1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005, 2006, 2007, 2008 Mellanox Technologies. All rights reserved.
4  * Copyright (c) 2006, 2007 Cisco Systems, Inc.  All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 #include <linux/errno.h>
36 #include <linux/export.h>
37 #include <linux/slab.h>
38 #include <linux/kernel.h>
39 #include <linux/vmalloc.h>
40 
41 #include <linux/mlx4/cmd.h>
42 
43 #include "mlx4.h"
44 #include "icm.h"
45 
46 static u32 mlx4_buddy_alloc(struct mlx4_buddy *buddy, int order)
47 {
48 	int o;
49 	int m;
50 	u32 seg;
51 
52 	spin_lock(&buddy->lock);
53 
54 	for (o = order; o <= buddy->max_order; ++o)
55 		if (buddy->num_free[o]) {
56 			m = 1 << (buddy->max_order - o);
57 			seg = find_first_bit(buddy->bits[o], m);
58 			if (seg < m)
59 				goto found;
60 		}
61 
62 	spin_unlock(&buddy->lock);
63 	return -1;
64 
65  found:
66 	clear_bit(seg, buddy->bits[o]);
67 	--buddy->num_free[o];
68 
69 	while (o > order) {
70 		--o;
71 		seg <<= 1;
72 		set_bit(seg ^ 1, buddy->bits[o]);
73 		++buddy->num_free[o];
74 	}
75 
76 	spin_unlock(&buddy->lock);
77 
78 	seg <<= order;
79 
80 	return seg;
81 }
82 
83 static void mlx4_buddy_free(struct mlx4_buddy *buddy, u32 seg, int order)
84 {
85 	seg >>= order;
86 
87 	spin_lock(&buddy->lock);
88 
89 	while (test_bit(seg ^ 1, buddy->bits[order])) {
90 		clear_bit(seg ^ 1, buddy->bits[order]);
91 		--buddy->num_free[order];
92 		seg >>= 1;
93 		++order;
94 	}
95 
96 	set_bit(seg, buddy->bits[order]);
97 	++buddy->num_free[order];
98 
99 	spin_unlock(&buddy->lock);
100 }
101 
102 static int mlx4_buddy_init(struct mlx4_buddy *buddy, int max_order)
103 {
104 	int i, s;
105 
106 	buddy->max_order = max_order;
107 	spin_lock_init(&buddy->lock);
108 
109 	buddy->bits = kcalloc(buddy->max_order + 1, sizeof(long *),
110 			      GFP_KERNEL);
111 	buddy->num_free = kcalloc(buddy->max_order + 1, sizeof(*buddy->num_free),
112 				  GFP_KERNEL);
113 	if (!buddy->bits || !buddy->num_free)
114 		goto err_out;
115 
116 	for (i = 0; i <= buddy->max_order; ++i) {
117 		s = BITS_TO_LONGS(1 << (buddy->max_order - i));
118 		buddy->bits[i] = kvmalloc_array(s, sizeof(long), GFP_KERNEL | __GFP_ZERO);
119 		if (!buddy->bits[i])
120 			goto err_out_free;
121 	}
122 
123 	set_bit(0, buddy->bits[buddy->max_order]);
124 	buddy->num_free[buddy->max_order] = 1;
125 
126 	return 0;
127 
128 err_out_free:
129 	for (i = 0; i <= buddy->max_order; ++i)
130 		kvfree(buddy->bits[i]);
131 
132 err_out:
133 	kfree(buddy->bits);
134 	kfree(buddy->num_free);
135 
136 	return -ENOMEM;
137 }
138 
139 static void mlx4_buddy_cleanup(struct mlx4_buddy *buddy)
140 {
141 	int i;
142 
143 	for (i = 0; i <= buddy->max_order; ++i)
144 		kvfree(buddy->bits[i]);
145 
146 	kfree(buddy->bits);
147 	kfree(buddy->num_free);
148 }
149 
150 u32 __mlx4_alloc_mtt_range(struct mlx4_dev *dev, int order)
151 {
152 	struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
153 	u32 seg;
154 	int seg_order;
155 	u32 offset;
156 
157 	seg_order = max_t(int, order - log_mtts_per_seg, 0);
158 
159 	seg = mlx4_buddy_alloc(&mr_table->mtt_buddy, seg_order);
160 	if (seg == -1)
161 		return -1;
162 
163 	offset = seg * (1 << log_mtts_per_seg);
164 
165 	if (mlx4_table_get_range(dev, &mr_table->mtt_table, offset,
166 				 offset + (1 << order) - 1)) {
167 		mlx4_buddy_free(&mr_table->mtt_buddy, seg, seg_order);
168 		return -1;
169 	}
170 
171 	return offset;
172 }
173 
174 static u32 mlx4_alloc_mtt_range(struct mlx4_dev *dev, int order)
175 {
176 	u64 in_param = 0;
177 	u64 out_param;
178 	int err;
179 
180 	if (mlx4_is_mfunc(dev)) {
181 		set_param_l(&in_param, order);
182 		err = mlx4_cmd_imm(dev, in_param, &out_param, RES_MTT,
183 						       RES_OP_RESERVE_AND_MAP,
184 						       MLX4_CMD_ALLOC_RES,
185 						       MLX4_CMD_TIME_CLASS_A,
186 						       MLX4_CMD_WRAPPED);
187 		if (err)
