Skip to contentSkip to conversation

RDMA/bnxt_re: driver update for supporting low latency push

PATCH Superseded Selvin Xavier 3 participants 17 messages Apr 10, 2023 – Apr 12, 2023

Subsystems (2)

v1 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Selvin Xavier

author maintainer
#0 lore raw

cover RDMA/bnxt_re: driver update for supporting low latency push

The series aims to add support for Low latency push path in some of the bnxt devices. The low latency implementation is supported only for the user applications. Also, the code is modified to use common mmap helper functions exported by IB core.

User library changes are added in the pull request
https://github.com/linux-rdma/rdma-core/pull/1321

Please review.

Thanks,
Selvin Xavier

Selvin Xavier (6):
RDMA/bnxt_re: Use the common mmap helper functions RDMA/bnxt_re: Add disassociate ucontext support RDMA/bnxt_re: Query function capabilities from firmware RDMA/bnxt_re: Move the interface version to chip context structure RDMA/bnxt_re: Reorg the bar mapping
RDMA/bnxt_re: Enable low latency push

 drivers/infiniband/hw/bnxt_re/ib_verbs.c   | 160 +++++++++++++++++++++---
 drivers/infiniband/hw/bnxt_re/ib_verbs.h   |  17 +++
 drivers/infiniband/hw/bnxt_re/main.c       | 123 +++++++++++++++++-
 drivers/infiniband/hw/bnxt_re/qplib_fp.c   |  14 +--
 drivers/infiniband/hw/bnxt_re/qplib_rcfw.c |   2 +-
 drivers/infiniband/hw/bnxt_re/qplib_rcfw.h |   1 +
 drivers/infiniband/hw/bnxt_re/qplib_res.c  | 192 +++++++++++++++++++----------
 drivers/infiniband/hw/bnxt_re/qplib_res.h  |  33 +++--
 drivers/infiniband/hw/bnxt_re/qplib_sp.c   |   3 +
 drivers/infiniband/hw/bnxt_re/qplib_sp.h   |   1 +
 include/uapi/rdma/bnxt_re-abi.h            |   9 ++
 11 files changed, 449 insertions(+), 106 deletions(-)
Signature · 1 line
2.5.5
author maintainer
#1 lore raw

PATCH 1/6 RDMA/bnxt_re: Use the common mmap helper functions

Replace the mmap handling function with common code in IB core. Create rdma_user_mmap_entry for each mmap resource and add to the ib_core mmap list. Add mmap_free verb support. Also, use rdma_user_mmap_io while mapping Doorbell pages.

---
 drivers/infiniband/hw/bnxt_re/ib_verbs.c  | 110 ++++++++++++++++++++++++++----
 drivers/infiniband/hw/bnxt_re/ib_verbs.h  |  15 ++++
 drivers/infiniband/hw/bnxt_re/main.c      |   1 +
 drivers/infiniband/hw/bnxt_re/qplib_fp.c  |  10 +--
 drivers/infiniband/hw/bnxt_re/qplib_res.c |   2 +-
 5 files changed, 115 insertions(+), 23 deletions(-)
drivers/infiniband/hw/bnxt_re/ib_verbs.c +95 -15
index e86afec..c9d134c 100644
@@ -533,12 +533,42 @@ static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd)
 	return rc;
 }
 
+static struct rdma_user_mmap_entry*
+bnxt_re_mmap_entry_insert(struct bnxt_re_ucontext *uctx, u64 mem_offset,
+			  enum bnxt_re_mmap_flag mmap_flag, u64 *offset)
+{
+	struct bnxt_re_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
+	int ret;
+
+	if (!entry)
+		return NULL;
+
+	entry->mem_offset = mem_offset;
+	entry->mmap_flag = mmap_flag;
+
+	ret = rdma_user_mmap_entry_insert(&uctx->ib_uctx,
+					  &entry->rdma_entry, PAGE_SIZE);
+	if (ret) {
+		kfree(entry);
+		return NULL;
+	}
+	if (offset)
+		*offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
+
+	return &entry->rdma_entry;
+}
+
 /* Protection Domains */
 int bnxt_re_dealloc_pd(struct ib_pd *ib_pd, struct ib_udata *udata)
 {
 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
 	struct bnxt_re_dev *rdev = pd->rdev;
 
+	if (udata) {
+		rdma_user_mmap_entry_remove(pd->pd_db_mmap);
+		pd->pd_db_mmap = NULL;
+	}
+
 	bnxt_re_destroy_fence_mr(pd);
 
 	if (pd->qplib_pd.id) {
@@ -557,7 +587,7 @@ int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 	struct bnxt_re_ucontext *ucntx = rdma_udata_to_drv_context(
 		udata, struct bnxt_re_ucontext, ib_uctx);
 	struct bnxt_re_pd *pd = container_of(ibpd, struct bnxt_re_pd, ib_pd);
-	int rc;
+	int rc = 0;
 
 	pd->rdev = rdev;
 	if (bnxt_qplib_alloc_pd(&rdev->qplib_res.pd_tbl, &pd->qplib_pd)) {
@@ -567,7 +597,7 @@ int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 	}
 
 	if (udata) {
-		struct bnxt_re_pd_resp resp;
+		struct bnxt_re_pd_resp resp = {};
 
 		if (!ucntx->dpi.dbr) {
 			/* Allocate DPI in alloc_pd to avoid failing of
@@ -584,12 +614,23 @@ int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 		resp.pdid = pd->qplib_pd.id;
 		/* Still allow mapping this DBR to the new user PD. */
 		resp.dpi = ucntx->dpi.dpi;
-		resp.dbr = (u64)ucntx->dpi.umdbr;
 
-		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
+		pd->pd_db_mmap = bnxt_re_mmap_entry_insert(ucntx, (u64)ucntx->dpi.umdbr,
+							   BNXT_RE_MMAP_UC_DB, &resp.dbr);
+
+		if (!pd->pd_db_mmap) {
+			ibdev_err(&rdev->ibdev,
+				  "Failed to insert mmap entry\n");
+			rc = -ENOMEM;
+			goto dbfail;
+		}
+
+		rc = ib_copy_to_udata(udata, &resp, min(sizeof(resp), udata->outlen));
 		if (rc) {
 			ibdev_err(&rdev->ibdev,
 				  "Failed to copy user response\n");
+			rdma_user_mmap_entry_remove(pd->pd_db_mmap);
+			rc = -EFAULT;
 			goto dbfail;
 		}
 	}
@@ -3994,6 +4035,13 @@ int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata)
 	resp.comp_mask |= BNXT_RE_UCNTX_CMASK_HAVE_MODE;
 	resp.mode = rdev->chip_ctx->modes.wqe_mode;
 
+	uctx->shpage_mmap = bnxt_re_mmap_entry_insert(uctx, 0, BNXT_RE_MMAP_SH_PAGE, NULL);
+	if (!uctx->shpage_mmap) {
+		ibdev_err(ibdev, "Failed to create mmap entry for shared page\n");
+		rc = -ENOMEM;
+		goto cfail;
+	}
+
 	rc = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
 	if (rc) {
 		ibdev_err(ibdev, "Failed to copy user context");
@@ -4017,6 +4065,8 @@ void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx)
 
 	struct bnxt_re_dev *rdev = uctx->rdev;
 
+	rdma_user_mmap_entry_remove(uctx->shpage_mmap);
+	uctx->shpage_mmap = NULL;
 	if (uctx->shpg)
 		free_page((unsigned long)uctx->shpg);
 
