore raw

Re: [PATCH 1/2] block/rbd: fix handling of holes in .bdrv_co_block_status

In reply to Peter Lieven
Quoted 98 lines · Peter Lieven

the assumption that we can't hit a hole if we do not diff against a snapshot was wrong.

We can see a hole in an image if we diff against base if there exists an older snapshot of the image and we have discarded blocks in the image where the snapshot has data.

Fixes 0347a8fd4c3faaedf119be04c197804be40a384b
---
 block/rbd.c | 55 +++++++++++++++++++++++++++++++++--------------------
 1 file changed, 34 insertions(+), 21 deletions(-)
block/rbd.c +34 -22
index def96292e0..5e9dc91d81 100644
@@ -1279,13 +1279,24 @@ static int qemu_rbd_diff_iterate_cb(uint64_t offs, size_t len,
     RBDDiffIterateReq *req = opaque;
 
     assert(req->offs + req->bytes <= offs);
-    /*
-     * we do not diff against a snapshot so we should never receive a callback
-     * for a hole.
-     */
-    assert(exists);
 
-    if (!req->exists && offs > req->offs) {
+    if (req->exists && offs > req->offs + req->bytes) {
+        /*
+         * we started in an allocated area and jumped over an unallocated area,
+         * req->bytes contains the length of the allocated area before the
+         * unallocated area. stop further processing.
+         */
+        return QEMU_RBD_EXIT_DIFF_ITERATE2;
+    }
+    if (req->exists && !exists) {
+        /*
+         * we started in an allocated area and reached a hole. req->bytes
+         * contains the length of the allocated area before the hole.
+         * stop further processing.
+         */
+        return QEMU_RBD_EXIT_DIFF_ITERATE2;
+    }
+    if (!req->exists && exists && offs > req->offs) {
         /*
          * we started in an unallocated area and hit the first allocated
          * block. req->bytes must be set to the length of the unallocated area
@@ -1295,17 +1306,19 @@ static int qemu_rbd_diff_iterate_cb(uint64_t offs, size_t len,
         return QEMU_RBD_EXIT_DIFF_ITERATE2;
     }
 
-    if (req->exists && offs > req->offs + req->bytes) {
-        /*
-         * we started in an allocated area and jumped over an unallocated area,
-         * req->bytes contains the length of the allocated area before the
-         * unallocated area. stop further processing.
-         */
-        return QEMU_RBD_EXIT_DIFF_ITERATE2;
-    }
+    /*
+     * assert that we caught all cases above and allocation state has not
+     * changed during callbacks.
+     */
+    assert(exists == req->exists || !req->bytes);
+    req->exists = exists;
 
-    req->bytes += len;
-    req->exists = true;
+    /*
+     * assert that we either return an unallocated block or have got callbacks
+     * for all allocated blocks present.
+     */
+    assert(!req->exists || offs == req->offs + req->bytes);
+    req->bytes = offs + len - req->offs;
 
     return 0;
 }
@@ -1354,13 +1367,13 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
     }
     assert(req.bytes <= bytes);
     if (!req.exists) {
-        if (r == 0) {
+        if (r == 0 && !req.bytes) {
             /*
-             * rbd_diff_iterate2 does not invoke callbacks for unallocated
-             * areas. This here catches the case where no callback was
-             * invoked at all (req.bytes == 0).
+             * rbd_diff_iterate2 does not invoke callbacks for unallocated areas
+             * except for the case where an overlay has a hole where the parent
+             * or an older snapshot of the image has not. This here catches the
+             * case where no callback was invoked at all.
              */
-            assert(req.bytes == 0);
             req.bytes = bytes;
         }
         status = BDRV_BLOCK_ZERO | BDRV_BLOCK_OFFSET_VALID;
--
Quoted 3 lines

2.25.1

Hi Peter,

Can we just skip these "holes" by replacing the existing assert with an if statement that would simply bail from the callback on !exists?

