1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
4 /* Copyright (c) 2008-2019, IBM Corporation */
6 #include <linux/errno.h>
7 #include <linux/types.h>
8 #include <linux/uaccess.h>
9 #include <linux/vmalloc.h>
10 #include <linux/xarray.h>
12 #include <rdma/iw_cm.h>
13 #include <rdma/ib_verbs.h>
14 #include <rdma/ib_user_verbs.h>
15 #include <rdma/uverbs_ioctl.h>
18 #include "siw_verbs.h"
21 static int ib_qp_state_to_siw_qp_state[IB_QPS_ERR + 1] = {
22 [IB_QPS_RESET] = SIW_QP_STATE_IDLE,
23 [IB_QPS_INIT] = SIW_QP_STATE_IDLE,
24 [IB_QPS_RTR] = SIW_QP_STATE_RTR,
25 [IB_QPS_RTS] = SIW_QP_STATE_RTS,
26 [IB_QPS_SQD] = SIW_QP_STATE_CLOSING,
27 [IB_QPS_SQE] = SIW_QP_STATE_TERMINATE,
28 [IB_QPS_ERR] = SIW_QP_STATE_ERROR
31 static char ib_qp_state_to_string[IB_QPS_ERR + 1][sizeof("RESET")] = {
32 [IB_QPS_RESET] = "RESET", [IB_QPS_INIT] = "INIT", [IB_QPS_RTR] = "RTR",
33 [IB_QPS_RTS] = "RTS", [IB_QPS_SQD] = "SQD", [IB_QPS_SQE] = "SQE",
37 void siw_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
39 struct siw_user_mmap_entry *entry = to_siw_mmap_entry(rdma_entry);
44 int siw_mmap(struct ib_ucontext *ctx, struct vm_area_struct *vma)
46 struct siw_ucontext *uctx = to_siw_ctx(ctx);
47 size_t size = vma->vm_end - vma->vm_start;
48 struct rdma_user_mmap_entry *rdma_entry;
49 struct siw_user_mmap_entry *entry;
53 * Must be page aligned
55 if (vma->vm_start & (PAGE_SIZE - 1)) {
56 pr_warn("siw: mmap not page aligned\n");
59 rdma_entry = rdma_user_mmap_entry_get(&uctx->base_ucontext, vma);
61 siw_dbg(&uctx->sdev->base_dev, "mmap lookup failed: %lu, %#zx\n",
65 entry = to_siw_mmap_entry(rdma_entry);
67 rv = remap_vmalloc_range(vma, entry->address, 0);
69 pr_warn("remap_vmalloc_range failed: %lu, %zu\n", vma->vm_pgoff,
74 rdma_user_mmap_entry_put(rdma_entry);
79 int siw_alloc_ucontext(struct ib_ucontext *base_ctx, struct ib_udata *udata)
81 struct siw_device *sdev = to_siw_dev(base_ctx->device);
82 struct siw_ucontext *ctx = to_siw_ctx(base_ctx);
83 struct siw_uresp_alloc_ctx uresp = {};
86 if (atomic_inc_return(&sdev->num_ctx) > SIW_MAX_CONTEXT) {
92 uresp.dev_id = sdev->vendor_part_id;
94 if (udata->outlen < sizeof(uresp)) {
98 rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
102 siw_dbg(base_ctx->device, "success. now %d context(s)\n",
103 atomic_read(&sdev->num_ctx));
108 atomic_dec(&sdev->num_ctx);
109 siw_dbg(base_ctx->device, "failure %d. now %d context(s)\n", rv,
110 atomic_read(&sdev->num_ctx));
115 void siw_dealloc_ucontext(struct ib_ucontext *base_ctx)
117 struct siw_ucontext *uctx = to_siw_ctx(base_ctx);
119 atomic_dec(&uctx->sdev->num_ctx);
122 int siw_query_device(struct ib_device *base_dev, struct ib_device_attr *attr,
123 struct ib_udata *udata)
125 struct siw_device *sdev = to_siw_dev(base_dev);
127 if (udata->inlen || udata->outlen)
130 memset(attr, 0, sizeof(*attr));
132 /* Revisit atomic caps if RFC 7306 gets supported */
133 attr->atomic_cap = 0;
134 attr->device_cap_flags =
135 IB_DEVICE_MEM_MGT_EXTENSIONS | IB_DEVICE_ALLOW_USER_UNREG;
136 attr->max_cq = sdev->attrs.max_cq;
137 attr->max_cqe = sdev->attrs.max_cqe;
138 attr->max_fast_reg_page_list_len = SIW_MAX_SGE_PBL;
139 attr->max_mr = sdev->attrs.max_mr;
140 attr->max_mw = sdev->attrs.max_mw;
141 attr->max_mr_size = ~0ull;
142 attr->max_pd = sdev->attrs.max_pd;
143 attr->max_qp = sdev->attrs.max_qp;
144 attr->max_qp_init_rd_atom = sdev->attrs.max_ird;
145 attr->max_qp_rd_atom = sdev->attrs.max_ord;
146 attr->max_qp_wr = sdev->attrs.max_qp_wr;
147 attr->max_recv_sge = sdev->attrs.max_sge;
148 attr->max_res_rd_atom = sdev->attrs.max_qp * sdev->attrs.max_ird;
149 attr->max_send_sge = sdev->attrs.max_sge;
150 attr->max_sge_rd = sdev->attrs.max_sge_rd;
151 attr->max_srq = sdev->attrs.max_srq;
152 attr->max_srq_sge = sdev->attrs.max_srq_sge;
153 attr->max_srq_wr = sdev->attrs.max_srq_wr;
154 attr->page_size_cap = PAGE_SIZE;
155 attr->vendor_id = SIW_VENDOR_ID;
156 attr->vendor_part_id = sdev->vendor_part_id;
158 memcpy(&attr->sys_image_guid, sdev->netdev->dev_addr, 6);
163 int siw_query_port(struct ib_device *base_dev, u8 port,
164 struct ib_port_attr *attr)
166 struct siw_device *sdev = to_siw_dev(base_dev);
169 memset(attr, 0, sizeof(*attr));
171 rv = ib_get_eth_speed(base_dev, port, &attr->active_speed,
172 &attr->active_width);
173 attr->gid_tbl_len = 1;
174 attr->max_msg_sz = -1;
175 attr->max_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
176 attr->active_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
177 attr->phys_state = sdev->state == IB_PORT_ACTIVE ?
