xref: /openbmc/linux/drivers/infiniband/sw/rxe/rxe_mr.c (revision 13dd8710)
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 /*
11  * lfsr (linear feedback shift register) with period 255
12  */
13 static u8 rxe_get_key(void)
14 {
15 	static u32 key = 1;
16 
17 	key = key << 1;
18 
19 	key |= (0 != (key & 0x100)) ^ (0 != (key & 0x10))
20 		^ (0 != (key & 0x80)) ^ (0 != (key & 0x40));
21 
22 	key &= 0xff;
23 
24 	return key;
25 }
26 
27 int mem_check_range(struct rxe_mem *mem, u64 iova, size_t length)
28 {
29 	switch (mem->type) {
30 	case RXE_MEM_TYPE_DMA:
31 		return 0;
32 
33 	case RXE_MEM_TYPE_MR:
34 	case RXE_MEM_TYPE_FMR:
35 		if (iova < mem->iova ||
36 		    length > mem->length ||
37 		    iova > mem->iova + mem->length - length)
38 			return -EFAULT;
39 		return 0;
40 
41 	default:
42 		return -EFAULT;
43 	}
44 }
45 
46 #define IB_ACCESS_REMOTE	(IB_ACCESS_REMOTE_READ		\
47 				| IB_ACCESS_REMOTE_WRITE	\
48 				| IB_ACCESS_REMOTE_ATOMIC)
49 
50 static void rxe_mem_init(int access, struct rxe_mem *mem)
51 {
52 	u32 lkey = mem->pelem.index << 8 | rxe_get_key();
53 	u32 rkey = (access & IB_ACCESS_REMOTE) ? lkey : 0;
54 
55 	mem->ibmr.lkey		= lkey;
56 	mem->ibmr.rkey		= rkey;
57 	mem->state		= RXE_MEM_STATE_INVALID;
58 	mem->type		= RXE_MEM_TYPE_NONE;
59 	mem->map_shift		= ilog2(RXE_BUF_PER_MAP);
60 }
61 
62 void rxe_mem_cleanup(struct rxe_pool_entry *arg)
63 {
64 	struct rxe_mem *mem = container_of(arg, typeof(*mem), pelem);
65 	int i;
66 
67 	ib_umem_release(mem->umem);
68 
69 	if (mem->map) {
70 		for (i = 0; i < mem->num_map; i++)
71 			kfree(mem->map[i]);
72 
73 		kfree(mem->map);
74 	}
75 }
76 
77 static int rxe_mem_alloc(struct rxe_mem *mem, int num_buf)
78 {
79 	int i;
80 	int num_map;
81 	struct rxe_map **map = mem->map;
82 
83 	num_map = (num_buf + RXE_BUF_PER_MAP - 1) / RXE_BUF_PER_MAP;
84 
85 	mem->map = kmalloc_array(num_map, sizeof(*map), GFP_KERNEL);
86 	if (!mem->map)
87 		goto err1;
88 
89 	for (i = 0; i < num_map; i++) {
90 		mem->map[i] = kmalloc(sizeof(**map), GFP_KERNEL);
91 		if (!mem->map[i])
92 			goto err2;
93 	}
94 
95 	BUILD_BUG_ON(!is_power_of_2(RXE_BUF_PER_MAP));
96 
97 	mem->map_shift	= ilog2(RXE_BUF_PER_MAP);
98 	mem->map_mask	= RXE_BUF_PER_MAP - 1;
99 
100 	mem->num_buf = num_buf;
101 	mem->num_map = num_map;
102 	mem->max_buf = num_map * RXE_BUF_PER_MAP;
103 
104 	return 0;
105 
106 err2:
107 	for (i--; i >= 0; i--)
108 		kfree(mem->map[i]);
109 
110 	kfree(mem->map);
111 err1:
112 	return -ENOMEM;
113 }
114 
115 void rxe_mem_init_dma(struct rxe_pd *pd,
116 		      int access, struct rxe_mem *mem)
117 {
118 	rxe_mem_init(access, mem);
119 
120 	mem->ibmr.pd		= &pd->ibpd;
121 	mem->access		= access;
122 	mem->state		= RXE_MEM_STATE_VALID;
123 	mem->type		= RXE_MEM_TYPE_DMA;
124 }
125 
126 int rxe_mem_init_user(struct rxe_pd *pd, u64 start,
127 		      u64 length, u64 iova, int access, struct ib_udata *udata,
128 		      struct rxe_mem *mem)
129 {
130 	struct rxe_map		**map;
131 	struct rxe_phys_buf	*buf = NULL;
132 	struct ib_umem		*umem;
133 	struct sg_page_iter	sg_iter;
134 	int			num_buf;
