From b414fa01c31318383ae29d9d23cb9ca4184bbd86 Mon Sep 17 00:00:00 2001 From: Steve Wise Date: Thu, 15 Dec 2016 08:09:35 -0800 Subject: [PATCH 01/15] iw_cxgb4: set correct FetchBurstMax for QPs The current QP FetchBurstMax value is 256B, which is incorrect since a WR can exceed that value. The result being a partial WR fetched by hardware, and a fatal "bad WR" error posted by the SGE. So bump the FetchBurstMax to 512B. Signed-off-by: Steve Wise Signed-off-by: Doug Ledford --- drivers/infiniband/hw/cxgb4/qp.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c index f57deba6717c..108c8837ba58 100644 --- a/drivers/infiniband/hw/cxgb4/qp.c +++ b/drivers/infiniband/hw/cxgb4/qp.c @@ -321,7 +321,8 @@ static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq, FW_RI_RES_WR_DCAEN_V(0) | FW_RI_RES_WR_DCACPU_V(0) | FW_RI_RES_WR_FBMIN_V(2) | - FW_RI_RES_WR_FBMAX_V(2) | + (t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_FBMAX_V(2) : + FW_RI_RES_WR_FBMAX_V(3)) | FW_RI_RES_WR_CIDXFTHRESHO_V(0) | FW_RI_RES_WR_CIDXFTHRESH_V(0) | FW_RI_RES_WR_EQSIZE_V(eqsize)); @@ -345,7 +346,7 @@ static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq, FW_RI_RES_WR_DCAEN_V(0) | FW_RI_RES_WR_DCACPU_V(0) | FW_RI_RES_WR_FBMIN_V(2) | - FW_RI_RES_WR_FBMAX_V(2) | + FW_RI_RES_WR_FBMAX_V(3) | FW_RI_RES_WR_CIDXFTHRESHO_V(0) | FW_RI_RES_WR_CIDXFTHRESH_V(0) | FW_RI_RES_WR_EQSIZE_V(eqsize)); From e259934d4df7f99f2a5c2c4f074f6a55bd4b1722 Mon Sep 17 00:00:00 2001 From: Bart Van Assche Date: Thu, 15 Dec 2016 17:15:07 +0100 Subject: [PATCH 02/15] IB/rxe: Fix a memory leak in rxe_qp_cleanup() A socket is associated with every QP by the rxe driver but sock_release() is never called. Add a call to sock_release() in rxe_qp_cleanup(). Fixes: commit 8700e3e7c48A5 ("Add Soft RoCE driver") Signed-off-by: Bart Van Assche Cc: Moni Shoua Cc: Kamal Heib Cc: Amir Vadai Cc: Haggai Eran Cc: Reviewed-by: Moni Shoua Signed-off-by: Doug Ledford --- drivers/infiniband/sw/rxe/rxe_qp.c | 1 + 1 file changed, 1 insertion(+) diff --git a/drivers/infiniband/sw/rxe/rxe_qp.c b/drivers/infiniband/sw/rxe/rxe_qp.c index b8036cfbce04..4b6e455e72a1 100644 --- a/drivers/infiniband/sw/rxe/rxe_qp.c +++ b/drivers/infiniband/sw/rxe/rxe_qp.c @@ -853,4 +853,5 @@ void rxe_qp_cleanup(void *arg) free_rd_atomic_resources(qp); kernel_sock_shutdown(qp->sk, SHUT_RDWR); + sock_release(qp->sk); } From fba332b079029c2f4f7e84c1c1cd8e3867310c90 Mon Sep 17 00:00:00 2001 From: Bart Van Assche Date: Mon, 19 Dec 2016 18:00:05 +0100 Subject: [PATCH 03/15] IB/cma: Fix a race condition in iboe_addr_get_sgid() Code that dereferences the struct net_device ip_ptr member must be protected with an in_dev_get() / in_dev_put() pair. Hence insert calls to these functions. Fixes: commit 7b85627b9f02 ("IB/cma: IBoE (RoCE) IP-based GID addressing") Signed-off-by: Bart Van Assche Reviewed-by: Moni Shoua Cc: Or Gerlitz Cc: Roland Dreier Cc: Signed-off-by: Doug Ledford --- include/rdma/ib_addr.h | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/include/rdma/ib_addr.h b/include/rdma/ib_addr.h index 931a47ba4571..1beab5532035 100644 --- a/include/rdma/ib_addr.h +++ b/include/rdma/ib_addr.h @@ -205,10 +205,12 @@ static inline void iboe_addr_get_sgid(struct rdma_dev_addr *dev_addr, dev = dev_get_by_index(&init_net, dev_addr->bound_dev_if); if (dev) { - ip4 = (struct in_device *)dev->ip_ptr; - if (ip4 && ip4->ifa_list && ip4->ifa_list->ifa_address) + ip4 = in_dev_get(dev); + if (ip4 && ip4->ifa_list && ip4->ifa_list->ifa_address) { ipv6_addr_set_v4mapped(ip4->ifa_list->ifa_address, (struct in6_addr *)gid); + in_dev_put(ip4); + } dev_put(dev); } } From 61f51b7b20f631ef8fe744bc0412d4eb5194b6a9 Mon Sep 17 00:00:00 2001 From: Chien Tin Tung Date: Wed, 21 Dec 2016 08:53:46 -0600 Subject: [PATCH 04/15] i40iw: Set 128B as the only supported RQ WQE size RQ WQE size other than 128B is not supported. Correct RQ size calculation to use 128B only. Since this breaks ABI, add additional code to provide compatibility with v4 user provider, libi40iw. Signed-off-by: Chien Tin Tung Signed-off-by: Henry Orosco Signed-off-by: Doug Ledford --- drivers/infiniband/hw/i40iw/i40iw_ctrl.c | 25 ++++++++++++++++---- drivers/infiniband/hw/i40iw/i40iw_puda.c | 2 +- drivers/infiniband/hw/i40iw/i40iw_type.h | 4 +++- drivers/infiniband/hw/i40iw/i40iw_ucontext.h | 4 ++-- drivers/infiniband/hw/i40iw/i40iw_uk.c | 17 +++++++++---- drivers/infiniband/hw/i40iw/i40iw_user.h | 4 +++- drivers/infiniband/hw/i40iw/i40iw_verbs.c | 21 ++++++++-------- drivers/infiniband/hw/i40iw/i40iw_verbs.h | 1 + 8 files changed, 53 insertions(+), 25 deletions(-) diff --git a/drivers/infiniband/hw/i40iw/i40iw_ctrl.c b/drivers/infiniband/hw/i40iw/i40iw_ctrl.c index 392f78384a60..98923a8cf86d 100644 --- a/drivers/infiniband/hw/i40iw/i40iw_ctrl.c +++ b/drivers/infiniband/hw/i40iw/i40iw_ctrl.c @@ -358,13 +358,16 @@ void i40iw_qp_add_qos(struct i40iw_sc_qp *qp) * @dev: sc device struct * @pd: sc pd ptr * @pd_id: pd_id for allocated pd + * @abi_ver: ABI version from user context, -1 if not valid */ static void i40iw_sc_pd_init(struct i40iw_sc_dev *dev, struct i40iw_sc_pd *pd, - u16 pd_id) + u16 pd_id, + int abi_ver) { pd->size = sizeof(*pd); pd->pd_id = pd_id; + pd->abi_ver = abi_ver; pd->dev = dev; } @@ -2252,6 +2255,7 @@ static enum i40iw_status_code i40iw_sc_qp_init(struct i40iw_sc_qp *qp, offset); info->qp_uk_init_info.wqe_alloc_reg = wqe_alloc_reg; + info->qp_uk_init_info.abi_ver = qp->pd->abi_ver; ret_code = i40iw_qp_uk_init(&qp->qp_uk, &info->qp_uk_init_info); if (ret_code) return ret_code; @@ -2270,10 +2274,21 @@ static enum i40iw_status_code i40iw_sc_qp_init(struct i40iw_sc_qp *qp, false); i40iw_debug(qp->dev, I40IW_DEBUG_WQE, "%s: hw_sq_size[%04d] sq_ring.size[%04d]\n", __func__, qp->hw_sq_size, qp->qp_uk.sq_ring.size); - ret_code = i40iw_fragcnt_to_wqesize_rq(qp->qp_uk.max_rq_frag_cnt, - &wqe_size); - if (ret_code) - return ret_code; + + switch (qp->pd->abi_ver) { + case 4: + ret_code = i40iw_fragcnt_to_wqesize_rq(qp->qp_uk.max_rq_frag_cnt, + &wqe_size); + if (ret_code) + return ret_code; + break; + case 5: /* fallthrough until next ABI version */ + default: + if (qp->qp_uk.max_rq_frag_cnt > I40IW_MAX_WQ_FRAGMENT_COUNT) + return I40IW_ERR_INVALID_FRAG_COUNT; + wqe_size = I40IW_MAX_WQE_SIZE_RQ; + break; + } qp->hw_rq_size = i40iw_get_encoded_wqe_size(qp->qp_uk.rq_size * (wqe_size / I40IW_QP_WQE_MIN_SIZE), false); i40iw_debug(qp->dev, I40IW_DEBUG_WQE, diff --git a/drivers/infiniband/hw/i40iw/i40iw_puda.c