Merge with /home/shaggy/git/linus-clean/
[linux-2.6] / drivers / infiniband / hw / mthca / mthca_memfree.c
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Cisco Systems.  All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  *
34  * $Id$
35  */
36
37 #include <linux/mm.h>
38
39 #include "mthca_memfree.h"
40 #include "mthca_dev.h"
41 #include "mthca_cmd.h"
42
43 /*
44  * We allocate in as big chunks as we can, up to a maximum of 256 KB
45  * per chunk.
46  */
47 enum {
48         MTHCA_ICM_ALLOC_SIZE   = 1 << 18,
49         MTHCA_TABLE_CHUNK_SIZE = 1 << 18
50 };
51
52 struct mthca_user_db_table {
53         struct mutex mutex;
54         struct {
55                 u64                uvirt;
56                 struct scatterlist mem;
57                 int                refcount;
58         }                page[0];
59 };
60
61 void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm)
62 {
63         struct mthca_icm_chunk *chunk, *tmp;
64         int i;
65
66         if (!icm)
67                 return;
68
69         list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
70                 if (chunk->nsg > 0)
71                         pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
72                                      PCI_DMA_BIDIRECTIONAL);
73
74                 for (i = 0; i < chunk->npages; ++i)
75                         __free_pages(chunk->mem[i].page,
76                                      get_order(chunk->mem[i].length));
77
78                 kfree(chunk);
79         }
80
81         kfree(icm);
82 }
83
84 struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
85                                   gfp_t gfp_mask)
86 {
87         struct mthca_icm *icm;
88         struct mthca_icm_chunk *chunk = NULL;
89         int cur_order;
90
91         icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
92         if (!icm)
93                 return icm;
94
95         icm->refcount = 0;
96         INIT_LIST_HEAD(&icm->chunk_list);
97
98         cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
99
100         while (npages > 0) {
101                 if (!chunk) {
102                         chunk = kmalloc(sizeof *chunk,
103                                         gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
104                         if (!chunk)
105                                 goto fail;
106
107                         chunk->npages = 0;
108                         chunk->nsg    = 0;
109                         list_add_tail(&chunk->list, &icm->chunk_list);
110                 }
111
112                 while (1 << cur_order > npages)
113                         --cur_order;
114
115                 chunk->mem[chunk->npages].page = alloc_pages(gfp_mask, cur_order);
116                 if (chunk->mem[chunk->npages].page) {
117                         chunk->mem[chunk->npages].length = PAGE_SIZE << cur_order;
118                         chunk->mem[chunk->npages].offset = 0;
119
120                         if (++chunk->npages == MTHCA_ICM_CHUNK_LEN) {
121                                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
122                                                         chunk->npages,
123                                                         PCI_DMA_BIDIRECTIONAL);
124
125                                 if (chunk->nsg <= 0)
126                                         goto fail;
127
128                                 chunk = NULL;
129                         }
130
131                         npages -= 1 << cur_order;
132                 } else {
133                         --cur_order;
134                         if (cur_order < 0)
135                                 goto fail;
136                 }
137         }
138
139         if (chunk) {
140                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
141                                         chunk->npages,
142                                         PCI_DMA_BIDIRECTIONAL);
143
144                 if (chunk->nsg <= 0)
145                         goto fail;
146         }
147
148         return icm;
149
150 fail:
151         mthca_free_icm(dev, icm);
152         return NULL;
153 }
154
155 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
156 {
157         int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
158         int ret = 0;
159         u8 status;
160
161         mutex_lock(&table->mutex);
162
163         if (table->icm[i]) {
164                 ++table->icm[i]->refcount;
165                 goto out;
166         }
167
168         table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
169                                         (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
170                                         __GFP_NOWARN);
171         if (!table->icm[i]) {
172                 ret = -ENOMEM;
173                 goto out;
174         }
175
176         if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
177                           &status) || status) {
178                 mthca_free_icm(dev, table->icm[i]);