178 IB_PORT_PHYS_STATE_LINK_UP : IB_PORT_PHYS_STATE_DISABLED;
179 attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_DEVICE_MGMT_SUP;
180 attr->state = sdev->state;
185 * attr->bad_pkey_cntr = 0;
186 * attr->qkey_viol_cntr = 0;
189 * attr->max_vl_num = 0;
191 * attr->subnet_timeout = 0;
192 * attr->init_type_repy = 0;
197 int siw_get_port_immutable(struct ib_device *base_dev, u8 port,
198 struct ib_port_immutable *port_immutable)
200 struct ib_port_attr attr;
201 int rv = siw_query_port(base_dev, port, &attr);
206 port_immutable->gid_tbl_len = attr.gid_tbl_len;
207 port_immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
212 int siw_query_gid(struct ib_device *base_dev, u8 port, int idx,
215 struct siw_device *sdev = to_siw_dev(base_dev);
217 /* subnet_prefix == interface_id == 0; */
218 memset(gid, 0, sizeof(*gid));
219 memcpy(&gid->raw[0], sdev->netdev->dev_addr, 6);
224 int siw_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
226 struct siw_device *sdev = to_siw_dev(pd->device);
228 if (atomic_inc_return(&sdev->num_pd) > SIW_MAX_PD) {
229 atomic_dec(&sdev->num_pd);
232 siw_dbg_pd(pd, "now %d PD's(s)\n", atomic_read(&sdev->num_pd));
237 int siw_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
239 struct siw_device *sdev = to_siw_dev(pd->device);
241 siw_dbg_pd(pd, "free PD\n");
242 atomic_dec(&sdev->num_pd);
246 void siw_qp_get_ref(struct ib_qp *base_qp)
248 siw_qp_get(to_siw_qp(base_qp));
251 void siw_qp_put_ref(struct ib_qp *base_qp)
253 siw_qp_put(to_siw_qp(base_qp));
256 static struct rdma_user_mmap_entry *
257 siw_mmap_entry_insert(struct siw_ucontext *uctx,
258 void *address, size_t length,
261 struct siw_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
264 *offset = SIW_INVAL_UOBJ_KEY;
268 entry->address = address;
270 rv = rdma_user_mmap_entry_insert(&uctx->base_ucontext,
278 *offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
280 return &entry->rdma_entry;
286 * Create QP of requested size on given device.
288 * @pd: Protection Domain
289 * @attrs: Initial QP attributes.
290 * @udata: used to provide QP ID, SQ and RQ size back to user.
293 struct ib_qp *siw_create_qp(struct ib_pd *pd,
294 struct ib_qp_init_attr *attrs,
295 struct ib_udata *udata)
297 struct siw_qp *qp = NULL;
298 struct ib_device *base_dev = pd->device;
299 struct siw_device *sdev = to_siw_dev(base_dev);
300 struct siw_ucontext *uctx =
301 rdma_udata_to_drv_context(udata, struct siw_ucontext,
303 struct siw_cq *scq = NULL, *rcq = NULL;
305 int num_sqe, num_rqe, rv = 0;
308 siw_dbg(base_dev, "create new QP\n");
310 if (attrs->create_flags)
311 return ERR_PTR(-EOPNOTSUPP);
313 if (atomic_inc_return(&sdev->num_qp) > SIW_MAX_QP) {
314 siw_dbg(base_dev, "too many QP's\n");
318 if (attrs->qp_type != IB_QPT_RC) {
319 siw_dbg(base_dev, "only RC QP's supported\n");
323 if ((attrs->cap.max_send_wr > SIW_MAX_QP_WR) ||
324 (attrs->cap.max_recv_wr > SIW_MAX_QP_WR) ||
325 (attrs->cap.max_send_sge > SIW_MAX_SGE) ||
326 (attrs->cap.max_recv_sge > SIW_MAX_SGE)) {
327 siw_dbg(base_dev, "QP size error\n");
331 if (attrs->cap.max_inline_data > SIW_MAX_INLINE) {
332 siw_dbg(base_dev, "max inline send: %d > %d\n",
333 attrs->cap.max_inline_data, (int)SIW_MAX_INLINE);
338 * NOTE: we allow for zero element SQ and RQ WQE's SGL's
339 * but not for a QP unable to hold any WQE (SQ + RQ)
341 if (attrs->cap.max_send_wr + attrs->cap.max_recv_wr == 0) {
342 siw_dbg(base_dev, "QP must have send or receive queue\n");
346 scq = to_siw_cq(attrs->send_cq);
347 rcq = to_siw_cq(attrs->recv_cq);
349 if (!scq || (!rcq && !attrs->srq)) {
350 siw_dbg(base_dev, "send CQ or receive CQ invalid\n");
354 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
359 init_rwsem(&qp->state_lock);
360 spin_lock_init(&qp->sq_lock);
361 spin_lock_init(&qp->rq_lock);
362 spin_lock_init(&qp->orq_lock);
364 rv = siw_qp_add(sdev, qp);
368 /* All queue indices are derived from modulo operations
369 * on a free running 'get' (consumer) and 'put' (producer)
370 * unsigned counter. Having queue sizes at power of two
371 * avoids handling counter wrap around.
