From: Selvin Xavier <selvin.xavier@broadcom.com>
Date: Wed, 2 Aug 2017 01:46:18 -0700
Subject: RDMA/bnxt_re: Allocate multiple notification queues
Patch-mainline: v4.14-rc1
Git-commit: 6a5df91baf2528e584bf4493c30bbafe2db74c9e
References: bsc#1050244 FATE#322915
Enables multiple Interrupt vectors. Driver is requesting the max
MSIX vectors based on the number of online cpus and creates upto
9 MSIx vectors (1 for control path and 8 for data path).
A tasklet is created for each of these vectors. NQs are assigned
to CQs in round robin fashion.
This patch also adds IRQ affinity hint for the MSIX vector of each NQ.
Signed-off-by: Ray Jui <ray.jui@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Acked-by: Thomas Bogendoerfer <tbogendoerfer@suse.de>
---
drivers/infiniband/hw/bnxt_re/bnxt_re.h | 5 -
drivers/infiniband/hw/bnxt_re/ib_verbs.c | 17 +++-
drivers/infiniband/hw/bnxt_re/main.c | 108 +++++++++++++++++++------------
drivers/infiniband/hw/bnxt_re/qplib_fp.c | 21 +++++-
drivers/infiniband/hw/bnxt_re/qplib_fp.h | 4 -
5 files changed, 105 insertions(+), 50 deletions(-)
--- a/drivers/infiniband/hw/bnxt_re/bnxt_re.h
+++ b/drivers/infiniband/hw/bnxt_re/bnxt_re.h
@@ -85,7 +85,7 @@ struct bnxt_re_sqp_entries {
};
#define BNXT_RE_MIN_MSIX 2
-#define BNXT_RE_MAX_MSIX 16
+#define BNXT_RE_MAX_MSIX 9
#define BNXT_RE_AEQ_IDX 0
#define BNXT_RE_NQ_IDX 1
@@ -116,7 +116,7 @@ struct bnxt_re_dev {
struct bnxt_qplib_rcfw rcfw;
/* NQ */
- struct bnxt_qplib_nq nq;
+ struct bnxt_qplib_nq nq[BNXT_RE_MAX_MSIX];
/* Device Resources */
struct bnxt_qplib_dev_attr dev_attr;
@@ -140,6 +140,7 @@ struct bnxt_re_dev {
struct bnxt_re_qp *qp1_sqp;
struct bnxt_re_ah *sqp_ah;
struct bnxt_re_sqp_entries sqp_tbl[1024];
+ atomic_t nq_alloc_cnt;
};
#define to_bnxt_re_dev(ptr, member) \
--- a/drivers/infiniband/hw/bnxt_re/ib_verbs.c
+++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c
@@ -2290,6 +2290,7 @@ int bnxt_re_destroy_cq(struct ib_cq *ib_
struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
struct bnxt_re_dev *rdev = cq->rdev;
int rc;
+ struct bnxt_qplib_nq *nq = cq->qplib_cq.nq;
rc = bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
if (rc) {
@@ -2304,7 +2305,7 @@ int bnxt_re_destroy_cq(struct ib_cq *ib_
kfree(cq);
}
atomic_dec(&rdev->cq_count);
- rdev->nq.budget--;
+ nq->budget--;
return 0;
}
@@ -2318,6 +2319,8 @@ struct ib_cq *bnxt_re_create_cq(struct i
struct bnxt_re_cq *cq = NULL;
int rc, entries;
int cqe = attr->cqe;
+ struct bnxt_qplib_nq *nq = NULL;
+ unsigned int nq_alloc_cnt;
/* Validate CQ fields */
if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
@@ -2369,9 +2372,15 @@ struct ib_cq *bnxt_re_create_cq(struct i
cq->qplib_cq.sghead = NULL;
cq->qplib_cq.nmap = 0;
}
+ /*
+ * Allocating the NQ in a round robin fashion. nq_alloc_cnt is a
+ * used for getting the NQ index.
