On Mon, 2011-06-27 at 12:03 +0100, Mel Gorman wrote: > On Fri, Jun 24, 2011 at 02:54:11PM -0400, Chris Mason wrote: > > Excerpts from Andi Kleen's message of 2011-06-24 14:44:12 -0400: > > > Andrew Lutomirski <luto@xxxxxxx> writes: > > > > > > [Putting the Intel graphics driver developers in cc.] > > > > > > > I'm back :-/ > > > > > > > > I just triggered the kswapd bug on 2.6.39.1, which has the > > > > cond_resched in shrink_slab. This time my system's still usable (I'm > > > > tying this email on it), but kswapd0 is taking 100% cpu. It *does* > > > > schedule (tested by setting its affinity the same as another CPU hog > > > > and confirming that each one gets 50%). > > > > > > > > It appears to be calling i915_gem_inactive_shrink in a loop. I have > > > > probes on entry and return of i915_gem_inactive_shrink and on return > > > > of shrink_slab. I see: > > > > > > > > kswapd0 47 [000] 59599.956573: mm_vmscan_kswapd_wake: nid=0 order=0 > > > > kswapd0 47 [000] 59599.956575: shrink_zone: > > > > (ffffffff810c848c) priority=12 zone=ffff8801005fe000 > > > > kswapd0 47 [000] 59599.956576: shrink_zone_return: > > > > (ffffffff810c848c <- ffffffff810c96c6) arg1=0 > > > > kswapd0 47 [000] 59599.956578: i915_gem_inactive_shrink: > > > > A similar trace came up a bunch of times in Jejb's NMI softlockup/kswapd > > consumes the machine thread. That one was tracked down to slub high > > order allocations. > > > > I'm sure that one is burned in on Mel's memory, but after a while the > > individual traces fell out of the thread, and I'm not sure the i915 part > > stuck out. > > > > I expect that Jejb's lockup is also fixed by "Stop kswapd consuming > 100% CPU when highest zone is small". i915 didn't help but at the end > of the day, kswapd shouldn't have been shrinking slab so aggressively. It will be a while before I can try this out, I'm afraid ... the laptop is currently on tour in Europe with its owner. I finally just downgraded it to FC13 which made most of the issues go away. James -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>