@@ -9,7 +9,7 @@ obj-$(CONFIG_RVU_ESWITCH) += rvu_rep.o
rvu_nicpf-y := otx2_pf.o otx2_common.o otx2_txrx.o otx2_ethtool.o \
otx2_flows.o otx2_tc.o cn10k.o otx2_dmac_flt.o \
- otx2_devlink.o qos_sq.o qos.o
+ otx2_devlink.o qos_sq.o qos.o otx2_xsk.o
rvu_nicvf-y := otx2_vf.o
rvu_rep-y := rep.o
@@ -112,9 +112,12 @@ int cn10k_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq)
struct otx2_nic *pfvf = dev;
int cnt = cq->pool_ptrs;
u64 ptrs[NPA_MAX_BURST];
+ struct otx2_pool *pool;
dma_addr_t bufptr;
int num_ptrs = 1;
+ pool = &pfvf->qset.pool[cq->cq_idx];
+
/* Refill pool with new buffers */
while (cq->pool_ptrs) {
if (otx2_alloc_buffer(pfvf, cq, &bufptr)) {
@@ -124,7 +127,8 @@ int cn10k_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq)
break;
}
cq->pool_ptrs--;
- ptrs[num_ptrs] = (u64)bufptr + OTX2_HEAD_ROOM;
+ ptrs[num_ptrs] = pool->xsk_pool ? (u64)bufptr : (u64)bufptr + OTX2_HEAD_ROOM;
+
num_ptrs++;
if (num_ptrs == NPA_MAX_BURST || cq->pool_ptrs == 0) {
__cn10k_aura_freeptr(pfvf, cq->cq_idx, ptrs,
@@ -17,6 +17,7 @@
#include "otx2_common.h"
#include "otx2_struct.h"
#include "cn10k.h"
+#include "otx2_xsk.h"
static bool otx2_is_pfc_enabled(struct otx2_nic *pfvf)
{
@@ -549,10 +550,13 @@ static int otx2_alloc_pool_buf(struct otx2_nic *pfvf, struct otx2_pool *pool,
}
static int __otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
- dma_addr_t *dma)
+ dma_addr_t *dma, int qidx, int idx)
{
u8 *buf;
+ if (pool->xsk_pool)
+ return otx2_xsk_pool_alloc_buf(pfvf, pool, dma, idx);
+
if (pool->page_pool)
return otx2_alloc_pool_buf(pfvf, pool, dma);
@@ -571,12 +575,12 @@ static int __otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
}
int otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
- dma_addr_t *dma)
+ dma_addr_t *dma, int qidx, int idx)
{
int ret;
local_bh_disable();
- ret = __otx2_alloc_rbuf(pfvf, pool, dma);
+ ret = __otx2_alloc_rbuf(pfvf, pool, dma, qidx, idx);
local_bh_enable();
return ret;
}
@@ -584,7 +588,8 @@ int otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
int otx2_alloc_buffer(struct otx2_nic *pfvf, struct otx2_cq_queue *cq,
dma_addr_t *dma)
{
- if (unlikely(__otx2_alloc_rbuf(pfvf, cq->rbpool, dma)))
+ if (unlikely(__otx2_alloc_rbuf(pfvf, cq->rbpool, dma,
+ cq->cq_idx, cq->pool_ptrs - 1)))
return -ENOMEM;
return 0;
}
@@ -884,7 +889,7 @@ void otx2_sqb_flush(struct otx2_nic *pfvf)
#define RQ_PASS_LVL_AURA (255 - ((95 * 256) / 100)) /* RED when 95% is full */
#define RQ_DROP_LVL_AURA (255 - ((99 * 256) / 100)) /* Drop when 99% is full */
-static int otx2_rq_init(struct otx2_nic *pfvf, u16 qidx, u16 lpb_aura)
+int otx2_rq_init(struct otx2_nic *pfvf, u16 qidx, u16 lpb_aura)
{
struct otx2_qset *qset = &pfvf->qset;
struct nix_aq_enq_req *aq;
@@ -1041,7 +1046,7 @@ int otx2_sq_init(struct otx2_nic *pfvf, u16 qidx, u16 sqb_aura)
}
-static int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx)
+int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx)
{
struct otx2_qset *qset = &pfvf->qset;
int err, pool_id, non_xdp_queues;
@@ -1057,11 +1062,18 @@ static int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx)
cq->cint_idx = qidx;
cq->cqe_cnt = qset->rqe_cnt;
if (pfvf->xdp_prog) {
- pool = &qset->pool[qidx];
xdp_rxq_info_reg(&cq->xdp_rxq, pfvf->netdev, qidx, 0);
- xdp_rxq_info_reg_mem_model(&cq->xdp_rxq,
- MEM_TYPE_PAGE_POOL,
- pool->page_pool);
+ pool = &qset->pool[qidx];
+ if (pool->xsk_pool) {
+ xdp_rxq_info_reg_mem_model(&cq->xdp_rxq,
+ MEM_TYPE_XSK_BUFF_POOL,
+ NULL);
+ xsk_pool_set_rxq_info(pool->xsk_pool, &cq->xdp_rxq);
+ } else if (pool->page_pool) {
+ xdp_rxq_info_reg_mem_model(&cq->xdp_rxq,
+ MEM_TYPE_PAGE_POOL,
+ pool->page_pool);
+ }
}
} else if (qidx < non_xdp_queues) {
cq->cq_type = CQ_TX;
@@ -1281,9 +1293,10 @@ void otx2_free_bufs(struct otx2_nic *pfvf, struct otx2_pool *pool,
pa = otx2_iova_to_phys(pfvf->iommu_domain, iova);
page = virt_to_head_page(phys_to_virt(pa));
-
if (pool->page_pool) {
page_pool_put_full_page(pool->page_pool, page, true);
+ } else if (pool->xsk_pool) {
+ /* Note: No way of identifying xdp_buff */
} else {
dma_unmap_page_attrs(pfvf->dev, iova, size,
DMA_FROM_DEVICE,
@@ -1298,6 +1311,7 @@ void otx2_free_aura_ptr(struct otx2_nic *pfvf, int type)
int pool_id, pool_start = 0, pool_end = 0, size = 0;
struct otx2_pool *pool;
u64 iova;
+ int idx;
if (type == AURA_NIX_SQ) {
pool_start = otx2_get_pool_idx(pfvf, type, 0);
@@ -1312,8 +1326,8 @@ void otx2_free_aura_ptr(struct otx2_nic *pfvf, int type)
/* Free SQB and RQB pointers from the aura pool */
for (pool_id = pool_start; pool_id < pool_end; pool_id++) {
- iova = otx2_aura_allocptr(pfvf, pool_id);
pool = &pfvf->qset.pool[pool_id];
+ iova = otx2_aura_allocptr(pfvf, pool_id);
while (iova) {
if (type == AURA_NIX_RQ)
iova -= OTX2_HEAD_ROOM;
@@ -1323,6 +1337,13 @@ void otx2_free_aura_ptr(struct otx2_nic *pfvf, int type)
iova = otx2_aura_allocptr(pfvf, pool_id);
}
}
+
+ for (idx = 0 ; idx < pool->xdp_cnt; idx++) {
+ if (!pool->xdp[idx])
+ continue;
+
+ xsk_buff_free(pool->xdp[idx]);
+ }
}
void otx2_aura_pool_free(struct otx2_nic *pfvf)
@@ -1337,8 +1358,12 @@ void otx2_aura_pool_free(struct otx2_nic *pfvf)
pool = &pfvf->qset.pool[pool_id];
qmem_free(pfvf->dev, pool->stack);
qmem_free(pfvf->dev, pool->fc_addr);
- page_pool_destroy(pool->page_pool);
- pool->page_pool = NULL;
+ if (pool->page_pool) {
+ page_pool_destroy(pool->page_pool);
+ pool->page_pool = NULL;
+ }
+ devm_kfree(pfvf->dev, pool->xdp);
+ pool->xsk_pool = NULL;
}
devm_kfree(pfvf->dev, pfvf->qset.pool);
pfvf->qset.pool = NULL;
@@ -1425,9 +1450,10 @@ int otx2_pool_init(struct otx2_nic *pfvf, u16 pool_id,
int stack_pages, int numptrs, int buf_size, int type)
{
struct page_pool_params pp_params = { 0 };
