xref: /openbmc/linux/drivers/infiniband/sw/rxe/rxe_mr.c (revision 36d26912)
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 void 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 
158 int rxe_mem_init_user(struct rxe_pd *pd, u64 start,
159 		      u64 length, u64 iova, int access, struct ib_udata *udata,
160 		      struct rxe_mem *mem)
161 {
162 	struct rxe_map		**map;
163 	struct rxe_phys_buf	*buf = NULL;
164 	struct ib_umem		*umem;
165 	struct sg_page_iter	sg_iter;
166 	int			num_buf;
167 	void			*vaddr;
168 	int err;
169 
170 	umem = ib_umem_get(pd->ibpd.device, start, length, access);
171 	if (IS_ERR(umem)) {
172 		pr_warn("err %d from rxe_umem_get\n",
173 			(int)PTR_ERR(umem));
174 		err = -EINVAL;
175 		goto err1;
176 	}
177 
178 	mem->umem = umem;
179 	num_buf = ib_umem_num_pages(umem);
180 
181 	rxe_mem_init(access, mem);
182 
183 	err = rxe_mem_alloc(mem, num_buf);
184 	if (err) {
185 		pr_warn("err %d from rxe_mem_alloc\n", err);
186 		ib_umem_release(umem);
187 		goto err1;
188 	}
189 
190 	mem->page_shift		= PAGE_SHIFT;
191 	mem->page_mask = PAGE_SIZE - 1;
192 
193 	num_buf			= 0;
194 	map			= mem->map;
195 	if (length > 0) {
196 		buf = map[0]->buf;
197 
198 		for_each_sg_page(umem->sg_head.sgl, &sg_iter, umem->nmap, 0) {
199 			if (num_buf >= RXE_BUF_PER_MAP) {
200 				map++;
201 				buf = map[0]->buf;
202 				num_buf = 0;
203 			}
204 
205 			vaddr = page_address(sg_page_iter_page(&sg_iter));
206 			if (!vaddr) {
207 				pr_warn("null vaddr\n");
208 				ib_umem_release(umem);
209 				err = -ENOMEM;
210 				goto err1;
211 			}
212 
213 			buf->addr = (uintptr_t)vaddr;
214 			buf->size = PAGE_SIZE;
215 			num_buf++;
216 			buf++;
217 
218 		}
219 	}
220 
221 	mem->pd			= pd;
222 	mem->umem		= umem;
223 	mem->access		= access;
224 	mem->length		= length;
225 	mem->iova		= iova;
226 	mem->va			= start;
227 	mem->offset		= ib_umem_offset(umem);
228 	mem->state		= RXE_MEM_STATE_VALID;
229 	mem->type		= RXE_MEM_TYPE_MR;
230 
231 	return 0;
232 
233 err1:
234 	return err;
235 }
236 
237 int rxe_mem_init_fast(struct rxe_pd *pd,
238 		      int max_pages, struct rxe_mem *mem)
239 {
240 	int err;
241 
242 	rxe_mem_init(0, mem);
243 
244 	/* In fastreg, we also set the rkey */
245 	mem->ibmr.rkey = mem->ibmr.lkey;
246 
247 	err = rxe_mem_alloc(mem, max_pages);
248 	if (err)
249 		goto err1;
250 
251 	mem->pd			= pd;
252 	mem->max_buf		= max_pages;
253 	mem->state		= RXE_MEM_STATE_FREE;
254 	mem->type		= RXE_MEM_TYPE_MR;
255 
256 	return 0;
257 
258 err1:
259 	return err;
260 }
261 
262 static void lookup_iova(
263 	struct rxe_mem	*mem,
264 	u64			iova,
265 	int			*m_out,
266 	int			*n_out,
267 	size_t			*offset_out)
268 {
269 	size_t			offset = iova - mem->iova + mem->offset;
270 	int			map_index;
271 	int			buf_index;
272 	u64			length;
273 
274 	if (likely(mem->page_shift)) {
275 		*offset_out = offset & mem->page_mask;
276 		offset >>= mem->page_shift;
277 		*n_out = offset & mem->map_mask;
278 		*m_out = offset >> mem->map_shift;
279 	} else {
280 		map_index = 0;
281 		buf_index = 0;
282 
283 		length = mem->map[map_index]->buf[buf_index].size;
284 
285 		while (offset >= length) {
286 			offset -= length;
287 			buf_index++;
288 
289 			if (buf_index == RXE_BUF_PER_MAP) {