Just trying to keep the logic as simple as possible since as it turns out we get to contend with ages-old librbd bugs here...

Thanks,

Ilya

#10 lore raw

Re: [PATCH 2/2] block/rbd: workaround for ceph issue #53784

In reply to Peter Lieven
Quoted 7 lines · Peter Lieven

librbd had a bug until early 2022 that affected all versions of ceph that supported fast-diff. This bug results in reporting of incorrect offsets if the offset parameter to rbd_diff_iterate2 is not object aligned. Work around this bug by rounding down the offset to object boundaries.

I don't think the Fixes tag is appropriate here. Linking librbd ticket is fine but this patch doesn't really fix anything.

Quoted 24 lines
---
 block/rbd.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)
block/rbd.c +3 -0
index 5e9dc91d81..260cb9f4b4 100644
@@ -1333,6 +1333,7 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
     int status, r;
     RBDDiffIterateReq req = { .offs = offset };
     uint64_t features, flags;
+    int64_t head;
 
     assert(offset + bytes <= s->image_size);
 
@@ -1360,6 +1361,19 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
         return status;
     }
 
+    /*
+     * librbd had a bug until early 2022 that affected all versions of ceph that

block/rbd.c static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs, · line 1375

Quoted 10 lines
block/rbd.c +10 -0
+     * supported fast-diff. This bug results in reporting of incorrect offsets
+     * if the offset parameter to rbd_diff_iterate2 is not object aligned.
+     * Work around this bug by rounding down the offset to object boundaries.
+     *
+     * See: https://tracker.ceph.com/issues/53784
+     */
+    head = offset & (s->object_size - 1);
+    offset -= head;
+    req.offs -= head;
+    bytes += head;

So it looks like the intention is to have more or less a permanent workaround since all librbd versions are affected, right? For that, I think we would need to also reject custom striping patterns and clones. For the above to be reliable, the image has to be standalone and have a default striping pattern (stripe_unit == object_size && stripe_count == 1). Otherwise, behave as if fast-diff is disabled or invalid.

block/rbd.c static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs, · line 1376

Quoted 1 line
block/rbd.c +1 -0
+

Nit: I'd replace { .offs = offset } initialization at the top with {} and assign to req.offs here, right before calling rbd_diff_iterate2().

block/rbd.c static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs, · line 1396

Quoted 9 lines
block/rbd.c +1 -1
     r = rbd_diff_iterate2(s->image, NULL, offset, bytes, true, true,
                           qemu_rbd_diff_iterate_cb, &req);
     if (r < 0 && r != QEMU_RBD_EXIT_DIFF_ITERATE2) {
@@ -1379,7 +1393,8 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
         status = BDRV_BLOCK_ZERO | BDRV_BLOCK_OFFSET_VALID;
     }
 
-    *pnum = req.bytes;
+    assert(req.bytes > head);

I'd expand the workaround comment with an explanation of why it's OK to round down the offset -- because rbd_diff_iterate2() is called with whole_object=true. If that wasn't the case, on top of inconsistent results for different offsets within an object, this assert could be triggered.

Thanks,

Ilya

#11 lore raw

Re: [PATCH 2/2] block/rbd: workaround for ceph issue #53784

Am 12.01.22 um 10:59 schrieb Ilya Dryomov:

Quoted 9 lines · Peter Lieven
Quoted 6 lines

librbd had a bug until early 2022 that affected all versions of ceph that supported fast-diff. This bug results in reporting of incorrect offsets if the offset parameter to rbd_diff_iterate2 is not object aligned. Work around this bug by rounding down the offset to object boundaries.

I don't think the Fixes tag is appropriate here. Linking librbd ticket is fine but this patch doesn't really fix anything.