+ struct xsk_buff_pool *xsk_pool;
struct npa_aq_enq_req *aq;
struct otx2_pool *pool;
- int err;
+ int err, sz;
pool = &pfvf->qset.pool[pool_id];
/* Alloc memory for stack which is used to store buffer pointers */
@@ -1468,21 +1494,36 @@ int otx2_pool_init(struct otx2_nic *pfvf, u16 pool_id,
aq->ctype = NPA_AQ_CTYPE_POOL;
aq->op = NPA_AQ_INSTOP_INIT;
- if (type != AURA_NIX_RQ) {
- pool->page_pool = NULL;
+ if (type != AURA_NIX_RQ)
+ return 0;
+
+ if (!test_bit(pool_id, pfvf->af_xdp_zc_qidx)) {
+ sz = ALIGN(ALIGN(SKB_DATA_ALIGN(buf_size), OTX2_ALIGN), PAGE_SIZE);
+ pp_params.order = (sz / PAGE_SIZE) - 1;
+ pp_params.flags = PP_FLAG_DMA_MAP;
+ pp_params.pool_size = min(OTX2_PAGE_POOL_SZ, numptrs);
+ pp_params.nid = NUMA_NO_NODE;
+ pp_params.dev = pfvf->dev;
+ pp_params.dma_dir = DMA_FROM_DEVICE;
+ pool->page_pool = page_pool_create(&pp_params);
+ if (IS_ERR(pool->page_pool)) {
+ netdev_err(pfvf->netdev, "Creation of page pool failed\n");
+ return PTR_ERR(pool->page_pool);
+ }
return 0;
}
- pp_params.order = get_order(buf_size);
- pp_params.flags = PP_FLAG_DMA_MAP;
- pp_params.pool_size = min(OTX2_PAGE_POOL_SZ, numptrs);
- pp_params.nid = NUMA_NO_NODE;
- pp_params.dev = pfvf->dev;
- pp_params.dma_dir = DMA_FROM_DEVICE;
- pool->page_pool = page_pool_create(&pp_params);
- if (IS_ERR(pool->page_pool)) {
- netdev_err(pfvf->netdev, "Creation of page pool failed\n");
- return PTR_ERR(pool->page_pool);
+ /* Set XSK pool to support AF_XDP zero-copy */
+ xsk_pool = xsk_get_pool_from_qid(pfvf->netdev, pool_id);
+ if (xsk_pool) {
+ pool->xsk_pool = xsk_pool;
+ pool->xdp_cnt = numptrs;
+ pool->xdp = devm_kcalloc(pfvf->dev,
+ numptrs, sizeof(struct xdp_buff *), GFP_KERNEL);
+ if (IS_ERR(pool->xdp)) {
+ netdev_err(pfvf->netdev, "Creation of xsk pool failed\n");
+ return PTR_ERR(pool->xdp);
+ }
}
return 0;
@@ -1543,9 +1584,18 @@ int otx2_sq_aura_pool_init(struct otx2_nic *pfvf)
}
for (ptr = 0; ptr < num_sqbs; ptr++) {
- err = otx2_alloc_rbuf(pfvf, pool, &bufptr);
- if (err)
+ err = otx2_alloc_rbuf(pfvf, pool, &bufptr, pool_id, ptr);
+ if (err) {
+ if (pool->xsk_pool) {
+ ptr--;
+ while (ptr >= 0) {
+ xsk_buff_free(pool->xdp[ptr]);
+ ptr--;
+ }
+ }
goto err_mem;
+ }
+
pfvf->hw_ops->aura_freeptr(pfvf, pool_id, bufptr);
sq->sqb_ptrs[sq->sqb_count++] = (u64)bufptr;
}
@@ -1595,11 +1645,19 @@ int otx2_rq_aura_pool_init(struct otx2_nic *pfvf)
/* Allocate pointers and free them to aura/pool */
for (pool_id = 0; pool_id < hw->rqpool_cnt; pool_id++) {
pool = &pfvf->qset.pool[pool_id];
+
for (ptr = 0; ptr < num_ptrs; ptr++) {
- err = otx2_alloc_rbuf(pfvf, pool, &bufptr);
- if (err)
+ err = otx2_alloc_rbuf(pfvf, pool, &bufptr, pool_id, ptr);
+ if (err) {
+ if (pool->xsk_pool) {
+ while (ptr)
+ xsk_buff_free(pool->xdp[--ptr]);
+ }
return -ENOMEM;
+ }
+
pfvf->hw_ops->aura_freeptr(pfvf, pool_id,
