2 * Copyright (c) 2006, 2007, 2008, 2010 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/err.h>
35 #include <linux/slab.h>
36 #include <linux/vmalloc.h>
38 #include "qib_verbs.h"
41 * qib_cq_enter - add a new entry to the completion queue
42 * @cq: completion queue
43 * @entry: work completion entry to add
44 * @sig: true if @entry is a solicitated entry
46 * This may be called with qp->s_lock held.
48 void qib_cq_enter(struct qib_cq *cq, struct ib_wc *entry, int solicited)
55 spin_lock_irqsave(&cq->lock, flags);
58 * Note that the head pointer might be writable by user processes.
59 * Take care to verify it is a sane value.
63 if (head >= (unsigned) cq->ibcq.cqe) {
68 if (unlikely(next == wc->tail)) {
69 spin_unlock_irqrestore(&cq->lock, flags);
70 if (cq->ibcq.event_handler) {
73 ev.device = cq->ibcq.device;
74 ev.element.cq = &cq->ibcq;
75 ev.event = IB_EVENT_CQ_ERR;
76 cq->ibcq.event_handler(&ev, cq->ibcq.cq_context);
81 wc->uqueue[head].wr_id = entry->wr_id;
82 wc->uqueue[head].status = entry->status;
83 wc->uqueue[head].opcode = entry->opcode;
84 wc->uqueue[head].vendor_err = entry->vendor_err;
85 wc->uqueue[head].byte_len = entry->byte_len;
86 wc->uqueue[head].ex.imm_data =
87 (__u32 __force)entry->ex.imm_data;
88 wc->uqueue[head].qp_num = entry->qp->qp_num;
89 wc->uqueue[head].src_qp = entry->src_qp;
90 wc->uqueue[head].wc_flags = entry->wc_flags;
91 wc->uqueue[head].pkey_index = entry->pkey_index;
92 wc->uqueue[head].slid = entry->slid;
93 wc->uqueue[head].sl = entry->sl;
94 wc->uqueue[head].dlid_path_bits = entry->dlid_path_bits;
95 wc->uqueue[head].port_num = entry->port_num;
96 /* Make sure entry is written before the head index. */
99 wc->kqueue[head] = *entry;
102 if (cq->notify == IB_CQ_NEXT_COMP ||
103 (cq->notify == IB_CQ_SOLICITED && solicited)) {
104 cq->notify = IB_CQ_NONE;
107 * This will cause send_complete() to be called in
110 queue_work(qib_cq_wq, &cq->comptask);
113 spin_unlock_irqrestore(&cq->lock, flags);
117 * qib_poll_cq - poll for work completion entries
118 * @ibcq: the completion queue to poll
119 * @num_entries: the maximum number of entries to return
120 * @entry: pointer to array where work completions are placed
122 * Returns the number of completion entries polled.
124 * This may be called from interrupt context. Also called by ib_poll_cq()
125 * in the generic verbs code.
127 int qib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
129 struct qib_cq *cq = to_icq(ibcq);
130 struct qib_cq_wc *wc;
135 /* The kernel can only poll a kernel completion queue */
141 spin_lock_irqsave(&cq->lock, flags);
145 if (tail > (u32) cq->ibcq.cqe)
146 tail = (u32) cq->ibcq.cqe;
147 for (npolled = 0; npolled < num_entries; ++npolled, ++entry) {
148 if (tail == wc->head)
150 /* The kernel doesn't need a RMB since it has the lock. */
151 *entry = wc->kqueue[tail];
152 if (tail >= cq->ibcq.cqe)
159 spin_unlock_irqrestore(&cq->lock, flags);
165 static void send_complete(struct work_struct *work)
167 struct qib_cq *cq = container_of(work, struct qib_cq, comptask);
170 * The completion handler will most likely rearm the notification
171 * and poll for all pending entries. If a new completion entry
172 * is added while we are in this routine, queue_work()
173 * won't call us again until we return so we check triggered to
174 * see if we need to call the handler again.
177 u8 triggered = cq->triggered;
180 * IPoIB connected mode assumes the callback is from a
181 * soft IRQ. We simulate this by blocking "bottom halves".
182 * See the implementation for ipoib_cm_handle_tx_wc(),
183 * netif_tx_lock_bh() and netif_tx_lock().
186 cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
189 if (cq->triggered == triggered)
195 * qib_create_cq - create a completion queue
196 * @ibdev: the device this completion queue is attached to
197 * @entries: the minimum size of the completion queue
198 * @context: unused by the QLogic_IB driver
199 * @udata: user data for libibverbs.so
201 * Returns a pointer to the completion queue or negative errno values
204 * Called by ib_create_cq() in the generic verbs code.
206 struct ib_cq *qib_create_cq(struct ib_device *ibdev, int entries,
207 int comp_vector, struct ib_ucontext *context,
208 struct ib_udata *udata)
210 struct qib_ibdev *dev = to_idev(ibdev);
212 struct qib_cq_wc *wc;
216 if (entries < 1 || entries > ib_qib_max_cqes) {
217 ret = ERR_PTR(-EINVAL);
221 /* Allocate the completion queue structure. */
222 cq = kmalloc(sizeof(*cq), GFP_KERNEL);
224 ret = ERR_PTR(-ENOMEM);
229 * Allocate the completion queue entries and head/tail pointers.
230 * This is allocated separately so that it can be resized and
231 * also mapped into user space.
