xref: /openbmc/linux/drivers/infiniband/hw/mlx5/mr.c (revision 046b212a)
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_odp.h>
43 #include "dm.h"
44 #include "mlx5_ib.h"
45 #include "umr.h"
46 
47 enum {
48 	MAX_PENDING_REG_MR = 8,
49 };
50 
51 #define MLX5_UMR_ALIGN 2048
52 
53 static void
54 create_mkey_callback(int status, struct mlx5_async_work *context);
55 static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem,
56 				     u64 iova, int access_flags,
57 				     unsigned int page_size, bool populate);
58 
59 static void set_mkc_access_pd_addr_fields(void *mkc, int acc, u64 start_addr,
60 					  struct ib_pd *pd)
61 {
62 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
63 
64 	MLX5_SET(mkc, mkc, a, !!(acc & IB_ACCESS_REMOTE_ATOMIC));
65 	MLX5_SET(mkc, mkc, rw, !!(acc & IB_ACCESS_REMOTE_WRITE));
66 	MLX5_SET(mkc, mkc, rr, !!(acc & IB_ACCESS_REMOTE_READ));
67 	MLX5_SET(mkc, mkc, lw, !!(acc & IB_ACCESS_LOCAL_WRITE));
68 	MLX5_SET(mkc, mkc, lr, 1);
69 
70 	if (acc & IB_ACCESS_RELAXED_ORDERING) {
71 		if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write))
72 			MLX5_SET(mkc, mkc, relaxed_ordering_write, 1);
73 
74 		if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read) ||
75 		    (MLX5_CAP_GEN(dev->mdev,
76 				  relaxed_ordering_read_pci_enabled) &&
77 		     pcie_relaxed_ordering_enabled(dev->mdev->pdev)))
78 			MLX5_SET(mkc, mkc, relaxed_ordering_read, 1);
79 	}
80 
81 	MLX5_SET(mkc, mkc, pd, to_mpd(pd)->pdn);
82 	MLX5_SET(mkc, mkc, qpn, 0xffffff);
83 	MLX5_SET64(mkc, mkc, start_addr, start_addr);
84 }
85 
86 static void assign_mkey_variant(struct mlx5_ib_dev *dev, u32 *mkey, u32 *in)
87 {
88 	u8 key = atomic_inc_return(&dev->mkey_var);
89 	void *mkc;
90 
91 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
92 	MLX5_SET(mkc, mkc, mkey_7_0, key);
93 	*mkey = key;
94 }
95 
96 static int mlx5_ib_create_mkey(struct mlx5_ib_dev *dev,
97 			       struct mlx5_ib_mkey *mkey, u32 *in, int inlen)
98 {
99 	int ret;
100 
101 	assign_mkey_variant(dev, &mkey->key, in);
102 	ret = mlx5_core_create_mkey(dev->mdev, &mkey->key, in, inlen);
103 	if (!ret)
104 		init_waitqueue_head(&mkey->wait);
105 
106 	return ret;
107 }
108 
109 static int mlx5_ib_create_mkey_cb(struct mlx5r_async_create_mkey *async_create)
110 {
111 	struct mlx5_ib_dev *dev = async_create->ent->dev;
112 	size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
113 	size_t outlen = MLX5_ST_SZ_BYTES(create_mkey_out);
114 
115 	MLX5_SET(create_mkey_in, async_create->in, opcode,
116 		 MLX5_CMD_OP_CREATE_MKEY);
117 	assign_mkey_variant(dev, &async_create->mkey, async_create->in);
118 	return mlx5_cmd_exec_cb(&dev->async_ctx, async_create->in, inlen,
119 				async_create->out, outlen, create_mkey_callback,
120 				&async_create->cb_work);
121 }
122 
123 static int mkey_cache_max_order(struct mlx5_ib_dev *dev);
124 static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent);
125 
126 static int destroy_mkey(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
127 {
128 	WARN_ON(xa_load(&dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key)));
129 
130 	return mlx5_core_destroy_mkey(dev->mdev, mr->mmkey.key);
131 }
132 
133 static void create_mkey_warn(struct mlx5_ib_dev *dev, int status, void *out)
134 {
135 	if (status == -ENXIO) /* core driver is not available */
136 		return;
137 
138 	mlx5_ib_warn(dev, "async reg mr failed. status %d\n", status);
139 	if (status != -EREMOTEIO) /* driver specific failure */
140 		return;
141 
142 	/* Failed in FW, print cmd out failure details */
143 	mlx5_cmd_out_err(dev->mdev, MLX5_CMD_OP_CREATE_MKEY, 0, out);
144 }
145 
146 static int push_mkey_locked(struct mlx5_cache_ent *ent, bool limit_pendings,
147 			    void *to_store)
148 {
149 	XA_STATE(xas, &ent->mkeys, 0);
150 	void *curr;
151 
152 	if (limit_pendings &&
153 	    (ent->reserved - ent->stored) > MAX_PENDING_REG_MR)
154 		return -EAGAIN;
155 
156 	while (1) {
157 		/*
158 		 * This is cmpxchg (NULL, XA_ZERO_ENTRY) however this version
159 		 * doesn't transparently unlock. Instead we set the xas index to
160 		 * the current value of reserved every iteration.
161 		 */
162 		xas_set(&xas, ent->reserved);
163 		curr = xas_load(&xas);
164 		if (!curr) {
165 			if (to_store && ent->stored == ent->reserved)
166 				xas_store(&xas, to_store);
167 			else
168 				xas_store(&xas, XA_ZERO_ENTRY);
169 			if (xas_valid(&xas)) {
170 				ent->reserved++;
171 				if (to_store) {
172 					if (ent->stored != ent->reserved)
173 						__xa_store(&ent->mkeys,
174 							   ent->stored,
175 							   to_store,
176 							   GFP_KERNEL);
177 					ent->stored++;
178 					queue_adjust_cache_locked(ent);
179 					WRITE_ONCE(ent->dev->cache.last_add,
180 						   jiffies);
181 				}
182 			}
183 		}
184 		xa_unlock_irq(&ent->mkeys);
185 
186 		/*
187 		 * Notice xas_nomem() must always be called as it cleans
188 		 * up any cached allocation.
189 		 */
190 		if (!xas_nomem(&xas, GFP_KERNEL))
191 			break;
192 		xa_lock_irq(&ent->mkeys);
193 	}
194 	xa_lock_irq(&ent->mkeys);
195 	if (xas_error(&xas))
196 		return xas_error(&xas);
197 	if (WARN_ON(curr))
198 		return -EINVAL;
199 	return 0;
200 }
201 
202 static int push_mkey(struct mlx5_cache_ent *ent, bool limit_pendings,
203 		     void *to_store)
204 {
205 	int ret;
206 
207 	xa_lock_irq(&ent->mkeys);
208 	ret = push_mkey_locked(ent, limit_pendings, to_store);
209 	xa_unlock_irq(&ent->mkeys);
210 	return ret;
211 }
212 
213 static void undo_push_reserve_mkey(struct mlx5_cache_ent *ent)
214 {
215 	void *old;
216 
217 	ent->reserved--;
218 	old = __xa_erase(&ent->mkeys, ent->reserved);
219 	WARN_ON(old);
220 }
221 
222 static void push_to_reserved(struct mlx5_cache_ent *ent, u32 mkey)
223 {
224 	void *old;
225 
226 	old = __xa_store(&ent->mkeys, ent->stored, xa_mk_value(mkey), 0);
227 	WARN_ON(old);
228 	ent->stored++;
229 }
230 
231 static u32 pop_stored_mkey(struct mlx5_cache_ent *ent)
232 {
233 	void *old, *xa_mkey;
234 
235 	ent->stored--;
236 	ent->reserved--;
237 
238 	if (ent->stored == ent->reserved) {
239 		xa_mkey = __xa_erase(&ent->mkeys, ent->stored);
240 		WARN_ON(!xa_mkey);
241 		return (u32)xa_to_value(xa_mkey);
242 	}
243 
244 	xa_mkey = __xa_store(&ent->mkeys, ent->stored, XA_ZERO_ENTRY,
245 			     GFP_KERNEL);
246 	WARN_ON(!xa_mkey || xa_is_err(xa_mkey));
247 	old = __xa_erase(&ent->mkeys, ent->reserved);
248 	WARN_ON(old);
249 	return (u32)xa_to_value(xa_mkey);
250 }
251 
252 static void create_mkey_callback(int status, struct mlx5_async_work *context)
253 {
254 	struct mlx5r_async_create_mkey *mkey_out =
255 		container_of(context, struct mlx5r_async_create_mkey, cb_work);
256 	struct mlx5_cache_ent *ent = mkey_out->ent;
257 	struct mlx5_ib_dev *dev = ent->dev;
258 	unsigned long flags;
259 
260 	if (status) {
261 		create_mkey_warn(dev, status, mkey_out->out);
262 		kfree(mkey_out);
263 		xa_lock_irqsave(&ent->mkeys, flags);
264 		undo_push_reserve_mkey(ent);
265 		WRITE_ONCE(dev->fill_delay, 1);
266 		xa_unlock_irqrestore(&ent->mkeys, flags);
267 		mod_timer(&dev->delay_timer, jiffies + HZ);
268 		return;
269 	}
270 
271 	mkey_out->mkey |= mlx5_idx_to_mkey(
272 		MLX5_GET(create_mkey_out, mkey_out->out, mkey_index));
273 	WRITE_ONCE(dev->cache.last_add, jiffies);
274 
275 	xa_lock_irqsave(&ent->mkeys, flags);
276 	push_to_reserved(ent, mkey_out->mkey);
277 	/* If we are doing fill_to_high_water then keep going. */
278 	queue_adjust_cache_locked(ent);
279 	xa_unlock_irqrestore(&ent->mkeys, flags);
280 	kfree(mkey_out);
281 }
282 
283 static int get_mkc_octo_size(unsigned int access_mode, unsigned int ndescs)
284 {
285 	int ret = 0;
286 
287 	switch (access_mode) {
288 	case MLX5_MKC_ACCESS_MODE_MTT:
289 		ret = DIV_ROUND_UP(ndescs, MLX5_IB_UMR_OCTOWORD /
290 						   sizeof(struct mlx5_mtt));
291 		break;
292 	case MLX5_MKC_ACCESS_MODE_KSM:
293 		ret = DIV_ROUND_UP(ndescs, MLX5_IB_UMR_OCTOWORD /
294 						   sizeof(struct mlx5_klm));
295 		break;
296 	default:
297 		WARN_ON(1);
298 	}
299 	return ret;
300 }
301 
302 static void set_cache_mkc(struct mlx5_cache_ent *ent, void *mkc)
303 {
304 	set_mkc_access_pd_addr_fields(mkc, 0, 0, ent->dev->umrc.pd);
305 	MLX5_SET(mkc, mkc, free, 1);
306 	MLX5_SET(mkc, mkc, umr_en, 1);
307 	MLX5_SET(mkc, mkc, access_mode_1_0, ent->rb_key.access_mode & 0x3);
308 	MLX5_SET(mkc, mkc, access_mode_4_2,
309 		(ent->rb_key.access_mode >> 2) & 0x7);
310 
311 	MLX5_SET(mkc, mkc, translations_octword_size,
312 		 get_mkc_octo_size(ent->rb_key.access_mode,
313 				   ent->rb_key.ndescs));
314 	MLX5_SET(mkc, mkc, log_page_size, PAGE_SHIFT);
315 }
316 
317 /* Asynchronously schedule new MRs to be populated in the cache. */
318 static int add_keys(struct mlx5_cache_ent *ent, unsigned int num)
319 {
320 	struct mlx5r_async_create_mkey *async_create;
321 	void *mkc;
322 	int err = 0;
323 	int i;
324 
325 	for (i = 0; i < num; i++) {
326 		async_create = kzalloc(sizeof(struct mlx5r_async_create_mkey),
327 				       GFP_KERNEL);
328 		if (!async_create)
329 			return -ENOMEM;
330 		mkc = MLX5_ADDR_OF(create_mkey_in, async_create->in,
331 				   memory_key_mkey_entry);
332 		set_cache_mkc(ent, mkc);
333 		async_create->ent = ent;
334 
335 		err = push_mkey(ent, true, NULL);
336 		if (err)
337 			goto free_async_create;
338 
339 		err = mlx5_ib_create_mkey_cb(async_create);
340 		if (err) {
341 			mlx5_ib_warn(ent->dev, "create mkey failed %d\n", err);
342 			goto err_undo_reserve;
343 		}
344 	}
345 
346 	return 0;
347 
348 err_undo_reserve:
349 	xa_lock_irq(&ent->mkeys);
350 	undo_push_reserve_mkey(ent);
351 	xa_unlock_irq(&ent->mkeys);
352 free_async_create:
353 	kfree(async_create);
354 	return err;
355 }
356 
357 /* Synchronously create a MR in the cache */
358 static int create_cache_mkey(struct mlx5_cache_ent *ent, u32 *mkey)
359 {
360 	size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
361 	void *mkc;
362 	u32 *in;
363 	int err;
364 
365 	in = kzalloc(inlen, GFP_KERNEL);
366 	if (!in)
367 		return -ENOMEM;
368 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
369 	set_cache_mkc(ent, mkc);
370 
371 	err = mlx5_core_create_mkey(ent->dev->mdev, mkey, in, inlen);
372 	if (err)
373 		goto free_in;
374 
375 	WRITE_ONCE(ent->dev->cache.last_add, jiffies);
376 free_in:
377 	kfree(in);
378 	return err;
379 }
380 
381 static void remove_cache_mr_locked(struct mlx5_cache_ent *ent)
382 {
383 	u32 mkey;
384 
385 	lockdep_assert_held(&ent->mkeys.xa_lock);
386 	if (!ent->stored)
387 		return;
388 	mkey = pop_stored_mkey(ent);
389 	xa_unlock_irq(&ent->mkeys);
390 	mlx5_core_destroy_mkey(ent->dev->mdev, mkey);
391 	xa_lock_irq(&ent->mkeys);
392 }
393 
394 static int resize_available_mrs(struct mlx5_cache_ent *ent, unsigned int target,
395 				bool limit_fill)
396 	 __acquires(&ent->mkeys) __releases(&ent->mkeys)
397 {
398 	int err;
399 
400 	lockdep_assert_held(&ent->mkeys.xa_lock);
401 
402 	while (true) {
403 		if (limit_fill)
404 			target = ent->limit * 2;
405 		if (target == ent->reserved)
406 			return 0;
407 		if (target > ent->reserved) {
408 			u32 todo = target - ent->reserved;
409 
410 			xa_unlock_irq(&ent->mkeys);
411 			err = add_keys(ent, todo);
412 			if (err == -EAGAIN)
413 				usleep_range(3000, 5000);
414 			xa_lock_irq(&ent->mkeys);
415 			if (err) {
416 				if (err != -EAGAIN)
417 					return err;
418 			} else
419 				return 0;
420 		} else {
421 			remove_cache_mr_locked(ent);
422 		}
423 	}
424 }
425 
426 static ssize_t size_write(struct file *filp, const char __user *buf,
427 			  size_t count, loff_t *pos)
428 {
429 	struct mlx5_cache_ent *ent = filp->private_data;
430 	u32 target;
431 	int err;
432 
433 	err = kstrtou32_from_user(buf, count, 0, &target);
434 	if (err)
435 		return err;
436 
437 	/*
438 	 * Target is the new value of total_mrs the user requests, however we
439 	 * cannot free MRs that are in use. Compute the target value for stored
440 	 * mkeys.