188 			return -1;
189 		return get_param_l(&out_param);
190 	}
191 	return __mlx4_alloc_mtt_range(dev, order);
192 }
193 
194 int mlx4_mtt_init(struct mlx4_dev *dev, int npages, int page_shift,
195 		  struct mlx4_mtt *mtt)
196 {
197 	int i;
198 
199 	if (!npages) {
200 		mtt->order      = -1;
201 		mtt->page_shift = MLX4_ICM_PAGE_SHIFT;
202 		return 0;
203 	} else
204 		mtt->page_shift = page_shift;
205 
206 	for (mtt->order = 0, i = 1; i < npages; i <<= 1)
207 		++mtt->order;
208 
209 	mtt->offset = mlx4_alloc_mtt_range(dev, mtt->order);
210 	if (mtt->offset == -1)
211 		return -ENOMEM;
212 
213 	return 0;
214 }
215 EXPORT_SYMBOL_GPL(mlx4_mtt_init);
216 
217 void __mlx4_free_mtt_range(struct mlx4_dev *dev, u32 offset, int order)
218 {
219 	u32 first_seg;
220 	int seg_order;
221 	struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
222 
223 	seg_order = max_t(int, order - log_mtts_per_seg, 0);
224 	first_seg = offset / (1 << log_mtts_per_seg);
225 
226 	mlx4_buddy_free(&mr_table->mtt_buddy, first_seg, seg_order);
227 	mlx4_table_put_range(dev, &mr_table->mtt_table, offset,
228 			     offset + (1 << order) - 1);
229 }
230 
231 static void mlx4_free_mtt_range(struct mlx4_dev *dev, u32 offset, int order)
232 {
233 	u64 in_param = 0;
234 	int err;
235 
236 	if (mlx4_is_mfunc(dev)) {
237 		set_param_l(&in_param, offset);
238 		set_param_h(&in_param, order);
239 		err = mlx4_cmd(dev, in_param, RES_MTT, RES_OP_RESERVE_AND_MAP,
240 						       MLX4_CMD_FREE_RES,
241 						       MLX4_CMD_TIME_CLASS_A,
242 						       MLX4_CMD_WRAPPED);
243 		if (err)
244 			mlx4_warn(dev, "Failed to free mtt range at:%d order:%d\n",
245 				  offset, order);
246 		return;
247 	}
248 	__mlx4_free_mtt_range(dev, offset, order);
249 }
250 
251 void mlx4_mtt_cleanup(struct mlx4_dev *dev, struct mlx4_mtt *mtt)
252 {
253 	if (mtt->order < 0)
254 		return;
255 
256 	mlx4_free_mtt_range(dev, mtt->offset, mtt->order);
257 }
258 EXPORT_SYMBOL_GPL(mlx4_mtt_cleanup);
259 
260 u64 mlx4_mtt_addr(struct mlx4_dev *dev, struct mlx4_mtt *mtt)
261 {
262 	return (u64) mtt->offset * dev->caps.mtt_entry_sz;
263 }
264 EXPORT_SYMBOL_GPL(mlx4_mtt_addr);
265 
266 static u32 hw_index_to_key(u32 ind)
267 {
268 	return (ind >> 24) | (ind << 8);
269 }
270 
271 static u32 key_to_hw_index(u32 key)
272 {
273 	return (key << 24) | (key >> 8);
274 }
275 
276 static int mlx4_SW2HW_MPT(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
277 			  int mpt_index)
278 {
279 	return mlx4_cmd(dev, mailbox->dma, mpt_index,
280 			0, MLX4_CMD_SW2HW_MPT, MLX4_CMD_TIME_CLASS_B,
281 			MLX4_CMD_WRAPPED);
282 }
283 
284 static int mlx4_HW2SW_MPT(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
285 			  int mpt_index)
286 {
287 	return mlx4_cmd_box(dev, 0, mailbox ? mailbox->dma : 0, mpt_index,
288 			    !mailbox, MLX4_CMD_HW2SW_MPT,
289 			    MLX4_CMD_TIME_CLASS_B, MLX4_CMD_WRAPPED);
290 }
291 
292 /* Must protect against concurrent access */
293 int mlx4_mr_hw_get_mpt(struct mlx4_dev *dev, struct mlx4_mr *mmr,
294 		       struct mlx4_mpt_entry ***mpt_entry)
295 {
296 	int err;
297 	int key = key_to_hw_index(mmr->key) & (dev->caps.num_mpts - 1);
298 	struct mlx4_cmd_mailbox *mailbox = NULL;
299 
300 	if (mmr->enabled != MLX4_MPT_EN_HW)
301 		return -EINVAL;
302 
303 	err = mlx4_HW2SW_MPT(dev, NULL, key);
304 	if (err) {
305 		mlx4_warn(dev, "HW2SW_MPT failed (%d).", err);
306 		mlx4_warn(dev, "Most likely the MR has MWs bound to it.\n");
307 		return err;
308 	}
309 
310 	mmr->enabled = MLX4_MPT_EN_SW;
311 
312 	if (!mlx4_is_mfunc(dev)) {
313 		**mpt_entry = mlx4_table_find(
314 				&mlx4_priv(dev)->mr_table.dmpt_table,
315 				key, NULL);
316 	} else {
317 		mailbox = mlx4_alloc_cmd_mailbox(dev);
318 		if (IS_ERR(mailbox))
319 			return PTR_ERR(mailbox);
320 
321 		err = mlx4_cmd_box(dev, 0, mailbox->dma, key,
322 				   0, MLX4_CMD_QUERY_MPT,