373 num_sqe = roundup_pow_of_two(attrs->cap.max_send_wr);
374 num_rqe = roundup_pow_of_two(attrs->cap.max_recv_wr);
377 qp->sendq = vmalloc_user(num_sqe * sizeof(struct siw_sqe));
379 qp->sendq = vzalloc(num_sqe * sizeof(struct siw_sqe));
381 if (qp->sendq == NULL) {
382 siw_dbg(base_dev, "SQ size %d alloc failed\n", num_sqe);
386 if (attrs->sq_sig_type != IB_SIGNAL_REQ_WR) {
387 if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
388 qp->attrs.flags |= SIW_SIGNAL_ALL_WR;
401 * Verbs 6.3.7: ignore RQ size, if SRQ present
402 * Verbs 6.3.5: do not check PD of SRQ against PD of QP
404 qp->srq = to_siw_srq(attrs->srq);
405 qp->attrs.rq_size = 0;
406 siw_dbg(base_dev, "QP [%u]: SRQ attached\n",
408 } else if (num_rqe) {
411 vmalloc_user(num_rqe * sizeof(struct siw_rqe));
413 qp->recvq = vzalloc(num_rqe * sizeof(struct siw_rqe));
415 if (qp->recvq == NULL) {
416 siw_dbg(base_dev, "RQ size %d alloc failed\n", num_rqe);
420 qp->attrs.rq_size = num_rqe;
422 qp->attrs.sq_size = num_sqe;
423 qp->attrs.sq_max_sges = attrs->cap.max_send_sge;
424 qp->attrs.rq_max_sges = attrs->cap.max_recv_sge;
426 /* Make those two tunables fixed for now. */
427 qp->tx_ctx.gso_seg_limit = 1;
428 qp->tx_ctx.zcopy_tx = zcopy_tx;
430 qp->attrs.state = SIW_QP_STATE_IDLE;
433 struct siw_uresp_create_qp uresp = {};
435 uresp.num_sqe = num_sqe;
436 uresp.num_rqe = num_rqe;
437 uresp.qp_id = qp_id(qp);
440 length = num_sqe * sizeof(struct siw_sqe);
442 siw_mmap_entry_insert(uctx, qp->sendq,
443 length, &uresp.sq_key);
451 length = num_rqe * sizeof(struct siw_rqe);
453 siw_mmap_entry_insert(uctx, qp->recvq,
454 length, &uresp.rq_key);
456 uresp.sq_key = SIW_INVAL_UOBJ_KEY;
462 if (udata->outlen < sizeof(uresp)) {
466 rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
470 qp->tx_cpu = siw_get_tx_cpu(sdev);
471 if (qp->tx_cpu < 0) {
475 INIT_LIST_HEAD(&qp->devq);
476 spin_lock_irqsave(&sdev->lock, flags);
477 list_add_tail(&qp->devq, &sdev->qp_list);
478 spin_unlock_irqrestore(&sdev->lock, flags);
483 xa_erase(&sdev->qp_xa, qp_id(qp));
487 rdma_user_mmap_entry_remove(qp->sq_entry);
488 rdma_user_mmap_entry_remove(qp->rq_entry);
494 atomic_dec(&sdev->num_qp);
500 * Minimum siw_query_qp() verb interface.
502 * @qp_attr_mask is not used but all available information is provided
504 int siw_query_qp(struct ib_qp *base_qp, struct ib_qp_attr *qp_attr,
505 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
508 struct siw_device *sdev;
510 if (base_qp && qp_attr && qp_init_attr) {
511 qp = to_siw_qp(base_qp);
512 sdev = to_siw_dev(base_qp->device);
516 qp_attr->cap.max_inline_data = SIW_MAX_INLINE;
517 qp_attr->cap.max_send_wr = qp->attrs.sq_size;
518 qp_attr->cap.max_send_sge = qp->attrs.sq_max_sges;
519 qp_attr->cap.max_recv_wr = qp->attrs.rq_size;
520 qp_attr->cap.max_recv_sge = qp->attrs.rq_max_sges;
521 qp_attr->path_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
522 qp_attr->max_rd_atomic = qp->attrs.irq_size;
523 qp_attr->max_dest_rd_atomic = qp->attrs.orq_size;
525 qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE |
526 IB_ACCESS_REMOTE_WRITE |
527 IB_ACCESS_REMOTE_READ;
529 qp_init_attr->qp_type = base_qp->qp_type;
530 qp_init_attr->send_cq = base_qp->send_cq;
531 qp_init_attr->recv_cq = base_qp->recv_cq;
532 qp_init_attr->srq = base_qp->srq;
534 qp_init_attr->cap = qp_attr->cap;
539 int siw_verbs_modify_qp(struct ib_qp *base_qp, struct ib_qp_attr *attr,
540 int attr_mask, struct ib_udata *udata)
542 struct siw_qp_attrs new_attrs;
543 enum siw_qp_attr_mask siw_attr_mask = 0;
544 struct siw_qp *qp = to_siw_qp(base_qp);
550 if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
553 memset(&new_attrs, 0, sizeof(new_attrs));
555 if (attr_mask & IB_QP_ACCESS_FLAGS) {
556 siw_attr_mask = SIW_QP_ATTR_ACCESS_FLAGS;
558 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
559 new_attrs.flags |= SIW_RDMA_READ_ENABLED;
560 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
561 new_attrs.flags |= SIW_RDMA_WRITE_ENABLED;
562 if (attr->qp_access_flags & IB_ACCESS_MW_BIND)
563 new_attrs.flags |= SIW_RDMA_BIND_ENABLED;
565 if (attr_mask & IB_QP_STATE) {
566 siw_dbg_qp(qp, "desired IB QP state: %s\n",
567 ib_qp_state_to_string[attr->qp_state]);
569 new_attrs.state = ib_qp_state_to_siw_qp_state[attr->qp_state];
571 if (new_attrs.state > SIW_QP_STATE_RTS)
572 qp->tx_ctx.tx_suspend = 1;
574 siw_attr_mask |= SIW_QP_ATTR_STATE;
579 down_write(&qp->state_lock);
581 rv = siw_qp_modify(qp, &new_attrs, siw_attr_mask);
583 up_write(&qp->state_lock);
588 int siw_destroy_qp(struct ib_qp *base_qp, struct ib_udata *udata)
590 struct siw_qp *qp = to_siw_qp(base_qp);
591 struct siw_ucontext *uctx =
592 rdma_udata_to_drv_context(udata, struct siw_ucontext,
594 struct siw_qp_attrs qp_attrs;
596 siw_dbg_qp(qp, "state %d\n", qp->attrs.state);
599 * Mark QP as in process of destruction to prevent from
600 * any async callbacks to RDMA core
602 qp->attrs.flags |= SIW_QP_IN_DESTROY;
603 qp->rx_stream.rx_suspend = 1;
606 rdma_user_mmap_entry_remove(qp->sq_entry);
607 rdma_user_mmap_entry_remove(qp->rq_entry);
610 down_write(&qp->state_lock);
612 qp_attrs.state = SIW_QP_STATE_ERROR;
613 siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE);
616 siw_cep_put(qp->cep);
619 up_write(&qp->state_lock);
621 kfree(qp->tx_ctx.mpa_crc_hd);
622 kfree(qp->rx_stream.mpa_crc_hd);
624 qp->scq = qp->rcq = NULL;
632 * siw_copy_inline_sgl()
634 * Prepare sgl of inlined data for sending. For userland callers
635 * function checks if given buffer addresses and len's are within
636 * process context bounds.
