Re: [Qemu-block] [PATCH v2] block: Fix bdrv_drain in coroutine

[Paolo Bonzini]

On 04/04/2016 13:57, Stefan Hajnoczi wrote:
> On Fri, Apr 01, 2016 at 09:57:38PM +0800, Fam Zheng wrote:
>> Using the nested aio_poll() in coroutine is a bad idea. This patch
>> replaces the aio_poll loop in bdrv_drain with a BH, if called in
>> coroutine.
>>
>> For example, the bdrv_drain() in mirror.c can hang when a guest issued
>> request is pending on it in qemu_co_mutex_lock().
>>
>> Mirror coroutine in this case has just finished a request, and the block
>> job is about to complete. It calls bdrv_drain() which waits for the
>> other coroutine to complete. The other coroutine is a scsi-disk request.
>> The deadlock happens when the latter is in turn pending on the former to
>> yield/terminate, in qemu_co_mutex_lock(). The state flow is as below
>> (assuming a qcow2 image):
>>
>>   mirror coroutine               scsi-disk coroutine
>>   -------------------------------------------------------------
>>   do last write
>>
>>     qcow2:qemu_co_mutex_lock()
>>     ...
>>                                  scsi disk read
>>
>>                                    tracked request begin
>>
>>                                    qcow2:qemu_co_mutex_lock.enter
>>
>>     qcow2:qemu_co_mutex_unlock()
>>
>>   bdrv_drain
>>     while (has tracked request)
>>       aio_poll()
>>
>> In the scsi-disk coroutine, the qemu_co_mutex_lock() will never return
>> because the mirror coroutine is blocked in the aio_poll(blocking=true).
>>
>> With this patch, the added qemu_coroutine_yield() allows the scsi-disk
>> coroutine to make progress as expected:
>>
>>   mirror coroutine               scsi-disk coroutine
>>   -------------------------------------------------------------
>>   do last write
>>
>>     qcow2:qemu_co_mutex_lock()
>>     ...
>>                                  scsi disk read
>>
>>                                    tracked request begin
>>
>>                                    qcow2:qemu_co_mutex_lock.enter
>>
>>     qcow2:qemu_co_mutex_unlock()
>>
>>   bdrv_drain.enter
>>>   schedule BH
>>>   qemu_coroutine_yield()
>>>                                  qcow2:qemu_co_mutex_lock.return
>>>                                  ...
>>                                    tracked request end
>>     ...
>>     (resumed from BH callback)
>>   bdrv_drain.return
>>   ...
>>
>> Reported-by: Laurent Vivier <address@hidden>
>> Suggested-by: Paolo Bonzini <address@hidden>
>> Signed-off-by: Fam Zheng <address@hidden>
>>
>> ---
>>
>> v2: Call qemu_bh_delete() in BH callback. [Paolo]
>>     Change the loop to an assertion. [Paolo]
>>     Elaborate a bit about the fix in commit log. [Paolo]
>> ---
>>  block/io.c | 41 +++++++++++++++++++++++++++++++++++++++++
>>  1 file changed, 41 insertions(+)
>>
>> diff --git a/block/io.c b/block/io.c
>> index c4869b9..ddcfb4c 100644
>> --- a/block/io.c
>> +++ b/block/io.c
>> @@ -253,6 +253,43 @@ static void bdrv_drain_recurse(BlockDriverState *bs)
>>      }
>>  }
>>  
>> +typedef struct {
>> +    Coroutine *co;
>> +    BlockDriverState *bs;
>> +    QEMUBH *bh;
>> +    bool done;
>> +} BdrvCoDrainData;
>> +
>> +static void bdrv_co_drain_bh_cb(void *opaque)
>> +{
>> +    BdrvCoDrainData *data = opaque;
>> +    Coroutine *co = data->co;
>> +
>> +    qemu_bh_delete(data->bh);
>> +    bdrv_drain(data->bs);
>> +    data->done = true;
>> +    qemu_coroutine_enter(co, NULL);
>> +}
>> +
>> +static void coroutine_fn bdrv_co_drain(BlockDriverState *bs)
>> +{
> 
> Please document why a BH is needed:
> 
> /* Calling bdrv_drain() from a BH ensures the
>  * current coroutine yields and other coroutines run if they were
>  * queued from qemu_co_queue_run_restart().
>  */
> 
>> +    BdrvCoDrainData data;
>> +
>> +    assert(qemu_in_coroutine());
>> +    data = (BdrvCoDrainData) {
>> +        .co = qemu_coroutine_self(),
>> +        .bs = bs,
>> +        .done = false,
>> +        .bh = aio_bh_new(bdrv_get_aio_context(bs), bdrv_co_drain_bh_cb, 
>> &data),
>> +    };
>> +    qemu_bh_schedule(data.bh);
>> +
>> +    qemu_coroutine_yield();
>> +    /* If we are resumed from some other event (such as an aio completion 
>> or a
>> +     * timer callback), it is a bug in the caller that should be fixed. */
>> +    assert(data.done);
>> +}
>> +
>>  /*
>>   * Wait for pending requests to complete on a single BlockDriverState 
>> subtree,
>>   * and suspend block driver's internal I/O until next request arrives.
>> @@ -269,6 +306,10 @@ void bdrv_drain(BlockDriverState *bs)
>>      bool busy = true;
>>  
>>      bdrv_drain_recurse(bs);
>> +    if (qemu_in_coroutine()) {
>> +        bdrv_co_drain(bs);
>> +        return;
>> +    }
>>      while (busy) {
>>          /* Keep iterating */
>>           bdrv_flush_io_queue(bs);
>> -- 
>> 2.7.4
> 
> block/mirror.c should call bdrv_co_drain() explicitly and bdrv_drain()
> should assert(!qemu_in_coroutine()).
> 
> The reason why existing bdrv_read() and friends detect coroutine context
> at runtime is because it is meant for legacy code that runs in both
> coroutine and non-coroutine contexts.
> 
> Modern coroutine code coroutine code the coroutine interfaces explicitly
> instead.

For what it's worth, I suspect Fam's patch removes the need for
http://article.gmane.org/gmane.comp.emulators.qemu/401375.  That's a
nice bonus. :)

Paolo