Re: [PATCH 1/5] mm: Introduce mm_struct.has_pinned

From: Jason Gunthorpe <jgg@ziepe.ca>
To: Peter Xu <peterx@redhat.com>
Cc: Jann Horn <jannh@google.com>, Linux-MM <linux-mm@kvack.org>,
	kernel list <linux-kernel@vger.kernel.org>,
	Andrew Morton <akpm@linux-foundation.org>,
	Jan Kara <jack@suse.cz>, Michal Hocko <mhocko@suse.com>,
	Kirill Tkhai <ktkhai@virtuozzo.com>,
	Kirill Shutemov <kirill@shutemov.name>,
	Hugh Dickins <hughd@google.com>, Christoph Hellwig <hch@lst.de>,
	Andrea Arcangeli <aarcange@redhat.com>,
	John Hubbard <jhubbard@nvidia.com>,
	Oleg Nesterov <oleg@redhat.com>,
	Leon Romanovsky <leonro@nvidia.com>,
	Linus Torvalds <torvalds@linux-foundation.org>
Subject: Re: [PATCH 1/5] mm: Introduce mm_struct.has_pinned
Date: Tue, 22 Sep 2020 12:56:04 -0300	[thread overview]
Message-ID: <20200922155604.GA731578@ziepe.ca> (raw)
In-Reply-To: <20200922142802.GC19098@xz-x1>

On Tue, Sep 22, 2020 at 10:28:02AM -0400, Peter Xu wrote:
> On Tue, Sep 22, 2020 at 08:54:36AM -0300, Jason Gunthorpe wrote:
> > On Tue, Sep 22, 2020 at 12:47:11AM +0200, Jann Horn wrote:
> > > On Tue, Sep 22, 2020 at 12:30 AM Peter Xu <peterx@redhat.com> wrote:
> > > > On Mon, Sep 21, 2020 at 11:43:38PM +0200, Jann Horn wrote:
> > > > > On Mon, Sep 21, 2020 at 11:17 PM Peter Xu <peterx@redhat.com> wrote:
> > > > > > (Commit message collected from Jason Gunthorpe)
> > > > > >
> > > > > > Reduce the chance of false positive from page_maybe_dma_pinned() by keeping
> > > > > > track if the mm_struct has ever been used with pin_user_pages(). mm_structs
> > > > > > that have never been passed to pin_user_pages() cannot have a positive
> > > > > > page_maybe_dma_pinned() by definition.
> > > > >
> > > > > There are some caveats here, right? E.g. this isn't necessarily true
> > > > > for pagecache pages, I think?
> > > >
> > > > Sorry I didn't follow here.  Could you help explain with some details?
> > > 
> > > The commit message says "mm_structs that have never been passed to
> > > pin_user_pages() cannot have a positive page_maybe_dma_pinned() by
> > > definition"; but that is not true for pages which may also be mapped
> > > in a second mm and may have been passed to pin_user_pages() through
> > > that second mm (meaning they must be writable over there and not
> > > shared with us via CoW).
> > 
> > The message does need a few more words to explain this trick can only
> > be used with COW'able pages.
> >  
> > > Process A:
> > > 
> > > fd_a = open("/foo/bar", O_RDWR);
> > > mapping_a = mmap(NULL, 0x1000, PROT_READ|PROT_WRITE, MAP_SHARED, fd_a, 0);
> > > pin_user_pages(mapping_a, 1, ...);
> > > 
> > > Process B:
> > > 
> > > fd_b = open("/foo/bar", O_RDONLY);
> > > mapping_b = mmap(NULL, 0x1000, PROT_READ|PROT_WRITE, MAP_PRIVATE, fd_b, 0);
> > > *(volatile char *)mapping_b;
> > > 
> > > At this point, process B has never called pin_user_pages(), but
> > > page_maybe_dma_pinned() on the page at mapping_b would return true.
> > 
> > My expectation is the pin_user_pages() should have already broken the
> > COW for the MAP_PRIVATE, so process B should not have a
> > page_maybe_dma_pinned()
> 
> When process B maps with PROT_READ only (w/o PROT_WRITE) then it seems the same
> page will be mapped.

I thought MAP_PRIVATE without PROT_WRITE was nonsensical, it only has
meaning for writes initiated by the mapping. MAP_SHARED/PROT_READ is
the same behavior on Linux, IIRC.

But, yes, you certainly can end up with B having
page_maybe_dma_pinned() pages in shared VMA, just not in COW'able
mappings.

> I think I get the point from Jann now.  Maybe it's easier I just remove the
> whole "mm_structs that have never been passed to pin_user_pages() cannot have a
> positive page_maybe_dma_pinned() by definition" sentence if that's misleading,
> because the rest seem to be clear enough on what this new field is used for.

"for COW" I think is still the important detail here, see for instance
my remark on the PUD/PMD splitting where it is necessary to test for
cow before using this.

Perhaps we should call it "has_pinned_for_cow" to place emphasis on
this detail? Due to the shared pages issue It really doesn't have any
broader utility, eg for file back pages or otherwise.

Jason