323 				   MLX4_CMD_TIME_CLASS_B,
324 				   MLX4_CMD_WRAPPED);
325 		if (err)
326 			goto free_mailbox;
327 
328 		*mpt_entry = (struct mlx4_mpt_entry **)&mailbox->buf;
329 	}
330 
331 	if (!(*mpt_entry) || !(**mpt_entry)) {
332 		err = -ENOMEM;
333 		goto free_mailbox;
334 	}
335 
336 	return 0;
337 
338 free_mailbox:
339 	mlx4_free_cmd_mailbox(dev, mailbox);
340 	return err;
341 }
342 EXPORT_SYMBOL_GPL(mlx4_mr_hw_get_mpt);
343 
344 int mlx4_mr_hw_write_mpt(struct mlx4_dev *dev, struct mlx4_mr *mmr,
345 			 struct mlx4_mpt_entry **mpt_entry)
346 {
347 	int err;
348 
349 	if (!mlx4_is_mfunc(dev)) {
350 		/* Make sure any changes to this entry are flushed */
351 		wmb();
352 
353 		*(u8 *)(*mpt_entry) = MLX4_MPT_STATUS_HW;
354 
355 		/* Make sure the new status is written */
356 		wmb();
357 
358 		err = mlx4_SYNC_TPT(dev);
359 	} else {
360 		int key = key_to_hw_index(mmr->key) & (dev->caps.num_mpts - 1);
361 
362 		struct mlx4_cmd_mailbox *mailbox =
363 			container_of((void *)mpt_entry, struct mlx4_cmd_mailbox,
364 				     buf);
365 
366 		err = mlx4_SW2HW_MPT(dev, mailbox, key);
367 	}
368 
369 	if (!err) {
370 		mmr->pd = be32_to_cpu((*mpt_entry)->pd_flags) & MLX4_MPT_PD_MASK;
371 		mmr->enabled = MLX4_MPT_EN_HW;
372 	}
373 	return err;
374 }
375 EXPORT_SYMBOL_GPL(mlx4_mr_hw_write_mpt);
376 
377 void mlx4_mr_hw_put_mpt(struct mlx4_dev *dev,
378 			struct mlx4_mpt_entry **mpt_entry)
379 {
380 	if (mlx4_is_mfunc(dev)) {
381 		struct mlx4_cmd_mailbox *mailbox =
382 			container_of((void *)mpt_entry, struct mlx4_cmd_mailbox,
383 				     buf);
384 		mlx4_free_cmd_mailbox(dev, mailbox);
385 	}
386 }
387 EXPORT_SYMBOL_GPL(mlx4_mr_hw_put_mpt);
388 
389 int mlx4_mr_hw_change_pd(struct mlx4_dev *dev, struct mlx4_mpt_entry *mpt_entry,
390 			 u32 pdn)
391 {
392 	u32 pd_flags = be32_to_cpu(mpt_entry->pd_flags) & ~MLX4_MPT_PD_MASK;
393 	/* The wrapper function will put the slave's id here */
394 	if (mlx4_is_mfunc(dev))
395 		pd_flags &= ~MLX4_MPT_PD_VF_MASK;
396 
397 	mpt_entry->pd_flags = cpu_to_be32(pd_flags |
398 					  (pdn & MLX4_MPT_PD_MASK)
399 					  | MLX4_MPT_PD_FLAG_EN_INV);
400 	return 0;
401 }
402 EXPORT_SYMBOL_GPL(mlx4_mr_hw_change_pd);
403 
404 int mlx4_mr_hw_change_access(struct mlx4_dev *dev,
405 			     struct mlx4_mpt_entry *mpt_entry,
406 			     u32 access)
407 {
408 	u32 flags = (be32_to_cpu(mpt_entry->flags) & ~MLX4_PERM_MASK) |
409 		    (access & MLX4_PERM_MASK);
410 
411 	mpt_entry->flags = cpu_to_be32(flags);
412 	return 0;
413 }
414 EXPORT_SYMBOL_GPL(mlx4_mr_hw_change_access);
415 
416 static int mlx4_mr_alloc_reserved(struct mlx4_dev *dev, u32 mridx, u32 pd,
417 			   u64 iova, u64 size, u32 access, int npages,
418 			   int page_shift, struct mlx4_mr *mr)
419 {
420 	mr->iova       = iova;
421 	mr->size       = size;
422 	mr->pd	       = pd;
423 	mr->access     = access;
424 	mr->enabled    = MLX4_MPT_DISABLED;
425 	mr->key	       = hw_index_to_key(mridx);
426 
427 	return mlx4_mtt_init(dev, npages, page_shift, &mr->mtt);
428 }
429 
430 static int mlx4_WRITE_MTT(struct mlx4_dev *dev,
431 			  struct mlx4_cmd_mailbox *mailbox,
432 			  int num_entries)
433 {
434 	return mlx4_cmd(dev, mailbox->dma, num_entries, 0, MLX4_CMD_WRITE_MTT,
435 			MLX4_CMD_TIME_CLASS_A,  MLX4_CMD_WRAPPED);
436 }
437 
438 int __mlx4_mpt_reserve(struct mlx4_dev *dev)
439 {
440 	struct mlx4_priv *priv = mlx4_priv(dev);
441 
442 	return mlx4_bitmap_alloc(&priv->mr_table.mpt_bitmap);
443 }
444 
445 static int mlx4_mpt_reserve(struct mlx4_dev *dev)
446 {
447 	u64 out_param;
448 
449 	if (mlx4_is_mfunc(dev)) {
450 		if (mlx4_cmd_imm(dev, 0, &out_param, RES_MPT, RES_OP_RESERVE,
451 				   MLX4_CMD_ALLOC_RES,
452 				   MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED))
453 			return -1;
454 		return get_param_l(&out_param);
455 	}
456 	return  __mlx4_mpt_reserve(dev);
457 }
458 
