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.
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:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
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.
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
36 #include <linux/scatterlist.h>
37 #include <linux/sched.h>
38 #include <linux/slab.h>
42 #include "mthca_memfree.h"
43 #include "mthca_dev.h"
44 #include "mthca_cmd.h"
47 * We allocate in as big chunks as we can, up to a maximum of 256 KB
51 MTHCA_ICM_ALLOC_SIZE = 1 << 18,
52 MTHCA_TABLE_CHUNK_SIZE = 1 << 18
55 struct mthca_user_db_table {
59 struct scatterlist mem;
64 static void mthca_free_icm_pages(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
69 pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
70 PCI_DMA_BIDIRECTIONAL);
72 for (i = 0; i < chunk->npages; ++i)
73 __free_pages(sg_page(&chunk->mem[i]),
74 get_order(chunk->mem[i].length));
77 static void mthca_free_icm_coherent(struct mthca_dev *dev, struct mthca_icm_chunk *chunk)
81 for (i = 0; i < chunk->npages; ++i) {
82 dma_free_coherent(&dev->pdev->dev, chunk->mem[i].length,
83 lowmem_page_address(sg_page(&chunk->mem[i])),
84 sg_dma_address(&chunk->mem[i]));
88 void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent)
90 struct mthca_icm_chunk *chunk, *tmp;
95 list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
97 mthca_free_icm_coherent(dev, chunk);
99 mthca_free_icm_pages(dev, chunk);
107 static int mthca_alloc_icm_pages(struct scatterlist *mem, int order, gfp_t gfp_mask)
112 * Use __GFP_ZERO because buggy firmware assumes ICM pages are
113 * cleared, and subtle failures are seen if they aren't.
115 page = alloc_pages(gfp_mask | __GFP_ZERO, order);
119 sg_set_page(mem, page, PAGE_SIZE << order, 0);
123 static int mthca_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
124 int order, gfp_t gfp_mask)
126 void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order, &sg_dma_address(mem),
131 sg_set_buf(mem, buf, PAGE_SIZE << order);
133 sg_dma_len(mem) = PAGE_SIZE << order;
137 struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
138 gfp_t gfp_mask, int coherent)
140 struct mthca_icm *icm;
141 struct mthca_icm_chunk *chunk = NULL;
145 /* We use sg_set_buf for coherent allocs, which assumes low memory */
146 BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
148 icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
153 INIT_LIST_HEAD(&icm->chunk_list);
155 cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
159 chunk = kmalloc(sizeof *chunk,
160 gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
164 sg_init_table(chunk->mem, MTHCA_ICM_CHUNK_LEN);
167 list_add_tail(&chunk->list, &icm->chunk_list);
170 while (1 << cur_order > npages)
174 ret = mthca_alloc_icm_coherent(&dev->pdev->dev,
175 &chunk->mem[chunk->npages],
176 cur_order, gfp_mask);
178 ret = mthca_alloc_icm_pages(&chunk->mem[chunk->npages],
179 cur_order, gfp_mask);
186 else if (chunk->npages == MTHCA_ICM_CHUNK_LEN) {
187 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
189 PCI_DMA_BIDIRECTIONAL);
195 if (chunk->npages == MTHCA_ICM_CHUNK_LEN)
198 npages -= 1 << cur_order;
206 if (!coherent && chunk) {
207 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
209 PCI_DMA_BIDIRECTIONAL);
218 mthca_free_icm(dev, icm, coherent);
222 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
224 int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
228 mutex_lock(&table->mutex);
231 ++table->icm[i]->refcount;
235 table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
236 (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
237 __GFP_NOWARN, table->coherent);
238 if (!table->icm[i]) {
243 if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
244 &status) || status) {
245 mthca_free_icm(dev, table->icm[i], table->coherent);
246 table->icm[i] = NULL;
251 ++table->icm[i]->refcount;
254 mutex_unlock(&table->mutex);
258 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
263 if (!mthca_is_memfree(dev))
266 i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
268 mutex_lock(&table->mutex);
270 if (--table->icm[i]->refcount == 0) {
271 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
272 MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
274 mthca_free_icm(dev, table->icm[i], table->coherent);
275 table->icm[i] = NULL;
278 mutex_unlock(&table->mutex);
281 void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle)
283 int idx, offset, dma_offset, i;
284 struct mthca_icm_chunk *chunk;
285 struct mthca_icm *icm;
