xref: /linux/drivers/net/ethernet/mellanox/mlx4/mr.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
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] = kcalloc(s, sizeof (long), GFP_KERNEL | __GFP_NOWARN);
119 		if (!buddy->bits[i]) {
120 			buddy->bits[i] = vzalloc(s * sizeof(long));
121 			if (!buddy->bits[i])
122 				goto err_out_free;
123 		}
124 	}
125 
126 	set_bit(0, buddy->bits[buddy->max_order]);
127 	buddy->num_free[buddy->max_order] = 1;
128 
129 	return 0;
130 
131 err_out_free:
132 	for (i = 0; i <= buddy->max_order; ++i)
133 		kvfree(buddy->bits[i]);
134 
135 err_out:
136 	kfree(buddy->bits);
137 	kfree(buddy->num_free);
138 
139 	return -ENOMEM;
140 }
141 
142 static void mlx4_buddy_cleanup(struct mlx4_buddy *buddy)
143 {
144 	int i;
145 
146 	for (i = 0; i <= buddy->max_order; ++i)
147 		kvfree(buddy->bits[i]);
148 
149 	kfree(buddy->bits);
150 	kfree(buddy->num_free);
151 }
152 
153 u32 __mlx4_alloc_mtt_range(struct mlx4_dev *dev, int order)
154 {
155 	struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
156 	u32 seg;
157 	int seg_order;
158 	u32 offset;
159 
160 	seg_order = max_t(int, order - log_mtts_per_seg, 0);
161 
162 	seg = mlx4_buddy_alloc(&mr_table->mtt_buddy, seg_order);
163 	if (seg == -1)
164 		return -1;
165 
166 	offset = seg * (1 << log_mtts_per_seg);
167 
168 	if (mlx4_table_get_range(dev, &mr_table->mtt_table, offset,
169 				 offset + (1 << order) - 1)) {
170 		mlx4_buddy_free(&mr_table->mtt_buddy, seg, seg_order);
171 		return -1;
172 	}
173 
174 	return offset;
175 }
176 
177 static u32 mlx4_alloc_mtt_range(struct mlx4_dev *dev, int order)
178 {
179 	u64 in_param = 0;
180 	u64 out_param;
181 	int err;
182 
183 	if (mlx4_is_mfunc(dev)) {
184 		set_param_l(&in_param, order);
185 		err = mlx4_cmd_imm(dev, in_param, &out_param, RES_MTT,
186 						       RES_OP_RESERVE_AND_MAP,
187 						       MLX4_CMD_ALLOC_RES,
188 						       MLX4_CMD_TIME_CLASS_A,
189 						       MLX4_CMD_WRAPPED);
190 		if (err)
191 			return -1;
192 		return get_param_l(&out_param);
193 	}
194 	return __mlx4_alloc_mtt_range(dev, order);
195 }
196 
197 int mlx4_mtt_init(struct mlx4_dev *dev, int npages, int page_shift,
198 		  struct mlx4_mtt *mtt)
199 {
200 	int i;
201 
202 	if (!npages) {
203 		mtt->order      = -1;
204 		mtt->page_shift = MLX4_ICM_PAGE_SHIFT;
205 		return 0;
206 	} else
207 		mtt->page_shift = page_shift;
208 
209 	for (mtt->order = 0, i = 1; i < npages; i <<= 1)
210 		++mtt->order;
211 
212 	mtt->offset = mlx4_alloc_mtt_range(dev, mtt->order);
213 	if (mtt->offset == -1)
214 		return -ENOMEM;
215 
216 	return 0;
217 }
218 EXPORT_SYMBOL_GPL(mlx4_mtt_init);
219 
220 void __mlx4_free_mtt_range(struct mlx4_dev *dev, u32 offset, int order)
221 {
222 	u32 first_seg;
223 	int seg_order;
224 	struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
225 
226 	seg_order = max_t(int, order - log_mtts_per_seg, 0);
227 	first_seg = offset / (1 << log_mtts_per_seg);
228 
229 	mlx4_buddy_free(&mr_table->mtt_buddy, first_seg, seg_order);
230 	mlx4_table_put_range(dev, &mr_table->mtt_table, offset,
231 			     offset + (1 << order) - 1);
232 }
233 
234 static void mlx4_free_mtt_range(struct mlx4_dev *dev, u32 offset, int order)
235 {
236 	u64 in_param = 0;
237 	int err;
238 
239 	if (mlx4_is_mfunc(dev)) {
240 		set_param_l(&in_param, offset);
241 		set_param_h(&in_param, order);
242 		err = mlx4_cmd(dev, in_param, RES_MTT, RES_OP_RESERVE_AND_MAP,
243 						       MLX4_CMD_FREE_RES,
244 						       MLX4_CMD_TIME_CLASS_A,
245 						       MLX4_CMD_WRAPPED);
246 		if (err)
247 			mlx4_warn(dev, "Failed to free mtt range at:%d order:%d\n",
248 				  offset, order);
249 		return;
250 	}
251 	__mlx4_free_mtt_range(dev, offset, order);
252 }
253 
254 void mlx4_mtt_cleanup(struct mlx4_dev *dev, struct mlx4_mtt *mtt)
255 {
256 	if (mtt->order < 0)
257 		return;
258 
259 	mlx4_free_mtt_range(dev, mtt->offset, mtt->order);
260 }
261 EXPORT_SYMBOL_GPL(mlx4_mtt_cleanup);
262 
263 u64 mlx4_mtt_addr(struct mlx4_dev *dev, struct mlx4_mtt *mtt)
264 {
265 	return (u64) mtt->offset * dev->caps.mtt_entry_sz;
266 }
267 EXPORT_SYMBOL_GPL(mlx4_mtt_addr);
268 
269 static u32 hw_index_to_key(u32 ind)
270 {
271 	return (ind >> 24) | (ind << 8);
272 }
273 
274 static u32 key_to_hw_index(u32 key)
275 {
276 	return (key << 24) | (key >> 8);
277 }
278 
279 static int mlx4_SW2HW_MPT(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
280 			  int mpt_index)
281 {
282 	return mlx4_cmd(dev, mailbox->dma, mpt_index,
283 			0, MLX4_CMD_SW2HW_MPT, MLX4_CMD_TIME_CLASS_B,
284 			MLX4_CMD_WRAPPED);
285 }
286 
287 static int mlx4_HW2SW_MPT(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
288 			  int mpt_index)
289 {
290 	return mlx4_cmd_box(dev, 0, mailbox ? mailbox->dma : 0, mpt_index,
291 			    !mailbox, MLX4_CMD_HW2SW_MPT,
292 			    MLX4_CMD_TIME_CLASS_B, MLX4_CMD_WRAPPED);
293 }
294 
295 /* Must protect against concurrent access */
296 int mlx4_mr_hw_get_mpt(struct mlx4_dev *dev, struct mlx4_mr *mmr,
297 		       struct mlx4_mpt_entry ***mpt_entry)
298 {
299 	int err;
300 	int key = key_to_hw_index(mmr->key) & (dev->caps.num_mpts - 1);
301 	struct mlx4_cmd_mailbox *mailbox = NULL;
302 
303 	if (mmr->enabled != MLX4_MPT_EN_HW)
304 		return -EINVAL;
305 
306 	err = mlx4_HW2SW_MPT(dev, NULL, key);
307 	if (err) {
308 		mlx4_warn(dev, "HW2SW_MPT failed (%d).", err);
309 		mlx4_warn(dev, "Most likely the MR has MWs bound to it.\n");
310 		return err;
311 	}
312 
313 	mmr->enabled = MLX4_MPT_EN_SW;
314 
315 	if (!mlx4_is_mfunc(dev)) {
316 		**mpt_entry = mlx4_table_find(
317 				&mlx4_priv(dev)->mr_table.dmpt_table,
318 				key, NULL);
319 	} else {
320 		mailbox = mlx4_alloc_cmd_mailbox(dev);
321 		if (IS_ERR(mailbox))
322 			return PTR_ERR(mailbox);
323 
324 		err = mlx4_cmd_box(dev, 0, mailbox->dma, key,
325 				   0, MLX4_CMD_QUERY_MPT,
326 				   MLX4_CMD_TIME_CLASS_B,
327 				   MLX4_CMD_WRAPPED);
328 		if (err)
329 			goto free_mailbox;
330 
331 		*mpt_entry = (struct mlx4_mpt_entry **)&mailbox->buf;
332 	}
333 
334 	if (!(*mpt_entry) || !(**mpt_entry)) {
335 		err = -ENOMEM;
336 		goto free_mailbox;
337 	}
338 
339 	return 0;
340 
341 free_mailbox:
342 	mlx4_free_cmd_mailbox(dev, mailbox);
343 	return err;
344 }
345 EXPORT_SYMBOL_GPL(mlx4_mr_hw_get_mpt);
346 
347 int mlx4_mr_hw_write_mpt(struct mlx4_dev *dev, struct mlx4_mr *mmr,
348 			 struct mlx4_mpt_entry **mpt_entry)
349 {
350 	int err;
351 
352 	if (!mlx4_is_mfunc(dev)) {
353 		/* Make sure any changes to this entry are flushed */
354 		wmb();
355 
356 		*(u8 *)(*mpt_entry) = MLX4_MPT_STATUS_HW;
357 
358 		/* Make sure the new status is written */
359 		wmb();
360 
361 		err = mlx4_SYNC_TPT(dev);
362 	} else {
363 		int key = key_to_hw_index(mmr->key) & (dev->caps.num_mpts - 1);
364 
365 		struct mlx4_cmd_mailbox *mailbox =
366 			container_of((void *)mpt_entry, struct mlx4_cmd_mailbox,
367 				     buf);
368 
369 		err = mlx4_SW2HW_MPT(dev, mailbox, key);
370 	}
371 
372 	if (!err) {
373 		mmr->pd = be32_to_cpu((*mpt_entry)->pd_flags) & MLX4_MPT_PD_MASK;
374 		mmr->enabled = MLX4_MPT_EN_HW;
375 	}
376 	return err;
377 }
378 EXPORT_SYMBOL_GPL(mlx4_mr_hw_write_mpt);
379 
380 void mlx4_mr_hw_put_mpt(struct mlx4_dev *dev,
381 			struct mlx4_mpt_entry **mpt_entry)
382 {
383 	if (mlx4_is_mfunc(dev)) {
384 		struct mlx4_cmd_mailbox *mailbox =
385 			container_of((void *)mpt_entry, struct mlx4_cmd_mailbox,
386 				     buf);
387 		mlx4_free_cmd_mailbox(dev, mailbox);
388 	}
389 }
390 EXPORT_SYMBOL_GPL(mlx4_mr_hw_put_mpt);
391 
392 int mlx4_mr_hw_change_pd(struct mlx4_dev *dev, struct mlx4_mpt_entry *mpt_entry,
393 			 u32 pdn)
394 {
395 	u32 pd_flags = be32_to_cpu(mpt_entry->pd_flags) & ~MLX4_MPT_PD_MASK;
396 	/* The wrapper function will put the slave's id here */
397 	if (mlx4_is_mfunc(dev))
398 		pd_flags &= ~MLX4_MPT_PD_VF_MASK;
399 
400 	mpt_entry->pd_flags = cpu_to_be32(pd_flags |
401 					  (pdn & MLX4_MPT_PD_MASK)
402 					  | MLX4_MPT_PD_FLAG_EN_INV);
403 	return 0;
404 }
405 EXPORT_SYMBOL_GPL(mlx4_mr_hw_change_pd);
406 
407 int mlx4_mr_hw_change_access(struct mlx4_dev *dev,
408 			     struct mlx4_mpt_entry *mpt_entry,
409 			     u32 access)
410 {
411 	u32 flags = (be32_to_cpu(mpt_entry->flags) & ~MLX4_PERM_MASK) |
412 		    (access & MLX4_PERM_MASK);
413 
414 	mpt_entry->flags = cpu_to_be32(flags);
415 	return 0;
416 }
417 EXPORT_SYMBOL_GPL(mlx4_mr_hw_change_access);
418 
419 static int mlx4_mr_alloc_reserved(struct mlx4_dev *dev, u32 mridx, u32 pd,
420 			   u64 iova, u64 size, u32 access, int npages,
421 			   int page_shift, struct mlx4_mr *mr)
422 {
423 	mr->iova       = iova;
424 	mr->size       = size;
425 	mr->pd	       = pd;
426 	mr->access     = access;
427 	mr->enabled    = MLX4_MPT_DISABLED;
428 	mr->key	       = hw_index_to_key(mridx);
429 
430 	return mlx4_mtt_init(dev, npages, page_shift, &mr->mtt);
431 }
432 
433 static int mlx4_WRITE_MTT(struct mlx4_dev *dev,
434 			  struct mlx4_cmd_mailbox *mailbox,
435 			  int num_entries)
436 {
437 	return mlx4_cmd(dev, mailbox->dma, num_entries, 0, MLX4_CMD_WRITE_MTT,
438 			MLX4_CMD_TIME_CLASS_A,  MLX4_CMD_WRAPPED);
439 }
440 
441 int __mlx4_mpt_reserve(struct mlx4_dev *dev)
442 {
443 	struct mlx4_priv *priv = mlx4_priv(dev);
444 
445 	return mlx4_bitmap_alloc(&priv->mr_table.mpt_bitmap);
446 }
447 
448 static int mlx4_mpt_reserve(struct mlx4_dev *dev)
449 {
450 	u64 out_param;
451 
452 	if (mlx4_is_mfunc(dev)) {
453 		if (mlx4_cmd_imm(dev, 0, &out_param, RES_MPT, RES_OP_RESERVE,
454 				   MLX4_CMD_ALLOC_RES,
455 				   MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED))
456 			return -1;
457 		return get_param_l(&out_param);
458 	}
459 	return  __mlx4_mpt_reserve(dev);
460 }
461 
462 void __mlx4_mpt_release(struct mlx4_dev *dev, u32 index)
463 {
464 	struct mlx4_priv *priv = mlx4_priv(dev);
465 
466 	mlx4_bitmap_free(&priv->mr_table.mpt_bitmap, index, MLX4_NO_RR);
467 }
468 
469 static void mlx4_mpt_release(struct mlx4_dev *dev, u32 index)
470 {
471 	u64 in_param = 0;
472 
473 	if (mlx4_is_mfunc(dev)) {
474 		set_param_l(&in_param, index);
475 		if (mlx4_cmd(dev, in_param, RES_MPT, RES_OP_RESERVE,
476 			       MLX4_CMD_FREE_RES,
477 			       MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED))
478 			mlx4_warn(dev, "Failed to release mr index:%d\n",
479 				  index);
480 		return;
481 	}
482 	__mlx4_mpt_release(dev, index);
483 }
484 
485 int __mlx4_mpt_alloc_icm(struct mlx4_dev *dev, u32 index, gfp_t gfp)
486 {
487 	struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
488 
489 	return mlx4_table_get(dev, &mr_table->dmpt_table, index, gfp);
490 }
491 
492 static int mlx4_mpt_alloc_icm(struct mlx4_dev *dev, u32 index, gfp_t gfp)
493 {
494 	u64 param = 0;
495 
496 	if (mlx4_is_mfunc(dev)) {
497 		set_param_l(&param, index);
498 		return mlx4_cmd_imm(dev, param, &param, RES_MPT, RES_OP_MAP_ICM,
499 							MLX4_CMD_ALLOC_RES,
500 							MLX4_CMD_TIME_CLASS_A,
501 							MLX4_CMD_WRAPPED);
502 	}
503 	return __mlx4_mpt_alloc_icm(dev, index, gfp);
504 }
505 
506 void __mlx4_mpt_free_icm(struct mlx4_dev *dev, u32 index)
507 {
508 	struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
509 
510 	mlx4_table_put(dev, &mr_table->dmpt_table, index);
511 }
512 
513 static void mlx4_mpt_free_icm(struct mlx4_dev *dev, u32 index)
514 {
515 	u64 in_param = 0;
516 
517 	if (mlx4_is_mfunc(dev)) {
518 		set_param_l(&in_param, index);
519 		if (mlx4_cmd(dev, in_param, RES_MPT, RES_OP_MAP_ICM,
520 			     MLX4_CMD_FREE_RES, MLX4_CMD_TIME_CLASS_A,
521 			     MLX4_CMD_WRAPPED))
522 			mlx4_warn(dev, "Failed to free icm of mr index:%d\n",
523 				  index);
524 		return;
525 	}
526 	return __mlx4_mpt_free_icm(dev, index);
527 }
528 
529 int mlx4_mr_alloc(struct mlx4_dev *dev, u32 pd, u64 iova, u64 size, u32 access,
530 		  int npages, int page_shift, struct mlx4_mr *mr)
531 {
532 	u32 index;
533 	int err;
534 
535 	index = mlx4_mpt_reserve(dev);
536 	if (index == -1)
537 		return -ENOMEM;
538 
539 	err = mlx4_mr_alloc_reserved(dev, index, pd, iova, size,
540 				     access, npages, page_shift, mr);
541 	if (err)
542 		mlx4_mpt_release(dev, index);
543 
544 	return err;
545 }
546 EXPORT_SYMBOL_GPL(mlx4_mr_alloc);
547 
548 static int mlx4_mr_free_reserved(struct mlx4_dev *dev, struct mlx4_mr *mr)
549 {
550 	int err;
551 
552 	if (mr->enabled == MLX4_MPT_EN_HW) {
553 		err = mlx4_HW2SW_MPT(dev, NULL,
554 				     key_to_hw_index(mr->key) &
555 				     (dev->caps.num_mpts - 1));
556 		if (err) {
557 			mlx4_warn(dev, "HW2SW_MPT failed (%d), MR has MWs bound to it\n",
558 				  err);
559 			return err;
560 		}
561 
562 		mr->enabled = MLX4_MPT_EN_SW;
563 	}
564 	mlx4_mtt_cleanup(dev, &mr->mtt);
565 
566 	return 0;
567 }
568 
569 int mlx4_mr_free(struct mlx4_dev *dev, struct mlx4_mr *mr)
570 {
571 	int ret;
572 
573 	ret = mlx4_mr_free_reserved(dev, mr);
574 	if (ret)
575 		return ret;
576 	if (mr->enabled)
577 		mlx4_mpt_free_icm(dev, key_to_hw_index(mr->key));
578 	mlx4_mpt_release(dev, key_to_hw_index(mr->key));
579 
580 	return 0;
581 }
582 EXPORT_SYMBOL_GPL(mlx4_mr_free);
583 
584 void mlx4_mr_rereg_mem_cleanup(struct mlx4_dev *dev, struct mlx4_mr *mr)
585 {
586 	mlx4_mtt_cleanup(dev, &mr->mtt);
587 	mr->mtt.order = -1;
588 }
589 EXPORT_SYMBOL_GPL(mlx4_mr_rereg_mem_cleanup);
590 
591 int mlx4_mr_rereg_mem_write(struct mlx4_dev *dev, struct mlx4_mr *mr,
592 			    u64 iova, u64 size, int npages,
593 			    int page_shift, struct mlx4_mpt_entry *mpt_entry)
594 {
595 	int err;
596 
597 	err = mlx4_mtt_init(dev, npages, page_shift, &mr->mtt);
598 	if (err)
599 		return err;
600 
601 	mpt_entry->start       = cpu_to_be64(iova);
602 	mpt_entry->length      = cpu_to_be64(size);
603 	mpt_entry->entity_size = cpu_to_be32(page_shift);
604 	mpt_entry->flags    &= ~(cpu_to_be32(MLX4_MPT_FLAG_FREE |
605 					   MLX4_MPT_FLAG_SW_OWNS));
606 	if (mr->mtt.order < 0) {
607 		mpt_entry->flags |= cpu_to_be32(MLX4_MPT_FLAG_PHYSICAL);
608 		mpt_entry->mtt_addr = 0;
609 	} else {
610 		mpt_entry->mtt_addr = cpu_to_be64(mlx4_mtt_addr(dev,
611 						  &mr->mtt));
612 		if (mr->mtt.page_shift == 0)
613 			mpt_entry->mtt_sz    = cpu_to_be32(1 << mr->mtt.order);
614 	}
615 	if (mr->mtt.order >= 0 && mr->mtt.page_shift == 0) {
616 		/* fast register MR in free state */
617 		mpt_entry->flags    |= cpu_to_be32(MLX4_MPT_FLAG_FREE);
618 		mpt_entry->pd_flags |= cpu_to_be32(MLX4_MPT_PD_FLAG_FAST_REG |
619 						   MLX4_MPT_PD_FLAG_RAE);
620 	} else {
621 		mpt_entry->flags    |= cpu_to_be32(MLX4_MPT_FLAG_SW_OWNS);
622 	}
623 	mr->enabled = MLX4_MPT_EN_SW;
624 
625 	return 0;
626 }
627 EXPORT_SYMBOL_GPL(mlx4_mr_rereg_mem_write);
628 
629 int mlx4_mr_enable(struct mlx4_dev *dev, struct mlx4_mr *mr)
630 {
631 	struct mlx4_cmd_mailbox *mailbox;
632 	struct mlx4_mpt_entry *mpt_entry;
633 	int err;
634 
635 	err = mlx4_mpt_alloc_icm(dev, key_to_hw_index(mr->key), GFP_KERNEL);
