1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2019 Mellanox Technologies. All rights reserved.
4  */
5 #include <rdma/ib_verbs.h>
6 #include <rdma/rdma_counter.h>
7 
8 #include "core_priv.h"
9 #include "restrack.h"
10 
11 #define ALL_AUTO_MODE_MASKS (RDMA_COUNTER_MASK_QP_TYPE | RDMA_COUNTER_MASK_PID)
12 
13 static int __counter_set_mode(struct rdma_counter_mode *curr,
14 			      enum rdma_nl_counter_mode new_mode,
15 			      enum rdma_nl_counter_mask new_mask)
16 {
17 	if ((new_mode == RDMA_COUNTER_MODE_AUTO) &&
18 	    ((new_mask & (~ALL_AUTO_MODE_MASKS)) ||
19 	     (curr->mode != RDMA_COUNTER_MODE_NONE)))
20 		return -EINVAL;
21 
22 	curr->mode = new_mode;
23 	curr->mask = new_mask;
24 	return 0;
25 }
26 
27 /**
28  * rdma_counter_set_auto_mode() - Turn on/off per-port auto mode
29  *
30  * When @on is true, the @mask must be set; When @on is false, it goes
31  * into manual mode if there's any counter, so that the user is able to
32  * manually access them.
33  */
34 int rdma_counter_set_auto_mode(struct ib_device *dev, u8 port,
35 			       bool on, enum rdma_nl_counter_mask mask)
36 {
37 	struct rdma_port_counter *port_counter;
38 	int ret;
39 
40 	port_counter = &dev->port_data[port].port_counter;
41 	if (!port_counter->hstats)
42 		return -EOPNOTSUPP;
43 
44 	mutex_lock(&port_counter->lock);
45 	if (on) {
46 		ret = __counter_set_mode(&port_counter->mode,
47 					 RDMA_COUNTER_MODE_AUTO, mask);
48 	} else {
49 		if (port_counter->mode.mode != RDMA_COUNTER_MODE_AUTO) {
50 			ret = -EINVAL;
51 			goto out;
52 		}
53 
54 		if (port_counter->num_counters)
55 			ret = __counter_set_mode(&port_counter->mode,
56 						 RDMA_COUNTER_MODE_MANUAL, 0);
57 		else
58 			ret = __counter_set_mode(&port_counter->mode,
59 						 RDMA_COUNTER_MODE_NONE, 0);
60 	}
61 
62 out:
63 	mutex_unlock(&port_counter->lock);
64 	return ret;
65 }
66 
67 static struct rdma_counter *rdma_counter_alloc(struct ib_device *dev, u8 port,
68 					       enum rdma_nl_counter_mode mode)
69 {
70 	struct rdma_port_counter *port_counter;
71 	struct rdma_counter *counter;
72 	int ret;
73 
74 	if (!dev->ops.counter_dealloc || !dev->ops.counter_alloc_stats)
75 		return NULL;
76 
77 	counter = kzalloc(sizeof(*counter), GFP_KERNEL);
78 	if (!counter)
79 		return NULL;
80 
81 	counter->device    = dev;
82 	counter->port      = port;
83 
84 	rdma_restrack_new(&counter->res, RDMA_RESTRACK_COUNTER);
85 	counter->stats = dev->ops.counter_alloc_stats(counter);
86 	if (!counter->stats)
87 		goto err_stats;
88 
89 	port_counter = &dev->port_data[port].port_counter;
90 	mutex_lock(&port_counter->lock);
91 	if (mode == RDMA_COUNTER_MODE_MANUAL) {
92 		ret = __counter_set_mode(&port_counter->mode,
93 					 RDMA_COUNTER_MODE_MANUAL, 0);
94 		if (ret)
95 			goto err_mode;
96 	}
97 
98 	port_counter->num_counters++;
99 	mutex_unlock(&port_counter->lock);
100 
101 	counter->mode.mode = mode;
102 	kref_init(&counter->kref);
103 	mutex_init(&counter->lock);
104 
105 	return counter;
106 
107 err_mode:
108 	mutex_unlock(&port_counter->lock);
109 	kfree(counter->stats);
110 err_stats:
111 	rdma_restrack_put(&counter->res);
112 	kfree(counter);
113 	return NULL;
114 }
115 
116 static void rdma_counter_free(struct rdma_counter *counter)
117 {
118 	struct rdma_port_counter *port_counter;
119 
120 	port_counter = &counter->device->port_data[counter->port].port_counter;
121 	mutex_lock(&port_counter->lock);
