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
38 #include <linux/scatterlist.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(chunk->mem[i].page,
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(chunk->mem[i].page),
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)
109 mem->page = alloc_pages(gfp_mask, order);
113 mem->length = PAGE_SIZE << order;
118 static int mthca_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
119 int order, gfp_t gfp_mask)
121 void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order, &sg_dma_address(mem),
126 sg_set_buf(mem, buf, PAGE_SIZE << order);
128 sg_dma_len(mem) = PAGE_SIZE << order;
132 struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
133 gfp_t gfp_mask, int coherent)
135 struct mthca_icm *icm;
136 struct mthca_icm_chunk *chunk = NULL;
140 /* We use sg_set_buf for coherent allocs, which assumes low memory */
141 BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
143 icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
148 INIT_LIST_HEAD(&icm->chunk_list);
150 cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
154 chunk = kmalloc(sizeof *chunk,
155 gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
161 list_add_tail(&chunk->list, &icm->chunk_list);
164 while (1 << cur_order > npages)
168 ret = mthca_alloc_icm_coherent(&dev->pdev->dev,
169 &chunk->mem[chunk->npages],
170 cur_order, gfp_mask);
172 ret = mthca_alloc_icm_pages(&chunk->mem[chunk->npages],
173 cur_order, gfp_mask);
178 if (!coherent && chunk->npages == MTHCA_ICM_CHUNK_LEN) {
179 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
181 PCI_DMA_BIDIRECTIONAL);
187 if (chunk->npages == MTHCA_ICM_CHUNK_LEN)
190 npages -= 1 << cur_order;
198 if (!coherent && chunk) {
199 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
201 PCI_DMA_BIDIRECTIONAL);
210 mthca_free_icm(dev, icm, coherent);
214 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
216 int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
220 mutex_lock(&table->mutex);
223 ++table->icm[i]->refcount;
227 table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
228 (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
229 __GFP_NOWARN, table->coherent);
230 if (!table->icm[i]) {
235 if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
236 &status) || status) {
237 mthca_free_icm(dev, table->icm[i], table->coherent);
238 table->icm[i] = NULL;
243 ++table->icm[i]->refcount;
246 mutex_unlock(&table->mutex);
250 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
255 if (!mthca_is_memfree(dev))
258 i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
260 mutex_lock(&table->mutex);
262 if (--table->icm[i]->refcount == 0) {
263 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
264 MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
266 mthca_free_icm(dev, table->icm[i], table->coherent);
267 table->icm[i] = NULL;
270 mutex_unlock(&table->mutex);
273 void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle)
275 int idx, offset, dma_offset, i;
276 struct mthca_icm_chunk *chunk;
277 struct mthca_icm *icm;
278 struct page *page = NULL;
283 mutex_lock(&table->mutex);
285 idx = (obj & (table->num_obj - 1)) * table->obj_size;
286 icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
287 dma_offset = offset = idx % MTHCA_TABLE_CHUNK_SIZE;
292 list_for_each_entry(chunk, &icm->chunk_list, list) {
293 for (i = 0; i < chunk->npages; ++i) {
294 if (dma_handle && dma_offset >= 0) {
295 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
296 *dma_handle = sg_dma_address(&chunk->mem[i]) +
298 dma_offset -= sg_dma_len(&chunk->mem[i]);
300 /* DMA mapping can merge pages but not split them,
301 * so if we found the page, dma_handle has already
302 * been assigned to. */
303 if (chunk->mem[i].length > offset) {
304 page = chunk->mem[i].page;
307 offset -= chunk->mem[i].length;
312 mutex_unlock(&table->mutex);
313 return page ? lowmem_page_address(page) + offset : NULL;
316 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
319 int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
322 for (i = start; i <= end; i += inc) {
323 err = mthca_table_get(dev, table, i);
333 mthca_table_put(dev, table, i);
339 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
344 if (!mthca_is_memfree(dev))
347 for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
348 mthca_table_put(dev, table, i);
351 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
352 u64 virt, int obj_size,
353 int nobj, int reserved,
354 int use_lowmem, int use_coherent)
356 struct mthca_icm_table *table;
362 num_icm = (obj_size * nobj + MTHCA_TABLE_CHUNK_SIZE - 1) / MTHCA_TABLE_CHUNK_SIZE;
364 table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
369 table->num_icm = num_icm;
370 table->num_obj = nobj;
371 table->obj_size = obj_size;
372 table->lowmem = use_lowmem;
373 table->coherent = use_coherent;
374 mutex_init(&table->mutex);
376 for (i = 0; i < num_icm; ++i)
377 table->icm[i] = NULL;
379 for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
380 chunk_size = MTHCA_TABLE_CHUNK_SIZE;
381 if ((i + 1) * MTHCA_TABLE_CHUNK_SIZE > nobj * obj_size)
382 chunk_size = nobj * obj_size - i * MTHCA_TABLE_CHUNK_SIZE;
384 table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
385 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
386 __GFP_NOWARN, use_coherent);
389 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
390 &status) || status) {
391 mthca_free_icm(dev, table->icm[i], table->coherent);
392 table->icm[i] = NULL;
397 * Add a reference to this ICM chunk so that it never
398 * gets freed (since it contains reserved firmware objects).
400 ++table->icm[i]->refcount;
406 for (i = 0; i < num_icm; ++i)
408 mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
409 MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
411 mthca_free_icm(dev, table->icm[i], table->coherent);
419 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
424 for (i = 0; i < table->num_icm; ++i)
426 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
427 MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE,
429 mthca_free_icm(dev, table->icm[i], table->coherent);
435 static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
437 return dev->uar_table.uarc_base +
438 uar->index * dev->uar_table.uarc_size +
439 page * MTHCA_ICM_PAGE_SIZE;
442 int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
443 struct mthca_user_db_table *db_tab, int index, u64 uaddr)
449 if (!mthca_is_memfree(dev))
452 if (index < 0 || index > dev->uar_table.uarc_size / 8)
455 mutex_lock(&db_tab->mutex);
457 i = index / MTHCA_DB_REC_PER_PAGE;
459 if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE) ||
460 (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
466 if (db_tab->page[i].refcount) {
467 ++db_tab->page[i].refcount;
471 ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
472 &db_tab->page[i].mem.page, NULL);
476 db_tab->page[i].mem.length = MTHCA_ICM_PAGE_SIZE;
477 db_tab->page[i].mem.offset = uaddr & ~PAGE_MASK;
479 ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
481 put_page(db_tab->page[i].mem.page);
485 ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
486 mthca_uarc_virt(dev, uar, i), &status);
490 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
491 put_page(db_tab->page[i].mem.page);
495 db_tab->page[i].uvirt = uaddr;
496 db_tab->page[i].refcount = 1;
499 mutex_unlock(&db_tab->mutex);
503 void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
504 struct mthca_user_db_table *db_tab, int index)
506 if (!mthca_is_memfree(dev))
510 * To make our bookkeeping simpler, we don't unmap DB
511 * pages until we clean up the whole db table.
514 mutex_lock(&db_tab->mutex);
516 --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
518 mutex_unlock(&db_tab->mutex);
521 struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
523 struct mthca_user_db_table *db_tab;
527 if (!mthca_is_memfree(dev))
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);
533 return ERR_PTR(-ENOMEM);
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;
544 void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
545 struct mthca_user_db_table *db_tab)
550 if (!mthca_is_memfree(dev))
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, &status);
556 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
557 put_page(db_tab->page[i].mem.page);
564 int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
570 struct mthca_db_page *page;
574 mutex_lock(&dev->db_tab->mutex);
577 case MTHCA_DB_TYPE_CQ_ARM:
578 case MTHCA_DB_TYPE_SQ:
581 end = dev->db_tab->max_group1;
585 case MTHCA_DB_TYPE_CQ_SET_CI:
586 case MTHCA_DB_TYPE_RQ:
587 case MTHCA_DB_TYPE_SRQ:
589 start = dev->db_tab->npages - 1;
590 end = dev->db_tab->min_group2;
599 for (i = start; i != end; i += dir)
600 if (dev->db_tab->page[i].db_rec &&
601 !bitmap_full(dev->db_tab->page[i].used,
602 MTHCA_DB_REC_PER_PAGE)) {
603 page = dev->db_tab->page + i;
607 for (i = start; i != end; i += dir)
608 if (!dev->db_tab->page[i].db_rec) {
609 page = dev->db_tab->page + i;
613 if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
619 ++dev->db_tab->max_group1;
621 --dev->db_tab->min_group2;
623 page = dev->db_tab->page + end;
626 page->db_rec = dma_alloc_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
627 &page->mapping, GFP_KERNEL);
632 memset(page->db_rec, 0, MTHCA_ICM_PAGE_SIZE);
634 ret = mthca_MAP_ICM_page(dev, page->mapping,
635 mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
639 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
640 page->db_rec, page->mapping);
644 bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
647 j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
648 set_bit(j, page->used);
651 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
653 ret = i * MTHCA_DB_REC_PER_PAGE + j;
655 page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
657 *db = (__be32 *) &page->db_rec[j];
660 mutex_unlock(&dev->db_tab->mutex);
665 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
668 struct mthca_db_page *page;
671 i = db_index / MTHCA_DB_REC_PER_PAGE;
672 j = db_index % MTHCA_DB_REC_PER_PAGE;
674 page = dev->db_tab->page + i;
676 mutex_lock(&dev->db_tab->mutex);
679 if (i >= dev->db_tab->min_group2)
680 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
681 clear_bit(j, page->used);
683 if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
684 i >= dev->db_tab->max_group1 - 1) {
685 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
687 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
688 page->db_rec, page->mapping);
691 if (i == dev->db_tab->max_group1) {
692 --dev->db_tab->max_group1;
693 /* XXX may be able to unmap more pages now */
695 if (i == dev->db_tab->min_group2)
696 ++dev->db_tab->min_group2;
699 mutex_unlock(&dev->db_tab->mutex);
702 int mthca_init_db_tab(struct mthca_dev *dev)
706 if (!mthca_is_memfree(dev))
709 dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
713 mutex_init(&dev->db_tab->mutex);
715 dev->db_tab->npages = dev->uar_table.uarc_size / MTHCA_ICM_PAGE_SIZE;
716 dev->db_tab->max_group1 = 0;
717 dev->db_tab->min_group2 = dev->db_tab->npages - 1;
719 dev->db_tab->page = kmalloc(dev->db_tab->npages *
720 sizeof *dev->db_tab->page,
722 if (!dev->db_tab->page) {
727 for (i = 0; i < dev->db_tab->npages; ++i)
728 dev->db_tab->page[i].db_rec = NULL;
733 void mthca_cleanup_db_tab(struct mthca_dev *dev)
738 if (!mthca_is_memfree(dev))
742 * Because we don't always free our UARC pages when they
743 * become empty to make mthca_free_db() simpler we need to
744 * make a sweep through the doorbell pages and free any
745 * leftover pages now.
747 for (i = 0; i < dev->db_tab->npages; ++i) {
748 if (!dev->db_tab->page[i].db_rec)
751 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
752 mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
754 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
756 dma_free_coherent(&dev->pdev->dev, MTHCA_ICM_PAGE_SIZE,
757 dev->db_tab->page[i].db_rec,
758 dev->db_tab->page[i].mapping);
761 kfree(dev->db_tab->page);