[Xrootd] Copy Response with correct stream ID to XrdXrootdAioTask xrootd#2592 #2593
Conversation
57d6539 to
ddd3c8e
Compare
|
This is a fix for #2592. The issue appeared to me to be incorrect stream ID placed in responses sent on bound (i.e. sub) Links, when the aio system is being used. (This happens for xcache, but not regular xrootd server, when the client is trying to use sub-streams). I think the problem is that a read/pgread command is read from XrdXrootdProtocol object for the main (login) Link, and that Protocol has the correct stream ID set in its Response (the one the client set in the the read/pgread command). The protocol object for the link of the correspoding return pathID is found ('pP'), and a XrdXrootdAioTask is allocated. But the AioTask is prepared with the stream ID in the protocol object for the bound link, which at this point may not be correct. The result is that response is eventually sent on the correct Link, but possibly with the wrong streamId; thus the client can not match the response to the request, discards it and keeps waiting for a response. |
abh3
left a comment
There was a problem hiding this comment.
Good call David. From a historical perspective, the uncorrected code should produced the same hang for a server as it does for Xcache. There is one big difference here. Regular servers no longer enable themselves for async I/O since the Linux implementation of disk async I/O is rather useless. However, Xcache servers do enable it because they necessarily bypass the Linux implementation and do the expected async operations. This happened in 5.3.0 2021-07-09. So, something between then and now must have broken the original design but we never noticed it because the feature was turned off for servers and, apparently, no one does multi-stream copies via xcache (though Atlas does do single stream copy to scratch).
|
Thanks for having a look @abh3 |
No description provided.