@@ -4036,27 +4086,57 @@ int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
 						   struct bnxt_re_ucontext,
 						   ib_uctx);
+	struct bnxt_re_user_mmap_entry *bnxt_entry;
+	struct rdma_user_mmap_entry *rdma_entry;
 	struct bnxt_re_dev *rdev = uctx->rdev;
+	int ret = 0;
 	u64 pfn;
 
-	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
+	rdma_entry = rdma_user_mmap_entry_get(&uctx->ib_uctx, vma);
+	if (!rdma_entry) {
+		ibdev_err(&rdev->ibdev,
+			  "%s: No valid rdma_entry pgoffset = 0x%lx\n",
+			  __func__, vma->vm_pgoff);
 		return -EINVAL;
+	}
 
-	if (vma->vm_pgoff) {
-		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
-		if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
-				       PAGE_SIZE, vma->vm_page_prot)) {
-			ibdev_err(&rdev->ibdev, "Failed to map DPI");
-			return -EAGAIN;
-		}
-	} else {
+	bnxt_entry = container_of(rdma_entry, struct bnxt_re_user_mmap_entry,
+				  rdma_entry);
+
+	switch (bnxt_entry->mmap_flag) {
+	case BNXT_RE_MMAP_UC_DB:
+		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
+		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
+					pgprot_noncached(vma->vm_page_prot),
+				rdma_entry);
+		if (ret)
+			ibdev_err(&rdev->ibdev, "Failed to map shared page");
+		break;
+	case BNXT_RE_MMAP_SH_PAGE:
 		pfn = virt_to_phys(uctx->shpg) >> PAGE_SHIFT;
 		if (remap_pfn_range(vma, vma->vm_start,
 				    pfn, PAGE_SIZE, vma->vm_page_prot)) {
 			ibdev_err(&rdev->ibdev, "Failed to map shared page");
-			return -EAGAIN;
+			ret =  -EAGAIN;
 		}
+		break;
+	default:
+		ibdev_err(&rdev->ibdev,
+			  "%s: invalid mmap flag = 0x%x\n", __func__, bnxt_entry->mmap_flag);
+		ret = -EINVAL;
+		break;
 	}
 
-	return 0;
+	rdma_user_mmap_entry_put(rdma_entry);
+	return ret;
+}
+
+void bnxt_re_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
+{
+	struct bnxt_re_user_mmap_entry *bnxt_entry;
+
+	bnxt_entry = container_of(rdma_entry, struct bnxt_re_user_mmap_entry,
+				  rdma_entry);
+
+	kfree(bnxt_entry);
 }
drivers/infiniband/hw/bnxt_re/ib_verbs.h +15 -0
index 31f7e34..dcd31ae 100644
@@ -60,6 +60,7 @@ struct bnxt_re_pd {
 	struct bnxt_re_dev	*rdev;
 	struct bnxt_qplib_pd	qplib_pd;
 	struct bnxt_re_fence_data fence;
+	struct rdma_user_mmap_entry *pd_db_mmap;
 };
 
 struct bnxt_re_ah {
@@ -136,6 +137,18 @@ struct bnxt_re_ucontext {
 	struct bnxt_qplib_dpi	dpi;
 	void			*shpg;
 	spinlock_t		sh_lock;	/* protect shpg */
+	struct rdma_user_mmap_entry *shpage_mmap;
+};
+
+enum bnxt_re_mmap_flag {
+	BNXT_RE_MMAP_SH_PAGE,
+	BNXT_RE_MMAP_UC_DB,
+};
+
+struct bnxt_re_user_mmap_entry {
+	struct rdma_user_mmap_entry rdma_entry;
+	u64 mem_offset;
+	u8 mmap_flag;
 };
 
 static inline u16 bnxt_re_get_swqe_size(int nsge)
@@ -213,6 +226,8 @@ struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata);
 void bnxt_re_dealloc_ucontext(struct ib_ucontext *context);
 int bnxt_re_mmap(struct ib_ucontext *context, struct vm_area_struct *vma);
+void bnxt_re_mmap_free(struct rdma_user_mmap_entry *rdma_entry);
+
 
 unsigned long bnxt_re_lock_cqs(struct bnxt_re_qp *qp);
 void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp, unsigned long flags);
drivers/infiniband/hw/bnxt_re/main.c +1 -0
index b9e2f89..1d361eb 100644
@@ -539,6 +539,7 @@ static const struct ib_device_ops bnxt_re_dev_ops = {
 	.get_port_immutable = bnxt_re_get_port_immutable,
 	.map_mr_sg = bnxt_re_map_mr_sg,
 	.mmap = bnxt_re_mmap,
+	.mmap_free = bnxt_re_mmap_free,
 	.modify_qp = bnxt_re_modify_qp,
 	.modify_srq = bnxt_re_modify_srq,
 	.poll_cq = bnxt_re_poll_cq,
drivers/infiniband/hw/bnxt_re/qplib_fp.c +3 -7
index f139d4c..ade858f 100644
@@ -471,7 +471,6 @@ static int bnxt_qplib_map_nq_db(struct bnxt_qplib_nq *nq,  u32 reg_offt)
 	resource_size_t reg_base;
 	struct bnxt_qplib_nq_db *nq_db;
 	struct pci_dev *pdev;
-	int rc = 0;
 
 	pdev = nq->pdev;
 	nq_db = &nq->nq_db;
@@ -481,8 +480,7 @@ static int bnxt_qplib_map_nq_db(struct bnxt_qplib_nq *nq,  u32 reg_offt)
 	if (!nq_db->reg.bar_base) {
 		dev_err(&pdev->dev, "QPLIB: NQ BAR region %d resc start is 0!",
 			nq_db->reg.bar_id);
-		rc = -ENOMEM;
-		goto fail;
+		return -ENOMEM;
 	}
 
 	reg_base = nq_db->reg.bar_base + reg_offt;
@@ -492,15 +490,13 @@ static int bnxt_qplib_map_nq_db(struct bnxt_qplib_nq *nq,  u32 reg_offt)
 	if (!nq_db->reg.bar_reg) {
 		dev_err(&pdev->dev, "QPLIB: NQ BAR region %d mapping failed",
 			nq_db->reg.bar_id);
-		rc = -ENOMEM;
-		goto fail;
+		return -ENOMEM;
 	}
 
 	nq_db->dbinfo.db = nq_db->reg.bar_reg;
 	nq_db->dbinfo.hwq = &nq->hwq;
 	nq_db->dbinfo.xid = nq->ring_id;
-fail:
-	return rc;
+	return 0;
 }
 
 int bnxt_qplib_enable_nq(struct pci_dev *pdev, struct bnxt_qplib_nq *nq,
drivers/infiniband/hw/bnxt_re/qplib_res.c +1 -1
index 126d4f2..920ab87 100644
@@ -813,7 +813,7 @@ static int bnxt_qplib_alloc_dpi_tbl(struct bnxt_qplib_res     *res,
 	return 0;
 
 unmap_io:
-	pci_iounmap(res->pdev, dpit->dbr_bar_reg_iomem);
+	iounmap(dpit->dbr_bar_reg_iomem);
 	dpit->dbr_bar_reg_iomem = NULL;
 	return -ENOMEM;
 }
Signature · 1 line
2.5.5
#2 lore raw

Re: [PATCH for-next 1/6] RDMA/bnxt_re: Use the common mmap helper functions

Quoted 20 lines · Selvin Xavier

Replace the mmap handling function with common code in IB core. Create rdma_user_mmap_entry for each mmap resource and add to the ib_core mmap list. Add mmap_free verb support. Also, use rdma_user_mmap_io while mapping Doorbell pages.