b/drivers/infiniband/hw/i40iw/i40iw_puda.c index 449ba8c81ce7..db41ab40da9c 100644 --- a/drivers/infiniband/hw/i40iw/i40iw_puda.c +++ b/drivers/infiniband/hw/i40iw/i40iw_puda.c @@ -930,7 +930,7 @@ enum i40iw_status_code i40iw_puda_create_rsrc(struct i40iw_sc_vsi *vsi, INIT_LIST_HEAD(&rsrc->txpend); rsrc->tx_wqe_avail_cnt = info->sq_size - 1; - dev->iw_pd_ops->pd_init(dev, &rsrc->sc_pd, info->pd_id); + dev->iw_pd_ops->pd_init(dev, &rsrc->sc_pd, info->pd_id, -1); rsrc->qp_id = info->qp_id; rsrc->cq_id = info->cq_id; rsrc->sq_size = info->sq_size; diff --git a/drivers/infiniband/hw/i40iw/i40iw_type.h b/drivers/infiniband/hw/i40iw/i40iw_type.h index f3f8e9cc3c05..7b76259752b0 100644 --- a/drivers/infiniband/hw/i40iw/i40iw_type.h +++ b/drivers/infiniband/hw/i40iw/i40iw_type.h @@ -280,6 +280,7 @@ struct i40iw_sc_pd { u32 size; struct i40iw_sc_dev *dev; u16 pd_id; + int abi_ver; }; struct i40iw_cqp_quanta { @@ -852,6 +853,7 @@ struct i40iw_qp_init_info { u64 host_ctx_pa; u64 q2_pa; u64 shadow_area_pa; + int abi_ver; u8 sq_tph_val; u8 rq_tph_val; u8 type; @@ -1051,7 +1053,7 @@ struct i40iw_aeq_ops { }; struct i40iw_pd_ops { - void (*pd_init)(struct i40iw_sc_dev *, struct i40iw_sc_pd *, u16); + void (*pd_init)(struct i40iw_sc_dev *, struct i40iw_sc_pd *, u16, int); }; struct i40iw_priv_qp_ops { diff --git a/drivers/infiniband/hw/i40iw/i40iw_ucontext.h b/drivers/infiniband/hw/i40iw/i40iw_ucontext.h index 12acd688def4..57d3f1d11ff1 100644 --- a/drivers/infiniband/hw/i40iw/i40iw_ucontext.h +++ b/drivers/infiniband/hw/i40iw/i40iw_ucontext.h @@ -39,8 +39,8 @@ #include -#define I40IW_ABI_USERSPACE_VER 4 -#define I40IW_ABI_KERNEL_VER 4 +#define I40IW_ABI_VER 5 + struct i40iw_alloc_ucontext_req { __u32 reserved32; __u8 userspace_ver; diff --git a/drivers/infiniband/hw/i40iw/i40iw_uk.c b/drivers/infiniband/hw/i40iw/i40iw_uk.c index 4376cd628774..2800f796271c 100644 --- a/drivers/infiniband/hw/i40iw/i40iw_uk.c +++ b/drivers/infiniband/hw/i40iw/i40iw_uk.c @@ -966,10 +966,6 @@ enum i40iw_status_code i40iw_qp_uk_init(struct i40iw_qp_uk *qp, if (ret_code) return ret_code; - ret_code = i40iw_get_wqe_shift(info->rq_size, info->max_rq_frag_cnt, 0, &rqshift); - if (ret_code) - return ret_code; - qp->sq_base = info->sq; qp->rq_base = info->rq; qp->shadow_area = info->shadow_area; @@ -998,8 +994,19 @@ enum i40iw_status_code i40iw_qp_uk_init(struct i40iw_qp_uk *qp, if (!qp->use_srq) { qp->rq_size = info->rq_size; qp->max_rq_frag_cnt = info->max_rq_frag_cnt; - qp->rq_wqe_size = rqshift; I40IW_RING_INIT(qp->rq_ring, qp->rq_size); + switch (info->abi_ver) { + case 4: + ret_code = i40iw_get_wqe_shift(info->rq_size, info->max_rq_frag_cnt, 0, &rqshift); + if (ret_code) + return ret_code; + break; + case 5: /* fallthrough until next ABI version */ + default: + rqshift = I40IW_MAX_RQ_WQE_SHIFT; + break; + } + qp->rq_wqe_size = rqshift; qp->rq_wqe_size_multiplier = 4 << rqshift; } qp->ops = iw_qp_uk_ops; diff --git a/drivers/infiniband/hw/i40iw/i40iw_user.h b/drivers/infiniband/hw/i40iw/i40iw_user.h index 80d9f464f65e..84be6f13b9c5 100644 --- a/drivers/infiniband/hw/i40iw/i40iw_user.h +++ b/drivers/infiniband/hw/i40iw/i40iw_user.h @@ -76,6 +76,7 @@ enum i40iw_device_capabilities_const { I40IW_MAX_ORD_SIZE = 127, I40IW_MAX_WQ_ENTRIES = 