637 * Data from all provided sge's are copied together into the wqe,
638 * referenced by a single sge.
640 static int siw_copy_inline_sgl(const struct ib_send_wr *core_wr,
643 struct ib_sge *core_sge = core_wr->sg_list;
644 void *kbuf = &sqe->sge[1];
645 int num_sge = core_wr->num_sge, bytes = 0;
647 sqe->sge[0].laddr = (uintptr_t)kbuf;
648 sqe->sge[0].lkey = 0;
651 if (!core_sge->length) {
655 bytes += core_sge->length;
656 if (bytes > SIW_MAX_INLINE) {
660 memcpy(kbuf, (void *)(uintptr_t)core_sge->addr,
663 kbuf += core_sge->length;
666 sqe->sge[0].length = bytes > 0 ? bytes : 0;
667 sqe->num_sge = bytes > 0 ? 1 : 0;
672 /* Complete SQ WR's without processing */
673 static int siw_sq_flush_wr(struct siw_qp *qp, const struct ib_send_wr *wr,
674 const struct ib_send_wr **bad_wr)
676 struct siw_sqe sqe = {};
681 sqe.opcode = wr->opcode;
682 rv = siw_sqe_complete(qp, &sqe, 0, SIW_WC_WR_FLUSH_ERR);
693 /* Complete RQ WR's without processing */
694 static int siw_rq_flush_wr(struct siw_qp *qp, const struct ib_recv_wr *wr,
695 const struct ib_recv_wr **bad_wr)
697 struct siw_rqe rqe = {};
702 rv = siw_rqe_complete(qp, &rqe, 0, 0, SIW_WC_WR_FLUSH_ERR);
716 * Post a list of S-WR's to a SQ.
718 * @base_qp: Base QP contained in siw QP
719 * @wr: Null terminated list of user WR's
720 * @bad_wr: Points to failing WR in case of synchronous failure.
722 int siw_post_send(struct ib_qp *base_qp, const struct ib_send_wr *wr,
723 const struct ib_send_wr **bad_wr)
725 struct siw_qp *qp = to_siw_qp(base_qp);
726 struct siw_wqe *wqe = tx_wqe(qp);
731 if (wr && !rdma_is_kernel_res(&qp->base_qp.res)) {
732 siw_dbg_qp(qp, "wr must be empty for user mapped sq\n");
738 * Try to acquire QP state lock. Must be non-blocking
739 * to accommodate kernel clients needs.
741 if (!down_read_trylock(&qp->state_lock)) {
742 if (qp->attrs.state == SIW_QP_STATE_ERROR) {
744 * ERROR state is final, so we can be sure
745 * this state will not change as long as the QP
748 * This handles an ib_drain_sq() call with
749 * a concurrent request to set the QP state
752 rv = siw_sq_flush_wr(qp, wr, bad_wr);
754 siw_dbg_qp(qp, "QP locked, state %d\n",
761 if (unlikely(qp->attrs.state != SIW_QP_STATE_RTS)) {
762 if (qp->attrs.state == SIW_QP_STATE_ERROR) {
764 * Immediately flush this WR to CQ, if QP
765 * is in ERROR state. SQ is guaranteed to
766 * be empty, so WR complets in-order.
768 * Typically triggered by ib_drain_sq().
770 rv = siw_sq_flush_wr(qp, wr, bad_wr);
772 siw_dbg_qp(qp, "QP out of state %d\n",
777 up_read(&qp->state_lock);
780 spin_lock_irqsave(&qp->sq_lock, flags);
783 u32 idx = qp->sq_put % qp->attrs.sq_size;
784 struct siw_sqe *sqe = &qp->sendq[idx];
787 siw_dbg_qp(qp, "sq full\n");
791 if (wr->num_sge > qp->attrs.sq_max_sges) {
792 siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
798 if ((wr->send_flags & IB_SEND_SIGNALED) ||
799 (qp->attrs.flags & SIW_SIGNAL_ALL_WR))
800 sqe->flags |= SIW_WQE_SIGNALLED;
802 if (wr->send_flags & IB_SEND_FENCE)
803 sqe->flags |= SIW_WQE_READ_FENCE;
805 switch (wr->opcode) {
807 case IB_WR_SEND_WITH_INV:
808 if (wr->send_flags & IB_SEND_SOLICITED)
809 sqe->flags |= SIW_WQE_SOLICITED;
811 if (!(wr->send_flags & IB_SEND_INLINE)) {
812 siw_copy_sgl(wr->sg_list, sqe->sge,
814 sqe->num_sge = wr->num_sge;
816 rv = siw_copy_inline_sgl(wr, sqe);
821 sqe->flags |= SIW_WQE_INLINE;
824 if (wr->opcode == IB_WR_SEND)
825 sqe->opcode = SIW_OP_SEND;
827 sqe->opcode = SIW_OP_SEND_REMOTE_INV;
828 sqe->rkey = wr->ex.invalidate_rkey;
832 case IB_WR_RDMA_READ_WITH_INV:
833 case IB_WR_RDMA_READ:
835 * iWarp restricts RREAD sink to SGL containing
836 * 1 SGE only. we could relax to SGL with multiple
837 * elements referring the SAME ltag or even sending
838 * a private per-rreq tag referring to a checked
839 * local sgl with MULTIPLE ltag's.
841 if (unlikely(wr->num_sge != 1)) {
845 siw_copy_sgl(wr->sg_list, &sqe->sge[0], 1);
847 * NOTE: zero length RREAD is allowed!