---
 drivers/infiniband/hw/bnxt_re/ib_verbs.c  | 110 ++++++++++++++++++++++++++----
 drivers/infiniband/hw/bnxt_re/ib_verbs.h  |  15 ++++
 drivers/infiniband/hw/bnxt_re/main.c      |   1 +
 drivers/infiniband/hw/bnxt_re/qplib_fp.c  |  10 +--
 drivers/infiniband/hw/bnxt_re/qplib_res.c |   2 +-
 5 files changed, 115 insertions(+), 23 deletions(-)
drivers/infiniband/hw/bnxt_re/ib_verbs.c +0 -0
index e86afec..c9d134c 100644

drivers/infiniband/hw/bnxt_re/ib_verbs.c static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd) · line 540

Quoted 8 lines
drivers/infiniband/hw/bnxt_re/ib_verbs.c +5 -0
@@ -533,12 +533,42 @@ static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd)
 	return rc;
 }
 
+static struct rdma_user_mmap_entry*
+bnxt_re_mmap_entry_insert(struct bnxt_re_ucontext *uctx, u64 mem_offset,
+			  enum bnxt_re_mmap_flag mmap_flag, u64 *offset)
+{
+	struct bnxt_re_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);

Please separate calls to kzalloc() and variable initialization.

Thanks

#3 lore raw

Re: [PATCH for-next 1/6] RDMA/bnxt_re: Use the common mmap helper functions

In reply to Selvin Xavier

drivers/infiniband/hw/bnxt_re/ib_verbs.c int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) · line 623

Quoted 7 lines · Selvin Xavier
drivers/infiniband/hw/bnxt_re/ib_verbs.c +6 -1
-		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
+		pd->pd_db_mmap = bnxt_re_mmap_entry_insert(ucntx, (u64)ucntx->dpi.umdbr,
+							   BNXT_RE_MMAP_UC_DB, &resp.dbr);
+
+		if (!pd->pd_db_mmap) {
+			ibdev_err(&rdev->ibdev,
+				  "Failed to insert mmap entry\n");

No prints from drivers on failure paths. dbg at worst.

Quoted 8 lines
drivers/infiniband/hw/bnxt_re/ib_verbs.c +8 -0
+	switch (bnxt_entry->mmap_flag) {
+	case BNXT_RE_MMAP_UC_DB:
+		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
+		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
+					pgprot_noncached(vma->vm_page_prot),
+				rdma_entry);
+		if (ret)
+			ibdev_err(&rdev->ibdev, "Failed to map shared page");

drivers/infiniband/hw/bnxt_re/ib_verbs.c int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma) · line 4122

Quoted 10 lines
drivers/infiniband/hw/bnxt_re/ib_verbs.c +4 -1
+		break;
+	case BNXT_RE_MMAP_SH_PAGE:
 		pfn = virt_to_phys(uctx->shpg) >> PAGE_SHIFT;
 		if (remap_pfn_range(vma, vma->vm_start,
 				    pfn, PAGE_SIZE, vma->vm_page_prot)) {
 			ibdev_err(&rdev->ibdev, "Failed to map shared page");
-			return -EAGAIN;
+			ret =  -EAGAIN;
 		}
+		break;

What is this? You can't enable disassociate support until all the mmap's in the driver have been fixed.

Jason

author maintainer
#4 lore raw

Re: [PATCH for-next 1/6] RDMA/bnxt_re: Use the common mmap helper functions

Quoted 31 lines · Jason Gunthorpe

drivers/infiniband/hw/bnxt_re/ib_verbs.c int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) · line 623

Quoted 7 lines · Selvin Xavier
drivers/infiniband/hw/bnxt_re/ib_verbs.c +6 -1
-             rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
+             pd->pd_db_mmap = bnxt_re_mmap_entry_insert(ucntx, (u64)ucntx->dpi.umdbr,
+                                                        BNXT_RE_MMAP_UC_DB, &resp.dbr);
+
+             if (!pd->pd_db_mmap) {
+                     ibdev_err(&rdev->ibdev,
+                               "Failed to insert mmap entry\n");

No prints from drivers on failure paths. dbg at worst.

Quoted 8 lines
drivers/infiniband/hw/bnxt_re/ib_verbs.c +8 -0
+     switch (bnxt_entry->mmap_flag) {
+     case BNXT_RE_MMAP_UC_DB:
+             pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
+             ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
+                                     pgprot_noncached(vma->vm_page_prot),
+                             rdma_entry);
+             if (ret)
+                     ibdev_err(&rdev->ibdev, "Failed to map shared page");

drivers/infiniband/hw/bnxt_re/ib_verbs.c int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma) · line 4122

Quoted 10 lines
drivers/infiniband/hw/bnxt_re/ib_verbs.c +4 -1
+             break;
+     case BNXT_RE_MMAP_SH_PAGE:
              pfn = virt_to_phys(uctx->shpg) >> PAGE_SHIFT;
              if (remap_pfn_range(vma, vma->vm_start,
                                  pfn, PAGE_SIZE, vma->vm_page_prot)) {
                      ibdev_err(&rdev->ibdev, "Failed to map shared page");
-                     return -EAGAIN;
+                     ret =  -EAGAIN;
              }
+             break;
Quoted 2 lines

What is this? You can't enable disassociate support until all the mmap's in the driver have been fixed.

Noted. Will modify this also to use rdma mmap apis.

Quoted 3 lines

Jason

author maintainer
#5 lore raw

PATCH 2/6 RDMA/bnxt_re: Add disassociate ucontext support

In reply to Selvin Xavier

Add empty stub for disassociate ucontext as done in other vendor drivers.

---
 drivers/infiniband/hw/bnxt_re/main.c | 5 +++++
 1 file changed, 5 insertions(+)
drivers/infiniband/hw/bnxt_re/main.c +5 -0
index 1d361eb..a866951 100644
@@ -466,6 +466,10 @@ static int bnxt_re_net_stats_ctx_alloc(struct bnxt_re_dev *rdev,
 	return rc;
 }
 
+static void bnxt_re_disassociate_ucontext(struct ib_ucontext *ibcontext)
+{
+}
+
 /* Device */
 
 static struct bnxt_re_dev *bnxt_re_from_netdev(struct net_device *netdev)
@@ -532,6 +536,7 @@ static const struct ib_device_ops bnxt_re_dev_ops = {
 	.destroy_qp = bnxt_re_destroy_qp,
 	.destroy_srq = bnxt_re_destroy_srq,
 	.device_group = &bnxt_re_dev_attr_group,
+	.disassociate_ucontext = bnxt_re_disassociate_ucontext,
 	.get_dev_fw_str = bnxt_re_query_fw_str,
 	.get_dma_mr = bnxt_re_get_dma_mr,
 	.get_hw_stats = bnxt_re_ib_get_hw_stats,
Signature · 1 line
2.5.5
#6 lore raw

Re: [PATCH for-next 2/6] RDMA/bnxt_re: Add disassociate ucontext support

Quoted 2 lines · Selvin Xavier

Add empty stub for disassociate ucontext as done in other vendor drivers.