2048, I40IW_Q2_BUFFER_SIZE = (248 + 100), + I40IW_MAX_WQE_SIZE_RQ = 128, I40IW_QP_CTX_SIZE = 248, I40IW_MAX_PDS = 32768 }; @@ -97,6 +98,7 @@ enum i40iw_device_capabilities_const { #define i40iw_address_list u64 * #define I40IW_MAX_MR_SIZE 0x10000000000L +#define I40IW_MAX_RQ_WQE_SHIFT 2 struct i40iw_qp_uk; struct i40iw_cq_uk; @@ -405,7 +407,7 @@ struct i40iw_qp_uk_init_info { u32 max_sq_frag_cnt; u32 max_rq_frag_cnt; u32 max_inline_data; - + int abi_ver; }; struct i40iw_cq_uk_init_info { diff --git a/drivers/infiniband/hw/i40iw/i40iw_verbs.c b/drivers/infiniband/hw/i40iw/i40iw_verbs.c index 7368a50bbdaa..29e97df9e1a7 100644 --- a/drivers/infiniband/hw/i40iw/i40iw_verbs.c +++ b/drivers/infiniband/hw/i40iw/i40iw_verbs.c @@ -145,9 +145,8 @@ static struct ib_ucontext *i40iw_alloc_ucontext(struct ib_device *ibdev, if (ib_copy_from_udata(&req, udata, sizeof(req))) return ERR_PTR(-EINVAL); - if (req.userspace_ver != I40IW_ABI_USERSPACE_VER) { - i40iw_pr_err("Invalid userspace driver version detected. Detected version %d, should be %d\n", - req.userspace_ver, I40IW_ABI_USERSPACE_VER); + if (req.userspace_ver < 4 || req.userspace_ver > I40IW_ABI_VER) { + i40iw_pr_err("Unsupported provider library version %u.\n", req.userspace_ver); return ERR_PTR(-EINVAL); } @@ -155,13 +154,14 @@ static struct ib_ucontext *i40iw_alloc_ucontext(struct ib_device *ibdev, uresp.max_qps = iwdev->max_qp; uresp.max_pds = iwdev->max_pd; uresp.wq_size = iwdev->max_qp_wr * 2; - uresp.kernel_ver = I40IW_ABI_KERNEL_VER; + uresp.kernel_ver = req.userspace_ver; ucontext = kzalloc(sizeof(*ucontext), GFP_KERNEL); if (!ucontext) return ERR_PTR(-ENOMEM); ucontext->iwdev = iwdev; + ucontext->abi_ver = req.userspace_ver; if (ib_copy_to_udata(udata, &uresp, sizeof(uresp))) { kfree(ucontext); @@ -333,6 +333,7 @@ static struct ib_pd *i40iw_alloc_pd(struct ib_device *ibdev, struct i40iw_sc_dev *dev = &iwdev->sc_dev; struct i40iw_alloc_pd_resp uresp; struct i40iw_sc_pd *sc_pd; + struct i40iw_ucontext *ucontext; u32 pd_id = 0; int err; @@ -353,15 +354,18 @@ static struct ib_pd *i40iw_alloc_pd(struct ib_device *ibdev, } sc_pd = &iwpd->sc_pd; - dev->iw_pd_ops->pd_init(dev, sc_pd, pd_id); if (context) { + ucontext = to_ucontext(context); + dev->iw_pd_ops->pd_init(dev, sc_pd, pd_id, ucontext->abi_ver); memset(&uresp, 0, sizeof(uresp)); uresp.pd_id = pd_id; if (ib_copy_to_udata(udata, &uresp, sizeof(uresp))) { err = -EFAULT; goto error; } + } else { + dev->iw_pd_ops->pd_init(dev, sc_pd, pd_id, -1); } i40iw_add_pdusecount(iwpd); @@ -518,7 +522,7 @@ static int i40iw_setup_kmode_qp(struct i40iw_device *iwdev, struct i40iw_dma_mem *mem = &iwqp->kqp.dma_mem; u32 sqdepth, rqdepth; u32 sq_size, rq_size; - u8 sqshift, rqshift; + u8 sqshift; u32 size; enum i40iw_status_code status; struct i40iw_qp_uk_init_info *ukinfo = &info->qp_uk_init_info; @@ -527,14 +531,11 @@ static int i40iw_setup_kmode_qp(struct i40iw_device *iwdev, rq_size = i40iw_qp_roundup(ukinfo->rq_size + 1); status = i40iw_get_wqe_shift(sq_size, ukinfo->max_sq_frag_cnt, ukinfo->max_inline_data, &sqshift); - if (!status) - status = i40iw_get_wqe_shift(rq_size, ukinfo->max_rq_frag_cnt, 0, &rqshift); - if (status) return -ENOMEM; sqdepth = sq_size << sqshift; - rqdepth = rq_size << rqshift; + rqdepth = rq_size << I40IW_MAX_RQ_WQE_SHIFT; size = sqdepth * sizeof(struct i40iw_sq_uk_wr_trk_info) + (rqdepth << 3); iwqp->kqp.wrid_mem = kzalloc(size, GFP_KERNEL); diff --git a/drivers/infiniband/hw/i40iw/i40iw_verbs.h b/drivers/infiniband/hw/i40iw/i40iw_verbs.h index 6549c939500f..07c3fec77de6 100644 --- a/drivers/infiniband/hw/i40iw/i40iw_verbs.h +++ b/drivers/infiniband/hw/i40iw/i40iw_verbs.h @@ -42,6 +42,7 @@ struct i40iw_ucontext { spinlock_t cq_reg_mem_list_lock; /* memory list for cq's */ struct list_head qp_reg_mem_list; spinlock_t qp_reg_mem_list_lock; /* memory list for qp's */ + int abi_ver; }; struct i40iw_pd { From c7eb3bced78fe83119b90d730ab58a572eb80e29 Mon Sep 17 00:00:00 2001 From: "Amrani, Ram" Date: Thu, 22 Dec 2016 14:40:33 +0200 Subject: [PATCH 05/15] qedr: configure the number of CQEs on CQ creation Configure ibcq->cqe when a CQ is created. Signed-off-by: Ram Amrani Reviewed-by: Michal Kalderon Signed-off-by: Doug Ledford --- drivers/infiniband/hw/qedr/verbs.c | 3 +++ 1 file changed, 3 insertions(+) diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c index 302fb05e6e6f..73df639755fa 100644 --- a/drivers/infiniband/hw/qedr/verbs.c +++ b/drivers/infiniband/hw/qedr/verbs.c @@ -890,6 +890,8 @@ struct ib_cq *qedr_create_cq(struct ib_device *ibdev, pbl_ptr = cq->q.pbl_tbl->pa; page_cnt = cq->q.pbl_info.num_pbes; + + cq->ibcq.cqe = chain_entries; } else { cq->cq_type = QEDR_CQ_TYPE_KERNEL; @@ -905,6 +907,7 @@ struct ib_cq *qedr_create_cq(struct ib_device *ibdev, page_cnt = qed_chain_get_page_cnt(&cq->pbl); pbl_ptr = qed_chain_get_pbl_phys(&cq->pbl); + cq->ibcq.cqe = cq->pbl.capacity; } qedr_init_cq_params(cq, ctx, dev, vector, chain_entries, page_cnt, From a121135973ca816821f4ff07aed523df82a91b8e Mon Sep 17 00:00:00 2001 From: "Amrani, Ram" Date: Thu, 22 Dec 2016 14:40:34 +0200 Subject: [PATCH 06/15] qedr: return error if destroy CQ failed Signed-off-by: Ram Amrani Reviewed-by: Michal Kalderon Signed-off-by: Doug Ledford --- drivers/infiniband/hw/qedr/verbs.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c index 73df639755fa..cbccd7d3ae1e 100644 --- a/drivers/infiniband/hw/qedr/verbs.c +++ b/drivers/infiniband/hw/qedr/verbs.c @@ -985,8 +985,13 @@ int qedr_destroy_cq(struct ib_cq *ibcq) /* GSIs CQs are handled by driver, so they don't exist in the FW */ if (cq->cq_type != QEDR_CQ_TYPE_GSI) { + int rc; + iparams.icid = cq->icid; - dev->ops->rdma_destroy_cq(dev->rdma_ctx, &iparams, &oparams); + rc = dev->ops->rdma_destroy_cq(dev->rdma_ctx, &iparams, + &oparams); + if (rc) + return rc; dev->ops->common->chain_free(dev->cdev, &cq->pbl); } From d6ebbf29c3015bdccef388a860ac4ef6772531f8 Mon Sep 17 00:00:00 2001 From: "Amrani, Ram" Date: Thu, 22 Dec 2016 14:40:35 +0200 Subject: [PATCH 07/15] qedr: return correct value on modify qp Signed-off-by: Ram Amrani Reviewed-by: Michal Kalderon Signed-off-by: Doug Ledford --- drivers/infiniband/hw/qedr/verbs.