On 8/13/17 11:50 PM, Malahal Naineni wrote:
>> That trace is the NSM clnt_dg clnt_call, the only use of outgoing UDP. It's
a mess, and has been a mess for a long time.
We get a file descriptor fd and then create "rec", but while destroying things, we close "fd" and then rpc_dplx_unref(). Re-arranging these in clnt_dg_destroy() (and other places) might help fix this issue, but I am not positive as I am not familiar with
this code.
I am also working on a blind replacement of "fd" by "struct gfd" where struct gfd has the "fd" as well as a "generation number". The generation number is incremented when ever such "fd" is created (e.g. accept() call or socket() call). The changes are many
but they are trivial.
Any thoughts?
It's not really interesting for the current code base. In V2.5, I've
already eliminated all the various copies of fd, and every SVCXPRT is
wrapped inside a dplx_rec, and they all use xp_fd, and it's in only one
tree (svc_rqst). So there's no longer any possibility of multiple
generations of fd.
That said, the last remaining problem is clnt_dg clnt_call, where the
fd can be passed to poll() at the same time as another copy is passed to
(or being removed from) epoll(). Requires a complete re-write.
I'd started doing the re-write long long ago, even made the rpc_ctx
transport independent (committed in V2.6/v1.6 Napalm rendezvous patch).
But there are still many problems redesigning with async callbacks.
I'm looking at the short-term fix I've mentioned earlier, that we should
try TCP before UDP, but given our current code base doesn't even compile,
I've given up until next week.
------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
Nfs-ganesha-devel mailing list
Nfs-ganesha-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/nfs-ganesha-devel
