On Mon, May 25, 2020 at 09:37:05PM -0300, Jason Gunthorpe wrote: > On Mon, May 25, 2020 at 01:56:28PM -0700, John Hubbard wrote: > > On 2020-05-25 09:56, Jason Gunthorpe wrote: > > > On Mon, May 25, 2020 at 11:11:42AM -0400, Peter Xu wrote: > > > > On Mon, May 25, 2020 at 11:46:51AM -0300, Jason Gunthorpe wrote: > > > > > On Mon, May 25, 2020 at 10:28:06AM -0400, Peter Xu wrote: > > > > > > On Mon, May 25, 2020 at 09:26:07AM -0300, Jason Gunthorpe wrote: > > > > > > > On Sat, May 23, 2020 at 07:52:57PM -0400, Peter Xu wrote: > > > > > > > > > > > > > > > For what I understand now, IMHO we should still need all those handlings of > > > > > > > > FAULT_FLAG_RETRY_NOWAIT like in the initial version. E.g., IIUC KVM gup will > > > > > > > > try with FOLL_NOWAIT when async is allowed, before the complete slow path. I'm > > > > > > > > not sure what would be the side effect of that if fault() blocked it. E.g., > > > > > > > > the caller could be in an atomic context. > > > > > > > > > > > > > > AFAICT FAULT_FLAG_RETRY_NOWAIT only impacts what happens when > > > > > > > VM_FAULT_RETRY is returned, which this doesn't do? > > > > > > > > > > > > Yes, that's why I think we should still properly return VM_FAULT_RETRY if > > > > > > needed.. because IMHO it is still possible that the caller calls with > > > > > > FAULT_FLAG_RETRY_NOWAIT. > > > > > > > > > > > > My understanding is that FAULT_FLAG_RETRY_NOWAIT majorly means: > > > > > > > > > > > > - We cannot release the mmap_sem, and, > > > > > > - We cannot sleep > > > > > > > > > > Sleeping looks fine, look at any FS implementation of fault, say, > > > > > xfs. The first thing it does is xfs_ilock() which does down_write(). > > > > > > > > Yeah. My wild guess is that maybe fs code will always be without > > > > FAULT_FLAG_RETRY_NOWAIT so it's safe to sleep unconditionally (e.g., I think > > > > the general #PF should be fine to sleep in fault(); gup should be special, but > > > > I didn't observe any gup code called upon file systems)? > > > > > > get_user_pages is called on filesystem backed pages. > > > > > > I have no idea what FAULT_FLAG_RETRY_NOWAIT is supposed to do. Maybe > > > John was able to guess when he reworked that stuff? > > > > > > > Although I didn't end up touching that particular area, I'm sure it's going > > to come up sometime soon, so I poked around just now, and found that > > FAULT_FLAG_RETRY_NOWAIT was added almost exactly 9 years ago. This flag was > > intended to make KVM and similar things behave better when doing GUP on > > file-backed pages that might, or might not be in memory. > > > > The idea is described in the changelog, but not in the code comments or > > Documentation, sigh: > > > > commit 318b275fbca1ab9ec0862de71420e0e92c3d1aa7 > > Author: Gleb Natapov <gleb@xxxxxxxxxx> > > Date: Tue Mar 22 16:30:51 2011 -0700 > > > > mm: allow GUP to fail instead of waiting on a page > > > > GUP user may want to try to acquire a reference to a page if it is already > > in memory, but not if IO, to bring it in, is needed. For example KVM may > > tell vcpu to schedule another guest process if current one is trying to > > access swapped out page. Meanwhile, the page will be swapped in and the > > guest process, that depends on it, will be able to run again. > > > > This patch adds FAULT_FLAG_RETRY_NOWAIT (suggested by Linus) and > > FOLL_NOWAIT follow_page flags. FAULT_FLAG_RETRY_NOWAIT, when used in > > conjunction with VM_FAULT_ALLOW_RETRY, indicates to handle_mm_fault that > > it shouldn't drop mmap_sem and wait on a page, but return VM_FAULT_RETRY > > instead. > > So, from kvm's perspective it was to avoid excessively long blocking in > common paths when it could rejoin the completed IO by somehow waiting > on a page itself? > > It all seems like it should not be used unless the page is going to go > to IO? I think NOWAIT is used as a common flag for kvm for its initial attempt to fault in a normal page, however... I just noticed another fact that actually __get_user_pages() won't work with PFNMAP (check_vma_flags should fail), but KVM just started to support fault() for PFNMAP from commit add6a0cd1c5b (2016) using fixup_user_fault(), where nvidia seems to have a similar request to have a fault handler on some mapped BARs. > > Certainly there is no reason to optimize the fringe case of vfio > sleeping if there is and incorrect concurrnent attempt to disable the > a BAR. If fixup_user_fault() (which is always with ALLOW_RETRY && !RETRY_NOWAIT) is the only path for the new fault(), then current way seems ok. Not sure whether this would worth a WARN_ON_ONCE(RETRY_NOWAIT) in the fault() to be clear of that fact. Thanks, -- Peter Xu
