xref: /openbmc/linux/drivers/infiniband/sw/rxe/rxe_mr.c (revision dc6a81c3)
1 /*
2  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
3  * Copyright (c) 2015 System Fabric Works, Inc. 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 "rxe.h"
35 #include "rxe_loc.h"
36 
37 /*
38  * lfsr (linear feedback shift register) with period 255
39  */
40 static u8 rxe_get_key(void)
41 {
42 	static u32 key = 1;
43 
44 	key = key << 1;
45 
46 	key |= (0 != (key & 0x100)) ^ (0 != (key & 0x10))
47 		^ (0 != (key & 0x80)) ^ (0 != (key & 0x40));
48 
49 	key &= 0xff;
50 
51 	return key;
52 }
53 
54 int mem_check_range(struct rxe_mem *mem, u64 iova, size_t length)
55 {
56 	switch (mem->type) {
57 	case RXE_MEM_TYPE_DMA:
58 		return 0;
59 
60 	case RXE_MEM_TYPE_MR:
61 	case RXE_MEM_TYPE_FMR:
62 		if (iova < mem->iova ||
63 		    length > mem->length ||
64 		    iova > mem->iova + mem->length - length)
65 			return -EFAULT;
66 		return 0;
67 
68 	default:
69 		return -EFAULT;
70 	}
71 }
72 
73 #define IB_ACCESS_REMOTE	(IB_ACCESS_REMOTE_READ		\
74 				| IB_ACCESS_REMOTE_WRITE	\
75 				| IB_ACCESS_REMOTE_ATOMIC)
76 
77 static void rxe_mem_init(int access, struct rxe_mem *mem)
78 {
79 	u32 lkey = mem->pelem.index << 8 | rxe_get_key();
80 	u32 rkey = (access & IB_ACCESS_REMOTE) ? lkey : 0;
81 
82 	if (mem->pelem.pool->type == RXE_TYPE_MR) {
83 		mem->ibmr.lkey		= lkey;
84 		mem->ibmr.rkey		= rkey;
85 	}
86 
87 	mem->lkey		= lkey;
88 	mem->rkey		= rkey;
89 	mem->state		= RXE_MEM_STATE_INVALID;
90 	mem->type		= RXE_MEM_TYPE_NONE;
91 	mem->map_shift		= ilog2(RXE_BUF_PER_MAP);
92 }
93 
94 void rxe_mem_cleanup(struct rxe_pool_entry *arg)
95 {
96 	struct rxe_mem *mem = container_of(arg, typeof(*mem), pelem);
97 	int i;
98 
99 	ib_umem_release(mem->umem);
100 
101 	if (mem->map) {
102 		for (i = 0; i < mem->num_map; i++)
103 			kfree(mem->map[i]);
104 
105 		kfree(mem->map);
106 	}
107 }
108 
109 static int rxe_mem_alloc(struct rxe_mem *mem, int num_buf)
110 {
111 	int i;
112 	int num_map;
113 	struct rxe_map **map = mem->map;
114 
115 	num_map = (num_buf + RXE_BUF_PER_MAP - 1) / RXE_BUF_PER_MAP;
116 
117 	mem->map = kmalloc_array(num_map, sizeof(*map), GFP_KERNEL);
118 	if (!mem->map)
119 		goto err1;
120 
121 	for (i = 0; i < num_map; i++) {
122 		mem->map[i] = kmalloc(sizeof(**map), GFP_KERNEL);
123 		if (!mem->map[i])
124 			goto err2;
125 	}
126 
127 	BUILD_BUG_ON(!is_power_of_2(RXE_BUF_PER_MAP));
128 
129 	mem->map_shift	= ilog2(RXE_BUF_PER_MAP);
130 	mem->map_mask	= RXE_BUF_PER_MAP - 1;
131 
132 	mem->num_buf = num_buf;
133 	mem->num_map = num_map;
134 	mem->max_buf = num_map * RXE_BUF_PER_MAP;
135 
136 	return 0;
137 
138 err2:
139 	for (i--; i >= 0; i--)
140 		kfree(mem->map[i]);
141 
142 	kfree(mem->map);
143 err1:
144 	return -ENOMEM;
145 }
146 
147 int rxe_mem_init_dma(struct rxe_pd *pd,
148 		     int access, struct rxe_mem *mem)
149 {
150 	rxe_mem_init(access, mem);
151 
152 	mem->pd			= pd;
153 	mem->access		= access;
154 	mem->state		= RXE_MEM_STATE_VALID;
155 	mem->type		= RXE_MEM_TYPE_DMA;
156 
157 	return 0;
158 }
159 
160 int rxe_mem_init_user(struct rxe_pd *pd, u64 start,
161 		      u64 length, u64 iova, int access, struct ib_udata *udata,
162 		      struct rxe_mem *mem)
163 {
164 	struct rxe_map		**map;
165 	struct rxe_phys_buf	*buf = NULL;
166 	struct ib_umem		*umem;
167 	struct sg_page_iter	sg_iter;
168 	int			num_buf;
169 	void			*vaddr;
170 	int err;
171 
172 	umem = ib_umem_get(pd->ibpd.device, start, length, access);
173 	if (IS_ERR(umem)) {
174 		pr_warn("err %d from rxe_umem_get\n",
175 			(int)PTR_ERR(umem));
176 		err = -EINVAL;
177 		goto err1;
178 	}
179 
180 	mem->umem = umem;
181 	num_buf = ib_umem_num_pages(umem);
182 
183 	rxe_mem_init(access, mem);
184 
185 	err = rxe_mem_alloc(mem, num_buf);
186 	if (err) {
187 		pr_warn("err %d from rxe_mem_alloc\n", err);
188 		ib_umem_release(umem);
189 		goto err1;
190 	}
191 
192 	mem->page_shift		= PAGE_SHIFT;
193 	mem->page_mask = PAGE_SIZE - 1;
194 
195 	num_buf			= 0;
196 	map			= mem->map;
197 	if (length > 0) {
198 		buf = map[0]->buf;
199 
200 		for_each_sg_page(umem->sg_head.sgl, &sg_iter, umem->nmap, 0) {
201 			if (num_buf >= RXE_BUF_PER_MAP) {
202 				map++;
203 				buf = map[0]->buf;
204 				num_buf = 0;
205 			}
206 
207 			vaddr = page_address(sg_page_iter_page(&sg_iter));
208 			if (!vaddr) {
209 				pr_warn("null vaddr\n");
210 				err = -ENOMEM;
211 				goto err1;
212 			}
213 
214 			buf->addr = (uintptr_t)vaddr;
215 			buf->size = PAGE_SIZE;
216 			num_buf++;
217 			buf++;
218 
219 		}
220 	}
221 
222 	mem->pd			= pd;
223 	mem->umem		= umem;
224 	mem->access		= access;
225 	mem->length		= length;
226 	mem->iova		= iova;
227 	mem->va			= start;
228 	mem->offset		= ib_umem_offset(umem);
229 	mem->state		= RXE_MEM_STATE_VALID;
230 	mem->type		= RXE_MEM_TYPE_MR;
231 
232 	return 0;
233 
234 err1:
235 	return err;
236 }
237 
238 int rxe_mem_init_fast(struct rxe_pd *pd,
239 		      int max_pages, struct rxe_mem *mem)
240 {
241 	int err;
242 
243 	rxe_mem_init(0, mem);
244 
245 	/* In fastreg, we also set the rkey */
246 	mem->ibmr.rkey = mem->ibmr.lkey;
247 
248 	err = rxe_mem_alloc(mem, max_pages);
249 	if (err)
250 		goto err1;
251 
252 	mem->pd			= pd;
253 	mem->max_buf		= max_pages;
254 	mem->state		= RXE_MEM_STATE_FREE;
255 	mem->type		= RXE_MEM_TYPE_MR;
256 
257 	return 0;
258 
259 err1:
260 	return err;
261 }
262 
263 static void lookup_iova(
264 	struct rxe_mem	*mem,
265 	u64			iova,
266 	int			*m_out,
267 	int			*n_out,
268 	size_t			*offset_out)
269 {
270 	size_t			offset = iova - mem->iova + mem->offset;
271 	int			map_index;
272 	int			buf_index;
273 	u64			length;
274 
275 	if (likely(mem->page_shift)) {
276 		*offset_out = offset & mem->page_mask;
277 		offset >>= mem->page_shift;
278 		*n_out = offset & mem->map_mask;
279 		*m_out = offset >> mem->map_shift;
280 	} else {
281 		map_index = 0;
282 		buf_index = 0;
283 
284 		length = mem->map[map_index]->buf[buf_index].size;
285 
286 		while (offset >= length) {
287 			offset -= length;
288 			buf_index++;
289 
290 			if (buf_index == RXE_BUF_PER_MAP) {
291 				map_index++;
292 				buf_index = 0;
293 			}
294 			length = mem->map[map_index]->buf[buf_index].size;
295 		}
296 
297 		*m_out = map_index;
298 		*n_out = buf_index;
299 		*offset_out = offset;
300 	}
301 }
302 
303 void *iova_to_vaddr(struct rxe_mem *mem, u64 iova, int length)
304 {
305 	size_t offset;
306 	int m, n;
307 	void *addr;
308 
309 	if (mem->state != RXE_MEM_STATE_VALID) {
310 		pr_warn("mem not in valid state\n");
311 		addr = NULL;
312 		goto out;
313 	}
314 
315 	if (!mem->map) {
316 		addr = (void *)(uintptr_t)iova;
317 		goto out;
318 	}
319 
320 	if (mem_check_range(mem, iova, length)) {
321 		pr_warn("range violation\n");
