Am 09.06.2017 um 16:16 hat Eric Blake geschrieben: > On 06/09/2017 07:58 AM, Kevin Wolf wrote: > > Am 09.06.2017 um 14:14 hat Eric Blake geschrieben: > >> On 06/09/2017 06:50 AM, Kevin Wolf wrote: > >>> When qemu is exited, all running jobs should be cancelled successfully. > >>> This adds a test for this for all types of block jobs that currently > >>> exist in qemu. > >>> > >>> Signed-off-by: Kevin Wolf <kw...@redhat.com> > > >>> +{"timestamp": {"seconds": TIMESTAMP, "microseconds": TIMESTAMP}, > >>> "event": "SHUTDOWN", "data": {"guest": false}} > >>> +{"timestamp": {"seconds": TIMESTAMP, "microseconds": TIMESTAMP}, > >>> "event": "BLOCK_JOB_CANCELLED", "data": {"device": "disk", "len": > >>> 67108864, "offset": 524288, "speed": 65536, "type": "commit"}} > >> > >> I'm worried that if you don't sanitize at least offset, you will still > >> be prone to some race conditions changing the output. You may want to > >> add in some additional filtering on the output to be safer. > > > > I considered that at first, but then I realised that these offsets are > > indeed predictable and we want to know if they change (it would likely > > mean that the throttling is broken). > > > > If you look at the individual cases, we have: > > > > * offset=512k for (intermediate) commit and streaming. This is exactly > > the buffer size for a single request and will be followed by a delay > > of eight seconds before the next chunk is copied, so we will never get > > a different value here. > > > > * offset=4M for active commit and mirror, because the mirror job has a > > larger buffer size by default, so one request completes it all. This > > number is already the maximum, so nothing is going to change here > > either. > > > > * offset=64k for backup, which works cluster by cluster. We know that > > the cluster size is exactly 64k, and while we have only one second > > of delay here, that's still plenty of time for the 'quit' command to > > arrive. > > These belong in comments in the test proper, because it is not obvious > otherwise. But with comments added (so someone debugging a theoretical > test failure down the road knows what they are up against),
# Note that the reference output intentionally includes the 'offset' field in # BLOCK_JOB_CANCELLED events for all of the following block jobs. They are # predictable and any change in the offsets would hint at a bug in the job # throttling code. # # In order to achieve these predictable offsets, all of the following tests # use speed=65536. Each job will perform exactly one iteration before it has # to sleep at least for a second, which is plenty of time for the 'quit' QMP # command to be received (after receiving the command, the rest runs # synchronously, so jobs can arbitrarily continue or complete). # # The buffer size for commit and streaming is 512k (waiting for 8 seconds after # the first request), for active commit and mirror it's large enough to cover # the full 4M, and for backup it's the qcow2 cluster size, which we know is # 64k. As all of these are at least as large as the speed, we are sure that the # offset doesn't advance after the first iteration before qemu exits. Does this look okay? Kevin
pgpxbq2jhFNKP.pgp
Description: PGP signature