Matthew Wilcox [off-list ref] wrote:

We're defeating the ondemand_readahead() algorithm here.  Let's suppose
userspace is doing 64kB reads, the filesystem is OrangeFS which only
wants to do 4MB reads, the page cache is initially empty and there's
only one thread doing a sequential read.  ondemand_readahead() calls
get_init_ra_size() which tells it to allocate 128kB and set the async
marker at 64kB.  Then orangefs calls readahead_expand() to allocate the
remainder of the 4MB.  After the app has read the first 64kB, it comes
back to read the next 64kB, sees the readahead marker and tries to trigger
the next batch of readahead, but it's already present, so it does nothing
(see page_cache_ra_unbounded() for what happens with pages present).

It sounds like Christoph is right on the right track and the vm needs to ask
the filesystem (and by extension, the cache) before doing the allocation and
before setting the trigger flag.  Then we don't need to call back into the vm
to expand the readahead.

Also, there's Steve's request to try and keep at least two requests in flight
for CIFS/SMB at the same time to consider.

David