Re: [PATCH v2 2/3] CMA: aggressively allocate the pages on cma reserved memory when not used

From: Joonsoo Kim
Date: Thu May 29 2014 - 03:45:45 EST


On Thu, May 29, 2014 at 04:24:58PM +0900, Gioh Kim wrote:
> I've not understand your code fully. Please let me ask some silly questions.
>
> 2014-05-28 ìí 4:04, Joonsoo Kim ì ê:
> > CMA is introduced to provide physically contiguous pages at runtime.
> > For this purpose, it reserves memory at boot time. Although it reserve
> > memory, this reserved memory can be used for movable memory allocation
> > request. This usecase is beneficial to the system that needs this CMA
> > reserved memory infrequently and it is one of main purpose of
> > introducing CMA.
> >
> > But, there is a problem in current implementation. The problem is that
> > it works like as just reserved memory approach. The pages on cma reserved
> > memory are hardly used for movable memory allocation. This is caused by
> > combination of allocation and reclaim policy.
> >
> > The pages on cma reserved memory are allocated if there is no movable
> > memory, that is, as fallback allocation. So the time this fallback
> > allocation is started is under heavy memory pressure. Although it is under
> > memory pressure, movable allocation easily succeed, since there would be
> > many pages on cma reserved memory. But this is not the case for unmovable
> > and reclaimable allocation, because they can't use the pages on cma
> > reserved memory. These allocations regard system's free memory as
> > (free pages - free cma pages) on watermark checking, that is, free
> > unmovable pages + free reclaimable pages + free movable pages. Because
> > we already exhausted movable pages, only free pages we have are unmovable
> > and reclaimable types and this would be really small amount. So watermark
> > checking would be failed. It will wake up kswapd to make enough free
> > memory for unmovable and reclaimable allocation and kswapd will do.
> > So before we fully utilize pages on cma reserved memory, kswapd start to
> > reclaim memory and try to make free memory over the high watermark. This
> > watermark checking by kswapd doesn't take care free cma pages so many
> > movable pages would be reclaimed. After then, we have a lot of movable
> > pages again, so fallback allocation doesn't happen again. To conclude,
> > amount of free memory on meminfo which includes free CMA pages is moving
> > around 512 MB if I reserve 512 MB memory for CMA.
> >
> > I found this problem on following experiment.
> >
> > 4 CPUs, 1024 MB, VIRTUAL MACHINE
> > make -j16
> >
> > CMA reserve: 0 MB 512 MB
> > Elapsed-time: 225.2 472.5
> > Average-MemFree: 322490 KB 630839 KB
> >
> > To solve this problem, I can think following 2 possible solutions.
> > 1. allocate the pages on cma reserved memory first, and if they are
> > exhausted, allocate movable pages.
> > 2. interleaved allocation: try to allocate specific amounts of memory
> > from cma reserved memory and then allocate from free movable memory.
> >
> > I tested #1 approach and found the problem. Although free memory on
> > meminfo can move around low watermark, there is large fluctuation on free
> > memory, because too many pages are reclaimed when kswapd is invoked.
> > Reason for this behaviour is that successive allocated CMA pages are
> > on the LRU list in that order and kswapd reclaim them in same order.
> > These memory doesn't help watermark checking from kwapd, so too many
> > pages are reclaimed, I guess.
> >
> > So, I implement #2 approach.