322 		addr = NULL;
323 		goto out;
324 	}
325 
326 	lookup_iova(mem, iova, &m, &n, &offset);
327 
328 	if (offset + length > mem->map[m]->buf[n].size) {
329 		pr_warn("crosses page boundary\n");
330 		addr = NULL;
331 		goto out;
332 	}
333 
334 	addr = (void *)(uintptr_t)mem->map[m]->buf[n].addr + offset;
335 
336 out:
337 	return addr;
338 }
339 
340 /* copy data from a range (vaddr, vaddr+length-1) to or from
341  * a mem object starting at iova. Compute incremental value of
342  * crc32 if crcp is not zero. caller must hold a reference to mem
343  */
344 int rxe_mem_copy(struct rxe_mem *mem, u64 iova, void *addr, int length,
345 		 enum copy_direction dir, u32 *crcp)
346 {
347 	int			err;
348 	int			bytes;
349 	u8			*va;
350 	struct rxe_map		**map;
351 	struct rxe_phys_buf	*buf;
352 	int			m;
353 	int			i;
354 	size_t			offset;
355 	u32			crc = crcp ? (*crcp) : 0;
356 
357 	if (length == 0)
358 		return 0;
359 
360 	if (mem->type == RXE_MEM_TYPE_DMA) {
361 		u8 *src, *dest;
362 
363 		src  = (dir == to_mem_obj) ?
364 			addr : ((void *)(uintptr_t)iova);
365 
366 		dest = (dir == to_mem_obj) ?
367 			((void *)(uintptr_t)iova) : addr;
368 
369 		memcpy(dest, src, length);
370 
371 		if (crcp)
372 			*crcp = rxe_crc32(to_rdev(mem->pd->ibpd.device),
373 					*crcp, dest, length);
374 
375 		return 0;
376 	}
377 
378 	WARN_ON_ONCE(!mem->map);
379 
380 	err = mem_check_range(mem, iova, length);
381 	if (err) {
382 		err = -EFAULT;
383 		goto err1;
384 	}
385 
386 	lookup_iova(mem, iova, &m, &i, &offset);
387 
388 	map	= mem->map + m;
389 	buf	= map[0]->buf + i;
390 
391 	while (length > 0) {
392 		u8 *src, *dest;
393 
394 		va	= (u8 *)(uintptr_t)buf->addr + offset;
395 		src  = (dir == to_mem_obj) ? addr : va;
396 		dest = (dir == to_mem_obj) ? va : addr;
397 
398 		bytes	= buf->size - offset;
399 
400 		if (bytes > length)
401 			bytes = length;
402 
403 		memcpy(dest, src, bytes);
404 
405 		if (crcp)
406 			crc = rxe_crc32(to_rdev(mem->pd->ibpd.device),
407 					crc, dest, bytes);
408 
409 		length	-= bytes;
410 		addr	+= bytes;
411 
412 		offset	= 0;
413 		buf++;
414 		i++;
415 
416 		if (i == RXE_BUF_PER_MAP) {
417 			i = 0;
418 			map++;
419 			buf = map[0]->buf;
420 		}
421 	}
422 
423 	if (crcp)
424 		*crcp = crc;
425 
426 	return 0;
427 
428 err1:
429 	return err;
430 }
431 
432 /* copy data in or out of a wqe, i.e. sg list
433  * under the control of a dma descriptor
434  */
435 int copy_data(
436 	struct rxe_pd		*pd,
437 	int			access,
438 	struct rxe_dma_info	*dma,
439 	void			*addr,
440 	int			length,
441 	enum copy_direction	dir,
442 	u32			*crcp)
443 {
444 	int			bytes;
445 	struct rxe_sge		*sge	= &dma->sge[dma->cur_sge];
446 	int			offset	= dma->sge_offset;
447 	int			resid	= dma->resid;
448 	struct rxe_mem		*mem	= NULL;
449 	u64			iova;
450 	int			err;
451 
452 	if (length == 0)
453 		return 0;
454 
455 	if (length > resid) {
456 		err = -EINVAL;
457 		goto err2;
458 	}
459 
460 	if (sge->length && (offset < sge->length)) {
461 		mem = lookup_mem(pd, access, sge->lkey, lookup_local);
462 		if (!mem) {
463 			err = -EINVAL;
464 			goto err1;
465 		}
466 	}
467 
468 	while (length > 0) {
469 		bytes = length;
470 
471 		if (offset >= sge->length) {
472 			if (mem) {
473 				rxe_drop_ref(mem);
474 				mem = NULL;
475 			}
476 			sge++;
477 			dma->cur_sge++;
478 			offset = 0;
479 
480 			if (dma->cur_sge >= dma->num_sge) {
