From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Date: Tue, 18 Dec 2018 12:09:54 +1100 From: Dave Chinner To: Matthew Wilcox Cc: Jerome Glisse , Jan Kara , John Hubbard , Dan Williams , John Hubbard , Andrew Morton , Linux MM , tom@talpey.com, Al Viro , benve@cisco.com, Christoph Hellwig , Christopher Lameter , "Dalessandro, Dennis" , Doug Ledford , Jason Gunthorpe , Michal Hocko , mike.marciniszyn@intel.com, rcampbell@nvidia.com, Linux Kernel Mailing List , linux-fsdevel Subject: Re: [PATCH 1/2] mm: introduce put_user_page*(), placeholder versions Message-ID: <20181218010954.GM6311@dastard> References: <20181207191620.GD3293@redhat.com> <3c4d46c0-aced-f96f-1bf3-725d02f11b60@nvidia.com> <20181208022445.GA7024@redhat.com> <20181210102846.GC29289@quack2.suse.cz> <20181212150319.GA3432@redhat.com> <20181212214641.GB29416@dastard> <20181214154321.GF8896@quack2.suse.cz> <20181216215819.GC10644@dastard> <20181217181148.GA3341@redhat.com> <20181217183443.GO10600@bombadil.infradead.org> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20181217183443.GO10600@bombadil.infradead.org> Sender: owner-linux-mm@kvack.org List-ID: On Mon, Dec 17, 2018 at 10:34:43AM -0800, Matthew Wilcox wrote: > On Mon, Dec 17, 2018 at 01:11:50PM -0500, Jerome Glisse wrote: > > On Mon, Dec 17, 2018 at 08:58:19AM +1100, Dave Chinner wrote: > > > Sure, that's a possibility, but that doesn't close off any race > > > conditions because there can be DMA into the page in progress while > > > the page is being bounced, right? AFAICT this ext3+DIF/DIX case is > > > different in that there is no 3rd-party access to the page while it > > > is under IO (ext3 arbitrates all access to it's metadata), and so > > > nothing can actually race for modification of the page between > > > submission and bouncing at the block layer. > > > > > > In this case, the moment the page is unlocked, anyone else can map > > > it and start (R)DMA on it, and that can happen before the bio is > > > bounced by the block layer. So AFAICT, block layer bouncing doesn't > > > solve the problem of racing writeback and DMA direct to the page we > > > are doing IO on. Yes, it reduces the race window substantially, but > > > it doesn't get rid of it. > > > > So the event flow is: > > - userspace create object that match a range of virtual address > > against a given kernel sub-system (let's say infiniband) and > > let's assume that the range is an mmap() of a regular file > > - device driver do GUP on the range (let's assume it is a write > > GUP) so if the page is not already map with write permission > > in the page table than a page fault is trigger and page_mkwrite > > happens > > - Once GUP return the page to the device driver and once the > > device driver as updated the hardware states to allow access > > to this page then from that point on hardware can write to the > > page at _any_ time, it is fully disconnected from any fs event > > like write back, it fully ignore things like page_mkclean > > > > This is how it is to day, we allowed people to push upstream such > > users of GUP. This is a fact we have to live with, we can not stop > > hardware access to the page, we can not force the hardware to follow > > page_mkclean and force a page_mkwrite once write back ends. This is > > the situation we are inheriting (and i am personnaly not happy with > > that). > > > > >From my point of view we are left with 2 choices: > > [C1] break all drivers that do not abide by the page_mkclean and > > page_mkwrite > > [C2] mitigate as much as possible the issue > > > > For [C2] the idea is to keep track of GUP per page so we know if we > > can expect the page to be written to at any time. Here is the event > > flow: > > - driver GUP the page and program the hardware, page is mark as > > GUPed > > ... > > - write back kicks in on the dirty page, lock the page and every > > thing as usual , sees it is GUPed and inform the block layer to > > use a bounce page > > No. The solution John, Dan & I have been looking at is to take the > dirty page off the LRU while it is pinned by GUP. It will never be > found for writeback. Pages are found for writeback by mapping tree lookup, not page LRU scans (i.e. write_cache_pages() from background writeback) Are suggesting that pages pinned by GUP are going to be removed from the page cache *and* the mapping tree while they are pinned? > That's not the end of the story though. Other parts of the kernel (eg > msync) also need to be taught to stay away from pages which are pinned > by GUP. But the idea is that no page gets written back to storage while > it's pinned by GUP. Only when the last GUP ends is the page returned > to the list of dirty pages. I think playing fast and loose with data integrity like this is fundamentally wrong. If this gets implemented, then I'll be sending every "I ran sync and then two hours later the system crashed but the data was lost when the system came back up" bug report directly to you. Cheers, Dave. -- Dave Chinner david@fromorbit.com