1 /*
2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Mellanox Technologies. 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 #include <linux/slab.h>
35 #include <linux/errno.h>
36 
37 #include "mthca_dev.h"
38 #include "mthca_cmd.h"
39 #include "mthca_memfree.h"
40 
41 struct mthca_mtt {
42 	struct mthca_buddy *buddy;
43 	int                 order;
44 	u32                 first_seg;
45 };
46 
47 /*
48  * Must be packed because mtt_seg is 64 bits but only aligned to 32 bits.
49  */
50 struct mthca_mpt_entry {
51 	__be32 flags;
52 	__be32 page_size;
53 	__be32 key;
54 	__be32 pd;
55 	__be64 start;
56 	__be64 length;
57 	__be32 lkey;
58 	__be32 window_count;
59 	__be32 window_count_limit;
60 	__be64 mtt_seg;
61 	__be32 mtt_sz;		/* Arbel only */
62 	u32    reserved[2];
63 } __packed;
64 
65 #define MTHCA_MPT_FLAG_SW_OWNS       (0xfUL << 28)
66 #define MTHCA_MPT_FLAG_MIO           (1 << 17)
67 #define MTHCA_MPT_FLAG_BIND_ENABLE   (1 << 15)
68 #define MTHCA_MPT_FLAG_PHYSICAL      (1 <<  9)
69 #define MTHCA_MPT_FLAG_REGION        (1 <<  8)
70 
71 #define MTHCA_MTT_FLAG_PRESENT       1
72 
73 #define MTHCA_MPT_STATUS_SW 0xF0
74 #define MTHCA_MPT_STATUS_HW 0x00
75 
76 #define SINAI_FMR_KEY_INC 0x1000000
77 
78 /*
79  * Buddy allocator for MTT segments (currently not very efficient
80  * since it doesn't keep a free list and just searches linearly
81  * through the bitmaps)
82  */
83 
84 static u32 mthca_buddy_alloc(struct mthca_buddy *buddy, int order)
85 {
86 	int o;
87 	int m;
88 	u32 seg;
89 
90 	spin_lock(&buddy->lock);
91 
92 	for (o = order; o <= buddy->max_order; ++o)
93 		if (buddy->num_free[o]) {
94 			m = 1 << (buddy->max_order - o);
95 			seg = find_first_bit(buddy->bits[o], m);
96 			if (seg < m)
97 				goto found;
98 		}
99 
100 	spin_unlock(&buddy->lock);
101 	return -1;
102 
103  found:
104 	clear_bit(seg, buddy->bits[o]);
105 	--buddy->num_free[o];
106 
107 	while (o > order) {
108 		--o;
109 		seg <<= 1;
110 		set_bit(seg ^ 1, buddy->bits[o]);
111 		++buddy->num_free[o];
112 	}
113 
114 	spin_unlock(&buddy->lock);
115 
116 	seg <<= order;
117 
118 	return seg;
119 }
120 
121 static void mthca_buddy_free(struct mthca_buddy *buddy, u32 seg, int order)
122 {
123 	seg >>= order;
124 
125 	spin_lock(&buddy->lock);
126 
127 	while (test_bit(seg ^ 1, buddy->bits[order])) {
128 		clear_bit(seg ^ 1, buddy->bits[order]);
129 		--buddy->num_free[order];
130 		seg >>= 1;
131 		++order;
132 	}
133 
134 	set_bit(seg, buddy->bits[order]);
135 	++buddy->num_free[order];
136 
137 	spin_unlock(&buddy->lock);
138 }
139 
140 static int mthca_buddy_init(struct mthca_buddy *buddy, int max_order)
141 {
142 	int i, s;
143 
144 	buddy->max_order = max_order;
145 	spin_lock_init(&buddy->lock);
146 
147 	buddy->bits = kcalloc(buddy->max_order + 1, sizeof(long *),
148 			      GFP_KERNEL);
149 	buddy->num_free = kcalloc((buddy->max_order + 1), sizeof *buddy->num_free,
150 				  GFP_KERNEL);
151 	if (!buddy->bits || !buddy->num_free)
152 		goto err_out;
153 
154 	for (i = 0; i <= buddy->max_order; ++i) {
155 		s = BITS_TO_LONGS(1 << (buddy->max_order - i));
156 		buddy->bits[i] = kmalloc_array(s, sizeof(long), GFP_KERNEL);
157 		if (!buddy->bits[i])
158 			goto err_out_free;
159 		bitmap_zero(buddy->bits[i],
160 			    1 << (buddy->max_order - i));
161 	}
162 
163 	set_bit(0, buddy->bits[buddy->max_order]);
164 	buddy->num_free[buddy->max_order] = 1;
165 
166 	return 0;
167 
168 err_out_free:
169 	for (i = 0; i <= buddy->max_order; ++i)
170 		kfree(buddy->bits[i]);
171 
172 err_out:
173 	kfree(buddy->bits);
174 	kfree(buddy->num_free);
175 
176 	return -ENOMEM;
177 }
178 
179 static void mthca_buddy_cleanup(struct mthca_buddy *buddy)
180 {
181 	int i;
182 
183 	for (i = 0; i <= buddy->max_order; ++i)
184 		kfree(buddy->bits[i]);
185 
186 	kfree(buddy->bits);
187 	kfree(buddy->num_free);
188 }
189 
190 static u32 mthca_alloc_mtt_range(struct mthca_dev *dev, int order,
191 				 struct mthca_buddy *buddy)
192 {
193 	u32 seg = mthca_buddy_alloc(buddy, order);
194 
195 	if (seg == -1)
196 		return -1;
197 
198 	if (mthca_is_memfree(dev))
199 		if (mthca_table_get_range(dev, dev->mr_table.mtt_table, seg,
200 					  seg + (1 << order) - 1)) {
201 			mthca_buddy_free(buddy, seg, order);
202 			seg = -1;
203 		}
204 
205 	return seg;
206 }
207 
208 static struct mthca_mtt *__mthca_alloc_mtt(struct mthca_dev *dev, int size,
209 					   struct mthca_buddy *buddy)
210 {
211 	struct mthca_mtt *mtt;
212 	int i;
213 
214 	if (size <= 0)
215 		return ERR_PTR(-EINVAL);
216 
217 	mtt = kmalloc(sizeof *mtt, GFP_KERNEL);
218 	if (!mtt)
219 		return ERR_PTR(-ENOMEM);
220 
221 	mtt->buddy = buddy;
222 	mtt->order = 0;
223 	for (i = dev->limits.mtt_seg_size / 8; i < size; i <<= 1)
224 		++mtt->order;
225 
226 	mtt->first_seg = mthca_alloc_mtt_range(dev, mtt->order, buddy);
227 	if (mtt->first_seg == -1) {
228 		kfree(mtt);
229 		return ERR_PTR(-ENOMEM);
230 	}
231 
232 	return mtt;
233 }
234 
235 struct mthca_mtt *mthca_alloc_mtt(struct mthca_dev *dev, int size)
236 {
237 	return __mthca_alloc_mtt(dev, size, &dev->mr_table.mtt_buddy);
238 }
239 
240 void mthca_free_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt)
241 {
242 	if (!mtt)
243 		return;
244 
245 	mthca_buddy_free(mtt->buddy, mtt->first_seg, mtt->order);
246 
247 	mthca_table_put_range(dev, dev->mr_table.mtt_table,
248 			      mtt->first_seg,
249 			      mtt->first_seg + (1 << mtt->order) - 1);
250 
251 	kfree(mtt);
252 }
253 
254 static int __mthca_write_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt,
255 			     int start_index, u64 *buffer_list, int list_len)
256 {
257 	struct mthca_mailbox *mailbox;
258 	__be64 *mtt_entry;
259 	int err = 0;
260 	int i;
261 
262 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
263 	if (IS_ERR(mailbox))
264 		return PTR_ERR(mailbox);
265 	mtt_entry = mailbox->buf;
266 
267 	while (list_len > 0) {
268 		mtt_entry[0] = cpu_to_be64(dev->mr_table.mtt_base +
269 					   mtt->first_seg * dev->limits.mtt_seg_size +
270 					   start_index * 8);
271 		mtt_entry[1] = 0;
272 		for (i = 0; i < list_len && i < MTHCA_MAILBOX_SIZE / 8 - 2; ++i)
273 			mtt_entry[i + 2] = cpu_to_be64(buffer_list[i] |
274 						       MTHCA_MTT_FLAG_PRESENT);
275 
276 		/*
277 		 * If we have an odd number of entries to write, add
278 		 * one more dummy entry for firmware efficiency.
279 		 */
280 		if (i & 1)
281 			mtt_entry[i + 2] = 0;
282 
283 		err = mthca_WRITE_MTT(dev, mailbox, (i + 1) & ~1);
284 		if (err) {
285 			mthca_warn(dev, "WRITE_MTT failed (%d)\n", err);
286 			goto out;
287 		}
288 
289 		list_len    -= i;
290 		start_index += i;
291 		buffer_list += i;
292 	}
293 
294 out:
295 	mthca_free_mailbox(dev, mailbox);
296 	return err;
297 }
298 
299 int mthca_write_mtt_size(struct mthca_dev *dev)
300 {
301 	if (dev->mr_table.fmr_mtt_buddy != &dev->mr_table.mtt_buddy ||
302 	    !(dev->mthca_flags & MTHCA_FLAG_FMR))
303 		/*
304 		 * Be friendly to WRITE_MTT command
305 		 * and leave two empty slots for the
306 		 * index and reserved fields of the
307 		 * mailbox.
