On Thu, Nov 22, 2018 at 07:16:32AM -0800, Matthew Wilcox wrote:

Yes, your allocations from the page_frag allocator have to have similar
lifetimes.  I thought that would be ideal for XFS though; as I understood
the problem, these were per-IO allocations, and IOs to the same filesystem
tend to take roughly the same amount of time.  Sure, in an error case,
some IOs will take a long time before timing out, but it should be OK
to have pages unavailable during that time in these rare situations.
What am I missing?

No, thee are allocations for meatada buffers, which can stay around
for a long time.  Worse still that depends on usage, so one buffer
allocated from ma page might basically stay around forever, while
another one might get reclaimed very quickly.