xref: /openbmc/linux/drivers/infiniband/sw/rxe/rxe_mr.c (revision 4b33b5ff)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5  */
6 
7 #include "rxe.h"
8 #include "rxe_loc.h"
9 
10 /* Return a random 8 bit key value that is
11  * different than the last_key. Set last_key to -1
12  * if this is the first key for an MR or MW
13  */
14 u8 rxe_get_next_key(u32 last_key)
15 {
16 	u8 key;
17 
18 	do {
19 		get_random_bytes(&key, 1);
20 	} while (key == last_key);
21 
22 	return key;
23 }
24 
25 int mr_check_range(struct rxe_mr *mr, u64 iova, size_t length)
26 {
27 	struct rxe_map_set *set = mr->cur_map_set;
28 
29 	switch (mr->type) {
30 	case IB_MR_TYPE_DMA:
31 		return 0;
32 
33 	case IB_MR_TYPE_USER:
34 	case IB_MR_TYPE_MEM_REG:
35 		if (iova < set->iova || length > set->length ||
36 		    iova > set->iova + set->length - length)
37 			return -EFAULT;
38 		return 0;
39 
40 	default:
41 		pr_warn("%s: mr type (%d) not supported\n",
42 			__func__, mr->type);
43 		return -EFAULT;
44 	}
45 }
46 
47 #define IB_ACCESS_REMOTE	(IB_ACCESS_REMOTE_READ		\
48 				| IB_ACCESS_REMOTE_WRITE	\
49 				| IB_ACCESS_REMOTE_ATOMIC)
50 
51 static void rxe_mr_init(int access, struct rxe_mr *mr)
52 {
53 	u32 lkey = mr->elem.index << 8 | rxe_get_next_key(-1);
54 	u32 rkey = (access & IB_ACCESS_REMOTE) ? lkey : 0;
55 
56 	/* set ibmr->l/rkey and also copy into private l/rkey
57 	 * for user MRs these will always be the same
58 	 * for cases where caller 'owns' the key portion
59 	 * they may be different until REG_MR WQE is executed.
60 	 */
61 	mr->lkey = mr->ibmr.lkey = lkey;
62 	mr->rkey = mr->ibmr.rkey = rkey;
63 
64 	mr->state = RXE_MR_STATE_INVALID;
65 	mr->map_shift = ilog2(RXE_BUF_PER_MAP);
66 }
67 
68 static void rxe_mr_free_map_set(int num_map, struct rxe_map_set *set)
69 {
70 	int i;
71 
72 	for (i = 0; i < num_map; i++)
73 		kfree(set->map[i]);
74 
75 	kfree(set->map);
76 	kfree(set);
77 }
78 
79 static int rxe_mr_alloc_map_set(int num_map, struct rxe_map_set **setp)
80 {
81 	int i;
82 	struct rxe_map_set *set;
83 
84 	set = kmalloc(sizeof(*set), GFP_KERNEL);
85 	if (!set)
86 		goto err_out;
87 
88 	set->map = kmalloc_array(num_map, sizeof(struct rxe_map *), GFP_KERNEL);
89 	if (!set->map)
90 		goto err_free_set;
91 
92 	for (i = 0; i < num_map; i++) {
93 		set->map[i] = kmalloc(sizeof(struct rxe_map), GFP_KERNEL);
94 		if (!set->map[i])
95 			goto err_free_map;
96 	}
97 
98 	*setp = set;
99 
100 	return 0;
101 
102 err_free_map:
103 	for (i--; i >= 0; i--)
104 		kfree(set->map[i]);
105 
106 	kfree(set->map);
107 err_free_set:
108 	kfree(set);
109 err_out:
110 	return -ENOMEM;
111 }
112 
113 /**
114  * rxe_mr_alloc() - Allocate memory map array(s) for MR
115  * @mr: Memory region
116  * @num_buf: Number of buffer descriptors to support
