On 7/24/21 12:22 PM, Matthew Wilcox wrote:

On Sun, Jul 11, 2021 at 07:44:07PM -0600, Jens Axboe wrote:

quoted

On 7/11/21 9:09 AM, Matthew Wilcox (Oracle) wrote:

quoted

I noticed a theoretical case where an IOCB_NOWAIT read could sleep:

filemap_get_pages
  filemap_get_read_batch
  page_cache_sync_readahead
    page_cache_sync_ra
      ondemand_readahead
        do_page_cache_ra
        page_cache_ra_unbounded
          gfp_t gfp_mask = readahead_gfp_mask(mapping);
          memalloc_nofs_save()
          __page_cache_alloc(gfp_mask);

We're in a nofs context, so we're not going to start new IO, but we might
wait for writeback to complete.  We generally don't want to sleep for IO,
particularly not for IO that isn't related to us.

Jens, can you run this through your test rig and see if it makes any
practical difference?

You bet, I'll see if I can trigger this condition and verify we're no
longer blocking on writeback. Thanks for hacking this up.

Did you have any success yet?

Sorry forgot to report back - I did run some testing last week, and
didn't manage to make it hit the blocking condition. Did various
read/write mix on the same file, made sure there was memory pressure,
etc. I'll give it another go, please let me know if you have an idea on
how to make this easier to hit... I know it's one of those things that
you hit all the time in certain workloads (hence why I would love to see
it get fixed), but I just didn't manage to provoke it when I tried.

-- 
Jens Axboe