441 	 */
442 	xa_lock_irq(&ent->mkeys);
443 	if (target < ent->in_use) {
444 		err = -EINVAL;
445 		goto err_unlock;
446 	}
447 	target = target - ent->in_use;
448 	if (target < ent->limit || target > ent->limit*2) {
449 		err = -EINVAL;
450 		goto err_unlock;
451 	}
452 	err = resize_available_mrs(ent, target, false);
453 	if (err)
454 		goto err_unlock;
455 	xa_unlock_irq(&ent->mkeys);
456 
457 	return count;
458 
459 err_unlock:
460 	xa_unlock_irq(&ent->mkeys);
461 	return err;
462 }
463 
464 static ssize_t size_read(struct file *filp, char __user *buf, size_t count,
465 			 loff_t *pos)
466 {
467 	struct mlx5_cache_ent *ent = filp->private_data;
468 	char lbuf[20];
469 	int err;
470 
471 	err = snprintf(lbuf, sizeof(lbuf), "%ld\n", ent->stored + ent->in_use);
472 	if (err < 0)
473 		return err;
474 
475 	return simple_read_from_buffer(buf, count, pos, lbuf, err);
476 }
477 
478 static const struct file_operations size_fops = {
479 	.owner	= THIS_MODULE,
480 	.open	= simple_open,
481 	.write	= size_write,
482 	.read	= size_read,
483 };
484 
485 static ssize_t limit_write(struct file *filp, const char __user *buf,
486 			   size_t count, loff_t *pos)
487 {
488 	struct mlx5_cache_ent *ent = filp->private_data;
489 	u32 var;
490 	int err;
491 
492 	err = kstrtou32_from_user(buf, count, 0, &var);
493 	if (err)
494 		return err;
495 
496 	/*
497 	 * Upon set we immediately fill the cache to high water mark implied by
498 	 * the limit.
499 	 */
500 	xa_lock_irq(&ent->mkeys);
501 	ent->limit = var;
502 	err = resize_available_mrs(ent, 0, true);
503 	xa_unlock_irq(&ent->mkeys);
504 	if (err)
505 		return err;
506 	return count;
507 }
508 
509 static ssize_t limit_read(struct file *filp, char __user *buf, size_t count,
510 			  loff_t *pos)
511 {
512 	struct mlx5_cache_ent *ent = filp->private_data;
513 	char lbuf[20];
514 	int err;
515 
516 	err = snprintf(lbuf, sizeof(lbuf), "%d\n", ent->limit);
517 	if (err < 0)
518 		return err;
519 
520 	return simple_read_from_buffer(buf, count, pos, lbuf, err);
521 }
522 
523 static const struct file_operations limit_fops = {
524 	.owner	= THIS_MODULE,
525 	.open	= simple_open,
526 	.write	= limit_write,
527 	.read	= limit_read,
528 };
529 
530 static bool someone_adding(struct mlx5_mkey_cache *cache)
531 {
532 	struct mlx5_cache_ent *ent;
533 	struct rb_node *node;
534 	bool ret;
535 
536 	mutex_lock(&cache->rb_lock);
537 	for (node = rb_first(&cache->rb_root); node; node = rb_next(node)) {
538 		ent = rb_entry(node, struct mlx5_cache_ent, node);
539 		xa_lock_irq(&ent->mkeys);
540 		ret = ent->stored < ent->limit;
541 		xa_unlock_irq(&ent->mkeys);
542 		if (ret) {
543 			mutex_unlock(&cache->rb_lock);
544 			return true;
545 		}
546 	}
547 	mutex_unlock(&cache->rb_lock);
548 	return false;
549 }
550 
551 /*
552  * Check if the bucket is outside the high/low water mark and schedule an async
553  * update. The cache refill has hysteresis, once the low water mark is hit it is
554  * refilled up to the high mark.
555  */
556 static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent)
557 {
558 	lockdep_assert_held(&ent->mkeys.xa_lock);
559 
560 	if (ent->disabled || READ_ONCE(ent->dev->fill_delay) || ent->is_tmp)
561 		return;
562 	if (ent->stored < ent->limit) {
563 		ent->fill_to_high_water = true;
564 		mod_delayed_work(ent->dev->cache.wq, &ent->dwork, 0);
565 	} else if (ent->fill_to_high_water &&
566 		   ent->reserved < 2 * ent->limit) {
567 		/*
568 		 * Once we start populating due to hitting a low water mark
569 		 * continue until we pass the high water mark.
570 		 */
571 		mod_delayed_work(ent->dev->cache.wq, &ent->dwork, 0);
572 	} else if (ent->stored == 2 * ent->limit) {
573 		ent->fill_to_high_water = false;
574 	} else if (ent->stored > 2 * ent->limit) {
575 		/* Queue deletion of excess entries */
576 		ent->fill_to_high_water = false;
577 		if (ent->stored != ent->reserved)
578 			queue_delayed_work(ent->dev->cache.wq, &ent->dwork,
579 					   msecs_to_jiffies(1000));
580 		else
581 			mod_delayed_work(ent->dev->cache.wq, &ent->dwork, 0);
582 	}
583 }
584 
585 static void __cache_work_func(struct mlx5_cache_ent *ent)
586 {
587 	struct mlx5_ib_dev *dev = ent->dev;
588 	struct mlx5_mkey_cache *cache = &dev->cache;
589 	int err;
590 
591 	xa_lock_irq(&ent->mkeys);
592 	if (ent->disabled)
593 		goto out;
594 
595 	if (ent->fill_to_high_water && ent->reserved < 2 * ent->limit &&
596 	    !READ_ONCE(dev->fill_delay)) {
597 		xa_unlock_irq(&ent->mkeys);
598 		err = add_keys(ent, 1);
599 		xa_lock_irq(&ent->mkeys);
600 		if (ent->disabled)
601 			goto out;
602 		if (err) {
603 			/*
604 			 * EAGAIN only happens if there are pending MRs, so we
605 			 * will be rescheduled when storing them. The only
606 			 * failure path here is ENOMEM.
607 			 */
608 			if (err != -EAGAIN) {
609 				mlx5_ib_warn(
610 					dev,
611 					"add keys command failed, err %d\n",
612 					err);
613 				queue_delayed_work(cache->wq, &ent->dwork,
614 						   msecs_to_jiffies(1000));
615 			}
616 		}
617 	} else if (ent->stored > 2 * ent->limit) {
618 		bool need_delay;
619 
620 		/*
621 		 * The remove_cache_mr() logic is performed as garbage
622 		 * collection task. Such task is intended to be run when no
623 		 * other active processes are running.
624 		 *
625 		 * The need_resched() will return TRUE if there are user tasks
626 		 * to be activated in near future.
627 		 *
628 		 * In such case, we don't execute remove_cache_mr() and postpone
629 		 * the garbage collection work to try to run in next cycle, in
630 		 * order to free CPU resources to other tasks.
631 		 */
632 		xa_unlock_irq(&ent->mkeys);
633 		need_delay = need_resched() || someone_adding(cache) ||
634 			     !time_after(jiffies,
635 					 READ_ONCE(cache->last_add) + 300 * HZ);
636 		xa_lock_irq(&ent->mkeys);
637 		if (ent->disabled)
638 			goto out;
639 		if (need_delay) {
640 			queue_delayed_work(cache->wq, &ent->dwork, 300 * HZ);
641 			goto out;
642 		}
643 		remove_cache_mr_locked(ent);
644 		queue_adjust_cache_locked(ent);
645 	}
646 out:
647 	xa_unlock_irq(&ent->mkeys);
648 }
649 
650 static void delayed_cache_work_func(struct work_struct *work)
651 {
652 	struct mlx5_cache_ent *ent;
653 
654 	ent = container_of(work, struct mlx5_cache_ent, dwork.work);
655 	__cache_work_func(ent);
656 }
657 
658 static int cache_ent_key_cmp(struct mlx5r_cache_rb_key key1,
659 			     struct mlx5r_cache_rb_key key2)
660 {
661 	int res;
662 
663 	res = key1.ats - key2.ats;
664 	if (res)
665 		return res;
666 
667 	res = key1.access_mode - key2.access_mode;
668 	if (res)
669 		return res;
670 
671 	res = key1.access_flags - key2.access_flags;
672 	if (res)
673 		return res;
674 
675 	/*
676 	 * keep ndescs the last in the compare table since the find function
677 	 * searches for an exact match on all properties and only closest
678 	 * match in size.
679 	 */
680 	return key1.ndescs - key2.ndescs;
681 }
682 
683 static int mlx5_cache_ent_insert(struct mlx5_mkey_cache *cache,
684 				 struct mlx5_cache_ent *ent)
685 {
686 	struct rb_node **new = &cache->rb_root.rb_node, *parent = NULL;
687 	struct mlx5_cache_ent *cur;
688 	int cmp;
689 
690 	/* Figure out where to put new node */
691 	while (*new) {
692 		cur = rb_entry(*new, struct mlx5_cache_ent, node);
693 		parent = *new;
694 		cmp = cache_ent_key_cmp(cur->rb_key, ent->rb_key);
695 		if (cmp > 0)
696 			new = &((*new)->rb_left);
697 		if (cmp < 0)
698 			new = &((*new)->rb_right);
699 		if (cmp == 0) {
700 			mutex_unlock(&cache->rb_lock);
701 			return -EEXIST;
702 		}
703 	}
704 
705 	/* Add new node and rebalance tree. */
706 	rb_link_node(&ent->node, parent, new);
707 	rb_insert_color(&ent->node, &cache->rb_root);
708 
709 	return 0;
710 }
711 
712 static struct mlx5_cache_ent *
713 mkey_cache_ent_from_rb_key(struct mlx5_ib_dev *dev,
714 			   struct mlx5r_cache_rb_key rb_key)
715 {
716 	struct rb_node *node = dev->cache.rb_root.rb_node;
717 	struct mlx5_cache_ent *cur, *smallest = NULL;
718 	int cmp;
719 
720 	/*
721 	 * Find the smallest ent with order >= requested_order.
722 	 */
723 	while (node) {
724 		cur = rb_entry(node, struct mlx5_cache_ent, node);
725 		cmp = cache_ent_key_cmp(cur->rb_key, rb_key);
726 		if (cmp > 0) {
727 			smallest = cur;
728 			node = node->rb_left;
729 		}
730 		if (cmp < 0)
731 			node = node->rb_right;
732 		if (cmp == 0)
733 			return cur;
734 	}
735 
736 	return (smallest &&
737 		smallest->rb_key.access_mode == rb_key.access_mode &&
738 		smallest->rb_key.access_flags == rb_key.access_flags &&
739 		smallest->rb_key.ats == rb_key.ats) ?
