Messages in this thread | | | From | "Rafael J. Wysocki" <> | Subject | Re: [RFC][PATCH] PM: Force GFP_NOIO during suspend/resume (was: Re: [linux-pm] Memory allocations in .suspend became very unreliable) | Date | Tue, 19 Jan 2010 21:37:35 +0100 |
| |
On Tuesday 19 January 2010, Oliver Neukum wrote: > Am Montag, 18. Januar 2010 21:41:49 schrieb Rafael J. Wysocki: > > On Monday 18 January 2010, Oliver Neukum wrote: > > > Am Sonntag, 17. Januar 2010 14:55:55 schrieb Rafael J. Wysocki: > > > > +void mm_force_noio_allocations(void) > > > > +{ > > > > + /* Wait for all slowpath allocations using the old mask to complete */ > > > > + down_write(&gfp_allowed_mask_sem); > > > > + saved_gfp_allowed_mask = gfp_allowed_mask; > > > > + gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS); > > > > + up_write(&gfp_allowed_mask_sem); > > > > +} > > > > > > In addition to this you probably want to exhaust all memory reserves > > > before you fail a memory allocation > > > > I'm not really sure what you mean. > > Forget it, it was foolish. Instead there's a different problem. > Suppose we are tight on memory. The problem is that we must not > exhaust all memory. If we are really out of memory we may be unable > to satisfy memory allocations in resume()
That doesn't make things any worse than the are already. If we block on I/O forever during resume, the gross result is pretty much the same.
That said, Maxim reported that in his test case the mm subsystem apparently attempted to use I/O even if there was a plenty of free memory available and I'd like prevent _that_ from happening.
Rafael
| |