net: mana: Add page pool for RX buffers
Subsystems (6)
Haiyang Zhang
maintainerPATCH v6 1/1 net: mana: Add page pool for RX buffers
Add page pool for RX buffers for faster buffer cycle and reduce CPU usage.
The standard page pool API is used.
With iperf and 128 threads test, this patch improved the throughput by 12-15%, and decreased the IRQ associated CPU's usage from 99-100% to 10-50%.
---
V6:
Added perf info as suggested by Jesper Dangaard Brouer
V5:
In err path, set page_pool_put_full_page(..., false) as suggested by Jakub Kicinski
V4:
Add nid setting, remove page_pool_nid_changed(), as suggested by Jesper Dangaard Brouer
V3:
Update xdp mem model, pool param, alloc as suggested by Jakub Kicinski V2:
Use the standard page pool API as suggested by Jesper Dangaard Brouer
--- drivers/net/ethernet/microsoft/mana/mana_en.c | 90 +++++++++++++++---- include/net/mana/mana.h | 3 + 2 files changed, 77 insertions(+), 16 deletions(-)
drivers/net/ethernet/microsoft/mana/mana_en.c +74 -16
@@ -1414,8 +1414,8 @@ static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va,return skb; } -static void mana_rx_skb (void *buf_va,struct mana_rxcomp_oob *cqe, -struct mana_rxq *rxq) +static void mana_rx_skb (void *buf_va,bool from_pool, +struct mana_rxcomp_oob *cqe,struct mana_rxq *rxq) {struct mana_stats_rx *rx_stats = &rxq->stats;struct net_device *ndev = rxq->ndev;
@@ -1448,6 +1448,9 @@ static void mana_rx_skb(void *buf_va, struct mana_rxcomp_oob *cqe,if (!skb) goto drop; +if (from_pool) +skb_mark_for_recycle (skb); + skb->dev = napi->dev; skb->protocol =eth_type_trans (skb, ndev);
@@ -1498,9 +1501,14 @@ static void mana_rx_skb(void *buf_va, struct mana_rxcomp_oob *cqe,u64_stats_update_end (&rx_stats->syncp); drop: -WARN_ON_ONCE (rxq->xdp_save_va); -/* Save for reuse */ - rxq->xdp_save_va = buf_va; +if (from_pool) { +page_pool_recycle_direct (rxq->page_pool, +virt_to_head_page (buf_va)); + }else { +WARN_ON_ONCE (rxq->xdp_save_va); +/* Save for reuse */ + rxq->xdp_save_va = buf_va; + } ++ndev->stats.rx_dropped;
@@ -1508,11 +1516,13 @@ static void mana_rx_skb(void *buf_va, struct mana_rxcomp_oob *cqe, }static void *mana_get_rxfrag (struct mana_rxq *rxq,struct device *dev, -dma_addr_t *da,bool is_napi) +dma_addr_t *da,bool *from_pool,bool is_napi) {struct page *page;void *va; + *from_pool = false; +/* Reuse XDP dropped page if available */ if (rxq->xdp_save_va) { va = rxq->xdp_save_va;
@@ -1533,17 +1543,22 @@ static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,return NULL ; } }else { - page =dev_alloc_page (); + page =page_pool_dev_alloc_pages (rxq->page_pool);if (!page)return NULL ; + *from_pool = true; va =page_to_virt (page); } *da =dma_map_single (dev, va + rxq->headroom, rxq->datasize,DMA_FROM_DEVICE );if (dma_mapping_error (dev, *da)) { -put_page (virt_to_head_page (va)); +if (*from_pool) +page_pool_put_full_page (rxq->page_pool, page, false); +else +put_page (virt_to_head_page (va)); +return NULL ; }
@@ -1552,21 +1567,25 @@ static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,/* Allocate frag for rx buffer, and save the old buf */ static void mana_refill_rx_oob (struct device *dev,struct mana_rxq *rxq, -struct mana_recv_buf_oob *rxoob,void **old_buf) +struct mana_recv_buf_oob *rxoob,void **old_buf, +bool *old_fp) { +bool from_pool;dma_addr_t da;void *va; - va =mana_get_rxfrag (rxq, dev, &da, true); + va =mana_get_rxfrag (rxq, dev, &da, &from_pool, true);if (!va)return ;dma_unmap_single (dev, rxoob->sgl[0 ].address, rxq->datasize,DMA_FROM_DEVICE ); *old_buf = rxoob->buf_va; + *old_fp = rxoob->from_pool; rxoob->buf_va = va; rxoob->sgl[0 ].address = da; + rxoob->from_pool = from_pool; } static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,
@@ -1580,6 +1599,7 @@ static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,struct device *dev = gc->dev;void *old_buf =NULL ;u32 curr, pktlen; +bool old_fp; apc =netdev_priv (ndev);
