This patchset adds frag page support in page pool and enable skb's page frag recycling based on page pool in hns3 drvier.
V1: 1. avoid atomic_long_read() in case of freeing or draining page frag, and drop RFC tag.
RFC v6: 1. Disable frag page support in system 32-bit arch and 64-bit DMA.
RFC v5: 1. Rename dma_addr[0] to pp_frag_count and adjust codes according to the rename.
RFC v4: 1. Use the dma_addr[1] to store bias. 2. Default to a pagecnt_bias of PAGE_SIZE - 1. 3. other minor comment suggested by Alexander.
RFC v3: 1. Implement the semantic of "page recycling only wait for the page pool user instead of all user of a page" 2. Support the frag allocation of different sizes 3. Merge patch 4 & 5 to one patch as it does not make sense to use page_pool_dev_alloc_pages() API directly with elevated refcnt. 4. other minor comment suggested by Alexander.
RFC v2: 1. Split patch 1 to more reviewable one. 2. Repurpose the lower 12 bits of the dma address to store the pagecnt_bias as suggested by Alexander. 3. support recycling to pool->alloc for elevated refcnt case too.
Yunsheng Lin (4): page_pool: keep pp info as long as page pool owns the page page_pool: add interface to manipulate frag count in page pool page_pool: add frag page recycling support in page pool net: hns3: support skb's frag page recycling based on page pool
drivers/net/ethernet/hisilicon/Kconfig | 1 + drivers/net/ethernet/hisilicon/hns3/hns3_enet.c | 79 +++++++++++++++-- drivers/net/ethernet/hisilicon/hns3/hns3_enet.h | 3 + drivers/net/ethernet/marvell/mvneta.c | 6 +- drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c | 2 +- drivers/net/ethernet/ti/cpsw.c | 2 +- drivers/net/ethernet/ti/cpsw_new.c | 2 +- include/linux/mm_types.h | 18 ++-- include/linux/skbuff.h | 4 +- include/net/page_pool.h | 68 +++++++++++--- net/core/page_pool.c | 112 +++++++++++++++++++++++- 11 files changed, 258 insertions(+), 39 deletions(-)
Currently, page->pp is cleared and set everytime the page is recycled, which is unnecessary.
So only set the page->pp when the page is added to the page pool and only clear it when the page is released from the page pool.
This is also a preparation to support allocating frag page in page pool.
Reviewed-by: Ilias Apalodimas ilias.apalodimas@linaro.org Signed-off-by: Yunsheng Lin linyunsheng@huawei.com --- drivers/net/ethernet/marvell/mvneta.c | 6 +----- drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c | 2 +- drivers/net/ethernet/ti/cpsw.c | 2 +- drivers/net/ethernet/ti/cpsw_new.c | 2 +- include/linux/skbuff.h | 4 +--- include/net/page_pool.h | 7 ------- net/core/page_pool.c | 21 +++++++++++++++++---- 7 files changed, 22 insertions(+), 22 deletions(-)
diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c index 361bc4f..89bf31fd 100644 --- a/drivers/net/ethernet/marvell/mvneta.c +++ b/drivers/net/ethernet/marvell/mvneta.c @@ -2327,7 +2327,7 @@ mvneta_swbm_build_skb(struct mvneta_port *pp, struct page_pool *pool, if (!skb) return ERR_PTR(-ENOMEM);
- skb_mark_for_recycle(skb, virt_to_page(xdp->data), pool); + skb_mark_for_recycle(skb);
skb_reserve(skb, xdp->data - xdp->data_hard_start); skb_put(skb, xdp->data_end - xdp->data); @@ -2339,10 +2339,6 @@ mvneta_swbm_build_skb(struct mvneta_port *pp, struct page_pool *pool, skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, skb_frag_page(frag), skb_frag_off(frag), skb_frag_size(frag), PAGE_SIZE); - /* We don't need to reset pp_recycle here. It's already set, so - * just mark fragments for recycling. - */ - page_pool_store_mem_info(skb_frag_page(frag), pool); }
return skb; diff --git a/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c b/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c index 3229baf..320eddb 100644 --- a/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c +++ b/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c @@ -3995,7 +3995,7 @@ static int mvpp2_rx(struct mvpp2_port *port, struct napi_struct *napi, }