122 	port_counter->num_counters--;
123 	if (!port_counter->num_counters &&
124 	    (port_counter->mode.mode == RDMA_COUNTER_MODE_MANUAL))
125 		__counter_set_mode(&port_counter->mode, RDMA_COUNTER_MODE_NONE,
126 				   0);
127 
128 	mutex_unlock(&port_counter->lock);
129 
130 	rdma_restrack_del(&counter->res);
131 	kfree(counter->stats);
132 	kfree(counter);
133 }
134 
135 static void auto_mode_init_counter(struct rdma_counter *counter,
136 				   const struct ib_qp *qp,
137 				   enum rdma_nl_counter_mask new_mask)
138 {
139 	struct auto_mode_param *param = &counter->mode.param;
140 
141 	counter->mode.mode = RDMA_COUNTER_MODE_AUTO;
142 	counter->mode.mask = new_mask;
143 
144 	if (new_mask & RDMA_COUNTER_MASK_QP_TYPE)
145 		param->qp_type = qp->qp_type;
146 }
147 
148 static bool auto_mode_match(struct ib_qp *qp, struct rdma_counter *counter,
149 			    enum rdma_nl_counter_mask auto_mask)
150 {
151 	struct auto_mode_param *param = &counter->mode.param;
152 	bool match = true;
153 
154 	if (auto_mask & RDMA_COUNTER_MASK_QP_TYPE)
155 		match &= (param->qp_type == qp->qp_type);
156 
157 	if (auto_mask & RDMA_COUNTER_MASK_PID)
158 		match &= (task_pid_nr(counter->res.task) ==
159 			  task_pid_nr(qp->res.task));
160 
161 	return match;
162 }
163 
164 static int __rdma_counter_bind_qp(struct rdma_counter *counter,
165 				  struct ib_qp *qp)
166 {
167 	int ret;
168 
169 	if (qp->counter)
170 		return -EINVAL;
171 
172 	if (!qp->device->ops.counter_bind_qp)
173 		return -EOPNOTSUPP;
174 
175 	mutex_lock(&counter->lock);
176 	ret = qp->device->ops.counter_bind_qp(counter, qp);
177 	mutex_unlock(&counter->lock);
178 
179 	return ret;
180 }
181 
182 static int __rdma_counter_unbind_qp(struct ib_qp *qp)
183 {
184 	struct rdma_counter *counter = qp->counter;
185 	int ret;
186 
187 	if (!qp->device->ops.counter_unbind_qp)
188 		return -EOPNOTSUPP;
189 
190 	mutex_lock(&counter->lock);
191 	ret = qp->device->ops.counter_unbind_qp(qp);
192 	mutex_unlock(&counter->lock);
193 
194 	return ret;
195 }
196 
197 static void counter_history_stat_update(struct rdma_counter *counter)
198 {
199 	struct ib_device *dev = counter->device;
200 	struct rdma_port_counter *port_counter;
201 	int i;
202 
203 	port_counter = &dev->port_data[counter->port].port_counter;
204 	if (!port_counter->hstats)
205 		return;
206 
207 	rdma_counter_query_stats(counter);
208 
209 	for (i = 0; i < counter->stats->num_counters; i++)
210 		port_counter->hstats->value[i] += counter->stats->value[i];
211 }
212 
213 /**
214  * rdma_get_counter_auto_mode - Find the counter that @qp should be bound
215  *     with in auto mode
216  *
217  * Return: The counter (with ref-count increased) if found
218  */
219 static struct rdma_counter *rdma_get_counter_auto_mode(struct ib_qp *qp,
220 						       u8 port)
221 {
222 	struct rdma_port_counter *port_counter;
223 	struct rdma_counter *counter = NULL;
224 	struct ib_device *dev = qp->device;
225 	struct rdma_restrack_entry *res;
226 	struct rdma_restrack_root *rt;
227 	unsigned long id = 0;
228 
229 	port_counter = &dev->port_data[port].port_counter;
230 	rt = &dev->res[RDMA_RESTRACK_COUNTER];
231 	xa_lock(&rt->xa);
232 	xa_for_each(&rt->xa, id, res) {
233 		counter = container_of(res, struct rdma_counter, res);
234 		if ((counter->device != qp->device) || (counter->port != port))
235 			goto next;
236 
237 		if (auto_mode_match(qp, counter, port_counter->mode.mask))
238 			break;
