Commit | Line | Data |
---|---|---|
77241056 MM |
1 | /* |
2 | * | |
3 | * This file is provided under a dual BSD/GPLv2 license. When using or | |
4 | * redistributing this file, you may do so under either license. | |
5 | * | |
6 | * GPL LICENSE SUMMARY | |
7 | * | |
8 | * Copyright(c) 2015 Intel Corporation. | |
9 | * | |
10 | * This program is free software; you can redistribute it and/or modify | |
11 | * it under the terms of version 2 of the GNU General Public License as | |
12 | * published by the Free Software Foundation. | |
13 | * | |
14 | * This program is distributed in the hope that it will be useful, but | |
15 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
17 | * General Public License for more details. | |
18 | * | |
19 | * BSD LICENSE | |
20 | * | |
21 | * Copyright(c) 2015 Intel Corporation. | |
22 | * | |
23 | * Redistribution and use in source and binary forms, with or without | |
24 | * modification, are permitted provided that the following conditions | |
25 | * are met: | |
26 | * | |
27 | * - Redistributions of source code must retain the above copyright | |
28 | * notice, this list of conditions and the following disclaimer. | |
29 | * - Redistributions in binary form must reproduce the above copyright | |
30 | * notice, this list of conditions and the following disclaimer in | |
31 | * the documentation and/or other materials provided with the | |
32 | * distribution. | |
33 | * - Neither the name of Intel Corporation nor the names of its | |
34 | * contributors may be used to endorse or promote products derived | |
35 | * from this software without specific prior written permission. | |
36 | * | |
37 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS | |
38 | * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT | |
39 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR | |
40 | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT | |
41 | * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, | |
42 | * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT | |
43 | * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, | |
44 | * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY | |
45 | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT | |
46 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE | |
47 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |
48 | * | |
49 | */ | |
50 | ||
51 | #include <linux/err.h> | |
52 | #include <linux/slab.h> | |
53 | #include <linux/vmalloc.h> | |
54 | #include <linux/kthread.h> | |
55 | ||
56 | #include "verbs.h" | |
57 | #include "hfi.h" | |
58 | ||
59 | /** | |
60 | * hfi1_cq_enter - add a new entry to the completion queue | |
61 | * @cq: completion queue | |
62 | * @entry: work completion entry to add | |
63 | * @sig: true if @entry is a solicited entry | |
64 | * | |
65 | * This may be called with qp->s_lock held. | |
66 | */ | |
67 | void hfi1_cq_enter(struct hfi1_cq *cq, struct ib_wc *entry, int solicited) | |
68 | { | |
69 | struct hfi1_cq_wc *wc; | |
70 | unsigned long flags; | |
71 | u32 head; | |
72 | u32 next; | |
73 | ||
74 | spin_lock_irqsave(&cq->lock, flags); | |
75 | ||
76 | /* | |
77 | * Note that the head pointer might be writable by user processes. | |
78 | * Take care to verify it is a sane value. | |
79 | */ | |
80 | wc = cq->queue; | |
81 | head = wc->head; | |
82 | if (head >= (unsigned) cq->ibcq.cqe) { | |
83 | head = cq->ibcq.cqe; | |
84 | next = 0; | |
85 | } else | |
86 | next = head + 1; | |
87 | if (unlikely(next == wc->tail)) { | |
88 | spin_unlock_irqrestore(&cq->lock, flags); | |
89 | if (cq->ibcq.event_handler) { | |
90 | struct ib_event ev; | |
91 | ||
92 | ev.device = cq->ibcq.device; | |
