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 <rdma/ib_user_verbs.h>
39 static u32 convert_access(int acc)
41 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX4_PERM_ATOMIC : 0) |
42 (acc & IB_ACCESS_REMOTE_WRITE ? MLX4_PERM_REMOTE_WRITE : 0) |
43 (acc & IB_ACCESS_REMOTE_READ ? MLX4_PERM_REMOTE_READ : 0) |
44 (acc & IB_ACCESS_LOCAL_WRITE ? MLX4_PERM_LOCAL_WRITE : 0) |
45 (acc & IB_ACCESS_MW_BIND ? MLX4_PERM_BIND_MW : 0) |
49 static enum mlx4_mw_type to_mlx4_type(enum ib_mw_type type)
52 case IB_MW_TYPE_1: return MLX4_MW_TYPE_1;
53 case IB_MW_TYPE_2: return MLX4_MW_TYPE_2;
58 struct ib_mr *mlx4_ib_get_dma_mr(struct ib_pd *pd, int acc)
60 struct mlx4_ib_mr *mr;
63 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
65 return ERR_PTR(-ENOMEM);
67 err = mlx4_mr_alloc(to_mdev(pd->device)->dev, to_mpd(pd)->pdn, 0,
68 ~0ull, convert_access(acc), 0, 0, &mr->mmr);
72 err = mlx4_mr_enable(to_mdev(pd->device)->dev, &mr->mmr);
76 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
82 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
91 MLX4_MAX_MTT_SHIFT = 31
94 static int mlx4_ib_umem_write_mtt_block(struct mlx4_ib_dev *dev,
96 u64 mtt_size, u64 mtt_shift, u64 len,
97 u64 cur_start_addr, u64 *pages,
98 int *start_index, int *npages)
100 u64 cur_end_addr = cur_start_addr + len;
101 u64 cur_end_addr_aligned = 0;
106 len += (cur_start_addr & (mtt_size - 1ULL));
107 cur_end_addr_aligned = round_up(cur_end_addr, mtt_size);
108 len += (cur_end_addr_aligned - cur_end_addr);
109 if (len & (mtt_size - 1ULL)) {
110 pr_warn("write_block: len %llx is not aligned to mtt_size %llx\n",
115 mtt_entries = (len >> mtt_shift);
118 * Align the MTT start address to the mtt_size.
119 * Required to handle cases when the MR starts in the middle of an MTT
120 * record. Was not required in old code since the physical addresses
121 * provided by the dma subsystem were page aligned, which was also the
124 cur_start_addr = round_down(cur_start_addr, mtt_size);
125 /* A new block is started ... */
126 for (k = 0; k < mtt_entries; ++k) {
127 pages[*npages] = cur_start_addr + (mtt_size * k);
130 * Be friendly to mlx4_write_mtt() and pass it chunks of
133 if (*npages == PAGE_SIZE / sizeof(u64)) {
134 err = mlx4_write_mtt(dev->dev, mtt, *start_index,
139 (*start_index) += *npages;
147 static inline u64 alignment_of(u64 ptr)
149 return ilog2(ptr & (~(ptr - 1)));
152 static int mlx4_ib_umem_calc_block_mtt(u64 next_block_start,
153 u64 current_block_end,
156 /* Check whether the alignment of the new block is aligned as well as
157 * the previous block.
158 * Block address must start with zeros till size of entity_size.
160 if ((next_block_start & ((1ULL << block_shift) - 1ULL)) != 0)
162 * It is not as well aligned as the previous block-reduce the
163 * mtt size accordingly. Here we take the last right bit which
166 block_shift = alignment_of(next_block_start);
169 * Check whether the alignment of the end of previous block - is it
170 * aligned as well as the start of the block
172 if (((current_block_end) & ((1ULL << block_shift) - 1ULL)) != 0)
174 * It is not as well aligned as the start of the block -
175 * reduce the mtt size accordingly.
177 block_shift = alignment_of(current_block_end);
182 int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt,
183 struct ib_umem *umem)
189 u64 cur_start_addr = 0;
193 struct scatterlist *sg;
196 pages = (u64 *) __get_free_page(GFP_KERNEL);
200 mtt_shift = mtt->page_shift;
201 mtt_size = 1ULL << mtt_shift;
203 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
204 if (cur_start_addr + len == sg_dma_address(sg)) {
205 /* still the same block */
206 len += sg_dma_len(sg);
210 * A new block is started ...