308 		 */
309 		return PAGE_SIZE / sizeof (u64) - 2;
310 
311 	/* For Arbel, all MTTs must fit in the same page. */
312 	return mthca_is_memfree(dev) ? (PAGE_SIZE / sizeof (u64)) : 0x7ffffff;
313 }
314 
315 static void mthca_tavor_write_mtt_seg(struct mthca_dev *dev,
316 				      struct mthca_mtt *mtt, int start_index,
317 				      u64 *buffer_list, int list_len)
318 {
319 	u64 __iomem *mtts;
320 	int i;
321 
322 	mtts = dev->mr_table.tavor_fmr.mtt_base + mtt->first_seg * dev->limits.mtt_seg_size +
323 		start_index * sizeof (u64);
324 	for (i = 0; i < list_len; ++i)
325 		mthca_write64_raw(cpu_to_be64(buffer_list[i] | MTHCA_MTT_FLAG_PRESENT),
326 				  mtts + i);
327 }
328 
329 static void mthca_arbel_write_mtt_seg(struct mthca_dev *dev,
330 				      struct mthca_mtt *mtt, int start_index,
331 				      u64 *buffer_list, int list_len)
332 {
333 	__be64 *mtts;
334 	dma_addr_t dma_handle;
335 	int i;
336 	int s = start_index * sizeof (u64);
337 
338 	/* For Arbel, all MTTs must fit in the same page. */
339 	BUG_ON(s / PAGE_SIZE != (s + list_len * sizeof(u64) - 1) / PAGE_SIZE);
340 	/* Require full segments */
341 	BUG_ON(s % dev->limits.mtt_seg_size);
342 
343 	mtts = mthca_table_find(dev->mr_table.mtt_table, mtt->first_seg +
344 				s / dev->limits.mtt_seg_size, &dma_handle);
345 
346 	BUG_ON(!mtts);
347 
348 	dma_sync_single_for_cpu(&dev->pdev->dev, dma_handle,
349 				list_len * sizeof (u64), DMA_TO_DEVICE);
350 
351 	for (i = 0; i < list_len; ++i)
352 		mtts[i] = cpu_to_be64(buffer_list[i] | MTHCA_MTT_FLAG_PRESENT);
353 
354 	dma_sync_single_for_device(&dev->pdev->dev, dma_handle,
355 				   list_len * sizeof (u64), DMA_TO_DEVICE);
356 }
357 
358 int mthca_write_mtt(struct mthca_dev *dev, struct mthca_mtt *mtt,
359 		    int start_index, u64 *buffer_list, int list_len)
360 {
361 	int size = mthca_write_mtt_size(dev);
362 	int chunk;
363 
364 	if (dev->mr_table.fmr_mtt_buddy != &dev->mr_table.mtt_buddy ||
365 	    !(dev->mthca_flags & MTHCA_FLAG_FMR))
366 		return __mthca_write_mtt(dev, mtt, start_index, buffer_list, list_len);
367 
368 	while (list_len > 0) {
369 		chunk = min(size, list_len);
370 		if (mthca_is_memfree(dev))
371 			mthca_arbel_write_mtt_seg(dev, mtt, start_index,
372 						  buffer_list, chunk);
373 		else
374 			mthca_tavor_write_mtt_seg(dev, mtt, start_index,
375 						  buffer_list, chunk);
376 
377 		list_len    -= chunk;
378 		start_index += chunk;
379 		buffer_list += chunk;
380 	}
381 
382 	return 0;
383 }
384 
385 static inline u32 tavor_hw_index_to_key(u32 ind)
386 {
387 	return ind;
388 }
389 
390 static inline u32 tavor_key_to_hw_index(u32 key)
391 {
392 	return key;
393 }
394 
395 static inline u32 arbel_hw_index_to_key(u32 ind)
396 {
397 	return (ind >> 24) | (ind << 8);
398 }
399 
400 static inline u32 arbel_key_to_hw_index(u32 key)
401 {
402 	return (key << 24) | (key >> 8);
403 }
404 
405 static inline u32 hw_index_to_key(struct mthca_dev *dev, u32 ind)
406 {
407 	if (mthca_is_memfree(dev))
408 		return arbel_hw_index_to_key(ind);
409 	else
410 		return tavor_hw_index_to_key(ind);
411 }
412 
413 static inline u32 key_to_hw_index(struct mthca_dev *dev, u32 key)
414 {
415 	if (mthca_is_memfree(dev))