Okay, I will change that to See:

Quoted 34 lines
Quoted 34 lines
---
 block/rbd.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)
block/rbd.c +12 -0
index 5e9dc91d81..260cb9f4b4 100644
@@ -1333,6 +1333,7 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
     int status, r;
     RBDDiffIterateReq req = { .offs = offset };
     uint64_t features, flags;
+    int64_t head;
 
     assert(offset + bytes <= s->image_size);
 
@@ -1360,6 +1361,19 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
         return status;
     }
 
+    /*
+     * librbd had a bug until early 2022 that affected all versions of ceph that
+     * supported fast-diff. This bug results in reporting of incorrect offsets
+     * if the offset parameter to rbd_diff_iterate2 is not object aligned.
+     * Work around this bug by rounding down the offset to object boundaries.
+     *
+     * See: https://tracker.ceph.com/issues/53784
+     */
+    head = offset & (s->object_size - 1);
+    offset -= head;
+    req.offs -= head;
Quoted 8 lines

block/rbd.c static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs, · line 1375

Quoted 1 line
block/rbd.c +1 -0
+    bytes += head;

So it looks like the intention is to have more or less a permanent workaround since all librbd versions are affected, right? For that, I think we would need to also reject custom striping patterns and clones. For the above to be reliable, the image has to be standalone and have a default striping pattern (stripe_unit == object_size && stripe_count == 1). Otherwise, behave as if fast-diff is disabled or invalid.

Do you have a fealing how many users use a different striping pattern than default?

What about EC backed pools?

Do you have another idea how we can detect if the librbd version is broken?

Quoted 12 lines

block/rbd.c static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs, · line 1376

Quoted 2 lines
block/rbd.c +1 -0
+

Nit: I'd replace { .offs = offset } initialization at the top with {} and assign to req.offs here, right before calling rbd_diff_iterate2().

Quoted 7 lines
block/rbd.c +0 -0
     r = rbd_diff_iterate2(s->image, NULL, offset, bytes, true, true,
                           qemu_rbd_diff_iterate_cb, &req);
     if (r < 0 && r != QEMU_RBD_EXIT_DIFF_ITERATE2) {
@@ -1379,7 +1393,8 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
         status = BDRV_BLOCK_ZERO | BDRV_BLOCK_OFFSET_VALID;
     }
 
Quoted 7 lines

block/rbd.c static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs, · line 1396

Quoted 2 lines
block/rbd.c +1 -1
-    *pnum = req.bytes;
+    assert(req.bytes > head);

I'd expand the workaround comment with an explanation of why it's OK to round down the offset -- because rbd_diff_iterate2() is called with whole_object=true. If that wasn't the case, on top of inconsistent results for different offsets within an object, this assert could be triggered.

Sure, you are right. I had this in mind. This also does not change complexity

since we stay with the offset in the same object. I will mention both.

Peter

#12 lore raw

Re: [PATCH 2/2] block/rbd: workaround for ceph issue #53784

Quoted 8 lines · Peter Lieven

Am 12.01.22 um 10:59 schrieb Ilya Dryomov:

Quoted 6 lines · Peter Lieven
Quoted 5 lines

librbd had a bug until early 2022 that affected all versions of ceph that supported fast-diff. This bug results in reporting of incorrect offsets if the offset parameter to rbd_diff_iterate2 is not object aligned. Work around this bug by rounding down the offset to object boundaries.

Quoted 6 lines
Quoted 3 lines
Quoted 1 line

I don't think the Fixes tag is appropriate here. Linking librbd ticket is fine but this patch doesn't really fix anything.

Okay, I will change that to See:

It's already linked in the source code, up to you if you also want to link it in the description.

