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>
11 #include <net/addrconf.h>
13 #include <rdma/iw_cm.h>
14 #include <rdma/ib_verbs.h>
15 #include <rdma/ib_user_verbs.h>
16 #include <rdma/uverbs_ioctl.h>
19 #include "siw_verbs.h"
22 static int ib_qp_state_to_siw_qp_state[IB_QPS_ERR + 1] = {
23 [IB_QPS_RESET] = SIW_QP_STATE_IDLE,
24 [IB_QPS_INIT] = SIW_QP_STATE_IDLE,
25 [IB_QPS_RTR] = SIW_QP_STATE_RTR,
26 [IB_QPS_RTS] = SIW_QP_STATE_RTS,
27 [IB_QPS_SQD] = SIW_QP_STATE_CLOSING,
28 [IB_QPS_SQE] = SIW_QP_STATE_TERMINATE,
29 [IB_QPS_ERR] = SIW_QP_STATE_ERROR
32 static char ib_qp_state_to_string[IB_QPS_ERR + 1][sizeof("RESET")] = {
33 [IB_QPS_RESET] = "RESET", [IB_QPS_INIT] = "INIT", [IB_QPS_RTR] = "RTR",
34 [IB_QPS_RTS] = "RTS", [IB_QPS_SQD] = "SQD", [IB_QPS_SQE] = "SQE",
38 void siw_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
40 struct siw_user_mmap_entry *entry = to_siw_mmap_entry(rdma_entry);
45 int siw_mmap(struct ib_ucontext *ctx, struct vm_area_struct *vma)
47 struct siw_ucontext *uctx = to_siw_ctx(ctx);
48 size_t size = vma->vm_end - vma->vm_start;
49 struct rdma_user_mmap_entry *rdma_entry;
50 struct siw_user_mmap_entry *entry;
54 * Must be page aligned
56 if (vma->vm_start & (PAGE_SIZE - 1)) {
57 pr_warn("siw: mmap not page aligned\n");
60 rdma_entry = rdma_user_mmap_entry_get(&uctx->base_ucontext, vma);
62 siw_dbg(&uctx->sdev->base_dev, "mmap lookup failed: %lu, %#zx\n",
66 entry = to_siw_mmap_entry(rdma_entry);
68 rv = remap_vmalloc_range(vma, entry->address, 0);
70 pr_warn("remap_vmalloc_range failed: %lu, %zu\n", vma->vm_pgoff,
75 rdma_user_mmap_entry_put(rdma_entry);
80 int siw_alloc_ucontext(struct ib_ucontext *base_ctx, struct ib_udata *udata)
82 struct siw_device *sdev = to_siw_dev(base_ctx->device);
83 struct siw_ucontext *ctx = to_siw_ctx(base_ctx);
84 struct siw_uresp_alloc_ctx uresp = {};
87 if (atomic_inc_return(&sdev->num_ctx) > SIW_MAX_CONTEXT) {
93 uresp.dev_id = sdev->vendor_part_id;
95 if (udata->outlen < sizeof(uresp)) {
99 rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
103 siw_dbg(base_ctx->device, "success. now %d context(s)\n",
104 atomic_read(&sdev->num_ctx));
109 atomic_dec(&sdev->num_ctx);
110 siw_dbg(base_ctx->device, "failure %d. now %d context(s)\n", rv,
111 atomic_read(&sdev->num_ctx));
116 void siw_dealloc_ucontext(struct ib_ucontext *base_ctx)
118 struct siw_ucontext *uctx = to_siw_ctx(base_ctx);
120 atomic_dec(&uctx->sdev->num_ctx);
123 int siw_query_device(struct ib_device *base_dev, struct ib_device_attr *attr,
124 struct ib_udata *udata)
126 struct siw_device *sdev = to_siw_dev(base_dev);
128 if (udata->inlen || udata->outlen)
131 memset(attr, 0, sizeof(*attr));
133 /* Revisit atomic caps if RFC 7306 gets supported */
134 attr->atomic_cap = 0;
135 attr->device_cap_flags =
136 IB_DEVICE_MEM_MGT_EXTENSIONS | IB_DEVICE_ALLOW_USER_UNREG;
137 attr->max_cq = sdev->attrs.max_cq;
138 attr->max_cqe = sdev->attrs.max_cqe;
139 attr->max_fast_reg_page_list_len = SIW_MAX_SGE_PBL;
140 attr->max_mr = sdev->attrs.max_mr;
141 attr->max_mw = sdev->attrs.max_mw;
142 attr->max_mr_size = ~0ull;
143 attr->max_pd = sdev->attrs.max_pd;
144 attr->max_qp = sdev->attrs.max_qp;
145 attr->max_qp_init_rd_atom = sdev->attrs.max_ird;
146 attr->max_qp_rd_atom = sdev->attrs.max_ord;
147 attr->max_qp_wr = sdev->attrs.max_qp_wr;
148 attr->max_recv_sge = sdev->attrs.max_sge;
149 attr->max_res_rd_atom = sdev->attrs.max_qp * sdev->attrs.max_ird;
150 attr->max_send_sge = sdev->attrs.max_sge;
151 attr->max_sge_rd = sdev->attrs.max_sge_rd;
152 attr->max_srq = sdev->attrs.max_srq;
153 attr->max_srq_sge = sdev->attrs.max_srq_sge;
154 attr->max_srq_wr = sdev->attrs.max_srq_wr;
155 attr->page_size_cap = PAGE_SIZE;
156 attr->vendor_id = SIW_VENDOR_ID;
157 attr->vendor_part_id = sdev->vendor_part_id;
159 addrconf_addr_eui48((u8 *)&attr->sys_image_guid,
160 sdev->netdev->dev_addr);
165 int siw_query_port(struct ib_device *base_dev, u32 port,
166 struct ib_port_attr *attr)
168 struct siw_device *sdev = to_siw_dev(base_dev);
171 memset(attr, 0, sizeof(*attr));
173 rv = ib_get_eth_speed(base_dev, port, &attr->active_speed,
174 &attr->active_width);
175 attr->gid_tbl_len = 1;
176 attr->max_msg_sz = -1;
177 attr->max_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
178 attr->active_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
179 attr->phys_state = sdev->state == IB_PORT_ACTIVE ?
180 IB_PORT_PHYS_STATE_LINK_UP : IB_PORT_PHYS_STATE_DISABLED;
181 attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_DEVICE_MGMT_SUP;
182 attr->state = sdev->state;
187 * attr->bad_pkey_cntr = 0;
188 * attr->qkey_viol_cntr = 0;
191 * attr->max_vl_num = 0;
193 * attr->subnet_timeout = 0;
194 * attr->init_type_repy = 0;
199 int siw_get_port_immutable(struct ib_device *base_dev, u32 port,
200 struct ib_port_immutable *port_immutable)
202 struct ib_port_attr attr;
203 int rv = siw_query_port(base_dev, port, &attr);
208 port_immutable->gid_tbl_len = attr.gid_tbl_len;
209 port_immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
214 int siw_query_gid(struct ib_device *base_dev, u32 port, int idx,
217 struct siw_device *sdev = to_siw_dev(base_dev);
219 /* subnet_prefix == interface_id == 0; */
220 memset(gid, 0, sizeof(*gid));
221 memcpy(&gid->raw[0], sdev->netdev->dev_addr, 6);
226 int siw_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
228 struct siw_device *sdev = to_siw_dev(pd->device);
230 if (atomic_inc_return(&sdev->num_pd) > SIW_MAX_PD) {
231 atomic_dec(&sdev->num_pd);
234 siw_dbg_pd(pd, "now %d PD's(s)\n", atomic_read(&sdev->num_pd));
239 int siw_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
241 struct siw_device *sdev = to_siw_dev(pd->device);
243 siw_dbg_pd(pd, "free PD\n");
244 atomic_dec(&sdev->num_pd);
248 void siw_qp_get_ref(struct ib_qp *base_qp)
250 siw_qp_get(to_siw_qp(base_qp));
253 void siw_qp_put_ref(struct ib_qp *base_qp)
255 siw_qp_put(to_siw_qp(base_qp));
258 static struct rdma_user_mmap_entry *
259 siw_mmap_entry_insert(struct siw_ucontext *uctx,
260 void *address, size_t length,
263 struct siw_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
266 *offset = SIW_INVAL_UOBJ_KEY;
270 entry->address = address;
272 rv = rdma_user_mmap_entry_insert(&uctx->base_ucontext,
280 *offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
282 return &entry->rdma_entry;
288 * Create QP of requested size on given device.
291 * @attrs: Initial QP attributes.
292 * @udata: used to provide QP ID, SQ and RQ size back to user.
295 int siw_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attrs,
296 struct ib_udata *udata)
298 struct ib_pd *pd = ibqp->pd;
299 struct siw_qp *qp = to_siw_qp(ibqp);
300 struct ib_device *base_dev = pd->device;
301 struct siw_device *sdev = to_siw_dev(base_dev);
302 struct siw_ucontext *uctx =
303 rdma_udata_to_drv_context(udata, struct siw_ucontext,
306 int num_sqe, num_rqe, rv = 0;
309 siw_dbg(base_dev, "create new QP\n");
311 if (attrs->create_flags)
314 if (atomic_inc_return(&sdev->num_qp) > SIW_MAX_QP) {
315 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");
347 if (!attrs->send_cq || (!attrs->recv_cq && !attrs->srq)) {
348 siw_dbg(base_dev, "send CQ or receive CQ invalid\n");
353 init_rwsem(&qp->state_lock);
354 spin_lock_init(&qp->sq_lock);
355 spin_lock_init(&qp->rq_lock);
356 spin_lock_init(&qp->orq_lock);
358 rv = siw_qp_add(sdev, qp);
362 num_sqe = attrs->cap.max_send_wr;
363 num_rqe = attrs->cap.max_recv_wr;
365 /* All queue indices are derived from modulo operations
366 * on a free running 'get' (consumer) and 'put' (producer)
367 * unsigned counter. Having queue sizes at power of two
368 * avoids handling counter wrap around.
371 num_sqe = roundup_pow_of_two(num_sqe);
373 /* Zero sized SQ is not supported */
378 num_rqe = roundup_pow_of_two(num_rqe);
381 qp->sendq = vmalloc_user(num_sqe * sizeof(struct siw_sqe));
383 qp->sendq = vzalloc(num_sqe * sizeof(struct siw_sqe));
385 if (qp->sendq == NULL) {
389 if (attrs->sq_sig_type != IB_SIGNAL_REQ_WR) {
390 if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
391 qp->attrs.flags |= SIW_SIGNAL_ALL_WR;
398 qp->scq = to_siw_cq(attrs->send_cq);
399 qp->rcq = to_siw_cq(attrs->recv_cq);
404 * Verbs 6.3.7: ignore RQ size, if SRQ present
405 * Verbs 6.3.5: do not check PD of SRQ against PD of QP
407 qp->srq = to_siw_srq(attrs->srq);
408 qp->attrs.rq_size = 0;
409 siw_dbg(base_dev, "QP [%u]: SRQ attached\n",
411 } else if (num_rqe) {
414 vmalloc_user(num_rqe * sizeof(struct siw_rqe));
416 qp->recvq = vzalloc(num_rqe * sizeof(struct siw_rqe));
418 if (qp->recvq == NULL) {
422 qp->attrs.rq_size = num_rqe;
424 qp->attrs.sq_size = num_sqe;
425 qp->attrs.sq_max_sges = attrs->cap.max_send_sge;
426 qp->attrs.rq_max_sges = attrs->cap.max_recv_sge;
428 /* Make those two tunables fixed for now. */
429 qp->tx_ctx.gso_seg_limit = 1;
430 qp->tx_ctx.zcopy_tx = zcopy_tx;
432 qp->attrs.state = SIW_QP_STATE_IDLE;
435 struct siw_uresp_create_qp uresp = {};
437 uresp.num_sqe = num_sqe;
438 uresp.num_rqe = num_rqe;
439 uresp.qp_id = qp_id(qp);
442 length = num_sqe * sizeof(struct siw_sqe);
444 siw_mmap_entry_insert(uctx, qp->sendq,
445 length, &uresp.sq_key);
453 length = num_rqe * sizeof(struct siw_rqe);
455 siw_mmap_entry_insert(uctx, qp->recvq,
456 length, &uresp.rq_key);
458 uresp.sq_key = SIW_INVAL_UOBJ_KEY;
464 if (udata->outlen < sizeof(uresp)) {
468 rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
472 qp->tx_cpu = siw_get_tx_cpu(sdev);
473 if (qp->tx_cpu < 0) {
477 INIT_LIST_HEAD(&qp->devq);
478 spin_lock_irqsave(&sdev->lock, flags);
479 list_add_tail(&qp->devq, &sdev->qp_list);
480 spin_unlock_irqrestore(&sdev->lock, flags);
485 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);
499 * Minimum siw_query_qp() verb interface.
