Loading...
1/*
2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3 * Copyright (c) 2007, 2008 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 <rdma/ib_user_verbs.h>
36
37#include "mlx4_ib.h"
38
39static u32 convert_access(int acc)
40{
41 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX4_PERM_ATOMIC : 0) |
42 (acc & IB_ACCESS_REMOTE_WRITE ? MLX4_PERM_REMOTE_WRITE : 0) |
43 (acc & IB_ACCESS_REMOTE_READ ? MLX4_PERM_REMOTE_READ : 0) |
44 (acc & IB_ACCESS_LOCAL_WRITE ? MLX4_PERM_LOCAL_WRITE : 0) |
45 (acc & IB_ACCESS_MW_BIND ? MLX4_PERM_BIND_MW : 0) |
46 MLX4_PERM_LOCAL_READ;
47}
48
49static enum mlx4_mw_type to_mlx4_type(enum ib_mw_type type)
50{
51 switch (type) {
52 case IB_MW_TYPE_1: return MLX4_MW_TYPE_1;
53 case IB_MW_TYPE_2: return MLX4_MW_TYPE_2;
54 default: return -1;
55 }
56}
57
58struct ib_mr *mlx4_ib_get_dma_mr(struct ib_pd *pd, int acc)
59{
60 struct mlx4_ib_mr *mr;
61 int err;
62
63 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
64 if (!mr)
65 return ERR_PTR(-ENOMEM);
66
67 err = mlx4_mr_alloc(to_mdev(pd->device)->dev, to_mpd(pd)->pdn, 0,
68 ~0ull, convert_access(acc), 0, 0, &mr->mmr);
69 if (err)
70 goto err_free;
71
72 err = mlx4_mr_enable(to_mdev(pd->device)->dev, &mr->mmr);
73 if (err)
74 goto err_mr;
75
76 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
77 mr->umem = NULL;
78
79 return &mr->ibmr;
80
81err_mr:
82 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
83
84err_free:
85 kfree(mr);
86
87 return ERR_PTR(err);
88}
89
90enum {
91 MLX4_MAX_MTT_SHIFT = 31
92};
93
94static int mlx4_ib_umem_write_mtt_block(struct mlx4_ib_dev *dev,
95 struct mlx4_mtt *mtt,
96 u64 mtt_size, u64 mtt_shift, u64 len,
97 u64 cur_start_addr, u64 *pages,
98 int *start_index, int *npages)
99{
100 u64 cur_end_addr = cur_start_addr + len;
101 u64 cur_end_addr_aligned = 0;
102 u64 mtt_entries;
103 int err = 0;
104 int k;
105
106 len += (cur_start_addr & (mtt_size - 1ULL));
107 cur_end_addr_aligned = round_up(cur_end_addr, mtt_size);
108 len += (cur_end_addr_aligned - cur_end_addr);
109 if (len & (mtt_size - 1ULL)) {
110 pr_warn("write_block: len %llx is not aligned to mtt_size %llx\n",
111 len, mtt_size);
112 return -EINVAL;
113 }
114
115 mtt_entries = (len >> mtt_shift);
116
117 /*
118 * Align the MTT start address to the mtt_size.
119 * Required to handle cases when the MR starts in the middle of an MTT
120 * record. Was not required in old code since the physical addresses
121 * provided by the dma subsystem were page aligned, which was also the
122 * MTT size.
123 */
124 cur_start_addr = round_down(cur_start_addr, mtt_size);
125 /* A new block is started ... */
126 for (k = 0; k < mtt_entries; ++k) {
127 pages[*npages] = cur_start_addr + (mtt_size * k);
128 (*npages)++;
129 /*
130 * Be friendly to mlx4_write_mtt() and pass it chunks of
131 * appropriate size.
132 */
133 if (*npages == PAGE_SIZE / sizeof(u64)) {
134 err = mlx4_write_mtt(dev->dev, mtt, *start_index,
135 *npages, pages);
136 if (err)
137 return err;
138
139 (*start_index) += *npages;
140 *npages = 0;
141 }
142 }
143
144 return 0;
145}
146
147static inline u64 alignment_of(u64 ptr)
148{
149 return ilog2(ptr & (~(ptr - 1)));
150}
151
152static int mlx4_ib_umem_calc_block_mtt(u64 next_block_start,
153 u64 current_block_end,
154 u64 block_shift)
155{
156 /* Check whether the alignment of the new block is aligned as well as
157 * the previous block.
158 * Block address must start with zeros till size of entity_size.
159 */
160 if ((next_block_start & ((1ULL << block_shift) - 1ULL)) != 0)
161 /*
162 * It is not as well aligned as the previous block-reduce the
163 * mtt size accordingly. Here we take the last right bit which
164 * is 1.
165 */
166 block_shift = alignment_of(next_block_start);
167
168 /*
169 * Check whether the alignment of the end of previous block - is it
170 * aligned as well as the start of the block
171 */
172 if (((current_block_end) & ((1ULL << block_shift) - 1ULL)) != 0)
173 /*
174 * It is not as well aligned as the start of the block -
175 * reduce the mtt size accordingly.
176 */
177 block_shift = alignment_of(current_block_end);
178
179 return block_shift;
180}
181
182int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt,
183 struct ib_umem *umem)
184{
185 u64 *pages;
186 u64 len = 0;
187 int err = 0;
188 u64 mtt_size;
189 u64 cur_start_addr = 0;
190 u64 mtt_shift;
191 int start_index = 0;
192 int npages = 0;
193 struct scatterlist *sg;
194 int i;
195
196 pages = (u64 *) __get_free_page(GFP_KERNEL);
197 if (!pages)
198 return -ENOMEM;
199
200 mtt_shift = mtt->page_shift;
201 mtt_size = 1ULL << mtt_shift;
202
203 for_each_sgtable_dma_sg(&umem->sgt_append.sgt, sg, i) {
204 if (cur_start_addr + len == sg_dma_address(sg)) {
205 /* still the same block */
206 len += sg_dma_len(sg);
207 continue;
208 }
209 /*
210 * A new block is started ...