117  * @both: If non zero allocate both mr->map and mr->next_map
118  *	  else just allocate mr->map. Used for fast MRs
119  *
120  * Return: 0 on success else an error
121  */
122 static int rxe_mr_alloc(struct rxe_mr *mr, int num_buf, int both)
123 {
124 	int ret;
125 	int num_map;
126 
127 	BUILD_BUG_ON(!is_power_of_2(RXE_BUF_PER_MAP));
128 	num_map = (num_buf + RXE_BUF_PER_MAP - 1) / RXE_BUF_PER_MAP;
129 
130 	mr->map_shift = ilog2(RXE_BUF_PER_MAP);
131 	mr->map_mask = RXE_BUF_PER_MAP - 1;
132 	mr->num_buf = num_buf;
133 	mr->max_buf = num_map * RXE_BUF_PER_MAP;
134 	mr->num_map = num_map;
135 
136 	ret = rxe_mr_alloc_map_set(num_map, &mr->cur_map_set);
137 	if (ret)
138 		return -ENOMEM;
139 
140 	if (both) {
141 		ret = rxe_mr_alloc_map_set(num_map, &mr->next_map_set);
142 		if (ret)
143 			goto err_free;
144 	}
145 
146 	return 0;
147 
148 err_free:
149 	rxe_mr_free_map_set(mr->num_map, mr->cur_map_set);
150 	mr->cur_map_set = NULL;
151 	return -ENOMEM;
152 }
153 
154 void rxe_mr_init_dma(struct rxe_pd *pd, int access, struct rxe_mr *mr)
155 {
156 	rxe_mr_init(access, mr);
157 
158 	mr->ibmr.pd = &pd->ibpd;
159 	mr->access = access;
160 	mr->state = RXE_MR_STATE_VALID;
161 	mr->type = IB_MR_TYPE_DMA;
162 }
163 
164 int rxe_mr_init_user(struct rxe_pd *pd, u64 start, u64 length, u64 iova,
165 		     int access, struct rxe_mr *mr)
166 {
167 	struct rxe_map_set	*set;
168 	struct rxe_map		**map;
169 	struct rxe_phys_buf	*buf = NULL;
170 	struct ib_umem		*umem;
171 	struct sg_page_iter	sg_iter;
172 	int			num_buf;
173 	void			*vaddr;
174 	int err;
175 
176 	umem = ib_umem_get(pd->ibpd.device, start, length, access);
177 	if (IS_ERR(umem)) {
178 		pr_warn("%s: Unable to pin memory region err = %d\n",
179 			__func__, (int)PTR_ERR(umem));
180 		err = PTR_ERR(umem);
181 		goto err_out;
182 	}
183 
184 	num_buf = ib_umem_num_pages(umem);
185 
186 	rxe_mr_init(access, mr);
187 
188 	err = rxe_mr_alloc(mr, num_buf, 0);
189 	if (err) {
190 		pr_warn("%s: Unable to allocate memory for map\n",
191 				__func__);
192 		goto err_release_umem;
193 	}
194 
195 	set = mr->cur_map_set;
196 	set->page_shift = PAGE_SHIFT;
197 	set->page_mask = PAGE_SIZE - 1;
198 
199 	num_buf = 0;
200 	map = set->map;
201 
202 	if (length > 0) {
203 		buf = map[0]->buf;
204 
205 		for_each_sgtable_page (&umem->sgt_append.sgt, &sg_iter, 0) {
206 			if (num_buf >= RXE_BUF_PER_MAP) {
207 				map++;
208 				buf = map[0]->buf;
209 				num_buf = 0;
210 			}
211 
212 			vaddr = page_address(sg_page_iter_page(&sg_iter));
213 			if (!vaddr) {
214 				pr_warn("%s: Unable to get virtual address\n",
215 						__func__);
216 				err = -ENOMEM;
217 				goto err_release_umem;
218 			}
219 
220 			buf->addr = (uintptr_t)vaddr;
221 			buf->size = PAGE_SIZE;
222 			num_buf++;
223 			buf++;
224 		}
