2 * Copyright(c) 2016 - 2018 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
48 #include <linux/slab.h>
49 #include <linux/vmalloc.h>
54 static struct workqueue_struct *comp_vector_wq;
57 * rvt_cq_enter - add a new entry to the completion queue
58 * @cq: completion queue
59 * @entry: work completion entry to add
60 * @solicited: true if @entry is solicited
62 * This may be called with qp->s_lock held.
64 * Return: return true on success, else return
65 * false if cq is full.
67 bool rvt_cq_enter(struct rvt_cq *cq, struct ib_wc *entry, bool solicited)
69 struct ib_uverbs_wc *uqueue = NULL;
70 struct ib_wc *kqueue = NULL;
71 struct rvt_cq_wc *u_wc = NULL;
72 struct rvt_k_cq_wc *k_wc = NULL;
78 spin_lock_irqsave(&cq->lock, flags);
82 uqueue = &u_wc->uqueue[0];
83 head = RDMA_READ_UAPI_ATOMIC(u_wc->head);
84 tail = RDMA_READ_UAPI_ATOMIC(u_wc->tail);
87 kqueue = &k_wc->kqueue[0];
93 * Note that the head pointer might be writable by
94 * user processes.Take care to verify it is a sane value.
96 if (head >= (unsigned)cq->ibcq.cqe) {
103 if (unlikely(next == tail || cq->cq_full)) {
104 struct rvt_dev_info *rdi = cq->rdi;
107 rvt_pr_err_ratelimited(rdi, "CQ is full!\n");
109 spin_unlock_irqrestore(&cq->lock, flags);
110 if (cq->ibcq.event_handler) {
113 ev.device = cq->ibcq.device;
114 ev.element.cq = &cq->ibcq;
115 ev.event = IB_EVENT_CQ_ERR;
116 cq->ibcq.event_handler(&ev, cq->ibcq.cq_context);
120 trace_rvt_cq_enter(cq, entry, head);
122 uqueue[head].wr_id = entry->wr_id;
123 uqueue[head].status = entry->status;
124 uqueue[head].opcode = entry->opcode;
125 uqueue[head].vendor_err = entry->vendor_err;
126 uqueue[head].byte_len = entry->byte_len;
127 uqueue[head].ex.imm_data = entry->ex.imm_data;
128 uqueue[head].qp_num = entry->qp->qp_num;
129 uqueue[head].src_qp = entry->src_qp;
130 uqueue[head].wc_flags = entry->wc_flags;
131 uqueue[head].pkey_index = entry->pkey_index;
132 uqueue[head].slid = ib_lid_cpu16(entry->slid);
133 uqueue[head].sl = entry->sl;
134 uqueue[head].dlid_path_bits = entry->dlid_path_bits;
135 uqueue[head].port_num = entry->port_num;
136 /* Make sure entry is written before the head index. */
137 RDMA_WRITE_UAPI_ATOMIC(u_wc->head, next);
139 kqueue[head] = *entry;
143 if (cq->notify == IB_CQ_NEXT_COMP ||
144 (cq->notify == IB_CQ_SOLICITED &&
145 (solicited || entry->status != IB_WC_SUCCESS))) {
147 * This will cause send_complete() to be called in
150 cq->notify = RVT_CQ_NONE;
152 queue_work_on(cq->comp_vector_cpu, comp_vector_wq,
156 spin_unlock_irqrestore(&cq->lock, flags);
159 EXPORT_SYMBOL(rvt_cq_enter);
161 static void send_complete(struct work_struct *work)
163 struct rvt_cq *cq = container_of(work, struct rvt_cq, comptask);
166 * The completion handler will most likely rearm the notification
167 * and poll for all pending entries. If a new completion entry
168 * is added while we are in this routine, queue_work()
169 * won't call us again until we return so we check triggered to
170 * see if we need to call the handler again.
173 u8 triggered = cq->triggered;
176 * IPoIB connected mode assumes the callback is from a
177 * soft IRQ. We simulate this by blocking "bottom halves".
178 * See the implementation for ipoib_cm_handle_tx_wc(),
179 * netif_tx_lock_bh() and netif_tx_lock().
182 cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
185 if (cq->triggered == triggered)
191 * rvt_create_cq - create a completion queue
192 * @ibcq: Allocated CQ
193 * @attr: creation attributes
194 * @udata: user data for libibverbs.so
196 * Called by ib_create_cq() in the generic verbs code.
