xref: /openbmc/linux/drivers/infiniband/hw/mlx5/mr.c (revision 2cf1c348)
1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  * Copyright (c) 2020, Intel Corporation. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 
35 #include <linux/kref.h>
36 #include <linux/random.h>
37 #include <linux/debugfs.h>
38 #include <linux/export.h>
39 #include <linux/delay.h>
40 #include <linux/dma-buf.h>
41 #include <linux/dma-resv.h>
42 #include <rdma/ib_umem.h>
43 #include <rdma/ib_umem_odp.h>
44 #include <rdma/ib_verbs.h>
45 #include "dm.h"
46 #include "mlx5_ib.h"
47 
48 /*
49  * We can't use an array for xlt_emergency_page because dma_map_single doesn't
50  * work on kernel modules memory
51  */
52 void *xlt_emergency_page;
53 static DEFINE_MUTEX(xlt_emergency_page_mutex);
54 
55 enum {
56 	MAX_PENDING_REG_MR = 8,
57 };
58 
59 #define MLX5_UMR_ALIGN 2048
60 
61 static void
62 create_mkey_callback(int status, struct mlx5_async_work *context);
63 static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem,
64 				     u64 iova, int access_flags,
65 				     unsigned int page_size, bool populate);
66 
67 static void set_mkc_access_pd_addr_fields(void *mkc, int acc, u64 start_addr,
68 					  struct ib_pd *pd)
69 {
70 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
71 	bool ro_pci_enabled = pcie_relaxed_ordering_enabled(dev->mdev->pdev);
72 
73 	MLX5_SET(mkc, mkc, a, !!(acc & IB_ACCESS_REMOTE_ATOMIC));
74 	MLX5_SET(mkc, mkc, rw, !!(acc & IB_ACCESS_REMOTE_WRITE));
75 	MLX5_SET(mkc, mkc, rr, !!(acc & IB_ACCESS_REMOTE_READ));
76 	MLX5_SET(mkc, mkc, lw, !!(acc & IB_ACCESS_LOCAL_WRITE));
77 	MLX5_SET(mkc, mkc, lr, 1);
78 
79 	if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write))
80 		MLX5_SET(mkc, mkc, relaxed_ordering_write,
81 			 (acc & IB_ACCESS_RELAXED_ORDERING) && ro_pci_enabled);
82 	if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read))
83 		MLX5_SET(mkc, mkc, relaxed_ordering_read,
84 			 (acc & IB_ACCESS_RELAXED_ORDERING) && ro_pci_enabled);
85 
86 	MLX5_SET(mkc, mkc, pd, to_mpd(pd)->pdn);
87 	MLX5_SET(mkc, mkc, qpn, 0xffffff);
88 	MLX5_SET64(mkc, mkc, start_addr, start_addr);
89 }
90 
91 static void assign_mkey_variant(struct mlx5_ib_dev *dev,
92 				struct mlx5_ib_mkey *mkey, u32 *in)
93 {
94 	u8 key = atomic_inc_return(&dev->mkey_var);
95 	void *mkc;
96 
97 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
98 	MLX5_SET(mkc, mkc, mkey_7_0, key);
99 	mkey->key = key;
100 }
101 
102 static int mlx5_ib_create_mkey(struct mlx5_ib_dev *dev,
103 			       struct mlx5_ib_mkey *mkey, u32 *in, int inlen)
104 {
105 	int ret;
106 
107 	assign_mkey_variant(dev, mkey, in);
108 	ret = mlx5_core_create_mkey(dev->mdev, &mkey->key, in, inlen);
109 	if (!ret)
110 		init_waitqueue_head(&mkey->wait);
111 
112 	return ret;
113 }
114 
115 static int
116 mlx5_ib_create_mkey_cb(struct mlx5_ib_dev *dev,
117 		       struct mlx5_ib_mkey *mkey,
118 		       struct mlx5_async_ctx *async_ctx,
119 		       u32 *in, int inlen, u32 *out, int outlen,
120 		       struct mlx5_async_work *context)
121 {
122 	MLX5_SET(create_mkey_in, in, opcode, MLX5_CMD_OP_CREATE_MKEY);
123 	assign_mkey_variant(dev, mkey, in);
124 	return mlx5_cmd_exec_cb(async_ctx, in, inlen, out, outlen,
125 				create_mkey_callback, context);
126 }
127 
128 static int mr_cache_max_order(struct mlx5_ib_dev *dev);
129 static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent);
130 
131 static bool umr_can_use_indirect_mkey(struct mlx5_ib_dev *dev)
132 {
133 	return !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled);
134 }
135 
136 static int destroy_mkey(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
137 {
138 	WARN_ON(xa_load(&dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key)));
139 
140 	return mlx5_core_destroy_mkey(dev->mdev, mr->mmkey.key);
141 }
142 
143 static void create_mkey_callback(int status, struct mlx5_async_work *context)
144 {
145 	struct mlx5_ib_mr *mr =
146 		container_of(context, struct mlx5_ib_mr, cb_work);
147 	struct mlx5_cache_ent *ent = mr->cache_ent;
148 	struct mlx5_ib_dev *dev = ent->dev;
149 	unsigned long flags;
150 
151 	if (status) {
152 		mlx5_ib_warn(dev, "async reg mr failed. status %d\n", status);
153 		kfree(mr);
154 		spin_lock_irqsave(&ent->lock, flags);
155 		ent->pending--;
156 		WRITE_ONCE(dev->fill_delay, 1);
157 		spin_unlock_irqrestore(&ent->lock, flags);
158 		mod_timer(&dev->delay_timer, jiffies + HZ);
159 		return;
160 	}
161 
162 	mr->mmkey.type = MLX5_MKEY_MR;
163 	mr->mmkey.key |= mlx5_idx_to_mkey(
164 		MLX5_GET(create_mkey_out, mr->out, mkey_index));
165 	init_waitqueue_head(&mr->mmkey.wait);
166 
167 	WRITE_ONCE(dev->cache.last_add, jiffies);
168 
169 	spin_lock_irqsave(&ent->lock, flags);
170 	list_add_tail(&mr->list, &ent->head);
171 	ent->available_mrs++;
172 	ent->total_mrs++;
173 	/* If we are doing fill_to_high_water then keep going. */
174 	queue_adjust_cache_locked(ent);
175 	ent->pending--;
176 	spin_unlock_irqrestore(&ent->lock, flags);
177 }
178 
179 static struct mlx5_ib_mr *alloc_cache_mr(struct mlx5_cache_ent *ent, void *mkc)
180 {
181 	struct mlx5_ib_mr *mr;
182 
183 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
184 	if (!mr)
185 		return NULL;
186 	mr->cache_ent = ent;
187 
188 	set_mkc_access_pd_addr_fields(mkc, 0, 0, ent->dev->umrc.pd);
189 	MLX5_SET(mkc, mkc, free, 1);
190 	MLX5_SET(mkc, mkc, umr_en, 1);
191 	MLX5_SET(mkc, mkc, access_mode_1_0, ent->access_mode & 0x3);
192 	MLX5_SET(mkc, mkc, access_mode_4_2, (ent->access_mode >> 2) & 0x7);
193 
194 	MLX5_SET(mkc, mkc, translations_octword_size, ent->xlt);
195 	MLX5_SET(mkc, mkc, log_page_size, ent->page);
196 	return mr;
197 }
198 
199 /* Asynchronously schedule new MRs to be populated in the cache. */
200 static int add_keys(struct mlx5_cache_ent *ent, unsigned int num)
201 {
202 	size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
203 	struct mlx5_ib_mr *mr;
204 	void *mkc;
205 	u32 *in;
206 	int err = 0;
207 	int i;
208 
209 	in = kzalloc(inlen, GFP_KERNEL);
210 	if (!in)
211 		return -ENOMEM;
212 
213 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
214 	for (i = 0; i < num; i++) {
215 		mr = alloc_cache_mr(ent, mkc);
216 		if (!mr) {
217 			err = -ENOMEM;
218 			break;
219 		}
220 		spin_lock_irq(&ent->lock);
221 		if (ent->pending >= MAX_PENDING_REG_MR) {
222 			err = -EAGAIN;
223 			spin_unlock_irq(&ent->lock);
224 			kfree(mr);
225 			break;
226 		}
227 		ent->pending++;
228 		spin_unlock_irq(&ent->lock);
229 		err = mlx5_ib_create_mkey_cb(ent->dev, &mr->mmkey,
230 					     &ent->dev->async_ctx, in, inlen,
231 					     mr->out, sizeof(mr->out),
232 					     &mr->cb_work);
233 		if (err) {
234 			spin_lock_irq(&ent->lock);
235 			ent->pending--;
236 			spin_unlock_irq(&ent->lock);
237 			mlx5_ib_warn(ent->dev, "create mkey failed %d\n", err);
238 			kfree(mr);
239 			break;
240 		}
241 	}
242 
243 	kfree(in);
244 	return err;
245 }
246 
247 /* Synchronously create a MR in the cache */
248 static struct mlx5_ib_mr *create_cache_mr(struct mlx5_cache_ent *ent)
249 {
250 	size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
251 	struct mlx5_ib_mr *mr;
252 	void *mkc;
253 	u32 *in;
254 	int err;
255 
256 	in = kzalloc(inlen, GFP_KERNEL);
257 	if (!in)
258 		return ERR_PTR(-ENOMEM);
259 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
260 
261 	mr = alloc_cache_mr(ent, mkc);
262 	if (!mr) {
263 		err = -ENOMEM;
264 		goto free_in;
265 	}
266 
267 	err = mlx5_core_create_mkey(ent->dev->mdev, &mr->mmkey.key, in, inlen);
268 	if (err)
269 		goto free_mr;
270 
271 	init_waitqueue_head(&mr->mmkey.wait);
272 	mr->mmkey.type = MLX5_MKEY_MR;
273 	WRITE_ONCE(ent->dev->cache.last_add, jiffies);
274 	spin_lock_irq(&ent->lock);
275 	ent->total_mrs++;
276 	spin_unlock_irq(&ent->lock);
277 	kfree(in);
278 	return mr;
279 free_mr:
280 	kfree(mr);
281 free_in:
282 	kfree(in);
283 	return ERR_PTR(err);
284 }
285 
286 static void remove_cache_mr_locked(struct mlx5_cache_ent *ent)
287 {
288 	struct mlx5_ib_mr *mr;
289 
290 	lockdep_assert_held(&ent->lock);
291 	if (list_empty(&ent->head))
292 		return;
293 	mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
294 	list_del(&mr->list);
295 	ent->available_mrs--;
296 	ent->total_mrs--;
297 	spin_unlock_irq(&ent->lock);
298 	mlx5_core_destroy_mkey(ent->dev->mdev, mr->mmkey.key);
299 	kfree(mr);
300 	spin_lock_irq(&ent->lock);
301 }
302 
303 static int resize_available_mrs(struct mlx5_cache_ent *ent, unsigned int target,
304 				bool limit_fill)
305 {
306 	int err;
307 
308 	lockdep_assert_held(&ent->lock);
309 
310 	while (true) {
311 		if (limit_fill)
312 			target = ent->limit * 2;
313 		if (target == ent->available_mrs + ent->pending)
314 			return 0;
315 		if (target > ent->available_mrs + ent->pending) {
316 			u32 todo = target - (ent->available_mrs + ent->pending);
317 
318 			spin_unlock_irq(&ent->lock);
319 			err = add_keys(ent, todo);
320 			if (err == -EAGAIN)
321 				usleep_range(3000, 5000);
322 			spin_lock_irq(&ent->lock);
323 			if (err) {
324 				if (err != -EAGAIN)
325 					return err;
326 			} else
327 				return 0;
328 		} else {
329 			remove_cache_mr_locked(ent);
330 		}
331 	}
332 }
333 
334 static ssize_t size_write(struct file *filp, const char __user *buf,
335 			  size_t count, loff_t *pos)
336 {
337 	struct mlx5_cache_ent *ent = filp->private_data;
338 	u32 target;
339 	int err;
340 
341 	err = kstrtou32_from_user(buf, count, 0, &target);
342 	if (err)
343 		return err;
344 
345 	/*
346 	 * Target is the new value of total_mrs the user requests, however we
347 	 * cannot free MRs that are in use. Compute the target value for
348 	 * available_mrs.
349 	 */
350 	spin_lock_irq(&ent->lock);
351 	if (target < ent->total_mrs - ent->available_mrs) {
352 		err = -EINVAL;
353 		goto err_unlock;
354 	}
355 	target = target - (ent->total_mrs - ent->available_mrs);
356 	if (target < ent->limit || target > ent->limit*2) {
357 		err = -EINVAL;
358 		goto err_unlock;
359 	}
360 	err = resize_available_mrs(ent, target, false);
361 	if (err)
362 		goto err_unlock;
363 	spin_unlock_irq(&ent->lock);
364 
365 	return count;
366 
367 err_unlock:
368 	spin_unlock_irq(&ent->lock);
369 	return err;
370 }
371 
372 static ssize_t size_read(struct file *filp, char __user *buf, size_t count,
373 			 loff_t *pos)
374 {
375 	struct mlx5_cache_ent *ent = filp->private_data;
376 	char lbuf[20];
377 	int err;
378 
379 	err = snprintf(lbuf, sizeof(lbuf), "%d\n", ent->total_mrs);
380 	if (err < 0)
381 		return err;
382 
383 	return simple_read_from_buffer(buf, count, pos, lbuf, err);
384 }
385 
386 static const struct file_operations size_fops = {
387 	.owner	= THIS_MODULE,
388 	.open	= simple_open,
389 	.write	= size_write,
390 	.read	= size_read,
391 };
392 
393 static ssize_t limit_write(struct file *filp, const char __user *buf,
394 			   size_t count, loff_t *pos)
395 {
396 	struct mlx5_cache_ent *ent = filp->private_data;
397 	u32 var;
398 	int err;
399 
400 	err = kstrtou32_from_user(buf, count, 0, &var);
401 	if (err)
402 		return err;
403 
404 	/*
405 	 * Upon set we immediately fill the cache to high water mark implied by
406 	 * the limit.
407 	 */
408 	spin_lock_irq(&ent->lock);
409 	ent->limit = var;
410 	err = resize_available_mrs(ent, 0, true);
411 	spin_unlock_irq(&ent->lock);
412 	if (err)
413 		return err;
414 	return count;
415 }
416 
417 static ssize_t limit_read(struct file *filp, char __user *buf, size_t count,
418 			  loff_t *pos)
419 {
420 	struct mlx5_cache_ent *ent = filp->private_data;
421 	char lbuf[20];
422 	int err;
423 
424 	err = snprintf(lbuf, sizeof(lbuf), "%d\n", ent->limit);
425 	if (err < 0)
426 		return err;
427 
428 	return simple_read_from_buffer(buf, count, pos, lbuf, err);
429 }
430 
431 static const struct file_operations limit_fops = {
432 	.owner	= THIS_MODULE,
433 	.open	= simple_open,
434 	.write	= limit_write,
435 	.read	= limit_read,
436 };
437 
438 static bool someone_adding(struct mlx5_mr_cache *cache)
439 {
440 	unsigned int i;
441 
442 	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
443 		struct mlx5_cache_ent *ent = &cache->ent[i];
444 		bool ret;
445 
446 		spin_lock_irq(&ent->lock);
447 		ret = ent->available_mrs < ent->limit;
448 		spin_unlock_irq(&ent->lock);
449 		if (ret)
450 			return true;
451 	}
452 	return false;
453 }
454 
455 /*
456  * Check if the bucket is outside the high/low water mark and schedule an async
457  * update. The cache refill has hysteresis, once the low water mark is hit it is
458  * refilled up to the high mark.
459  */
460 static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent)
461 {
462 	lockdep_assert_held(&ent->lock);
463 
464 	if (ent->disabled || READ_ONCE(ent->dev->fill_delay))
465 		return;
466 	if (ent->available_mrs < ent->limit) {
467 		ent->fill_to_high_water = true;
468 		queue_work(ent->dev->cache.wq, &ent->work);
469 	} else if (ent->fill_to_high_water &&
470 		   ent->available_mrs + ent->pending < 2 * ent->limit) {
471 		/*
472 		 * Once we start populating due to hitting a low water mark
473 		 * continue until we pass the high water mark.
474 		 */
475 		queue_work(ent->dev->cache.wq, &ent->work);
476 	} else if (ent->available_mrs == 2 * ent->limit) {
477 		ent->fill_to_high_water = false;
478 	} else if (ent->available_mrs > 2 * ent->limit) {
479 		/* Queue deletion of excess entries */
480 		ent->fill_to_high_water = false;
481 		if (ent->pending)
482 			queue_delayed_work(ent->dev->cache.wq, &ent->dwork,
483 					   msecs_to_jiffies(1000));
484 		else
485 			queue_work(ent->dev->cache.wq, &ent->work);
486 	}
487 }
488 
489 static void __cache_work_func(struct mlx5_cache_ent *ent)
490 {
491 	struct mlx5_ib_dev *dev = ent->dev;
492 	struct mlx5_mr_cache *cache = &dev->cache;
493 	int err;
494 
495 	spin_lock_irq(&ent->lock);
496 	if (ent->disabled)
497 		goto out;
498 
499 	if (ent->fill_to_high_water &&
500 	    ent->available_mrs + ent->pending < 2 * ent->limit &&
501 	    !READ_ONCE(dev->fill_delay)) {
502 		spin_unlock_irq(&ent->lock);
503 		err = add_keys(ent, 1);
504 		spin_lock_irq(&ent->lock);
505 		if (ent->disabled)
506 			goto out;
507 		if (err) {
508 			/*
509 			 * EAGAIN only happens if pending is positive, so we
510 			 * will be rescheduled from reg_mr_callback(). The only
511 			 * failure path here is ENOMEM.
512 			 */
513 			if (err != -EAGAIN) {
514 				mlx5_ib_warn(
515 					dev,
516 					"command failed order %d, err %d\n",
517 					ent->order, err);
518 				queue_delayed_work(cache->wq, &ent->dwork,
519 						   msecs_to_jiffies(1000));
520 			}
521 		}
522 	} else if (ent->available_mrs > 2 * ent->limit) {
523 		bool need_delay;
524 
525 		/*
526 		 * The remove_cache_mr() logic is performed as garbage
527 		 * collection task. Such task is intended to be run when no
528 		 * other active processes are running.
529 		 *
530 		 * The need_resched() will return TRUE if there are user tasks
531 		 * to be activated in near future.
532 		 *
533 		 * In such case, we don't execute remove_cache_mr() and postpone
534 		 * the garbage collection work to try to run in next cycle, in
535 		 * order to free CPU resources to other tasks.
536 		 */
537 		spin_unlock_irq(&ent->lock);
538 		need_delay = need_resched() || someone_adding(cache) ||
539 			     !time_after(jiffies,
540 					 READ_ONCE(cache->last_add) + 300 * HZ);
541 		spin_lock_irq(&ent->lock);
542 		if (ent->disabled)
543 			goto out;
544 		if (need_delay)
545 			queue_delayed_work(cache->wq, &ent->dwork, 300 * HZ);
546 		remove_cache_mr_locked(ent);
547 		queue_adjust_cache_locked(ent);
548 	}
549 out:
550 	spin_unlock_irq(&ent->lock);
551 }
552 
553 static void delayed_cache_work_func(struct work_struct *work)
554 {
555 	struct mlx5_cache_ent *ent;
556 
557 	ent = container_of(work, struct mlx5_cache_ent, dwork.work);
558 	__cache_work_func(ent);
559 }
560 
561 static void cache_work_func(struct work_struct *work)
562 {
563 	struct mlx5_cache_ent *ent;
564 
565 	ent = container_of(work, struct mlx5_cache_ent, work);
566 	__cache_work_func(ent);
567 }
568 
569 /* Allocate a special entry from the cache */
570 struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
571 				       unsigned int entry, int access_flags)
572 {
573 	struct mlx5_mr_cache *cache = &dev->cache;
574 	struct mlx5_cache_ent *ent;
575 	struct mlx5_ib_mr *mr;
576 
577 	if (WARN_ON(entry <= MR_CACHE_LAST_STD_ENTRY ||
578 		    entry >= ARRAY_SIZE(cache->ent)))
579 		return ERR_PTR(-EINVAL);
580 
581 	/* Matches access in alloc_cache_mr() */
582 	if (!mlx5_ib_can_reconfig_with_umr(dev, 0, access_flags))
583 		return ERR_PTR(-EOPNOTSUPP);
584 
585 	ent = &cache->ent[entry];
586 	spin_lock_irq(&ent->lock);
587 	if (list_empty(&ent->head)) {
588 		spin_unlock_irq(&ent->lock);
589 		mr = create_cache_mr(ent);
590 		if (IS_ERR(mr))
591 			return mr;
592 	} else {
593 		mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
594 		list_del(&mr->list);
595 		ent->available_mrs--;
596 		queue_adjust_cache_locked(ent);
597 		spin_unlock_irq(&ent->lock);
598 
599 		mlx5_clear_mr(mr);
600 	}
601 	mr->access_flags = access_flags;
602 	return mr;
603 }
604 
605 /* Return a MR already available in the cache */
606 static struct mlx5_ib_mr *get_cache_mr(struct mlx5_cache_ent *req_ent)
607 {
608 	struct mlx5_ib_mr *mr = NULL;
609 	struct mlx5_cache_ent *ent = req_ent;
610 
611 	spin_lock_irq(&ent->lock);
612 	if (!list_empty(&ent->head)) {
613 		mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
614 		list_del(&mr->list);
615 		ent->available_mrs--;
616 		queue_adjust_cache_locked(ent);
617 		spin_unlock_irq(&ent->lock);
618 		mlx5_clear_mr(mr);
619 		return mr;
620 	}
621 	queue_adjust_cache_locked(ent);
622 	spin_unlock_irq(&ent->lock);
623 	req_ent->miss++;
624 	return NULL;
625 }
626 
627 static void mlx5_mr_cache_free(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
628 {
629 	struct mlx5_cache_ent *ent = mr->cache_ent;
630 
631 	spin_lock_irq(&ent->lock);
632 	list_add_tail(&mr->list, &ent->head);
633 	ent->available_mrs++;
634 	queue_adjust_cache_locked(ent);
635 	spin_unlock_irq(&ent->lock);
636 }
637 
638 static void clean_keys(struct mlx5_ib_dev *dev, int c)
639 {
640 	struct mlx5_mr_cache *cache = &dev->cache;
641 	struct mlx5_cache_ent *ent = &cache->ent[c];
642 	struct mlx5_ib_mr *tmp_mr;
643 	struct mlx5_ib_mr *mr;
644 	LIST_HEAD(del_list);
645 
646 	cancel_delayed_work(&ent->dwork);
647 	while (1) {
648 		spin_lock_irq(&ent->lock);
649 		if (list_empty(&ent->head)) {
650 			spin_unlock_irq(&ent->lock);
651 			break;
652 		}
653 		mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
654 		list_move(&mr->list, &del_list);
655 		ent->available_mrs--;
656 		ent->total_mrs--;
657 		spin_unlock_irq(&ent->lock);
658 		mlx5_core_destroy_mkey(dev->mdev, mr->mmkey.key);
659 	}
660 
661 	list_for_each_entry_safe(mr, tmp_mr, &del_list, list) {
662 		list_del(&mr->list);
663 		kfree(mr);
664 	}
665 }
666 
667 static void mlx5_mr_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
668 {
669 	if (!mlx5_debugfs_root || dev->is_rep)
670 		return;
671 
672 	debugfs_remove_recursive(dev->cache.root);
673 	dev->cache.root = NULL;
674 }
675 
676 static void mlx5_mr_cache_debugfs_init(struct mlx5_ib_dev *dev)
677 {
678 	struct mlx5_mr_cache *cache = &dev->cache;
679 	struct mlx5_cache_ent *ent;
680 	struct dentry *dir;
681 	int i;
682 
683 	if (!mlx5_debugfs_root || dev->is_rep)
684 		return;
685 
686 	cache->root = debugfs_create_dir("mr_cache", dev->mdev->priv.dbg_root);
687 
688 	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
689 		ent = &cache->ent[i];
690 		sprintf(ent->name, "%d", ent->order);
691 		dir = debugfs_create_dir(ent->name, cache->root);
692 		debugfs_create_file("size", 0600, dir, ent, &size_fops);
693 		debugfs_create_file("limit", 0600, dir, ent, &limit_fops);
694 		debugfs_create_u32("cur", 0400, dir, &ent->available_mrs);
695 		debugfs_create_u32("miss", 0600, dir, &ent->miss);
696 	}
697 }
698 
699 static void delay_time_func(struct timer_list *t)
700 {
701 	struct mlx5_ib_dev *dev = from_timer(dev, t, delay_timer);
702 
703 	WRITE_ONCE(dev->fill_delay, 0);
704 }
705 
706 int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
707 {
708 	struct mlx5_mr_cache *cache = &dev->cache;
709 	struct mlx5_cache_ent *ent;
710 	int i;
711 
712 	mutex_init(&dev->slow_path_mutex);
713 	cache->wq = alloc_ordered_workqueue("mkey_cache", WQ_MEM_RECLAIM);
714 	if (!cache->wq) {
715 		mlx5_ib_warn(dev, "failed to create work queue\n");
716 		return -ENOMEM;
717 	}
718 
719 	mlx5_cmd_init_async_ctx(dev->mdev, &dev->async_ctx);
720 	timer_setup(&dev->delay_timer, delay_time_func, 0);
721 	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
722 		ent = &cache->ent[i];
723 		INIT_LIST_HEAD(&ent->head);
724 		spin_lock_init(&ent->lock);
725 		ent->order = i + 2;
726 		ent->dev = dev;
727 		ent->limit = 0;
728 
729 		INIT_WORK(&ent->work, cache_work_func);
730 		INIT_DELAYED_WORK(&ent->dwork, delayed_cache_work_func);
731 
732 		if (i > MR_CACHE_LAST_STD_ENTRY) {
733 			mlx5_odp_init_mr_cache_entry(ent);
734 			continue;
735 		}
736 
737 		if (ent->order > mr_cache_max_order(dev))
738 			continue;
739 
740 		ent->page = PAGE_SHIFT;
741 		ent->xlt = (1 << ent->order) * sizeof(struct mlx5_mtt) /
742 			   MLX5_IB_UMR_OCTOWORD;
743 		ent->access_mode = MLX5_MKC_ACCESS_MODE_MTT;
744 		if ((dev->mdev->profile.mask & MLX5_PROF_MASK_MR_CACHE) &&
745 		    !dev->is_rep && mlx5_core_is_pf(dev->mdev) &&
746 		    mlx5_ib_can_load_pas_with_umr(dev, 0))
747 			ent->limit = dev->mdev->profile.mr_cache[i].limit;
748 		else
749 			ent->limit = 0;
750 		spin_lock_irq(&ent->lock);
751 		queue_adjust_cache_locked(ent);
752 		spin_unlock_irq(&ent->lock);
753 	}
754 
755 	mlx5_mr_cache_debugfs_init(dev);
756 
757 	return 0;
758 }
759 
760 int mlx5_mr_cache_cleanup(struct mlx5_ib_dev *dev)
761 {
762 	unsigned int i;
763 
764 	if (!dev->cache.wq)
765 		return 0;
766 
767 	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
768 		struct mlx5_cache_ent *ent = &dev->cache.ent[i];
769 
770 		spin_lock_irq(&ent->lock);
771 		ent->disabled = true;
772 		spin_unlock_irq(&ent->lock);
773 		cancel_work_sync(&ent->work);
774 		cancel_delayed_work_sync(&ent->dwork);
775 	}
776 
777 	mlx5_mr_cache_debugfs_cleanup(dev);
778 	mlx5_cmd_cleanup_async_ctx(&dev->async_ctx);
779 
780 	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++)
781 		clean_keys(dev, i);
782 
783 	destroy_workqueue(dev->cache.wq);
784 	del_timer_sync(&dev->delay_timer);
785 
786 	return 0;
787 }
788 
789 struct ib_mr *mlx5_ib_get_dma_mr(struct ib_pd *pd, int acc)
790 {
791 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
792 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
793 	struct mlx5_ib_mr *mr;
794 	void *mkc;
795 	u32 *in;
796 	int err;
797 
798 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
799 	if (!mr)
800 		return ERR_PTR(-ENOMEM);
801 
802 	in = kzalloc(inlen, GFP_KERNEL);
803 	if (!in) {
804 		err = -ENOMEM;
805 		goto err_free;
806 	}
807 
808 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
809 
810 	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_PA);
811 	MLX5_SET(mkc, mkc, length64, 1);
812 	set_mkc_access_pd_addr_fields(mkc, acc | IB_ACCESS_RELAXED_ORDERING, 0,
813 				      pd);
814 
815 	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
816 	if (err)
817 		goto err_in;
818 
819 	kfree(in);
820 	mr->mmkey.type = MLX5_MKEY_MR;
821 	mr->ibmr.lkey = mr->mmkey.key;
822 	mr->ibmr.rkey = mr->mmkey.key;
823 	mr->umem = NULL;
824 
825 	return &mr->ibmr;
826 
827 err_in:
828 	kfree(in);
829 
830 err_free:
831 	kfree(mr);
832 
833 	return ERR_PTR(err);
834 }
835 
836 static int get_octo_len(u64 addr, u64 len, int page_shift)
837 {
838 	u64 page_size = 1ULL << page_shift;
839 	u64 offset;
840 	int npages;
841 
842 	offset = addr & (page_size - 1);
843 	npages = ALIGN(len + offset, page_size) >> page_shift;
844 	return (npages + 1) / 2;
845 }
846 
847 static int mr_cache_max_order(struct mlx5_ib_dev *dev)
848 {
849 	if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
850 		return MR_CACHE_LAST_STD_ENTRY + 2;
851 	return MLX5_MAX_UMR_SHIFT;
852 }
853 
854 static void mlx5_ib_umr_done(struct ib_cq *cq, struct ib_wc *wc)
855 {
856 	struct mlx5_ib_umr_context *context =
857 		container_of(wc->wr_cqe, struct mlx5_ib_umr_context, cqe);
858 
859 	context->status = wc->status;
860 	complete(&context->done);
861 }
862 
863 static inline void mlx5_ib_init_umr_context(struct mlx5_ib_umr_context *context)
864 {
865 	context->cqe.done = mlx5_ib_umr_done;
866 	context->status = -1;
867 	init_completion(&context->done);
868 }
869 
870 static int mlx5_ib_post_send_wait(struct mlx5_ib_dev *dev,
871 				  struct mlx5_umr_wr *umrwr)
872 {
873 	struct umr_common *umrc = &dev->umrc;
874 	const struct ib_send_wr *bad;
875 	int err;
876 	struct mlx5_ib_umr_context umr_context;
877 
878 	mlx5_ib_init_umr_context(&umr_context);
879 	umrwr->wr.wr_cqe = &umr_context.cqe;
880 
881 	down(&umrc->sem);
882 	err = ib_post_send(umrc->qp, &umrwr->wr, &bad);
883 	if (err) {
884 		mlx5_ib_warn(dev, "UMR post send failed, err %d\n", err);
885 	} else {
886 		wait_for_completion(&umr_context.done);
887 		if (umr_context.status != IB_WC_SUCCESS) {
888 			mlx5_ib_warn(dev, "reg umr failed (%u)\n",
889 				     umr_context.status);
890 			err = -EFAULT;
891 		}
892 	}
893 	up(&umrc->sem);
894 	return err;
895 }
896 
897 static struct mlx5_cache_ent *mr_cache_ent_from_order(struct mlx5_ib_dev *dev,
898 						      unsigned int order)
899 {
900 	struct mlx5_mr_cache *cache = &dev->cache;
901 
902 	if (order < cache->ent[0].order)
903 		return &cache->ent[0];
904 	order = order - cache->ent[0].order;
905 	if (order > MR_CACHE_LAST_STD_ENTRY)
906 		return NULL;
907 	return &cache->ent[order];
908 }
909 
910 static void set_mr_fields(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr,
911 			  u64 length, int access_flags, u64 iova)
912 {
913 	mr->ibmr.lkey = mr->mmkey.key;
914 	mr->ibmr.rkey = mr->mmkey.key;
915 	mr->ibmr.length = length;
916 	mr->ibmr.device = &dev->ib_dev;
917 	mr->ibmr.iova = iova;
918 	mr->access_flags = access_flags;
919 }
920 
921 static unsigned int mlx5_umem_dmabuf_default_pgsz(struct ib_umem *umem,
922 						  u64 iova)
923 {
924 	/*
925 	 * The alignment of iova has already been checked upon entering
926 	 * UVERBS_METHOD_REG_DMABUF_MR
927 	 */
928 	umem->iova = iova;
929 	return PAGE_SIZE;
930 }
931 
932 static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd,
933 					     struct ib_umem *umem, u64 iova,
934 					     int access_flags)
935 {
936 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
937 	struct mlx5_cache_ent *ent;
938 	struct mlx5_ib_mr *mr;
939 	unsigned int page_size;
940 
941 	if (umem->is_dmabuf)
942 		page_size = mlx5_umem_dmabuf_default_pgsz(umem, iova);
943 	else
944 		page_size = mlx5_umem_find_best_pgsz(umem, mkc, log_page_size,
945 						     0, iova);
946 	if (WARN_ON(!page_size))
947 		return ERR_PTR(-EINVAL);
948 	ent = mr_cache_ent_from_order(
949 		dev, order_base_2(ib_umem_num_dma_blocks(umem, page_size)));
950 	/*
951 	 * Matches access in alloc_cache_mr(). If the MR can't come from the
952 	 * cache then synchronously create an uncached one.
953 	 */
954 	if (!ent || ent->limit == 0 ||
955 	    !mlx5_ib_can_reconfig_with_umr(dev, 0, access_flags)) {
956 		mutex_lock(&dev->slow_path_mutex);
957 		mr = reg_create(pd, umem, iova, access_flags, page_size, false);
958 		mutex_unlock(&dev->slow_path_mutex);
959 		return mr;
960 	}
961 
962 	mr = get_cache_mr(ent);
963 	if (!mr) {
964 		mr = create_cache_mr(ent);
965 		/*
966 		 * The above already tried to do the same stuff as reg_create(),
967 		 * no reason to try it again.
968 		 */
969 		if (IS_ERR(mr))
970 			return mr;
971 	}
972 
973 	mr->ibmr.pd = pd;
974 	mr->umem = umem;
975 	mr->page_shift = order_base_2(page_size);
976 	set_mr_fields(dev, mr, umem->length, access_flags, iova);
977 
978 	return mr;
979 }
980 
981 #define MLX5_MAX_UMR_CHUNK ((1 << (MLX5_MAX_UMR_SHIFT + 4)) - \
982 			    MLX5_UMR_MTT_ALIGNMENT)
983 #define MLX5_SPARE_UMR_CHUNK 0x10000
984 
985 /*
986  * Allocate a temporary buffer to hold the per-page information to transfer to
987  * HW. For efficiency this should be as large as it can be, but buffer
988  * allocation failure is not allowed, so try smaller sizes.
989  */
990 static void *mlx5_ib_alloc_xlt(size_t *nents, size_t ent_size, gfp_t gfp_mask)
991 {
992 	const size_t xlt_chunk_align =
993 		MLX5_UMR_MTT_ALIGNMENT / ent_size;
994 	size_t size;
995 	void *res = NULL;
996 
997 	static_assert(PAGE_SIZE % MLX5_UMR_MTT_ALIGNMENT == 0);
998 
999 	/*
1000 	 * MLX5_IB_UPD_XLT_ATOMIC doesn't signal an atomic context just that the
1001 	 * allocation can't trigger any kind of reclaim.
1002 	 */
1003 	might_sleep();
1004 
1005 	gfp_mask |= __GFP_ZERO | __GFP_NORETRY;
1006 
1007 	/*
1008 	 * If the system already has a suitable high order page then just use
1009 	 * that, but don't try hard to create one. This max is about 1M, so a
1010 	 * free x86 huge page will satisfy it.
1011 	 */
1012 	size = min_t(size_t, ent_size * ALIGN(*nents, xlt_chunk_align),
1013 		     MLX5_MAX_UMR_CHUNK);
1014 	*nents = size / ent_size;
1015 	res = (void *)__get_free_pages(gfp_mask | __GFP_NOWARN,
1016 				       get_order(size));
1017 	if (res)
1018 		return res;
1019 
1020 	if (size > MLX5_SPARE_UMR_CHUNK) {
1021 		size = MLX5_SPARE_UMR_CHUNK;
1022 		*nents = size / ent_size;
1023 		res = (void *)__get_free_pages(gfp_mask | __GFP_NOWARN,
1024 					       get_order(size));
1025 		if (res)
1026 			return res;
1027 	}
1028 
1029 	*nents = PAGE_SIZE / ent_size;
1030 	res = (void *)__get_free_page(gfp_mask);
1031 	if (res)
1032 		return res;
1033 
1034 	mutex_lock(&xlt_emergency_page_mutex);
1035 	memset(xlt_emergency_page, 0, PAGE_SIZE);
1036 	return xlt_emergency_page;
1037 }
1038 
1039 static void mlx5_ib_free_xlt(void *xlt, size_t length)
1040 {
1041 	if (xlt == xlt_emergency_page) {
1042 		mutex_unlock(&xlt_emergency_page_mutex);
1043 		return;
1044 	}
1045 
1046 	free_pages((unsigned long)xlt, get_order(length));
1047 }
1048 
1049 /*
1050  * Create a MLX5_IB_SEND_UMR_UPDATE_XLT work request and XLT buffer ready for
1051  * submission.
1052  */
1053 static void *mlx5_ib_create_xlt_wr(struct mlx5_ib_mr *mr,
1054 				   struct mlx5_umr_wr *wr, struct ib_sge *sg,
1055 				   size_t nents, size_t ent_size,
1056 				   unsigned int flags)
1057 {
1058 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
1059 	struct device *ddev = &dev->mdev->pdev->dev;
1060 	dma_addr_t dma;
1061 	void *xlt;
1062 
1063 	xlt = mlx5_ib_alloc_xlt(&nents, ent_size,
1064 				flags & MLX5_IB_UPD_XLT_ATOMIC ? GFP_ATOMIC :
1065 								 GFP_KERNEL);
1066 	sg->length = nents * ent_size;
1067 	dma = dma_map_single(ddev, xlt, sg->length, DMA_TO_DEVICE);
1068 	if (dma_mapping_error(ddev, dma)) {
1069 		mlx5_ib_err(dev, "unable to map DMA during XLT update.\n");
1070 		mlx5_ib_free_xlt(xlt, sg->length);
1071 		return NULL;
1072 	}
1073 	sg->addr = dma;
1074 	sg->lkey = dev->umrc.pd->local_dma_lkey;
1075 
1076 	memset(wr, 0, sizeof(*wr));
1077 	wr->wr.send_flags = MLX5_IB_SEND_UMR_UPDATE_XLT;
1078 	if (!(flags & MLX5_IB_UPD_XLT_ENABLE))
1079 		wr->wr.send_flags |= MLX5_IB_SEND_UMR_FAIL_IF_FREE;
1080 	wr->wr.sg_list = sg;
1081 	wr->wr.num_sge = 1;
1082 	wr->wr.opcode = MLX5_IB_WR_UMR;
1083 	wr->pd = mr->ibmr.pd;
1084 	wr->mkey = mr->mmkey.key;
1085 	wr->length = mr->ibmr.length;
1086 	wr->virt_addr = mr->ibmr.iova;
1087 	wr->access_flags = mr->access_flags;
1088 	wr->page_shift = mr->page_shift;
1089 	wr->xlt_size = sg->length;
1090 	return xlt;
1091 }
1092 
1093 static void mlx5_ib_unmap_free_xlt(struct mlx5_ib_dev *dev, void *xlt,
1094 				   struct ib_sge *sg)
1095 {
1096 	struct device *ddev = &dev->mdev->pdev->dev;
1097 
1098 	dma_unmap_single(ddev, sg->addr, sg->length, DMA_TO_DEVICE);
1099 	mlx5_ib_free_xlt(xlt, sg->length);
1100 }
1101 
1102 static unsigned int xlt_wr_final_send_flags(unsigned int flags)
1103 {
1104 	unsigned int res = 0;
1105 
1106 	if (flags & MLX5_IB_UPD_XLT_ENABLE)
1107 		res |= MLX5_IB_SEND_UMR_ENABLE_MR |
1108 		       MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS |
1109 		       MLX5_IB_SEND_UMR_UPDATE_TRANSLATION;
1110 	if (flags & MLX5_IB_UPD_XLT_PD || flags & MLX5_IB_UPD_XLT_ACCESS)
1111 		res |= MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS;
1112 	if (flags & MLX5_IB_UPD_XLT_ADDR)
1113 		res |= MLX5_IB_SEND_UMR_UPDATE_TRANSLATION;
1114 	return res;
1115 }
1116 
1117 int mlx5_ib_update_xlt(struct mlx5_ib_mr *mr, u64 idx, int npages,
1118 		       int page_shift, int flags)
1119 {
1120 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
1121 	struct device *ddev = &dev->mdev->pdev->dev;
1122 	void *xlt;
1123 	struct mlx5_umr_wr wr;
1124 	struct ib_sge sg;
1125 	int err = 0;
1126 	int desc_size = (flags & MLX5_IB_UPD_XLT_INDIRECT)
1127 			       ? sizeof(struct mlx5_klm)
1128 			       : sizeof(struct mlx5_mtt);
1129 	const int page_align = MLX5_UMR_MTT_ALIGNMENT / desc_size;
1130 	const int page_mask = page_align - 1;
1131 	size_t pages_mapped = 0;
1132 	size_t pages_to_map = 0;
1133 	size_t pages_iter;
1134 	size_t size_to_map = 0;
1135 	size_t orig_sg_length;
1136 
1137 	if ((flags & MLX5_IB_UPD_XLT_INDIRECT) &&
1138 	    !umr_can_use_indirect_mkey(dev))
1139 		return -EPERM;
1140 
1141 	if (WARN_ON(!mr->umem->is_odp))
1142 		return -EINVAL;
1143 
1144 	/* UMR copies MTTs in units of MLX5_UMR_MTT_ALIGNMENT bytes,
1145 	 * so we need to align the offset and length accordingly
1146 	 */
1147 	if (idx & page_mask) {
1148 		npages += idx & page_mask;
1149 		idx &= ~page_mask;
1150 	}
1151 	pages_to_map = ALIGN(npages, page_align);
1152 
1153 	xlt = mlx5_ib_create_xlt_wr(mr, &wr, &sg, npages, desc_size, flags);
1154 	if (!xlt)
1155 		return -ENOMEM;
1156 	pages_iter = sg.length / desc_size;
1157 	orig_sg_length = sg.length;
1158 
1159 	if (!(flags & MLX5_IB_UPD_XLT_INDIRECT)) {
1160 		struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
1161 		size_t max_pages = ib_umem_odp_num_pages(odp) - idx;
1162 
1163 		pages_to_map = min_t(size_t, pages_to_map, max_pages);
1164 	}
1165 
1166 	wr.page_shift = page_shift;
1167 
1168 	for (pages_mapped = 0;
1169 	     pages_mapped < pages_to_map && !err;
1170 	     pages_mapped += pages_iter, idx += pages_iter) {
1171 		npages = min_t(int, pages_iter, pages_to_map - pages_mapped);
1172 		size_to_map = npages * desc_size;
1173 		dma_sync_single_for_cpu(ddev, sg.addr, sg.length,
1174 					DMA_TO_DEVICE);
1175 		mlx5_odp_populate_xlt(xlt, idx, npages, mr, flags);
1176 		dma_sync_single_for_device(ddev, sg.addr, sg.length,
1177 					   DMA_TO_DEVICE);
1178 
1179 		sg.length = ALIGN(size_to_map, MLX5_UMR_MTT_ALIGNMENT);
1180 
1181 		if (pages_mapped + pages_iter >= pages_to_map)
1182 			wr.wr.send_flags |= xlt_wr_final_send_flags(flags);
1183 
1184 		wr.offset = idx * desc_size;
1185 		wr.xlt_size = sg.length;
1186 
1187 		err = mlx5_ib_post_send_wait(dev, &wr);
1188 	}
1189 	sg.length = orig_sg_length;
1190 	mlx5_ib_unmap_free_xlt(dev, xlt, &sg);
1191 	return err;
1192 }
1193 
1194 /*
1195  * Send the DMA list to the HW for a normal MR using UMR.
1196  * Dmabuf MR is handled in a similar way, except that the MLX5_IB_UPD_XLT_ZAP
1197  * flag may be used.
1198  */
1199 int mlx5_ib_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
1200 {
1201 	struct mlx5_ib_dev *dev = mr_to_mdev(mr);
1202 	struct device *ddev = &dev->mdev->pdev->dev;
1203 	struct ib_block_iter biter;
1204 	struct mlx5_mtt *cur_mtt;
1205 	struct mlx5_umr_wr wr;
1206 	size_t orig_sg_length;
1207 	struct mlx5_mtt *mtt;
1208 	size_t final_size;
1209 	struct ib_sge sg;
1210 	int err = 0;
1211 
1212 	if (WARN_ON(mr->umem->is_odp))
1213 		return -EINVAL;
1214 
1215 	mtt = mlx5_ib_create_xlt_wr(mr, &wr, &sg,
1216 				    ib_umem_num_dma_blocks(mr->umem,
1217 							   1 << mr->page_shift),
1218 				    sizeof(*mtt), flags);
1219 	if (!mtt)
1220 		return -ENOMEM;
1221 	orig_sg_length = sg.length;
1222 
1223 	cur_mtt = mtt;
1224 	rdma_for_each_block (mr->umem->sgt_append.sgt.sgl, &biter,
1225 			     mr->umem->sgt_append.sgt.nents,
1226 			     BIT(mr->page_shift)) {
1227 		if (cur_mtt == (void *)mtt + sg.length) {
1228 			dma_sync_single_for_device(ddev, sg.addr, sg.length,
1229 						   DMA_TO_DEVICE);
1230 			err = mlx5_ib_post_send_wait(dev, &wr);
1231 			if (err)
1232 				goto err;
1233 			dma_sync_single_for_cpu(ddev, sg.addr, sg.length,
1234 						DMA_TO_DEVICE);
1235 			wr.offset += sg.length;
1236 			cur_mtt = mtt;
1237 		}
1238 
1239 		cur_mtt->ptag =
1240 			cpu_to_be64(rdma_block_iter_dma_address(&biter) |
1241 				    MLX5_IB_MTT_PRESENT);
1242 
1243 		if (mr->umem->is_dmabuf && (flags & MLX5_IB_UPD_XLT_ZAP))
1244 			cur_mtt->ptag = 0;
1245 
1246 		cur_mtt++;
1247 	}
1248 
1249 	final_size = (void *)cur_mtt - (void *)mtt;
1250 	sg.length = ALIGN(final_size, MLX5_UMR_MTT_ALIGNMENT);
1251 	memset(cur_mtt, 0, sg.length - final_size);
1252 	wr.wr.send_flags |= xlt_wr_final_send_flags(flags);
1253 	wr.xlt_size = sg.length;
1254 
1255 	dma_sync_single_for_device(ddev, sg.addr, sg.length, DMA_TO_DEVICE);
1256 	err = mlx5_ib_post_send_wait(dev, &wr);
1257 
1258 err:
1259 	sg.length = orig_sg_length;
1260 	mlx5_ib_unmap_free_xlt(dev, mtt, &sg);
1261 	return err;
1262 }
1263 
1264 /*
1265  * If ibmr is NULL it will be allocated by reg_create.
1266  * Else, the given ibmr will be used.
1267  */
1268 static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem,
1269 				     u64 iova, int access_flags,
1270 				     unsigned int page_size, bool populate)
1271 {
1272 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1273 	struct mlx5_ib_mr *mr;
1274 	__be64 *pas;
1275 	void *mkc;
1276 	int inlen;
1277 	u32 *in;
1278 	int err;
1279 	bool pg_cap = !!(MLX5_CAP_GEN(dev->mdev, pg));
1280 
1281 	if (!page_size)
1282 		return ERR_PTR(-EINVAL);
1283 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1284 	if (!mr)
1285 		return ERR_PTR(-ENOMEM);
1286 
1287 	mr->ibmr.pd = pd;
1288 	mr->access_flags = access_flags;
1289 	mr->page_shift = order_base_2(page_size);
1290 
1291 	inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1292 	if (populate)
1293 		inlen += sizeof(*pas) *
1294 			 roundup(ib_umem_num_dma_blocks(umem, page_size), 2);
1295 	in = kvzalloc(inlen, GFP_KERNEL);
1296 	if (!in) {
1297 		err = -ENOMEM;
1298 		goto err_1;
1299 	}
1300 	pas = (__be64 *)MLX5_ADDR_OF(create_mkey_in, in, klm_pas_mtt);
1301 	if (populate) {
1302 		if (WARN_ON(access_flags & IB_ACCESS_ON_DEMAND)) {
1303 			err = -EINVAL;
1304 			goto err_2;
1305 		}
1306 		mlx5_ib_populate_pas(umem, 1UL << mr->page_shift, pas,
1307 				     pg_cap ? MLX5_IB_MTT_PRESENT : 0);
1308 	}
1309 
1310 	/* The pg_access bit allows setting the access flags
1311 	 * in the page list submitted with the command. */
1312 	MLX5_SET(create_mkey_in, in, pg_access, !!(pg_cap));
1313 
1314 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1315 	set_mkc_access_pd_addr_fields(mkc, access_flags, iova,
1316 				      populate ? pd : dev->umrc.pd);
1317 	MLX5_SET(mkc, mkc, free, !populate);
1318 	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_MTT);
1319 	MLX5_SET(mkc, mkc, umr_en, 1);
1320 
1321 	MLX5_SET64(mkc, mkc, len, umem->length);
1322 	MLX5_SET(mkc, mkc, bsf_octword_size, 0);
1323 	MLX5_SET(mkc, mkc, translations_octword_size,
1324 		 get_octo_len(iova, umem->length, mr->page_shift));
1325 	MLX5_SET(mkc, mkc, log_page_size, mr->page_shift);
1326 	if (populate) {
1327 		MLX5_SET(create_mkey_in, in, translations_octword_actual_size,
1328 			 get_octo_len(iova, umem->length, mr->page_shift));
1329 	}
1330 
1331 	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
1332 	if (err) {
1333 		mlx5_ib_warn(dev, "create mkey failed\n");
1334 		goto err_2;
1335 	}
1336 	mr->mmkey.type = MLX5_MKEY_MR;
1337 	mr->umem = umem;
1338 	set_mr_fields(dev, mr, umem->length, access_flags, iova);
1339 	kvfree(in);
1340 
1341 	mlx5_ib_dbg(dev, "mkey = 0x%x\n", mr->mmkey.key);
1342 
1343 	return mr;
1344 
1345 err_2:
1346 	kvfree(in);
1347 err_1:
1348 	kfree(mr);
1349 	return ERR_PTR(err);
1350 }
1351 
1352 static struct ib_mr *mlx5_ib_get_dm_mr(struct ib_pd *pd, u64 start_addr,
1353 				       u64 length, int acc, int mode)
1354 {
1355 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1356 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1357 	struct mlx5_ib_mr *mr;
1358 	void *mkc;
1359 	u32 *in;
1360 	int err;
1361 
1362 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1363 	if (!mr)
1364 		return ERR_PTR(-ENOMEM);
1365 
1366 	in = kzalloc(inlen, GFP_KERNEL);
1367 	if (!in) {
1368 		err = -ENOMEM;
1369 		goto err_free;
1370 	}
1371 
1372 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1373 
1374 	MLX5_SET(mkc, mkc, access_mode_1_0, mode & 0x3);
1375 	MLX5_SET(mkc, mkc, access_mode_4_2, (mode >> 2) & 0x7);
1376 	MLX5_SET64(mkc, mkc, len, length);
1377 	set_mkc_access_pd_addr_fields(mkc, acc, start_addr, pd);
1378 
1379 	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
1380 	if (err)
1381 		goto err_in;
1382 
1383 	kfree(in);
1384 
1385 	set_mr_fields(dev, mr, length, acc, start_addr);
1386 
1387 	return &mr->ibmr;
1388 
1389 err_in:
1390 	kfree(in);
1391 
1392 err_free:
1393 	kfree(mr);
1394 
1395 	return ERR_PTR(err);
1396 }
1397 
1398 int mlx5_ib_advise_mr(struct ib_pd *pd,
1399 		      enum ib_uverbs_advise_mr_advice advice,
1400 		      u32 flags,
1401 		      struct ib_sge *sg_list,
1402 		      u32 num_sge,
1403 		      struct uverbs_attr_bundle *attrs)
1404 {
1405 	if (advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH &&
1406 	    advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
1407 	    advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT)
1408 		return -EOPNOTSUPP;
1409 
1410 	return mlx5_ib_advise_mr_prefetch(pd, advice, flags,
1411 					 sg_list, num_sge);
1412 }
1413 
1414 struct ib_mr *mlx5_ib_reg_dm_mr(struct ib_pd *pd, struct ib_dm *dm,
1415 				struct ib_dm_mr_attr *attr,
1416 				struct uverbs_attr_bundle *attrs)
1417 {
1418 	struct mlx5_ib_dm *mdm = to_mdm(dm);
1419 	struct mlx5_core_dev *dev = to_mdev(dm->device)->mdev;
1420 	u64 start_addr = mdm->dev_addr + attr->offset;
1421 	int mode;
1422 
1423 	switch (mdm->type) {
1424 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
1425 		if (attr->access_flags & ~MLX5_IB_DM_MEMIC_ALLOWED_ACCESS)
1426 			return ERR_PTR(-EINVAL);
1427 
1428 		mode = MLX5_MKC_ACCESS_MODE_MEMIC;
1429 		start_addr -= pci_resource_start(dev->pdev, 0);
1430 		break;
1431 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
1432 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
1433 		if (attr->access_flags & ~MLX5_IB_DM_SW_ICM_ALLOWED_ACCESS)
1434 			return ERR_PTR(-EINVAL);
1435 
1436 		mode = MLX5_MKC_ACCESS_MODE_SW_ICM;
1437 		break;
1438 	default:
1439 		return ERR_PTR(-EINVAL);
1440 	}
1441 
1442 	return mlx5_ib_get_dm_mr(pd, start_addr, attr->length,
1443 				 attr->access_flags, mode);
1444 }
1445 
1446 static struct ib_mr *create_real_mr(struct ib_pd *pd, struct ib_umem *umem,
1447 				    u64 iova, int access_flags)
1448 {
1449 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1450 	struct mlx5_ib_mr *mr = NULL;
1451 	bool xlt_with_umr;
1452 	int err;
1453 
1454 	xlt_with_umr = mlx5_ib_can_load_pas_with_umr(dev, umem->length);
1455 	if (xlt_with_umr) {
1456 		mr = alloc_cacheable_mr(pd, umem, iova, access_flags);
1457 	} else {
1458 		unsigned int page_size = mlx5_umem_find_best_pgsz(
1459 			umem, mkc, log_page_size, 0, iova);
1460 
1461 		mutex_lock(&dev->slow_path_mutex);
1462 		mr = reg_create(pd, umem, iova, access_flags, page_size, true);
1463 		mutex_unlock(&dev->slow_path_mutex);
1464 	}
1465 	if (IS_ERR(mr)) {
1466 		ib_umem_release(umem);
1467 		return ERR_CAST(mr);
1468 	}
1469 
1470 	mlx5_ib_dbg(dev, "mkey 0x%x\n", mr->mmkey.key);
1471 
1472 	atomic_add(ib_umem_num_pages(umem), &dev->mdev->priv.reg_pages);
1473 
1474 	if (xlt_with_umr) {
1475 		/*
1476 		 * If the MR was created with reg_create then it will be
1477 		 * configured properly but left disabled. It is safe to go ahead
1478 		 * and configure it again via UMR while enabling it.
1479 		 */
1480 		err = mlx5_ib_update_mr_pas(mr, MLX5_IB_UPD_XLT_ENABLE);
1481 		if (err) {
1482 			mlx5_ib_dereg_mr(&mr->ibmr, NULL);
1483 			return ERR_PTR(err);
1484 		}
1485 	}
1486 	return &mr->ibmr;
1487 }
1488 
1489 static struct ib_mr *create_user_odp_mr(struct ib_pd *pd, u64 start, u64 length,
1490 					u64 iova, int access_flags,
1491 					struct ib_udata *udata)
1492 {
1493 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1494 	struct ib_umem_odp *odp;
1495 	struct mlx5_ib_mr *mr;
1496 	int err;
1497 
1498 	if (!IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING))
1499 		return ERR_PTR(-EOPNOTSUPP);
1500 
1501 	err = mlx5r_odp_create_eq(dev, &dev->odp_pf_eq);
1502 	if (err)
1503 		return ERR_PTR(err);
1504 	if (!start && length == U64_MAX) {
1505 		if (iova != 0)
1506 			return ERR_PTR(-EINVAL);
1507 		if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1508 			return ERR_PTR(-EINVAL);
1509 
1510 		mr = mlx5_ib_alloc_implicit_mr(to_mpd(pd), access_flags);
1511 		if (IS_ERR(mr))
1512 			return ERR_CAST(mr);
1513 		return &mr->ibmr;
1514 	}
1515 
1516 	/* ODP requires xlt update via umr to work. */
1517 	if (!mlx5_ib_can_load_pas_with_umr(dev, length))
1518 		return ERR_PTR(-EINVAL);
1519 
1520 	odp = ib_umem_odp_get(&dev->ib_dev, start, length, access_flags,
1521 			      &mlx5_mn_ops);
1522 	if (IS_ERR(odp))
1523 		return ERR_CAST(odp);
1524 
1525 	mr = alloc_cacheable_mr(pd, &odp->umem, iova, access_flags);
1526 	if (IS_ERR(mr)) {
1527 		ib_umem_release(&odp->umem);
1528 		return ERR_CAST(mr);
1529 	}
1530 	xa_init(&mr->implicit_children);
1531 
1532 	odp->private = mr;
1533 	err = mlx5r_store_odp_mkey(dev, &mr->mmkey);
1534 	if (err)
1535 		goto err_dereg_mr;
1536 
1537 	err = mlx5_ib_init_odp_mr(mr);
1538 	if (err)
1539 		goto err_dereg_mr;
1540 	return &mr->ibmr;
1541 
1542 err_dereg_mr:
1543 	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
1544 	return ERR_PTR(err);
1545 }
1546 
1547 struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
1548 				  u64 iova, int access_flags,
1549 				  struct ib_udata *udata)
1550 {
1551 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1552 	struct ib_umem *umem;
1553 
1554 	if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM))
1555 		return ERR_PTR(-EOPNOTSUPP);
1556 
1557 	mlx5_ib_dbg(dev, "start 0x%llx, iova 0x%llx, length 0x%llx, access_flags 0x%x\n",
1558 		    start, iova, length, access_flags);
1559 
1560 	if (access_flags & IB_ACCESS_ON_DEMAND)
1561 		return create_user_odp_mr(pd, start, length, iova, access_flags,
1562 					  udata);
1563 	umem = ib_umem_get(&dev->ib_dev, start, length, access_flags);
1564 	if (IS_ERR(umem))
1565 		return ERR_CAST(umem);
1566 	return create_real_mr(pd, umem, iova, access_flags);
1567 }
1568 
1569 static void mlx5_ib_dmabuf_invalidate_cb(struct dma_buf_attachment *attach)
1570 {
1571 	struct ib_umem_dmabuf *umem_dmabuf = attach->importer_priv;
1572 	struct mlx5_ib_mr *mr = umem_dmabuf->private;
1573 
1574 	dma_resv_assert_held(umem_dmabuf->attach->dmabuf->resv);
1575 
1576 	if (!umem_dmabuf->sgt)
1577 		return;
1578 
1579 	mlx5_ib_update_mr_pas(mr, MLX5_IB_UPD_XLT_ZAP);
1580 	ib_umem_dmabuf_unmap_pages(umem_dmabuf);
1581 }
1582 
1583 static struct dma_buf_attach_ops mlx5_ib_dmabuf_attach_ops = {
1584 	.allow_peer2peer = 1,
1585 	.move_notify = mlx5_ib_dmabuf_invalidate_cb,
1586 };
1587 
1588 struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset,
1589 					 u64 length, u64 virt_addr,
1590 					 int fd, int access_flags,
1591 					 struct ib_udata *udata)
1592 {
1593 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1594 	struct mlx5_ib_mr *mr = NULL;
1595 	struct ib_umem_dmabuf *umem_dmabuf;
1596 	int err;
1597 
1598 	if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM) ||
1599 	    !IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING))
1600 		return ERR_PTR(-EOPNOTSUPP);
1601 
1602 	mlx5_ib_dbg(dev,
1603 		    "offset 0x%llx, virt_addr 0x%llx, length 0x%llx, fd %d, access_flags 0x%x\n",
1604 		    offset, virt_addr, length, fd, access_flags);
1605 
1606 	/* dmabuf requires xlt update via umr to work. */
1607 	if (!mlx5_ib_can_load_pas_with_umr(dev, length))
1608 		return ERR_PTR(-EINVAL);
1609 
1610 	umem_dmabuf = ib_umem_dmabuf_get(&dev->ib_dev, offset, length, fd,
1611 					 access_flags,
1612 					 &mlx5_ib_dmabuf_attach_ops);
1613 	if (IS_ERR(umem_dmabuf)) {
1614 		mlx5_ib_dbg(dev, "umem_dmabuf get failed (%ld)\n",
1615 			    PTR_ERR(umem_dmabuf));
1616 		return ERR_CAST(umem_dmabuf);
1617 	}
1618 
1619 	mr = alloc_cacheable_mr(pd, &umem_dmabuf->umem, virt_addr,
1620 				access_flags);
1621 	if (IS_ERR(mr)) {
1622 		ib_umem_release(&umem_dmabuf->umem);
1623 		return ERR_CAST(mr);
1624 	}
1625 
1626 	mlx5_ib_dbg(dev, "mkey 0x%x\n", mr->mmkey.key);
1627 
1628 	atomic_add(ib_umem_num_pages(mr->umem), &dev->mdev->priv.reg_pages);
1629 	umem_dmabuf->private = mr;
1630 	err = mlx5r_store_odp_mkey(dev, &mr->mmkey);
1631 	if (err)
1632 		goto err_dereg_mr;
1633 
1634 	err = mlx5_ib_init_dmabuf_mr(mr);
1635 	if (err)
1636 		goto err_dereg_mr;
1637 	return &mr->ibmr;
1638 
1639 err_dereg_mr:
1640 	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
1641 	return ERR_PTR(err);
1642 }
1643 
1644 /**
1645  * revoke_mr - Fence all DMA on the MR
1646  * @mr: The MR to fence
1647  *
1648  * Upon return the NIC will not be doing any DMA to the pages under the MR,
1649  * and any DMA in progress will be completed. Failure of this function
1650  * indicates the HW has failed catastrophically.
1651  */
1652 static int revoke_mr(struct mlx5_ib_mr *mr)
1653 {
1654 	struct mlx5_umr_wr umrwr = {};
1655 
1656 	if (mr_to_mdev(mr)->mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)
1657 		return 0;
1658 
1659 	umrwr.wr.send_flags = MLX5_IB_SEND_UMR_DISABLE_MR |
1660 			      MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS;
1661 	umrwr.wr.opcode = MLX5_IB_WR_UMR;
1662 	umrwr.pd = mr_to_mdev(mr)->umrc.pd;
1663 	umrwr.mkey = mr->mmkey.key;
1664 	umrwr.ignore_free_state = 1;
1665 
1666 	return mlx5_ib_post_send_wait(mr_to_mdev(mr), &umrwr);
1667 }
1668 
1669 /*
1670  * True if the change in access flags can be done via UMR, only some access
1671  * flags can be updated.
1672  */
1673 static bool can_use_umr_rereg_access(struct mlx5_ib_dev *dev,
1674 				     unsigned int current_access_flags,
1675 				     unsigned int target_access_flags)
1676 {
1677 	unsigned int diffs = current_access_flags ^ target_access_flags;
1678 
1679 	if (diffs & ~(IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE |
1680 		      IB_ACCESS_REMOTE_READ | IB_ACCESS_RELAXED_ORDERING))
1681 		return false;
1682 	return mlx5_ib_can_reconfig_with_umr(dev, current_access_flags,
1683 					     target_access_flags);
1684 }
1685 
1686 static int umr_rereg_pd_access(struct mlx5_ib_mr *mr, struct ib_pd *pd,
1687 			       int access_flags)
1688 {
1689 	struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device);
1690 	struct mlx5_umr_wr umrwr = {
1691 		.wr = {
1692 			.send_flags = MLX5_IB_SEND_UMR_FAIL_IF_FREE |
1693 				      MLX5_IB_SEND_UMR_UPDATE_PD_ACCESS,
1694 			.opcode = MLX5_IB_WR_UMR,
1695 		},
1696 		.mkey = mr->mmkey.key,
1697 		.pd = pd,
1698 		.access_flags = access_flags,
1699 	};
1700 	int err;
1701 
1702 	err = mlx5_ib_post_send_wait(dev, &umrwr);
1703 	if (err)
1704 		return err;
1705 
1706 	mr->access_flags = access_flags;
1707 	return 0;
1708 }
1709 
1710 static bool can_use_umr_rereg_pas(struct mlx5_ib_mr *mr,
1711 				  struct ib_umem *new_umem,
1712 				  int new_access_flags, u64 iova,
1713 				  unsigned long *page_size)
1714 {
1715 	struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device);
1716 
1717 	/* We only track the allocated sizes of MRs from the cache */
1718 	if (!mr->cache_ent)
1719 		return false;
1720 	if (!mlx5_ib_can_load_pas_with_umr(dev, new_umem->length))
1721 		return false;
1722 
1723 	*page_size =
1724 		mlx5_umem_find_best_pgsz(new_umem, mkc, log_page_size, 0, iova);
1725 	if (WARN_ON(!*page_size))
1726 		return false;
1727 	return (1ULL << mr->cache_ent->order) >=
1728 	       ib_umem_num_dma_blocks(new_umem, *page_size);
1729 }
1730 
1731 static int umr_rereg_pas(struct mlx5_ib_mr *mr, struct ib_pd *pd,
1732 			 int access_flags, int flags, struct ib_umem *new_umem,
1733 			 u64 iova, unsigned long page_size)
1734 {
1735 	struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device);
1736 	int upd_flags = MLX5_IB_UPD_XLT_ADDR | MLX5_IB_UPD_XLT_ENABLE;
1737 	struct ib_umem *old_umem = mr->umem;
1738 	int err;
1739 
1740 	/*
1741 	 * To keep everything simple the MR is revoked before we start to mess
1742 	 * with it. This ensure the change is atomic relative to any use of the
1743 	 * MR.
1744 	 */
1745 	err = revoke_mr(mr);
1746 	if (err)
1747 		return err;
1748 
1749 	if (flags & IB_MR_REREG_PD) {
1750 		mr->ibmr.pd = pd;
1751 		upd_flags |= MLX5_IB_UPD_XLT_PD;
1752 	}
1753 	if (flags & IB_MR_REREG_ACCESS) {
1754 		mr->access_flags = access_flags;
1755 		upd_flags |= MLX5_IB_UPD_XLT_ACCESS;
1756 	}
1757 
1758 	mr->ibmr.length = new_umem->length;
1759 	mr->ibmr.iova = iova;
1760 	mr->ibmr.length = new_umem->length;
1761 	mr->page_shift = order_base_2(page_size);
1762 	mr->umem = new_umem;
1763 	err = mlx5_ib_update_mr_pas(mr, upd_flags);
1764 	if (err) {
1765 		/*
1766 		 * The MR is revoked at this point so there is no issue to free
1767 		 * new_umem.
1768 		 */
1769 		mr->umem = old_umem;
1770 		return err;
1771 	}
1772 
1773 	atomic_sub(ib_umem_num_pages(old_umem), &dev->mdev->priv.reg_pages);
1774 	ib_umem_release(old_umem);
1775 	atomic_add(ib_umem_num_pages(new_umem), &dev->mdev->priv.reg_pages);
1776 	return 0;
1777 }
1778 
1779 struct ib_mr *mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start,
1780 				    u64 length, u64 iova, int new_access_flags,
1781 				    struct ib_pd *new_pd,
1782 				    struct ib_udata *udata)
1783 {
1784 	struct mlx5_ib_dev *dev = to_mdev(ib_mr->device);
1785 	struct mlx5_ib_mr *mr = to_mmr(ib_mr);
1786 	int err;
1787 
1788 	if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM))
1789 		return ERR_PTR(-EOPNOTSUPP);
1790 
1791 	mlx5_ib_dbg(
1792 		dev,
1793 		"start 0x%llx, iova 0x%llx, length 0x%llx, access_flags 0x%x\n",
1794 		start, iova, length, new_access_flags);
1795 
1796 	if (flags & ~(IB_MR_REREG_TRANS | IB_MR_REREG_PD | IB_MR_REREG_ACCESS))
1797 		return ERR_PTR(-EOPNOTSUPP);
1798 
1799 	if (!(flags & IB_MR_REREG_ACCESS))
1800 		new_access_flags = mr->access_flags;
1801 	if (!(flags & IB_MR_REREG_PD))
1802 		new_pd = ib_mr->pd;
1803 
1804 	if (!(flags & IB_MR_REREG_TRANS)) {
1805 		struct ib_umem *umem;
1806 
1807 		/* Fast path for PD/access change */
1808 		if (can_use_umr_rereg_access(dev, mr->access_flags,
1809 					     new_access_flags)) {
1810 			err = umr_rereg_pd_access(mr, new_pd, new_access_flags);
1811 			if (err)
1812 				return ERR_PTR(err);
1813 			return NULL;
1814 		}
1815 		/* DM or ODP MR's don't have a normal umem so we can't re-use it */
1816 		if (!mr->umem || is_odp_mr(mr) || is_dmabuf_mr(mr))
1817 			goto recreate;
1818 
1819 		/*
1820 		 * Only one active MR can refer to a umem at one time, revoke
1821 		 * the old MR before assigning the umem to the new one.
1822 		 */
1823 		err = revoke_mr(mr);
1824 		if (err)
1825 			return ERR_PTR(err);
1826 		umem = mr->umem;
1827 		mr->umem = NULL;
1828 		atomic_sub(ib_umem_num_pages(umem), &dev->mdev->priv.reg_pages);
1829 
1830 		return create_real_mr(new_pd, umem, mr->ibmr.iova,
1831 				      new_access_flags);
1832 	}
1833 
1834 	/*
1835 	 * DM doesn't have a PAS list so we can't re-use it, odp/dmabuf does
1836 	 * but the logic around releasing the umem is different
1837 	 */
1838 	if (!mr->umem || is_odp_mr(mr) || is_dmabuf_mr(mr))
1839 		goto recreate;
1840 
1841 	if (!(new_access_flags & IB_ACCESS_ON_DEMAND) &&
1842 	    can_use_umr_rereg_access(dev, mr->access_flags, new_access_flags)) {
1843 		struct ib_umem *new_umem;
1844 		unsigned long page_size;
1845 
1846 		new_umem = ib_umem_get(&dev->ib_dev, start, length,
1847 				       new_access_flags);
1848 		if (IS_ERR(new_umem))
1849 			return ERR_CAST(new_umem);
1850 
1851 		/* Fast path for PAS change */
1852 		if (can_use_umr_rereg_pas(mr, new_umem, new_access_flags, iova,
1853 					  &page_size)) {
1854 			err = umr_rereg_pas(mr, new_pd, new_access_flags, flags,
1855 					    new_umem, iova, page_size);
1856 			if (err) {
1857 				ib_umem_release(new_umem);
1858 				return ERR_PTR(err);
1859 			}
1860 			return NULL;
1861 		}
1862 		return create_real_mr(new_pd, new_umem, iova, new_access_flags);
1863 	}
1864 
1865 	/*
1866 	 * Everything else has no state we can preserve, just create a new MR
1867 	 * from scratch
1868 	 */
1869 recreate:
1870 	return mlx5_ib_reg_user_mr(new_pd, start, length, iova,
1871 				   new_access_flags, udata);
1872 }
1873 
1874 static int
1875 mlx5_alloc_priv_descs(struct ib_device *device,
1876 		      struct mlx5_ib_mr *mr,
1877 		      int ndescs,
1878 		      int desc_size)
1879 {
1880 	struct mlx5_ib_dev *dev = to_mdev(device);
1881 	struct device *ddev = &dev->mdev->pdev->dev;
1882 	int size = ndescs * desc_size;
1883 	int add_size;
1884 	int ret;
1885 
1886 	add_size = max_t(int, MLX5_UMR_ALIGN - ARCH_KMALLOC_MINALIGN, 0);
1887 
1888 	mr->descs_alloc = kzalloc(size + add_size, GFP_KERNEL);
1889 	if (!mr->descs_alloc)
1890 		return -ENOMEM;
1891 
1892 	mr->descs = PTR_ALIGN(mr->descs_alloc, MLX5_UMR_ALIGN);
1893 
1894 	mr->desc_map = dma_map_single(ddev, mr->descs, size, DMA_TO_DEVICE);
1895 	if (dma_mapping_error(ddev, mr->desc_map)) {
1896 		ret = -ENOMEM;
1897 		goto err;
1898 	}
1899 
1900 	return 0;
1901 err:
1902 	kfree(mr->descs_alloc);
1903 
1904 	return ret;
1905 }
1906 
1907 static void
1908 mlx5_free_priv_descs(struct mlx5_ib_mr *mr)
1909 {
1910 	if (!mr->umem && mr->descs) {
1911 		struct ib_device *device = mr->ibmr.device;
1912 		int size = mr->max_descs * mr->desc_size;
1913 		struct mlx5_ib_dev *dev = to_mdev(device);
1914 
1915 		dma_unmap_single(&dev->mdev->pdev->dev, mr->desc_map, size,
1916 				 DMA_TO_DEVICE);
1917 		kfree(mr->descs_alloc);
1918 		mr->descs = NULL;
1919 	}
1920 }
1921 
1922 int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1923 {
1924 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
1925 	struct mlx5_ib_dev *dev = to_mdev(ibmr->device);
1926 	int rc;
1927 
1928 	/*
1929 	 * Any async use of the mr must hold the refcount, once the refcount
1930 	 * goes to zero no other thread, such as ODP page faults, prefetch, any
1931 	 * UMR activity, etc can touch the mkey. Thus it is safe to destroy it.
1932 	 */
1933 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) &&
1934 	    refcount_read(&mr->mmkey.usecount) != 0 &&
1935 	    xa_erase(&mr_to_mdev(mr)->odp_mkeys, mlx5_base_mkey(mr->mmkey.key)))
1936 		mlx5r_deref_wait_odp_mkey(&mr->mmkey);
1937 
1938 	if (ibmr->type == IB_MR_TYPE_INTEGRITY) {
1939 		xa_cmpxchg(&dev->sig_mrs, mlx5_base_mkey(mr->mmkey.key),
1940 			   mr->sig, NULL, GFP_KERNEL);
1941 
1942 		if (mr->mtt_mr) {
1943 			rc = mlx5_ib_dereg_mr(&mr->mtt_mr->ibmr, NULL);
1944 			if (rc)
1945 				return rc;
1946 			mr->mtt_mr = NULL;
1947 		}
1948 		if (mr->klm_mr) {
1949 			rc = mlx5_ib_dereg_mr(&mr->klm_mr->ibmr, NULL);
1950 			if (rc)
1951 				return rc;
1952 			mr->klm_mr = NULL;
1953 		}
1954 
1955 		if (mlx5_core_destroy_psv(dev->mdev,
1956 					  mr->sig->psv_memory.psv_idx))
1957 			mlx5_ib_warn(dev, "failed to destroy mem psv %d\n",
1958 				     mr->sig->psv_memory.psv_idx);
1959 		if (mlx5_core_destroy_psv(dev->mdev, mr->sig->psv_wire.psv_idx))
1960 			mlx5_ib_warn(dev, "failed to destroy wire psv %d\n",
1961 				     mr->sig->psv_wire.psv_idx);
1962 		kfree(mr->sig);
1963 		mr->sig = NULL;
1964 	}
1965 
1966 	/* Stop DMA */
1967 	if (mr->cache_ent) {
1968 		if (revoke_mr(mr)) {
1969 			spin_lock_irq(&mr->cache_ent->lock);
1970 			mr->cache_ent->total_mrs--;
1971 			spin_unlock_irq(&mr->cache_ent->lock);
1972 			mr->cache_ent = NULL;
1973 		}
1974 	}
1975 	if (!mr->cache_ent) {
1976 		rc = destroy_mkey(to_mdev(mr->ibmr.device), mr);
1977 		if (rc)
1978 			return rc;
1979 	}
1980 
1981 	if (mr->umem) {
1982 		bool is_odp = is_odp_mr(mr);
1983 
1984 		if (!is_odp)
1985 			atomic_sub(ib_umem_num_pages(mr->umem),
1986 				   &dev->mdev->priv.reg_pages);
1987 		ib_umem_release(mr->umem);
1988 		if (is_odp)
1989 			mlx5_ib_free_odp_mr(mr);
1990 	}
1991 
1992 	if (mr->cache_ent) {
1993 		mlx5_mr_cache_free(dev, mr);
1994 	} else {
1995 		mlx5_free_priv_descs(mr);
1996 		kfree(mr);
1997 	}
1998 	return 0;
1999 }
2000 
2001 static void mlx5_set_umr_free_mkey(struct ib_pd *pd, u32 *in, int ndescs,
2002 				   int access_mode, int page_shift)
2003 {
2004 	void *mkc;
2005 
2006 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
2007 
2008 	/* This is only used from the kernel, so setting the PD is OK. */
2009 	set_mkc_access_pd_addr_fields(mkc, IB_ACCESS_RELAXED_ORDERING, 0, pd);
2010 	MLX5_SET(mkc, mkc, free, 1);
2011 	MLX5_SET(mkc, mkc, translations_octword_size, ndescs);
2012 	MLX5_SET(mkc, mkc, access_mode_1_0, access_mode & 0x3);
2013 	MLX5_SET(mkc, mkc, access_mode_4_2, (access_mode >> 2) & 0x7);
2014 	MLX5_SET(mkc, mkc, umr_en, 1);
2015 	MLX5_SET(mkc, mkc, log_page_size, page_shift);
2016 }
2017 
2018 static int _mlx5_alloc_mkey_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
2019 				  int ndescs, int desc_size, int page_shift,
2020 				  int access_mode, u32 *in, int inlen)
2021 {
2022 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
2023 	int err;
2024 
2025 	mr->access_mode = access_mode;
2026 	mr->desc_size = desc_size;
2027 	mr->max_descs = ndescs;
2028 
2029 	err = mlx5_alloc_priv_descs(pd->device, mr, ndescs, desc_size);
2030 	if (err)
2031 		return err;
2032 
2033 	mlx5_set_umr_free_mkey(pd, in, ndescs, access_mode, page_shift);
2034 
2035 	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
2036 	if (err)
2037 		goto err_free_descs;
2038 
2039 	mr->mmkey.type = MLX5_MKEY_MR;
2040 	mr->ibmr.lkey = mr->mmkey.key;
2041 	mr->ibmr.rkey = mr->mmkey.key;
2042 
2043 	return 0;
2044 
2045 err_free_descs:
2046 	mlx5_free_priv_descs(mr);
2047 	return err;
2048 }
2049 
2050 static struct mlx5_ib_mr *mlx5_ib_alloc_pi_mr(struct ib_pd *pd,
2051 				u32 max_num_sg, u32 max_num_meta_sg,
2052 				int desc_size, int access_mode)
2053 {
2054 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
2055 	int ndescs = ALIGN(max_num_sg + max_num_meta_sg, 4);
2056 	int page_shift = 0;
2057 	struct mlx5_ib_mr *mr;
2058 	u32 *in;
2059 	int err;
2060 
2061 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2062 	if (!mr)
2063 		return ERR_PTR(-ENOMEM);
2064 
2065 	mr->ibmr.pd = pd;
2066 	mr->ibmr.device = pd->device;
2067 
2068 	in = kzalloc(inlen, GFP_KERNEL);
2069 	if (!in) {
2070 		err = -ENOMEM;
2071 		goto err_free;
2072 	}
2073 
2074 	if (access_mode == MLX5_MKC_ACCESS_MODE_MTT)
2075 		page_shift = PAGE_SHIFT;
2076 
2077 	err = _mlx5_alloc_mkey_descs(pd, mr, ndescs, desc_size, page_shift,
2078 				     access_mode, in, inlen);
2079 	if (err)
2080 		goto err_free_in;
2081 
2082 	mr->umem = NULL;
2083 	kfree(in);
2084 
2085 	return mr;
2086 
2087 err_free_in:
2088 	kfree(in);
2089 err_free:
2090 	kfree(mr);
2091 	return ERR_PTR(err);
2092 }
2093 
2094 static int mlx5_alloc_mem_reg_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
2095 				    int ndescs, u32 *in, int inlen)
2096 {
2097 	return _mlx5_alloc_mkey_descs(pd, mr, ndescs, sizeof(struct mlx5_mtt),
2098 				      PAGE_SHIFT, MLX5_MKC_ACCESS_MODE_MTT, in,
2099 				      inlen);
2100 }
2101 
2102 static int mlx5_alloc_sg_gaps_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
2103 				    int ndescs, u32 *in, int inlen)
2104 {
2105 	return _mlx5_alloc_mkey_descs(pd, mr, ndescs, sizeof(struct mlx5_klm),
2106 				      0, MLX5_MKC_ACCESS_MODE_KLMS, in, inlen);
2107 }
2108 
2109 static int mlx5_alloc_integrity_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
2110 				      int max_num_sg, int max_num_meta_sg,
2111 				      u32 *in, int inlen)
2112 {
2113 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
2114 	u32 psv_index[2];
2115 	void *mkc;
2116 	int err;
2117 
2118 	mr->sig = kzalloc(sizeof(*mr->sig), GFP_KERNEL);
2119 	if (!mr->sig)
2120 		return -ENOMEM;
2121 
2122 	/* create mem & wire PSVs */
2123 	err = mlx5_core_create_psv(dev->mdev, to_mpd(pd)->pdn, 2, psv_index);
2124 	if (err)
2125 		goto err_free_sig;
2126 
2127 	mr->sig->psv_memory.psv_idx = psv_index[0];
2128 	mr->sig->psv_wire.psv_idx = psv_index[1];
2129 
2130 	mr->sig->sig_status_checked = true;
2131 	mr->sig->sig_err_exists = false;
2132 	/* Next UMR, Arm SIGERR */
2133 	++mr->sig->sigerr_count;
2134 	mr->klm_mr = mlx5_ib_alloc_pi_mr(pd, max_num_sg, max_num_meta_sg,
2135 					 sizeof(struct mlx5_klm),
2136 					 MLX5_MKC_ACCESS_MODE_KLMS);
2137 	if (IS_ERR(mr->klm_mr)) {
2138 		err = PTR_ERR(mr->klm_mr);
2139 		goto err_destroy_psv;
2140 	}
2141 	mr->mtt_mr = mlx5_ib_alloc_pi_mr(pd, max_num_sg, max_num_meta_sg,
2142 					 sizeof(struct mlx5_mtt),
2143 					 MLX5_MKC_ACCESS_MODE_MTT);
2144 	if (IS_ERR(mr->mtt_mr)) {
2145 		err = PTR_ERR(mr->mtt_mr);
2146 		goto err_free_klm_mr;
2147 	}
2148 
2149 	/* Set bsf descriptors for mkey */
2150 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
2151 	MLX5_SET(mkc, mkc, bsf_en, 1);
2152 	MLX5_SET(mkc, mkc, bsf_octword_size, MLX5_MKEY_BSF_OCTO_SIZE);
2153 
2154 	err = _mlx5_alloc_mkey_descs(pd, mr, 4, sizeof(struct mlx5_klm), 0,
2155 				     MLX5_MKC_ACCESS_MODE_KLMS, in, inlen);
2156 	if (err)
2157 		goto err_free_mtt_mr;
2158 
2159 	err = xa_err(xa_store(&dev->sig_mrs, mlx5_base_mkey(mr->mmkey.key),
2160 			      mr->sig, GFP_KERNEL));
2161 	if (err)
2162 		goto err_free_descs;
2163 	return 0;
2164 
2165 err_free_descs:
2166 	destroy_mkey(dev, mr);
2167 	mlx5_free_priv_descs(mr);
2168 err_free_mtt_mr:
2169 	mlx5_ib_dereg_mr(&mr->mtt_mr->ibmr, NULL);
2170 	mr->mtt_mr = NULL;
2171 err_free_klm_mr:
2172 	mlx5_ib_dereg_mr(&mr->klm_mr->ibmr, NULL);
2173 	mr->klm_mr = NULL;
2174 err_destroy_psv:
2175 	if (mlx5_core_destroy_psv(dev->mdev, mr->sig->psv_memory.psv_idx))
2176 		mlx5_ib_warn(dev, "failed to destroy mem psv %d\n",
2177 			     mr->sig->psv_memory.psv_idx);
2178 	if (mlx5_core_destroy_psv(dev->mdev, mr->sig->psv_wire.psv_idx))
2179 		mlx5_ib_warn(dev, "failed to destroy wire psv %d\n",
2180 			     mr->sig->psv_wire.psv_idx);
2181 err_free_sig:
2182 	kfree(mr->sig);
2183 
2184 	return err;
2185 }
2186 
2187 static struct ib_mr *__mlx5_ib_alloc_mr(struct ib_pd *pd,
2188 					enum ib_mr_type mr_type, u32 max_num_sg,
2189 					u32 max_num_meta_sg)
2190 {
2191 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
2192 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
2193 	int ndescs = ALIGN(max_num_sg, 4);
2194 	struct mlx5_ib_mr *mr;
2195 	u32 *in;
2196 	int err;
2197 
2198 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2199 	if (!mr)
2200 		return ERR_PTR(-ENOMEM);
2201 
2202 	in = kzalloc(inlen, GFP_KERNEL);
2203 	if (!in) {
2204 		err = -ENOMEM;
2205 		goto err_free;
2206 	}
2207 
2208 	mr->ibmr.device = pd->device;
2209 	mr->umem = NULL;
2210 
2211 	switch (mr_type) {
2212 	case IB_MR_TYPE_MEM_REG:
2213 		err = mlx5_alloc_mem_reg_descs(pd, mr, ndescs, in, inlen);
2214 		break;
2215 	case IB_MR_TYPE_SG_GAPS:
2216 		err = mlx5_alloc_sg_gaps_descs(pd, mr, ndescs, in, inlen);
2217 		break;
2218 	case IB_MR_TYPE_INTEGRITY:
2219 		err = mlx5_alloc_integrity_descs(pd, mr, max_num_sg,
2220 						 max_num_meta_sg, in, inlen);
2221 		break;
2222 	default:
2223 		mlx5_ib_warn(dev, "Invalid mr type %d\n", mr_type);
2224 		err = -EINVAL;
2225 	}
2226 
2227 	if (err)
2228 		goto err_free_in;
2229 
2230 	kfree(in);
2231 
2232 	return &mr->ibmr;
2233 
2234 err_free_in:
2235 	kfree(in);
2236 err_free:
2237 	kfree(mr);
2238 	return ERR_PTR(err);
2239 }
2240 
2241 struct ib_mr *mlx5_ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2242 			       u32 max_num_sg)
2243 {
2244 	return __mlx5_ib_alloc_mr(pd, mr_type, max_num_sg, 0);
2245 }
2246 
2247 struct ib_mr *mlx5_ib_alloc_mr_integrity(struct ib_pd *pd,
2248 					 u32 max_num_sg, u32 max_num_meta_sg)
2249 {
2250 	return __mlx5_ib_alloc_mr(pd, IB_MR_TYPE_INTEGRITY, max_num_sg,
2251 				  max_num_meta_sg);
2252 }
2253 
2254 int mlx5_ib_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
2255 {
2256 	struct mlx5_ib_dev *dev = to_mdev(ibmw->device);
2257 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
2258 	struct mlx5_ib_mw *mw = to_mmw(ibmw);
2259 	unsigned int ndescs;
2260 	u32 *in = NULL;
2261 	void *mkc;
2262 	int err;
2263 	struct mlx5_ib_alloc_mw req = {};
2264 	struct {
2265 		__u32	comp_mask;
2266 		__u32	response_length;
2267 	} resp = {};
2268 
2269 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
2270 	if (err)
2271 		return err;
2272 
2273 	if (req.comp_mask || req.reserved1 || req.reserved2)
2274 		return -EOPNOTSUPP;
2275 
2276 	if (udata->inlen > sizeof(req) &&
2277 	    !ib_is_udata_cleared(udata, sizeof(req),
2278 				 udata->inlen - sizeof(req)))
2279 		return -EOPNOTSUPP;
2280 
2281 	ndescs = req.num_klms ? roundup(req.num_klms, 4) : roundup(1, 4);
2282 
2283 	in = kzalloc(inlen, GFP_KERNEL);
2284 	if (!in) {
2285 		err = -ENOMEM;
2286 		goto free;
2287 	}
2288 
2289 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
2290 
2291 	MLX5_SET(mkc, mkc, free, 1);
2292 	MLX5_SET(mkc, mkc, translations_octword_size, ndescs);
2293 	MLX5_SET(mkc, mkc, pd, to_mpd(ibmw->pd)->pdn);
2294 	MLX5_SET(mkc, mkc, umr_en, 1);
2295 	MLX5_SET(mkc, mkc, lr, 1);
2296 	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_KLMS);
2297 	MLX5_SET(mkc, mkc, en_rinval, !!((ibmw->type == IB_MW_TYPE_2)));
2298 	MLX5_SET(mkc, mkc, qpn, 0xffffff);
2299 
2300 	err = mlx5_ib_create_mkey(dev, &mw->mmkey, in, inlen);
2301 	if (err)
2302 		goto free;
2303 
2304 	mw->mmkey.type = MLX5_MKEY_MW;
2305 	ibmw->rkey = mw->mmkey.key;
2306 	mw->mmkey.ndescs = ndescs;
2307 
2308 	resp.response_length =
2309 		min(offsetofend(typeof(resp), response_length), udata->outlen);
2310 	if (resp.response_length) {
2311 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
2312 		if (err)
2313 			goto free_mkey;
2314 	}
2315 
2316 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
2317 		err = mlx5r_store_odp_mkey(dev, &mw->mmkey);
2318 		if (err)
2319 			goto free_mkey;
2320 	}
2321 
2322 	kfree(in);
2323 	return 0;
2324 
2325 free_mkey:
2326 	mlx5_core_destroy_mkey(dev->mdev, mw->mmkey.key);
2327 free:
2328 	kfree(in);
2329 	return err;
2330 }
2331 
2332 int mlx5_ib_dealloc_mw(struct ib_mw *mw)
2333 {
2334 	struct mlx5_ib_dev *dev = to_mdev(mw->device);
2335 	struct mlx5_ib_mw *mmw = to_mmw(mw);
2336 
2337 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) &&
2338 	    xa_erase(&dev->odp_mkeys, mlx5_base_mkey(mmw->mmkey.key)))
2339 		/*
2340 		 * pagefault_single_data_segment() may be accessing mmw
2341 		 * if the user bound an ODP MR to this MW.
2342 		 */
2343 		mlx5r_deref_wait_odp_mkey(&mmw->mmkey);
2344 
2345 	return mlx5_core_destroy_mkey(dev->mdev, mmw->mmkey.key);
2346 }
2347 
2348 int mlx5_ib_check_mr_status(struct ib_mr *ibmr, u32 check_mask,
2349 			    struct ib_mr_status *mr_status)
2350 {
2351 	struct mlx5_ib_mr *mmr = to_mmr(ibmr);
2352 	int ret = 0;
2353 
2354 	if (check_mask & ~IB_MR_CHECK_SIG_STATUS) {
2355 		pr_err("Invalid status check mask\n");
2356 		ret = -EINVAL;
2357 		goto done;
2358 	}
2359 
2360 	mr_status->fail_status = 0;
2361 	if (check_mask & IB_MR_CHECK_SIG_STATUS) {
2362 		if (!mmr->sig) {
2363 			ret = -EINVAL;
2364 			pr_err("signature status check requested on a non-signature enabled MR\n");
2365 			goto done;
2366 		}
2367 
2368 		mmr->sig->sig_status_checked = true;
2369 		if (!mmr->sig->sig_err_exists)
2370 			goto done;
2371 
2372 		if (ibmr->lkey == mmr->sig->err_item.key)
2373 			memcpy(&mr_status->sig_err, &mmr->sig->err_item,
2374 			       sizeof(mr_status->sig_err));
2375 		else {
2376 			mr_status->sig_err.err_type = IB_SIG_BAD_GUARD;
2377 			mr_status->sig_err.sig_err_offset = 0;
2378 			mr_status->sig_err.key = mmr->sig->err_item.key;
2379 		}
2380 
2381 		mmr->sig->sig_err_exists = false;
2382 		mr_status->fail_status |= IB_MR_CHECK_SIG_STATUS;
2383 	}
2384 
2385 done:
2386 	return ret;
2387 }
2388 
2389 static int
2390 mlx5_ib_map_pa_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2391 			int data_sg_nents, unsigned int *data_sg_offset,
2392 			struct scatterlist *meta_sg, int meta_sg_nents,
2393 			unsigned int *meta_sg_offset)
2394 {
2395 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2396 	unsigned int sg_offset = 0;
2397 	int n = 0;
2398 
2399 	mr->meta_length = 0;
2400 	if (data_sg_nents == 1) {
2401 		n++;
2402 		mr->mmkey.ndescs = 1;
2403 		if (data_sg_offset)
2404 			sg_offset = *data_sg_offset;
2405 		mr->data_length = sg_dma_len(data_sg) - sg_offset;
2406 		mr->data_iova = sg_dma_address(data_sg) + sg_offset;
2407 		if (meta_sg_nents == 1) {
2408 			n++;
2409 			mr->meta_ndescs = 1;
2410 			if (meta_sg_offset)
2411 				sg_offset = *meta_sg_offset;
2412 			else
2413 				sg_offset = 0;
2414 			mr->meta_length = sg_dma_len(meta_sg) - sg_offset;
2415 			mr->pi_iova = sg_dma_address(meta_sg) + sg_offset;
2416 		}
2417 		ibmr->length = mr->data_length + mr->meta_length;
2418 	}
2419 
2420 	return n;
2421 }
2422 
2423 static int
2424 mlx5_ib_sg_to_klms(struct mlx5_ib_mr *mr,
2425 		   struct scatterlist *sgl,
2426 		   unsigned short sg_nents,
2427 		   unsigned int *sg_offset_p,
2428 		   struct scatterlist *meta_sgl,
2429 		   unsigned short meta_sg_nents,
2430 		   unsigned int *meta_sg_offset_p)
2431 {
2432 	struct scatterlist *sg = sgl;
2433 	struct mlx5_klm *klms = mr->descs;
2434 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2435 	u32 lkey = mr->ibmr.pd->local_dma_lkey;
2436 	int i, j = 0;
2437 
2438 	mr->ibmr.iova = sg_dma_address(sg) + sg_offset;
2439 	mr->ibmr.length = 0;
2440 
2441 	for_each_sg(sgl, sg, sg_nents, i) {
2442 		if (unlikely(i >= mr->max_descs))
2443 			break;
2444 		klms[i].va = cpu_to_be64(sg_dma_address(sg) + sg_offset);
2445 		klms[i].bcount = cpu_to_be32(sg_dma_len(sg) - sg_offset);
2446 		klms[i].key = cpu_to_be32(lkey);
2447 		mr->ibmr.length += sg_dma_len(sg) - sg_offset;
2448 
2449 		sg_offset = 0;
2450 	}
2451 
2452 	if (sg_offset_p)
2453 		*sg_offset_p = sg_offset;
2454 
2455 	mr->mmkey.ndescs = i;
2456 	mr->data_length = mr->ibmr.length;
2457 
2458 	if (meta_sg_nents) {
2459 		sg = meta_sgl;
2460 		sg_offset = meta_sg_offset_p ? *meta_sg_offset_p : 0;
2461 		for_each_sg(meta_sgl, sg, meta_sg_nents, j) {
2462 			if (unlikely(i + j >= mr->max_descs))
2463 				break;
2464 			klms[i + j].va = cpu_to_be64(sg_dma_address(sg) +
2465 						     sg_offset);
2466 			klms[i + j].bcount = cpu_to_be32(sg_dma_len(sg) -
2467 							 sg_offset);
2468 			klms[i + j].key = cpu_to_be32(lkey);
2469 			mr->ibmr.length += sg_dma_len(sg) - sg_offset;
2470 
2471 			sg_offset = 0;
2472 		}
2473 		if (meta_sg_offset_p)
2474 			*meta_sg_offset_p = sg_offset;
2475 
2476 		mr->meta_ndescs = j;
2477 		mr->meta_length = mr->ibmr.length - mr->data_length;
2478 	}
2479 
2480 	return i + j;
2481 }
2482 
2483 static int mlx5_set_page(struct ib_mr *ibmr, u64 addr)
2484 {
2485 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2486 	__be64 *descs;
2487 
2488 	if (unlikely(mr->mmkey.ndescs == mr->max_descs))
2489 		return -ENOMEM;
2490 
2491 	descs = mr->descs;
2492 	descs[mr->mmkey.ndescs++] = cpu_to_be64(addr | MLX5_EN_RD | MLX5_EN_WR);
2493 
2494 	return 0;
2495 }
2496 
2497 static int mlx5_set_page_pi(struct ib_mr *ibmr, u64 addr)
2498 {
2499 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2500 	__be64 *descs;
2501 
2502 	if (unlikely(mr->mmkey.ndescs + mr->meta_ndescs == mr->max_descs))
2503 		return -ENOMEM;
2504 
2505 	descs = mr->descs;
2506 	descs[mr->mmkey.ndescs + mr->meta_ndescs++] =
2507 		cpu_to_be64(addr | MLX5_EN_RD | MLX5_EN_WR);
2508 
2509 	return 0;
2510 }
2511 
2512 static int
2513 mlx5_ib_map_mtt_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2514 			 int data_sg_nents, unsigned int *data_sg_offset,
2515 			 struct scatterlist *meta_sg, int meta_sg_nents,
2516 			 unsigned int *meta_sg_offset)
2517 {
2518 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2519 	struct mlx5_ib_mr *pi_mr = mr->mtt_mr;
2520 	int n;
2521 
2522 	pi_mr->mmkey.ndescs = 0;
2523 	pi_mr->meta_ndescs = 0;
2524 	pi_mr->meta_length = 0;
2525 
2526 	ib_dma_sync_single_for_cpu(ibmr->device, pi_mr->desc_map,
2527 				   pi_mr->desc_size * pi_mr->max_descs,
2528 				   DMA_TO_DEVICE);
2529 
2530 	pi_mr->ibmr.page_size = ibmr->page_size;
2531 	n = ib_sg_to_pages(&pi_mr->ibmr, data_sg, data_sg_nents, data_sg_offset,
2532 			   mlx5_set_page);
2533 	if (n != data_sg_nents)
2534 		return n;
2535 
2536 	pi_mr->data_iova = pi_mr->ibmr.iova;
2537 	pi_mr->data_length = pi_mr->ibmr.length;
2538 	pi_mr->ibmr.length = pi_mr->data_length;
2539 	ibmr->length = pi_mr->data_length;
2540 
2541 	if (meta_sg_nents) {
2542 		u64 page_mask = ~((u64)ibmr->page_size - 1);
2543 		u64 iova = pi_mr->data_iova;
2544 
2545 		n += ib_sg_to_pages(&pi_mr->ibmr, meta_sg, meta_sg_nents,
2546 				    meta_sg_offset, mlx5_set_page_pi);
2547 
2548 		pi_mr->meta_length = pi_mr->ibmr.length;
2549 		/*
2550 		 * PI address for the HW is the offset of the metadata address
2551 		 * relative to the first data page address.
2552 		 * It equals to first data page address + size of data pages +
2553 		 * metadata offset at the first metadata page
2554 		 */
2555 		pi_mr->pi_iova = (iova & page_mask) +
2556 				 pi_mr->mmkey.ndescs * ibmr->page_size +
2557 				 (pi_mr->ibmr.iova & ~page_mask);
2558 		/*
2559 		 * In order to use one MTT MR for data and metadata, we register
2560 		 * also the gaps between the end of the data and the start of
2561 		 * the metadata (the sig MR will verify that the HW will access
2562 		 * to right addresses). This mapping is safe because we use
2563 		 * internal mkey for the registration.
2564 		 */
2565 		pi_mr->ibmr.length = pi_mr->pi_iova + pi_mr->meta_length - iova;
2566 		pi_mr->ibmr.iova = iova;
2567 		ibmr->length += pi_mr->meta_length;
2568 	}
2569 
2570 	ib_dma_sync_single_for_device(ibmr->device, pi_mr->desc_map,
2571 				      pi_mr->desc_size * pi_mr->max_descs,
2572 				      DMA_TO_DEVICE);
2573 
2574 	return n;
2575 }
2576 
2577 static int
2578 mlx5_ib_map_klm_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2579 			 int data_sg_nents, unsigned int *data_sg_offset,
2580 			 struct scatterlist *meta_sg, int meta_sg_nents,
2581 			 unsigned int *meta_sg_offset)
2582 {
2583 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2584 	struct mlx5_ib_mr *pi_mr = mr->klm_mr;
2585 	int n;
2586 
2587 	pi_mr->mmkey.ndescs = 0;
2588 	pi_mr->meta_ndescs = 0;
2589 	pi_mr->meta_length = 0;
2590 
2591 	ib_dma_sync_single_for_cpu(ibmr->device, pi_mr->desc_map,
2592 				   pi_mr->desc_size * pi_mr->max_descs,
2593 				   DMA_TO_DEVICE);
2594 
2595 	n = mlx5_ib_sg_to_klms(pi_mr, data_sg, data_sg_nents, data_sg_offset,
2596 			       meta_sg, meta_sg_nents, meta_sg_offset);
2597 
2598 	ib_dma_sync_single_for_device(ibmr->device, pi_mr->desc_map,
2599 				      pi_mr->desc_size * pi_mr->max_descs,
2600 				      DMA_TO_DEVICE);
2601 
2602 	/* This is zero-based memory region */
2603 	pi_mr->data_iova = 0;
2604 	pi_mr->ibmr.iova = 0;
2605 	pi_mr->pi_iova = pi_mr->data_length;
2606 	ibmr->length = pi_mr->ibmr.length;
2607 
2608 	return n;
2609 }
2610 
2611 int mlx5_ib_map_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2612 			 int data_sg_nents, unsigned int *data_sg_offset,
2613 			 struct scatterlist *meta_sg, int meta_sg_nents,
2614 			 unsigned int *meta_sg_offset)
2615 {
2616 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2617 	struct mlx5_ib_mr *pi_mr = NULL;
2618 	int n;
2619 
2620 	WARN_ON(ibmr->type != IB_MR_TYPE_INTEGRITY);
2621 
2622 	mr->mmkey.ndescs = 0;
2623 	mr->data_length = 0;
2624 	mr->data_iova = 0;
2625 	mr->meta_ndescs = 0;
2626 	mr->pi_iova = 0;
2627 	/*
2628 	 * As a performance optimization, if possible, there is no need to
2629 	 * perform UMR operation to register the data/metadata buffers.
2630 	 * First try to map the sg lists to PA descriptors with local_dma_lkey.
2631 	 * Fallback to UMR only in case of a failure.
2632 	 */
2633 	n = mlx5_ib_map_pa_mr_sg_pi(ibmr, data_sg, data_sg_nents,
2634 				    data_sg_offset, meta_sg, meta_sg_nents,
2635 				    meta_sg_offset);
2636 	if (n == data_sg_nents + meta_sg_nents)
2637 		goto out;
2638 	/*
2639 	 * As a performance optimization, if possible, there is no need to map
2640 	 * the sg lists to KLM descriptors. First try to map the sg lists to MTT
2641 	 * descriptors and fallback to KLM only in case of a failure.
2642 	 * It's more efficient for the HW to work with MTT descriptors
2643 	 * (especially in high load).
2644 	 * Use KLM (indirect access) only if it's mandatory.
2645 	 */
2646 	pi_mr = mr->mtt_mr;
2647 	n = mlx5_ib_map_mtt_mr_sg_pi(ibmr, data_sg, data_sg_nents,
2648 				     data_sg_offset, meta_sg, meta_sg_nents,
2649 				     meta_sg_offset);
2650 	if (n == data_sg_nents + meta_sg_nents)
2651 		goto out;
2652 
2653 	pi_mr = mr->klm_mr;
2654 	n = mlx5_ib_map_klm_mr_sg_pi(ibmr, data_sg, data_sg_nents,
2655 				     data_sg_offset, meta_sg, meta_sg_nents,
2656 				     meta_sg_offset);
2657 	if (unlikely(n != data_sg_nents + meta_sg_nents))
2658 		return -ENOMEM;
2659 
2660 out:
2661 	/* This is zero-based memory region */
2662 	ibmr->iova = 0;
2663 	mr->pi_mr = pi_mr;
2664 	if (pi_mr)
2665 		ibmr->sig_attrs->meta_length = pi_mr->meta_length;
2666 	else
2667 		ibmr->sig_attrs->meta_length = mr->meta_length;
2668 
2669 	return 0;
2670 }
2671 
2672 int mlx5_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
2673 		      unsigned int *sg_offset)
2674 {
2675 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2676 	int n;
2677 
2678 	mr->mmkey.ndescs = 0;
2679 
2680 	ib_dma_sync_single_for_cpu(ibmr->device, mr->desc_map,
2681 				   mr->desc_size * mr->max_descs,
2682 				   DMA_TO_DEVICE);
2683 
2684 	if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
2685 		n = mlx5_ib_sg_to_klms(mr, sg, sg_nents, sg_offset, NULL, 0,
2686 				       NULL);
2687 	else
2688 		n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset,
2689 				mlx5_set_page);
2690 
2691 	ib_dma_sync_single_for_device(ibmr->device, mr->desc_map,
2692 				      mr->desc_size * mr->max_descs,
2693 				      DMA_TO_DEVICE);
2694 
2695 	return n;
2696 }
2697