On Wed, Mar 30, 2022 at 08:54:09AM -0700, Christoph Hellwig wrote:

On Wed, Mar 30, 2022 at 11:45:56AM -0400, Johannes Weiner wrote:

quoted

quoted

FYI, I started redoing that version and I think with all the cleanups
to filemap.c and the readahead code this can be done fairly nicely now:

http://git.infradead.org/users/hch/block.git/commitdiff/666abb29c6db870d3941acc5ac19e83fbc72cfd4

Yes, it's definitely much nicer now with the MM instantiating the
pages for ->readpage(s).

But AFAICS this breaks compressed btrfs (and erofs?) because those
still do additional add_to_page_cache_lru() and bio submissions.

In btrfs, add_ra_bio_pages only passed freshly allocated pages to
add_to_page_cache_lru.  These can't really have PageWorkingSet set,
can they?  In erofs they can also come from a local page pool, but
I think otherwise the same applies.

It's add_to_page_cache_lru() that sets the flag.

Basically, when a PageWorkingset (hot) page gets reclaimed, the bit is
stored in the vacated tree slot. When the entry is brought back in,
add_to_page_cache_lru() transfers it to the newly allocated page.

add_to_page_cache_lru()
  filemap_add_folio()
    __filemap_add_folio(mapping, folio, index, gfp, &shadow)
      *shadow = *slot
      *slot = folio
  if (shadow)
    workingset_refault(folio, shadow)
      folio_set_workingset()