Date:	Wed, 31 Oct 2012 18:15:33 -0700
From:	Paul Turner <pjt@...gle.com>
To:	KOSAKI Motohiro <kosaki.motohiro@...il.com>
Cc:	Andrew Morton <akpm@...ux-foundation.org>,
	Minchan Kim <minchan@...nel.org>, linux-kernel@...r.kernel.org,
	linux-mm@...ck.org, John Stultz <john.stultz@...aro.org>,
	Christoph Lameter <cl@...ux.com>,
	Android Kernel Team <kernel-team@...roid.com>,
	Robert Love <rlove@...gle.com>, Mel Gorman <mel@....ul.ie>,
	Hugh Dickins <hughd@...gle.com>,
	Dave Hansen <dave@...ux.vnet.ibm.com>,
	Rik van Riel <riel@...hat.com>,
	Dave Chinner <david@...morbit.com>, Neil Brown <neilb@...e.de>,
	Mike Hommey <mh@...ndium.org>, Taras Glek <tglek@...illa.com>,
	KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>,
	sanjay@...gle.com, David Rientjes <rientjes@...gle.com>
Subject: Re: [RFC v2] Support volatile range for anon vma

On Wed, Oct 31, 2012 at 3:56 PM, KOSAKI Motohiro
<kosaki.motohiro@...il.com> wrote:
>>> > Allocator should call madvise(MADV_NOVOLATILE) before reusing for
>>> > allocating that area to user. Otherwise, accessing of volatile range
>>> > will meet SIGBUS error.
>>>
>>> Well, why?  It would be easy enough for the fault handler to give
>>> userspace a new, zeroed page at that address.
>>
>> Note: MADV_DONTNEED already has this (nice) property.
>
> I don't think I strictly understand this patch. but maybe I can answer why
> userland and malloc folks don't like MADV_DONTNEED.
>
> glibc malloc discard freed memory by using MADV_DONTNEED
> as tcmalloc. and it is often a source of large performance decrease.
> because of MADV_DONTNEED discard memory immediately and
> right after malloc() call fall into page fault and pagesize memset() path.
> then, using DONTNEED increased zero fill and cache miss rate.
>
> At called free() time, malloc don't have a knowledge when next big malloc()
> is called. then, immediate discarding may or may not get good performance
> gain. (Ah, ok, the rate is not 5:5. then usually it is worth. but not everytime)
>

Ah; In tcmalloc allocations (and their associated free-lists) are
binned into separate lists as a function of object-size which helps to
mitigate this.

I'd make a separate more general argument here:
If I'm allocating a large (multi-kilobyte object) the cost of what I'm
about to do with that object is likely fairly large -- The fault/zero
cost a probably fairly small proportional cost, which limits the
optimization value.

>
> In past, several developers tryied to avoid such situation, likes
>
> - making zero page daemon and avoid pagesize zero fill at page fault
> - making new vma or page flags and mark as discardable w/o swap and
>   vmscan treat it. (like this and/or MADV_FREE)
> - making new process option and avoid page zero fill from page fault path.
>   (yes, it is big incompatibility and insecure. but some embedded folks thought
>    they are acceptable downside)
> - etc
>
>
> btw, I'm not sure this patch is better for malloc because current MADV_DONTNEED
> don't need mmap_sem and works very effectively when a lot of threads case.
> taking mmap_sem might bring worse performance than DONTNEED. dunno.

MADV_VOLATILE also seems to end up looking quite similar to a
user-visible (range-based) cleancache.

A second popular use-case for such semantics is the case of
discardable cache elements (e.g. web browser).  I suspect we'd want to
at least mention these in the changelog.  (Alternatively, what does a
cleancache-backed-fs exposing these semantics look like?)
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists