Re: [PATCH v4 08/10] mm/gup: limit number of gup migration failures, honor failures

From: Pavel Tatashin <pasha.tatashin@soleen.com>
To: Jason Gunthorpe <jgg@ziepe.ca>
Cc: LKML <linux-kernel@vger.kernel.org>,
	linux-mm <linux-mm@kvack.org>,
	Andrew Morton <akpm@linux-foundation.org>,
	Vlastimil Babka <vbabka@suse.cz>, Michal Hocko <mhocko@suse.com>,
	David Hildenbrand <david@redhat.com>,
	Oscar Salvador <osalvador@suse.de>,
	Dan Williams <dan.j.williams@intel.com>,
	Sasha Levin <sashal@kernel.org>,
	Tyler Hicks <tyhicks@linux.microsoft.com>,
	Joonsoo Kim <iamjoonsoo.kim@lge.com>,
	mike.kravetz@oracle.com, Steven Rostedt <rostedt@goodmis.org>,
	Ingo Molnar <mingo@redhat.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Mel Gorman <mgorman@suse.de>,
	Matthew Wilcox <willy@infradead.org>,
	David Rientjes <rientjes@google.com>,
	John Hubbard <jhubbard@nvidia.com>,
	Linux Doc Mailing List <linux-doc@vger.kernel.org>,
	Ira Weiny <ira.weiny@intel.com>,
	linux-kselftest@vger.kernel.org
Subject: Re: [PATCH v4 08/10] mm/gup: limit number of gup migration failures, honor failures
Date: Wed, 13 Jan 2021 15:05:41 -0500	[thread overview]
Message-ID: <CA+CK2bDDUMOeCH8rQBL7fBdHCAUZBOykyXNL2N=hmxq7xi0giQ@mail.gmail.com> (raw)
In-Reply-To: <20210113195528.GD4605@ziepe.ca>

> > > Oh, that existing logic is wrong too :( Another bug.
> >
> > I do not think there is a bug.
> >
> > > You can't skip pages in the pages[] array under the assumption they
> > > are contiguous. ie the i+=step is wrong.
> >
> > If pages[i] is part of a compound page, the other parts of this page
> > must be sequential in this array for this compound page
>
> That is true only if the PMD points to the page. If the PTE points to
> a tail page then there is no requirement that other PTEs are
> contiguous with the compount page.
>
> At this point we have no idea if the GUP logic got this compound page
> as a head page in a PMD or as a tail page from a PTE, so we can't
> assume a contiguous run of addresses.

I see, I will fix this bug in an upstream as a separate patch in my
series, and keep the fix when my fixes are applied.

>
> Look at split_huge_pmd() - it doesn't break up the compound page it
> just converts the PMD to a PTE array and scatters the tail pages to
> the PTE.

Got it, unfortunately the fix will deoptimize the code by having to
check every page if it is part of a previous compound page or not.

>
> I understand Matt is pushing on this idea more by having compound
> pages in the page cache, but still mapping tail pages when required.
>
> > This is actually standard migration procedure, elsewhere in the kernel
> > we migrate pages in exactly the same fashion: isolate and later
> > migrate. The isolation works for LRU only pages.
>
> But do other places cause a userspace visible random failure when LRU
> isolation fails?

Makes sense, I will remove maximum retries for isolation, and retry
indefinitely, the same as it is done during memory hot-remove. So, we
will fail only when migration fails.

>
> I don't like it at all, what is the user supposed to do?
>
> Jason