2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
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:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
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.
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
34 #include <linux/slab.h>
35 #include <linux/module.h>
36 #include <linux/sched.h>
38 #include <asm/atomic64.h>
42 static u32 convert_access(int acc)
44 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX4_PERM_ATOMIC : 0) |
45 (acc & IB_ACCESS_REMOTE_WRITE ? MLX4_PERM_REMOTE_WRITE : 0) |
46 (acc & IB_ACCESS_REMOTE_READ ? MLX4_PERM_REMOTE_READ : 0) |
47 (acc & IB_ACCESS_LOCAL_WRITE ? MLX4_PERM_LOCAL_WRITE : 0) |
48 (acc & IB_ACCESS_MW_BIND ? MLX4_PERM_BIND_MW : 0) |
52 static enum mlx4_mw_type to_mlx4_type(enum ib_mw_type type)
55 case IB_MW_TYPE_1: return MLX4_MW_TYPE_1;
56 case IB_MW_TYPE_2: return MLX4_MW_TYPE_2;
61 struct ib_mr *mlx4_ib_get_dma_mr(struct ib_pd *pd, int acc)
63 struct mlx4_ib_mr *mr;
66 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
68 return ERR_PTR(-ENOMEM);
70 err = mlx4_mr_alloc(to_mdev(pd->device)->dev, to_mpd(pd)->pdn, 0,
71 ~0ull, convert_access(acc), 0, 0, &mr->mmr);
75 err = mlx4_mr_enable(to_mdev(pd->device)->dev, &mr->mmr);
79 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
85 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
93 int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt,
101 struct scatterlist *sg;
103 pages = (u64 *) __get_free_page(GFP_KERNEL);
109 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
110 len = sg_dma_len(sg) >> mtt->page_shift;
111 for (k = 0; k < len; ++k) {
112 pages[i++] = sg_dma_address(sg) +
115 * Be friendly to mlx4_write_mtt() and
116 * pass it chunks of appropriate size.
118 if (i == PAGE_SIZE / sizeof (u64)) {
119 err = mlx4_write_mtt(dev->dev, mtt, n,
130 err = mlx4_write_mtt(dev->dev, mtt, n, i, pages);
133 free_page((unsigned long) pages);
137 struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
138 u64 virt_addr, int access_flags,
139 struct ib_udata *udata)
141 struct mlx4_ib_dev *dev = to_mdev(pd->device);
142 struct mlx4_ib_mr *mr;
147 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
149 return ERR_PTR(-ENOMEM);
151 /* Force registering the memory as writable. */
152 /* Used for memory re-registeration. HCA protects the access */
153 mr->umem = ib_umem_get(pd->uobject->context, start, length,
154 access_flags | IB_ACCESS_LOCAL_WRITE, 0);
155 if (IS_ERR(mr->umem)) {
156 err = PTR_ERR(mr->umem);
160 n = ib_umem_page_count(mr->umem);
161 shift = ilog2(mr->umem->page_size);
163 err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length,
164 convert_access(access_flags), n, shift, &mr->mmr);
168 err = mlx4_ib_umem_write_mtt(dev, &mr->mmr.mtt, mr->umem);
172 err = mlx4_mr_enable(dev->dev, &mr->mmr);
176 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
181 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
184 ib_umem_release(mr->umem);
192 int mlx4_ib_rereg_user_mr(struct ib_mr *mr, int flags,
193 u64 start, u64 length, u64 virt_addr,
194 int mr_access_flags, struct ib_pd *pd,
195 struct ib_udata *udata)
197 struct mlx4_ib_dev *dev = to_mdev(mr->device);
198 struct mlx4_ib_mr *mmr = to_mmr(mr);
199 struct mlx4_mpt_entry *mpt_entry;