239 next:
240 		counter = NULL;
241 	}
242 
243 	if (counter && !kref_get_unless_zero(&counter->kref))
244 		counter = NULL;
245 
246 	xa_unlock(&rt->xa);
247 	return counter;
248 }
249 
250 static void rdma_counter_res_add(struct rdma_counter *counter,
251 				 struct ib_qp *qp)
252 {
253 	rdma_restrack_parent_name(&counter->res, &qp->res);
254 	rdma_restrack_add(&counter->res);
255 }
256 
257 static void counter_release(struct kref *kref)
258 {
259 	struct rdma_counter *counter;
260 
261 	counter = container_of(kref, struct rdma_counter, kref);
262 	counter_history_stat_update(counter);
263 	counter->device->ops.counter_dealloc(counter);
264 	rdma_counter_free(counter);
265 }
266 
267 /**
268  * rdma_counter_bind_qp_auto - Check and bind the QP to a counter base on
269  *   the auto-mode rule
270  */
271 int rdma_counter_bind_qp_auto(struct ib_qp *qp, u8 port)
272 {
273 	struct rdma_port_counter *port_counter;
274 	struct ib_device *dev = qp->device;
275 	struct rdma_counter *counter;
276 	int ret;
277 
278 	if (!qp->res.valid || rdma_is_kernel_res(&qp->res))
279 		return 0;
280 
281 	if (!rdma_is_port_valid(dev, port))
282 		return -EINVAL;
283 
284 	port_counter = &dev->port_data[port].port_counter;
285 	if (port_counter->mode.mode != RDMA_COUNTER_MODE_AUTO)
286 		return 0;
287 
288 	counter = rdma_get_counter_auto_mode(qp, port);
289 	if (counter) {
290 		ret = __rdma_counter_bind_qp(counter, qp);
291 		if (ret) {
292 			kref_put(&counter->kref, counter_release);
293 			return ret;
294 		}
295 	} else {
296 		counter = rdma_counter_alloc(dev, port, RDMA_COUNTER_MODE_AUTO);
297 		if (!counter)
298 			return -ENOMEM;
299 
300 		auto_mode_init_counter(counter, qp, port_counter->mode.mask);
301 
302 		ret = __rdma_counter_bind_qp(counter, qp);
303 		if (ret) {
304 			rdma_counter_free(counter);
305 			return ret;
306 		}
307 
308 		rdma_counter_res_add(counter, qp);
309 	}
310 
311 	return 0;
312 }
313 
314 /**
315  * rdma_counter_unbind_qp - Unbind a qp from a counter
316  * @force:
317  *   true - Decrease the counter ref-count anyway (e.g., qp destroy)
318  */
319 int rdma_counter_unbind_qp(struct ib_qp *qp, bool force)
320 {
321 	struct rdma_counter *counter = qp->counter;
322 	int ret;
323 
324 	if (!counter)
325 		return -EINVAL;
326 
327 	ret = __rdma_counter_unbind_qp(qp);
328 	if (ret && !force)
329 		return ret;
330 
331 	kref_put(&counter->kref, counter_release);
332 	return 0;
333 }
334 
335 int rdma_counter_query_stats(struct rdma_counter *counter)
336 {
337 	struct ib_device *dev = counter->device;
338 	int ret;
339 
340 	if (!dev->ops.counter_update_stats)
341 		return -EINVAL;
342 
343 	mutex_lock(&counter->lock);
344 	ret = dev->ops.counter_update_stats(counter);
345 	mutex_unlock(&counter->lock);
346 
347 	return ret;
348 }
349 
350 static u64 get_running_counters_hwstat_sum(struct ib_device *dev,
351 					   u8 port, u32 index)
352 {
353 	struct rdma_restrack_entry *res;
354 	struct rdma_restrack_root *rt;
355 	struct rdma_counter *counter;
356 	unsigned long id = 0;
357 	u64 sum = 0;
358 
359 	rt = &dev->res[RDMA_RESTRACK_COUNTER];
360 	xa_lock(&rt->xa);
361 	xa_for_each(&rt->xa, id, res) {
362 		if (!rdma_restrack_get(res))
363 			continue;
364 
365 		xa_unlock(&rt->xa);
366 
367 		counter = container_of(res, struct rdma_counter, res);
368 		if ((counter->device != dev) || (counter->port != port) ||
369 		    rdma_counter_query_stats(counter))
370 			goto next;
371 
372 		sum += counter->stats->value[index];
373 
374 next:
375 		xa_lock(&rt->xa);
376 		rdma_restrack_put(res);
377 	}
378 
379 	xa_unlock(&rt->xa);
380 	return sum;
381 }
382 
383 /**
384  * rdma_counter_get_hwstat_value() - Get the sum value of all counters on a
385  *   specific port, including the running ones and history data
386  */
387 u64 rdma_counter_get_hwstat_value(struct ib_device *dev, u8 port, u32 index)
388 {
389 	struct rdma_port_counter *port_counter;
390 	u64 sum;
391 
392 	port_counter = &dev->port_data[port].port_counter;
393 	if (!port_counter->hstats)
394 		return 0;
395 
396 	sum = get_running_counters_hwstat_sum(dev, port, index);
397 	sum += port_counter->hstats->value[index];
398 
399 	return sum;
400 }
401 
402 static struct ib_qp *rdma_counter_get_qp(struct ib_device *dev, u32 qp_num)
403 {
404 	struct rdma_restrack_entry *res = NULL;
405 	struct ib_qp *qp = NULL;
406 
407 	res = rdma_restrack_get_byid(dev, RDMA_RESTRACK_QP, qp_num);
408 	if (IS_ERR(res))
409 		return NULL;
410 
411 	qp = container_of(res, struct ib_qp, res);
412 	if (qp->qp_type == IB_QPT_RAW_PACKET && !capable(CAP_NET_RAW))
413 		goto err;
414 
415 	return qp;
416 
417 err:
418 	rdma_restrack_put(res);
419 	return NULL;
420 }
421 
422 static int rdma_counter_bind_qp_manual(struct rdma_counter *counter,
423 				       struct ib_qp *qp)
424 {
425 	if ((counter->device != qp->device) || (counter->port != qp->port))
426 		return -EINVAL;
427 
428 	return __rdma_counter_bind_qp(counter, qp);
429 }
430 
431 static struct rdma_counter *rdma_get_counter_by_id(struct ib_device *dev,
432 						   u32 counter_id)
433 {
434 	struct rdma_restrack_entry *res;
435 	struct rdma_counter *counter;
436 
437 	res = rdma_restrack_get_byid(dev, RDMA_RESTRACK_COUNTER, counter_id);
438 	if (IS_ERR(res))
439 		return NULL;
440 
441 	counter = container_of(res, struct rdma_counter, res);
442 	kref_get(&counter->kref);
443 	rdma_restrack_put(res);
444 
445 	return counter;
446 }
447 
448 /**
449  * rdma_counter_bind_qpn() - Bind QP @qp_num to counter @counter_id
450  */
451 int rdma_counter_bind_qpn(struct ib_device *dev, u8 port,
452 			  u32 qp_num, u32 counter_id)
453 {
454 	struct rdma_port_counter *port_counter;
455 	struct rdma_counter *counter;
456 	struct ib_qp *qp;
457 	int ret;
458 
459 	port_counter = &dev->port_data[port].port_counter;
460 	if (port_counter->mode.mode == RDMA_COUNTER_MODE_AUTO)
461 		return -EINVAL;
462 
463 	qp = rdma_counter_get_qp(dev, qp_num);
464 	if (!qp)
465 		return -ENOENT;
466 
467 	counter = rdma_get_counter_by_id(dev, counter_id);
468 	if (!counter) {
469 		ret = -ENOENT;
470 		goto err;
471 	}
472 
473 	if (rdma_is_kernel_res(&counter->res) != rdma_is_kernel_res(&qp->res)) {
474 		ret = -EINVAL;
475 		goto err_task;
476 	}
477 
478 	ret = rdma_counter_bind_qp_manual(counter, qp);
479 	if (ret)
480 		goto err_task;
481 
482 	rdma_restrack_put(&qp->res);
483 	return 0;
484 
485 err_task:
486 	kref_put(&counter->kref, counter_release);
487 err:
488 	rdma_restrack_put(&qp->res);
489 	return ret;
490 }
491 
492 /**
493  * rdma_counter_bind_qpn_alloc() - Alloc a counter and bind QP @qp_num to it
494  *   The id of new counter is returned in @counter_id
495  */
496 int rdma_counter_bind_qpn_alloc(struct ib_device *dev, u8 port,
497 				u32 qp_num, u32 *counter_id)
498 {
499 	struct rdma_port_counter *port_counter;
500 	struct rdma_counter *counter;
