Re: [PATCH, RFC] Don't do page stablization if !CONFIG_BLKDEV_INTEGRITY

[Andy Lutomirski <luto@amacapital.net>]

On 03/08/2012 08:43 AM, Sage Weil wrote:
> On Thu, 8 Mar 2012, Ted Ts'o wrote:
>> On Wed, Mar 07, 2012 at 10:27:43PM -0800, Sage Weil wrote:
>>>
>>> This avoids the problem for devices that don't need stable pages, but 
>>> doesn't help for those that do (btrfs, raid, iscsi, dif/dix, etc.).  It 
>>> seems to me like a more elegant solution would be to COW the page in the 
>>> address_space so that you get stable writeback pages without blocking.  
>>> That's clearly more complex, and I'm sure there are a range of issues 
>>> involved in making that work, but I would hope that it would be doable 
>>> with generic MM infrastructure so that everyone would benefit.
>>
>> Well, even doing a COW (or anything that involves messing with page
>> tables) is not free.  So even if we can make the cost of stable
>> writeback pages cheaper, if we can completely avoid the cost, this
>> would be good.  I'd also rather fix the performance regression sooner
>> rather than later, and I suspect the COW solution is not something
>> that could be prepared in time for the upcoming merge window.
> 
> Definitely.  This patch looks like a fine approach for your situation. I 
> just don't want the subject to come up without talking about a general 
> solution.  And it's very interesting to hear about a (simple) workload 
> that is affected by the wait_on_page_writeback().

I'll add a simple workload.  I have a soft real-time program that has
two threads.  One of them fallocates some files, mmaps them, mlocks
them, and touches all the pages to prefault them.  (This thread has no
real-time constraints -- it just needs to keep up.)  The other thread
writes to the files.

On Windows, this works very well.  On Linux without stable pages, it
almost works.  With stable pages, it's a complete disaster.  No amount
of minimizing the amount of time that pages under writeback can cause
writers to sleep will help -- writers *must not wait for io* when
writing mlocked, prefaulted pages for my code to work.

(The other issue involves file_update_time.  I'll send a fix eventually.)

FWIW, it would be really nice if there was a way to lock a mapping so
hard that accesses are guaranteed to not even cause soft faults.  We're
far from being able to do that now, though.

--Andy