501 * @qp_attr_mask is not used but all available information is provided
503 int siw_query_qp(struct ib_qp *base_qp, struct ib_qp_attr *qp_attr,
504 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
507 struct siw_device *sdev;
509 if (base_qp && qp_attr && qp_init_attr) {
510 qp = to_siw_qp(base_qp);
511 sdev = to_siw_dev(base_qp->device);
515 qp_attr->cap.max_inline_data = SIW_MAX_INLINE;
516 qp_attr->cap.max_send_wr = qp->attrs.sq_size;
517 qp_attr->cap.max_send_sge = qp->attrs.sq_max_sges;
518 qp_attr->cap.max_recv_wr = qp->attrs.rq_size;
519 qp_attr->cap.max_recv_sge = qp->attrs.rq_max_sges;
520 qp_attr->path_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
521 qp_attr->max_rd_atomic = qp->attrs.irq_size;
522 qp_attr->max_dest_rd_atomic = qp->attrs.orq_size;
524 qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE |
525 IB_ACCESS_REMOTE_WRITE |
526 IB_ACCESS_REMOTE_READ;
528 qp_init_attr->qp_type = base_qp->qp_type;
529 qp_init_attr->send_cq = base_qp->send_cq;
530 qp_init_attr->recv_cq = base_qp->recv_cq;
531 qp_init_attr->srq = base_qp->srq;
533 qp_init_attr->cap = qp_attr->cap;
538 int siw_verbs_modify_qp(struct ib_qp *base_qp, struct ib_qp_attr *attr,
539 int attr_mask, struct ib_udata *udata)
541 struct siw_qp_attrs new_attrs;
542 enum siw_qp_attr_mask siw_attr_mask = 0;
543 struct siw_qp *qp = to_siw_qp(base_qp);
549 if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
552 memset(&new_attrs, 0, sizeof(new_attrs));
554 if (attr_mask & IB_QP_ACCESS_FLAGS) {
555 siw_attr_mask = SIW_QP_ATTR_ACCESS_FLAGS;
557 if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
558 new_attrs.flags |= SIW_RDMA_READ_ENABLED;
559 if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
560 new_attrs.flags |= SIW_RDMA_WRITE_ENABLED;
561 if (attr->qp_access_flags & IB_ACCESS_MW_BIND)
562 new_attrs.flags |= SIW_RDMA_BIND_ENABLED;
564 if (attr_mask & IB_QP_STATE) {
565 siw_dbg_qp(qp, "desired IB QP state: %s\n",
566 ib_qp_state_to_string[attr->qp_state]);
568 new_attrs.state = ib_qp_state_to_siw_qp_state[attr->qp_state];
570 if (new_attrs.state > SIW_QP_STATE_RTS)
571 qp->tx_ctx.tx_suspend = 1;
573 siw_attr_mask |= SIW_QP_ATTR_STATE;
578 down_write(&qp->state_lock);
580 rv = siw_qp_modify(qp, &new_attrs, siw_attr_mask);
582 up_write(&qp->state_lock);
587 int siw_destroy_qp(struct ib_qp *base_qp, struct ib_udata *udata)
589 struct siw_qp *qp = to_siw_qp(base_qp);
590 struct siw_ucontext *uctx =
591 rdma_udata_to_drv_context(udata, struct siw_ucontext,
593 struct siw_qp_attrs qp_attrs;
595 siw_dbg_qp(qp, "state %d\n", qp->attrs.state);
598 * Mark QP as in process of destruction to prevent from
599 * any async callbacks to RDMA core
601 qp->attrs.flags |= SIW_QP_IN_DESTROY;
602 qp->rx_stream.rx_suspend = 1;
605 rdma_user_mmap_entry_remove(qp->sq_entry);
606 rdma_user_mmap_entry_remove(qp->rq_entry);
609 down_write(&qp->state_lock);
611 qp_attrs.state = SIW_QP_STATE_ERROR;
612 siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE);
615 siw_cep_put(qp->cep);
618 up_write(&qp->state_lock);
620 kfree(qp->tx_ctx.mpa_crc_hd);
621 kfree(qp->rx_stream.mpa_crc_hd);
623 qp->scq = qp->rcq = NULL;
631 * siw_copy_inline_sgl()
633 * Prepare sgl of inlined data for sending. For userland callers
634 * function checks if given buffer addresses and len's are within
635 * process context bounds.
636 * Data from all provided sge's are copied together into the wqe,
637 * referenced by a single sge.