740 		       smallest :
741 		       NULL;
742 }
743 
744 static struct mlx5_ib_mr *_mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
745 					struct mlx5_cache_ent *ent,
746 					int access_flags)
747 {
748 	struct mlx5_ib_mr *mr;
749 	int err;
750 
751 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
752 	if (!mr)
753 		return ERR_PTR(-ENOMEM);
754 
755 	xa_lock_irq(&ent->mkeys);
756 	ent->in_use++;
757 
758 	if (!ent->stored) {
759 		queue_adjust_cache_locked(ent);
760 		ent->miss++;
761 		xa_unlock_irq(&ent->mkeys);
762 		err = create_cache_mkey(ent, &mr->mmkey.key);
763 		if (err) {
764 			xa_lock_irq(&ent->mkeys);
765 			ent->in_use--;
766 			xa_unlock_irq(&ent->mkeys);
767 			kfree(mr);
768 			return ERR_PTR(err);
769 		}
770 	} else {
771 		mr->mmkey.key = pop_stored_mkey(ent);
772 		queue_adjust_cache_locked(ent);
773 		xa_unlock_irq(&ent->mkeys);
774 	}
775 	mr->mmkey.cache_ent = ent;
776 	mr->mmkey.type = MLX5_MKEY_MR;
777 	init_waitqueue_head(&mr->mmkey.wait);
778 	return mr;
779 }
780 
781 static int get_unchangeable_access_flags(struct mlx5_ib_dev *dev,
782 					 int access_flags)
783 {
784 	int ret = 0;
785 
786 	if ((access_flags & IB_ACCESS_REMOTE_ATOMIC) &&
787 	    MLX5_CAP_GEN(dev->mdev, atomic) &&
788 	    MLX5_CAP_GEN(dev->mdev, umr_modify_atomic_disabled))
789 		ret |= IB_ACCESS_REMOTE_ATOMIC;
790 
791 	if ((access_flags & IB_ACCESS_RELAXED_ORDERING) &&
792 	    MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write) &&
793 	    !MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write_umr))
794 		ret |= IB_ACCESS_RELAXED_ORDERING;
795 
796 	if ((access_flags & IB_ACCESS_RELAXED_ORDERING) &&
797 	    (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read) ||
798 	     MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read_pci_enabled)) &&
799 	    !MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read_umr))
800 		ret |= IB_ACCESS_RELAXED_ORDERING;
801 
802 	return ret;
803 }
804 
805 struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
806 				       int access_flags, int access_mode,
807 				       int ndescs)
808 {
809 	struct mlx5r_cache_rb_key rb_key = {
810 		.ndescs = ndescs,
811 		.access_mode = access_mode,
812 		.access_flags = get_unchangeable_access_flags(dev, access_flags)
813 	};
814 	struct mlx5_cache_ent *ent = mkey_cache_ent_from_rb_key(dev, rb_key);
815 
816 	if (!ent)
817 		return ERR_PTR(-EOPNOTSUPP);
818 
819 	return _mlx5_mr_cache_alloc(dev, ent, access_flags);
820 }
821 
822 static void clean_keys(struct mlx5_ib_dev *dev, struct mlx5_cache_ent *ent)
823 {
824 	u32 mkey;
825 
826 	cancel_delayed_work(&ent->dwork);
827 	xa_lock_irq(&ent->mkeys);
828 	while (ent->stored) {
829 		mkey = pop_stored_mkey(ent);
830 		xa_unlock_irq(&ent->mkeys);
831 		mlx5_core_destroy_mkey(dev->mdev, mkey);
832 		xa_lock_irq(&ent->mkeys);
833 	}
834 	xa_unlock_irq(&ent->mkeys);
835 }
836 
837 static void mlx5_mkey_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
838 {
839 	if (!mlx5_debugfs_root || dev->is_rep)
840 		return;
841 
842 	debugfs_remove_recursive(dev->cache.fs_root);
843 	dev->cache.fs_root = NULL;
844 }
845 
846 static void mlx5_mkey_cache_debugfs_add_ent(struct mlx5_ib_dev *dev,
847 					    struct mlx5_cache_ent *ent)
848 {
849 	int order = order_base_2(ent->rb_key.ndescs);
850 	struct dentry *dir;
851 
852 	if (!mlx5_debugfs_root || dev->is_rep)
853 		return;
854 
855 	if (ent->rb_key.access_mode == MLX5_MKC_ACCESS_MODE_KSM)
856 		order = MLX5_IMR_KSM_CACHE_ENTRY + 2;
857 
858 	sprintf(ent->name, "%d", order);
859 	dir = debugfs_create_dir(ent->name, dev->cache.fs_root);
860 	debugfs_create_file("size", 0600, dir, ent, &size_fops);
861 	debugfs_create_file("limit", 0600, dir, ent, &limit_fops);
862 	debugfs_create_ulong("cur", 0400, dir, &ent->stored);
863 	debugfs_create_u32("miss", 0600, dir, &ent->miss);
864 }
865 
866 static void mlx5_mkey_cache_debugfs_init(struct mlx5_ib_dev *dev)
867 {
868 	struct dentry *dbg_root = mlx5_debugfs_get_dev_root(dev->mdev);
869 	struct mlx5_mkey_cache *cache = &dev->cache;
870 
871 	if (!mlx5_debugfs_root || dev->is_rep)
872 		return;
873 
874 	cache->fs_root = debugfs_create_dir("mr_cache", dbg_root);
875 }
876 
877 static void delay_time_func(struct timer_list *t)
878 {
879 	struct mlx5_ib_dev *dev = from_timer(dev, t, delay_timer);
880 
881 	WRITE_ONCE(dev->fill_delay, 0);
882 }
883 
884 struct mlx5_cache_ent *
885 mlx5r_cache_create_ent_locked(struct mlx5_ib_dev *dev,
886 			      struct mlx5r_cache_rb_key rb_key,
887 			      bool persistent_entry)
888 {
889 	struct mlx5_cache_ent *ent;
890 	int order;
891 	int ret;
892 
893 	ent = kzalloc(sizeof(*ent), GFP_KERNEL);
894 	if (!ent)
895 		return ERR_PTR(-ENOMEM);
896 
897 	xa_init_flags(&ent->mkeys, XA_FLAGS_LOCK_IRQ);
898 	ent->rb_key = rb_key;
899 	ent->dev = dev;
900 	ent->is_tmp = !persistent_entry;
901 
902 	INIT_DELAYED_WORK(&ent->dwork, delayed_cache_work_func);
903 
904 	ret = mlx5_cache_ent_insert(&dev->cache, ent);
905 	if (ret) {
906 		kfree(ent);
907 		return ERR_PTR(ret);
908 	}
909 
910 	if (persistent_entry) {
911 		if (rb_key.access_mode == MLX5_MKC_ACCESS_MODE_KSM)
912 			order = MLX5_IMR_KSM_CACHE_ENTRY;
913 		else
914 			order = order_base_2(rb_key.ndescs) - 2;
915 
916 		if ((dev->mdev->profile.mask & MLX5_PROF_MASK_MR_CACHE) &&
917 		    !dev->is_rep && mlx5_core_is_pf(dev->mdev) &&
918 		    mlx5r_umr_can_load_pas(dev, 0))
919 			ent->limit = dev->mdev->profile.mr_cache[order].limit;
920 		else
921 			ent->limit = 0;
922 
923 		mlx5_mkey_cache_debugfs_add_ent(dev, ent);
924 	} else {
925 		mod_delayed_work(ent->dev->cache.wq,
926 				 &ent->dev->cache.remove_ent_dwork,
927 				 msecs_to_jiffies(30 * 1000));
928 	}
929 
930 	return ent;
931 }
932 
933 static void remove_ent_work_func(struct work_struct *work)
934 {
935 	struct mlx5_mkey_cache *cache;
936 	struct mlx5_cache_ent *ent;
937 	struct rb_node *cur;
938 
939 	cache = container_of(work, struct mlx5_mkey_cache,
940 			     remove_ent_dwork.work);
941 	mutex_lock(&cache->rb_lock);
942 	cur = rb_last(&cache->rb_root);
943 	while (cur) {
944 		ent = rb_entry(cur, struct mlx5_cache_ent, node);
945 		cur = rb_prev(cur);
946 		mutex_unlock(&cache->rb_lock);
947 
948 		xa_lock_irq(&ent->mkeys);
949 		if (!ent->is_tmp) {
950 			xa_unlock_irq(&ent->mkeys);
951 			mutex_lock(&cache->rb_lock);
952 			continue;
953 		}
954 		xa_unlock_irq(&ent->mkeys);
955 
956 		clean_keys(ent->dev, ent);
957 		mutex_lock(&cache->rb_lock);
958 	}
959 	mutex_unlock(&cache->rb_lock);
960 }
961 
962 int mlx5_mkey_cache_init(struct mlx5_ib_dev *dev)
963 {
964 	struct mlx5_mkey_cache *cache = &dev->cache;
965 	struct rb_root *root = &dev->cache.rb_root;
966 	struct mlx5r_cache_rb_key rb_key = {
967 		.access_mode = MLX5_MKC_ACCESS_MODE_MTT,
968 	};
969 	struct mlx5_cache_ent *ent;
970 	struct rb_node *node;
971 	int ret;
972 	int i;
973 
974 	mutex_init(&dev->slow_path_mutex);
975 	mutex_init(&dev->cache.rb_lock);
976 	dev->cache.rb_root = RB_ROOT;
977 	INIT_DELAYED_WORK(&dev->cache.remove_ent_dwork, remove_ent_work_func);
978 	cache->wq = alloc_ordered_workqueue("mkey_cache", WQ_MEM_RECLAIM);
979 	if (!cache->wq) {
980 		mlx5_ib_warn(dev, "failed to create work queue\n");
981 		return -ENOMEM;
982 	}
983 
984 	mlx5_cmd_init_async_ctx(dev->mdev, &dev->async_ctx);
985 	timer_setup(&dev->delay_timer, delay_time_func, 0);
986 	mlx5_mkey_cache_debugfs_init(dev);
987 	mutex_lock(&cache->rb_lock);
988 	for (i = 0; i <= mkey_cache_max_order(dev); i++) {
989 		rb_key.ndescs = 1 << (i + 2);
990 		ent = mlx5r_cache_create_ent_locked(dev, rb_key, true);
991 		if (IS_ERR(ent)) {
992 			ret = PTR_ERR(ent);
993 			goto err;
994 		}
995 	}
996 
997 	ret = mlx5_odp_init_mkey_cache(dev);
998 	if (ret)
999 		goto err;
1000 
1001 	mutex_unlock(&cache->rb_lock);
1002 	for (node = rb_first(root); node; node = rb_next(node)) {
1003 		ent = rb_entry(node, struct mlx5_cache_ent, node);
1004 		xa_lock_irq(&ent->mkeys);
1005 		queue_adjust_cache_locked(ent);
1006 		xa_unlock_irq(&ent->mkeys);
1007 	}
1008 
1009 	return 0;
1010 
1011 err:
1012 	mutex_unlock(&cache->rb_lock);
1013 	mlx5_mkey_cache_debugfs_cleanup(dev);
1014 	mlx5_ib_warn(dev, "failed to create mkey cache entry\n");
1015 	return ret;
1016 }
1017 
1018 void mlx5_mkey_cache_cleanup(struct mlx5_ib_dev *dev)
1019 {
1020 	struct rb_root *root = &dev->cache.rb_root;
1021 	struct mlx5_cache_ent *ent;
1022 	struct rb_node *node;
1023 
1024 	if (!dev->cache.wq)
1025 		return;
1026 
1027 	cancel_delayed_work_sync(&dev->cache.remove_ent_dwork);
1028 	mutex_lock(&dev->cache.rb_lock);
1029 	for (node = rb_first(root); node; node = rb_next(node)) {
1030 		ent = rb_entry(node, struct mlx5_cache_ent, node);
1031 		xa_lock_irq(&ent->mkeys);
1032 		ent->disabled = true;
1033 		xa_unlock_irq(&ent->mkeys);
1034 		cancel_delayed_work_sync(&ent->dwork);
1035 	}
1036 
1037 	mlx5_mkey_cache_debugfs_cleanup(dev);
1038 	mlx5_cmd_cleanup_async_ctx(&dev->async_ctx);
1039 
1040 	node = rb_first(root);
1041 	while (node) {
1042 		ent = rb_entry(node, struct mlx5_cache_ent, node);
1043 		node = rb_next(node);
1044 		clean_keys(dev, ent);
1045 		rb_erase(&ent->node, root);
1046 		kfree(ent);
1047 	}
1048 	mutex_unlock(&dev->cache.rb_lock);
1049 
1050 	destroy_workqueue(dev->cache.wq);
1051 	del_timer_sync(&dev->delay_timer);
1052 }
1053 
1054 struct ib_mr *mlx5_ib_get_dma_mr(struct ib_pd *pd, int acc)
1055 {
1056 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1057 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1058 	struct mlx5_ib_mr *mr;
1059 	void *mkc;
1060 	u32 *in;
1061 	int err;
1062 
1063 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1064 	if (!mr)
1065 		return ERR_PTR(-ENOMEM);
1066 
1067 	in = kzalloc(inlen, GFP_KERNEL);
1068 	if (!in) {
1069 		err = -ENOMEM;
1070 		goto err_free;
1071 	}
1072 
1073 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1074 
1075 	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_PA);
1076 	MLX5_SET(mkc, mkc, length64, 1);
1077 	set_mkc_access_pd_addr_fields(mkc, acc | IB_ACCESS_RELAXED_ORDERING, 0,
1078 				      pd);
1079 
1080 	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
1081 	if (err)
1082 		goto err_in;
1083 
1084 	kfree(in);
1085 	mr->mmkey.type = MLX5_MKEY_MR;
1086 	mr->ibmr.lkey = mr->mmkey.key;
1087 	mr->ibmr.rkey = mr->mmkey.key;
1088 	mr->umem = NULL;
1089 
1090 	return &mr->ibmr;
1091 
1092 err_in:
1093 	kfree(in);
1094 
1095 err_free:
1096 	kfree(mr);
1097 
1098 	return ERR_PTR(err);
1099 }
1100 
1101 static int get_octo_len(u64 addr, u64 len, int page_shift)
1102 {
1103 	u64 page_size = 1ULL << page_shift;
1104 	u64 offset;
1105 	int npages;
1106 
1107 	offset = addr & (page_size - 1);
1108 	npages = ALIGN(len + offset, page_size) >> page_shift;
1109 	return (npages + 1) / 2;
1110 }
1111 
1112 static int mkey_cache_max_order(struct mlx5_ib_dev *dev)
1113 {
1114 	if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
1115 		return MKEY_CACHE_LAST_STD_ENTRY;
1116 	return MLX5_MAX_UMR_SHIFT;
1117 }
1118 
1119 static void set_mr_fields(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr,
1120 			  u64 length, int access_flags, u64 iova)
1121 {
1122 	mr->ibmr.lkey = mr->mmkey.key;
1123 	mr->ibmr.rkey = mr->mmkey.key;
1124 	mr->ibmr.length = length;
1125 	mr->ibmr.device = &dev->ib_dev;
1126 	mr->ibmr.iova = iova;
1127 	mr->access_flags = access_flags;
1128 }
1129 
1130 static unsigned int mlx5_umem_dmabuf_default_pgsz(struct ib_umem *umem,
1131 						  u64 iova)
1132 {
1133 	/*
1134 	 * The alignment of iova has already been checked upon entering
1135 	 * UVERBS_METHOD_REG_DMABUF_MR
1136 	 */
1137 	umem->iova = iova;
1138 	return PAGE_SIZE;
1139 }
1140 
1141 static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd,
1142 					     struct ib_umem *umem, u64 iova,
1143 					     int access_flags)
1144 {
1145 	struct mlx5r_cache_rb_key rb_key = {
1146 		.access_mode = MLX5_MKC_ACCESS_MODE_MTT,
1147 	};
1148 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1149 	struct mlx5_cache_ent *ent;
1150 	struct mlx5_ib_mr *mr;
1151 	unsigned int page_size;
1152 
1153 	if (umem->is_dmabuf)
1154 		page_size = mlx5_umem_dmabuf_default_pgsz(umem, iova);
1155 	else
1156 		page_size = mlx5_umem_find_best_pgsz(umem, mkc, log_page_size,
1157 						     0, iova);
1158 	if (WARN_ON(!page_size))
1159 		return ERR_PTR(-EINVAL);
1160 
1161 	rb_key.ndescs = ib_umem_num_dma_blocks(umem, page_size);
1162 	rb_key.ats = mlx5_umem_needs_ats(dev, umem, access_flags);
1163 	rb_key.access_flags = get_unchangeable_access_flags(dev, access_flags);
1164 	ent = mkey_cache_ent_from_rb_key(dev, rb_key);
1165 	/*
1166 	 * If the MR can't come from the cache then synchronously create an uncached
1167 	 * one.
1168 	 */
1169 	if (!ent) {
1170 		mutex_lock(&dev->slow_path_mutex);
1171 		mr = reg_create(pd, umem, iova, access_flags, page_size, false);
1172 		mutex_unlock(&dev->slow_path_mutex);
1173 		if (IS_ERR(mr))
1174 			return mr;
1175 		mr->mmkey.rb_key = rb_key;
1176 		return mr;
1177 	}
1178 
1179 	mr = _mlx5_mr_cache_alloc(dev, ent, access_flags);
1180 	if (IS_ERR(mr))
1181 		return mr;
1182 
1183 	mr->ibmr.pd = pd;
1184 	mr->umem = umem;
1185 	mr->page_shift = order_base_2(page_size);
1186 	set_mr_fields(dev, mr, umem->length, access_flags, iova);
1187 
1188 	return mr;
1189 }
1190 
1191 /*
1192  * If ibmr is NULL it will be allocated by reg_create.
1193  * Else, the given ibmr will be used.
1194  */
1195 static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem,
1196 				     u64 iova, int access_flags,
1197 				     unsigned int page_size, bool populate)
1198 {
1199 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1200 	struct mlx5_ib_mr *mr;
1201 	__be64 *pas;
1202 	void *mkc;
1203 	int inlen;
1204 	u32 *in;
1205 	int err;
1206 	bool pg_cap = !!(MLX5_CAP_GEN(dev->mdev, pg));
1207 
1208 	if (!page_size)
1209 		return ERR_PTR(-EINVAL);
1210 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1211 	if (!mr)
1212 		return ERR_PTR(-ENOMEM);
1213 
1214 	mr->ibmr.pd = pd;
1215 	mr->access_flags = access_flags;
1216 	mr->page_shift = order_base_2(page_size);
1217 
1218 	inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1219 	if (populate)
1220 		inlen += sizeof(*pas) *
1221 			 roundup(ib_umem_num_dma_blocks(umem, page_size), 2);
1222 	in = kvzalloc(inlen, GFP_KERNEL);
1223 	if (!in) {
1224 		err = -ENOMEM;
1225 		goto err_1;
1226 	}
1227 	pas = (__be64 *)MLX5_ADDR_OF(create_mkey_in, in, klm_pas_mtt);
1228 	if (populate) {
1229 		if (WARN_ON(access_flags & IB_ACCESS_ON_DEMAND)) {
1230 			err = -EINVAL;
1231 			goto err_2;
1232 		}
1233 		mlx5_ib_populate_pas(umem, 1UL << mr->page_shift, pas,
1234 				     pg_cap ? MLX5_IB_MTT_PRESENT : 0);
1235 	}
1236 
1237 	/* The pg_access bit allows setting the access flags
1238 	 * in the page list submitted with the command.
1239 	 */
1240 	MLX5_SET(create_mkey_in, in, pg_access, !!(pg_cap));
1241 
1242 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1243 	set_mkc_access_pd_addr_fields(mkc, access_flags, iova,
1244 				      populate ? pd : dev->umrc.pd);
1245 	MLX5_SET(mkc, mkc, free, !populate);
1246 	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_MTT);
1247 	MLX5_SET(mkc, mkc, umr_en, 1);
1248 
1249 	MLX5_SET64(mkc, mkc, len, umem->length);
1250 	MLX5_SET(mkc, mkc, bsf_octword_size, 0);
1251 	MLX5_SET(mkc, mkc, translations_octword_size,
1252 		 get_octo_len(iova, umem->length, mr->page_shift));
1253 	MLX5_SET(mkc, mkc, log_page_size, mr->page_shift);
1254 	if (mlx5_umem_needs_ats(dev, umem, access_flags))
1255 		MLX5_SET(mkc, mkc, ma_translation_mode, 1);
1256 	if (populate) {
1257 		MLX5_SET(create_mkey_in, in, translations_octword_actual_size,
1258 			 get_octo_len(iova, umem->length, mr->page_shift));
1259 	}
1260 
1261 	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
1262 	if (err) {
1263 		mlx5_ib_warn(dev, "create mkey failed\n");
1264 		goto err_2;
1265 	}
1266 	mr->mmkey.type = MLX5_MKEY_MR;
1267 	mr->mmkey.ndescs = get_octo_len(iova, umem->length, mr->page_shift);
1268 	mr->umem = umem;
1269 	set_mr_fields(dev, mr, umem->length, access_flags, iova);
1270 	kvfree(in);
1271 
1272 	mlx5_ib_dbg(dev, "mkey = 0x%x\n", mr->mmkey.key);
1273 
1274 	return mr;
1275 
1276 err_2:
1277 	kvfree(in);
1278 err_1:
1279 	kfree(mr);
1280 	return ERR_PTR(err);
1281 }
1282 
1283 static struct ib_mr *mlx5_ib_get_dm_mr(struct ib_pd *pd, u64 start_addr,
1284 				       u64 length, int acc, int mode)
1285 {
1286 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1287 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1288 	struct mlx5_ib_mr *mr;
1289 	void *mkc;
1290 	u32 *in;
1291 	int err;
1292 
1293 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1294 	if (!mr)
1295 		return ERR_PTR(-ENOMEM);
1296 
1297 	in = kzalloc(inlen, GFP_KERNEL);
1298 	if (!in) {
1299 		err = -ENOMEM;
1300 		goto err_free;
1301 	}
1302 
1303 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1304 
1305 	MLX5_SET(mkc, mkc, access_mode_1_0, mode & 0x3);
1306 	MLX5_SET(mkc, mkc, access_mode_4_2, (mode >> 2) & 0x7);
1307 	MLX5_SET64(mkc, mkc, len, length);
1308 	set_mkc_access_pd_addr_fields(mkc, acc, start_addr, pd);
1309 
1310 	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
1311 	if (err)
1312 		goto err_in;
1313 
1314 	kfree(in);
1315 
1316 	set_mr_fields(dev, mr, length, acc, start_addr);
1317 
1318 	return &mr->ibmr;
1319 
1320 err_in:
1321 	kfree(in);
1322 
1323 err_free:
1324 	kfree(mr);
1325 
1326 	return ERR_PTR(err);
1327 }
1328 
1329 int mlx5_ib_advise_mr(struct ib_pd *pd,
1330 		      enum ib_uverbs_advise_mr_advice advice,
1331 		      u32 flags,
1332 		      struct ib_sge *sg_list,
1333 		      u32 num_sge,
1334 		      struct uverbs_attr_bundle *attrs)
1335 {
1336 	if (advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH &&
1337 	    advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
1338 	    advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT)
1339 		return -EOPNOTSUPP;
1340 
1341 	return mlx5_ib_advise_mr_prefetch(pd, advice, flags,
1342 					 sg_list, num_sge);
1343 }
1344 
1345 struct ib_mr *mlx5_ib_reg_dm_mr(struct ib_pd *pd, struct ib_dm *dm,
1346 				struct ib_dm_mr_attr *attr,
1347 				struct uverbs_attr_bundle *attrs)
1348 {
1349 	struct mlx5_ib_dm *mdm = to_mdm(dm);
1350 	struct mlx5_core_dev *dev = to_mdev(dm->device)->mdev;
1351 	u64 start_addr = mdm->dev_addr + attr->offset;
1352 	int mode;
1353 
1354 	switch (mdm->type) {
1355 	case MLX5_IB_UAPI_DM_TYPE_MEMIC:
1356 		if (attr->access_flags & ~MLX5_IB_DM_MEMIC_ALLOWED_ACCESS)
1357 			return ERR_PTR(-EINVAL);
1358 
1359 		mode = MLX5_MKC_ACCESS_MODE_MEMIC;
1360 		start_addr -= pci_resource_start(dev->pdev, 0);
1361 		break;
1362 	case MLX5_IB_UAPI_DM_TYPE_STEERING_SW_ICM:
1363 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_SW_ICM:
1364 	case MLX5_IB_UAPI_DM_TYPE_HEADER_MODIFY_PATTERN_SW_ICM:
1365 		if (attr->access_flags & ~MLX5_IB_DM_SW_ICM_ALLOWED_ACCESS)
1366 			return ERR_PTR(-EINVAL);
1367 
1368 		mode = MLX5_MKC_ACCESS_MODE_SW_ICM;
1369 		break;
1370 	default:
1371 		return ERR_PTR(-EINVAL);
1372 	}
1373 
1374 	return mlx5_ib_get_dm_mr(pd, start_addr, attr->length,
1375 				 attr->access_flags, mode);
1376 }
1377 
1378 static struct ib_mr *create_real_mr(struct ib_pd *pd, struct ib_umem *umem,
1379 				    u64 iova, int access_flags)
1380 {
1381 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1382 	struct mlx5_ib_mr *mr = NULL;
1383 	bool xlt_with_umr;
1384 	int err;
1385 
1386 	xlt_with_umr = mlx5r_umr_can_load_pas(dev, umem->length);
1387 	if (xlt_with_umr) {
1388 		mr = alloc_cacheable_mr(pd, umem, iova, access_flags);
1389 	} else {
1390 		unsigned int page_size = mlx5_umem_find_best_pgsz(
1391 			umem, mkc, log_page_size, 0, iova);
1392 
1393 		mutex_lock(&dev->slow_path_mutex);
1394 		mr = reg_create(pd, umem, iova, access_flags, page_size, true);
1395 		mutex_unlock(&dev->slow_path_mutex);
1396 	}
1397 	if (IS_ERR(mr)) {
1398 		ib_umem_release(umem);
1399 		return ERR_CAST(mr);
1400 	}
1401 
1402 	mlx5_ib_dbg(dev, "mkey 0x%x\n", mr->mmkey.key);
1403 
1404 	atomic_add(ib_umem_num_pages(umem), &dev->mdev->priv.reg_pages);
1405 
1406 	if (xlt_with_umr) {
1407 		/*
1408 		 * If the MR was created with reg_create then it will be
1409 		 * configured properly but left disabled. It is safe to go ahead
1410 		 * and configure it again via UMR while enabling it.
1411 		 */
1412 		err = mlx5r_umr_update_mr_pas(mr, MLX5_IB_UPD_XLT_ENABLE);
1413 		if (err) {
1414 			mlx5_ib_dereg_mr(&mr->ibmr, NULL);
1415 			return ERR_PTR(err);
1416 		}
1417 	}
1418 	return &mr->ibmr;
1419 }
1420 
1421 static struct ib_mr *create_user_odp_mr(struct ib_pd *pd, u64 start, u64 length,
1422 					u64 iova, int access_flags,
1423 					struct ib_udata *udata)
1424 {
1425 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1426 	struct ib_umem_odp *odp;
1427 	struct mlx5_ib_mr *mr;
1428 	int err;
1429 
1430 	if (!IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING))
1431 		return ERR_PTR(-EOPNOTSUPP);
1432 
1433 	err = mlx5r_odp_create_eq(dev, &dev->odp_pf_eq);
1434 	if (err)
1435 		return ERR_PTR(err);
1436 	if (!start && length == U64_MAX) {
1437 		if (iova != 0)
1438 			return ERR_PTR(-EINVAL);
1439 		if (!(dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
1440 			return ERR_PTR(-EINVAL);
1441 
1442 		mr = mlx5_ib_alloc_implicit_mr(to_mpd(pd), access_flags);
1443 		if (IS_ERR(mr))
1444 			return ERR_CAST(mr);
1445 		return &mr->ibmr;
1446 	}
1447 
1448 	/* ODP requires xlt update via umr to work. */
1449 	if (!mlx5r_umr_can_load_pas(dev, length))
1450 		return ERR_PTR(-EINVAL);
1451 
1452 	odp = ib_umem_odp_get(&dev->ib_dev, start, length, access_flags,
1453 			      &mlx5_mn_ops);
1454 	if (IS_ERR(odp))
1455 		return ERR_CAST(odp);
1456 
1457 	mr = alloc_cacheable_mr(pd, &odp->umem, iova, access_flags);
1458 	if (IS_ERR(mr)) {
1459 		ib_umem_release(&odp->umem);
1460 		return ERR_CAST(mr);
1461 	}
1462 	xa_init(&mr->implicit_children);
1463 
1464 	odp->private = mr;
1465 	err = mlx5r_store_odp_mkey(dev, &mr->mmkey);
1466 	if (err)
1467 		goto err_dereg_mr;
1468 
1469 	err = mlx5_ib_init_odp_mr(mr);
1470 	if (err)
1471 		goto err_dereg_mr;
1472 	return &mr->ibmr;
1473 
1474 err_dereg_mr:
1475 	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
1476 	return ERR_PTR(err);
1477 }
1478 
1479 struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
1480 				  u64 iova, int access_flags,
1481 				  struct ib_udata *udata)
1482 {
1483 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1484 	struct ib_umem *umem;
1485 
1486 	if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM))
1487 		return ERR_PTR(-EOPNOTSUPP);
1488 
1489 	mlx5_ib_dbg(dev, "start 0x%llx, iova 0x%llx, length 0x%llx, access_flags 0x%x\n",
1490 		    start, iova, length, access_flags);
1491 
1492 	if (access_flags & IB_ACCESS_ON_DEMAND)
1493 		return create_user_odp_mr(pd, start, length, iova, access_flags,
1494 					  udata);
1495 	umem = ib_umem_get(&dev->ib_dev, start, length, access_flags);
1496 	if (IS_ERR(umem))
1497 		return ERR_CAST(umem);
1498 	return create_real_mr(pd, umem, iova, access_flags);
1499 }
1500 
1501 static void mlx5_ib_dmabuf_invalidate_cb(struct dma_buf_attachment *attach)
1502 {
1503 	struct ib_umem_dmabuf *umem_dmabuf = attach->importer_priv;
1504 	struct mlx5_ib_mr *mr = umem_dmabuf->private;
1505 
1506 	dma_resv_assert_held(umem_dmabuf->attach->dmabuf->resv);
1507 
1508 	if (!umem_dmabuf->sgt)
1509 		return;
1510 
1511 	mlx5r_umr_update_mr_pas(mr, MLX5_IB_UPD_XLT_ZAP);
1512 	ib_umem_dmabuf_unmap_pages(umem_dmabuf);
1513 }
1514 
1515 static struct dma_buf_attach_ops mlx5_ib_dmabuf_attach_ops = {
1516 	.allow_peer2peer = 1,
1517 	.move_notify = mlx5_ib_dmabuf_invalidate_cb,
1518 };
1519 
1520 struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset,
1521 					 u64 length, u64 virt_addr,
1522 					 int fd, int access_flags,
1523 					 struct ib_udata *udata)
1524 {
1525 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1526 	struct mlx5_ib_mr *mr = NULL;
1527 	struct ib_umem_dmabuf *umem_dmabuf;
1528 	int err;
1529 
1530 	if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM) ||
1531 	    !IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING))
1532 		return ERR_PTR(-EOPNOTSUPP);
1533 
1534 	mlx5_ib_dbg(dev,
1535 		    "offset 0x%llx, virt_addr 0x%llx, length 0x%llx, fd %d, access_flags 0x%x\n",
1536 		    offset, virt_addr, length, fd, access_flags);
1537 
1538 	/* dmabuf requires xlt update via umr to work. */
1539 	if (!mlx5r_umr_can_load_pas(dev, length))
1540 		return ERR_PTR(-EINVAL);
1541 
1542 	umem_dmabuf = ib_umem_dmabuf_get(&dev->ib_dev, offset, length, fd,
1543 					 access_flags,
1544 					 &mlx5_ib_dmabuf_attach_ops);
1545 	if (IS_ERR(umem_dmabuf)) {
1546 		mlx5_ib_dbg(dev, "umem_dmabuf get failed (%ld)\n",
1547 			    PTR_ERR(umem_dmabuf));
1548 		return ERR_CAST(umem_dmabuf);
1549 	}
1550 
1551 	mr = alloc_cacheable_mr(pd, &umem_dmabuf->umem, virt_addr,
1552 				access_flags);
1553 	if (IS_ERR(mr)) {
1554 		ib_umem_release(&umem_dmabuf->umem);
1555 		return ERR_CAST(mr);
1556 	}
1557 
1558 	mlx5_ib_dbg(dev, "mkey 0x%x\n", mr->mmkey.key);
1559 
1560 	atomic_add(ib_umem_num_pages(mr->umem), &dev->mdev->priv.reg_pages);
1561 	umem_dmabuf->private = mr;
1562 	err = mlx5r_store_odp_mkey(dev, &mr->mmkey);
1563 	if (err)
1564 		goto err_dereg_mr;
1565 
1566 	err = mlx5_ib_init_dmabuf_mr(mr);
1567 	if (err)
1568 		goto err_dereg_mr;
1569 	return &mr->ibmr;
1570 
1571 err_dereg_mr:
1572 	mlx5_ib_dereg_mr(&mr->ibmr, NULL);
1573 	return ERR_PTR(err);
1574 }
1575 
1576 /*
1577  * True if the change in access flags can be done via UMR, only some access
1578  * flags can be updated.
1579  */
1580 static bool can_use_umr_rereg_access(struct mlx5_ib_dev *dev,
1581 				     unsigned int current_access_flags,
1582 				     unsigned int target_access_flags)
1583 {
1584 	unsigned int diffs = current_access_flags ^ target_access_flags;
1585 
1586 	if (diffs & ~(IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE |
1587 		      IB_ACCESS_REMOTE_READ | IB_ACCESS_RELAXED_ORDERING))
1588 		return false;
1589 	return mlx5r_umr_can_reconfig(dev, current_access_flags,
1590 				      target_access_flags);
1591 }
1592 
1593 static bool can_use_umr_rereg_pas(struct mlx5_ib_mr *mr,
1594 				  struct ib_umem *new_umem,
1595 				  int new_access_flags, u64 iova,
1596 				  unsigned long *page_size)
1597 {
1598 	struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device);
1599 
1600 	/* We only track the allocated sizes of MRs from the cache */
1601 	if (!mr->mmkey.cache_ent)
1602 		return false;
1603 	if (!mlx5r_umr_can_load_pas(dev, new_umem->length))
1604 		return false;
1605 
1606 	*page_size =
1607 		mlx5_umem_find_best_pgsz(new_umem, mkc, log_page_size, 0, iova);
1608 	if (WARN_ON(!*page_size))
1609 		return false;
1610 	return (mr->mmkey.cache_ent->rb_key.ndescs) >=
1611 	       ib_umem_num_dma_blocks(new_umem, *page_size);
1612 }
1613 
1614 static int umr_rereg_pas(struct mlx5_ib_mr *mr, struct ib_pd *pd,
1615 			 int access_flags, int flags, struct ib_umem *new_umem,
1616 			 u64 iova, unsigned long page_size)
1617 {
1618 	struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device);
1619 	int upd_flags = MLX5_IB_UPD_XLT_ADDR | MLX5_IB_UPD_XLT_ENABLE;
1620 	struct ib_umem *old_umem = mr->umem;
1621 	int err;
1622 
1623 	/*
1624 	 * To keep everything simple the MR is revoked before we start to mess
1625 	 * with it. This ensure the change is atomic relative to any use of the
1626 	 * MR.
1627 	 */
1628 	err = mlx5r_umr_revoke_mr(mr);
1629 	if (err)
1630 		return err;
1631 
1632 	if (flags & IB_MR_REREG_PD) {
1633 		mr->ibmr.pd = pd;
1634 		upd_flags |= MLX5_IB_UPD_XLT_PD;
1635 	}
1636 	if (flags & IB_MR_REREG_ACCESS) {
1637 		mr->access_flags = access_flags;
1638 		upd_flags |= MLX5_IB_UPD_XLT_ACCESS;
1639 	}
1640 
1641 	mr->ibmr.iova = iova;
1642 	mr->ibmr.length = new_umem->length;
1643 	mr->page_shift = order_base_2(page_size);
1644 	mr->umem = new_umem;
1645 	err = mlx5r_umr_update_mr_pas(mr, upd_flags);
1646 	if (err) {
1647 		/*
1648 		 * The MR is revoked at this point so there is no issue to free
1649 		 * new_umem.
1650 		 */
1651 		mr->umem = old_umem;
1652 		return err;
1653 	}
1654 
1655 	atomic_sub(ib_umem_num_pages(old_umem), &dev->mdev->priv.reg_pages);
1656 	ib_umem_release(old_umem);
1657 	atomic_add(ib_umem_num_pages(new_umem), &dev->mdev->priv.reg_pages);
1658 	return 0;
1659 }
1660 
1661 struct ib_mr *mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start,
1662 				    u64 length, u64 iova, int new_access_flags,
1663 				    struct ib_pd *new_pd,
1664 				    struct ib_udata *udata)
1665 {
1666 	struct mlx5_ib_dev *dev = to_mdev(ib_mr->device);
1667 	struct mlx5_ib_mr *mr = to_mmr(ib_mr);
1668 	int err;
1669 
1670 	if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM))
1671 		return ERR_PTR(-EOPNOTSUPP);
1672 
1673 	mlx5_ib_dbg(
1674 		dev,
1675 		"start 0x%llx, iova 0x%llx, length 0x%llx, access_flags 0x%x\n",
1676 		start, iova, length, new_access_flags);
1677 
1678 	if (flags & ~(IB_MR_REREG_TRANS | IB_MR_REREG_PD | IB_MR_REREG_ACCESS))
1679 		return ERR_PTR(-EOPNOTSUPP);
1680 
1681 	if (!(flags & IB_MR_REREG_ACCESS))
1682 		new_access_flags = mr->access_flags;
1683 	if (!(flags & IB_MR_REREG_PD))
1684 		new_pd = ib_mr->pd;
1685 
1686 	if (!(flags & IB_MR_REREG_TRANS)) {
1687 		struct ib_umem *umem;
1688 
1689 		/* Fast path for PD/access change */
1690 		if (can_use_umr_rereg_access(dev, mr->access_flags,
1691 					     new_access_flags)) {
1692 			err = mlx5r_umr_rereg_pd_access(mr, new_pd,
1693 							new_access_flags);
1694 			if (err)
1695 				return ERR_PTR(err);
1696 			return NULL;
1697 		}
1698 		/* DM or ODP MR's don't have a normal umem so we can't re-use it */
1699 		if (!mr->umem || is_odp_mr(mr) || is_dmabuf_mr(mr))
1700 			goto recreate;
1701 
1702 		/*
1703 		 * Only one active MR can refer to a umem at one time, revoke
1704 		 * the old MR before assigning the umem to the new one.
1705 		 */
1706 		err = mlx5r_umr_revoke_mr(mr);
1707 		if (err)
1708 			return ERR_PTR(err);
1709 		umem = mr->umem;
1710 		mr->umem = NULL;
1711 		atomic_sub(ib_umem_num_pages(umem), &dev->mdev->priv.reg_pages);
1712 
1713 		return create_real_mr(new_pd, umem, mr->ibmr.iova,
1714 				      new_access_flags);
1715 	}
1716 
1717 	/*
1718 	 * DM doesn't have a PAS list so we can't re-use it, odp/dmabuf does
1719 	 * but the logic around releasing the umem is different
1720 	 */
1721 	if (!mr->umem || is_odp_mr(mr) || is_dmabuf_mr(mr))
1722 		goto recreate;
1723 
1724 	if (!(new_access_flags & IB_ACCESS_ON_DEMAND) &&
1725 	    can_use_umr_rereg_access(dev, mr->access_flags, new_access_flags)) {
1726 		struct ib_umem *new_umem;
1727 		unsigned long page_size;
1728 
1729 		new_umem = ib_umem_get(&dev->ib_dev, start, length,
1730 				       new_access_flags);
1731 		if (IS_ERR(new_umem))
1732 			return ERR_CAST(new_umem);
1733 
1734 		/* Fast path for PAS change */
1735 		if (can_use_umr_rereg_pas(mr, new_umem, new_access_flags, iova,
1736 					  &page_size)) {
1737 			err = umr_rereg_pas(mr, new_pd, new_access_flags, flags,
1738 					    new_umem, iova, page_size);
1739 			if (err) {
1740 				ib_umem_release(new_umem);
1741 				return ERR_PTR(err);
1742 			}
1743 			return NULL;
1744 		}
1745 		return create_real_mr(new_pd, new_umem, iova, new_access_flags);
1746 	}
1747 
1748 	/*
1749 	 * Everything else has no state we can preserve, just create a new MR
1750 	 * from scratch
1751 	 */
1752 recreate:
1753 	return mlx5_ib_reg_user_mr(new_pd, start, length, iova,
1754 				   new_access_flags, udata);
1755 }
1756 
1757 static int
1758 mlx5_alloc_priv_descs(struct ib_device *device,
1759 		      struct mlx5_ib_mr *mr,
1760 		      int ndescs,
1761 		      int desc_size)
1762 {
1763 	struct mlx5_ib_dev *dev = to_mdev(device);
1764 	struct device *ddev = &dev->mdev->pdev->dev;
1765 	int size = ndescs * desc_size;
1766 	int add_size;
1767 	int ret;
1768 
1769 	add_size = max_t(int, MLX5_UMR_ALIGN - ARCH_KMALLOC_MINALIGN, 0);
1770 	if (is_power_of_2(MLX5_UMR_ALIGN) && add_size) {
1771 		int end = max_t(int, MLX5_UMR_ALIGN, roundup_pow_of_two(size));
1772 
1773 		add_size = min_t(int, end - size, add_size);
1774 	}
1775 
1776 	mr->descs_alloc = kzalloc(size + add_size, GFP_KERNEL);
1777 	if (!mr->descs_alloc)
1778 		return -ENOMEM;
1779 
1780 	mr->descs = PTR_ALIGN(mr->descs_alloc, MLX5_UMR_ALIGN);
1781 
1782 	mr->desc_map = dma_map_single(ddev, mr->descs, size, DMA_TO_DEVICE);
1783 	if (dma_mapping_error(ddev, mr->desc_map)) {
1784 		ret = -ENOMEM;
1785 		goto err;
1786 	}
1787 
1788 	return 0;
1789 err:
1790 	kfree(mr->descs_alloc);
1791 
1792 	return ret;
1793 }
1794 
1795 static void
1796 mlx5_free_priv_descs(struct mlx5_ib_mr *mr)
1797 {
1798 	if (!mr->umem && mr->descs) {
1799 		struct ib_device *device = mr->ibmr.device;
1800 		int size = mr->max_descs * mr->desc_size;
1801 		struct mlx5_ib_dev *dev = to_mdev(device);
1802 
1803 		dma_unmap_single(&dev->mdev->pdev->dev, mr->desc_map, size,
1804 				 DMA_TO_DEVICE);
1805 		kfree(mr->descs_alloc);
1806 		mr->descs = NULL;
1807 	}
1808 }
1809 
1810 static int cache_ent_find_and_store(struct mlx5_ib_dev *dev,
1811 				    struct mlx5_ib_mr *mr)
1812 {
1813 	struct mlx5_mkey_cache *cache = &dev->cache;
1814 	struct mlx5_cache_ent *ent;
1815 	int ret;
1816 
1817 	if (mr->mmkey.cache_ent) {
1818 		xa_lock_irq(&mr->mmkey.cache_ent->mkeys);
1819 		mr->mmkey.cache_ent->in_use--;
1820 		goto end;
1821 	}
1822 
1823 	mutex_lock(&cache->rb_lock);
1824 	ent = mkey_cache_ent_from_rb_key(dev, mr->mmkey.rb_key);
1825 	if (ent) {
1826 		if (ent->rb_key.ndescs == mr->mmkey.rb_key.ndescs) {
1827 			if (ent->disabled) {
1828 				mutex_unlock(&cache->rb_lock);
1829 				return -EOPNOTSUPP;
1830 			}
1831 			mr->mmkey.cache_ent = ent;
1832 			xa_lock_irq(&mr->mmkey.cache_ent->mkeys);
1833 			mutex_unlock(&cache->rb_lock);
1834 			goto end;
1835 		}
1836 	}
1837 
1838 	ent = mlx5r_cache_create_ent_locked(dev, mr->mmkey.rb_key, false);
1839 	mutex_unlock(&cache->rb_lock);
1840 	if (IS_ERR(ent))
1841 		return PTR_ERR(ent);
1842 
1843 	mr->mmkey.cache_ent = ent;
1844 	xa_lock_irq(&mr->mmkey.cache_ent->mkeys);
1845 
1846 end:
1847 	ret = push_mkey_locked(mr->mmkey.cache_ent, false,
1848 			       xa_mk_value(mr->mmkey.key));
1849 	xa_unlock_irq(&mr->mmkey.cache_ent->mkeys);
1850 	return ret;
1851 }
1852 
1853 int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1854 {
1855 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
1856 	struct mlx5_ib_dev *dev = to_mdev(ibmr->device);
1857 	int rc;
1858 
1859 	/*
1860 	 * Any async use of the mr must hold the refcount, once the refcount
1861 	 * goes to zero no other thread, such as ODP page faults, prefetch, any
1862 	 * UMR activity, etc can touch the mkey. Thus it is safe to destroy it.
1863 	 */
1864 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) &&
1865 	    refcount_read(&mr->mmkey.usecount) != 0 &&
1866 	    xa_erase(&mr_to_mdev(mr)->odp_mkeys, mlx5_base_mkey(mr->mmkey.key)))
1867 		mlx5r_deref_wait_odp_mkey(&mr->mmkey);
1868 
1869 	if (ibmr->type == IB_MR_TYPE_INTEGRITY) {
1870 		xa_cmpxchg(&dev->sig_mrs, mlx5_base_mkey(mr->mmkey.key),
1871 			   mr->sig, NULL, GFP_KERNEL);
1872 
1873 		if (mr->mtt_mr) {
1874 			rc = mlx5_ib_dereg_mr(&mr->mtt_mr->ibmr, NULL);
1875 			if (rc)
1876 				return rc;
1877 			mr->mtt_mr = NULL;
1878 		}
1879 		if (mr->klm_mr) {
1880 			rc = mlx5_ib_dereg_mr(&mr->klm_mr->ibmr, NULL);
1881 			if (rc)
1882 				return rc;
1883 			mr->klm_mr = NULL;
1884 		}
1885 
1886 		if (mlx5_core_destroy_psv(dev->mdev,
1887 					  mr->sig->psv_memory.psv_idx))
1888 			mlx5_ib_warn(dev, "failed to destroy mem psv %d\n",
1889 				     mr->sig->psv_memory.psv_idx);
1890 		if (mlx5_core_destroy_psv(dev->mdev, mr->sig->psv_wire.psv_idx))
1891 			mlx5_ib_warn(dev, "failed to destroy wire psv %d\n",
1892 				     mr->sig->psv_wire.psv_idx);
1893 		kfree(mr->sig);
1894 		mr->sig = NULL;
1895 	}
1896 
1897 	/* Stop DMA */
1898 	if (mr->umem && mlx5r_umr_can_load_pas(dev, mr->umem->length))
1899 		if (mlx5r_umr_revoke_mr(mr) ||
1900 		    cache_ent_find_and_store(dev, mr))
1901 			mr->mmkey.cache_ent = NULL;
1902 
1903 	if (!mr->mmkey.cache_ent) {
1904 		rc = destroy_mkey(to_mdev(mr->ibmr.device), mr);
1905 		if (rc)
1906 			return rc;
1907 	}
1908 
1909 	if (mr->umem) {
1910 		bool is_odp = is_odp_mr(mr);
1911 
1912 		if (!is_odp)
1913 			atomic_sub(ib_umem_num_pages(mr->umem),
1914 				   &dev->mdev->priv.reg_pages);
1915 		ib_umem_release(mr->umem);
1916 		if (is_odp)
1917 			mlx5_ib_free_odp_mr(mr);
1918 	}
1919 
1920 	if (!mr->mmkey.cache_ent)
1921 		mlx5_free_priv_descs(mr);
1922 
1923 	kfree(mr);
1924 	return 0;
1925 }
1926 
1927 static void mlx5_set_umr_free_mkey(struct ib_pd *pd, u32 *in, int ndescs,
1928 				   int access_mode, int page_shift)
1929 {
1930 	void *mkc;
1931 
1932 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
1933 
1934 	/* This is only used from the kernel, so setting the PD is OK. */
1935 	set_mkc_access_pd_addr_fields(mkc, IB_ACCESS_RELAXED_ORDERING, 0, pd);
1936 	MLX5_SET(mkc, mkc, free, 1);
1937 	MLX5_SET(mkc, mkc, translations_octword_size, ndescs);
1938 	MLX5_SET(mkc, mkc, access_mode_1_0, access_mode & 0x3);
1939 	MLX5_SET(mkc, mkc, access_mode_4_2, (access_mode >> 2) & 0x7);
1940 	MLX5_SET(mkc, mkc, umr_en, 1);
1941 	MLX5_SET(mkc, mkc, log_page_size, page_shift);
1942 }
1943 
1944 static int _mlx5_alloc_mkey_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
1945 				  int ndescs, int desc_size, int page_shift,
1946 				  int access_mode, u32 *in, int inlen)
1947 {
1948 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
1949 	int err;
1950 
1951 	mr->access_mode = access_mode;
1952 	mr->desc_size = desc_size;
1953 	mr->max_descs = ndescs;
1954 
1955 	err = mlx5_alloc_priv_descs(pd->device, mr, ndescs, desc_size);
1956 	if (err)
1957 		return err;
1958 
1959 	mlx5_set_umr_free_mkey(pd, in, ndescs, access_mode, page_shift);
1960 
1961 	err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen);
1962 	if (err)
1963 		goto err_free_descs;
1964 
1965 	mr->mmkey.type = MLX5_MKEY_MR;
1966 	mr->ibmr.lkey = mr->mmkey.key;
1967 	mr->ibmr.rkey = mr->mmkey.key;
1968 
1969 	return 0;
1970 
1971 err_free_descs:
1972 	mlx5_free_priv_descs(mr);
1973 	return err;
1974 }
1975 
1976 static struct mlx5_ib_mr *mlx5_ib_alloc_pi_mr(struct ib_pd *pd,
1977 				u32 max_num_sg, u32 max_num_meta_sg,
1978 				int desc_size, int access_mode)
1979 {
1980 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
1981 	int ndescs = ALIGN(max_num_sg + max_num_meta_sg, 4);
1982 	int page_shift = 0;
1983 	struct mlx5_ib_mr *mr;
1984 	u32 *in;
1985 	int err;
1986 
1987 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1988 	if (!mr)
1989 		return ERR_PTR(-ENOMEM);
1990 
1991 	mr->ibmr.pd = pd;
1992 	mr->ibmr.device = pd->device;
1993 
1994 	in = kzalloc(inlen, GFP_KERNEL);
1995 	if (!in) {
1996 		err = -ENOMEM;
1997 		goto err_free;
1998 	}
1999 
2000 	if (access_mode == MLX5_MKC_ACCESS_MODE_MTT)
2001 		page_shift = PAGE_SHIFT;
2002 
2003 	err = _mlx5_alloc_mkey_descs(pd, mr, ndescs, desc_size, page_shift,
2004 				     access_mode, in, inlen);
2005 	if (err)
2006 		goto err_free_in;
2007 
2008 	mr->umem = NULL;
2009 	kfree(in);
2010 
2011 	return mr;
2012 
2013 err_free_in:
2014 	kfree(in);
2015 err_free:
2016 	kfree(mr);
2017 	return ERR_PTR(err);
2018 }
2019 
2020 static int mlx5_alloc_mem_reg_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
2021 				    int ndescs, u32 *in, int inlen)
2022 {
2023 	return _mlx5_alloc_mkey_descs(pd, mr, ndescs, sizeof(struct mlx5_mtt),
2024 				      PAGE_SHIFT, MLX5_MKC_ACCESS_MODE_MTT, in,
2025 				      inlen);
2026 }
2027 
2028 static int mlx5_alloc_sg_gaps_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
2029 				    int ndescs, u32 *in, int inlen)
2030 {
2031 	return _mlx5_alloc_mkey_descs(pd, mr, ndescs, sizeof(struct mlx5_klm),
2032 				      0, MLX5_MKC_ACCESS_MODE_KLMS, in, inlen);
2033 }
2034 
2035 static int mlx5_alloc_integrity_descs(struct ib_pd *pd, struct mlx5_ib_mr *mr,
2036 				      int max_num_sg, int max_num_meta_sg,
2037 				      u32 *in, int inlen)
2038 {
2039 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
2040 	u32 psv_index[2];
2041 	void *mkc;
2042 	int err;
2043 
2044 	mr->sig = kzalloc(sizeof(*mr->sig), GFP_KERNEL);
2045 	if (!mr->sig)
2046 		return -ENOMEM;
2047 
2048 	/* create mem & wire PSVs */
2049 	err = mlx5_core_create_psv(dev->mdev, to_mpd(pd)->pdn, 2, psv_index);
2050 	if (err)
2051 		goto err_free_sig;
2052 
2053 	mr->sig->psv_memory.psv_idx = psv_index[0];
2054 	mr->sig->psv_wire.psv_idx = psv_index[1];
2055 
2056 	mr->sig->sig_status_checked = true;
2057 	mr->sig->sig_err_exists = false;
2058 	/* Next UMR, Arm SIGERR */
2059 	++mr->sig->sigerr_count;
2060 	mr->klm_mr = mlx5_ib_alloc_pi_mr(pd, max_num_sg, max_num_meta_sg,
2061 					 sizeof(struct mlx5_klm),
2062 					 MLX5_MKC_ACCESS_MODE_KLMS);
2063 	if (IS_ERR(mr->klm_mr)) {
2064 		err = PTR_ERR(mr->klm_mr);
2065 		goto err_destroy_psv;
2066 	}
2067 	mr->mtt_mr = mlx5_ib_alloc_pi_mr(pd, max_num_sg, max_num_meta_sg,
2068 					 sizeof(struct mlx5_mtt),
2069 					 MLX5_MKC_ACCESS_MODE_MTT);
2070 	if (IS_ERR(mr->mtt_mr)) {
2071 		err = PTR_ERR(mr->mtt_mr);
2072 		goto err_free_klm_mr;
2073 	}
2074 
2075 	/* Set bsf descriptors for mkey */
2076 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
2077 	MLX5_SET(mkc, mkc, bsf_en, 1);
2078 	MLX5_SET(mkc, mkc, bsf_octword_size, MLX5_MKEY_BSF_OCTO_SIZE);
2079 
2080 	err = _mlx5_alloc_mkey_descs(pd, mr, 4, sizeof(struct mlx5_klm), 0,
2081 				     MLX5_MKC_ACCESS_MODE_KLMS, in, inlen);
2082 	if (err)
2083 		goto err_free_mtt_mr;
2084 
2085 	err = xa_err(xa_store(&dev->sig_mrs, mlx5_base_mkey(mr->mmkey.key),
2086 			      mr->sig, GFP_KERNEL));
2087 	if (err)
2088 		goto err_free_descs;
2089 	return 0;
2090 
2091 err_free_descs:
2092 	destroy_mkey(dev, mr);
2093 	mlx5_free_priv_descs(mr);
2094 err_free_mtt_mr:
2095 	mlx5_ib_dereg_mr(&mr->mtt_mr->ibmr, NULL);
2096 	mr->mtt_mr = NULL;
2097 err_free_klm_mr:
2098 	mlx5_ib_dereg_mr(&mr->klm_mr->ibmr, NULL);
2099 	mr->klm_mr = NULL;
2100 err_destroy_psv:
2101 	if (mlx5_core_destroy_psv(dev->mdev, mr->sig->psv_memory.psv_idx))
2102 		mlx5_ib_warn(dev, "failed to destroy mem psv %d\n",
2103 			     mr->sig->psv_memory.psv_idx);
2104 	if (mlx5_core_destroy_psv(dev->mdev, mr->sig->psv_wire.psv_idx))
2105 		mlx5_ib_warn(dev, "failed to destroy wire psv %d\n",
2106 			     mr->sig->psv_wire.psv_idx);
2107 err_free_sig:
2108 	kfree(mr->sig);
2109 
2110 	return err;
2111 }
2112 
2113 static struct ib_mr *__mlx5_ib_alloc_mr(struct ib_pd *pd,
2114 					enum ib_mr_type mr_type, u32 max_num_sg,
2115 					u32 max_num_meta_sg)
2116 {
2117 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
2118 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
2119 	int ndescs = ALIGN(max_num_sg, 4);
2120 	struct mlx5_ib_mr *mr;
2121 	u32 *in;
2122 	int err;
2123 
2124 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2125 	if (!mr)
2126 		return ERR_PTR(-ENOMEM);
2127 
2128 	in = kzalloc(inlen, GFP_KERNEL);
2129 	if (!in) {
2130 		err = -ENOMEM;
2131 		goto err_free;
2132 	}
2133 
2134 	mr->ibmr.device = pd->device;
2135 	mr->umem = NULL;
2136 
2137 	switch (mr_type) {
2138 	case IB_MR_TYPE_MEM_REG:
2139 		err = mlx5_alloc_mem_reg_descs(pd, mr, ndescs, in, inlen);
2140 		break;
2141 	case IB_MR_TYPE_SG_GAPS:
2142 		err = mlx5_alloc_sg_gaps_descs(pd, mr, ndescs, in, inlen);
2143 		break;
2144 	case IB_MR_TYPE_INTEGRITY:
2145 		err = mlx5_alloc_integrity_descs(pd, mr, max_num_sg,
2146 						 max_num_meta_sg, in, inlen);
2147 		break;
2148 	default:
2149 		mlx5_ib_warn(dev, "Invalid mr type %d\n", mr_type);
2150 		err = -EINVAL;
2151 	}
2152 
2153 	if (err)
2154 		goto err_free_in;
2155 
2156 	kfree(in);
2157 
2158 	return &mr->ibmr;
2159 
2160 err_free_in:
2161 	kfree(in);
2162 err_free:
2163 	kfree(mr);
2164 	return ERR_PTR(err);
2165 }
2166 
2167 struct ib_mr *mlx5_ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2168 			       u32 max_num_sg)
2169 {
2170 	return __mlx5_ib_alloc_mr(pd, mr_type, max_num_sg, 0);
2171 }
2172 
2173 struct ib_mr *mlx5_ib_alloc_mr_integrity(struct ib_pd *pd,
2174 					 u32 max_num_sg, u32 max_num_meta_sg)
2175 {
2176 	return __mlx5_ib_alloc_mr(pd, IB_MR_TYPE_INTEGRITY, max_num_sg,
2177 				  max_num_meta_sg);
2178 }
2179 
2180 int mlx5_ib_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
2181 {
2182 	struct mlx5_ib_dev *dev = to_mdev(ibmw->device);
2183 	int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
2184 	struct mlx5_ib_mw *mw = to_mmw(ibmw);
2185 	unsigned int ndescs;
2186 	u32 *in = NULL;
2187 	void *mkc;
2188 	int err;
2189 	struct mlx5_ib_alloc_mw req = {};
2190 	struct {
2191 		__u32	comp_mask;
2192 		__u32	response_length;
2193 	} resp = {};
2194 
2195 	err = ib_copy_from_udata(&req, udata, min(udata->inlen, sizeof(req)));
2196 	if (err)
2197 		return err;
2198 
2199 	if (req.comp_mask || req.reserved1 || req.reserved2)
2200 		return -EOPNOTSUPP;
2201 
2202 	if (udata->inlen > sizeof(req) &&
2203 	    !ib_is_udata_cleared(udata, sizeof(req),
2204 				 udata->inlen - sizeof(req)))
2205 		return -EOPNOTSUPP;
2206 
2207 	ndescs = req.num_klms ? roundup(req.num_klms, 4) : roundup(1, 4);
2208 
2209 	in = kzalloc(inlen, GFP_KERNEL);
2210 	if (!in)
2211 		return -ENOMEM;
2212 
2213 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
2214 
2215 	MLX5_SET(mkc, mkc, free, 1);
2216 	MLX5_SET(mkc, mkc, translations_octword_size, ndescs);
2217 	MLX5_SET(mkc, mkc, pd, to_mpd(ibmw->pd)->pdn);
2218 	MLX5_SET(mkc, mkc, umr_en, 1);
2219 	MLX5_SET(mkc, mkc, lr, 1);
2220 	MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_KLMS);
2221 	MLX5_SET(mkc, mkc, en_rinval, !!((ibmw->type == IB_MW_TYPE_2)));
2222 	MLX5_SET(mkc, mkc, qpn, 0xffffff);
2223 
2224 	err = mlx5_ib_create_mkey(dev, &mw->mmkey, in, inlen);
2225 	if (err)
2226 		goto free;
2227 
2228 	mw->mmkey.type = MLX5_MKEY_MW;
2229 	ibmw->rkey = mw->mmkey.key;
2230 	mw->mmkey.ndescs = ndescs;
2231 
2232 	resp.response_length =
2233 		min(offsetofend(typeof(resp), response_length), udata->outlen);
2234 	if (resp.response_length) {
2235 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
2236 		if (err)
2237 			goto free_mkey;
2238 	}
2239 
2240 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
2241 		err = mlx5r_store_odp_mkey(dev, &mw->mmkey);
2242 		if (err)
2243 			goto free_mkey;
2244 	}
2245 
2246 	kfree(in);
2247 	return 0;
2248 
2249 free_mkey:
2250 	mlx5_core_destroy_mkey(dev->mdev, mw->mmkey.key);
2251 free:
2252 	kfree(in);
2253 	return err;
2254 }
2255 
2256 int mlx5_ib_dealloc_mw(struct ib_mw *mw)
2257 {
2258 	struct mlx5_ib_dev *dev = to_mdev(mw->device);
2259 	struct mlx5_ib_mw *mmw = to_mmw(mw);
2260 
2261 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) &&
2262 	    xa_erase(&dev->odp_mkeys, mlx5_base_mkey(mmw->mmkey.key)))
2263 		/*
2264 		 * pagefault_single_data_segment() may be accessing mmw
2265 		 * if the user bound an ODP MR to this MW.