636 	if (err)
637 		return err;
638 
639 	mailbox = mlx4_alloc_cmd_mailbox(dev);
640 	if (IS_ERR(mailbox)) {
641 		err = PTR_ERR(mailbox);
642 		goto err_table;
643 	}
644 	mpt_entry = mailbox->buf;
645 	mpt_entry->flags = cpu_to_be32(MLX4_MPT_FLAG_MIO	 |
646 				       MLX4_MPT_FLAG_REGION	 |
647 				       mr->access);
648 
649 	mpt_entry->key	       = cpu_to_be32(key_to_hw_index(mr->key));
650 	mpt_entry->pd_flags    = cpu_to_be32(mr->pd | MLX4_MPT_PD_FLAG_EN_INV);
651 	mpt_entry->start       = cpu_to_be64(mr->iova);
652 	mpt_entry->length      = cpu_to_be64(mr->size);
653 	mpt_entry->entity_size = cpu_to_be32(mr->mtt.page_shift);
654 
655 	if (mr->mtt.order < 0) {
656 		mpt_entry->flags |= cpu_to_be32(MLX4_MPT_FLAG_PHYSICAL);
657 		mpt_entry->mtt_addr = 0;
658 	} else {
659 		mpt_entry->mtt_addr = cpu_to_be64(mlx4_mtt_addr(dev,
660 						  &mr->mtt));
661 	}
662 
663 	if (mr->mtt.order >= 0 && mr->mtt.page_shift == 0) {
664 		/* fast register MR in free state */
665 		mpt_entry->flags    |= cpu_to_be32(MLX4_MPT_FLAG_FREE);
666 		mpt_entry->pd_flags |= cpu_to_be32(MLX4_MPT_PD_FLAG_FAST_REG |
667 						   MLX4_MPT_PD_FLAG_RAE);
668 		mpt_entry->mtt_sz    = cpu_to_be32(1 << mr->mtt.order);
669 	} else {
670 		mpt_entry->flags    |= cpu_to_be32(MLX4_MPT_FLAG_SW_OWNS);
671 	}
672 
673 	err = mlx4_SW2HW_MPT(dev, mailbox,
674 			     key_to_hw_index(mr->key) & (dev->caps.num_mpts - 1));
675 	if (err) {
676 		mlx4_warn(dev, "SW2HW_MPT failed (%d)\n", err);
677 		goto err_cmd;
678 	}
679 	mr->enabled = MLX4_MPT_EN_HW;
680 
681 	mlx4_free_cmd_mailbox(dev, mailbox);
682 
683 	return 0;
684 
685 err_cmd:
686 	mlx4_free_cmd_mailbox(dev, mailbox);
687 
688 err_table:
689 	mlx4_mpt_free_icm(dev, key_to_hw_index(mr->key));
690 	return err;
691 }
692 EXPORT_SYMBOL_GPL(mlx4_mr_enable);
693 
694 static int mlx4_write_mtt_chunk(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
695 				int start_index, int npages, u64 *page_list)
696 {
697 	struct mlx4_priv *priv = mlx4_priv(dev);
698 	__be64 *mtts;
699 	dma_addr_t dma_handle;
700 	int i;
701 
702 	mtts = mlx4_table_find(&priv->mr_table.mtt_table, mtt->offset +
703 			       start_index, &dma_handle);
704 
705 	if (!mtts)
706 		return -ENOMEM;
707 
708 	dma_sync_single_for_cpu(&dev->persist->pdev->dev, dma_handle,
709 				npages * sizeof (u64), DMA_TO_DEVICE);
710 
711 	for (i = 0; i < npages; ++i)
712 		mtts[i] = cpu_to_be64(page_list[i] | MLX4_MTT_FLAG_PRESENT);
713 
714 	dma_sync_single_for_device(&dev->persist->pdev->dev, dma_handle,
715 				   npages * sizeof (u64), DMA_TO_DEVICE);
716 
717 	return 0;
718 }
719 
720 int __mlx4_write_mtt(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
721 		     int start_index, int npages, u64 *page_list)
722 {
723 	int err = 0;
724 	int chunk;
725 	int mtts_per_page;
726 	int max_mtts_first_page;
727 
728 	/* compute how may mtts fit in the first page */
729 	mtts_per_page = PAGE_SIZE / sizeof(u64);
730 	max_mtts_first_page = mtts_per_page - (mtt->offset + start_index)
731 			      % mtts_per_page;
732 
733 	chunk = min_t(int, max_mtts_first_page, npages);
734 
735 	while (npages > 0) {
736 		err = mlx4_write_mtt_chunk(dev, mtt, start_index, chunk, page_list);
737 		if (err)
738 			return err;
739 		npages      -= chunk;
740 		start_index += chunk;
741 		page_list   += chunk;
742 
743 		chunk = min_t(int, mtts_per_page, npages);
744 	}
745 	return err;
746 }
747 
748 int mlx4_write_mtt(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
749 		   int start_index, int npages, u64 *page_list)
750 {
751 	struct mlx4_cmd_mailbox *mailbox = NULL;
752 	__be64 *inbox = NULL;
753 	int chunk;
754 	int err = 0;
755 	int i;
756 
757 	if (mtt->order < 0)