211 * If len is malaligned, write an extra mtt entry to cover the
212 * misaligned area (round up the division)
213 */
214 err = mlx4_ib_umem_write_mtt_block(dev, mtt, mtt_size,
215 mtt_shift, len,
216 cur_start_addr,
217 pages, &start_index,
218 &npages);
219 if (err)
220 goto out;
221
222 cur_start_addr = sg_dma_address(sg);
223 len = sg_dma_len(sg);
224 }
225
226 /* Handle the last block */
227 if (len > 0) {
228 /*
229 * If len is malaligned, write an extra mtt entry to cover
230 * the misaligned area (round up the division)
231 */
232 err = mlx4_ib_umem_write_mtt_block(dev, mtt, mtt_size,
233 mtt_shift, len,
234 cur_start_addr, pages,
235 &start_index, &npages);
236 if (err)
237 goto out;
238 }
239
240 if (npages)
241 err = mlx4_write_mtt(dev->dev, mtt, start_index, npages, pages);
242
243out:
244 free_page((unsigned long) pages);
245 return err;
246}
247
248/*
249 * Calculate optimal mtt size based on contiguous pages.
250 * Function will return also the number of pages that are not aligned to the
251 * calculated mtt_size to be added to total number of pages. For that we should
252 * check the first chunk length & last chunk length and if not aligned to
253 * mtt_size we should increment the non_aligned_pages number. All chunks in the
254 * middle already handled as part of mtt shift calculation for both their start
255 * & end addresses.
256 */
257int mlx4_ib_umem_calc_optimal_mtt_size(struct ib_umem *umem, u64 start_va,
258 int *num_of_mtts)
259{
260 u64 block_shift = MLX4_MAX_MTT_SHIFT;
261 u64 min_shift = PAGE_SHIFT;
262 u64 last_block_aligned_end = 0;
263 u64 current_block_start = 0;
264 u64 first_block_start = 0;
265 u64 current_block_len = 0;
266 u64 last_block_end = 0;
267 struct scatterlist *sg;
268 u64 current_block_end;
269 u64 misalignment_bits;
270 u64 next_block_start;
271 u64 total_len = 0;
272 int i;
273
274 *num_of_mtts = ib_umem_num_dma_blocks(umem, PAGE_SIZE);
275
276 for_each_sgtable_dma_sg(&umem->sgt_append.sgt, sg, i) {
277 /*
278 * Initialization - save the first chunk start as the
279 * current_block_start - block means contiguous pages.
280 */
281 if (current_block_len == 0 && current_block_start == 0) {
282 current_block_start = sg_dma_address(sg);
283 first_block_start = current_block_start;
284 /*
285 * Find the bits that are different between the physical
286 * address and the virtual address for the start of the
287 * MR.
288 * umem_get aligned the start_va to a page boundary.
289 * Therefore, we need to align the start va to the same
290 * boundary.
291 * misalignment_bits is needed to handle the case of a
292 * single memory region. In this case, the rest of the
293 * logic will not reduce the block size. If we use a
294 * block size which is bigger than the alignment of the
295 * misalignment bits, we might use the virtual page
296 * number instead of the physical page number, resulting
297 * in access to the wrong data.
298 */
299 misalignment_bits =
300 (start_va & (~(((u64)(PAGE_SIZE)) - 1ULL))) ^
301 current_block_start;
302 block_shift = min(alignment_of(misalignment_bits),
303 block_shift);
304 }
305
306 /*
307 * Go over the scatter entries and check if they continue the
308 * previous scatter entry.
309 */
310 next_block_start = sg_dma_address(sg);
311 current_block_end = current_block_start + current_block_len;
312 /* If we have a split (non-contig.) between two blocks */
313 if (current_block_end != next_block_start) {
314 block_shift = mlx4_ib_umem_calc_block_mtt
315 (next_block_start,
316 current_block_end,
317 block_shift);
318
319 /*
320 * If we reached the minimum shift for 4k page we stop
321 * the loop.
322 */
323 if (block_shift <= min_shift)
324 goto end;
325
326 /*
327 * If not saved yet we are in first block - we save the
328 * length of first block to calculate the
329 * non_aligned_pages number at the end.
330 */
331 total_len += current_block_len;
332
333 /* Start a new block */
334 current_block_start = next_block_start;
335 current_block_len = sg_dma_len(sg);
336 continue;
337 }
338 /* The scatter entry is another part of the current block,
339 * increase the block size.
340 * An entry in the scatter can be larger than 4k (page) as of
341 * dma mapping which merge some blocks together.
342 */
343 current_block_len += sg_dma_len(sg);
344 }
345
346 /* Account for the last block in the total len */
347 total_len += current_block_len;
348 /* Add to the first block the misalignment that it suffers from. */
349 total_len += (first_block_start & ((1ULL << block_shift) - 1ULL));
350 last_block_end = current_block_start + current_block_len;
351 last_block_aligned_end = round_up(last_block_end, 1ULL << block_shift);
352 total_len += (last_block_aligned_end - last_block_end);
353
354 if (total_len & ((1ULL << block_shift) - 1ULL))
355 pr_warn("misaligned total length detected (%llu, %llu)!",
356 total_len, block_shift);
357
358 *num_of_mtts = total_len >> block_shift;
359end:
360 if (block_shift < min_shift) {
361 /*
362 * If shift is less than the min we set a warning and return the
363 * min shift.
364 */
365 pr_warn("umem_calc_optimal_mtt_size - unexpected shift %lld\n", block_shift);
366
367 block_shift = min_shift;
368 }
369 return block_shift;
370}
371
372static struct ib_umem *mlx4_get_umem_mr(struct ib_device *device, u64 start,
373 u64 length, int access_flags)
374{
375 /*
376 * Force registering the memory as writable if the underlying pages
377 * are writable. This is so rereg can change the access permissions
378 * from readable to writable without having to run through ib_umem_get
379 * again
380 */
381 if (!ib_access_writable(access_flags)) {
382 unsigned long untagged_start = untagged_addr(start);
383 struct vm_area_struct *vma;
384
385 mmap_read_lock(current->mm);
386 /*
387 * FIXME: Ideally this would iterate over all the vmas that
388 * cover the memory, but for now it requires a single vma to
389 * entirely cover the MR to support RO mappings.
