]>
Commit | Line | Data |
---|---|---|
f931551b | 1 | /* |
85caafe3 | 2 | * Copyright (c) 2013 Intel Corporation. All rights reserved. |
f931551b RC |
3 | * Copyright (c) 2006, 2007, 2008, 2010 QLogic Corporation. All rights reserved. |
4 | * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved. | |
5 | * | |
6 | * This software is available to you under a choice of one of two | |
7 | * licenses. You may choose to be licensed under the terms of the GNU | |
8 | * General Public License (GPL) Version 2, available from the file | |
9 | * COPYING in the main directory of this source tree, or the | |
10 | * OpenIB.org BSD license below: | |
11 | * | |
12 | * Redistribution and use in source and binary forms, with or | |
13 | * without modification, are permitted provided that the following | |
14 | * conditions are met: | |
15 | * | |
16 | * - Redistributions of source code must retain the above | |
17 | * copyright notice, this list of conditions and the following | |
18 | * disclaimer. | |
19 | * | |
20 | * - Redistributions in binary form must reproduce the above | |
21 | * copyright notice, this list of conditions and the following | |
22 | * disclaimer in the documentation and/or other materials | |
23 | * provided with the distribution. | |
24 | * | |
25 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, | |
26 | * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF | |
27 | * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND | |
28 | * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS | |
29 | * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN | |
30 | * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN | |
31 | * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE | |
32 | * SOFTWARE. | |
33 | */ | |
34 | ||
35 | #include <linux/err.h> | |
36 | #include <linux/slab.h> | |
37 | #include <linux/vmalloc.h> | |
85caafe3 | 38 | #include <linux/kthread.h> |
f931551b RC |
39 | |
40 | #include "qib_verbs.h" | |
85caafe3 | 41 | #include "qib.h" |
f931551b RC |
42 | |
43 | /** | |
44 | * qib_cq_enter - add a new entry to the completion queue | |
45 | * @cq: completion queue | |
46 | * @entry: work completion entry to add | |
47 | * @sig: true if @entry is a solicitated entry | |
48 | * | |
49 | * This may be called with qp->s_lock held. | |
50 | */ | |
51 | void qib_cq_enter(struct qib_cq *cq, struct ib_wc *entry, int solicited) | |
52 | { | |
53 | struct qib_cq_wc *wc; | |
54 | unsigned long flags; | |
55 | u32 head; | |
56 | u32 next; | |
57 | ||
58 | spin_lock_irqsave(&cq->lock, flags); | |
59 | ||
60 | /* | |
61 | * Note that the head pointer might be writable by user processes. | |
62 | * Take care to verify it is a sane value. | |
63 | */ | |
64 | wc = cq->queue; | |
65 | head = wc->head; | |
66 | if (head >= (unsigned) cq->ibcq.cqe) { | |
67 | head = cq->ibcq.cqe; | |
68 | next = 0; | |
69 | } else | |
70 | next = head + 1; | |
71 | if (unlikely(next == wc->tail)) { | |
72 | spin_unlock_irqrestore(&cq->lock, flags); | |
73 | if (cq->ibcq.event_handler) { | |
74 | struct ib_event ev; | |
75 | ||
76 | ev.device = cq->ibcq.device; | |
77 | ev.element.cq = &cq->ibcq; | |
78 | ev.event = IB_EVENT_CQ_ERR; | |
79 | cq->ibcq.event_handler(&ev, cq->ibcq.cq_context); | |
80 | } | |
81 | return; | |
82 | } | |
83 | if (cq->ip) { | |
84 | wc->uqueue[head].wr_id = entry->wr_id; | |
85 | wc->uqueue[head].status = entry->status; | |
86 | wc->uqueue[head].opcode = entry->opcode; | |
87 | wc->uqueue[head].vendor_err = entry->vendor_err; | |
88 | wc->uqueue[head].byte_len = entry->byte_len; | |
89 | wc->uqueue[head].ex.imm_data = | |
