Re: [Qemu-devel] [PATCH v3 10/15] block: add block job transactions

[John Snow <jsnow@redhat.com>]

On 07/13/2015 11:04 PM, Fam Zheng wrote:
> On Mon, 07/13 19:12, John Snow wrote:
>>
>>
>> On 07/09/2015 11:46 PM, Fam Zheng wrote:
>>> From: Stefan Hajnoczi <stefanha@redhat.com>
>>>
>>> Sometimes block jobs must execute as a transaction group.  Finishing
>>> jobs wait until all other jobs are ready to complete successfully.
>>> Failure or cancellation of one job cancels the other jobs in the group.
>>>
>>> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
>>> [Rewrite the implementation which is now contained in block_job_completed.
>>> --Fam]
>>> Signed-off-by: Fam Zheng <famz@redhat.com>
>>> ---
>>>  blockjob.c               | 86 ++++++++++++++++++++++++++++++++++++++++++++++++
>>>  include/block/block.h    |  1 +
>>>  include/block/blockjob.h | 26 +++++++++++++++
>>>  3 files changed, 113 insertions(+)
>>>
>>> diff --git a/blockjob.c b/blockjob.c
>>> index e057dd5..7b59b53 100644
>>> --- a/blockjob.c
>>> +++ b/blockjob.c
>>> @@ -36,6 +36,16 @@
>>>  #include "qemu/timer.h"
>>>  #include "qapi-event.h"
>>>  
>>> +/* Transactional group of block jobs */
>>> +struct BlockJobTxn {
>>> +
>>> +    /* Is this txn being cancelled? */
>>> +    bool aborting;
>>> +
>>> +    /* List of jobs */
>>> +    QLIST_HEAD(, BlockJob) jobs;
>>> +};
>>> +
>>>  void *block_job_create(const BlockJobDriver *driver, BlockDriverState *bs,
>>>                         int64_t speed, BlockCompletionFunc *cb,
>>>                         void *opaque, Error **errp)
>>> @@ -84,6 +94,59 @@ void block_job_release(BlockDriverState *bs)
>>>      g_free(job);
>>>  }
>>>  
>>> +static void block_job_completed_txn(BlockJobTxn *txn, BlockJob *job, int ret)
>>> +{
>>> +    AioContext *ctx;
>>> +    BlockJob *other_job, *next;
>>> +    if (ret < 0 || block_job_is_cancelled(job)) {
>>> +        if (!txn->aborting) {
>>> +            /* We are the first failed job. Cancel other jobs. */
>>
>> What guarantee against a race is there at this point? Something to do
>> with deferring back to the main loop before we call completion?
> 
> This is always in the main thread, so there can't be any race.
>>
>> Do jobs always complete in the main thread?
> 
> Yes.
> 

Thought as much, now I know, thanks.

>>
>>> +            txn->aborting = true;
>>> +            QLIST_FOREACH_SAFE(other_job, &txn->jobs, txn_list, next) {
>>> +                if (other_job == job || other_job->completed) {
>>> +                    continue;
>>> +                }
>>> +                ctx = bdrv_get_aio_context(other_job->bs);
>>> +                aio_context_acquire(ctx);
>>> +                block_job_cancel_sync(other_job);
>>> +                assert(other_job->completed);
>>> +                aio_context_release(ctx);
>>> +            }
>>> +            QLIST_FOREACH_SAFE(other_job, &txn->jobs, txn_list, next) {
>>> +                if (other_job->driver->txn_abort) {
>>> +                    other_job->driver->txn_abort(other_job);
>>> +                }
>>> +                other_job->cb(other_job->opaque,
>>> +                              other_job->ret ? : -ECANCELED);
>>> +                block_job_release(other_job->bs);
>>> +            }
>>> +        } else {
>>> +            /*
>>> +             * We are cancelled by another job, who will handle everything.
>>> +             */
>>> +            return;
>>> +        }
>>> +    } else {
>>> +        /*
>>> +         * Successful completion, see if there is other running jobs in this
>>> +         * txn. */
>>
>> s/is/are/
> 
> OK.
> 
>>
>>> +        QLIST_FOREACH(other_job, &txn->jobs, txn_list) {
>>> +            if (!other_job->completed) {
>>> +                return;
>>> +            }
>>> +        }
>>
>> Clever (though quadratic!) :~)
> 
> If someone suggest there could be 1000+ jobs in a txn, I can add a counter.
> 

My idea of a joke. Forgive me :)

>>
>>> +        /* We are the last completed job, commit the transaction. */
>>> +        QLIST_FOREACH_SAFE(other_job, &txn->jobs, txn_list, next) {
>>> +            if (other_job->driver->txn_commit) {
>>> +                other_job->driver->txn_commit(other_job);
>>> +            }
>>> +            assert(other_job->ret == 0);
>>> +            other_job->cb(other_job->opaque, 0);
>>> +            block_job_release(other_job->bs);
>>
>> Worth creating some static local/inline routine and sharing it with
>> block_job_completed?
> 
> Mayb when a 3rd common line comes?