if (pp) - skb_mark_for_recycle(skb, page, pp); + skb_mark_for_recycle(skb); else dma_unmap_single_attrs(dev->dev.parent, dma_addr, bm_pool->buf_size, DMA_FROM_DEVICE, diff --git a/drivers/net/ethernet/ti/cpsw.c b/drivers/net/ethernet/ti/cpsw.c index cbbd0f6..9d59143 100644 --- a/drivers/net/ethernet/ti/cpsw.c +++ b/drivers/net/ethernet/ti/cpsw.c @@ -431,7 +431,7 @@ static void cpsw_rx_handler(void *token, int len, int status) skb->protocol = eth_type_trans(skb, ndev);
/* mark skb for recycling */ - skb_mark_for_recycle(skb, page, pool); + skb_mark_for_recycle(skb); netif_receive_skb(skb);
ndev->stats.rx_bytes += len; diff --git a/drivers/net/ethernet/ti/cpsw_new.c b/drivers/net/ethernet/ti/cpsw_new.c index 57d279f..a4234a3 100644 --- a/drivers/net/ethernet/ti/cpsw_new.c +++ b/drivers/net/ethernet/ti/cpsw_new.c @@ -374,7 +374,7 @@ static void cpsw_rx_handler(void *token, int len, int status) skb->protocol = eth_type_trans(skb, ndev);
/* mark skb for recycling */ - skb_mark_for_recycle(skb, page, pool); + skb_mark_for_recycle(skb); netif_receive_skb(skb);
ndev->stats.rx_bytes += len; diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h index b2db9cd..7795979 100644 --- a/include/linux/skbuff.h +++ b/include/linux/skbuff.h @@ -4711,11 +4711,9 @@ static inline u64 skb_get_kcov_handle(struct sk_buff *skb) }
#ifdef CONFIG_PAGE_POOL -static inline void skb_mark_for_recycle(struct sk_buff *skb, struct page *page, - struct page_pool *pp) +static inline void skb_mark_for_recycle(struct sk_buff *skb) { skb->pp_recycle = 1; - page_pool_store_mem_info(page, pp); } #endif
diff --git a/include/net/page_pool.h b/include/net/page_pool.h index 3dd62dd..8d7744d 100644 --- a/include/net/page_pool.h +++ b/include/net/page_pool.h @@ -253,11 +253,4 @@ static inline void page_pool_ring_unlock(struct page_pool *pool) spin_unlock_bh(&pool->ring.producer_lock); }
-/* Store mem_info on struct page and use it while recycling skb frags */ -static inline -void page_pool_store_mem_info(struct page *page, struct page_pool *pp) -{ - page->pp = pp; -} - #endif /* _NET_PAGE_POOL_H */ diff --git a/net/core/page_pool.c b/net/core/page_pool.c index 33b7dd7..bee33f6 100644 --- a/net/core/page_pool.c +++ b/net/core/page_pool.c @@ -206,6 +206,19 @@ static bool page_pool_dma_map(struct page_pool *pool, struct page *page) return true; }
+static void page_pool_set_pp_info(struct page_pool *pool, + struct page *page) +{ + page->pp = pool; + page->pp_magic |= PP_SIGNATURE; +} + +static void page_pool_clear_pp_info(struct page *page) +{ + page->pp_magic = 0; + page->pp = NULL; +} + static struct page *__page_pool_alloc_page_order(struct page_pool *pool, gfp_t gfp) { @@ -222,7 +235,7 @@ static struct page *__page_pool_alloc_page_order(struct page_pool *pool, return NULL; }
- page->pp_magic |= PP_SIGNATURE; + page_pool_set_pp_info(pool, page);
/* Track how many pages are held 'in-flight' */ pool->pages_state_hold_cnt++; @@ -266,7 +279,8 @@ static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool, put_page(page); continue; } - page->pp_magic |= PP_SIGNATURE; + + page_pool_set_pp_info(pool, page); pool->alloc.cache[pool->alloc.count++] = page; /* Track how many pages are held 'in-flight' */ pool->pages_state_hold_cnt++; @@ -345,7 +359,7 @@ void page_pool_release_page(struct page_pool *pool, struct page *page) DMA_ATTR_SKIP_CPU_SYNC); page_pool_set_dma_addr(page, 0); skip_dma_unmap: - page->pp_magic = 0; + page_pool_clear_pp_info(page);
/* This may be the last page returned, releasing the pool, so * it is not safe to reference pool afterwards. @@ -644,7 +658,6 @@ bool page_pool_return_skb_page(struct page *page) * The page will be returned to the pool here regardless of the * 'flipped' fragment being in use or not. */ - page->pp = NULL; page_pool_put_full_page(pp, page, false);
return true;
For 32 bit systems with 64 bit dma, dma_addr[1] is used to store the upper 32 bit dma addr, those system should be rare those days.
