ofiwg / libfabric

Open Fabric Interfaces
http://libfabric.org/
Other
530 stars 371 forks source link

EAGAIN endless loop #10150

Closed wangjiale01 closed 1 week ago

wangjiale01 commented 2 weeks ago

https://github.com/ofiwg/libfabric/blob/4171ebc98d6298bfb0c1f78bba001af275a0ddf7/prov/verbs/src/verbs_ep.c#L186 Why choose rx_cq instead of tx_cq in the vrb_post_send function, which would use the intermediate lethal loop

chien-intel commented 2 weeks ago

Looks like a bug to me. If we are out of SQ credit, we should poll CQ for SQ to see if we can increase credit for the post_send. Nice catch.