vserver 2.0 rc7
[linux-2.6.git] / drivers / infiniband / hw / mthca / mthca_memfree.c
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  *
32  * $Id$
33  */
34
35 #include <linux/mm.h>
36
37 #include "mthca_memfree.h"
38 #include "mthca_dev.h"
39 #include "mthca_cmd.h"
40
41 /*
42  * We allocate in as big chunks as we can, up to a maximum of 256 KB
43  * per chunk.
44  */
45 enum {
46         MTHCA_ICM_ALLOC_SIZE   = 1 << 18,
47         MTHCA_TABLE_CHUNK_SIZE = 1 << 18
48 };
49
50 void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm)
51 {
52         struct mthca_icm_chunk *chunk, *tmp;
53         int i;
54
55         if (!icm)
56                 return;
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 mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
74                                   unsigned int gfp_mask)
75 {
76         struct mthca_icm *icm;
77         struct mthca_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(MTHCA_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 == MTHCA_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         mthca_free_icm(dev, icm);
141         return NULL;
142 }
143
144 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
145 {
146         int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
147         int ret = 0;
148         u8 status;
149
150         down(&table->mutex);
151
152         if (table->icm[i]) {
153                 ++table->icm[i]->refcount;
154                 goto out;
155         }
156
157         table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
158                                         (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
159                                         __GFP_NOWARN);
160         if (!table->icm[i]) {
161                 ret = -ENOMEM;
162                 goto out;
163         }
164
165         if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
166                           &status) || status) {
167                 mthca_free_icm(dev, table->icm[i]);
168                 table->icm[i] = NULL;
169                 ret = -ENOMEM;
170                 goto out;
171         }
172
173         ++table->icm[i]->refcount;
174
175 out:
176         up(&table->mutex);
177         return ret;
178 }
179
180 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
181 {
182         int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
183         u8 status;
184
185         down(&table->mutex);
186
187         if (--table->icm[i]->refcount == 0) {
188                 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
189                                 MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
190                 mthca_free_icm(dev, table->icm[i]);
191                 table->icm[i] = NULL;
192         }
193
194         up(&table->mutex);
195 }
196
197 void *mthca_table_find(struct mthca_icm_table *table, int obj)
198 {
199         int idx, offset, i;
200         struct mthca_icm_chunk *chunk;
201         struct mthca_icm *icm;
202         struct page *page = NULL;
203
204         if (!table->lowmem)
205                 return NULL;
206
207         down(&table->mutex);
208
209         idx = (obj & (table->num_obj - 1)) * table->obj_size;
210         icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
211         offset = idx % MTHCA_TABLE_CHUNK_SIZE;
212
213         if (!icm)
214                 goto out;
215
216         list_for_each_entry(chunk, &icm->chunk_list, list) {
217                 for (i = 0; i < chunk->npages; ++i) {
218                         if (chunk->mem[i].length >= offset) {
219                                 page = chunk->mem[i].page;
220                                 break;
221                         }
222                         offset -= chunk->mem[i].length;
223                 }
224         }
225
226 out:
227         up(&table->mutex);
228         return page ? lowmem_page_address(page) + offset : NULL;
229 }
230
231 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
232                           int start, int end)
233 {
234         int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
235         int i, err;
236
237         for (i = start; i <= end; i += inc) {
238                 err = mthca_table_get(dev, table, i);
239                 if (err)
240                         goto fail;
241         }
242
243         return 0;
244
245 fail:
246         while (i > start) {
247                 i -= inc;
248                 mthca_table_put(dev, table, i);
249         }
250
251         return err;
252 }
253
254 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
255                            int start, int end)
256 {
257         int i;
258
259         for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
260                 mthca_table_put(dev, table, i);
261 }
262
263 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
264                                               u64 virt, int obj_size,
265                                               int nobj, int reserved,
266                                               int use_lowmem)
267 {
268         struct mthca_icm_table *table;
269         int num_icm;
270         int i;
271         u8 status;
272
273         num_icm = obj_size * nobj / MTHCA_TABLE_CHUNK_SIZE;
274
275         table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
276         if (!table)
277                 return NULL;
278
279         table->virt     = virt;
280         table->num_icm  = num_icm;
281         table->num_obj  = nobj;
282         table->obj_size = obj_size;
283         table->lowmem   = use_lowmem;
284         init_MUTEX(&table->mutex);
285
286         for (i = 0; i < num_icm; ++i)
287                 table->icm[i] = NULL;
288
289         for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
290                 table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
291                                                 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
292                                                 __GFP_NOWARN);
293                 if (!table->icm[i])
294                         goto err;
295                 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
296                                   &status) || status) {
297                         mthca_free_icm(dev, table->icm[i]);
298                         table->icm[i] = NULL;
299                         goto err;
300                 }
301
302                 /*
303                  * Add a reference to this ICM chunk so that it never
304                  * gets freed (since it contains reserved firmware objects).