849 sqe->raddr = rdma_wr(wr)->remote_addr;
850 sqe->rkey = rdma_wr(wr)->rkey;
853 if (wr->opcode == IB_WR_RDMA_READ)
854 sqe->opcode = SIW_OP_READ;
856 sqe->opcode = SIW_OP_READ_LOCAL_INV;
859 case IB_WR_RDMA_WRITE:
860 if (!(wr->send_flags & IB_SEND_INLINE)) {
861 siw_copy_sgl(wr->sg_list, &sqe->sge[0],
863 sqe->num_sge = wr->num_sge;
865 rv = siw_copy_inline_sgl(wr, sqe);
866 if (unlikely(rv < 0)) {
870 sqe->flags |= SIW_WQE_INLINE;
873 sqe->raddr = rdma_wr(wr)->remote_addr;
874 sqe->rkey = rdma_wr(wr)->rkey;
875 sqe->opcode = SIW_OP_WRITE;
879 sqe->base_mr = (uintptr_t)reg_wr(wr)->mr;
880 sqe->rkey = reg_wr(wr)->key;
881 sqe->access = reg_wr(wr)->access & IWARP_ACCESS_MASK;
882 sqe->opcode = SIW_OP_REG_MR;
885 case IB_WR_LOCAL_INV:
886 sqe->rkey = wr->ex.invalidate_rkey;
887 sqe->opcode = SIW_OP_INVAL_STAG;
891 siw_dbg_qp(qp, "ib wr type %d unsupported\n",
896 siw_dbg_qp(qp, "opcode %d, flags 0x%x, wr_id 0x%pK\n",
897 sqe->opcode, sqe->flags,
898 (void *)(uintptr_t)sqe->id);
900 if (unlikely(rv < 0))
903 /* make SQE only valid after completely written */
905 sqe->flags |= SIW_WQE_VALID;
912 * Send directly if SQ processing is not in progress.
913 * Eventual immediate errors (rv < 0) do not affect the involved
914 * RI resources (Verbs, 8.3.1) and thus do not prevent from SQ
915 * processing, if new work is already pending. But rv must be passed
918 if (wqe->wr_status != SIW_WR_IDLE) {
919 spin_unlock_irqrestore(&qp->sq_lock, flags);
920 goto skip_direct_sending;
922 rv = siw_activate_tx(qp);
923 spin_unlock_irqrestore(&qp->sq_lock, flags);
926 goto skip_direct_sending;
928 if (rdma_is_kernel_res(&qp->base_qp.res)) {
929 rv = siw_sq_start(qp);
931 qp->tx_ctx.in_syscall = 1;
933 if (siw_qp_sq_process(qp) != 0 && !(qp->tx_ctx.tx_suspend))
934 siw_qp_cm_drop(qp, 0);
936 qp->tx_ctx.in_syscall = 0;
940 up_read(&qp->state_lock);
947 siw_dbg_qp(qp, "error %d\n", rv);
956 * Post a list of R-WR's to a RQ.
958 * @base_qp: Base QP contained in siw QP
959 * @wr: Null terminated list of user WR's
960 * @bad_wr: Points to failing WR in case of synchronous failure.
962 int siw_post_receive(struct ib_qp *base_qp, const struct ib_recv_wr *wr,
963 const struct ib_recv_wr **bad_wr)
965 struct siw_qp *qp = to_siw_qp(base_qp);
971 return -EOPNOTSUPP; /* what else from errno.h? */
973 if (!rdma_is_kernel_res(&qp->base_qp.res)) {
974 siw_dbg_qp(qp, "no kernel post_recv for user mapped rq\n");
980 * Try to acquire QP state lock. Must be non-blocking
981 * to accommodate kernel clients needs.
983 if (!down_read_trylock(&qp->state_lock)) {
984 if (qp->attrs.state == SIW_QP_STATE_ERROR) {
986 * ERROR state is final, so we can be sure
987 * this state will not change as long as the QP
990 * This handles an ib_drain_rq() call with
991 * a concurrent request to set the QP state
994 rv = siw_rq_flush_wr(qp, wr, bad_wr);
996 siw_dbg_qp(qp, "QP locked, state %d\n",
1003 if (qp->attrs.state > SIW_QP_STATE_RTS) {
1004 if (qp->attrs.state == SIW_QP_STATE_ERROR) {
1006 * Immediately flush this WR to CQ, if QP
1007 * is in ERROR state. RQ is guaranteed to
1008 * be empty, so WR complets in-order.
1010 * Typically triggered by ib_drain_rq().
1012 rv = siw_rq_flush_wr(qp, wr, bad_wr);
1014 siw_dbg_qp(qp, "QP out of state %d\n",
1019 up_read(&qp->state_lock);
1023 * Serialize potentially multiple producers.
1024 * Not needed for single threaded consumer side.
1026 spin_lock_irqsave(&qp->rq_lock, flags);
1029 u32 idx = qp->rq_put % qp->attrs.rq_size;
1030 struct siw_rqe *rqe = &qp->recvq[idx];
1033 siw_dbg_qp(qp, "RQ full\n");
1037 if (wr->num_sge > qp->attrs.rq_max_sges) {
1038 siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
1042 rqe->id = wr->wr_id;
1043 rqe->num_sge = wr->num_sge;
1044 siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1046 /* make sure RQE is completely written before valid */
1049 rqe->flags = SIW_WQE_VALID;
1054 spin_unlock_irqrestore(&qp->rq_lock, flags);
1056 up_read(&qp->state_lock);
1059 siw_dbg_qp(qp, "error %d\n", rv);
1062 return rv > 0 ? 0 : rv;
1065 int siw_destroy_cq(struct ib_cq *base_cq, struct ib_udata *udata)
1067 struct siw_cq *cq = to_siw_cq(base_cq);
1068 struct siw_device *sdev = to_siw_dev(base_cq->device);
1069 struct siw_ucontext *ctx =
1070 rdma_udata_to_drv_context(udata, struct siw_ucontext,
1073 siw_dbg_cq(cq, "free CQ resources\n");
1078 rdma_user_mmap_entry_remove(cq->cq_entry);
1080 atomic_dec(&sdev->num_cq);
1089 * Populate CQ of requested size
1091 * @base_cq: CQ as allocated by RDMA midlayer
1092 * @attr: Initial CQ attributes
1093 * @udata: relates to user context
1096 int siw_create_cq(struct ib_cq *base_cq, const struct ib_cq_init_attr *attr,
1097 struct ib_udata *udata)
1099 struct siw_device *sdev = to_siw_dev(base_cq->device);
1100 struct siw_cq *cq = to_siw_cq(base_cq);
1101 int rv, size = attr->cqe;
1106 if (atomic_inc_return(&sdev->num_cq) > SIW_MAX_CQ) {
1107 siw_dbg(base_cq->device, "too many CQ's\n");
1111 if (size < 1 || size > sdev->attrs.max_cqe) {
1112 siw_dbg(base_cq->device, "CQ size error: %d\n", size);
1116 size = roundup_pow_of_two(size);
1117 cq->base_cq.cqe = size;
1121 cq->queue = vmalloc_user(size * sizeof(struct siw_cqe) +
1122 sizeof(struct siw_cq_ctrl));
1124 cq->queue = vzalloc(size * sizeof(struct siw_cqe) +
1125 sizeof(struct siw_cq_ctrl));
1127 if (cq->queue == NULL) {
1131 get_random_bytes(&cq->id, 4);
1132 siw_dbg(base_cq->device, "new CQ [%u]\n", cq->id);
1134 spin_lock_init(&cq->lock);
1136 cq->notify = (struct siw_cq_ctrl *)&cq->queue[size];
1139 struct siw_uresp_create_cq uresp = {};
1140 struct siw_ucontext *ctx =
1141 rdma_udata_to_drv_context(udata, struct siw_ucontext,
1143 size_t length = size * sizeof(struct siw_cqe) +
1144 sizeof(struct siw_cq_ctrl);
1147 siw_mmap_entry_insert(ctx, cq->queue,
1148 length, &uresp.cq_key);
1149 if (!cq->cq_entry) {
1154 uresp.cq_id = cq->id;
1155 uresp.num_cqe = size;
1157 if (udata->outlen < sizeof(uresp)) {
1161 rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1168 siw_dbg(base_cq->device, "CQ creation failed: %d", rv);
1170 if (cq && cq->queue) {
1171 struct siw_ucontext *ctx =
1172 rdma_udata_to_drv_context(udata, struct siw_ucontext,
1175 rdma_user_mmap_entry_remove(cq->cq_entry);
1178 atomic_dec(&sdev->num_cq);
1186 * Reap CQ entries if available and copy work completion status into
1187 * array of WC's provided by caller. Returns number of reaped CQE's.