> > One thing I should note is that we should not change allocation target
> > (movable list or cma) on each allocation attempt, since this prevent
> > allocated pages to be in physically succession, so some I/O devices can
> > be hurt their performance. To solve this, I keep allocation target
> > in at least pageblock_nr_pages attempts and make this number reflect
> > ratio, free pages without free cma pages to free cma pages. With this
> > approach, system works very smoothly and fully utilize the pages on
> > cma reserved memory.
> >
> > Following is the experimental result of this patch.
> >
> > 4 CPUs, 1024 MB, VIRTUAL MACHINE
> > make -j16
> >
> > <Before>
> > CMA reserve: 0 MB 512 MB
> > Elapsed-time: 225.2 472.5
> > Average-MemFree: 322490 KB 630839 KB
> > nr_free_cma: 0 131068
> > pswpin: 0 261666
> > pswpout: 75 1241363
> >
> > <After>
> > CMA reserve: 0 MB 512 MB
> > Elapsed-time: 222.7 224
> > Average-MemFree: 325595 KB 393033 KB
> > nr_free_cma: 0 61001
> > pswpin: 0 6
> > pswpout: 44 502
> >
> > There is no difference if we don't have cma reserved memory (0 MB case).
> > But, with cma reserved memory (512 MB case), we fully utilize these
> > reserved memory through this patch and the system behaves like as
> > it doesn't reserve any memory.
> >
> > With this patch, we aggressively allocate the pages on cma reserved memory
> > so latency of CMA can arise. Below is the experimental result about
> > latency.
> >
> > 4 CPUs, 1024 MB, VIRTUAL MACHINE
> > CMA reserve: 512 MB
> > Backgound Workload: make -jN
> > Real Workload: 8 MB CMA allocation/free 20 times with 5 sec interval
> >
> > N: 1 4 8 16
> > Elapsed-time(Before): 4309.75 9511.09 12276.1 77103.5
> > Elapsed-time(After): 5391.69 16114.1 19380.3 34879.2
> >
> > So generally we can see latency increase. Ratio of this increase
> > is rather big - up to 70%. But, under the heavy workload, it shows
> > latency decrease - up to 55%. This may be worst-case scenario, but
> > reducing it would be important for some system, so, I can say that
> > this patch have advantages and disadvantages in terms of latency.
> >
> > Although I think that this patch is right direction for CMA, there is
> > side-effect in following case. If there is small memory zone and CMA
> > occupys most of them, LRU for this zone would have many CMA pages. When
> > reclaim is started, these CMA pages would be reclaimed, but not counted
> > for watermark checking, so too many CMA pages could be reclaimed
> > unnecessarily. Until now, this can't happen because free CMA pages aren't
> > used easily. But, with this patch, free CMA pages are used easily so
> > this problem can be possible. I will handle it on another patchset
> > after some investigating.
> >
> > v2: In fastpath, just replenish counters. Calculation is done whenver
> > cma area is varied
> >
> > Acked-by: Michal Nazarewicz <mina86@xxxxxxxxxx>
> > Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx>
> >
> > diff --git a/arch/powerpc/kvm/book3s_hv_cma.c b/arch/powerpc/kvm/book3s_hv_cma.c
> > index d9d3d85..84a7582 100644
> > --- a/arch/powerpc/kvm/book3s_hv_cma.c
> > +++ b/arch/powerpc/kvm/book3s_hv_cma.c
> > @@ -132,6 +132,8 @@ struct page *kvm_alloc_cma(unsigned long nr_pages, unsigned long align_pages)
> > if (ret == 0) {
> > bitmap_set(cma->bitmap, pageno, nr_chunk);
> > page = pfn_to_page(pfn);
> > + adjust_managed_cma_page_count(page_zone(page),
> > + nr_pages);
>
> I think it should be -nr_pages to decrease the managed_cma_pages variable.
> But it is not. I think there is a reason.
> Why the managed_cma_pages is increased by allocation?