93 | ev.element.cq = &cq->ibcq; | |
94 | ev.event = IB_EVENT_CQ_ERR; | |
95 | cq->ibcq.event_handler(&ev, cq->ibcq.cq_context); | |
96 | } | |
97 | return; | |
98 | } | |
99 | if (cq->ip) { | |
100 | wc->uqueue[head].wr_id = entry->wr_id; | |
101 | wc->uqueue[head].status = entry->status; | |
102 | wc->uqueue[head].opcode = entry->opcode; | |
103 | wc->uqueue[head].vendor_err = entry->vendor_err; | |
104 | wc->uqueue[head].byte_len = entry->byte_len; | |
105 | wc->uqueue[head].ex.imm_data = | |
106 | (__u32 __force)entry->ex.imm_data; | |
107 | wc->uqueue[head].qp_num = entry->qp->qp_num; | |
108 | wc->uqueue[head].src_qp = entry->src_qp; | |
109 | wc->uqueue[head].wc_flags = entry->wc_flags; | |
110 | wc->uqueue[head].pkey_index = entry->pkey_index; | |
111 | wc->uqueue[head].slid = entry->slid; | |
112 | wc->uqueue[head].sl = entry->sl; | |
113 | wc->uqueue[head].dlid_path_bits = entry->dlid_path_bits; | |
114 | wc->uqueue[head].port_num = entry->port_num; | |
115 | /* Make sure entry is written before the head index. */ | |
116 | smp_wmb(); | |
117 | } else | |
118 | wc->kqueue[head] = *entry; | |
119 | wc->head = next; | |
120 | ||
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; | |
125 | /* | |
126 | * This will cause send_complete() to be called in | |
127 | * another thread. | |
128 | */ | |
129 | smp_read_barrier_depends(); /* see hfi1_cq_exit */ | |
130 | worker = cq->dd->worker; | |
131 | if (likely(worker)) { | |
132 | cq->notify = IB_CQ_NONE; | |
133 | cq->triggered++; | |
134 | queue_kthread_work(worker, &cq->comptask); | |
135 | } | |
136 | } | |
137 | ||
138 | spin_unlock_irqrestore(&cq->lock, flags); | |
139 | } | |
140 | ||
141 | /** | |
142 | * hfi1_poll_cq - poll for work completion entries | |
143 | * @ibcq: the completion queue to poll | |
144 | * @num_entries: the maximum number of entries to return | |
145 | * @entry: pointer to array where work completions are placed | |
146 | * | |
147 | * Returns the number of completion entries polled. | |
148 | * | |
149 | * This may be called from interrupt context. Also called by ib_poll_cq() | |
150 | * in the generic verbs code. | |
151 | */ | |
152 | int hfi1_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry) | |
153 | { | |
154 | struct hfi1_cq *cq = to_icq(ibcq); | |
155 | struct hfi1_cq_wc *wc; | |
156 | unsigned long flags; | |
157 | int npolled; | |
158 | u32 tail; | |
159 | ||
160 | /* The kernel can only poll a kernel completion queue */ | |
161 | if (cq->ip) { | |
162 | npolled = -EINVAL; | |
163 | goto bail; | |
164 | } | |
165 | ||
166 | spin_lock_irqsave(&cq->lock, flags); | |
167 | ||
168 | wc = cq->queue; | |
169 | tail = wc->tail; | |
170 | if (tail > (u32) cq->ibcq.cqe) | |
171 | tail = (u32) cq->ibcq.cqe; | |
172 | for (npolled = 0; npolled < num_entries; ++npolled, ++entry) { | |
173 | if (tail == wc->head) | |
174 | break; | |
175 | /* The kernel doesn't need a RMB since it has the lock. */ | |
176 | *entry = wc->kqueue[tail]; | |
177 | if (tail >= cq->ibcq.cqe) | |
178 | tail = 0; | |
179 | else | |
180 | tail++; | |
181 | } | |
182 | wc->tail = tail; | |
183 | ||
184 | spin_unlock_irqrestore(&cq->lock, flags); | |
185 | ||
186 | bail: | |
187 | return npolled; | |
188 | } | |
189 | ||
190 | static void send_complete(struct kthread_work *work) | |
191 | { | |
192 | struct hfi1_cq *cq = container_of(work, struct hfi1_cq, comptask); | |
193 | ||
194 | /* | |
195 | * The completion handler will most likely rearm the notification | |
196 | * and poll for all pending entries. If a new completion entry | |
197 | * is added while we are in this routine, queue_work() | |
198 | * won't call us again until we return so we check triggered to | |
