1/*
2 * Copyright (c) 2006, 2007, 2008, 2010 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4 *
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:
10 *
11 *     Redistribution and use in source and binary forms, with or
12 *     without modification, are permitted provided that the following
13 *     conditions are met:
14 *
15 *      - Redistributions of source code must retain the above
16 *        copyright notice, this list of conditions and the following
17 *        disclaimer.
18 *
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.
23 *
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
31 * SOFTWARE.
32 */
33
34#include <linux/err.h>
35#include <linux/slab.h>
36#include <linux/vmalloc.h>
37
38#include "qib_verbs.h"
39
40/**
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
45 *
46 * This may be called with qp->s_lock held.
47 */
48void qib_cq_enter(struct qib_cq *cq, struct ib_wc *entry, int solicited)
49{
50	struct qib_cq_wc *wc;
51	unsigned long flags;
52	u32 head;
53	u32 next;
54
55	spin_lock_irqsave(&cq->lock, flags);
56
57	/*
58	 * Note that the head pointer might be writable by user processes.
59	 * Take care to verify it is a sane value.
60	 */
61	wc = cq->queue;
62	head = wc->head;
63	if (head >= (unsigned) cq->ibcq.cqe) {
64		head = cq->ibcq.cqe;
65		next = 0;
66	} else
67		next = head + 1;
68	if (unlikely(next == wc->tail)) {
69		spin_unlock_irqrestore(&cq->lock, flags);
70		if (cq->ibcq.event_handler) {
71			struct ib_event ev;
72
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);
77		}
78		return;
79	}
80	if (cq->ip) {
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. */
97		smp_wmb();
98	} else
99		wc->kqueue[head] = *entry;
100	wc->head = next;
101
102	if (cq->notify == IB_CQ_NEXT_COMP ||
103	    (cq->notify == IB_CQ_SOLICITED &&
104	     (solicited || entry->status != IB_WC_SUCCESS))) {
105		cq->notify = IB_CQ_NONE;
106		cq->triggered++;
107		/*
108		 * This will cause send_complete() to be called in
109		 * another thread.
110		 */
111		queue_work(qib_cq_wq, &cq->comptask);
112	}
113
114	spin_unlock_irqrestore(&cq->lock, flags);
115}
116
117/**
118 * qib_poll_cq - poll for work completion entries
119 * @ibcq: the completion queue to poll
120 * @num_entries: the maximum number of entries to return
121 * @entry: pointer to array where work completions are placed
122 *
123 * Returns the number of completion entries polled.
124 *
125 * This may be called from interrupt context.  Also called by ib_poll_cq()
126 * in the generic verbs code.
127 */
128int qib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
129{
130	struct qib_cq *cq = to_icq(ibcq);
131	struct qib_cq_wc *wc;
132	unsigned long flags;
133	int npolled;
134	u32 tail;
135
136	/* The kernel can only poll a kernel completion queue */
137	if (cq->ip) {
138		npolled = -EINVAL;
139		goto bail;
140	}
141
142	spin_lock_irqsave(&cq->lock, flags);
143
144	wc = cq->queue;
145	tail = wc->tail;
146	if (tail > (u32) cq->ibcq.cqe)
147		tail = (u32) cq->ibcq.cqe;
148	for (npolled = 0; npolled < num_entries; ++npolled, ++entry) {
149		if (tail == wc->head)
150			break;
151		/* The kernel doesn't need a RMB since it has the lock. */
152		*entry = wc->kqueue[tail];
153		if (tail >= cq->ibcq.cqe)
154			tail = 0;
155		else
156			tail++;
157	}
158	wc->tail = tail;
159
160	spin_unlock_irqrestore(&cq->lock, flags);
161
162bail:
163	return npolled;
164}
165
166static void send_complete(struct work_struct *work)
167{
168	struct qib_cq *cq = container_of(work, struct qib_cq, comptask);
169
170	/*
171	 * The completion handler will most likely rearm the notification
172	 * and poll for all pending entries.  If a new completion entry
173	 * is added while we are in this routine, queue_work()
174	 * won't call us again until we return so we check triggered to
175	 * see if we need to call the handler again.
176	 */
177	for (;;) {
178		u8 triggered = cq->triggered;
179
180		/*
181		 * IPoIB connected mode assumes the callback is from a
182		 * soft IRQ. We simulate this by blocking "bottom halves".
183		 * See the implementation for ipoib_cm_handle_tx_wc(),
184		 * netif_tx_lock_bh() and netif_tx_lock().