459 void __mlx4_mpt_release(struct mlx4_dev *dev, u32 index)
460 {
461 	struct mlx4_priv *priv = mlx4_priv(dev);
462 
463 	mlx4_bitmap_free(&priv->mr_table.mpt_bitmap, index, MLX4_NO_RR);
464 }
465 
466 static void mlx4_mpt_release(struct mlx4_dev *dev, u32 index)
467 {
468 	u64 in_param = 0;
469 
470 	if (mlx4_is_mfunc(dev)) {
471 		set_param_l(&in_param, index);
472 		if (mlx4_cmd(dev, in_param, RES_MPT, RES_OP_RESERVE,
473 			       MLX4_CMD_FREE_RES,
474 			       MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED))
475 			mlx4_warn(dev, "Failed to release mr index:%d\n",
476 				  index);
477 		return;
478 	}
479 	__mlx4_mpt_release(dev, index);
480 }
481 
482 int __mlx4_mpt_alloc_icm(struct mlx4_dev *dev, u32 index)
483 {
484 	struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
485 
486 	return mlx4_table_get(dev, &mr_table->dmpt_table, index);
487 }
488 
489 static int mlx4_mpt_alloc_icm(struct mlx4_dev *dev, u32 index)
490 {
491 	u64 param = 0;
492 
493 	if (mlx4_is_mfunc(dev)) {
494 		set_param_l(&param, index);
495 		return mlx4_cmd_imm(dev, param, &param, RES_MPT, RES_OP_MAP_ICM,
496 							MLX4_CMD_ALLOC_RES,
497 							MLX4_CMD_TIME_CLASS_A,
498 							MLX4_CMD_WRAPPED);
499 	}
500 	return __mlx4_mpt_alloc_icm(dev, index);
501 }
502 
503 void __mlx4_mpt_free_icm(struct mlx4_dev *dev, u32 index)
504 {
505 	struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
506 
507 	mlx4_table_put(dev, &mr_table->dmpt_table, index);
508 }
509 
510 static void mlx4_mpt_free_icm(struct mlx4_dev *dev, u32 index)
511 {
512 	u64 in_param = 0;
513 
514 	if (mlx4_is_mfunc(dev)) {
515 		set_param_l(&in_param, index);
516 		if (mlx4_cmd(dev, in_param, RES_MPT, RES_OP_MAP_ICM,
517 			     MLX4_CMD_FREE_RES, MLX4_CMD_TIME_CLASS_A,
518 			     MLX4_CMD_WRAPPED))
519 			mlx4_warn(dev, "Failed to free icm of mr index:%d\n",
520 				  index);
521 		return;
522 	}
523 	return __mlx4_mpt_free_icm(dev, index);
524 }
525 
526 int mlx4_mr_alloc(struct mlx4_dev *dev, u32 pd, u64 iova, u64 size, u32 access,
527 		  int npages, int page_shift, struct mlx4_mr *mr)
528 {
529 	u32 index;
530 	int err;
531 
532 	index = mlx4_mpt_reserve(dev);
533 	if (index == -1)
534 		return -ENOMEM;
535 
536 	err = mlx4_mr_alloc_reserved(dev, index, pd, iova, size,
537 				     access, npages, page_shift, mr);
538 	if (err)
539 		mlx4_mpt_release(dev, index);
540 
541 	return err;
542 }
543 EXPORT_SYMBOL_GPL(mlx4_mr_alloc);
544 
545 static int mlx4_mr_free_reserved(struct mlx4_dev *dev, struct mlx4_mr *mr)
546 {
547 	int err;
548 
549 	if (mr->enabled == MLX4_MPT_EN_HW) {
550 		err = mlx4_HW2SW_MPT(dev, NULL,
551 				     key_to_hw_index(mr->key) &
552 				     (dev->caps.num_mpts - 1));
553 		if (err) {
554 			mlx4_warn(dev, "HW2SW_MPT failed (%d), MR has MWs bound to it\n",
555 				  err);
556 			return err;
557 		}
558 
559 		mr->enabled = MLX4_MPT_EN_SW;
560 	}
561 	mlx4_mtt_cleanup(dev, &mr->mtt);
562 
563 	return 0;
564 }
565 
566 int mlx4_mr_free(struct mlx4_dev *dev, struct mlx4_mr *mr)
567 {
568 	int ret;
569 
570 	ret = mlx4_mr_free_reserved(dev, mr);
571 	if (ret)
572 		return ret;
573 	if (mr->enabled)
574 		mlx4_mpt_free_icm(dev, key_to_hw_index(mr->key));
575 	mlx4_mpt_release(dev, key_to_hw_index(mr->key));
576 
577 	return 0;
578 }
579 EXPORT_SYMBOL_GPL(mlx4_mr_free);
580 
581 void mlx4_mr_rereg_mem_cleanup(struct mlx4_dev *dev, struct mlx4_mr *mr)
582 {
583 	mlx4_mtt_cleanup(dev, &mr->mtt);
584 	mr->mtt.order = -1;
585 }
586 EXPORT_SYMBOL_GPL(mlx4_mr_rereg_mem_cleanup);
587 
588 int mlx4_mr_rereg_mem_write(struct mlx4_dev *dev, struct mlx4_mr *mr,
589 			    u64 iova, u64 size, int npages,
590 			    int page_shift, struct mlx4_mpt_entry *mpt_entry)
591 {
592 	int err;
593 
594 	err = mlx4_mtt_init(dev, npages, page_shift, &mr->mtt);
595 	if (err)
596 		return err;
597 
598 	mpt_entry->start       = cpu_to_be64(iova);
