On Wed, 2021-07-21 at 19:54 +0100, David Howells wrote:

Jeff Layton [off-list ref] wrote:

quoted

The above two deltas seem like they should have been in patch #2.

Yeah.  Looks like at least partially so.

quoted

quoted

@@ -635,15 +625,8 @@ void netfs_subreq_terminated(struct netfs_read_subrequest *subreq,
 		goto failed;
 	}
 
-	if (WARN(transferred_or_error > subreq->len - subreq->transferred,
-		 "Subreq overread: R%x[%x] %zd > %zu - %zu",
-		 rreq->debug_id, subreq->debug_index,
-		 transferred_or_error, subreq->len, subreq->transferred))
-		transferred_or_error = subreq->len - subreq->transferred;
-
 	subreq->error = 0;
-	subreq->transferred += transferred_or_error;
-	if (subreq->transferred < subreq->len)
+	if (iov_iter_count(&subreq->iter))
 		goto incomplete;
 

I must be missing it, but where does subreq->iter get advanced to the
end of the current read? If you're getting rid of subreq->transferred
then I think that has to happen above, no?

For afs, afs_req_issue_op() points fsreq->iter at the subrequest iterator and
calls afs_fetch_data().  Thereafter, we wend our way to
afs_deliver_fs_fetch_data() or yfs_deliver_fs_fetch_data() which set
call->iter to point to that iterator and then call afs_extract_data() which
passes it to rxrpc_kernel_recv_data(), which eventually passes it to
skb_copy_datagram_iter(), which advances the iterator.

For the cache, the subrequest iterator is passed to the cache backend by
netfs_read_from_cache().  This would be cachefiles_read() which calls
vfs_iocb_iter_read() which I thought advances the iterator (leastways,
filemap_read() keeps going until iov_iter_count() reaches 0 or some other stop
condition occurs and doesn't thereafter call iov_iter_revert()).

Ok, this will probably regress ceph then. We don't really have anything
to do with the subreq->iter at this point and this patch doesn't change
that. If you're going to make this change, it'd be cleaner to also fix
up ceph_netfs_issue_op to advance the iter at the same time.
-- 
Jeff Layton [off-list ref]