Am 23/05/2022 um 15:15 schrieb Stefan Hajnoczi:
> On Mon, May 23, 2022 at 10:48:48AM +0200, Emanuele Giuseppe Esposito wrote:
>>
>>
>> Am 22/05/2022 um 17:06 schrieb Stefan Hajnoczi:
>>> On Wed, May 18, 2022 at 06:14:17PM +0200, Kevin Wolf wrote:
>>>> Am 18.05.2022 um 14:43 hat Paolo Bonzini geschrieben:
>>>>> On 5/18/22 14:28, Emanuele Giuseppe Esposito wrote:
>>>>>> For example, all callers of bdrv_open() always take the AioContext lock.
>>>>>> Often it is taken very high in the call stack, but it's always taken.
>>>>>
>>>>> I think it's actually not a problem of who takes the AioContext lock or
>>>>> where; the requirements are contradictory:
>>>>>
>>>>> * IO_OR_GS_CODE() functions, when called from coroutine context, expect to
>>>>> be called with the AioContext lock taken (example: bdrv_co_yield_to_drain)
>>>>>
>>>>> * to call these functions with the lock taken, the code has to run in the
>>>>> BDS's home iothread. Attempts to do otherwise results in deadlocks (the
>>>>> main loop's AIO_WAIT_WHILEs expect progress from the iothread, that cannot
>>>>> happen without releasing the aiocontext lock)
>>>>>
>>>>> * running the code in the BDS's home iothread is not possible for
>>>>> GLOBAL_STATE_CODE() functions (unless the BDS home iothread is the main
>>>>> thread, but that cannot be guaranteed in general)
>>>>>
>>>>>> We might suppose that many callbacks are called under drain and in
>>>>>> GLOBAL_STATE, which should be enough, but from our experimentation in
>>>>>> the previous series we saw that currently not everything is under drain,
>>>>>> leaving some operations unprotected (remember assert_graph_writable
>>>>>> temporarily disabled, since drain coverage for bdrv_replace_child_noperm
>>>>>> was not 100%?).
>>>>>> Therefore we need to add more drains. But isn't drain what we decided to
>>>>>> drop at the beginning? Why isn't drain good?
>>>>>
>>>>> To sum up the patch ordering deadlock that we have right now:
>>>>>
>>>>> * in some cases, graph manipulations are protected by the AioContext lock
>>>>>
>>>>> * eliminating the AioContext lock is needed to move callbacks to coroutine
>>>>> contexts (see above for the deadlock scenario)
>>>>>
>>>>> * moving callbacks to coroutine context is needed by the graph rwlock
>>>>> implementation
>>>>>
>>>>> On one hand, we cannot protect the graph across manipulations with a graph
>>>>> rwlock without removing the AioContext lock; on the other hand, the
>>>>> AioContext lock is what _right now_ protects the graph.
>>>>>
>>>>> So I'd rather go back to Emanuele's draining approach. It may not be
>>>>> beautiful, but it allows progress. Once that is in place, we can remove
>>>>> the
>>>>> AioContext lock (which mostly protects virtio-blk/virtio-scsi code right
>>>>> now) and reevaluate our next steps.
>>>>
>>>> If we want to use drain for locking, we need to make sure that drain
>>>> actually does the job correctly. I see two major problems with it:
>>>>
>>>> The first one is that drain only covers I/O paths, but we need to
>>>> protect against _anything_ touching block nodes. This might mean a
>>>> massive audit and making sure that everything in QEMU that could
>>>> possibly touch a block node is integrated with drain.
>>>>
>>>> I think Emanuele has argued before that because writes to the graph only
>>>> happen in the main thread and we believe that currently only I/O
>>>> requests are processed in iothreads, this is safe and we don't actually
>>>> need to audit everything.
>>>>
>>>> This is true as long as the assumption holds true (how do we ensure that
>>>> nobody ever introduces non-I/O code touching a block node in an
>>>> iothread?) and as long as the graph writer never yields or polls. I
>>>> think the latter condition is violated today, a random example is that
>>>> adjusting drain counts in bdrv_replace_child_noperm() does poll. Without
>>>> cooperation from all relevant places, the effectively locked code
>>>> section ends right there, even if the drained section continues. Even if
>>>> we can fix this, verifying that the conditions are met everywhere seems
>>>> not trivial.
>>>>
>>>> And that's exactly my second major concern: Even if we manage to
>>>> correctly implement things with drain, I don't see a way to meaningfully
>>>> review it. I just don't know how to verify with some confidence that
>>>> it's actually correct and covering everything that needs to be covered.
