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
50 MLX4_ICM_ALLOC_SIZE
= 1 << 18,
51 MLX4_TABLE_CHUNK_SIZE
= 1 << 18
54 static void mlx4_free_icm_pages(struct mlx4_dev
*dev
, struct mlx4_icm_chunk
*chunk
)
59 pci_unmap_sg(dev
->pdev
, chunk
->mem
, chunk
->npages
,
60 PCI_DMA_BIDIRECTIONAL
);
62 for (i
= 0; i
< chunk
->npages
; ++i
)
63 __free_pages(sg_page(&chunk
->mem
[i
]),
64 get_order(chunk
->mem
[i
].length
));
67 static void mlx4_free_icm_coherent(struct mlx4_dev
*dev
, struct mlx4_icm_chunk
*chunk
)
71 for (i
= 0; i
< chunk
->npages
; ++i
)
72 dma_free_coherent(&dev
->pdev
->dev
, chunk
->mem
[i
].length
,
73 lowmem_page_address(sg_page(&chunk
->mem
[i
])),
74 sg_dma_address(&chunk
->mem
[i
]));
77 void mlx4_free_icm(struct mlx4_dev
*dev
, struct mlx4_icm
*icm
, int coherent
)
79 struct mlx4_icm_chunk
*chunk
, *tmp
;
84 list_for_each_entry_safe(chunk
, tmp
, &icm
->chunk_list
, list
) {
86 mlx4_free_icm_coherent(dev
, chunk
);
88 mlx4_free_icm_pages(dev
, chunk
);
96 static int mlx4_alloc_icm_pages(struct scatterlist
*mem
, int order
,
97 gfp_t gfp_mask
, int node
)
101 page
= alloc_pages_node(node
, gfp_mask
, order
);
103 page
= alloc_pages(gfp_mask
, order
);
108 sg_set_page(mem
, page
, PAGE_SIZE
<< order
, 0);
112 static int mlx4_alloc_icm_coherent(struct device
*dev
, struct scatterlist
*mem
,
113 int order
, gfp_t gfp_mask
)
115 void *buf
= dma_alloc_coherent(dev
, PAGE_SIZE
<< order
,
116 &sg_dma_address(mem
), gfp_mask
);
120 sg_set_buf(mem
, buf
, PAGE_SIZE
<< order
);
122 sg_dma_len(mem
) = PAGE_SIZE
<< order
;
126 struct mlx4_icm
*mlx4_alloc_icm(struct mlx4_dev
*dev
, int npages
,
127 gfp_t gfp_mask
, int coherent
)
129 struct mlx4_icm
*icm
;
130 struct mlx4_icm_chunk
*chunk
= NULL
;
134 /* We use sg_set_buf for coherent allocs, which assumes low memory */
135 BUG_ON(coherent
&& (gfp_mask
& __GFP_HIGHMEM
));
137 icm
= kmalloc_node(sizeof(*icm
),
138 gfp_mask
& ~(__GFP_HIGHMEM
| __GFP_NOWARN
),
141 icm
= kmalloc(sizeof(*icm
),
142 gfp_mask
& ~(__GFP_HIGHMEM
| __GFP_NOWARN
));
148 INIT_LIST_HEAD(&icm
->chunk_list
);
150 cur_order
= get_order(MLX4_ICM_ALLOC_SIZE
);
154 chunk
= kmalloc_node(sizeof(*chunk
),
155 gfp_mask
& ~(__GFP_HIGHMEM
|
159 chunk
= kmalloc(sizeof(*chunk
),
160 gfp_mask
& ~(__GFP_HIGHMEM
|
166 sg_init_table(chunk
->mem
, MLX4_ICM_CHUNK_LEN
);
169 list_add_tail(&chunk
->list
, &icm
->chunk_list
);
172 while (1 << cur_order
> npages
)
176 ret
= mlx4_alloc_icm_coherent(&dev
->pdev
->dev
,
177 &chunk
->mem
[chunk
->npages
],
178 cur_order
, gfp_mask
);
180 ret
= mlx4_alloc_icm_pages(&chunk
->mem
[chunk
->npages
],
195 else if (chunk
->npages
== MLX4_ICM_CHUNK_LEN
) {
196 chunk
->nsg
= pci_map_sg(dev
->pdev
, chunk
->mem
,
198 PCI_DMA_BIDIRECTIONAL
);
204 if (chunk
->npages