Quoted 46 lines
Quoted 36 lines
Quoted 36 lines
---
 block/rbd.c | 17 ++++++++++++++++-
 1 file changed, 16 insertions(+), 1 deletion(-)
block/rbd.c +12 -0
index 5e9dc91d81..260cb9f4b4 100644
@@ -1333,6 +1333,7 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
     int status, r;
     RBDDiffIterateReq req = { .offs = offset };
     uint64_t features, flags;
+    int64_t head;
 
     assert(offset + bytes <= s->image_size);
 
@@ -1360,6 +1361,19 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
         return status;
     }
 
+    /*
+     * librbd had a bug until early 2022 that affected all versions of ceph that
+     * supported fast-diff. This bug results in reporting of incorrect offsets
+     * if the offset parameter to rbd_diff_iterate2 is not object aligned.
+     * Work around this bug by rounding down the offset to object boundaries.
+     *
+     * See: https://tracker.ceph.com/issues/53784
+     */
+    head = offset & (s->object_size - 1);
+    offset -= head;
+    req.offs -= head;
Quoted 8 lines

block/rbd.c static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs, · line 1375

Quoted 1 line
block/rbd.c +1 -0
+    bytes += head;

So it looks like the intention is to have more or less a permanent workaround since all librbd versions are affected, right? For that, I think we would need to also reject custom striping patterns and clones. For the above to be reliable, the image has to be standalone and have a default striping pattern (stripe_unit == object_size && stripe_count == 1). Otherwise, behave as if fast-diff is disabled or invalid.

Quoted 1 line

Do you have a fealing how many users use a different striping pattern than default?

Very few.

Quoted 2 lines

What about EC backed pools?

In this context EC pools behave exactly the same as replicated pools.

Quoted 2 lines

Do you have another idea how we can detect if the librbd version is broken?

No. Initially I wanted to just fix these bugs in librbd, relying on the assumption that setups with a new QEMU should also have a fairly new librbd. But after looking at various distros and realizing the extent of rbd_diff_iterate2() issues, I think a long-term workaround in QEMU makes sense. A configure-time check for known good versions of librbd can be added later if someone feels like it.

Thanks,

Ilya

#3 lore raw

Re: [PATCH 1/2] block/rbd: fix handling of holes in .bdrv_co_block_status

In reply to Ilya Dryomov

Am 12.01.22 um 10:05 schrieb Ilya Dryomov:

Quoted 103 lines · Peter Lieven
Quoted 97 lines

the assumption that we can't hit a hole if we do not diff against a snapshot was wrong.

We can see a hole in an image if we diff against base if there exists an older snapshot of the image and we have discarded blocks in the image where the snapshot has data.

Fixes 0347a8fd4c3faaedf119be04c197804be40a384b
---
 block/rbd.c | 55 +++++++++++++++++++++++++++++++++--------------------
 1 file changed, 34 insertions(+), 21 deletions(-)
block/rbd.c +34 -22
index def96292e0..5e9dc91d81 100644
@@ -1279,13 +1279,24 @@ static int qemu_rbd_diff_iterate_cb(uint64_t offs, size_t len,
     RBDDiffIterateReq *req = opaque;
 
     assert(req->offs + req->bytes <= offs);
-    /*
-     * we do not diff against a snapshot so we should never receive a callback
-     * for a hole.
-     */
-    assert(exists);
 
-    if (!req->exists && offs > req->offs) {
+    if (req->exists && offs > req->offs + req->bytes) {
+        /*
+         * we started in an allocated area and jumped over an unallocated area,
+         * req->bytes contains the length of the allocated area before the
+         * unallocated area. stop further processing.
+         */
+        return QEMU_RBD_EXIT_DIFF_ITERATE2;
+    }
+    if (req->exists && !exists) {
+        /*
+         * we started in an allocated area and reached a hole. req->bytes
+         * contains the length of the allocated area before the hole.
+         * stop further processing.
+         */
+        return QEMU_RBD_EXIT_DIFF_ITERATE2;
+    }
+    if (!req->exists && exists && offs > req->offs) {
         /*
          * we started in an unallocated area and hit the first allocated
          * block. req->bytes must be set to the length of the unallocated area
@@ -1295,17 +1306,19 @@ static int qemu_rbd_diff_iterate_cb(uint64_t offs, size_t len,
         return QEMU_RBD_EXIT_DIFF_ITERATE2;
     }
 