90 | (__u32 __force)entry->ex.imm_data; | |
91 | wc->uqueue[head].qp_num = entry->qp->qp_num; | |
92 | wc->uqueue[head].src_qp = entry->src_qp; | |
93 | wc->uqueue[head].wc_flags = entry->wc_flags; | |
94 | wc->uqueue[head].pkey_index = entry->pkey_index; | |
95 | wc->uqueue[head].slid = entry->slid; | |
96 | wc->uqueue[head].sl = entry->sl; | |
97 | wc->uqueue[head].dlid_path_bits = entry->dlid_path_bits; | |
98 | wc->uqueue[head].port_num = entry->port_num; | |
99 | /* Make sure entry is written before the head index. */ | |
100 | smp_wmb(); | |
101 | } else | |
102 | wc->kqueue[head] = *entry; | |
103 | wc->head = next; | |
104 | ||
105 | if (cq->notify == IB_CQ_NEXT_COMP || | |
a377acd1 MM |
106 | (cq->notify == IB_CQ_SOLICITED && |
107 | (solicited || entry->status != IB_WC_SUCCESS))) { | |
85caafe3 | 108 | struct kthread_worker *worker; |
f931551b RC |
109 | /* |
110 | * This will cause send_complete() to be called in | |
111 | * another thread. | |
112 | */ | |
85caafe3 MM |
113 | smp_rmb(); |
114 | worker = cq->dd->worker; | |
115 | if (likely(worker)) { | |
116 | cq->notify = IB_CQ_NONE; | |
117 | cq->triggered++; | |
118 | queue_kthread_work(worker, &cq->comptask); | |
119 | } | |
f931551b RC |
120 | } |
121 | ||
122 | spin_unlock_irqrestore(&cq->lock, flags); | |
123 | } | |
124 | ||
125 | /** | |
126 | * qib_poll_cq - poll for work completion entries | |
127 | * @ibcq: the completion queue to poll | |
128 | * @num_entries: the maximum number of entries to return | |
129 | * @entry: pointer to array where work completions are placed | |
130 | * | |
131 | * Returns the number of completion entries polled. | |
132 | * | |
133 | * This may be called from interrupt context. Also called by ib_poll_cq() | |
134 | * in the generic verbs code. | |
135 | */ | |
136 | int qib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry) | |
137 | { | |
138 | struct qib_cq *cq = to_icq(ibcq); | |
139 | struct qib_cq_wc *wc; | |
140 | unsigned long flags; | |
141 | int npolled; | |
142 | u32 tail; | |
143 | ||
144 | /* The kernel can only poll a kernel completion queue */ | |
145 | if (cq->ip) { | |
146 | npolled = -EINVAL; | |
147 | goto bail; | |
148 | } | |
149 | ||
150 | spin_lock_irqsave(&cq->lock, flags); | |
151 | ||
152 | wc = cq->queue; | |
153 | tail = wc->tail; | |
154 | if (tail > (u32) cq->ibcq.cqe) | |
155 | tail = (u32) cq->ibcq.cqe; | |
156 | for (npolled = 0; npolled < num_entries; ++npolled, ++entry) { | |
157 | if (tail == wc->head) | |
158 | break; | |
159 | /* The kernel doesn't need a RMB since it has the lock. */ | |
160 | *entry = wc->kqueue[tail]; | |
161 | if (tail >= cq->ibcq.cqe) | |
162 | tail = 0; | |
163 | else | |
164 | tail++; | |
165 | } | |
166 | wc->tail = tail; | |
167 | ||
168 | spin_unlock_irqrestore(&cq->lock, flags); | |
169 | ||
170 | bail: | |
171 | return npolled; | |
172 | } | |
173 | ||
85caafe3 | 174 | static void send_complete(struct kthread_work *work) |
f931551b RC |
175 | { |
176 | struct qib_cq *cq = container_of(work, struct qib_cq, comptask); | |
177 | ||
178 | /* | |
179 | * The completion handler will most likely rearm the notification | |
180 | * and poll for all pending entries. If a new completion entry | |
181 | * is added while we are in this routine, queue_work() | |
182 | * won't call us again until we return so we check triggered to | |
183 | * see if we need to call the handler again. | |
184 | */ | |
185 | for (;;) { | |
186 | u8 triggered = cq->triggered; | |
187 | ||
188 | /* | |
189 | * IPoIB connected mode assumes the callback is from a | |
190 | * soft IRQ. We simulate this by blocking "bottom halves". | |
