On Thu, Apr 4, 2013 at 12:00 AM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> On Wed, Apr 03, 2013 at 03:33:11PM -0400, Peter Watkins wrote:
>> Wondering if anyone has a suggestion for when
>> xfs_iomap_write_unwritten gets into congestion_wait.
> Do less IO?
>> In this case the system has almost half of normal zone pages in
>> NR_WRITEBACK with pretty much everybody held up in either
>> congestion_wait or balance_dirty_pages.
> Which is excessive - how are you getting to the point of having that
> many pages under IO at once? Writeback depth is limited by the IO
> elevator queue depths, so this shouldn't happen unless you've been
> tweaking block device parameters (i.e. nr_requests/max_sectors_kb)...
>> Since there are some free pages, seems like we'd be better off just
>> using a little more memory to finish this IO and in turn reduce pages
>> under write-back and add to free memory, rather than holding up here.
>> So maybe PF_MEMALLOC?
> Definitely not. Unwritten extent conversion can require hundreds of
> kilobytes of memory to complete, so all this will do is trigger even
> further exhaustion of memory reserves before we block on IO.
>> It also looks like this path allocates log vectors with KM_SLEEP but
>> lv_buf's with KM_SLEEP|KM_NOFS. Why is that?
> The transaction commit is copying the changes made into separate
> buffers to insert into the CIL for a later checkpoint to write to
> disk. This is normal behaviour - we can sleep there, but we cannot
> allow memory reclaim to recurse into the filesystem (for obvious
> Dave Chinner
Thanks for the help.
There are other clues the VM system was rather quickly overwhelmed,
i.e. it couldn't even get bdi flush threads started without sending
threadd into congestion_wait.
So indeed there is a big multi-threaded writer which starts all at
once, and that can be smoothed out.
And nr_requests is dialed up from 128 to 1024. Is anyone really able
to resist that temptation?