416 		return arbel_key_to_hw_index(key);
417 	else
418 		return tavor_key_to_hw_index(key);
419 }
420 
421 static inline u32 adjust_key(struct mthca_dev *dev, u32 key)
422 {
423 	if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
424 		return ((key << 20) & 0x800000) | (key & 0x7fffff);
425 	else
426 		return key;
427 }
428 
429 int mthca_mr_alloc(struct mthca_dev *dev, u32 pd, int buffer_size_shift,
430 		   u64 iova, u64 total_size, u32 access, struct mthca_mr *mr)
431 {
432 	struct mthca_mailbox *mailbox;
433 	struct mthca_mpt_entry *mpt_entry;
434 	u32 key;
435 	int i;
436 	int err;
437 
438 	WARN_ON(buffer_size_shift >= 32);
439 
440 	key = mthca_alloc(&dev->mr_table.mpt_alloc);
441 	if (key == -1)
442 		return -ENOMEM;
443 	key = adjust_key(dev, key);
444 	mr->ibmr.rkey = mr->ibmr.lkey = hw_index_to_key(dev, key);
445 
446 	if (mthca_is_memfree(dev)) {
447 		err = mthca_table_get(dev, dev->mr_table.mpt_table, key);
448 		if (err)
449 			goto err_out_mpt_free;
450 	}
451 
452 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
453 	if (IS_ERR(mailbox)) {
454 		err = PTR_ERR(mailbox);
455 		goto err_out_table;
456 	}
457 	mpt_entry = mailbox->buf;
458 
459 	mpt_entry->flags = cpu_to_be32(MTHCA_MPT_FLAG_SW_OWNS     |
460 				       MTHCA_MPT_FLAG_MIO         |
461 				       MTHCA_MPT_FLAG_REGION      |
462 				       access);
463 	if (!mr->mtt)
464 		mpt_entry->flags |= cpu_to_be32(MTHCA_MPT_FLAG_PHYSICAL);
465 
466 	mpt_entry->page_size = cpu_to_be32(buffer_size_shift - 12);
467 	mpt_entry->key       = cpu_to_be32(key);
468 	mpt_entry->pd        = cpu_to_be32(pd);
469 	mpt_entry->start     = cpu_to_be64(iova);
470 	mpt_entry->length    = cpu_to_be64(total_size);
471 
472 	memset(&mpt_entry->lkey, 0,
473 	       sizeof *mpt_entry - offsetof(struct mthca_mpt_entry, lkey));
474 
475 	if (mr->mtt)
476 		mpt_entry->mtt_seg =
477 			cpu_to_be64(dev->mr_table.mtt_base +
478 				    mr->mtt->first_seg * dev->limits.mtt_seg_size);
479 
480 	if (0) {
481 		mthca_dbg(dev, "Dumping MPT entry %08x:\n", mr->ibmr.lkey);
482 		for (i = 0; i < sizeof (struct mthca_mpt_entry) / 4; ++i) {
483 			if (i % 4 == 0)
484 				printk("[%02x] ", i * 4);
485 			printk(" %08x", be32_to_cpu(((__be32 *) mpt_entry)[i]));
486 			if ((i + 1) % 4 == 0)
487 				printk("\n");
488 		}
489 	}
490 
491 	err = mthca_SW2HW_MPT(dev, mailbox,
492 			      key & (dev->limits.num_mpts - 1));
493 	if (err) {
494 		mthca_warn(dev, "SW2HW_MPT failed (%d)\n", err);
495 		goto err_out_mailbox;
496 	}
497 
498 	mthca_free_mailbox(dev, mailbox);
499 	return err;
500 
501 err_out_mailbox:
502 	mthca_free_mailbox(dev, mailbox);
503 
504 err_out_table:
505 	mthca_table_put(dev, dev->mr_table.mpt_table, key);
506 
507 err_out_mpt_free:
508 	mthca_free(&dev->mr_table.mpt_alloc, key);
509 	return err;
510 }
511 
512 int mthca_mr_alloc_notrans(struct mthca_dev *dev, u32 pd,
513 			   u32 access, struct mthca_mr *mr)
514 {
515 	mr->mtt = NULL;
516 	return mthca_mr_alloc(dev, pd, 12, 0, ~0ULL, access, mr);
517 }
518 
519 int mthca_mr_alloc_phys(struct mthca_dev *dev, u32 pd,
520 			u64 *buffer_list, int buffer_size_shift,
521 			int list_len, u64 iova, u64 total_size,
522 			u32 access, struct mthca_mr *mr)
523 {
524 	int err;
