xref: /openbmc/linux/drivers/infiniband/hw/mlx5/mr.c (revision 07c7c6bf)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 
34 #include <linux/kref.h>
35 #include <linux/random.h>
36 #include <linux/debugfs.h>
37 #include <linux/export.h>
38 #include <linux/delay.h>
39 #include <rdma/ib_umem.h>
40 #include <rdma/ib_umem_odp.h>
41 #include <rdma/ib_verbs.h>
42 #include "mlx5_ib.h"
43 
44 enum {
45 	MAX_PENDING_REG_MR = 8,
46 };
47 
48 #define MLX5_UMR_ALIGN 2048
49 
50 static void clean_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr);
51 static void dereg_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr);
52 static int mr_cache_max_order(struct mlx5_ib_dev *dev);
53 static int unreg_umr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr);
54 static bool umr_can_modify_entity_size(struct mlx5_ib_dev *dev)
55 {
56 	return !MLX5_CAP_GEN(dev->mdev, umr_modify_entity_size_disabled);
57 }
58 
59 static bool umr_can_use_indirect_mkey(struct mlx5_ib_dev *dev)
60 {
61 	return !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled);
62 }
63 
64 static bool use_umr(struct mlx5_ib_dev *dev, int order)
65 {
66 	return order <= mr_cache_max_order(dev) &&
67 		umr_can_modify_entity_size(dev);
68 }
69 
70 static int destroy_mkey(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
71 {
72 	int err = mlx5_core_destroy_mkey(dev->mdev, &mr->mmkey);
73 
74 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING))
75 		/* Wait until all page fault handlers using the mr complete. */
76 		synchronize_srcu(&dev->mr_srcu);
77 
78 	return err;
79 }
80 
81 static int order2idx(struct mlx5_ib_dev *dev, int order)
82 {
83 	struct mlx5_mr_cache *cache = &dev->cache;
84 
85 	if (order < cache->ent[0].order)
86 		return 0;
87 	else
88 		return order - cache->ent[0].order;
89 }
90 
91 static bool use_umr_mtt_update(struct mlx5_ib_mr *mr, u64 start, u64 length)
92 {
93 	return ((u64)1 << mr->order) * MLX5_ADAPTER_PAGE_SIZE >=
94 		length + (start & (MLX5_ADAPTER_PAGE_SIZE - 1));
95 }
96 
97 static void update_odp_mr(struct mlx5_ib_mr *mr)
98 {
99 	if (is_odp_mr(mr)) {
100 		/*
101 		 * This barrier prevents the compiler from moving the
102 		 * setting of umem->odp_data->private to point to our
103 		 * MR, before reg_umr finished, to ensure that the MR
104 		 * initialization have finished before starting to
105 		 * handle invalidations.
106 		 */
107 		smp_wmb();
108 		to_ib_umem_odp(mr->umem)->private = mr;
109 		/*
110 		 * Make sure we will see the new
111 		 * umem->odp_data->private value in the invalidation
112 		 * routines, before we can get page faults on the
113 		 * MR. Page faults can happen once we put the MR in
114 		 * the tree, below this line. Without the barrier,
115 		 * there can be a fault handling and an invalidation
116 		 * before umem->odp_data->private == mr is visible to
117 		 * the invalidation handler.
118 		 */
119 		smp_wmb();
120 	}
121 }
122 
123 static void reg_mr_callback(int status, struct mlx5_async_work *context)
124 {
125 	struct mlx5_ib_mr *mr =
126 		container_of(context, struct mlx5_ib_mr, cb_work);
127 	struct mlx5_ib_dev *dev = mr->dev;
128 	struct mlx5_mr_cache *cache = &dev->cache;
129 	int c = order2idx(dev, mr->order);
130 	struct mlx5_cache_ent *ent = &cache->ent[c];
131 	u8 key;
132 	unsigned long flags;
133 	struct mlx5_mkey_table *table = &dev->mdev->priv.mkey_table;
134 	int err;
135 
136 	spin_lock_irqsave(&ent->lock, flags);
137 	ent->pending--;
138 	spin_unlock_irqrestore(&ent->lock, flags);
139 	if (status) {
140 		mlx5_ib_warn(dev, "async reg mr failed. status %d\n", status);
141 		kfree(mr);
142 		dev->fill_delay = 1;
143 		mod_timer(&dev->delay_timer, jiffies + HZ);
144 		return;
145 	}
146 
147 	mr->mmkey.type = MLX5_MKEY_MR;
148 	spin_lock_irqsave(&dev->mdev->priv.mkey_lock, flags);
149 	key = dev->mdev->priv.mkey_key++;
150 	spin_unlock_irqrestore(&dev->mdev->priv.mkey_lock, flags);
151 	mr->mmkey.key = mlx5_idx_to_mkey(MLX5_GET(create_mkey_out, mr->out, mkey_index)) | key;
152 
153 	cache->last_add = jiffies;
154 
155 	spin_lock_irqsave(&ent->lock, flags);
156 	list_add_tail(&mr->list, &ent->head);
157 	ent->cur++;
158 	ent->size++;
159 	spin_unlock_irqrestore(&ent->lock, flags);
160 
161 	write_lock_irqsave(&table->lock, flags);
162 	err = radix_tree_insert(&table->tree, mlx5_base_mkey(mr->mmkey.key),
163 				&mr->mmkey);
164 	if (err)
165 		pr_err("Error inserting to mkey tree. 0x%x\n", -err);
166 	write_unlock_irqrestore(&table->lock, flags);
167 
168 	if (!completion_done(&ent->compl))
169 		complete(&ent->compl);
170 }
171 
172 static int add_keys(struct mlx5_ib_dev *dev, int c, int num)
173 {
174 	struct mlx5_mr_cache *cache = &dev->cache;
175 	struct mlx5_cache_ent *ent = &cache->ent[c];
176 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
177 	struct mlx5_ib_mr *mr;
178 	void *mkc;
179 	u32 *in;
180 	int err = 0;
181 	int i;
182 
183 	in = kzalloc(inlen, GFP_KERNEL);
184 	if (!in)
185 		return -ENOMEM;
186 
187 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
188 	for (i = 0; i < num; i++) {
189 		if (ent->pending >= MAX_PENDING_REG_MR) {
190 			err = -EAGAIN;
191 			break;
192 		}
193 
194 		mr = kzalloc(sizeof(*mr), GFP_KERNEL);
195 		if (!mr) {
196 			err = -ENOMEM;
197 			break;
198 		}
199 		mr->order = ent->order;
200 		mr->allocated_from_cache = 1;
201 		mr->dev = dev;
202 
203 		MLX5_SET(mkc, mkc, free, 1);
204 		MLX5_SET(mkc, mkc, umr_en, 1);
205 		MLX5_SET(mkc, mkc, access_mode_1_0, ent->access_mode & 0x3);
206 		MLX5_SET(mkc, mkc, access_mode_4_2,
207 			 (ent->access_mode >> 2) & 0x7);
208 
209 		MLX5_SET(mkc, mkc, qpn, 0xffffff);
210 		MLX5_SET(mkc, mkc, translations_octword_size, ent->xlt);
211 		MLX5_SET(mkc, mkc, log_page_size, ent->page);
212 
213 		spin_lock_irq(&ent->lock);
214 		ent->pending++;
215 		spin_unlock_irq(&ent->lock);
216 		err = mlx5_core_create_mkey_cb(dev->mdev, &mr->mmkey,
217 					       &dev->async_ctx, in, inlen,
218 					       mr->out, sizeof(mr->out),
219 					       reg_mr_callback, &mr->cb_work);
220 		if (err) {
221 			spin_lock_irq(&ent->lock);
222 			ent->pending--;
223 			spin_unlock_irq(&ent->lock);
224 			mlx5_ib_warn(dev, "create mkey failed %d\n", err);
225 			kfree(mr);
226 			break;
227 		}
228 	}
229 
230 	kfree(in);
231 	return err;
232 }
233 
234 static void remove_keys(struct mlx5_ib_dev *dev, int c, int num)
235 {
236 	struct mlx5_mr_cache *cache = &dev->cache;
237 	struct mlx5_cache_ent *ent = &cache->ent[c];
238 	struct mlx5_ib_mr *tmp_mr;
239 	struct mlx5_ib_mr *mr;
240 	LIST_HEAD(del_list);
241 	int i;
242 
243 	for (i = 0; i < num; i++) {
244 		spin_lock_irq(&ent->lock);
245 		if (list_empty(&ent->head)) {
246 			spin_unlock_irq(&ent->lock);
247 			break;
248 		}
249 		mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
250 		list_move(&mr->list, &del_list);
251 		ent->cur--;
252 		ent->size--;
253 		spin_unlock_irq(&ent->lock);
254 		mlx5_core_destroy_mkey(dev->mdev, &mr->mmkey);
255 	}
256 
257 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING))
258 		synchronize_srcu(&dev->mr_srcu);
259 
260 	list_for_each_entry_safe(mr, tmp_mr, &del_list, list) {
261 		list_del(&mr->list);
262 		kfree(mr);
263 	}
264 }
265 
266 static ssize_t size_write(struct file *filp, const char __user *buf,
267 			  size_t count, loff_t *pos)
268 {
269 	struct mlx5_cache_ent *ent = filp->private_data;
270 	struct mlx5_ib_dev *dev = ent->dev;
271 	char lbuf[20] = {0};
272 	u32 var;
273 	int err;
274 	int c;
275 
