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