525 
526 	mr->mtt = mthca_alloc_mtt(dev, list_len);
527 	if (IS_ERR(mr->mtt))
528 		return PTR_ERR(mr->mtt);
529 
530 	err = mthca_write_mtt(dev, mr->mtt, 0, buffer_list, list_len);
531 	if (err) {
532 		mthca_free_mtt(dev, mr->mtt);
533 		return err;
534 	}
535 
536 	err = mthca_mr_alloc(dev, pd, buffer_size_shift, iova,
537 			     total_size, access, mr);
538 	if (err)
539 		mthca_free_mtt(dev, mr->mtt);
540 
541 	return err;
542 }
543 
544 /* Free mr or fmr */
545 static void mthca_free_region(struct mthca_dev *dev, u32 lkey)
546 {
547 	mthca_table_put(dev, dev->mr_table.mpt_table,
548 			key_to_hw_index(dev, lkey));
549 
550 	mthca_free(&dev->mr_table.mpt_alloc, key_to_hw_index(dev, lkey));
551 }
552 
553 void mthca_free_mr(struct mthca_dev *dev, struct mthca_mr *mr)
554 {
555 	int err;
556 
557 	err = mthca_HW2SW_MPT(dev, NULL,
558 			      key_to_hw_index(dev, mr->ibmr.lkey) &
559 			      (dev->limits.num_mpts - 1));
560 	if (err)
561 		mthca_warn(dev, "HW2SW_MPT failed (%d)\n", err);
562 
563 	mthca_free_region(dev, mr->ibmr.lkey);
564 	mthca_free_mtt(dev, mr->mtt);
565 }
566 
567 int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd,
568 		    u32 access, struct mthca_fmr *mr)
569 {
570 	struct mthca_mpt_entry *mpt_entry;
571 	struct mthca_mailbox *mailbox;
572 	u64 mtt_seg;
573 	u32 key, idx;
574 	int list_len = mr->attr.max_pages;
575 	int err = -ENOMEM;
576 	int i;
577 
578 	if (mr->attr.page_shift < 12 || mr->attr.page_shift >= 32)
579 		return -EINVAL;
580 
581 	/* For Arbel, all MTTs must fit in the same page. */
582 	if (mthca_is_memfree(dev) &&
583 	    mr->attr.max_pages * sizeof *mr->mem.arbel.mtts > PAGE_SIZE)
584 		return -EINVAL;
585 
586 	mr->maps = 0;
587 
588 	key = mthca_alloc(&dev->mr_table.mpt_alloc);
589 	if (key == -1)
590 		return -ENOMEM;
591 	key = adjust_key(dev, key);
592 
593 	idx = key & (dev->limits.num_mpts - 1);
594 	mr->ibmr.rkey = mr->ibmr.lkey = hw_index_to_key(dev, key);
595 
596 	if (mthca_is_memfree(dev)) {
597 		err = mthca_table_get(dev, dev->mr_table.mpt_table, key);
598 		if (err)
599 			goto err_out_mpt_free;
600 
601 		mr->mem.arbel.mpt = mthca_table_find(dev->mr_table.mpt_table, key, NULL);
602 		BUG_ON(!mr->mem.arbel.mpt);
603 	} else
604 		mr->mem.tavor.mpt = dev->mr_table.tavor_fmr.mpt_base +
605 			sizeof *(mr->mem.tavor.mpt) * idx;
606 
607 	mr->mtt = __mthca_alloc_mtt(dev, list_len, dev->mr_table.fmr_mtt_buddy);
608 	if (IS_ERR(mr->mtt)) {
609 		err = PTR_ERR(mr->mtt);
610 		goto err_out_table;
611 	}
612 
613 	mtt_seg = mr->mtt->first_seg * dev->limits.mtt_seg_size;
614 
615 	if (mthca_is_memfree(dev)) {
616 		mr->mem.arbel.mtts = mthca_table_find(dev->mr_table.mtt_table,
617 						      mr->mtt->first_seg,
618 						      &mr->mem.arbel.dma_handle);
619 		BUG_ON(!mr->mem.arbel.mtts);
620 	} else
621 		mr->mem.tavor.mtts = dev->mr_table.tavor_fmr.mtt_base + mtt_seg;
622 
623 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
624 	if (IS_ERR(mailbox)) {
625 		err = PTR_ERR(mailbox);
626 		goto err_out_free_mtt;
627 	}
628 
629 	mpt_entry = mailbox->buf;
630 
631 	mpt_entry->flags = cpu_to_be32(MTHCA_MPT_FLAG_SW_OWNS     |
632 				       MTHCA_MPT_FLAG_MIO         |
633 				       MTHCA_MPT_FLAG_REGION      |
