Pull d-states into release branch
[linux-2.6] / drivers / net / mlx4 / icm.c
1 /*
2  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
3  * Copyright (c) 2006, 2007 Cisco Systems, Inc.  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/init.h>
35 #include <linux/errno.h>
36 #include <linux/mm.h>
37
38 #include <linux/mlx4/cmd.h>
39
40 #include "mlx4.h"
41 #include "icm.h"
42 #include "fw.h"
43
44 /*
45  * We allocate in as big chunks as we can, up to a maximum of 256 KB
46  * per chunk.
47  */
48 enum {
49         MLX4_ICM_ALLOC_SIZE     = 1 << 18,
50         MLX4_TABLE_CHUNK_SIZE   = 1 << 18
51 };
52
53 void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm)
54 {
55         struct mlx4_icm_chunk *chunk, *tmp;
56         int i;
57
58         list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
59                 if (chunk->nsg > 0)
60                         pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
61                                      PCI_DMA_BIDIRECTIONAL);
62
63                 for (i = 0; i < chunk->npages; ++i)
64                         __free_pages(chunk->mem[i].page,
65                                      get_order(chunk->mem[i].length));
66
67                 kfree(chunk);
68         }
69
70         kfree(icm);
71 }
72
73 struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
74                                 gfp_t gfp_mask)
75 {
76         struct mlx4_icm *icm;
77         struct mlx4_icm_chunk *chunk = NULL;
78         int cur_order;
79
80         icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
81         if (!icm)
82                 return icm;
83
84         icm->refcount = 0;
85         INIT_LIST_HEAD(&icm->chunk_list);
86
87         cur_order = get_order(MLX4_ICM_ALLOC_SIZE);
88
89         while (npages > 0) {
90                 if (!chunk) {
91                         chunk = kmalloc(sizeof *chunk,
92                                         gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
93                         if (!chunk)
94                                 goto fail;
95
96                         chunk->npages = 0;
97                         chunk->nsg    = 0;
98                         list_add_tail(&chunk->list, &icm->chunk_list);
99                 }
100
101                 while (1 << cur_order > npages)
102                         --cur_order;
103
104                 chunk->mem[chunk->npages].page = alloc_pages(gfp_mask, cur_order);
105                 if (chunk->mem[chunk->npages].page) {
106                         chunk->mem[chunk->npages].length = PAGE_SIZE << cur_order;
107                         chunk->mem[chunk->npages].offset = 0;
108
109                         if (++chunk->npages == MLX4_ICM_CHUNK_LEN) {
110                                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
111                                                         chunk->npages,
112                                                         PCI_DMA_BIDIRECTIONAL);
113
114                                 if (chunk->nsg <= 0)
115                                         goto fail;
116
117                                 chunk = NULL;
118                         }
119
120                         npages -= 1 << cur_order;
121                 } else {
122                         --cur_order;
123                         if (cur_order < 0)
124                                 goto fail;
125                 }
126         }
127
128         if (chunk) {
129                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
130                                         chunk->npages,
131                                         PCI_DMA_BIDIRECTIONAL);
132
133                 if (chunk->nsg <= 0)
134                         goto fail;
135         }
136
137         return icm;
138
139 fail:
140         mlx4_free_icm(dev, icm);
141         return NULL;
142 }
143
144 static int mlx4_MAP_ICM(struct mlx4_dev *dev, struct mlx4_icm *icm, u64 virt)
145 {
146         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM, icm, virt);
147 }
148
149 int mlx4_UNMAP_ICM(struct mlx4_dev *dev, u64 virt, u32 page_count)
150 {
151         return mlx4_cmd(dev, virt, page_count, 0, MLX4_CMD_UNMAP_ICM,
152                         MLX4_CMD_TIME_CLASS_B);
153 }
154
155 int mlx4_MAP_ICM_page(struct mlx4_dev *dev, u64 dma_addr, u64 virt)
156 {
157         struct mlx4_cmd_mailbox *mailbox;
158         __be64 *inbox;
159         int err;
160
161         mailbox = mlx4_alloc_cmd_mailbox(dev);
162         if (IS_ERR(mailbox))
163                 return PTR_ERR(mailbox);
164         inbox = mailbox->buf;
165
166         inbox[0] = cpu_to_be64(virt);
167         inbox[1] = cpu_to_be64(dma_addr);
168
169         err = mlx4_cmd(dev, mailbox->dma, 1, 0, MLX4_CMD_MAP_ICM,
170                        MLX4_CMD_TIME_CLASS_B);
171
172         mlx4_free_cmd_mailbox(dev, mailbox);
173
174         if (!err)
175                 mlx4_dbg(dev, "Mapped page at %llx to %llx for ICM.\n",
176                           (unsigned long long) dma_addr, (unsigned long long) virt);
177
178         return err;
179 }
180
181 int mlx4_MAP_ICM_AUX(struct mlx4_dev *dev, struct mlx4_icm *icm)
182 {
183         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM_AUX, icm, -1);
184 }
185
186 int mlx4_UNMAP_ICM_AUX(struct mlx4_dev *dev)
187 {
188         return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_UNMAP_ICM_AUX, MLX4_CMD_TIME_CLASS_B);
189 }
190
191 int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj)
192 {
193         int i = (obj & (table->num_obj - 1)) / (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
194         int ret = 0;
195
196         mutex_lock(&table->mutex);
197
198         if (table->icm[i]) {
199                 ++table->icm[i]->refcount;
200                 goto out;
201         }
202
203         table->icm[i] = mlx4_alloc_icm(dev, MLX4_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
204                                        (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
205                                        __GFP_NOWARN);
206         if (!table->icm[i]) {
207                 ret = -ENOMEM;
208                 goto out;
209         }
210
211         if (mlx4_MAP_ICM(dev, table->icm[i], table->virt +
212                          (u64) i * MLX4_TABLE_CHUNK_SIZE)) {
213                 mlx4_free_icm(dev, table->icm[i]);
214                 table->icm[i] = NULL;
215                 ret = -ENOMEM;
216                 goto out;
217         }
218
219         ++table->icm[i]->refcount;
220
221 out:
222         mutex_unlock(&table->mutex);
223         return ret;
224 }
225
226 void mlx4_table_put(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj)
227 {
228         int i;
229
230         i = (obj & (table->num_obj - 1)) / (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
231
232         mutex_lock(&table->mutex);
233
234         if (--table->icm[i]->refcount == 0) {
235                 mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
236                                MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
237                 mlx4_free_icm(dev, table->icm[i]);
238                 table->icm[i] = NULL;
239         }
240
241         mutex_unlock(&table->mutex);
242 }
243
244 void *mlx4_table_find(struct mlx4_icm_table *table, int obj)
245 {
246         int idx, offset, i;
247         struct mlx4_icm_chunk *chunk;
248         struct mlx4_icm *icm;
249         struct page *page = NULL;
250
251         if (!table->lowmem)
252                 return NULL;
253
254         mutex_lock(&table->mutex);
255
256         idx = obj & (table->num_obj - 1);
257         icm = table->icm[idx / (MLX4_TABLE_CHUNK_SIZE / table->obj_size)];
258         offset = idx % (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
259
260         if (!icm)
261                 goto out;
262
263         list_for_each_entry(chunk, &icm->chunk_list, list) {
264                 for (i = 0; i < chunk->npages; ++i) {
265                         if (chunk->mem[i].length > offset) {
266                                 page = chunk->mem[i].page;
267                                 goto out;
268                         }
269                         offset -= chunk->mem[i].length;
270                 }
271         }
272
273 out:
274         mutex_unlock(&table->mutex);
275         return page ? lowmem_page_address(page) + offset : NULL;
276 }
277
278 int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
279                          int start, int end)
280 {
281         int inc = MLX4_TABLE_CHUNK_SIZE / table->obj_size;
282         int i, err;
283
284         for (i = start; i <= end; i += inc) {
285                 err = mlx4_table_get(dev, table, i);
286                 if (err)
287                         goto fail;
288         }
289
290         return 0;
291
292 fail:
293         while (i > start) {
294                 i -= inc;
295                 mlx4_table_put(dev, table, i);
296         }
297
298         return err;
299 }
300
301 void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
302                           int start, int end)
303 {
304         int i;
305
306         for (i = start; i <= end; i += MLX4_TABLE_CHUNK_SIZE / table->obj_size)
307                 mlx4_table_put(dev, table, i);
308 }
309
310 int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
311                         u64 virt, int obj_size, int nobj, int reserved,
312                         int use_lowmem)
313 {
314         int obj_per_chunk;
315         int num_icm;
316         unsigned chunk_size;
317         int i;
318
319         obj_per_chunk = MLX4_TABLE_CHUNK_SIZE / obj_size;
320         num_icm = (nobj + obj_per_chunk - 1) / obj_per_chunk;
321
322         table->icm      = kcalloc(num_icm, sizeof *table->icm, GFP_KERNEL);
323         if (!table->icm)
324                 return -ENOMEM;
325         table->virt     = virt;
326         table->num_icm  = num_icm;
327         table->num_obj  = nobj;
328         table->obj_size = obj_size;
329         table->lowmem   = use_lowmem;
330         mutex_init(&table->mutex);
331
332         for (i = 0; i * MLX4_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
333                 chunk_size = MLX4_TABLE_CHUNK_SIZE;
334                 if ((i + 1) * MLX4_TABLE_CHUNK_SIZE > nobj * obj_size)
335                         chunk_size = PAGE_ALIGN(nobj * obj_size - i * MLX4_TABLE_CHUNK_SIZE);
336
337                 table->icm[i] = mlx4_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
338                                                (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
339                                                __GFP_NOWARN);
340                 if (!table->icm[i])
341                         goto err;
342                 if (mlx4_MAP_ICM(dev, table->icm[i], virt + i * MLX4_TABLE_CHUNK_SIZE)) {
343                         mlx4_free_icm(dev, table->icm[i]);
344                         table->icm[i] = NULL;
345                         goto err;
346                 }
347
348                 /*
349                  * Add a reference to this ICM chunk so that it never
350                  * gets freed (since it contains reserved firmware objects).
351                  */
352                 ++table->icm[i]->refcount;
353         }
354
355         return 0;
356
357 err:
358         for (i = 0; i < num_icm; ++i)
359                 if (table->icm[i]) {
360                         mlx4_UNMAP_ICM(dev, virt + i * MLX4_TABLE_CHUNK_SIZE,
361                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
362                         mlx4_free_icm(dev, table->icm[i]);
363                 }
364
365         return -ENOMEM;
366 }
367
368 void mlx4_cleanup_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table)
369 {
370         int i;
371
372         for (i = 0; i < table->num_icm; ++i)
373                 if (table->icm[i]) {
374                         mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
375                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
376                         mlx4_free_icm(dev, table->icm[i]);
377                 }
378
379         kfree(table->icm);
380 }