-    if (req->exists && offs > req->offs + req->bytes) {
-        /*
-         * we started in an allocated area and jumped over an unallocated area,
-         * req->bytes contains the length of the allocated area before the
-         * unallocated area. stop further processing.
-         */
-        return QEMU_RBD_EXIT_DIFF_ITERATE2;
-    }
+    /*
+     * assert that we caught all cases above and allocation state has not
+     * changed during callbacks.
+     */
+    assert(exists == req->exists || !req->bytes);
+    req->exists = exists;
 
-    req->bytes += len;
-    req->exists = true;
+    /*
+     * assert that we either return an unallocated block or have got callbacks
+     * for all allocated blocks present.
+     */
+    assert(!req->exists || offs == req->offs + req->bytes);
+    req->bytes = offs + len - req->offs;
 
     return 0;
 }
@@ -1354,13 +1367,13 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
     }
     assert(req.bytes <= bytes);
     if (!req.exists) {
-        if (r == 0) {
+        if (r == 0 && !req.bytes) {
             /*
-             * rbd_diff_iterate2 does not invoke callbacks for unallocated
-             * areas. This here catches the case where no callback was
-             * invoked at all (req.bytes == 0).
+             * rbd_diff_iterate2 does not invoke callbacks for unallocated areas
+             * except for the case where an overlay has a hole where the parent
+             * or an older snapshot of the image has not. This here catches the
+             * case where no callback was invoked at all.
              */
-            assert(req.bytes == 0);
             req.bytes = bytes;
         }
         status = BDRV_BLOCK_ZERO | BDRV_BLOCK_OFFSET_VALID;
--
Quoted 3 lines

2.25.1

Hi Peter,

Quoted 5 lines

Can we just skip these "holes" by replacing the existing assert with an if statement that would simply bail from the callback on !exists?

Just trying to keep the logic as simple as possible since as it turns out we get to contend with ages-old librbd bugs here...

I'm afraid I think this would not work. Consider qemu-img convert.

If we bail out we would immediately call get_block_status with the offset

where we stopped and hit the !exist again.

Peter

#4 lore raw

Re: [PATCH 1/2] block/rbd: fix handling of holes in .bdrv_co_block_status

Quoted 112 lines · Peter Lieven

Am 12.01.22 um 10:05 schrieb Ilya Dryomov:

Quoted 103 lines · Peter Lieven
Quoted 97 lines

the assumption that we can't hit a hole if we do not diff against a snapshot was wrong.

We can see a hole in an image if we diff against base if there exists an older snapshot of the image and we have discarded blocks in the image where the snapshot has data.

Fixes 0347a8fd4c3faaedf119be04c197804be40a384b
---
 block/rbd.c | 55 +++++++++++++++++++++++++++++++++--------------------
 1 file changed, 34 insertions(+), 21 deletions(-)
block/rbd.c +34 -22
index def96292e0..5e9dc91d81 100644
@@ -1279,13 +1279,24 @@ static int qemu_rbd_diff_iterate_cb(uint64_t offs, size_t len,
     RBDDiffIterateReq *req = opaque;
 
     assert(req->offs + req->bytes <= offs);
-    /*
-     * we do not diff against a snapshot so we should never receive a callback
-     * for a hole.
-     */
-    assert(exists);
 