290 				map_index++;
291 				buf_index = 0;
292 			}
293 			length = mem->map[map_index]->buf[buf_index].size;
294 		}
295 
296 		*m_out = map_index;
297 		*n_out = buf_index;
298 		*offset_out = offset;
299 	}
300 }
301 
302 void *iova_to_vaddr(struct rxe_mem *mem, u64 iova, int length)
303 {
304 	size_t offset;
305 	int m, n;
306 	void *addr;
307 
308 	if (mem->state != RXE_MEM_STATE_VALID) {
309 		pr_warn("mem not in valid state\n");
310 		addr = NULL;
311 		goto out;
312 	}
313 
314 	if (!mem->map) {
315 		addr = (void *)(uintptr_t)iova;
316 		goto out;
317 	}
318 
319 	if (mem_check_range(mem, iova, length)) {
320 		pr_warn("range violation\n");
321 		addr = NULL;
322 		goto out;
323 	}
324 
325 	lookup_iova(mem, iova, &m, &n, &offset);
326 
327 	if (offset + length > mem->map[m]->buf[n].size) {
328 		pr_warn("crosses page boundary\n");
329 		addr = NULL;
330 		goto out;
331 	}
332 
333 	addr = (void *)(uintptr_t)mem->map[m]->buf[n].addr + offset;
334 
335 out:
336 	return addr;
337 }
338 
339 /* copy data from a range (vaddr, vaddr+length-1) to or from
340  * a mem object starting at iova. Compute incremental value of
341  * crc32 if crcp is not zero. caller must hold a reference to mem
342  */
343 int rxe_mem_copy(struct rxe_mem *mem, u64 iova, void *addr, int length,
344 		 enum copy_direction dir, u32 *crcp)
345 {
346 	int			err;
347 	int			bytes;
348 	u8			*va;
349 	struct rxe_map		**map;
350 	struct rxe_phys_buf	*buf;
351 	int			m;
352 	int			i;
353 	size_t			offset;
354 	u32			crc = crcp ? (*crcp) : 0;
355 
356 	if (length == 0)
357 		return 0;
358 
359 	if (mem->type == RXE_MEM_TYPE_DMA) {
360 		u8 *src, *dest;
361 
362 		src  = (dir == to_mem_obj) ?
363 			addr : ((void *)(uintptr_t)iova);
364 
365 		dest = (dir == to_mem_obj) ?
366 			((void *)(uintptr_t)iova) : addr;
367 
368 		memcpy(dest, src, length);
369 
370 		if (crcp)
371 			*crcp = rxe_crc32(to_rdev(mem->pd->ibpd.device),
372 					*crcp, dest, length);
373 
374 		return 0;
375 	}
376 
377 	WARN_ON_ONCE(!mem->map);
378 
379 	err = mem_check_range(mem, iova, length);
380 	if (err) {
381 		err = -EFAULT;
382 		goto err1;
383 	}
384 
385 	lookup_iova(mem, iova, &m, &i, &offset);
386 
387 	map	= mem->map + m;
388 	buf	= map[0]->buf + i;
389 
390 	while (length > 0) {
391 		u8 *src, *dest;
392 
393 		va	= (u8 *)(uintptr_t)buf->addr + offset;
394 		src  = (dir == to_mem_obj) ? addr : va;
395 		dest = (dir == to_mem_obj) ? va : addr;
396 
397 		bytes	= buf->size - offset;
398 
399 		if (bytes > length)
400 			bytes = length;
401 
402 		memcpy(dest, src, bytes);
403 
404 		if (crcp)
405 			crc = rxe_crc32(to_rdev(mem->pd->ibpd.device),
406 					crc, dest, bytes);
407 
408 		length	-= bytes;
409 		addr	+= bytes;
410 
411 		offset	= 0;
412 		buf++;
413 		i++;
414 
415 		if (i == RXE_BUF_PER_MAP) {
416 			i = 0;
417 			map++;
418 			buf = map[0]->buf;
419 		}
420 	}
421 
422 	if (crcp)
423 		*crcp = crc;
424 
425 	return 0;
426 
427 err1:
428 	return err;
429 }
430 
431 /* copy data in or out of a wqe, i.e. sg list
432  * under the control of a dma descriptor
433  */
434 int copy_data(
435 	struct rxe_pd		*pd,
436 	int			access,
437 	struct rxe_dma_info	*dma,
438 	void			*addr,
439 	int			length,
440 	enum copy_direction	dir,
441 	u32			*crcp)