390 */
391 vma = find_vma(current->mm, untagged_start);
392 if (vma && vma->vm_end >= untagged_start + length &&
393 vma->vm_start <= untagged_start) {
394 if (vma->vm_flags & VM_WRITE)
395 access_flags |= IB_ACCESS_LOCAL_WRITE;
396 } else {
397 access_flags |= IB_ACCESS_LOCAL_WRITE;
398 }
399
400 mmap_read_unlock(current->mm);
401 }
402
403 return ib_umem_get(device, start, length, access_flags);
404}
405
406struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
407 u64 virt_addr, int access_flags,
408 struct ib_udata *udata)
409{
410 struct mlx4_ib_dev *dev = to_mdev(pd->device);
411 struct mlx4_ib_mr *mr;
412 int shift;
413 int err;
414 int n;
415
416 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
417 if (!mr)
418 return ERR_PTR(-ENOMEM);
419
420 mr->umem = mlx4_get_umem_mr(pd->device, start, length, access_flags);
421 if (IS_ERR(mr->umem)) {
422 err = PTR_ERR(mr->umem);
423 goto err_free;
424 }
425
426 shift = mlx4_ib_umem_calc_optimal_mtt_size(mr->umem, start, &n);
427
428 err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length,
429 convert_access(access_flags), n, shift, &mr->mmr);
430 if (err)
431 goto err_umem;
432
433 err = mlx4_ib_umem_write_mtt(dev, &mr->mmr.mtt, mr->umem);
434 if (err)
435 goto err_mr;
436
437 err = mlx4_mr_enable(dev->dev, &mr->mmr);
438 if (err)
439 goto err_mr;
440
441 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
442 mr->ibmr.page_size = 1U << shift;
443
444 return &mr->ibmr;
445
446err_mr:
447 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
448
449err_umem:
450 ib_umem_release(mr->umem);
451
452err_free:
453 kfree(mr);
454
455 return ERR_PTR(err);
456}
457
458struct ib_mr *mlx4_ib_rereg_user_mr(struct ib_mr *mr, int flags, u64 start,
459 u64 length, u64 virt_addr,
460 int mr_access_flags, struct ib_pd *pd,
461 struct ib_udata *udata)
462{
463 struct mlx4_ib_dev *dev = to_mdev(mr->device);
464 struct mlx4_ib_mr *mmr = to_mmr(mr);
465 struct mlx4_mpt_entry *mpt_entry;
466 struct mlx4_mpt_entry **pmpt_entry = &mpt_entry;
467 int err;
468
469 /* Since we synchronize this call and mlx4_ib_dereg_mr via uverbs,
470 * we assume that the calls can't run concurrently. Otherwise, a
471 * race exists.
472 */
473 err = mlx4_mr_hw_get_mpt(dev->dev, &mmr->mmr, &pmpt_entry);
474 if (err)
475 return ERR_PTR(err);
476
477 if (flags & IB_MR_REREG_PD) {
478 err = mlx4_mr_hw_change_pd(dev->dev, *pmpt_entry,
479 to_mpd(pd)->pdn);
480
481 if (err)
482 goto release_mpt_entry;
483 }
484
485 if (flags & IB_MR_REREG_ACCESS) {
486 if (ib_access_writable(mr_access_flags) &&
487 !mmr->umem->writable) {
488 err = -EPERM;
489 goto release_mpt_entry;
490 }
491
492 err = mlx4_mr_hw_change_access(dev->dev, *pmpt_entry,
493 convert_access(mr_access_flags));
494
495 if (err)
496 goto release_mpt_entry;
497 }
498
499 if (flags & IB_MR_REREG_TRANS) {
500 int shift;
501 int n;
502
503 mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
504 ib_umem_release(mmr->umem);
505 mmr->umem = mlx4_get_umem_mr(mr->device, start, length,
506 mr_access_flags);
507 if (IS_ERR(mmr->umem)) {
508 err = PTR_ERR(mmr->umem);
509 /* Prevent mlx4_ib_dereg_mr from free'ing invalid pointer */
510 mmr->umem = NULL;
511 goto release_mpt_entry;
512 }
513 n = ib_umem_num_dma_blocks(mmr->umem, PAGE_SIZE);
514 shift = PAGE_SHIFT;
515
516 err = mlx4_mr_rereg_mem_write(dev->dev, &mmr->mmr,
517 virt_addr, length, n, shift,
518 *pmpt_entry);
519 if (err) {
520 ib_umem_release(mmr->umem);
521 goto release_mpt_entry;
522 }
523 mmr->mmr.iova = virt_addr;
524 mmr->mmr.size = length;
525
526 err = mlx4_ib_umem_write_mtt(dev, &mmr->mmr.mtt, mmr->umem);
527 if (err) {
528 mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
529 ib_umem_release(mmr->umem);
530 goto release_mpt_entry;
531 }
532 }
533
534 /* If we couldn't transfer the MR to the HCA, just remember to
535 * return a failure. But dereg_mr will free the resources.
536 */
537 err = mlx4_mr_hw_write_mpt(dev->dev, &mmr->mmr, pmpt_entry);
538 if (!err && flags & IB_MR_REREG_ACCESS)
539 mmr->mmr.access = mr_access_flags;
540
541release_mpt_entry:
542 mlx4_mr_hw_put_mpt(dev->dev, pmpt_entry);
543 if (err)
544 return ERR_PTR(err);
545 return NULL;
546}
547
548static int
549mlx4_alloc_priv_pages(struct ib_device *device,
550 struct mlx4_ib_mr *mr,
551 int max_pages)
552{
553 int ret;
554
555 /* Ensure that size is aligned to DMA cacheline
556 * requirements.
557 * max_pages is limited to MLX4_MAX_FAST_REG_PAGES
558 * so page_map_size will never cross PAGE_SIZE.