-    if (!req->exists && offs > req->offs) {
+    if (req->exists && offs > req->offs + req->bytes) {
+        /*
+         * we started in an allocated area and jumped over an unallocated area,
+         * req->bytes contains the length of the allocated area before the
+         * unallocated area. stop further processing.
+         */
+        return QEMU_RBD_EXIT_DIFF_ITERATE2;
+    }
+    if (req->exists && !exists) {
+        /*
+         * we started in an allocated area and reached a hole. req->bytes
+         * contains the length of the allocated area before the hole.
+         * stop further processing.
+         */
+        return QEMU_RBD_EXIT_DIFF_ITERATE2;
+    }
+    if (!req->exists && exists && offs > req->offs) {
         /*
          * we started in an unallocated area and hit the first allocated
          * block. req->bytes must be set to the length of the unallocated area
@@ -1295,17 +1306,19 @@ static int qemu_rbd_diff_iterate_cb(uint64_t offs, size_t len,
         return QEMU_RBD_EXIT_DIFF_ITERATE2;
     }
 
-    if (req->exists && offs > req->offs + req->bytes) {
-        /*
-         * we started in an allocated area and jumped over an unallocated area,
-         * req->bytes contains the length of the allocated area before the
-         * unallocated area. stop further processing.
-         */
-        return QEMU_RBD_EXIT_DIFF_ITERATE2;
-    }
+    /*
+     * assert that we caught all cases above and allocation state has not
+     * changed during callbacks.
+     */
+    assert(exists == req->exists || !req->bytes);
+    req->exists = exists;
 
-    req->bytes += len;
-    req->exists = true;
+    /*
+     * assert that we either return an unallocated block or have got callbacks
+     * for all allocated blocks present.
+     */
+    assert(!req->exists || offs == req->offs + req->bytes);
+    req->bytes = offs + len - req->offs;
 
     return 0;
 }
@@ -1354,13 +1367,13 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
     }
     assert(req.bytes <= bytes);
     if (!req.exists) {
-        if (r == 0) {
+        if (r == 0 && !req.bytes) {
             /*
-             * rbd_diff_iterate2 does not invoke callbacks for unallocated
-             * areas. This here catches the case where no callback was
-             * invoked at all (req.bytes == 0).
+             * rbd_diff_iterate2 does not invoke callbacks for unallocated areas
+             * except for the case where an overlay has a hole where the parent
+             * or an older snapshot of the image has not. This here catches the
+             * case where no callback was invoked at all.
              */
-            assert(req.bytes == 0);
             req.bytes = bytes;
         }
         status = BDRV_BLOCK_ZERO | BDRV_BLOCK_OFFSET_VALID;
--
Quoted 3 lines

2.25.1

Hi Peter,

Quoted 5 lines

Can we just skip these "holes" by replacing the existing assert with an if statement that would simply bail from the callback on !exists?

Just trying to keep the logic as simple as possible since as it turns out we get to contend with ages-old librbd bugs here...

Quoted 5 lines

I'm afraid I think this would not work. Consider qemu-img convert.

If we bail out we would immediately call get_block_status with the offset

where we stopped and hit the !exist again.

I'm suggesting bailing from the callback (i.e. return 0), not from the entire rbd_diff_iterate2() instance. The iteration would move on and either stumble upon an allocated area within the requested range or run off the end of the requested range. Both of these cases are already handled by the existing code.

Thanks,

Ilya

#5 lore raw

Re: [PATCH 1/2] block/rbd: fix handling of holes in .bdrv_co_block_status

Quoted 120 lines

Am 12.01.2022 um 22:06 schrieb Ilya Dryomov <[email protected]>:

Quoted 116 lines · Peter Lieven
Quoted 105 lines

Am 12.01.22 um 10:05 schrieb Ilya Dryomov:

Quoted 97 lines · Peter Lieven

the assumption that we can't hit a hole if we do not diff against a snapshot was wrong.

We can see a hole in an image if we diff against base if there exists an older snapshot of the image and we have discarded blocks in the image where the snapshot has data.

