Messages in this thread | | | Date | Wed, 22 Apr 2009 15:13:37 +0900 | From | KAMEZAWA Hiroyuki <> | Subject | Re: [PATCH] memcg: remove trylock_page_cgroup |
| |
On Tue, 21 Apr 2009 23:01:47 -0700 Andrew Morton <akpm@linux-foundation.org> wrote:
> On Wed, 22 Apr 2009 13:41:08 +0900 KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote: > > > > I expect that it will reliably fail if the caller is running as > > > SCHED_FIFO and the machine is single-CPU, or if we're trying to yield > > > to a SCHED_OTHER task which is pinned to this CPU, etc. The cond_resched() > > > won't work. > > > > > Hm, signal_pending() is supported now (so special user scan use alaram()) > > I used yield() before cond_resched() but I was told don't use it. > > Should I replace cond_resched() with congestion_wait(HZ/10) or some ? > > msleep(1) would be typical. That can also be used to give a > predictable number of seconds for the timeout. > > If 1 millisecond is too coarse then it's possible to sleep for much > shorter intervals if the platform implements hi-res timers. We don't > appear to have a handy interface to that (usleep, microsleep, > nanosleep, etc?). > > And an attempt to sleep for 1us will fall back to 1/HZ if the platform > doesn't implement hi-res timers, so that loop will need to be turned > into a do {} while(!timer_after(jiffies, start))) thing. Probably it > should be converted to that anyway, to be better behaved/predictable, > etc. > I'm now consdiering to make this loop _short_ and call lru_add_drain_all() when the loop finds a page is busy. (lru_add_drain_all() will sleep until the end of keventd.) And yes, checking jiffies and do sleep if it seems necessary.
I'll write some RFC today.
Thanks, -Kame
| |