501 	struct ib_qp *qp;
502 	int ret;
503 
504 	if (!rdma_is_port_valid(dev, port))
505 		return -EINVAL;
506 
507 	port_counter = &dev->port_data[port].port_counter;
508 	if (!port_counter->hstats)
509 		return -EOPNOTSUPP;
510 
511 	if (port_counter->mode.mode == RDMA_COUNTER_MODE_AUTO)
512 		return -EINVAL;
513 
514 	qp = rdma_counter_get_qp(dev, qp_num);
515 	if (!qp)
516 		return -ENOENT;
517 
518 	if (rdma_is_port_valid(dev, qp->port) && (qp->port != port)) {
519 		ret = -EINVAL;
520 		goto err;
521 	}
522 
523 	counter = rdma_counter_alloc(dev, port, RDMA_COUNTER_MODE_MANUAL);
524 	if (!counter) {
525 		ret = -ENOMEM;
526 		goto err;
527 	}
528 
529 	ret = rdma_counter_bind_qp_manual(counter, qp);
530 	if (ret)
531 		goto err_bind;
532 
533 	if (counter_id)
534 		*counter_id = counter->id;
535 
536 	rdma_counter_res_add(counter, qp);
537 
538 	rdma_restrack_put(&qp->res);
539 	return ret;
540 
541 err_bind:
542 	rdma_counter_free(counter);
543 err:
544 	rdma_restrack_put(&qp->res);
545 	return ret;
546 }
547 
548 /**
549  * rdma_counter_unbind_qpn() - Unbind QP @qp_num from a counter
550  */
551 int rdma_counter_unbind_qpn(struct ib_device *dev, u8 port,
552 			    u32 qp_num, u32 counter_id)
553 {
554 	struct rdma_port_counter *port_counter;
555 	struct ib_qp *qp;
556 	int ret;
557 
558 	if (!rdma_is_port_valid(dev, port))
559 		return -EINVAL;
560 
561 	qp = rdma_counter_get_qp(dev, qp_num);
562 	if (!qp)
563 		return -ENOENT;
564 
565 	if (rdma_is_port_valid(dev, qp->port) && (qp->port != port)) {
566 		ret = -EINVAL;
567 		goto out;
568 	}
569 
570 	port_counter = &dev->port_data[port].port_counter;
571 	if (!qp->counter || qp->counter->id != counter_id ||
572 	    port_counter->mode.mode != RDMA_COUNTER_MODE_MANUAL) {
573 		ret = -EINVAL;
574 		goto out;
575 	}
576 
577 	ret = rdma_counter_unbind_qp(qp, false);
578 
579 out:
580 	rdma_restrack_put(&qp->res);
581 	return ret;
582 }
583 
584 int rdma_counter_get_mode(struct ib_device *dev, u8 port,
585 			  enum rdma_nl_counter_mode *mode,
586 			  enum rdma_nl_counter_mask *mask)
587 {
588 	struct rdma_port_counter *port_counter;
589 
590 	port_counter = &dev->port_data[port].port_counter;
591 	*mode = port_counter->mode.mode;
592 	*mask = port_counter->mode.mask;
593 
594 	return 0;
595 }
596 
597 void rdma_counter_init(struct ib_device *dev)
598 {
599 	struct rdma_port_counter *port_counter;
600 	u32 port, i;
601 
602 	if (!dev->port_data)
603 		return;
604 
605 	rdma_for_each_port(dev, port) {
606 		port_counter = &dev->port_data[port].port_counter;
607 		port_counter->mode.mode = RDMA_COUNTER_MODE_NONE;
608 		mutex_init(&port_counter->lock);
609 
610 		if (!dev->ops.alloc_hw_stats)
611 			continue;
612 
613 		port_counter->hstats = dev->ops.alloc_hw_stats(dev, port);
614 		if (!port_counter->hstats)
615 			goto fail;
616 	}
617 
618 	return;
619 
620 fail:
621 	for (i = port; i >= rdma_start_port(dev); i--) {
622 		port_counter = &dev->port_data[port].port_counter;
623 		kfree(port_counter->hstats);
624 		port_counter->hstats = NULL;
625 		mutex_destroy(&port_counter->lock);
626 	}
627 }
628 
629 void rdma_counter_release(struct ib_device *dev)
630 {
631 	struct rdma_port_counter *port_counter;
632 	u32 port;
633 
634 	rdma_for_each_port(dev, port) {
635 		port_counter = &dev->port_data[port].port_counter;
636 		kfree(port_counter->hstats);
637 		mutex_destroy(&port_counter->lock);
638 	}
639 }
640