Re: Writes blocked on wait_for_stable_page (Writes of less than page size sometimes take too long)

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Wed, Jan 28, 2015 at 03:36:02PM -0800, Nikhilesh Reddy wrote:
> Also can someone please confirm if the following patch  help address
> this issue That I described below?
> 
> https://kernel.googlesource.com/pub/scm/linux/kernel/git/tytso/ext4/+/7afe5aa59ed3da7b6161617e7f157c7c680dc41e
> 
> Quick Summary:
> 
> in the function
> wait_on_page_writeback(page)  inside ext4_da_write_begin seems to
> take long and the process is stuck waiting.
> 
> The writeback just seems to take long.. but only when we write to the
> same page over again...
> The grab_cache_page_write_begin gives the same page until we  write
> enough...
> Is there a wait to not wait for the writeback?

Oh, right.  <smacks head>

Yes, that wait_for_page_writeback() was replaced with a call to
wait_for_stable_page() upstream, so it'll probably work for your 3.10
kernel.

(I also wonder why not jump to a newer LTS kernel, but that's neither
here nor there.)

<more below>

> On 01/28/2015 03:23 PM, Nikhilesh Reddy wrote:
> >
> >Hi Darrick
> >Thanks so much for your reply.
> >I apologize  the stable  page wait seems to be an odd outlier in that
> >run ... I ran multiple traces again.
> >
> >Please find my answers inline below.
> >
> >On 01/28/2015 01:39 PM, Darrick J. Wong wrote:
> >>On Wed, Jan 28, 2015 at 11:27:13AM -0800, Nikhilesh Reddy wrote:
> >>>Hi
> >>>I am working on a 64 bit Android device and have been trying to
> >>>improve performance for stream based data download (for example an
> >>>ftp)
> >>>The device has 3GB of ram and the dirty_ratio and
> >>>dirty_background_ratio are set to 5 and 1 respectively.
> >>>
> >>>Kernel 3.10 , Highmem is not enabled and the backing device is a
> >>>emmc and checksumming is not enabled
> >>Ok, 3.10 kernel is new enough that stable page writes only apply to
> >>devices that demand it, and apparently your eMMC demands it.
> >I tried checking my logs again .. and yes the stable pages dont apply
> >... that seems to be an odd instance where it took a while.
> >/My apologies.
> >
> >So On running many more trace runs it appears that the wait is actually
> >on in the function
> >
> >wait_on_page_writeback(page) ext4_da_write_begin
> >
> >Snippet below
> >
> >     /*
> >      * grab_cache_page_write_begin() can take a long time if the
> >      * system is thrashing due to memory pressure, or if the page
> >      * is being written back.  So grab it first before we start
> >      * the transaction handle.  This also allows us to allocate
> >      * the page (if needed) without using GFP_NOFS.
> >      */
> >retry_grab:
> >     page = grab_cache_page_write_begin(mapping, index, flags);
> >     if (!page)
> >         return -ENOMEM;
> >     unlock_page(page);
> >
> >retry_journal:
> >     handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
> >     if (IS_ERR(handle)) {
> >         page_cache_release(page);
> >         return PTR_ERR(handle);
> >     }
> >
> >     lock_page(page);
> >     if (page->mapping != mapping) {
> >         /* The page got truncated from under us */
> >         unlock_page(page);
> >         page_cache_release(page);
> >         ext4_journal_stop(handle);
> >         goto retry_grab;
> >     }
> >     wait_on_page_writeback(page); <<<<<<<<<<<<<<<<<<This is where the
> >wait is!
> >
> >
> >
> >The writeback just seems to take long.. but only when we write to the
> >same page over again...
> >The grab_cache_page_write_begin gives the same page until we  write
> >enough...
> >
> >Is there a wait to not wait for the writeback?
> >
> >
> >>
> >>>I noticed when profiling writes that if we dont use streamed IO (ie.
> >>>use write of whatever size data was read on the tcp stream) there
> >>>are some writes that seem to get blocked on
> >>>wait_for_stable_page.
> >>>
> >>>If I force the writes to be buffered in the userspace and ensure
> >>>writing 4k chunks the writes never seem to stall.
> >>That's consistent with a page being partially dirtied, written out,
> >>and partially dirtied again before write-out finishes.  If you buffer
> >>the incoming data such that a page is only dirtied once, you'll never
> >>notice wait_for_stable_page.
> >>
> >>Are you explicitly forcing writeout (i.e. fsync()) after every chunk
> >>arrives?  Or, is the rate of incoming data high enough such that we
> >>hit either dirty*ratio limit?  It isn't too hard to hit 30MB these
> >>days.  Why are you lowering the ratios from their defaults?
>
> >Using the higher values seems to cause longer stall ...when we do stall...
> >Smaller values are cause the write out to happen often but each one
> >takes less time.
> >
> >Setting it to the defaults causes a longer stall which seems to be
> >impacting TCP due to the delay in reading.
> >Additional info : dirty_expire_centisecs was set to 200 and the data
> >rate is close to 300Mbps. so yea 30MB should be less than a second.

Ah, you're concerned about the writeout latency, and it's important
that incoming data end up on flash as quickly as is convenient.  Hm.

That patch you found will probably get rid of the symptoms; I'd also
suggest fsync()ing after each page boundary instead of lowering the
writeback ratios since that increases the write load systemwide,
which will make the writeouts for your app take more time and wear out
the flash sooner.

--D

> >
> >Am I missing something obvious here?
> >
> >Please do let me know.
> >
> >Thanks so much for you help
> >
> >
> >>
> >>>I noticed there was earlier discussion on this and idea were
> >>>proposed to use snapshotting of the pages to avoid stalls...
> >>>For example: https://lwn.net/Articles/546658/
> >>>
> >>>But this seems to only snapshot ext3 ... (unless i misunderstood
> >>>what the patch is doing)
> >>>
> >>>Is there a similar patch to snapshot the buffers to not stall the
> >>>writes for ext4?
> >>No, there is not -- the problem with the snapshot solution is that it
> >>requires page allocations when the FS is (potentially) trying to
> >>reclaim memory by writing out dirty pages.
> >>
> >>--D
> >>
> >>>Please let me know.
> >>>
> >>>I would really appreciate any help you can give me.
> >>>
> >>>
> >>>--
> >>>Thanks
> >>>Nikhilesh Reddy
> >>>
> >>>Qualcomm Innovation Center, Inc.
> >>>The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
> >>>Forum,
> >>>a Linux Foundation Collaborative Project.
> >>>--
> >>>To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
> >>>the body of a message to majordomo@vger.kernel.org
> >>>More majordomo info at  http://vger.kernel.org/majordomo-info.html
> >
> >
> 
> 
> -- 
> Thanks
> Nikhilesh Reddy
> 
> Qualcomm Innovation Center, Inc.
> The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
> a Linux Foundation Collaborative Project.