>>>>
>>>> Drain is not really a lock. But if you use it as one, the best it can
>>>> provide is advisory locking (callers, inside and outside the block
>>>> layer, need to explicitly support drain instead of having the lock
>>>> applied somewhere in the block layer functions). And even if all
>>>> relevant pieces actually make use of it, it still has an awkward
>>>> interface for locking:
>>>>
>>>> /* Similar to rdlock(), but doesn't wait for writers to finish. It is
>>>> * the callers responsibility to make sure that there are no writers. */
>>>> bdrv_inc_in_flight()
>>>>
>>>> /* Similar to wrlock(). Waits for readers to finish. New readers are not
>>>> * prevented from starting after it returns. Third parties are politely
>>>> * asked not to touch the block node while it is drained. */
>>>> bdrv_drained_begin()
>>>>
>>>> (I think the unlock counterparts actually behave as expected from a real
>>>> lock.)
>>>>
>>>> Having an actual rdlock() (that waits for writers), and using static
>>>> analysis to verify that all relevant places use it (so that wrlock()
>>>> doesn't rely on politely asking third parties to leave the node alone)
>>>> gave me some confidence that we could verify the result.
>>>>
>>>> I'm not sure at all how to achieve the same with the drain interface. In
>>>> theory, it's possible. But it complicates the conditions so much that
>>>> I'm pretty much sure that the review wouldn't only be very time
>>>> consuming, but I would make mistakes during the review, rendering it
>>>> useless.
>>>>
>>>> Maybe throwing some more static analysis on the code can help, not sure.
>>>> It's going to be a bit more complex than with the other approach,
>>>> though.
>>>
>>> Hi,
>>> Sorry for the long email. I've included three topics that may help us
>>> discuss
>>> draining and AioContext removal further.
>>>
>>> The shortcomings of drain
>>> -------------------------
>>> I wanted to explore the logical argument that making graph modifications
>>> within
>>> a drained section is correct:
>>> - Graph modifications and BB/BDS lookup are Global State (GS).
>>> - Graph traversal from a BB/BDS pointer is IO.
>>> - Only one thread executes GS code at a time.
>>> - IO is quiesced within a drained section.
>>> - Therefore a drained section in GS code suspends graph traversal, other
>>> graph
>>> modifications, and BB/BDS lookup.
>>> - Therefore it is safe to modify the graph from a GS drained section.
>>>
>>> However, I hit on a problem that I think Emanuele and Paolo have already
>>> pointed out: draining is GS & IO. This might have worked under the 1
>>> IOThread
>>> model but it does not make sense for multi-queue. It is possible to submit
>>> I/O
>>> requests in drained sections. How can multiple threads be in drained
>>> sections
>>> simultaneously and possibly submit further I/O requests in their drained
>>> sections? Those sections wouldn't be "drained" in any useful sense of the
>>> word.
>>>
>>> It is necessary to adjust how recursive drained sections work:
>>> bdrv_drained_begin() must not return while there are deeper nested drained
>>> sections.
>>>
>>> This is allowed:
>>>
>>> Monitor command Block job
>>> --------------- ---------
>>> > bdrv_drained_begin()
>>> . > bdrv_drained_begin()
>>> . < bdrv_drained_begin()
>>> . .
>>> . .
>>> . > bdrv_drained_end()
>>> . < bdrv_drained_end()
>>> < bdrv_drained_begin()
>>> .
>>> .
>>> > bdrv_drained_end()
>>> < bdrv_drained_end()
>>>
>>> This is not allowed:
>>>
>>> Monitor command Block job
>>> --------------- ---------
>>> > bdrv_drained_begin()
>>> . > bdrv_drained_begin()
>>> . < bdrv_drained_begin()
>>> . .
>>> . .
>>> < bdrv_drained_begin() .
>>> . .
>>> . > bdrv_drained_end()
>>> . < bdrv_drained_end()
>>> > bdrv_drained_end()
>>> < bdrv_drained_end()
>>>
>>> This restriction creates an ordering between bdrv_drained_begin() callers.
>>> In
>>> this example the block job must not depend on the monitor command completing
>>> first. Otherwise there would be a deadlock just like with two threads wait
>>> for
>>> each other while holding a mutex.
>>>
>>> For sanity I think draining should be GS-only. IO code should not invoke
>>> bdrv_drained_begin() to avoid ordering problems when multiple threads drain
>>> simultaneously and have dependencies on each other.
>>>
>>> block/mirror.c needs to be modified because it currently drains from IO when
>>> mirroring is about to end.
>>
>> Yes, mirror seems to be the only clear place where draining is performed
>> from IO, namely in mirror_run. It's also a little bit weird because just
>> drained_begin is invoked from IO, while drained_end is in the main loop.
