RDMA/bnxt_re: Share a page to expose per CQ info with userspace
authorSelvin Xavier <selvin.xavier@broadcom.com>
Thu, 14 Dec 2023 06:31:24 +0000 (22:31 -0800)
committerLeon Romanovsky <leon@kernel.org>
Sun, 17 Dec 2023 13:35:58 +0000 (15:35 +0200)
Gen P7 adapters needs to share a toggle bits information received
in kernel driver with the user space. User space needs this
info during the request notify call back to arm the CQ.

User space application can get this page using the
UAPI routines. Library will mmap this page and get the
toggle bits to be used in the next ARM Doorbell.

Uses a hash list to map the CQ structure from the CQ ID.
CQ structure is retrieved from the hash list while the
library calls the UAPI routine to get the toggle page
mapping. Currently the full page is mapped per CQ. This
can be optimized to enable multiple CQs from the same
application share the same page and different offsets
in the page.

Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Link: https://lore.kernel.org/r/1702535484-26844-3-git-send-email-selvin.xavier@broadcom.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
drivers/infiniband/hw/bnxt_re/bnxt_re.h
drivers/infiniband/hw/bnxt_re/ib_verbs.c
drivers/infiniband/hw/bnxt_re/ib_verbs.h
drivers/infiniband/hw/bnxt_re/main.c
drivers/infiniband/hw/bnxt_re/qplib_res.h
include/uapi/rdma/bnxt_re-abi.h

index 9fd9849ebdd1428813e5d0b0a90b217629f0b077..9dca451ed5221a7bab7eb303258895974017911a 100644 (file)
@@ -41,6 +41,7 @@
 #define __BNXT_RE_H__
 #include <rdma/uverbs_ioctl.h>
 #include "hw_counters.h"
+#include <linux/hashtable.h>
 #define ROCE_DRV_MODULE_NAME           "bnxt_re"
 
 #define BNXT_RE_DESC   "Broadcom NetXtreme-C/E RoCE Driver"
@@ -135,6 +136,7 @@ struct bnxt_re_pacing {
 #define BNXT_RE_DB_FIFO_ROOM_SHIFT 15
 #define BNXT_RE_GRC_FIFO_REG_BASE 0x2000
 
+#define MAX_CQ_HASH_BITS               (16)
 struct bnxt_re_dev {
        struct ib_device                ibdev;
        struct list_head                list;
@@ -189,6 +191,7 @@ struct bnxt_re_dev {
        struct bnxt_re_pacing pacing;
        struct work_struct dbq_fifo_check_work;
        struct delayed_work dbq_pacing_work;
+       DECLARE_HASHTABLE(cq_hash, MAX_CQ_HASH_BITS);
 };
 
 #define to_bnxt_re_dev(ptr, member)    \
index 758ea02e1d139c34effc9dcd6b2e908eb944305e..7213dc7574d01fb5e7fb4b65fcfdcae142067ec4 100644 (file)
@@ -50,6 +50,7 @@
 #include <rdma/ib_mad.h>
 #include <rdma/ib_cache.h>
 #include <rdma/uverbs_ioctl.h>
+#include <linux/hashtable.h>
 
 #include "bnxt_ulp.h"
 
@@ -2910,14 +2911,20 @@ int bnxt_re_post_recv(struct ib_qp *ib_qp, const struct ib_recv_wr *wr,
 /* Completion Queues */
 int bnxt_re_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
 {
-       struct bnxt_re_cq *cq;
+       struct bnxt_qplib_chip_ctx *cctx;
        struct bnxt_qplib_nq *nq;
        struct bnxt_re_dev *rdev;
+       struct bnxt_re_cq *cq;
 
        cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
        rdev = cq->rdev;
        nq = cq->qplib_cq.nq;
+       cctx = rdev->chip_ctx;
 
+       if (cctx->modes.toggle_bits & BNXT_QPLIB_CQ_TOGGLE_BIT) {
+               free_page((unsigned long)cq->uctx_cq_page);
+               hash_del(&cq->hash_entry);
+       }
        bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
        ib_umem_release(cq->umem);
 
@@ -2935,10 +2942,11 @@ int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
        struct bnxt_re_ucontext *uctx =
                rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
        struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
-       int rc, entries;
-       int cqe = attr->cqe;
+       struct bnxt_qplib_chip_ctx *cctx;
        struct bnxt_qplib_nq *nq = NULL;
+       int rc = -ENOMEM, entries;
        unsigned int nq_alloc_cnt;
+       int cqe = attr->cqe;
        u32 active_cqs;
 
        if (attr->flags)
@@ -2951,6 +2959,7 @@ int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
        }
 
        cq->rdev = rdev;
+       cctx = rdev->chip_ctx;
        cq->qplib_cq.cq_handle = (u64)(unsigned long)(&cq->qplib_cq);
 
        entries = bnxt_re_init_depth(cqe + 1, uctx);
@@ -3012,22 +3021,32 @@ int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
        spin_lock_init(&cq->cq_lock);
 
        if (udata) {
-               struct bnxt_re_cq_resp resp;
-
+               struct bnxt_re_cq_resp resp = {};
+
+               if (cctx->modes.toggle_bits & BNXT_QPLIB_CQ_TOGGLE_BIT) {
+                       hash_add(rdev->cq_hash, &cq->hash_entry, cq->qplib_cq.id);
+                       /* Allocate a page */
+                       cq->uctx_cq_page = (void *)get_zeroed_page(GFP_KERNEL);
+                       if (!cq->uctx_cq_page)
+                               goto c2fail;
+                       resp.comp_mask |= BNXT_RE_CQ_TOGGLE_PAGE_SUPPORT;
+               }
                resp.cqid = cq->qplib_cq.id;
                resp.tail = cq->qplib_cq.hwq.cons;
                resp.phase = cq->qplib_cq.period;
                resp.rsvd = 0;
-               rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
+               rc = ib_copy_to_udata(udata, &resp, min(sizeof(resp), udata->outlen));
                if (rc) {
                        ibdev_err(&rdev->ibdev, "Failed to copy CQ udata");
                        bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
-                       goto c2fail;
+                       goto free_mem;
                }
        }
 
        return 0;
 
+free_mem:
+       free_page((unsigned long)cq->uctx_cq_page);
 c2fail:
        ib_umem_release(cq->umem);
 fail:
@@ -4214,6 +4233,19 @@ void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx)
        }
 }
 
+static struct bnxt_re_cq *bnxt_re_search_for_cq(struct bnxt_re_dev *rdev, u32 cq_id)
+{
+       struct bnxt_re_cq *cq = NULL, *tmp_cq;
+
+       hash_for_each_possible(rdev->cq_hash, tmp_cq, hash_entry, cq_id) {
+               if (tmp_cq->qplib_cq.id == cq_id) {
+                       cq = tmp_cq;
+                       break;
+               }
+       }
+       return cq;
+}
+
 /* Helper function to mmap the virtual memory from user app */
 int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
 {
@@ -4442,10 +4474,12 @@ static int UVERBS_HANDLER(BNXT_RE_METHOD_GET_TOGGLE_MEM)(struct uverbs_attr_bund
        struct bnxt_re_ucontext *uctx;
        struct ib_ucontext *ib_uctx;
        struct bnxt_re_dev *rdev;
+       struct bnxt_re_cq *cq;
        u64 mem_offset;
        u64 addr = 0;
        u32 length;
        u32 offset;
+       u32 cq_id;
        int err;
 
        ib_uctx = ib_uverbs_get_ucontext(attrs);
@@ -4461,6 +4495,19 @@ static int UVERBS_HANDLER(BNXT_RE_METHOD_GET_TOGGLE_MEM)(struct uverbs_attr_bund
 
        switch (res_type) {
        case BNXT_RE_CQ_TOGGLE_MEM:
+               err = uverbs_copy_from(&cq_id, attrs, BNXT_RE_TOGGLE_MEM_RES_ID);
+               if (err)
+                       return err;
+
+               cq = bnxt_re_search_for_cq(rdev, cq_id);
+               if (!cq)
+                       return -EINVAL;
+
+               length = PAGE_SIZE;
+               addr = (u64)cq->uctx_cq_page;
+               mmap_flag = BNXT_RE_MMAP_TOGGLE_PAGE;
+               offset = 0;
+               break;
        case BNXT_RE_SRQ_TOGGLE_MEM:
                break;
 