211 * If len is malaligned, write an extra mtt entry to cover the
212 * misaligned area (round up the division)
214 err = mlx4_ib_umem_write_mtt_block(dev, mtt, mtt_size,
222 cur_start_addr = sg_dma_address(sg);
223 len = sg_dma_len(sg);
226 /* Handle the last block */
229 * If len is malaligned, write an extra mtt entry to cover
230 * the misaligned area (round up the division)
232 err = mlx4_ib_umem_write_mtt_block(dev, mtt, mtt_size,
234 cur_start_addr, pages,
235 &start_index, &npages);
241 err = mlx4_write_mtt(dev->dev, mtt, start_index, npages, pages);
244 free_page((unsigned long) pages);
249 * Calculate optimal mtt size based on contiguous pages.
250 * Function will return also the number of pages that are not aligned to the
251 * calculated mtt_size to be added to total number of pages. For that we should
252 * check the first chunk length & last chunk length and if not aligned to
253 * mtt_size we should increment the non_aligned_pages number. All chunks in the
254 * middle already handled as part of mtt shift calculation for both their start
257 int mlx4_ib_umem_calc_optimal_mtt_size(struct ib_umem *umem, u64 start_va,
260 u64 block_shift = MLX4_MAX_MTT_SHIFT;
261 u64 min_shift = PAGE_SHIFT;
262 u64 last_block_aligned_end = 0;
263 u64 current_block_start = 0;
264 u64 first_block_start = 0;
265 u64 current_block_len = 0;
266 u64 last_block_end = 0;
267 struct scatterlist *sg;
268 u64 current_block_end;
269 u64 misalignment_bits;
270 u64 next_block_start;
274 *num_of_mtts = ib_umem_num_dma_blocks(umem, PAGE_SIZE);
276 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
278 * Initialization - save the first chunk start as the
279 * current_block_start - block means contiguous pages.
281 if (current_block_len == 0 && current_block_start == 0) {
282 current_block_start = sg_dma_address(sg);
283 first_block_start = current_block_start;
285 * Find the bits that are different between the physical
286 * address and the virtual address for the start of the
288 * umem_get aligned the start_va to a page boundary.
289 * Therefore, we need to align the start va to the same
291 * misalignment_bits is needed to handle the case of a
292 * single memory region. In this case, the rest of the
293 * logic will not reduce the block size. If we use a
294 * block size which is bigger than the alignment of the
295 * misalignment bits, we might use the virtual page
296 * number instead of the physical page number, resulting
297 * in access to the wrong data.
300 (start_va & (~(((u64)(PAGE_SIZE)) - 1ULL))) ^
302 block_shift = min(alignment_of(misalignment_bits),
307 * Go over the scatter entries and check if they continue the
308 * previous scatter entry.
310 next_block_start = sg_dma_address(sg);
311 current_block_end = current_block_start + current_block_len;
312 /* If we have a split (non-contig.) between two blocks */
313 if (current_block_end != next_block_start) {
314 block_shift = mlx4_ib_umem_calc_block_mtt
320 * If we reached the minimum shift for 4k page we stop
323 if (block_shift <= min_shift)
327 * If not saved yet we are in first block - we save the
328 * length of first block to calculate the
329 * non_aligned_pages number at the end.
331 total_len += current_block_len;
333 /* Start a new block */
334 current_block_start = next_block_start;
335 current_block_len = sg_dma_len(sg);
338 /* The scatter entry is another part of the current block,
339 * increase the block size.
340 * An entry in the scatter can be larger than 4k (page) as of
341 * dma mapping which merge some blocks together.
343 current_block_len += sg_dma_len(sg);
346 /* Account for the last block in the total len */
347 total_len += current_block_len;
348 /* Add to the first block the misalignment that it suffers from. */
349 total_len += (first_block_start & ((1ULL << block_shift) - 1ULL));
350 last_block_end = current_block_start + current_block_len;
351 last_block_aligned_end = round_up(last_block_end, 1ULL << block_shift);
352 total_len += (last_block_aligned_end - last_block_end);
354 if (total_len & ((1ULL << block_shift) - 1ULL))
355 pr_warn("misaligned total length detected (%llu, %llu)!",
356 total_len, block_shift);
358 *num_of_mtts = total_len >> block_shift;
360 if (block_shift < min_shift) {
362 * If shift is less than the min we set a warning and return the
365 pr_warn("umem_calc_optimal_mtt_size - unexpected shift %lld\n", block_shift);
367 block_shift = min_shift;
372 static struct ib_umem *mlx4_get_umem_mr(struct ib_device *device, u64 start,
373 u64 length, int access_flags)
376 * Force registering the memory as writable if the underlying pages
377 * are writable. This is so rereg can change the access permissions
378 * from readable to writable without having to run through ib_umem_get
381 if (!ib_access_writable(access_flags)) {
382 unsigned long untagged_start = untagged_addr(start);
383 struct vm_area_struct *vma;
385 mmap_read_lock(current->mm);
387 * FIXME: Ideally this would iterate over all the vmas that
388 * cover the memory, but for now it requires a single vma to
389 * entirely cover the MR to support RO mappings.