It will be great to mention in commit message that the reason to this stub is because you use rdma_user_mmap_io().

Thanks

Quoted 32 lines
---
 drivers/infiniband/hw/bnxt_re/main.c | 5 +++++
 1 file changed, 5 insertions(+)
drivers/infiniband/hw/bnxt_re/main.c +5 -0
index 1d361eb..a866951 100644
@@ -466,6 +466,10 @@ static int bnxt_re_net_stats_ctx_alloc(struct bnxt_re_dev *rdev,
 	return rc;
 }
 
+static void bnxt_re_disassociate_ucontext(struct ib_ucontext *ibcontext)
+{
+}
+
 /* Device */
 
 static struct bnxt_re_dev *bnxt_re_from_netdev(struct net_device *netdev)
@@ -532,6 +536,7 @@ static const struct ib_device_ops bnxt_re_dev_ops = {
 	.destroy_qp = bnxt_re_destroy_qp,
 	.destroy_srq = bnxt_re_destroy_srq,
 	.device_group = &bnxt_re_dev_attr_group,
+	.disassociate_ucontext = bnxt_re_disassociate_ucontext,
 	.get_dev_fw_str = bnxt_re_query_fw_str,
 	.get_dma_mr = bnxt_re_get_dma_mr,
 	.get_hw_stats = bnxt_re_ib_get_hw_stats,
Signature · 1 line
2.5.5
#7 lore raw

Re: [PATCH for-next 2/6] RDMA/bnxt_re: Add disassociate ucontext support

Quoted 6 lines · Leon Romanovsky
Quoted 2 lines · Selvin Xavier

Add empty stub for disassociate ucontext as done in other vendor drivers.

It will be great to mention in commit message that the reason to this stub is because you use rdma_user_mmap_io().

That isn't the reason to do this stub..

It is some yet-to-be-cleaned-up way for the driver to indicate it *fully* supports driver disassociation.

rdma_user_mmap_io() is part of that support, but the whole functionality should be tested before setting this.

Jason

author maintainer
#8 lore raw

PATCH 3/6 RDMA/bnxt_re: Query function capabilities from firmware

In reply to Selvin Xavier

Query Function capabilities to enable advanced features.

---
 drivers/infiniband/hw/bnxt_re/main.c | 30 ++++++++++++++++++++++++++++++
 1 file changed, 30 insertions(+)
drivers/infiniband/hw/bnxt_re/main.c +30 -0
index a866951..1b69198 100644
@@ -83,6 +83,7 @@ static int bnxt_re_netdev_event(struct notifier_block *notifier,
 				unsigned long event, void *ptr);
 static struct bnxt_re_dev *bnxt_re_from_netdev(struct net_device *netdev);
 static void bnxt_re_dev_uninit(struct bnxt_re_dev *rdev);
+static int bnxt_re_hwrm_qcaps(struct bnxt_re_dev *rdev);
 
 static void bnxt_re_set_drv_mode(struct bnxt_re_dev *rdev, u8 mode)
 {
@@ -91,6 +92,9 @@ static void bnxt_re_set_drv_mode(struct bnxt_re_dev *rdev, u8 mode)
 	cctx = rdev->chip_ctx;
 	cctx->modes.wqe_mode = bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx) ?
 			       mode : BNXT_QPLIB_WQE_MODE_STATIC;
+	if (bnxt_re_hwrm_qcaps(rdev))
+		dev_err(rdev_to_dev(rdev),
+			"Failed to query hwrm qcaps\n");
 }
 
 static void bnxt_re_destroy_chip_ctx(struct bnxt_re_dev *rdev)
@@ -334,6 +338,32 @@ static void bnxt_re_fill_fw_msg(struct bnxt_fw_msg *fw_msg, void *msg,
 	fw_msg->timeout = timeout;
 }
 
+/* Query function capabilities using common hwrm */
+int bnxt_re_hwrm_qcaps(struct bnxt_re_dev *rdev)
+{
+	struct bnxt_en_dev *en_dev = rdev->en_dev;
+	struct hwrm_func_qcaps_output resp = {0};
+	struct hwrm_func_qcaps_input req = {0};
+	struct bnxt_qplib_chip_ctx *cctx;
+	struct bnxt_fw_msg fw_msg;
+	int rc;
+
+	cctx = rdev->chip_ctx;
+	memset(&fw_msg, 0, sizeof(fw_msg));
+	bnxt_re_init_hwrm_hdr(rdev, (void *)&req,
+			      HWRM_FUNC_QCAPS, -1, -1);
+	req.fid = cpu_to_le16(0xffff);
+	bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&resp,
+			    sizeof(resp), DFLT_HWRM_CMD_TIMEOUT);
+	rc = bnxt_send_msg(en_dev, &fw_msg);
+	if (rc) {
+		dev_err(rdev_to_dev(rdev),
+			"Failed to query capabilities, rc = %#x", rc);
+		return rc;
+	}
+	return 0;
+}
+
 static int bnxt_re_net_ring_free(struct bnxt_re_dev *rdev,
 				 u16 fw_ring_id, int type)
 {
Signature · 1 line
2.5.5
#9 lore raw

Re: [PATCH for-next 3/6] RDMA/bnxt_re: Query function capabilities from firmware

Quoted 20 lines · Selvin Xavier

Query Function capabilities to enable advanced features.

---
 drivers/infiniband/hw/bnxt_re/main.c | 30 ++++++++++++++++++++++++++++++
 1 file changed, 30 insertions(+)
drivers/infiniband/hw/bnxt_re/main.c +1 -0
index a866951..1b69198 100644
@@ -83,6 +83,7 @@ static int bnxt_re_netdev_event(struct notifier_block *notifier,
 				unsigned long event, void *ptr);
 static struct bnxt_re_dev *bnxt_re_from_netdev(struct net_device *netdev);
 static void bnxt_re_dev_uninit(struct bnxt_re_dev *rdev);
+static int bnxt_re_hwrm_qcaps(struct bnxt_re_dev *rdev);
 
 static void bnxt_re_set_drv_mode(struct bnxt_re_dev *rdev, u8 mode)
 {

drivers/infiniband/hw/bnxt_re/main.c static void bnxt_re_set_drv_mode(struct bnxt_re_dev *rdev, u8 mode) · line 97

Quoted 6 lines
drivers/infiniband/hw/bnxt_re/main.c +3 -0
@@ -91,6 +92,9 @@ static void bnxt_re_set_drv_mode(struct bnxt_re_dev *rdev, u8 mode)
 	cctx = rdev->chip_ctx;
 	cctx->modes.wqe_mode = bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx) ?
 			       mode : BNXT_QPLIB_WQE_MODE_STATIC;
+	if (bnxt_re_hwrm_qcaps(rdev))
+		dev_err(rdev_to_dev(rdev),
+			"Failed to query hwrm qcaps\n");

You already print error in bnxt_re_hwrm_qcaps()

Quoted 1 line
drivers/infiniband/hw/bnxt_re/main.c +0 -0
 }

<...>

Quoted 2 lines
drivers/infiniband/hw/bnxt_re/main.c +2 -0
+/* Query function capabilities using common hwrm */
+int bnxt_re_hwrm_qcaps(struct bnxt_re_dev *rdev)

<...>

Quoted 14 lines
drivers/infiniband/hw/bnxt_re/main.c +8 -0
+	if (rc) {
+		dev_err(rdev_to_dev(rdev),
+			"Failed to query capabilities, rc = %#x", rc);
+		return rc;
+	}
+	return 0;
+}
+
 static int bnxt_re_net_ring_free(struct bnxt_re_dev *rdev,
 				 u16 fw_ring_id, int type)
 {
Signature · 1 line
2.5.5
author maintainer
#10 lore raw

PATCH 4/6 RDMA/bnxt_re: Move the interface version to chip context structure

In reply to Selvin Xavier

FW interface version check is required for multiple features. Moving the interface version to chip context structure.