225 	}
226 
227 	mr->ibmr.pd = &pd->ibpd;
228 	mr->umem = umem;
229 	mr->access = access;
230 	mr->state = RXE_MR_STATE_VALID;
231 	mr->type = IB_MR_TYPE_USER;
232 
233 	set->length = length;
234 	set->iova = iova;
235 	set->va = start;
236 	set->offset = ib_umem_offset(umem);
237 
238 	return 0;
239 
240 err_release_umem:
241 	ib_umem_release(umem);
242 err_out:
243 	return err;
244 }
245 
246 int rxe_mr_init_fast(struct rxe_pd *pd, int max_pages, struct rxe_mr *mr)
247 {
248 	int err;
249 
250 	/* always allow remote access for FMRs */
251 	rxe_mr_init(IB_ACCESS_REMOTE, mr);
252 
253 	err = rxe_mr_alloc(mr, max_pages, 1);
254 	if (err)
255 		goto err1;
256 
257 	mr->ibmr.pd = &pd->ibpd;
258 	mr->max_buf = max_pages;
259 	mr->state = RXE_MR_STATE_FREE;
260 	mr->type = IB_MR_TYPE_MEM_REG;
261 
262 	return 0;
263 
264 err1:
265 	return err;
266 }
267 
268 static void lookup_iova(struct rxe_mr *mr, u64 iova, int *m_out, int *n_out,
269 			size_t *offset_out)
270 {
271 	struct rxe_map_set *set = mr->cur_map_set;
272 	size_t offset = iova - set->iova + set->offset;
273 	int			map_index;
274 	int			buf_index;
275 	u64			length;
276 	struct rxe_map *map;
277 
278 	if (likely(set->page_shift)) {
279 		*offset_out = offset & set->page_mask;
280 		offset >>= set->page_shift;
281 		*n_out = offset & mr->map_mask;
282 		*m_out = offset >> mr->map_shift;
283 	} else {
284 		map_index = 0;
285 		buf_index = 0;
286 
287 		map = set->map[map_index];
288 		length = map->buf[buf_index].size;
289 
290 		while (offset >= length) {
291 			offset -= length;
292 			buf_index++;
293 
294 			if (buf_index == RXE_BUF_PER_MAP) {
295 				map_index++;
296 				buf_index = 0;
297 			}
298 			map = set->map[map_index];
299 			length = map->buf[buf_index].size;
300 		}
301 
302 		*m_out = map_index;
303 		*n_out = buf_index;
304 		*offset_out = offset;
305 	}
306 }
307 
308 void *iova_to_vaddr(struct rxe_mr *mr, u64 iova, int length)
309 {
310 	size_t offset;
311 	int m, n;
312 	void *addr;
313 
314 	if (mr->state != RXE_MR_STATE_VALID) {
315 		pr_warn("mr not in valid state\n");
316 		addr = NULL;
317 		goto out;
318 	}
319 
320 	if (!mr->cur_map_set) {
321 		addr = (void *)(uintptr_t)iova;
322 		goto out;
323 	}
324 
325 	if (mr_check_range(mr, iova, length)) {
326 		pr_warn("range violation\n");
327 		addr = NULL;
328 		goto out;
329 	}
330 
331 	lookup_iova(mr, iova, &m, &n, &offset);
332 
333 	if (offset + length > mr->cur_map_set->map[m]->buf[n].size) {
334 		pr_warn("crosses page boundary\n");
335 		addr = NULL;
336 		goto out;
337 	}
338 
339 	addr = (void *)(uintptr_t)mr->cur_map_set->map[m]->buf[n].addr + offset;
340 
341 out:
342 	return addr;
343 }
344 
345 /* copy data from a range (vaddr, vaddr+length-1) to or from
346  * a mr object starting at iova.