599 	mpt_entry->length      = cpu_to_be64(size);
600 	mpt_entry->entity_size = cpu_to_be32(page_shift);
601 	mpt_entry->flags    &= ~(cpu_to_be32(MLX4_MPT_FLAG_FREE |
602 					   MLX4_MPT_FLAG_SW_OWNS));
603 	if (mr->mtt.order < 0) {
604 		mpt_entry->flags |= cpu_to_be32(MLX4_MPT_FLAG_PHYSICAL);
605 		mpt_entry->mtt_addr = 0;
606 	} else {
607 		mpt_entry->mtt_addr = cpu_to_be64(mlx4_mtt_addr(dev,
608 						  &mr->mtt));
609 		if (mr->mtt.page_shift == 0)
610 			mpt_entry->mtt_sz    = cpu_to_be32(1 << mr->mtt.order);
611 	}
612 	if (mr->mtt.order >= 0 && mr->mtt.page_shift == 0) {
613 		/* fast register MR in free state */
614 		mpt_entry->flags    |= cpu_to_be32(MLX4_MPT_FLAG_FREE);
615 		mpt_entry->pd_flags |= cpu_to_be32(MLX4_MPT_PD_FLAG_FAST_REG |
616 						   MLX4_MPT_PD_FLAG_RAE);
617 	} else {
618 		mpt_entry->flags    |= cpu_to_be32(MLX4_MPT_FLAG_SW_OWNS);
619 	}
620 	mr->enabled = MLX4_MPT_EN_SW;
621 
622 	return 0;
623 }
624 EXPORT_SYMBOL_GPL(mlx4_mr_rereg_mem_write);
625 
626 int mlx4_mr_enable(struct mlx4_dev *dev, struct mlx4_mr *mr)
627 {
628 	struct mlx4_cmd_mailbox *mailbox;
629 	struct mlx4_mpt_entry *mpt_entry;
630 	int err;
631 
632 	err = mlx4_mpt_alloc_icm(dev, key_to_hw_index(mr->key));
633 	if (err)
634 		return err;
635 
636 	mailbox = mlx4_alloc_cmd_mailbox(dev);
637 	if (IS_ERR(mailbox)) {
638 		err = PTR_ERR(mailbox);
639 		goto err_table;
640 	}
641 	mpt_entry = mailbox->buf;
642 	mpt_entry->flags = cpu_to_be32(MLX4_MPT_FLAG_MIO	 |
643 				       MLX4_MPT_FLAG_REGION	 |
644 				       mr->access);
645 
646 	mpt_entry->key	       = cpu_to_be32(key_to_hw_index(mr->key));
647 	mpt_entry->pd_flags    = cpu_to_be32(mr->pd | MLX4_MPT_PD_FLAG_EN_INV);
648 	mpt_entry->start       = cpu_to_be64(mr->iova);
649 	mpt_entry->length      = cpu_to_be64(mr->size);
650 	mpt_entry->entity_size = cpu_to_be32(mr->mtt.page_shift);
651 
652 	if (mr->mtt.order < 0) {
653 		mpt_entry->flags |= cpu_to_be32(MLX4_MPT_FLAG_PHYSICAL);
654 		mpt_entry->mtt_addr = 0;
655 	} else {
656 		mpt_entry->mtt_addr = cpu_to_be64(mlx4_mtt_addr(dev,
657 						  &mr->mtt));
658 	}
659 
660 	if (mr->mtt.order >= 0 && mr->mtt.page_shift == 0) {
661 		/* fast register MR in free state */
662 		mpt_entry->flags    |= cpu_to_be32(MLX4_MPT_FLAG_FREE);
663 		mpt_entry->pd_flags |= cpu_to_be32(MLX4_MPT_PD_FLAG_FAST_REG |
664 						   MLX4_MPT_PD_FLAG_RAE);
665 		mpt_entry->mtt_sz    = cpu_to_be32(1 << mr->mtt.order);
666 	} else {
667 		mpt_entry->flags    |= cpu_to_be32(MLX4_MPT_FLAG_SW_OWNS);
668 	}
669 
670 	err = mlx4_SW2HW_MPT(dev, mailbox,
671 			     key_to_hw_index(mr->key) & (dev->caps.num_mpts - 1));
672 	if (err) {
673 		mlx4_warn(dev, "SW2HW_MPT failed (%d)\n", err);
674 		goto err_cmd;
675 	}
676 	mr->enabled = MLX4_MPT_EN_HW;
677 
678 	mlx4_free_cmd_mailbox(dev, mailbox);
679 
680 	return 0;
681 
682 err_cmd:
683 	mlx4_free_cmd_mailbox(dev, mailbox);
684 
685 err_table:
686 	mlx4_mpt_free_icm(dev, key_to_hw_index(mr->key));
687 	return err;
688 }
689 EXPORT_SYMBOL_GPL(mlx4_mr_enable);
690 
691 static int mlx4_write_mtt_chunk(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
692 				int start_index, int npages, u64 *page_list)
693 {
694 	struct mlx4_priv *priv = mlx4_priv(dev);
695 	__be64 *mtts;
696 	dma_addr_t dma_handle;
697 	int i;
698 
699 	mtts = mlx4_table_find(&priv->mr_table.mtt_table, mtt->offset +
700 			       start_index, &dma_handle);
701 
702 	if (!mtts)
703 		return -ENOMEM;
704 
705 	dma_sync_single_for_cpu(&dev->persist->pdev->dev, dma_handle,
706 				npages * sizeof(u64), DMA_TO_DEVICE);
707 
708 	for (i = 0; i < npages; ++i)
709 		mtts[i] = cpu_to_be64(page_list[i] | MLX4_MTT_FLAG_PRESENT);
710 
711 	dma_sync_single_for_device(&dev->persist->pdev->dev, dma_handle,
712 				   npages * sizeof(u64), DMA_TO_DEVICE);
713 
714 	return 0;
715 }
716 
717 int __mlx4_write_mtt(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
