2 * Copyright (c) 2005, 2006, 2007, 2008 Mellanox Technologies. All rights reserved.
3 * Copyright (c) 2006, 2007 Cisco Systems, Inc. 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/errno.h>
36 #include <linux/scatterlist.h>
37 #include <linux/slab.h>
39 #include <linux/mlx4/cmd.h>
46 * We allocate in as big chunks as we can, up to a maximum of 256 KB
47 * per chunk. Note that the chunks are not necessarily in contiguous
51 MLX4_ICM_ALLOC_SIZE = 1 << 18,
52 MLX4_TABLE_CHUNK_SIZE = 1 << 18,
55 static void mlx4_free_icm_pages(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk)
60 pci_unmap_sg(dev->persist->pdev, chunk->mem, chunk->npages,
61 PCI_DMA_BIDIRECTIONAL);
63 for (i = 0; i < chunk->npages; ++i)
64 __free_pages(sg_page(&chunk->mem[i]),
65 get_order(chunk->mem[i].length));
68 static void mlx4_free_icm_coherent(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk)
72 for (i = 0; i < chunk->npages; ++i)
73 dma_free_coherent(&dev->persist->pdev->dev,
75 lowmem_page_address(sg_page(&chunk->mem[i])),
76 sg_dma_address(&chunk->mem[i]));
79 void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm, int coherent)
81 struct mlx4_icm_chunk *chunk, *tmp;
86 list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
88 mlx4_free_icm_coherent(dev, chunk);
90 mlx4_free_icm_pages(dev, chunk);
98 static int mlx4_alloc_icm_pages(struct scatterlist *mem, int order,
99 gfp_t gfp_mask, int node)
103 page = alloc_pages_node(node, gfp_mask, order);
105 page = alloc_pages(gfp_mask, order);
110 sg_set_page(mem, page, PAGE_SIZE << order, 0);
114 static int mlx4_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
115 int order, gfp_t gfp_mask)
117 void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order,
118 &sg_dma_address(mem), gfp_mask);
122 if (offset_in_page(buf)) {
123 dma_free_coherent(dev, PAGE_SIZE << order,
124 buf, sg_dma_address(mem));
128 sg_set_buf(mem, buf, PAGE_SIZE << order);
129 sg_dma_len(mem) = PAGE_SIZE << order;
133 struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
134 gfp_t gfp_mask, int coherent)
136 struct mlx4_icm *icm;
137 struct mlx4_icm_chunk *chunk = NULL;
142 /* We use sg_set_buf for coherent allocs, which assumes low memory */
143 BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
145 icm = kmalloc_node(sizeof(*icm),
146 gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN),
149 icm = kmalloc(sizeof(*icm),
150 gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
156 INIT_LIST_HEAD(&icm->chunk_list);
158 cur_order = get_order(MLX4_ICM_ALLOC_SIZE);
162 chunk = kmalloc_node(sizeof(*chunk),
163 gfp_mask & ~(__GFP_HIGHMEM |
167 chunk = kmalloc(sizeof(*chunk),
168 gfp_mask & ~(__GFP_HIGHMEM |
174 sg_init_table(chunk->mem, MLX4_ICM_CHUNK_LEN);
177 list_add_tail(&chunk->list, &icm->chunk_list);
180 while (1 << cur_order > npages)
185 mask &= ~__GFP_DIRECT_RECLAIM;
188 ret = mlx4_alloc_icm_coherent(&dev->persist->pdev->dev,
189 &chunk->mem[chunk->npages],
192 ret = mlx4_alloc_icm_pages(&chunk->mem[chunk->npages],
207 else if (chunk->npages == MLX4_ICM_CHUNK_LEN) {
208 chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->mem,
210 PCI_DMA_BIDIRECTIONAL);
216 if (chunk->npages == MLX4_ICM_CHUNK_LEN)
219 npages -= 1 << cur_order;
222 if (!coherent && chunk) {
223 chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->mem,
225 PCI_DMA_BIDIRECTIONAL);
234 mlx4_free_icm(dev, icm, coherent);
238 static int mlx4_MAP_ICM(struct mlx4_dev *dev, struct mlx4_icm *icm, u64 virt)
240 return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM, icm, virt);
243 static int mlx4_UNMAP_ICM(struct mlx4_dev *dev, u64 virt, u32 page_count)
245 return mlx4_cmd(dev, virt, page_count, 0, MLX4_CMD_UNMAP_ICM,
246 MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE);
249 int mlx4_MAP_ICM_AUX(struct mlx4_dev *dev, struct mlx4_icm *icm)
251 return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM_AUX, icm, -1);
254 int mlx4_UNMAP_ICM_AUX(struct mlx4_dev *dev)
256 return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_UNMAP_ICM_AUX,
257 MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE);
260 int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, u32 obj)
262 u32 i = (obj & (table->num_obj - 1)) /
263 (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
266 mutex_lock(&table->mutex);
269 ++table->icm[i]->refcount;
273 table->icm[i] = mlx4_alloc_icm(dev, MLX4_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
274 (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
275 __GFP_NOWARN, table->coherent);
276 if (!table->icm[i]) {
281 if (mlx4_MAP_ICM(dev, table->icm[i], table->virt +
282 (u64) i * MLX4_TABLE_CHUNK_SIZE)) {
283 mlx4_free_icm(dev, table->icm[i], table->coherent);
284 table->icm[i] = NULL;
289 ++table->icm[i]->refcount;
292 mutex_unlock(&table->mutex);
296 void mlx4_table_put(struct mlx4_dev *dev, struct mlx4_icm_table *table, u32 obj)
301 i = (obj & (table->num_obj - 1)) / (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
303 mutex_lock(&table->mutex);
305 if (--table->icm[i]->refcount == 0) {
306 offset = (u64) i * MLX4_TABLE_CHUNK_SIZE;
307 mlx4_UNMAP_ICM(dev, table->virt + offset,
308 MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
309 mlx4_free_icm(dev, table->icm[i], table->coherent);
310 table->icm[i] = NULL;
313 mutex_unlock(&table->mutex);
316 void *mlx4_table_find(struct mlx4_icm_table *table, u32 obj,
317 dma_addr_t *dma_handle)
319 int offset, dma_offset, i;
321 struct mlx4_icm_chunk *chunk;
322 struct mlx4_icm *icm;
323 struct page *page = NULL;
328 mutex_lock(&table->mutex);
330 idx = (u64) (obj & (table->num_obj - 1)) * table->obj_size;
331 icm = table->icm[idx / MLX4_TABLE_CHUNK_SIZE];
332 dma_offset = offset = idx % MLX4_TABLE_CHUNK_SIZE;
337 list_for_each_entry(chunk, &icm->chunk_list, list) {
338 for (i = 0; i < chunk->npages; ++i) {
339 if (dma_handle && dma_offset >= 0) {
340 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
341 *dma_handle = sg_dma_address(&chunk->mem[i]) +
343 dma_offset -= sg_dma_len(&chunk->mem[i]);
346 * DMA mapping can merge pages but not split them,
347 * so if we found the page, dma_handle has already
350 if (chunk->mem[i].length > offset) {
351 page = sg_page(&chunk->mem[i]);
354 offset -= chunk->mem[i].length;
359 mutex_unlock(&table->mutex);
360 return page ? lowmem_page_address(page) + offset : NULL;
363 int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
366 int inc = MLX4_TABLE_CHUNK_SIZE / table->obj_size;
370 for (i = start; i <= end; i += inc) {
371 err = mlx4_table_get(dev, table, i);
381 mlx4_table_put(dev, table, i);
387 void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
392 for (i = start; i <= end; i += MLX4_TABLE_CHUNK_SIZE / table->obj_size)
393 mlx4_table_put(dev, table, i);
396 int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
397 u64 virt, int obj_size, u32 nobj, int reserved,
398 int use_lowmem, int use_coherent)
406 obj_per_chunk = MLX4_TABLE_CHUNK_SIZE / obj_size;
407 if (WARN_ON(!obj_per_chunk))
409 num_icm = DIV_ROUND_UP(nobj, obj_per_chunk);
411 table->icm = kvcalloc(num_icm, sizeof(*table->icm), GFP_KERNEL);
415 table->num_icm = num_icm;
416 table->num_obj = nobj;
417 table->obj_size = obj_size;
418 table->lowmem = use_lowmem;
419 table->coherent = use_coherent;
420 mutex_init(&table->mutex);
422 size = (u64) nobj * obj_size;
423 for (i = 0; i * MLX4_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
424 chunk_size = MLX4_TABLE_CHUNK_SIZE;
425 if ((i + 1) * MLX4_TABLE_CHUNK_SIZE > size)
426 chunk_size = PAGE_ALIGN(size -
427 i * MLX4_TABLE_CHUNK_SIZE);
429 table->icm[i] = mlx4_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
430 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
431 __GFP_NOWARN, use_coherent);
434 if (mlx4_MAP_ICM(dev, table->icm[i], virt + i * MLX4_TABLE_CHUNK_SIZE)) {
435 mlx4_free_icm(dev, table->icm[i], use_coherent);
436 table->icm[i] = NULL;
441 * Add a reference to this ICM chunk so that it never
442 * gets freed (since it contains reserved firmware objects).
444 ++table->icm[i]->refcount;
450 for (i = 0; i < num_icm; ++i)
452 mlx4_UNMAP_ICM(dev, virt + i * MLX4_TABLE_CHUNK_SIZE,
453 MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
454 mlx4_free_icm(dev, table->icm[i], use_coherent);
462 void mlx4_cleanup_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table)
466 for (i = 0; i < table->num_icm; ++i)
468 mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
469 MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
470 mlx4_free_icm(dev, table->icm[i], table->coherent);