276 	count = min(count, sizeof(lbuf) - 1);
277 	if (copy_from_user(lbuf, buf, count))
278 		return -EFAULT;
279 
280 	c = order2idx(dev, ent->order);
281 
282 	if (sscanf(lbuf, "%u", &var) != 1)
283 		return -EINVAL;
284 
285 	if (var < ent->limit)
286 		return -EINVAL;
287 
288 	if (var > ent->size) {
289 		do {
290 			err = add_keys(dev, c, var - ent->size);
291 			if (err && err != -EAGAIN)
292 				return err;
293 
294 			usleep_range(3000, 5000);
295 		} while (err);
296 	} else if (var < ent->size) {
297 		remove_keys(dev, c, ent->size - var);
298 	}
299 
300 	return count;
301 }
302 
303 static ssize_t size_read(struct file *filp, char __user *buf, size_t count,
304 			 loff_t *pos)
305 {
306 	struct mlx5_cache_ent *ent = filp->private_data;
307 	char lbuf[20];
308 	int err;
309 
310 	err = snprintf(lbuf, sizeof(lbuf), "%d\n", ent->size);
311 	if (err < 0)
312 		return err;
313 
314 	return simple_read_from_buffer(buf, count, pos, lbuf, err);
315 }
316 
317 static const struct file_operations size_fops = {
318 	.owner	= THIS_MODULE,
319 	.open	= simple_open,
320 	.write	= size_write,
321 	.read	= size_read,
322 };
323 
324 static ssize_t limit_write(struct file *filp, const char __user *buf,
325 			   size_t count, loff_t *pos)
326 {
327 	struct mlx5_cache_ent *ent = filp->private_data;
328 	struct mlx5_ib_dev *dev = ent->dev;
329 	char lbuf[20] = {0};
330 	u32 var;
331 	int err;
332 	int c;
333 
334 	count = min(count, sizeof(lbuf) - 1);
335 	if (copy_from_user(lbuf, buf, count))
336 		return -EFAULT;
337 
338 	c = order2idx(dev, ent->order);
339 
340 	if (sscanf(lbuf, "%u", &var) != 1)
341 		return -EINVAL;
342 
343 	if (var > ent->size)
344 		return -EINVAL;
345 
346 	ent->limit = var;
347 
348 	if (ent->cur < ent->limit) {
349 		err = add_keys(dev, c, 2 * ent->limit - ent->cur);
350 		if (err)
351 			return err;
352 	}
353 
354 	return count;
355 }
356 
357 static ssize_t limit_read(struct file *filp, char __user *buf, size_t count,
358 			  loff_t *pos)
359 {
360 	struct mlx5_cache_ent *ent = filp->private_data;
361 	char lbuf[20];
362 	int err;
363 
364 	err = snprintf(lbuf, sizeof(lbuf), "%d\n", ent->limit);
365 	if (err < 0)
366 		return err;
367 
368 	return simple_read_from_buffer(buf, count, pos, lbuf, err);
369 }
370 
371 static const struct file_operations limit_fops = {
372 	.owner	= THIS_MODULE,
373 	.open	= simple_open,
374 	.write	= limit_write,
375 	.read	= limit_read,
376 };
377 
378 static int someone_adding(struct mlx5_mr_cache *cache)
379 {
380 	int i;
381 
382 	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
383 		if (cache->ent[i].cur < cache->ent[i].limit)
384 			return 1;
385 	}
386 
387 	return 0;
388 }
389 
390 static void __cache_work_func(struct mlx5_cache_ent *ent)
391 {
392 	struct mlx5_ib_dev *dev = ent->dev;
393 	struct mlx5_mr_cache *cache = &dev->cache;
394 	int i = order2idx(dev, ent->order);
395 	int err;
396 
397 	if (cache->stopped)
398 		return;
399 
400 	ent = &dev->cache.ent[i];
401 	if (ent->cur < 2 * ent->limit && !dev->fill_delay) {
402 		err = add_keys(dev, i, 1);
403 		if (ent->cur < 2 * ent->limit) {
404 			if (err == -EAGAIN) {
405 				mlx5_ib_dbg(dev, "returned eagain, order %d\n",
406 					    i + 2);
407 				queue_delayed_work(cache->wq, &ent->dwork,
408 						   msecs_to_jiffies(3));
409 			} else if (err) {
410 				mlx5_ib_warn(dev, "command failed order %d, err %d\n",
411 					     i + 2, err);
412 				queue_delayed_work(cache->wq, &ent->dwork,
413 						   msecs_to_jiffies(1000));
414 			} else {
415 				queue_work(cache->wq, &ent->work);
416 			}
417 		}
418 	} else if (ent->cur > 2 * ent->limit) {
419 		/*
420 		 * The remove_keys() logic is performed as garbage collection
421 		 * task. Such task is intended to be run when no other active
422 		 * processes are running.
423 		 *
424 		 * The need_resched() will return TRUE if there are user tasks
425 		 * to be activated in near future.
426 		 *
427 		 * In such case, we don't execute remove_keys() and postpone
428 		 * the garbage collection work to try to run in next cycle,
429 		 * in order to free CPU resources to other tasks.
430 		 */
431 		if (!need_resched() && !someone_adding(cache) &&
432 		    time_after(jiffies, cache->last_add + 300 * HZ)) {
433 			remove_keys(dev, i, 1);
434 			if (ent->cur > ent->limit)
435 				queue_work(cache->wq, &ent->work);
436 		} else {
437 			queue_delayed_work(cache->wq, &ent->dwork, 300 * HZ);
438 		}
439 	}
440 }
441 
442 static void delayed_cache_work_func(struct work_struct *work)
443 {
444 	struct mlx5_cache_ent *ent;
445 
446 	ent = container_of(work, struct mlx5_cache_ent, dwork.work);
447 	__cache_work_func(ent);
448 }
449 
450 static void cache_work_func(struct work_struct *work)
451 {
452 	struct mlx5_cache_ent *ent;
453 
454 	ent = container_of(work, struct mlx5_cache_ent, work);
455 	__cache_work_func(ent);
456 }
457 
458 struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev, int entry)
459 {
460 	struct mlx5_mr_cache *cache = &dev->cache;
461 	struct mlx5_cache_ent *ent;
462 	struct mlx5_ib_mr *mr;
463 	int err;
464 
465 	if (entry < 0 || entry >= MAX_MR_CACHE_ENTRIES) {
466 		mlx5_ib_err(dev, "cache entry %d is out of range\n", entry);
467 		return NULL;
468 	}
469 
470 	ent = &cache->ent[entry];
471 	while (1) {
472 		spin_lock_irq(&ent->lock);
473 		if (list_empty(&ent->head)) {
474 			spin_unlock_irq(&ent->lock);
475 
476 			err = add_keys(dev, entry, 1);
477 			if (err && err != -EAGAIN)
478 				return ERR_PTR(err);
479 
480 			wait_for_completion(&ent->compl);
481 		} else {
482 			mr = list_first_entry(&ent->head, struct mlx5_ib_mr,
483 					      list);
484 			list_del(&mr->list);
485 			ent->cur--;
486 			spin_unlock_irq(&ent->lock);
487 			if (ent->cur < ent->limit)
488 				queue_work(cache->wq, &ent->work);
489 			return mr;
490 		}
491 	}
492 }
493 
494 static struct mlx5_ib_mr *alloc_cached_mr(struct mlx5_ib_dev *dev, int order)
495 {
496 	struct mlx5_mr_cache *cache = &dev->cache;
497 	struct mlx5_ib_mr *mr = NULL;
498 	struct mlx5_cache_ent *ent;
499 	int last_umr_cache_entry;
500 	int c;
501 	int i;
502 
503 	c = order2idx(dev, order);
504 	last_umr_cache_entry = order2idx(dev, mr_cache_max_order(dev));
505 	if (c < 0 || c > last_umr_cache_entry) {
506 		mlx5_ib_warn(dev, "order %d, cache index %d\n", order, c);
507 		return NULL;
508 	}
509 
510 	for (i = c; i <= last_umr_cache_entry; i++) {
511 		ent = &cache->ent[i];
512 
513 		mlx5_ib_dbg(dev, "order %d, cache index %d\n", ent->order, i);
514 
515 		spin_lock_irq(&ent->lock);
516 		if (!list_empty(&ent->head)) {
517 			mr = list_first_entry(&ent->head, struct mlx5_ib_mr,
518 					      list);
519 			list_del(&mr->list);
520 			ent->cur--;
521 			spin_unlock_irq(&ent->lock);
522 			if (ent->cur < ent->limit)
523 				queue_work(cache->wq, &ent->work);
524 			break;
525 		}
526 		spin_unlock_irq(&ent->lock);
527 
528 		queue_work(cache->wq, &ent->work);
529 	}
530 
531 	if (!mr)
532 		cache->ent[c].miss++;
533 
534 	return mr;
535 }
536 
537 void mlx5_mr_cache_free(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
538 {
539 	struct mlx5_mr_cache *cache = &dev->cache;
540 	struct mlx5_cache_ent *ent;
541 	int shrink = 0;
542 	int c;
543 
544 	if (!mr->allocated_from_cache)
545 		return;
546 
547 	c = order2idx(dev, mr->order);
548 	if (c < 0 || c >= MAX_MR_CACHE_ENTRIES) {
549 		mlx5_ib_warn(dev, "order %d, cache index %d\n", mr->order, c);
550 		return;
551 	}
552 
553 	if (unreg_umr(dev, mr))