198 * Return: 0 on success
200 int rvt_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
201 struct ib_udata *udata)
203 struct ib_device *ibdev = ibcq->device;
204 struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
205 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
206 struct rvt_cq_wc *u_wc = NULL;
207 struct rvt_k_cq_wc *k_wc = NULL;
209 unsigned int entries = attr->cqe;
210 int comp_vector = attr->comp_vector;
216 if (entries < 1 || entries > rdi->dparms.props.max_cqe)
222 comp_vector = comp_vector % rdi->ibdev.num_comp_vectors;
225 * Allocate the completion queue entries and head/tail pointers.
226 * This is allocated separately so that it can be resized and
227 * also mapped into user space.
228 * We need to use vmalloc() in order to support mmap and large
229 * numbers of entries.
231 if (udata && udata->outlen >= sizeof(__u64)) {
232 sz = sizeof(struct ib_uverbs_wc) * (entries + 1);
234 u_wc = vmalloc_user(sz);
238 sz = sizeof(struct ib_wc) * (entries + 1);
240 k_wc = vzalloc_node(sz, rdi->dparms.node);
246 * Return the address of the WC as the offset to mmap.
247 * See rvt_mmap() for details.
249 if (udata && udata->outlen >= sizeof(__u64)) {
250 cq->ip = rvt_create_mmap_info(rdi, sz, udata, u_wc);
256 err = ib_copy_to_udata(udata, &cq->ip->offset,
257 sizeof(cq->ip->offset));
262 spin_lock_irq(&rdi->n_cqs_lock);
263 if (rdi->n_cqs_allocated == rdi->dparms.props.max_cq) {
264 spin_unlock_irq(&rdi->n_cqs_lock);
269 rdi->n_cqs_allocated++;
270 spin_unlock_irq(&rdi->n_cqs_lock);
273 spin_lock_irq(&rdi->pending_lock);
274 list_add(&cq->ip->pending_mmaps, &rdi->pending_mmaps);
275 spin_unlock_irq(&rdi->pending_lock);
279 * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe.
280 * The number of entries should be >= the number requested or return
284 if (rdi->driver_f.comp_vect_cpu_lookup)
285 cq->comp_vector_cpu =
286 rdi->driver_f.comp_vect_cpu_lookup(rdi, comp_vector);
288 cq->comp_vector_cpu =
289 cpumask_first(cpumask_of_node(rdi->dparms.node));
291 cq->ibcq.cqe = entries;
292 cq->notify = RVT_CQ_NONE;
293 spin_lock_init(&cq->lock);
294 INIT_WORK(&cq->comptask, send_complete);
300 trace_rvt_create_cq(cq, attr);
312 * rvt_destroy_cq - destroy a completion queue
313 * @ibcq: the completion queue to destroy.
314 * @udata: user data or NULL for kernel object
316 * Called by ib_destroy_cq() in the generic verbs code.
318 void rvt_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
320 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
321 struct rvt_dev_info *rdi = cq->rdi;
323 flush_work(&cq->comptask);
324 spin_lock_irq(&rdi->n_cqs_lock);
325 rdi->n_cqs_allocated--;
326 spin_unlock_irq(&rdi->n_cqs_lock);
328 kref_put(&cq->ip->ref, rvt_release_mmap_info);
334 * rvt_req_notify_cq - change the notification type for a completion queue
335 * @ibcq: the completion queue
336 * @notify_flags: the type of notification to request
338 * This may be called from interrupt context. Also called by
339 * ib_req_notify_cq() in the generic verbs code.
341 * Return: 0 for success.
343 int rvt_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
345 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
349 spin_lock_irqsave(&cq->lock, flags);
351 * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow
352 * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2).
354 if (cq->notify != IB_CQ_NEXT_COMP)
355 cq->notify = notify_flags & IB_CQ_SOLICITED_MASK;
357 if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) {
359 if (RDMA_READ_UAPI_ATOMIC(cq->queue->head) !=
360 RDMA_READ_UAPI_ATOMIC(cq->queue->tail))
363 if (cq->kqueue->head != cq->kqueue->tail)
368 spin_unlock_irqrestore(&cq->lock, flags);
374 * rvt_resize_cq - change the size of the CQ
375 * @ibcq: the completion queue
377 * Return: 0 for success.
