block/copy-on-read: Fix permissions for inactive node
PATCH 1/1 block/copy-on-read: Fix permissions for inactive node
The copy-on-read drive must not request the WRITE_UNCHANGED permission for its child if the node is inactive, otherwise starting a migration destination with -incoming will fail because the child cannot provide write access yet:
qemu-system-x86_64: -blockdev copy-on-read,file=img,node-name=cor: Block node is read-only
Earlier QEMU versions additionally ran into an abort() on the migration source side: bdrv_inactivate_recurse() failed to update permissions. This is silently ignored today because it was only supposed to loosen restrictions. This is the symptom that was originally reported here:
https://bugzilla.redhat.com/show_bug.cgi?id=1733022
--- block/copy-on-read.c | 16 +++++++--------- 1 file changed, 7 insertions(+), 9 deletions(-)
block/copy-on-read.c +7 -9
@@ -56,16 +56,14 @@ static void cor_child_perm(BlockDriverState *bs, BdrvChild *c, uint64_t perm, uint64_t shared, uint64_t *nperm, uint64_t *nshared) { - if (c == NULL) { - *nperm = (perm &PERM_PASSTHROUGH ) |BLK_PERM_WRITE_UNCHANGED ; - *nshared = (shared &PERM_PASSTHROUGH ) |PERM_UNCHANGED ; -return ; - } + *nperm = perm &PERM_PASSTHROUGH ; + *nshared = (shared &PERM_PASSTHROUGH ) |PERM_UNCHANGED ; - *nperm = (perm &PERM_PASSTHROUGH ) | - (c->perm &PERM_UNCHANGED ); - *nshared = (shared &PERM_PASSTHROUGH ) | - (c->shared_perm &PERM_UNCHANGED ); +/* We must not request write permissions for an inactive node, the child +* cannot provide it. */ +if (!(bs->open_flags &BDRV_O_INACTIVE )) { + *nperm |=BLK_PERM_WRITE_UNCHANGED ; + } }
Signature · 1 line
2.20.1
Quoted 12 lines · Kevin Wolf
The copy-on-read drive must not request the
WRITE_UNCHANGEDpermission for its child if the node is inactive, otherwise starting a migration destination with -incoming will fail because the child cannot provide write access yet:qemu-system-x86_64: -blockdev copy-on-read,file=img,node-name=cor: Block node is read-only
Earlier QEMU versions additionally ran into an
abort()on the migration source side:bdrv_inactivate_recurse()failed to update permissions. This is silently ignored today because it was only supposed to loosen restrictions. This is the symptom that was originally reported here:
Quoted 6 lines
https://bugzilla.redhat.com/show_bug.cgi?id=1733022
Signed-off-by Kevin Wolf <[email protected]>--- block/copy-on-read.c | 16 +++++++--------- 1 file changed, 7 insertions(+), 9 deletions(-)
Do any of the iotests cover this? Should they, especially if you are trying to get this in for -rc3 tomorrow?
Quoted 11 lines
block/copy-on-read.c +0
-2@@ -56,16 +56,14 @@ static void cor_child_perm(BlockDriverState *bs, BdrvChild *c, uint64_t perm, uint64_t shared, uint64_t *nperm, uint64_t *nshared) { - if (c == NULL) { - *nperm = (perm &PERM_PASSTHROUGH ) |BLK_PERM_WRITE_UNCHANGED ;
block/copy-on-read.c static void cor_child_perm(BlockDriverState *bs, BdrvChild *c, · line 68
Quoted 10 lines
block/copy-on-read.c +2
-7- *nshared = (shared &PERM_PASSTHROUGH ) |PERM_UNCHANGED ; -return ; - } + *nperm = perm &PERM_PASSTHROUGH ; + *nshared = (shared &PERM_PASSTHROUGH ) |PERM_UNCHANGED ; - *nperm = (perm &PERM_PASSTHROUGH ) | - (c->perm &PERM_UNCHANGED ); - *nshared = (shared &PERM_PASSTHROUGH ) | - (c->shared_perm &PERM_UNCHANGED );
The old code unconditionally returned one set of permissions when c == NULL, or made a choice based on c's existing permissions on whether to pass in those two bits.
block/copy-on-read.c static void cor_child_perm(BlockDriverState *bs, BdrvChild *c, · line 66
Quoted 5 lines
block/copy-on-read.c +5
-0+/* We must not request write permissions for an inactive node, the child +* cannot provide it. */ +if (!(bs->open_flags &BDRV_O_INACTIVE )) { + *nperm |=BLK_PERM_WRITE_UNCHANGED ; + }
The new code changes the condition for or'ing in WRITE_UNCHANGED to *nperm (it is no longer dependent on whether c == NULL, but whether the drive is inactive), which matches your commit message.
But the new code also changes to always pass in the PERM_UNCHANGED to *nshared; that used to be skipped if c was non-NULL and did not already have the permission. I don't follow that change from the commit message, am I missing something?
