2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Cisco Systems. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
38 #include "mthca_memfree.h"
39 #include "mthca_dev.h"
40 #include "mthca_cmd.h"
43 * We allocate in as big chunks as we can, up to a maximum of 256 KB
47 MTHCA_ICM_ALLOC_SIZE = 1 << 18,
48 MTHCA_TABLE_CHUNK_SIZE = 1 << 18
51 struct mthca_user_db_table {
52 struct semaphore mutex;
55 struct scatterlist mem;
60 void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm)
62 struct mthca_icm_chunk *chunk, *tmp;
68 list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
70 pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
71 PCI_DMA_BIDIRECTIONAL);
73 for (i = 0; i < chunk->npages; ++i)
74 __free_pages(chunk->mem[i].page,
75 get_order(chunk->mem[i].length));
83 struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
84 unsigned int gfp_mask)
86 struct mthca_icm *icm;
87 struct mthca_icm_chunk *chunk = NULL;
90 icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
95 INIT_LIST_HEAD(&icm->chunk_list);
97 cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
101 chunk = kmalloc(sizeof *chunk,
102 gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
108 list_add_tail(&chunk->list, &icm->chunk_list);
111 while (1 << cur_order > npages)
114 chunk->mem[chunk->npages].page = alloc_pages(gfp_mask, cur_order);
115 if (chunk->mem[chunk->npages].page) {
116 chunk->mem[chunk->npages].length = PAGE_SIZE << cur_order;
117 chunk->mem[chunk->npages].offset = 0;
119 if (++chunk->npages == MTHCA_ICM_CHUNK_LEN) {
120 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
122 PCI_DMA_BIDIRECTIONAL);
130 npages -= 1 << cur_order;
139 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
141 PCI_DMA_BIDIRECTIONAL);
150 mthca_free_icm(dev, icm);
154 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
156 int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
163 ++table->icm[i]->refcount;
167 table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
168 (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
170 if (!table->icm[i]) {
175 if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
176 &status) || status) {
177 mthca_free_icm(dev, table->icm[i]);
178 table->icm[i] = NULL;
183 ++table->icm[i]->refcount;
190 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
195 if (!mthca_is_memfree(dev))
198 i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
202 if (--table->icm[i]->refcount == 0) {
203 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
204 MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
205 mthca_free_icm(dev, table->icm[i]);
206 table->icm[i] = NULL;
212 void *mthca_table_find(struct mthca_icm_table *table, int obj)
215 struct mthca_icm_chunk *chunk;
216 struct mthca_icm *icm;
217 struct page *page = NULL;
224 idx = (obj & (table->num_obj - 1)) * table->obj_size;
225 icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
226 offset = idx % MTHCA_TABLE_CHUNK_SIZE;
231 list_for_each_entry(chunk, &icm->chunk_list, list) {
232 for (i = 0; i < chunk->npages; ++i) {
233 if (chunk->mem[i].length >= offset) {
234 page = chunk->mem[i].page;
237 offset -= chunk->mem[i].length;
243 return page ? lowmem_page_address(page) + offset : NULL;
246 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
249 int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
252 for (i = start; i <= end; i += inc) {
253 err = mthca_table_get(dev, table, i);
263 mthca_table_put(dev, table, i);
269 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
274 if (!mthca_is_memfree(dev))
277 for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
278 mthca_table_put(dev, table, i);
281 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
282 u64 virt, int obj_size,
283 int nobj, int reserved,
286 struct mthca_icm_table *table;
291 num_icm = obj_size * nobj / MTHCA_TABLE_CHUNK_SIZE;
293 table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
298 table->num_icm = num_icm;
299 table->num_obj = nobj;
300 table->obj_size = obj_size;
301 table->lowmem = use_lowmem;
302 init_MUTEX(&table->mutex);
304 for (i = 0; i < num_icm; ++i)
305 table->icm[i] = NULL;
307 for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
308 table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
309 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
313 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
314 &status) || status) {
315 mthca_free_icm(dev, table->icm[i]);
316 table->icm[i] = NULL;
321 * Add a reference to this ICM chunk so that it never
322 * gets freed (since it contains reserved firmware objects).
324 ++table->icm[i]->refcount;
330 for (i = 0; i < num_icm; ++i)
332 mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
333 MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
334 mthca_free_icm(dev, table->icm[i]);
342 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
347 for (i = 0; i < table->num_icm; ++i)
349 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
350 MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
351 mthca_free_icm(dev, table->icm[i]);
357 static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
359 return dev->uar_table.uarc_base +
360 uar->index * dev->uar_table.uarc_size +
364 int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
365 struct mthca_user_db_table *db_tab, int index, u64 uaddr)
371 if (!mthca_is_memfree(dev))
374 if (index < 0 || index > dev->uar_table.uarc_size / 8)
377 down(&db_tab->mutex);
379 i = index / MTHCA_DB_REC_PER_PAGE;
381 if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE) ||
382 (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
388 if (db_tab->page[i].refcount) {
389 ++db_tab->page[i].refcount;
393 ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
394 &db_tab->page[i].mem.page, NULL);
398 db_tab->page[i].mem.length = 4096;
399 db_tab->page[i].mem.offset = uaddr & ~PAGE_MASK;
401 ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
403 put_page(db_tab->page[i].mem.page);
407 ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
408 mthca_uarc_virt(dev, uar, i), &status);
412 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
413 put_page(db_tab->page[i].mem.page);
417 db_tab->page[i].uvirt = uaddr;
418 db_tab->page[i].refcount = 1;
425 void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
426 struct mthca_user_db_table *db_tab, int index)
428 if (!mthca_is_memfree(dev))
432 * To make our bookkeeping simpler, we don't unmap DB
433 * pages until we clean up the whole db table.