135 	void			*vaddr;
136 	int err;
137 
138 	umem = ib_umem_get(pd->ibpd.device, start, length, access);
139 	if (IS_ERR(umem)) {
140 		pr_warn("err %d from rxe_umem_get\n",
141 			(int)PTR_ERR(umem));
142 		err = -EINVAL;
143 		goto err1;
144 	}
145 
146 	mem->umem = umem;
147 	num_buf = ib_umem_num_pages(umem);
148 
149 	rxe_mem_init(access, mem);
150 
151 	err = rxe_mem_alloc(mem, num_buf);
152 	if (err) {
153 		pr_warn("err %d from rxe_mem_alloc\n", err);
154 		ib_umem_release(umem);
155 		goto err1;
156 	}
157 
158 	mem->page_shift		= PAGE_SHIFT;
159 	mem->page_mask = PAGE_SIZE - 1;
160 
161 	num_buf			= 0;
162 	map			= mem->map;
163 	if (length > 0) {
164 		buf = map[0]->buf;
165 
166 		for_each_sg_page(umem->sg_head.sgl, &sg_iter, umem->nmap, 0) {
167 			if (num_buf >= RXE_BUF_PER_MAP) {
168 				map++;
169 				buf = map[0]->buf;
170 				num_buf = 0;
171 			}
172 
173 			vaddr = page_address(sg_page_iter_page(&sg_iter));
174 			if (!vaddr) {
175 				pr_warn("null vaddr\n");
176 				ib_umem_release(umem);
177 				err = -ENOMEM;
178 				goto err1;
179 			}
180 
181 			buf->addr = (uintptr_t)vaddr;
182 			buf->size = PAGE_SIZE;
183 			num_buf++;
184 			buf++;
185 
186 		}
187 	}
188 
189 	mem->ibmr.pd		= &pd->ibpd;
190 	mem->umem		= umem;
191 	mem->access		= access;
192 	mem->length		= length;
193 	mem->iova		= iova;
194 	mem->va			= start;
195 	mem->offset		= ib_umem_offset(umem);
196 	mem->state		= RXE_MEM_STATE_VALID;
197 	mem->type		= RXE_MEM_TYPE_MR;
198 
199 	return 0;
200 
201 err1:
202 	return err;
203 }
204 
205 int rxe_mem_init_fast(struct rxe_pd *pd,
206 		      int max_pages, struct rxe_mem *mem)
207 {
208 	int err;
209 
210 	rxe_mem_init(0, mem);
211 
212 	/* In fastreg, we also set the rkey */
213 	mem->ibmr.rkey = mem->ibmr.lkey;
214 
215 	err = rxe_mem_alloc(mem, max_pages);
216 	if (err)
217 		goto err1;
218 
219 	mem->ibmr.pd		= &pd->ibpd;
220 	mem->max_buf		= max_pages;
221 	mem->state		= RXE_MEM_STATE_FREE;
222 	mem->type		= RXE_MEM_TYPE_MR;
223 
224 	return 0;
225 
226 err1:
227 	return err;
228 }
229 
230 static void lookup_iova(
231 	struct rxe_mem	*mem,
232 	u64			iova,
233 	int			*m_out,
234 	int			*n_out,
235 	size_t			*offset_out)
236 {
237 	size_t			offset = iova - mem->iova + mem->offset;
238 	int			map_index;
239 	int			buf_index;
240 	u64			length;
241 
242 	if (likely(mem->page_shift)) {
243 		*offset_out = offset & mem->page_mask;
244 		offset >>= mem->page_shift;
245 		*n_out = offset & mem->map_mask;
246 		*m_out = offset >> mem->map_shift;
247 	} else {
248 		map_index = 0;
249 		buf_index = 0;
250 
251 		length = mem->map[map_index]->buf[buf_index].size;
252 
253 		while (offset >= length) {
254 			offset -= length;
255 			buf_index++;
256 
257 			if (buf_index == RXE_BUF_PER_MAP) {
258 				map_index++;
259 				buf_index = 0;
260 			}
261 			length = mem->map[map_index]->buf[buf_index].size;
262 		}
263 
264 		*m_out = map_index;
265 		*n_out = buf_index;
266 		*offset_out = offset;
267 	}
268 }
269 
270 void *iova_to_vaddr(struct rxe_mem *mem, u64 iova, int length)
271 {
272 	size_t offset;
273 	int m, n;
274 	void *addr;
275 
276 	if (mem->state != RXE_MEM_STATE_VALID) {
277 		pr_warn("mem not in valid state\n");
278 		addr = NULL;
279 		goto out;
280 	}
281 
282 	if (!mem->map) {
283 		addr = (void *)(uintptr_t)iova;
284 		goto out;
285 	}
286 
287 	if (mem_check_range(mem, iova, length)) {
288 		pr_warn("range violation\n");
289 		addr = NULL;
290 		goto out;
291 	}
292 
293 	lookup_iova(mem, iova, &m, &n, &offset);
294 
295 	if (offset + length > mem->map[m]->buf[n].size) {
296 		pr_warn("crosses page boundary\n");
297 		addr = NULL;
298 		goto out;
299 	}
300 
301 	addr = (void *)(uintptr_t)mem->map[m]->buf[n].addr + offset;
302 
303 out:
304 	return addr;
305 }
306 
307 /* copy data from a range (vaddr, vaddr+length-1) to or from
308  * a mem object starting at iova. Compute incremental value of
309  * crc32 if crcp is not zero. caller must hold a reference to mem
310  */
311 int rxe_mem_copy(struct rxe_mem *mem, u64 iova, void *addr, int length,
312 		 enum copy_direction dir, u32 *crcp)
313 {
314 	int			err;
315 	int			bytes;
316 	u8			*va;
317 	struct rxe_map		**map;
318 	struct rxe_phys_buf	*buf;
319 	int			m;
320 	int			i;
321 	size_t			offset;
322 	u32			crc = crcp ? (*crcp) : 0;
323 
324 	if (length == 0)
325 		return 0;
326 
327 	if (mem->type == RXE_MEM_TYPE_DMA) {
328 		u8 *src, *dest;
329 
330 		src  = (dir == to_mem_obj) ?
331 			addr : ((void *)(uintptr_t)iova);
332 
333 		dest = (dir == to_mem_obj) ?
334 			((void *)(uintptr_t)iova) : addr;
335 
336 		memcpy(dest, src, length);
337 
338 		if (crcp)
339 			*crcp = rxe_crc32(to_rdev(mem->ibmr.device),
340 					*crcp, dest, length);
341 
342 		return 0;
343 	}
344 
345 	WARN_ON_ONCE(!mem->map);
346 
347 	err = mem_check_range(mem, iova, length);
348 	if (err) {
349 		err = -EFAULT;
350 		goto err1;
351 	}
352 
353 	lookup_iova(mem, iova, &m, &i, &offset);
354 
355 	map	= mem->map + m;
356 	buf	= map[0]->buf + i;
357 
358 	while (length > 0) {
359 		u8 *src, *dest;
360 
361 		va	= (u8 *)(uintptr_t)buf->addr + offset;
362 		src  = (dir == to_mem_obj) ? addr : va;
363 		dest = (dir == to_mem_obj) ? va : addr;
364 
365 		bytes	= buf->size - offset;
366 
367 		if (bytes > length)
368 			bytes = length;
369 
370 		memcpy(dest, src, bytes);
371 
372 		if (crcp)
373 			crc = rxe_crc32(to_rdev(mem->ibmr.device),
374 					crc, dest, bytes);
375 
376 		length	-= bytes;
377 		addr	+= bytes;
378 
379 		offset	= 0;
380 		buf++;
381 		i++;
382 
383 		if (i == RXE_BUF_PER_MAP) {
384 			i = 0;
385 			map++;
386 			buf = map[0]->buf;
387 		}
388 	}
389 
390 	if (crcp)
391 		*crcp = crc;
392 
393 	return 0;
394 
395 err1:
396 	return err;
397 }
398 
399 /* copy data in or out of a wqe, i.e. sg list
400  * under the control of a dma descriptor
401  */
402 int copy_data(
403 	struct rxe_pd		*pd,
404 	int			access,
405 	struct rxe_dma_info	*dma,
406 	void			*addr,
407 	int			length,
408 	enum copy_direction	dir,
409 	u32			*crcp)
410 {
411 	int			bytes;
412 	struct rxe_sge		*sge	= &dma->sge[dma->cur_sge];
413 	int			offset	= dma->sge_offset;
414 	int			resid	= dma->resid;
415 	struct rxe_mem		*mem	= NULL;
416 	u64			iova;
417 	int			err;
418 
419 	if (length == 0)
420 		return 0;
421 
422 	if (length > resid) {
423 		err = -EINVAL;
424 		goto err2;
425 	}
426 
427 	if (sge->length && (offset < sge->length)) {
428 		mem = lookup_mem(pd, access, sge->lkey, lookup_local);
429 		if (!mem) {
430 			err = -EINVAL;
431 			goto err1;
432 		}
433 	}
434 
435 	while (length > 0) {
436 		bytes = length;
437 
438 		if (offset >= sge->length) {
439 			if (mem) {
440 				rxe_drop_ref(mem);
441 				mem = NULL;
442 			}
443 			sge++;
444 			dma->cur_sge++;
445 			offset = 0;
446 
447 			if (dma->cur_sge >= dma->num_sge) {
448 				err = -ENOSPC;
449 				goto err2;
450 			}
451 
452 			if (sge->length) {
453 				mem = lookup_mem(pd, access, sge->lkey,
454 						 lookup_local);
455 				if (!mem) {
456 					err = -EINVAL;
457 					goto err1;
458 				}
459 			} else {
460 				continue;
461 			}
462 		}
463 
464 		if (bytes > sge->length - offset)
465 			bytes = sge->length - offset;
466 
467 		if (bytes > 0) {
468 			iova = sge->addr + offset;
469 
470 			err = rxe_mem_copy(mem, iova, addr, bytes, dir, crcp);
471 			if (err)
472 				goto err2;
473 
474 			offset	+= bytes;
475 			resid	-= bytes;
476 			length	-= bytes;
477 			addr	+= bytes;
478 		}
479 	}
480 
481 	dma->sge_offset = offset;
482 	dma->resid	= resid;
483 
484 	if (mem)
485 		rxe_drop_ref(mem);
486 
487 	return 0;
488 
489 err2:
490 	if (mem)
491 		rxe_drop_ref(mem);
492 err1:
493 	return err;
494 }
495 
496 int advance_dma_data(struct rxe_dma_info *dma, unsigned int length)
497 {
498 	struct rxe_sge		*sge	= &dma->sge[dma->cur_sge];
499 	int			offset	= dma->sge_offset;
500 	int			resid	= dma->resid;
501 
502 	while (length) {
503 		unsigned int bytes;
504 
505 		if (offset >= sge->length) {
506 			sge++;
507 			dma->cur_sge++;
508 			offset = 0;
509 			if (dma->cur_sge >= dma->num_sge)
510 				return -ENOSPC;
511 		}
512 
513 		bytes = length;
514 
515 		if (bytes > sge->length - offset)
516 			bytes = sge->length - offset;
517 
518 		offset	+= bytes;
519 		resid	-= bytes;
520 		length	-= bytes;
521 	}
522 
523 	dma->sge_offset = offset;
524 	dma->resid	= resid;
525 
526 	return 0;
527 }
528 
529 /* (1) find the mem (mr or mw) corresponding to lkey/rkey
530  *     depending on lookup_type
531  * (2) verify that the (qp) pd matches the mem pd
532  * (3) verify that the mem can support the requested access
533  * (4) verify that mem state is valid
534  */
535 struct rxe_mem *lookup_mem(struct rxe_pd *pd, int access, u32 key,
536 			   enum lookup_type type)
537 {
538 	struct rxe_mem *mem;
539 	struct rxe_dev *rxe = to_rdev(pd->ibpd.device);
540 	int index = key >> 8;
541 
542 	mem = rxe_pool_get_index(&rxe->mr_pool, index);
543 	if (!mem)
544 		return NULL;
545 
546 	if (unlikely((type == lookup_local && mr_lkey(mem) != key) ||
547 		     (type == lookup_remote && mr_rkey(mem) != key) ||
548 		     mr_pd(mem) != pd ||
549 		     (access && !(access & mem->access)) ||
550 		     mem->state != RXE_MEM_STATE_VALID)) {
551 		rxe_drop_ref(mem);
552 		mem = NULL;
553 	}
554 
555 	return mem;
556 }
557