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c index cbccd7d3ae1e..5d1980dd6c48 100644 --- a/drivers/infiniband/hw/qedr/verbs.c +++ b/drivers/infiniband/hw/qedr/verbs.c @@ -1974,7 +1974,7 @@ int qedr_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, if (attr_mask & IB_QP_STATE) { if ((qp->qp_type != IB_QPT_GSI) && (!udata)) - qedr_update_qp_state(dev, qp, qp_params.new_state); + rc = qedr_update_qp_state(dev, qp, qp_params.new_state); qp->state = qp_params.new_state; } From b4c2cc48aa0be767281669bff9f230e81ef27c56 Mon Sep 17 00:00:00 2001 From: "Amrani, Ram" Date: Thu, 22 Dec 2016 14:40:36 +0200 Subject: [PATCH 08/15] qedr: modify QP state to error when destroying it Current code didn't modify the QP state to error because it queried the QP state as a bitmap while it isn't. So the code never got executed. This patch fixes this and queries for each QP state respectively and not at once via a bitmask. Signed-off-by: Ram Amrani Reviewed-by: Michal Kalderon Signed-off-by: Doug Ledford --- drivers/infiniband/hw/qedr/verbs.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c index 5d1980dd6c48..0237ff986829 100644 --- a/drivers/infiniband/hw/qedr/verbs.c +++ b/drivers/infiniband/hw/qedr/verbs.c @@ -2078,8 +2078,10 @@ int qedr_destroy_qp(struct ib_qp *ibqp) DP_DEBUG(dev, QEDR_MSG_QP, "destroy qp: destroying %p, qp type=%d\n", qp, qp->qp_type); - if (qp->state != (QED_ROCE_QP_STATE_RESET | QED_ROCE_QP_STATE_ERR | - QED_ROCE_QP_STATE_INIT)) { + if ((qp->state != QED_ROCE_QP_STATE_RESET) && + (qp->state != QED_ROCE_QP_STATE_ERR) && + (qp->state != QED_ROCE_QP_STATE_INIT)) { + attr.qp_state = IB_QPS_ERR; attr_mask |= IB_QP_STATE; From 8b0cabc650a95a4f44de99aae6e8c128d70a40cd Mon Sep 17 00:00:00 2001 From: "Amrani, Ram" Date: Thu, 22 Dec 2016 14:40:37 +0200 Subject: [PATCH 09/15] qedr: ignore inline flag in read verbs In the current implementation a read verb with IB_SEND_INLINE may be illegally configured. In this fix we ignore the inline bit in the case of a read verb. Signed-off-by: Ram Amrani Reviewed-by: Michal Kalderon Signed-off-by: Doug Ledford --- drivers/infiniband/hw/qedr/verbs.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c index 0237ff986829..9af2b4f3b62f 100644 --- a/drivers/infiniband/hw/qedr/verbs.c +++ b/drivers/infiniband/hw/qedr/verbs.c @@ -2636,7 +2636,9 @@ static u32 qedr_prepare_sq_rdma_data(struct qedr_dev *dev, rwqe2->r_key = cpu_to_le32(rdma_wr(wr)->rkey); DMA_REGPAIR_LE(rwqe2->remote_va, rdma_wr(wr)->remote_addr); - if (wr->send_flags & IB_SEND_INLINE) { + if (wr->send_flags & IB_SEND_INLINE && + (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM || + wr->opcode == IB_WR_RDMA_WRITE)) { u8 flags = 0; SET_FIELD2(flags, RDMA_SQ_RDMA_WQE_1ST_INLINE_FLG, 1); From 922d9a40d3baeb30bfecb59b2c083ccec4b349d7 Mon Sep 17 00:00:00 2001 From: "Amrani, Ram" Date: Thu, 22 Dec 2016 14:40:38 +0200 Subject: [PATCH 10/15] qedr: post_send/recv according to QP state Enable posting to SQ only in RTS, ERR and SQD QP state. Enable posting to RQ in ERR QP state. Signed-off-by: Ram Amrani Reviewed-by: Michal Kalderon Signed-off-by: Doug Ledford --- drivers/infiniband/hw/qedr/verbs.