436 down(&db_tab->mutex);
438 --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
443 struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
445 struct mthca_user_db_table *db_tab;
449 if (!mthca_is_memfree(dev))
452 npages = dev->uar_table.uarc_size / 4096;
453 db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
455 return ERR_PTR(-ENOMEM);
457 init_MUTEX(&db_tab->mutex);
458 for (i = 0; i < npages; ++i) {
459 db_tab->page[i].refcount = 0;
460 db_tab->page[i].uvirt = 0;
466 void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
467 struct mthca_user_db_table *db_tab)
472 if (!mthca_is_memfree(dev))
475 for (i = 0; i < dev->uar_table.uarc_size / 4096; ++i) {
476 if (db_tab->page[i].uvirt) {
477 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1, &status);
478 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
479 put_page(db_tab->page[i].mem.page);
484 int mthca_alloc_db(struct mthca_dev *dev, int type, u32 qn, u32 **db)
489 struct mthca_db_page *page;
493 down(&dev->db_tab->mutex);
496 case MTHCA_DB_TYPE_CQ_ARM:
497 case MTHCA_DB_TYPE_SQ:
500 end = dev->db_tab->max_group1;
504 case MTHCA_DB_TYPE_CQ_SET_CI:
505 case MTHCA_DB_TYPE_RQ:
506 case MTHCA_DB_TYPE_SRQ:
508 start = dev->db_tab->npages - 1;
509 end = dev->db_tab->min_group2;
518 for (i = start; i != end; i += dir)
519 if (dev->db_tab->page[i].db_rec &&
520 !bitmap_full(dev->db_tab->page[i].used,
521 MTHCA_DB_REC_PER_PAGE)) {
522 page = dev->db_tab->page + i;
526 if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
531 page = dev->db_tab->page + end;
532 page->db_rec = dma_alloc_coherent(&dev->pdev->dev, 4096,
533 &page->mapping, GFP_KERNEL);
538 memset(page->db_rec, 0, 4096);
540 ret = mthca_MAP_ICM_page(dev, page->mapping,
541 mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
545 dma_free_coherent(&dev->pdev->dev, 4096,
546 page->db_rec, page->mapping);
550 bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
552 ++dev->db_tab->max_group1;
554 --dev->db_tab->min_group2;
557 j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
558 set_bit(j, page->used);
561 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
563 ret = i * MTHCA_DB_REC_PER_PAGE + j;
565 page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
567 *db = (u32 *) &page->db_rec[j];
570 up(&dev->db_tab->mutex);
575 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
578 struct mthca_db_page *page;
581 i = db_index / MTHCA_DB_REC_PER_PAGE;
582 j = db_index % MTHCA_DB_REC_PER_PAGE;
584 page = dev->db_tab->page + i;
586 down(&dev->db_tab->mutex);
589 if (i >= dev->db_tab->min_group2)
590 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
591 clear_bit(j, page->used);
593 if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
594 i >= dev->db_tab->max_group1 - 1) {
595 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
597 dma_free_coherent(&dev->pdev->dev, 4096,
598 page->db_rec, page->mapping);
601 if (i == dev->db_tab->max_group1) {
602 --dev->db_tab->max_group1;
603 /* XXX may be able to unmap more pages now */
605 if (i == dev->db_tab->min_group2)
606 ++dev->db_tab->min_group2;
609 up(&dev->db_tab->mutex);
612 int mthca_init_db_tab(struct mthca_dev *dev)
616 if (!mthca_is_memfree(dev))
619 dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
623 init_MUTEX(&dev->db_tab->mutex);
625 dev->db_tab->npages = dev->uar_table.uarc_size / 4096;
626 dev->db_tab->max_group1 = 0;
627 dev->db_tab->min_group2 = dev->db_tab->npages - 1;
629 dev->db_tab->page = kmalloc(dev->db_tab->npages *
630 sizeof *dev->db_tab->page,
632 if (!dev->db_tab->page) {
637 for (i = 0; i < dev->db_tab->npages; ++i)
638 dev->db_tab->page[i].db_rec = NULL;
643 void mthca_cleanup_db_tab(struct mthca_dev *dev)
648 if (!mthca_is_memfree(dev))
652 * Because we don't always free our UARC pages when they
653 * become empty to make mthca_free_db() simpler we need to
654 * make a sweep through the doorbell pages and free any
655 * leftover pages now.
657 for (i = 0; i < dev->db_tab->npages; ++i) {
658 if (!dev->db_tab->page[i].db_rec)
661 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
662 mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
664 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
666 dma_free_coherent(&dev->pdev->dev, 4096,
667 dev->db_tab->page[i].db_rec,
668 dev->db_tab->page[i].mapping);
671 kfree(dev->db_tab->page);