559 */
560 mr->page_map_size = roundup(max_pages * sizeof(u64),
561 MLX4_MR_PAGES_ALIGN);
562
563 /* Prevent cross page boundary allocation. */
564 mr->pages = (__be64 *)get_zeroed_page(GFP_KERNEL);
565 if (!mr->pages)
566 return -ENOMEM;
567
568 mr->page_map = dma_map_single(device->dev.parent, mr->pages,
569 mr->page_map_size, DMA_TO_DEVICE);
570
571 if (dma_mapping_error(device->dev.parent, mr->page_map)) {
572 ret = -ENOMEM;
573 goto err;
574 }
575
576 return 0;
577
578err:
579 free_page((unsigned long)mr->pages);
580 return ret;
581}
582
583static void
584mlx4_free_priv_pages(struct mlx4_ib_mr *mr)
585{
586 if (mr->pages) {
587 struct ib_device *device = mr->ibmr.device;
588
589 dma_unmap_single(device->dev.parent, mr->page_map,
590 mr->page_map_size, DMA_TO_DEVICE);
591 free_page((unsigned long)mr->pages);
592 mr->pages = NULL;
593 }
594}
595
596int mlx4_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
597{
598 struct mlx4_ib_mr *mr = to_mmr(ibmr);
599 int ret;
600
601 mlx4_free_priv_pages(mr);
602
603 ret = mlx4_mr_free(to_mdev(ibmr->device)->dev, &mr->mmr);
604 if (ret)
605 return ret;
606 if (mr->umem)
607 ib_umem_release(mr->umem);
608 kfree(mr);
609
610 return 0;
611}
612
613int mlx4_ib_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
614{
615 struct mlx4_ib_dev *dev = to_mdev(ibmw->device);
616 struct mlx4_ib_mw *mw = to_mmw(ibmw);
617 int err;
618
619 err = mlx4_mw_alloc(dev->dev, to_mpd(ibmw->pd)->pdn,
620 to_mlx4_type(ibmw->type), &mw->mmw);
621 if (err)
622 return err;
623
624 err = mlx4_mw_enable(dev->dev, &mw->mmw);
625 if (err)
626 goto err_mw;
627
628 ibmw->rkey = mw->mmw.key;
629 return 0;
630
631err_mw:
632 mlx4_mw_free(dev->dev, &mw->mmw);
633 return err;
634}
635
636int mlx4_ib_dealloc_mw(struct ib_mw *ibmw)
637{
638 struct mlx4_ib_mw *mw = to_mmw(ibmw);
639
640 mlx4_mw_free(to_mdev(ibmw->device)->dev, &mw->mmw);
641 return 0;
642}
643
644struct ib_mr *mlx4_ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
645 u32 max_num_sg)
646{
647 struct mlx4_ib_dev *dev = to_mdev(pd->device);
648 struct mlx4_ib_mr *mr;
649 int err;
650
651 if (mr_type != IB_MR_TYPE_MEM_REG ||
652 max_num_sg > MLX4_MAX_FAST_REG_PAGES)
653 return ERR_PTR(-EINVAL);
654
655 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
656 if (!mr)
657 return ERR_PTR(-ENOMEM);
658
659 err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, 0, 0, 0,
660 max_num_sg, 0, &mr->mmr);
661 if (err)
662 goto err_free;
663
664 err = mlx4_alloc_priv_pages(pd->device, mr, max_num_sg);
665 if (err)
666 goto err_free_mr;
667
668 mr->max_pages = max_num_sg;
669 err = mlx4_mr_enable(dev->dev, &mr->mmr);
670 if (err)
671 goto err_free_pl;
672
673 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
674 mr->umem = NULL;
675
676 return &mr->ibmr;
677
678err_free_pl:
679 mr->ibmr.device = pd->device;
680 mlx4_free_priv_pages(mr);
681err_free_mr:
682 (void) mlx4_mr_free(dev->dev, &mr->mmr);
683err_free:
684 kfree(mr);
685 return ERR_PTR(err);
686}
687
688static int mlx4_set_page(struct ib_mr *ibmr, u64 addr)
689{
690 struct mlx4_ib_mr *mr = to_mmr(ibmr);
691
692 if (unlikely(mr->npages == mr->max_pages))
693 return -ENOMEM;
694
695 mr->pages[mr->npages++] = cpu_to_be64(addr | MLX4_MTT_FLAG_PRESENT);
696
697 return 0;
698}
699
700int mlx4_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
701 unsigned int *sg_offset)
702{
703 struct mlx4_ib_mr *mr = to_mmr(ibmr);
704 int rc;
705
706 mr->npages = 0;
707
708 ib_dma_sync_single_for_cpu(ibmr->device, mr->page_map,
709 mr->page_map_size, DMA_TO_DEVICE);
710
711 rc = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, mlx4_set_page);
712
713 ib_dma_sync_single_for_device(ibmr->device, mr->page_map,
714 mr->page_map_size, DMA_TO_DEVICE);
715
716 return rc;
717}
1/*
2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3 * Copyright (c) 2007, 2008 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 <rdma/ib_user_verbs.h>
36
37#include "mlx4_ib.h"
38
39static u32 convert_access(int acc)
40{
41 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX4_PERM_ATOMIC : 0) |
42 (acc & IB_ACCESS_REMOTE_WRITE ? MLX4_PERM_REMOTE_WRITE : 0) |
43 (acc & IB_ACCESS_REMOTE_READ ? MLX4_PERM_REMOTE_READ : 0) |
44 (acc & IB_ACCESS_LOCAL_WRITE ? MLX4_PERM_LOCAL_WRITE : 0) |
45 (acc & IB_ACCESS_MW_BIND ? MLX4_PERM_BIND_MW : 0) |
46 MLX4_PERM_LOCAL_READ;
47}
48
49static enum mlx4_mw_type to_mlx4_type(enum ib_mw_type type)
50{
51 switch (type) {
52 case IB_MW_TYPE_1: return MLX4_MW_TYPE_1;
53 case IB_MW_TYPE_2: return MLX4_MW_TYPE_2;
54 default: return -1;
55 }
56}
57
58struct ib_mr *mlx4_ib_get_dma_mr(struct ib_pd *pd, int acc)
59{
60 struct mlx4_ib_mr *mr;
61 int err;
62
63 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
64 if (!mr)
65 return ERR_PTR(-ENOMEM);
66
67 err = mlx4_mr_alloc(to_mdev(pd->device)->dev, to_mpd(pd)->pdn, 0,
68 ~0ull, convert_access(acc), 0, 0, &mr->mmr);
69 if (err)
70 goto err_free;
71
72 err = mlx4_mr_enable(to_mdev(pd->device)->dev, &mr->mmr);
73 if (err)
74 goto err_mr;
75
76 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
77 mr->umem = NULL;
78
79 return &mr->ibmr;
80
81err_mr:
82 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
83
84err_free:
85 kfree(mr);
86
87 return ERR_PTR(err);
88}
89
90enum {
91 MLX4_MAX_MTT_SHIFT = 31
92};
93
94static int mlx4_ib_umem_write_mtt_block(struct mlx4_ib_dev *dev,
95 struct mlx4_mtt *mtt,
96 u64 mtt_size, u64 mtt_shift, u64 len,
97 u64 cur_start_addr, u64 *pages,
98 int *start_index, int *npages)
99{
100 u64 cur_end_addr = cur_start_addr + len;
101 u64 cur_end_addr_aligned = 0;
102 u64 mtt_entries;
103 int err = 0;
104 int k;
105
106 len += (cur_start_addr & (mtt_size - 1ULL));
107 cur_end_addr_aligned = round_up(cur_end_addr, mtt_size);
108 len += (cur_end_addr_aligned - cur_end_addr);
109 if (len & (mtt_size - 1ULL)) {
110 pr_warn("write_block: len %llx is not aligned to mtt_size %llx\n",
111 len, mtt_size);
112 return -EINVAL;
113 }
114
115 mtt_entries = (len >> mtt_shift);
116
117 /*
118 * Align the MTT start address to the mtt_size.