+ pool->xsk_pool ? bufptr :
bufptr + OTX2_HEAD_ROOM);
}
}
@@ -532,6 +532,7 @@ struct otx2_nic {
/* Inline ipsec */
struct cn10k_ipsec ipsec;
+ unsigned long *af_xdp_zc_qidx;
};
static inline bool is_otx2_lbkvf(struct pci_dev *pdev)
@@ -1003,7 +1004,7 @@ void otx2_txschq_free_one(struct otx2_nic *pfvf, u16 lvl, u16 schq);
void otx2_free_pending_sqe(struct otx2_nic *pfvf);
void otx2_sqb_flush(struct otx2_nic *pfvf);
int otx2_alloc_rbuf(struct otx2_nic *pfvf, struct otx2_pool *pool,
- dma_addr_t *dma);
+ dma_addr_t *dma, int qidx, int idx);
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);
@@ -1033,6 +1034,14 @@ void otx2_pfaf_mbox_destroy(struct otx2_nic *pf);
void otx2_disable_mbox_intr(struct otx2_nic *pf);
void otx2_disable_napi(struct otx2_nic *pf);
irqreturn_t otx2_cq_intr_handler(int irq, void *cq_irq);
+int otx2_init_rsrc(struct pci_dev *pdev, struct otx2_nic *pf);
+int otx2_aura_aq_init(struct otx2_nic *pfvf, int aura_id,
+ int pool_id, int numptrs);
+int otx2_pool_aq_init(struct otx2_nic *pfvf, u16 pool_id,
+ int stack_pages, int numptrs, int buf_size, int type);
+int otx2_set_hw_capabilities(struct otx2_nic *pfvf);
+int otx2_rq_init(struct otx2_nic *pfvf, u16 qidx, u16 lpb_aura);
+int otx2_cq_init(struct otx2_nic *pfvf, u16 qidx);
/* RSS configuration APIs*/
int otx2_rss_init(struct otx2_nic *pfvf);
@@ -27,6 +27,7 @@
#include "qos.h"
#include <rvu_trace.h>
#include "cn10k_ipsec.h"
+#include "otx2_xsk.h"
#define DRV_NAME "rvu_nicpf"
#define DRV_STRING "Marvell RVU NIC Physical Function Driver"
@@ -1662,9 +1663,7 @@ 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 */
@@ -1705,13 +1704,6 @@ 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 */
@@ -2789,6 +2781,8 @@ static int otx2_xdp(struct net_device *netdev, struct netdev_bpf *xdp)
switch (xdp->command) {
case XDP_SETUP_PROG:
return otx2_xdp_setup(pf, xdp->prog);
+ case XDP_SETUP_XSK_POOL:
+ return otx2_xsk_pool_setup(pf, xdp->xsk.pool, xdp->xsk.queue_id);
default:
return -EINVAL;
}
@@ -2866,6 +2860,7 @@ static const struct net_device_ops otx2_netdev_ops = {
.ndo_set_vf_vlan = otx2_set_vf_vlan,
.ndo_get_vf_config = otx2_get_vf_config,
.ndo_bpf = otx2_xdp,
+ .ndo_xsk_wakeup = otx2_xsk_wakeup,
.ndo_xdp_xmit = otx2_xdp_xmit,
.ndo_setup_tc = otx2_setup_tc,
.ndo_set_vf_trust = otx2_ndo_set_vf_trust,
@@ -3204,6 +3199,10 @@ static int otx2_probe(struct pci_dev *pdev, const struct pci_device_id *id)
/* Enable link notifications */
otx2_cgx_config_linkevents(pf, true);
+ pf->af_xdp_zc_qidx = bitmap_zalloc(qcount, GFP_KERNEL);
+ if (!pf->af_xdp_zc_qidx)
+ goto err_pf_sriov_init;
+
#ifdef CONFIG_DCB
err = otx2_dcbnl_set_ops(netdev);
if (err)
@@ -12,6 +12,7 @@
#include <linux/bpf_trace.h>
#include <net/ip6_checksum.h>
#include <net/xfrm.h>
+#include <net/xdp.h>
#include "otx2_reg.h"
#include "otx2_common.h"
