On Mon, Jul 12, 2021 at 5:17 AM Yunsheng Lin linyunsheng@huawei.com wrote:

As suggested by Alexander, "A DMA mapping should be page aligned anyway so the lower 12 bits would be reserved 0", so it might make more sense to repurpose the lower 12 bits of the dma address to store the pagecnt_bias for frag page support in page pool.

As newly added page_pool_get_pagecnt_bias() may be called outside of the softirq context, so annotate the access to page->dma_addr[0] with READ_ONCE() and WRITE_ONCE().

And page_pool_get_pagecnt_bias_ptr() is added to implement the pagecnt_bias atomic updating when a page is passsed to the user.

Other three interfaces using page->dma_addr[0] is only called in the softirq context during normal rx processing, hopefully the barrier in the rx processing will ensure the correct order between getting and setting pagecnt_bias.

Signed-off-by: Yunsheng Lin linyunsheng@huawei.com

include/net/page_pool.h | 29 +++++++++++++++++++++++++++-- net/core/page_pool.c | 8 +++++++- 2 files changed, 34 insertions(+), 3 deletions(-)

diff --git a/include/net/page_pool.h b/include/net/page_pool.h index 8d7744d..84cd972 100644 --- a/include/net/page_pool.h +++ b/include/net/page_pool.h @@ -200,17 +200,42 @@ static inline void page_pool_recycle_direct(struct page_pool *pool,

static inline dma_addr_t page_pool_get_dma_addr(struct page *page) {

•   dma_addr_t ret = page->dma_addr[0];
  

•   dma_addr_t ret = READ_ONCE(page->dma_addr[0]) & PAGE_MASK;
    if (sizeof(dma_addr_t) > sizeof(unsigned long))
            ret |= (dma_addr_t)page->dma_addr[1] << 16 << 16;
    return ret;
  

}

-static inline void page_pool_set_dma_addr(struct page *page, dma_addr_t addr) +static inline bool page_pool_set_dma_addr(struct page *page, dma_addr_t addr) {

•   if (WARN_ON(addr & ~PAGE_MASK))
  

•           return false;
  

•   page->dma_addr[0] = addr;
    if (sizeof(dma_addr_t) > sizeof(unsigned long))
            page->dma_addr[1] = upper_32_bits(addr);
  

•   return true;
  

+}

Rather than making this a part of the check here it might make more sense to pull this out and perform the WARN_ON after the check for dma_mapping_error.

Also it occurs to me that we only really have to do this in the case where dma_addr_t is larger than the size of a long. Otherwise we could just have the code split things so that dma_addr[0] is the dma_addr and dma_addr[1] is our pagecnt_bias value in which case we could probably just skip the check.

+static inline int page_pool_get_pagecnt_bias(struct page *page) +{

•   return READ_ONCE(page->dma_addr[0]) & ~PAGE_MASK;
  

+}

+static inline unsigned long *page_pool_pagecnt_bias_ptr(struct page *page) +{

•   return page->dma_addr;
  

+}

+static inline void page_pool_set_pagecnt_bias(struct page *page, int bias) +{

•   unsigned long dma_addr_0 = READ_ONCE(page->dma_addr[0]);
  

•   dma_addr_0 &= PAGE_MASK;
  

•   dma_addr_0 |= bias;
  

•   WRITE_ONCE(page->dma_addr[0], dma_addr_0);
  

}

static inline bool is_page_pool_compiled_in(void) diff --git a/net/core/page_pool.c b/net/core/page_pool.c index 78838c6..1abefc6 100644 --- a/net/core/page_pool.c +++ b/net/core/page_pool.c @@ -198,7 +198,13 @@ static bool page_pool_dma_map(struct page_pool *pool, struct page *page) if (dma_mapping_error(pool->p.dev, dma)) return false;

So instead of adding to the function below you could just add your WARN_ON check here with the unmapping call.

•   page_pool_set_dma_addr(page, dma);
  

•   if (unlikely(!page_pool_set_dma_addr(page, dma))) {
  

•           dma_unmap_page_attrs(pool->p.dev, dma,
  

•                                PAGE_SIZE << pool->p.order,
  

•                                pool->p.dma_dir,
  

•                                DMA_ATTR_SKIP_CPU_SYNC);
  

•           return false;
  

•   }
  
    if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
            page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
  

-- 2.7.4

[...]

...

...

+static inline bool page_pool_set_dma_addr(struct page *page, dma_addr_t addr) {

•   if (WARN_ON(addr & ~PAGE_MASK))
  

•           return false;
  

•    page->dma_addr[0] = addr;
     if (sizeof(dma_addr_t) > sizeof(unsigned long))
             page->dma_addr[1] = upper_32_bits(addr);
  

•   return true;
  

+}

Rather than making this a part of the check here it might make more sense to pull this out and perform the WARN_ON after the check for dma_mapping_error.

I need to point out that I don't like WARN_ON and BUG_ON code in fast-path code, because compiler adds 'ud2' assembler instructions that influences the instruction-cache fetching in the CPU. Yes, I have seen a measuresable impact from this before.

...

Also it occurs to me that we only really have to do this in the case where dma_addr_t is larger than the size of a long. Otherwise we could just have the code split things so that dma_addr[0] is the dma_addr and dma_addr[1] is our pagecnt_bias value in which case we could probably just skip the check.

The dance to get 64-bit DMA addr on 32-bit systems is rather ugly and confusing, sadly. We could take advantage of this, I just hope this will not make it uglier.

Note here that we can only use this because dma_addr is not aliased to compound page anymore (after the initial page_pool recycling patchset). We must keep this in mind if we even restructure struct page.

Can we do something more radical for this? The 64/32 bit dance is only there for 32 bit systems with 64 bit dma. Since the last time we asked about this no one seemed to care about these, and I really doubt we'll get an ethernet driver for them (that needs recycling....), can we *only* support frag allocation and recycling for 'normal' systems? We could always just r e-purpose dma_addr[1] for those.

Regards /Ilias

...

...

+static inline int page_pool_get_pagecnt_bias(struct page *page) +{

•   return READ_ONCE(page->dma_addr[0]) & ~PAGE_MASK;
  

+}

+static inline unsigned long *page_pool_pagecnt_bias_ptr(struct page *page) +{

•   return page->dma_addr;
  

+}

+static inline void page_pool_set_pagecnt_bias(struct page *page, int bias) +{

•   unsigned long dma_addr_0 = READ_ONCE(page->dma_addr[0]);
  

•   dma_addr_0 &= PAGE_MASK;
  

•   dma_addr_0 |= bias;
  

•   WRITE_ONCE(page->dma_addr[0], dma_addr_0);
  

  }

  static inline bool is_page_pool_compiled_in(void)

diff --git a/net/core/page_pool.c b/net/core/page_pool.c index 78838c6..1abefc6 100644 --- a/net/core/page_pool.c +++ b/net/core/page_pool.c @@ -198,7 +198,13 @@ static bool page_pool_dma_map(struct page_pool *pool, struct page *page) if (dma_mapping_error(pool->p.dev, dma)) return false;

So instead of adding to the function below you could just add your WARN_ON check here with the unmapping call.

...

•   page_pool_set_dma_addr(page, dma);
  

•   if (unlikely(!page_pool_set_dma_addr(page, dma))) {
  

•           dma_unmap_page_attrs(pool->p.dev, dma,
  

•                                PAGE_SIZE << pool->p.order,
  

•                                pool->p.dma_dir,
  

•                                DMA_ATTR_SKIP_CPU_SYNC);
  

•           return false;
  

•   }
  
     if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
             page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
  

-- 2.7.4

On 2021/7/13 3:46, Alexander Duyck wrote:

On Mon, Jul 12, 2021 at 5:17 AM Yunsheng Lin linyunsheng@huawei.com wrote:

...

Currently page pool only support page recycling only when there is only one user of the page, and the split page reusing implemented in the most driver can not use the page pool as bing-pong way of reusing requires the multi user support in page pool.

Those reusing or recycling has below limitations:

1. page from page pool can only be used be one user in order for the page recycling to happen.
2. Bing-pong way of reusing in most driver does not support multi desc using different part of the same page in order to save memory.

So add multi-users support and frag page recycling in page pool to overcome the above limitation.

Signed-off-by: Yunsheng Lin linyunsheng@huawei.com

include/net/page_pool.h | 22 ++++++++- net/core/page_pool.c | 126 +++++++++++++++++++++++++++++++++++++++++++----- 2 files changed, 134 insertions(+), 14 deletions(-)

diff --git a/include/net/page_pool.h b/include/net/page_pool.h index 84cd972..d9a736f 100644 --- a/include/net/page_pool.h +++ b/include/net/page_pool.h @@ -45,7 +45,10 @@ * Please note DMA-sync-for-CPU is still * device driver responsibility */ -#define PP_FLAG_ALL (PP_FLAG_DMA_MAP | PP_FLAG_DMA_SYNC_DEV) +#define PP_FLAG_PAGE_FRAG BIT(2) /* for page frag feature */ +#define PP_FLAG_ALL (PP_FLAG_DMA_MAP |\

•                            PP_FLAG_DMA_SYNC_DEV |\
  

•                            PP_FLAG_PAGE_FRAG)
  
  

/*

• Fast allocation side cache array/stack

@@ -88,6 +91,9 @@ struct page_pool { unsigned long defer_warn;

    u32 pages_state_hold_cnt;