347  */
348 int rxe_mr_copy(struct rxe_mr *mr, u64 iova, void *addr, int length,
349 		enum rxe_mr_copy_dir dir)
350 {
351 	int			err;
352 	int			bytes;
353 	u8			*va;
354 	struct rxe_map		**map;
355 	struct rxe_phys_buf	*buf;
356 	int			m;
357 	int			i;
358 	size_t			offset;
359 
360 	if (length == 0)
361 		return 0;
362 
363 	if (mr->type == IB_MR_TYPE_DMA) {
364 		u8 *src, *dest;
365 
366 		src = (dir == RXE_TO_MR_OBJ) ? addr : ((void *)(uintptr_t)iova);
367 
368 		dest = (dir == RXE_TO_MR_OBJ) ? ((void *)(uintptr_t)iova) : addr;
369 
370 		memcpy(dest, src, length);
371 
372 		return 0;
373 	}
374 
375 	WARN_ON_ONCE(!mr->cur_map_set);
376 
377 	err = mr_check_range(mr, iova, length);
378 	if (err) {
379 		err = -EFAULT;
380 		goto err1;
381 	}
382 
383 	lookup_iova(mr, iova, &m, &i, &offset);
384 
385 	map = mr->cur_map_set->map + m;
386 	buf	= map[0]->buf + i;
387 
388 	while (length > 0) {
389 		u8 *src, *dest;
390 
391 		va	= (u8 *)(uintptr_t)buf->addr + offset;
392 		src = (dir == RXE_TO_MR_OBJ) ? addr : va;
393 		dest = (dir == RXE_TO_MR_OBJ) ? va : addr;
394 
395 		bytes	= buf->size - offset;
396 
397 		if (bytes > length)
398 			bytes = length;
399 
400 		memcpy(dest, src, bytes);
401 
402 		length	-= bytes;
403 		addr	+= bytes;
404 
405 		offset	= 0;
406 		buf++;
407 		i++;
408 
409 		if (i == RXE_BUF_PER_MAP) {
410 			i = 0;
411 			map++;
412 			buf = map[0]->buf;
413 		}
414 	}
415 
416 	return 0;
417 
418 err1:
419 	return err;
420 }
421 
422 /* copy data in or out of a wqe, i.e. sg list
423  * under the control of a dma descriptor
424  */
425 int copy_data(
426 	struct rxe_pd		*pd,
427 	int			access,
428 	struct rxe_dma_info	*dma,
429 	void			*addr,
430 	int			length,
431 	enum rxe_mr_copy_dir	dir)
432 {
433 	int			bytes;
434 	struct rxe_sge		*sge	= &dma->sge[dma->cur_sge];
435 	int			offset	= dma->sge_offset;
436 	int			resid	= dma->resid;
437 	struct rxe_mr		*mr	= NULL;
438 	u64			iova;
439 	int			err;
440 
441 	if (length == 0)
442 		return 0;
443 
444 	if (length > resid) {
445 		err = -EINVAL;
446 		goto err2;
447 	}
448 
449 	if (sge->length && (offset < sge->length)) {
450 		mr = lookup_mr(pd, access, sge->lkey, RXE_LOOKUP_LOCAL);
451 		if (!mr) {
452 			err = -EINVAL;
453 			goto err1;
454 		}
455 	}
456 
457 	while (length > 0) {
458 		bytes = length;
459 
460 		if (offset >= sge->length) {
461 			if (mr) {
462 				rxe_put(mr);
463 				mr = NULL;
464 			}
465 			sge++;
466 			dma->cur_sge++;
467 			offset = 0;
468 
469 			if (dma->cur_sge >= dma->num_sge) {
470 				err = -ENOSPC;
471 				goto err2;
472 			}
473 
474 			if (sge->length) {
475 				mr = lookup_mr(pd, access, sge->lkey,
476 					       RXE_LOOKUP_LOCAL);
477 				if (!mr) {
478 					err = -EINVAL;
479 					goto err1;
480 				}
481 			} else {
482 				continue;
483 			}
484 		}
485 
486 		if (bytes > sge->length - offset)
487 			bytes = sge->length - offset;
488 
489 		if (bytes > 0) {
490 			iova = sge->addr + offset;
491 
492 			err = rxe_mr_copy(mr, iova, addr, bytes, dir);
