xref: /linux/drivers/infiniband/sw/rdmavt/cq.c (revision dd6d7f85)
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /*
3  * Copyright(c) 2016 - 2018 Intel Corporation.
4  */
5 
6 #include <linux/slab.h>
7 #include <linux/vmalloc.h>
8 #include <rdma/uverbs_ioctl.h>
9 #include "cq.h"
10 #include "vt.h"
11 #include "trace.h"
12 
13 static struct workqueue_struct *comp_vector_wq;
14 
15 /**
16  * rvt_cq_enter - add a new entry to the completion queue
17  * @cq: completion queue
18  * @entry: work completion entry to add
19  * @solicited: true if @entry is solicited
20  *
21  * This may be called with qp->s_lock held.
22  *
23  * Return: return true on success, else return
24  * false if cq is full.
25  */
rvt_cq_enter(struct rvt_cq * cq,struct ib_wc * entry,bool solicited)26 bool rvt_cq_enter(struct rvt_cq *cq, struct ib_wc *entry, bool solicited)
27 {
28 	struct ib_uverbs_wc *uqueue = NULL;
29 	struct ib_wc *kqueue = NULL;
30 	struct rvt_cq_wc *u_wc = NULL;
31 	struct rvt_k_cq_wc *k_wc = NULL;
32 	unsigned long flags;
33 	u32 head;
34 	u32 next;
35 	u32 tail;
36 
37 	spin_lock_irqsave(&cq->lock, flags);
38 
39 	if (cq->ip) {
40 		u_wc = cq->queue;
41 		uqueue = &u_wc->uqueue[0];
42 		head = RDMA_READ_UAPI_ATOMIC(u_wc->head);
43 		tail = RDMA_READ_UAPI_ATOMIC(u_wc->tail);
44 	} else {
45 		k_wc = cq->kqueue;
46 		kqueue = &k_wc->kqueue[0];
47 		head = k_wc->head;
48 		tail = k_wc->tail;
49 	}
50 
51 	/*
52 	 * Note that the head pointer might be writable by
53 	 * user processes.Take care to verify it is a sane value.
54 	 */
55 	if (head >= (unsigned)cq->ibcq.cqe) {
56 		head = cq->ibcq.cqe;
57 		next = 0;
58 	} else {
59 		next = head + 1;
60 	}
61 
62 	if (unlikely(next == tail || cq->cq_full)) {
63 		struct rvt_dev_info *rdi = cq->rdi;
64 
65 		if (!cq->cq_full)
66 			rvt_pr_err_ratelimited(rdi, "CQ is full!\n");
67 		cq->cq_full = true;
68 		spin_unlock_irqrestore(&cq->lock, flags);
69 		if (cq->ibcq.event_handler) {
70 			struct ib_event ev;
71 
72 			ev.device = cq->ibcq.device;
73 			ev.element.cq = &cq->ibcq;
74 			ev.event = IB_EVENT_CQ_ERR;
75 			cq->ibcq.event_handler(&ev, cq->ibcq.cq_context);
76 		}
77 		return false;
78 	}
79 	trace_rvt_cq_enter(cq, entry, head);
80 	if (uqueue) {
81 		uqueue[head].wr_id = entry->wr_id;
82 		uqueue[head].status = entry->status;
83 		uqueue[head].opcode = entry->opcode;
84 		uqueue[head].vendor_err = entry->vendor_err;
85 		uqueue[head].byte_len = entry->byte_len;
86 		uqueue[head].ex.imm_data = entry->ex.imm_data;
87 		uqueue[head].qp_num = entry->qp->qp_num;
88 		uqueue[head].src_qp = entry->src_qp;
89 		uqueue[head].wc_flags = entry->wc_flags;
90 		uqueue[head].pkey_index = entry->pkey_index;
91 		uqueue[head].slid = ib_lid_cpu16(entry->slid);
92 		uqueue[head].sl = entry->sl;
93 		uqueue[head].dlid_path_bits = entry->dlid_path_bits;
94 		uqueue[head].port_num = entry->port_num;
95 		/* Make sure entry is written before the head index. */
96 		RDMA_WRITE_UAPI_ATOMIC(u_wc->head, next);
97 	} else {
98 		kqueue[head] = *entry;
99 		k_wc->head = next;
100 	}
101 
102 	if (cq->notify == IB_CQ_NEXT_COMP ||
103 	    (cq->notify == IB_CQ_SOLICITED &&
104 	     (solicited || entry->status != IB_WC_SUCCESS))) {
105 		/*
106 		 * This will cause send_complete() to be called in
107 		 * another thread.
108 		 */
109 		cq->notify = RVT_CQ_NONE;
110 		cq->triggered++;
111 		queue_work_on(cq->comp_vector_cpu, comp_vector_wq,
112 			      &cq->comptask);
113 	}
114 
115 	spin_unlock_irqrestore(&cq->lock, flags);
116 	return true;
117 }
118 EXPORT_SYMBOL(rvt_cq_enter);
119 
send_complete(struct work_struct * work)120 static void send_complete(struct work_struct *work)
121 {
122 	struct rvt_cq *cq = container_of(work, struct rvt_cq, comptask);
123 
124 	/*
125 	 * The completion handler will most likely rearm the notification
126 	 * and poll for all pending entries.  If a new completion entry
127 	 * is added while we are in this routine, queue_work()
128 	 * won't call us again until we return so we check triggered to
129 	 * see if we need to call the handler again.
130 	 */
131 	for (;;) {
132 		u8 triggered = cq->triggered;
133 
134 		/*
135 		 * IPoIB connected mode assumes the callback is from a
136 		 * soft IRQ. We simulate this by blocking "bottom halves".
137 		 * See the implementation for ipoib_cm_handle_tx_wc(),
138 		 * netif_tx_lock_bh() and netif_tx_lock().
139 		 */
140 		local_bh_disable();
141 		cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
142 		local_bh_enable();
143 
144 		if (cq->triggered == triggered)
145 			return;
146 	}
147 }
148 
149 /**
150  * rvt_create_cq - create a completion queue
151  * @ibcq: Allocated CQ
152  * @attr: creation attributes
153  * @attrs: uverbs bundle
154  *
155  * Called by ib_create_cq() in the generic verbs code.
156  *
157  * Return: 0 on success
158  */
rvt_create_cq(struct ib_cq * ibcq,const struct ib_cq_init_attr * attr,struct uverbs_attr_bundle * attrs)159 int rvt_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
160 		  struct uverbs_attr_bundle *attrs)
161 {
162 	struct ib_udata *udata = &attrs->driver_udata;
163 	struct ib_device *ibdev = ibcq->device;
164 	struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
165 	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
166 	struct rvt_cq_wc *u_wc = NULL;
167 	struct rvt_k_cq_wc *k_wc = NULL;
168 	u32 sz;
169 	unsigned int entries = attr->cqe;
170 	int comp_vector = attr->comp_vector;
171 	int err;
172 
173 	if (attr->flags)
174 		return -EOPNOTSUPP;
175 
176 	if (entries < 1 || entries > rdi->dparms.props.max_cqe)
177 		return -EINVAL;
178 
179 	if (comp_vector < 0)
180 		comp_vector = 0;
181 
182 	comp_vector = comp_vector % rdi->ibdev.num_comp_vectors;
183 
184 	/*
185 	 * Allocate the completion queue entries and head/tail pointers.
186 	 * This is allocated separately so that it can be resized and
187 	 * also mapped into user space.
188 	 * We need to use vmalloc() in order to support mmap and large
189 	 * numbers of entries.
190 	 */
191 	if (udata && udata->outlen >= sizeof(__u64)) {
192 		sz = sizeof(struct ib_uverbs_wc) * (entries + 1);
193 		sz += sizeof(*u_wc);
194 		u_wc = vmalloc_user(sz);
195 		if (!u_wc)
196 			return -ENOMEM;
197 	} else {
198 		sz = sizeof(struct ib_wc) * (entries + 1);
199 		sz += sizeof(*k_wc);
200 		k_wc = vzalloc_node(sz, rdi->dparms.node);
201 		if (!k_wc)
202 			return -ENOMEM;
203 	}
204 
205 	/*
206 	 * Return the address of the WC as the offset to mmap.
