From: Peter Xu <peterx@redhat.com>
To: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Guo Ren <guoren@kernel.org>,
Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
Gerald Schaefer <gerald.schaefer@de.ibm.com>,
Andrew Morton <akpm@linux-foundation.org>,
Andrea Arcangeli <aarcange@redhat.com>,
linux-csky@vger.kernel.org
Subject: Re: [PATCH 07/25] mm/csky: Use mm_fault_accounting()
Date: Thu, 18 Jun 2020 10:38:01 -0400 [thread overview]
Message-ID: <20200618143801.GK76766@xz-x1> (raw)
In-Reply-To: <CAHk-=wj_V2Tps2QrMn20_W0OJF9xqNh52XSGA42s-ZJ8Y+GyKw@mail.gmail.com>
On Wed, Jun 17, 2020 at 01:15:31PM -0700, Linus Torvalds wrote:
> On Wed, Jun 17, 2020 at 12:58 PM Peter Xu <peterx@redhat.com> wrote:
> >
> > But currently remote GUP will still do the page fault accounting on the remote
> > task_struct, am I right? E.g., when the get_user_pages_remote() is called with
> > "tsk != current", it seems the faultin_page() will still do maj_flt/min_flt
> > accounting for that remote task/thread?
>
> Well, that would be a data race and fundamentally buggy.
>
> It would be ok with something like ptrace (which only works when the
> target is quiescent), but is completely wrong otherwise.
>
> I guess it works fine in practice, and it's only statistics so even if
> you were to have a data race it doesn't much matter, but it's
> definitely conceptually very very wrong.
>
> The fault stats should be about who does the fault (they are about the
> _thread_) not about who the fault is done to (which is about the
> _mm_).
>
> Allocating the fault data to somebody else sounds frankly silly and
> stupid to me, exactly because it's (a) racy and (b) not even
> conceptually correct. The other thread literally _isn't_ doing a major
> page fault, for crissake!
>
> Now, there are some actual per-mm statistics too (the rss stuff etc),
> and it's fundamentally harder exactly because of the shared data. See
> the mm_counter stuff etc. Those are not about who does soemthing, they
> are about the resulting MM state.
I see. How about I move another small step further to cover GUP (to be
explicit, faultin_page() and fixup_user_fault())?
GUP needs the per-task accounting, but not the perf events. We can do that by
slightly changing the new approach into:
bool major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
if (major)
current->maj_flt++;
else
current->min_flt++;
if (!regs)
return ret;
if (major)
perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
else
perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
With the change we will always account that onto current task. Since there're
GUP calls that are not even accounted at all, e.g., get_user_pages_remote()
with tsk==NULL: for these calls, we also account these page faults onto current
task.
Another major benefit I can see (besides a completely cleaned up accounting) is
that we can remove the task_struct pointer in the whole GUP code, because
AFAICT that's only used for this pf accounting either in faultin_page() or
fixup_user_fault(), which seems questionable itself now to not use current...
Any opinions?
Thanks,
--
Peter Xu
next prev parent reply other threads:[~2020-06-18 14:38 UTC|newest]
Thread overview: 11+ messages / expand[flat|nested] mbox.gz Atom feed top
[not found] <20200615221607.7764-1-peterx@redhat.com>
2020-06-15 22:15 ` [PATCH 07/25] mm/csky: Use mm_fault_accounting() Peter Xu
2020-06-17 7:04 ` Guo Ren
2020-06-17 15:49 ` Peter Xu
2020-06-17 17:53 ` Linus Torvalds
2020-06-17 19:58 ` Peter Xu
2020-06-17 20:15 ` Linus Torvalds
2020-06-18 14:38 ` Peter Xu [this message]
2020-06-18 17:15 ` Linus Torvalds
2020-06-18 21:24 ` Peter Xu
2020-06-18 22:28 ` Peter Xu
2020-06-18 22:59 ` Linus Torvalds
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20200618143801.GK76766@xz-x1 \
--to=peterx@redhat.com \
--cc=aarcange@redhat.com \
--cc=akpm@linux-foundation.org \
--cc=gerald.schaefer@de.ibm.com \
--cc=guoren@kernel.org \
--cc=linux-csky@vger.kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=torvalds@linux-foundation.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).