232 * We need to use vmalloc() in order to support mmap and large
233 * numbers of entries.
236 if (udata && udata->outlen >= sizeof(__u64))
237 sz += sizeof(struct ib_uverbs_wc) * (entries + 1);
239 sz += sizeof(struct ib_wc) * (entries + 1);
240 wc = vmalloc_user(sz);
242 ret = ERR_PTR(-ENOMEM);
247 * Return the address of the WC as the offset to mmap.
248 * See qib_mmap() for details.
250 if (udata && udata->outlen >= sizeof(__u64)) {
253 cq->ip = qib_create_mmap_info(dev, sz, context, wc);
255 ret = ERR_PTR(-ENOMEM);
259 err = ib_copy_to_udata(udata, &cq->ip->offset,
260 sizeof(cq->ip->offset));
268 spin_lock(&dev->n_cqs_lock);
269 if (dev->n_cqs_allocated == ib_qib_max_cqs) {
270 spin_unlock(&dev->n_cqs_lock);
271 ret = ERR_PTR(-ENOMEM);
275 dev->n_cqs_allocated++;
276 spin_unlock(&dev->n_cqs_lock);
279 spin_lock_irq(&dev->pending_lock);
280 list_add(&cq->ip->pending_mmaps, &dev->pending_mmaps);
281 spin_unlock_irq(&dev->pending_lock);
285 * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe.
286 * The number of entries should be >= the number requested or return
289 cq->ibcq.cqe = entries;
290 cq->notify = IB_CQ_NONE;
292 spin_lock_init(&cq->lock);
293 INIT_WORK(&cq->comptask, send_complete);
313 * qib_destroy_cq - destroy a completion queue
314 * @ibcq: the completion queue to destroy.
316 * Returns 0 for success.
318 * Called by ib_destroy_cq() in the generic verbs code.
320 int qib_destroy_cq(struct ib_cq *ibcq)
322 struct qib_ibdev *dev = to_idev(ibcq->device);
323 struct qib_cq *cq = to_icq(ibcq);
325 flush_work(&cq->comptask);
326 spin_lock(&dev->n_cqs_lock);
327 dev->n_cqs_allocated--;
328 spin_unlock(&dev->n_cqs_lock);
330 kref_put(&cq->ip->ref, qib_release_mmap_info);
339 * qib_req_notify_cq - change the notification type for a completion queue
340 * @ibcq: the completion queue
341 * @notify_flags: the type of notification to request
343 * Returns 0 for success.
345 * This may be called from interrupt context. Also called by
346 * ib_req_notify_cq() in the generic verbs code.
348 int qib_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
350 struct qib_cq *cq = to_icq(ibcq);
354 spin_lock_irqsave(&cq->lock, flags);
356 * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow
357 * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2).
359 if (cq->notify != IB_CQ_NEXT_COMP)
360 cq->notify = notify_flags & IB_CQ_SOLICITED_MASK;
362 if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
363 cq->queue->head != cq->queue->tail)
366 spin_unlock_irqrestore(&cq->lock, flags);
372 * qib_resize_cq - change the size of the CQ
373 * @ibcq: the completion queue
375 * Returns 0 for success.
377 int qib_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
379 struct qib_cq *cq = to_icq(ibcq);
380 struct qib_cq_wc *old_wc;
381 struct qib_cq_wc *wc;
386 if (cqe < 1 || cqe > ib_qib_max_cqes) {
392 * Need to use vmalloc() if we want to support large #s of entries.
395 if (udata && udata->outlen >= sizeof(__u64))
396 sz += sizeof(struct ib_uverbs_wc) * (cqe + 1);
398 sz += sizeof(struct ib_wc) * (cqe + 1);
399 wc = vmalloc_user(sz);
405 /* Check that we can write the offset to mmap. */
406 if (udata && udata->outlen >= sizeof(__u64)) {
409 ret = ib_copy_to_udata(udata, &offset, sizeof(offset));
414 spin_lock_irq(&cq->lock);
416 * Make sure head and tail are sane since they
417 * might be user writable.
421 if (head > (u32) cq->ibcq.cqe)
422 head = (u32) cq->ibcq.cqe;
424 if (tail > (u32) cq->ibcq.cqe)
425 tail = (u32) cq->ibcq.cqe;
427 n = cq->ibcq.cqe + 1 + head - tail;
430 if (unlikely((u32)cqe < n)) {
434 for (n = 0; tail != head; n++) {
436 wc->uqueue[n] = old_wc->uqueue[tail];
438 wc->kqueue[n] = old_wc->kqueue[tail];
439 if (tail == (u32) cq->ibcq.cqe)
448 spin_unlock_irq(&cq->lock);
453 struct qib_ibdev *dev = to_idev(ibcq->device);
454 struct qib_mmap_info *ip = cq->ip;
456 qib_update_mmap_info(dev, ip, sz, wc);
459 * Return the offset to mmap.
460 * See qib_mmap() for details.
462 if (udata && udata->outlen >= sizeof(__u64)) {
463 ret = ib_copy_to_udata(udata, &ip->offset,
469 spin_lock_irq(&dev->pending_lock);
470 if (list_empty(&ip->pending_mmaps))
471 list_add(&ip->pending_mmaps, &dev->pending_mmaps);
472 spin_unlock_irq(&dev->pending_lock);
479 spin_unlock_irq(&cq->lock);