2 * Copyright(c) 2016 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>
50 #include <linux/kthread.h>
56 * rvt_cq_enter - add a new entry to the completion queue
57 * @cq: completion queue
58 * @entry: work completion entry to add
59 * @sig: true if @entry is solicited
61 * This may be called with qp->s_lock held.
63 void rvt_cq_enter(struct rvt_cq *cq, struct ib_wc *entry, bool solicited)
70 spin_lock_irqsave(&cq->lock, flags);
73 * Note that the head pointer might be writable by user processes.
74 * Take care to verify it is a sane value.
78 if (head >= (unsigned)cq->ibcq.cqe) {
85 if (unlikely(next == wc->tail)) {
86 spin_unlock_irqrestore(&cq->lock, flags);
87 if (cq->ibcq.event_handler) {
90 ev.device = cq->ibcq.device;
91 ev.element.cq = &cq->ibcq;
92 ev.event = IB_EVENT_CQ_ERR;
93 cq->ibcq.event_handler(&ev, cq->ibcq.cq_context);
97 trace_rvt_cq_enter(cq, entry, head);
99 wc->uqueue[head].wr_id = entry->wr_id;
100 wc->uqueue[head].status = entry->status;
101 wc->uqueue[head].opcode = entry->opcode;
102 wc->uqueue[head].vendor_err = entry->vendor_err;
103 wc->uqueue[head].byte_len = entry->byte_len;
104 wc->uqueue[head].ex.imm_data =
105 (__u32 __force)entry->ex.imm_data;
106 wc->uqueue[head].qp_num = entry->qp->qp_num;
107 wc->uqueue[head].src_qp = entry->src_qp;
108 wc->uqueue[head].wc_flags = entry->wc_flags;
109 wc->uqueue[head].pkey_index = entry->pkey_index;
110 wc->uqueue[head].slid = ib_lid_cpu16(entry->slid);
111 wc->uqueue[head].sl = entry->sl;
112 wc->uqueue[head].dlid_path_bits = entry->dlid_path_bits;
113 wc->uqueue[head].port_num = entry->port_num;
114 /* Make sure entry is written before the head index. */
117 wc->kqueue[head] = *entry;
121 if (cq->notify == IB_CQ_NEXT_COMP ||
122 (cq->notify == IB_CQ_SOLICITED &&
123 (solicited || entry->status != IB_WC_SUCCESS))) {
124 struct kthread_worker *worker;
127 * This will cause send_complete() to be called in
131 worker = rcu_dereference(cq->rdi->worker);
132 if (likely(worker)) {
133 cq->notify = RVT_CQ_NONE;
135 kthread_queue_work(worker, &cq->comptask);
140 spin_unlock_irqrestore(&cq->lock, flags);
142 EXPORT_SYMBOL(rvt_cq_enter);
144 static void send_complete(struct kthread_work *work)
146 struct rvt_cq *cq = container_of(work, struct rvt_cq, comptask);
149 * The completion handler will most likely rearm the notification
150 * and poll for all pending entries. If a new completion entry
151 * is added while we are in this routine, queue_work()
152 * won't call us again until we return so we check triggered to
153 * see if we need to call the handler again.
156 u8 triggered = cq->triggered;
159 * IPoIB connected mode assumes the callback is from a
160 * soft IRQ. We simulate this by blocking "bottom halves".
161 * See the implementation for ipoib_cm_handle_tx_wc(),
162 * netif_tx_lock_bh() and netif_tx_lock().
165 cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
168 if (cq->triggered == triggered)
174 * rvt_create_cq - create a completion queue
175 * @ibdev: the device this completion queue is attached to
176 * @attr: creation attributes
177 * @context: unused by the QLogic_IB driver
178 * @udata: user data for libibverbs.so
180 * Called by ib_create_cq() in the generic verbs code.