@@ -535,6 +536,7 @@ int otx2_napi_handler(struct napi_struct *napi, int budget)
struct otx2_cq_poll *cq_poll;
int workdone = 0, cq_idx, i;
struct otx2_cq_queue *cq;
+ struct otx2_pool *pool;
struct otx2_qset *qset;
struct otx2_nic *pfvf;
int filled_cnt = -1;
@@ -559,6 +561,7 @@ int otx2_napi_handler(struct napi_struct *napi, int budget)
if (rx_cq && rx_cq->pool_ptrs)
filled_cnt = pfvf->hw_ops->refill_pool_ptrs(pfvf, rx_cq);
+
/* Clear the IRQ */
otx2_write64(pfvf, NIX_LF_CINTX_INT(cq_poll->cint_idx), BIT_ULL(0));
@@ -571,6 +574,7 @@ int otx2_napi_handler(struct napi_struct *napi, int budget)
if (pfvf->flags & OTX2_FLAG_ADPTV_INT_COAL_ENABLED)
otx2_adjust_adaptive_coalese(pfvf, cq_poll);
+ pool = &pfvf->qset.pool[cq->cq_idx];
if (unlikely(!filled_cnt)) {
struct refill_work *work;
struct delayed_work *dwork;
@@ -584,7 +588,13 @@ int otx2_napi_handler(struct napi_struct *napi, int budget)
schedule_delayed_work(dwork,
msecs_to_jiffies(100));
}
+ /* Call for wake-up for not able to fill buffers */
+ if (pool->xsk_pool)
+ xsk_set_rx_need_wakeup(pool->xsk_pool);
} else {
+ /* Clear wake-up, since buffers are filled successfully */
+ if (pool->xsk_pool)
+ xsk_clear_rx_need_wakeup(pool->xsk_pool);
/* Re-enable interrupts */
otx2_write64(pfvf,
NIX_LF_CINTX_ENA_W1S(cq_poll->cint_idx),
@@ -1235,15 +1245,19 @@ void otx2_cleanup_rx_cqes(struct otx2_nic *pfvf, struct otx2_cq_queue *cq, int q
u16 pool_id;
u64 iova;
- if (pfvf->xdp_prog)
+ pool_id = otx2_get_pool_idx(pfvf, AURA_NIX_RQ, qidx);
+ pool = &pfvf->qset.pool[pool_id];
+
+ if (pfvf->xdp_prog) {
+ if (pool->page_pool)
+ xdp_rxq_info_unreg_mem_model(&cq->xdp_rxq);
+
xdp_rxq_info_unreg(&cq->xdp_rxq);
+ }
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++;
@@ -1426,13 +1440,24 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
unsigned char *hard_start;
struct otx2_pool *pool;
int qidx = cq->cq_idx;
- struct xdp_buff xdp;
+ struct xdp_buff xdp, *xsk_buff = NULL;
struct page *page;
u64 iova, pa;
u32 act;
int err;
pool = &pfvf->qset.pool[qidx];
+
+ if (pool->xsk_pool) {
+ xsk_buff = pool->xdp[--cq->rbpool->xdp_top];
+ if (!xsk_buff)
+ return false;
+
+ xsk_buff->data_end = xsk_buff->data + cqe->sg.seg_size;
+ act = bpf_prog_run_xdp(prog, xsk_buff);
+ goto handle_xdp_verdict;
+ }
+
iova = cqe->sg.seg_addr - OTX2_HEAD_ROOM;
pa = otx2_iova_to_phys(pfvf->iommu_domain, iova);
page = virt_to_page(phys_to_virt(pa));
@@ -1445,6 +1470,7 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
act = bpf_prog_run_xdp(prog, &xdp);
+handle_xdp_verdict:
switch (act) {
case XDP_PASS:
break;
@@ -1455,8 +1481,16 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
cqe->sg.seg_size, qidx);
case XDP_REDIRECT:
cq->pool_ptrs++;
- err = xdp_do_redirect(pfvf->netdev, &xdp, prog);
+ if (xsk_buff) {
+ err = xdp_do_redirect(pfvf->netdev, xsk_buff, prog);
+ if (!err) {
+ *need_xdp_flush = true;