554 		return;
555 
556 	ent = &cache->ent[c];
557 	spin_lock_irq(&ent->lock);
558 	list_add_tail(&mr->list, &ent->head);
559 	ent->cur++;
560 	if (ent->cur > 2 * ent->limit)
561 		shrink = 1;
562 	spin_unlock_irq(&ent->lock);
563 
564 	if (shrink)
565 		queue_work(cache->wq, &ent->work);
566 }
567 
568 static void clean_keys(struct mlx5_ib_dev *dev, int c)
569 {
570 	struct mlx5_mr_cache *cache = &dev->cache;
571 	struct mlx5_cache_ent *ent = &cache->ent[c];
572 	struct mlx5_ib_mr *tmp_mr;
573 	struct mlx5_ib_mr *mr;
574 	LIST_HEAD(del_list);
575 
576 	cancel_delayed_work(&ent->dwork);
577 	while (1) {
578 		spin_lock_irq(&ent->lock);
579 		if (list_empty(&ent->head)) {
580 			spin_unlock_irq(&ent->lock);
581 			break;
582 		}
583 		mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
584 		list_move(&mr->list, &del_list);
585 		ent->cur--;
586 		ent->size--;
587 		spin_unlock_irq(&ent->lock);
588 		mlx5_core_destroy_mkey(dev->mdev, &mr->mmkey);
589 	}
590 
591 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
592 	synchronize_srcu(&dev->mr_srcu);
593 #endif
594 
595 	list_for_each_entry_safe(mr, tmp_mr, &del_list, list) {
596 		list_del(&mr->list);
597 		kfree(mr);
598 	}
599 }
600 
601 static void mlx5_mr_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
602 {
603 	if (!mlx5_debugfs_root || dev->is_rep)
604 		return;
605 
606 	debugfs_remove_recursive(dev->cache.root);
607 	dev->cache.root = NULL;
608 }
609 
610 static void mlx5_mr_cache_debugfs_init(struct mlx5_ib_dev *dev)
611 {
612 	struct mlx5_mr_cache *cache = &dev->cache;
613 	struct mlx5_cache_ent *ent;
614 	struct dentry *dir;
615 	int i;
616 
617 	if (!mlx5_debugfs_root || dev->is_rep)
618 		return;
619 
620 	cache->root = debugfs_create_dir("mr_cache", dev->mdev->priv.dbg_root);
621 
622 	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
623 		ent = &cache->ent[i];
624 		sprintf(ent->name, "%d", ent->order);
625 		dir = debugfs_create_dir(ent->name, cache->root);
626 		debugfs_create_file("size", 0600, dir, ent, &size_fops);
627 		debugfs_create_file("limit", 0600, dir, ent, &limit_fops);
628 		debugfs_create_u32("cur", 0400, dir, &ent->cur);
629 		debugfs_create_u32("miss", 0600, dir, &ent->miss);
630 	}
631 }
632 
633 static void delay_time_func(struct timer_list *t)
634 {
635 	struct mlx5_ib_dev *dev = from_timer(dev, t, delay_timer);
636 
637 	dev->fill_delay = 0;
638 }
639 
640 int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
641 {
642 	struct mlx5_mr_cache *cache = &dev->cache;
643 	struct mlx5_cache_ent *ent;
644 	int i;
645 
646 	mutex_init(&dev->slow_path_mutex);
647 	cache->wq = alloc_ordered_workqueue("mkey_cache", WQ_MEM_RECLAIM);
648 	if (!cache->wq) {
649 		mlx5_ib_warn(dev, "failed to create work queue\n");
650 		return -ENOMEM;
651 	}
652 
653 	mlx5_cmd_init_async_ctx(dev->mdev, &dev->async_ctx);
654 	timer_setup(&dev->delay_timer, delay_time_func, 0);
655 	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
656 		ent = &cache->ent[i];
657 		INIT_LIST_HEAD(&ent->head);
658 		spin_lock_init(&ent->lock);
659 		ent->order = i + 2;
660 		ent->dev = dev;
661 		ent->limit = 0;
662 
663 		init_completion(&ent->compl);
664 		INIT_WORK(&ent->work, cache_work_func);
665 		INIT_DELAYED_WORK(&ent->dwork, delayed_cache_work_func);
666 
667 		if (i > MR_CACHE_LAST_STD_ENTRY) {
668 			mlx5_odp_init_mr_cache_entry(ent);
669 			continue;
670 		}
671 
672 		if (ent->order > mr_cache_max_order(dev))
673 			continue;
674 
675 		ent->page = PAGE_SHIFT;
676 		ent->xlt = (1 << ent->order) * sizeof(struct mlx5_mtt) /
677 			   MLX5_IB_UMR_OCTOWORD;
678 		ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
679 		if ((dev->mdev->profile->mask & MLX5_PROF_MASK_MR_CACHE) &&
680 		    !dev->is_rep &&
681 		    mlx5_core_is_pf(dev->mdev))
682 			ent->limit = dev->mdev->profile->mr_cache[i].limit;
683 		else
684 			ent->limit = 0;
685 		queue_work(cache->wq, &ent->work);
686 	}
687 
688 	mlx5_mr_cache_debugfs_init(dev);
689 
690 	return 0;
691 }
692 
693 int mlx5_mr_cache_cleanup(struct mlx5_ib_dev *dev)
694 {
695 	int i;
696 
697 	if (!dev->cache.wq)
698 		return 0;
699 
700 	dev->cache.stopped = 1;
701 	flush_workqueue(dev->cache.wq);
702 
703 	mlx5_mr_cache_debugfs_cleanup(dev);
704 	mlx5_cmd_cleanup_async_ctx(&dev->async_ctx);
705 
706 	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++)
707 		clean_keys(dev, i);
708 
709 	destroy_workqueue(dev->cache.wq);
710 	del_timer_sync(&dev->delay_timer);
711 
712 	return 0;
713 }
714 
715 struct ib_mr *mlx5_ib_get_dma_mr(struct ib_pd *pd, int acc)
716 {
717 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
718 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
719 	struct mlx5_core_dev *mdev = dev->mdev;
720 	struct mlx5_ib_mr *mr;
721 	void *mkc;
722 	u32 *in;
723 	int err;
724 
725 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
726 	if (!mr)
727 		return ERR_PTR(-ENOMEM);
728 
729 	in = kzalloc(inlen, GFP_KERNEL);
730 	if (!in) {
731 		err = -ENOMEM;
732 		goto err_free;
733 	}
734 
735 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
736 
737 	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_PA);
738 	MLX5_SET(mkc, mkc, a, !!(acc & IB_ACCESS_REMOTE_ATOMIC));
739 	MLX5_SET(mkc, mkc, rw, !!(acc & IB_ACCESS_REMOTE_WRITE));
740 	MLX5_SET(mkc, mkc, rr, !!(acc & IB_ACCESS_REMOTE_READ));
741 	MLX5_SET(mkc, mkc, lw, !!(acc & IB_ACCESS_LOCAL_WRITE));
742 	MLX5_SET(mkc, mkc, lr, 1);
743 
744 	MLX5_SET(mkc, mkc, length64, 1);
745 	MLX5_SET(mkc, mkc, pd, to_mpd(pd)->pdn);
746 	MLX5_SET(mkc, mkc, qpn, 0xffffff);
747 	MLX5_SET64(mkc, mkc, start_addr, 0);
748 
749 	err = mlx5_core_create_mkey(mdev, &mr->mmkey, in, inlen);
750 	if (err)
751 		goto err_in;
752 
753 	kfree(in);
754 	mr->mmkey.type = MLX5_MKEY_MR;
755 	mr->ibmr.lkey = mr->mmkey.key;
756 	mr->ibmr.rkey = mr->mmkey.key;
757 	mr->umem = NULL;
758 
759 	return &mr->ibmr;
760 
761 err_in:
762 	kfree(in);
763 
764 err_free:
765 	kfree(mr);
766 
767 	return ERR_PTR(err);
768 }
769 
770 static int get_octo_len(u64 addr, u64 len, int page_shift)
771 {
772 	u64 page_size = 1ULL << page_shift;
773 	u64 offset;
774 	int npages;
775 
776 	offset = addr & (page_size - 1);
777 	npages = ALIGN(len + offset, page_size) >> page_shift;
778 	return (npages + 1) / 2;
779 }
780 
781 static int mr_cache_max_order(struct mlx5_ib_dev *dev)
782 {
783 	if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
784 		return MR_CACHE_LAST_STD_ENTRY + 2;
785 	return MLX5_MAX_UMR_SHIFT;
786 }
787 
788 static int mr_umem_get(struct mlx5_ib_dev *dev, struct ib_udata *udata,
789 		       u64 start, u64 length, int access_flags,
790 		       struct ib_umem **umem, int *npages, int *page_shift,
791 		       int *ncont, int *order)
792 {
793 	struct ib_umem *u;
794 	int err;
795 
796 	*umem = NULL;
797 
798 	u = ib_umem_get(udata, start, length, access_flags, 0);
799 	err = PTR_ERR_OR_ZERO(u);
800 	if (err) {
801 		mlx5_ib_dbg(dev, "umem get failed (%d)\n", err);
802 		return err;
803 	}
804 
805 	mlx5_ib_cont_pages(u, start, MLX5_MKEY_PAGE_SHIFT_MASK, npages,
806 			   page_shift, ncont, order);
807 	if (!*npages) {
808 		mlx5_ib_warn(dev, "avoid zero region\n");
809 		ib_umem_release(u);
810 		return -EINVAL;
811 	}
812 
813 	*umem = u;
814 
815 	mlx5_ib_dbg(dev, "npages %d, ncont %d, order %d, page_shift %d\n",
816 		    *npages, *ncont, *order, *page_shift);
817 
818 	return 0;
819 }
820 
821 static void mlx5_ib_umr_done(struct ib_cq *cq, struct ib_wc *wc)
822 {