---
 drivers/infiniband/hw/bnxt_re/main.c       | 2 +-
 drivers/infiniband/hw/bnxt_re/qplib_rcfw.c | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
drivers/infiniband/hw/bnxt_re/main.c +1 -1
index 1b69198..ecaa69cb 100644
@@ -1094,7 +1094,7 @@ static void bnxt_re_query_hwrm_intf_version(struct bnxt_re_dev *rdev)
 			  rc);
 		return;
 	}
-	rdev->qplib_ctx.hwrm_intf_ver =
+	rdev->chip_ctx->hwrm_intf_ver =
 		(u64)le16_to_cpu(resp.hwrm_intf_major) << 48 |
 		(u64)le16_to_cpu(resp.hwrm_intf_minor) << 32 |
 		(u64)le16_to_cpu(resp.hwrm_intf_build) << 16 |
drivers/infiniband/hw/bnxt_re/qplib_rcfw.c +1 -1
index de90691..f610199 100644
@@ -593,7 +593,7 @@ int bnxt_qplib_alloc_rcfw_channel(struct bnxt_qplib_res *res,
 			"HW channel CREQ allocation failed\n");
 		goto fail;
 	}
-	if (ctx->hwrm_intf_ver < HWRM_VERSION_RCFW_CMDQ_DEPTH_CHECK)
+	if (res->cctx->hwrm_intf_ver < HWRM_VERSION_RCFW_CMDQ_DEPTH_CHECK)
 		rcfw->cmdq_depth = BNXT_QPLIB_CMDQE_MAX_CNT_256;
 	else
 		rcfw->cmdq_depth = BNXT_QPLIB_CMDQE_MAX_CNT_8192;
Signature · 1 line
2.5.5
author maintainer
#11 lore raw

PATCH 5/6 RDMA/bnxt_re: Reorg the bar mapping

In reply to Selvin Xavier

Reorganize the code for allocation and mapping of Doorbell pages. Implements new HW command to get the BAR length used by L2 driver. These changes are used by the future patch which maps the WC Doorbell pages.

Also, introduced a new lock dpi_tbl_lock for synchronize the DB page allocation between differnt users.

---
 drivers/infiniband/hw/bnxt_re/ib_verbs.c   |   7 +-
 drivers/infiniband/hw/bnxt_re/main.c       |  75 +++++++++++-
 drivers/infiniband/hw/bnxt_re/qplib_fp.c   |   4 +-
 drivers/infiniband/hw/bnxt_re/qplib_rcfw.h |   1 +
 drivers/infiniband/hw/bnxt_re/qplib_res.c  | 189 +++++++++++++++++++----------
 drivers/infiniband/hw/bnxt_re/qplib_res.h  |  30 +++--
 drivers/infiniband/hw/bnxt_re/qplib_sp.c   |   3 +
 drivers/infiniband/hw/bnxt_re/qplib_sp.h   |   1 +
 8 files changed, 229 insertions(+), 81 deletions(-)
drivers/infiniband/hw/bnxt_re/ib_verbs.c +3 -4
index c9d134c..e2468b8 100644
@@ -604,8 +604,8 @@ int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 			 * ibv_devinfo and family of application when DPIs
 			 * are depleted.
 			 */
-			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res.dpi_tbl,
-						 &ucntx->dpi, ucntx)) {
+			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res,
+						 &ucntx->dpi, ucntx, BNXT_QPLIB_DPI_TYPE_UC)) {
 				rc = -ENOMEM;
 				goto dbfail;
 			}
@@ -4074,8 +4074,7 @@ void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx)
 		/* Free DPI only if this is the first PD allocated by the
 		 * application and mark the context dpi as NULL
 		 */
-		bnxt_qplib_dealloc_dpi(&rdev->qplib_res,
-				       &rdev->qplib_res.dpi_tbl, &uctx->dpi);
+		bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->dpi);
 		uctx->dpi.dbr = NULL;
 	}
 }
drivers/infiniband/hw/bnxt_re/main.c +71 -4

118 lines not on this page · raw message

drivers/infiniband/hw/bnxt_re/qplib_fp.c +2 -2
index ade858f..719bdc6 100644
@@ -655,7 +655,7 @@ int bnxt_qplib_create_srq(struct bnxt_qplib_res *res,
 	srq->dbinfo.xid = srq->id;
 	srq->dbinfo.db = srq->dpi->dbr;
 	srq->dbinfo.max_slot = 1;
-	srq->dbinfo.priv_db = res->dpi_tbl.dbr_bar_reg_iomem;
+	srq->dbinfo.priv_db = res->dpi_tbl.priv_db;
 	if (srq->threshold)
 		bnxt_qplib_armen_db(&srq->dbinfo, DBC_DBC_TYPE_SRQ_ARMENA);
 	srq->arm_req = false;
@@ -2099,7 +2099,7 @@ int bnxt_qplib_create_cq(struct bnxt_qplib_res *res, struct bnxt_qplib_cq *cq)
 	cq->dbinfo.hwq = &cq->hwq;
 	cq->dbinfo.xid = cq->id;
 	cq->dbinfo.db = cq->dpi->dbr;
-	cq->dbinfo.priv_db = res->dpi_tbl.dbr_bar_reg_iomem;
+	cq->dbinfo.priv_db = res->dpi_tbl.priv_db;
 
 	bnxt_qplib_armen_db(&cq->dbinfo, DBC_DBC_TYPE_CQ_ARMENA);
 
drivers/infiniband/hw/bnxt_re/qplib_rcfw.h +1 -0
index dd56514..4d1ccf1 100644
@@ -111,6 +111,7 @@ static inline u32 bnxt_qplib_set_cmd_slots(struct cmdq_base *req)
 #define RCFW_BLOCKED_CMD_WAIT_COUNT	20000000UL /* 20 sec */
 
 #define HWRM_VERSION_RCFW_CMDQ_DEPTH_CHECK 0x1000900020011ULL
+#define HWRM_VERSION_DEV_ATTR_MAX_DPI  0x1000A0000000DULL
 
 /* Crsq buf is 1024-Byte */
 struct bnxt_qplib_crsbe {
drivers/infiniband/hw/bnxt_re/qplib_res.c +125 -64

253 lines not on this page · raw message

drivers/infiniband/hw/bnxt_re/qplib_res.h +23 -7
index 982e2c9..95b1d6c 100644
@@ -56,8 +56,12 @@ struct bnxt_qplib_chip_ctx {
 	u8	chip_metal;
 	u16	hw_stats_size;
 	struct bnxt_qplib_drv_modes modes;
+	u64 hwrm_intf_ver;
 };
 
