Re: [Qemu-devel] [RFC V8 03/13] quorum: Add quorum_aio_writev and its dependencies.

[Kevin Wolf]

Am 26.09.2013 um 18:29 hat Benoît Canet geschrieben:
> Le Friday 08 Feb 2013 à 11:38:38 (+0100), Kevin Wolf a écrit :
> > Am 28.01.2013 18:07, schrieb Benoît Canet:
> > > Signed-off-by: Benoit Canet <address@hidden>
> > > ---
> > >  block/quorum.c |  111 
> > > ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> > >  1 file changed, 111 insertions(+)
> > > 
> > > diff --git a/block/quorum.c b/block/quorum.c
> > > index d8fffbe..5d8470b 100644
> > > --- a/block/quorum.c
> > > +++ b/block/quorum.c
> > > @@ -52,11 +52,122 @@ struct QuorumAIOCB {
> > >      int vote_ret;
> > >  };
> > >  
> > > +static void quorum_aio_cancel(BlockDriverAIOCB *blockacb)
> > > +{
> > > +    QuorumAIOCB *acb = container_of(blockacb, QuorumAIOCB, common);
> > > +    bool finished = false;
> > > +
> > > +    /* Wait for the request to finish */
> > > +    acb->finished = &finished;
> > > +    while (!finished) {
> > > +        qemu_aio_wait();
> > > +    }
> > > +}
> > > +
> > > +static AIOCBInfo quorum_aiocb_info = {
> > > +    .aiocb_size         = sizeof(QuorumAIOCB),
> > > +    .cancel             = quorum_aio_cancel,
> > > +};
> > > +
> > > +static void quorum_aio_bh(void *opaque)
> > > +{
> > > +    QuorumAIOCB *acb = opaque;
> > > +    BDRVQuorumState *s = acb->bqs;
> > > +    int ret;
> > > +
> > > +    ret = s->threshold <= acb->success_count ? 0 : -EIO;
> > 
> > It would be very much preferable if you stored the actual error code
> > instead of turning everything into -EIO.
> > 
> > > +
> > > +    qemu_bh_delete(acb->bh);
> > > +    acb->common.cb(acb->common.opaque, ret);
> > > +    if (acb->finished) {
> > > +        *acb->finished = true;
> > > +    }
> > > +    g_free(acb->aios);
> > > +    qemu_aio_release(acb);
> > > +}
> > 
> > Move this down so that it's next to the function using the bottom half.
> > 
> > > +
> > > +static QuorumAIOCB *quorum_aio_get(BDRVQuorumState *s,
> > > +                                   BlockDriverState *bs,
> > > +                                   QEMUIOVector *qiov,
> > > +                                   uint64_t sector_num,
> > > +                                   int nb_sectors,
> > > +                                   BlockDriverCompletionFunc *cb,
> > > +                                   void *opaque)
> > > +{
> > > +    QuorumAIOCB *acb = qemu_aio_get(&quorum_aiocb_info, bs, cb, opaque);
> > > +    int i;
> > > +
> > > +    acb->aios = g_new0(QuorumSingleAIOCB, s->total);
> > > +
> > > +    acb->bqs = s;
> > > +    acb->qiov = qiov;
> > > +    acb->bh = NULL;
> > > +    acb->count = 0;
> > > +    acb->success_count = 0;
> > > +    acb->sector_num = sector_num;
> > > +    acb->nb_sectors = nb_sectors;
> > > +    acb->vote = NULL;
> > > +    acb->vote_ret = 0;
> > > +    acb->finished = NULL;
> > > +
> > > +    for (i = 0; i < s->total; i++) {
> > > +        acb->aios[i].buf = NULL;
> > > +        acb->aios[i].ret = 0;
> > > +        acb->aios[i].parent = acb;
> > > +    }
> > 
> > Would you mind to reorder the initialisation of the fields according to
> > the order that is used in the struct definition?
> > 
> > > +
> > > +    return acb;
> > > +}
> > > +
> > > +static void quorum_aio_cb(void *opaque, int ret)
> > > +{
> > > +    QuorumSingleAIOCB *sacb = opaque;
> > > +    QuorumAIOCB *acb = sacb->parent;
> > > +    BDRVQuorumState *s = acb->bqs;
> > > +
> > > +    sacb->ret = ret;
> > > +    acb->count++;
> > > +    if (ret == 0) {
> > > +        acb->success_count++;
> > > +    }
> > > +    assert(acb->count <= s->total);
> > > +    assert(acb->success_count <= s->total);
> > > +    if (acb->count < s->total) {
> > > +        return;
> > > +    }
> > > +
> > > +    acb->bh = qemu_bh_new(quorum_aio_bh, acb);
> > > +    qemu_bh_schedule(acb->bh);
> > 
> > What's the reason for using a bottom half here? Worth a comment?
> > 
> > multiwrite_cb() in block.c doesn't use one to achieve something similar.
> > Is it buggy when you need one here?
> > 
> 
> I tried the code without bh and it doesn't work.

It's long ago tbat I wrote that comment, but the remark about
multiwrite_cb() concerns me. Do you know _why_ it doesn't work without
the BH, and whether the same problem affects multiwrite_cb()? I'd prefer
if we understood what we're doing over just basing the code on
experiments.

Kevin