207 	 * See rvt_mmap() for details.
208 	 */
209 	if (udata && udata->outlen >= sizeof(__u64)) {
210 		cq->ip = rvt_create_mmap_info(rdi, sz, udata, u_wc);
211 		if (IS_ERR(cq->ip)) {
212 			err = PTR_ERR(cq->ip);
213 			goto bail_wc;
214 		}
215 
216 		err = ib_copy_to_udata(udata, &cq->ip->offset,
217 				       sizeof(cq->ip->offset));
218 		if (err)
219 			goto bail_ip;
220 	}
221 
222 	spin_lock_irq(&rdi->n_cqs_lock);
223 	if (rdi->n_cqs_allocated == rdi->dparms.props.max_cq) {
224 		spin_unlock_irq(&rdi->n_cqs_lock);
225 		err = -ENOMEM;
226 		goto bail_ip;
227 	}
228 
229 	rdi->n_cqs_allocated++;
230 	spin_unlock_irq(&rdi->n_cqs_lock);
231 
232 	if (cq->ip) {
233 		spin_lock_irq(&rdi->pending_lock);
234 		list_add(&cq->ip->pending_mmaps, &rdi->pending_mmaps);
235 		spin_unlock_irq(&rdi->pending_lock);
236 	}
237 
238 	/*
239 	 * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe.
240 	 * The number of entries should be >= the number requested or return
241 	 * an error.
242 	 */
243 	cq->rdi = rdi;
244 	if (rdi->driver_f.comp_vect_cpu_lookup)
245 		cq->comp_vector_cpu =
246 			rdi->driver_f.comp_vect_cpu_lookup(rdi, comp_vector);
247 	else
248 		cq->comp_vector_cpu =
249 			cpumask_first(cpumask_of_node(rdi->dparms.node));
250 
251 	cq->ibcq.cqe = entries;
252 	cq->notify = RVT_CQ_NONE;
253 	spin_lock_init(&cq->lock);
254 	INIT_WORK(&cq->comptask, send_complete);
255 	if (u_wc)
256 		cq->queue = u_wc;
257 	else
258 		cq->kqueue = k_wc;
259 
260 	trace_rvt_create_cq(cq, attr);
261 	return 0;
262 
263 bail_ip:
264 	kfree(cq->ip);
265 bail_wc:
266 	vfree(u_wc);
267 	vfree(k_wc);
268 	return err;
269 }
270 
271 /**
272  * rvt_destroy_cq - destroy a completion queue
273  * @ibcq: the completion queue to destroy.
274  * @udata: user data or NULL for kernel object
275  *
276  * Called by ib_destroy_cq() in the generic verbs code.
277  */
rvt_destroy_cq(struct ib_cq * ibcq,struct ib_udata * udata)278 int rvt_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
279 {
280 	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
281 	struct rvt_dev_info *rdi = cq->rdi;
282 
283 	flush_work(&cq->comptask);
284 	spin_lock_irq(&rdi->n_cqs_lock);
285 	rdi->n_cqs_allocated--;
286 	spin_unlock_irq(&rdi->n_cqs_lock);
287 	if (cq->ip)
288 		kref_put(&cq->ip->ref, rvt_release_mmap_info);
289 	else
290 		vfree(cq->kqueue);
291 	return 0;
292 }
293 
294 /**
295  * rvt_req_notify_cq - change the notification type for a completion queue
296  * @ibcq: the completion queue
297  * @notify_flags: the type of notification to request
298  *
299  * This may be called from interrupt context.  Also called by
300  * ib_req_notify_cq() in the generic verbs code.
301  *
302  * Return: 0 for success.
303  */
rvt_req_notify_cq(struct ib_cq * ibcq,enum ib_cq_notify_flags notify_flags)304 int rvt_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
305 {
306 	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
307 	unsigned long flags;
308 	int ret = 0;
309 
310 	spin_lock_irqsave(&cq->lock, flags);
311 	/*
312 	 * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow
313 	 * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2).
