qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH 2/2] nbd/server: Quiesce coroutines on context switch


From: Sergio Lopez
Subject: Re: [PATCH 2/2] nbd/server: Quiesce coroutines on context switch
Date: Wed, 9 Dec 2020 18:31:59 +0100

On Fri, Dec 04, 2020 at 12:39:07PM -0600, Eric Blake wrote:
> On 12/4/20 10:53 AM, Sergio Lopez wrote:
> > When switching between AIO contexts we need to me make sure that both
> > recv_coroutine and send_coroutine are not scheduled to run. Otherwise,
> > QEMU may crash while attaching the new context with an error like
> > this one:
> > 
> > aio_co_schedule: Co-routine was already scheduled in 'aio_co_schedule'
> > 
> > To achieve this we need a local implementation of
> > 'qio_channel_readv_all_eof' named 'nbd_read_eof' (a trick already done
> > by 'nbd/client.c') that allows us to interrupt the operation and to
> > know when recv_coroutine is yielding.
> > 
> > With this in place, we delegate detaching the AIO context to the
> > owning context with a BH ('nbd_aio_detach_bh') scheduled using
> > 'aio_wait_bh_oneshot'. This BH signals that we need to quiesce the
> > channel by setting 'client->quiescing' to 'true', and either waits for
> > the coroutine to finish using AIO_WAIT_WHILE or, if it's yielding in
> > 'nbd_read_eof', actively enters the coroutine to interrupt it.
> > 
> > RHBZ: https://bugzilla.redhat.com/show_bug.cgi?id=1900326
> > Signed-off-by: Sergio Lopez <slp@redhat.com>
> > ---
> >  nbd/server.c | 120 +++++++++++++++++++++++++++++++++++++++++++++------
> >  1 file changed, 106 insertions(+), 14 deletions(-)
> 
> A complex patch, so I'd appreciate a second set of eyes.
> 
> > 
> > diff --git a/nbd/server.c b/nbd/server.c
> > index 613ed2634a..7229f487d2 100644
> > --- a/nbd/server.c
> > +++ b/nbd/server.c
> > @@ -132,6 +132,9 @@ struct NBDClient {
> >      CoMutex send_lock;
> >      Coroutine *send_coroutine;
> >  
> > +    bool read_yielding;
> > +    bool quiescing;
> 
> Will either of these fields need to be accessed atomically once the
> 'yank' code is added, or are we still safe with direct access because
> coroutines are not multithreaded?

Yes, those are only accessed from coroutines, which will be scheduled
on the same thread.

> > +
> >      QTAILQ_ENTRY(NBDClient) next;
> >      int nb_requests;
> >      bool closing;
> > @@ -1352,14 +1355,60 @@ static coroutine_fn int nbd_negotiate(NBDClient 
> > *client, Error **errp)
> >      return 0;
> >  }
> >  
> > -static int nbd_receive_request(QIOChannel *ioc, NBDRequest *request,
> > +/* nbd_read_eof
> > + * Tries to read @size bytes from @ioc. This is a local implementation of
> > + * qio_channel_readv_all_eof. We have it here because we need it to be
> > + * interruptible and to know when the coroutine is yielding.
> > + * Returns 1 on success
> > + *         0 on eof, when no data was read (errp is not set)
> > + *         negative errno on failure (errp is set)
> > + */
> > +static inline int coroutine_fn
> > +nbd_read_eof(NBDClient *client, void *buffer, size_t size, Error **errp)
> > +{
> > +    bool partial = false;
> > +
> > +    assert(size);
> > +    while (size > 0) {
> > +        struct iovec iov = { .iov_base = buffer, .iov_len = size };
> > +        ssize_t len;
> > +
> > +        len = qio_channel_readv(client->ioc, &iov, 1, errp);
> > +        if (len == QIO_CHANNEL_ERR_BLOCK) {
> > +            client->read_yielding = true;
> > +            qio_channel_yield(client->ioc, G_IO_IN);
> > +            client->read_yielding = false;
> 
> nbd/client.c:nbd_read_eof() uses bdrv_dec/inc_in_flight instead of
> read_yielding...
> 
> > +            if (client->quiescing) {
> > +                return -EAGAIN;
> > +            }
> 
> and the quiescing check is new; otherwise, these two functions look
> identical.  Having two static functions with the same name makes gdb a
> bit more annoying (which one of the two did you want your breakpoint
> on?).  Is there any way we could write this code only once in
> nbd/common.c for reuse by both client and server?  But I can live with
> it as written.