493 			if (err)
494 				goto err2;
495 
496 			offset	+= bytes;
497 			resid	-= bytes;
498 			length	-= bytes;
499 			addr	+= bytes;
500 		}
501 	}
502 
503 	dma->sge_offset = offset;
504 	dma->resid	= resid;
505 
506 	if (mr)
507 		rxe_put(mr);
508 
509 	return 0;
510 
511 err2:
512 	if (mr)
513 		rxe_put(mr);
514 err1:
515 	return err;
516 }
517 
518 int advance_dma_data(struct rxe_dma_info *dma, unsigned int length)
519 {
520 	struct rxe_sge		*sge	= &dma->sge[dma->cur_sge];
521 	int			offset	= dma->sge_offset;
522 	int			resid	= dma->resid;
523 
524 	while (length) {
525 		unsigned int bytes;
526 
527 		if (offset >= sge->length) {
528 			sge++;
529 			dma->cur_sge++;
530 			offset = 0;
531 			if (dma->cur_sge >= dma->num_sge)
532 				return -ENOSPC;
533 		}
534 
535 		bytes = length;
536 
537 		if (bytes > sge->length - offset)
538 			bytes = sge->length - offset;
539 
540 		offset	+= bytes;
541 		resid	-= bytes;
542 		length	-= bytes;
543 	}
544 
545 	dma->sge_offset = offset;
546 	dma->resid	= resid;
547 
548 	return 0;
549 }
550 
551 /* (1) find the mr corresponding to lkey/rkey
552  *     depending on lookup_type
553  * (2) verify that the (qp) pd matches the mr pd
554  * (3) verify that the mr can support the requested access
555  * (4) verify that mr state is valid
556  */
557 struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,
558 			 enum rxe_mr_lookup_type type)
559 {
560 	struct rxe_mr *mr;
561 	struct rxe_dev *rxe = to_rdev(pd->ibpd.device);
562 	int index = key >> 8;
563 
564 	mr = rxe_pool_get_index(&rxe->mr_pool, index);
565 	if (!mr)
566 		return NULL;
567 
568 	if (unlikely((type == RXE_LOOKUP_LOCAL && mr->lkey != key) ||
569 		     (type == RXE_LOOKUP_REMOTE && mr->rkey != key) ||
570 		     mr_pd(mr) != pd || (access && !(access & mr->access)) ||
571 		     mr->state != RXE_MR_STATE_VALID)) {
572 		rxe_put(mr);
573 		mr = NULL;
574 	}
575 
576 	return mr;
577 }
578 
579 int rxe_invalidate_mr(struct rxe_qp *qp, u32 rkey)
580 {
581 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
582 	struct rxe_mr *mr;
583 	int ret;
584 
585 	mr = rxe_pool_get_index(&rxe->mr_pool, rkey >> 8);
586 	if (!mr) {
587 		pr_err("%s: No MR for rkey %#x\n", __func__, rkey);
588 		ret = -EINVAL;
589 		goto err;
590 	}
591 
592 	if (rkey != mr->rkey) {
593 		pr_err("%s: rkey (%#x) doesn't match mr->rkey (%#x)\n",
594 			__func__, rkey, mr->rkey);
595 		ret = -EINVAL;
596 		goto err_drop_ref;
597 	}
598 
599 	if (atomic_read(&mr->num_mw) > 0) {
600 		pr_warn("%s: Attempt to invalidate an MR while bound to MWs\n",
601 			__func__);
602 		ret = -EINVAL;
603 		goto err_drop_ref;
604 	}
605 
606 	if (unlikely(mr->type != IB_MR_TYPE_MEM_REG)) {
607 		pr_warn("%s: mr->type (%d) is wrong type\n", __func__, mr->type);
608 		ret = -EINVAL;
609 		goto err_drop_ref;
610 	}
611 
612 	mr->state = RXE_MR_STATE_FREE;
613 	ret = 0;
614 
615 err_drop_ref:
616 	rxe_put(mr);
617 err:
618 	return ret;
619 }
620 
621 /* user can (re)register fast MR by executing a REG_MR WQE.