Hello, Gioh.

It's my mistake. It should be -nr_pages.
Thanks for pointing out.

>
> > memset(pfn_to_kaddr(pfn), 0, nr_pages << PAGE_SHIFT);
> > break;
> > } else if (ret != -EBUSY) {
> > @@ -180,6 +182,7 @@ bool kvm_release_cma(struct page *pages, unsigned long nr_pages)
> > (pfn - cma->base_pfn) >> (KVM_CMA_CHUNK_ORDER - PAGE_SHIFT),
> > nr_chunk);
> > free_contig_range(pfn, nr_pages);
> > + adjust_managed_cma_page_count(page_zone(pages), nr_pages);
> > mutex_unlock(&kvm_cma_mutex);
> >
> > return true;
> > @@ -210,6 +213,8 @@ static int __init kvm_cma_activate_area(unsigned long base_pfn,
> > }
> > init_cma_reserved_pageblock(pfn_to_page(base_pfn));
> > } while (--i);
> > + adjust_managed_cma_page_count(zone, count);
> > +
> > return 0;
> > }
> >
> > diff --git a/drivers/base/dma-contiguous.c b/drivers/base/dma-contiguous.c
> > index 165c2c2..c578d5a 100644
> > --- a/drivers/base/dma-contiguous.c
> > +++ b/drivers/base/dma-contiguous.c
> > @@ -160,6 +160,7 @@ static int __init cma_activate_area(struct cma *cma)
> > }
> > init_cma_reserved_pageblock(pfn_to_page(base_pfn));
> > } while (--i);
> > + adjust_managed_cma_page_count(zone, cma->count);
> >
> > return 0;
> > }
> > @@ -307,6 +308,7 @@ struct page *dma_alloc_from_contiguous(struct device *dev, int count,
> > if (ret == 0) {
> > bitmap_set(cma->bitmap, pageno, count);
> > page = pfn_to_page(pfn);
> > + adjust_managed_cma_page_count(page_zone(page), count);
>
> I think this also should be -count.

Ditto.

>
> > break;
> > } else if (ret != -EBUSY) {
> > break;
> > @@ -353,6 +355,7 @@ bool dma_release_from_contiguous(struct device *dev, struct page *pages,
> > mutex_lock(&cma_mutex);
> > bitmap_clear(cma->bitmap, pfn - cma->base_pfn, count);
> > free_contig_range(pfn, count);
> > + adjust_managed_cma_page_count(page_zone(pages), count);
> > mutex_unlock(&cma_mutex);
> >
> > return true;
> > diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> > index 39b81dc..51cffc1 100644
> > --- a/include/linux/gfp.h
> > +++ b/include/linux/gfp.h
> > @@ -415,6 +415,7 @@ extern int alloc_contig_range(unsigned long start, unsigned long end,
> > extern void free_contig_range(unsigned long pfn, unsigned nr_pages);
> >
> > /* CMA stuff */
> > +extern void adjust_managed_cma_page_count(struct zone *zone, long count);
> > extern void init_cma_reserved_pageblock(struct page *page);
> >
> > #endif
> > diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> > index fac5509..f52cb96 100644
> > --- a/include/linux/mmzone.h
> > +++ b/include/linux/mmzone.h
> > @@ -389,6 +389,20 @@ struct zone {
> > int compact_order_failed;
> > #endif
> >
> > +#ifdef CONFIG_CMA
> > + unsigned long managed_cma_pages;
> > + /*
> > + * Number of allocation attempt on each movable/cma type
> > + * without switching type. max_try(movable/cma) maintain
> > + * predefined calculated counter and replenish nr_try_(movable/cma)
> > + * with each of them whenever both of them are 0.
> > + */
> > + int nr_try_movable;
> > + int nr_try_cma;
> > + int max_try_movable;
> > + int max_try_cma;
> > +#endif
> > +
> > ZONE_PADDING(_pad1_)
> >
> > /* Fields commonly accessed by the page reclaim scanner */
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index 674ade7..ca678b6 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -788,6 +788,56 @@ void __init __free_pages_bootmem(struct page *page, unsigned int order)
> > }
> >
> > #ifdef CONFIG_CMA
> > +void adjust_managed_cma_page_count(struct zone *zone, long count)
> > +{
> > + unsigned long flags;
> > + long total, cma, movable;
> > +
> > + spin_lock_irqsave(&zone->lock, flags);
> > + zone->managed_cma_pages += count;
> > +
> > + total = zone->managed_pages;
> > + cma = zone->managed_cma_pages;
> > + movable = total - cma - high_wmark_pages(zone);
>
> If cma can be negative value, above calcuation increase movable value becuase -cma becomes positive value.
> Does it need a sign check?

