octeontx2-pf: Add support for page pool
Ratheesh Kannoth
maintainerPATCH v3 1/1 octeontx2-pf: Add support for page pool
Page pool for each rx queue enhance rx side performance by reclaiming buffers back to each queue specific pool. DMA mapping is done only for first allocation of buffers. As subsequent buffers allocation avoid DMA mapping, it results in performance improvement.
Image | Performance
------------ | ------------
Vannila | 3Mpps
|
with this | 42Mppschange |
---------------------------
---
ChangeLog
v2 -> v3:
* Modified commit message
* Fixed nit comments.
v1 -> v2:
* Removed GFP_DMA flag
* Returned correct err value
v0 -> v1:
* Removed CONFIG_PAGE_POOL #ifdefs in code
* Used compound page APIs
* Replaced page_pool_put_page API with page_pool_put_full_page API
--- .../net/ethernet/marvell/octeontx2/Kconfig | 1 + .../marvell/octeontx2/nic/otx2_common.c | 74 ++++++++++++++++--- .../marvell/octeontx2/nic/otx2_common.h | 6 +- .../ethernet/marvell/octeontx2/nic/otx2_pf.c | 11 ++- .../marvell/octeontx2/nic/otx2_txrx.c | 19 +++-- .../marvell/octeontx2/nic/otx2_txrx.h | 1 + .../ethernet/marvell/octeontx2/nic/qos_sq.c | 2 +- 7 files changed, 93 insertions(+), 21 deletions(-)
drivers/net/ethernet/marvell/octeontx2/Kconfig +1 -0
@@ -32,6 +32,7 @@ config OCTEONTX2_PF tristate "Marvell OcteonTX2 NIC Physical Function driver" select OCTEONTX2_MBOX select NET_DEVLINK + select PAGE_POOL depends on (64BIT && COMPILE_TEST) || ARM64 select DIMLIB depends on PCI
drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c +65 -9
135 lines not on this page · raw message
drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h +4 -2
@@ -976,7 +976,7 @@ int otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,int otx2_rxtx_enable (struct otx2_nic *pfvf,bool enable);void otx2_ctx_disable (struct mbox *mbox,int type,bool npa);int otx2_nix_config_bp (struct otx2_nic *pfvf,bool enable); -void otx2_cleanup_rx_cqes (struct otx2_nic *pfvf,struct otx2_cq_queue *cq); +void otx2_cleanup_rx_cqes (struct otx2_nic *pfvf,struct otx2_cq_queue *cq,int qidx);void otx2_cleanup_tx_cqes (struct otx2_nic *pfvf,struct otx2_cq_queue *cq);int otx2_sq_init (struct otx2_nic *pfvf,u16 qidx,u16 sqb_aura);int otx2_sq_aq_init (void *dev,u16 qidx,u16 sqb_aura);
@@ -984,7 +984,7 @@ int cn10k_sq_aq_init(void *dev, u16 qidx, u16 sqb_aura);int otx2_alloc_buffer (struct otx2_nic *pfvf,struct otx2_cq_queue *cq,dma_addr_t *dma);int otx2_pool_init (struct otx2_nic *pfvf,u16 pool_id, -int stack_pages,int numptrs,int buf_size); +int stack_pages,int numptrs,int buf_size,int type);int otx2_aura_init (struct otx2_nic *pfvf,int aura_id,int pool_id,int numptrs);
@@ -1054,6 +1054,8 @@ u16 otx2_get_max_mtu(struct otx2_nic *pfvf);int otx2_handle_ntuple_tc_features (struct net_device *netdev,netdev_features_t features);int otx2_smq_flush (struct otx2_nic *pfvf,int smq); +void otx2_free_bufs (struct otx2_nic *pfvf,struct otx2_pool *pool, +u64 iova,int size);/* tc support */ int otx2_init_tc (struct otx2_nic *nic);
drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c +10 -1
@@ -1555,7 +1555,9 @@ static void otx2_free_hw_resources(struct otx2_nic *pf)struct nix_lf_free_req *free_req;struct mbox *mbox = &pf->mbox;struct otx2_cq_queue *cq; +struct otx2_pool *pool;struct msg_req *req; +int pool_id;int qidx;/* Ensure all SQE are processed */