@@ -1622,12 +1642,12 @@ static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq, rxbuf_oob = &rxq->rx_oobs[curr];WARN_ON_ONCE (rxbuf_oob->wqe_inf.wqe_size_in_bu !=1 ); -mana_refill_rx_oob (dev, rxq, rxbuf_oob, &old_buf); +mana_refill_rx_oob (dev, rxq, rxbuf_oob, &old_buf, &old_fp);/* Unsuccessful refill will have old_buf == NULL. * In this case, mana_rx_skb() will drop the packet. */ -mana_rx_skb (old_buf, oob, rxq); +mana_rx_skb (old_buf, old_fp, oob, rxq); drop:mana_move_wq_tail (rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu);
@@ -1887,6 +1907,7 @@ static void mana_destroy_rxq(struct mana_port_context *apc,struct mana_recv_buf_oob *rx_oob;struct device *dev = gc->dev;struct napi_struct *napi; +struct page *page;int i;if (!rxq)
@@ -1919,10 +1940,18 @@ static void mana_destroy_rxq(struct mana_port_context *apc,dma_unmap_single (dev, rx_oob->sgl[0 ].address, rx_oob->sgl[0 ].size,DMA_FROM_DEVICE ); -put_page (virt_to_head_page (rx_oob->buf_va)); + page =virt_to_head_page (rx_oob->buf_va); + +if (rx_oob->from_pool) +page_pool_put_full_page (rxq->page_pool, page, false); +else +put_page (page); + rx_oob->buf_va =NULL ; } +page_pool_destroy (rxq->page_pool); +if (rxq->gdma_rq)mana_gd_destroy_queue (gc, rxq->gdma_rq);
@@ -1933,18 +1962,20 @@ static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key, struct mana_rxq *rxq, struct device *dev) { struct mana_port_context *mpc = netdev_priv(rxq->ndev); +bool from_pool = false;dma_addr_t da;void *va;if (mpc->rxbufs_pre) va =mana_get_rxbuf_pre (rxq, &da);else - va =mana_get_rxfrag (rxq, dev, &da, false); + va =mana_get_rxfrag (rxq, dev, &da, &from_pool, false);if (!va)return -ENOMEM ; rx_oob->buf_va = va; + rx_oob->from_pool = from_pool; rx_oob->sgl[0 ].address = da; rx_oob->sgl[0 ].size = rxq->datasize;
@@ -2014,6 +2045,26 @@ static int mana_push_wqe(struct mana_rxq *rxq)return 0 ; } +static int mana_create_page_pool (struct mana_rxq *rxq,struct gdma_context *gc) +{ +struct page_pool_params pprm = {}; +int ret; + + pprm.pool_size =RX_BUFFERS_PER_QUEUE ; + pprm.nid = gc->numa_node; + pprm.napi = &rxq->rx_cq.napi; + + rxq->page_pool =page_pool_create (&pprm); + +if (IS_ERR (rxq->page_pool)) { + ret =PTR_ERR (rxq->page_pool); + rxq->page_pool =NULL ; +return ret; + } + +return 0 ; +} +static struct mana_rxq *mana_create_rxq (struct mana_port_context *apc,u32 rxq_idx,struct mana_eq *eq,struct net_device *ndev)
@@ -2043,6 +2094,13 @@ static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,mana_get_rxbuf_cfg (ndev->mtu, &rxq->datasize, &rxq->alloc_size, &rxq->headroom); +/* Create page pool for RX queue */ + err =mana_create_page_pool (rxq, gc); +if (err) { +netdev_err (ndev,"Create page pool err:%d\n" , err); + goto out; + } + err =mana_alloc_rx_wqe (apc, rxq, &rq_size, &cq_size);if (err) goto out;
@@ -2114,8 +2172,8 @@ static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,WARN_ON (xdp_rxq_info_reg (&rxq->xdp_rxq, ndev, rxq_idx, cq->napi.napi_id)); -WARN_ON (xdp_rxq_info_reg_mem_model (&rxq->xdp_rxq, -MEM_TYPE_PAGE_SHARED ,NULL )); +WARN_ON (xdp_rxq_info_reg_mem_model (&rxq->xdp_rxq,MEM_TYPE_PAGE_POOL , + rxq->page_pool));napi_enable (&cq->napi);
include/net/mana/mana.h +3 -0
@@ -280,6 +280,7 @@ struct mana_recv_buf_oob {struct gdma_wqe_request wqe_req;void *buf_va; +bool from_pool;/* allocated from a page pool */ /* SGL of the buffer going to be sent has part of the work request. */ u32 num_sge;
@@ -330,6 +331,8 @@ struct mana_rxq {bool xdp_flush;int xdp_rc;/* XDP redirect return code */ +struct page_pool *page_pool; +/* MUST BE THE LAST MEMBER: * Each receive buffer has an associated mana_recv_buf_oob. */
Signature · 1 line
2.25.1
Quoted 8 lines · Haiyang Zhang
Add page pool for RX buffers for faster buffer cycle and reduce CPU usage.
