On Tue, Jun 22, 2010 at 09:27:35PM +0800, Jan Kara wrote:

On Tue 22-06-10 12:07:27, Wu Fengguang wrote:

quoted

On Mon, Jun 21, 2010 at 03:42:39PM +0200, Jan Kara wrote:

quoted

On Fri 18-06-10 12:21:37, Peter Zijlstra wrote:

quoted

On Thu, 2010-06-17 at 20:04 +0200, Jan Kara wrote:

quoted

+               if (bdi_stat(bdi, BDI_WRITTEN) >= bdi->wb_written_head)
+                       bdi_wakeup_writers(bdi); 

For the paranoid amongst us you could make wb_written_head s64 and write
the above as:

  if (bdi_stat(bdi, BDI_WRITTEN) - bdi->wb_written_head > 0)

Which, if you assume both are monotonic and wb_written_head is always
within 2^63 of the actual bdi_stat() value, should give the same end
result and deal with wrap-around.

For when we manage to create a device that can write 2^64 pages in our
uptime :-)

  OK, the fix is simple enough so I've changed it, although I'm not
paranoic enough ;) (I actually did the math before writing that test).

a bit more change :)

type:

-       u64 wb_written_head
+       s64 wb_written_head

resetting:

-                       bdi->wb_written_head = ~(u64)0;
+                       bdi->wb_written_head = 0;

setting:

                bdi->wb_written_head = bdi_stat(bdi, BDI_WRITTEN) + wc->written;
+               bdi->wb_written_head |= 1;

testing:

        if (bdi->wb_written_head &&
            bdi_stat(bdi, BDI_WRITTEN) - bdi->wb_written_head > 0)

This avoids calling into bdi_wakeup_writers() pointlessly when no one
is being throttled (which is the normal case).

  Actually, I've already changed wb_written_head to s64. I kept setting
wb_written_head to s64 maximum. That also avoids calling into
bdi_wakeup_writers() unnecessarily...

Ah OK, I forgot bdi_stat() calls percpu_counter_read_positive() which
is always in range [0, s64 max].

Thanks,
Fengguang

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>