Message-ID: <20190418153047.GN28541@quack2.suse.cz>
Date:   Thu, 18 Apr 2019 17:30:47 +0200
From:   Jan Kara <jack@...e.cz>
To:     Jerome Glisse <jglisse@...hat.com>
Cc:     Jan Kara <jack@...e.cz>, Dan Williams <dan.j.williams@...el.com>,
        Kent Overstreet <kent.overstreet@...il.com>,
        Boaz Harrosh <boaz@...xistor.com>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        linux-fsdevel <linux-fsdevel@...r.kernel.org>,
        linux-block@...r.kernel.org, Linux MM <linux-mm@...ck.org>,
        John Hubbard <jhubbard@...dia.com>,
        Alexander Viro <viro@...iv.linux.org.uk>,
        Johannes Thumshirn <jthumshirn@...e.de>,
        Christoph Hellwig <hch@....de>, Jens Axboe <axboe@...nel.dk>,
        Ming Lei <ming.lei@...hat.com>, Jason Gunthorpe <jgg@...pe.ca>,
        Matthew Wilcox <willy@...radead.org>,
        Steve French <sfrench@...ba.org>, linux-cifs@...r.kernel.org,
        samba-technical@...ts.samba.org, Yan Zheng <zyan@...hat.com>,
        Sage Weil <sage@...hat.com>, Ilya Dryomov <idryomov@...il.com>,
        Alex Elder <elder@...nel.org>, ceph-devel@...r.kernel.org,
        Eric Van Hensbergen <ericvh@...il.com>,
        Latchesar Ionkov <lucho@...kov.net>,
        Mike Marshall <hubcap@...ibond.com>,
        Martin Brandenburg <martin@...ibond.com>,
        devel@...ts.orangefs.org,
        Dominique Martinet <asmadeus@...ewreck.org>,
        v9fs-developer@...ts.sourceforge.net, Coly Li <colyli@...e.de>,
        linux-bcache@...r.kernel.org,
        Ernesto A. Fernández 
        <ernesto.mnd.fernandez@...il.com>
Subject: Re: [PATCH v1 00/15] Keep track of GUPed pages in fs and block

