Messages in this thread Patches in this message | | | Date | Tue, 17 May 2011 08:50:44 +0900 | Subject | Re: [PATCH 4/4] mm: vmscan: If kswapd has been running too long, allow it to sleep | From | Minchan Kim <> |
| |
On Mon, May 16, 2011 at 7:27 PM, Mel Gorman <mgorman@suse.de> wrote: > On Mon, May 16, 2011 at 05:58:59PM +0900, Minchan Kim wrote: >> On Mon, May 16, 2011 at 5:45 PM, Mel Gorman <mgorman@suse.de> wrote: >> > On Mon, May 16, 2011 at 02:04:00PM +0900, Minchan Kim wrote: >> >> On Mon, May 16, 2011 at 1:21 PM, James Bottomley >> >> <James.Bottomley@hansenpartnership.com> wrote: >> >> > On Sun, 2011-05-15 at 19:27 +0900, KOSAKI Motohiro wrote: >> >> >> (2011/05/13 23:03), Mel Gorman wrote: >> >> >> > Under constant allocation pressure, kswapd can be in the situation where >> >> >> > sleeping_prematurely() will always return true even if kswapd has been >> >> >> > running a long time. Check if kswapd needs to be scheduled. >> >> >> > >> >> >> > Signed-off-by: Mel Gorman<mgorman@suse.de> >> >> >> > --- >> >> >> > mm/vmscan.c | 4 ++++ >> >> >> > 1 files changed, 4 insertions(+), 0 deletions(-) >> >> >> > >> >> >> > diff --git a/mm/vmscan.c b/mm/vmscan.c >> >> >> > index af24d1e..4d24828 100644 >> >> >> > --- a/mm/vmscan.c >> >> >> > +++ b/mm/vmscan.c >> >> >> > @@ -2251,6 +2251,10 @@ static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining, >> >> >> > unsigned long balanced = 0; >> >> >> > bool all_zones_ok = true; >> >> >> > >> >> >> > + /* If kswapd has been running too long, just sleep */ >> >> >> > + if (need_resched()) >> >> >> > + return false; >> >> >> > + >> >> >> >> >> >> Hmm... I don't like this patch so much. because this code does >> >> >> >> >> >> - don't sleep if kswapd got context switch at shrink_inactive_list >> >> > >> >> > This isn't entirely true: need_resched() will be false, so we'll follow >> >> > the normal path for determining whether to sleep or not, in effect >> >> > leaving the current behaviour unchanged. >> >> > >> >> >> - sleep if kswapd didn't >> >> > >> >> > This also isn't entirely true: whether need_resched() is true at this >> >> > point depends on a whole lot more that whether we did a context switch >> >> > in shrink_inactive. It mostly depends on how long we've been running >> >> > without giving up the CPU. Generally that will mean we've been round >> >> > the shrinker loop hundreds to thousands of times without sleeping. >> >> > >> >> >> It seems to be semi random behavior. >> >> > >> >> > Well, we have to do something. Chris Mason first suspected the hang was >> >> > a kswapd rescheduling problem a while ago. We tried putting >> >> > cond_rescheds() in several places in the vmscan code, but to no avail. >> >> >> >> Is it a result of test with patch of Hannes(ie, !pgdat_balanced)? >> >> >> >> If it isn't, it would be nop regardless of putting cond_reshed at vmscan.c. >> >> Because, although we complete zone balancing, kswapd doesn't sleep as >> >> pgdat_balance returns wrong result. And at last VM calls >> >> balance_pgdat. In this case, balance_pgdat returns without any work as >> >> kswap couldn't find zones which have not enough free pages and goto >> >> out. kswapd could repeat this work infinitely. So you don't have a >> >> chance to call cond_resched. >> >> >> >> But if your test was with Hanne's patch, I am very curious how come >> >> kswapd consumes CPU a lot. >> >> >> >> > The need_resched() in sleeping_prematurely() seems to be about the best >> >> > option. The other option might be just to put a cond_resched() in >> >> > kswapd_try_to_sleep(), but that will really have about the same effect. >> >> >> >> I don't oppose it but before that, I think we have to know why kswapd >> >> consumes CPU a lot although we applied Hannes' patch. >> >> >> > >> > Because it's still possible for processes to allocate pages at the same >> > rate kswapd is freeing them leading to a situation where kswapd does not >> > consider the zone balanced for prolonged periods of time. >> >> We have cond_resched in shrink_page_list, shrink_slab and balance_pgdat. >> So I think kswapd can be scheduled out although it's scheduled in >> after a short time as task scheduled also need page reclaim. Although >> all task in system need reclaim, kswapd cpu 99% consumption is a >> natural result, I think. >> Do I miss something? >> > > Lets see; > > shrink_page_list() only applies if inactive pages were isolated > which in turn may not happen if all_unreclaimable is set in > shrink_zones(). If for whatver reason, all_unreclaimable is > set on all zones, we can miss calling cond_resched(). > > shrink_slab only applies if we are reclaiming slab pages. If the first > shrinker returns -1, we do not call cond_resched(). If that > first shrinker is dcache and __GFP_FS is not set, direct > reclaimers will not shrink at all. However, if there are > enough of them running or if one of the other shrinkers > is running for a very long time, kswapd could be starved > acquiring the shrinker_rwsem and never reaching the > cond_resched().
Don't we have to move cond_resched?
diff --git a/mm/vmscan.c b/mm/vmscan.c index 292582c..633e761 100644 --- a/mm/vmscan.c +++ b/mm/vmscan.c @@ -231,8 +231,10 @@ unsigned long shrink_slab(struct shrink_control *shrink, if (scanned == 0) scanned = SWAP_CLUSTER_MAX;
- if (!down_read_trylock(&shrinker_rwsem)) - return 1; /* Assume we'll be able to shrink next time */ + if (!down_read_trylock(&shrinker_rwsem)) { + ret = 1; + goto out; /* Assume we'll be able to shrink next time */ + }
list_for_each_entry(shrinker, &shrinker_list, list) { unsigned long long delta; @@ -280,12 +282,14 @@ unsigned long shrink_slab(struct shrink_control *shrink, count_vm_events(SLABS_SCANNED, this_scan); total_scan -= this_scan;
- cond_resched(); }
shrinker->nr += total_scan; + cond_resched(); } up_read(&shrinker_rwsem); +out: + cond_resched(); return ret; }
> > balance_pgdat() only calls cond_resched if the zones are not > balanced. For a high-order allocation that is balanced, it > checks order-0 again. During that window, order-0 might have > become unbalanced so it loops again for order-0 and returns > that was reclaiming for order-0 to kswapd(). It can then find > that a caller has rewoken kswapd for a high-order and re-enters > balance_pgdat() without ever have called cond_resched().
If kswapd reclaims order-o followed by high order, it would have a chance to call cond_resched in shrink_page_list. But if all zones are all_unreclaimable is set, balance_pgdat could return any work. Okay. It does make sense. By your scenario, someone wakes up kswapd with higher order, again. So re-enters balance_pgdat without ever have called cond_resched. But if someone wakes up higher order again, we can't have a chance to call kswapd_try_to_sleep. So your patch effect would be nop, too.
It would be better to put cond_resched after balance_pgdat?
diff --git a/mm/vmscan.c b/mm/vmscan.c index 292582c..61c45d0 100644 --- a/mm/vmscan.c +++ b/mm/vmscan.c @@ -2753,6 +2753,7 @@ static int kswapd(void *p) if (!ret) { trace_mm_vmscan_kswapd_wake(pgdat->node_id, order); order = balance_pgdat(pgdat, order, &classzone_idx); + cond_resched(); } } return 0; > > While it appears unlikely, there are bad conditions which can result > in cond_resched() being avoided.
> > -- > Mel Gorman > SUSE Labs >
-- Kind regards, Minchan Kim -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |