RDMA/i40iw: Add base memory management extensions
authorIsmail, Mustafa <mustafa.ismail@intel.com>
Mon, 18 Apr 2016 15:33:06 +0000 (10:33 -0500)
committerDoug Ledford <dledford@redhat.com>
Thu, 28 Apr 2016 20:32:54 +0000 (16:32 -0400)
Implement fast register mr, Local invalidate, send with
invalidate and RDMA read with invalidate.

Signed-off-by: Mustafa Ismail <mustafa.ismail@intel.com>
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/hw/i40iw/i40iw_ctrl.c
drivers/infiniband/hw/i40iw/i40iw_type.h
drivers/infiniband/hw/i40iw/i40iw_verbs.c
drivers/infiniband/hw/i40iw/i40iw_verbs.h

index f05802bf6ca04dbbe5e60bc00b044dd1faab4cb1..437cb8603540f9b5fe3037357be3dbcef580b6e6 100644 (file)
@@ -2908,6 +2908,65 @@ static enum i40iw_status_code i40iw_sc_mw_alloc(
        return 0;
 }
 
+/**
+ * i40iw_sc_mr_fast_register - Posts RDMA fast register mr WR to iwarp qp
+ * @qp: sc qp struct
+ * @info: fast mr info
+ * @post_sq: flag for cqp db to ring
+ */
+enum i40iw_status_code i40iw_sc_mr_fast_register(
+                               struct i40iw_sc_qp *qp,
+                               struct i40iw_fast_reg_stag_info *info,
+                               bool post_sq)
+{
+       u64 temp, header;
+       u64 *wqe;
+       u32 wqe_idx;
+
+       wqe = i40iw_qp_get_next_send_wqe(&qp->qp_uk, &wqe_idx, I40IW_QP_WQE_MIN_SIZE,
+                                        0, info->wr_id);
+       if (!wqe)
+               return I40IW_ERR_QP_TOOMANY_WRS_POSTED;
+
+       i40iw_debug(qp->dev, I40IW_DEBUG_MR, "%s: wr_id[%llxh] wqe_idx[%04d] location[%p]\n",
+                   __func__, info->wr_id, wqe_idx,
+                   &qp->qp_uk.sq_wrtrk_array[wqe_idx].wrid);
+       temp = (info->addr_type == I40IW_ADDR_TYPE_VA_BASED) ? (uintptr_t)info->va : info->fbo;
+       set_64bit_val(wqe, 0, temp);
+
+       temp = RS_64(info->first_pm_pbl_index >> 16, I40IWQPSQ_FIRSTPMPBLIDXHI);
+       set_64bit_val(wqe,
+                     8,
+                     LS_64(temp, I40IWQPSQ_FIRSTPMPBLIDXHI) |
+                     LS_64(info->reg_addr_pa >> I40IWQPSQ_PBLADDR_SHIFT, I40IWQPSQ_PBLADDR));
+
+       set_64bit_val(wqe,
+                     16,
+                     info->total_len |
+                     LS_64(info->first_pm_pbl_index, I40IWQPSQ_FIRSTPMPBLIDXLO));
+
+       header = LS_64(info->stag_key, I40IWQPSQ_STAGKEY) |
+                LS_64(info->stag_idx, I40IWQPSQ_STAGINDEX) |
+                LS_64(I40IWQP_OP_FAST_REGISTER, I40IWQPSQ_OPCODE) |
+                LS_64(info->chunk_size, I40IWQPSQ_LPBLSIZE) |
+                LS_64(info->page_size, I40IWQPSQ_HPAGESIZE) |
+                LS_64(info->access_rights, I40IWQPSQ_STAGRIGHTS) |
+                LS_64(info->addr_type, I40IWQPSQ_VABASEDTO) |
+                LS_64(info->read_fence, I40IWQPSQ_READFENCE) |
+                LS_64(info->local_fence, I40IWQPSQ_LOCALFENCE) |
+                LS_64(info->signaled, I40IWQPSQ_SIGCOMPL) |
+                LS_64(qp->qp_uk.swqe_polarity, I40IWQPSQ_VALID);
+
+       i40iw_insert_wqe_hdr(wqe, header);
+
+       i40iw_debug_buf(qp->dev, I40IW_DEBUG_WQE, "FAST_REG WQE",
+                       wqe, I40IW_QP_WQE_MIN_SIZE);
+
+       if (post_sq)
+               i40iw_qp_post_wr(&qp->qp_uk);
+       return 0;
+}
+
 /**
  * i40iw_sc_send_lsmm - send last streaming mode message
  * @qp: sc qp struct
@@ -4559,17 +4618,18 @@ static struct i40iw_pd_ops iw_pd_ops = {
 };
 
 static struct i40iw_priv_qp_ops iw_priv_qp_ops = {
-       i40iw_sc_qp_init,
-       i40iw_sc_qp_create,
-       i40iw_sc_qp_modify,
-       i40iw_sc_qp_destroy,
-       i40iw_sc_qp_flush_wqes,
-       i40iw_sc_qp_upload_context,
-       i40iw_sc_qp_setctx,
-       i40iw_sc_send_lsmm,
-       i40iw_sc_send_lsmm_nostag,
-       i40iw_sc_send_rtt,
-       i40iw_sc_post_wqe0,
+       .qp_init = i40iw_sc_qp_init,
+       .qp_create = i40iw_sc_qp_create,
+       .qp_modify = i40iw_sc_qp_modify,
+       .qp_destroy = i40iw_sc_qp_destroy,
+       .qp_flush_wqes = i40iw_sc_qp_flush_wqes,
+       .qp_upload_context = i40iw_sc_qp_upload_context,
+       .qp_setctx = i40iw_sc_qp_setctx,
+       .qp_send_lsmm = i40iw_sc_send_lsmm,
+       .qp_send_lsmm_nostag = i40iw_sc_send_lsmm_nostag,
+       .qp_send_rtt = i40iw_sc_send_rtt,
+       .qp_post_wqe0 = i40iw_sc_post_wqe0,
+       .iw_mr_fast_register = i40iw_sc_mr_fast_register
 };
 
 static struct i40iw_priv_cq_ops iw_priv_cq_ops = {
index 5b6a49143a92b8736877e26c69b84a19aee81825..c9261981da0b4e8f08eb3a481fee5d78829a3a7f 100644 (file)
@@ -1041,6 +1041,9 @@ struct i40iw_priv_qp_ops {
        void (*qp_send_lsmm_nostag)(struct i40iw_sc_qp *, void *, u32);
        void (*qp_send_rtt)(struct i40iw_sc_qp *, bool);
        enum i40iw_status_code (*qp_post_wqe0)(struct i40iw_sc_qp *, u8);
+       enum i40iw_status_code (*iw_mr_fast_register)(struct i40iw_sc_qp *,
+                                                     struct i40iw_fast_reg_stag_info *,
+                                                     bool);
 };
 
 struct i40iw_priv_cq_ops {
index 6359a3ebb05a680a27a9e3a9051c134d0f29c867..196d6f006392df1ec4470994c472e32f0a23b31f 100644 (file)
@@ -725,8 +725,10 @@ static struct ib_qp *i40iw_create_qp(struct ib_pd *ibpd,
        iwarp_info = &iwqp->iwarp_info;
        iwarp_info->rd_enable = true;
        iwarp_info->wr_rdresp_en = true;
-       if (!iwqp->user_mode)
+       if (!iwqp->user_mode) {
+               iwarp_info->fast_reg_en = true;
                iwarp_info->priv_mode_en = true;
+       }
        iwarp_info->ddp_ver = 1;
        iwarp_info->rdmap_ver = 1;
 
@@ -1446,6 +1448,139 @@ static int i40iw_handle_q_mem(struct i40iw_device *iwdev,
        return err;
 }
 