+#define BNXT_QPLIB_DBR_PF_DB_OFFSET     0x10000
+#define BNXT_QPLIB_DBR_VF_DB_OFFSET     0x4000
+
 #define PTR_CNT_PER_PG		(PAGE_SIZE / sizeof(void *))
 #define PTR_MAX_IDX_PER_PG	(PTR_CNT_PER_PG - 1)
 #define PTR_PG(x)		(((x) & ~PTR_MAX_IDX_PER_PG) / PTR_CNT_PER_PG)
@@ -109,6 +113,7 @@ enum bnxt_qplib_hwrm_pg_size {
 struct bnxt_qplib_reg_desc {
 	u8		bar_id;
 	resource_size_t	bar_base;
+	unsigned long	offset;
 	void __iomem	*bar_reg;
 	size_t		len;
 };
@@ -185,18 +190,26 @@ struct bnxt_qplib_sgid_tbl {
 	u8				*vlan;
 };
 
+enum {
+	BNXT_QPLIB_DPI_TYPE_KERNEL      = 0,
+	BNXT_QPLIB_DPI_TYPE_UC          = 1,
+};
+
 struct bnxt_qplib_dpi {
 	u32				dpi;
+	u32				bit;
 	void __iomem			*dbr;
 	u64				umdbr;
+	u8				type;
 };
 
 struct bnxt_qplib_dpi_tbl {
 	void				**app_tbl;
 	unsigned long			*tbl;
 	u16				max;
-	void __iomem			*dbr_bar_reg_iomem;
-	u64				unmapped_dbr;
+	struct bnxt_qplib_reg_desc	ucreg; /* Hold entire DB bar. */
+	struct bnxt_qplib_reg_desc	wcreg;
+	void __iomem			*priv_db;
 };
 
 struct bnxt_qplib_stats {
@@ -241,7 +254,6 @@ struct bnxt_qplib_ctx {
 	struct bnxt_qplib_tqm_ctx	tqm_ctx;
 	struct bnxt_qplib_stats		stats;
 	struct bnxt_qplib_vf_res	vf_res;
-	u64				hwrm_intf_ver;
 };
 
 struct bnxt_qplib_res {
@@ -253,6 +265,8 @@ struct bnxt_qplib_res {
 	struct bnxt_qplib_pd_tbl	pd_tbl;
 	struct bnxt_qplib_sgid_tbl	sgid_tbl;
 	struct bnxt_qplib_dpi_tbl	dpi_tbl;
+	/* To protect the dpi table bit map */
+	struct mutex                    dpi_tbl_lock;
 	bool				prio;
 	bool                            is_vf;
 };
@@ -344,11 +358,10 @@ int bnxt_qplib_alloc_pd(struct bnxt_qplib_pd_tbl *pd_tbl,
 int bnxt_qplib_dealloc_pd(struct bnxt_qplib_res *res,
 			  struct bnxt_qplib_pd_tbl *pd_tbl,
 			  struct bnxt_qplib_pd *pd);
-int bnxt_qplib_alloc_dpi(struct bnxt_qplib_dpi_tbl *dpit,
-			 struct bnxt_qplib_dpi     *dpi,
-			 void                      *app);
+int bnxt_qplib_alloc_dpi(struct bnxt_qplib_res *res,
+			 struct bnxt_qplib_dpi *dpi,
+			 void *app, u8 type);
 int bnxt_qplib_dealloc_dpi(struct bnxt_qplib_res *res,
-			   struct bnxt_qplib_dpi_tbl *dpi_tbl,
 			   struct bnxt_qplib_dpi *dpi);
 void bnxt_qplib_cleanup_res(struct bnxt_qplib_res *res);
 int bnxt_qplib_init_res(struct bnxt_qplib_res *res);
@@ -361,6 +374,9 @@ void bnxt_qplib_free_ctx(struct bnxt_qplib_res *res,
 int bnxt_qplib_alloc_ctx(struct bnxt_qplib_res *res,
 			 struct bnxt_qplib_ctx *ctx,
 			 bool virt_fn, bool is_p5);
+int bnxt_qplib_map_db_bar(struct bnxt_qplib_res *res);
+void bnxt_qplib_unmap_db_bar(struct bnxt_qplib_res *res);
+
 int bnxt_qplib_determine_atomics(struct pci_dev *dev);
 
 static inline void bnxt_qplib_hwq_incr_prod(struct bnxt_qplib_hwq *hwq, u32 cnt)
drivers/infiniband/hw/bnxt_re/qplib_sp.c +3 -0
index 1714a1e..fe3a22d 100644
@@ -170,6 +170,9 @@ int bnxt_qplib_get_dev_attr(struct bnxt_qplib_rcfw *rcfw,
 		attr->tqm_alloc_reqs[i * 4 + 3] = *(++tqm_alloc);
 	}
 
+	if (rcfw->res->cctx->hwrm_intf_ver >= HWRM_VERSION_DEV_ATTR_MAX_DPI)
+		attr->max_dpi = le32_to_cpu(sb->max_dpi);
+
 	attr->is_atomic = bnxt_qplib_is_atomic_cap(rcfw);
 bail:
 	bnxt_qplib_rcfw_free_sbuf(rcfw, sbuf);
drivers/infiniband/hw/bnxt_re/qplib_sp.h +1 -0
index 5de87465..0efd464 100644
@@ -72,6 +72,7 @@ struct bnxt_qplib_dev_attr {
 	u8				tqm_alloc_reqs[MAX_TQM_ALLOC_REQ];
 	bool				is_atomic;
 	u16                             dev_cap_flags;
+	u32                             max_dpi;
 };
 
 struct bnxt_qplib_pd {
Signature · 1 line
2.5.5
author maintainer
#12 lore raw

PATCH 6/6 RDMA/bnxt_re: Enable low latency push

In reply to Selvin Xavier

Enabling the low latency push in Gen P5 adapters for small packets. This is supported only for the user space QPs. Introduce new mmap flag for write combine buffers. Allocate separate Write Combine pages from BAR when Low latency push mode is enabled in HW.

---
 drivers/infiniband/hw/bnxt_re/ib_verbs.c  | 43 +++++++++++++++++++++++++++++++
 drivers/infiniband/hw/bnxt_re/ib_verbs.h  |  4 ++-
 drivers/infiniband/hw/bnxt_re/main.c      | 10 +++++++
 drivers/infiniband/hw/bnxt_re/qplib_res.c |  3 +++
 drivers/infiniband/hw/bnxt_re/qplib_res.h |  3 ++-
 include/uapi/rdma/bnxt_re-abi.h           |  9 +++++++
 6 files changed, 70 insertions(+), 2 deletions(-)
drivers/infiniband/hw/bnxt_re/ib_verbs.c +43 -0
index e2468b8..591ee82 100644
@@ -566,6 +566,9 @@ int bnxt_re_dealloc_pd(struct ib_pd *ib_pd, struct ib_udata *udata)
 
 	if (udata) {
 		rdma_user_mmap_entry_remove(pd->pd_db_mmap);
+		if (pd->pd_wcdb_mmap)
+			rdma_user_mmap_entry_remove(pd->pd_wcdb_mmap);
+		pd->pd_wcdb_mmap = NULL;
 		pd->pd_db_mmap = NULL;
 	}
 