The standard page pool API is used.
With iperf and 128 threads test, this patch improved the throughput by 12-15%, and decreased the IRQ associated CPU's usage from 99-100% to 10-50%.
https://www.spinics.net/lists/netdev/msg584734.html
The performance of throughput and cpu utility is very good. I have a similar patch series with this in the above link. And David Miller had the following comments:
"
The system is supposed to hold onto enough atomic memory to absorb all
reasonable situations like this.
If anything a solution to this problem belongs generically somewhere, not in a driver. And furthermore looping over an allocation attempt with a delay is strongly discouraged.
"
Hope your commit can be merged into upstream (pray).
Zhu Yanjun
Quoted 278 lines
Signed-off-by Haiyang Zhang <[email protected]>Reviewed-by Jesse Brandeburg <[email protected]>---
V6:
Added perf info as suggested by Jesper Dangaard Brouer
V5:
In err path, setpage_pool_put_full_page(..., false)as suggested by Jakub Kicinski
V4:
Add nid setting, removepage_pool_nid_changed(), as suggested by Jesper Dangaard Brouer
V3:
Update xdp mem model, pool param, alloc as suggested by Jakub Kicinski V2:
Use the standard page pool API as suggested by Jesper Dangaard Brouer--- drivers/net/ethernet/microsoft/mana/mana_en.c | 90 +++++++++++++++---- include/net/mana/mana.h | 3 + 2 files changed, 77 insertions(+), 16 deletions(-)drivers/net/ethernet/microsoft/mana/mana_en.c +74
-16@@ -1414,8 +1414,8 @@ static struct sk_buff *mana_build_skb(struct mana_rxq *rxq, void *buf_va,return skb; } -static void mana_rx_skb (void *buf_va,struct mana_rxcomp_oob *cqe, -struct mana_rxq *rxq) +static void mana_rx_skb (void *buf_va,bool from_pool, +struct mana_rxcomp_oob *cqe,struct mana_rxq *rxq) {struct mana_stats_rx *rx_stats = &rxq->stats;struct net_device *ndev = rxq->ndev;@@ -1448,6 +1448,9 @@ static void mana_rx_skb(void *buf_va, struct mana_rxcomp_oob *cqe,if (!skb) goto drop; +if (from_pool) +skb_mark_for_recycle (skb); + skb->dev = napi->dev; skb->protocol =eth_type_trans (skb, ndev);@@ -1498,9 +1501,14 @@ static void mana_rx_skb(void *buf_va, struct mana_rxcomp_oob *cqe,u64_stats_update_end (&rx_stats->syncp); drop: -WARN_ON_ONCE (rxq->xdp_save_va); -/* Save for reuse */ - rxq->xdp_save_va = buf_va; +if (from_pool) { +page_pool_recycle_direct (rxq->page_pool, +virt_to_head_page (buf_va)); + }else { +WARN_ON_ONCE (rxq->xdp_save_va); +/* Save for reuse */ + rxq->xdp_save_va = buf_va; + } ++ndev->stats.rx_dropped;@@ -1508,11 +1516,13 @@ static void mana_rx_skb(void *buf_va, struct mana_rxcomp_oob *cqe, }static void *mana_get_rxfrag (struct mana_rxq *rxq,struct device *dev, -dma_addr_t *da,bool is_napi) +dma_addr_t *da,bool *from_pool,bool is_napi) {struct page *page;void *va; + *from_pool = false; +/* Reuse XDP dropped page if available */ if (rxq->xdp_save_va) { va = rxq->xdp_save_va;@@ -1533,17 +1543,22 @@ static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,return NULL ; } }else { - page =dev_alloc_page (); + page =page_pool_dev_alloc_pages (rxq->page_pool);if (!page)return NULL ; + *from_pool = true; va =page_to_virt (page); } *da =dma_map_single (dev, va + rxq->headroom, rxq->datasize,DMA_FROM_DEVICE );if (dma_mapping_error (dev, *da)) { -put_page (virt_to_head_page (va)); +if (*from_pool) +page_pool_put_full_page (rxq->page_pool, page, false); +else +put_page (virt_to_head_page (va)); +return NULL ; }@@ -1552,21 +1567,25 @@ static void *mana_get_rxfrag(struct mana_rxq *rxq, struct device *dev,/* Allocate frag for rx buffer, and save the old buf */ static void mana_refill_rx_oob (struct device *dev,struct mana_rxq *rxq, -struct mana_recv_buf_oob *rxoob,void **old_buf) +struct mana_recv_buf_oob *rxoob,void **old_buf, +bool *old_fp) { +bool from_pool;dma_addr_t da;void *va; - va =mana_get_rxfrag (rxq, dev, &da, true); + va =mana_get_rxfrag (rxq, dev, &da, &from_pool, true);if (!va)return ;dma_unmap_single (dev, rxoob->sgl[0 ].address, rxq->datasize,DMA_FROM_DEVICE ); *old_buf = rxoob->buf_va; + *old_fp = rxoob->from_pool; rxoob->buf_va = va; rxoob->sgl[0 ].address = da; + rxoob->from_pool = from_pool; } static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,@@ -1580,6 +1599,7 @@ static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,struct device *dev = gc->dev;void *old_buf =NULL ;u32 curr, pktlen; +bool old_fp; apc =netdev_priv (ndev);@@ -1622,12 +1642,12 @@ static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq, rxbuf_oob = &rxq->rx_oobs[curr];WARN_ON_ONCE (rxbuf_oob->wqe_inf.wqe_size_in_bu !=1 ); -mana_refill_rx_oob (dev, rxq, rxbuf_oob, &old_buf); +mana_refill_rx_oob (dev, rxq, rxbuf_oob, &old_buf, &old_fp);/* Unsuccessful refill will have old_buf == NULL. * In this case, mana_rx_skb() will drop the packet. */ -mana_rx_skb (old_buf, oob, rxq); +mana_rx_skb (old_buf, old_fp, oob, rxq); drop:mana_move_wq_tail (rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu);@@ -1887,6 +1907,7 @@ static void mana_destroy_rxq(struct mana_port_context *apc,struct mana_recv_buf_oob *rx_oob;struct device *dev = gc->dev;struct napi_struct *napi; +struct page *page;int i;if (!rxq)@@ -1919,10 +1940,18 @@ static void mana_destroy_rxq(struct mana_port_context *apc,dma_unmap_single (dev, rx_oob->sgl[0 ].address, rx_oob->sgl[0 ].size,DMA_FROM_DEVICE ); -put_page (virt_to_head_page (rx_oob->buf_va)); + page =virt_to_head_page (rx_oob->buf_va); + +if (rx_oob->from_pool) +page_pool_put_full_page (rxq->page_pool, page, false); +else +put_page (page); + rx_oob->buf_va =NULL ; } +page_pool_destroy (rxq->page_pool); +if (rxq->gdma_rq)mana_gd_destroy_queue (gc, rxq->gdma_rq);@@ -1933,18 +1962,20 @@ static int mana_fill_rx_oob(struct mana_recv_buf_oob *rx_oob, u32 mem_key, struct mana_rxq *rxq, struct device *dev) { struct mana_port_context *mpc = netdev_priv(rxq->ndev); +bool from_pool = false;dma_addr_t da;void *va;if (mpc->rxbufs_pre) va =mana_get_rxbuf_pre (rxq, &da);else - va =mana_get_rxfrag (rxq, dev, &da, false); + va =mana_get_rxfrag (rxq, dev, &da, &from_pool, false);if (!va)return -ENOMEM ; rx_oob->buf_va = va; + rx_oob->from_pool = from_pool; rx_oob->sgl[0 ].address = da; rx_oob->sgl[0 ].size = rxq->datasize;@@ -2014,6 +2045,26 @@ static int mana_push_wqe(struct mana_rxq *rxq)return 