For normal system, the dma_addr[1] in 'struct page' is not used, so we can reuse dma_addr[1] for storing frag count, which means how many frags this page might be splited to.
In order to simplify the page frag support in the page pool, the PAGE_POOL_DMA_USE_PP_FRAG_COUNT macro is added to indicate the 32 bit systems with 64 bit dma, and the page frag support in page pool is disabled for such system.
The newly added page_pool_set_frag_count() is called to reserve the maximum frag count before any page frag is passed to the user. The page_pool_atomic_sub_frag_count_return() is called when user is done with the page frag.
Signed-off-by: Yunsheng Lin linyunsheng@huawei.com --- include/linux/mm_types.h | 18 +++++++++++++----- include/net/page_pool.h | 46 +++++++++++++++++++++++++++++++++++++++------- net/core/page_pool.c | 4 ++++ 3 files changed, 56 insertions(+), 12 deletions(-)
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h index d33d97c..3e339dd 100644 --- a/include/linux/mm_types.h +++ b/include/linux/mm_types.h @@ -103,11 +103,19 @@ struct page { unsigned long pp_magic; struct page_pool *pp; unsigned long _pp_mapping_pad; - /** - * @dma_addr: might require a 64-bit value on - * 32-bit architectures. - */ - unsigned long dma_addr[2]; + unsigned long dma_addr; + union { + /** + * dma_addr_upper: might require a 64-bit + * value on 32-bit architectures. + */ + unsigned long dma_addr_upper; + /** + * For frag page support, not supported in + * 32-bit architectures with 64-bit DMA. + */ + atomic_long_t pp_frag_count; + }; }; struct { /* slab, slob and slub */ union { diff --git a/include/net/page_pool.h b/include/net/page_pool.h index 8d7744d..42e6997 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 @@ -198,19 +201,48 @@ static inline void page_pool_recycle_direct(struct page_pool *pool, page_pool_put_full_page(pool, page, true); }
+#define PAGE_POOL_DMA_USE_PP_FRAG_COUNT \ + (sizeof(dma_addr_t) > sizeof(unsigned long)) + static inline dma_addr_t page_pool_get_dma_addr(struct page *page) { - dma_addr_t ret = page->dma_addr[0]; - if (sizeof(dma_addr_t) > sizeof(unsigned long)) - ret |= (dma_addr_t)page->dma_addr[1] << 16 << 16; + dma_addr_t ret = page->dma_addr; + + if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT) + ret |= (dma_addr_t)page->dma_addr_upper << 16 << 16; + return ret; }
static inline void page_pool_set_dma_addr(struct page *page, dma_addr_t addr) { - page->dma_addr[0] = addr; - if (sizeof(dma_addr_t) > sizeof(unsigned long)) - page->dma_addr[1] = upper_32_bits(addr); + page->dma_addr = addr; + if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT) + page->dma_addr_upper = upper_32_bits(addr); +} + +static inline void page_pool_set_frag_count(struct page *page, long nr) +{ + atomic_long_set(&page->pp_frag_count, nr); +} + +static inline long page_pool_atomic_sub_frag_count_return(struct page *page, + long nr) +{ + long ret; + + /* As suggested by Alexander, atomic_long_read() may cover up the + * reference count errors, so avoid calling atomic_long_read() in + * the cases of freeing or draining the page_frags, where we would + * not expect it to match or that are slowpath anyway. + */ + if (__builtin_constant_p(nr) && + atomic_long_read(&page->pp_frag_count) == nr) + return 0; + + ret = atomic_long_sub_return(nr, &page->pp_frag_count); + WARN_ON(ret < 0); + return ret; }
