From: Andy Lutomirski <luto@amacapital.net>
To: Ingo Molnar <mingo@kernel.org>
Cc: linux-mm@kvack.org, linux-kernel@vger.kernel.org,
Andrew Morton <akpm@linux-foundation.org>,
Al Viro <viro@zeniv.linux.org.uk>,
Michel Lespinasse <walken@google.com>,
Hugh Dickins <hughd@google.com>, "J??rn Engel" <joern@logfs.org>,
Linus Torvalds <torvalds@linux-foundation.org>
Subject: Re: [PATCH v2] mm: Downgrade mmap_sem before locking or populating on mmap
Date: Sun, 16 Dec 2012 09:52:44 -0800 [thread overview]
Message-ID: <CALCETrX=3oQMKMNF2L3K7ur35KpeiqUN12RMq3XvtRChh9OJkg@mail.gmail.com> (raw)
In-Reply-To: <20121216090026.GB21690@gmail.com>
On Sun, Dec 16, 2012 at 1:00 AM, Ingo Molnar <mingo@kernel.org> wrote:
>
> * Andy Lutomirski <luto@amacapital.net> wrote:
>
>> This is a serious cause of mmap_sem contention. MAP_POPULATE
>> and MCL_FUTURE, in particular, are disastrous in multithreaded programs.
>>
>> Signed-off-by: Andy Lutomirski <luto@amacapital.net>
>> ---
>>
>> Changes from v1:
>>
>> The non-unlocking versions of do_mmap_pgoff and mmap_region are still
>> available for aio_setup_ring's benefit. In theory, aio_setup_ring
>> would do better with a lock-downgrading version, but that would be
>> somewhat ugly and doesn't help my workload.
>>
>> arch/tile/mm/elf.c | 9 +++---
>> fs/aio.c | 4 +++
>> include/linux/mm.h | 19 ++++++++++--
>> ipc/shm.c | 6 ++--
>> mm/fremap.c | 10 ++++--
>> mm/mmap.c | 89 ++++++++++++++++++++++++++++++++++++++++++++++++------
>> mm/util.c | 3 +-
>> 7 files changed, 117 insertions(+), 23 deletions(-)
>
>> +unsigned long mmap_region(struct file *file, unsigned long addr,
>> + unsigned long len, unsigned long flags,
>> + vm_flags_t vm_flags, unsigned long pgoff)
>> +{
>> + return mmap_region_helper(file, addr, len, flags, vm_flags, pgoff, 0);
>> +}
>> +
>
> That 0 really wants to be NULL ...
Sigh. I blame C++11 -- I wanted to type nullptr, but that's no good :)
>
> Also, with your patch applied there's no user of mmap_region()
> left anymore.
>
> More fundamentally, while I agree with the optimization,
> couldn't we de-uglify it a bit more?
>
> In particular, instead of this wrappery:
>
>> +unsigned long mmap_region_unlock(struct file *file, unsigned long addr,
>> + unsigned long len, unsigned long flags,
>> + vm_flags_t vm_flags, unsigned long pgoff)
>> +{
>> + int downgraded = 0;
>> + unsigned long ret = mmap_region_helper(file, addr, len,
>> + flags, vm_flags, pgoff, &downgraded);
>> +
>> + if (downgraded)
>> + up_read(¤t->mm->mmap_sem);
>> + else
>> + up_write(¤t->mm->mmap_sem);
>> +
>> + return ret;
>> +}
>
> 1)
>
> We could at minimum wrap up the conditional unlocking as:
>
> up_read_write(&mm->mmap_sem, read_locked);
>
> With that I'd also suggest to rename 'downgraded' to
> 'read_locked', which more clearly expresses the locking state.
>
> 2)
>
> More aggressively, we could just make it the _rule_ that the mm
> lock gets downgraded to read in mmap_region_helper(), no matter
> what.
>
> From a quick look I *think* all the usage sites (including
> sys_aio_setup()) are fine with that unlocking - but I could be
> wrong.
They are.
>
> There's a couple of shorter codepaths that would now see an
> extra op of downgrading:
>
> down_write(&mm->mmap_sem);
> ...
> downgrade_write(&mm->mmap_sem);
> ...
> up_read(&mm->mmap_sem);
>
> with not much work done with the lock read-locked - but I think
> they are all fine and mostly affect error paths. So there's no
> real value in keeping the conditional nature of the unlocking I
> think.
There's also the normal (i.e. neither lock nor populate) success path.
Does this matter? Presumably downgrade_write + up_read isn't much
slower than up_write.
--Andy
next prev parent reply other threads:[~2012-12-16 17:53 UTC|newest]
Thread overview: 22+ messages / expand[flat|nested] mbox.gz Atom feed top
2012-12-14 5:49 [PATCH] mm: Downgrade mmap_sem before locking or populating on mmap Andy Lutomirski
2012-12-14 7:27 ` Al Viro
2012-12-14 11:14 ` Andy Lutomirski
2012-12-14 14:49 ` Al Viro
2012-12-14 16:12 ` Andy Lutomirski
2012-12-16 8:41 ` Ingo Molnar
2012-12-16 17:04 ` Al Viro
2012-12-16 17:48 ` Al Viro
2012-12-16 18:49 ` Johannes Weiner
2012-12-16 19:53 ` Al Viro
2012-12-16 20:16 ` Al Viro
2012-12-15 2:17 ` [PATCH v2] " Andy Lutomirski
2012-12-16 9:00 ` Ingo Molnar
2012-12-16 17:52 ` Andy Lutomirski [this message]
2012-12-17 9:52 ` Ingo Molnar
2012-12-18 0:54 ` [PATCH v3] " Andy Lutomirski
2012-12-20 2:22 ` Simon Jeons
2012-12-16 12:39 ` [PATCH v2] " Michel Lespinasse
2012-12-16 18:05 ` Andy Lutomirski
2012-12-17 3:29 ` Michel Lespinasse
2012-12-17 22:01 ` Andy Lutomirski
2012-12-16 19:58 ` 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='CALCETrX=3oQMKMNF2L3K7ur35KpeiqUN12RMq3XvtRChh9OJkg@mail.gmail.com' \
--to=luto@amacapital.net \
--cc=akpm@linux-foundation.org \
--cc=hughd@google.com \
--cc=joern@logfs.org \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-mm@kvack.org \
--cc=mingo@kernel.org \
--cc=torvalds@linux-foundation.org \
--cc=viro@zeniv.linux.org.uk \
--cc=walken@google.com \
/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).