191 | * See the implementation for ipoib_cm_handle_tx_wc(), | |
192 | * netif_tx_lock_bh() and netif_tx_lock(). | |
193 | */ | |
194 | local_bh_disable(); | |
195 | cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context); | |
196 | local_bh_enable(); | |
197 | ||
198 | if (cq->triggered == triggered) | |
199 | return; | |
200 | } | |
201 | } | |
202 | ||
203 | /** | |
204 | * qib_create_cq - create a completion queue | |
205 | * @ibdev: the device this completion queue is attached to | |
bcf4c1ea | 206 | * @attr: creation attributes |
f931551b RC |
207 | * @context: unused by the QLogic_IB driver |
208 | * @udata: user data for libibverbs.so | |
209 | * | |
210 | * Returns a pointer to the completion queue or negative errno values | |
211 | * for failure. | |
212 | * | |
213 | * Called by ib_create_cq() in the generic verbs code. | |
214 | */ | |
bcf4c1ea MB |
215 | struct ib_cq *qib_create_cq(struct ib_device *ibdev, |
216 | const struct ib_cq_init_attr *attr, | |
217 | struct ib_ucontext *context, | |
f931551b RC |
218 | struct ib_udata *udata) |
219 | { | |
bcf4c1ea | 220 | int entries = attr->cqe; |
f931551b RC |
221 | struct qib_ibdev *dev = to_idev(ibdev); |
222 | struct qib_cq *cq; | |
223 | struct qib_cq_wc *wc; | |
224 | struct ib_cq *ret; | |
225 | u32 sz; | |
226 | ||
bcf4c1ea MB |
227 | if (attr->flags) |
228 | return ERR_PTR(-EINVAL); | |
229 | ||
f931551b RC |
230 | if (entries < 1 || entries > ib_qib_max_cqes) { |
231 | ret = ERR_PTR(-EINVAL); | |
232 | goto done; | |
233 | } | |
234 | ||
235 | /* Allocate the completion queue structure. */ | |
236 | cq = kmalloc(sizeof(*cq), GFP_KERNEL); | |
237 | if (!cq) { | |
238 | ret = ERR_PTR(-ENOMEM); | |
239 | goto done; | |
240 | } | |
241 | ||
242 | /* | |
243 | * Allocate the completion queue entries and head/tail pointers. | |
244 | * This is allocated separately so that it can be resized and | |
245 | * also mapped into user space. | |
246 | * We need to use vmalloc() in order to support mmap and large | |
247 | * numbers of entries. | |
248 | */ | |
249 | sz = sizeof(*wc); | |
250 | if (udata && udata->outlen >= sizeof(__u64)) | |
251 | sz += sizeof(struct ib_uverbs_wc) * (entries + 1); | |
252 | else | |
253 | sz += sizeof(struct ib_wc) * (entries + 1); | |
254 | wc = vmalloc_user(sz); | |
255 | if (!wc) { | |
256 | ret = ERR_PTR(-ENOMEM); | |
257 | goto bail_cq; | |
258 | } | |
259 | ||
260 | /* | |
261 | * Return the address of the WC as the offset to mmap. | |
262 | * See qib_mmap() for details. | |
263 | */ | |
264 | if (udata && udata->outlen >= sizeof(__u64)) { | |
265 | int err; | |
266 | ||
267 | cq->ip = qib_create_mmap_info(dev, sz, context, wc); | |
268 | if (!cq->ip) { | |
269 | ret = ERR_PTR(-ENOMEM); | |
270 | goto bail_wc; | |
271 | } | |
272 | ||
273 | err = ib_copy_to_udata(udata, &cq->ip->offset, | |
274 | sizeof(cq->ip->offset)); | |
275 | if (err) { | |
276 | ret = ERR_PTR(err); | |
277 | goto bail_ip; | |
278 | } | |
279 | } else | |
280 | cq->ip = NULL; | |
281 | ||
282 | spin_lock(&dev->n_cqs_lock); | |
283 | if (dev->n_cqs_allocated == ib_qib_max_cqs) { | |
284 | spin_unlock(&dev->n_cqs_lock); | |
285 | ret = ERR_PTR(-ENOMEM); | |
286 | goto bail_ip; | |
287 | } | |
288 | ||
289 | dev->n_cqs_allocated++; | |
290 | spin_unlock(&dev->n_cqs_lock); | |
291 | ||
292 | if (cq->ip) { | |
293 | spin_lock_irq(&dev->pending_lock); | |
294 | list_add(&cq->ip->pending_mmaps, &dev->pending_mmaps); | |
295 | spin_unlock_irq(&dev->pending_lock); | |
296 | } | |
297 | ||
298 | /* | |
299 | * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe. | |
300 | * The number of entries should be >= the number requested or return | |