1189 * @base_cq: Base CQ contained in siw CQ.
1190 * @num_cqe: Maximum number of CQE's to reap.
1191 * @wc: Array of work completions to be filled by siw.
1193 int siw_poll_cq(struct ib_cq *base_cq, int num_cqe, struct ib_wc *wc)
1195 struct siw_cq *cq = to_siw_cq(base_cq);
1198 for (i = 0; i < num_cqe; i++) {
1199 if (!siw_reap_cqe(cq, wc))
1207 * siw_req_notify_cq()
1209 * Request notification for new CQE's added to that CQ.
1211 * o SIW_CQ_NOTIFY_SOLICITED lets siw trigger a notification
1212 * event if a WQE with notification flag set enters the CQ
1213 * o SIW_CQ_NOTIFY_NEXT_COMP lets siw trigger a notification
1214 * event if a WQE enters the CQ.
1215 * o IB_CQ_REPORT_MISSED_EVENTS: return value will provide the
1216 * number of not reaped CQE's regardless of its notification
1217 * type and current or new CQ notification settings.
1219 * @base_cq: Base CQ contained in siw CQ.
1220 * @flags: Requested notification flags.
1222 int siw_req_notify_cq(struct ib_cq *base_cq, enum ib_cq_notify_flags flags)
1224 struct siw_cq *cq = to_siw_cq(base_cq);
1226 siw_dbg_cq(cq, "flags: 0x%02x\n", flags);
1228 if ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED)
1230 * Enable CQ event for next solicited completion.
1231 * and make it visible to all associated producers.
1233 smp_store_mb(cq->notify->flags, SIW_NOTIFY_SOLICITED);
1236 * Enable CQ event for any signalled completion.
1237 * and make it visible to all associated producers.
1239 smp_store_mb(cq->notify->flags, SIW_NOTIFY_ALL);
1241 if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1242 return cq->cq_put - cq->cq_get;
1250 * Release Memory Region.
1252 * @base_mr: Base MR contained in siw MR.
1253 * @udata: points to user context, unused.
1255 int siw_dereg_mr(struct ib_mr *base_mr, struct ib_udata *udata)
1257 struct siw_mr *mr = to_siw_mr(base_mr);
1258 struct siw_device *sdev = to_siw_dev(base_mr->device);
1260 siw_dbg_mem(mr->mem, "deregister MR\n");
1262 atomic_dec(&sdev->num_mr);
1264 siw_mr_drop_mem(mr);
1273 * Register Memory Region.
1275 * @pd: Protection Domain
1276 * @start: starting address of MR (virtual address)
1278 * @rnic_va: not used by siw
1279 * @rights: MR access rights
1280 * @udata: user buffer to communicate STag and Key.