179                 table->icm[i] = NULL;
180                 ret = -ENOMEM;
181                 goto out;
182         }
183
184         ++table->icm[i]->refcount;
185
186 out:
187         mutex_unlock(&table->mutex);
188         return ret;
189 }
190
191 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
192 {
193         int i;
194         u8 status;
195
196         if (!mthca_is_memfree(dev))
197                 return;
198
199         i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
200
201         mutex_lock(&table->mutex);
202
203         if (--table->icm[i]->refcount == 0) {
204                 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
205                                 MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
206                 mthca_free_icm(dev, table->icm[i]);
207                 table->icm[i] = NULL;
208         }
209
210         mutex_unlock(&table->mutex);
211 }
212
213 void *mthca_table_find(struct mthca_icm_table *table, int obj)
214 {
215         int idx, offset, i;
216         struct mthca_icm_chunk *chunk;
217         struct mthca_icm *icm;
218         struct page *page = NULL;
219
220         if (!table->lowmem)
221                 return NULL;
222
223         mutex_lock(&table->mutex);
224
225         idx = (obj & (table->num_obj - 1)) * table->obj_size;
226         icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
227         offset = idx % MTHCA_TABLE_CHUNK_SIZE;
228
229         if (!icm)
230                 goto out;
231
232         list_for_each_entry(chunk, &icm->chunk_list, list) {
233                 for (i = 0; i < chunk->npages; ++i) {
234                         if (chunk->mem[i].length >= offset) {
235                                 page = chunk->mem[i].page;
236                                 goto out;
237                         }
238                         offset -= chunk->mem[i].length;
239                 }
240         }
241
242 out:
243         mutex_unlock(&table->mutex);
244         return page ? lowmem_page_address(page) + offset : NULL;
245 }
246
247 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
248                           int start, int end)
249 {
250         int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
251         int i, err;
252
253         for (i = start; i <= end; i += inc) {
254                 err = mthca_table_get(dev, table, i);
255                 if (err)
256                         goto fail;
257         }
258
259         return 0;
260
261 fail:
262         while (i > start) {
263                 i -= inc;
264                 mthca_table_put(dev, table, i);
265         }
266
267         return err;
268 }
269
270 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
271                            int start, int end)
272 {
273         int i;
274
275         if (!mthca_is_memfree(dev))
276                 return;
277
278         for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
279                 mthca_table_put(dev, table, i);
280 }
281
282 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
283                                               u64 virt, int obj_size,
284                                               int nobj, int reserved,
285                                               int use_lowmem)
286 {
287         struct mthca_icm_table *table;
288         int num_icm;
289         unsigned chunk_size;
290         int i;
291         u8 status;
292
293         num_icm = (obj_size * nobj + MTHCA_TABLE_CHUNK_SIZE - 1) / MTHCA_TABLE_CHUNK_SIZE;
294
295         table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
296         if (!table)
297                 return NULL;
298
299         table->virt     = virt;
300         table->num_icm  = num_icm;
301         table->num_obj  = nobj;
302         table->obj_size = obj_size;
303         table->lowmem   = use_lowmem;
304         mutex_init(&table->mutex);
305
306         for (i = 0; i < num_icm; ++i)
307                 table->icm[i] = NULL;
308
309         for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
310                 chunk_size = MTHCA_TABLE_CHUNK_SIZE;
311                 if ((i + 1) * MTHCA_TABLE_CHUNK_SIZE > nobj * obj_size)
312                         chunk_size = nobj * obj_size - i * MTHCA_TABLE_CHUNK_SIZE;
313
314                 table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
315                                                 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
316                                                 __GFP_NOWARN);
317                 if (!table->icm[i])
318                         goto err;
319                 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
320                                   &status) || status) {
321                         mthca_free_icm(dev, table->icm[i]);
322                         table->icm[i] = NULL;
323                         goto err;
324                 }
325
326                 /*
327                  * Add a reference to this ICM chunk so that it never
328                  * gets freed (since it contains reserved firmware objects).
329                  */
330                 ++table->icm[i]->refcount;
331         }
332
333         return table;
334
335 err:
336         for (i = 0; i < num_icm; ++i)
337                 if (table->icm[i]) {
338                         mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
339                                         MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
340                         mthca_free_icm(dev, table->icm[i]);
341                 }
342
343         kfree(table);
344
345         return NULL;
346 }
347
348 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
349 {
350         int i;
351         u8 status;
352
353         for (i = 0; i < table->num_icm; ++i)
354                 if (table->icm[i]) {
355                         mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
356                                         MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
357                         mthca_free_icm(dev, table->icm[i]);
358                 }
359
360         kfree(table);
361 }
362
363 static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
364 {
365         return dev->uar_table.uarc_base +
366                 uar->index * dev->uar_table.uarc_size +
367                 page * 4096;
368 }
369
370 int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
371                       struct mthca_user_db_table *db_tab, int index, u64 uaddr)
372 {
373         int ret = 0;
374         u8 status;
375         int i;
376
377         if (!mthca_is_memfree(dev))
378                 return 0;
379
380         if (index < 0 || index > dev->uar_table.uarc_size / 8)
381                 return -EINVAL;
382
383         mutex_lock(&db_tab->mutex);
384
385         i = index / MTHCA_DB_REC_PER_PAGE;
386
387         if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE)       ||
388             (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
389             (uaddr & 4095)) {
390                 ret = -EINVAL;
391                 goto out;
392         }
393
394         if (db_tab->page[i].refcount) {
395                 ++db_tab->page[i].refcount;
396                 goto out;
397         }
398
399         ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
400                              &db_tab->page[i].mem.page, NULL);
401         if (ret < 0)
402                 goto out;
403
404         db_tab->page[i].mem.length = 4096;
405         db_tab->page[i].mem.offset = uaddr & ~PAGE_MASK;
406
407         ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
408         if (ret < 0) {
409                 put_page(db_tab->page[i].mem.page);
410                 goto out;
411         }
412
413         ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
414                                  mthca_uarc_virt(dev, uar, i), &status);
415         if (!ret && status)
416                 ret = -EINVAL;
417         if (ret) {
418                 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
419                 put_page(db_tab->page[i].mem.page);
420                 goto out;
421         }
422
423         db_tab->page[i].uvirt    = uaddr;
424         db_tab->page[i].refcount = 1;
425
426 out:
427         mutex_unlock(&db_tab->mutex);
428         return ret;
429 }
430
431 void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
432                          struct mthca_user_db_table *db_tab, int index)
433 {
434         if (!mthca_is_memfree(dev))
435                 return;
436
437         /*
438          * To make our bookkeeping simpler, we don't unmap DB
439          * pages until we clean up the whole db table.
440          */
441
442         mutex_lock(&db_tab->mutex);
443
444         --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
445
446         mutex_unlock(&db_tab->mutex);
447 }
448
449 struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
450 {
451         struct mthca_user_db_table *db_tab;
452         int npages;
453         int i;
454
455         if (!mthca_is_memfree(dev))
456                 return NULL;
457
458         npages = dev->uar_table.uarc_size / 4096;
459         db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
460         if (!db_tab)
461                 return ERR_PTR(-ENOMEM);
462
463         mutex_init(&db_tab->mutex);
464         for (i = 0; i < npages; ++i) {
465                 db_tab->page[i].refcount = 0;
466                 db_tab->page[i].uvirt    = 0;
467         }
468
469         return db_tab;
470 }
471
472 void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
473                                struct mthca_user_db_table *db_tab)
474 {
475         int i;
476         u8 status;
477
478         if (!mthca_is_memfree(dev))
479                 return;
480
481         for (i = 0; i < dev->uar_table.uarc_size / 4096; ++i) {
482                 if (db_tab->page[i].uvirt) {
483                         mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1, &status);
484                         pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
485                         put_page(db_tab->page[i].mem.page);
486                 }
487         }
488
489         kfree(db_tab);
490 }
491
492 int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
493                    u32 qn, __be32 **db)
494 {
495         int group;
496         int start, end, dir;
497         int i, j;
498         struct mthca_db_page *page;
499         int ret = 0;
500         u8 status;
501
502         mutex_lock(&dev->db_tab->mutex);
503
504         switch (type) {
505         case MTHCA_DB_TYPE_CQ_ARM:
506         case MTHCA_DB_TYPE_SQ:
507                 group = 0;
508                 start = 0;
509                 end   = dev->db_tab->max_group1;
510                 dir   = 1;
511                 break;
512
513         case MTHCA_DB_TYPE_CQ_SET_CI:
514         case MTHCA_DB_TYPE_RQ:
515         case MTHCA_DB_TYPE_SRQ:
516                 group = 1;
517                 start = dev->db_tab->npages - 1;
518                 end   = dev->db_tab->min_group2;
519                 dir   = -1;
520                 break;
521
522         default:
523                 ret = -EINVAL;
524                 goto out;
525         }
526
527         for (i = start; i != end; i += dir)
528                 if (dev->db_tab->page[i].db_rec &&
529                     !bitmap_full(dev->db_tab->page[i].used,
530                                  MTHCA_DB_REC_PER_PAGE)) {
531                         page = dev->db_tab->page + i;
532                         goto found;
533                 }
534
535         for (i = start; i != end; i += dir)
536                 if (!dev->db_tab->page[i].db_rec) {
537                         page = dev->db_tab->page + i;
538                         goto alloc;
539                 }
540
541         if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
542                 ret = -ENOMEM;
543                 goto out;
544         }
545
546         if (group == 0)
547                 ++dev->db_tab->max_group1;
548         else
549                 --dev->db_tab->min_group2;
550
551         page = dev->db_tab->page + end;
552
553 alloc:
554         page->db_rec = dma_alloc_coherent(&dev->pdev->dev, 4096,
555                                           &page->mapping, GFP_KERNEL);
556         if (!page->db_rec) {
557                 ret = -ENOMEM;
558                 goto out;
559         }
560         memset(page->db_rec, 0, 4096);
561
562         ret = mthca_MAP_ICM_page(dev, page->mapping,
563                                  mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
564         if (!ret && status)
565                 ret = -EINVAL;
566         if (ret) {
567                 dma_free_coherent(&dev->pdev->dev, 4096,
568                                   page->db_rec, page->mapping);
569                 goto out;
570         }
571
572         bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
573
574 found:
575         j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
576         set_bit(j, page->used);
577
578         if (group == 1)
579                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
580
581         ret = i * MTHCA_DB_REC_PER_PAGE + j;
582
583         page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
584
585         *db = (__be32 *) &page->db_rec[j];
586
587 out:
588         mutex_unlock(&dev->db_tab->mutex);
589
590         return ret;
591 }
592
593 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
594 {
595         int i, j;
596         struct mthca_db_page *page;
597         u8 status;
598
599         i = db_index / MTHCA_DB_REC_PER_PAGE;
600         j = db_index % MTHCA_DB_REC_PER_PAGE;
601
602         page = dev->db_tab->page + i;
603
604         mutex_lock(&dev->db_tab->mutex);
605
606         page->db_rec[j] = 0;
607         if (i >= dev->db_tab->min_group2)
608                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
609         clear_bit(j, page->used);
610
611         if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
612             i >= dev->db_tab->max_group1 - 1) {
613                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
614
615                 dma_free_coherent(&dev->pdev->dev, 4096,
616                                   page->db_rec, page->mapping);
617                 page->db_rec = NULL;
618
619                 if (i == dev->db_tab->max_group1) {
620                         --dev->db_tab->max_group1;
621                         /* XXX may be able to unmap more pages now */
622                 }
623                 if (i == dev->db_tab->min_group2)
624                         ++dev->db_tab->min_group2;
625         }
626
627         mutex_unlock(&dev->db_tab->mutex);
628 }
629
630 int mthca_init_db_tab(struct mthca_dev *dev)
631 {
632         int i;
633
634         if (!mthca_is_memfree(dev))
635                 return 0;
636
637         dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
638         if (!dev->db_tab)
639                 return -ENOMEM;
640
641         mutex_init(&dev->db_tab->mutex);
642
643         dev->db_tab->npages     = dev->uar_table.uarc_size / 4096;
644         dev->db_tab->max_group1 = 0;
645         dev->db_tab->min_group2 = dev->db_tab->npages - 1;
646
647         dev->db_tab->page = kmalloc(dev->db_tab->npages *
648                                     sizeof *dev->db_tab->page,
649                                     GFP_KERNEL);
650         if (!dev->db_tab->page) {
651                 kfree(dev->db_tab);
652                 return -ENOMEM;
653         }
654
655         for (i = 0; i < dev->db_tab->npages; ++i)
656                 dev->db_tab->page[i].db_rec = NULL;
657
658         return 0;
659 }
660
661 void mthca_cleanup_db_tab(struct mthca_dev *dev)
662 {
663         int i;
664         u8 status;
665
666         if (!mthca_is_memfree(dev))
667                 return;
668
669         /*
670          * Because we don't always free our UARC pages when they
671          * become empty to make mthca_free_db() simpler we need to
672          * make a sweep through the doorbell pages and free any
673          * leftover pages now.
674          */
675         for (i = 0; i < dev->db_tab->npages; ++i) {
676                 if (!dev->db_tab->page[i].db_rec)
677                         continue;
678
679                 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
680                         mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
681
682                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
683
684                 dma_free_coherent(&dev->pdev->dev, 4096,
685                                   dev->db_tab->page[i].db_rec,
686                                   dev->db_tab->page[i].mapping);
687         }
688
689         kfree(dev->db_tab->page);
690         kfree(dev->db_tab);
691 }