758 		return -EINVAL;
759 
760 	if (mlx4_is_mfunc(dev)) {
761 		mailbox = mlx4_alloc_cmd_mailbox(dev);
762 		if (IS_ERR(mailbox))
763 			return PTR_ERR(mailbox);
764 		inbox = mailbox->buf;
765 
766 		while (npages > 0) {
767 			chunk = min_t(int, MLX4_MAILBOX_SIZE / sizeof(u64) - 2,
768 				      npages);
769 			inbox[0] = cpu_to_be64(mtt->offset + start_index);
770 			inbox[1] = 0;
771 			for (i = 0; i < chunk; ++i)
772 				inbox[i + 2] = cpu_to_be64(page_list[i] |
773 					       MLX4_MTT_FLAG_PRESENT);
774 			err = mlx4_WRITE_MTT(dev, mailbox, chunk);
775 			if (err) {
776 				mlx4_free_cmd_mailbox(dev, mailbox);
777 				return err;
778 			}
779 
780 			npages      -= chunk;
781 			start_index += chunk;
782 			page_list   += chunk;
783 		}
784 		mlx4_free_cmd_mailbox(dev, mailbox);
785 		return err;
786 	}
787 
788 	return __mlx4_write_mtt(dev, mtt, start_index, npages, page_list);
789 }
790 EXPORT_SYMBOL_GPL(mlx4_write_mtt);
791 
792 int mlx4_buf_write_mtt(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
793 		       struct mlx4_buf *buf, gfp_t gfp)
794 {
795 	u64 *page_list;
796 	int err;
797 	int i;
798 
799 	page_list = kmalloc(buf->npages * sizeof *page_list,
800 			    gfp);
801 	if (!page_list)
802 		return -ENOMEM;
803 
804 	for (i = 0; i < buf->npages; ++i)
805 		if (buf->nbufs == 1)
806 			page_list[i] = buf->direct.map + (i << buf->page_shift);
807 		else
808 			page_list[i] = buf->page_list[i].map;
809 
810 	err = mlx4_write_mtt(dev, mtt, 0, buf->npages, page_list);
811 
812 	kfree(page_list);
813 	return err;
814 }
815 EXPORT_SYMBOL_GPL(mlx4_buf_write_mtt);
816 
817 int mlx4_mw_alloc(struct mlx4_dev *dev, u32 pd, enum mlx4_mw_type type,
818 		  struct mlx4_mw *mw)
819 {
820 	u32 index;
821 
822 	if ((type == MLX4_MW_TYPE_1 &&
823 	     !(dev->caps.flags & MLX4_DEV_CAP_FLAG_MEM_WINDOW)) ||
824 	     (type == MLX4_MW_TYPE_2 &&
825 	     !(dev->caps.bmme_flags & MLX4_BMME_FLAG_TYPE_2_WIN)))
826 		return -EOPNOTSUPP;
827 
828 	index = mlx4_mpt_reserve(dev);
829 	if (index == -1)
830 		return -ENOMEM;
831 
832 	mw->key	    = hw_index_to_key(index);
833 	mw->pd      = pd;
834 	mw->type    = type;
835 	mw->enabled = MLX4_MPT_DISABLED;
836 
837 	return 0;
838 }
839 EXPORT_SYMBOL_GPL(mlx4_mw_alloc);
840 
841 int mlx4_mw_enable(struct mlx4_dev *dev, struct mlx4_mw *mw)
842 {
843 	struct mlx4_cmd_mailbox *mailbox;
844 	struct mlx4_mpt_entry *mpt_entry;
845 	int err;
846 
847 	err = mlx4_mpt_alloc_icm(dev, key_to_hw_index(mw->key), GFP_KERNEL);
848 	if (err)
849 		return err;
850 
851 	mailbox = mlx4_alloc_cmd_mailbox(dev);
852 	if (IS_ERR(mailbox)) {
853 		err = PTR_ERR(mailbox);
854 		goto err_table;
855 	}
856 	mpt_entry = mailbox->buf;
857 
858 	/* Note that the MLX4_MPT_FLAG_REGION bit in mpt_entry->flags is turned
859 	 * off, thus creating a memory window and not a memory region.
860 	 */
861 	mpt_entry->key	       = cpu_to_be32(key_to_hw_index(mw->key));
862 	mpt_entry->pd_flags    = cpu_to_be32(mw->pd);
863 	if (mw->type == MLX4_MW_TYPE_2) {
864 		mpt_entry->flags    |= cpu_to_be32(MLX4_MPT_FLAG_FREE);
865 		mpt_entry->qpn       = cpu_to_be32(MLX4_MPT_QP_FLAG_BOUND_QP);
866 		mpt_entry->pd_flags |= cpu_to_be32(MLX4_MPT_PD_FLAG_EN_INV);
867 	}
868 
869 	err = mlx4_SW2HW_MPT(dev, mailbox,
870 			     key_to_hw_index(mw->key) &
871 			     (dev->caps.num_mpts - 1));
872 	if (err) {
873 		mlx4_warn(dev, "SW2HW_MPT failed (%d)\n", err);
874 		goto err_cmd;
875 	}
876 	mw->enabled = MLX4_MPT_EN_HW;
877 
878 	mlx4_free_cmd_mailbox(dev, mailbox);
879 
880 	return 0;
881 
882 err_cmd:
883 	mlx4_free_cmd_mailbox(dev, mailbox);
884 
885 err_table:
886 	mlx4_mpt_free_icm(dev, key_to_hw_index(mw->key));
887 	return err;
888 }