718 		     int start_index, int npages, u64 *page_list)
719 {
720 	int err = 0;
721 	int chunk;
722 	int mtts_per_page;
723 	int max_mtts_first_page;
724 
725 	/* compute how may mtts fit in the first page */
726 	mtts_per_page = PAGE_SIZE / sizeof(u64);
727 	max_mtts_first_page = mtts_per_page - (mtt->offset + start_index)
728 			      % mtts_per_page;
729 
730 	chunk = min_t(int, max_mtts_first_page, npages);
731 
732 	while (npages > 0) {
733 		err = mlx4_write_mtt_chunk(dev, mtt, start_index, chunk, page_list);
734 		if (err)
735 			return err;
736 		npages      -= chunk;
737 		start_index += chunk;
738 		page_list   += chunk;
739 
740 		chunk = min_t(int, mtts_per_page, npages);
741 	}
742 	return err;
743 }
744 
745 int mlx4_write_mtt(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
746 		   int start_index, int npages, u64 *page_list)
747 {
748 	struct mlx4_cmd_mailbox *mailbox = NULL;
749 	__be64 *inbox = NULL;
750 	int chunk;
751 	int err = 0;
752 	int i;
753 
754 	if (mtt->order < 0)
755 		return -EINVAL;
756 
757 	if (mlx4_is_mfunc(dev)) {
758 		mailbox = mlx4_alloc_cmd_mailbox(dev);
759 		if (IS_ERR(mailbox))
760 			return PTR_ERR(mailbox);
761 		inbox = mailbox->buf;
762 
763 		while (npages > 0) {
764 			chunk = min_t(int, MLX4_MAILBOX_SIZE / sizeof(u64) - 2,
765 				      npages);
766 			inbox[0] = cpu_to_be64(mtt->offset + start_index);
767 			inbox[1] = 0;
768 			for (i = 0; i < chunk; ++i)
769 				inbox[i + 2] = cpu_to_be64(page_list[i] |
770 					       MLX4_MTT_FLAG_PRESENT);
771 			err = mlx4_WRITE_MTT(dev, mailbox, chunk);
772 			if (err) {
773 				mlx4_free_cmd_mailbox(dev, mailbox);
774 				return err;
775 			}
776 
777 			npages      -= chunk;
778 			start_index += chunk;
779 			page_list   += chunk;
780 		}
781 		mlx4_free_cmd_mailbox(dev, mailbox);
782 		return err;
783 	}
784 
785 	return __mlx4_write_mtt(dev, mtt, start_index, npages, page_list);
786 }
787 EXPORT_SYMBOL_GPL(mlx4_write_mtt);
788 
789 int mlx4_buf_write_mtt(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
790 		       struct mlx4_buf *buf)
791 {
792 	u64 *page_list;
793 	int err;
794 	int i;
795 
796 	page_list = kcalloc(buf->npages, sizeof(*page_list), GFP_KERNEL);
797 	if (!page_list)
798 		return -ENOMEM;
799 
800 	for (i = 0; i < buf->npages; ++i)
801 		if (buf->nbufs == 1)
802 			page_list[i] = buf->direct.map + (i << buf->page_shift);
803 		else
804 			page_list[i] = buf->page_list[i].map;
805 
806 	err = mlx4_write_mtt(dev, mtt, 0, buf->npages, page_list);
807 
808 	kfree(page_list);
809 	return err;
810 }
811 EXPORT_SYMBOL_GPL(mlx4_buf_write_mtt);
812 
813 int mlx4_mw_alloc(struct mlx4_dev *dev, u32 pd, enum mlx4_mw_type type,
814 		  struct mlx4_mw *mw)
815 {
816 	u32 index;
817 
818 	if ((type == MLX4_MW_TYPE_1 &&
819 	     !(dev->caps.flags & MLX4_DEV_CAP_FLAG_MEM_WINDOW)) ||
820 	     (type == MLX4_MW_TYPE_2 &&
821 	     !(dev->caps.bmme_flags & MLX4_BMME_FLAG_TYPE_2_WIN)))
822 		return -EOPNOTSUPP;
823 
824 	index = mlx4_mpt_reserve(dev);
825 	if (index == -1)
826 		return -ENOMEM;
827 
828 	mw->key	    = hw_index_to_key(index);
829 	mw->pd      = pd;
830 	mw->type    = type;
831 	mw->enabled = MLX4_MPT_DISABLED;
832 
833 	return 0;
834 }
835 EXPORT_SYMBOL_GPL(mlx4_mw_alloc);
836 
837 int mlx4_mw_enable(struct mlx4_dev *dev, struct mlx4_mw *mw)
838 {
839 	struct mlx4_cmd_mailbox *mailbox;
840 	struct mlx4_mpt_entry *mpt_entry;
841 	int err;
842 
843 	err = mlx4_mpt_alloc_icm(dev, key_to_hw_index(mw->key));
844 	if (err)
845 		return err;
846 
847 	mailbox = mlx4_alloc_cmd_mailbox(dev);
848 	if (IS_ERR(mailbox)) {
849 		err = PTR_ERR(mailbox);
850 		goto err_table;
851 	}
852 	mpt_entry = mailbox->buf;
853 
854 	/* Note that the MLX4_MPT_FLAG_REGION bit in mpt_entry->flags is turned
855 	 * off, thus creating a memory window and not a memory region.
