[libvirt] [PATCH v3 08/18] snapshot: Avoid latent use-after-free when cleaning snapshots

[Eric Blake]

On 3/7/19 8:02 AM, John Ferlan wrote:
> 
> 
> On 3/4/19 10:34 PM, Eric Blake wrote:
>> Right now, the only callers of qemuDomainSnapshotDiscardAllMetadata()
>> are right before freeing the virDomainSnapshotObjList, so it did not
>> matter if the list's metaroot (which points to all the defined root
>> snapshots) is left inconsistent. But an upcoming patch will want to
>> clear all snapshots if a bulk redefine fails partway through, in
>> which case things must be reset.  Make this work by teaching the
>> existing virDomainSnapshotUpdateRelations() to be safe regardless of
>> the incoming state of the metaroot (since we don't want to leak that
>> internal detail into qemu code), then fixing the qemu code to use
>> it after deleting all snapshots.
>>
>> Signed-off-by: Eric Blake <eblake at redhat.com>
>> ---
>>  src/conf/snapshot_conf.c | 7 +++++--
>>  src/qemu/qemu_domain.c   | 4 ++++
>>  2 files changed, 9 insertions(+), 2 deletions(-)
>>
> 
> I have some random personal grumbling about the lack of comments for
> existing code. What's described in this patch as being done would seem
> correct, but my knowledge of snapshot's is 'cursory'.

Yeah, and you already tasked me with refactoring virDomainSnapshotObj
and virDomainCheckpointObj to share a common, better-commented, parent
class. So hopefully there are patches in the next few days to improve
the situation (or at least isolate it into one file instead of the
current leaky abstraction where src/qemu is poking at .first_child
contents in the first place).

> 
>> diff --git a/src/conf/snapshot_conf.c b/src/conf/snapshot_conf.c
>> index 206b05c172..386ec82d15 100644
>> --- a/src/conf/snapshot_conf.c
>> +++ b/src/conf/snapshot_conf.c
>> @@ -1209,13 +1209,16 @@ virDomainSnapshotSetRelations(void *payload,
>>  }
>>
>>  /* Populate parent link and child count of all snapshots, with all
>> - * relations starting as 0/NULL.  Return 0 on success, -1 if a parent
>> - * is missing or if a circular relationship was requested.  */
>> + * assigned defs having relations starting as 0/NULL.  Return 0 on
>> + * success, -1 if a parent is missing or if a circular relationship
>> + * was requested.  */
>                     ^^
> nit: there's an extra space here

Pre-existing, and habitual, but I'll clean up the ones you or I notice.


>> +++ b/src/qemu/qemu_domain.c
>> @@ -8625,6 +8625,10 @@ qemuDomainSnapshotDiscardAllMetadata(virQEMUDriverPtr driver,
>>      rem.err = 0;
>>      virDomainSnapshotForEach(vm->snapshots, qemuDomainSnapshotDiscardAll,
>>                               &rem);
>> +    if (rem.current)
>> +        vm->current_snapshot = NULL;
> 
> If rem.err != 0, is this still something that should be done

Yes. rem.current is set only if we removed the current snapshot's
metadata; even if we later failed halfway through before removing ALL
snapshots, keeping the current_snapshot pointer around would be pointing
into free'd memory.  So the pointer must be cleared whether we succeed
or fail to match that we did remove that metadata (and since we don't
have any good way to undo a partial failure).

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 488 bytes
Desc: OpenPGP digital signature
URL: <http://listman.redhat.com/archives/libvir-list/attachments/20190307/165a6e55/attachment-0001.sig>