639 static int siw_copy_inline_sgl(const struct ib_send_wr *core_wr,
642 struct ib_sge *core_sge = core_wr->sg_list;
643 void *kbuf = &sqe->sge[1];
644 int num_sge = core_wr->num_sge, bytes = 0;
646 sqe->sge[0].laddr = (uintptr_t)kbuf;
647 sqe->sge[0].lkey = 0;
650 if (!core_sge->length) {
654 bytes += core_sge->length;
655 if (bytes > SIW_MAX_INLINE) {
659 memcpy(kbuf, (void *)(uintptr_t)core_sge->addr,
662 kbuf += core_sge->length;
665 sqe->sge[0].length = max(bytes, 0);
666 sqe->num_sge = bytes > 0 ? 1 : 0;
671 /* Complete SQ WR's without processing */
672 static int siw_sq_flush_wr(struct siw_qp *qp, const struct ib_send_wr *wr,
673 const struct ib_send_wr **bad_wr)
675 struct siw_sqe sqe = {};
680 sqe.opcode = wr->opcode;
681 rv = siw_sqe_complete(qp, &sqe, 0, SIW_WC_WR_FLUSH_ERR);
692 /* Complete RQ WR's without processing */
693 static int siw_rq_flush_wr(struct siw_qp *qp, const struct ib_recv_wr *wr,
694 const struct ib_recv_wr **bad_wr)
696 struct siw_rqe rqe = {};
701 rv = siw_rqe_complete(qp, &rqe, 0, 0, SIW_WC_WR_FLUSH_ERR);
715 * Post a list of S-WR's to a SQ.
717 * @base_qp: Base QP contained in siw QP
718 * @wr: Null terminated list of user WR's
719 * @bad_wr: Points to failing WR in case of synchronous failure.
721 int siw_post_send(struct ib_qp *base_qp, const struct ib_send_wr *wr,
722 const struct ib_send_wr **bad_wr)
724 struct siw_qp *qp = to_siw_qp(base_qp);
725 struct siw_wqe *wqe = tx_wqe(qp);
730 if (wr && !rdma_is_kernel_res(&qp->base_qp.res)) {
731 siw_dbg_qp(qp, "wr must be empty for user mapped sq\n");
737 * Try to acquire QP state lock. Must be non-blocking
738 * to accommodate kernel clients needs.
740 if (!down_read_trylock(&qp->state_lock)) {
741 if (qp->attrs.state == SIW_QP_STATE_ERROR) {
743 * ERROR state is final, so we can be sure
744 * this state will not change as long as the QP
747 * This handles an ib_drain_sq() call with
748 * a concurrent request to set the QP state
751 rv = siw_sq_flush_wr(qp, wr, bad_wr);
753 siw_dbg_qp(qp, "QP locked, state %d\n",
760 if (unlikely(qp->attrs.state != SIW_QP_STATE_RTS)) {
761 if (qp->attrs.state == SIW_QP_STATE_ERROR) {
763 * Immediately flush this WR to CQ, if QP
764 * is in ERROR state. SQ is guaranteed to
765 * be empty, so WR complets in-order.
767 * Typically triggered by ib_drain_sq().
769 rv = siw_sq_flush_wr(qp, wr, bad_wr);
771 siw_dbg_qp(qp, "QP out of state %d\n",
776 up_read(&qp->state_lock);
779 spin_lock_irqsave(&qp->sq_lock, flags);
782 u32 idx = qp->sq_put % qp->attrs.sq_size;
783 struct siw_sqe *sqe = &qp->sendq[idx];
786 siw_dbg_qp(qp, "sq full\n");
790 if (wr->num_sge > qp->attrs.sq_max_sges) {
791 siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
797 if ((wr->send_flags & IB_SEND_SIGNALED) ||
798 (qp->attrs.flags & SIW_SIGNAL_ALL_WR))
799 sqe->flags |= SIW_WQE_SIGNALLED;
801 if (wr->send_flags & IB_SEND_FENCE)
802 sqe->flags |= SIW_WQE_READ_FENCE;
804 switch (wr->opcode) {
806 case IB_WR_SEND_WITH_INV:
807 if (wr->send_flags & IB_SEND_SOLICITED)
808 sqe->flags |= SIW_WQE_SOLICITED;
810 if (!(wr->send_flags & IB_SEND_INLINE)) {
811 siw_copy_sgl(wr->sg_list, sqe->sge,
813 sqe->num_sge = wr->num_sge;
815 rv = siw_copy_inline_sgl(wr, sqe);
820 sqe->flags |= SIW_WQE_INLINE;
823 if (wr->opcode == IB_WR_SEND)
824 sqe->opcode = SIW_OP_SEND;
826 sqe->opcode = SIW_OP_SEND_REMOTE_INV;
827 sqe->rkey = wr->ex.invalidate_rkey;
831 case IB_WR_RDMA_READ_WITH_INV:
832 case IB_WR_RDMA_READ:
834 * iWarp restricts RREAD sink to SGL containing
835 * 1 SGE only. we could relax to SGL with multiple
836 * elements referring the SAME ltag or even sending
837 * a private per-rreq tag referring to a checked
838 * local sgl with MULTIPLE ltag's.
840 if (unlikely(wr->num_sge != 1)) {
844 siw_copy_sgl(wr->sg_list, &sqe->sge[0], 1);
846 * NOTE: zero length RREAD is allowed!
848 sqe->raddr = rdma_wr(wr)->remote_addr;
849 sqe->rkey = rdma_wr(wr)->rkey;
852 if (wr->opcode == IB_WR_RDMA_READ)
853 sqe->opcode = SIW_OP_READ;
855 sqe->opcode = SIW_OP_READ_LOCAL_INV;
858 case IB_WR_RDMA_WRITE:
859 if (!(wr->send_flags & IB_SEND_INLINE)) {
860 siw_copy_sgl(wr->sg_list, &sqe->sge[0],
862 sqe->num_sge = wr->num_sge;
864 rv = siw_copy_inline_sgl(wr, sqe);
865 if (unlikely(rv < 0)) {
869 sqe->flags |= SIW_WQE_INLINE;
872 sqe->raddr = rdma_wr(wr)->remote_addr;
873 sqe->rkey = rdma_wr(wr)->rkey;
874 sqe->opcode = SIW_OP_WRITE;
878 sqe->base_mr = (uintptr_t)reg_wr(wr)->mr;
879 sqe->rkey = reg_wr(wr)->key;
880 sqe->access = reg_wr(wr)->access & IWARP_ACCESS_MASK;
881 sqe->opcode = SIW_OP_REG_MR;
884 case IB_WR_LOCAL_INV:
885 sqe->rkey = wr->ex.invalidate_rkey;
886 sqe->opcode = SIW_OP_INVAL_STAG;
890 siw_dbg_qp(qp, "ib wr type %d unsupported\n",
895 siw_dbg_qp(qp, "opcode %d, flags 0x%x, wr_id 0x%pK\n",
896 sqe->opcode, sqe->flags,
897 (void *)(uintptr_t)sqe->id);
899 if (unlikely(rv < 0))
902 /* make SQE only valid after completely written */
904 sqe->flags |= SIW_WQE_VALID;
911 * Send directly if SQ processing is not in progress.