823 	struct mlx5_ib_umr_context *context =
824 		container_of(wc->wr_cqe, struct mlx5_ib_umr_context, cqe);
825 
826 	context->status = wc->status;
827 	complete(&context->done);
828 }
829 
830 static inline void mlx5_ib_init_umr_context(struct mlx5_ib_umr_context *context)
831 {
832 	context->cqe.done = mlx5_ib_umr_done;
833 	context->status = -1;
834 	init_completion(&context->done);
835 }
836 
837 static int mlx5_ib_post_send_wait(struct mlx5_ib_dev *dev,
838 				  struct mlx5_umr_wr *umrwr)
839 {
840 	struct umr_common *umrc = &dev->umrc;
841 	const struct ib_send_wr *bad;
842 	int err;
843 	struct mlx5_ib_umr_context umr_context;
844 
845 	mlx5_ib_init_umr_context(&umr_context);
846 	umrwr->wr.wr_cqe = &umr_context.cqe;
847 
848 	down(&umrc->sem);
849 	err = ib_post_send(umrc->qp, &umrwr->wr, &bad);
850 	if (err) {
851 		mlx5_ib_warn(dev, "UMR post send failed, err %d\n", err);
852 	} else {
853 		wait_for_completion(&umr_context.done);
854 		if (umr_context.status != IB_WC_SUCCESS) {
855 			mlx5_ib_warn(dev, "reg umr failed (%u)\n",
856 				     umr_context.status);
857 			err = -EFAULT;
858 		}
859 	}
860 	up(&umrc->sem);
861 	return err;
862 }
863 
864 static struct mlx5_ib_mr *alloc_mr_from_cache(
865 				  struct ib_pd *pd, struct ib_umem *umem,
866 				  u64 virt_addr, u64 len, int npages,
867 				  int page_shift, int order, int access_flags)
868 {
869 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
870 	struct mlx5_ib_mr *mr;
871 	int err = 0;
872 	int i;
873 
874 	for (i = 0; i < 1; i++) {
875 		mr = alloc_cached_mr(dev, order);
876 		if (mr)
877 			break;
878 
879 		err = add_keys(dev, order2idx(dev, order), 1);
880 		if (err && err != -EAGAIN) {
881 			mlx5_ib_warn(dev, "add_keys failed, err %d\n", err);
882 			break;
883 		}
884 	}
885 
886 	if (!mr)
887 		return ERR_PTR(-EAGAIN);
888 
889 	mr->ibmr.pd = pd;
890 	mr->umem = umem;
891 	mr->access_flags = access_flags;
892 	mr->desc_size = sizeof(struct mlx5_mtt);
893 	mr->mmkey.iova = virt_addr;
894 	mr->mmkey.size = len;
895 	mr->mmkey.pd = to_mpd(pd)->pdn;
896 
897 	return mr;
898 }
899 
900 static inline int populate_xlt(struct mlx5_ib_mr *mr, int idx, int npages,
901 			       void *xlt, int page_shift, size_t size,
902 			       int flags)
903 {
904 	struct mlx5_ib_dev *dev = mr->dev;
905 	struct ib_umem *umem = mr->umem;
906 
907 	if (flags & MLX5_IB_UPD_XLT_INDIRECT) {
908 		if (!umr_can_use_indirect_mkey(dev))
909 			return -EPERM;
910 		mlx5_odp_populate_klm(xlt, idx, npages, mr, flags);
911 		return npages;
912 	}
913 
914 	npages = min_t(size_t, npages, ib_umem_num_pages(umem) - idx);
915 
916 	if (!(flags & MLX5_IB_UPD_XLT_ZAP)) {
917 		__mlx5_ib_populate_pas(dev, umem, page_shift,
918 				       idx, npages, xlt,
919 				       MLX5_IB_MTT_PRESENT);
920 		/* Clear padding after the pages
921 		 * brought from the umem.
922 		 */
923 		memset(xlt + (npages * sizeof(struct mlx5_mtt)), 0,
924 		       size - npages * sizeof(struct mlx5_mtt));
925 	}
926 
927 	return npages;
928 }
929 
930 #define MLX5_MAX_UMR_CHUNK ((1 << (MLX5_MAX_UMR_SHIFT + 4)) - \
931 			    MLX5_UMR_MTT_ALIGNMENT)
932 #define MLX5_SPARE_UMR_CHUNK 0x10000
933 
934 int mlx5_ib_update_xlt(struct mlx5_ib_mr *mr, u64 idx, int npages,
935 		       int page_shift, int flags)
936 {
937 	struct mlx5_ib_dev *dev = mr->dev;
938 	struct device *ddev = dev->ib_dev.dev.parent;
939 	int size;
940 	void *xlt;
941 	dma_addr_t dma;
942 	struct mlx5_umr_wr wr;
943 	struct ib_sge sg;
944 	int err = 0;
945 	int desc_size = (flags & MLX5_IB_UPD_XLT_INDIRECT)
946 			       ? sizeof(struct mlx5_klm)
947 			       : sizeof(struct mlx5_mtt);
948 	const int page_align = MLX5_UMR_MTT_ALIGNMENT / desc_size;
949 	const int page_mask = page_align - 1;
950 	size_t pages_mapped = 0;
951 	size_t pages_to_map = 0;
952 	size_t pages_iter = 0;
953 	gfp_t gfp;
954 	bool use_emergency_page = false;
955 
956 	if ((flags & MLX5_IB_UPD_XLT_INDIRECT) &&
957 	    !umr_can_use_indirect_mkey(dev))
958 		return -EPERM;
959 
960 	/* UMR copies MTTs in units of MLX5_UMR_MTT_ALIGNMENT bytes,
961 	 * so we need to align the offset and length accordingly
962 	 */
963 	if (idx & page_mask) {
964 		npages += idx & page_mask;
965 		idx &= ~page_mask;
966 	}
967 
968 	gfp = flags & MLX5_IB_UPD_XLT_ATOMIC ? GFP_ATOMIC : GFP_KERNEL;
969 	gfp |= __GFP_ZERO | __GFP_NOWARN;
970 
971 	pages_to_map = ALIGN(npages, page_align);
972 	size = desc_size * pages_to_map;
973 	size = min_t(int, size, MLX5_MAX_UMR_CHUNK);
974 
975 	xlt = (void *)__get_free_pages(gfp, get_order(size));
976 	if (!xlt && size > MLX5_SPARE_UMR_CHUNK) {
977 		mlx5_ib_dbg(dev, "Failed to allocate %d bytes of order %d. fallback to spare UMR allocation od %d bytes\n",
978 			    size, get_order(size), MLX5_SPARE_UMR_CHUNK);
979 
980 		size = MLX5_SPARE_UMR_CHUNK;
981 		xlt = (void *)__get_free_pages(gfp, get_order(size));
982 	}
983 
984 	if (!xlt) {
985 		mlx5_ib_warn(dev, "Using XLT emergency buffer\n");
986 		xlt = (void *)mlx5_ib_get_xlt_emergency_page();
987 		size = PAGE_SIZE;
988 		memset(xlt, 0, size);
989 		use_emergency_page = true;
990 	}
991 	pages_iter = size / desc_size;
992 	dma = dma_map_single(ddev, xlt, size, DMA_TO_DEVICE);
993 	if (dma_mapping_error(ddev, dma)) {
994 		mlx5_ib_err(dev, "unable to map DMA during XLT update.\n");
995 		err = -ENOMEM;
996 		goto free_xlt;
997 	}
998 
999 	sg.addr = dma;
1000 	sg.lkey = dev->umrc.pd->local_dma_lkey;
1001 
1002 	memset(&wr, 0, sizeof(wr));
1003 	wr.wr.send_flags = MLX5_IB_SEND_UMR_UPDATE_XLT;
1004 	if (!(flags & MLX5_IB_UPD_XLT_ENABLE))
1005 		wr.wr.send_flags |= MLX5_IB_SEND_UMR_FAIL_IF_FREE;
1006 	wr.wr.sg_list = &sg;
1007 	wr.wr.num_sge = 1;
1008 	wr.wr.opcode = MLX5_IB_WR_UMR;
1009 
1010 	wr.pd = mr->ibmr.pd;
1011 	wr.mkey = mr->mmkey.key;
1012 	wr.length = mr->mmkey.size;
1013 	wr.virt_addr = mr->mmkey.iova;
1014 	wr.access_flags = mr->access_flags;
1015 	wr.page_shift = page_shift;
1016 
1017 	for (pages_mapped = 0;
1018 	     pages_mapped < pages_to_map && !err;
1019 	     pages_mapped += pages_iter, idx += pages_iter) {
1020 		npages = min_t(int, pages_iter, pages_to_map - pages_mapped);
1021 		dma_sync_single_for_cpu(ddev, dma, size, DMA_TO_DEVICE);
1022 		npages = populate_xlt(mr, idx, npages, xlt,
1023 				      page_shift, size, flags);
1024 
1025 		dma_sync_single_for_device(ddev, dma, size, DMA_TO_DEVICE);
1026 
1027 		sg.length = ALIGN(npages * desc_size,
1028 				  MLX5_UMR_MTT_ALIGNMENT);
1029 
1030 		if (pages_mapped + pages_iter >= pages_to_map) {
1031 			if (flags & MLX5_IB_UPD_XLT_ENABLE)
1032 				wr.wr.send_flags |=
1033 					MLX5_IB_SEND_UMR_ENABLE_MR |
1034 					MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS |
1035 					MLX5_IB_SEND_UMR_UPDATE_TRANSLATION;
1036 			if (flags & MLX5_IB_UPD_XLT_PD ||
1037 			    flags & MLX5_IB_UPD_XLT_ACCESS)
1038 				wr.wr.send_flags |=
1039 					MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS;
1040 			if (flags & MLX5_IB_UPD_XLT_ADDR)
1041 				wr.wr.send_flags |=
1042 					MLX5_IB_SEND_UMR_UPDATE_TRANSLATION;
1043 		}
1044 
1045 		wr.offset = idx * desc_size;
1046 		wr.xlt_size = sg.length;
1047 
1048 		err = mlx5_ib_post_send_wait(dev, &wr);
1049 	}
1050 	dma_unmap_single(ddev, dma, size, DMA_TO_DEVICE);
1051 
1052 free_xlt:
1053 	if (use_emergency_page)
1054 		mlx5_ib_put_xlt_emergency_page();
1055 	else
1056 		free_pages((unsigned long)xlt, get_order(size));
1057 
1058 	return err;
1059 }
1060 
1061 /*
1062  * If ibmr is NULL it will be allocated by reg_create.
1063  * Else, the given ibmr will be used.
1064  */
1065 static struct mlx5_ib_mr *reg_create(struct ib_mr *ibmr, struct ib_pd *pd,
1066 				     u64 virt_addr, u64 length,
1067 				     struct ib_umem *umem, int npages,
1068 				     int page_shift, int access_flags,
1069 				     bool populate)
1070 {
1071 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1072 	struct mlx5_ib_mr *mr;
1073 	__be64 *pas;
1074 	void *mkc;
1075 	int inlen;
1076 	u32 *in;
1077 	int err;
1078 	bool pg_cap = !!(MLX5_CAP_GEN(dev->mdev, pg));
1079 
1080 	mr = ibmr ? to_mmr(ibmr) : kzalloc(sizeof(*mr), GFP_KERNEL);
1081 	if (!mr)
1082 		return ERR_PTR(-ENOMEM);
1083 
1084 	mr->ibmr.pd = pd;
1085 	mr->access_flags = access_flags;
1086 
1087 	inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1088 	if (populate)
1089 		inlen += sizeof(*pas) * roundup(npages, 2);
1090 	in = kvzalloc(inlen, GFP_KERNEL);
1091 	if (!in) {
1092 		err = -ENOMEM;
1093 		goto err_1;
1094 	}
1095 	pas = (__be64 *)MLX5_ADDR_OF(create_mkey_in, in, klm_pas_mtt);
1096 	if (populate && !(access_flags & IB_ACCESS_ON_DEMAND))
1097 		mlx5_ib_populate_pas(dev, umem, page_shift, pas,
1098 				     pg_cap ? MLX5_IB_MTT_PRESENT : 0);
1099 
1100 	/* The pg_access bit allows setting the access flags
1101 	 * in the page list submitted with the command. */
1102 	MLX5_SET(create_mkey_in, in, pg_access, !!(pg_cap));
1103 
1104 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1105 	MLX5_SET(mkc, mkc, free, !populate);
1106 	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_MTT);
1107 	MLX5_SET(mkc, mkc, a, !!(access_flags & IB_ACCESS_REMOTE_ATOMIC));
1108 	MLX5_SET(mkc, mkc, rw, !!(access_flags & IB_ACCESS_REMOTE_WRITE));
1109 	MLX5_SET(mkc, mkc, rr, !!(access_flags & IB_ACCESS_REMOTE_READ));
1110 	MLX5_SET(mkc, mkc, lw, !!(access_flags & IB_ACCESS_LOCAL_WRITE));
1111 	MLX5_SET(mkc, mkc, lr, 1);
1112 	MLX5_SET(mkc, mkc, umr_en, 1);
1113 
1114 	MLX5_SET64(mkc, mkc, start_addr, virt_addr);
1115 	MLX5_SET64(mkc, mkc, len, length);
1116 	MLX5_SET(mkc, mkc, pd, to_mpd(pd)->pdn);
1117 	MLX5_SET(mkc, mkc, bsf_octword_size, 0);
1118 	MLX5_SET(mkc, mkc, translations_octword_size,
1119 		 get_octo_len(virt_addr, length, page_shift));
1120 	MLX5_SET(mkc, mkc, log_page_size, page_shift);
1121 	MLX5_SET(mkc, mkc, qpn, 0xffffff);
1122 	if (populate) {
1123 		MLX5_SET(create_mkey_in, in, translations_octword_actual_size,
1124 			 get_octo_len(virt_addr, length, page_shift));
1125 	}
1126 
1127 	err = mlx5_core_create_mkey(dev->mdev, &mr->mmkey, in, inlen);
1128 	if (err) {
1129 		mlx5_ib_warn(dev, "create mkey failed\n");
1130 		goto err_2;
1131 	}
1132 	mr->mmkey.type = MLX5_MKEY_MR;
1133 	mr->desc_size = sizeof(struct mlx5_mtt);
1134 	mr->dev = dev;
1135 	kvfree(in);
1136 
1137 	mlx5_ib_dbg(dev, "mkey = 0x%x\n", mr->mmkey.key);
1138 
1139 	return mr;
1140 
1141 err_2:
1142 	kvfree(in);
1143 
1144 err_1:
1145 	if (!ibmr)
1146 		kfree(mr);
1147 
1148 	return ERR_PTR(err);
1149 }
1150 
1151 static void set_mr_fields(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr,
1152 			  int npages, u64 length, int access_flags)
1153 {
1154 	mr->npages = npages;
1155 	atomic_add(npages, &dev->mdev->priv.reg_pages);
1156 	mr->ibmr.lkey = mr->mmkey.key;
1157 	mr->ibmr.rkey = mr->mmkey.key;
1158 	mr->ibmr.length = length;
1159 	mr->access_flags = access_flags;
1160 }
1161 
1162 static struct ib_mr *mlx5_ib_get_dm_mr(struct ib_pd *pd, u64 start_addr,
1163 				       u64 length, int acc, int mode)
1164 {
1165 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1166 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1167 	struct mlx5_core_dev *mdev = dev->mdev;
1168 	struct mlx5_ib_mr *mr;
1169 	void *mkc;
1170 	u32 *in;
1171 	int err;
1172 
1173 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1174 	if (!mr)
1175 		return ERR_PTR(-ENOMEM);
1176 
1177 	in = kzalloc(inlen, GFP_KERNEL);
1178 	if (!in) {
1179 		err = -ENOMEM;
1180 		goto err_free;
1181 	}
1182 
1183 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1184 
1185 	MLX5_SET(mkc, mkc, access_mode_1_0, mode & 0x3);
1186 	MLX5_SET(mkc, mkc, access_mode_4_2, (mode >> 2) & 0x7);
1187 	MLX5_SET(mkc, mkc, a, !!(acc & IB_ACCESS_REMOTE_ATOMIC));
1188 	MLX5_SET(mkc, mkc, rw, !!(acc & IB_ACCESS_REMOTE_WRITE));
1189 	MLX5_SET(mkc, mkc, rr, !!(acc & IB_ACCESS_REMOTE_READ));
1190 	MLX5_SET(mkc, mkc, lw, !!(acc & IB_ACCESS_LOCAL_WRITE));
1191 	MLX5_SET(mkc, mkc, lr, 1);
1192 
1193 	MLX5_SET64(mkc, mkc, len, length);
1194 	MLX5_SET(mkc, mkc, pd, to_mpd(pd)->pdn);
1195 	MLX5_SET(mkc, mkc, qpn, 0xffffff);
1196 	MLX5_SET64(mkc, mkc, start_addr, start_addr);
1197 
1198 	err = mlx5_core_create_mkey(mdev, &mr->mmkey, in, inlen);
1199 	if (err)
1200 		goto err_in;
1201 
1202 	kfree(in);
1203 
1204 	mr->umem = NULL;
1205 	set_mr_fields(dev, mr, 0, length, acc);
1206 
1207 	return &mr->ibmr;
1208 
1209 err_in:
1210 	kfree(in);
1211 
1212 err_free:
1213 	kfree(mr);
1214 
1215 	return ERR_PTR(err);
1216 }
1217 
1218 int mlx5_ib_advise_mr(struct ib_pd *pd,
1219 		      enum ib_uverbs_advise_mr_advice advice,
1220 		      u32 flags,
1221 		      struct ib_sge *sg_list,
1222 		      u32 num_sge,
1223 		      struct uverbs_attr_bundle *attrs)
1224 {
1225 	if (advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH &&
1226 	    advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE)
1227 		return -EOPNOTSUPP;
1228 
1229 	return mlx5_ib_advise_mr_prefetch(pd, advice, flags,
1230 					 sg_list, num_sge);
1231 }
1232 
1233 struct ib_mr *mlx5_ib_reg_dm_mr(struct ib_pd *pd, struct ib_dm *dm,
1234 				struct ib_dm_mr_attr *attr,
1235 				struct uverbs_attr_bundle *attrs)
1236 {
1237 	struct mlx5_ib_dm *mdm = to_mdm(dm);
1238 	struct mlx5_core_dev *dev = to_mdev(dm->device)->mdev;
1239 	u64 start_addr = mdm->dev_addr + attr->offset;
1240 	int mode;
1241 
1242 	switch (mdm->type) {
1243 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
1244 		if (attr->access_flags & ~MLX5_IB_DM_MEMIC_ALLOWED_ACCESS)
1245 			return ERR_PTR(-EINVAL);
1246 
1247 		mode = MLX5_MKC_ACCESS_MODE_MEMIC;
1248 		start_addr -= pci_resource_start(dev->pdev, 0);
1249 		break;
1250 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
1251 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
1252 		if (attr->access_flags & ~MLX5_IB_DM_SW_ICM_ALLOWED_ACCESS)
1253 			return ERR_PTR(-EINVAL);
1254 
1255 		mode = MLX5_MKC_ACCESS_MODE_SW_ICM;
1256 		break;
1257 	default:
1258 		return ERR_PTR(-EINVAL);
1259 	}
1260 
1261 	return mlx5_ib_get_dm_mr(pd, start_addr, attr->length,
1262 				 attr->access_flags, mode);
1263 }
1264 
1265 struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
1266 				  u64 virt_addr, int access_flags,
1267 				  struct ib_udata *udata)
1268 {
1269 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1270 	struct mlx5_ib_mr *mr = NULL;
1271 	bool populate_mtts = false;
1272 	struct ib_umem *umem;
1273 	int page_shift;
1274 	int npages;
1275 	int ncont;
1276 	int order;
1277 	int err;
1278 
1279 	if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM))
1280 		return ERR_PTR(-EOPNOTSUPP);
1281 
1282 	mlx5_ib_dbg(dev, "start 0x%llx, virt_addr 0x%llx, length 0x%llx, access_flags 0x%x\n",
1283 		    start, virt_addr, length, access_flags);
1284 
1285 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) && !start &&
1286 	    length == U64_MAX) {
1287 		if (!(access_flags & IB_ACCESS_ON_DEMAND) ||
1288 		    !(dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1289 			return ERR_PTR(-EINVAL);
1290 
1291 		mr = mlx5_ib_alloc_implicit_mr(to_mpd(pd), udata, access_flags);
1292 		if (IS_ERR(mr))
1293 			return ERR_CAST(mr);
1294 		return &mr->ibmr;
1295 	}
1296 
1297 	err = mr_umem_get(dev, udata, start, length, access_flags, &umem,
1298 			  &npages, &page_shift, &ncont, &order);
1299 
1300 	if (err < 0)
1301 		return ERR_PTR(err);
1302 
1303 	if (use_umr(dev, order)) {
1304 		mr = alloc_mr_from_cache(pd, umem, virt_addr, length, ncont,
1305 					 page_shift, order, access_flags);
1306 		if (PTR_ERR(mr) == -EAGAIN) {
1307 			mlx5_ib_dbg(dev, "cache empty for order %d\n", order);
1308 			mr = NULL;
1309 		}
1310 		populate_mtts = false;
1311 	} else if (!MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset)) {
1312 		if (access_flags & IB_ACCESS_ON_DEMAND) {
1313 			err = -EINVAL;
1314 			pr_err("Got MR registration for ODP MR > 512MB, not supported for Connect-IB\n");
1315 			goto error;
1316 		}
1317 		populate_mtts = true;
1318 	}
1319 
1320 	if (!mr) {
1321 		if (!umr_can_modify_entity_size(dev))
1322 			populate_mtts = true;
1323 		mutex_lock(&dev->slow_path_mutex);
1324 		mr = reg_create(NULL, pd, virt_addr, length, umem, ncont,
1325 				page_shift, access_flags, populate_mtts);
1326 		mutex_unlock(&dev->slow_path_mutex);
1327 	}
1328 
1329 	if (IS_ERR(mr)) {
1330 		err = PTR_ERR(mr);
1331 		goto error;
1332 	}
1333 
1334 	mlx5_ib_dbg(dev, "mkey 0x%x\n", mr->mmkey.key);
1335 
1336 	mr->umem = umem;
1337 	set_mr_fields(dev, mr, npages, length, access_flags);
1338 
1339 	update_odp_mr(mr);
1340 
1341 	if (!populate_mtts) {
1342 		int update_xlt_flags = MLX5_IB_UPD_XLT_ENABLE;
1343 
1344 		if (access_flags & IB_ACCESS_ON_DEMAND)
1345 			update_xlt_flags |= MLX5_IB_UPD_XLT_ZAP;
1346 
1347 		err = mlx5_ib_update_xlt(mr, 0, ncont, page_shift,
1348 					 update_xlt_flags);
1349 
1350 		if (err) {
1351 			dereg_mr(dev, mr);
1352 			return ERR_PTR(err);
1353 		}
1354 	}
1355 
1356 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
1357 		mr->live = 1;
1358 		atomic_set(&mr->num_pending_prefetch, 0);
1359 	}
1360 
1361 	return &mr->ibmr;
1362 error:
1363 	ib_umem_release(umem);
1364 	return ERR_PTR(err);
1365 }
1366 
1367 static int unreg_umr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
1368 {
1369 	struct mlx5_core_dev *mdev = dev->mdev;
1370 	struct mlx5_umr_wr umrwr = {};
1371 
1372 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)
1373 		return 0;
1374 
1375 	umrwr.wr.send_flags = MLX5_IB_SEND_UMR_DISABLE_MR |
1376 			      MLX5_IB_SEND_UMR_FAIL_IF_FREE;
1377 	umrwr.wr.opcode = MLX5_IB_WR_UMR;
1378 	umrwr.mkey = mr->mmkey.key;
1379 
1380 	return mlx5_ib_post_send_wait(dev, &umrwr);
1381 }
1382 
1383 static int rereg_umr(struct ib_pd *pd, struct mlx5_ib_mr *mr,
1384 		     int access_flags, int flags)
1385 {
1386 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1387 	struct mlx5_umr_wr umrwr = {};
1388 	int err;
1389 
1390 	umrwr.wr.send_flags = MLX5_IB_SEND_UMR_FAIL_IF_FREE;
1391 
1392 	umrwr.wr.opcode = MLX5_IB_WR_UMR;
1393 	umrwr.mkey = mr->mmkey.key;
1394 
1395 	if (flags & IB_MR_REREG_PD || flags & IB_MR_REREG_ACCESS) {
1396 		umrwr.pd = pd;
1397 		umrwr.access_flags = access_flags;
1398 		umrwr.wr.send_flags |= MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS;
1399 	}
1400 
1401 	err = mlx5_ib_post_send_wait(dev, &umrwr);
1402 
1403 	return err;
1404 }
1405 
1406 int mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start,
1407 			  u64 length, u64 virt_addr, int new_access_flags,
1408 			  struct ib_pd *new_pd, struct ib_udata *udata)
1409 {
1410 	struct mlx5_ib_dev *dev = to_mdev(ib_mr->device);
1411 	struct mlx5_ib_mr *mr = to_mmr(ib_mr);
1412 	struct ib_pd *pd = (flags & IB_MR_REREG_PD) ? new_pd : ib_mr->pd;
1413 	int access_flags = flags & IB_MR_REREG_ACCESS ?
1414 			    new_access_flags :
1415 			    mr->access_flags;
1416 	int page_shift = 0;
1417 	int upd_flags = 0;
1418 	int npages = 0;
1419 	int ncont = 0;
1420 	int order = 0;
1421 	u64 addr, len;
1422 	int err;
1423 
1424 	mlx5_ib_dbg(dev, "start 0x%llx, virt_addr 0x%llx, length 0x%llx, access_flags 0x%x\n",
1425 		    start, virt_addr, length, access_flags);
1426 
1427 	atomic_sub(mr->npages, &dev->mdev->priv.reg_pages);
1428 
1429 	if (!mr->umem)
1430 		return -EINVAL;
1431 
1432 	if (flags & IB_MR_REREG_TRANS) {
1433 		addr = virt_addr;
1434 		len = length;
1435 	} else {
1436 		addr = mr->umem->address;
1437 		len = mr->umem->length;
1438 	}
1439 
1440 	if (flags != IB_MR_REREG_PD) {
1441 		/*
1442 		 * Replace umem. This needs to be done whether or not UMR is
1443 		 * used.
1444 		 */
1445 		flags |= IB_MR_REREG_TRANS;
1446 		ib_umem_release(mr->umem);
1447 		mr->umem = NULL;
1448 		err = mr_umem_get(dev, udata, addr, len, access_flags,
1449 				  &mr->umem, &npages, &page_shift, &ncont,
1450 				  &order);
1451 		if (err)
1452 			goto err;
1453 	}
1454 
1455 	if (flags & IB_MR_REREG_TRANS && !use_umr_mtt_update(mr, addr, len)) {
1456 		/*
1457 		 * UMR can't be used - MKey needs to be replaced.
1458 		 */
1459 		if (mr->allocated_from_cache)
1460 			err = unreg_umr(dev, mr);
1461 		else
1462 			err = destroy_mkey(dev, mr);
1463 		if (err)
1464 			goto err;
1465 
1466 		mr = reg_create(ib_mr, pd, addr, len, mr->umem, ncont,
1467 				page_shift, access_flags, true);
1468 
1469 		if (IS_ERR(mr)) {
1470 			err = PTR_ERR(mr);
1471 			mr = to_mmr(ib_mr);
1472 			goto err;
1473 		}
1474 
1475 		mr->allocated_from_cache = 0;
1476 		if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING))
1477 			mr->live = 1;
1478 	} else {
1479 		/*
1480 		 * Send a UMR WQE
1481 		 */
1482 		mr->ibmr.pd = pd;
1483 		mr->access_flags = access_flags;
1484 		mr->mmkey.iova = addr;
1485 		mr->mmkey.size = len;
1486 		mr->mmkey.pd = to_mpd(pd)->pdn;
1487 
1488 		if (flags & IB_MR_REREG_TRANS) {
1489 			upd_flags = MLX5_IB_UPD_XLT_ADDR;
1490 			if (flags & IB_MR_REREG_PD)
1491 				upd_flags |= MLX5_IB_UPD_XLT_PD;
1492 			if (flags & IB_MR_REREG_ACCESS)
1493 				upd_flags |= MLX5_IB_UPD_XLT_ACCESS;
1494 			err = mlx5_ib_update_xlt(mr, 0, npages, page_shift,
1495 						 upd_flags);
1496 		} else {
1497 			err = rereg_umr(pd, mr, access_flags, flags);
1498 		}
1499 
1500 		if (err)
1501 			goto err;
1502 	}
1503 
1504 	set_mr_fields(dev, mr, npages, len, access_flags);
1505 
1506 	update_odp_mr(mr);
1507 	return 0;
1508 
1509 err:
1510 	if (mr->umem) {
1511 		ib_umem_release(mr->umem);
1512 		mr->umem = NULL;
1513 	}
1514 	clean_mr(dev, mr);
1515 	return err;
1516 }
1517 
1518 static int
1519 mlx5_alloc_priv_descs(struct ib_device *device,
1520 		      struct mlx5_ib_mr *mr,
1521 		      int ndescs,
1522 		      int desc_size)
1523 {
1524 	int size = ndescs * desc_size;
1525 	int add_size;
1526 	int ret;
1527 
1528 	add_size = max_t(int, MLX5_UMR_ALIGN - ARCH_KMALLOC_MINALIGN, 0);
1529 
1530 	mr->descs_alloc = kzalloc(size + add_size, GFP_KERNEL);
1531 	if (!mr->descs_alloc)
1532 		return -ENOMEM;
1533 
1534 	mr->descs = PTR_ALIGN(mr->descs_alloc, MLX5_UMR_ALIGN);
1535 
1536 	mr->desc_map = dma_map_single(device->dev.parent, mr->descs,
1537 				      size, DMA_TO_DEVICE);
1538 	if (dma_mapping_error(device->dev.parent, mr->desc_map)) {
1539 		ret = -ENOMEM;
1540 		goto err;
1541 	}
1542 
1543 	return 0;
1544 err:
1545 	kfree(mr->descs_alloc);
1546 
1547 	return ret;
1548 }
1549 
1550 static void
1551 mlx5_free_priv_descs(struct mlx5_ib_mr *mr)
1552 {
1553 	if (mr->descs) {
1554 		struct ib_device *device = mr->ibmr.device;
1555 		int size = mr->max_descs * mr->desc_size;
1556 
1557 		dma_unmap_single(device->dev.parent, mr->desc_map,
1558 				 size, DMA_TO_DEVICE);
1559 		kfree(mr->descs_alloc);
1560 		mr->descs = NULL;
1561 	}
1562 }
1563 
1564 static void clean_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
1565 {
1566 	int allocated_from_cache = mr->allocated_from_cache;
1567 
1568 	if (mr->sig) {
1569 		if (mlx5_core_destroy_psv(dev->mdev,
1570 					  mr->sig->psv_memory.psv_idx))
1571 			mlx5_ib_warn(dev, "failed to destroy mem psv %d\n",
1572 				     mr->sig->psv_memory.psv_idx);
1573 		if (mlx5_core_destroy_psv(dev->mdev,
1574 					  mr->sig->psv_wire.psv_idx))
1575 			mlx5_ib_warn(dev, "failed to destroy wire psv %d\n",
1576 				     mr->sig->psv_wire.psv_idx);
1577 		kfree(mr->sig);
1578 		mr->sig = NULL;
1579 	}
1580 
1581 	mlx5_free_priv_descs(mr);
1582 
1583 	if (!allocated_from_cache)
1584 		destroy_mkey(dev, mr);
1585 }
1586 
1587 static void dereg_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
1588 {
1589 	int npages = mr->npages;
1590 	struct ib_umem *umem = mr->umem;
1591 
1592 	if (is_odp_mr(mr)) {
1593 		struct ib_umem_odp *umem_odp = to_ib_umem_odp(umem);
1594 
1595 		/* Prevent new page faults and
1596 		 * prefetch requests from succeeding
1597 		 */
1598 		mr->live = 0;
1599 
1600 		/* dequeue pending prefetch requests for the mr */
1601 		if (atomic_read(&mr->num_pending_prefetch))
1602 			flush_workqueue(system_unbound_wq);
1603 		WARN_ON(atomic_read(&mr->num_pending_prefetch));
1604 
1605 		/* Wait for all running page-fault handlers to finish. */
1606 		synchronize_srcu(&dev->mr_srcu);
1607 		/* Destroy all page mappings */
1608 		if (umem_odp->page_list)
1609 			mlx5_ib_invalidate_range(umem_odp, ib_umem_start(umem),
1610 						 ib_umem_end(umem));
1611 		else
1612 			mlx5_ib_free_implicit_mr(mr);
1613 		/*
1614 		 * We kill the umem before the MR for ODP,
1615 		 * so that there will not be any invalidations in
1616 		 * flight, looking at the *mr struct.
1617 		 */
1618 		ib_umem_release(umem);
1619 		atomic_sub(npages, &dev->mdev->priv.reg_pages);
1620 
1621 		/* Avoid double-freeing the umem. */
1622 		umem = NULL;
1623 	}
1624 
1625 	clean_mr(dev, mr);
1626 
1627 	/*
1628 	 * We should unregister the DMA address from the HCA before
1629 	 * remove the DMA mapping.
1630 	 */
1631 	mlx5_mr_cache_free(dev, mr);
1632 	if (umem) {
1633 		ib_umem_release(umem);
1634 		atomic_sub(npages, &dev->mdev->priv.reg_pages);
1635 	}
1636 	if (!mr->allocated_from_cache)
1637 		kfree(mr);
1638 }
1639 
1640 int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1641 {
1642 	dereg_mr(to_mdev(ibmr->device), to_mmr(ibmr));
1643 	return 0;
1644 }
1645 
1646 struct ib_mr *mlx5_ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
1647 			       u32 max_num_sg, struct ib_udata *udata)
1648 {
1649 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1650 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1651 	int ndescs = ALIGN(max_num_sg, 4);
1652 	struct mlx5_ib_mr *mr;
1653 	void *mkc;
1654 	u32 *in;
1655 	int err;
1656 
1657 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1658 	if (!mr)
1659 		return ERR_PTR(-ENOMEM);
1660 
1661 	in = kzalloc(inlen, GFP_KERNEL);
1662 	if (!in) {
1663 		err = -ENOMEM;
1664 		goto err_free;
1665 	}
1666 
1667 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1668 	MLX5_SET(mkc, mkc, free, 1);
1669 	MLX5_SET(mkc, mkc, translations_octword_size, ndescs);
1670 	MLX5_SET(mkc, mkc, qpn, 0xffffff);
1671 	MLX5_SET(mkc, mkc, pd, to_mpd(pd)->pdn);
1672 
1673 	if (mr_type == IB_MR_TYPE_MEM_REG) {
1674 		mr->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
1675 		MLX5_SET(mkc, mkc, log_page_size, PAGE_SHIFT);
1676 		err = mlx5_alloc_priv_descs(pd->device, mr,
1677 					    ndescs, sizeof(struct mlx5_mtt));
1678 		if (err)
1679 			goto err_free_in;
1680 
1681 		mr->desc_size = sizeof(struct mlx5_mtt);
1682 		mr->max_descs = ndescs;
1683 	} else if (mr_type == IB_MR_TYPE_SG_GAPS) {
1684 		mr->access_mode = MLX5_MKC_ACCESS_MODE_KLMS;
1685 
1686 		err = mlx5_alloc_priv_descs(pd->device, mr,
1687 					    ndescs, sizeof(struct mlx5_klm));
1688 		if (err)
1689 			goto err_free_in;
1690 		mr->desc_size = sizeof(struct mlx5_klm);
1691 		mr->max_descs = ndescs;
1692 	} else if (mr_type == IB_MR_TYPE_SIGNATURE) {
1693 		u32 psv_index[2];
1694 
1695 		MLX5_SET(mkc, mkc, bsf_en, 1);
1696 		MLX5_SET(mkc, mkc, bsf_octword_size, MLX5_MKEY_BSF_OCTO_SIZE);
1697 		mr->sig = kzalloc(sizeof(*mr->sig), GFP_KERNEL);
1698 		if (!mr->sig) {
1699 			err = -ENOMEM;
1700 			goto err_free_in;
1701 		}
1702 
1703 		/* create mem & wire PSVs */
1704 		err = mlx5_core_create_psv(dev->mdev, to_mpd(pd)->pdn,
1705 					   2, psv_index);
1706 		if (err)
1707 			goto err_free_sig;
1708 
1709 		mr->access_mode = MLX5_MKC_ACCESS_MODE_KLMS;