119 * Required to handle cases when the MR starts in the middle of an MTT
120 * record. Was not required in old code since the physical addresses
121 * provided by the dma subsystem were page aligned, which was also the
122 * MTT size.
123 */
124 cur_start_addr = round_down(cur_start_addr, mtt_size);
125 /* A new block is started ... */
126 for (k = 0; k < mtt_entries; ++k) {
127 pages[*npages] = cur_start_addr + (mtt_size * k);
128 (*npages)++;
129 /*
130 * Be friendly to mlx4_write_mtt() and pass it chunks of
131 * appropriate size.
132 */
133 if (*npages == PAGE_SIZE / sizeof(u64)) {
134 err = mlx4_write_mtt(dev->dev, mtt, *start_index,
135 *npages, pages);
136 if (err)
137 return err;
138
139 (*start_index) += *npages;
140 *npages = 0;
141 }
142 }
143
144 return 0;
145}
146
147static inline u64 alignment_of(u64 ptr)
148{
149 return ilog2(ptr & (~(ptr - 1)));
150}
151
152static int mlx4_ib_umem_calc_block_mtt(u64 next_block_start,
153 u64 current_block_end,
154 u64 block_shift)
155{
156 /* Check whether the alignment of the new block is aligned as well as
157 * the previous block.
158 * Block address must start with zeros till size of entity_size.
159 */
160 if ((next_block_start & ((1ULL << block_shift) - 1ULL)) != 0)
161 /*
162 * It is not as well aligned as the previous block-reduce the
163 * mtt size accordingly. Here we take the last right bit which
164 * is 1.
165 */
166 block_shift = alignment_of(next_block_start);
167
168 /*
169 * Check whether the alignment of the end of previous block - is it
170 * aligned as well as the start of the block
171 */
172 if (((current_block_end) & ((1ULL << block_shift) - 1ULL)) != 0)
173 /*
174 * It is not as well aligned as the start of the block -
175 * reduce the mtt size accordingly.
176 */
177 block_shift = alignment_of(current_block_end);
178
179 return block_shift;
180}
181
182int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt,
183 struct ib_umem *umem)
184{
185 u64 *pages;
186 u64 len = 0;
187 int err = 0;
188 u64 mtt_size;
189 u64 cur_start_addr = 0;
190 u64 mtt_shift;
191 int start_index = 0;
192 int npages = 0;
193 struct scatterlist *sg;
194 int i;
195
196 pages = (u64 *) __get_free_page(GFP_KERNEL);
197 if (!pages)
198 return -ENOMEM;
199
200 mtt_shift = mtt->page_shift;
201 mtt_size = 1ULL << mtt_shift;
202
203 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
204 if (cur_start_addr + len == sg_dma_address(sg)) {
205 /* still the same block */
206 len += sg_dma_len(sg);
207 continue;
208 }
209 /*
210 * A new block is started ...
211 * If len is malaligned, write an extra mtt entry to cover the
212 * misaligned area (round up the division)
213 */
214 err = mlx4_ib_umem_write_mtt_block(dev, mtt, mtt_size,
215 mtt_shift, len,
216 cur_start_addr,
217 pages, &start_index,
218 &npages);
219 if (err)
220 goto out;
221
222 cur_start_addr = sg_dma_address(sg);
223 len = sg_dma_len(sg);
224 }
225
226 /* Handle the last block */
227 if (len > 0) {
228 /*
229 * If len is malaligned, write an extra mtt entry to cover
230 * the misaligned area (round up the division)
231 */
232 err = mlx4_ib_umem_write_mtt_block(dev, mtt, mtt_size,
233 mtt_shift, len,
234 cur_start_addr, pages,
235 &start_index, &npages);
236 if (err)
237 goto out;
238 }
239
240 if (npages)
241 err = mlx4_write_mtt(dev->dev, mtt, start_index, npages, pages);
242
243out:
244 free_page((unsigned long) pages);
245 return err;
246}
247
248/*
249 * Calculate optimal mtt size based on contiguous pages.
250 * Function will return also the number of pages that are not aligned to the
251 * calculated mtt_size to be added to total number of pages. For that we should
252 * check the first chunk length & last chunk length and if not aligned to
253 * mtt_size we should increment the non_aligned_pages number. All chunks in the
254 * middle already handled as part of mtt shift calculation for both their start
255 * & end addresses.