286 struct page *page = NULL;
291 mutex_lock(&table->mutex);
293 idx = (obj & (table->num_obj - 1)) * table->obj_size;
294 icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
295 dma_offset = offset = idx % MTHCA_TABLE_CHUNK_SIZE;
300 list_for_each_entry(chunk, &icm->chunk_list, list) {
301 for (i = 0; i < chunk->npages; ++i) {
302 if (dma_handle && dma_offset >= 0) {
303 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
304 *dma_handle = sg_dma_address(&chunk->mem[i]) +
306 dma_offset -= sg_dma_len(&chunk->mem[i]);
308 /* DMA mapping can merge pages but not split them,
309 * so if we found the page, dma_handle has already
310 * been assigned to. */
311 if (chunk->mem[i].length > offset) {
312 page = sg_page(&chunk->mem[i]);
315 offset -= chunk->mem[i].length;
320 mutex_unlock(&table->mutex);
321 return page ? lowmem_page_address(page) + offset : NULL;
324 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
327 int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
330 for (i = start; i <= end; i += inc) {
331 err = mthca_table_get(dev, table, i);
341 mthca_table_put(dev, table, i);
347 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
352 if (!mthca_is_memfree(dev))
355 for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
356 mthca_table_put(dev, table, i);
359 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
360 u64 virt, int obj_size,
361 int nobj, int reserved,
362 int use_lowmem, int use_coherent)
364 struct mthca_icm_table *table;
371 obj_per_chunk = MTHCA_TABLE_CHUNK_SIZE / obj_size;
372 num_icm = DIV_ROUND_UP(nobj, obj_per_chunk);
374 table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
379 table->num_icm = num_icm;
380 table->num_obj = nobj;
381 table->obj_size = obj_size;
382 table->lowmem = use_lowmem;
383 table->coherent = use_coherent;
384 mutex_init(&table->mutex);
386 for (i = 0; i < num_icm; ++i)
387 table->icm[i] = NULL;
389 for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
390 chunk_size = MTHCA_TABLE_CHUNK_SIZE;
391 if ((i + 1) * MTHCA_TABLE_CHUNK_SIZE > nobj * obj_size)
392 chunk_size = nobj * obj_size - i * MTHCA_TABLE_CHUNK_SIZE;
394 table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
395 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
396 __GFP_NOWARN, use_coherent);
399 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
400 &status) || status) {
401 mthca_free_icm(dev, table->icm[i], table->coherent);
402 table->icm[i] = NULL;
407 * Add a reference to this ICM chunk so that it never
408 * gets freed (since it contains reserved firmware objects).
410 ++table->icm[i]->refcount;
416 for (i = 0; i < num_icm; ++i)
418 mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
419 MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
421 mthca_free_icm(dev, table->icm[i], table->coherent);
429 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
434 for (i = 0; i < table->num_icm; ++i)
436 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
437 MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
439 mthca_free_icm(dev, table->icm[i], table->coherent);
445 static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
447 return dev->uar_table.uarc_base +
448 uar->index * dev->uar_table.uarc_size +
449 page * MTHCA_ICM_PAGE_SIZE;
452 int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
453 struct mthca_user_db_table *db_tab, int index, u64 uaddr)
455 struct page *pages[1];
460 if (!mthca_is_memfree(dev))
463 if (index < 0 || index > dev->uar_table.uarc_size / 8)
466 mutex_lock(&db_tab->mutex);
468 i = index / MTHCA_DB_REC_PER_PAGE;
470 if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE) ||
471 (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
477 if (db_tab->page[i].refcount) {
478 ++db_tab->page[i].refcount;
482 ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
487 sg_set_page(&db_tab->page[i].mem, pages[0], MTHCA_ICM_PAGE_SIZE,
490 ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
496 ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
497 mthca_uarc_virt(dev, uar, i), &status);
501 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
502 put_page(sg_page(&db_tab->page[i].mem));
506 db_tab->page[i].uvirt = uaddr;
507 db_tab->page[i].refcount = 1;
510 mutex_unlock(&db_tab->mutex);
514 void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
515 struct mthca_user_db_table *db_tab, int index)
517 if (!mthca_is_memfree(dev))
521 * To make our bookkeeping simpler, we don't unmap DB
522 * pages until we clean up the whole db table.
