lkml.org 
[lkml]   [2019]   [Dec]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: Re: [Xen-devel] [PATCH v12 2/5] xenbus/backend: Protect xenbus callback with lock
Date
On Wed, 18 Dec 2019 16:11:51 +0100 "Jürgen Groß" <jgross@suse.com> wrote:

> On 18.12.19 15:40, SeongJae Park wrote:
> > On Wed, 18 Dec 2019 14:30:44 +0100 "Jürgen Groß" <jgross@suse.com> wrote:
> >
> >> On 18.12.19 13:42, SeongJae Park wrote:
> >>> On Wed, 18 Dec 2019 13:27:37 +0100 "Jürgen Groß" <jgross@suse.com> wrote:
> >>>
> >>>> On 18.12.19 11:42, SeongJae Park wrote:
> >>>>> From: SeongJae Park <sjpark@amazon.de>
> >>>>>
> >>>>> 'reclaim_memory' callback can race with a driver code as this callback
> >>>>> will be called from any memory pressure detected context. To deal with
> >>>>> the case, this commit adds a spinlock in the 'xenbus_device'. Whenever
> >>>>> 'reclaim_memory' callback is called, the lock of the device which passed
> >>>>> to the callback as its argument is locked. Thus, drivers registering
> >>>>> their 'reclaim_memory' callback should protect the data that might race
> >>>>> with the callback with the lock by themselves.
> >>>>
> >>>> Any reason you don't take the lock around the .probe() and .remove()
> >>>> calls of the backend (xenbus_dev_probe() and xenbus_dev_remove())? This
> >>>> would eliminate the need to do that in each backend instead.
> >>>
> >>> First of all, I would like to keep the critical section as small as possible.
> >>> With my small test, I could see slightly increasing memory pressure as the
> >>> critical section becomes wider. Also, some drivers might share the data their
> >>> 'reclaim_memory' callback touches with other functions. I think only the
> >>> driver owners can know what data is shared and what is the minimum critical
> >>> section to protect it.
> >>
> >> But this kind of serialization can still be added on top.
> >
> > I'm still worrying about the unnecessarily large critical section, but it might
> > be small enough to be ignored. If no others have strong objection, I will take
> > the lock around the '->probe()' and '->remove()'.
>
> The lock is per device, so contention is possible only for the
> reclaim case. In case probe or remove are running reclaim will have
> nothing to free (in probe case nothing is allocated yet, in remove
> case everything should be freed anyway). So the larger critical section
> is no problem at all IMO.

Agreed. I think I was worried about nothing really existing now.

>
> >> And with the trylock in the reclaim path I believe you can even avoid
> >> the irq variants of the spinlock. But I might be wrong, so you should
> >> try that with lockdep enabled. If it is working there is no harm done
> >> when making the critical section larger, as memory allocations will
> >> work as before.
> >
> > Yes, you're right. I will try test with lockdep.
>
> Thanks,

Good news, lockdep says it's okay :)

Will post next version soon.


Thanks,
SeongJae Park

>
>
> Juergen

\
 
 \ /
  Last update: 2019-12-18 18:33    [W:0.036 / U:24.184 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site