On Sun, Aug 04, 2019 at 07:48:01PM +0300, Nikolay Borisov wrote: > > > On 1.08.19 г. 5:17 ч., Dave Chinner wrote: > > From: Dave Chinner <dchinner@xxxxxxxxxx> > > > > Right now deferred work is picked up by whatever GFP_KERNEL context > > reclaimer that wins the race to empty the node's deferred work > > counter. However, if there are lots of direct reclaimers, that > > work might be continually picked up by contexts taht can't do any > > work and so the opportunities to do the work are missed by contexts > > that could do them. > > > > A further problem with the current code is that the deferred work > > can be picked up by a random direct reclaimer, resulting in that > > specific process having to do all the deferred reclaim work and > > hence can take extremely long latencies if the reclaim work blocks > > regularly. This is not good for direct reclaim fairness or for > > minimising long tail latency events. > > > > To avoid these problems, simply limit deferred work to kswapd > > contexts. We know kswapd is a context that can always do reclaim > > work, and hence deferring work to kswapd allows the deferred work to > > be done in the background and not adversely affect any specific > > process context doing direct reclaim. > > > > The advantage of this is that amount of work to be done in direct > > reclaim is now bound and predictable - it is entirely based on > > the cache's freeable objects and the reclaim priority. hence all > > direct reclaimers running at the same time should be doing > > relatively equal amounts of work, thereby reducing the incidence of > > long tail latencies due to uneven reclaim workloads. > > > > Signed-off-by: Dave Chinner <dchinner@xxxxxxxxxx> > > --- > > mm/vmscan.c | 93 ++++++++++++++++++++++++++++------------------------- > > 1 file changed, 50 insertions(+), 43 deletions(-) > > > > diff --git a/mm/vmscan.c b/mm/vmscan.c > > index b7472953b0e6..c583b4efb9bf 100644 > > --- a/mm/vmscan.c > > +++ b/mm/vmscan.c > > @@ -500,15 +500,15 @@ static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, > > struct shrinker *shrinker, int priority) > > { > > unsigned long freed = 0; > > - long total_scan; > > int64_t freeable_objects = 0; > > int64_t scan_count; > > - long nr; > > + int64_t scanned_objects = 0; > > + int64_t next_deferred = 0; > > + int64_t deferred_count = 0; > > long new_nr; > > int nid = shrinkctl->nid; > > long batch_size = shrinker->batch ? shrinker->batch > > : SHRINK_BATCH; > > - long scanned = 0, next_deferred; > > > > if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) > > nid = 0; > > @@ -519,47 +519,53 @@ static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, > > return scan_count; > > > > /* > > - * copy the current shrinker scan count into a local variable > > - * and zero it so that other concurrent shrinker invocations > > - * don't also do this scanning work. > > + * If kswapd, we take all the deferred work and do it here. We don't let > > + * direct reclaim do this, because then it means some poor sod is going > > + * to have to do somebody else's GFP_NOFS reclaim, and it hides the real > > + * amount of reclaim work from concurrent kswapd operations. Hence we do > > + * the work in the wrong place, at the wrong time, and it's largely > > + * unpredictable. > > + * > > + * By doing the deferred work only in kswapd, we can schedule the work > > + * according the the reclaim priority - low priority reclaim will do > > + * less deferred work, hence we'll do more of the deferred work the more > > + * desperate we become for free memory. This avoids the need for needing > > + * to specifically avoid deferred work windup as low amount os memory > > + * pressure won't excessive trim caches anymore. > > */ > > - nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); > > + if (current_is_kswapd()) { > > + int64_t deferred_scan; > > > > - total_scan = nr + scan_count; > > - if (total_scan < 0) { > > - pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n", > > - shrinker->scan_objects, total_scan); > > - total_scan = scan_count; > > - next_deferred = nr; > > - } else > > - next_deferred = total_scan; > > + deferred_count = atomic64_xchg(&shrinker->nr_deferred[nid], 0); > > > > - /* > > - * We need to avoid excessive windup on filesystem shrinkers > > - * due to large numbers of GFP_NOFS allocations causing the > > - * shrinkers to return -1 all the time. This results in a large > > - * nr being built up so when a shrink that can do some work > > - * comes along it empties the entire cache due to nr >>> > > - * freeable. This is bad for sustaining a working set in > > - * memory. > > - * > > - * Hence only allow the shrinker to scan the entire cache when > > - * a large delta change is calculated directly. > > - */ > > - if (scan_count < freeable_objects / 4) > > - total_scan = min_t(long, total_scan, freeable_objects / 2); > > + /* we want to scan 5-10% of the deferred work here at minimum */ > > + deferred_scan = deferred_count; > > + if (priority) > > + do_div(deferred_scan, priority); > > + scan_count += deferred_scan; > > + > > + /* > > + * If there is more deferred work than the number of freeable > > + * items in the cache, limit the amount of work we will carry > > + * over to the next kswapd run on this cache. This prevents > > + * deferred work windup. > > + */ > > + if (deferred_count > freeable_objects * 2) > > + deferred_count = freeable_objects * 2; > > nit : deferred_count = min(deferred_count, freeable_objects * 2). *nod* > How can we have more deferred objects than are currently on the LRU? deferred work is aggregated. Put enough direct reclaimers in action in GFP_NOFS context (e.g. fsmark create workload) and it will wind up the deferred count much faster than kswapd can drain it. > Aren't deferred objects always some part of freeable objects. For a single scan, yes. In aggregate, no. Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx