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 elevated refcnt case 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.

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; +} + +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;

- 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);

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 elevated refcnt support.

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 elevated refcnt support in page pool to in order to overcome the above limitation.

This is a preparation to support allocating page frag in page pool.

Signed-off-by: Yunsheng Lin linyunsheng@huawei.com --- include/net/page_pool.h | 22 ++++++++- net/core/page_pool.c | 121 ++++++++++++++++++++++++++++++++++++++++++------ 2 files changed, 129 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..9f518dc 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; +} + /* 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 +539,77 @@ 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) + +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; +} + +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; +} + +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) { + frag_page = page_pool_alloc_pages(pool, gfp); + if (unlikely(!frag_page)) { + pool->frag_page = NULL; + return NULL; + } + + pool->frag_page = frag_page; + 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 +715,8 @@ 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;

This patch adds skb's frag page recycling support based on the elevated refcnt support in page pool.

The performance improves above 10~20% with IOMMU disabled. The performance improves about 200% when IOMMU is enabled and iperf server shares the same cpu with irq/NAPI.

Signed-off-by: Yunsheng Lin linyunsheng@huawei.com --- drivers/net/ethernet/hisilicon/hns3/hns3_enet.c | 79 +++++++++++++++++++++++-- drivers/net/ethernet/hisilicon/hns3/hns3_enet.h | 3 + 2 files changed, 77 insertions(+), 5 deletions(-)

diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c index cdb5f14..c799129 100644 --- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c +++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c @@ -3205,6 +3205,21 @@ static int hns3_alloc_buffer(struct hns3_enet_ring *ring, unsigned int order = hns3_page_order(ring); struct page *p;

+ if (ring->page_pool) { + p = page_pool_dev_alloc_frag(ring->page_pool, + &cb->page_offset, + hns3_buf_size(ring)); + if (unlikely(!p)) + return -ENOMEM; + + cb->priv = p; + cb->buf = page_address(p); + cb->dma = page_pool_get_dma_addr(p); + cb->type = DESC_TYPE_FRAG; + cb->reuse_flag = 0; + return 0; + } + p = dev_alloc_pages(order); if (!p) return -ENOMEM; @@ -3227,8 +3242,12 @@ static void hns3_free_buffer(struct hns3_enet_ring *ring, if (cb->type & (DESC_TYPE_SKB | DESC_TYPE_BOUNCE_HEAD | DESC_TYPE_BOUNCE_ALL | DESC_TYPE_SGL_SKB)) napi_consume_skb(cb->priv, budget); - else if (!HNAE3_IS_TX_RING(ring) && cb->pagecnt_bias) - __page_frag_cache_drain(cb->priv, cb->pagecnt_bias); + else if (!HNAE3_IS_TX_RING(ring)) { + if (cb->type & DESC_TYPE_PAGE && cb->pagecnt_bias) + __page_frag_cache_drain(cb->priv, cb->pagecnt_bias); + else if (cb->type & DESC_TYPE_FRAG) + page_pool_put_full_page(ring->page_pool, cb->priv, false); + } memset(cb, 0, sizeof(*cb)); }

@@ -3315,7 +3334,7 @@ static int hns3_alloc_and_map_buffer(struct hns3_enet_ring *ring, int ret;

ret = hns3_alloc_buffer(ring, cb); - if (ret) + if (ret || ring->page_pool) goto out;

ret = hns3_map_buffer(ring, cb); @@ -3337,7 +3356,8 @@ static int hns3_alloc_and_attach_buffer(struct hns3_enet_ring *ring, int i) if (ret) return ret;

- ring->desc[i].addr = cpu_to_le64(ring->desc_cb[i].dma); + ring->desc[i].addr = cpu_to_le64(ring->desc_cb[i].dma + + ring->desc_cb[i].page_offset);

return 0; } @@ -3367,7 +3387,8 @@ static void hns3_replace_buffer(struct hns3_enet_ring *ring, int i, { hns3_unmap_buffer(ring, &ring->desc_cb[i]); ring->desc_cb[i] = *res_cb; - ring->desc[i].addr = cpu_to_le64(ring->desc_cb[i].dma); + ring->desc[i].addr = cpu_to_le64(ring->desc_cb[i].dma + + ring->desc_cb[i].page_offset); ring->desc[i].rx.bd_base_info = 0; }

@@ -3539,6 +3560,12 @@ static void hns3_nic_reuse_page(struct sk_buff *skb, int i, u32 frag_size = size - pull_len; bool reused;

+ if (ring->page_pool) { + skb_add_rx_frag(skb, i, desc_cb->priv, frag_offset, + frag_size, truesize); + return; + } + /* Avoid re-using remote or pfmem page */ if (unlikely(!dev_page_is_reusable(desc_cb->priv))) goto out; @@ -3856,6 +3883,9 @@ static int hns3_alloc_skb(struct hns3_enet_ring *ring, unsigned int length, /* We can reuse buffer as-is, just make sure it is reusable */ if (dev_page_is_reusable(desc_cb->priv)) desc_cb->reuse_flag = 1; + else if (desc_cb->type & DESC_TYPE_FRAG) + page_pool_put_full_page(ring->page_pool, desc_cb->priv, + false); else /* This page cannot be reused so discard it */ __page_frag_cache_drain(desc_cb->priv, desc_cb->pagecnt_bias); @@ -3863,6 +3893,10 @@ static int hns3_alloc_skb(struct hns3_enet_ring *ring, unsigned int length, hns3_rx_ring_move_fw(ring); return 0; } + + if (ring->page_pool) + skb_mark_for_recycle(skb); + u64_stats_update_begin(&ring->syncp); ring->stats.seg_pkt_cnt++; u64_stats_update_end(&ring->syncp); @@ -3901,6 +3935,10 @@ static int hns3_add_frag(struct hns3_enet_ring *ring) "alloc rx fraglist skb fail\n"); return -ENXIO; } + + if (ring->page_pool) + skb_mark_for_recycle(new_skb); + ring->frag_num = 0;

if (ring->tail_skb) { @@ -4705,6 +4743,29 @@ static void hns3_put_ring_config(struct hns3_nic_priv *priv) priv->ring = NULL; }

+static void hns3_alloc_page_pool(struct hns3_enet_ring *ring) +{ + struct page_pool_params pp_params = { + .flags = PP_FLAG_DMA_MAP | PP_FLAG_PAGE_FRAG, + .order = hns3_page_order(ring), + .pool_size = ring->desc_num * hns3_buf_size(ring) / PAGE_SIZE, + .nid = dev_to_node(ring_to_dev(ring)), + .dev = ring_to_dev(ring), + .dma_dir = DMA_FROM_DEVICE, + .offset = 0, + .max_len = 0, + }; + + ring->page_pool = page_pool_create(&pp_params); + if (IS_ERR(ring->page_pool)) { + dev_warn(ring_to_dev(ring), "page pool creation failed: %ld\n", + PTR_ERR(ring->page_pool)); + ring->page_pool = NULL; + } else { + dev_info(ring_to_dev(ring), "page pool creation succeeded\n"); + } +} + static int hns3_alloc_ring_memory(struct hns3_enet_ring *ring) { int ret; @@ -4724,6 +4785,8 @@ static int hns3_alloc_ring_memory(struct hns3_enet_ring *ring) goto out_with_desc_cb;

if (!HNAE3_IS_TX_RING(ring)) { + hns3_alloc_page_pool(ring); + ret = hns3_alloc_ring_buffers(ring); if (ret) goto out_with_desc; @@ -4764,6 +4827,12 @@ void hns3_fini_ring(struct hns3_enet_ring *ring) devm_kfree(ring_to_dev(ring), tx_spare); ring->tx_spare = NULL; } + + if (!HNAE3_IS_TX_RING(ring) && ring->page_pool) { + page_pool_destroy(ring->page_pool); + ring->page_pool = NULL; + dev_info(ring_to_dev(ring), "page pool destroyed\n"); + } }

static int hns3_buf_size2type(u32 buf_size) diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h index 15af3d9..115c0ce 100644 --- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h +++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.h @@ -6,6 +6,7 @@

#include <linux/dim.h> #include <linux/if_vlan.h> +#include <net/page_pool.h>

#include "hnae3.h"

@@ -307,6 +308,7 @@ enum hns3_desc_type { DESC_TYPE_BOUNCE_ALL = 1 << 3, DESC_TYPE_BOUNCE_HEAD = 1 << 4, DESC_TYPE_SGL_SKB = 1 << 5, + DESC_TYPE_FRAG = 1 << 6, };

struct hns3_desc_cb { @@ -451,6 +453,7 @@ struct hns3_enet_ring { struct hnae3_queue *tqp; int queue_index; struct device *dev; /* will be used for DMA mapping of descriptors */ + struct page_pool *page_pool;

/* statistic */ struct ring_stats stats;