Signature · 3 lines
Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org
Am 29.07.2019 um 15:35 hat Eric Blake geschrieben:
Quoted 15 lines · Kevin Wolf
Quoted 14 lines
The copy-on-read drive must not request the
WRITE_UNCHANGEDpermission for its child if the node is inactive, otherwise starting a migration destination with -incoming will fail because the child cannot provide write access yet:qemu-system-x86_64: -blockdev copy-on-read,file=img,node-name=cor: Block node is read-only
Earlier QEMU versions additionally ran into an
abort()on the migration source side:bdrv_inactivate_recurse()failed to update permissions. This is silently ignored today because it was only supposed to loosen restrictions. This is the symptom that was originally reported here:
Quoted 7 lines
Quoted 4 lines
Signed-off-by Kevin Wolf <[email protected]>--- block/copy-on-read.c | 16 +++++++--------- 1 file changed, 7 insertions(+), 9 deletions(-)Do any of the iotests cover this? Should they, especially if you are trying to get this in for -rc3 tomorrow?
No, we don't have any iotests for migration with filter drivers yet. We probably should, but I didn't want to miss -rc3 with the fix because I was busy writing a test case.
Quoted 32 lines
Quoted 11 lines
block/copy-on-read.c +0
-2@@ -56,16 +56,14 @@ static void cor_child_perm(BlockDriverState *bs, BdrvChild *c, uint64_t perm, uint64_t shared, uint64_t *nperm, uint64_t *nshared) { - if (c == NULL) { - *nperm = (perm &PERM_PASSTHROUGH ) |BLK_PERM_WRITE_UNCHANGED ;block/copy-on-read.c static void cor_child_perm(BlockDriverState *bs, BdrvChild *c, · line 68
Quoted 10 lines
block/copy-on-read.c +2
-7- *nshared = (shared &PERM_PASSTHROUGH ) |PERM_UNCHANGED ; -return ; - } + *nperm = perm &PERM_PASSTHROUGH ; + *nshared = (shared &PERM_PASSTHROUGH ) |PERM_UNCHANGED ; - *nperm = (perm &PERM_PASSTHROUGH ) | - (c->perm &PERM_UNCHANGED ); - *nshared = (shared &PERM_PASSTHROUGH ) | - (c->shared_perm &PERM_UNCHANGED );The old code unconditionally returned one set of permissions when c == NULL, or made a choice based on c's existing permissions on whether to pass in those two bits.
block/copy-on-read.c static void cor_child_perm(BlockDriverState *bs, BdrvChild *c, · line 66
Quoted 5 lines
block/copy-on-read.c +5
-0+/* We must not request write permissions for an inactive node, the child +* cannot provide it. */ +if (!(bs->open_flags &BDRV_O_INACTIVE )) { + *nperm |=BLK_PERM_WRITE_UNCHANGED ; + }
Quoted 8 lines
The new code changes the condition for or'ing in
WRITE_UNCHANGEDto *nperm (it is no longer dependent on whether c == NULL, but whether the drive is inactive), which matches your commit message.But the new code also changes to always pass in the
PERM_UNCHANGEDto *nshared; that used to be skipped if c was non-NULL and did not already have the permission. I don't follow that change from the commit message, am I missing something?
The old code didn't actually do anything that should have a different result (apart from WRITE_UNCHANGED for inactive images), just everything in a more complicated way for no apparent reason. Or at least that's what Max and I concluded after looking at this.
Taking the PERM_UNCHANGED bits from the old value effectively means that they are taken from the very first call, which had c == NULL. So we can just use the same code to set them instead of referring to the old values of c->perm and c->shared_perm (which is really something a .bdrv_child_perm implementation shouldn't do - there are more cases, but we can clean them up for 4.2).
Not cleaning this up would mean that I'd have to explicitly clear the WRITE_UNCHANGED bit after uselessly copying from the old state. This would be further complication of already unnecessarily complicated code, so I decided that cleaning it up so that its correctness becomes very obvious (request everything the parent nodes need, plus WRITE_UNCHANGED for the copy on read functionality if the node is active) makes more sense.
Kevin
Quoted 12 lines · Kevin Wolf
The copy-on-read drive must not request the
WRITE_UNCHANGEDpermission for its child if the node is inactive, otherwise starting a migration destination with -incoming will fail because the child cannot provide write access yet:qemu-system-x86_64: -blockdev copy-on-read,file=img,node-name=cor: Block node is read-only
Earlier QEMU versions additionally ran into an
abort()on the migration source side:bdrv_inactivate_recurse()failed to update permissions. This is silently ignored today because it was only supposed to loosen restrictions. This is the symptom that was originally reported here:
Quoted 6 lines
https://bugzilla.redhat.com/show_bug.cgi?id=1733022
Signed-off-by Kevin Wolf <[email protected]>--- block/copy-on-read.c | 16 +++++++--------- 1 file changed, 7 insertions(+), 9 deletions(-)