301 | * an error. | |
302 | */ | |
85caafe3 | 303 | cq->dd = dd_from_dev(dev); |
f931551b RC |
304 | cq->ibcq.cqe = entries; |
305 | cq->notify = IB_CQ_NONE; | |
306 | cq->triggered = 0; | |
307 | spin_lock_init(&cq->lock); | |
85caafe3 | 308 | init_kthread_work(&cq->comptask, send_complete); |
f931551b RC |
309 | wc->head = 0; |
310 | wc->tail = 0; | |
311 | cq->queue = wc; | |
312 | ||
313 | ret = &cq->ibcq; | |
314 | ||
315 | goto done; | |
316 | ||
317 | bail_ip: | |
318 | kfree(cq->ip); | |
319 | bail_wc: | |
320 | vfree(wc); | |
321 | bail_cq: | |
322 | kfree(cq); | |
323 | done: | |
324 | return ret; | |
325 | } | |
326 | ||
327 | /** | |
328 | * qib_destroy_cq - destroy a completion queue | |
329 | * @ibcq: the completion queue to destroy. | |
330 | * | |
331 | * Returns 0 for success. | |
332 | * | |
333 | * Called by ib_destroy_cq() in the generic verbs code. | |
334 | */ | |
335 | int qib_destroy_cq(struct ib_cq *ibcq) | |
336 | { | |
337 | struct qib_ibdev *dev = to_idev(ibcq->device); | |
338 | struct qib_cq *cq = to_icq(ibcq); | |
339 | ||
85caafe3 | 340 | flush_kthread_work(&cq->comptask); |
f931551b RC |
341 | spin_lock(&dev->n_cqs_lock); |
342 | dev->n_cqs_allocated--; | |
343 | spin_unlock(&dev->n_cqs_lock); | |
344 | if (cq->ip) | |
345 | kref_put(&cq->ip->ref, qib_release_mmap_info); | |
346 | else | |
347 | vfree(cq->queue); | |
348 | kfree(cq); | |
349 | ||
350 | return 0; | |
351 | } | |
352 | ||
353 | /** | |
354 | * qib_req_notify_cq - change the notification type for a completion queue | |
355 | * @ibcq: the completion queue | |
356 | * @notify_flags: the type of notification to request | |
357 | * | |
358 | * Returns 0 for success. | |
359 | * | |
360 | * This may be called from interrupt context. Also called by | |
361 | * ib_req_notify_cq() in the generic verbs code. | |
362 | */ | |
363 | int qib_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags) | |
364 | { | |
365 | struct qib_cq *cq = to_icq(ibcq); | |
366 | unsigned long flags; | |
367 | int ret = 0; | |
368 | ||
369 | spin_lock_irqsave(&cq->lock, flags); | |
370 | /* | |
371 | * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow | |
372 | * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2). | |
373 | */ | |
374 | if (cq->notify != IB_CQ_NEXT_COMP) | |
375 | cq->notify = notify_flags & IB_CQ_SOLICITED_MASK; | |
376 | ||
377 | if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) && | |
378 | cq->queue->head != cq->queue->tail) | |
379 | ret = 1; | |
380 | ||
381 | spin_unlock_irqrestore(&cq->lock, flags); | |
382 | ||
383 | return ret; | |
384 | } | |
385 | ||
386 | /** | |
387 | * qib_resize_cq - change the size of the CQ | |
388 | * @ibcq: the completion queue | |
389 | * | |
390 | * Returns 0 for success. | |
391 | */ | |
392 | int qib_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata) | |
393 | { | |
394 | struct qib_cq *cq = to_icq(ibcq); | |
395 | struct qib_cq_wc *old_wc; | |
396 | struct qib_cq_wc *wc; | |
397 | u32 head, tail, n; | |
398 | int ret; | |
399 | u32 sz; | |
400 | ||
401 | if (cqe < 1 || cqe > ib_qib_max_cqes) { | |
402 | ret = -EINVAL; | |
403 | goto bail; | |
404 | } | |
405 | ||
406 | /* | |
407 | * Need to use vmalloc() if we want to support large #s of entries. | |
408 | */ | |
409 | sz = sizeof(*wc); | |
410 | if (udata && udata->outlen >= sizeof(__u64)) | |
411 | sz += sizeof(struct ib_uverbs_wc) * (cqe + 1); | |
412 | else | |
413 | sz += sizeof(struct ib_wc) * (cqe + 1); | |
414 | wc = vmalloc_user(sz); | |
415 | if (!wc) { | |
416 | ret = -ENOMEM; | |
417 | goto bail; | |
418 | } | |