Fixes 0347a8fd4c3faaedf119be04c197804be40a384b

---
block/rbd.c | 55 +++++++++++++++++++++++++++++++++--------------------

1 file changed, 34 insertions(+), 21 deletions(-)
block/rbd.c +34 -22
index def96292e0..5e9dc91d81 100644
@@ -1279,13 +1279,24 @@ static int qemu_rbd_diff_iterate_cb(uint64_t offs, size_t len,
    RBDDiffIterateReq *req = opaque;
 
    assert(req->offs + req->bytes <= offs);
-    /*
-     * we do not diff against a snapshot so we should never receive a callback
-     * for a hole.
-     */
-    assert(exists);
 
-    if (!req->exists && offs > req->offs) {
+    if (req->exists && offs > req->offs + req->bytes) {
+        /*
+         * we started in an allocated area and jumped over an unallocated area,
+         * req->bytes contains the length of the allocated area before the
+         * unallocated area. stop further processing.
+         */
+        return QEMU_RBD_EXIT_DIFF_ITERATE2;
+    }
+    if (req->exists && !exists) {
+        /*
+         * we started in an allocated area and reached a hole. req->bytes
+         * contains the length of the allocated area before the hole.
+         * stop further processing.
+         */
+        return QEMU_RBD_EXIT_DIFF_ITERATE2;
+    }
+    if (!req->exists && exists && offs > req->offs) {
        /*
         * we started in an unallocated area and hit the first allocated
         * block. req->bytes must be set to the length of the unallocated area
@@ -1295,17 +1306,19 @@ static int qemu_rbd_diff_iterate_cb(uint64_t offs, size_t len,
        return QEMU_RBD_EXIT_DIFF_ITERATE2;
    }
 
-    if (req->exists && offs > req->offs + req->bytes) {
-        /*
-         * we started in an allocated area and jumped over an unallocated area,
-         * req->bytes contains the length of the allocated area before the
-         * unallocated area. stop further processing.
-         */
-        return QEMU_RBD_EXIT_DIFF_ITERATE2;
-    }
+    /*
+     * assert that we caught all cases above and allocation state has not
+     * changed during callbacks.
+     */
+    assert(exists == req->exists || !req->bytes);
+    req->exists = exists;
 
-    req->bytes += len;
-    req->exists = true;
+    /*
+     * assert that we either return an unallocated block or have got callbacks
+     * for all allocated blocks present.
+     */
+    assert(!req->exists || offs == req->offs + req->bytes);
+    req->bytes = offs + len - req->offs;
 
    return 0;
 }
@@ -1354,13 +1367,13 @@ static int coroutine_fn qemu_rbd_co_block_status(BlockDriverState *bs,
    }
    assert(req.bytes <= bytes);
    if (!req.exists) {
-        if (r == 0) {
+        if (r == 0 && !req.bytes) {
            /*
-             * rbd_diff_iterate2 does not invoke callbacks for unallocated
-             * areas. This here catches the case where no callback was
-             * invoked at all (req.bytes == 0).
+             * rbd_diff_iterate2 does not invoke callbacks for unallocated areas
+             * except for the case where an overlay has a hole where the parent
+             * or an older snapshot of the image has not. This here catches the
+             * case where no callback was invoked at all.
             */
-            assert(req.bytes == 0);
            req.bytes = bytes;
        }
        status = BDRV_BLOCK_ZERO | BDRV_BLOCK_OFFSET_VALID;
--
Quoted 3 lines

2.25.1

Hi Peter,

Quoted 5 lines

Can we just skip these "holes" by replacing the existing assert with an if statement that would simply bail from the callback on !exists?

Just trying to keep the logic as simple as possible since as it turns out we get to contend with ages-old librbd bugs here...

I'm afraid I think this would not work. Consider qemu-img convert.

If we bail out we would immediately call get_block_status with the offset

Quoted 7 lines
Quoted 1 line

where we stopped and hit the !exist again.

I'm suggesting bailing from the callback (i.e. return 0), not from the entire rbd_diff_iterate2() instance. The iteration would move on and either stumble upon an allocated area within the requested range or run off the end of the requested range. Both of these cases are already handled by the existing code.

Ah, got it. That’s a smart solution!

Peter