912 * Eventual immediate errors (rv < 0) do not affect the involved
913 * RI resources (Verbs, 8.3.1) and thus do not prevent from SQ
914 * processing, if new work is already pending. But rv must be passed
917 if (wqe->wr_status != SIW_WR_IDLE) {
918 spin_unlock_irqrestore(&qp->sq_lock, flags);
919 goto skip_direct_sending;
921 rv = siw_activate_tx(qp);
922 spin_unlock_irqrestore(&qp->sq_lock, flags);
925 goto skip_direct_sending;
927 if (rdma_is_kernel_res(&qp->base_qp.res)) {
928 rv = siw_sq_start(qp);
930 qp->tx_ctx.in_syscall = 1;
932 if (siw_qp_sq_process(qp) != 0 && !(qp->tx_ctx.tx_suspend))
933 siw_qp_cm_drop(qp, 0);
935 qp->tx_ctx.in_syscall = 0;
939 up_read(&qp->state_lock);
946 siw_dbg_qp(qp, "error %d\n", rv);
955 * Post a list of R-WR's to a RQ.
957 * @base_qp: Base QP contained in siw QP
958 * @wr: Null terminated list of user WR's
959 * @bad_wr: Points to failing WR in case of synchronous failure.
961 int siw_post_receive(struct ib_qp *base_qp, const struct ib_recv_wr *wr,
962 const struct ib_recv_wr **bad_wr)
964 struct siw_qp *qp = to_siw_qp(base_qp);
968 if (qp->srq || qp->attrs.rq_size == 0) {
972 if (!rdma_is_kernel_res(&qp->base_qp.res)) {
973 siw_dbg_qp(qp, "no kernel post_recv for user mapped rq\n");
979 * Try to acquire QP state lock. Must be non-blocking
980 * to accommodate kernel clients needs.
982 if (!down_read_trylock(&qp->state_lock)) {
983 if (qp->attrs.state == SIW_QP_STATE_ERROR) {
985 * ERROR state is final, so we can be sure
986 * this state will not change as long as the QP
989 * This handles an ib_drain_rq() call with
990 * a concurrent request to set the QP state
993 rv = siw_rq_flush_wr(qp, wr, bad_wr);
995 siw_dbg_qp(qp, "QP locked, state %d\n",
1002 if (qp->attrs.state > SIW_QP_STATE_RTS) {
1003 if (qp->attrs.state == SIW_QP_STATE_ERROR) {
1005 * Immediately flush this WR to CQ, if QP
1006 * is in ERROR state. RQ is guaranteed to
1007 * be empty, so WR complets in-order.
1009 * Typically triggered by ib_drain_rq().
1011 rv = siw_rq_flush_wr(qp, wr, bad_wr);
1013 siw_dbg_qp(qp, "QP out of state %d\n",
1018 up_read(&qp->state_lock);
1022 * Serialize potentially multiple producers.
1023 * Not needed for single threaded consumer side.
1025 spin_lock_irqsave(&qp->rq_lock, flags);
1028 u32 idx = qp->rq_put % qp->attrs.rq_size;
1029 struct siw_rqe *rqe = &qp->recvq[idx];
1032 siw_dbg_qp(qp, "RQ full\n");
1036 if (wr->num_sge > qp->attrs.rq_max_sges) {
1037 siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
1041 rqe->id = wr->wr_id;
1042 rqe->num_sge = wr->num_sge;
1043 siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1045 /* make sure RQE is completely written before valid */
1048 rqe->flags = SIW_WQE_VALID;
1053 spin_unlock_irqrestore(&qp->rq_lock, flags);
1055 up_read(&qp->state_lock);
1058 siw_dbg_qp(qp, "error %d\n", rv);
1061 return rv > 0 ? 0 : rv;
1064 int siw_destroy_cq(struct ib_cq *base_cq, struct ib_udata *udata)
1066 struct siw_cq *cq = to_siw_cq(base_cq);
1067 struct siw_device *sdev = to_siw_dev(base_cq->device);
1068 struct siw_ucontext *ctx =
1069 rdma_udata_to_drv_context(udata, struct siw_ucontext,
1072 siw_dbg_cq(cq, "free CQ resources\n");
1077 rdma_user_mmap_entry_remove(cq->cq_entry);
1079 atomic_dec(&sdev->num_cq);
1088 * Populate CQ of requested size
1090 * @base_cq: CQ as allocated by RDMA midlayer
1091 * @attr: Initial CQ attributes
1092 * @udata: relates to user context
1095 int siw_create_cq(struct ib_cq *base_cq, const struct ib_cq_init_attr *attr,
1096 struct ib_udata *udata)
1098 struct siw_device *sdev = to_siw_dev(base_cq->device);
1099 struct siw_cq *cq = to_siw_cq(base_cq);
1100 int rv, size = attr->cqe;
1105 if (atomic_inc_return(&sdev->num_cq) > SIW_MAX_CQ) {
1106 siw_dbg(base_cq->device, "too many CQ's\n");
1110 if (size < 1 || size > sdev->attrs.max_cqe) {
1111 siw_dbg(base_cq->device, "CQ size error: %d\n", size);
1115 size = roundup_pow_of_two(size);
1116 cq->base_cq.cqe = size;
1120 cq->queue = vmalloc_user(size * sizeof(struct siw_cqe) +
1121 sizeof(struct siw_cq_ctrl));
1123 cq->queue = vzalloc(size * sizeof(struct siw_cqe) +
1124 sizeof(struct siw_cq_ctrl));
1126 if (cq->queue == NULL) {
1130 get_random_bytes(&cq->id, 4);
1131 siw_dbg(base_cq->device, "new CQ [%u]\n", cq->id);
1133 spin_lock_init(&cq->lock);
1135 cq->notify = (struct siw_cq_ctrl *)&cq->queue[size];
1138 struct siw_uresp_create_cq uresp = {};
1139 struct siw_ucontext *ctx =
1140 rdma_udata_to_drv_context(udata, struct siw_ucontext,
1142 size_t length = size * sizeof(struct siw_cqe) +
1143 sizeof(struct siw_cq_ctrl);
1146 siw_mmap_entry_insert(ctx, cq->queue,
1147 length, &uresp.cq_key);
1148 if (!cq->cq_entry) {
1153 uresp.cq_id = cq->id;
1154 uresp.num_cqe = size;
1156 if (udata->outlen < sizeof(uresp)) {
1160 rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1167 siw_dbg(base_cq->device, "CQ creation failed: %d", rv);
1169 if (cq && cq->queue) {
1170 struct siw_ucontext *ctx =
1171 rdma_udata_to_drv_context(udata, struct siw_ucontext,
1174 rdma_user_mmap_entry_remove(cq->cq_entry);
1177 atomic_dec(&sdev->num_cq);
1185 * Reap CQ entries if available and copy work completion status into
1186 * array of WC's provided by caller. Returns number of reaped CQE's.