I'm not happy with this either, but on the first implementation I've
tried to come up with a unique function for both use cases, and it
looked terrible.

We can easily use a different name, though.

> > @@ -2151,20 +2223,23 @@ static int nbd_co_send_bitmap(NBDClient *client, 
> > uint64_t handle,
> >  
> >  /* nbd_co_receive_request
> >   * Collect a client request. Return 0 if request looks valid, -EIO to drop
> > - * connection right away, and any other negative value to report an error 
> > to
> > - * the client (although the caller may still need to disconnect after 
> > reporting
> > - * the error).
> > + * connection right away, -EAGAIN to indicate we were interrupted and the
> > + * channel should be quiesced, and any other negative value to report an 
> > error
> > + * to the client (although the caller may still need to disconnect after
> > + * reporting the error).
> >   */
> >  static int nbd_co_receive_request(NBDRequestData *req, NBDRequest *request,
> >                                    Error **errp)
> >  {
> >      NBDClient *client = req->client;
> >      int valid_flags;
> > +    int ret;
> >  
> >      g_assert(qemu_in_coroutine());
> >      assert(client->recv_coroutine == qemu_coroutine_self());
> > -    if (nbd_receive_request(client->ioc, request, errp) < 0) {
> > -        return -EIO;
> > +    ret = nbd_receive_request(client, request, errp);
> > +    if (ret < 0) {
> > +        return  ret;
> 
> Why the double space?

Ouch, copy/paste mistake.

> The old code slams to EIO, you preserve errors.  Is that going to bite
> us by causing us to see a different errno leaked through?

When reading from the channel, nbd_read_eof hides the actual errno
behind EIO, so the only actual difference is that, with this change,
nbd_receive_request may send this error...

    if (magic != NBD_REQUEST_MAGIC) {
        error_setg(errp, "invalid magic (got 0x%" PRIx32 ")", magic);
        return -EINVAL;
    }

... to the client (via server.c:2624), which I think is the right
thing to do.

> >      }
> >  
> >      trace_nbd_co_receive_request_decode_type(request->handle, 
> > request->type,
> > @@ -2507,6 +2582,17 @@ static coroutine_fn void nbd_trip(void *opaque)
> >          return;
> >      }
> >  
> > +    if (client->quiescing) {
> > +        /*
> > +         * We're switching between AIO contexts. Don't attempt to receive 
> > a new
> > +         * request and kick the main context which may be waiting for us.
> 
> s/request/request,/

Thanks, will fix this comment.

> > +         */
> > +        nbd_client_put(client);
> > +        client->recv_coroutine = NULL;
> > +        aio_wait_kick();
> > +        return;
> > +    }
> > +
> >      req = nbd_request_get(client);
> >      ret = nbd_co_receive_request(req, &request, &local_err);
> >      client->recv_coroutine = NULL;
> > @@ -2519,6 +2605,11 @@ static coroutine_fn void nbd_trip(void *opaque)
> >          goto done;
> >      }
> >  
> > +    if (ret == -EAGAIN) {
> > +        assert(client->quiescing);
> > +        goto done;
> > +    }
> > +
> >      nbd_client_receive_next_request(client);
> >      if (ret == -EIO) {
> >          goto disconnect;
> > @@ -2565,7 +2656,8 @@ disconnect:
> >  
> >  static void nbd_client_receive_next_request(NBDClient *client)
> >  {
> > -    if (!client->recv_coroutine && client->nb_requests < MAX_NBD_REQUESTS) 
> > {
> > +    if (!client->recv_coroutine && client->nb_requests < MAX_NBD_REQUESTS 
> > &&
> > +        !client->quiescing) {
> >          nbd_client_get(client);
> >          client->recv_coroutine = qemu_coroutine_create(nbd_trip, client);
> >          aio_co_schedule(client->exp->common.ctx, client->recv_coroutine);
> > 
> 
> Overall looks okay to me,
> Reviewed-by: Eric Blake <eblake@redhat.com>
> 
> -- 
> Eric Blake, Principal Software Engineer
> Red Hat, Inc.           +1-919-301-3226
> Virtualization:  qemu.org | libvirt.org
> 

Attachment: signature.asc
Description: PGP signature


reply via email to

[Prev in Thread] Current Thread [Next in Thread]