SG: Change sg_set_page() to take length and offset argument
[linux-3.10.git] / 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 #include <linux/scatterlist.h>
39 #include <linux/sched.h>
40
41 #include <asm/page.h>
42
43 #include "mthca_memfree.h"
44 #include "mthca_dev.h"
45 #include "mthca_cmd.h"
46
47 /*
48  * We allocate in as big chunks as we can, up to a maximum of 256 KB
49  * per chunk.
50  */
51 enum {
52         MTHCA_ICM_ALLOC_SIZE   = 1 << 18,
53         MTHCA_TABLE_CHUNK_SIZE = 1 << 18
54 };
55
56 struct mthca_user_db_table {
57         struct mutex mutex;
58         struct {
59                 u64                uvirt;
60                 struct scatterlist mem;
61                 int                refcount;
62         }                page[0];
63 };
64
65 static void mthca_free_icm_pages(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
66 {
67         int i;
68
69         if (chunk->nsg > 0)
70                 pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
71                              PCI_DMA_BIDIRECTIONAL);
72
73         for (i = 0; i < chunk->npages; ++i)
74                 __free_pages(sg_page(&chunk->mem[i]),
75                              get_order(chunk->mem[i].length));
76 }
77
78 static void mthca_free_icm_coherent(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
79 {
80         int i;
81
82         for (i = 0; i < chunk->npages; ++i) {
83                 dma_free_coherent(&dev->pdev->dev, chunk->mem[i].length,
84                                   lowmem_page_address(sg_page(&chunk->mem[i])),
85                                   sg_dma_address(&chunk->mem[i]));
86         }
87 }
88
89 void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent)
90 {
91         struct mthca_icm_chunk *chunk, *tmp;
92
93         if (!icm)
94                 return;
95
96         list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
97                 if (coherent)
98                         mthca_free_icm_coherent(dev, chunk);
99                 else
100                         mthca_free_icm_pages(dev, chunk);
101
102                 kfree(chunk);
103         }
104
105         kfree(icm);
106 }
107
108 static int mthca_alloc_icm_pages(struct scatterlist *mem, int order, gfp_t gfp_mask)
109 {
110         struct page *page;
111
112         page = alloc_pages(gfp_mask, order);
113         if (!page)
114                 return -ENOMEM;
115
116         sg_set_page(mem, page, PAGE_SIZE << order, 0);
117         return 0;
118 }
119
120 static int mthca_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
121                                     int order, gfp_t gfp_mask)
122 {
123         void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order, &sg_dma_address(mem),
124                                        gfp_mask);
125         if (!buf)
126                 return -ENOMEM;
127
128         sg_set_buf(mem, buf, PAGE_SIZE << order);
129         BUG_ON(mem->offset);
130         sg_dma_len(mem) = PAGE_SIZE << order;
131         return 0;
132 }
133
134 struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
135                                   gfp_t gfp_mask, int coherent)
136 {
137         struct mthca_icm *icm;
138         struct mthca_icm_chunk *chunk = NULL;
139         int cur_order;
140         int ret;
141
142         /* We use sg_set_buf for coherent allocs, which assumes low memory */
143         BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
144
145         icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
146         if (!icm)
147                 return icm;
148
149         icm->refcount = 0;
150         INIT_LIST_HEAD(&icm->chunk_list);
151
152         cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
153
154         while (npages > 0) {
155                 if (!chunk) {
156                         chunk = kmalloc(sizeof *chunk,
157                                         gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
158                         if (!chunk)
159                                 goto fail;
160
161                         sg_init_table(chunk->mem, MTHCA_ICM_CHUNK_LEN);
162                         chunk->npages = 0;
163                         chunk->nsg    = 0;
164                         list_add_tail(&chunk->list, &icm->chunk_list);
165                 }
166
167                 while (1 << cur_order > npages)
168                         --cur_order;
169
170                 if (coherent)
171                         ret = mthca_alloc_icm_coherent(&dev->pdev->dev,
172                                                        &chunk->mem[chunk->npages],
173                                                        cur_order, gfp_mask);
174                 else
175                         ret = mthca_alloc_icm_pages(&chunk->mem[chunk->npages],
176                                                     cur_order, gfp_mask);
177
178                 if (!ret) {
179                         ++chunk->npages;
180
181                         if (coherent)
182                                 ++chunk->nsg;
183                         else if (chunk->npages == MTHCA_ICM_CHUNK_LEN) {
184                                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
185                                                         chunk->npages,
186                                                         PCI_DMA_BIDIRECTIONAL);
187
188                                 if (chunk->nsg <= 0)
189                                         goto fail;
190                         }
191
192                         if (chunk->npages == MTHCA_ICM_CHUNK_LEN)
193                                 chunk = NULL;
194
195                         npages -= 1 << cur_order;
196                 } else {
197                         --cur_order;
198                         if (cur_order < 0)
199                                 goto fail;
200                 }
201         }
202
203         if (!coherent && chunk) {
204                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
205                                         chunk->npages,
206                                         PCI_DMA_BIDIRECTIONAL);
207
208                 if (chunk->nsg <= 0)
209                         goto fail;
210         }
211
212         return icm;
213
214 fail:
215         mthca_free_icm(dev, icm, coherent);
216         return NULL;
217 }
218
219 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
220 {
221         int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
222         int ret = 0;
223         u8 status;
224
225         mutex_lock(&table->mutex);
226
227         if (table->icm[i]) {
228                 ++table->icm[i]->refcount;
229                 goto out;
230         }
231
232         table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
233                                         (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
234                                         __GFP_NOWARN, table->coherent);
235         if (!table->icm[i]) {
236                 ret = -ENOMEM;
237                 goto out;
238         }
239
240         if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
241                           &status) || status) {
242                 mthca_free_icm(dev, table->icm[i], table->coherent);
243                 table->icm[i] = NULL;
244                 ret = -ENOMEM;
245                 goto out;
246         }
247
248         ++table->icm[i]->refcount;
249
250 out:
251         mutex_unlock(&table->mutex);
252         return ret;
253 }
254
255 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
256 {
257         int i;
258         u8 status;
259
260         if (!mthca_is_memfree(dev))
261                 return;
262
263         i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
264
265         mutex_lock(&table->mutex);
266
267         if (--table->icm[i]->refcount == 0) {
268                 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
269                                 MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
270                                 &status);
271                 mthca_free_icm(dev, table->icm[i], table->coherent);
272                 table->icm[i] = NULL;
273         }
274
275         mutex_unlock(&table->mutex);
276 }
277
278 void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle)
279 {
280         int idx, offset, dma_offset, i;
281         struct mthca_icm_chunk *chunk;
282         struct mthca_icm *icm;
283         struct page *page = NULL;
284
285         if (!table->lowmem)
286                 return NULL;
287
288         mutex_lock(&table->mutex);
289
290         idx = (obj & (table->num_obj - 1)) * table->obj_size;
291         icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
292         dma_offset = offset = idx % MTHCA_TABLE_CHUNK_SIZE;
293
294         if (!icm)
295                 goto out;
296
297         list_for_each_entry(chunk, &icm->chunk_list, list) {
298                 for (i = 0; i < chunk->npages; ++i) {
299                         if (dma_handle && dma_offset >= 0) {
300                                 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
301                                         *dma_handle = sg_dma_address(&chunk->mem[i]) +
302                                                 dma_offset;
303                                 dma_offset -= sg_dma_len(&chunk->mem[i]);
304                         }
305                         /* DMA mapping can merge pages but not split them,
306                          * so if we found the page, dma_handle has already
307                          * been assigned to. */
308                         if (chunk->mem[i].length > offset) {
309                                 page = sg_page(&chunk->mem[i]);
310                                 goto out;
311                         }
312                         offset -= chunk->mem[i].length;
313                 }
314         }
315
316 out:
317         mutex_unlock(&table->mutex);
318         return page ? lowmem_page_address(page) + offset : NULL;
319 }
320
321 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
322                           int start, int end)
323 {
324         int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
325         int i, err;
326
327         for (i = start; i <= end; i += inc) {
328                 err = mthca_table_get(dev, table, i);
329                 if (err)
330                         goto fail;
331         }
332
333         return 0;
334
335 fail:
336         while (i > start) {
337                 i -= inc;
338                 mthca_table_put(dev, table, i);
339         }
340
341         return err;
342 }
343
344 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
345                            int start, int end)
346 {
347         int i;
348
349         if (!mthca_is_memfree(dev))
350                 return;
351
352         for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
353                 mthca_table_put(dev, table, i);
354 }
355
356 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
357                                               u64 virt, int obj_size,
358                                               int nobj, int reserved,
359                                               int use_lowmem, int use_coherent)
360 {
361         struct mthca_icm_table *table;
362         int num_icm;
363         unsigned chunk_size;
364         int i;
365         u8 status;
366
367         num_icm = (obj_size * nobj + MTHCA_TABLE_CHUNK_SIZE - 1) / MTHCA_TABLE_CHUNK_SIZE;
368
369         table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
370         if (!table)
371                 return NULL;
372
373         table->virt     = virt;
374         table->num_icm  = num_icm;
375         table->num_obj  = nobj;
376         table->obj_size = obj_size;
377         table->lowmem   = use_lowmem;
378         table->coherent = use_coherent;
379         mutex_init(&table->mutex);
380
381         for (i = 0; i < num_icm; ++i)
382                 table->icm[i] = NULL;
383
384         for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
385                 chunk_size = MTHCA_TABLE_CHUNK_SIZE;
386                 if ((i + 1) * MTHCA_TABLE_CHUNK_SIZE > nobj * obj_size)
387                         chunk_size = nobj * obj_size - i * MTHCA_TABLE_CHUNK_SIZE;
388
389                 table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
390                                                 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
391                                                 __GFP_NOWARN, use_coherent);
392                 if (!table->icm[i])
393                         goto err;
394                 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
395                                   &status) || status) {
396                         mthca_free_icm(dev, table->icm[i], table->coherent);
397                         table->icm[i] = NULL;
398                         goto err;
399                 }
400
401                 /*
402                  * Add a reference to this ICM chunk so that it never
403                  * gets freed (since it contains reserved firmware objects).
404                  */
405                 ++table->icm[i]->refcount;
406         }
407
408         return table;
409
410 err:
411         for (i = 0; i < num_icm; ++i)
412                 if (table->icm[i]) {
413                         mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
414                                         MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
415                                         &status);
416                         mthca_free_icm(dev, table->icm[i], table->coherent);
417                 }
418
419         kfree(table);
420
421         return NULL;
422 }
423
424 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
425 {
426         int i;
427         u8 status;
428
429         for (i = 0; i < table->num_icm; ++i)
430                 if (table->icm[i]) {
431                         mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
432                                         MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
433                                         &status);
434                         mthca_free_icm(dev, table->icm[i], table->coherent);
435                 }
436
437         kfree(table);
438 }
439
440 static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
441 {
442         return dev->uar_table.uarc_base +
443                 uar->index * dev->uar_table.uarc_size +
444                 page * MTHCA_ICM_PAGE_SIZE;
445 }
446
447 int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
448                       struct mthca_user_db_table *db_tab, int index, u64 uaddr)
449 {
450         struct page *pages[1];
451         int ret = 0;
452         u8 status;
453         int i;
454
455         if (!mthca_is_memfree(dev))
456                 return 0;
457
458         if (index < 0 || index > dev->uar_table.uarc_size / 8)
459                 return -EINVAL;
460
461         mutex_lock(&db_tab->mutex);
462
463         i = index / MTHCA_DB_REC_PER_PAGE;
464
465         if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE)       ||
466             (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
467             (uaddr & 4095)) {
468                 ret = -EINVAL;
469                 goto out;
470         }
471
472         if (db_tab->page[i].refcount) {
473                 ++db_tab->page[i].refcount;
474                 goto out;
475         }
476
477         ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
478                              pages, NULL);
479         if (ret < 0)
480                 goto out;
481
482         sg_set_page(&db_tab->page[i].mem, pages[0], MTHCA_ICM_PAGE_SIZE,
483                         uaddr & ~PAGE_MASK);
484
485         ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
486         if (ret < 0) {
487                 put_page(pages[0]);
488                 goto out;
489         }
490
491         ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
492                                  mthca_uarc_virt(dev, uar, i), &status);
493         if (!ret && status)
494                 ret = -EINVAL;
495         if (ret) {
496                 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
497                 put_page(sg_page(&db_tab->page[i].mem));
498                 goto out;
499         }
500
501         db_tab->page[i].uvirt    = uaddr;
502         db_tab->page[i].refcount = 1;
503
504 out:
505         mutex_unlock(&db_tab->mutex);
506         return ret;
507 }
508
509 void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
510                          struct mthca_user_db_table *db_tab, int index)
511 {
512         if (!mthca_is_memfree(dev))
513                 return;
514
515         /*
516          * To make our bookkeeping simpler, we don't unmap DB
517          * pages until we clean up the whole db table.
518          */
519
520         mutex_lock(&db_tab->mutex);
521
522         --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
523
524         mutex_unlock(&db_tab->mutex);
525 }
526
527 struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
528 {
529         struct mthca_user_db_table *db_tab;
530         int npages;
531         int i;
532
533         if (!mthca_is_memfree(dev))
534                 return NULL;
535
536         npages = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
537         db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
538         if (!db_tab)
539                 return ERR_PTR(-ENOMEM);
540
541         mutex_init(&db_tab->mutex);
542         for (i = 0; i < npages; ++i) {
543                 db_tab->page[i].refcount = 0;
544                 db_tab->page[i].uvirt    = 0;
545         }
546
547         return db_tab;
548 }
549
550 void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
551                                struct mthca_user_db_table *db_tab)
552 {
553         int i;
554         u8 status;
555
556         if (!mthca_is_memfree(dev))
557                 return;
558
559         for (i = 0; i < dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE; ++i) {
560                 if (db_tab->page[i].uvirt) {
561                         mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1, &status);
562                         pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
563                         put_page(sg_page(&db_tab->page[i].mem));
564                 }
565         }
566
567         kfree(db_tab);
568 }
569
570 int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
571                    u32 qn, __be32 **db)
572 {
573         int group;
574         int start, end, dir;
575         int i, j;
576         struct mthca_db_page *page;
577         int ret = 0;
578         u8 status;
579
580         mutex_lock(&dev->db_tab->mutex);
581
582         switch (type) {
583         case MTHCA_DB_TYPE_CQ_ARM:
584         case MTHCA_DB_TYPE_SQ:
585                 group = 0;
586                 start = 0;
587                 end   = dev->db_tab->max_group1;
588                 dir   = 1;
589                 break;
590
591         case MTHCA_DB_TYPE_CQ_SET_CI:
592         case MTHCA_DB_TYPE_RQ:
593         case MTHCA_DB_TYPE_SRQ:
594                 group = 1;
595                 start = dev->db_tab->npages - 1;
596                 end   = dev->db_tab->min_group2;
597                 dir   = -1;
598                 break;
599
600         default:
601                 ret = -EINVAL;
602                 goto out;
603         }
604
605         for (i = start; i != end; i += dir)
606                 if (dev->db_tab->page[i].db_rec &&
607                     !bitmap_full(dev->db_tab->page[i].used,
608                                  MTHCA_DB_REC_PER_PAGE)) {
609                         page = dev->db_tab->page + i;
610                         goto found;
611                 }
612
613         for (i = start; i != end; i += dir)
614                 if (!dev->db_tab->page[i].db_rec) {
615                         page = dev->db_tab->page + i;
616                         goto alloc;
617                 }
618
619         if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
620                 ret = -ENOMEM;
621                 goto out;
622         }
623
624         if (group == 0)
625                 ++dev->db_tab->max_group1;
626         else
627                 --dev->db_tab->min_group2;
628
629         page = dev->db_tab->page + end;
630
631 alloc:
632         page->db_rec = dma_alloc_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
633                                           &page->mapping, GFP_KERNEL);
634         if (!page->db_rec) {
635                 ret = -ENOMEM;
636                 goto out;
637         }
638         memset(page->db_rec, 0, MTHCA_ICM_PAGE_SIZE);
639
640         ret = mthca_MAP_ICM_page(dev, page->mapping,
641                                  mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
642         if (!ret && status)
643                 ret = -EINVAL;
644         if (ret) {
645                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
646                                   page->db_rec, page->mapping);
647                 goto out;
648         }
649
650         bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
651
652 found:
653         j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
654         set_bit(j, page->used);
655
656         if (group == 1)
657                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
658
659         ret = i * MTHCA_DB_REC_PER_PAGE + j;
660
661         page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
662
663         *db = (__be32 *) &page->db_rec[j];
664
665 out:
666         mutex_unlock(&dev->db_tab->mutex);
667
668         return ret;
669 }
670
671 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
672 {
673         int i, j;
674         struct mthca_db_page *page;
675         u8 status;
676
677         i = db_index / MTHCA_DB_REC_PER_PAGE;
678         j = db_index % MTHCA_DB_REC_PER_PAGE;
679
680         page = dev->db_tab->page + i;
681
682         mutex_lock(&dev->db_tab->mutex);
683
684         page->db_rec[j] = 0;
685         if (i >= dev->db_tab->min_group2)
686                 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
687         clear_bit(j, page->used);
688
689         if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
690             i >= dev->db_tab->max_group1 - 1) {
691                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
692
693                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
694                                   page->db_rec, page->mapping);
695                 page->db_rec = NULL;
696
697                 if (i == dev->db_tab->max_group1) {
698                         --dev->db_tab->max_group1;
699                         /* XXX may be able to unmap more pages now */
700                 }
701                 if (i == dev->db_tab->min_group2)
702                         ++dev->db_tab->min_group2;
703         }
704
705         mutex_unlock(&dev->db_tab->mutex);
706 }
707
708 int mthca_init_db_tab(struct mthca_dev *dev)
709 {
710         int i;
711
712         if (!mthca_is_memfree(dev))
713                 return 0;
714
715         dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
716         if (!dev->db_tab)
717                 return -ENOMEM;
718
719         mutex_init(&dev->db_tab->mutex);
720
721         dev->db_tab->npages     = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
722         dev->db_tab->max_group1 = 0;
723         dev->db_tab->min_group2 = dev->db_tab->npages - 1;
724
725         dev->db_tab->page = kmalloc(dev->db_tab->npages *
726                                     sizeof *dev->db_tab->page,
727                                     GFP_KERNEL);
728         if (!dev->db_tab->page) {
729                 kfree(dev->db_tab);
730                 return -ENOMEM;
731         }
732
733         for (i = 0; i < dev->db_tab->npages; ++i)
734                 dev->db_tab->page[i].db_rec = NULL;
735
736         return 0;
737 }
738
739 void mthca_cleanup_db_tab(struct mthca_dev *dev)
740 {
741         int i;
742         u8 status;
743
744         if (!mthca_is_memfree(dev))
745                 return;
746
747         /*
748          * Because we don't always free our UARC pages when they
749          * become empty to make mthca_free_db() simpler we need to
750          * make a sweep through the doorbell pages and free any
751          * leftover pages now.
752          */
753         for (i = 0; i < dev->db_tab->npages; ++i) {
754                 if (!dev->db_tab->page[i].db_rec)
755                         continue;
756
757                 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
758                         mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
759
760                 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
761
762                 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
763                                   dev->db_tab->page[i].db_rec,
764                                   dev->db_tab->page[i].mapping);
765         }
766
767         kfree(dev->db_tab->page);
768         kfree(dev->db_tab);
769 }