•   unsigned int frag_offset;
  

•   int frag_bias;
  

•   struct page *frag_page;
  
    /*
     * Data structure for allocation side
  

@@ -137,6 +143,20 @@ static inline struct page *page_pool_dev_alloc_pages(struct page_pool *pool) return page_pool_alloc_pages(pool, gfp); }

+struct page *page_pool_alloc_frag(struct page_pool *pool,

•                             unsigned int *offset,
  

•                             unsigned int size,
  

•                             gfp_t gfp);
  

+static inline struct page *page_pool_dev_alloc_frag(struct page_pool *pool,

•                                               unsigned int *offset,
  

•                                               unsigned int size)
  

+{

•   gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN);
  

•   return page_pool_alloc_frag(pool, offset, size, gfp);
  

+}

/* get the stored dma direction. A driver might decide to treat this locally and

• avoid the extra cache line from page_pool to determine the direction

*/ diff --git a/net/core/page_pool.c b/net/core/page_pool.c index 1abefc6..39d5156 100644 --- a/net/core/page_pool.c +++ b/net/core/page_pool.c @@ -24,6 +24,8 @@ #define DEFER_TIME (msecs_to_jiffies(1000)) #define DEFER_WARN_INTERVAL (60 * HZ)

+#define BIAS_MAX (PAGE_SIZE - 1)

static int page_pool_init(struct page_pool *pool, const struct page_pool_params *params) { @@ -304,6 +306,33 @@ static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool, return page; }

+/* nr could be negative */ +static int page_pool_atomic_add_bias(struct page *page, int nr) +{

•   unsigned long *bias_ptr = page_pool_pagecnt_bias_ptr(page);
  

•   unsigned long old_bias = READ_ONCE(*bias_ptr);
  

•   unsigned long new_bias;
  

•   do {
  

•           int bias = (int)(old_bias & ~PAGE_MASK);
  

•           /* Warn when page_pool_dev_alloc_pages() is called
  

•            * with PP_FLAG_PAGE_FRAG flag in driver.
  

•            */
  

•           WARN_ON(!bias);
  

•           /* already the last user */
  

•           if (!(bias + nr))
  

•                   return 0;
  

•           new_bias = old_bias + nr;
  

•   } while (!try_cmpxchg(bias_ptr, &old_bias, new_bias));
  

•   WARN_ON((new_bias & PAGE_MASK) != (old_bias & PAGE_MASK));
  

•   return new_bias & ~PAGE_MASK;
  

+}

So instead of having a function to add bias it might make more sense to have a function that does a subtract and test since that is what we are really doing. So in a way you could think of it as a countdown until unmap or recycle.

ok

For the frags case we could probably default to one of two values. For the variable sized frags we default to a pagecnt_bias of PAGE_SIZE - 1 and then eventually do the subtraction at the end to free up unused count when the page has been fully used. For fixed sized frags you could theoretically just store off the PAGE_SIZE / frag_size - 1 and use that to initialize the pagecnt_bias of the pages as you use them.

Will default to a pagecnt_bias of PAGE_SIZE - 1.

Also one thing we may want to do is look at renaming this since we aren't messing with the page_ref_count anymore. It might make more sense to refer to this as something such as pp_frag_refcount. Also as per the other patch if we can get away from having to share the same space as DMA that would be even better.

page_pool_pagecnt_bias_ptr() is gone in the new version when we get away from having to share the same space as DMA.

...

/* For using page_pool replace: alloc_pages() API calls, but provide

• synchronization guarantee for allocation side.

*/ @@ -425,6 +454,11 @@ static __always_inline struct page * __page_pool_put_page(struct page_pool *pool, struct page *page, unsigned int dma_sync_size, bool allow_direct) {

•   /* It is not the last user for the page frag case */
  

•   if (pool->p.flags & PP_FLAG_PAGE_FRAG &&
  

•       page_pool_atomic_add_bias(page, -1))
  

•           return NULL;
  

•   /* This allocator is optimized for the XDP mode that uses
     * one-frame-per-page, but have fallbacks that act like the
     * regular page allocator APIs.
  

@@ -448,19 +482,7 @@ __page_pool_put_page(struct page_pool *pool, struct page *page, /* Page found as candidate for recycling */ return page; }

•   /* Fallback/non-XDP mode: API user have elevated refcnt.
  

•    *
  

•    * Many drivers split up the page into fragments, and some
  

•    * want to keep doing this to save memory and do refcnt based
  

•    * recycling. Support this use case too, to ease drivers
  

•    * switching between XDP/non-XDP.
  

•    *
  

•    * In-case page_pool maintains the DMA mapping, API user must
  

•    * call page_pool_put_page once.  In this elevated refcnt
  

•    * case, the DMA is unmapped/released, as driver is likely
  

•    * doing refcnt based recycle tricks, meaning another process
  

•    * will be invoking put_page.
  

•    */
  

•   /* Do not replace this with page_pool_return_page() */
    page_pool_release_page(pool, page);
    put_page(page);
  

@@ -517,6 +539,82 @@ void page_pool_put_page_bulk(struct page_pool *pool, void **data, } EXPORT_SYMBOL(page_pool_put_page_bulk);

+/* When BIAS_RESERVE to avoid frag page being recycled back to

• • page pool while the frag page is still in pool->frag_page
• • waiting for more user. As minimum align size for DMA seems to
• • be 32, so we support max size of 2047 * 32 for 4K page size.
• */

+#define BIAS_RESERVE ((int)(BIAS_MAX / 2 + 1)) +#define BIAS_NEGATIVE_RESERVE (0 - BIAS_RESERVE)

The explanation and these values don't make sense to me. BIAS_RESERVE would be 4095 / 2 + 1 which should come out to 2048, and BIAS_NEGATIVE_REVERSE is -2048. Should these values have been 127/-127 which would have been 4096 / 32 - 1?

The idea is to prevent the bias from being subtracted to zeor while the frag page is sitting in the page pool waiting for more users.

Anyway, the above comment is not needed if I take the "default to a pagecnt_bias of PAGE_SIZE - 1" suggestion.

...

+static struct page *page_pool_drain_frag(struct page_pool *pool,

•                                    struct page *page)
  

+{

•   /* page pool is not the last user */
  

•   if (page_pool_atomic_add_bias(page, pool->frag_bias +
  

•                                 BIAS_NEGATIVE_RESERVE))
  

•           return NULL;
  

•   else
  

•           return page;
  

+}

So this check isn't quite complete, and I probably wouldn't use it this way. To be complete you would need to check the page_ref_count and pfmemalloc flags. Also, we already have recycling into the page pool so we might as well just use that rather than trying to introduce yet another layer of recycling.

Yes, will add the the pfmemalloc and page_ref_count checking.

As introducing another layer of recycling, it seems we can not call page_pool_put_page() directly, and if the page is reusable, adding back to the pool and taking one from the pool seems unnecessary.

...

+static void page_pool_free_frag(struct page_pool *pool) +{

•   struct page *page = pool->frag_page;
  

•   if (!page ||
  

•       page_pool_atomic_add_bias(page, pool->frag_bias +
  

•                                 BIAS_NEGATIVE_RESERVE))
  

•           return;
  

•   page_pool_return_page(pool, page);
  

•   pool->frag_page = NULL;
  

+}

If we make the add_bias into a sub_bias then we are just needing to flip the values so that you are subtracting BIAS_RESERVE - pool->frag_bias which is much more readable in my opinion.

ok.

...

+struct page *page_pool_alloc_frag(struct page_pool *pool,

•                             unsigned int *offset,
  

•                             unsigned int size,
  

•                             gfp_t gfp)
  

+{

•   unsigned int max_size = PAGE_SIZE << pool->p.order;
  

•   unsigned int frag_offset = pool->frag_offset;
  

•   struct page *frag_page = pool->frag_page;
  

•   if (WARN_ON(!(pool->p.flags & PP_FLAG_PAGE_FRAG) ||
  

•               size > max_size))
  

•           return NULL;
  

•   size = ALIGN(size, dma_get_cache_alignment());
  

•   if (frag_page && frag_offset + size > max_size) {
  

•           frag_page = page_pool_drain_frag(pool, frag_page);
  

•           if (frag_page)
  

•                   goto frag_reset;
  

•   }
  

•   if (!frag_page) {
  

•           frag_page = page_pool_alloc_pages(pool, gfp);
  

•           if (unlikely(!frag_page)) {
  

•                   pool->frag_page = NULL;
  

•                   return NULL;
  

•           }
  

•           pool->frag_page = frag_page;
  

+frag_reset:

•           pool->frag_bias = 0;
  

•           frag_offset = 0;
  

•           page_pool_set_pagecnt_bias(frag_page, BIAS_RESERVE);
  

•   }
  

•   pool->frag_bias++;
  

•   *offset = frag_offset;
  

•   pool->frag_offset = frag_offset + size;
  

•   return frag_page;
  

+} +EXPORT_SYMBOL(page_pool_alloc_frag);

static void page_pool_empty_ring(struct page_pool *pool) { struct page *page; @@ -622,6 +720,8 @@ void page_pool_destroy(struct page_pool *pool) if (!page_pool_put(pool)) return;

•   page_pool_free_frag(pool);
  

•   if (!page_pool_release(pool))
            return;
  
  

-- 2.7.4

.