199 | * see if we need to call the handler again. | |
200 | */ | |
201 | for (;;) { | |
202 | u8 triggered = cq->triggered; | |
203 | ||
204 | /* | |
205 | * IPoIB connected mode assumes the callback is from a | |
206 | * soft IRQ. We simulate this by blocking "bottom halves". | |
207 | * See the implementation for ipoib_cm_handle_tx_wc(), | |
208 | * netif_tx_lock_bh() and netif_tx_lock(). | |
209 | */ | |
210 | local_bh_disable(); | |
211 | cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context); | |
212 | local_bh_enable(); | |
213 | ||
214 | if (cq->triggered == triggered) | |
215 | return; | |
216 | } | |
217 | } | |
218 | ||
219 | /** | |
220 | * hfi1_create_cq - create a completion queue | |
221 | * @ibdev: the device this completion queue is attached to | |
222 | * @attr: creation attributes | |
223 | * @context: unused by the driver | |
224 | * @udata: user data for libibverbs.so | |
225 | * | |
226 | * Returns a pointer to the completion queue or negative errno values | |
227 | * for failure. | |
228 | * | |
229 | * Called by ib_create_cq() in the generic verbs code. | |
230 | */ | |
231 | struct ib_cq *hfi1_create_cq( | |
232 | struct ib_device *ibdev, | |
233 | const struct ib_cq_init_attr *attr, | |
234 | struct ib_ucontext *context, | |
235 | struct ib_udata *udata) | |
236 | { | |
237 | struct hfi1_ibdev *dev = to_idev(ibdev); | |
238 | struct hfi1_cq *cq; | |
239 | struct hfi1_cq_wc *wc; | |
240 | struct ib_cq *ret; | |
241 | u32 sz; | |
242 | unsigned int entries = attr->cqe; | |
243 | ||
244 | if (attr->flags) | |
245 | return ERR_PTR(-EINVAL); | |
246 | ||
247 | if (entries < 1 || entries > hfi1_max_cqes) | |
248 | return ERR_PTR(-EINVAL); | |
249 | ||
250 | /* Allocate the completion queue structure. */ | |
251 | cq = kmalloc(sizeof(*cq), GFP_KERNEL); | |
252 | if (!cq) | |
253 | return ERR_PTR(-ENOMEM); | |
254 | ||
255 | /* | |
256 | * Allocate the completion queue entries and head/tail pointers. | |
257 | * This is allocated separately so that it can be resized and | |
258 | * also mapped into user space. | |
259 | * We need to use vmalloc() in order to support mmap and large | |
260 | * numbers of entries. | |
261 | */ | |
262 | sz = sizeof(*wc); | |
263 | if (udata && udata->outlen >= sizeof(__u64)) | |
264 | sz += sizeof(struct ib_uverbs_wc) * (entries + 1); | |
265 | else | |
266 | sz += sizeof(struct ib_wc) * (entries + 1); | |
267 | wc = vmalloc_user(sz); | |
268 | if (!wc) { | |
269 | ret = ERR_PTR(-ENOMEM); | |
270 | goto bail_cq; | |
271 | } | |
272 | ||
273 | /* | |
274 | * Return the address of the WC as the offset to mmap. | |
275 | * See hfi1_mmap() for details. | |
276 | */ | |
277 | if (udata && udata->outlen >= sizeof(__u64)) { | |
278 | int err; | |
279 | ||
280 | cq->ip = hfi1_create_mmap_info(dev, sz, context, wc); | |
281 | if (!cq->ip) { | |
282 | ret = ERR_PTR(-ENOMEM); | |
283 | goto bail_wc; | |
284 | } | |
285 | ||
286 | err = ib_copy_to_udata(udata, &cq->ip->offset, | |
287 | sizeof(cq->ip->offset)); | |
288 | if (err) { | |
289 | ret = ERR_PTR(err); | |
290 | goto bail_ip; | |
291 | } | |
292 | } else | |
293 | cq->ip = NULL; | |
294 | ||
295 | spin_lock(&dev->n_cqs_lock); | |
296 | if (dev->n_cqs_allocated == hfi1_max_cqs) { | |
297 | spin_unlock(&dev->n_cqs_lock); | |
298 | ret = ERR_PTR(-ENOMEM); | |
299 | goto bail_ip; | |
300 | } | |
301 | ||
302 | dev->n_cqs_allocated++; | |
303 | spin_unlock(&dev->n_cqs_lock); | |
304 | ||
305 | if (cq->ip) { | |
306 | spin_lock_irq(&dev->pending_lock); | |
307 | list_add(&cq->ip->pending_mmaps, &dev->pending_mmaps); | |
308 | spin_unlock_irq(&dev->pending_lock); | |
309 | } | |