1710 		mr->sig->psv_memory.psv_idx = psv_index[0];
1711 		mr->sig->psv_wire.psv_idx = psv_index[1];
1712 
1713 		mr->sig->sig_status_checked = true;
1714 		mr->sig->sig_err_exists = false;
1715 		/* Next UMR, Arm SIGERR */
1716 		++mr->sig->sigerr_count;
1717 	} else {
1718 		mlx5_ib_warn(dev, "Invalid mr type %d\n", mr_type);
1719 		err = -EINVAL;
1720 		goto err_free_in;
1721 	}
1722 
1723 	MLX5_SET(mkc, mkc, access_mode_1_0, mr->access_mode & 0x3);
1724 	MLX5_SET(mkc, mkc, access_mode_4_2, (mr->access_mode >> 2) & 0x7);
1725 	MLX5_SET(mkc, mkc, umr_en, 1);
1726 
1727 	mr->ibmr.device = pd->device;
1728 	err = mlx5_core_create_mkey(dev->mdev, &mr->mmkey, in, inlen);
1729 	if (err)
1730 		goto err_destroy_psv;
1731 
1732 	mr->mmkey.type = MLX5_MKEY_MR;
1733 	mr->ibmr.lkey = mr->mmkey.key;
1734 	mr->ibmr.rkey = mr->mmkey.key;
1735 	mr->umem = NULL;
1736 	kfree(in);
1737 
1738 	return &mr->ibmr;
1739 
1740 err_destroy_psv:
1741 	if (mr->sig) {
1742 		if (mlx5_core_destroy_psv(dev->mdev,
1743 					  mr->sig->psv_memory.psv_idx))
1744 			mlx5_ib_warn(dev, "failed to destroy mem psv %d\n",
1745 				     mr->sig->psv_memory.psv_idx);
1746 		if (mlx5_core_destroy_psv(dev->mdev,
1747 					  mr->sig->psv_wire.psv_idx))
1748 			mlx5_ib_warn(dev, "failed to destroy wire psv %d\n",
1749 				     mr->sig->psv_wire.psv_idx);
1750 	}
1751 	mlx5_free_priv_descs(mr);
1752 err_free_sig:
1753 	kfree(mr->sig);
1754 err_free_in:
1755 	kfree(in);
1756 err_free:
1757 	kfree(mr);
1758 	return ERR_PTR(err);
1759 }
1760 
1761 struct ib_mw *mlx5_ib_alloc_mw(struct ib_pd *pd, enum ib_mw_type type,
1762 			       struct ib_udata *udata)
1763 {
1764 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1765 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1766 	struct mlx5_ib_mw *mw = NULL;
1767 	u32 *in = NULL;
1768 	void *mkc;
1769 	int ndescs;
1770 	int err;
1771 	struct mlx5_ib_alloc_mw req = {};
1772 	struct {
1773 		__u32	comp_mask;
1774 		__u32	response_length;
1775 	} resp = {};
1776 
1777 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
1778 	if (err)
1779 		return ERR_PTR(err);
1780 
1781 	if (req.comp_mask || req.reserved1 || req.reserved2)
1782 		return ERR_PTR(-EOPNOTSUPP);
1783 
1784 	if (udata->inlen > sizeof(req) &&
1785 	    !ib_is_udata_cleared(udata, sizeof(req),
1786 				 udata->inlen - sizeof(req)))
1787 		return ERR_PTR(-EOPNOTSUPP);
1788 
1789 	ndescs = req.num_klms ? roundup(req.num_klms, 4) : roundup(1, 4);
1790 
1791 	mw = kzalloc(sizeof(*mw), GFP_KERNEL);
1792 	in = kzalloc(inlen, GFP_KERNEL);
1793 	if (!mw || !in) {
1794 		err = -ENOMEM;
1795 		goto free;
1796 	}
1797 
1798 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1799 
1800 	MLX5_SET(mkc, mkc, free, 1);
1801 	MLX5_SET(mkc, mkc, translations_octword_size, ndescs);
1802 	MLX5_SET(mkc, mkc, pd, to_mpd(pd)->pdn);
1803 	MLX5_SET(mkc, mkc, umr_en, 1);
1804 	MLX5_SET(mkc, mkc, lr, 1);
1805 	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_KLMS);
1806 	MLX5_SET(mkc, mkc, en_rinval, !!((type == IB_MW_TYPE_2)));
1807 	MLX5_SET(mkc, mkc, qpn, 0xffffff);
1808 
1809 	err = mlx5_core_create_mkey(dev->mdev, &mw->mmkey, in, inlen);
1810 	if (err)
1811 		goto free;
1812 
1813 	mw->mmkey.type = MLX5_MKEY_MW;
1814 	mw->ibmw.rkey = mw->mmkey.key;
1815 	mw->ndescs = ndescs;
1816 
1817 	resp.response_length = min(offsetof(typeof(resp), response_length) +
1818 				   sizeof(resp.response_length), udata->outlen);
1819 	if (resp.response_length) {
1820 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
1821 		if (err) {
1822 			mlx5_core_destroy_mkey(dev->mdev, &mw->mmkey);
1823 			goto free;
1824 		}
1825 	}
1826 
1827 	kfree(in);
1828 	return &mw->ibmw;
1829 
1830 free:
1831 	kfree(mw);
1832 	kfree(in);
1833 	return ERR_PTR(err);
1834 }
1835 
1836 int mlx5_ib_dealloc_mw(struct ib_mw *mw)
1837 {
1838 	struct mlx5_ib_mw *mmw = to_mmw(mw);
1839 	int err;
1840 
1841 	err =  mlx5_core_destroy_mkey((to_mdev(mw->device))->mdev,
1842 				      &mmw->mmkey);
1843 	if (!err)
1844 		kfree(mmw);
1845 	return err;
1846 }
1847 
1848 int mlx5_ib_check_mr_status(struct ib_mr *ibmr, u32 check_mask,
1849 			    struct ib_mr_status *mr_status)
1850 {
1851 	struct mlx5_ib_mr *mmr = to_mmr(ibmr);
1852 	int ret = 0;
1853 
1854 	if (check_mask & ~IB_MR_CHECK_SIG_STATUS) {
1855 		pr_err("Invalid status check mask\n");
1856 		ret = -EINVAL;
1857 		goto done;
1858 	}
1859 
1860 	mr_status->fail_status = 0;
1861 	if (check_mask & IB_MR_CHECK_SIG_STATUS) {
1862 		if (!mmr->sig) {
1863 			ret = -EINVAL;
1864 			pr_err("signature status check requested on a non-signature enabled MR\n");
1865 			goto done;
1866 		}
1867 
1868 		mmr->sig->sig_status_checked = true;
1869 		if (!mmr->sig->sig_err_exists)
1870 			goto done;
1871 
1872 		if (ibmr->lkey == mmr->sig->err_item.key)
1873 			memcpy(&mr_status->sig_err, &mmr->sig->err_item,
1874 			       sizeof(mr_status->sig_err));
1875 		else {
1876 			mr_status->sig_err.err_type = IB_SIG_BAD_GUARD;
1877 			mr_status->sig_err.sig_err_offset = 0;
1878 			mr_status->sig_err.key = mmr->sig->err_item.key;
1879 		}
1880 
1881 		mmr->sig->sig_err_exists = false;
1882 		mr_status->fail_status |= IB_MR_CHECK_SIG_STATUS;
1883 	}
1884 
1885 done:
1886 	return ret;
1887 }
1888 
1889 static int
1890 mlx5_ib_sg_to_klms(struct mlx5_ib_mr *mr,
1891 		   struct scatterlist *sgl,
1892 		   unsigned short sg_nents,
1893 		   unsigned int *sg_offset_p)
1894 {
1895 	struct scatterlist *sg = sgl;
1896 	struct mlx5_klm *klms = mr->descs;
1897 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
1898 	u32 lkey = mr->ibmr.pd->local_dma_lkey;
1899 	int i;
1900 
1901 	mr->ibmr.iova = sg_dma_address(sg) + sg_offset;
1902 	mr->ibmr.length = 0;
1903 
1904 	for_each_sg(sgl, sg, sg_nents, i) {
1905 		if (unlikely(i >= mr->max_descs))
1906 			break;
1907 		klms[i].va = cpu_to_be64(sg_dma_address(sg) + sg_offset);
1908 		klms[i].bcount = cpu_to_be32(sg_dma_len(sg) - sg_offset);
1909 		klms[i].key = cpu_to_be32(lkey);
1910 		mr->ibmr.length += sg_dma_len(sg) - sg_offset;
1911 
1912 		sg_offset = 0;
1913 	}
1914 	mr->ndescs = i;
1915 
1916 	if (sg_offset_p)
1917 		*sg_offset_p = sg_offset;
1918 
1919 	return i;
1920 }
1921 
1922 static int mlx5_set_page(struct ib_mr *ibmr, u64 addr)
1923 {
1924 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
1925 	__be64 *descs;
1926 
1927 	if (unlikely(mr->ndescs == mr->max_descs))
1928 		return -ENOMEM;
1929 
1930 	descs = mr->descs;
1931 	descs[mr->ndescs++] = cpu_to_be64(addr | MLX5_EN_RD | MLX5_EN_WR);
1932 
1933 	return 0;
1934 }
1935 
1936 int mlx5_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
1937 		      unsigned int *sg_offset)
1938 {
1939 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
1940 	int n;
1941 
1942 	mr->ndescs = 0;
1943 
1944 	ib_dma_sync_single_for_cpu(ibmr->device, mr->desc_map,
1945 				   mr->desc_size * mr->max_descs,
1946 				   DMA_TO_DEVICE);
1947 
1948 	if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
1949 		n = mlx5_ib_sg_to_klms(mr, sg, sg_nents, sg_offset);
1950 	else
1951 		n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset,
1952 				mlx5_set_page);
1953 
1954 	ib_dma_sync_single_for_device(ibmr->device, mr->desc_map,
1955 				      mr->desc_size * mr->max_descs,
1956 				      DMA_TO_DEVICE);
1957 
1958 	return n;
1959 }
1960