Re: [Qemu-devel] [PATCH] ISCSI: Only set up the read-event if we are actually waiting for data to come back in from the target.

[Paolo Bonzini]

Il 11/05/2012 12:22, Ronnie Sahlberg ha scritto:
> Signed-off-by: Ronnie Sahlberg <address@hidden>
> ---
>  block/iscsi.c |    4 +++-
>  1 files changed, 3 insertions(+), 1 deletions(-)
> 
> diff --git a/block/iscsi.c b/block/iscsi.c
> index d37c4ee..989b5e9 100644
> --- a/block/iscsi.c
> +++ b/block/iscsi.c
> @@ -105,7 +105,9 @@ iscsi_set_events(IscsiLun *iscsilun)
>  {
>      struct iscsi_context *iscsi = iscsilun->iscsi;
>  
> -    qemu_aio_set_fd_handler(iscsi_get_fd(iscsi), iscsi_process_read,
> +    qemu_aio_set_fd_handler(iscsi_get_fd(iscsi),
> +                           (iscsi_queue_length(iscsi) > 0)
> +                           ? iscsi_process_read : NULL,
>                             (iscsi_which_events(iscsi) & POLLOUT)
>                             ? iscsi_process_write : NULL,
>                             iscsi_process_flush, iscsilun);

I wonder if iscsi is also susceptible to the same race condition I saw
with NBD, where you can have:

1) select in the iothread exiting and reporting readability

2) the iothread subsequently blocking on the mutex

3) a VCPU thread's qemu_aio_wait() calling iscsi_process_read

4) when the VCPU releases the mutex, the iothread will call
iscsi_process_read again.

This should be easily reproducible with IDE drives, but the above patch
would not fix it.  Perhaps it's better to call iscsi_queue_length in
iscsi_process_read instead.

Paolo