310 | ||
311 | /* | |
312 | * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe. | |
313 | * The number of entries should be >= the number requested or return | |
314 | * an error. | |
315 | */ | |
316 | cq->dd = dd_from_dev(dev); | |
317 | cq->ibcq.cqe = entries; | |
318 | cq->notify = IB_CQ_NONE; | |
319 | cq->triggered = 0; | |
320 | spin_lock_init(&cq->lock); | |
321 | init_kthread_work(&cq->comptask, send_complete); | |
322 | wc->head = 0; | |
323 | wc->tail = 0; | |
324 | cq->queue = wc; | |
325 | ||
326 | ret = &cq->ibcq; | |
327 | ||
328 | goto done; | |
329 | ||
330 | bail_ip: | |
331 | kfree(cq->ip); | |
332 | bail_wc: | |
333 | vfree(wc); | |
334 | bail_cq: | |
335 | kfree(cq); | |
336 | done: | |
337 | return ret; | |
338 | } | |
339 | ||
340 | /** | |
341 | * hfi1_destroy_cq - destroy a completion queue | |
342 | * @ibcq: the completion queue to destroy. | |
343 | * | |
344 | * Returns 0 for success. | |
345 | * | |
346 | * Called by ib_destroy_cq() in the generic verbs code. | |
347 | */ | |
348 | int hfi1_destroy_cq(struct ib_cq *ibcq) | |
349 | { | |
350 | struct hfi1_ibdev *dev = to_idev(ibcq->device); | |
351 | struct hfi1_cq *cq = to_icq(ibcq); | |
352 | ||
353 | flush_kthread_work(&cq->comptask); | |
354 | spin_lock(&dev->n_cqs_lock); | |
355 | dev->n_cqs_allocated--; | |
356 | spin_unlock(&dev->n_cqs_lock); | |
357 | if (cq->ip) | |
358 | kref_put(&cq->ip->ref, hfi1_release_mmap_info); | |
359 | else | |
360 | vfree(cq->queue); | |
361 | kfree(cq); | |
362 | ||
363 | return 0; | |
364 | } | |
365 | ||
366 | /** | |
367 | * hfi1_req_notify_cq - change the notification type for a completion queue | |
368 | * @ibcq: the completion queue | |
369 | * @notify_flags: the type of notification to request | |
370 | * | |
371 | * Returns 0 for success. | |
372 | * | |
373 | * This may be called from interrupt context. Also called by | |
374 | * ib_req_notify_cq() in the generic verbs code. | |
375 | */ | |
376 | int hfi1_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags) | |
377 | { | |
378 | struct hfi1_cq *cq = to_icq(ibcq); | |
379 | unsigned long flags; | |
380 | int ret = 0; | |
381 | ||
382 | spin_lock_irqsave(&cq->lock, flags); | |
383 | /* | |
384 | * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow | |
385 | * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2). | |
386 | */ | |
387 | if (cq->notify != IB_CQ_NEXT_COMP) | |
388 | cq->notify = notify_flags & IB_CQ_SOLICITED_MASK; | |
389 | ||
390 | if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) && | |
391 | cq->queue->head != cq->queue->tail) | |
392 | ret = 1; | |
393 | ||
394 | spin_unlock_irqrestore(&cq->lock, flags); | |
395 | ||
396 | return ret; | |
397 | } | |
398 | ||
399 | /** | |
400 | * hfi1_resize_cq - change the size of the CQ | |
401 | * @ibcq: the completion queue | |
402 | * | |
403 | * Returns 0 for success. | |
404 | */ | |
405 | int hfi1_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata) | |
406 | { | |
407 | struct hfi1_cq *cq = to_icq(ibcq); | |
408 | struct hfi1_cq_wc *old_wc; | |
409 | struct hfi1_cq_wc *wc; | |
410 | u32 head, tail, n; | |
411 | int ret; | |
412 | u32 sz; | |
413 | ||
414 | if (cqe < 1 || cqe > hfi1_max_cqes) { | |
415 | ret = -EINVAL; | |
416 | goto bail; | |
417 | } | |
418 | ||
419 | /* | |
420 | * Need to use vmalloc() if we want to support large #s of entries. | |
421 | */ | |
422 | sz = sizeof(*wc); | |
423 | if (udata && udata->outlen >= sizeof(__u64)) | |
424 | sz += sizeof(struct ib_uverbs_wc) * (cqe + 1); | |
425 | else | |
426 | sz += sizeof(struct ib_wc) * (cqe + 1); | |
427 | wc = vmalloc_user(sz); | |
428 | if (!wc) { | |
429 | ret = -ENOMEM; | |