+ */
+ nq_alloc_cnt = atomic_inc_return(&rdev->nq_alloc_cnt);
+ nq = &rdev->nq[nq_alloc_cnt % (rdev->num_msix - 1)];
cq->qplib_cq.max_wqe = entries;
- cq->qplib_cq.cnq_hw_ring_id = rdev->nq.ring_id;
- cq->qplib_cq.nq = &rdev->nq;
+ cq->qplib_cq.cnq_hw_ring_id = nq->ring_id;
+ cq->qplib_cq.nq = nq;
rc = bnxt_qplib_create_cq(&rdev->qplib_res, &cq->qplib_cq);
if (rc) {
@@ -2381,7 +2390,7 @@ struct ib_cq *bnxt_re_create_cq(struct i
cq->ib_cq.cqe = entries;
cq->cq_period = cq->qplib_cq.period;
- rdev->nq.budget++;
+ nq->budget++;
atomic_inc(&rdev->cq_count);
--- a/drivers/infiniband/hw/bnxt_re/main.c
+++ b/drivers/infiniband/hw/bnxt_re/main.c
@@ -161,7 +161,7 @@ static int bnxt_re_free_msix(struct bnxt
static int bnxt_re_request_msix(struct bnxt_re_dev *rdev)
{
- int rc = 0, num_msix_want = BNXT_RE_MIN_MSIX, num_msix_got;
+ int rc = 0, num_msix_want = BNXT_RE_MAX_MSIX, num_msix_got;
struct bnxt_en_dev *en_dev;
if (!rdev)
@@ -169,6 +169,8 @@ static int bnxt_re_request_msix(struct b
en_dev = rdev->en_dev;
+ num_msix_want = min_t(u32, BNXT_RE_MAX_MSIX, num_online_cpus());
+
rtnl_lock();
num_msix_got = en_dev->en_ops->bnxt_request_msix(en_dev, BNXT_ROCE_ULP,
rdev->msix_entries,
@@ -651,8 +653,12 @@ static int bnxt_re_cqn_handler(struct bn
static void bnxt_re_cleanup_res(struct bnxt_re_dev *rdev)
{
- if (rdev->nq.hwq.max_elements)
- bnxt_qplib_disable_nq(&rdev->nq);
+ int i;
+
+ if (rdev->nq[0].hwq.max_elements) {
+ for (i = 1; i < rdev->num_msix; i++)
+ bnxt_qplib_disable_nq(&rdev->nq[i - 1]);
+ }
if (rdev->qplib_res.rcfw)
bnxt_qplib_cleanup_res(&rdev->qplib_res);
@@ -660,31 +666,41 @@ static void bnxt_re_cleanup_res(struct b
static int bnxt_re_init_res(struct bnxt_re_dev *rdev)
{
- int rc = 0;
+ int rc = 0, i;
bnxt_qplib_init_res(&rdev->qplib_res);
- if (rdev->msix_entries[BNXT_RE_NQ_IDX].vector <= 0)
- return -EINVAL;
-
- rc = bnxt_qplib_enable_nq(rdev->en_dev->pdev, &rdev->nq,
- rdev->msix_entries[BNXT_RE_NQ_IDX].vector,
- rdev->msix_entries[BNXT_RE_NQ_IDX].db_offset,
- &bnxt_re_cqn_handler,
- NULL);
+ for (i = 1; i < rdev->num_msix ; i++) {
+ rc = bnxt_qplib_enable_nq(rdev->en_dev->pdev, &rdev->nq[i - 1],
+ i - 1, rdev->msix_entries[i].vector,
+ rdev->msix_entries[i].db_offset,
+ &bnxt_re_cqn_handler, NULL);
+
+ if (rc) {
+ dev_err(rdev_to_dev(rdev),
+ "Failed to enable NQ with rc = 0x%x", rc);
+ goto fail;
+ }
+ }
+ return 0;
+fail:
+ return rc;
+}
- if (rc)
- dev_err(rdev_to_dev(rdev), "Failed to enable NQ: %#x", rc);
+static void bnxt_re_free_nq_res(struct bnxt_re_dev *rdev, bool lock_wait)