1188 * @base_cq: Base CQ contained in siw CQ.
1189 * @num_cqe: Maximum number of CQE's to reap.
1190 * @wc: Array of work completions to be filled by siw.
1192 int siw_poll_cq(struct ib_cq *base_cq, int num_cqe, struct ib_wc *wc)
1194 struct siw_cq *cq = to_siw_cq(base_cq);
1197 for (i = 0; i < num_cqe; i++) {
1198 if (!siw_reap_cqe(cq, wc))
1206 * siw_req_notify_cq()
1208 * Request notification for new CQE's added to that CQ.
1210 * o SIW_CQ_NOTIFY_SOLICITED lets siw trigger a notification
1211 * event if a WQE with notification flag set enters the CQ
1212 * o SIW_CQ_NOTIFY_NEXT_COMP lets siw trigger a notification
1213 * event if a WQE enters the CQ.
1214 * o IB_CQ_REPORT_MISSED_EVENTS: return value will provide the
1215 * number of not reaped CQE's regardless of its notification
1216 * type and current or new CQ notification settings.
1218 * @base_cq: Base CQ contained in siw CQ.
1219 * @flags: Requested notification flags.
1221 int siw_req_notify_cq(struct ib_cq *base_cq, enum ib_cq_notify_flags flags)
1223 struct siw_cq *cq = to_siw_cq(base_cq);
1225 siw_dbg_cq(cq, "flags: 0x%02x\n", flags);
1227 if ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED)
1229 * Enable CQ event for next solicited completion.
1230 * and make it visible to all associated producers.
1232 smp_store_mb(cq->notify->flags, SIW_NOTIFY_SOLICITED);
1235 * Enable CQ event for any signalled completion.
1236 * and make it visible to all associated producers.
1238 smp_store_mb(cq->notify->flags, SIW_NOTIFY_ALL);
1240 if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1241 return cq->cq_put - cq->cq_get;
1249 * Release Memory Region.
1251 * @base_mr: Base MR contained in siw MR.
1252 * @udata: points to user context, unused.
1254 int siw_dereg_mr(struct ib_mr *base_mr, struct ib_udata *udata)
1256 struct siw_mr *mr = to_siw_mr(base_mr);
1257 struct siw_device *sdev = to_siw_dev(base_mr->device);
1259 siw_dbg_mem(mr->mem, "deregister MR\n");
1261 atomic_dec(&sdev->num_mr);
1263 siw_mr_drop_mem(mr);
1272 * Register Memory Region.
1274 * @pd: Protection Domain
1275 * @start: starting address of MR (virtual address)
1277 * @rnic_va: not used by siw
1278 * @rights: MR access rights
1279 * @udata: user buffer to communicate STag and Key.
1281 struct ib_mr *siw_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
1282 u64 rnic_va, int rights, struct ib_udata *udata)
1284 struct siw_mr *mr = NULL;
1285 struct siw_umem *umem = NULL;
1286 struct siw_ureq_reg_mr ureq;
1287 struct siw_device *sdev = to_siw_dev(pd->device);
1289 unsigned long mem_limit = rlimit(RLIMIT_MEMLOCK);
1292 siw_dbg_pd(pd, "start: 0x%pK, va: 0x%pK, len: %llu\n",
1293 (void *)(uintptr_t)start, (void *)(uintptr_t)rnic_va,
1294 (unsigned long long)len);
1296 if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1297 siw_dbg_pd(pd, "too many mr's\n");
1305 if (mem_limit != RLIM_INFINITY) {
1306 unsigned long num_pages =
1307 (PAGE_ALIGN(len + (start & ~PAGE_MASK))) >> PAGE_SHIFT;
1308 mem_limit >>= PAGE_SHIFT;
1310 if (num_pages > mem_limit - current->mm->locked_vm) {
1311 siw_dbg_pd(pd, "pages req %lu, max %lu, lock %lu\n",
1312 num_pages, mem_limit,
1313 current->mm->locked_vm);
1318 umem = siw_umem_get(start, len, ib_access_writable(rights));
1321 siw_dbg_pd(pd, "getting user memory failed: %d\n", rv);
1325 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1330 rv = siw_mr_add_mem(mr, pd, umem, start, len, rights);
1335 struct siw_uresp_reg_mr uresp = {};
1336 struct siw_mem *mem = mr->mem;
1338 if (udata->inlen < sizeof(ureq)) {
1342 rv = ib_copy_from_udata(&ureq, udata, sizeof(ureq));
1346 mr->base_mr.lkey |= ureq.stag_key;
1347 mr->base_mr.rkey |= ureq.stag_key;
1348 mem->stag |= ureq.stag_key;
1349 uresp.stag = mem->stag;
1351 if (udata->outlen < sizeof(uresp)) {
1355 rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1359 mr->mem->stag_valid = 1;
1361 return &mr->base_mr;
1364 atomic_dec(&sdev->num_mr);
1367 siw_mr_drop_mem(mr);
1371 siw_umem_release(umem, false);
1376 struct ib_mr *siw_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
1379 struct siw_device *sdev = to_siw_dev(pd->device);
1380 struct siw_mr *mr = NULL;
1381 struct siw_pbl *pbl = NULL;
1384 if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1385 siw_dbg_pd(pd, "too many mr's\n");
1389 if (mr_type != IB_MR_TYPE_MEM_REG) {
1390 siw_dbg_pd(pd, "mr type %d unsupported\n", mr_type);