481 				err = -ENOSPC;
482 				goto err2;
483 			}
484 
485 			if (sge->length) {
486 				mem = lookup_mem(pd, access, sge->lkey,
487 						 lookup_local);
488 				if (!mem) {
489 					err = -EINVAL;
490 					goto err1;
491 				}
492 			} else {
493 				continue;
494 			}
495 		}
496 
497 		if (bytes > sge->length - offset)
498 			bytes = sge->length - offset;
499 
500 		if (bytes > 0) {
501 			iova = sge->addr + offset;
502 
503 			err = rxe_mem_copy(mem, iova, addr, bytes, dir, crcp);
504 			if (err)
505 				goto err2;
506 
507 			offset	+= bytes;
508 			resid	-= bytes;
509 			length	-= bytes;
510 			addr	+= bytes;
511 		}
512 	}
513 
514 	dma->sge_offset = offset;
515 	dma->resid	= resid;
516 
517 	if (mem)
518 		rxe_drop_ref(mem);
519 
520 	return 0;
521 
522 err2:
523 	if (mem)
524 		rxe_drop_ref(mem);
525 err1:
526 	return err;
527 }
528 
529 int advance_dma_data(struct rxe_dma_info *dma, unsigned int length)
530 {
531 	struct rxe_sge		*sge	= &dma->sge[dma->cur_sge];
532 	int			offset	= dma->sge_offset;
533 	int			resid	= dma->resid;
534 
535 	while (length) {
536 		unsigned int bytes;
537 
538 		if (offset >= sge->length) {
539 			sge++;
540 			dma->cur_sge++;
541 			offset = 0;
542 			if (dma->cur_sge >= dma->num_sge)
543 				return -ENOSPC;
544 		}
545 
546 		bytes = length;
547 
548 		if (bytes > sge->length - offset)
549 			bytes = sge->length - offset;
550 
551 		offset	+= bytes;
552 		resid	-= bytes;
553 		length	-= bytes;
554 	}
555 
556 	dma->sge_offset = offset;
557 	dma->resid	= resid;
558 
559 	return 0;
560 }
561 
562 /* (1) find the mem (mr or mw) corresponding to lkey/rkey
563  *     depending on lookup_type
564  * (2) verify that the (qp) pd matches the mem pd
565  * (3) verify that the mem can support the requested access
566  * (4) verify that mem state is valid
567  */
568 struct rxe_mem *lookup_mem(struct rxe_pd *pd, int access, u32 key,
569 			   enum lookup_type type)
570 {
571 	struct rxe_mem *mem;
572 	struct rxe_dev *rxe = to_rdev(pd->ibpd.device);
573 	int index = key >> 8;
574 
575 	mem = rxe_pool_get_index(&rxe->mr_pool, index);
576 	if (!mem)
577 		return NULL;
578 
579 	if (unlikely((type == lookup_local && mem->lkey != key) ||
580 		     (type == lookup_remote && mem->rkey != key) ||
581 		     mem->pd != pd ||
582 		     (access && !(access & mem->access)) ||
583 		     mem->state != RXE_MEM_STATE_VALID)) {
584 		rxe_drop_ref(mem);
585 		mem = NULL;
586 	}
587 
588 	return mem;
589 }
590 
591 int rxe_mem_map_pages(struct rxe_dev *rxe, struct rxe_mem *mem,
592 		      u64 *page, int num_pages, u64 iova)
593 {
594 	int i;
595 	int num_buf;
596 	int err;
597 	struct rxe_map **map;
598 	struct rxe_phys_buf *buf;
599 	int page_size;
600 
601 	if (num_pages > mem->max_buf) {
602 		err = -EINVAL;
603 		goto err1;
604 	}
605 
606 	num_buf		= 0;
607 	page_size	= 1 << mem->page_shift;
608 	map		= mem->map;
609 	buf		= map[0]->buf;
610 
611 	for (i = 0; i < num_pages; i++) {
612 		buf->addr = *page++;
613 		buf->size = page_size;
614 		buf++;
615 		num_buf++;
616 
617 		if (num_buf == RXE_BUF_PER_MAP) {
618 			map++;
619 			buf = map[0]->buf;
620 			num_buf = 0;
621 		}
622 	}
623 
624 	mem->iova	= iova;
625 	mem->va		= iova;
626 	mem->length	= num_pages << mem->page_shift;
627 	mem->state	= RXE_MEM_STATE_VALID;
628 
629 	return 0;
630 
631 err1:
632 	return err;
633 }
634