@@ -1584,7 +1586,7 @@ static void otx2_free_hw_resources(struct otx2_nic *pf)for (qidx =0 ; qidx < qset->cq_cnt; qidx++) { cq = &qset->cq[qidx];if (cq->cq_type ==CQ_RX ) -otx2_cleanup_rx_cqes (pf, cq); +otx2_cleanup_rx_cqes (pf, cq, qidx);else otx2_cleanup_tx_cqes (pf, cq); }
@@ -1594,6 +1596,13 @@ static void otx2_free_hw_resources(struct otx2_nic *pf)/* Free RQ buffer pointers*/ otx2_free_aura_ptr (pf,AURA_NIX_RQ ); +for (qidx =0 ; qidx < pf->hw.rx_queues; qidx++) { + pool_id =otx2_get_pool_idx (pf,AURA_NIX_RQ , qidx); + pool = &pf->qset.pool[pool_id]; +page_pool_destroy (pool->page_pool); + pool->page_pool =NULL ; + } +otx2_free_cq_res (pf);/* Free all ingress bandwidth profiles allocated */
drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c +11 -8
@@ -217,9 +217,6 @@ static bool otx2_skb_add_frag(struct otx2_nic *pfvf, struct sk_buff *skb,skb_add_rx_frag (skb,skb_shinfo (skb)->nr_frags, page, va -page_address (page) + off, len - off, pfvf->rbsize); - -otx2_dma_unmap_page (pfvf, iova -OTX2_HEAD_ROOM , - pfvf->rbsize,DMA_FROM_DEVICE );return true; }
@@ -382,6 +379,8 @@ static void otx2_rcv_pkt_handler(struct otx2_nic *pfvf,if (pfvf->netdev->features &NETIF_F_RXCSUM ) skb->ip_summed =CHECKSUM_UNNECESSARY ; +skb_mark_for_recycle (skb); +napi_gro_frags (napi); }
@@ -1186,11 +1185,13 @@ bool otx2_sq_append_skb(struct net_device *netdev, struct otx2_snd_queue *sq, }EXPORT_SYMBOL (otx2_sq_append_skb); -void otx2_cleanup_rx_cqes (struct otx2_nic *pfvf,struct otx2_cq_queue *cq) +void otx2_cleanup_rx_cqes (struct otx2_nic *pfvf,struct otx2_cq_queue *cq,int qidx) {struct nix_cqe_rx_s *cqe;int processed_cqe =0 ; -u64 iova, pa; +struct otx2_pool *pool; +u16 pool_id; +u64 iova;if (pfvf->xdp_prog)xdp_rxq_info_unreg (&cq->xdp_rxq);
@@ -1198,6 +1199,9 @@ void otx2_cleanup_rx_cqes(struct otx2_nic *pfvf, struct otx2_cq_queue *cq)if (otx2_nix_cq_op_status (pfvf, cq) || !cq->pend_cqe)return ; + pool_id =otx2_get_pool_idx (pfvf,AURA_NIX_RQ , qidx); + pool = &pfvf->qset.pool[pool_id]; +while (cq->pend_cqe) { cqe = (struct nix_cqe_rx_s *)otx2_get_next_cqe (cq); processed_cqe++;
@@ -1210,9 +1214,8 @@ void otx2_cleanup_rx_cqes(struct otx2_nic *pfvf, struct otx2_cq_queue *cq)continue ; } iova = cqe->sg.seg_addr -OTX2_HEAD_ROOM ; - pa =otx2_iova_to_phys (pfvf->iommu_domain, iova); -otx2_dma_unmap_page (pfvf, iova, pfvf->rbsize,DMA_FROM_DEVICE ); -put_page (virt_to_page (phys_to_virt (pa))); + +otx2_free_bufs (pfvf, pool, iova, pfvf->rbsize); }/* Free CQEs to HW */
drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.h +1 -0
@@ -118,6 +118,7 @@ struct otx2_cq_poll {struct otx2_pool {struct qmem *stack;struct qmem *fc_addr; +struct page_pool *page_pool;u16 rbsize; };
drivers/net/ethernet/marvell/octeontx2/nic/qos_sq.c +1 -1
@@ -63,7 +63,7 @@ static int otx2_qos_sq_aura_pool_init(struct otx2_nic *pfvf, int qidx)/* Initialize pool context */ err =otx2_pool_init (pfvf, pool_id, stack_pages, - num_sqbs, hw->sqb_size); + num_sqbs, hw->sqb_size,AURA_NIX_SQ );if (err) goto aura_free;
Signature · 1 line
2.25.1
Quoted 6 lines · Ratheesh Kannoth
Page pool for each rx queue enhance rx side performance by reclaiming buffers back to each queue specific pool. DMA mapping is done only for first allocation of buffers. As subsequent buffers allocation avoid DMA mapping, it results in performance improvement.