+/**
+ * i40iw_hw_alloc_stag - cqp command to allocate stag
+ * @iwdev: iwarp device
+ * @iwmr: iwarp mr pointer
+ */
+static int i40iw_hw_alloc_stag(struct i40iw_device *iwdev, struct i40iw_mr *iwmr)
+{
+       struct i40iw_allocate_stag_info *info;
+       struct i40iw_pd *iwpd = to_iwpd(iwmr->ibmr.pd);
+       enum i40iw_status_code status;
+       int err = 0;
+       struct i40iw_cqp_request *cqp_request;
+       struct cqp_commands_info *cqp_info;
+
+       cqp_request = i40iw_get_cqp_request(&iwdev->cqp, true);
+       if (!cqp_request)
+               return -ENOMEM;
+
+       cqp_info = &cqp_request->info;
+       info = &cqp_info->in.u.alloc_stag.info;
+       memset(info, 0, sizeof(*info));
+       info->page_size = PAGE_SIZE;
+       info->stag_idx = iwmr->stag >> I40IW_CQPSQ_STAG_IDX_SHIFT;
+       info->pd_id = iwpd->sc_pd.pd_id;
+       info->total_len = iwmr->length;
+       cqp_info->cqp_cmd = OP_ALLOC_STAG;
+       cqp_info->post_sq = 1;
+       cqp_info->in.u.alloc_stag.dev = &iwdev->sc_dev;
+       cqp_info->in.u.alloc_stag.scratch = (uintptr_t)cqp_request;
+
+       status = i40iw_handle_cqp_op(iwdev, cqp_request);
+       if (status) {
+               err = -ENOMEM;
+               i40iw_pr_err("CQP-OP MR Reg fail");
+       }
+       return err;
+}
+
+/**
+ * i40iw_alloc_mr - register stag for fast memory registration
+ * @pd: ibpd pointer
+ * @mr_type: memory for stag registrion
+ * @max_num_sg: man number of pages
+ */
+static struct ib_mr *i40iw_alloc_mr(struct ib_pd *pd,
+                                   enum ib_mr_type mr_type,
+                                   u32 max_num_sg)
+{
+       struct i40iw_pd *iwpd = to_iwpd(pd);
+       struct i40iw_device *iwdev = to_iwdev(pd->device);
+       struct i40iw_pble_alloc *palloc;
+       struct i40iw_pbl *iwpbl;
+       struct i40iw_mr *iwmr;
+       enum i40iw_status_code status;
+       u32 stag;
+       int err_code = -ENOMEM;
+
+       iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
+       if (!iwmr)
+               return ERR_PTR(-ENOMEM);
+
+       stag = i40iw_create_stag(iwdev);
+       if (!stag) {
+               err_code = -EOVERFLOW;
+               goto err;
+       }
+       iwmr->stag = stag;
+       iwmr->ibmr.rkey = stag;
+       iwmr->ibmr.lkey = stag;
+       iwmr->ibmr.pd = pd;
+       iwmr->ibmr.device = pd->device;
+       iwpbl = &iwmr->iwpbl;
+       iwpbl->iwmr = iwmr;
+       iwmr->type = IW_MEMREG_TYPE_MEM;
+       palloc = &iwpbl->pble_alloc;
+       iwmr->page_cnt = max_num_sg;
+       mutex_lock(&iwdev->pbl_mutex);
+       status = i40iw_get_pble(&iwdev->sc_dev, iwdev->pble_rsrc, palloc, iwmr->page_cnt);
+       mutex_unlock(&iwdev->pbl_mutex);
+       if (!status)
+               goto err1;
+
+       if (palloc->level != I40IW_LEVEL_1)
+               goto err2;
+       err_code = i40iw_hw_alloc_stag(iwdev, iwmr);
+       if (err_code)
+               goto err2;
+       iwpbl->pbl_allocated = true;
+       i40iw_add_pdusecount(iwpd);
+       return &iwmr->ibmr;
+err2:
+       i40iw_free_pble(iwdev->pble_rsrc, palloc);
+err1:
+       i40iw_free_stag(iwdev, stag);
+err:
+       kfree(iwmr);
+       return ERR_PTR(err_code);
+}
+
+/**
+ * i40iw_set_page - populate pbl list for fmr
+ * @ibmr: ib mem to access iwarp mr pointer
+ * @addr: page dma address fro pbl list
+ */
+static int i40iw_set_page(struct ib_mr *ibmr, u64 addr)
+{
+       struct i40iw_mr *iwmr = to_iwmr(ibmr);
+       struct i40iw_pbl *iwpbl = &iwmr->iwpbl;
+       struct i40iw_pble_alloc *palloc = &iwpbl->pble_alloc;
+       u64 *pbl;
+
+       if (unlikely(iwmr->npages == iwmr->page_cnt))
+               return -ENOMEM;
+
+       pbl = (u64 *)palloc->level1.addr;
+       pbl[iwmr->npages++] = cpu_to_le64(addr);
+       return 0;
+}
+
+/**
+ * i40iw_map_mr_sg - map of sg list for fmr
+ * @ibmr: ib mem to access iwarp mr pointer
+ * @sg: scatter gather list for fmr
+ * @sg_nents: number of sg pages
+ */
+static int i40iw_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents)
+{
+       struct i40iw_mr *iwmr = to_iwmr(ibmr);
+
+       iwmr->npages = 0;
+       return ib_sg_to_pages(ibmr, sg, sg_nents, i40iw_set_page);
+}
+
 /**
  * i40iw_hwreg_mr - send cqp command for memory registration
  * @iwdev: iwarp device
@@ -1886,12 +2021,14 @@ static int i40iw_post_send(struct ib_qp *ibqp,
        enum i40iw_status_code ret;
        int err = 0;
        unsigned long flags;
+       bool inv_stag;
 
        iwqp = (struct i40iw_qp *)ibqp;
        ukqp = &iwqp->sc_qp.qp_uk;
 
        spin_lock_irqsave(&iwqp->lock, flags);
        while (ib_wr) {
+               inv_stag = false;
                memset(&info, 0, sizeof(info));
                info.wr_id = (u64)(ib_wr->wr_id);
                if ((ib_wr->send_flags & IB_SEND_SIGNALED) || iwqp->sig_all)
@@ -1901,19 +2038,28 @@ static int i40iw_post_send(struct ib_qp *ibqp,
 
                switch (ib_wr->opcode) {
                case IB_WR_SEND:
-                       if (ib_wr->send_flags & IB_SEND_SOLICITED)
-                               info.op_type = I40IW_OP_TYPE_SEND_SOL;
-                       else
-                               info.op_type = I40IW_OP_TYPE_SEND;
+                       /* fall-through */
+               case IB_WR_SEND_WITH_INV:
+                       if (ib_wr->opcode == IB_WR_SEND) {
+                               if (ib_wr->send_flags & IB_SEND_SOLICITED)
+                                       info.op_type = I40IW_OP_TYPE_SEND_SOL;
+                               else
+                                       info.op_type = I40IW_OP_TYPE_SEND;
+                       } else {
+                               if (ib_wr->send_flags & IB_SEND_SOLICITED)
+                                       info.op_type = I40IW_OP_TYPE_SEND_SOL_INV;
+                               else
+                                       info.op_type = I40IW_OP_TYPE_SEND_INV;
+                       }
 
                        if (ib_wr->send_flags & IB_SEND_INLINE) {
                                info.op.inline_send.data = (void *)(unsigned long)ib_wr->sg_list[0].addr;
                                info.op.inline_send.len = ib_wr->sg_list[0].length;
-                               ret = ukqp->ops.iw_inline_send(ukqp, &info, rdma_wr(ib_wr)->rkey, false);
+                               ret = ukqp->ops.iw_inline_send(ukqp, &info, ib_wr->ex.invalidate_rkey, false);
                        } else {
                                info.op.send.num_sges = ib_wr->num_sge;
                                info.op.send.sg_list = (struct i40iw_sge *)ib_wr->sg_list;
-                               ret = ukqp->ops.iw_send(ukqp, &info, rdma_wr(ib_wr)->rkey, false);
+                               ret = ukqp->ops.iw_send(ukqp, &info, ib_wr->ex.invalidate_rkey, false);
                        }
 