+ return true;
+ }
+ return false;
+ }
+ err = xdp_do_redirect(pfvf->netdev, &xdp, prog);
otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize,
DMA_FROM_DEVICE);
if (!err) {
@@ -1469,13 +1503,21 @@ static bool otx2_xdp_rcv_pkt_handler(struct otx2_nic *pfvf,
bpf_warn_invalid_xdp_action(pfvf->netdev, prog, act);
break;
case XDP_ABORTED:
+ if (xsk_buff)
+ xsk_buff_free(xsk_buff);
trace_xdp_exception(pfvf->netdev, prog, act);
break;
case XDP_DROP:
- otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize,
- DMA_FROM_DEVICE);
- page_pool_recycle_direct(pool->page_pool, page);
cq->pool_ptrs++;
+ if (xsk_buff) {
+ xsk_buff_free(xsk_buff);
+ } else if (page->pp) {
+ page_pool_recycle_direct(pool->page_pool, page);
+ } else {
+ otx2_dma_unmap_page(pfvf, iova, pfvf->rbsize,
+ DMA_FROM_DEVICE);
+ put_page(page);
+ }
return true;
}
return false;
@@ -12,6 +12,7 @@
#include <linux/iommu.h>
#include <linux/if_vlan.h>
#include <net/xdp.h>
+#include <net/xdp_sock_drv.h>
#define LBK_CHAN_BASE 0x000
#define SDP_CHAN_BASE 0x700
@@ -127,7 +128,11 @@ struct otx2_pool {
struct qmem *stack;
struct qmem *fc_addr;
struct page_pool *page_pool;
+ struct xsk_buff_pool *xsk_pool;
+ struct xdp_buff **xdp;
+ u16 xdp_cnt;
u16 rbsize;
+ u16 xdp_top;
};
struct otx2_cq_queue {
@@ -144,6 +149,7 @@ struct otx2_cq_queue {
void *cqe_base;
struct qmem *cqe;
struct otx2_pool *rbpool;
+ bool xsk_zc_en;
struct xdp_rxq_info xdp_rxq;
} ____cacheline_aligned_in_smp;
@@ -722,6 +722,10 @@ static int otx2vf_probe(struct pci_dev *pdev, const struct pci_device_id *id)
if (err)
goto err_shutdown_tc;
+ vf->af_xdp_zc_qidx = bitmap_zalloc(qcount, GFP_KERNEL);
+ if (!vf->af_xdp_zc_qidx)
+ goto err_shutdown_tc;
+
#ifdef CONFIG_DCB
err = otx2_dcbnl_set_ops(netdev);
if (err)
new file mode 100644
@@ -0,0 +1,182 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Marvell RVU Ethernet driver
+ *
+ * Copyright (C) 2024 Marvell.
+ *
+ */
+
+#include <linux/bpf_trace.h>
+#include <linux/stringify.h>
+#include <net/xdp_sock_drv.h>
+#include <net/xdp.h>
+
+#include "otx2_common.h"
+#include "otx2_xsk.h"
+
+int otx2_xsk_pool_alloc_buf(struct otx2_nic *pfvf, struct otx2_pool *pool,
+ dma_addr_t *dma, int idx)
+{
+ struct xdp_buff *xdp;
+ int delta;
+
+ xdp = xsk_buff_alloc(pool->xsk_pool);
+ if (!xdp)
+ return -ENOMEM;
+
+ pool->xdp[pool->xdp_top++] = xdp;
+ *dma = OTX2_DATA_ALIGN(xsk_buff_xdp_get_dma(xdp));
+ /* Adjust xdp->data for unaligned addresses */
+ delta = *dma - xsk_buff_xdp_get_dma(xdp);
+ xdp->data += delta;
+
+ return 0;
+}
+
+static int otx2_xsk_ctx_disable(struct otx2_nic *pfvf, u16 qidx, int aura_id)
+{
+ struct nix_cn10k_aq_enq_req *cn10k_rq_aq;
+ struct npa_aq_enq_req *aura_aq;
+ struct npa_aq_enq_req *pool_aq;
+ struct nix_aq_enq_req *rq_aq;
+
+ if (test_bit(CN10K_LMTST, &pfvf->hw.cap_flag)) {