256 */
257int mlx4_ib_umem_calc_optimal_mtt_size(struct ib_umem *umem, u64 start_va,
258 int *num_of_mtts)
259{
260 u64 block_shift = MLX4_MAX_MTT_SHIFT;
261 u64 min_shift = PAGE_SHIFT;
262 u64 last_block_aligned_end = 0;
263 u64 current_block_start = 0;
264 u64 first_block_start = 0;
265 u64 current_block_len = 0;
266 u64 last_block_end = 0;
267 struct scatterlist *sg;
268 u64 current_block_end;
269 u64 misalignment_bits;
270 u64 next_block_start;
271 u64 total_len = 0;
272 int i;
273
274 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
275 /*
276 * Initialization - save the first chunk start as the
277 * current_block_start - block means contiguous pages.
278 */
279 if (current_block_len == 0 && current_block_start == 0) {
280 current_block_start = sg_dma_address(sg);
281 first_block_start = current_block_start;
282 /*
283 * Find the bits that are different between the physical
284 * address and the virtual address for the start of the
285 * MR.
286 * umem_get aligned the start_va to a page boundary.
287 * Therefore, we need to align the start va to the same
288 * boundary.
289 * misalignment_bits is needed to handle the case of a
290 * single memory region. In this case, the rest of the
291 * logic will not reduce the block size. If we use a
292 * block size which is bigger than the alignment of the
293 * misalignment bits, we might use the virtual page
294 * number instead of the physical page number, resulting
295 * in access to the wrong data.
296 */
297 misalignment_bits =
298 (start_va & (~(((u64)(PAGE_SIZE)) - 1ULL))) ^
299 current_block_start;
300 block_shift = min(alignment_of(misalignment_bits),
301 block_shift);
302 }
303
304 /*
305 * Go over the scatter entries and check if they continue the
306 * previous scatter entry.
307 */
308 next_block_start = sg_dma_address(sg);
309 current_block_end = current_block_start + current_block_len;
310 /* If we have a split (non-contig.) between two blocks */
311 if (current_block_end != next_block_start) {
312 block_shift = mlx4_ib_umem_calc_block_mtt
313 (next_block_start,
314 current_block_end,
315 block_shift);
316
317 /*
318 * If we reached the minimum shift for 4k page we stop
319 * the loop.
320 */
321 if (block_shift <= min_shift)
322 goto end;
323
324 /*
325 * If not saved yet we are in first block - we save the
326 * length of first block to calculate the
327 * non_aligned_pages number at the end.
328 */
329 total_len += current_block_len;
330
331 /* Start a new block */
332 current_block_start = next_block_start;
333 current_block_len = sg_dma_len(sg);
334 continue;
335 }
336 /* The scatter entry is another part of the current block,
337 * increase the block size.
338 * An entry in the scatter can be larger than 4k (page) as of
339 * dma mapping which merge some blocks together.
340 */
341 current_block_len += sg_dma_len(sg);
342 }
343
344 /* Account for the last block in the total len */
345 total_len += current_block_len;
346 /* Add to the first block the misalignment that it suffers from. */
347 total_len += (first_block_start & ((1ULL << block_shift) - 1ULL));
348 last_block_end = current_block_start + current_block_len;
349 last_block_aligned_end = round_up(last_block_end, 1ULL << block_shift);
350 total_len += (last_block_aligned_end - last_block_end);
351
352 if (total_len & ((1ULL << block_shift) - 1ULL))
353 pr_warn("misaligned total length detected (%llu, %llu)!",
354 total_len, block_shift);
355
356 *num_of_mtts = total_len >> block_shift;
357end:
358 if (block_shift < min_shift) {
359 /*
360 * If shift is less than the min we set a warning and return the
361 * min shift.
362 */
363 pr_warn("umem_calc_optimal_mtt_size - unexpected shift %lld\n", block_shift);
364
365 block_shift = min_shift;
366 }
367 return block_shift;
368}
369
370static struct ib_umem *mlx4_get_umem_mr(struct ib_udata *udata, u64 start,
371 u64 length, int access_flags)
372{
373 /*
374 * Force registering the memory as writable if the underlying pages
375 * are writable. This is so rereg can change the access permissions
376 * from readable to writable without having to run through ib_umem_get
377 * again
378 */
379 if (!ib_access_writable(access_flags)) {
380 unsigned long untagged_start = untagged_addr(start);
381 struct vm_area_struct *vma;
382
383 down_read(¤t->mm->mmap_sem);
384 /*
385 * FIXME: Ideally this would iterate over all the vmas that
386 * cover the memory, but for now it requires a single vma to
387 * entirely cover the MR to support RO mappings.
388 */
389 vma = find_vma(current->mm, untagged_start);
390 if (vma && vma->vm_end >= untagged_start + length &&
391 vma->vm_start <= untagged_start) {
392 if (vma->vm_flags & VM_WRITE)
393 access_flags |= IB_ACCESS_LOCAL_WRITE;
394 } else {
395 access_flags |= IB_ACCESS_LOCAL_WRITE;
396 }
397
398 up_read(¤t->mm->mmap_sem);
399 }
400
401 return ib_umem_get(udata, start, length, access_flags, 0);
402}
403
404struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
405 u64 virt_addr, int access_flags,
406 struct ib_udata *udata)
407{
408 struct mlx4_ib_dev *dev = to_mdev(pd->device);
409 struct mlx4_ib_mr *mr;
410 int shift;
411 int err;
412 int n;
413
414 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
415 if (!mr)
416 return ERR_PTR(-ENOMEM);
417
418 mr->umem = mlx4_get_umem_mr(udata, start, length, access_flags);
419 if (IS_ERR(mr->umem)) {
420 err = PTR_ERR(mr->umem);
421 goto err_free;
422 }
423
424 n = ib_umem_page_count(mr->umem);
425 shift = mlx4_ib_umem_calc_optimal_mtt_size(mr->umem, start, &n);
426
427 err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length,
428 convert_access(access_flags), n, shift, &mr->mmr);
429 if (err)
430 goto err_umem;
431
432 err = mlx4_ib_umem_write_mtt(dev, &mr->mmr.mtt, mr->umem);
433 if (err)
434 goto err_mr;
435
436 err = mlx4_mr_enable(dev->dev, &mr->mmr);
437 if (err)
438 goto err_mr;
439
440 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
441 mr->ibmr.length = length;
442 mr->ibmr.iova = virt_addr;
443 mr->ibmr.page_size = 1U << shift;
444
445 return &mr->ibmr;
446
447err_mr:
448 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
449
450err_umem:
451 ib_umem_release(mr->umem);
452
453err_free:
454 kfree(mr);
455
456 return ERR_PTR(err);
457}
458
459int mlx4_ib_rereg_user_mr(struct ib_mr *mr, int flags,
460 u64 start, u64 length, u64 virt_addr,
461 int mr_access_flags, struct ib_pd *pd,
462 struct ib_udata *udata)
463{
464 struct mlx4_ib_dev *dev = to_mdev(mr->device);
465 struct mlx4_ib_mr *mmr = to_mmr(mr);
466 struct mlx4_mpt_entry *mpt_entry;
467 struct mlx4_mpt_entry **pmpt_entry = &mpt_entry;
468 int err;
469
470 /* Since we synchronize this call and mlx4_ib_dereg_mr via uverbs,
471 * we assume that the calls can't run concurrently. Otherwise, a
472 * race exists.