889 EXPORT_SYMBOL_GPL(mlx4_mw_enable);
890 
891 void mlx4_mw_free(struct mlx4_dev *dev, struct mlx4_mw *mw)
892 {
893 	int err;
894 
895 	if (mw->enabled == MLX4_MPT_EN_HW) {
896 		err = mlx4_HW2SW_MPT(dev, NULL,
897 				     key_to_hw_index(mw->key) &
898 				     (dev->caps.num_mpts - 1));
899 		if (err)
900 			mlx4_warn(dev, "xxx HW2SW_MPT failed (%d)\n", err);
901 
902 		mw->enabled = MLX4_MPT_EN_SW;
903 	}
904 	if (mw->enabled)
905 		mlx4_mpt_free_icm(dev, key_to_hw_index(mw->key));
906 	mlx4_mpt_release(dev, key_to_hw_index(mw->key));
907 }
908 EXPORT_SYMBOL_GPL(mlx4_mw_free);
909 
910 int mlx4_init_mr_table(struct mlx4_dev *dev)
911 {
912 	struct mlx4_priv *priv = mlx4_priv(dev);
913 	struct mlx4_mr_table *mr_table = &priv->mr_table;
914 	int err;
915 
916 	/* Nothing to do for slaves - all MR handling is forwarded
917 	* to the master */
918 	if (mlx4_is_slave(dev))
919 		return 0;
920 
921 	if (!is_power_of_2(dev->caps.num_mpts))
922 		return -EINVAL;
923 
924 	err = mlx4_bitmap_init(&mr_table->mpt_bitmap, dev->caps.num_mpts,
925 			       ~0, dev->caps.reserved_mrws, 0);
926 	if (err)
927 		return err;
928 
929 	err = mlx4_buddy_init(&mr_table->mtt_buddy,
930 			      ilog2((u32)dev->caps.num_mtts /
931 			      (1 << log_mtts_per_seg)));
932 	if (err)
933 		goto err_buddy;
934 
935 	if (dev->caps.reserved_mtts) {
936 		priv->reserved_mtts =
937 			mlx4_alloc_mtt_range(dev,
938 					     fls(dev->caps.reserved_mtts - 1));
939 		if (priv->reserved_mtts < 0) {
940 			mlx4_warn(dev, "MTT table of order %u is too small\n",
941 				  mr_table->mtt_buddy.max_order);
942 			err = -ENOMEM;
943 			goto err_reserve_mtts;
944 		}
945 	}
946 
947 	return 0;
948 
949 err_reserve_mtts:
950 	mlx4_buddy_cleanup(&mr_table->mtt_buddy);
951 
952 err_buddy:
953 	mlx4_bitmap_cleanup(&mr_table->mpt_bitmap);
954 
955 	return err;
956 }
957 
958 void mlx4_cleanup_mr_table(struct mlx4_dev *dev)
959 {
960 	struct mlx4_priv *priv = mlx4_priv(dev);
961 	struct mlx4_mr_table *mr_table = &priv->mr_table;
962 
963 	if (mlx4_is_slave(dev))
964 		return;
965 	if (priv->reserved_mtts >= 0)
966 		mlx4_free_mtt_range(dev, priv->reserved_mtts,
967 				    fls(dev->caps.reserved_mtts - 1));
968 	mlx4_buddy_cleanup(&mr_table->mtt_buddy);
969 	mlx4_bitmap_cleanup(&mr_table->mpt_bitmap);
970 }
971 
972 static inline int mlx4_check_fmr(struct mlx4_fmr *fmr, u64 *page_list,
973 				  int npages, u64 iova)
974 {
975 	int i, page_mask;
976 
977 	if (npages > fmr->max_pages)
978 		return -EINVAL;
979 
980 	page_mask = (1 << fmr->page_shift) - 1;
981 
982 	/* We are getting page lists, so va must be page aligned. */
983 	if (iova & page_mask)
984 		return -EINVAL;
985 
986 	/* Trust the user not to pass misaligned data in page_list */
987 	if (0)
988 		for (i = 0; i < npages; ++i) {
989 			if (page_list[i] & ~page_mask)
990 				return -EINVAL;
991 		}
992 
993 	if (fmr->maps >= fmr->max_maps)
994 		return -EINVAL;
995 
996 	return 0;
997 }
998 
999 int mlx4_map_phys_fmr(struct mlx4_dev *dev, struct mlx4_fmr *fmr, u64 *page_list,
1000 		      int npages, u64 iova, u32 *lkey, u32 *rkey)
1001 {
1002 	u32 key;
1003 	int i, err;
1004 
1005 	err = mlx4_check_fmr(fmr, page_list, npages, iova);
1006 	if (err)
1007 		return err;
1008 
1009 	++fmr->maps;
1010 
1011 	key = key_to_hw_index(fmr->mr.key);
1012 	key += dev->caps.num_mpts;
1013 	*lkey = *rkey = fmr->mr.key = hw_index_to_key(key);
1014 
1015 	*(u8 *) fmr->mpt = MLX4_MPT_STATUS_SW;
1016 
1017 	/* Make sure MPT status is visible before writing MTT entries */
1018 	wmb();
1019 
1020 	dma_sync_single_for_cpu(&dev->persist->pdev->dev, fmr->dma_handle,
1021 				npages * sizeof(u64), DMA_TO_DEVICE);
1022 
1023 	for (i = 0; i < npages; ++i)