== MLX4_ICM_CHUNK_LEN
)
207 npages
-= 1 << cur_order
;
210 if (!coherent
&& chunk
) {
211 chunk
->nsg
= pci_map_sg(dev
->pdev
, chunk
->mem
,
213 PCI_DMA_BIDIRECTIONAL
);
222 mlx4_free_icm(dev
, icm
, coherent
);
226 static int mlx4_MAP_ICM(struct mlx4_dev
*dev
, struct mlx4_icm
*icm
, u64 virt
)
228 return mlx4_map_cmd(dev
, MLX4_CMD_MAP_ICM
, icm
, virt
);
231 static int mlx4_UNMAP_ICM(struct mlx4_dev
*dev
, u64 virt
, u32 page_count
)
233 return mlx4_cmd(dev
, virt
, page_count
, 0, MLX4_CMD_UNMAP_ICM
,
234 MLX4_CMD_TIME_CLASS_B
, MLX4_CMD_NATIVE
);
237 int mlx4_MAP_ICM_AUX(struct mlx4_dev
*dev
, struct mlx4_icm
*icm
)
239 return mlx4_map_cmd(dev
, MLX4_CMD_MAP_ICM_AUX
, icm
, -1);
242 int mlx4_UNMAP_ICM_AUX(struct mlx4_dev
*dev
)
244 return mlx4_cmd(dev
, 0, 0, 0, MLX4_CMD_UNMAP_ICM_AUX
,
245 MLX4_CMD_TIME_CLASS_B
, MLX4_CMD_NATIVE
);
248 int mlx4_table_get(struct mlx4_dev
*dev
, struct mlx4_icm_table
*table
, u32 obj
,
251 u32 i
= (obj
& (table
->num_obj
- 1)) /
252 (MLX4_TABLE_CHUNK_SIZE
/ table
->obj_size
);
255 mutex_lock(&table
->mutex
);
258 ++table
->icm
[i
]->refcount
;
262 table
->icm
[i
] = mlx4_alloc_icm(dev
, MLX4_TABLE_CHUNK_SIZE
>> PAGE_SHIFT
,
263 (table
->lowmem
? gfp
: GFP_HIGHUSER
) |
264 __GFP_NOWARN
, table
->coherent
);
265 if (!table
->icm
[i
]) {
270 if (mlx4_MAP_ICM(dev
, table
->icm
[i
], table
->virt
+
271 (u64
) i
* MLX4_TABLE_CHUNK_SIZE
)) {
272 mlx4_free_icm(dev
, table
->icm
[i
], table
->coherent
);
273 table
->icm
[i
] = NULL
;
278 ++table
->icm
[i
]->refcount
;
281 mutex_unlock(&table
->mutex
);
285 void mlx4_table_put(struct mlx4_dev
*dev
, struct mlx4_icm_table
*table
, u32 obj
)
290 i
= (obj
& (table
->num_obj
- 1)) / (MLX4_TABLE_CHUNK_SIZE
/ table
->obj_size
);
292 mutex_lock(&table
->mutex
);
294 if (--table
->icm
[i
]->refcount
== 0) {
295 offset
= (u64
) i
* MLX4_TABLE_CHUNK_SIZE
;
296 mlx4_UNMAP_ICM(dev
, table
->virt
+ offset
,
297 MLX4_TABLE_CHUNK_SIZE
/ MLX4_ICM_PAGE_SIZE
);
298 mlx4_free_icm(dev
, table
->icm
[i
], table
->coherent
);
299 table
->icm
[i
] = NULL
;
302 mutex_unlock(&table
->mutex
);
305 void *mlx4_table_find(struct mlx4_icm_table
*table
, u32 obj
,
306 dma_addr_t
*dma_handle
)
308 int offset
, dma_offset
, i
;
310 struct mlx4_icm_chunk
*chunk
;
311 struct mlx4_icm
*icm
;
312 struct page
*page
= NULL
;
317 mutex_lock(&table
->mutex
);
319 idx
= (u64
) (obj
& (table
->num_obj
- 1)) * table
->obj_size
;
320 icm
= table
->icm
[idx
/ MLX4_TABLE_CHUNK_SIZE
];
321 dma_offset
= offset
= idx
% MLX4_TABLE_CHUNK_SIZE
;
326 list_for_each_entry(chunk
, &icm
->chunk_list
, list
) {
327 for (i
= 0; i
< chunk
->npages
; ++i
) {
328 if (dma_handle
&& dma_offset
>= 0) {
329 if (sg_dma_len(&chunk
->mem
[i
]) > dma_offset
)
330 *dma_handle
= sg_dma_address(&chunk
->mem