200 struct mlx4_mpt_entry **pmpt_entry = &mpt_entry;
203 /* Since we synchronize this call and mlx4_ib_dereg_mr via uverbs,
204 * we assume that the calls can't run concurrently. Otherwise, a
207 err = mlx4_mr_hw_get_mpt(dev->dev, &mmr->mmr, &pmpt_entry);
212 if (flags & IB_MR_REREG_PD) {
213 err = mlx4_mr_hw_change_pd(dev->dev, *pmpt_entry,
217 goto release_mpt_entry;
220 if (flags & IB_MR_REREG_ACCESS) {
221 err = mlx4_mr_hw_change_access(dev->dev, *pmpt_entry,
222 convert_access(mr_access_flags));
225 goto release_mpt_entry;
228 if (flags & IB_MR_REREG_TRANS) {
232 mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
233 ib_umem_release(mmr->umem);
234 mmr->umem = ib_umem_get(mr->uobject->context, start, length,
236 IB_ACCESS_LOCAL_WRITE,
238 if (IS_ERR(mmr->umem)) {
239 err = PTR_ERR(mmr->umem);
240 /* Prevent mlx4_ib_dereg_mr from free'ing invalid pointer */
242 goto release_mpt_entry;
244 n = ib_umem_page_count(mmr->umem);
245 shift = ilog2(mmr->umem->page_size);
247 err = mlx4_mr_rereg_mem_write(dev->dev, &mmr->mmr,
248 virt_addr, length, n, shift,
251 ib_umem_release(mmr->umem);
252 goto release_mpt_entry;
254 mmr->mmr.iova = virt_addr;
255 mmr->mmr.size = length;
257 err = mlx4_ib_umem_write_mtt(dev, &mmr->mmr.mtt, mmr->umem);
259 mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
260 ib_umem_release(mmr->umem);
261 goto release_mpt_entry;
265 /* If we couldn't transfer the MR to the HCA, just remember to
266 * return a failure. But dereg_mr will free the resources.
268 err = mlx4_mr_hw_write_mpt(dev->dev, &mmr->mmr, pmpt_entry);
269 if (!err && flags & IB_MR_REREG_ACCESS)
270 mmr->mmr.access = mr_access_flags;
273 mlx4_mr_hw_put_mpt(dev->dev, pmpt_entry);
279 mlx4_alloc_priv_pages(struct ib_device *device,
280 struct mlx4_ib_mr *mr,
285 /* Ensure that size is aligned to DMA cacheline
287 * max_pages is limited to MLX4_MAX_FAST_REG_PAGES
288 * so page_map_size will never cross PAGE_SIZE.
290 mr->page_map_size = roundup(max_pages * sizeof(u64),
291 MLX4_MR_PAGES_ALIGN);
293 /* Prevent cross page boundary allocation. */
294 mr->pages = (__be64 *)get_zeroed_page(GFP_KERNEL);
298 mr->page_map = dma_map_single(device->dma_device, mr->pages,
299 mr->page_map_size, DMA_TO_DEVICE);
301 if (dma_mapping_error(device->dma_device, mr->page_map)) {
309 free_page((unsigned long)mr->pages);
314 mlx4_free_priv_pages(struct mlx4_ib_mr *mr)
317 struct ib_device *device = mr->ibmr.device;
319 dma_unmap_single(device->dma_device, mr->page_map,
320 mr->page_map_size, DMA_TO_DEVICE);
321 free_page((unsigned long)mr->pages);
326 int mlx4_ib_dereg_mr(struct ib_mr *ibmr)
328 struct mlx4_ib_mr *mr = to_mmr(ibmr);
331 mlx4_free_priv_pages(mr);
333 ret = mlx4_mr_free(to_mdev(ibmr->device)->dev, &mr->mmr);
337 ib_umem_release(mr->umem);
343 struct ib_mw *mlx4_ib_alloc_mw(struct ib_pd *pd, enum ib_mw_type type,
344 struct ib_udata *udata)
346 struct mlx4_ib_dev *dev = to_mdev(pd->device);
347 struct mlx4_ib_mw *mw;
350 mw = kmalloc(sizeof(*mw), GFP_KERNEL);
352 return ERR_PTR(-ENOMEM);
354 err = mlx4_mw_alloc(dev->dev, to_mpd(pd)->pdn,
355 to_mlx4_type(type), &mw->mmw);
359 err = mlx4_mw_enable(dev->dev, &mw->mmw);
363 mw->ibmw.rkey = mw->mmw.key;
368 mlx4_mw_free(dev->dev, &mw->mmw);
376 int mlx4_ib_dealloc_mw(struct ib_mw *ibmw)
378 struct mlx4_ib_mw *mw = to_mmw(ibmw);