Quoted 7 lines
Image | Performance ------------ | ------------ Vannila | 3Mpps | with this | 42Mppschange |
---------------------------
LGTM.
Quoted 3 lines
Signed-off-by Ratheesh Kannoth <[email protected]>---
Quoted 15 lines · Ratheesh Kannoth
Page pool for each rx queue enhance rx side performance by reclaiming buffers back to each queue specific pool. DMA mapping is done only for first allocation of buffers. As subsequent buffers allocation avoid DMA mapping, it results in performance improvement.
Image | Performance ------------ | ------------ Vannila | 3Mpps | with this | 42Mppschange |
---------------------------Signed-off-by Ratheesh Kannoth <[email protected]>
...
drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c void otx2_sq_free_sqbs(struct otx2_nic *pfvf) · line 1233
Quoted 9 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c +5
-0@@ -1205,10 +1226,28 @@ void otx2_sq_free_sqbs(struct otx2_nic *pfvf) } } +void otx2_free_bufs (struct otx2_nic *pfvf,struct otx2_pool *pool, +u64 iova,int size) +{ +u64 pa =otx2_iova_to_phys (pfvf->iommu_domain, iova); +struct page *page =virt_to_head_page (phys_to_virt (pa));
nit: please arrange local variables in networking code in reverse xmas tree
order - longest line to shortest.
u64 pa = otx2_iova_to_phys(pfvf->iommu_domain, iova);
struct page *page;
page = virt_to_head_page(phys_to_virt(pa));
The following tool can check this:
https://github.com/ecree-solarflare/xmastree...
drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c bool otx2_sq_append_skb(struct net_device *netdev, struct otx2_snd_queue *sq, · line 1194
Quoted 13 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c +4
-2@@ -1186,11 +1185,13 @@ bool otx2_sq_append_skb(struct net_device *netdev, struct otx2_snd_queue *sq, }EXPORT_SYMBOL (otx2_sq_append_skb); -void otx2_cleanup_rx_cqes (struct otx2_nic *pfvf,struct otx2_cq_queue *cq) +void otx2_cleanup_rx_cqes (struct otx2_nic *pfvf,struct otx2_cq_queue *cq,int qidx) {struct nix_cqe_rx_s *cqe;int processed_cqe =0 ; -u64 iova, pa; +struct otx2_pool *pool; +u16 pool_id; +u64 iova;
Likewise here.
Quoted 3 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c +0
-0if (pfvf->xdp_prog)xdp_rxq_info_unreg (&cq->xdp_rxq);
...
Ratheesh Kannoth
maintainerQuoted 41 lines
-----Original Message-----
From: Simon Horman <[email protected]>
Sent: Friday, May 19, 2023 3:33 PM
To: Ratheesh Kannoth <[email protected]>
Cc: [email protected]; [email protected]; Sunil Kovvuri Goutham <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; Subbaraya Sundeep Bhatta <[email protected]>; Geethasowjanya Akula <[email protected]>; Srujana Challa <[email protected]>; Hariprasad Kelam <[email protected]>
Subject: [EXT] Re: [PATCH net-next v3] octeontx2-pf: Add support for page poolExternal Email
----------------------------------------------------------------------
Quoted 15 lines · Ratheesh Kannoth
Page pool for each rx queue enhance rx side performance by reclaiming buffers back to each queue specific pool. DMA mapping is done only for first allocation of buffers.
As subsequent buffers allocation avoid DMA mapping, it results in performance improvement.Image | Performance ------------ | ------------ Vannila | 3Mpps | with this | 42Mppschange |
---------------------------Signed-off-by Ratheesh Kannoth <[email protected]>...
Quoted 4 lines
(quoted diff) +0
-0@@ -1205,10 +1226,28 @@ void otx2_sq_free_sqbs(struct otx2_nic *pfvf) } }
Quoted 8 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c void otx2_sq_free_sqbs(struct otx2_nic *pfvf) · line 1233
Quoted 5 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c +5
-0+void otx2_free_bufs (struct otx2_nic *pfvf,struct otx2_pool *pool, +u64 iova,int size) +{ +u64 pa =otx2_iova_to_phys (pfvf->iommu_domain, iova); +struct page *page =virt_to_head_page (phys_to_virt (pa));nit: please arrange local variables in networking code in reverse xmas tree
order - longest line to shortest.