2266 		 */
2267 		mlx5r_deref_wait_odp_mkey(&mmw->mmkey);
2268 
2269 	return mlx5_core_destroy_mkey(dev->mdev, mmw->mmkey.key);
2270 }
2271 
2272 int mlx5_ib_check_mr_status(struct ib_mr *ibmr, u32 check_mask,
2273 			    struct ib_mr_status *mr_status)
2274 {
2275 	struct mlx5_ib_mr *mmr = to_mmr(ibmr);
2276 	int ret = 0;
2277 
2278 	if (check_mask & ~IB_MR_CHECK_SIG_STATUS) {
2279 		pr_err("Invalid status check mask\n");
2280 		ret = -EINVAL;
2281 		goto done;
2282 	}
2283 
2284 	mr_status->fail_status = 0;
2285 	if (check_mask & IB_MR_CHECK_SIG_STATUS) {
2286 		if (!mmr->sig) {
2287 			ret = -EINVAL;
2288 			pr_err("signature status check requested on a non-signature enabled MR\n");
2289 			goto done;
2290 		}
2291 
2292 		mmr->sig->sig_status_checked = true;
2293 		if (!mmr->sig->sig_err_exists)
2294 			goto done;
2295 
2296 		if (ibmr->lkey == mmr->sig->err_item.key)
2297 			memcpy(&mr_status->sig_err, &mmr->sig->err_item,
2298 			       sizeof(mr_status->sig_err));
2299 		else {
2300 			mr_status->sig_err.err_type = IB_SIG_BAD_GUARD;
2301 			mr_status->sig_err.sig_err_offset = 0;
2302 			mr_status->sig_err.key = mmr->sig->err_item.key;
2303 		}
2304 
2305 		mmr->sig->sig_err_exists = false;
2306 		mr_status->fail_status |= IB_MR_CHECK_SIG_STATUS;
2307 	}
2308 
2309 done:
2310 	return ret;
2311 }
2312 
2313 static int
2314 mlx5_ib_map_pa_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2315 			int data_sg_nents, unsigned int *data_sg_offset,
2316 			struct scatterlist *meta_sg, int meta_sg_nents,
2317 			unsigned int *meta_sg_offset)
2318 {
2319 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2320 	unsigned int sg_offset = 0;
2321 	int n = 0;
2322 
2323 	mr->meta_length = 0;
2324 	if (data_sg_nents == 1) {
2325 		n++;
2326 		mr->mmkey.ndescs = 1;
2327 		if (data_sg_offset)
2328 			sg_offset = *data_sg_offset;
2329 		mr->data_length = sg_dma_len(data_sg) - sg_offset;
2330 		mr->data_iova = sg_dma_address(data_sg) + sg_offset;
2331 		if (meta_sg_nents == 1) {
2332 			n++;
2333 			mr->meta_ndescs = 1;
2334 			if (meta_sg_offset)
2335 				sg_offset = *meta_sg_offset;
2336 			else
2337 				sg_offset = 0;
2338 			mr->meta_length = sg_dma_len(meta_sg) - sg_offset;
2339 			mr->pi_iova = sg_dma_address(meta_sg) + sg_offset;
2340 		}
2341 		ibmr->length = mr->data_length + mr->meta_length;
2342 	}
2343 
2344 	return n;
2345 }
2346 
2347 static int
2348 mlx5_ib_sg_to_klms(struct mlx5_ib_mr *mr,
2349 		   struct scatterlist *sgl,
2350 		   unsigned short sg_nents,
2351 		   unsigned int *sg_offset_p,
2352 		   struct scatterlist *meta_sgl,
2353 		   unsigned short meta_sg_nents,
2354 		   unsigned int *meta_sg_offset_p)
2355 {
2356 	struct scatterlist *sg = sgl;
2357 	struct mlx5_klm *klms = mr->descs;
2358 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2359 	u32 lkey = mr->ibmr.pd->local_dma_lkey;
2360 	int i, j = 0;
2361 
2362 	mr->ibmr.iova = sg_dma_address(sg) + sg_offset;
2363 	mr->ibmr.length = 0;
2364 
2365 	for_each_sg(sgl, sg, sg_nents, i) {
2366 		if (unlikely(i >= mr->max_descs))
2367 			break;
2368 		klms[i].va = cpu_to_be64(sg_dma_address(sg) + sg_offset);
2369 		klms[i].bcount = cpu_to_be32(sg_dma_len(sg) - sg_offset);
2370 		klms[i].key = cpu_to_be32(lkey);
2371 		mr->ibmr.length += sg_dma_len(sg) - sg_offset;
2372 
2373 		sg_offset = 0;
2374 	}
2375 
2376 	if (sg_offset_p)
2377 		*sg_offset_p = sg_offset;
2378 
2379 	mr->mmkey.ndescs = i;
2380 	mr->data_length = mr->ibmr.length;
2381 
2382 	if (meta_sg_nents) {
2383 		sg = meta_sgl;
2384 		sg_offset = meta_sg_offset_p ? *meta_sg_offset_p : 0;
2385 		for_each_sg(meta_sgl, sg, meta_sg_nents, j) {
2386 			if (unlikely(i + j >= mr->max_descs))
2387 				break;
2388 			klms[i + j].va = cpu_to_be64(sg_dma_address(sg) +
2389 						     sg_offset);
2390 			klms[i + j].bcount = cpu_to_be32(sg_dma_len(sg) -
2391 							 sg_offset);
2392 			klms[i + j].key = cpu_to_be32(lkey);
2393 			mr->ibmr.length += sg_dma_len(sg) - sg_offset;
2394 
2395 			sg_offset = 0;
2396 		}
2397 		if (meta_sg_offset_p)
2398 			*meta_sg_offset_p = sg_offset;
2399 
2400 		mr->meta_ndescs = j;
2401 		mr->meta_length = mr->ibmr.length - mr->data_length;
2402 	}
2403 
2404 	return i + j;
2405 }
2406 
2407 static int mlx5_set_page(struct ib_mr *ibmr, u64 addr)
2408 {
2409 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2410 	__be64 *descs;
2411 
2412 	if (unlikely(mr->mmkey.ndescs == mr->max_descs))
2413 		return -ENOMEM;
2414 
2415 	descs = mr->descs;
2416 	descs[mr->mmkey.ndescs++] = cpu_to_be64(addr | MLX5_EN_RD | MLX5_EN_WR);
2417 
2418 	return 0;
2419 }
2420 
2421 static int mlx5_set_page_pi(struct ib_mr *ibmr, u64 addr)
2422 {
2423 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2424 	__be64 *descs;
2425 
2426 	if (unlikely(mr->mmkey.ndescs + mr->meta_ndescs == mr->max_descs))
2427 		return -ENOMEM;
2428 
2429 	descs = mr->descs;
2430 	descs[mr->mmkey.ndescs + mr->meta_ndescs++] =
2431 		cpu_to_be64(addr | MLX5_EN_RD | MLX5_EN_WR);
2432 
2433 	return 0;
2434 }
2435 
2436 static int
2437 mlx5_ib_map_mtt_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2438 			 int data_sg_nents, unsigned int *data_sg_offset,
2439 			 struct scatterlist *meta_sg, int meta_sg_nents,
2440 			 unsigned int *meta_sg_offset)
2441 {
2442 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2443 	struct mlx5_ib_mr *pi_mr = mr->mtt_mr;
2444 	int n;
2445 
2446 	pi_mr->mmkey.ndescs = 0;
2447 	pi_mr->meta_ndescs = 0;
2448 	pi_mr->meta_length = 0;
2449 
2450 	ib_dma_sync_single_for_cpu(ibmr->device, pi_mr->desc_map,
2451 				   pi_mr->desc_size * pi_mr->max_descs,
2452 				   DMA_TO_DEVICE);
2453 
2454 	pi_mr->ibmr.page_size = ibmr->page_size;
2455 	n = ib_sg_to_pages(&pi_mr->ibmr, data_sg, data_sg_nents, data_sg_offset,
2456 			   mlx5_set_page);
2457 	if (n != data_sg_nents)
2458 		return n;
2459 
2460 	pi_mr->data_iova = pi_mr->ibmr.iova;
2461 	pi_mr->data_length = pi_mr->ibmr.length;
2462 	pi_mr->ibmr.length = pi_mr->data_length;
2463 	ibmr->length = pi_mr->data_length;
2464 
2465 	if (meta_sg_nents) {
2466 		u64 page_mask = ~((u64)ibmr->page_size - 1);
2467 		u64 iova = pi_mr->data_iova;
2468 
2469 		n += ib_sg_to_pages(&pi_mr->ibmr, meta_sg, meta_sg_nents,
2470 				    meta_sg_offset, mlx5_set_page_pi);
2471 
2472 		pi_mr->meta_length = pi_mr->ibmr.length;
2473 		/*
2474 		 * PI address for the HW is the offset of the metadata address
2475 		 * relative to the first data page address.
2476 		 * It equals to first data page address + size of data pages +
2477 		 * metadata offset at the first metadata page
2478 		 */
2479 		pi_mr->pi_iova = (iova & page_mask) +
2480 				 pi_mr->mmkey.ndescs * ibmr->page_size +
2481 				 (pi_mr->ibmr.iova & ~page_mask);
2482 		/*
2483 		 * In order to use one MTT MR for data and metadata, we register
2484 		 * also the gaps between the end of the data and the start of
2485 		 * the metadata (the sig MR will verify that the HW will access
2486 		 * to right addresses). This mapping is safe because we use
2487 		 * internal mkey for the registration.
2488 		 */
2489 		pi_mr->ibmr.length = pi_mr->pi_iova + pi_mr->meta_length - iova;
2490 		pi_mr->ibmr.iova = iova;
2491 		ibmr->length += pi_mr->meta_length;
2492 	}
2493 
2494 	ib_dma_sync_single_for_device(ibmr->device, pi_mr->desc_map,
2495 				      pi_mr->desc_size * pi_mr->max_descs,
2496 				      DMA_TO_DEVICE);
2497 
2498 	return n;
2499 }
2500 
2501 static int
2502 mlx5_ib_map_klm_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2503 			 int data_sg_nents, unsigned int *data_sg_offset,
2504 			 struct scatterlist *meta_sg, int meta_sg_nents,
2505 			 unsigned int *meta_sg_offset)
2506 {
2507 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2508 	struct mlx5_ib_mr *pi_mr = mr->klm_mr;
2509 	int n;
2510 
2511 	pi_mr->mmkey.ndescs = 0;
2512 	pi_mr->meta_ndescs = 0;
2513 	pi_mr->meta_length = 0;
2514 
2515 	ib_dma_sync_single_for_cpu(ibmr->device, pi_mr->desc_map,
2516 				   pi_mr->desc_size * pi_mr->max_descs,
2517 				   DMA_TO_DEVICE);
2518 
2519 	n = mlx5_ib_sg_to_klms(pi_mr, data_sg, data_sg_nents, data_sg_offset,
2520 			       meta_sg, meta_sg_nents, meta_sg_offset);
2521 
2522 	ib_dma_sync_single_for_device(ibmr->device, pi_mr->desc_map,
2523 				      pi_mr->desc_size * pi_mr->max_descs,
2524 				      DMA_TO_DEVICE);
2525 
2526 	/* This is zero-based memory region */
2527 	pi_mr->data_iova = 0;
2528 	pi_mr->ibmr.iova = 0;
2529 	pi_mr->pi_iova = pi_mr->data_length;
2530 	ibmr->length = pi_mr->ibmr.length;
2531 
2532 	return n;
2533 }
2534 
2535 int mlx5_ib_map_mr_sg_pi(struct ib_mr *ibmr, struct scatterlist *data_sg,
2536 			 int data_sg_nents, unsigned int *data_sg_offset,
2537 			 struct scatterlist *meta_sg, int meta_sg_nents,
2538 			 unsigned int *meta_sg_offset)
2539 {
2540 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2541 	struct mlx5_ib_mr *pi_mr = NULL;
2542 	int n;
2543 
2544 	WARN_ON(ibmr->type != IB_MR_TYPE_INTEGRITY);
2545 
2546 	mr->mmkey.ndescs = 0;
2547 	mr->data_length = 0;
2548 	mr->data_iova = 0;
2549 	mr->meta_ndescs = 0;
2550 	mr->pi_iova = 0;
2551 	/*
2552 	 * As a performance optimization, if possible, there is no need to
2553 	 * perform UMR operation to register the data/metadata buffers.
2554 	 * First try to map the sg lists to PA descriptors with local_dma_lkey.
2555 	 * Fallback to UMR only in case of a failure.
2556 	 */
2557 	n = mlx5_ib_map_pa_mr_sg_pi(ibmr, data_sg, data_sg_nents,
2558 				    data_sg_offset, meta_sg, meta_sg_nents,
2559 				    meta_sg_offset);
2560 	if (n == data_sg_nents + meta_sg_nents)
2561 		goto out;
2562 	/*
2563 	 * As a performance optimization, if possible, there is no need to map
2564 	 * the sg lists to KLM descriptors. First try to map the sg lists to MTT
2565 	 * descriptors and fallback to KLM only in case of a failure.
2566 	 * It's more efficient for the HW to work with MTT descriptors
2567 	 * (especially in high load).
2568 	 * Use KLM (indirect access) only if it's mandatory.
2569 	 */
2570 	pi_mr = mr->mtt_mr;
2571 	n = mlx5_ib_map_mtt_mr_sg_pi(ibmr, data_sg, data_sg_nents,
2572 				     data_sg_offset, meta_sg, meta_sg_nents,
2573 				     meta_sg_offset);
2574 	if (n == data_sg_nents + meta_sg_nents)
2575 		goto out;
2576 
2577 	pi_mr = mr->klm_mr;
2578 	n = mlx5_ib_map_klm_mr_sg_pi(ibmr, data_sg, data_sg_nents,
2579 				     data_sg_offset, meta_sg, meta_sg_nents,
2580 				     meta_sg_offset);
2581 	if (unlikely(n != data_sg_nents + meta_sg_nents))
2582 		return -ENOMEM;
2583 
2584 out:
2585 	/* This is zero-based memory region */
2586 	ibmr->iova = 0;
2587 	mr->pi_mr = pi_mr;
2588 	if (pi_mr)
2589 		ibmr->sig_attrs->meta_length = pi_mr->meta_length;
2590 	else
2591 		ibmr->sig_attrs->meta_length = mr->meta_length;
2592 
2593 	return 0;
2594 }
2595 
2596 int mlx5_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
2597 		      unsigned int *sg_offset)
2598 {
2599 	struct mlx5_ib_mr *mr = to_mmr(ibmr);
2600 	int n;
2601 
2602 	mr->mmkey.ndescs = 0;
2603 
2604 	ib_dma_sync_single_for_cpu(ibmr->device, mr->desc_map,
2605 				   mr->desc_size * mr->max_descs,
2606 				   DMA_TO_DEVICE);
2607 
2608 	if (mr->access_mode == MLX5_MKC_ACCESS_MODE_KLMS)
2609 		n = mlx5_ib_sg_to_klms(mr, sg, sg_nents, sg_offset, NULL, 0,
2610 				       NULL);
2611 	else
2612 		n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset,
2613 				mlx5_set_page);
2614 
2615 	ib_dma_sync_single_for_device(ibmr->device, mr->desc_map,
2616 				      mr->desc_size * mr->max_descs,
2617 				      DMA_TO_DEVICE);
2618 
2619 	return n;
2620 }
2621