lkml.org 
[lkml]   [2019]   [Apr]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v1 10/15] block: add gup flag to bio_add_page()/bio_add_pc_page()/__bio_add_page()
On Tue, Apr 16, 2019 at 06:52:06PM +0200, Jan Kara wrote:
> On Mon 15-04-19 20:22:04, Jerome Glisse wrote:
> > On Mon, Apr 15, 2019 at 04:59:52PM +0200, Jan Kara wrote:
> > > Hi Jerome!
> > >
> > > On Thu 11-04-19 17:08:29, jglisse@redhat.com wrote:
> > > > From: Jérôme Glisse <jglisse@redhat.com>
> > > >
> > > > We want to keep track of how we got a reference on page added to bio_vec
> > > > ie wether the page was reference through GUP (get_user_page*) or not. So
> > > > add a flag to bio_add_page()/bio_add_pc_page()/__bio_add_page() to that
> > > > effect.
> > >
> > > Thanks for writing this patch set! Looking through patches like this one,
> > > I'm a bit concerned. With so many bio_add_page() callers it's difficult to
> > > get things right and not regress in the future. I'm wondering whether the
> > > things won't be less error-prone if we required that all page reference
> > > from bio are gup-like (not necessarily taken by GUP, if creator of the bio
> > > gets to struct page he needs via some other means (e.g. page cache lookup),
> > > he could just use get_gup_pin() helper we'd provide). After all, a page
> > > reference in bio means that the page is pinned for the duration of IO and
> > > can be DMAed to/from so it even makes some sense to track the reference
> > > like that. Then bio_put() would just unconditionally do put_user_page() and
> > > we won't have to propagate the information in the bio.
> > >
> > > Do you think this would be workable and easier?
> >
> > Thinking again on this, i can drop that patch and just add a new
> > bio_add_page_from_gup() and then it would be much more obvious that
> > only very few places need to use that new version and they are mostly
> > obvious places. It is usualy GUP then right away add the pages to bio
> > or bvec.
>
> Yes, that's another option. Probably second preferred by me after my own
> proposal ;)
>
> > We can probably add documentation around GUP explaining that if you
> > want to build a bio or bvec from GUP you must pay attention to which
> > function you use.
>
> Yes, although we both know how careful people are in reading
> documentation...

Yes i know this is a sad state, but if enough people see comments in
enough places we should end up with more eyes aware of the gotcha and
hopefully increase the likelyhood of catching any new user.

>
> > Also pages going in a bio are not necessarily written too, they can
> > be use as source (writting to block) or as destination (reading from
> > block). So having all of them with refcount bias as GUP would muddy
> > the water somemore between pages we can no longer clean (ie GUPed)
> > and those that are just being use in regular read or write operation.
>
> Why would the difference matter here?

Restricting GUP like status to GUP insure that we only ever back-off
because of GUP and not because of some innocuous I/O.

I am working on a v2 that just add a new variant to add page, but i
will have to run (x)fstest before re-posting.

I also have the scatterlist convertion mostly ready:

https://cgit.freedesktop.org/~glisse/linux/log/?h=gup-scatterlist-v1

After that GUP is mostly isolated to individual driver and much easier
to track and update.

Cheers,
Jérôme

\
 
 \ /
  Last update: 2019-04-16 20:33    [W:0.100 / U:5.052 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site