634 				       access);
635 
636 	mpt_entry->page_size = cpu_to_be32(mr->attr.page_shift - 12);
637 	mpt_entry->key       = cpu_to_be32(key);
638 	mpt_entry->pd        = cpu_to_be32(pd);
639 	memset(&mpt_entry->start, 0,
640 	       sizeof *mpt_entry - offsetof(struct mthca_mpt_entry, start));
641 	mpt_entry->mtt_seg   = cpu_to_be64(dev->mr_table.mtt_base + mtt_seg);
642 
643 	if (0) {
644 		mthca_dbg(dev, "Dumping MPT entry %08x:\n", mr->ibmr.lkey);
645 		for (i = 0; i < sizeof (struct mthca_mpt_entry) / 4; ++i) {
646 			if (i % 4 == 0)
647 				printk("[%02x] ", i * 4);
648 			printk(" %08x", be32_to_cpu(((__be32 *) mpt_entry)[i]));
649 			if ((i + 1) % 4 == 0)
650 				printk("\n");
651 		}
652 	}
653 
654 	err = mthca_SW2HW_MPT(dev, mailbox,
655 			      key & (dev->limits.num_mpts - 1));
656 	if (err) {
657 		mthca_warn(dev, "SW2HW_MPT failed (%d)\n", err);
658 		goto err_out_mailbox_free;
659 	}
660 
661 	mthca_free_mailbox(dev, mailbox);
662 	return 0;
663 
664 err_out_mailbox_free:
665 	mthca_free_mailbox(dev, mailbox);
666 
667 err_out_free_mtt:
668 	mthca_free_mtt(dev, mr->mtt);
669 
670 err_out_table:
671 	mthca_table_put(dev, dev->mr_table.mpt_table, key);
672 
673 err_out_mpt_free:
674 	mthca_free(&dev->mr_table.mpt_alloc, key);
675 	return err;
676 }
677 
678 int mthca_free_fmr(struct mthca_dev *dev, struct mthca_fmr *fmr)
679 {
680 	if (fmr->maps)
681 		return -EBUSY;
682 
683 	mthca_free_region(dev, fmr->ibmr.lkey);
684 	mthca_free_mtt(dev, fmr->mtt);
685 
686 	return 0;
687 }
688 
689 static inline int mthca_check_fmr(struct mthca_fmr *fmr, u64 *page_list,
690 				  int list_len, u64 iova)
691 {
692 	int i, page_mask;
693 
694 	if (list_len > fmr->attr.max_pages)
695 		return -EINVAL;
696 
697 	page_mask = (1 << fmr->attr.page_shift) - 1;
698 
699 	/* We are getting page lists, so va must be page aligned. */
700 	if (iova & page_mask)
701 		return -EINVAL;
702 
703 	/* Trust the user not to pass misaligned data in page_list */
704 	if (0)
705 		for (i = 0; i < list_len; ++i) {
706 			if (page_list[i] & ~page_mask)
707 				return -EINVAL;
708 		}
709 
710 	if (fmr->maps >= fmr->attr.max_maps)
711 		return -EINVAL;
712 
713 	return 0;
714 }
715 
716 
717 int mthca_tavor_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
718 			     int list_len, u64 iova)
719 {
720 	struct mthca_fmr *fmr = to_mfmr(ibfmr);
721 	struct mthca_dev *dev = to_mdev(ibfmr->device);
722 	struct mthca_mpt_entry mpt_entry;
723 	u32 key;
724 	int i, err;
725 
726 	err = mthca_check_fmr(fmr, page_list, list_len, iova);
727 	if (err)
728 		return err;
729 
730 	++fmr->maps;
731 
732 	key = tavor_key_to_hw_index(fmr->ibmr.lkey);
733 	key += dev->limits.num_mpts;
734 	fmr->ibmr.lkey = fmr->ibmr.rkey = tavor_hw_index_to_key(key);
735 
736 	writeb(MTHCA_MPT_STATUS_SW, fmr->mem.tavor.mpt);
737 
738 	for (i = 0; i < list_len; ++i) {
739 		__be64 mtt_entry = cpu_to_be64(page_list[i] |
740 					       MTHCA_MTT_FLAG_PRESENT);
741 		mthca_write64_raw(mtt_entry, fmr->mem.tavor.mtts + i);
742 	}
743 
744 	mpt_entry.lkey   = cpu_to_be32(key);
745 	mpt_entry.length = cpu_to_be64(list_len * (1ull << fmr->attr.page_shift));
746 	mpt_entry.start  = cpu_to_be64(iova);
