On 4/10/2018 10:17 PM, Burakov, Anatoly wrote:
On 10-Apr-18 2:53 PM, Tan, Jianfeng wrote:

On 4/10/2018 6:03 PM, Anatoly Burakov wrote:
EAL did not stop processing further asynchronous requests on
encountering a request that should trigger the callback. This
resulted in erasing valid requests but not triggering them.

That means one wakeup could process multiple replies, and following process_async_request() will erase some valid requests?


Fix this by stopping the loop once we have a request that we
can trigger. Also, remove unnecessary check for trigger
request being NULL.

Fixes: f05e26051c15 ("eal: add IPC asynchronous request")
Cc: anatoly.bura...@intel.com

Signed-off-by: Anatoly Burakov <anatoly.bura...@intel.com>

Acked-by: Jianfeng Tan <jianfeng....@intel.com>

  lib/librte_eal/common/eal_common_proc.c | 4 ++--
  1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/lib/librte_eal/common/eal_common_proc.c b/lib/librte_eal/common/eal_common_proc.c
index f98622f..1ea3b58 100644
--- a/lib/librte_eal/common/eal_common_proc.c
+++ b/lib/librte_eal/common/eal_common_proc.c
@@ -510,11 +510,11 @@ async_reply_handle(void *arg __rte_unused)
                              sr, next);
-                } else if (action == ACTION_TRIGGER &&
-                        trigger == NULL) {
+                } else if (action == ACTION_TRIGGER) {
                              sr, next);
                      trigger = sr;
+                    break;

If I understand it correctly above, break here, we will trigger an async action, and then go back to sleep with some ready requests not handled? Seems that we shall unlock, process, and lock here. Right?

Well, we won't go back to sleep - we'll just loop around and come back.

See eal_common_proc.c:472:

    /* sometimes, we don't even wait */
    if (sr->reply_received) {
        nowait = true;

Followed by line 478:

    if (nowait)
        ret = 0;

Followed by line 495:

    if (ret == 0 || ret == ETIMEDOUT) {

So, having messages with replies already in the queue will cause wait to be cancelled. It's not much slower than unlocking, triggering, and locking again.

Ah, sorry, I overlooked that fact that every iteration we re-scan the request list.

However, if you're OK with lock-loop-unlock-trigger-lock-loop-unlock-... sequence until we run out of triggers, then sure, i can add that.

Don't see the reason for that.


Reply via email to