[i
]) +
332 dma_offset
-= sg_dma_len(&chunk
->mem
[i
]);
335 * DMA mapping can merge pages but not split them,
336 * so if we found the page, dma_handle has already
339 if (chunk
->mem
[i
].length
> offset
) {
340 page
= sg_page(&chunk
->mem
[i
]);
343 offset
-= chunk
->mem
[i
].length
;
348 mutex_unlock(&table
->mutex
);
349 return page
? lowmem_page_address(page
) + offset
: NULL
;
352 int mlx4_table_get_range(struct mlx4_dev
*dev
, struct mlx4_icm_table
*table
,
355 int inc
= MLX4_TABLE_CHUNK_SIZE
/ table
->obj_size
;
359 for (i
= start
; i
<= end
; i
+= inc
) {
360 err
= mlx4_table_get(dev
, table
, i
, GFP_KERNEL
);
370 mlx4_table_put(dev
, table
, i
);
376 void mlx4_table_put_range(struct mlx4_dev
*dev
, struct mlx4_icm_table
*table
,
381 for (i
= start
; i
<= end
; i
+= MLX4_TABLE_CHUNK_SIZE
/ table
->obj_size
)
382 mlx4_table_put(dev
, table
, i
);
385 int mlx4_init_icm_table(struct mlx4_dev
*dev
, struct mlx4_icm_table
*table
,
386 u64 virt
, int obj_size
, u32 nobj
, int reserved
,
387 int use_lowmem
, int use_coherent
)
395 obj_per_chunk
= MLX4_TABLE_CHUNK_SIZE
/ obj_size
;
396 num_icm
= (nobj
+ obj_per_chunk
- 1) / obj_per_chunk
;
398 table
->icm
= kcalloc(num_icm
, sizeof *table
->icm
, GFP_KERNEL
);
402 table
->num_icm
= num_icm
;
403 table
->num_obj
= nobj
;
404 table
->obj_size
= obj_size
;
405 table
->lowmem
= use_lowmem
;
406 table
->coherent
= use_coherent
;
407 mutex_init(&table
->mutex
);
409 size
= (u64
) nobj
* obj_size
;
410 for (i
= 0; i
* MLX4_TABLE_CHUNK_SIZE
< reserved
* obj_size
; ++i
) {
411 chunk_size
= MLX4_TABLE_CHUNK_SIZE
;
412 if ((i
+ 1) * MLX4_TABLE_CHUNK_SIZE
> size
)
413 chunk_size
= PAGE_ALIGN(size
-
414 i
* MLX4_TABLE_CHUNK_SIZE
);
416 table
->icm
[i
] = mlx4_alloc_icm(dev
, chunk_size
>> PAGE_SHIFT
,
417 (use_lowmem
? GFP_KERNEL
: GFP_HIGHUSER
) |
418 __GFP_NOWARN
, use_coherent
);
421 if (mlx4_MAP_ICM(dev
, table
->icm
[i
], virt
+ i
* MLX4_TABLE_CHUNK_SIZE
)) {
422 mlx4_free_icm(dev
, table
->icm
[i
], use_coherent
);
423 table
->icm
[i
] = NULL
;
428 * Add a reference to this ICM chunk so that it never
429 * gets freed (since it contains reserved firmware objects).
431 ++table
->icm
[i
]->refcount
;
437 for (i
= 0; i
< num_icm
; ++i
)
439 mlx4_UNMAP_ICM(dev
, virt
+ i
* MLX4_TABLE_CHUNK_SIZE
,
440 MLX4_TABLE_CHUNK_SIZE
/ MLX4_ICM_PAGE_SIZE
);
441 mlx4_free_icm(dev
, table
->icm
[i
], use_coherent
);
449 void mlx4_cleanup_icm_table(struct mlx4_dev
*dev
, struct mlx4_icm_table
*table
)
453 for (i
= 0; i
< table
->num_icm
; ++i
)
455 mlx4_UNMAP_ICM(dev
, table
->virt
+ i
* MLX4_TABLE_CHUNK_SIZE
,
456 MLX4_TABLE_CHUNK_SIZE
/ MLX4_ICM_PAGE_SIZE
);
457 mlx4_free_icm(dev
, table
->icm
[i
], table
->coherent
);
This page took 0.051147 seconds and 5 git commands to generate.