379 int rvt_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
381 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
385 struct rvt_dev_info *rdi = cq->rdi;
386 struct rvt_cq_wc *u_wc = NULL;
387 struct rvt_cq_wc *old_u_wc = NULL;
388 struct rvt_k_cq_wc *k_wc = NULL;
389 struct rvt_k_cq_wc *old_k_wc = NULL;
391 if (cqe < 1 || cqe > rdi->dparms.props.max_cqe)
395 * Need to use vmalloc() if we want to support large #s of entries.
397 if (udata && udata->outlen >= sizeof(__u64)) {
398 sz = sizeof(struct ib_uverbs_wc) * (cqe + 1);
400 u_wc = vmalloc_user(sz);
404 sz = sizeof(struct ib_wc) * (cqe + 1);
406 k_wc = vzalloc_node(sz, rdi->dparms.node);
410 /* Check that we can write the offset to mmap. */
411 if (udata && udata->outlen >= sizeof(__u64)) {
414 ret = ib_copy_to_udata(udata, &offset, sizeof(offset));
419 spin_lock_irq(&cq->lock);
421 * Make sure head and tail are sane since they
422 * might be user writable.
425 old_u_wc = cq->queue;
426 head = RDMA_READ_UAPI_ATOMIC(old_u_wc->head);
427 tail = RDMA_READ_UAPI_ATOMIC(old_u_wc->tail);
429 old_k_wc = cq->kqueue;
430 head = old_k_wc->head;
431 tail = old_k_wc->tail;
434 if (head > (u32)cq->ibcq.cqe)
435 head = (u32)cq->ibcq.cqe;
436 if (tail > (u32)cq->ibcq.cqe)
437 tail = (u32)cq->ibcq.cqe;
439 n = cq->ibcq.cqe + 1 + head - tail;
442 if (unlikely((u32)cqe < n)) {
446 for (n = 0; tail != head; n++) {
448 u_wc->uqueue[n] = old_u_wc->uqueue[tail];
450 k_wc->kqueue[n] = old_k_wc->kqueue[tail];
451 if (tail == (u32)cq->ibcq.cqe)
458 RDMA_WRITE_UAPI_ATOMIC(u_wc->head, n);
459 RDMA_WRITE_UAPI_ATOMIC(u_wc->tail, 0);
466 spin_unlock_irq(&cq->lock);
474 struct rvt_mmap_info *ip = cq->ip;
476 rvt_update_mmap_info(rdi, ip, sz, u_wc);
479 * Return the offset to mmap.
480 * See rvt_mmap() for details.
482 if (udata && udata->outlen >= sizeof(__u64)) {
483 ret = ib_copy_to_udata(udata, &ip->offset,
489 spin_lock_irq(&rdi->pending_lock);
490 if (list_empty(&ip->pending_mmaps))
491 list_add(&ip->pending_mmaps, &rdi->pending_mmaps);
492 spin_unlock_irq(&rdi->pending_lock);
498 spin_unlock_irq(&cq->lock);
507 * rvt_poll_cq - poll for work completion entries
508 * @ibcq: the completion queue to poll
509 * @num_entries: the maximum number of entries to return
510 * @entry: pointer to array where work completions are placed
512 * This may be called from interrupt context. Also called by ib_poll_cq()
513 * in the generic verbs code.
515 * Return: the number of completion entries polled.
517 int rvt_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
519 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
520 struct rvt_k_cq_wc *wc;
525 /* The kernel can only poll a kernel completion queue */
529 spin_lock_irqsave(&cq->lock, flags);
533 if (tail > (u32)cq->ibcq.cqe)
534 tail = (u32)cq->ibcq.cqe;
535 for (npolled = 0; npolled < num_entries; ++npolled, ++entry) {
536 if (tail == wc->head)
538 /* The kernel doesn't need a RMB since it has the lock. */
539 trace_rvt_cq_poll(cq, &wc->kqueue[tail], npolled);
540 *entry = wc->kqueue[tail];
541 if (tail >= cq->ibcq.cqe)
548 spin_unlock_irqrestore(&cq->lock, flags);
554 * rvt_driver_cq_init - Init cq resources on behalf of driver
555 * @rdi: rvt dev structure
557 * Return: 0 on success
559 int rvt_driver_cq_init(void)
561 comp_vector_wq = alloc_workqueue("%s", WQ_HIGHPRI | WQ_CPU_INTENSIVE,
570 * rvt_cq_exit - tear down cq reources
571 * @rdi: rvt dev structure
573 void rvt_cq_exit(void)
575 destroy_workqueue(comp_vector_wq);
576 comp_vector_wq = NULL;