1024 		fmr->mtts[i] = cpu_to_be64(page_list[i] | MLX4_MTT_FLAG_PRESENT);
1025 
1026 	dma_sync_single_for_device(&dev->persist->pdev->dev, fmr->dma_handle,
1027 				   npages * sizeof(u64), DMA_TO_DEVICE);
1028 
1029 	fmr->mpt->key    = cpu_to_be32(key);
1030 	fmr->mpt->lkey   = cpu_to_be32(key);
1031 	fmr->mpt->length = cpu_to_be64(npages * (1ull << fmr->page_shift));
1032 	fmr->mpt->start  = cpu_to_be64(iova);
1033 
1034 	/* Make MTT entries are visible before setting MPT status */
1035 	wmb();
1036 
1037 	*(u8 *) fmr->mpt = MLX4_MPT_STATUS_HW;
1038 
1039 	/* Make sure MPT status is visible before consumer can use FMR */
1040 	wmb();
1041 
1042 	return 0;
1043 }
1044 EXPORT_SYMBOL_GPL(mlx4_map_phys_fmr);
1045 
1046 int mlx4_fmr_alloc(struct mlx4_dev *dev, u32 pd, u32 access, int max_pages,
1047 		   int max_maps, u8 page_shift, struct mlx4_fmr *fmr)
1048 {
1049 	struct mlx4_priv *priv = mlx4_priv(dev);
1050 	int err = -ENOMEM;
1051 
1052 	if (max_maps > dev->caps.max_fmr_maps)
1053 		return -EINVAL;
1054 
1055 	if (page_shift < (ffs(dev->caps.page_size_cap) - 1) || page_shift >= 32)
1056 		return -EINVAL;
1057 
1058 	/* All MTTs must fit in the same page */
1059 	if (max_pages * sizeof *fmr->mtts > PAGE_SIZE)
1060 		return -EINVAL;
1061 
1062 	fmr->page_shift = page_shift;
1063 	fmr->max_pages  = max_pages;
1064 	fmr->max_maps   = max_maps;
1065 	fmr->maps = 0;
1066 
1067 	err = mlx4_mr_alloc(dev, pd, 0, 0, access, max_pages,
1068 			    page_shift, &fmr->mr);
1069 	if (err)
1070 		return err;
1071 
1072 	fmr->mtts = mlx4_table_find(&priv->mr_table.mtt_table,
1073 				    fmr->mr.mtt.offset,
1074 				    &fmr->dma_handle);
1075 
1076 	if (!fmr->mtts) {
1077 		err = -ENOMEM;
1078 		goto err_free;
1079 	}
1080 
1081 	return 0;
1082 
1083 err_free:
1084 	(void) mlx4_mr_free(dev, &fmr->mr);
1085 	return err;
1086 }
1087 EXPORT_SYMBOL_GPL(mlx4_fmr_alloc);
1088 
1089 int mlx4_fmr_enable(struct mlx4_dev *dev, struct mlx4_fmr *fmr)
1090 {
1091 	struct mlx4_priv *priv = mlx4_priv(dev);
1092 	int err;
1093 
1094 	err = mlx4_mr_enable(dev, &fmr->mr);
1095 	if (err)
1096 		return err;
1097 
1098 	fmr->mpt = mlx4_table_find(&priv->mr_table.dmpt_table,
1099 				    key_to_hw_index(fmr->mr.key), NULL);
1100 	if (!fmr->mpt)
1101 		return -ENOMEM;
1102 
1103 	return 0;
1104 }
1105 EXPORT_SYMBOL_GPL(mlx4_fmr_enable);
1106 
1107 void mlx4_fmr_unmap(struct mlx4_dev *dev, struct mlx4_fmr *fmr,
1108 		    u32 *lkey, u32 *rkey)
1109 {
1110 	struct mlx4_cmd_mailbox *mailbox;
1111 	int err;
1112 
1113 	if (!fmr->maps)
1114 		return;
1115 
1116 	fmr->maps = 0;
1117 
1118 	mailbox = mlx4_alloc_cmd_mailbox(dev);
1119 	if (IS_ERR(mailbox)) {
1120 		err = PTR_ERR(mailbox);
1121 		pr_warn("mlx4_ib: mlx4_alloc_cmd_mailbox failed (%d)\n", err);
1122 		return;
1123 	}
1124 
1125 	err = mlx4_HW2SW_MPT(dev, NULL,
1126 			     key_to_hw_index(fmr->mr.key) &
1127 			     (dev->caps.num_mpts - 1));
1128 	mlx4_free_cmd_mailbox(dev, mailbox);
1129 	if (err) {
1130 		pr_warn("mlx4_ib: mlx4_HW2SW_MPT failed (%d)\n", err);
1131 		return;
1132 	}
1133 	fmr->mr.enabled = MLX4_MPT_EN_SW;
1134 }
1135 EXPORT_SYMBOL_GPL(mlx4_fmr_unmap);
1136 
1137 int mlx4_fmr_free(struct mlx4_dev *dev, struct mlx4_fmr *fmr)
1138 {
1139 	int ret;
1140 
1141 	if (fmr->maps)
1142 		return -EBUSY;
1143 
1144 	ret = mlx4_mr_free(dev, &fmr->mr);
1145 	if (ret)
1146 		return ret;
1147 	fmr->mr.enabled = MLX4_MPT_DISABLED;
1148 
1149 	return 0;
1150 }
1151 EXPORT_SYMBOL_GPL(mlx4_fmr_free);
1152 
1153 int mlx4_SYNC_TPT(struct mlx4_dev *dev)
1154 {
1155 	return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_SYNC_TPT,
1156 			MLX4_CMD_TIME_CLASS_A, MLX4_CMD_NATIVE);
1157 }
1158 EXPORT_SYMBOL_GPL(mlx4_SYNC_TPT);
1159