442 {
443 	int			bytes;
444 	struct rxe_sge		*sge	= &dma->sge[dma->cur_sge];
445 	int			offset	= dma->sge_offset;
446 	int			resid	= dma->resid;
447 	struct rxe_mem		*mem	= NULL;
448 	u64			iova;
449 	int			err;
450 
451 	if (length == 0)
452 		return 0;
453 
454 	if (length > resid) {
455 		err = -EINVAL;
456 		goto err2;
457 	}
458 
459 	if (sge->length && (offset < sge->length)) {
460 		mem = lookup_mem(pd, access, sge->lkey, lookup_local);
461 		if (!mem) {
462 			err = -EINVAL;
463 			goto err1;
464 		}
465 	}
466 
467 	while (length > 0) {
468 		bytes = length;
469 
470 		if (offset >= sge->length) {
471 			if (mem) {
472 				rxe_drop_ref(mem);
473 				mem = NULL;
474 			}
475 			sge++;
476 			dma->cur_sge++;
477 			offset = 0;
478 
479 			if (dma->cur_sge >= dma->num_sge) {
480 				err = -ENOSPC;
481 				goto err2;
482 			}
483 
484 			if (sge->length) {
485 				mem = lookup_mem(pd, access, sge->lkey,
486 						 lookup_local);
487 				if (!mem) {
488 					err = -EINVAL;
489 					goto err1;
490 				}
491 			} else {
492 				continue;
493 			}
494 		}
495 
496 		if (bytes > sge->length - offset)
497 			bytes = sge->length - offset;
498 
499 		if (bytes > 0) {
500 			iova = sge->addr + offset;
501 
502 			err = rxe_mem_copy(mem, iova, addr, bytes, dir, crcp);
503 			if (err)
504 				goto err2;
505 
506 			offset	+= bytes;
507 			resid	-= bytes;
508 			length	-= bytes;
509 			addr	+= bytes;
510 		}
511 	}
512 
513 	dma->sge_offset = offset;
514 	dma->resid	= resid;
515 
516 	if (mem)
517 		rxe_drop_ref(mem);
518 
519 	return 0;
520 
521 err2:
522 	if (mem)
523 		rxe_drop_ref(mem);
524 err1:
525 	return err;
526 }
527 
528 int advance_dma_data(struct rxe_dma_info *dma, unsigned int length)
529 {
530 	struct rxe_sge		*sge	= &dma->sge[dma->cur_sge];
531 	int			offset	= dma->sge_offset;
532 	int			resid	= dma->resid;
533 
534 	while (length) {
535 		unsigned int bytes;
536 
537 		if (offset >= sge->length) {
538 			sge++;
539 			dma->cur_sge++;
540 			offset = 0;
541 			if (dma->cur_sge >= dma->num_sge)
542 				return -ENOSPC;
543 		}
544 
545 		bytes = length;
546 
547 		if (bytes > sge->length - offset)
548 			bytes = sge->length - offset;
549 
550 		offset	+= bytes;
551 		resid	-= bytes;
552 		length	-= bytes;
553 	}
554 
555 	dma->sge_offset = offset;
556 	dma->resid	= resid;
557 
558 	return 0;
559 }
560 
561 /* (1) find the mem (mr or mw) corresponding to lkey/rkey
562  *     depending on lookup_type
563  * (2) verify that the (qp) pd matches the mem pd
564  * (3) verify that the mem can support the requested access
565  * (4) verify that mem state is valid
566  */
567 struct rxe_mem *lookup_mem(struct rxe_pd *pd, int access, u32 key,
568 			   enum lookup_type type)
569 {
570 	struct rxe_mem *mem;
571 	struct rxe_dev *rxe = to_rdev(pd->ibpd.device);
572 	int index = key >> 8;
573 
574 	mem = rxe_pool_get_index(&rxe->mr_pool, index);
575 	if (!mem)
576 		return NULL;
577 
578 	if (unlikely((type == lookup_local && mem->lkey != key) ||
579 		     (type == lookup_remote && mem->rkey != key) ||
580 		     mem->pd != pd ||
581 		     (access && !(access & mem->access)) ||
582 		     mem->state != RXE_MEM_STATE_VALID)) {
583 		rxe_drop_ref(mem);
584 		mem = NULL;
585 	}
586 
587 	return mem;
588 }
589