在 2024/12/13 15:56, Barry Song 写道:

On Fri, Dec 13, 2024 at 3:37 PM yangge1116@126.com wrote:

...

From: yangge yangge1116@126.com

Since commit 984fdba6a32e ("mm, compaction: use proper alloc_flags in __compaction_suitable()") allow compaction to proceed when free pages required for compaction reside in the CMA pageblocks, it's possible that __compaction_suitable() always returns true, and in some cases, it's not acceptable.

There are 4 NUMA nodes on my machine, and each NUMA node has 32GB of memory. I have configured 16GB of CMA memory on each NUMA node, and starting a 32GB virtual machine with device passthrough is extremely slow, taking almost an hour.

I don't fully understand why each node has a 16GB CMA. As I recall, I designed the per-NUMA CMA to support devices that are not behind the IOMMU, such as the IOMMU itself or certain device drivers which are not having IOMMU and need contiguous memory for DMA. These devices don't seem to require that much memory.

Our hardware supports setting specific protection for contiguous memory block, but the granularity of protection is relatively large, exceeding 4MB, which makes it unsuitable for allocation from buddy. Therefore, during system startup, a certain percentage of memory on each node is reserved as CMA memory, allowing for the allocation of large contiguous memory blocks through cma_alloc.

...

During the start-up of the virtual machine, it will call pin_user_pages_remote(..., FOLL_LONGTERM, ...) to allocate memory. Long term GUP cannot allocate memory from CMA area, so a maximum of 16 GB of no-CMA memory on a NUMA node can be used as virtual machine memory. Since there is 16G of free CMA memory on the NUMA node, watermark for order-0 always be met for compaction, so __compaction_suitable() always returns true, even if the node is unable to allocate non-CMA memory for the virtual machine.

For costly allocations, because __compaction_suitable() always returns true, __alloc_pages_slowpath() can't exit at the appropriate place, resulting in excessively long virtual machine startup times. Call trace: __alloc_pages_slowpath if (compact_result == COMPACT_SKIPPED || compact_result == COMPACT_DEFERRED) goto nopage; // should exit __alloc_pages_slowpath() from here

To sum up, during long term GUP flow, we should remove ALLOC_CMA both in __compaction_suitable() and __isolate_free_page().

What’s the outcome after your fix? Will it quickly fall back to remote NUMA nodes for the pin?

Starting a 32GB virtual machine with device passthrough takes only a free seconds. Yes, it will quickly fall back to remote NUMA nodes.

...

Fixes: 984fdba6a32e ("mm, compaction: use proper alloc_flags in __compaction_suitable()") Cc: stable@vger.kernel.org Signed-off-by: yangge yangge1116@126.com

---

mm/compaction.c | 8 +++++--- mm/page_alloc.c | 4 +++- 2 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/mm/compaction.c b/mm/compaction.c index 07bd227..044c2247 100644 --- a/mm/compaction.c +++ b/mm/compaction.c @@ -2384,6 +2384,7 @@ static bool __compaction_suitable(struct zone *zone, int order, unsigned long wmark_target) { unsigned long watermark;

•   bool pin;
     /*
      * Watermarks for order-0 must be met for compaction to be able to
      * isolate free pages for migration targets. This means that the
  

@@ -2395,14 +2396,15 @@ static bool __compaction_suitable(struct zone *zone, int order, * even if compaction succeeds. * For costly orders, we require low watermark instead of min for * compaction to proceed to increase its chances.

•    * ALLOC_CMA is used, as pages in CMA pageblocks are considered
  

•    * suitable migration targets
  

•    * In addition to long term GUP flow, ALLOC_CMA is used, as pages in
  

•    * CMA pageblocks are considered suitable migration targets
      */
     watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
                             low_wmark_pages(zone) : min_wmark_pages(zone);
     watermark += compact_gap(order);
  

•   pin = !!(current->flags & PF_MEMALLOC_PIN);
     return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
  

•                              ALLOC_CMA, wmark_target);
  

•                              pin ? 0 : ALLOC_CMA, wmark_target);
  

  }

  /*

diff --git a/mm/page_alloc.c b/mm/page_alloc.c index dde19db..9a5dfda 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -2813,6 +2813,7 @@ int __isolate_free_page(struct page *page, unsigned int order) { struct zone *zone = page_zone(page); int mt = get_pageblock_migratetype(page);

•   bool pin;
  
     if (!is_migrate_isolate(mt)) {
             unsigned long watermark;
  

@@ -2823,7 +2824,8 @@ int __isolate_free_page(struct page *page, unsigned int order) * exists. */ watermark = zone->_watermark[WMARK_MIN] + (1UL << order);

•           if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
  

•           pin = !!(current->flags & PF_MEMALLOC_PIN);
  

•           if (!zone_watermark_ok(zone, 0, watermark, 0, pin ? 0 : ALLOC_CMA))
                     return 0;
     }
  
  

-- 2.7.4

Thanks Barry

在 2024/12/13 16:23, Baolin Wang 写道:

On 2024/12/13 15:37, yangge1116@126.com wrote:

...