185		 */
186		local_bh_disable();
187		cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
188		local_bh_enable();
189
190		if (cq->triggered == triggered)
191			return;
192	}
193}
194
195/**
196 * qib_create_cq - create a completion queue
197 * @ibdev: the device this completion queue is attached to
198 * @entries: the minimum size of the completion queue
199 * @context: unused by the QLogic_IB driver
200 * @udata: user data for libibverbs.so
201 *
202 * Returns a pointer to the completion queue or negative errno values
203 * for failure.
204 *
205 * Called by ib_create_cq() in the generic verbs code.
206 */
207struct ib_cq *qib_create_cq(struct ib_device *ibdev, int entries,
208			    int comp_vector, struct ib_ucontext *context,
209			    struct ib_udata *udata)
210{
211	struct qib_ibdev *dev = to_idev(ibdev);
212	struct qib_cq *cq;
213	struct qib_cq_wc *wc;
214	struct ib_cq *ret;
215	u32 sz;
216
217	if (entries < 1 || entries > ib_qib_max_cqes) {
218		ret = ERR_PTR(-EINVAL);
219		goto done;
220	}
221
222	/* Allocate the completion queue structure. */
223	cq = kmalloc(sizeof(*cq), GFP_KERNEL);
224	if (!cq) {
225		ret = ERR_PTR(-ENOMEM);
226		goto done;
227	}
228
229	/*
230	 * Allocate the completion queue entries and head/tail pointers.
231	 * This is allocated separately so that it can be resized and
232	 * also mapped into user space.
233	 * We need to use vmalloc() in order to support mmap and large
234	 * numbers of entries.
235	 */
236	sz = sizeof(*wc);
237	if (udata && udata->outlen >= sizeof(__u64))
238		sz += sizeof(struct ib_uverbs_wc) * (entries + 1);
239	else
240		sz += sizeof(struct ib_wc) * (entries + 1);
241	wc = vmalloc_user(sz);
242	if (!wc) {
243		ret = ERR_PTR(-ENOMEM);
244		goto bail_cq;
245	}
246
247	/*
248	 * Return the address of the WC as the offset to mmap.
249	 * See qib_mmap() for details.
250	 */
251	if (udata && udata->outlen >= sizeof(__u64)) {
252		int err;
253
254		cq->ip = qib_create_mmap_info(dev, sz, context, wc);
255		if (!cq->ip) {
256			ret = ERR_PTR(-ENOMEM);
257			goto bail_wc;
258		}
259
260		err = ib_copy_to_udata(udata, &cq->ip->offset,
261				       sizeof(cq->ip->offset));
262		if (err) {
263			ret = ERR_PTR(err);
264			goto bail_ip;
265		}
266	} else
267		cq->ip = NULL;
268
269	spin_lock(&dev->n_cqs_lock);
270	if (dev->n_cqs_allocated == ib_qib_max_cqs) {
271		spin_unlock(&dev->n_cqs_lock);
272		ret = ERR_PTR(-ENOMEM);
273		goto bail_ip;
274	}
275
276	dev->n_cqs_allocated++;
277	spin_unlock(&dev->n_cqs_lock);
278
279	if (cq->ip) {
280		spin_lock_irq(&dev->pending_lock);
281		list_add(&cq->ip->pending_mmaps, &dev->pending_mmaps);
282		spin_unlock_irq(&dev->pending_lock);
283	}
284
285	/*
286	 * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe.
287	 * The number of entries should be >= the number requested or return
288	 * an error.
289	 */
290	cq->ibcq.cqe = entries;
291	cq->notify = IB_CQ_NONE;
292	cq->triggered = 0;
293	spin_lock_init(&cq->lock);
294	INIT_WORK(&cq->comptask, send_complete);
295	wc->head = 0;
296	wc->tail = 0;
297	cq->queue = wc;
298
299	ret = &cq->ibcq;
300
301	goto done;
302
303bail_ip:
304	kfree(cq->ip);
305bail_wc:
306	vfree(wc);
307bail_cq:
308	kfree(cq);
309done:
310	return ret;
311}
312
313/**
314 * qib_destroy_cq - destroy a completion queue
315 * @ibcq: the completion queue to destroy.
316 *
317 * Returns 0 for success.
318 *
319 * Called by ib_destroy_cq() in the generic verbs code.
320 */
321int qib_destroy_cq(struct ib_cq *ibcq)
322{
323	struct qib_ibdev *dev = to_idev(ibcq->device);
324	struct qib_cq *cq = to_icq(ibcq);
325
326	flush_work(&cq->comptask);
327	spin_lock(&dev->n_cqs_lock);
328	dev->n_cqs_allocated--;
329	spin_unlock(&dev->n_cqs_lock);
330	if (cq->ip)
331		kref_put(&cq->ip->ref, qib_release_mmap_info);
332	else
333		vfree(cq->queue);
334	kfree(cq);
335
336	return 0;
337}
338
339/**
340 * qib_req_notify_cq - change the notification type for a completion queue
341 * @ibcq: the completion queue
342 * @notify_flags: the type of notification to request
343 *
344 * Returns 0 for success.