On Thu 18-04-19 10:27:29, Jerome Glisse wrote:
> On Thu, Apr 18, 2019 at 12:42:05PM +0200, Jan Kara wrote:
> > On Wed 17-04-19 18:28:58, Jerome Glisse wrote:
> > > On Wed, Apr 17, 2019 at 02:53:28PM -0700, Dan Williams wrote:
> > > > On Tue, Apr 16, 2019 at 12:50 PM Jerome Glisse <jglisse@...hat.com> wrote:
> > > > >
> > > > > On Tue, Apr 16, 2019 at 12:12:27PM -0700, Dan Williams wrote:
> > > > > > On Tue, Apr 16, 2019 at 11:59 AM Kent Overstreet
> > > > > > <kent.overstreet@...il.com> wrote:
> > > > > > >
> > > > > > > On Tue, Apr 16, 2019 at 09:35:04PM +0300, Boaz Harrosh wrote:
> > > > > > > > On Thu, Apr 11, 2019 at 05:08:19PM -0400, jglisse@...hat.com wrote:
> > > > > > > > > From: Jérôme Glisse <jglisse@...hat.com>
> > > > > > > > >
> > > > > > > > > This patchset depends on various small fixes [1] and also on patchset
> > > > > > > > > which introduce put_user_page*() [2] and thus is 5.3 material as those
> > > > > > > > > pre-requisite will get in 5.2 at best. Nonetheless i am posting it now
> > > > > > > > > so that it can get review and comments on how and what should be done
> > > > > > > > > to test things.
> > > > > > > > >
> > > > > > > > > For various reasons [2] [3] we want to track page reference through GUP
> > > > > > > > > differently than "regular" page reference. Thus we need to keep track
> > > > > > > > > of how we got a page within the block and fs layer. To do so this patch-
> > > > > > > > > set change the bio_bvec struct to store a pfn and flags instead of a
> > > > > > > > > direct pointer to a page. This way we can flag page that are coming from
> > > > > > > > > GUP.
> > > > > > > > >
> > > > > > > > > This patchset is divided as follow:
> > > > > > > > >     - First part of the patchset is just small cleanup i believe they
> > > > > > > > >       can go in as his assuming people are ok with them.
> > > > > > > >
> > > > > > > >
> > > > > > > > >     - Second part convert bio_vec->bv_page to bio_vec->bv_pfn this is
> > > > > > > > >       done in multi-step, first we replace all direct dereference of
> > > > > > > > >       the field by call to inline helper, then we introduce macro for
> > > > > > > > >       bio_bvec that are initialized on the stack. Finaly we change the
> > > > > > > > >       bv_page field to bv_pfn.
> > > > > > > >
> > > > > > > > Why do we need a bv_pfn. Why not just use the lowest bit of the page-ptr
> > > > > > > > as a flag (pointer always aligned to 64 bytes in our case).
> > > > > > > >
> > > > > > > > So yes we need an inline helper for reference of the page but is it not clearer
> > > > > > > > that we assume a page* and not any kind of pfn ?
> > > > > > > > It will not be the first place using low bits of a pointer for flags.
> > > > > > > >
> > > > > > > > That said. Why we need it at all? I mean why not have it as a bio flag. If it exist
> > > > > > > > at all that a user has a GUP and none-GUP pages to IO at the same request he/she
> > > > > > > > can just submit them as two separate BIOs (chained at the block layer).
> > > > > > > >
> > > > > > > > Many users just submit one page bios and let elevator merge them any way.
> > > > > > >
> > > > > > > Let's please not add additional flags and weirdness to struct bio - "if this
> > > > > > > flag is set interpret one way, if not interpret another" - or eventually bios
> > > > > > > will be as bad as skbuffs. I would much prefer just changing bv_page to bv_pfn.
> > > > > >
> > > > > > This all reminds of the failed attempt to teach the block layer to
> > > > > > operate without pages:
> > > > > >
> > > > > > https://lore.kernel.org/lkml/20150316201640.33102.33761.stgit@dwillia2-desk3.amr.corp.intel.com/
> > > > > >
> > > > > > >
> > > > > > > Question though - why do we need a flag for whether a page is a GUP page or not?
> > > > > > > Couldn't the needed information just be determined by what range the pfn is not
> > > > > > > (i.e. whether or not it has a struct page associated with it)?
> > > > > >
> > > > > > That amounts to a pfn_valid() check which is a bit heavier than if we
> > > > > > can store a flag in the bv_pfn entry directly.
> > > > > >
> > > > > > I'd say create a new PFN_* flag, and make bv_pfn a 'pfn_t' rather than
> > > > > > an 'unsigned long'.
> > > > > >
> > > > > > That said, I'm still in favor of Jan's proposal to just make the
> > > > > > bv_page semantics uniform. Otherwise we're complicating this core
> > > > > > infrastructure for some yet to be implemented GPU memory management
> > > > > > capabilities with yet to be determined value. Circle back when that
> > > > > > value is clear, but in the meantime fix the GUP bug.
> > > > >
> > > > > This has nothing to do with GPU, what make you think so ? Here i am
> > > > > trying to solve GUP and to keep the value of knowing wether a page
> > > > > has been GUP or not. I argue that if we bias every page in every bio
> > > > > then we loose that information and thus the value.
> > > > >
> > > > > I gave the page protection mechanisms as an example that would be
> > > > > impacted but it is not the only one. Knowing if a page has been GUP
> > > > > can be useful for memory reclaimation, compaction, NUMA balancing,
> > > > 
> > > > Right, this is what I was reacting to in your pushback to Jan's
> > > > proposal. You're claiming value for not doing the simple thing for
> > > > some future "may be useful in these contexts". To my knowledge those
> > > > things are not broken today. You're asking for the complexity to be
> > > > carried today for some future benefit, and I'm asking for the
> > > > simplicity to be maintained as much as possible today and let the
> > > > value of future changes stand on their own to push for more complexity
> > > > later.
> > > > 
> > > > Effectively don't use this bug fix to push complexity for a future
> > > > agenda where the value has yet to be quantified.
> > > 
> > > Except that this solution (biasing everyone in bio) would _more complex_
> > > it is only conceptualy appealing. The changes are on the other hand much
> > > deeper and much riskier but you decided to ignore that and focus on some-
> > > thing i was just giving as an example.
> > 
> > Yeah, after going and reading several places like fs/iomap.c, fs/mpage.c,
> > drivers/md/dm-io.c I agree with you. The places that are not doing direct
> > IO usually just don't hold any page reference that could be directly
> > attributed to the bio (and they don't drop it when bio finishes). They
> > rather use other means (like PageLocked, PageWriteback) to make sure the
> > page stays alive so mandating gup-pin reference for all pages attached to a
> > bio would require a lot of reworking of places that are not related to our
> > problem and currently work just fine. So I withdraw my suggestion. Nice in
> > theory, too much work in practice ;).
> 
> Have you seem Boaz proposal ? I have started it and it does not look to
> bad (but you knwo taste and color :)) You can take a peek:
> 
> https://cgit.freedesktop.org/~glisse/linux/log/?h=gup-bio-v2
> 
> I need to finish that and run fstests on bunch of different fs before
> posting. Dunno if i will have enough time to do that before LSF/MM.

Yes, I've seen it. I just wasn't sure how the result will look like. What
you have in your tree looks pretty clean so far. BTW (I know I'm repeating
myself ;) what if we made iov_iter_get_pages() & iov_iter_get_pages_alloc()
always return gup-pin reference? That would get rid of the need for two
ioend handlers for each call site...

								Honza
-- 
Jan Kara <jack@...e.com>
SUSE Labs, CR

Powered by blists - more mailing lists