473 */
474 err = mlx4_mr_hw_get_mpt(dev->dev, &mmr->mmr, &pmpt_entry);
475
476 if (err)
477 return err;
478
479 if (flags & IB_MR_REREG_PD) {
480 err = mlx4_mr_hw_change_pd(dev->dev, *pmpt_entry,
481 to_mpd(pd)->pdn);
482
483 if (err)
484 goto release_mpt_entry;
485 }
486
487 if (flags & IB_MR_REREG_ACCESS) {
488 if (ib_access_writable(mr_access_flags) &&
489 !mmr->umem->writable) {
490 err = -EPERM;
491 goto release_mpt_entry;
492 }
493
494 err = mlx4_mr_hw_change_access(dev->dev, *pmpt_entry,
495 convert_access(mr_access_flags));
496
497 if (err)
498 goto release_mpt_entry;
499 }
500
501 if (flags & IB_MR_REREG_TRANS) {
502 int shift;
503 int n;
504
505 mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
506 ib_umem_release(mmr->umem);
507 mmr->umem = mlx4_get_umem_mr(udata, start, length,
508 mr_access_flags);
509 if (IS_ERR(mmr->umem)) {
510 err = PTR_ERR(mmr->umem);
511 /* Prevent mlx4_ib_dereg_mr from free'ing invalid pointer */
512 mmr->umem = NULL;
513 goto release_mpt_entry;
514 }
515 n = ib_umem_page_count(mmr->umem);
516 shift = PAGE_SHIFT;
517
518 err = mlx4_mr_rereg_mem_write(dev->dev, &mmr->mmr,
519 virt_addr, length, n, shift,
520 *pmpt_entry);
521 if (err) {
522 ib_umem_release(mmr->umem);
523 goto release_mpt_entry;
524 }
525 mmr->mmr.iova = virt_addr;
526 mmr->mmr.size = length;
527
528 err = mlx4_ib_umem_write_mtt(dev, &mmr->mmr.mtt, mmr->umem);
529 if (err) {
530 mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
531 ib_umem_release(mmr->umem);
532 goto release_mpt_entry;
533 }
534 }
535
536 /* If we couldn't transfer the MR to the HCA, just remember to
537 * return a failure. But dereg_mr will free the resources.
538 */
539 err = mlx4_mr_hw_write_mpt(dev->dev, &mmr->mmr, pmpt_entry);
540 if (!err && flags & IB_MR_REREG_ACCESS)
541 mmr->mmr.access = mr_access_flags;
542
543release_mpt_entry:
544 mlx4_mr_hw_put_mpt(dev->dev, pmpt_entry);
545
546 return err;
547}
548
549static int
550mlx4_alloc_priv_pages(struct ib_device *device,
551 struct mlx4_ib_mr *mr,
552 int max_pages)
553{
554 int ret;
555
556 /* Ensure that size is aligned to DMA cacheline
557 * requirements.
558 * max_pages is limited to MLX4_MAX_FAST_REG_PAGES
559 * so page_map_size will never cross PAGE_SIZE.
560 */
561 mr->page_map_size = roundup(max_pages * sizeof(u64),
562 MLX4_MR_PAGES_ALIGN);
563
564 /* Prevent cross page boundary allocation. */
565 mr->pages = (__be64 *)get_zeroed_page(GFP_KERNEL);
566 if (!mr->pages)
567 return -ENOMEM;
568
569 mr->page_map = dma_map_single(device->dev.parent, mr->pages,
570 mr->page_map_size, DMA_TO_DEVICE);
571
572 if (dma_mapping_error(device->dev.parent, mr->page_map)) {
573 ret = -ENOMEM;
574 goto err;
575 }
576
577 return 0;
578
579err:
580 free_page((unsigned long)mr->pages);
581 return ret;
582}
583
584static void
585mlx4_free_priv_pages(struct mlx4_ib_mr *mr)
586{
587 if (mr->pages) {
588 struct ib_device *device = mr->ibmr.device;
589
590 dma_unmap_single(device->dev.parent, mr->page_map,
591 mr->page_map_size, DMA_TO_DEVICE);
592 free_page((unsigned long)mr->pages);
593 mr->pages = NULL;
594 }
595}
596
597int mlx4_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
598{
599 struct mlx4_ib_mr *mr = to_mmr(ibmr);
600 int ret;
601
602 mlx4_free_priv_pages(mr);
603
604 ret = mlx4_mr_free(to_mdev(ibmr->device)->dev, &mr->mmr);
605 if (ret)
606 return ret;
607 if (mr->umem)
608 ib_umem_release(mr->umem);
609 kfree(mr);
610
611 return 0;
612}
613
614struct ib_mw *mlx4_ib_alloc_mw(struct ib_pd *pd, enum ib_mw_type type,
615 struct ib_udata *udata)
616{
617 struct mlx4_ib_dev *dev = to_mdev(pd->device);
618 struct mlx4_ib_mw *mw;
619 int err;
620
621 mw = kmalloc(sizeof(*mw), GFP_KERNEL);
622 if (!mw)
623 return ERR_PTR(-ENOMEM);
624
625 err = mlx4_mw_alloc(dev->dev, to_mpd(pd)->pdn,
626 to_mlx4_type(type), &mw->mmw);
627 if (err)
628 goto err_free;
629
630 err = mlx4_mw_enable(dev->dev, &mw->mmw);
631 if (err)
632 goto err_mw;
633
634 mw->ibmw.rkey = mw->mmw.key;
635
636 return &mw->ibmw;
637
638err_mw:
639 mlx4_mw_free(dev->dev, &mw->mmw);
640
641err_free:
642 kfree(mw);
643
644 return ERR_PTR(err);
645}
646
647int mlx4_ib_dealloc_mw(struct ib_mw *ibmw)
648{
649 struct mlx4_ib_mw *mw = to_mmw(ibmw);
650
651 mlx4_mw_free(to_mdev(ibmw->device)->dev, &mw->mmw);