1282 struct ib_mr *siw_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
1283 u64 rnic_va, int rights, struct ib_udata *udata)
1285 struct siw_mr *mr = NULL;
1286 struct siw_umem *umem = NULL;
1287 struct siw_ureq_reg_mr ureq;
1288 struct siw_device *sdev = to_siw_dev(pd->device);
1290 unsigned long mem_limit = rlimit(RLIMIT_MEMLOCK);
1293 siw_dbg_pd(pd, "start: 0x%pK, va: 0x%pK, len: %llu\n",
1294 (void *)(uintptr_t)start, (void *)(uintptr_t)rnic_va,
1295 (unsigned long long)len);
1297 if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1298 siw_dbg_pd(pd, "too many mr's\n");
1306 if (mem_limit != RLIM_INFINITY) {
1307 unsigned long num_pages =
1308 (PAGE_ALIGN(len + (start & ~PAGE_MASK))) >> PAGE_SHIFT;
1309 mem_limit >>= PAGE_SHIFT;
1311 if (num_pages > mem_limit - current->mm->locked_vm) {
1312 siw_dbg_pd(pd, "pages req %lu, max %lu, lock %lu\n",
1313 num_pages, mem_limit,
1314 current->mm->locked_vm);
1319 umem = siw_umem_get(start, len, ib_access_writable(rights));
1322 siw_dbg_pd(pd, "getting user memory failed: %d\n", rv);
1326 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1331 rv = siw_mr_add_mem(mr, pd, umem, start, len, rights);
1336 struct siw_uresp_reg_mr uresp = {};
1337 struct siw_mem *mem = mr->mem;
1339 if (udata->inlen < sizeof(ureq)) {
1343 rv = ib_copy_from_udata(&ureq, udata, sizeof(ureq));
1347 mr->base_mr.lkey |= ureq.stag_key;
1348 mr->base_mr.rkey |= ureq.stag_key;
1349 mem->stag |= ureq.stag_key;
1350 uresp.stag = mem->stag;
1352 if (udata->outlen < sizeof(uresp)) {
1356 rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1360 mr->mem->stag_valid = 1;
1362 return &mr->base_mr;
1365 atomic_dec(&sdev->num_mr);
1368 siw_mr_drop_mem(mr);
1372 siw_umem_release(umem, false);
1377 struct ib_mr *siw_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
1380 struct siw_device *sdev = to_siw_dev(pd->device);
1381 struct siw_mr *mr = NULL;
1382 struct siw_pbl *pbl = NULL;
1385 if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1386 siw_dbg_pd(pd, "too many mr's\n");
1390 if (mr_type != IB_MR_TYPE_MEM_REG) {
1391 siw_dbg_pd(pd, "mr type %d unsupported\n", mr_type);
1395 if (max_sge > SIW_MAX_SGE_PBL) {
1396 siw_dbg_pd(pd, "too many sge's: %d\n", max_sge);
1400 pbl = siw_pbl_alloc(max_sge);
1403 siw_dbg_pd(pd, "pbl allocation failed: %d\n", rv);
1407 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1412 rv = siw_mr_add_mem(mr, pd, pbl, 0, max_sge * PAGE_SIZE, 0);
1416 mr->mem->is_pbl = 1;
1418 siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1420 return &mr->base_mr;
1423 atomic_dec(&sdev->num_mr);
1429 siw_mr_drop_mem(mr);
1432 siw_dbg_pd(pd, "failed: %d\n", rv);
1437 /* Just used to count number of pages being mapped */
1438 static int siw_set_pbl_page(struct ib_mr *base_mr, u64 buf_addr)
1443 int siw_map_mr_sg(struct ib_mr *base_mr, struct scatterlist *sl, int num_sle,
1444 unsigned int *sg_off)
1446 struct scatterlist *slp;
1447 struct siw_mr *mr = to_siw_mr(base_mr);
1448 struct siw_mem *mem = mr->mem;
1449 struct siw_pbl *pbl = mem->pbl;
1450 struct siw_pble *pble;
1451 unsigned long pbl_size;
1455 siw_dbg_mem(mem, "no PBL allocated\n");
1460 if (pbl->max_buf < num_sle) {
1461 siw_dbg_mem(mem, "too many SGE's: %d > %d\n",
1462 mem->pbl->max_buf, num_sle);
1465 for_each_sg(sl, slp, num_sle, i) {
1466 if (sg_dma_len(slp) == 0) {
1467 siw_dbg_mem(mem, "empty SGE\n");
1471 pble->addr = sg_dma_address(slp);
1472 pble->size = sg_dma_len(slp);
1474 pbl_size = pble->size;
1477 /* Merge PBL entries if adjacent */
1478 if (pble->addr + pble->size == sg_dma_address(slp)) {
1479 pble->size += sg_dma_len(slp);
1483 pble->addr = sg_dma_address(slp);
1484 pble->size = sg_dma_len(slp);
1485 pble->pbl_off = pbl_size;
1487 pbl_size += sg_dma_len(slp);
1490 "sge[%d], size %u, addr 0x%p, total %lu\n",
1491 i, pble->size, (void *)(uintptr_t)pble->addr,
1494 rv = ib_sg_to_pages(base_mr, sl, num_sle, sg_off, siw_set_pbl_page);
1496 mem->len = base_mr->length;
1497 mem->va = base_mr->iova;
1499 "%llu bytes, start 0x%pK, %u SLE to %u entries\n",
1500 mem->len, (void *)(uintptr_t)mem->va, num_sle,
1509 * Create a (empty) DMA memory region, where no umem is attached.
1511 struct ib_mr *siw_get_dma_mr(struct ib_pd *pd, int rights)
1513 struct siw_device *sdev = to_siw_dev(pd->device);
1514 struct siw_mr *mr = NULL;
1517 if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1518 siw_dbg_pd(pd, "too many mr's\n");
1522 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1527 rv = siw_mr_add_mem(mr, pd, NULL, 0, ULONG_MAX, rights);
1531 mr->mem->stag_valid = 1;
1533 siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1535 return &mr->base_mr;
1541 atomic_dec(&sdev->num_mr);
1549 * Create Shared Receive Queue of attributes @init_attrs
1550 * within protection domain given by @pd.
1552 * @base_srq: Base SRQ contained in siw SRQ.
1553 * @init_attrs: SRQ init attributes.
1554 * @udata: points to user context
1556 int siw_create_srq(struct ib_srq *base_srq,
1557 struct ib_srq_init_attr *init_attrs, struct ib_udata *udata)
1559 struct siw_srq *srq = to_siw_srq(base_srq);
1560 struct ib_srq_attr *attrs = &init_attrs->attr;
1561 struct siw_device *sdev = to_siw_dev(base_srq->device);
1562 struct siw_ucontext *ctx =
1563 rdma_udata_to_drv_context(udata, struct siw_ucontext,
1567 if (init_attrs->srq_type != IB_SRQT_BASIC)
1570 if (atomic_inc_return(&sdev->num_srq) > SIW_MAX_SRQ) {
1571 siw_dbg_pd(base_srq->pd, "too many SRQ's\n");
1575 if (attrs->max_wr == 0 || attrs->max_wr > SIW_MAX_SRQ_WR ||
1576 attrs->max_sge > SIW_MAX_SGE || attrs->srq_limit > attrs->max_wr) {
1580 srq->max_sge = attrs->max_sge;
1581 srq->num_rqe = roundup_pow_of_two(attrs->max_wr);
1582 srq->limit = attrs->srq_limit;
1586 srq->is_kernel_res = !udata;
1590 vmalloc_user(srq->num_rqe * sizeof(struct siw_rqe));
1592 srq->recvq = vzalloc(srq->num_rqe * sizeof(struct siw_rqe));
1594 if (srq->recvq == NULL) {
1599 struct siw_uresp_create_srq uresp = {};
1600 size_t length = srq->num_rqe * sizeof(struct siw_rqe);
1603 siw_mmap_entry_insert(ctx, srq->recvq,
1604 length, &uresp.srq_key);
1605 if (!srq->srq_entry) {
1610 uresp.num_rqe = srq->num_rqe;
1612 if (udata->outlen < sizeof(uresp)) {
1616 rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1620 spin_lock_init(&srq->lock);
1622 siw_dbg_pd(base_srq->pd, "[SRQ]: success\n");
1629 rdma_user_mmap_entry_remove(srq->srq_entry);
1632 atomic_dec(&sdev->num_srq);
1640 * Modify SRQ. The caller may resize SRQ and/or set/reset notification
1641 * limit and (re)arm IB_EVENT_SRQ_LIMIT_REACHED notification.