525 mutex_lock(&db_tab->mutex);
527 --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
529 mutex_unlock(&db_tab->mutex);
532 struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
534 struct mthca_user_db_table *db_tab;
538 if (!mthca_is_memfree(dev))
541 npages = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
542 db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
544 return ERR_PTR(-ENOMEM);
546 mutex_init(&db_tab->mutex);
547 for (i = 0; i < npages; ++i) {
548 db_tab->page[i].refcount = 0;
549 db_tab->page[i].uvirt = 0;
550 sg_init_table(&db_tab->page[i].mem, 1);
556 void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
557 struct mthca_user_db_table *db_tab)
562 if (!mthca_is_memfree(dev))
565 for (i = 0; i < dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE; ++i) {
566 if (db_tab->page[i].uvirt) {
567 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1, &status);
568 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
569 put_page(sg_page(&db_tab->page[i].mem));
576 int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
582 struct mthca_db_page *page;
586 mutex_lock(&dev->db_tab->mutex);
589 case MTHCA_DB_TYPE_CQ_ARM:
590 case MTHCA_DB_TYPE_SQ:
593 end = dev->db_tab->max_group1;
597 case MTHCA_DB_TYPE_CQ_SET_CI:
598 case MTHCA_DB_TYPE_RQ:
599 case MTHCA_DB_TYPE_SRQ:
601 start = dev->db_tab->npages - 1;
602 end = dev->db_tab->min_group2;
611 for (i = start; i != end; i += dir)
612 if (dev->db_tab->page[i].db_rec &&
613 !bitmap_full(dev->db_tab->page[i].used,
614 MTHCA_DB_REC_PER_PAGE)) {
615 page = dev->db_tab->page + i;
619 for (i = start; i != end; i += dir)
620 if (!dev->db_tab->page[i].db_rec) {
621 page = dev->db_tab->page + i;
625 if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
631 ++dev->db_tab->max_group1;
633 --dev->db_tab->min_group2;
635 page = dev->db_tab->page + end;
638 page->db_rec = dma_alloc_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
639 &page->mapping, GFP_KERNEL);
644 memset(page->db_rec, 0, MTHCA_ICM_PAGE_SIZE);
646 ret = mthca_MAP_ICM_page(dev, page->mapping,
647 mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
651 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
652 page->db_rec, page->mapping);
656 bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
659 j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
660 set_bit(j, page->used);
663 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
665 ret = i * MTHCA_DB_REC_PER_PAGE + j;
667 page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
669 *db = (__be32 *) &page->db_rec[j];
672 mutex_unlock(&dev->db_tab->mutex);
677 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
680 struct mthca_db_page *page;
683 i = db_index / MTHCA_DB_REC_PER_PAGE;
684 j = db_index % MTHCA_DB_REC_PER_PAGE;
686 page = dev->db_tab->page + i;
688 mutex_lock(&dev->db_tab->mutex);
691 if (i >= dev->db_tab->min_group2)
692 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
693 clear_bit(j, page->used);
695 if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
696 i >= dev->db_tab->max_group1 - 1) {
697 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
699 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
700 page->db_rec, page->mapping);
703 if (i == dev->db_tab->max_group1) {
704 --dev->db_tab->max_group1;
705 /* XXX may be able to unmap more pages now */
707 if (i == dev->db_tab->min_group2)
708 ++dev->db_tab->min_group2;
711 mutex_unlock(&dev->db_tab->mutex);
714 int mthca_init_db_tab(struct mthca_dev *dev)
718 if (!mthca_is_memfree(dev))
721 dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
725 mutex_init(&dev->db_tab->mutex);
727 dev->db_tab->npages = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
728 dev->db_tab->max_group1 = 0;
729 dev->db_tab->min_group2 = dev->db_tab->npages - 1;
731 dev->db_tab->page = kmalloc(dev->db_tab->npages *
732 sizeof *dev->db_tab->page,
734 if (!dev->db_tab->page) {
739 for (i = 0; i < dev->db_tab->npages; ++i)
740 dev->db_tab->page[i].db_rec = NULL;
745 void mthca_cleanup_db_tab(struct mthca_dev *dev)
750 if (!mthca_is_memfree(dev))
754 * Because we don't always free our UARC pages when they
755 * become empty to make mthca_free_db() simpler we need to
756 * make a sweep through the doorbell pages and free any
757 * leftover pages now.
759 for (i = 0; i < dev->db_tab->npages; ++i) {
760 if (!dev->db_tab->page[i].db_rec)
763 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
764 mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
766 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
768 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
769 dev->db_tab->page[i].db_rec,
770 dev->db_tab->page[i].mapping);
773 kfree(dev->db_tab->page);