+{
+ int i;
- return rc;
+ for (i = 0; i < rdev->num_msix - 1; i++) {
+ bnxt_re_net_ring_free(rdev, rdev->nq[i].ring_id, lock_wait);
+ bnxt_qplib_free_nq(&rdev->nq[i]);
+ }
}
static void bnxt_re_free_res(struct bnxt_re_dev *rdev, bool lock_wait)
{
- if (rdev->nq.hwq.max_elements) {
- bnxt_re_net_ring_free(rdev, rdev->nq.ring_id, lock_wait);
- bnxt_qplib_free_nq(&rdev->nq);
- }
+ bnxt_re_free_nq_res(rdev, lock_wait);
+
if (rdev->qplib_res.dpi_tbl.max) {
bnxt_qplib_dealloc_dpi(&rdev->qplib_res,
&rdev->qplib_res.dpi_tbl,
@@ -698,7 +714,7 @@ static void bnxt_re_free_res(struct bnxt
static int bnxt_re_alloc_res(struct bnxt_re_dev *rdev)
{
- int rc = 0;
+ int rc = 0, i;
/* Configure and allocate resources for qplib */
rdev->qplib_res.rcfw = &rdev->rcfw;
@@ -715,30 +731,42 @@ static int bnxt_re_alloc_res(struct bnxt
&rdev->dpi_privileged,
rdev);
if (rc)
- goto fail;
+ goto dealloc_res;
- rdev->nq.hwq.max_elements = BNXT_RE_MAX_CQ_COUNT +
- BNXT_RE_MAX_SRQC_COUNT + 2;
- rc = bnxt_qplib_alloc_nq(rdev->en_dev->pdev, &rdev->nq);
- if (rc) {
- dev_err(rdev_to_dev(rdev),
- "Failed to allocate NQ memory: %#x", rc);
- goto fail;
- }
- rc = bnxt_re_net_ring_alloc
- (rdev, rdev->nq.hwq.pbl[PBL_LVL_0].pg_map_arr,
- rdev->nq.hwq.pbl[rdev->nq.hwq.level].pg_count,
- HWRM_RING_ALLOC_CMPL, BNXT_QPLIB_NQE_MAX_CNT - 1,
- rdev->msix_entries[BNXT_RE_NQ_IDX].ring_idx,
- &rdev->nq.ring_id);
- if (rc) {
- dev_err(rdev_to_dev(rdev),
- "Failed to allocate NQ ring: %#x", rc);
- goto free_nq;
+ for (i = 0; i < rdev->num_msix - 1; i++) {
+ rdev->nq[i].hwq.max_elements = BNXT_RE_MAX_CQ_COUNT +
+ BNXT_RE_MAX_SRQC_COUNT + 2;
+ rc = bnxt_qplib_alloc_nq(rdev->en_dev->pdev, &rdev->nq[i]);
+ if (rc) {
+ dev_err(rdev_to_dev(rdev), "Alloc Failed NQ%d rc:%#x",
+ i, rc);
+ goto dealloc_dpi;
+ }
+ rc = bnxt_re_net_ring_alloc
+ (rdev, rdev->nq[i].hwq.pbl[PBL_LVL_0].pg_map_arr,
+ rdev->nq[i].hwq.pbl[rdev->nq[i].hwq.level].pg_count,
+ HWRM_RING_ALLOC_CMPL,
+ BNXT_QPLIB_NQE_MAX_CNT - 1,
+ rdev->msix_entries[i + 1].ring_idx,
+ &rdev->nq[i].ring_id);
+ if (rc) {
+ dev_err(rdev_to_dev(rdev),
+ "Failed to allocate NQ fw id with rc = 0x%x",
+ rc);
+ goto free_nq;
+ }
}
return 0;
free_nq:
- bnxt_qplib_free_nq(&rdev->nq);
+ for (i = 0; i < rdev->num_msix - 1; i++)
+ bnxt_qplib_free_nq(&rdev->nq[i]);
+dealloc_dpi:
+ bnxt_qplib_dealloc_dpi(&rdev->qplib_res,
+ &rdev->qplib_res.dpi_tbl,
+ &rdev->dpi_privileged);
+dealloc_res:
+ bnxt_qplib_free_res(&rdev->qplib_res);
+
fail:
