Date:   Sun, 29 Oct 2017 16:46:44 -0700
From:   Dan Williams <dan.j.williams@...il.com>
To:     Jan Kara <jack@...e.cz>
Cc:     Christoph Hellwig <hch@....de>, Michal Hocko <mhocko@...e.com>,
        Benjamin Herrenschmidt <benh@...nel.crashing.org>,
        Dave Hansen <dave.hansen@...ux.intel.com>,
        Heiko Carstens <heiko.carstens@...ibm.com>,
        "J. Bruce Fields" <bfields@...ldses.org>,
        linux-mm <linux-mm@...ck.org>, Paul Mackerras <paulus@...ba.org>,
        Sean Hefty <sean.hefty@...el.com>,
        Jeff Layton <jlayton@...chiereds.net>,
        Matthew Wilcox <mawilcox@...rosoft.com>,
        linux-rdma@...r.kernel.org, Michael Ellerman <mpe@...erman.id.au>,
        Jason Gunthorpe <jgunthorpe@...idianresearch.com>,
        Doug Ledford <dledford@...hat.com>,
        Hal Rosenstock <hal.rosenstock@...il.com>,
        Dave Chinner <david@...morbit.com>,
        linux-fsdevel <linux-fsdevel@...r.kernel.org>,
        Alexander Viro <viro@...iv.linux.org.uk>,
        Gerald Schaefer <gerald.schaefer@...ibm.com>,
        "linux-nvdimm@...ts.01.org" <linux-nvdimm@...ts.01.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        linux-xfs@...r.kernel.org,
        Martin Schwidefsky <schwidefsky@...ibm.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        "Darrick J. Wong" <darrick.wong@...cle.com>,
        "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>
Subject: Re: [PATCH v3 00/13] dax: fix dma vs truncate and remove 'page-less' support

On Thu, Oct 26, 2017 at 3:58 AM, Jan Kara <jack@...e.cz> wrote:
> On Fri 20-10-17 11:31:48, Christoph Hellwig wrote:
>> On Fri, Oct 20, 2017 at 09:47:50AM +0200, Christoph Hellwig wrote:
>> > I'd like to brainstorm how we can do something better.
>> >
>> > How about:
>> >
>> > If we hit a page with an elevated refcount in truncate / hole puch
>> > etc for a DAX file system we do not free the blocks in the file system,
>> > but add it to the extent busy list.  We mark the page as delayed
>> > free (e.g. page flag?) so that when it finally hits refcount zero we
>> > call back into the file system to remove it from the busy list.
>>
>> Brainstorming some more:
>>
>> Given that on a DAX file there shouldn't be any long-term page
>> references after we unmap it from the page table and don't allow
>> get_user_pages calls why not wait for the references for all
>> DAX pages to go away first?  E.g. if we find a DAX page in
>> truncate_inode_pages_range that has an elevated refcount we set
>> a new flag to prevent new references from showing up, and then
>> simply wait for it to go away.  Instead of a busy way we can
>> do this through a few hashed waitqueued in dev_pagemap.  And in
>> fact put_zone_device_page already gets called when putting the
>> last page so we can handle the wakeup from there.
>>
>> In fact if we can't find a page flag for the stop new callers
>> things we could probably come up with a way to do that through
>> dev_pagemap somehow, but I'm not sure how efficient that would
>> be.
>
> We were talking about this yesterday with Dan so some more brainstorming
> from us. We can implement the solution with extent busy list in ext4
> relatively easily - we already have such list currently similarly to XFS.
> There would be some modifications needed but nothing too complex. The
> biggest downside of this solution I see is that it requires per-filesystem
> solution for busy extents - ext4 and XFS are reasonably fine, however btrfs
> may have problems and ext2 definitely will need some modifications.
> Invisible used blocks may be surprising to users at times although given
> page refs should be relatively short term, that should not be a big issue.
> But are we guaranteed page refs are short term? E.g. if someone creates
> v4l2 videobuf in MAP_SHARED mapping of a file on DAX filesystem, page refs
> can be rather long-term similarly as in RDMA case. Also freeing of blocks
> on page reference drop is another async entry point into the filesystem
> which could unpleasantly surprise us but I guess workqueues would solve
> that reasonably fine.
>
> WRT waiting for page refs to be dropped before proceeding with truncate (or
> punch hole for that matter - that case is even nastier since we don't have
> i_size to guard us). What I like about this solution is that it is very
> visible there's something unusual going on with the file being truncated /
> punched and so problems are easier to diagnose / fix from the admin side.
> So far we have guarded hole punching from concurrent faults (and
> get_user_pages() does fault once you do unmap_mapping_range()) with
> I_MMAP_LOCK (or its equivalent in ext4). We cannot easily wait for page
> refs to be dropped under I_MMAP_LOCK as that could deadlock - the most
> obvious case Dan came up with is when GUP obtains ref to page A, then hole
> punch comes grabbing I_MMAP_LOCK and waiting for page ref on A to be
> dropped, and then GUP blocks on trying to fault in another page.
>
> I think we cannot easily prevent new page references to be grabbed as you
> write above since nobody expects stuff like get_page() to fail. But I
> think that unmapping relevant pages and then preventing them to be faulted
> in again is workable and stops GUP as well. The problem with that is though
> what to do with page faults to such pages - you cannot just fail them for
> hole punch, and you cannot easily allocate new blocks either. So we are
> back at a situation where we need to detach blocks from the inode and then
> wait for page refs to be dropped - so some form of busy extents. Am I
> missing something?

Coming back to this since Dave has made clear that new locking to
coordinate get_user_pages() is a no-go.

We can unmap to force new get_user_pages() attempts to block on the
per-fs mmap lock, but if punch-hole finds any elevated pages it needs
to drop the mmap lock and wait. We need this lock dropped to get
around the problem that the driver will not start to drop page
references until it has elevated the page references on all the pages
in the I/O. If we need to drop the mmap lock that makes it impossible
to coordinate this unlock/retry loop within truncate_inode_pages_range
which would otherwise be the natural place to land this code.

Would it be palatable to unmap and drain dma in any path that needs to
detach blocks from an inode? Something like the following that builds
on dax_wait_dma() tried to achieve, but does not introduce a new lock
for the fs to manage:

retry:
    per_fs_mmap_lock(inode);
    unmap_mapping_range(mapping, start, end); /* new page references
cannot be established */
    if ((dax_page = dax_dma_busy_page(mapping, start, end)) != NULL) {
        per_fs_mmap_unlock(inode); /* new page references can happen,
so we need to start over */
        wait_for_page_idle(dax_page);
        goto retry;
    }
    truncate_inode_pages_range(mapping, start, end);
    per_fs_mmap_unlock(inode);

Given how far away taking the mmap lock occurs from where we know we
are actually performing a punch-hole operation this may lead to
unnecessary unmapping and dma flushing.

As far as I can see the extent busy mechanism does not simplify the
solution. If we have code to wait for the pages to go idle might as
well have the truncate/punch-hole wait on that event in the first
instance.

Powered by blists - more mailing lists