380 mlx4_mw_free(to_mdev(ibmw->device)->dev, &mw->mmw);
386 struct ib_mr *mlx4_ib_alloc_mr(struct ib_pd *pd,
387 enum ib_mr_type mr_type,
390 struct mlx4_ib_dev *dev = to_mdev(pd->device);
391 struct mlx4_ib_mr *mr;
394 if (mr_type != IB_MR_TYPE_MEM_REG ||
395 max_num_sg > MLX4_MAX_FAST_REG_PAGES)
396 return ERR_PTR(-EINVAL);
398 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
400 return ERR_PTR(-ENOMEM);
402 err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, 0, 0, 0,
403 max_num_sg, 0, &mr->mmr);
407 err = mlx4_alloc_priv_pages(pd->device, mr, max_num_sg);
411 mr->max_pages = max_num_sg;
413 err = mlx4_mr_enable(dev->dev, &mr->mmr);
417 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
423 mlx4_free_priv_pages(mr);
425 (void) mlx4_mr_free(dev->dev, &mr->mmr);
431 struct ib_fmr *mlx4_ib_fmr_alloc(struct ib_pd *pd, int acc,
432 struct ib_fmr_attr *fmr_attr)
434 struct mlx4_ib_dev *dev = to_mdev(pd->device);
435 struct mlx4_ib_fmr *fmr;
438 fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
440 return ERR_PTR(-ENOMEM);
442 err = mlx4_fmr_alloc(dev->dev, to_mpd(pd)->pdn, convert_access(acc),
443 fmr_attr->max_pages, fmr_attr->max_maps,
444 fmr_attr->page_shift, &fmr->mfmr);
448 err = mlx4_fmr_enable(to_mdev(pd->device)->dev, &fmr->mfmr);
452 fmr->ibfmr.rkey = fmr->ibfmr.lkey = fmr->mfmr.mr.key;
457 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &fmr->mfmr.mr);
465 int mlx4_ib_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
466 int npages, u64 iova)
468 struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
469 struct mlx4_ib_dev *dev = to_mdev(ifmr->ibfmr.device);
471 return mlx4_map_phys_fmr(dev->dev, &ifmr->mfmr, page_list, npages, iova,
472 &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey);
475 int mlx4_ib_unmap_fmr(struct list_head *fmr_list)
477 struct ib_fmr *ibfmr;
479 struct mlx4_dev *mdev = NULL;
481 list_for_each_entry(ibfmr, fmr_list, list) {
482 if (mdev && to_mdev(ibfmr->device)->dev != mdev)
484 mdev = to_mdev(ibfmr->device)->dev;
490 list_for_each_entry(ibfmr, fmr_list, list) {
491 struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
493 mlx4_fmr_unmap(mdev, &ifmr->mfmr, &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey);
497 * Make sure all MPT status updates are visible before issuing
498 * SYNC_TPT firmware command.
502 err = mlx4_SYNC_TPT(mdev);
504 pr_warn("SYNC_TPT error %d when "
505 "unmapping FMRs\n", err);
510 int mlx4_ib_fmr_dealloc(struct ib_fmr *ibfmr)
512 struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr);
513 struct mlx4_ib_dev *dev = to_mdev(ibfmr->device);
516 err = mlx4_fmr_free(dev->dev, &ifmr->mfmr);
524 static int mlx4_set_page(struct ib_mr *ibmr, u64 addr)
526 struct mlx4_ib_mr *mr = to_mmr(ibmr);
528 if (unlikely(mr->npages == mr->max_pages))
531 mr->pages[mr->npages++] = cpu_to_be64(addr | MLX4_MTT_FLAG_PRESENT);
536 int mlx4_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
537 unsigned int *sg_offset)
539 struct mlx4_ib_mr *mr = to_mmr(ibmr);
544 ib_dma_sync_single_for_cpu(ibmr->device, mr->page_map,
545 mr->page_map_size, DMA_TO_DEVICE);
547 rc = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, mlx4_set_page);
549 ib_dma_sync_single_for_device(ibmr->device, mr->page_map,
550 mr->page_map_size, DMA_TO_DEVICE);