static inline bool is_page_pool_compiled_in(void) diff --git a/net/core/page_pool.c b/net/core/page_pool.c index bee33f6..d40c11c 100644 --- a/net/core/page_pool.c +++ b/net/core/page_pool.c @@ -67,6 +67,10 @@ static int page_pool_init(struct page_pool *pool, */ }
+ if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT && + pool->p.flags & PP_FLAG_PAGE_FRAG) + return -EINVAL; + if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0) return -ENOMEM;
Currently page pool only support page recycling 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 | 15 +++++++++ net/core/page_pool.c | 87 +++++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 102 insertions(+)
diff --git a/include/net/page_pool.h b/include/net/page_pool.h index 42e6997..a408240 100644 --- a/include/net/page_pool.h +++ b/include/net/page_pool.h @@ -91,6 +91,9 @@ struct page_pool { unsigned long defer_warn;
u32 pages_state_hold_cnt; + unsigned int frag_offset; + struct page *frag_page; + long frag_users;
/* * Data structure for allocation side @@ -140,6 +143,18 @@ 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 d40c11c..7f82465 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 LONG_MAX + static int page_pool_init(struct page_pool *pool, const struct page_pool_params *params) { @@ -423,6 +425,11 @@ 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_sub_frag_count_return(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. @@ -515,6 +522,84 @@ void page_pool_put_page_bulk(struct page_pool *pool, void **data, } EXPORT_SYMBOL(page_pool_put_page_bulk);
+static struct page *page_pool_drain_frag(struct page_pool *pool, + struct page *page) +{ + long drain_count = BIAS_MAX - pool->frag_users; + + /* Some user is still using the page frag */ + if (likely(page_pool_atomic_sub_frag_count_return(page, + drain_count))) + return NULL; + + if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) { + if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) + page_pool_dma_sync_for_device(pool, page, -1); + + return page; + } + + page_pool_return_page(pool, page); + return NULL; +} + +static void page_pool_free_frag(struct page_pool *pool) +{ + long drain_count = BIAS_MAX - pool->frag_users; + struct page *page = pool->frag_page; + + pool->frag_page = NULL; + + if (!page || + page_pool_atomic_sub_frag_count_return(page, drain_count)) + return; + + page_pool_return_page(pool, page); +} + +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; + struct page *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()); + *offset = pool->frag_offset; + + if (page && *offset + size > max_size) { + page = page_pool_drain_frag(pool, page); + if (page) + goto frag_reset; + } + + if (!page) { + page = page_pool_alloc_pages(pool, gfp); + if (unlikely(!page)) { + pool->frag_page = NULL; + return NULL; + } + + pool->frag_page = page; + +frag_reset: + pool->frag_users = 1; + *offset = 0; + pool->frag_offset = size; + page_pool_set_frag_count(page, BIAS_MAX); + return page; + } + + pool->frag_users++; + pool->frag_offset = *offset + size; + return page; +} +EXPORT_SYMBOL(page_pool_alloc_frag); + static void page_pool_empty_ring(struct page_pool *pool) { struct page *page; @@ -620,6 +705,8 @@ 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 frag page support in page pool.
The performance improves above 10~20% for single thread iperf TCP flow with IOMMU disabled when iperf server and irq/NAPI have a different CPU.