391 vma = find_vma(current->mm, untagged_start);
392 if (vma && vma->vm_end >= untagged_start + length &&
393 vma->vm_start <= untagged_start) {
394 if (vma->vm_flags & VM_WRITE)
395 access_flags |= IB_ACCESS_LOCAL_WRITE;
397 access_flags |= IB_ACCESS_LOCAL_WRITE;
400 mmap_read_unlock(current->mm);
403 return ib_umem_get(device, start, length, access_flags);
406 struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
407 u64 virt_addr, int access_flags,
408 struct ib_udata *udata)
410 struct mlx4_ib_dev *dev = to_mdev(pd->device);
411 struct mlx4_ib_mr *mr;
416 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
418 return ERR_PTR(-ENOMEM);
420 mr->umem = mlx4_get_umem_mr(pd->device, start, length, access_flags);
421 if (IS_ERR(mr->umem)) {
422 err = PTR_ERR(mr->umem);
426 shift = mlx4_ib_umem_calc_optimal_mtt_size(mr->umem, start, &n);
428 err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length,
429 convert_access(access_flags), n, shift, &mr->mmr);
433 err = mlx4_ib_umem_write_mtt(dev, &mr->mmr.mtt, mr->umem);
437 err = mlx4_mr_enable(dev->dev, &mr->mmr);
441 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
442 mr->ibmr.page_size = 1U << shift;
447 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
450 ib_umem_release(mr->umem);
458 int mlx4_ib_rereg_user_mr(struct ib_mr *mr, int flags,
459 u64 start, u64 length, u64 virt_addr,
460 int mr_access_flags, struct ib_pd *pd,
461 struct ib_udata *udata)
463 struct mlx4_ib_dev *dev = to_mdev(mr->device);
464 struct mlx4_ib_mr *mmr = to_mmr(mr);
465 struct mlx4_mpt_entry *mpt_entry;
466 struct mlx4_mpt_entry **pmpt_entry = &mpt_entry;
469 /* Since we synchronize this call and mlx4_ib_dereg_mr via uverbs,
470 * we assume that the calls can't run concurrently. Otherwise, a
473 err = mlx4_mr_hw_get_mpt(dev->dev, &mmr->mmr, &pmpt_entry);
478 if (flags & IB_MR_REREG_PD) {
479 err = mlx4_mr_hw_change_pd(dev->dev, *pmpt_entry,
483 goto release_mpt_entry;
486 if (flags & IB_MR_REREG_ACCESS) {
487 if (ib_access_writable(mr_access_flags) &&
488 !mmr->umem->writable) {
490 goto release_mpt_entry;
493 err = mlx4_mr_hw_change_access(dev->dev, *pmpt_entry,
494 convert_access(mr_access_flags));
497 goto release_mpt_entry;
500 if (flags & IB_MR_REREG_TRANS) {
504 mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
505 ib_umem_release(mmr->umem);
506 mmr->umem = mlx4_get_umem_mr(mr->device, start, length,
508 if (IS_ERR(mmr->umem)) {
509 err = PTR_ERR(mmr->umem);
510 /* Prevent mlx4_ib_dereg_mr from free'ing invalid pointer */
512 goto release_mpt_entry;
514 n = ib_umem_num_dma_blocks(mmr->umem, PAGE_SIZE);
517 err = mlx4_mr_rereg_mem_write(dev->dev, &mmr->mmr,
518 virt_addr, length, n, shift,
521 ib_umem_release(mmr->umem);
522 goto release_mpt_entry;
524 mmr->mmr.iova = virt_addr;
525 mmr->mmr.size = length;
527 err = mlx4_ib_umem_write_mtt(dev, &mmr->mmr.mtt, mmr->umem);
529 mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
530 ib_umem_release(mmr->umem);
531 goto release_mpt_entry;
535 /* If we couldn't transfer the MR to the HCA, just remember to
536 * return a failure. But dereg_mr will free the resources.