rdev->qplib_res.rcfw = NULL;
return rc;
--- a/drivers/infiniband/hw/bnxt_re/qplib_fp.c
+++ b/drivers/infiniband/hw/bnxt_re/qplib_fp.c
@@ -365,6 +365,7 @@ void bnxt_qplib_disable_nq(struct bnxt_q
tasklet_kill(&nq->worker);
if (nq->requested) {
+ irq_set_affinity_hint(nq->vector, NULL);
free_irq(nq->vector, nq);
nq->requested = false;
}
@@ -378,7 +379,7 @@ void bnxt_qplib_disable_nq(struct bnxt_q
}
int bnxt_qplib_enable_nq(struct pci_dev *pdev, struct bnxt_qplib_nq *nq,
- int msix_vector, int bar_reg_offset,
+ int nq_idx, int msix_vector, int bar_reg_offset,
int (*cqn_handler)(struct bnxt_qplib_nq *nq,
struct bnxt_qplib_cq *),
int (*srqn_handler)(struct bnxt_qplib_nq *nq,
@@ -402,13 +403,25 @@ int bnxt_qplib_enable_nq(struct pci_dev
goto fail;
nq->requested = false;
- rc = request_irq(nq->vector, bnxt_qplib_nq_irq, 0, "bnxt_qplib_nq", nq);
+ memset(nq->name, 0, 32);
+ sprintf(nq->name, "bnxt_qplib_nq-%d", nq_idx);
+ rc = request_irq(nq->vector, bnxt_qplib_nq_irq, 0, nq->name, nq);
if (rc) {
dev_err(&nq->pdev->dev,
"Failed to request IRQ for NQ: %#x", rc);
bnxt_qplib_disable_nq(nq);
goto fail;
}
+
+ cpumask_clear(&nq->mask);
+ cpumask_set_cpu(nq_idx, &nq->mask);
+ rc = irq_set_affinity_hint(nq->vector, &nq->mask);
+ if (rc) {
+ dev_warn(&nq->pdev->dev,
+ "QPLIB: set affinity failed; vector: %d nq_idx: %d\n",
+ nq->vector, nq_idx);
+ }
+
nq->requested = true;
nq->bar_reg = NQ_CONS_PCI_BAR_REGION;
nq->bar_reg_off = bar_reg_offset;
@@ -432,8 +445,10 @@ fail:
void bnxt_qplib_free_nq(struct bnxt_qplib_nq *nq)
{
- if (nq->hwq.max_elements)
+ if (nq->hwq.max_elements) {
bnxt_qplib_free_hwq(nq->pdev, &nq->hwq);
+ nq->hwq.max_elements = 0;
+ }
}
int bnxt_qplib_alloc_nq(struct pci_dev *pdev, struct bnxt_qplib_nq *nq)
--- a/drivers/infiniband/hw/bnxt_re/qplib_fp.h
+++ b/drivers/infiniband/hw/bnxt_re/qplib_fp.h
@@ -407,6 +407,7 @@ struct bnxt_qplib_nq {
struct pci_dev *pdev;
int vector;
+ cpumask_t mask;
int budget;
bool requested;
struct tasklet_struct worker;
@@ -425,6 +426,7 @@ struct bnxt_qplib_nq {
void *srq,
u8 event);
struct workqueue_struct *cqn_wq;
+ char name[32];
};
struct bnxt_qplib_nq_work {
@@ -435,7 +437,7 @@ struct bnxt_qplib_nq_work {
void bnxt_qplib_disable_nq(struct bnxt_qplib_nq *nq);
int bnxt_qplib_enable_nq(struct pci_dev *pdev, struct bnxt_qplib_nq *nq,
- int msix_vector, int bar_reg_offset,
+ int nq_idx, int msix_vector, int bar_reg_offset,
int (*cqn_handler)(struct bnxt_qplib_nq *nq,
struct bnxt_qplib_cq *cq),
int (*srqn_handler)(struct bnxt_qplib_nq *nq,