@@ -597,6 +600,7 @@ int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 	}
 
 	if (udata) {
+		struct bnxt_qplib_chip_ctx *cctx = rdev->chip_ctx;
 		struct bnxt_re_pd_resp resp = {};
 
 		if (!ucntx->dpi.dbr) {
@@ -606,9 +610,23 @@ int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 			 */
 			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res,
 						 &ucntx->dpi, ucntx, BNXT_QPLIB_DPI_TYPE_UC)) {
+				dev_err(rdev_to_dev(rdev), "DP alloc failed");
 				rc = -ENOMEM;
 				goto dbfail;
 			}
+			if (cctx->modes.db_push) {
+				rc = bnxt_qplib_alloc_dpi(&rdev->qplib_res, &ucntx->wcdpi,
+							  ucntx, BNXT_QPLIB_DPI_TYPE_WC);
+				if (rc) {
+					dev_err(rdev_to_dev(rdev), "push dp alloc failed");
+					bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &ucntx->dpi);
+					ucntx->dpi.dbr = NULL;
+					rc = -ENOMEM;
+					goto dbfail;
+				}
+				resp.wcdpi = ucntx->wcdpi.dpi;
+				resp.comp_mask = BNXT_RE_COMP_MASK_PD_HAS_WC_DPI;
+			}
 		}
 
 		resp.pdid = pd->qplib_pd.id;
@@ -625,6 +643,16 @@ int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 			goto dbfail;
 		}
 
+		pd->pd_wcdb_mmap = bnxt_re_mmap_entry_insert(ucntx, (u64)ucntx->wcdpi.umdbr,
+							     BNXT_RE_MMAP_WC_DB, &resp.wcdbr);
+
+		if (!pd->pd_wcdb_mmap) {
+			ibdev_err(&rdev->ibdev,
+				  "Failed to insert mmap entry\n");
+			rc = -ENOMEM;
+			goto dbfail;
+		}
+
 		rc = ib_copy_to_udata(udata, &resp, min(sizeof(resp), udata->outlen));
 		if (rc) {
 			ibdev_err(&rdev->ibdev,
@@ -4035,6 +4063,9 @@ int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata)
 	resp.comp_mask |= BNXT_RE_UCNTX_CMASK_HAVE_MODE;
 	resp.mode = rdev->chip_ctx->modes.wqe_mode;
 
+	if (rdev->chip_ctx->modes.db_push)
+		resp.comp_mask |= BNXT_RE_UCNTX_CMASK_WC_DPI_ENABLED;
+
 	uctx->shpage_mmap = bnxt_re_mmap_entry_insert(uctx, 0, BNXT_RE_MMAP_SH_PAGE, NULL);
 	if (!uctx->shpage_mmap) {
 		ibdev_err(ibdev, "Failed to create mmap entry for shared page\n");
@@ -4074,6 +4105,10 @@ void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx)
 		/* Free DPI only if this is the first PD allocated by the
 		 * application and mark the context dpi as NULL
 		 */
+		if (uctx->wcdpi.dbr) {
+			bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->wcdpi);
+			uctx->wcdpi.dbr = NULL;
+		}
 		bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->dpi);
 		uctx->dpi.dbr = NULL;
 	}
@@ -4103,6 +4138,14 @@ int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
 				  rdma_entry);
 
 	switch (bnxt_entry->mmap_flag) {
+	case BNXT_RE_MMAP_WC_DB:
+		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
+		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
+					pgprot_writecombine(vma->vm_page_prot),
+					rdma_entry);
+		if (ret)
+			ibdev_err(&rdev->ibdev, "Failed to map WC DB");
+		break;
 	case BNXT_RE_MMAP_UC_DB:
 		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
 		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
drivers/infiniband/hw/bnxt_re/ib_verbs.h +3 -1
index dcd31ae..faf1481 100644
@@ -61,6 +61,7 @@ struct bnxt_re_pd {
 	struct bnxt_qplib_pd	qplib_pd;
 	struct bnxt_re_fence_data fence;
 	struct rdma_user_mmap_entry *pd_db_mmap;
+	struct rdma_user_mmap_entry *pd_wcdb_mmap;
 };
 
 struct bnxt_re_ah {
@@ -135,6 +136,7 @@ struct bnxt_re_ucontext {
 	struct ib_ucontext      ib_uctx;
 	struct bnxt_re_dev	*rdev;
 	struct bnxt_qplib_dpi	dpi;
+	struct bnxt_qplib_dpi   wcdpi;
 	void			*shpg;
 	spinlock_t		sh_lock;	/* protect shpg */
 	struct rdma_user_mmap_entry *shpage_mmap;
@@ -143,6 +145,7 @@ struct bnxt_re_ucontext {
 enum bnxt_re_mmap_flag {
 	BNXT_RE_MMAP_SH_PAGE,
 	BNXT_RE_MMAP_UC_DB,
+	BNXT_RE_MMAP_WC_DB,
 };
 
 struct bnxt_re_user_mmap_entry {
@@ -228,7 +231,6 @@ void bnxt_re_dealloc_ucontext(struct ib_ucontext *context);
 int bnxt_re_mmap(struct ib_ucontext *context, struct vm_area_struct *vma);
 void bnxt_re_mmap_free(struct rdma_user_mmap_entry *rdma_entry);
 
-
 unsigned long bnxt_re_lock_cqs(struct bnxt_re_qp *qp);
 void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp, unsigned long flags);
 #endif /* __BNXT_RE_IB_VERBS_H__ */
drivers/infiniband/hw/bnxt_re/main.c +10 -0
index cfd3708..824dc3b 100644
@@ -117,6 +117,11 @@ static void bnxt_re_set_db_offset(struct bnxt_re_dev *rdev)
 	 * in such cases and DB-push will be disabled.
 	 */
 	barlen = pci_resource_len(res->pdev, RCFW_DBR_PCI_BAR_REGION);
+	if (cctx->modes.db_push && l2db_len && en_dev->l2_db_size != barlen) {
+		res->dpi_tbl.wcreg.offset = en_dev->l2_db_size;
+		dev_info(rdev_to_dev(rdev),
+			 "Low latency framework is enabled\n");
+	}
 }
 
 static void bnxt_re_set_drv_mode(struct bnxt_re_dev *rdev, u8 mode)
@@ -430,6 +435,11 @@ int bnxt_re_hwrm_qcaps(struct bnxt_re_dev *rdev)
 			"Failed to query capabilities, rc = %#x", rc);
 		return rc;
 	}
+	cctx->modes.db_push = le32_to_cpu(resp.flags) & FUNC_QCAPS_RESP_FLAGS_WCB_PUSH_MODE;
+
+	if (cctx->modes.db_push)
+		ibdev_dbg(&rdev->ibdev, "DB push enabled");
+
 	return 0;
 }
 
drivers/infiniband/hw/bnxt_re/qplib_res.c +3 -0
index bb3087d..5483778 100644
@@ -741,6 +741,9 @@ int bnxt_qplib_alloc_dpi(struct bnxt_qplib_res *res,
 		dpi->dbr = dpit->priv_db;
 		dpi->dpi = dpi->bit;
 		break;
+	case BNXT_QPLIB_DPI_TYPE_WC:
+		dpi->dbr = ioremap_wc(umaddr, PAGE_SIZE);
+		break;
 	default:
 		dpi->dbr = ioremap(umaddr, PAGE_SIZE);
 		break;
drivers/infiniband/hw/bnxt_re/qplib_res.h +2 -1
index 95b1d6c..dc39f67 100644
@@ -47,7 +47,7 @@ extern const struct bnxt_qplib_gid bnxt_qplib_gid_zero;
 
 struct bnxt_qplib_drv_modes {
 	u8	wqe_mode;
-	/* Other modes to follow here */
+	bool db_push;
 };
 