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c index 9af2b4f3b62f..e31141be5b04 100644 --- a/drivers/infiniband/hw/qedr/verbs.c +++ b/drivers/infiniband/hw/qedr/verbs.c @@ -2989,8 +2989,9 @@ int qedr_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, spin_lock_irqsave(&qp->q_lock, flags); - if ((qp->state == QED_ROCE_QP_STATE_RESET) || - (qp->state == QED_ROCE_QP_STATE_ERR)) { + if ((qp->state != QED_ROCE_QP_STATE_RTS) && + (qp->state != QED_ROCE_QP_STATE_ERR) && + (qp->state != QED_ROCE_QP_STATE_SQD)) { spin_unlock_irqrestore(&qp->q_lock, flags); *bad_wr = wr; DP_DEBUG(dev, QEDR_MSG_CQ, @@ -3043,8 +3044,7 @@ int qedr_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, spin_lock_irqsave(&qp->q_lock, flags); - if ((qp->state == QED_ROCE_QP_STATE_RESET) || - (qp->state == QED_ROCE_QP_STATE_ERR)) { + if (qp->state == QED_ROCE_QP_STATE_RESET) { spin_unlock_irqrestore(&qp->q_lock, flags); *bad_wr = wr; return -EINVAL; From 27035a1b37fc284b59a2bca4cf2f910ebf03717f Mon Sep 17 00:00:00 2001 From: "Amrani, Ram" Date: Thu, 22 Dec 2016 14:40:39 +0200 Subject: [PATCH 11/15] qedr: clear the vendor error field in the work completion We clear the vendor error field in the work completion so that if a work completion is erroneous the field won't confuse the caller. Signed-off-by: Ram Amrani Reviewed-by: Michal Kalderon Signed-off-by: Doug Ledford --- drivers/infiniband/hw/qedr/verbs.c | 3 +++ 1 file changed, 3 insertions(+) diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c index e31141be5b04..4017315ae734 100644 --- a/drivers/infiniband/hw/qedr/verbs.c +++ b/drivers/infiniband/hw/qedr/verbs.c @@ -3186,6 +3186,7 @@ static int process_req(struct qedr_dev *dev, struct qedr_qp *qp, /* fill WC */ wc->status = status; + wc->vendor_err = 0; wc->wc_flags = 0; wc->src_qp = qp->id; wc->qp = &qp->ibqp; @@ -3375,6 +3376,7 @@ static void __process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp, /* fill WC */ wc->status = wc_status; + wc->vendor_err = 0; wc->src_qp = qp->id; wc->qp = &qp->ibqp; wc->wr_id = wr_id; @@ -3403,6 +3405,7 @@ static int process_resp_flush(struct qedr_qp *qp, struct qedr_cq *cq, while (num_entries && qp->rq.wqe_cons != hw_cons) { /* fill WC */ wc->status = IB_WC_WR_FLUSH_ERR; + wc->vendor_err = 0; wc->wc_flags = 0; wc->src_qp = qp->id; wc->byte_len = 0; From 74c3875c3d9aad356ae27fc4f4176d4dc89c457b Mon Sep 17 00:00:00 2001 From: "Amrani, Ram" Date: Thu, 22 Dec 2016 14:40:40 +0200 Subject: [PATCH 12/15] qedr: Always notify the verb consumer of flushed CQEs Signed-off-by: Ram Amrani Reviewed-by: Michal Kalderon Signed-off-by: Doug Ledford --- drivers/infiniband/hw/qedr/verbs.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c index 4017315ae734..57c8de208077 100644 --- a/drivers/infiniband/hw/qedr/verbs.c +++ b/drivers/infiniband/hw/qedr/verbs.c @@ -3238,7 +3238,7 @@ static int qedr_poll_cq_req(struct qedr_dev *dev, "Error: POLL CQ with RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR. CQ icid=0x%x, QP icid=0x%x\n", cq->icid, qp->icid); cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons, - IB_WC_WR_FLUSH_ERR, 0); + IB_WC_WR_FLUSH_ERR, 1); break; default: /* process all WQE before the cosumer */ From 37b36193946e4fe7af2c3965e394eb311ab6601d Mon Sep 17 00:00:00 2001 From: Andrew Boyer Date: Thu, 22 Dec 2016 08:54:36 -0500 Subject: [PATCH 13/15] IB/rxe: Use BTH_PSN_MASK when ACKing duplicate sends Signed-off-by: Andrew Boyer Signed-off-by: Doug Ledford --- drivers/infiniband/sw/rxe/rxe_resp.