This is leftover from version 1. They (totla, cma, movable)
can not be negative on this v2. I will fix it on v3.

>
>
>
> > +
> > + /* No cma pages, so do only movable allocation */
> > + if (cma <= 0) {
> > + zone->max_try_movable = pageblock_nr_pages;
> > + zone->max_try_cma = 0;
> > + goto out;
> > + }
> > +
> > + /*
> > + * We want to consume cma pages with well balanced ratio so that
> > + * we have consumed enough cma pages before the reclaim. For this
> > + * purpose, we can use the ratio, movable : cma. And we doesn't
> > + * want to switch too frequently, because it prevent allocated pages
> > + * from beging successive and it is bad for some sorts of devices.
> > + * I choose pageblock_nr_pages for the minimum amount of successive
> > + * allocation because it is the size of a huge page and fragmentation
> > + * avoidance is implemented based on this size.
> > + *
> > + * To meet above criteria, I derive following equation.
> > + *
> > + * if (movable > cma) then; movable : cma = X : pageblock_nr_pages
> > + * else (movable <= cma) then; movable : cma = pageblock_nr_pages : X
> > + */
> > + if (movable > cma) {
> > + zone->max_try_movable =
> > + (movable * pageblock_nr_pages) / cma;
>
> I think you assume that cma value cannot be negative. If cma can be negative, the resule of dividing by cma becomes negative. Right?

It cannot be negative.

>
> > + zone->max_try_cma = pageblock_nr_pages;
> > + } else {
> > + zone->max_try_movable = pageblock_nr_pages;
> > + zone->max_try_cma = cma * pageblock_nr_pages / movable;
> > + }
> > +
> > +out:
> > + zone->nr_try_movable = zone->max_try_movable;
> > + zone->nr_try_cma = zone->max_try_cma;
> > +
> > + spin_unlock_irqrestore(&zone->lock, flags);
> > +}
> > +
> > /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
> > void __init init_cma_reserved_pageblock(struct page *page)
> > {
> > @@ -1136,6 +1186,36 @@ __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
> > return NULL;
> > }
> >
> > +#ifdef CONFIG_CMA
> > +static struct page *__rmqueue_cma(struct zone *zone, unsigned int order)
> > +{
> > + struct page *page;
> > +
> > + if (zone->nr_try_movable > 0)
> > + goto alloc_movable;
> > +
> > + if (zone->nr_try_cma > 0) {
> > + /* Okay. Now, we can try to allocate the page from cma region */
> > + zone->nr_try_cma -= 1 << order;
> > + page = __rmqueue_smallest(zone, order, MIGRATE_CMA);
> > +
> > + /* CMA pages can vanish through CMA allocation */
> > + if (unlikely(!page && order == 0))
> > + zone->nr_try_cma = 0;
> > +
> > + return page;
> > + }
> > +
> > + /* Reset counter */
> > + zone->nr_try_movable = zone->max_try_movable;
> > + zone->nr_try_cma = zone->max_try_cma;
> > +
> > +alloc_movable:
> > + zone->nr_try_movable -= 1 << order;
> > + return NULL;
> > +}
> > +#endif
> > +
> > /*
> > * Do the hard work of removing an element from the buddy allocator.
> > * Call me with the zone->lock already held.
> > @@ -1143,10 +1223,15 @@ __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
> > static struct page *__rmqueue(struct zone *zone, unsigned int order,
> > int migratetype)
> > {
> > - struct page *page;
> > + struct page *page = NULL;
> > +
> > + if (IS_ENABLED(CONFIG_CMA) &&
>
> You might know that CONFIG_CMA is enabled and there is no CMA memory, because CONFIG_CMA_SIZE_MBYTES can be zero.
> Is IS_ENABLED(CONFIG_CMA) alright in that case?

next line checks whether zone->managed_cma_pages is positive or not.
If there is no CMA memory, zone->managed_cma_pages will be zero and
we will skip to call __rmqueue_cma().

Thanks for review!!!

Thanks.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/