419 | ||
420 | /* Check that we can write the offset to mmap. */ | |
421 | if (udata && udata->outlen >= sizeof(__u64)) { | |
422 | __u64 offset = 0; | |
423 | ||
424 | ret = ib_copy_to_udata(udata, &offset, sizeof(offset)); | |
425 | if (ret) | |
426 | goto bail_free; | |
427 | } | |
428 | ||
429 | spin_lock_irq(&cq->lock); | |
430 | /* | |
431 | * Make sure head and tail are sane since they | |
432 | * might be user writable. | |
433 | */ | |
434 | old_wc = cq->queue; | |
435 | head = old_wc->head; | |
436 | if (head > (u32) cq->ibcq.cqe) | |
437 | head = (u32) cq->ibcq.cqe; | |
438 | tail = old_wc->tail; | |
439 | if (tail > (u32) cq->ibcq.cqe) | |
440 | tail = (u32) cq->ibcq.cqe; | |
441 | if (head < tail) | |
442 | n = cq->ibcq.cqe + 1 + head - tail; | |
443 | else | |
444 | n = head - tail; | |
445 | if (unlikely((u32)cqe < n)) { | |
446 | ret = -EINVAL; | |
447 | goto bail_unlock; | |
448 | } | |
449 | for (n = 0; tail != head; n++) { | |
450 | if (cq->ip) | |
451 | wc->uqueue[n] = old_wc->uqueue[tail]; | |
452 | else | |
453 | wc->kqueue[n] = old_wc->kqueue[tail]; | |
454 | if (tail == (u32) cq->ibcq.cqe) | |
455 | tail = 0; | |
456 | else | |
457 | tail++; | |
458 | } | |
459 | cq->ibcq.cqe = cqe; | |
460 | wc->head = n; | |
461 | wc->tail = 0; | |
462 | cq->queue = wc; | |
463 | spin_unlock_irq(&cq->lock); | |
464 | ||
465 | vfree(old_wc); | |
466 | ||
467 | if (cq->ip) { | |
468 | struct qib_ibdev *dev = to_idev(ibcq->device); | |
469 | struct qib_mmap_info *ip = cq->ip; | |
470 | ||
471 | qib_update_mmap_info(dev, ip, sz, wc); | |
472 | ||
473 | /* | |
474 | * Return the offset to mmap. | |
475 | * See qib_mmap() for details. | |
476 | */ | |
477 | if (udata && udata->outlen >= sizeof(__u64)) { | |
478 | ret = ib_copy_to_udata(udata, &ip->offset, | |
479 | sizeof(ip->offset)); | |
480 | if (ret) | |
481 | goto bail; | |
482 | } | |
483 | ||
484 | spin_lock_irq(&dev->pending_lock); | |
485 | if (list_empty(&ip->pending_mmaps)) | |
486 | list_add(&ip->pending_mmaps, &dev->pending_mmaps); | |
487 | spin_unlock_irq(&dev->pending_lock); | |
488 | } | |
489 | ||
490 | ret = 0; | |
491 | goto bail; | |
492 | ||
493 | bail_unlock: | |
494 | spin_unlock_irq(&cq->lock); | |
495 | bail_free: | |
496 | vfree(wc); | |
497 | bail: | |
498 | return ret; | |
499 | } | |
85caafe3 MM |
500 | |
501 | int qib_cq_init(struct qib_devdata *dd) | |
502 | { | |
503 | int ret = 0; | |
504 | int cpu; | |
505 | struct task_struct *task; | |
506 | ||
507 | if (dd->worker) | |
508 | return 0; | |
509 | dd->worker = kzalloc(sizeof(*dd->worker), GFP_KERNEL); | |
510 | if (!dd->worker) | |
511 | return -ENOMEM; | |
512 | init_kthread_worker(dd->worker); | |
513 | task = kthread_create_on_node( | |
514 | kthread_worker_fn, | |
515 | dd->worker, | |
516 | dd->assigned_node_id, | |
517 | "qib_cq%d", dd->unit); | |
518 | if (IS_ERR(task)) | |
519 | goto task_fail; | |
520 | cpu = cpumask_first(cpumask_of_node(dd->assigned_node_id)); | |
521 | kthread_bind(task, cpu); | |
522 | wake_up_process(task); | |
523 | out: | |
524 | return ret; | |
525 | task_fail: | |
526 | ret = PTR_ERR(task); | |
527 | kfree(dd->worker); | |
528 | dd->worker = NULL; | |
529 | goto out; | |
530 | } | |
531 | ||
532 | void qib_cq_exit(struct qib_devdata *dd) | |
533 | { | |
534 | struct kthread_worker *worker; | |
535 | ||
536 | worker = dd->worker; | |
537 | if (!worker) | |
538 | return; | |
539 | /* blocks future queuing from send_complete() */ | |
540 | dd->worker = NULL; | |
541 | smp_wmb(); | |
542 | flush_kthread_worker(worker); | |
543 | kthread_stop(worker->task); | |
544 | kfree(worker); | |
545 | } |