1394 if (max_sge > SIW_MAX_SGE_PBL) {
1395 siw_dbg_pd(pd, "too many sge's: %d\n", max_sge);
1399 pbl = siw_pbl_alloc(max_sge);
1402 siw_dbg_pd(pd, "pbl allocation failed: %d\n", rv);
1406 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1411 rv = siw_mr_add_mem(mr, pd, pbl, 0, max_sge * PAGE_SIZE, 0);
1415 mr->mem->is_pbl = 1;
1417 siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1419 return &mr->base_mr;
1422 atomic_dec(&sdev->num_mr);
1428 siw_mr_drop_mem(mr);
1431 siw_dbg_pd(pd, "failed: %d\n", rv);
1436 /* Just used to count number of pages being mapped */
1437 static int siw_set_pbl_page(struct ib_mr *base_mr, u64 buf_addr)
1442 int siw_map_mr_sg(struct ib_mr *base_mr, struct scatterlist *sl, int num_sle,
1443 unsigned int *sg_off)
1445 struct scatterlist *slp;
1446 struct siw_mr *mr = to_siw_mr(base_mr);
1447 struct siw_mem *mem = mr->mem;
1448 struct siw_pbl *pbl = mem->pbl;
1449 struct siw_pble *pble;
1450 unsigned long pbl_size;
1454 siw_dbg_mem(mem, "no PBL allocated\n");
1459 if (pbl->max_buf < num_sle) {
1460 siw_dbg_mem(mem, "too many SGE's: %d > %d\n",
1461 mem->pbl->max_buf, num_sle);
1464 for_each_sg(sl, slp, num_sle, i) {
1465 if (sg_dma_len(slp) == 0) {
1466 siw_dbg_mem(mem, "empty SGE\n");
1470 pble->addr = sg_dma_address(slp);
1471 pble->size = sg_dma_len(slp);
1473 pbl_size = pble->size;
1476 /* Merge PBL entries if adjacent */
1477 if (pble->addr + pble->size == sg_dma_address(slp)) {
1478 pble->size += sg_dma_len(slp);
1482 pble->addr = sg_dma_address(slp);
1483 pble->size = sg_dma_len(slp);
1484 pble->pbl_off = pbl_size;
1486 pbl_size += sg_dma_len(slp);
1489 "sge[%d], size %u, addr 0x%p, total %lu\n",
1490 i, pble->size, (void *)(uintptr_t)pble->addr,
1493 rv = ib_sg_to_pages(base_mr, sl, num_sle, sg_off, siw_set_pbl_page);
1495 mem->len = base_mr->length;
1496 mem->va = base_mr->iova;
1498 "%llu bytes, start 0x%pK, %u SLE to %u entries\n",
1499 mem->len, (void *)(uintptr_t)mem->va, num_sle,
1508 * Create a (empty) DMA memory region, where no umem is attached.
1510 struct ib_mr *siw_get_dma_mr(struct ib_pd *pd, int rights)
1512 struct siw_device *sdev = to_siw_dev(pd->device);
1513 struct siw_mr *mr = NULL;
1516 if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1517 siw_dbg_pd(pd, "too many mr's\n");
1521 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1526 rv = siw_mr_add_mem(mr, pd, NULL, 0, ULONG_MAX, rights);
1530 mr->mem->stag_valid = 1;
1532 siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1534 return &mr->base_mr;
1540 atomic_dec(&sdev->num_mr);
1548 * Create Shared Receive Queue of attributes @init_attrs
1549 * within protection domain given by @pd.
1551 * @base_srq: Base SRQ contained in siw SRQ.
1552 * @init_attrs: SRQ init attributes.
1553 * @udata: points to user context
1555 int siw_create_srq(struct ib_srq *base_srq,
1556 struct ib_srq_init_attr *init_attrs, struct ib_udata *udata)
1558 struct siw_srq *srq = to_siw_srq(base_srq);
1559 struct ib_srq_attr *attrs = &init_attrs->attr;
1560 struct siw_device *sdev = to_siw_dev(base_srq->device);
1561 struct siw_ucontext *ctx =
1562 rdma_udata_to_drv_context(udata, struct siw_ucontext,
1566 if (init_attrs->srq_type != IB_SRQT_BASIC)
1569 if (atomic_inc_return(&sdev->num_srq) > SIW_MAX_SRQ) {
1570 siw_dbg_pd(base_srq->pd, "too many SRQ's\n");
1574 if (attrs->max_wr == 0 || attrs->max_wr > SIW_MAX_SRQ_WR ||
1575 attrs->max_sge > SIW_MAX_SGE || attrs->srq_limit > attrs->max_wr) {
1579 srq->max_sge = attrs->max_sge;
1580 srq->num_rqe = roundup_pow_of_two(attrs->max_wr);
1581 srq->limit = attrs->srq_limit;
1585 srq->is_kernel_res = !udata;
1589 vmalloc_user(srq->num_rqe * sizeof(struct siw_rqe));
1591 srq->recvq = vzalloc(srq->num_rqe * sizeof(struct siw_rqe));
1593 if (srq->recvq == NULL) {
1598 struct siw_uresp_create_srq uresp = {};
1599 size_t length = srq->num_rqe * sizeof(struct siw_rqe);
1602 siw_mmap_entry_insert(ctx, srq->recvq,
1603 length, &uresp.srq_key);
1604 if (!srq->srq_entry) {
1609 uresp.num_rqe = srq->num_rqe;
1611 if (udata->outlen < sizeof(uresp)) {
1615 rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1619 spin_lock_init(&srq->lock);
1621 siw_dbg_pd(base_srq->pd, "[SRQ]: success\n");
1628 rdma_user_mmap_entry_remove(srq->srq_entry);
1631 atomic_dec(&sdev->num_srq);
1639 * Modify SRQ. The caller may resize SRQ and/or set/reset notification
1640 * limit and (re)arm IB_EVENT_SRQ_LIMIT_REACHED notification.