 struct bnxt_qplib_chip_ctx {
@@ -193,6 +193,7 @@ struct bnxt_qplib_sgid_tbl {
 enum {
 	BNXT_QPLIB_DPI_TYPE_KERNEL      = 0,
 	BNXT_QPLIB_DPI_TYPE_UC          = 1,
+	BNXT_QPLIB_DPI_TYPE_WC          = 2
 };
 
 struct bnxt_qplib_dpi {
include/uapi/rdma/bnxt_re-abi.h +9 -0
index c4e9077..719f1fe 100644
@@ -51,6 +51,7 @@
 enum {
 	BNXT_RE_UCNTX_CMASK_HAVE_CCTX = 0x1ULL,
 	BNXT_RE_UCNTX_CMASK_HAVE_MODE = 0x02ULL,
+	BNXT_RE_UCNTX_CMASK_WC_DPI_ENABLED = 0x04ULL,
 };
 
 enum bnxt_re_wqe_mode {
@@ -78,10 +79,18 @@ struct bnxt_re_uctx_resp {
  * not 8 byted aligned. To avoid undesired padding in various cases we have to
  * set this struct to packed.
  */
+enum {
+	BNXT_RE_COMP_MASK_PD_HAS_WC_DPI = 0x01,
+};
+
 struct bnxt_re_pd_resp {
 	__u32 pdid;
 	__u32 dpi;
 	__u64 dbr;
+	__u64 comp_mask;
+	__u32 rsvd;
+	__u32 wcdpi;
+	__u64 wcdbr;
 } __attribute__((packed, aligned(4)));
 
 struct bnxt_re_cq_req {
Signature · 1 line
2.5.5
#13 lore raw

Re: [PATCH for-next 6/6] RDMA/bnxt_re: Enable low latency push

Quoted 5 lines · Selvin Xavier

Enabling the low latency push in Gen P5 adapters for small packets. This is supported only for the user space QPs. Introduce new mmap flag for write combine buffers. Allocate separate Write Combine pages from BAR when Low latency push mode is enabled in HW.

No new mmap flags.

This looks like it is basically the same as mlx5, I expect you'll run into all the same issues and needs as mlx5 did.

Follow the modern mlx5 design, add a new API to allocate one of these pages and have it return back the mmap cookie to use. Let userspace allocate as many as it needs.

Jason

author maintainer
#14 lore raw

Re: [PATCH for-next 6/6] RDMA/bnxt_re: Enable low latency push

Quoted 8 lines · Jason Gunthorpe
Quoted 5 lines · Selvin Xavier

Enabling the low latency push in Gen P5 adapters for small packets. This is supported only for the user space QPs. Introduce new mmap flag for write combine buffers. Allocate separate Write Combine pages from BAR when Low latency push mode is enabled in HW.

Quoted 8 lines

No new mmap flags.

This looks like it is basically the same as mlx5, I expect you'll run into all the same issues and needs as mlx5 did.

Follow the modern mlx5 design, add a new API to allocate one of these pages and have it return back the mmap cookie to use. Let userspace allocate as many as it needs.

Thanks Jason for the pointers. Will review this and rework on the patch

Quoted 2 lines

Jason

#15 lore raw
In reply to Selvin Xavier
Quoted 5 lines · Selvin Xavier

The series aims to add support for Low latency push path in some of the bnxt devices. The low latency implementation is supported only for the user applications. Also, the code is modified to use common mmap helper functions exported by IB core.

What does it mean "low latency push"?

Thanks

Quoted 33 lines

User library changes are added in the pull request
https://github.com/linux-rdma/rdma-core/pull/1321

Please review.

Thanks,
Selvin Xavier

Selvin Xavier (6):
RDMA/bnxt_re: Use the common mmap helper functions RDMA/bnxt_re: Add disassociate ucontext support RDMA/bnxt_re: Query function capabilities from firmware RDMA/bnxt_re: Move the interface version to chip context structure RDMA/bnxt_re: Reorg the bar mapping
RDMA/bnxt_re: Enable low latency push

 drivers/infiniband/hw/bnxt_re/ib_verbs.c   | 160 +++++++++++++++++++++---
 drivers/infiniband/hw/bnxt_re/ib_verbs.h   |  17 +++
 drivers/infiniband/hw/bnxt_re/main.c       | 123 +++++++++++++++++-
 drivers/infiniband/hw/bnxt_re/qplib_fp.c   |  14 +--
 drivers/infiniband/hw/bnxt_re/qplib_rcfw.c |   2 +-
 drivers/infiniband/hw/bnxt_re/qplib_rcfw.h |   1 +
 drivers/infiniband/hw/bnxt_re/qplib_res.c  | 192 +++++++++++++++++++----------
 drivers/infiniband/hw/bnxt_re/qplib_res.h  |  33 +++--
 drivers/infiniband/hw/bnxt_re/qplib_sp.c   |   3 +
 drivers/infiniband/hw/bnxt_re/qplib_sp.h   |   1 +
 include/uapi/rdma/bnxt_re-abi.h            |   9 ++
 11 files changed, 449 insertions(+), 106 deletions(-)
Signature · 1 line
2.5.5
author maintainer
#16 lore raw
Quoted 9 lines · Leon Romanovsky
Quoted 5 lines · Selvin Xavier

The series aims to add support for Low latency push path in some of the bnxt devices. The low latency implementation is supported only for the user applications. Also, the code is modified to use common mmap helper functions exported by IB core.

What does it mean "low latency push"?

It is a hardware channel that can be used for transmitting smaller buffers (<=96 bytes) with lower latency. This is by pushing the data to HW buffers directly from the host (instead of DMA by the HW to get the data to be transmitted). Hence the name "low latency push".

Quoted 38 lines

Thanks

Quoted 33 lines

User library changes are added in the pull request
https://github.com/linux-rdma/rdma-core/pull/1321

Please review.

Thanks,
Selvin Xavier

Selvin Xavier (6):
RDMA/bnxt_re: Use the common mmap helper functions RDMA/bnxt_re: Add disassociate ucontext support RDMA/bnxt_re: Query function capabilities from firmware RDMA/bnxt_re: Move the interface version to chip context structure RDMA/bnxt_re: Reorg the bar mapping
RDMA/bnxt_re: Enable low latency push

 drivers/infiniband/hw/bnxt_re/ib_verbs.c   | 160 +++++++++++++++++++++---
 drivers/infiniband/hw/bnxt_re/ib_verbs.h   |  17 +++
 drivers/infiniband/hw/bnxt_re/main.c       | 123 +++++++++++++++++-
 drivers/infiniband/hw/bnxt_re/qplib_fp.c   |  14 +--
 drivers/infiniband/hw/bnxt_re/qplib_rcfw.c |   2 +-
 drivers/infiniband/hw/bnxt_re/qplib_rcfw.h |   1 +
 drivers/infiniband/hw/bnxt_re/qplib_res.c  | 192 +++++++++++++++++++----------
 drivers/infiniband/hw/bnxt_re/qplib_res.h  |  33 +++--
 drivers/infiniband/hw/bnxt_re/qplib_sp.c   |   3 +
 drivers/infiniband/hw/bnxt_re/qplib_sp.h   |   1 +
 include/uapi/rdma/bnxt_re-abi.h            |   9 ++
 11 files changed, 449 insertions(+), 106 deletions(-)

--
2.5.5