345 *
346 * This may be called from interrupt context.  Also called by
347 * ib_req_notify_cq() in the generic verbs code.
348 */
349int qib_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
350{
351	struct qib_cq *cq = to_icq(ibcq);
352	unsigned long flags;
353	int ret = 0;
354
355	spin_lock_irqsave(&cq->lock, flags);
356	/*
357	 * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow
358	 * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2).
359	 */
360	if (cq->notify != IB_CQ_NEXT_COMP)
361		cq->notify = notify_flags & IB_CQ_SOLICITED_MASK;
362
363	if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
364	    cq->queue->head != cq->queue->tail)
365		ret = 1;
366
367	spin_unlock_irqrestore(&cq->lock, flags);
368
369	return ret;
370}
371
372/**
373 * qib_resize_cq - change the size of the CQ
374 * @ibcq: the completion queue
375 *
376 * Returns 0 for success.
377 */
378int qib_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
379{
380	struct qib_cq *cq = to_icq(ibcq);
381	struct qib_cq_wc *old_wc;
382	struct qib_cq_wc *wc;
383	u32 head, tail, n;
384	int ret;
385	u32 sz;
386
387	if (cqe < 1 || cqe > ib_qib_max_cqes) {
388		ret = -EINVAL;
389		goto bail;
390	}
391
392	/*
393	 * Need to use vmalloc() if we want to support large #s of entries.
394	 */
395	sz = sizeof(*wc);
396	if (udata && udata->outlen >= sizeof(__u64))
397		sz += sizeof(struct ib_uverbs_wc) * (cqe + 1);
398	else
399		sz += sizeof(struct ib_wc) * (cqe + 1);
400	wc = vmalloc_user(sz);
401	if (!wc) {
402		ret = -ENOMEM;
403		goto bail;
404	}
405
406	/* Check that we can write the offset to mmap. */
407	if (udata && udata->outlen >= sizeof(__u64)) {
408		__u64 offset = 0;
409
410		ret = ib_copy_to_udata(udata, &offset, sizeof(offset));
411		if (ret)
412			goto bail_free;
413	}
414
415	spin_lock_irq(&cq->lock);
416	/*
417	 * Make sure head and tail are sane since they
418	 * might be user writable.
419	 */
420	old_wc = cq->queue;
421	head = old_wc->head;
422	if (head > (u32) cq->ibcq.cqe)
423		head = (u32) cq->ibcq.cqe;
424	tail = old_wc->tail;
425	if (tail > (u32) cq->ibcq.cqe)
426		tail = (u32) cq->ibcq.cqe;
427	if (head < tail)
428		n = cq->ibcq.cqe + 1 + head - tail;
429	else
430		n = head - tail;
431	if (unlikely((u32)cqe < n)) {
432		ret = -EINVAL;
433		goto bail_unlock;
434	}
435	for (n = 0; tail != head; n++) {
436		if (cq->ip)
437			wc->uqueue[n] = old_wc->uqueue[tail];
438		else
439			wc->kqueue[n] = old_wc->kqueue[tail];
440		if (tail == (u32) cq->ibcq.cqe)
441			tail = 0;
442		else
443			tail++;
444	}
445	cq->ibcq.cqe = cqe;
446	wc->head = n;
447	wc->tail = 0;
448	cq->queue = wc;
449	spin_unlock_irq(&cq->lock);
450
451	vfree(old_wc);
452
453	if (cq->ip) {
454		struct qib_ibdev *dev = to_idev(ibcq->device);
455		struct qib_mmap_info *ip = cq->ip;
456
457		qib_update_mmap_info(dev, ip, sz, wc);
458
459		/*
460		 * Return the offset to mmap.
461		 * See qib_mmap() for details.
462		 */
463		if (udata && udata->outlen >= sizeof(__u64)) {
464			ret = ib_copy_to_udata(udata, &ip->offset,
465					       sizeof(ip->offset));
466			if (ret)
467				goto bail;
468		}
469
470		spin_lock_irq(&dev->pending_lock);
471		if (list_empty(&ip->pending_mmaps))
472			list_add(&ip->pending_mmaps, &dev->pending_mmaps);
473		spin_unlock_irq(&dev->pending_lock);
474	}
475
476	ret = 0;
477	goto bail;
478
479bail_unlock:
480	spin_unlock_irq(&cq->lock);
481bail_free:
482	vfree(wc);
483bail:
484	return ret;
485}
486