314 	 */
315 	if (cq->notify != IB_CQ_NEXT_COMP)
316 		cq->notify = notify_flags & IB_CQ_SOLICITED_MASK;
317 
318 	if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) {
319 		if (cq->queue) {
320 			if (RDMA_READ_UAPI_ATOMIC(cq->queue->head) !=
321 				RDMA_READ_UAPI_ATOMIC(cq->queue->tail))
322 				ret = 1;
323 		} else {
324 			if (cq->kqueue->head != cq->kqueue->tail)
325 				ret = 1;
326 		}
327 	}
328 
329 	spin_unlock_irqrestore(&cq->lock, flags);
330 
331 	return ret;
332 }
333 
334 /*
335  * rvt_resize_cq - change the size of the CQ
336  * @ibcq: the completion queue
337  *
338  * Return: 0 for success.
339  */
rvt_resize_cq(struct ib_cq * ibcq,int cqe,struct ib_udata * udata)340 int rvt_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
341 {
342 	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
343 	u32 head, tail, n;
344 	int ret;
345 	u32 sz;
346 	struct rvt_dev_info *rdi = cq->rdi;
347 	struct rvt_cq_wc *u_wc = NULL;
348 	struct rvt_cq_wc *old_u_wc = NULL;
349 	struct rvt_k_cq_wc *k_wc = NULL;
350 	struct rvt_k_cq_wc *old_k_wc = NULL;
351 
352 	if (cqe < 1 || cqe > rdi->dparms.props.max_cqe)
353 		return -EINVAL;
354 
355 	/*
356 	 * Need to use vmalloc() if we want to support large #s of entries.
357 	 */
358 	if (udata && udata->outlen >= sizeof(__u64)) {
359 		sz = sizeof(struct ib_uverbs_wc) * (cqe + 1);
360 		sz += sizeof(*u_wc);
361 		u_wc = vmalloc_user(sz);
362 		if (!u_wc)
363 			return -ENOMEM;
364 	} else {
365 		sz = sizeof(struct ib_wc) * (cqe + 1);
366 		sz += sizeof(*k_wc);
367 		k_wc = vzalloc_node(sz, rdi->dparms.node);
368 		if (!k_wc)
369 			return -ENOMEM;
370 	}
371 	/* Check that we can write the offset to mmap. */
372 	if (udata && udata->outlen >= sizeof(__u64)) {
373 		__u64 offset = 0;
374 
375 		ret = ib_copy_to_udata(udata, &offset, sizeof(offset));
376 		if (ret)
377 			goto bail_free;
378 	}
379 
380 	spin_lock_irq(&cq->lock);
381 	/*
382 	 * Make sure head and tail are sane since they
383 	 * might be user writable.
384 	 */
385 	if (u_wc) {
386 		old_u_wc = cq->queue;
387 		head = RDMA_READ_UAPI_ATOMIC(old_u_wc->head);
388 		tail = RDMA_READ_UAPI_ATOMIC(old_u_wc->tail);
389 	} else {
390 		old_k_wc = cq->kqueue;
391 		head = old_k_wc->head;
392 		tail = old_k_wc->tail;
393 	}
394 
395 	if (head > (u32)cq->ibcq.cqe)
396 		head = (u32)cq->ibcq.cqe;
397 	if (tail > (u32)cq->ibcq.cqe)
398 		tail = (u32)cq->ibcq.cqe;
399 	if (head < tail)
400 		n = cq->ibcq.cqe + 1 + head - tail;
401 	else
402 		n = head - tail;
403 	if (unlikely((u32)cqe < n)) {
404 		ret = -EINVAL;
405 		goto bail_unlock;
406 	}
407 	for (n = 0; tail != head; n++) {
408 		if (u_wc)
409 			u_wc->uqueue[n] = old_u_wc->uqueue[tail];
410 		else
411 			k_wc->kqueue[n] = old_k_wc->kqueue[tail];
412 		if (tail == (u32)cq->ibcq.cqe)
413 			tail = 0;
414 		else
415 			tail++;
416 	}
417 	cq->ibcq.cqe = cqe;
418 	if (u_wc) {
419 		RDMA_WRITE_UAPI_ATOMIC(u_wc->head, n);
420 		RDMA_WRITE_UAPI_ATOMIC(u_wc->tail, 0);
421 		cq->queue = u_wc;
422 	} else {
423 		k_wc->head = n;
424 		k_wc->tail = 0;
425 		cq->kqueue = k_wc;
426 	}
427 	spin_unlock_irq(&cq->lock);
428 
429 	if (u_wc)
430 		vfree(old_u_wc);
431 	else
432 		vfree(old_k_wc);
433 
434 	if (cq->ip) {
435 		struct rvt_mmap_info *ip = cq->ip;
436 
437 		rvt_update_mmap_info(rdi, ip, sz, u_wc);
438 
439 		/*
440 		 * Return the offset to mmap.
