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
39 #include "mthca_memfree.h"
40 #include "mthca_dev.h"
41 #include "mthca_cmd.h"
44 * We allocate in as big chunks as we can, up to a maximum of 256 KB
48 MTHCA_ICM_ALLOC_SIZE = 1 << 18,
49 MTHCA_TABLE_CHUNK_SIZE = 1 << 18
52 struct mthca_user_db_table {
53 struct semaphore mutex;
56 struct scatterlist mem;
61 void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm)
63 struct mthca_icm_chunk *chunk, *tmp;
69 list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
71 pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
72 PCI_DMA_BIDIRECTIONAL);
74 for (i = 0; i < chunk->npages; ++i)
75 __free_pages(chunk->mem[i].page,
76 get_order(chunk->mem[i].length));
84 struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages,
85 unsigned int gfp_mask)
87 struct mthca_icm *icm;
88 struct mthca_icm_chunk *chunk = NULL;
91 icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
96 INIT_LIST_HEAD(&icm->chunk_list);
98 cur_order = get_order(MTHCA_ICM_ALLOC_SIZE);
102 chunk = kmalloc(sizeof *chunk,
103 gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
109 list_add_tail(&chunk->list, &icm->chunk_list);
112 while (1 << cur_order > npages)
115 chunk->mem[chunk->npages].page = alloc_pages(gfp_mask, cur_order);
116 if (chunk->mem[chunk->npages].page) {
117 chunk->mem[chunk->npages].length = PAGE_SIZE << cur_order;
118 chunk->mem[chunk->npages].offset = 0;
120 if (++chunk->npages == MTHCA_ICM_CHUNK_LEN) {
121 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
123 PCI_DMA_BIDIRECTIONAL);
131 npages -= 1 << cur_order;
140 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
142 PCI_DMA_BIDIRECTIONAL);
151 mthca_free_icm(dev, icm);
155 int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
157 int i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
164 ++table->icm[i]->refcount;
168 table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
169 (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
171 if (!table->icm[i]) {
176 if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
177 &status) || status) {
178 mthca_free_icm(dev, table->icm[i]);
179 table->icm[i] = NULL;
184 ++table->icm[i]->refcount;
191 void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj)
196 if (!mthca_is_memfree(dev))
199 i = (obj & (table->num_obj - 1)) * table->obj_size / MTHCA_TABLE_CHUNK_SIZE;
203 if (--table->icm[i]->refcount == 0) {
204 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
205 MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
206 mthca_free_icm(dev, table->icm[i]);
207 table->icm[i] = NULL;
213 void *mthca_table_find(struct mthca_icm_table *table, int obj)
216 struct mthca_icm_chunk *chunk;
217 struct mthca_icm *icm;
218 struct page *page = NULL;
225 idx = (obj & (table->num_obj - 1)) * table->obj_size;
226 icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE];
227 offset = idx % MTHCA_TABLE_CHUNK_SIZE;
232 list_for_each_entry(chunk, &icm->chunk_list, list) {
233 for (i = 0; i < chunk->npages; ++i) {
234 if (chunk->mem[i].length >= offset) {
235 page = chunk->mem[i].page;
238 offset -= chunk->mem[i].length;
244 return page ? lowmem_page_address(page) + offset : NULL;
247 int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table,
250 int inc = MTHCA_TABLE_CHUNK_SIZE / table->obj_size;
253 for (i = start; i <= end; i += inc) {
254 err = mthca_table_get(dev, table, i);
264 mthca_table_put(dev, table, i);
270 void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table,
275 if (!mthca_is_memfree(dev))
278 for (i = start; i <= end; i += MTHCA_TABLE_CHUNK_SIZE / table->obj_size)
279 mthca_table_put(dev, table, i);
282 struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev,
283 u64 virt, int obj_size,
284 int nobj, int reserved,
287 struct mthca_icm_table *table;
292 num_icm = obj_size * nobj / MTHCA_TABLE_CHUNK_SIZE;
294 table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
299 table->num_icm = num_icm;
300 table->num_obj = nobj;
301 table->obj_size = obj_size;
302 table->lowmem = use_lowmem;
303 init_MUTEX(&table->mutex);
305 for (i = 0; i < num_icm; ++i)
306 table->icm[i] = NULL;
308 for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
309 table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
310 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
314 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
315 &status) || status) {
316 mthca_free_icm(dev, table->icm[i]);
317 table->icm[i] = NULL;
322 * Add a reference to this ICM chunk so that it never
323 * gets freed (since it contains reserved firmware objects).
325 ++table->icm[i]->refcount;
331 for (i = 0; i < num_icm; ++i)
333 mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
334 MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
335 mthca_free_icm(dev, table->icm[i]);
343 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
348 for (i = 0; i < table->num_icm; ++i)
350 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
351 MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
352 mthca_free_icm(dev, table->icm[i]);
358 static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
360 return dev->uar_table.uarc_base +
361 uar->index * dev->uar_table.uarc_size +
365 int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
366 struct mthca_user_db_table *db_tab, int index, u64 uaddr)
372 if (!mthca_is_memfree(dev))
375 if (index < 0 || index > dev->uar_table.uarc_size / 8)
378 down(&db_tab->mutex);
380 i = index / MTHCA_DB_REC_PER_PAGE;
382 if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE) ||
383 (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
389 if (db_tab->page[i].refcount) {
390 ++db_tab->page[i].refcount;
394 ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
395 &db_tab->page[i].mem.page, NULL);
399 db_tab->page[i].mem.length = 4096;
400 db_tab->page[i].mem.offset = uaddr & ~PAGE_MASK;
402 ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
404 put_page(db_tab->page[i].mem.page);
408 ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
409 mthca_uarc_virt(dev, uar, i), &status);
413 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
414 put_page(db_tab->page[i].mem.page);
418 db_tab->page[i].uvirt = uaddr;
419 db_tab->page[i].refcount = 1;
426 void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
427 struct mthca_user_db_table *db_tab, int index)
429 if (!mthca_is_memfree(dev))
433 * To make our bookkeeping simpler, we don't unmap DB
434 * pages until we clean up the whole db table.