1642 * NOTE: it is unclear if RDMA core allows for changing the MAX_SGE
1643 * parameter. siw_modify_srq() does not check the attrs->max_sge param.
1645 int siw_modify_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs,
1646 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
1648 struct siw_srq *srq = to_siw_srq(base_srq);
1649 unsigned long flags;
1652 spin_lock_irqsave(&srq->lock, flags);
1654 if (attr_mask & IB_SRQ_MAX_WR) {
1655 /* resize request not yet supported */
1659 if (attr_mask & IB_SRQ_LIMIT) {
1660 if (attrs->srq_limit) {
1661 if (unlikely(attrs->srq_limit > srq->num_rqe)) {
1669 srq->limit = attrs->srq_limit;
1672 spin_unlock_irqrestore(&srq->lock, flags);
1680 * Query SRQ attributes.
1682 int siw_query_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs)
1684 struct siw_srq *srq = to_siw_srq(base_srq);
1685 unsigned long flags;
1687 spin_lock_irqsave(&srq->lock, flags);
1689 attrs->max_wr = srq->num_rqe;
1690 attrs->max_sge = srq->max_sge;
1691 attrs->srq_limit = srq->limit;
1693 spin_unlock_irqrestore(&srq->lock, flags);
1702 * It is assumed that the SRQ is not referenced by any
1703 * QP anymore - the code trusts the RDMA core environment to keep track
1706 int siw_destroy_srq(struct ib_srq *base_srq, struct ib_udata *udata)
1708 struct siw_srq *srq = to_siw_srq(base_srq);
1709 struct siw_device *sdev = to_siw_dev(base_srq->device);
1710 struct siw_ucontext *ctx =
1711 rdma_udata_to_drv_context(udata, struct siw_ucontext,
1715 rdma_user_mmap_entry_remove(srq->srq_entry);
1717 atomic_dec(&sdev->num_srq);
1722 * siw_post_srq_recv()
1724 * Post a list of receive queue elements to SRQ.
1725 * NOTE: The function does not check or lock a certain SRQ state
1726 * during the post operation. The code simply trusts the
1727 * RDMA core environment.
1729 * @base_srq: Base SRQ contained in siw SRQ
1730 * @wr: List of R-WR's
1731 * @bad_wr: Updated to failing WR if posting fails.
1733 int siw_post_srq_recv(struct ib_srq *base_srq, const struct ib_recv_wr *wr,
1734 const struct ib_recv_wr **bad_wr)
1736 struct siw_srq *srq = to_siw_srq(base_srq);
1737 unsigned long flags;
1740 if (unlikely(!srq->is_kernel_res)) {
1741 siw_dbg_pd(base_srq->pd,
1742 "[SRQ]: no kernel post_recv for mapped srq\n");
1747 * Serialize potentially multiple producers.
1748 * Also needed to serialize potentially multiple
1751 spin_lock_irqsave(&srq->lock, flags);
1754 u32 idx = srq->rq_put % srq->num_rqe;
1755 struct siw_rqe *rqe = &srq->recvq[idx];
1758 siw_dbg_pd(base_srq->pd, "SRQ full\n");
1762 if (unlikely(wr->num_sge > srq->max_sge)) {
1763 siw_dbg_pd(base_srq->pd,
1764 "[SRQ]: too many sge's: %d\n", wr->num_sge);
1768 rqe->id = wr->wr_id;
1769 rqe->num_sge = wr->num_sge;
1770 siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1772 /* Make sure S-RQE is completely written before valid */
1775 rqe->flags = SIW_WQE_VALID;
1780 spin_unlock_irqrestore(&srq->lock, flags);
1782 if (unlikely(rv < 0)) {
1783 siw_dbg_pd(base_srq->pd, "[SRQ]: error %d\n", rv);
1789 void siw_qp_event(struct siw_qp *qp, enum ib_event_type etype)
1791 struct ib_event event;
1792 struct ib_qp *base_qp = &qp->base_qp;
1795 * Do not report asynchronous errors on QP which gets
1796 * destroyed via verbs interface (siw_destroy_qp())
1798 if (qp->attrs.flags & SIW_QP_IN_DESTROY)
1801 event.event = etype;
1802 event.device = base_qp->device;
1803 event.element.qp = base_qp;
1805 if (base_qp->event_handler) {
1806 siw_dbg_qp(qp, "reporting event %d\n", etype);
1807 base_qp->event_handler(&event, base_qp->qp_context);
1811 void siw_cq_event(struct siw_cq *cq, enum ib_event_type etype)
1813 struct ib_event event;
1814 struct ib_cq *base_cq = &cq->base_cq;
1816 event.event = etype;
1817 event.device = base_cq->device;
1818 event.element.cq = base_cq;
1820 if (base_cq->event_handler) {
1821 siw_dbg_cq(cq, "reporting CQ event %d\n", etype);
1822 base_cq->event_handler(&event, base_cq->cq_context);
1826 void siw_srq_event(struct siw_srq *srq, enum ib_event_type etype)
1828 struct ib_event event;
1829 struct ib_srq *base_srq = &srq->base_srq;
1831 event.event = etype;
1832 event.device = base_srq->device;
1833 event.element.srq = base_srq;
1835 if (base_srq->event_handler) {
1836 siw_dbg_pd(srq->base_srq.pd,
1837 "reporting SRQ event %d\n", etype);
1838 base_srq->event_handler(&event, base_srq->srq_context);
1842 void siw_port_event(struct siw_device *sdev, u32 port, enum ib_event_type etype)
1844 struct ib_event event;
1846 event.event = etype;
1847 event.device = &sdev->base_dev;
1848 event.element.port_num = port;
1850 siw_dbg(&sdev->base_dev, "reporting port event %d\n", etype);
1852 ib_dispatch_event(&event);