182 * Return: pointer to the completion queue or negative errno values
185 struct ib_cq *rvt_create_cq(struct ib_device *ibdev,
186 const struct ib_cq_init_attr *attr,
187 struct ib_ucontext *context,
188 struct ib_udata *udata)
190 struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
192 struct rvt_cq_wc *wc;
195 unsigned int entries = attr->cqe;
198 return ERR_PTR(-EINVAL);
200 if (entries < 1 || entries > rdi->dparms.props.max_cqe)
201 return ERR_PTR(-EINVAL);
203 /* Allocate the completion queue structure. */
204 cq = kzalloc_node(sizeof(*cq), GFP_KERNEL, rdi->dparms.node);
206 return ERR_PTR(-ENOMEM);
209 * Allocate the completion queue entries and head/tail pointers.
210 * This is allocated separately so that it can be resized and
211 * also mapped into user space.
212 * We need to use vmalloc() in order to support mmap and large
213 * numbers of entries.
216 if (udata && udata->outlen >= sizeof(__u64))
217 sz += sizeof(struct ib_uverbs_wc) * (entries + 1);
219 sz += sizeof(struct ib_wc) * (entries + 1);
222 vzalloc_node(sz, rdi->dparms.node);
224 ret = ERR_PTR(-ENOMEM);
229 * Return the address of the WC as the offset to mmap.
230 * See rvt_mmap() for details.
232 if (udata && udata->outlen >= sizeof(__u64)) {
235 cq->ip = rvt_create_mmap_info(rdi, sz, context, wc);
237 ret = ERR_PTR(-ENOMEM);
241 err = ib_copy_to_udata(udata, &cq->ip->offset,
242 sizeof(cq->ip->offset));
249 spin_lock_irq(&rdi->n_cqs_lock);
250 if (rdi->n_cqs_allocated == rdi->dparms.props.max_cq) {
251 spin_unlock_irq(&rdi->n_cqs_lock);
252 ret = ERR_PTR(-ENOMEM);
256 rdi->n_cqs_allocated++;
257 spin_unlock_irq(&rdi->n_cqs_lock);
260 spin_lock_irq(&rdi->pending_lock);
261 list_add(&cq->ip->pending_mmaps, &rdi->pending_mmaps);
262 spin_unlock_irq(&rdi->pending_lock);
266 * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe.
267 * The number of entries should be >= the number requested or return
271 cq->ibcq.cqe = entries;
272 cq->notify = RVT_CQ_NONE;
273 spin_lock_init(&cq->lock);
274 kthread_init_work(&cq->comptask, send_complete);
292 * rvt_destroy_cq - destroy a completion queue
293 * @ibcq: the completion queue to destroy.
295 * Called by ib_destroy_cq() in the generic verbs code.
299 int rvt_destroy_cq(struct ib_cq *ibcq)
301 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
302 struct rvt_dev_info *rdi = cq->rdi;
304 kthread_flush_work(&cq->comptask);
305 spin_lock_irq(&rdi->n_cqs_lock);
306 rdi->n_cqs_allocated--;
307 spin_unlock_irq(&rdi->n_cqs_lock);
309 kref_put(&cq->ip->ref, rvt_release_mmap_info);
318 * rvt_req_notify_cq - change the notification type for a completion queue
319 * @ibcq: the completion queue
320 * @notify_flags: the type of notification to request
322 * This may be called from interrupt context. Also called by
323 * ib_req_notify_cq() in the generic verbs code.
325 * Return: 0 for success.
327 int rvt_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
329 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
333 spin_lock_irqsave(&cq->lock, flags);
335 * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow
336 * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2).
338 if (cq->notify != IB_CQ_NEXT_COMP)
339 cq->notify = notify_flags & IB_CQ_SOLICITED_MASK;
341 if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
342 cq->queue->head != cq->queue->tail)
345 spin_unlock_irqrestore(&cq->lock, flags);
351 * rvt_resize_cq - change the size of the CQ
352 * @ibcq: the completion queue
354 * Return: 0 for success.
356 int rvt_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
358 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
359 struct rvt_cq_wc *old_wc;
360 struct rvt_cq_wc *wc;
364 struct rvt_dev_info *rdi = cq->rdi;
366 if (cqe < 1 || cqe > rdi->dparms.props.max_cqe)
370 * Need to use vmalloc() if we want to support large #s of entries.