437 down(&db_tab->mutex);
439 --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
444 struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
446 struct mthca_user_db_table *db_tab;
450 if (!mthca_is_memfree(dev))
453 npages = dev->uar_table.uarc_size / 4096;
454 db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
456 return ERR_PTR(-ENOMEM);
458 init_MUTEX(&db_tab->mutex);
459 for (i = 0; i < npages; ++i) {
460 db_tab->page[i].refcount = 0;
461 db_tab->page[i].uvirt = 0;
467 void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
468 struct mthca_user_db_table *db_tab)
473 if (!mthca_is_memfree(dev))
476 for (i = 0; i < dev->uar_table.uarc_size / 4096; ++i) {
477 if (db_tab->page[i].uvirt) {
478 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1, &status);
479 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
480 put_page(db_tab->page[i].mem.page);
485 int mthca_alloc_db(struct mthca_dev *dev, int type, u32 qn, __be32 **db)
490 struct mthca_db_page *page;
494 down(&dev->db_tab->mutex);
497 case MTHCA_DB_TYPE_CQ_ARM:
498 case MTHCA_DB_TYPE_SQ:
501 end = dev->db_tab->max_group1;
505 case MTHCA_DB_TYPE_CQ_SET_CI:
506 case MTHCA_DB_TYPE_RQ:
507 case MTHCA_DB_TYPE_SRQ:
509 start = dev->db_tab->npages - 1;
510 end = dev->db_tab->min_group2;
519 for (i = start; i != end; i += dir)
520 if (dev->db_tab->page[i].db_rec &&
521 !bitmap_full(dev->db_tab->page[i].used,
522 MTHCA_DB_REC_PER_PAGE)) {
523 page = dev->db_tab->page + i;
527 if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
532 page = dev->db_tab->page + end;
533 page->db_rec = dma_alloc_coherent(&dev->pdev->dev, 4096,
534 &page->mapping, GFP_KERNEL);
539 memset(page->db_rec, 0, 4096);
541 ret = mthca_MAP_ICM_page(dev, page->mapping,
542 mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
546 dma_free_coherent(&dev->pdev->dev, 4096,
547 page->db_rec, page->mapping);
551 bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
553 ++dev->db_tab->max_group1;
555 --dev->db_tab->min_group2;
558 j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
559 set_bit(j, page->used);
562 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
564 ret = i * MTHCA_DB_REC_PER_PAGE + j;
566 page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
568 *db = (__be32 *) &page->db_rec[j];
571 up(&dev->db_tab->mutex);
576 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
579 struct mthca_db_page *page;
582 i = db_index / MTHCA_DB_REC_PER_PAGE;
583 j = db_index % MTHCA_DB_REC_PER_PAGE;
585 page = dev->db_tab->page + i;
587 down(&dev->db_tab->mutex);
590 if (i >= dev->db_tab->min_group2)
591 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
592 clear_bit(j, page->used);
594 if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
595 i >= dev->db_tab->max_group1 - 1) {
596 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
598 dma_free_coherent(&dev->pdev->dev, 4096,
599 page->db_rec, page->mapping);
602 if (i == dev->db_tab->max_group1) {
603 --dev->db_tab->max_group1;
604 /* XXX may be able to unmap more pages now */
606 if (i == dev->db_tab->min_group2)
607 ++dev->db_tab->min_group2;
610 up(&dev->db_tab->mutex);
613 int mthca_init_db_tab(struct mthca_dev *dev)
617 if (!mthca_is_memfree(dev))
620 dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
624 init_MUTEX(&dev->db_tab->mutex);
626 dev->db_tab->npages = dev->uar_table.uarc_size / 4096;
627 dev->db_tab->max_group1 = 0;
628 dev->db_tab->min_group2 = dev->db_tab->npages - 1;
630 dev->db_tab->page = kmalloc(dev->db_tab->npages *
631 sizeof *dev->db_tab->page,
633 if (!dev->db_tab->page) {
638 for (i = 0; i < dev->db_tab->npages; ++i)
639 dev->db_tab->page[i].db_rec = NULL;
644 void mthca_cleanup_db_tab(struct mthca_dev *dev)
649 if (!mthca_is_memfree(dev))
653 * Because we don't always free our UARC pages when they
654 * become empty to make mthca_free_db() simpler we need to
655 * make a sweep through the doorbell pages and free any
656 * leftover pages now.
658 for (i = 0; i < dev->db_tab->npages; ++i) {
659 if (!dev->db_tab->page[i].db_rec)
662 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
663 mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
665 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
667 dma_free_coherent(&dev->pdev->dev, 4096,
668 dev->db_tab->page[i].db_rec,
669 dev->db_tab->page[i].mapping);
672 kfree(dev->db_tab->page);