From: yangge yangge1116@126.com

Since commit 984fdba6a32e ("mm, compaction: use proper alloc_flags in __compaction_suitable()") allow compaction to proceed when free pages required for compaction reside in the CMA pageblocks, it's possible that __compaction_suitable() always returns true, and in some cases, it's not acceptable.

There are 4 NUMA nodes on my machine, and each NUMA node has 32GB of memory. I have configured 16GB of CMA memory on each NUMA node, and starting a 32GB virtual machine with device passthrough is extremely slow, taking almost an hour.

During the start-up of the virtual machine, it will call pin_user_pages_remote(..., FOLL_LONGTERM, ...) to allocate memory. Long term GUP cannot allocate memory from CMA area, so a maximum of 16 GB of no-CMA memory on a NUMA node can be used as virtual machine memory. Since there is 16G of free CMA memory on the NUMA node, watermark for order-0 always be met for compaction, so __compaction_suitable() always returns true, even if the node is unable to allocate non-CMA memory for the virtual machine.

For costly allocations, because __compaction_suitable() always returns true, __alloc_pages_slowpath() can't exit at the appropriate place, resulting in excessively long virtual machine startup times. Call trace: __alloc_pages_slowpath      if (compact_result == COMPACT_SKIPPED ||          compact_result == COMPACT_DEFERRED)          goto nopage; // should exit __alloc_pages_slowpath() from here

To sum up, during long term GUP flow, we should remove ALLOC_CMA both in __compaction_suitable() and __isolate_free_page().

Fixes: 984fdba6a32e ("mm, compaction: use proper alloc_flags in __compaction_suitable()") Cc: stable@vger.kernel.org Signed-off-by: yangge yangge1116@126.com

---

mm/compaction.c | 8 +++++---   mm/page_alloc.c | 4 +++-   2 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/mm/compaction.c b/mm/compaction.c index 07bd227..044c2247 100644 --- a/mm/compaction.c +++ b/mm/compaction.c @@ -2384,6 +2384,7 @@ static bool __compaction_suitable(struct zone *zone, int order,                     unsigned long wmark_target)   {       unsigned long watermark; +    bool pin;       /*        * Watermarks for order-0 must be met for compaction to be able to        * isolate free pages for migration targets. This means that the @@ -2395,14 +2396,15 @@ static bool __compaction_suitable(struct zone *zone, int order,        * even if compaction succeeds.        * For costly orders, we require low watermark instead of min for        * compaction to proceed to increase its chances. -     * ALLOC_CMA is used, as pages in CMA pageblocks are considered -     * suitable migration targets +     * In addition to long term GUP flow, ALLOC_CMA is used, as pages in +     * CMA pageblocks are considered suitable migration targets        */       watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?                   low_wmark_pages(zone) : min_wmark_pages(zone);       watermark += compact_gap(order); +    pin = !!(current->flags & PF_MEMALLOC_PIN);       return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx, -                   ALLOC_CMA, wmark_target); +                   pin ? 0 : ALLOC_CMA, wmark_target);   }

Seems a little hack for me. Using the 'cc->alloc_flags' passed from the caller to determin if ‘ALLOC_CMA’ is needed looks more reasonable to me.

Ok, thanks.

...

/* diff --git a/mm/page_alloc.c b/mm/page_alloc.c index dde19db..9a5dfda 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -2813,6 +2813,7 @@ int __isolate_free_page(struct page *page, unsigned int order)   {       struct zone *zone = page_zone(page);       int mt = get_pageblock_migratetype(page); +    bool pin;       if (!is_migrate_isolate(mt)) {           unsigned long watermark; @@ -2823,7 +2824,8 @@ int __isolate_free_page(struct page *page, unsigned int order)            * exists.            */           watermark = zone->_watermark[WMARK_MIN] + (1UL << order); -        if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA)) +        pin = !!(current->flags & PF_MEMALLOC_PIN); +        if (!zone_watermark_ok(zone, 0, watermark, 0, pin ? 0 : ALLOC_CMA))               return 0;       }