index da3fe018f2557313cf07515abb789e9e58e1eb83..b267d6d5975f7304da54d5a6b7ba46fed4bb6f9e 100644 (file)
@@ -108,6 +108,8 @@ struct bnxt_re_cq {
        struct ib_umem          *umem;
        struct ib_umem          *resize_umem;
        int                     resize_cqe;
+       void                    *uctx_cq_page;
+       struct hlist_node       hash_entry;
 };
 
 struct bnxt_re_mr {
index 7f4f6db1392c8e360d91f67ad46ce4ce1fad70c6..eb03ebad2e5adbcd33d3dd7d6b0c4c7d614ed1a6 100644 (file)
@@ -54,6 +54,7 @@
 #include <rdma/ib_user_verbs.h>
 #include <rdma/ib_umem.h>
 #include <rdma/ib_addr.h>
+#include <linux/hashtable.h>
 
 #include "bnxt_ulp.h"
 #include "roce_hsi.h"
@@ -136,6 +137,8 @@ static void bnxt_re_set_drv_mode(struct bnxt_re_dev *rdev, u8 mode)
        if (bnxt_re_hwrm_qcaps(rdev))
                dev_err(rdev_to_dev(rdev),
                        "Failed to query hwrm qcaps\n");
+       if (bnxt_qplib_is_chip_gen_p7(rdev->chip_ctx))
+               cctx->modes.toggle_bits |= BNXT_QPLIB_CQ_TOGGLE_BIT;
 }
 
 static void bnxt_re_destroy_chip_ctx(struct bnxt_re_dev *rdev)
@@ -1206,9 +1209,13 @@ static int bnxt_re_cqn_handler(struct bnxt_qplib_nq *nq,
 {
        struct bnxt_re_cq *cq = container_of(handle, struct bnxt_re_cq,
                                             qplib_cq);
+       u32 *cq_ptr;
 
        if (cq->ib_cq.comp_handler) {
-               /* Lock comp_handler? */
+               if (cq->uctx_cq_page) {
+                       cq_ptr = (u32 *)cq->uctx_cq_page;
+                       *cq_ptr = cq->qplib_cq.toggle;
+               }
                (*cq->ib_cq.comp_handler)(&cq->ib_cq, cq->ib_cq.cq_context);
        }
 
@@ -1730,6 +1737,7 @@ static int bnxt_re_dev_init(struct bnxt_re_dev *rdev, u8 wqe_mode)
                 */
                bnxt_re_vf_res_config(rdev);
        }
+       hash_init(rdev->cq_hash);
 
        return 0;
 free_sctx:
index 382d89fa7d16d682c1594441fbe456f5c5e98fb4..61628f7f125326cfe1eb478e551388666673b36b 100644 (file)
@@ -55,6 +55,12 @@ struct bnxt_qplib_drv_modes {
        u8      wqe_mode;
        bool db_push;
        bool dbr_pacing;
+       u32 toggle_bits;
+};
+
+enum bnxt_re_toggle_modes {
+       BNXT_QPLIB_CQ_TOGGLE_BIT = 0x1,
+       BNXT_QPLIB_SRQ_TOGGLE_BIT = 0x2,
 };
 
 struct bnxt_qplib_chip_ctx {
index 9b9eb10cb5e5e6cb8d359798a3ea39b79c7ff6a9..c0c34aca90ecb3a45eb1579e7df7f74bd2c580e1 100644 (file)
@@ -102,11 +102,16 @@ struct bnxt_re_cq_req {
        __aligned_u64 cq_handle;
 };
 
+enum bnxt_re_cq_mask {
+       BNXT_RE_CQ_TOGGLE_PAGE_SUPPORT = 0x1,
+};
+
 struct bnxt_re_cq_resp {
        __u32 cqid;
        __u32 tail;
        __u32 phase;
        __u32 rsvd;
+       __aligned_u64 comp_mask;
 };
 
 struct bnxt_re_resize_cq_req {