+ cn10k_rq_aq = otx2_mbox_alloc_msg_nix_cn10k_aq_enq(&pfvf->mbox);
+ if (!cn10k_rq_aq)
+ return -ENOMEM;
+ cn10k_rq_aq->qidx = qidx;
+ cn10k_rq_aq->rq.ena = 0;
+ cn10k_rq_aq->rq_mask.ena = 1;
+ cn10k_rq_aq->ctype = NIX_AQ_CTYPE_RQ;
+ cn10k_rq_aq->op = NIX_AQ_INSTOP_WRITE;
+ } else {
+ rq_aq = otx2_mbox_alloc_msg_nix_aq_enq(&pfvf->mbox);
+ if (!rq_aq)
+ return -ENOMEM;
+ rq_aq->qidx = qidx;
+ rq_aq->sq.ena = 0;
+ rq_aq->sq_mask.ena = 1;
+ rq_aq->ctype = NIX_AQ_CTYPE_RQ;
+ rq_aq->op = NIX_AQ_INSTOP_WRITE;
+ }
+
+ aura_aq = otx2_mbox_alloc_msg_npa_aq_enq(&pfvf->mbox);
+ if (!aura_aq) {
+ otx2_mbox_reset(&pfvf->mbox.mbox, 0);
+ return -ENOMEM;
+ }
+
+ aura_aq->aura_id = aura_id;
+ aura_aq->aura.ena = 0;
+ aura_aq->aura_mask.ena = 1;
+ aura_aq->ctype = NPA_AQ_CTYPE_AURA;
+ aura_aq->op = NPA_AQ_INSTOP_WRITE;
+
+ pool_aq = otx2_mbox_alloc_msg_npa_aq_enq(&pfvf->mbox);
+ if (!pool_aq) {
+ otx2_mbox_reset(&pfvf->mbox.mbox, 0);
+ return -ENOMEM;
+ }
+
+ pool_aq->aura_id = aura_id;
+ pool_aq->pool.ena = 0;
+ pool_aq->pool_mask.ena = 1;
+
+ pool_aq->ctype = NPA_AQ_CTYPE_POOL;
+ pool_aq->op = NPA_AQ_INSTOP_WRITE;
+
+ return otx2_sync_mbox_msg(&pfvf->mbox);
+}
+
+static void otx2_clean_up_rq(struct otx2_nic *pfvf, int qidx)
+{
+ struct otx2_qset *qset = &pfvf->qset;
+ struct otx2_cq_queue *cq;
+ struct otx2_pool *pool;
+ u64 iova;
+
+ /* If the DOWN flag is set SQs are already freed */
+ if (pfvf->flags & OTX2_FLAG_INTF_DOWN)
+ return;
+
+ cq = &qset->cq[qidx];
+ if (cq)
+ otx2_cleanup_rx_cqes(pfvf, cq, qidx);
+
+ pool = &pfvf->qset.pool[qidx];
+ iova = otx2_aura_allocptr(pfvf, qidx);
+ while (iova) {
+ iova -= OTX2_HEAD_ROOM;
+ otx2_free_bufs(pfvf, pool, iova, pfvf->rbsize);
+ iova = otx2_aura_allocptr(pfvf, qidx);
+ }
+
+ mutex_lock(&pfvf->mbox.lock);
+ otx2_xsk_ctx_disable(pfvf, qidx, qidx);
+ mutex_unlock(&pfvf->mbox.lock);
+}
+
+int otx2_xsk_pool_enable(struct otx2_nic *pf, struct xsk_buff_pool *pool, u16 qidx)
+{
+ u16 rx_queues = pf->hw.rx_queues;
+ u16 tx_queues = pf->hw.tx_queues;
+ int err;
+
+ if (qidx >= rx_queues || qidx >= tx_queues)
+ return -EINVAL;
+
+ err = xsk_pool_dma_map(pool, pf->dev, DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
+ if (err)
+ return err;
+
+ set_bit(qidx, pf->af_xdp_zc_qidx);
+ otx2_clean_up_rq(pf, qidx);
+ /* Kick start the NAPI context so that receiving will start */
+ return otx2_xsk_wakeup(pf->netdev, qidx, XDP_WAKEUP_RX);
+}
+
+int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qidx)
+{
+ struct net_device *netdev = pf->netdev;
+ struct xsk_buff_pool *pool;
+
+ pool = xsk_get_pool_from_qid(netdev, qidx);
+ if (!pool)
+ return -EINVAL;
+
+ otx2_clean_up_rq(pf, qidx);
+ clear_bit(qidx, pf->af_xdp_zc_qidx);
+ xsk_pool_dma_unmap(pool, DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
+
+ return 0;
+}
+
+int otx2_xsk_pool_setup(struct otx2_nic *pf, struct xsk_buff_pool *pool, u16 qidx)