Variable "pa" is used in second line. Are you suggesting to defer assignment later; and only declare variables here in reverse xmas tree style ?
Quoted 21 lines
u64 pa = otx2_iova_to_phys(pfvf->iommu_domain, iova); struct page *page; page = virt_to_head_page(phys_to_virt(pa)); The following tool can check this: https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_ecree- 2Dsolarflare_xmastree&d=DwIBAg&c=nKjWec2b6R0mOyPaz7xtfQ&r=aekcsy BCH00_LewrEDcQBzsRw8KCpUR0vZb_auTHk4M&m=EAM7f8UtMNRC57jfDp gqNsIClttVw6HHyCL7XecFBync-LgDwxIm9_45n-w- WvjY&s=x_0EQ5wG1cZv01ySk_vzRX3kc8Bd2OALcpI8NgxYXTI&e=
...
Quoted 4 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c +0
-0@@ -1186,11 +1185,13 @@ bool otx2_sq_append_skb(struct net_device *netdev, struct otx2_snd_queue *sq, }EXPORT_SYMBOL (otx2_sq_append_skb);
Quoted 13 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c bool otx2_sq_append_skb(struct net_device *netdev, struct otx2_snd_queue *sq, · line 1194
Quoted 11 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c +5
-2-void otx2_cleanup_rx_cqes (struct otx2_nic *pfvf,struct otx2_cq_queue *cq) +void otx2_cleanup_rx_cqes (struct otx2_nic *pfvf,struct otx2_cq_queue +*cq,int qidx) {struct nix_cqe_rx_s *cqe;int processed_cqe =0 ; -u64 iova, pa; +struct otx2_pool *pool; +u16 pool_id; +u64 iova;Likewise here.
Okay
Quoted 5 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c bool otx2_sq_append_skb(struct net_device *netdev, struct otx2_snd_queue *sq, · line 1197
Quoted 4 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_txrx.c +0
-0if (pfvf->xdp_prog)xdp_rxq_info_unreg (&cq->xdp_rxq);
-Ratheesh
Quoted 49 lines · Ratheesh Kannoth
Quoted 41 lines
-----Original Message-----
From: Simon Horman <[email protected]>
Sent: Friday, May 19, 2023 3:33 PM
To: Ratheesh Kannoth <[email protected]>
Cc: [email protected]; [email protected]; Sunil Kovvuri Goutham <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; Subbaraya Sundeep Bhatta <[email protected]>; Geethasowjanya Akula <[email protected]>; Srujana Challa <[email protected]>; Hariprasad Kelam <[email protected]>
Subject: [EXT] Re: [PATCH net-next v3] octeontx2-pf: Add support for page poolExternal Email
----------------------------------------------------------------------
Quoted 15 lines · Ratheesh Kannoth
Page pool for each rx queue enhance rx side performance by reclaiming buffers back to each queue specific pool. DMA mapping is done only for first allocation of buffers.
As subsequent buffers allocation avoid DMA mapping, it results in performance improvement.Image | Performance ------------ | ------------ Vannila | 3Mpps | with this | 42Mppschange |
---------------------------Signed-off-by Ratheesh Kannoth <[email protected]>...
Quoted 4 lines
(quoted diff) +0
-0@@ -1205,10 +1226,28 @@ void otx2_sq_free_sqbs(struct otx2_nic *pfvf) } }Quoted 8 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c void otx2_sq_free_sqbs(struct otx2_nic *pfvf) · line 1233
Quoted 5 lines
drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c +5
-0+void otx2_free_bufs (struct otx2_nic *pfvf,struct otx2_pool *pool, +u64 iova,int size) +{ +u64 pa =otx2_iova_to_phys (pfvf->iommu_domain, iova); +struct page *page =virt_to_head_page (phys_to_virt (pa));nit: please arrange local variables in networking code in reverse xmas tree
order - longest line to shortest.
Quoted 1 line
Variable "pa" is used in second line. Are you suggesting to defer assignment later; and only declare variables here in reverse xmas tree style ?
Yes, that is my suggestion.