The performance improves about 135%(14Gbit to 33Gbit) for single thread iperf TCP flow IOMMU is in strict mode and iperf server shares the same cpu with irq/NAPI.
Signed-off-by: Yunsheng Lin linyunsheng@huawei.com --- drivers/net/ethernet/hisilicon/Kconfig | 1 + drivers/net/ethernet/hisilicon/hns3/hns3_enet.c | 79 +++++++++++++++++++++++-- drivers/net/ethernet/hisilicon/hns3/hns3_enet.h | 3 + 3 files changed, 78 insertions(+), 5 deletions(-)
diff --git a/drivers/net/ethernet/hisilicon/Kconfig b/drivers/net/ethernet/hisilicon/Kconfig index bb062b0..4e114f4 100644 --- a/drivers/net/ethernet/hisilicon/Kconfig +++ b/drivers/net/ethernet/hisilicon/Kconfig @@ -90,6 +90,7 @@ config HNS_ENET config HNS3 tristate "Hisilicon Network Subsystem Support HNS3 (Framework)" depends on PCI + select PAGE_POOL help This selects the framework support for Hisilicon Network Subsystem 3. This layer facilitates clients like ENET, RoCE and user-space ethernet diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c index cdb5f14..90597fe 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_PP_FRAG; + cb->reuse_flag = 0; + return 0; + } + p = dev_alloc_pages(order); if (!p) return -ENOMEM; @@ -3227,8 +3242,13 @@ 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_PP_FRAG) + page_pool_put_full_page(ring->page_pool, cb->priv, + false); + } memset(cb, 0, sizeof(*cb)); }
@@ -3315,7 +3335,7 @@ 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 +3357,8 @@ 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 +3388,8 @@ 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 +3561,12 @@ 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 +3884,9 @@ 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_PP_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 +3894,10 @@ 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 +3936,10 @@ 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 +4744,29 @@ 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 | + PP_FLAG_DMA_SYNC_DEV, + .order = hns3_page_order(ring), + .pool_size = ring->desc_num * hns3_buf_size(ring) / + (PAGE_SIZE << hns3_page_order(ring)), + .nid = dev_to_node(ring_to_dev(ring)), + .dev = ring_to_dev(ring), + .dma_dir = DMA_FROM_DEVICE, + .offset = 0, + .max_len = PAGE_SIZE << hns3_page_order(ring), + }; + + 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; + } +} + static int hns3_alloc_ring_memory(struct hns3_enet_ring *ring) { int ret; @@ -4724,6 +4786,8 @@ 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 +4828,11 @@ 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; + } }
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..27809d6 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_PP_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;
On Thu, 5 Aug 2021 19:05:26 +0800 Yunsheng Lin wrote:
This patch adds skb's frag page recycling support based on the frag page support in page pool.
The performance improves above 10~20% for single thread iperf TCP flow with IOMMU disabled when iperf server and irq/NAPI have a different CPU.
The performance improves about 135%(14Gbit to 33Gbit) for single thread iperf TCP flow IOMMU is in strict mode and iperf server shares the same cpu with irq/NAPI.
Signed-off-by: Yunsheng Lin linyunsheng@huawei.com
This patch does not apply cleanly to net-next, please rebase if you're targeting that tree.
On 2021/8/6 9:20, Jakub Kicinski wrote:
On Thu, 5 Aug 2021 19:05:26 +0800 Yunsheng Lin wrote:
This patch adds skb's frag page recycling support based on the frag page support in page pool.
The performance improves above 10~20% for single thread iperf TCP flow with IOMMU disabled when iperf server and irq/NAPI have a different CPU.
The performance improves about 135%(14Gbit to 33Gbit) for single thread iperf TCP flow IOMMU is in strict mode and iperf server shares the same cpu with irq/NAPI.
Signed-off-by: Yunsheng Lin linyunsheng@huawei.com
This patch does not apply cleanly to net-next, please rebase if you're targeting that tree.
It seems I forgot to rebase the net-next tree before doing "git format-patch", thanks for mentioning that.
.
-
Jakub Kicinski -
Yunsheng Lin