Fine by me, just my habit of asking out loud. The answer can always
simply be "I don't feel like it."

>>
>>> +        }
>>> +    }
>>> +}
>>> +
>>>  void block_job_completed(BlockJob *job, int ret)
>>>  {
>>>      BlockDriverState *bs = job->bs;
>>> @@ -96,6 +159,10 @@ void block_job_completed(BlockJob *job, int ret)
>>>          qemu_bh_delete(job->defer_to_main_loop_data.bh);
>>>          job->defer_to_main_loop_data.bh = NULL;
>>>      }
>>> +    if (job->txn) {
>>> +        block_job_completed_txn(job->txn, job, ret);
>>> +        return;
>>> +    }
>>>      job->cb(job->opaque, ret);
>>>      block_job_release(bs);
>>>  }
>>> @@ -384,3 +451,22 @@ void block_job_defer_to_main_loop(BlockJob *job,
>>>  
>>>      qemu_bh_schedule(data->bh);
>>>  }
>>> +
>>> +BlockJobTxn *block_job_txn_new(void)
>>> +{
>>> +    BlockJobTxn *txn = g_new0(BlockJobTxn, 1);
>>> +    QLIST_INIT(&txn->jobs);
>>> +    return txn;
>>> +}
>>> +
>>> +void block_job_txn_add_job(BlockJobTxn *txn, BlockJob *job)
>>> +{
>>> +    if (!txn) {
>>> +        return;
>>> +    }
>>> +
>>> +    assert(!job->txn);
>>> +    job->txn = txn;
>>> +
>>> +    QLIST_INSERT_HEAD(&txn->jobs, job, txn_list);
>>> +}
>>> diff --git a/include/block/block.h b/include/block/block.h
>>> index 7437590..c7fc5b6 100644
>>> --- a/include/block/block.h
>>> +++ b/include/block/block.h
>>> @@ -13,6 +13,7 @@
>>>  typedef struct BlockDriver BlockDriver;
>>>  typedef struct BlockJob BlockJob;
>>>  typedef struct BdrvChildRole BdrvChildRole;
>>> +typedef struct BlockJobTxn BlockJobTxn;
>>>  
>>>  typedef struct BlockDriverInfo {
>>>      /* in bytes, 0 if irrelevant */
>>> diff --git a/include/block/blockjob.h b/include/block/blockjob.h
>>> index 5bac2e2..d854ee0 100644
>>> --- a/include/block/blockjob.h
>>> +++ b/include/block/blockjob.h
>>> @@ -157,6 +157,9 @@ struct BlockJob {
>>>       */
>>>      int ret;
>>>  
>>> +    /** Non-NULL if this job is part of a transaction */
>>> +    BlockJobTxn *txn;
>>> +    QLIST_ENTRY(BlockJob) txn_list;
>>>  };
>>>  
>>>  /**
>>> @@ -389,4 +392,27 @@ void block_job_defer_to_main_loop(BlockJob *job,
>>>                                    BlockJobDeferToMainLoopFn *fn,
>>>                                    void *opaque);
>>>  
>>> +/**
>>> + * block_job_txn_new:
>>> + *
>>> + * Allocate and return a new block job transaction.  Jobs can be added to the
>>> + * transaction using block_job_txn_add_job().
>>> + *
>>> + * All jobs in the transaction either complete successfully or fail/cancel as a
>>> + * group.  Jobs wait for each other before completing.  Cancelling one job
>>> + * cancels all jobs in the transaction.
>>> + */
>>> +BlockJobTxn *block_job_txn_new(void);
>>> +
>>> +/**
>>> + * block_job_txn_add_job:
>>> + * @txn: The transaction (may be NULL)
>>> + * @job: Job to add to the transaction
>>> + *
>>> + * Add @job to the transaction.  The @job must not already be in a transaction.
>>> + * The block job driver must call block_job_txn_prepare_to_complete() before
>>> + * final cleanup and completion.
>>> + */
>>> +void block_job_txn_add_job(BlockJobTxn *txn, BlockJob *job);
>>> +
>>>  #endif
>>>
>>
>> This definitely combines the best of our approaches AND has less code to
>> boot. Great!
>>
>> Nitpicks and comment spelling aside:
>> Reviewed-by: John Snow <jsnow@redhat.com>
> 
> Thanks!
> 
>>
>> --
>> (linguistic postscript: I don't know how to say "Stefan and I's patches"
>> because I think 'I' is incorrect here because the subject of the
>> sentence is the patches, not Stefan or myself. I think it's some variant
>> on "me" "my" or "mine," but can't quite work it out. English is hard, or
>> I'm stupid, or both.)
>>
> 
> LOL!
> 
> Fam
>