430 | goto bail; | |
431 | } | |
432 | ||
433 | /* Check that we can write the offset to mmap. */ | |
434 | if (udata && udata->outlen >= sizeof(__u64)) { | |
435 | __u64 offset = 0; | |
436 | ||
437 | ret = ib_copy_to_udata(udata, &offset, sizeof(offset)); | |
438 | if (ret) | |
439 | goto bail_free; | |
440 | } | |
441 | ||
442 | spin_lock_irq(&cq->lock); | |
443 | /* | |
444 | * Make sure head and tail are sane since they | |
445 | * might be user writable. | |
446 | */ | |
447 | old_wc = cq->queue; | |
448 | head = old_wc->head; | |
449 | if (head > (u32) cq->ibcq.cqe) | |
450 | head = (u32) cq->ibcq.cqe; | |
451 | tail = old_wc->tail; | |
452 | if (tail > (u32) cq->ibcq.cqe) | |
453 | tail = (u32) cq->ibcq.cqe; | |
454 | if (head < tail) | |
455 | n = cq->ibcq.cqe + 1 + head - tail; | |
456 | else | |
457 | n = head - tail; | |
458 | if (unlikely((u32)cqe < n)) { | |
459 | ret = -EINVAL; | |
460 | goto bail_unlock; | |
461 | } | |
462 | for (n = 0; tail != head; n++) { | |
463 | if (cq->ip) | |
464 | wc->uqueue[n] = old_wc->uqueue[tail]; | |
465 | else | |
466 | wc->kqueue[n] = old_wc->kqueue[tail]; | |
467 | if (tail == (u32) cq->ibcq.cqe) | |
468 | tail = 0; | |
469 | else | |
470 | tail++; | |
471 | } | |
472 | cq->ibcq.cqe = cqe; | |
473 | wc->head = n; | |
474 | wc->tail = 0; | |
475 | cq->queue = wc; | |
476 | spin_unlock_irq(&cq->lock); | |
477 | ||
478 | vfree(old_wc); | |
479 | ||
480 | if (cq->ip) { | |
481 | struct hfi1_ibdev *dev = to_idev(ibcq->device); | |
482 | struct hfi1_mmap_info *ip = cq->ip; | |
483 | ||
484 | hfi1_update_mmap_info(dev, ip, sz, wc); | |
485 | ||
486 | /* | |
487 | * Return the offset to mmap. | |
488 | * See hfi1_mmap() for details. | |
489 | */ | |
490 | if (udata && udata->outlen >= sizeof(__u64)) { | |
491 | ret = ib_copy_to_udata(udata, &ip->offset, | |
492 | sizeof(ip->offset)); | |
493 | if (ret) | |
494 | goto bail; | |
495 | } | |
496 | ||
497 | spin_lock_irq(&dev->pending_lock); | |
498 | if (list_empty(&ip->pending_mmaps)) | |
499 | list_add(&ip->pending_mmaps, &dev->pending_mmaps); | |
500 | spin_unlock_irq(&dev->pending_lock); | |
501 | } | |
502 | ||
503 | ret = 0; | |
504 | goto bail; | |
505 | ||
506 | bail_unlock: | |
507 | spin_unlock_irq(&cq->lock); | |
508 | bail_free: | |
509 | vfree(wc); | |
510 | bail: | |
511 | return ret; | |
512 | } | |
513 | ||
514 | int hfi1_cq_init(struct hfi1_devdata *dd) | |
515 | { | |
516 | int ret = 0; | |
517 | int cpu; | |
518 | struct task_struct *task; | |
519 | ||
520 | if (dd->worker) | |
521 | return 0; | |
522 | dd->worker = kzalloc(sizeof(*dd->worker), GFP_KERNEL); | |
523 | if (!dd->worker) | |
524 | return -ENOMEM; | |
525 | init_kthread_worker(dd->worker); | |
526 | task = kthread_create_on_node( | |
527 | kthread_worker_fn, | |
528 | dd->worker, | |
529 | dd->assigned_node_id, | |
530 | "hfi1_cq%d", dd->unit); | |
531 | if (IS_ERR(task)) | |
532 | goto task_fail; | |
533 | cpu = cpumask_first(cpumask_of_node(dd->assigned_node_id)); | |
534 | kthread_bind(task, cpu); | |
535 | wake_up_process(task); | |
536 | out: | |
537 | return ret; | |
538 | task_fail: | |
539 | ret = PTR_ERR(task); | |
540 | kfree(dd->worker); | |
541 | dd->worker = NULL; | |
542 | goto out; | |
543 | } | |
544 | ||
545 | void hfi1_cq_exit(struct hfi1_devdata *dd) | |
546 | { | |
547 | struct kthread_worker *worker; | |
548 | ||
549 | worker = dd->worker; | |
550 | if (!worker) | |
551 | return; | |
552 | /* blocks future queuing from send_complete() */ | |
553 | dd->worker = NULL; | |
554 | smp_wmb(); /* See hfi1_cq_enter */ | |
555 | flush_kthread_worker(worker); | |
556 | kthread_stop(worker->task); | |
557 | kfree(worker); | |
558 | } |