652 kfree(mw);
653
654 return 0;
655}
656
657struct ib_mr *mlx4_ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
658 u32 max_num_sg, struct ib_udata *udata)
659{
660 struct mlx4_ib_dev *dev = to_mdev(pd->device);
661 struct mlx4_ib_mr *mr;
662 int err;
663
664 if (mr_type != IB_MR_TYPE_MEM_REG ||
665 max_num_sg > MLX4_MAX_FAST_REG_PAGES)
666 return ERR_PTR(-EINVAL);
667
668 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
669 if (!mr)
670 return ERR_PTR(-ENOMEM);
671
672 err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, 0, 0, 0,
673 max_num_sg, 0, &mr->mmr);
674 if (err)
675 goto err_free;
676
677 err = mlx4_alloc_priv_pages(pd->device, mr, max_num_sg);
678 if (err)
679 goto err_free_mr;
680
681 mr->max_pages = max_num_sg;
682 err = mlx4_mr_enable(dev->dev, &mr->mmr);
683 if (err)
684 goto err_free_pl;
685
686 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
687 mr->umem = NULL;
688
689 return &mr->ibmr;
690
691err_free_pl:
692 mr->ibmr.device = pd->device;
693 mlx4_free_priv_pages(mr);
694err_free_mr:
695 (void) mlx4_mr_free(dev->dev, &mr->mmr);
696err_free:
697 kfree(mr);
698 return ERR_PTR(err);
699}
700
701struct ib_fmr *mlx4_ib_fmr_alloc(struct ib_pd *pd, int acc,
702 struct ib_fmr_attr *fmr_attr)
703{
704 struct mlx4_ib_dev *dev = to_mdev(pd->device);
705 struct mlx4_ib_fmr *fmr;
706 int err = -ENOMEM;
707
708 fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
709 if (!fmr)
710 return ERR_PTR(-ENOMEM);
711
712 err = mlx4_fmr_alloc(dev->dev, to_mpd(pd)->pdn, convert_access(acc),
713 fmr_attr->max_pages, fmr_attr->max_maps,
714 fmr_attr->page_shift, &fmr->mfmr);
715 if (err)
716 goto err_free;
717
718 err = mlx4_fmr_enable(to_mdev(pd->device)->dev, &fmr->mfmr);
719 if (err)
720 goto err_mr;
721
722 fmr->ibfmr.rkey = fmr->ibfmr.lkey = fmr->mfmr.mr.key;
723
724 return &fmr->ibfmr;
725
726err_mr:
727 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &fmr->mfmr.mr);
728
729err_free:
730 kfree(fmr);
731
732 return ERR_PTR(err);
733}
734
735int mlx4_ib_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
736 int npages, u64 iova)
737{
738 struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
739 struct mlx4_ib_dev *dev = to_mdev(ifmr->ibfmr.device);
740
741 return mlx4_map_phys_fmr(dev->dev, &ifmr->mfmr, page_list, npages, iova,
742 &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey);
743}
744
745int mlx4_ib_unmap_fmr(struct list_head *fmr_list)
746{
747 struct ib_fmr *ibfmr;
748 int err;
749 struct mlx4_dev *mdev = NULL;
750
751 list_for_each_entry(ibfmr, fmr_list, list) {
752 if (mdev && to_mdev(ibfmr->device)->dev != mdev)
753 return -EINVAL;
754 mdev = to_mdev(ibfmr->device)->dev;
755 }
756
757 if (!mdev)
758 return 0;
759
760 list_for_each_entry(ibfmr, fmr_list, list) {
761 struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
762
763 mlx4_fmr_unmap(mdev, &ifmr->mfmr, &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey);
764 }
765
766 /*
767 * Make sure all MPT status updates are visible before issuing
768 * SYNC_TPT firmware command.
769 */
770 wmb();
771
772 err = mlx4_SYNC_TPT(mdev);
773 if (err)
774 pr_warn("SYNC_TPT error %d when "
775 "unmapping FMRs\n", err);
776
777 return 0;
778}
779
780int mlx4_ib_fmr_dealloc(struct ib_fmr *ibfmr)
781{
782 struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
783 struct mlx4_ib_dev *dev = to_mdev(ibfmr->device);
784 int err;
785
786 err = mlx4_fmr_free(dev->dev, &ifmr->mfmr);
787
788 if (!err)
789 kfree(ifmr);
790
791 return err;
792}
793
794static int mlx4_set_page(struct ib_mr *ibmr, u64 addr)
795{
796 struct mlx4_ib_mr *mr = to_mmr(ibmr);
797
798 if (unlikely(mr->npages == mr->max_pages))
799 return -ENOMEM;
800
801 mr->pages[mr->npages++] = cpu_to_be64(addr | MLX4_MTT_FLAG_PRESENT);
802
803 return 0;
804}
805
806int mlx4_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
807 unsigned int *sg_offset)
808{
809 struct mlx4_ib_mr *mr = to_mmr(ibmr);
810 int rc;
811
812 mr->npages = 0;
813
814 ib_dma_sync_single_for_cpu(ibmr->device, mr->page_map,
815 mr->page_map_size, DMA_TO_DEVICE);
816
817 rc = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, mlx4_set_page);
818
819 ib_dma_sync_single_for_device(ibmr->device, mr->page_map,
820 mr->page_map_size, DMA_TO_DEVICE);
821
822 return rc;
823}