373 if (udata && udata->outlen >= sizeof(__u64))
374 sz += sizeof(struct ib_uverbs_wc) * (cqe + 1);
376 sz += sizeof(struct ib_wc) * (cqe + 1);
379 vzalloc_node(sz, rdi->dparms.node);
383 /* Check that we can write the offset to mmap. */
384 if (udata && udata->outlen >= sizeof(__u64)) {
387 ret = ib_copy_to_udata(udata, &offset, sizeof(offset));
392 spin_lock_irq(&cq->lock);
394 * Make sure head and tail are sane since they
395 * might be user writable.
399 if (head > (u32)cq->ibcq.cqe)
400 head = (u32)cq->ibcq.cqe;
402 if (tail > (u32)cq->ibcq.cqe)
403 tail = (u32)cq->ibcq.cqe;
405 n = cq->ibcq.cqe + 1 + head - tail;
408 if (unlikely((u32)cqe < n)) {
412 for (n = 0; tail != head; n++) {
414 wc->uqueue[n] = old_wc->uqueue[tail];
416 wc->kqueue[n] = old_wc->kqueue[tail];
417 if (tail == (u32)cq->ibcq.cqe)
426 spin_unlock_irq(&cq->lock);
431 struct rvt_mmap_info *ip = cq->ip;
433 rvt_update_mmap_info(rdi, ip, sz, wc);
436 * Return the offset to mmap.
437 * See rvt_mmap() for details.
439 if (udata && udata->outlen >= sizeof(__u64)) {
440 ret = ib_copy_to_udata(udata, &ip->offset,
446 spin_lock_irq(&rdi->pending_lock);
447 if (list_empty(&ip->pending_mmaps))
448 list_add(&ip->pending_mmaps, &rdi->pending_mmaps);
449 spin_unlock_irq(&rdi->pending_lock);
455 spin_unlock_irq(&cq->lock);
462 * rvt_poll_cq - poll for work completion entries
463 * @ibcq: the completion queue to poll
464 * @num_entries: the maximum number of entries to return
465 * @entry: pointer to array where work completions are placed
467 * This may be called from interrupt context. Also called by ib_poll_cq()
468 * in the generic verbs code.
470 * Return: the number of completion entries polled.
472 int rvt_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
474 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
475 struct rvt_cq_wc *wc;
480 /* The kernel can only poll a kernel completion queue */
484 spin_lock_irqsave(&cq->lock, flags);
488 if (tail > (u32)cq->ibcq.cqe)
489 tail = (u32)cq->ibcq.cqe;
490 for (npolled = 0; npolled < num_entries; ++npolled, ++entry) {
491 if (tail == wc->head)
493 /* The kernel doesn't need a RMB since it has the lock. */
494 trace_rvt_cq_poll(cq, &wc->kqueue[tail], npolled);
495 *entry = wc->kqueue[tail];
496 if (tail >= cq->ibcq.cqe)
503 spin_unlock_irqrestore(&cq->lock, flags);
509 * rvt_driver_cq_init - Init cq resources on behalf of driver
510 * @rdi: rvt dev structure
512 * Return: 0 on success
514 int rvt_driver_cq_init(struct rvt_dev_info *rdi)
517 struct kthread_worker *worker;
519 if (rcu_access_pointer(rdi->worker))
522 spin_lock_init(&rdi->n_cqs_lock);
524 cpu = cpumask_first(cpumask_of_node(rdi->dparms.node));
525 worker = kthread_create_worker_on_cpu(cpu, 0,
526 "%s", rdi->dparms.cq_name);
528 return PTR_ERR(worker);
530 set_user_nice(worker->task, MIN_NICE);
531 RCU_INIT_POINTER(rdi->worker, worker);
536 * rvt_cq_exit - tear down cq reources
537 * @rdi: rvt dev structure
539 void rvt_cq_exit(struct rvt_dev_info *rdi)
541 struct kthread_worker *worker;
543 if (!rcu_access_pointer(rdi->worker))
546 spin_lock(&rdi->n_cqs_lock);
547 worker = rcu_dereference_protected(rdi->worker,
548 lockdep_is_held(&rdi->n_cqs_lock));
550 spin_unlock(&rdi->n_cqs_lock);
553 RCU_INIT_POINTER(rdi->worker, NULL);
554 spin_unlock(&rdi->n_cqs_lock);
557 kthread_destroy_worker(worker);