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/drivers/infiniband/sw/rxe/rxe_resp.c b/drivers/infiniband/sw/rxe/rxe_resp.c index 7a36ec9dbc0c..3435efff8799 100644 --- a/drivers/infiniband/sw/rxe/rxe_resp.c +++ b/drivers/infiniband/sw/rxe/rxe_resp.c @@ -1070,12 +1070,13 @@ static enum resp_states duplicate_request(struct rxe_qp *qp, struct rxe_pkt_info *pkt) { enum resp_states rc; + u32 prev_psn = (qp->resp.psn - 1) & BTH_PSN_MASK; if (pkt->mask & RXE_SEND_MASK || pkt->mask & RXE_WRITE_MASK) { /* SEND. Ack again and cleanup. C9-105. */ if (bth_ack(pkt)) - send_ack(qp, pkt, AETH_ACK_UNLIMITED, qp->resp.psn - 1); + send_ack(qp, pkt, AETH_ACK_UNLIMITED, prev_psn); rc = RESPST_CLEANUP; goto out; } else if (pkt->mask & RXE_READ_MASK) { From cbf1f9a46c9c5903e34b0b110ca40b1e7d4b2ab1 Mon Sep 17 00:00:00 2001 From: Andrew Boyer Date: Thu, 22 Dec 2016 08:54:37 -0500 Subject: [PATCH 14/15] IB/rxe: Drop future atomic/read packets rather than retrying If the completer is in the middle of a large read operation, one lost packet can cause havoc. Going to COMPST_ERROR_RETRY will cause the requester to resend the request. After that, any packet from the first attempt still in the receive queue will be interpreted as an error, restarting the error/retry sequence. The transfer will quickly exhaust its retries. This behavior is very noticeable when doing 512KB reads on a QEMU system configured with 1500B MTU. Also, a resent request here will prompt the responder on the other side to immediately start resending, but the resent packets will get stuck in the already-loaded receive queue and will never be processed. Rather than erroring out every time an unexpected future packet arrives, just drop it. Eventually the retry timer will send a duplicate request; the completer will be able to make progress since the queue will start relatively empty. Signed-off-by: Andrew Boyer Signed-off-by: Doug Ledford --- drivers/infiniband/sw/rxe/rxe_comp.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/drivers/infiniband/sw/rxe/rxe_comp.c b/drivers/infiniband/sw/rxe/rxe_comp.c index cd27cbde7652..d369f24425f9 100644 --- a/drivers/infiniband/sw/rxe/rxe_comp.c +++ b/drivers/infiniband/sw/rxe/rxe_comp.c @@ -224,7 +224,7 @@ static inline enum comp_state check_psn(struct rxe_qp *qp, else return COMPST_DONE; } else if ((diff > 0) && (wqe->mask & WR_ATOMIC_OR_READ_MASK)) { - return COMPST_ERROR_RETRY; + return COMPST_DONE; } else { return COMPST_CHECK_ACK; } From 5cc8fabc5e4c588c75a5ec21423e7c3425f69f48 Mon Sep 17 00:00:00 2001 From: Andrew Boyer Date: Thu, 22 Dec 2016 08:54:38 -0500 Subject: [PATCH 15/15] IB/rxe: Don't check for null ptr in send() pkt->qp was already dereferenced earlier in the function. Fixes Smatch complaint: drivers/infiniband/sw/rxe/rxe_net.c:458 send() warn: variable dereferenced before check 'pkt->qp' (see line 441) Signed-off-by: Andrew Boyer Signed-off-by: Doug Ledford --- drivers/infiniband/sw/rxe/rxe_net.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/drivers/infiniband/sw/rxe/rxe_net.c b/drivers/infiniband/sw/rxe/rxe_net.c index a576603304f7..10ec699834fa 100644 --- a/drivers/infiniband/sw/rxe/rxe_net.c +++ b/drivers/infiniband/sw/rxe/rxe_net.c @@ -459,8 +459,7 @@ static int send(struct rxe_dev *rxe, struct rxe_pkt_info *pkt, return -EAGAIN; } - if (pkt->qp) - atomic_inc(&pkt->qp->skb_out); + atomic_inc(&pkt->qp->skb_out); kfree_skb(skb); return 0;