305                  */
306                 ++table->icm[i]->refcount;
307         }
308
309         return table;
310
311 err:
312         for (i = 0; i < num_icm; ++i)
313                 if (table->icm[i]) {
314                         mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
315                                         MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
316                         mthca_free_icm(dev, table->icm[i]);
317                 }
318
319         kfree(table);
320
321         return NULL;
322 }
323
324 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
325 {
326         int i;
327         u8 status;
328
329         for (i = 0; i < table->num_icm; ++i)
330                 if (table->icm[i]) {
331                         mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
332                                         MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
333                         mthca_free_icm(dev, table->icm[i]);
334                 }
335
336         kfree(table);
337 }
338
339 static u64 mthca_uarc_virt(struct mthca_dev *dev, int page)
340 {
341         return dev->uar_table.uarc_base +
342                 dev->driver_uar.index * dev->uar_table.uarc_size +
343                 page * 4096;
344 }
345
346 int mthca_alloc_db(struct mthca_dev *dev, int type, u32 qn, u32 **db)
347 {
348         int group;
349         int start, end, dir;
350         int i, j;
351         struct mthca_db_page *page;
352         int ret = 0;
353         u8 status;
354
355         down(&dev->db_tab->mutex);
356
357         switch (type) {
358         case MTHCA_DB_TYPE_CQ_ARM:
359         case MTHCA_DB_TYPE_SQ:
360                 group = 0;
361                 start = 0;
362                 end   = dev->db_tab->max_group1;
363                 dir   = 1;
364                 break;
365
366         case MTHCA_DB_TYPE_CQ_SET_CI:
367         case MTHCA_DB_TYPE_RQ:
368         case MTHCA_DB_TYPE_SRQ:
369                 group = 1;
370                 start = dev->db_tab->npages - 1;
371                 end   = dev->db_tab->min_group2;
372                 dir   = -1;
373                 break;
374
375         default:
376                 ret = -EINVAL;
377                 goto out;
378         }
379
380         for (i = start; i != end; i += dir)
381                 if (dev->db_tab->page[i].db_rec &&
382                     !bitmap_full(dev->db_tab->page[i].used,
383                                  MTHCA_DB_REC_PER_PAGE)) {
384                         page = dev->db_tab->page + i;
385                         goto found;
386                 }
387
388         if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
389                 ret = -ENOMEM;
390                 goto out;
391         }
392
393         page = dev->db_tab->page + end;
394         page->db_rec = dma_alloc_coherent(&dev->pdev->dev, 4096,
395                                           &page->mapping, GFP_KERNEL);
396         if (!page->db_rec) {
397                 ret = -ENOMEM;
398                 goto out;
399         }
400         memset(page->db_rec, 0, 4096);
401
402         ret = mthca_MAP_ICM_page(dev, page->mapping, mthca_uarc_virt(dev, i), &status);
403         if (!ret && status)
404                 ret = -EINVAL;
405         if (ret) {
406                 dma_free_coherent(&dev->pdev->dev, 4096,
407                                   page->db_rec, page->mapping);
408                 goto out;
409         }
410
411         bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
412         if (group == 0)
413                 ++dev->db_tab->max_group1;
414         else
415                 --dev->db_tab->min_group2;
416
417 found:
418         j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
419         set_bit(j, page->used);
420
421         if (group == 1)
422                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
423
424         ret = i * MTHCA_DB_REC_PER_PAGE + j;
425
426         page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
427
428         *db = (u32 *) &page->db_rec[j];
429
430 out:
431         up(&dev->db_tab->mutex);
432
433         return ret;
434 }
435
436 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
437 {
438         int i, j;
439         struct mthca_db_page *page;
440         u8 status;
441
442         i = db_index / MTHCA_DB_REC_PER_PAGE;
443         j = db_index % MTHCA_DB_REC_PER_PAGE;
444
445         page = dev->db_tab->page + i;
446
447         down(&dev->db_tab->mutex);
448
449         page->db_rec[j] = 0;
450         if (i >= dev->db_tab->min_group2)
451                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
452         clear_bit(j, page->used);
453
454         if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
455             i >= dev->db_tab->max_group1 - 1) {
456                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, i), 1, &status);
457
458                 dma_free_coherent(&dev->pdev->dev, 4096,
459                                   page->db_rec, page->mapping);
460                 page->db_rec = NULL;
461
462                 if (i == dev->db_tab->max_group1) {
463                         --dev->db_tab->max_group1;
464                         /* XXX may be able to unmap more pages now */
465                 }
466                 if (i == dev->db_tab->min_group2)
467                         ++dev->db_tab->min_group2;
468         }
469
470         up(&dev->db_tab->mutex);
471 }
472
473 int mthca_init_db_tab(struct mthca_dev *dev)
474 {
475         int i;
476
477         if (!mthca_is_memfree(dev))
478                 return 0;
479
480         dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
481         if (!dev->db_tab)
482                 return -ENOMEM;
483
484         init_MUTEX(&dev->db_tab->mutex);
485
486         dev->db_tab->npages     = dev->uar_table.uarc_size / 4096;
487         dev->db_tab->max_group1 = 0;
488         dev->db_tab->min_group2 = dev->db_tab->npages - 1;
489
490         dev->db_tab->page = kmalloc(dev->db_tab->npages *
491                                     sizeof *dev->db_tab->page,
492                                     GFP_KERNEL);
493         if (!dev->db_tab->page) {
494                 kfree(dev->db_tab);
495                 return -ENOMEM;
496         }
497
498         for (i = 0; i < dev->db_tab->npages; ++i)
499                 dev->db_tab->page[i].db_rec = NULL;
500
501         return 0;
502 }
503
504 void mthca_cleanup_db_tab(struct mthca_dev *dev)
505 {
506         int i;
507         u8 status;
508
509         if (!mthca_is_memfree(dev))
510                 return;
511
512         /*
513          * Because we don't always free our UARC pages when they
514          * become empty to make mthca_free_db() simpler we need to
515          * make a sweep through the doorbell pages and free any
516          * leftover pages now.
517          */
518         for (i = 0; i < dev->db_tab->npages; ++i) {
519                 if (!dev->db_tab->page[i].db_rec)
520                         continue;
521
522                 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
523                         mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
524
525                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, i), 1, &status);
526
527                 dma_free_coherent(&dev->pdev->dev, 4096,
528                                   dev->db_tab->page[i].db_rec,
529                                   dev->db_tab->page[i].mapping);
530         }
531
532         kfree(dev->db_tab->page);
533         kfree(dev->db_tab);
534 }