747 
748 	__raw_writel((__force u32) mpt_entry.lkey, &fmr->mem.tavor.mpt->key);
749 	memcpy_toio(&fmr->mem.tavor.mpt->start, &mpt_entry.start,
750 		    offsetof(struct mthca_mpt_entry, window_count) -
751 		    offsetof(struct mthca_mpt_entry, start));
752 
753 	writeb(MTHCA_MPT_STATUS_HW, fmr->mem.tavor.mpt);
754 
755 	return 0;
756 }
757 
758 int mthca_arbel_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
759 			     int list_len, u64 iova)
760 {
761 	struct mthca_fmr *fmr = to_mfmr(ibfmr);
762 	struct mthca_dev *dev = to_mdev(ibfmr->device);
763 	u32 key;
764 	int i, err;
765 
766 	err = mthca_check_fmr(fmr, page_list, list_len, iova);
767 	if (err)
768 		return err;
769 
770 	++fmr->maps;
771 
772 	key = arbel_key_to_hw_index(fmr->ibmr.lkey);
773 	if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
774 		key += SINAI_FMR_KEY_INC;
775 	else
776 		key += dev->limits.num_mpts;
777 	fmr->ibmr.lkey = fmr->ibmr.rkey = arbel_hw_index_to_key(key);
778 
779 	*(u8 *) fmr->mem.arbel.mpt = MTHCA_MPT_STATUS_SW;
780 
781 	wmb();
782 
783 	dma_sync_single_for_cpu(&dev->pdev->dev, fmr->mem.arbel.dma_handle,
784 				list_len * sizeof(u64), DMA_TO_DEVICE);
785 
786 	for (i = 0; i < list_len; ++i)
787 		fmr->mem.arbel.mtts[i] = cpu_to_be64(page_list[i] |
788 						     MTHCA_MTT_FLAG_PRESENT);
789 
790 	dma_sync_single_for_device(&dev->pdev->dev, fmr->mem.arbel.dma_handle,
791 				   list_len * sizeof(u64), DMA_TO_DEVICE);
792 
793 	fmr->mem.arbel.mpt->key    = cpu_to_be32(key);
794 	fmr->mem.arbel.mpt->lkey   = cpu_to_be32(key);
795 	fmr->mem.arbel.mpt->length = cpu_to_be64(list_len * (1ull << fmr->attr.page_shift));
796 	fmr->mem.arbel.mpt->start  = cpu_to_be64(iova);
797 
798 	wmb();
799 
800 	*(u8 *) fmr->mem.arbel.mpt = MTHCA_MPT_STATUS_HW;
801 
802 	wmb();
803 
804 	return 0;
805 }
806 
807 void mthca_tavor_fmr_unmap(struct mthca_dev *dev, struct mthca_fmr *fmr)
808 {
809 	if (!fmr->maps)
810 		return;
811 
812 	fmr->maps = 0;
813 
814 	writeb(MTHCA_MPT_STATUS_SW, fmr->mem.tavor.mpt);
815 }
816 
817 void mthca_arbel_fmr_unmap(struct mthca_dev *dev, struct mthca_fmr *fmr)
818 {
819 	if (!fmr->maps)
820 		return;
821 
822 	fmr->maps = 0;
823 
824 	*(u8 *) fmr->mem.arbel.mpt = MTHCA_MPT_STATUS_SW;
825 }
826 
827 int mthca_init_mr_table(struct mthca_dev *dev)
828 {
829 	phys_addr_t addr;
830 	int mpts, mtts, err, i;
831 
832 	err = mthca_alloc_init(&dev->mr_table.mpt_alloc,
833 			       dev->limits.num_mpts,
834 			       ~0, dev->limits.reserved_mrws);
835 	if (err)
836 		return err;
837 
838 	if (!mthca_is_memfree(dev) &&
839 	    (dev->mthca_flags & MTHCA_FLAG_DDR_HIDDEN))
840 		dev->limits.fmr_reserved_mtts = 0;
841 	else
842 		dev->mthca_flags |= MTHCA_FLAG_FMR;
843 
844 	if (dev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
845 		mthca_dbg(dev, "Memory key throughput optimization activated.\n");
846 
847 	err = mthca_buddy_init(&dev->mr_table.mtt_buddy,
848 			       fls(dev->limits.num_mtt_segs - 1));
849 
850 	if (err)
851 		goto err_mtt_buddy;
852 
853 	dev->mr_table.tavor_fmr.mpt_base = NULL;
854 	dev->mr_table.tavor_fmr.mtt_base = NULL;
855 
856 	if (dev->limits.fmr_reserved_mtts) {
857 		i = fls(dev->limits.fmr_reserved_mtts - 1);
858 