+{
+ if (pool)
+ return otx2_xsk_pool_enable(pf, pool, qidx);
+
+ return otx2_xsk_pool_disable(pf, qidx);
+}
+
+int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
+{
+ struct otx2_nic *pf = netdev_priv(dev);
+ struct otx2_cq_poll *cq_poll = NULL;
+ struct otx2_qset *qset = &pf->qset;
+
+ if (pf->flags & OTX2_FLAG_INTF_DOWN)
+ return -ENETDOWN;
+
+ if (queue_id >= pf->hw.rx_queues)
+ return -EINVAL;
+
+ cq_poll = &qset->napi[queue_id];
+ if (!cq_poll)
+ return -EINVAL;
+
+ /* Trigger interrupt */
+ if (!napi_if_scheduled_mark_missed(&cq_poll->napi))
+ otx2_write64(pf, NIX_LF_CINTX_ENA_W1S(cq_poll->cint_idx), BIT_ULL(0));
+
+ return 0;
+}
new file mode 100644
@@ -0,0 +1,21 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* Marvell RVU PF/VF Netdev Devlink
+ *
+ * Copyright (C) 2024 Marvell.
+ *
+ */
+
+#ifndef OTX2_XSK_H
+#define OTX2_XSK_H
+
+struct otx2_nic;
+struct xsk_buff_pool;
+
+int otx2_xsk_pool_setup(struct otx2_nic *pf, struct xsk_buff_pool *pool, u16 qid);
+int otx2_xsk_pool_enable(struct otx2_nic *pf, struct xsk_buff_pool *pool, u16 qid);
+int otx2_xsk_pool_disable(struct otx2_nic *pf, u16 qid);
+int otx2_xsk_pool_alloc_buf(struct otx2_nic *pfvf, struct otx2_pool *pool,
+ dma_addr_t *dma, int idx);
+int otx2_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags);
+
+#endif /* OTX2_XSK_H */
@@ -82,7 +82,7 @@ static int otx2_qos_sq_aura_pool_init(struct otx2_nic *pfvf, int qidx)
}
for (ptr = 0; ptr < num_sqbs; ptr++) {
- err = otx2_alloc_rbuf(pfvf, pool, &bufptr);
+ err = otx2_alloc_rbuf(pfvf, pool, &bufptr, pool_id, ptr);
if (err)
goto sqb_free;
pfvf->hw_ops->aura_freeptr(pfvf, pool_id, bufptr);
This patch adds support to AF_XDP zero copy support for CN10K. This patch specifically adds receive side support. In this approach once a xdp program with zero copy support on a specific rx queue is enabled, then that receive queue is disabled/detached from the existing kernel queue and re-assigned to the umem memory. Signed-off-by: Suman Ghosh <sumang@marvell.com> --- .../ethernet/marvell/octeontx2/nic/Makefile | 2 +- .../ethernet/marvell/octeontx2/nic/cn10k.c | 6 +- .../marvell/octeontx2/nic/otx2_common.c | 120 +++++++++--- .../marvell/octeontx2/nic/otx2_common.h | 11 +- .../ethernet/marvell/octeontx2/nic/otx2_pf.c | 17 +- .../marvell/octeontx2/nic/otx2_txrx.c | 60 +++++- .../marvell/octeontx2/nic/otx2_txrx.h | 6 + .../ethernet/marvell/octeontx2/nic/otx2_vf.c | 4 + .../ethernet/marvell/octeontx2/nic/otx2_xsk.c | 182 ++++++++++++++++++ .../ethernet/marvell/octeontx2/nic/otx2_xsk.h | 21 ++ .../ethernet/marvell/octeontx2/nic/qos_sq.c | 2 +- 11 files changed, 378 insertions(+), 53 deletions(-) create mode 100644 drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.c create mode 100644 drivers/net/ethernet/marvell/octeontx2/nic/otx2_xsk.h