On Mon, 04 Mar 2024, Matthew Wilcox wrote:
> On Mon, Mar 04, 2024 at 09:45:48AM +1100, NeilBrown wrote:
> > I have in mind a more explicit statement of how much waiting is
> > acceptable.
> >
> > GFP_NOFAIL - wait indefinitely
> > GFP_KILLABLE - wait indefinitely unless fatal signal is pending.
> > GFP_RETRY - may retry but deadlock, though unlikely, is possible. So
> > don't wait indefinitely. May abort more quickly if fatal
> > signal is pending.
> > GFP_NO_RETRY - only try things once. This may sleep, but will give up
> > fairly quickly. Either deadlock is a significant
> > possibility, or alternate strategy is fairly cheap.
> > GFP_ATOMIC - don't sleep - same as current.
>
> I don't think these should be GFP flags. Rather, these should be
> context flags (and indeed, they're mutually exclusive, so this is a
> small integer to represent where we are on the spectrum). That is
> we want code to do
>
> void *alloc_foo(void)
> {
> return init_foo(kmalloc(256, GFP_MOVABLE));
> }
>
> void submit_foo(void)
> {
> spin_lock(&submit_lock);
> flags = memalloc_set_atomic();
> __submit_foo(alloc_foo());
> memalloc_restore_flags(flags);
> spin_unlock(&submit_lock);
> }
>
> struct foo *prealloc_foo(void)
> {
> return alloc_foo();
> }
>
> ... for various degrees of complexity. That is, the _location_ of memory
> is allocation site knowledge, but how hard to try is _path_ dependent,
> and not known at the allocation site because it doesn't know what locks
> are held.
>
I'm not convinced. While there is a path dependency there is also
location dependency.
The code at the call-site determines what happens in response to
failure.
For GFP_NOFAIL, failure is not possible. We cannot allow context to
turn NOFAIL into NOSLEEP because context cannot add error handling.
Consider mempool_alloc(). That requests a NORETRY allocation because
there is an easy fall-back. Is that a location dependency or a path
dependency? I would say it is location. Of course a location is just a
very short path so it is a path dependency too - but is that perspective
really helpful?
Certainly I could accept a GFP_WHATEVER which uses NOSLEEP if path
context demands that, and NO_OOM otherwise.
Thanks,
NeilBrown
