Re: [patch 03/13] io-accounting: write accounting

[Suleiman Souhlal <ssouhlal@FreeBSD.org>]

Suleiman Souhlal wrote:
> Andrew Morton wrote:
> 
>> On Wed, 13 Dec 2006 00:45:50 -0800
>> Suleiman Souhlal <ssouhlal@FreeBSD.org> wrote:
>>
>>
>>> akpm@osdl.org wrote:
>>>
>>>> From: Andrew Morton <akpm@osdl.org>
>>>>
>>>> Accounting writes is fairly simple: whenever a process flips a page 
>>>> from clean
>>>> to dirty, we accuse it of having caused a write to underlying 
>>>> storage of
>>>> PAGE_CACHE_SIZE bytes.
>>>
>>>
>>> On architectures where dirtying a page doesn't cause a page fault 
>>> (like i386), couldn't you end up billing the wrong process (in fact, 
>>> I think that even on other archituctures set_page_dirty() doesn't get 
>>> called immediately in the page fault handler)? 
>>
>>
>>
>> Yes, that would be a problem in 2.6.18 and earlier.
>>
>> In 2.6.19 and later, we do take a fault when transitioning a page from
>> pte-clean to pte-dirty.  That was done to get the dirty-page accounting
>> right - to avoid the 
>> all-of-memory-is-dirty-but-the-kernel-doesn't-know-it
>> problem.
> 
> 
> Ah yes indeed. I'm unable to keep up with all the new developments. :-(
> 
> However, if my understanding of this code is correct, it seems that the
> page fault is only done for shared writable VMAs, so you still can't
> rely on set_page_dirty() always being called by the process that
> dirtied the page in the first place.
> 
> Am I wrong?

Yes I am.
The only I/O non-shared VMAs might cause is from swapping, and I'm not
sure if the io accounting patches actually care about that.
My confusion came from the term "shared": A VMA is considered shared
whenever MAP_SHARED is specified, even if it only has only one single
"user".

-- Suleiman