Re: [Qemu-devel] [PATCH 05/12] aio: introduce aio_{disable, enable}_clients

[Fam Zheng <famz@redhat.com>]

On Mon, 10/12 10:31, Kevin Wolf wrote:
> Am 09.10.2015 um 18:27 hat Fam Zheng geschrieben:
> > On Fri, 10/09 16:31, Kevin Wolf wrote:
> > > Am 09.10.2015 um 07:45 hat Fam Zheng geschrieben:
> > > > Signed-off-by: Fam Zheng <famz@redhat.com>
> > > > ---
> > > >  aio-posix.c         |  3 ++-
> > > >  aio-win32.c         |  3 ++-
> > > >  async.c             | 42 ++++++++++++++++++++++++++++++++++++++++++
> > > >  include/block/aio.h | 30 ++++++++++++++++++++++++++++++
> > > >  4 files changed, 76 insertions(+), 2 deletions(-)
> > > > 
> > > > diff --git a/aio-posix.c b/aio-posix.c
> > > > index d25fcfc..a261892 100644
> > > > --- a/aio-posix.c
> > > > +++ b/aio-posix.c
> > > > @@ -261,7 +261,8 @@ bool aio_poll(AioContext *ctx, bool blocking)
> > > >  
> > > >      /* fill pollfds */
> > > >      QLIST_FOREACH(node, &ctx->aio_handlers, node) {
> > > > -        if (!node->deleted && node->pfd.events) {
> > > > +        if (!node->deleted && node->pfd.events
> > > > +            && !aio_type_disabled(ctx, node->type)) {
> > > >              add_pollfd(node);
> > > >          }
> > > >      }
> > > > diff --git a/aio-win32.c b/aio-win32.c
> > > > index f5ecf57..66cff60 100644
> > > > --- a/aio-win32.c
> > > > +++ b/aio-win32.c
> > > > @@ -309,7 +309,8 @@ bool aio_poll(AioContext *ctx, bool blocking)
> > > >      /* fill fd sets */
> > > >      count = 0;
> > > >      QLIST_FOREACH(node, &ctx->aio_handlers, node) {
> > > > -        if (!node->deleted && node->io_notify) {
> > > > +        if (!node->deleted && node->io_notify
> > > > +            && !aio_type_disabled(ctx, node->type)) {
> > > >              events[count++] = event_notifier_get_handle(node->e);
> > > >          }
> > > >      }
> > > > diff --git a/async.c b/async.c
> > > > index 244bf79..855b9d5 100644
> > > > --- a/async.c
> > > > +++ b/async.c
> > > > @@ -361,3 +361,45 @@ void aio_context_release(AioContext *ctx)
> > > >  {
> > > >      rfifolock_unlock(&ctx->lock);
> > > >  }
> > > > +
> > > > +bool aio_type_disabled(AioContext *ctx, int type)
> > > > +{
> > > > +    int i = 1;
> > > > +    int n = 0;
> > > > +
> > > > +    while (type) {
> > > > +        bool b = type & 0x1;
> > > > +        type >>= 1;
> > > > +        n++;
> > > 
> > > Any specific reason for leaving client_disable_counters[0] unused?
> > 
> > No, I should have started from 0.
> > 
> > > 
> > > > +        i <<= 1;
> > > 
> > > i is never read.
> > > 
> > > > +        if (!b) {
> > > > +            continue;
> > > > +        }
> > > > +        if (ctx->client_disable_counters[n]) {
> > > > +            return true;
> > > > +        }
> > > > +    }
> > > > +    return false;
> > > > +}
> > > 
> > > In general I wonder whether this function really needs to take a mask
> > > with possibly multiple set bits instead of just a single type.
> > 
> > Previous versions used to have more types than "internal" and "external", so it
> > has been a mask. So yes, I think a single type will be better now.
> > 
> > > 
> > > > +void aio_disable_enable_clients(AioContext *ctx, int clients_mask,
> > > > +                                bool is_disable)
> > > > +{
> > > > +    int i = 1;
> > > > +    int n = 0;
> > > > +    aio_context_acquire(ctx);
> > > > +
> > > > +    while (clients_mask) {
> > > > +        bool b = clients_mask & 0x1;
> > > > +        clients_mask >>= 1;
> > > > +        n++;
> > > > +        i <<= 1;
> > > 
> > > This i isn't used either.
> > > 
> > > > +        if (!b) {
> > > > +            continue;
> > > > +        }
> > > > +        if (ctx->client_disable_counters[n]) {
> > > > +            return true;
> > > > +        }
> > > 
> > > Wait, why are you checking the state instead of setting it?
> > 
> > Oops, apparent I screwed my workspaces as I do remember coding this assignment.
> > And I must have used a wrong command when building the tree so that I don't
> > even catch the compiling error. :(
> > 
> > > 
> > > How did you test this series?
> > 
> > So far only smoke testing and qemu-iotests, because I don't have a good idea of
> > testifying the transaction's atomicity. Any suggestions?
> 
> Perhaps you could use blkdebug to delay something in the middle of the
> transaction while your guest keeps writing stuff? That should result in
> 100% reproducability.
> 
> I guess you actually need to make sure that your guest doesn't do any
> I/O, then set the blkdebug breakpoint, send the transaction, and once a
> request is stopped, you start some I/O in the guest. Resume as soon as
> you know that something bad happened.
> 
> Possibly you need to add a new blkdebug event to find a good place to
> suspend a transaction request.
> 

It's difficult to "start some I/O" in the guest in the middle of transaction,
even with help of blkdebug, because BQL is hold during the whole transaction.

I think it would be a bit easier to program a VCPU to constantly submit I/O
requests to the vq, but that's far from enough.

Anyway I'll start by writing some unit test code instead, in tests/test-aio.c.

Fam