0 ; } +static int mana_create_page_pool (struct mana_rxq *rxq,struct gdma_context *gc) +{ +struct page_pool_params pprm = {}; +int ret; + + pprm.pool_size =RX_BUFFERS_PER_QUEUE ; + pprm.nid = gc->numa_node; + pprm.napi = &rxq->rx_cq.napi; + + rxq->page_pool =page_pool_create (&pprm); + +if (IS_ERR (rxq->page_pool)) { + ret =PTR_ERR (rxq->page_pool); + rxq->page_pool =NULL ; +return ret; + } + +return 0 ; +} +static struct mana_rxq *mana_create_rxq (struct mana_port_context *apc,u32 rxq_idx,struct mana_eq *eq,struct net_device *ndev)@@ -2043,6 +2094,13 @@ static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,mana_get_rxbuf_cfg (ndev->mtu, &rxq->datasize, &rxq->alloc_size, &rxq->headroom); +/* Create page pool for RX queue */ + err =mana_create_page_pool (rxq, gc); +if (err) { +netdev_err (ndev,"Create page pool err:%d\n" , err); + goto out; + } + err =mana_alloc_rx_wqe (apc, rxq, &rq_size, &cq_size);if (err) goto out;@@ -2114,8 +2172,8 @@ static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc,WARN_ON (xdp_rxq_info_reg (&rxq->xdp_rxq, ndev, rxq_idx, cq->napi.napi_id)); -WARN_ON (xdp_rxq_info_reg_mem_model (&rxq->xdp_rxq, -MEM_TYPE_PAGE_SHARED ,NULL )); +WARN_ON (xdp_rxq_info_reg_mem_model (&rxq->xdp_rxq,MEM_TYPE_PAGE_POOL , + rxq->page_pool));napi_enable (&cq->napi);include/net/mana/mana.h +3
-0@@ -280,6 +280,7 @@ struct mana_recv_buf_oob {struct gdma_wqe_request wqe_req;void *buf_va; +bool from_pool;/* allocated from a page pool */ /* SGL of the buffer going to be sent has part of the work request. */ u32 num_sge;@@ -330,6 +331,8 @@ struct mana_rxq {bool xdp_flush;int xdp_rc;/* XDP redirect return code */ +struct page_pool *page_pool; +/* MUST BE THE LAST MEMBER: * Each receive buffer has an associated mana_recv_buf_oob. */
This patch was applied to netdev/net-next.git (main) by David S. Miller <[email protected]>:
Hello:
This patch was applied to netdev/net-next.git (main) by David S. Miller <[email protected]>:
Quoted 10 lines · On Fri, 4 Aug 2023 13:33:53 -0700 you wrote:
Add page pool for RX buffers for faster buffer cycle and reduce CPU usage.
The standard page pool API is used.
With iperf and 128 threads test, this patch improved the throughput by 12-15%, and decreased the IRQ associated CPU's usage from 99-100% to 10-50%.
[...]
Here is the summary with links:
- [V6,net-next] net: mana: Add page pool for RX buffers
https://git.kernel.org/netdev/net-next/c/b1d13f7a3b53
You are awesome, thank you!
Signature · 2 lines
Deet-doot-dot, I am a bot. https://korg.docs.kernel.org/patchwork/pwbot.html
Quoted 11 lines · Haiyang Zhang
Add page pool for RX buffers for faster buffer cycle and reduce CPU usage.
The standard page pool API is used.
With iperf and 128 threads test, this patch improved the throughput by 12-15%, and decreased the IRQ associated CPU's usage from 99-100% to 10-50%.
Signed-off-by Haiyang Zhang<[email protected]>Reviewed-by Jesse Brandeburg<[email protected]>
For the record I want to provide my ACK as page_pool maintainer:
As patch was applied Sunday, my ACK will not reach the git tree
https://git.kernel.org/netdev/net-next/c/b1d13f7a3b53
--Jesper