>>
>> If I understand correctly, the reason for that is that we want to
>> prevent bdrv_get_dirty_count to be modified (ie that additional
>> modifications come) after we just finished mirroring.
>>
>> Not sure how to deal with this though.
>
> I don't know the code so I don't have a concrete solution in mind.
>
>>>
>>> With this change to draining I think the logical argument for correctness
>>> with
>>> graph modifications holds.
>>>
>>> Enforcing GS/IO separation at compile time
>>> ------------------------------------------
>>> Right now GS/IO asserts check assumptions at runtime. The next step may be
>>> using the type system to separate GS and IO APIs so it's impossible for IO
>>> code
>>> to accidentally call GS code, for example.
>>>
>>> typedef struct {
>>> BlockDriverState *bs;
>>> } BlockDriverStateIOPtr;
>>>
>>> typedef struct {
>>> BlockDriverState *bs;
>>> } BlockDriverStateGSPtr;
>>>
>>> Then APIs can be protected as follows:
>>>
>>> void bdrv_set_aio_context_ignore(BlockDriverStateGSPtr bs, ...);
>>>
>>> A function that only has a BlockDriverStateIOPtr cannot call
>>> bdrv_set_aio_context_ignore() by mistake since the compiler will complain
>>> that
>>> the first argument must be a BlockDriverStateGSPtr.
>>
>> What about functions that do not need a BlockDriverState *, and for
>> example they use BdrvChild? I would assume that we need to replicate it
>> also for that.
>>
>> And what about GS & IO functions? Not only drain, but also all the
>> generated_co_wrapper should be GS & IO.
>
> What is the meaning of GS & IO in a multi-queue world? I guess it's
> effectively IO except it's well-behaved if called with the BQL held?
Good question. I think so. But will need to be protected, as they will
be probably graph traversing while being either GS or IO.
>
>>
>>>
>>> Possible steps for AioContext removal
>>> -------------------------------------
>>> I also wanted to share my assumptions about multi-queue and AioContext
>>> removal.
>>> Please let me know if anything seems wrong or questionable:
>>>
>>> - IO code can execute in any thread that has an AioContext.
>>> - Multiple threads may execute a IO code at the same time.
>>> - GS code only execute under the BQL.
>>>
>>> For AioContext removal this means:
>>>
>>> - bdrv_get_aio_context() becomes mostly meaningless since there is no need
>>> for
>>> a special "home" AioContext.
>>
>> Makes sense
>>
>>> - bdrv_coroutine_enter() becomes mostly meaningless because there is no
>>> need to
>>> run a coroutine in the BDS's AioContext.
>>
>> Why is there no need?
>
> The coroutine can execute in the current thread since the BDS must be
> thread-safe.
>
> If bdrv_coroutine_enter() is used in cases like block jobs to mean "run
> in the same AioContext as the job coroutine" then the code should
> probably be changed to explicitly use the job AioContext instead of the
> more vague BDS AioContext.
Ok makes sense
>
>>> - aio_disable_external(bdrv_get_aio_context(bs)) no longer works because
>>> many
>>> threads/AioContexts may submit new I/O requests.
>>> BlockDevOps.drained_begin()
>>> may be used instead (e.g. to temporarily disable ioeventfds on a
>>> multi-queue
>>> virtio-blk device).
>>
>> Ok
>>
>>> - AIO_WAIT_WHILE() simplifies to
>>>
>>> while ((cond)) {
>>> aio_poll(qemu_get_current_aio_context(), true);
>>> ...
>>> }
>>>
>>> and the distinction between home AioContext and non-home context is
>>> eliminated. AioContext unlocking is dropped.
>>
>> I guess this implies no coroutine in BDS's AioContext.
>
> I'm not sure what you mean but coroutines accessing the BDS can run in
> any AioContext with multi-queue. If they have any locking requirements
> amongst each other then that should be explicit instead of just throwing
> them all into the BDS AioContext.
Ok
>
>>>
>>> Does this make sense? I haven't seen these things in recent patch series.
>>
>> You haven't seen them because there's no way to do it if we don't even
>> agree on what to replace the AioContext lock with.
>> Getting to this point would imply firstly having something (drain,
>> rwlock, whatever) together with AioContext, and then gradually remove
>> it. At least, that's how I understood it.
>
> What, besides graph modification and the things I listed in this email,
> still needs to be untangled from AioContext?
Nothing else (I hope), but as Kevin also said, this requires a lot of
preliminary work before getting to this state.
Emanuele
>
> Stefan
>