xref: /openbmc/linux/drivers/infiniband/hw/mlx4/mr.c (revision 9cfc5c90)
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 
36 #include "mlx4_ib.h"
37 
38 static u32 convert_access(int acc)
39 {
40 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX4_PERM_ATOMIC       : 0) |
41 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MLX4_PERM_REMOTE_WRITE : 0) |
42 	       (acc & IB_ACCESS_REMOTE_READ   ? MLX4_PERM_REMOTE_READ  : 0) |
43 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MLX4_PERM_LOCAL_WRITE  : 0) |
44 	       (acc & IB_ACCESS_MW_BIND	      ? MLX4_PERM_BIND_MW      : 0) |
45 	       MLX4_PERM_LOCAL_READ;
46 }
47 
48 static enum mlx4_mw_type to_mlx4_type(enum ib_mw_type type)
49 {
50 	switch (type) {
51 	case IB_MW_TYPE_1:	return MLX4_MW_TYPE_1;
52 	case IB_MW_TYPE_2:	return MLX4_MW_TYPE_2;
53 	default:		return -1;
54 	}
55 }
56 
57 struct ib_mr *mlx4_ib_get_dma_mr(struct ib_pd *pd, int acc)
58 {
59 	struct mlx4_ib_mr *mr;
60 	int err;
61 
62 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
63 	if (!mr)
64 		return ERR_PTR(-ENOMEM);
65 
66 	err = mlx4_mr_alloc(to_mdev(pd->device)->dev, to_mpd(pd)->pdn, 0,
67 			    ~0ull, convert_access(acc), 0, 0, &mr->mmr);
68 	if (err)
69 		goto err_free;
70 
71 	err = mlx4_mr_enable(to_mdev(pd->device)->dev, &mr->mmr);
72 	if (err)
73 		goto err_mr;
74 
75 	mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
76 	mr->umem = NULL;
77 
78 	return &mr->ibmr;
79 
80 err_mr:
81 	(void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
82 
83 err_free:
84 	kfree(mr);
85 
86 	return ERR_PTR(err);
87 }
88 
89 int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt,
90 			   struct ib_umem *umem)
91 {
92 	u64 *pages;
93 	int i, k, entry;
94 	int n;
95 	int len;
96 	int err = 0;
97 	struct scatterlist *sg;
98 
99 	pages = (u64 *) __get_free_page(GFP_KERNEL);
100 	if (!pages)
101 		return -ENOMEM;
102 
103 	i = n = 0;
104 
105 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
106 		len = sg_dma_len(sg) >> mtt->page_shift;
107 		for (k = 0; k < len; ++k) {
108 			pages[i++] = sg_dma_address(sg) +
109 				umem->page_size * k;
110 			/*
111 			 * Be friendly to mlx4_write_mtt() and
112 			 * pass it chunks of appropriate size.
113 			 */
114 			if (i == PAGE_SIZE / sizeof (u64)) {
115 				err = mlx4_write_mtt(dev->dev, mtt, n,
116 						     i, pages);
117 				if (err)
118 					goto out;
119 				n += i;
120 				i = 0;
121 			}
122 		}
123 	}
124 
125 	if (i)
126 		err = mlx4_write_mtt(dev->dev, mtt, n, i, pages);
127 
128 out:
129 	free_page((unsigned long) pages);
130 	return err;
131 }
132 
133 struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
134 				  u64 virt_addr, int access_flags,
135 				  struct ib_udata *udata)
136 {
137 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
138 	struct mlx4_ib_mr *mr;
139 	int shift;
140 	int err;
141 	int n;
142 
143 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
144 	if (!mr)
145 		return ERR_PTR(-ENOMEM);
146 
147 	/* Force registering the memory as writable. */
148 	/* Used for memory re-registeration. HCA protects the access */
149 	mr->umem = ib_umem_get(pd->uobject->context, start, length,
150 			       access_flags | IB_ACCESS_LOCAL_WRITE, 0);
151 	if (IS_ERR(mr->umem)) {
152 		err = PTR_ERR(mr->umem);
153 		goto err_free;
154 	}
155 
156 	n = ib_umem_page_count(mr->umem);
157 	shift = ilog2(mr->umem->page_size);
158 
159 	err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length,
160 			    convert_access(access_flags), n, shift, &mr->mmr);
161 	if (err)
162 		goto err_umem;
163 
164 	err = mlx4_ib_umem_write_mtt(dev, &mr->mmr.mtt, mr->umem);
165 	if (err)
166 		goto err_mr;
167 
168 	err = mlx4_mr_enable(dev->dev, &mr->mmr);
169 	if (err)
170 		goto err_mr;
171 
172 	mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
173 
174 	return &mr->ibmr;
175 
176 err_mr:
177 	(void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
178 
179 err_umem:
180 	ib_umem_release(mr->umem);
181 
182 err_free:
183 	kfree(mr);
184 
185 	return ERR_PTR(err);
186 }
187 
188 int mlx4_ib_rereg_user_mr(struct ib_mr *mr, int flags,
189 			  u64 start, u64 length, u64 virt_addr,
190 			  int mr_access_flags, struct ib_pd *pd,
191 			  struct ib_udata *udata)
192 {
193 	struct mlx4_ib_dev *dev = to_mdev(mr->device);
194 	struct mlx4_ib_mr *mmr = to_mmr(mr);
195 	struct mlx4_mpt_entry *mpt_entry;
196 	struct mlx4_mpt_entry **pmpt_entry = &mpt_entry;
197 	int err;
198 
199 	/* Since we synchronize this call and mlx4_ib_dereg_mr via uverbs,
200 	 * we assume that the calls can't run concurrently. Otherwise, a
201 	 * race exists.
202 	 */
203 	err =  mlx4_mr_hw_get_mpt(dev->dev, &mmr->mmr, &pmpt_entry);
204 
205 	if (err)
206 		return err;
207 
208 	if (flags & IB_MR_REREG_PD) {
209 		err = mlx4_mr_hw_change_pd(dev->dev, *pmpt_entry,
210 					   to_mpd(pd)->pdn);
211 
212 		if (err)
213 			goto release_mpt_entry;
214 	}
215 
216 	if (flags & IB_MR_REREG_ACCESS) {
217 		err = mlx4_mr_hw_change_access(dev->dev, *pmpt_entry,
218 					       convert_access(mr_access_flags));
219 
220 		if (err)
221 			goto release_mpt_entry;
222 	}
223 
224 	if (flags & IB_MR_REREG_TRANS) {
225 		int shift;
226 		int n;
227 
228 		mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
229 		ib_umem_release(mmr->umem);
230 		mmr->umem = ib_umem_get(mr->uobject->context, start, length,
231 					mr_access_flags |
232 					IB_ACCESS_LOCAL_WRITE,
233 					0);
234 		if (IS_ERR(mmr->umem)) {
235 			err = PTR_ERR(mmr->umem);
236 			/* Prevent mlx4_ib_dereg_mr from free'ing invalid pointer */
237 			mmr->umem = NULL;
238 			goto release_mpt_entry;
239 		}
240 		n = ib_umem_page_count(mmr->umem);
241 		shift = ilog2(mmr->umem->page_size);
242 
243 		err = mlx4_mr_rereg_mem_write(dev->dev, &mmr->mmr,
244 					      virt_addr, length, n, shift,
245 					      *pmpt_entry);
246 		if (err) {
247 			ib_umem_release(mmr->umem);
248 			goto release_mpt_entry;
249 		}
250 		mmr->mmr.iova       = virt_addr;
251 		mmr->mmr.size       = length;
252 
253 		err = mlx4_ib_umem_write_mtt(dev, &mmr->mmr.mtt, mmr->umem);
254 		if (err) {
255 			mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
256 			ib_umem_release(mmr->umem);
257 			goto release_mpt_entry;
258 		}
259 	}
260 
261 	/* If we couldn't transfer the MR to the HCA, just remember to
262 	 * return a failure. But dereg_mr will free the resources.
263 	 */
264 	err = mlx4_mr_hw_write_mpt(dev->dev, &mmr->mmr, pmpt_entry);
265 	if (!err && flags & IB_MR_REREG_ACCESS)
266 		mmr->mmr.access = mr_access_flags;
267 
268 release_mpt_entry:
269 	mlx4_mr_hw_put_mpt(dev->dev, pmpt_entry);
270 
271 	return err;
272 }
273 
274 static int
275 mlx4_alloc_priv_pages(struct ib_device *device,
276 		      struct mlx4_ib_mr *mr,
277 		      int max_pages)
278 {
279 	int size = max_pages * sizeof(u64);
280 	int add_size;
281 	int ret;
282 
283 	add_size = max_t(int, MLX4_MR_PAGES_ALIGN - ARCH_KMALLOC_MINALIGN, 0);
284 
285 	mr->pages_alloc = kzalloc(size + add_size, GFP_KERNEL);
286 	if (!mr->pages_alloc)
287 		return -ENOMEM;
288 
289 	mr->pages = PTR_ALIGN(mr->pages_alloc, MLX4_MR_PAGES_ALIGN);
290 
291 	mr->page_map = dma_map_single(device->dma_device, mr->pages,
292 				      size, DMA_TO_DEVICE);
293 
294 	if (dma_mapping_error(device->dma_device, mr->page_map)) {
295 		ret = -ENOMEM;
296 		goto err;
297 	}
298 
299 	return 0;
300 err:
301 	kfree(mr->pages_alloc);
302 
303 	return ret;
304 }
305 
306 static void
307 mlx4_free_priv_pages(struct mlx4_ib_mr *mr)
308 {
309 	if (mr->pages) {
310 		struct ib_device *device = mr->ibmr.device;
311 		int size = mr->max_pages * sizeof(u64);
312 
313 		dma_unmap_single(device->dma_device, mr->page_map,
314 				 size, DMA_TO_DEVICE);
315 		kfree(mr->pages_alloc);
316 		mr->pages = NULL;
317 	}
318 }
319 
320 int mlx4_ib_dereg_mr(struct ib_mr *ibmr)
321 {
322 	struct mlx4_ib_mr *mr = to_mmr(ibmr);
323 	int ret;
324 
325 	mlx4_free_priv_pages(mr);
326 
327 	ret = mlx4_mr_free(to_mdev(ibmr->device)->dev, &mr->mmr);
328 	if (ret)
329 		return ret;
330 	if (mr->umem)
331 		ib_umem_release(mr->umem);
332 	kfree(mr);
333 
334 	return 0;
335 }
336 
337 struct ib_mw *mlx4_ib_alloc_mw(struct ib_pd *pd, enum ib_mw_type type)
338 {
339 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
340 	struct mlx4_ib_mw *mw;
341 	int err;
342 
343 	mw = kmalloc(sizeof(*mw), GFP_KERNEL);
344 	if (!mw)
345 		return ERR_PTR(-ENOMEM);
346 
347 	err = mlx4_mw_alloc(dev->dev, to_mpd(pd)->pdn,
348 			    to_mlx4_type(type), &mw->mmw);
349 	if (err)
350 		goto err_free;
351 
352 	err = mlx4_mw_enable(dev->dev, &mw->mmw);
353 	if (err)
354 		goto err_mw;
355 
356 	mw->ibmw.rkey = mw->mmw.key;
357 
358 	return &mw->ibmw;
359 
360 err_mw:
361 	mlx4_mw_free(dev->dev, &mw->mmw);
362 
363 err_free:
364 	kfree(mw);
365 
366 	return ERR_PTR(err);
367 }
368 
369 int mlx4_ib_bind_mw(struct ib_qp *qp, struct ib_mw *mw,
370 		    struct ib_mw_bind *mw_bind)
371 {
372 	struct ib_bind_mw_wr  wr;
373 	struct ib_send_wr *bad_wr;
374 	int ret;
375 
376 	memset(&wr, 0, sizeof(wr));
377 	wr.wr.opcode		= IB_WR_BIND_MW;
378 	wr.wr.wr_id		= mw_bind->wr_id;
379 	wr.wr.send_flags	= mw_bind->send_flags;
380 	wr.mw			= mw;
381 	wr.bind_info		= mw_bind->bind_info;
382 	wr.rkey			= ib_inc_rkey(mw->rkey);
383 
384 	ret = mlx4_ib_post_send(qp, &wr.wr, &bad_wr);
385 	if (!ret)
386 		mw->rkey = wr.rkey;
387 
388 	return ret;
389 }
390 
391 int mlx4_ib_dealloc_mw(struct ib_mw *ibmw)
392 {
393 	struct mlx4_ib_mw *mw = to_mmw(ibmw);
394 
395 	mlx4_mw_free(to_mdev(ibmw->device)->dev, &mw->mmw);
396 	kfree(mw);
397 
398 	return 0;
399 }
400 
401 struct ib_mr *mlx4_ib_alloc_mr(struct ib_pd *pd,
402 			       enum ib_mr_type mr_type,
403 			       u32 max_num_sg)
404 {
405 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
406 	struct mlx4_ib_mr *mr;
407 	int err;
408 
409 	if (mr_type != IB_MR_TYPE_MEM_REG ||
410 	    max_num_sg > MLX4_MAX_FAST_REG_PAGES)
411 		return ERR_PTR(-EINVAL);
412 
413 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
414 	if (!mr)
415 		return ERR_PTR(-ENOMEM);
416 
417 	err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, 0, 0, 0,
418 			    max_num_sg, 0, &mr->mmr);
419 	if (err)
420 		goto err_free;
421 
422 	err = mlx4_alloc_priv_pages(pd->device, mr, max_num_sg);
423 	if (err)
424 		goto err_free_mr;
425 
426 	mr->max_pages = max_num_sg;
427 
428 	err = mlx4_mr_enable(dev->dev, &mr->mmr);
429 	if (err)
430 		goto err_free_pl;
431 
432 	mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
433 	mr->umem = NULL;
434 
435 	return &mr->ibmr;
436 
437 err_free_pl:
438 	mlx4_free_priv_pages(mr);
439 err_free_mr:
440 	(void) mlx4_mr_free(dev->dev, &mr->mmr);
441 err_free:
442 	kfree(mr);
443 	return ERR_PTR(err);
444 }
445 
446 struct ib_fmr *mlx4_ib_fmr_alloc(struct ib_pd *pd, int acc,
447 				 struct ib_fmr_attr *fmr_attr)
448 {
449 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
450 	struct mlx4_ib_fmr *fmr;
451 	int err = -ENOMEM;
452 
453 	fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
454 	if (!fmr)
455 		return ERR_PTR(-ENOMEM);
456 
457 	err = mlx4_fmr_alloc(dev->dev, to_mpd(pd)->pdn, convert_access(acc),
458 			     fmr_attr->max_pages, fmr_attr->max_maps,
459 			     fmr_attr->page_shift, &fmr->mfmr);
460 	if (err)
461 		goto err_free;
462 
463 	err = mlx4_fmr_enable(to_mdev(pd->device)->dev, &fmr->mfmr);
464 	if (err)
465 		goto err_mr;
466 
467 	fmr->ibfmr.rkey = fmr->ibfmr.lkey = fmr->mfmr.mr.key;
468 
469 	return &fmr->ibfmr;
470 
471 err_mr:
472 	(void) mlx4_mr_free(to_mdev(pd->device)->dev, &fmr->mfmr.mr);
473 
474 err_free:
475 	kfree(fmr);
476 
477 	return ERR_PTR(err);
478 }
479 
480 int mlx4_ib_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
481 		      int npages, u64 iova)
482 {
483 	struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
484 	struct mlx4_ib_dev *dev = to_mdev(ifmr->ibfmr.device);
485 
486 	return mlx4_map_phys_fmr(dev->dev, &ifmr->mfmr, page_list, npages, iova,
487 				 &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey);
488 }
489 
490 int mlx4_ib_unmap_fmr(struct list_head *fmr_list)
491 {
492 	struct ib_fmr *ibfmr;
493 	int err;
494 	struct mlx4_dev *mdev = NULL;
495 
496 	list_for_each_entry(ibfmr, fmr_list, list) {
497 		if (mdev && to_mdev(ibfmr->device)->dev != mdev)
498 			return -EINVAL;
499 		mdev = to_mdev(ibfmr->device)->dev;
500 	}
501 
502 	if (!mdev)
503 		return 0;
504 
505 	list_for_each_entry(ibfmr, fmr_list, list) {
506 		struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
507 
508 		mlx4_fmr_unmap(mdev, &ifmr->mfmr, &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey);
509 	}
510 
511 	/*
512 	 * Make sure all MPT status updates are visible before issuing
513 	 * SYNC_TPT firmware command.
514 	 */
515 	wmb();
516 
517 	err = mlx4_SYNC_TPT(mdev);
518 	if (err)
519 		pr_warn("SYNC_TPT error %d when "
520 		       "unmapping FMRs\n", err);
521 
522 	return 0;
523 }
524 
525 int mlx4_ib_fmr_dealloc(struct ib_fmr *ibfmr)
526 {
527 	struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
528 	struct mlx4_ib_dev *dev = to_mdev(ibfmr->device);
529 	int err;
530 
531 	err = mlx4_fmr_free(dev->dev, &ifmr->mfmr);
532 
533 	if (!err)
534 		kfree(ifmr);
535 
536 	return err;
537 }
538 
539 static int mlx4_set_page(struct ib_mr *ibmr, u64 addr)
540 {
541 	struct mlx4_ib_mr *mr = to_mmr(ibmr);
542 
543 	if (unlikely(mr->npages == mr->max_pages))
544 		return -ENOMEM;
545 
546 	mr->pages[mr->npages++] = cpu_to_be64(addr | MLX4_MTT_FLAG_PRESENT);
547 
548 	return 0;
549 }
550 
551 int mlx4_ib_map_mr_sg(struct ib_mr *ibmr,
552 		      struct scatterlist *sg,
553 		      int sg_nents)
554 {
555 	struct mlx4_ib_mr *mr = to_mmr(ibmr);
556 	int rc;
557 
558 	mr->npages = 0;
559 
560 	ib_dma_sync_single_for_cpu(ibmr->device, mr->page_map,
561 				   sizeof(u64) * mr->max_pages,
562 				   DMA_TO_DEVICE);
563 
564 	rc = ib_sg_to_pages(ibmr, sg, sg_nents, mlx4_set_page);
565 
566 	ib_dma_sync_single_for_device(ibmr->device, mr->page_map,
567 				      sizeof(u64) * mr->max_pages,
568 				      DMA_TO_DEVICE);
569 
570 	return rc;
571 }
572