                        if (ret)
@@ -1941,6 +2087,9 @@ static int i40iw_post_send(struct ib_qp *ibqp,
                        if (ret)
                                err = -EIO;
                        break;
+               case IB_WR_RDMA_READ_WITH_INV:
+                       inv_stag = true;
+                       /* fall-through*/
                case IB_WR_RDMA_READ:
                        info.op_type = I40IW_OP_TYPE_RDMA_READ;
                        info.op.rdma_read.rem_addr.tag_off = rdma_wr(ib_wr)->remote_addr;
@@ -1949,10 +2098,47 @@ static int i40iw_post_send(struct ib_qp *ibqp,
                        info.op.rdma_read.lo_addr.tag_off = ib_wr->sg_list->addr;
                        info.op.rdma_read.lo_addr.stag = ib_wr->sg_list->lkey;
                        info.op.rdma_read.lo_addr.len = ib_wr->sg_list->length;
-                       ret = ukqp->ops.iw_rdma_read(ukqp, &info, false, false);
+                       ret = ukqp->ops.iw_rdma_read(ukqp, &info, inv_stag, false);
                        if (ret)
                                err = -EIO;
                        break;
+               case IB_WR_LOCAL_INV:
+                       info.op_type = I40IW_OP_TYPE_INV_STAG;
+                       info.op.inv_local_stag.target_stag = ib_wr->ex.invalidate_rkey;
+                       ret = ukqp->ops.iw_stag_local_invalidate(ukqp, &info, true);
+                       if (ret)
+                               err = -EIO;
+                       break;
+               case IB_WR_REG_MR:
+               {
+                       struct i40iw_mr *iwmr = to_iwmr(reg_wr(ib_wr)->mr);
+                       int page_shift = ilog2(reg_wr(ib_wr)->mr->page_size);
+                       int flags = reg_wr(ib_wr)->access;
+                       struct i40iw_pble_alloc *palloc = &iwmr->iwpbl.pble_alloc;
+                       struct i40iw_sc_dev *dev = &iwqp->iwdev->sc_dev;
+                       struct i40iw_fast_reg_stag_info info;
+
+                       info.access_rights = I40IW_ACCESS_FLAGS_LOCALREAD;
+                       info.access_rights |= i40iw_get_user_access(flags);
+                       info.stag_key = reg_wr(ib_wr)->key & 0xff;
+                       info.stag_idx = reg_wr(ib_wr)->key >> 8;
+                       info.wr_id = ib_wr->wr_id;
+
+                       info.addr_type = I40IW_ADDR_TYPE_VA_BASED;
+                       info.va = (void *)(uintptr_t)iwmr->ibmr.iova;
+                       info.total_len = iwmr->ibmr.length;
+                       info.first_pm_pbl_index = palloc->level1.idx;
+                       info.local_fence = ib_wr->send_flags & IB_SEND_FENCE;
+                       info.signaled = ib_wr->send_flags & IB_SEND_SIGNALED;
+
+                       if (page_shift == 21)
+                               info.page_size = 1; /* 2M page */
+
+                       ret = dev->iw_priv_qp_ops->iw_mr_fast_register(&iwqp->sc_qp, &info, true);
+                       if (ret)
+                               err = -EIO;
+                       break;
+               }
                default:
                        err = -EINVAL;
                        i40iw_pr_err(" upost_send bad opcode = 0x%x\n",
@@ -2328,6 +2514,8 @@ static struct i40iw_ib_device *i40iw_init_rdma_device(struct i40iw_device *iwdev
        iwibdev->ibdev.query_device = i40iw_query_device;
        iwibdev->ibdev.create_ah = i40iw_create_ah;
        iwibdev->ibdev.destroy_ah = i40iw_destroy_ah;
+       iwibdev->ibdev.alloc_mr = i40iw_alloc_mr;
+       iwibdev->ibdev.map_mr_sg = i40iw_map_mr_sg;
        iwibdev->ibdev.iwcm = kzalloc(sizeof(*iwibdev->ibdev.iwcm), GFP_KERNEL);
        if (!iwibdev->ibdev.iwcm) {
                ib_dealloc_device(&iwibdev->ibdev);
index 1101f77080e62fd98aff799c8c90c19f175d953b..0acb6c8fe0f00fe83a9aafb82e9408d450fb0ac7 100644 (file)
@@ -92,6 +92,7 @@ struct i40iw_mr {
        struct ib_umem *region;
        u16 type;
        u32 page_cnt;
+       u32 npages;
        u32 stag;
        u64 length;
        u64 pgaddrmem[MAX_SAVE_PAGE_ADDRS];