On Thu, 2007-10-04 at 16:09 -0700, Andrew Morton wrote:
quoted text
> On Fri, 05 Oct 2007 00:39:16 +0200
> Miklos Szeredi <miklos@szeredi.hu> wrote:
>=20
> > > throttle_vm_writeout() should be a per-zone thing, I guess.  Perhaps =
fixing
quoted text
> > > that would fix your deadlock.  That's doubtful, but I don't know anyt=
hing
quoted text
> > > about your deadlock so I cannot say.
> >=20
> > No, doing the throttling per-zone won't in itself fix the deadlock.
> >=20
> > Here's a deadlock example:
> >=20
> > Total memory =3D 32M
> > /proc/sys/vm/dirty_ratio =3D 10
> > dirty_threshold =3D 3M
> > ratelimit_pages =3D 1M
> >=20
> > Some program dirties 4M (dirty_threshold + ratelimit_pages) of mmap on
> > a fuse fs.  Page balancing is called which turns all these into
> > writeback pages.
> >=20
> > Then userspace filesystem gets a write request, and tries to allocate
> > memory needed to complete the writeout.
> >=20
> > That will possibly trigger direct reclaim, and throttle_vm_writeout()
> > will be called.  That will block until nr_writeback goes below 3.3M
> > (dirty_threshold + 10%).  But since all 4M of writeback is from the
> > fuse fs, that will never happen.
> >=20
> > Does that explain it better?
> >=20
>=20
> yup, thanks.
>=20
> This is a somewhat general problem: a userspace process is in the IO path=
.=20
quoted text
> Userspace block drivers, for example - pretty much anything which involve=
s
quoted text
> kernel->userspace upcalls for storage applications.
>=20
> I solved it once in the past by marking the userspace process as
> PF_MEMALLOC and I beleive that others have implemented the same hack.
>=20
> I suspect that what we need is a general solution, and that the solution
> will involve explicitly telling the kernel that this process is one which
> actually cleans memory and needs special treatment.
>=20
> Because I bet there will be other corner-cases where such a process needs
> kernel help, and there might be optimisation opportunities as well.
>=20
> Problem is, any such mark-me-as-special syscall would need to be
> privileged, and FUSE servers presently don't require special perms (do
> they?)

I think just adding nr_cpus * ratelimit_pages to the dirth_thresh in
throttle_vm_writeout() will also solve the problem