lkml.org 
[lkml]   [2011]   [Nov]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: Latency writing to an mlocked ext4 mapping
On Tue, Nov 1, 2011 at 4:03 PM, Jan Kara <jack@suse.cz> wrote:
> On Mon 31-10-11 16:14:47, Andy Lutomirski wrote:
>> On Mon, Oct 31, 2011 at 4:10 PM, Jan Kara <jack@suse.cz> wrote:
>> > On Fri 28-10-11 16:37:03, Andy Lutomirski wrote:
>> >> On Tue, Oct 25, 2011 at 5:26 AM, Jan Kara <jack@suse.cz> wrote:
>> >> >>  - Why are we calling file_update_time at all?  Presumably we also
>> >> >> update the time when the page is written back (if not, that sounds
>> >> >> like a bug, since the contents may be changed after something saw the
>> >> >> mtime update), and, if so, why bother updating it on the first write?
>> >> >> Anything that relies on this behavior is, I think, unreliable, because
>> >> >> the page could be made writable arbitrarily early by another program
>> >> >> that changes nothing.
>> >> >  We don't update timestamp when the page is written back. I believe this
>> >> > is mostly because we don't know whether the data has been changed by a
>> >> > write syscall, which already updated the timestamp, or by mmap. That is
>> >> > also the reason why we update the timestamp at page fault time.
>> >> >
>> >> >  The reason why file_update_time() blocks for you is probably that it
>> >> > needs to get access to buffer where inode is stored on disk and because a
>> >> > transaction including this buffer is committing at the moment, your thread
>> >> > has to wait until the transaction commit finishes. This is mostly a problem
>> >> > specific to how ext4 works so e.g. xfs shouldn't have it.
>> >> >
>> >> >  Generally I believe the attempts to achieve any RT-like latencies when
>> >> > writing to a filesystem are rather hopeless. How much hopeless depends on
>> >> > the load of the filesystem (e.g., in your case of mostly idle filesystem I
>> >> > can imagine some tweaks could reduce your latencies to an acceptable level
>> >> > but once the disk gets loaded you'll be screwed). So I'd suggest that
>> >> > having RT thread just store log in memory (or write to a pipe) and have
>> >> > another non-RT thread write the data to disk would be a much more robust
>> >> > design.
>> >>
>> >> Windows seems to do pretty well at this, and I think it should be fixable on
>> >> Linux too.  "All" that needs to be done is to remove the pte_wrprotect from
>> >> page_mkclean_one.  The fallout from that might be unpleasant, though, but
>> >> it would probably speed up a number of workloads.
>> >  Well, but Linux's mm pretty much depends the pte_wrprotect() so that's
>> > unlikely to go away in a forseeable future. The reason is that we need to
>> > reliably account the number of dirty pages so that we can throttle
>> > processes that dirty too much of memory and also protect agaist system
>> > going into out-of-memory problems when too many pages would be dirty (and
>> > thus hard to reclaim). Thus we create clean pages as write-protected, when
>> > they are first written to, we account them as dirtied and unprotect them.
>> > When pages are cleaned by writeback, we decrement number of dirty pages
>> > accordingly and write-protect them again.
>>
>> What about skipping pte_wrprotect for mlocked pages and continuing to
>> account them dirty even if they're actually clean?  This should be a
>> straightforward patch except for the effect on stable pages for
>> writeback.  (It would also have unfortunate side effects on
>> ctime/mtime without my other patch to rearrange that code.)
>  Well, doing proper dirty accounting would be a mess (you'd have to
> unaccount dirty pages during munlock etc.) and I'm not sure what all would
> break when page writes would not be coupled with page faults. So I don't
> think it's really worth it.

I'll add it to my back burner. I haven't figured out all (any?) of
the accounting yet.

>
> Avoiding IO during a minor fault would be a decent thing which might be
> worth pursuing. As you properly noted "stable pages during writeback"
> requirement is one obstacle which won't be that trivial to avoid though...

There's an easy solution that would be good enough for me: add a mount
option to turn off stable pages.

Is the other problem just a race, perhaps? __block_page_mkwrite calls
__block_write_begin (which calls get_block, which I think is where the
latency comes from) *before* wait_on_page_writeback, which means that
there might not be any space allocated yet.

--Andy
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2011-11-02 00:13    [W:0.101 / U:43.564 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site