Re: [PATCH v2 4/4] migration/rdma: source: poll cm_event from return path

["Dr. David Alan Gilbert" <dgilbert@redhat.com>]

* Li Zhijian (lizhijian@cn.fujitsu.com) wrote:
> source side always blocks if postcopy is only enabled at source side.
> users are not able to cancel this migration in this case.
> 
> Let source side have chance to cancel this migration
> 
> Signed-off-by: Li Zhijian <lizhijian@cn.fujitsu.com>
> ---
> V2: utilize poll to check cm event
> ---
>  migration/rdma.c | 42 ++++++++++++++++++++++++++++++++++++++----
>  1 file changed, 38 insertions(+), 4 deletions(-)
> 
> diff --git a/migration/rdma.c b/migration/rdma.c
> index d829d08d076..f67e21b4f54 100644
> --- a/migration/rdma.c
> +++ b/migration/rdma.c
> @@ -36,6 +36,7 @@
>  #include <rdma/rdma_cma.h>
>  #include "trace.h"
>  #include "qom/object.h"
> +#include <poll.h>
>  
>  /*
>   * Print and error on both the Monitor and the Log file.
> @@ -2460,7 +2461,36 @@ err_rdma_source_init:
>      return -1;
>  }
>  
> -static int qemu_rdma_connect(RDMAContext *rdma, Error **errp)
> +static int qemu_get_cm_event_timeout(RDMAContext *rdma,
> +                                     struct rdma_cm_event **cm_event,
> +                                     long msec, Error **errp)
> +{
> +    int ret;
> +    struct pollfd poll_fd = {
> +                                .fd = rdma->channel->fd,
> +                                .events = POLLIN,
> +                                .revents = 0
> +                            };
> +
> +    do {
> +        ret = poll(&poll_fd, 1, msec);
> +    } while (ret < 0 && errno == EINTR);
> +
> +    if (ret == 0) {
> +        ERROR(errp, "poll cm event timeout");
> +        return -1;
> +    } else if (ret < 0) {
> +        ERROR(errp, "failed to pull cm event, errno=%i", errno);

Typo: 'poll' - I can fix that.

> +        return -1;
> +    } else if (poll_fd.revents & POLLIN) {
> +        return rdma_get_cm_event(rdma->channel, cm_event);
> +    } else {
> +        ERROR(errp, "no POLLIN event, revent=%x", poll_fd.revents);
> +        return -1;
> +    }
> +}
> +
> +static int qemu_rdma_connect(RDMAContext *rdma, Error **errp, bool return_path)
>  {
>      RDMACapabilities cap = {
>                                  .version = RDMA_CONTROL_VERSION_CURRENT,
> @@ -2498,7 +2528,11 @@ static int qemu_rdma_connect(RDMAContext *rdma, Error **errp)
>          goto err_rdma_source_connect;
>      }
>  
> -    ret = rdma_get_cm_event(rdma->channel, &cm_event);
> +    if (return_path) {
> +        ret = qemu_get_cm_event_timeout(rdma, &cm_event, 5000, errp);

Fixed timeouts are not a great fix; but I can't think of anything
better; the only alternative would be to register the fd on the main
thread's poll and get it to be called back when the event happened.

But for now;

Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>

> +    } else {
> +        ret = rdma_get_cm_event(rdma->channel, &cm_event);
> +    }
>      if (ret) {
>          perror("rdma_get_cm_event after rdma_connect");
>          ERROR(errp, "connecting to destination!");
> @@ -4111,7 +4145,7 @@ void rdma_start_outgoing_migration(void *opaque,
>      }
>  
>      trace_rdma_start_outgoing_migration_after_rdma_source_init();
> -    ret = qemu_rdma_connect(rdma, errp);
> +    ret = qemu_rdma_connect(rdma, errp, false);
>  
>      if (ret) {
>          goto err;
> @@ -4132,7 +4166,7 @@ void rdma_start_outgoing_migration(void *opaque,
>              goto return_path_err;
>          }
>  
> -        ret = qemu_rdma_connect(rdma_return_path, errp);
> +        ret = qemu_rdma_connect(rdma_return_path, errp, true);
>  
>          if (ret) {
>              goto return_path_err;
> -- 
> 2.30.2
> 
> 
> 
-- 
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK