xref: /linux/drivers/infiniband/hw/mlx4/mr.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 = kmalloc(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 = kmalloc(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 int mlx4_ib_dereg_mr(struct ib_mr *ibmr)
275 {
276 	struct mlx4_ib_mr *mr = to_mmr(ibmr);
277 	int ret;
278 
279 	ret = mlx4_mr_free(to_mdev(ibmr->device)->dev, &mr->mmr);
280 	if (ret)
281 		return ret;
282 	if (mr->umem)
283 		ib_umem_release(mr->umem);
284 	kfree(mr);
285 
286 	return 0;
287 }
288 
289 struct ib_mw *mlx4_ib_alloc_mw(struct ib_pd *pd, enum ib_mw_type type)
290 {
291 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
292 	struct mlx4_ib_mw *mw;
293 	int err;
294 
295 	mw = kmalloc(sizeof(*mw), GFP_KERNEL);
296 	if (!mw)
297 		return ERR_PTR(-ENOMEM);
298 
299 	err = mlx4_mw_alloc(dev->dev, to_mpd(pd)->pdn,
300 			    to_mlx4_type(type), &mw->mmw);
301 	if (err)
302 		goto err_free;
303 
304 	err = mlx4_mw_enable(dev->dev, &mw->mmw);
305 	if (err)
306 		goto err_mw;
307 
308 	mw->ibmw.rkey = mw->mmw.key;
309 
310 	return &mw->ibmw;
311 
312 err_mw:
313 	mlx4_mw_free(dev->dev, &mw->mmw);
314 
315 err_free:
316 	kfree(mw);
317 
318 	return ERR_PTR(err);
319 }
320 
321 int mlx4_ib_bind_mw(struct ib_qp *qp, struct ib_mw *mw,
322 		    struct ib_mw_bind *mw_bind)
323 {
324 	struct ib_send_wr  wr;
325 	struct ib_send_wr *bad_wr;
326 	int ret;
327 
328 	memset(&wr, 0, sizeof(wr));
329 	wr.opcode               = IB_WR_BIND_MW;
330 	wr.wr_id                = mw_bind->wr_id;
331 	wr.send_flags           = mw_bind->send_flags;
332 	wr.wr.bind_mw.mw        = mw;
333 	wr.wr.bind_mw.bind_info = mw_bind->bind_info;
334 	wr.wr.bind_mw.rkey      = ib_inc_rkey(mw->rkey);
335 
336 	ret = mlx4_ib_post_send(qp, &wr, &bad_wr);
337 	if (!ret)
338 		mw->rkey = wr.wr.bind_mw.rkey;
339 
340 	return ret;
341 }
342 
343 int mlx4_ib_dealloc_mw(struct ib_mw *ibmw)
344 {
345 	struct mlx4_ib_mw *mw = to_mmw(ibmw);
346 
347 	mlx4_mw_free(to_mdev(ibmw->device)->dev, &mw->mmw);
348 	kfree(mw);
349 
350 	return 0;
351 }
352 
353 struct ib_mr *mlx4_ib_alloc_mr(struct ib_pd *pd,
354 			       enum ib_mr_type mr_type,
355 			       u32 max_num_sg)
356 {
357 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
358 	struct mlx4_ib_mr *mr;
359 	int err;
360 
361 	if (mr_type != IB_MR_TYPE_MEM_REG ||
362 	    max_num_sg > MLX4_MAX_FAST_REG_PAGES)
363 		return ERR_PTR(-EINVAL);
364 
365 	mr = kmalloc(sizeof *mr, GFP_KERNEL);
366 	if (!mr)
367 		return ERR_PTR(-ENOMEM);
368 
369 	err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, 0, 0, 0,
370 			    max_num_sg, 0, &mr->mmr);
371 	if (err)
372 		goto err_free;
373 
374 	err = mlx4_mr_enable(dev->dev, &mr->mmr);
375 	if (err)
376 		goto err_mr;
377 
378 	mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
379 	mr->umem = NULL;
380 
381 	return &mr->ibmr;
382 
383 err_mr:
384 	(void) mlx4_mr_free(dev->dev, &mr->mmr);
385 
386 err_free:
387 	kfree(mr);
388 	return ERR_PTR(err);
389 }
390 
391 struct ib_fast_reg_page_list *mlx4_ib_alloc_fast_reg_page_list(struct ib_device *ibdev,
392 							       int page_list_len)
393 {
394 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
395 	struct mlx4_ib_fast_reg_page_list *mfrpl;
396 	int size = page_list_len * sizeof (u64);
397 
398 	if (page_list_len > MLX4_MAX_FAST_REG_PAGES)
399 		return ERR_PTR(-EINVAL);
400 
401 	mfrpl = kmalloc(sizeof *mfrpl, GFP_KERNEL);
402 	if (!mfrpl)
403 		return ERR_PTR(-ENOMEM);
404 
405 	mfrpl->ibfrpl.page_list = kmalloc(size, GFP_KERNEL);
406 	if (!mfrpl->ibfrpl.page_list)
407 		goto err_free;
408 
409 	mfrpl->mapped_page_list = dma_alloc_coherent(&dev->dev->persist->
410 						     pdev->dev,
411 						     size, &mfrpl->map,
412 						     GFP_KERNEL);
413 	if (!mfrpl->mapped_page_list)
414 		goto err_free;
415 
416 	WARN_ON(mfrpl->map & 0x3f);
417 
418 	return &mfrpl->ibfrpl;
419 
420 err_free:
421 	kfree(mfrpl->ibfrpl.page_list);
422 	kfree(mfrpl);
423 	return ERR_PTR(-ENOMEM);
424 }
425 
426 void mlx4_ib_free_fast_reg_page_list(struct ib_fast_reg_page_list *page_list)
427 {
428 	struct mlx4_ib_dev *dev = to_mdev(page_list->device);
429 	struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(page_list);
430 	int size = page_list->max_page_list_len * sizeof (u64);
431 
432 	dma_free_coherent(&dev->dev->persist->pdev->dev, size,
433 			  mfrpl->mapped_page_list,
434 			  mfrpl->map);
435 	kfree(mfrpl->ibfrpl.page_list);
436 	kfree(mfrpl);
437 }
438 
439 struct ib_fmr *mlx4_ib_fmr_alloc(struct ib_pd *pd, int acc,
440 				 struct ib_fmr_attr *fmr_attr)
441 {
442 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
443 	struct mlx4_ib_fmr *fmr;
444 	int err = -ENOMEM;
445 
446 	fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
447 	if (!fmr)
448 		return ERR_PTR(-ENOMEM);
449 
450 	err = mlx4_fmr_alloc(dev->dev, to_mpd(pd)->pdn, convert_access(acc),
451 			     fmr_attr->max_pages, fmr_attr->max_maps,
452 			     fmr_attr->page_shift, &fmr->mfmr);
453 	if (err)
454 		goto err_free;
455 
456 	err = mlx4_fmr_enable(to_mdev(pd->device)->dev, &fmr->mfmr);
457 	if (err)
458 		goto err_mr;
459 
460 	fmr->ibfmr.rkey = fmr->ibfmr.lkey = fmr->mfmr.mr.key;
461 
462 	return &fmr->ibfmr;
463 
464 err_mr:
465 	(void) mlx4_mr_free(to_mdev(pd->device)->dev, &fmr->mfmr.mr);
466 
467 err_free:
468 	kfree(fmr);
469 
470 	return ERR_PTR(err);
471 }
472 
473 int mlx4_ib_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
474 		      int npages, u64 iova)
475 {
476 	struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
477 	struct mlx4_ib_dev *dev = to_mdev(ifmr->ibfmr.device);
478 
479 	return mlx4_map_phys_fmr(dev->dev, &ifmr->mfmr, page_list, npages, iova,
480 				 &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey);
481 }
482 
483 int mlx4_ib_unmap_fmr(struct list_head *fmr_list)
484 {
485 	struct ib_fmr *ibfmr;
486 	int err;
487 	struct mlx4_dev *mdev = NULL;
488 
489 	list_for_each_entry(ibfmr, fmr_list, list) {
490 		if (mdev && to_mdev(ibfmr->device)->dev != mdev)
491 			return -EINVAL;
492 		mdev = to_mdev(ibfmr->device)->dev;
493 	}
494 
495 	if (!mdev)
496 		return 0;
497 
498 	list_for_each_entry(ibfmr, fmr_list, list) {
499 		struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
500 
501 		mlx4_fmr_unmap(mdev, &ifmr->mfmr, &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey);
502 	}
503 
504 	/*
505 	 * Make sure all MPT status updates are visible before issuing
506 	 * SYNC_TPT firmware command.
507 	 */
508 	wmb();
509 
510 	err = mlx4_SYNC_TPT(mdev);
511 	if (err)
512 		pr_warn("SYNC_TPT error %d when "
513 		       "unmapping FMRs\n", err);
514 
515 	return 0;
516 }
517 
518 int mlx4_ib_fmr_dealloc(struct ib_fmr *ibfmr)
519 {
520 	struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
521 	struct mlx4_ib_dev *dev = to_mdev(ibfmr->device);
522 	int err;
523 
524 	err = mlx4_fmr_free(dev->dev, &ifmr->mfmr);
525 
526 	if (!err)
527 		kfree(ifmr);
528 
529 	return err;
530 }
531