441 		 * See rvt_mmap() for details.
442 		 */
443 		if (udata && udata->outlen >= sizeof(__u64)) {
444 			ret = ib_copy_to_udata(udata, &ip->offset,
445 					       sizeof(ip->offset));
446 			if (ret)
447 				return ret;
448 		}
449 
450 		spin_lock_irq(&rdi->pending_lock);
451 		if (list_empty(&ip->pending_mmaps))
452 			list_add(&ip->pending_mmaps, &rdi->pending_mmaps);
453 		spin_unlock_irq(&rdi->pending_lock);
454 	}
455 
456 	return 0;
457 
458 bail_unlock:
459 	spin_unlock_irq(&cq->lock);
460 bail_free:
461 	vfree(u_wc);
462 	vfree(k_wc);
463 
464 	return ret;
465 }
466 
467 /**
468  * rvt_poll_cq - poll for work completion entries
469  * @ibcq: the completion queue to poll
470  * @num_entries: the maximum number of entries to return
471  * @entry: pointer to array where work completions are placed
472  *
473  * This may be called from interrupt context.  Also called by ib_poll_cq()
474  * in the generic verbs code.
475  *
476  * Return: the number of completion entries polled.
477  */
rvt_poll_cq(struct ib_cq * ibcq,int num_entries,struct ib_wc * entry)478 int rvt_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
479 {
480 	struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
481 	struct rvt_k_cq_wc *wc;
482 	unsigned long flags;
483 	int npolled;
484 	u32 tail;
485 
486 	/* The kernel can only poll a kernel completion queue */
487 	if (cq->ip)
488 		return -EINVAL;
489 
490 	spin_lock_irqsave(&cq->lock, flags);
491 
492 	wc = cq->kqueue;
493 	tail = wc->tail;
494 	if (tail > (u32)cq->ibcq.cqe)
495 		tail = (u32)cq->ibcq.cqe;
496 	for (npolled = 0; npolled < num_entries; ++npolled, ++entry) {
497 		if (tail == wc->head)
498 			break;
499 		/* The kernel doesn't need a RMB since it has the lock. */
500 		trace_rvt_cq_poll(cq, &wc->kqueue[tail], npolled);
501 		*entry = wc->kqueue[tail];
502 		if (tail >= cq->ibcq.cqe)
503 			tail = 0;
504 		else
505 			tail++;
506 	}
507 	wc->tail = tail;
508 
509 	spin_unlock_irqrestore(&cq->lock, flags);
510 
511 	return npolled;
512 }
513 
514 /**
515  * rvt_driver_cq_init - Init cq resources on behalf of driver
516  *
517  * Return: 0 on success
518  */
rvt_driver_cq_init(void)519 int rvt_driver_cq_init(void)
520 {
521 	comp_vector_wq = alloc_workqueue("%s", WQ_HIGHPRI | WQ_CPU_INTENSIVE,
522 					 0, "rdmavt_cq");
523 	if (!comp_vector_wq)
524 		return -ENOMEM;
525 
526 	return 0;
527 }
528 
529 /**
530  * rvt_cq_exit - tear down cq reources
531  */
rvt_cq_exit(void)532 void rvt_cq_exit(void)
533 {
534 	destroy_workqueue(comp_vector_wq);
535 	comp_vector_wq = NULL;
536 }
537