856 	 */
857 	mpt_entry->key	       = cpu_to_be32(key_to_hw_index(mw->key));
858 	mpt_entry->pd_flags    = cpu_to_be32(mw->pd);
859 	if (mw->type == MLX4_MW_TYPE_2) {
860 		mpt_entry->flags    |= cpu_to_be32(MLX4_MPT_FLAG_FREE);
861 		mpt_entry->qpn       = cpu_to_be32(MLX4_MPT_QP_FLAG_BOUND_QP);
862 		mpt_entry->pd_flags |= cpu_to_be32(MLX4_MPT_PD_FLAG_EN_INV);
863 	}
864 
865 	err = mlx4_SW2HW_MPT(dev, mailbox,
866 			     key_to_hw_index(mw->key) &
867 			     (dev->caps.num_mpts - 1));
868 	if (err) {
869 		mlx4_warn(dev, "SW2HW_MPT failed (%d)\n", err);
870 		goto err_cmd;
871 	}
872 	mw->enabled = MLX4_MPT_EN_HW;
873 
874 	mlx4_free_cmd_mailbox(dev, mailbox);
875 
876 	return 0;
877 
878 err_cmd:
879 	mlx4_free_cmd_mailbox(dev, mailbox);
880 
881 err_table:
882 	mlx4_mpt_free_icm(dev, key_to_hw_index(mw->key));
883 	return err;
884 }
885 EXPORT_SYMBOL_GPL(mlx4_mw_enable);
886 
887 void mlx4_mw_free(struct mlx4_dev *dev, struct mlx4_mw *mw)
888 {
889 	int err;
890 
891 	if (mw->enabled == MLX4_MPT_EN_HW) {
892 		err = mlx4_HW2SW_MPT(dev, NULL,
893 				     key_to_hw_index(mw->key) &
894 				     (dev->caps.num_mpts - 1));
895 		if (err)
896 			mlx4_warn(dev, "xxx HW2SW_MPT failed (%d)\n", err);
897 
898 		mw->enabled = MLX4_MPT_EN_SW;
899 	}
900 	if (mw->enabled)
901 		mlx4_mpt_free_icm(dev, key_to_hw_index(mw->key));
902 	mlx4_mpt_release(dev, key_to_hw_index(mw->key));
903 }
904 EXPORT_SYMBOL_GPL(mlx4_mw_free);
905 
906 int mlx4_init_mr_table(struct mlx4_dev *dev)
907 {
908 	struct mlx4_priv *priv = mlx4_priv(dev);
909 	struct mlx4_mr_table *mr_table = &priv->mr_table;
910 	int err;
911 
912 	/* Nothing to do for slaves - all MR handling is forwarded
913 	* to the master */
914 	if (mlx4_is_slave(dev))
915 		return 0;
916 
917 	if (!is_power_of_2(dev->caps.num_mpts))
918 		return -EINVAL;
919 
920 	err = mlx4_bitmap_init(&mr_table->mpt_bitmap, dev->caps.num_mpts,
921 			       ~0, dev->caps.reserved_mrws, 0);
922 	if (err)
923 		return err;
924 
925 	err = mlx4_buddy_init(&mr_table->mtt_buddy,
926 			      ilog2((u32)dev->caps.num_mtts /
927 			      (1 << log_mtts_per_seg)));
928 	if (err)
929 		goto err_buddy;
930 
931 	if (dev->caps.reserved_mtts) {
932 		priv->reserved_mtts =
933 			mlx4_alloc_mtt_range(dev,
934 					     fls(dev->caps.reserved_mtts - 1));
935 		if (priv->reserved_mtts < 0) {
936 			mlx4_warn(dev, "MTT table of order %u is too small\n",
937 				  mr_table->mtt_buddy.max_order);
938 			err = -ENOMEM;
939 			goto err_reserve_mtts;
940 		}
941 	}
942 
943 	return 0;
944 
945 err_reserve_mtts:
946 	mlx4_buddy_cleanup(&mr_table->mtt_buddy);
947 
948 err_buddy:
949 	mlx4_bitmap_cleanup(&mr_table->mpt_bitmap);
950 
951 	return err;
952 }
953 
954 void mlx4_cleanup_mr_table(struct mlx4_dev *dev)
955 {
956 	struct mlx4_priv *priv = mlx4_priv(dev);
957 	struct mlx4_mr_table *mr_table = &priv->mr_table;
958 
959 	if (mlx4_is_slave(dev))
960 		return;
961 	if (priv->reserved_mtts >= 0)
962 		mlx4_free_mtt_range(dev, priv->reserved_mtts,
963 				    fls(dev->caps.reserved_mtts - 1));
964 	mlx4_buddy_cleanup(&mr_table->mtt_buddy);
965 	mlx4_bitmap_cleanup(&mr_table->mpt_bitmap);
966 }
967 
968 static inline int mlx4_check_fmr(struct mlx4_fmr *fmr, u64 *page_list,
969 				  int npages, u64 iova)
970 {
971 	int i, page_mask;
972 
973 	if (npages > fmr->max_pages)
974 		return -EINVAL;
975 
976 	page_mask = (1 << fmr->page_shift) - 1;
977 
978 	/* We are getting page lists, so va must be page aligned. */
979 	if (iova & page_mask)
980 		return -EINVAL;
981 
982 	/* Trust the user not to pass misaligned data in page_list */
983 	if (0)
984 		for (i = 0; i < npages; ++i) {
985 			if (page_list[i] & ~page_mask)
986 				return -EINVAL;
987 		}
988 
989 	if (fmr->maps >= fmr->max_maps)
990 		return -EINVAL;
991 
992 	return 0;
993 }
994 
995 int mlx4_map_phys_fmr(struct mlx4_dev *dev, struct mlx4_fmr *fmr, u64 *page_list,
996 		      int npages, u64 iova, u32 *lkey, u32 *rkey)
997 {
998 	u32 key;
999 	int i, err;
1000 
1001 	err = mlx4_check_fmr(fmr, page_list, npages, iova);
1002 	if (err)
1003 		return err;
1004 
1005 	++fmr->maps;
1006 