1643 * NOTE: it is unclear if RDMA core allows for changing the MAX_SGE
1644 * parameter. siw_modify_srq() does not check the attrs->max_sge param.
1646 int siw_modify_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs,
1647 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
1649 struct siw_srq *srq = to_siw_srq(base_srq);
1650 unsigned long flags;
1653 spin_lock_irqsave(&srq->lock, flags);
1655 if (attr_mask & IB_SRQ_MAX_WR) {
1656 /* resize request not yet supported */
1660 if (attr_mask & IB_SRQ_LIMIT) {
1661 if (attrs->srq_limit) {
1662 if (unlikely(attrs->srq_limit > srq->num_rqe)) {
1670 srq->limit = attrs->srq_limit;
1673 spin_unlock_irqrestore(&srq->lock, flags);
1681 * Query SRQ attributes.
1683 int siw_query_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs)
1685 struct siw_srq *srq = to_siw_srq(base_srq);
1686 unsigned long flags;
1688 spin_lock_irqsave(&srq->lock, flags);
1690 attrs->max_wr = srq->num_rqe;
1691 attrs->max_sge = srq->max_sge;
1692 attrs->srq_limit = srq->limit;
1694 spin_unlock_irqrestore(&srq->lock, flags);
1703 * It is assumed that the SRQ is not referenced by any
1704 * QP anymore - the code trusts the RDMA core environment to keep track
1707 int siw_destroy_srq(struct ib_srq *base_srq, struct ib_udata *udata)
1709 struct siw_srq *srq = to_siw_srq(base_srq);
1710 struct siw_device *sdev = to_siw_dev(base_srq->device);
1711 struct siw_ucontext *ctx =
1712 rdma_udata_to_drv_context(udata, struct siw_ucontext,
1716 rdma_user_mmap_entry_remove(srq->srq_entry);
1718 atomic_dec(&sdev->num_srq);
1723 * siw_post_srq_recv()
1725 * Post a list of receive queue elements to SRQ.
1726 * NOTE: The function does not check or lock a certain SRQ state
1727 * during the post operation. The code simply trusts the
1728 * RDMA core environment.
1730 * @base_srq: Base SRQ contained in siw SRQ
1731 * @wr: List of R-WR's
1732 * @bad_wr: Updated to failing WR if posting fails.
1734 int siw_post_srq_recv(struct ib_srq *base_srq, const struct ib_recv_wr *wr,
1735 const struct ib_recv_wr **bad_wr)
1737 struct siw_srq *srq = to_siw_srq(base_srq);
1738 unsigned long flags;
1741 if (unlikely(!srq->is_kernel_res)) {
1742 siw_dbg_pd(base_srq->pd,
1743 "[SRQ]: no kernel post_recv for mapped srq\n");
1748 * Serialize potentially multiple producers.
1749 * Also needed to serialize potentially multiple
1752 spin_lock_irqsave(&srq->lock, flags);
1755 u32 idx = srq->rq_put % srq->num_rqe;
1756 struct siw_rqe *rqe = &srq->recvq[idx];
1759 siw_dbg_pd(base_srq->pd, "SRQ full\n");
1763 if (unlikely(wr->num_sge > srq->max_sge)) {
1764 siw_dbg_pd(base_srq->pd,
1765 "[SRQ]: too many sge's: %d\n", wr->num_sge);
1769 rqe->id = wr->wr_id;
1770 rqe->num_sge = wr->num_sge;
1771 siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1773 /* Make sure S-RQE is completely written before valid */
1776 rqe->flags = SIW_WQE_VALID;
1781 spin_unlock_irqrestore(&srq->lock, flags);
1783 if (unlikely(rv < 0)) {
1784 siw_dbg_pd(base_srq->pd, "[SRQ]: error %d\n", rv);
1790 void siw_qp_event(struct siw_qp *qp, enum ib_event_type etype)
1792 struct ib_event event;
1793 struct ib_qp *base_qp = &qp->base_qp;
1796 * Do not report asynchronous errors on QP which gets
1797 * destroyed via verbs interface (siw_destroy_qp())
1799 if (qp->attrs.flags & SIW_QP_IN_DESTROY)
1802 event.event = etype;
1803 event.device = base_qp->device;
1804 event.element.qp = base_qp;
1806 if (base_qp->event_handler) {
1807 siw_dbg_qp(qp, "reporting event %d\n", etype);
1808 base_qp->event_handler(&event, base_qp->qp_context);
1812 void siw_cq_event(struct siw_cq *cq, enum ib_event_type etype)
1814 struct ib_event event;
1815 struct ib_cq *base_cq = &cq->base_cq;
1817 event.event = etype;
1818 event.device = base_cq->device;
1819 event.element.cq = base_cq;
1821 if (base_cq->event_handler) {
1822 siw_dbg_cq(cq, "reporting CQ event %d\n", etype);
1823 base_cq->event_handler(&event, base_cq->cq_context);
1827 void siw_srq_event(struct siw_srq *srq, enum ib_event_type etype)
1829 struct ib_event event;
1830 struct ib_srq *base_srq = &srq->base_srq;
1832 event.event = etype;
1833 event.device = base_srq->device;
1834 event.element.srq = base_srq;
1836 if (base_srq->event_handler) {
1837 siw_dbg_pd(srq->base_srq.pd,
1838 "reporting SRQ event %d\n", etype);
1839 base_srq->event_handler(&event, base_srq->srq_context);
1843 void siw_port_event(struct siw_device *sdev, u8 port, enum ib_event_type etype)
1845 struct ib_event event;
1847 event.event = etype;
1848 event.device = &sdev->base_dev;
1849 event.element.port_num = port;
1851 siw_dbg(&sdev->base_dev, "reporting port event %d\n", etype);
1853 ib_dispatch_event(&event);