859 		if (i >= 31) {
860 			mthca_warn(dev, "Unable to reserve 2^31 FMR MTTs.\n");
861 			err = -EINVAL;
862 			goto err_fmr_mpt;
863 		}
864 		mpts = mtts = 1 << i;
865 	} else {
866 		mtts = dev->limits.num_mtt_segs;
867 		mpts = dev->limits.num_mpts;
868 	}
869 
870 	if (!mthca_is_memfree(dev) &&
871 	    (dev->mthca_flags & MTHCA_FLAG_FMR)) {
872 
873 		addr = pci_resource_start(dev->pdev, 4) +
874 			((pci_resource_len(dev->pdev, 4) - 1) &
875 			 dev->mr_table.mpt_base);
876 
877 		dev->mr_table.tavor_fmr.mpt_base =
878 			ioremap(addr, mpts * sizeof(struct mthca_mpt_entry));
879 
880 		if (!dev->mr_table.tavor_fmr.mpt_base) {
881 			mthca_warn(dev, "MPT ioremap for FMR failed.\n");
882 			err = -ENOMEM;
883 			goto err_fmr_mpt;
884 		}
885 
886 		addr = pci_resource_start(dev->pdev, 4) +
887 			((pci_resource_len(dev->pdev, 4) - 1) &
888 			 dev->mr_table.mtt_base);
889 
890 		dev->mr_table.tavor_fmr.mtt_base =
891 			ioremap(addr, mtts * dev->limits.mtt_seg_size);
892 		if (!dev->mr_table.tavor_fmr.mtt_base) {
893 			mthca_warn(dev, "MTT ioremap for FMR failed.\n");
894 			err = -ENOMEM;
895 			goto err_fmr_mtt;
896 		}
897 	}
898 
899 	if (dev->limits.fmr_reserved_mtts) {
900 		err = mthca_buddy_init(&dev->mr_table.tavor_fmr.mtt_buddy, fls(mtts - 1));
901 		if (err)
902 			goto err_fmr_mtt_buddy;
903 
904 		/* Prevent regular MRs from using FMR keys */
905 		err = mthca_buddy_alloc(&dev->mr_table.mtt_buddy, fls(mtts - 1));
906 		if (err)
907 			goto err_reserve_fmr;
908 
909 		dev->mr_table.fmr_mtt_buddy =
910 			&dev->mr_table.tavor_fmr.mtt_buddy;
911 	} else
912 		dev->mr_table.fmr_mtt_buddy = &dev->mr_table.mtt_buddy;
913 
914 	/* FMR table is always the first, take reserved MTTs out of there */
915 	if (dev->limits.reserved_mtts) {
916 		i = fls(dev->limits.reserved_mtts - 1);
917 
918 		if (mthca_alloc_mtt_range(dev, i,
919 					  dev->mr_table.fmr_mtt_buddy) == -1) {
920 			mthca_warn(dev, "MTT table of order %d is too small.\n",
921 				  dev->mr_table.fmr_mtt_buddy->max_order);
922 			err = -ENOMEM;
923 			goto err_reserve_mtts;
924 		}
925 	}
926 
927 	return 0;
928 
929 err_reserve_mtts:
930 err_reserve_fmr:
931 	if (dev->limits.fmr_reserved_mtts)
932 		mthca_buddy_cleanup(&dev->mr_table.tavor_fmr.mtt_buddy);
933 
934 err_fmr_mtt_buddy:
935 	if (dev->mr_table.tavor_fmr.mtt_base)
936 		iounmap(dev->mr_table.tavor_fmr.mtt_base);
937 
938 err_fmr_mtt:
939 	if (dev->mr_table.tavor_fmr.mpt_base)
940 		iounmap(dev->mr_table.tavor_fmr.mpt_base);
941 
942 err_fmr_mpt:
943 	mthca_buddy_cleanup(&dev->mr_table.mtt_buddy);
944 
945 err_mtt_buddy:
946 	mthca_alloc_cleanup(&dev->mr_table.mpt_alloc);
947 
948 	return err;
949 }
950 
951 void mthca_cleanup_mr_table(struct mthca_dev *dev)
952 {
953 	/* XXX check if any MRs are still allocated? */
954 	if (dev->limits.fmr_reserved_mtts)
955 		mthca_buddy_cleanup(&dev->mr_table.tavor_fmr.mtt_buddy);
956 
957 	mthca_buddy_cleanup(&dev->mr_table.mtt_buddy);
958 
959 	if (dev->mr_table.tavor_fmr.mtt_base)
960 		iounmap(dev->mr_table.tavor_fmr.mtt_base);
961 	if (dev->mr_table.tavor_fmr.mpt_base)
962 		iounmap(dev->mr_table.tavor_fmr.mpt_base);
963 
964 	mthca_alloc_cleanup(&dev->mr_table.mpt_alloc);
965 }
966