538 err = mlx4_mr_hw_write_mpt(dev->dev, &mmr->mmr, pmpt_entry);
539 if (!err && flags & IB_MR_REREG_ACCESS)
540 mmr->mmr.access = mr_access_flags;
543 mlx4_mr_hw_put_mpt(dev->dev, pmpt_entry);
549 mlx4_alloc_priv_pages(struct ib_device *device,
550 struct mlx4_ib_mr *mr,
555 /* Ensure that size is aligned to DMA cacheline
557 * max_pages is limited to MLX4_MAX_FAST_REG_PAGES
558 * so page_map_size will never cross PAGE_SIZE.
560 mr->page_map_size = roundup(max_pages * sizeof(u64),
561 MLX4_MR_PAGES_ALIGN);
563 /* Prevent cross page boundary allocation. */
564 mr->pages = (__be64 *)get_zeroed_page(GFP_KERNEL);
568 mr->page_map = dma_map_single(device->dev.parent, mr->pages,
569 mr->page_map_size, DMA_TO_DEVICE);
571 if (dma_mapping_error(device->dev.parent, mr->page_map)) {
579 free_page((unsigned long)mr->pages);
584 mlx4_free_priv_pages(struct mlx4_ib_mr *mr)
587 struct ib_device *device = mr->ibmr.device;
589 dma_unmap_single(device->dev.parent, mr->page_map,
590 mr->page_map_size, DMA_TO_DEVICE);
591 free_page((unsigned long)mr->pages);
596 int mlx4_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
598 struct mlx4_ib_mr *mr = to_mmr(ibmr);
601 mlx4_free_priv_pages(mr);
603 ret = mlx4_mr_free(to_mdev(ibmr->device)->dev, &mr->mmr);
607 ib_umem_release(mr->umem);
613 int mlx4_ib_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
615 struct mlx4_ib_dev *dev = to_mdev(ibmw->device);
616 struct mlx4_ib_mw *mw = to_mmw(ibmw);
619 err = mlx4_mw_alloc(dev->dev, to_mpd(ibmw->pd)->pdn,
620 to_mlx4_type(ibmw->type), &mw->mmw);
624 err = mlx4_mw_enable(dev->dev, &mw->mmw);
628 ibmw->rkey = mw->mmw.key;
632 mlx4_mw_free(dev->dev, &mw->mmw);
636 int mlx4_ib_dealloc_mw(struct ib_mw *ibmw)
638 struct mlx4_ib_mw *mw = to_mmw(ibmw);
640 mlx4_mw_free(to_mdev(ibmw->device)->dev, &mw->mmw);
644 struct ib_mr *mlx4_ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
647 struct mlx4_ib_dev *dev = to_mdev(pd->device);
648 struct mlx4_ib_mr *mr;
651 if (mr_type != IB_MR_TYPE_MEM_REG ||
652 max_num_sg > MLX4_MAX_FAST_REG_PAGES)
653 return ERR_PTR(-EINVAL);
655 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
657 return ERR_PTR(-ENOMEM);
659 err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, 0, 0, 0,
660 max_num_sg, 0, &mr->mmr);
664 err = mlx4_alloc_priv_pages(pd->device, mr, max_num_sg);
668 mr->max_pages = max_num_sg;
669 err = mlx4_mr_enable(dev->dev, &mr->mmr);
673 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key;
679 mr->ibmr.device = pd->device;
680 mlx4_free_priv_pages(mr);
682 (void) mlx4_mr_free(dev->dev, &mr->mmr);
688 static int mlx4_set_page(struct ib_mr *ibmr, u64 addr)
690 struct mlx4_ib_mr *mr = to_mmr(ibmr);
692 if (unlikely(mr->npages == mr->max_pages))
695 mr->pages[mr->npages++] = cpu_to_be64(addr | MLX4_MTT_FLAG_PRESENT);
700 int mlx4_ib_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
701 unsigned int *sg_offset)
703 struct mlx4_ib_mr *mr = to_mmr(ibmr);
708 ib_dma_sync_single_for_cpu(ibmr->device, mr->page_map,
709 mr->page_map_size, DMA_TO_DEVICE);
711 rc = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, mlx4_set_page);
713 ib_dma_sync_single_for_device(ibmr->device, mr->page_map,
714 mr->page_map_size, DMA_TO_DEVICE);