1007 	key = key_to_hw_index(fmr->mr.key);
1008 	key += dev->caps.num_mpts;
1009 	*lkey = *rkey = fmr->mr.key = hw_index_to_key(key);
1010 
1011 	*(u8 *) fmr->mpt = MLX4_MPT_STATUS_SW;
1012 
1013 	/* Make sure MPT status is visible before writing MTT entries */
1014 	wmb();
1015 
1016 	dma_sync_single_for_cpu(&dev->persist->pdev->dev, fmr->dma_handle,
1017 				npages * sizeof(u64), DMA_TO_DEVICE);
1018 
1019 	for (i = 0; i < npages; ++i)
1020 		fmr->mtts[i] = cpu_to_be64(page_list[i] | MLX4_MTT_FLAG_PRESENT);
1021 
1022 	dma_sync_single_for_device(&dev->persist->pdev->dev, fmr->dma_handle,
1023 				   npages * sizeof(u64), DMA_TO_DEVICE);
1024 
1025 	fmr->mpt->key    = cpu_to_be32(key);
1026 	fmr->mpt->lkey   = cpu_to_be32(key);
1027 	fmr->mpt->length = cpu_to_be64(npages * (1ull << fmr->page_shift));
1028 	fmr->mpt->start  = cpu_to_be64(iova);
1029 
1030 	/* Make MTT entries are visible before setting MPT status */
1031 	wmb();
1032 
1033 	*(u8 *) fmr->mpt = MLX4_MPT_STATUS_HW;
1034 
1035 	/* Make sure MPT status is visible before consumer can use FMR */
1036 	wmb();
1037 
1038 	return 0;
1039 }
1040 EXPORT_SYMBOL_GPL(mlx4_map_phys_fmr);
1041 
1042 int mlx4_fmr_alloc(struct mlx4_dev *dev, u32 pd, u32 access, int max_pages,
1043 		   int max_maps, u8 page_shift, struct mlx4_fmr *fmr)
1044 {
1045 	struct mlx4_priv *priv = mlx4_priv(dev);
1046 	int err = -ENOMEM;
1047 
1048 	if (max_maps > dev->caps.max_fmr_maps)
1049 		return -EINVAL;
1050 
1051 	if (page_shift < (ffs(dev->caps.page_size_cap) - 1) || page_shift >= 32)
1052 		return -EINVAL;
1053 
1054 	/* All MTTs must fit in the same page */
1055 	if (max_pages * sizeof(*fmr->mtts) > PAGE_SIZE)
1056 		return -EINVAL;
1057 
1058 	fmr->page_shift = page_shift;
1059 	fmr->max_pages  = max_pages;
1060 	fmr->max_maps   = max_maps;
1061 	fmr->maps = 0;
1062 
1063 	err = mlx4_mr_alloc(dev, pd, 0, 0, access, max_pages,
1064 			    page_shift, &fmr->mr);
1065 	if (err)
1066 		return err;
1067 
1068 	fmr->mtts = mlx4_table_find(&priv->mr_table.mtt_table,
1069 				    fmr->mr.mtt.offset,
1070 				    &fmr->dma_handle);
1071 
1072 	if (!fmr->mtts) {
1073 		err = -ENOMEM;
1074 		goto err_free;
1075 	}
1076 
1077 	return 0;
1078 
1079 err_free:
1080 	(void) mlx4_mr_free(dev, &fmr->mr);
1081 	return err;
1082 }
1083 EXPORT_SYMBOL_GPL(mlx4_fmr_alloc);
1084 
1085 int mlx4_fmr_enable(struct mlx4_dev *dev, struct mlx4_fmr *fmr)
1086 {
1087 	struct mlx4_priv *priv = mlx4_priv(dev);
1088 	int err;
1089 
1090 	err = mlx4_mr_enable(dev, &fmr->mr);
1091 	if (err)
1092 		return err;
1093 
1094 	fmr->mpt = mlx4_table_find(&priv->mr_table.dmpt_table,
1095 				    key_to_hw_index(fmr->mr.key), NULL);
1096 	if (!fmr->mpt)
1097 		return -ENOMEM;
1098 
1099 	return 0;
1100 }
1101 EXPORT_SYMBOL_GPL(mlx4_fmr_enable);
1102 
1103 void mlx4_fmr_unmap(struct mlx4_dev *dev, struct mlx4_fmr *fmr,
1104 		    u32 *lkey, u32 *rkey)
1105 {
1106 	struct mlx4_cmd_mailbox *mailbox;
1107 	int err;
1108 
1109 	if (!fmr->maps)
1110 		return;
1111 
1112 	fmr->maps = 0;
1113 
1114 	mailbox = mlx4_alloc_cmd_mailbox(dev);
1115 	if (IS_ERR(mailbox)) {
1116 		err = PTR_ERR(mailbox);
1117 		pr_warn("mlx4_ib: mlx4_alloc_cmd_mailbox failed (%d)\n", err);
1118 		return;
1119 	}
1120 
1121 	err = mlx4_HW2SW_MPT(dev, NULL,
1122 			     key_to_hw_index(fmr->mr.key) &
1123 			     (dev->caps.num_mpts - 1));
1124 	mlx4_free_cmd_mailbox(dev, mailbox);
1125 	if (err) {
1126 		pr_warn("mlx4_ib: mlx4_HW2SW_MPT failed (%d)\n", err);
1127 		return;
1128 	}
1129 	fmr->mr.enabled = MLX4_MPT_EN_SW;
1130 }
1131 EXPORT_SYMBOL_GPL(mlx4_fmr_unmap);
1132 
1133 int mlx4_fmr_free(struct mlx4_dev *dev, struct mlx4_fmr *fmr)
1134 {
1135 	int ret;
1136 
1137 	if (fmr->maps)
1138 		return -EBUSY;
1139 
1140 	ret = mlx4_mr_free(dev, &fmr->mr);
1141 	if (ret)
1142 		return ret;
1143 	fmr->mr.enabled = MLX4_MPT_DISABLED;
1144 
1145 	return 0;
1146 }
1147 EXPORT_SYMBOL_GPL(mlx4_fmr_free);
1148 
1149 int mlx4_SYNC_TPT(struct mlx4_dev *dev)
1150 {
1151 	return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_SYNC_TPT,
1152 			MLX4_CMD_TIME_CLASS_A, MLX4_CMD_NATIVE);
1153 }
1154 EXPORT_SYMBOL_GPL(mlx4_SYNC_TPT);
1155