On Mon 02-09-24 18:32:33, Kent Overstreet wrote:
> On Mon, Sep 02, 2024 at 02:52:52PM GMT, Andrew Morton wrote:
> > On Mon, 2 Sep 2024 05:53:59 -0400 Kent Overstreet <kent.overstreet@xxxxxxxxx> wrote:
> >
> > > On Mon, Sep 02, 2024 at 11:51:48AM GMT, Michal Hocko wrote:
> > > > The previous version has been posted in [1]. Based on the review feedback
> > > > I have sent v2 of patches in the same threat but it seems that the
> > > > review has mostly settled on these patches. There is still an open
> > > > discussion on whether having a NORECLAIM allocator semantic (compare to
> > > > atomic) is worthwhile or how to deal with broken GFP_NOFAIL users but
> > > > those are not really relevant to this particular patchset as it 1)
> > > > doesn't aim to implement either of the two and 2) it aims at spreading
> > > > PF_MEMALLOC_NORECLAIM use while it doesn't have a properly defined
> > > > semantic now that it is not widely used and much harder to fix.
> > > >
> > > > I have collected Reviewed-bys and reposting here. These patches are
> > > > touching bcachefs, VFS and core MM so I am not sure which tree to merge
> > > > this through but I guess going through Andrew makes the most sense.
> > > >
> > > > Changes since v1;
> > > > - compile fixes
> > > > - rather than dropping PF_MEMALLOC_NORECLAIM alone reverted eab0af905bfc
> > > > ("mm: introduce PF_MEMALLOC_NORECLAIM, PF_MEMALLOC_NOWARN") suggested
> > > > by Matthew.
> > >
> > > To reiterate:
> > >
> >
> > It would be helpful to summarize your concerns.
> >
> > What runtime impact do you expect this change will have upon bcachefs?
>
> For bcachefs: I try really hard to minimize tail latency and make
> performance robust in extreme scenarios - thrashing. A large part of
> that is that btree locks must be held for no longer than necessary.
>
> We definitely don't want to recurse into other parts of the kernel,
> taking other locks (i.e. in memory reclaim) while holding btree locks;
> that's a great way to stack up (and potentially multiply) latencies.
OK, these two patches do not fail to do that. The only existing user is
turned into GFP_NOWAIT so the final code works the same way. Right?
> But gfp flags don't work with vmalloc allocations (and that's unlikely
> to change), and we require vmalloc fallbacks for e.g. btree node
> allocation. That's the big reason we want MEMALLOC_PF_NORECLAIM.
Have you even tried to reach out to vmalloc maintainers and asked for
GFP_NOWAIT support for vmalloc? Because I do not remember that. Sure
kernel page tables are have hardcoded GFP_KERNEL context which slightly
complicates that but that doesn't really mean the only potential
solution is to use a per task flag to override that. Just from top of my
head we can consider pre-allocating virtual address space for
non-sleeping allocations. Maybe there are other options that only people
deeply familiar with the vmalloc internals can see.
This requires discussions not pushing a very particular solution
through.
--
Michal Hocko
SUSE Labs