622  * user is expected to hold a reference on the ib mr until the
623  * WQE completes.
624  * Once a fast MR is created this is the only way to change the
625  * private keys. It is the responsibility of the user to maintain
626  * the ib mr keys in sync with rxe mr keys.
627  */
628 int rxe_reg_fast_mr(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
629 {
630 	struct rxe_mr *mr = to_rmr(wqe->wr.wr.reg.mr);
631 	u32 key = wqe->wr.wr.reg.key & 0xff;
632 	u32 access = wqe->wr.wr.reg.access;
633 	struct rxe_map_set *set;
634 
635 	/* user can only register MR in free state */
636 	if (unlikely(mr->state != RXE_MR_STATE_FREE)) {
637 		pr_warn("%s: mr->lkey = 0x%x not free\n",
638 			__func__, mr->lkey);
639 		return -EINVAL;
640 	}
641 
642 	/* user can only register mr with qp in same protection domain */
643 	if (unlikely(qp->ibqp.pd != mr->ibmr.pd)) {
644 		pr_warn("%s: qp->pd and mr->pd don't match\n",
645 			__func__);
646 		return -EINVAL;
647 	}
648 
649 	mr->access = access;
650 	mr->lkey = (mr->lkey & ~0xff) | key;
651 	mr->rkey = (access & IB_ACCESS_REMOTE) ? mr->lkey : 0;
652 	mr->state = RXE_MR_STATE_VALID;
653 
654 	set = mr->cur_map_set;
655 	mr->cur_map_set = mr->next_map_set;
656 	mr->cur_map_set->iova = wqe->wr.wr.reg.mr->iova;
657 	mr->next_map_set = set;
658 
659 	return 0;
660 }
661 
662 int rxe_mr_set_page(struct ib_mr *ibmr, u64 addr)
663 {
664 	struct rxe_mr *mr = to_rmr(ibmr);
665 	struct rxe_map_set *set = mr->next_map_set;
666 	struct rxe_map *map;
667 	struct rxe_phys_buf *buf;
668 
669 	if (unlikely(set->nbuf == mr->num_buf))
670 		return -ENOMEM;
671 
672 	map = set->map[set->nbuf / RXE_BUF_PER_MAP];
673 	buf = &map->buf[set->nbuf % RXE_BUF_PER_MAP];
674 
675 	buf->addr = addr;
676 	buf->size = ibmr->page_size;
677 	set->nbuf++;
678 
679 	return 0;
680 }
681 
682 int rxe_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
683 {
684 	struct rxe_mr *mr = to_rmr(ibmr);
685 
686 	if (atomic_read(&mr->num_mw) > 0) {
687 		pr_warn("%s: Attempt to deregister an MR while bound to MWs\n",
688 			__func__);
689 		return -EINVAL;
690 	}
691 
692 	mr->state = RXE_MR_STATE_INVALID;
693 	rxe_put(mr_pd(mr));
694 	rxe_put(mr);
695 
696 	return 0;
697 }
698 
699 void rxe_mr_cleanup(struct rxe_pool_elem *elem)
700 {
701 	struct rxe_mr *mr = container_of(elem, typeof(*mr), elem);
702 
703 	ib_umem_release(mr->umem);
704 
705 	if (mr->cur_map_set)
706 		rxe_mr_free_map_set(mr->num_map, mr->cur_map_set);
707 
708 	if (mr->next_map_set)
709 		rxe_mr_free_map_set(mr->num_map, mr->next_map_set);
710 }
711