Re: [PATCH] mm: fix page cache convergence regression

From: Linus Torvalds
Date: Mon Jun 24 2019 - 17:52:23 EST

On Mon, Jun 24, 2019 at 11:19 PM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> Hey, it's three weeks later and we're about to miss 5.2.
> This sucks, Matthew.


And I do think that having a real gfp field there would be better than
the very odd xa_flags that is *marked* as being gfp_t, but isn't
really a gfp_t at all.

So how about we apply Johannes' patch, and then work on making that
xa_flags field be a proper type of its own. Because it really isn't a
gfp_t, and never has been, even if there might be some very limited
and hacky overlap right now.

Alternatrively, the subset of bits that _can_ be used as a gfp should
actually be used as such, in xas_alloc/xas_nomem. So that you can do

xa_init_flags(&mapping->i_pages, XA_FLAGS_LOCK_IRQ | __GFP_ACCOUNT);

in __address_space_init_once() and it would do what it is supposed to do.