[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [PATCH v9 19/21] replay: add BH oneshot event for block
Re: [Qemu-devel] [PATCH v9 19/21] replay: add BH oneshot event for block layer
Mon, 14 Jan 2019 14:48:11 +0300
> From: Kevin Wolf [mailto:address@hidden
> Am 14.01.2019 um 12:10 hat Pavel Dovgalyuk geschrieben:
> > > From: Kevin Wolf [mailto:address@hidden
> > > Am 09.01.2019 um 13:13 hat Pavel Dovgalyuk geschrieben:
> > > > Replay is capable of recording normal BH events, but sometimes
> > > > there are single use callbacks scheduled with aio_bh_schedule_oneshot
> > > > function. This patch enables recording and replaying such callbacks.
> > > > Block layer uses these events for calling the completion function.
> > > > Replaying these calls makes the execution deterministic.
> > > >
> > > > Signed-off-by: Pavel Dovgalyuk <address@hidden>
> > >
> > > This still doesn't come even close to catching all BHs that need to be
> > > caught. While you managed to show a few BHs that actually don't need to
> > > be considered for recording when I asked for this in v7, most BHs in the
> > > block layer can in some way lead to device callbacks and must therefore
> > > be recorded.
> > Let's have a brief review. I can change all the places, but how
> > should I make a test case to be sure, that all of them are working ok?
> The list is changing all the time. This is why I am so concerned about
> special-casing a few callers instead of having a generic solution. I
> don't know how we could make sure that we call the right function
> > aio_bh_schedule_oneshot is used in:
> > - blk_abort_aio_request
> > - bdrv_co_yield_to_drain
> > - iscsi_co_generic_cb
> > - nfs_co_generic_cb
> > - null_aio_common
> > - nvme_process_completion
> > - nvme_rw_cb
> > - rbd_finish_aiocb
> > - vxhs_iio_callback
> > - (and couple of others not in the block layer)
> In addition to these, we have at least a few functions that just resume
> block layer coroutines rather than directly scheduling a BH with a
> callback somewhere in the block layer.
> > We must change this call to replay_bh_schedule_oneshot_event when
> > the result of the BH execution affects the replayed guest state
> > (e.g., interrupt request is generated or memory is written)
> > If you think that all of these can do that, then I should change
> > such function calls.
> I haven't reviewed the code, but these names look like all of them can
> eventually call back into the guest devices. They won't do that always,
> but potentially.
Ok, then I'll try making all of them deterministic.
> > > How bad would it be to record some BHs even if recording them isn't
> > > necessary? I'd definitely try to err on the safe side here. Having two
> > > different sets of BH functions, you can't expect that people always use
> > > the right one (especially if you don't even make the existing code base
> > > consistently use the right one intially).
> > There are two possible options:
> > 1. Execution hangs when recording. Kind of deadlock caused by the incorrect
> > management of the events. E.g., adding stopping the VM and trying to
> > flush
> > the block layer queue.
> > 2. Execution hangs when replaying.
> > One of the events that affect the guest state is missed or generated
> > at the other moment (e.g., when BH is not linked to the execution step).
> > Then the guest behaves differently and the order of the events in the log
> > does not match the guest state (e.g., interrupt processing is not
> > matched).
> So basically when you have two events that are kind of nested?
Can you give a concrete example?
> A triggers event A, but in order to complete the operation, you call
> operation B with event B internally, which isn't available yet because
> we're still handling event A?
> Could this be solved by not having an order of events, but an order of
> sets of events?
There are two kind of operations: synchronous and not.
E.g., querying the disk contents through the PIO waits until the disk operation
is finished, and does not need other synchronization. DMA operations are
by coroutines and bottom halves. Synchronization of these objects should be
recorded, because the moment of invocation may be different, and it affects
the guest state.
[Qemu-devel] [PATCH v9 20/21] replay: init rtc after enabling the replay, Pavel Dovgalyuk, 2019/01/09
[Qemu-devel] [PATCH v9 21/21] replay: document development rules, Pavel Dovgalyuk, 2019/01/09
Re: [Qemu-devel] [PATCH v9 00/21] Fixing record/replay and adding reverse debugging, no-reply, 2019/01/13
- Re: [Qemu-devel] [PATCH v9 13/21] replay: implement replay-seek command to proceed to the desired step, (continued)
- [Qemu-devel] [PATCH v9 14/21] replay: refine replay-time module, Pavel Dovgalyuk, 2019/01/09
- [Qemu-devel] [PATCH v9 15/21] replay: flush rr queue before loading the vmstate, Pavel Dovgalyuk, 2019/01/09
- [Qemu-devel] [PATCH v9 16/21] gdbstub: add reverse step support in replay mode, Pavel Dovgalyuk, 2019/01/09
- [Qemu-devel] [PATCH v9 17/21] gdbstub: add reverse continue support in replay mode, Pavel Dovgalyuk, 2019/01/09
- [Qemu-devel] [PATCH v9 18/21] replay: describe reverse debugging in docs/replay.txt, Pavel Dovgalyuk, 2019/01/09
- [Qemu-devel] [PATCH v9 19/21] replay: add BH oneshot event for block layer, Pavel Dovgalyuk, 2019/01/09