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,
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;
293 num_icm = (obj_size * nobj + MTHCA_TABLE_CHUNK_SIZE - 1) / MTHCA_TABLE_CHUNK_SIZE;
295 table = kmalloc(sizeof *table + num_icm * sizeof *table->icm, GFP_KERNEL);
300 table->num_icm = num_icm;
301 table->num_obj = nobj;
302 table->obj_size = obj_size;
303 table->lowmem = use_lowmem;
304 init_MUTEX(&table->mutex);
306 for (i = 0; i < num_icm; ++i)
307 table->icm[i] = NULL;
309 for (i = 0; i * MTHCA_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
310 chunk_size = MTHCA_TABLE_CHUNK_SIZE;
311 if ((i + 1) * MTHCA_TABLE_CHUNK_SIZE > nobj * obj_size)
312 chunk_size = nobj * obj_size - i * MTHCA_TABLE_CHUNK_SIZE;
314 table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
315 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
319 if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE,
320 &status) || status) {
321 mthca_free_icm(dev, table->icm[i]);
322 table->icm[i] = NULL;
327 * Add a reference to this ICM chunk so that it never
328 * gets freed (since it contains reserved firmware objects).
330 ++table->icm[i]->refcount;
336 for (i = 0; i < num_icm; ++i)
338 mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE,
339 MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
340 mthca_free_icm(dev, table->icm[i]);
348 void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table)
353 for (i = 0; i < table->num_icm; ++i)
355 mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE,
356 MTHCA_TABLE_CHUNK_SIZE >> 12, &status);
357 mthca_free_icm(dev, table->icm[i]);
363 static u64 mthca_uarc_virt(struct mthca_dev *dev, struct mthca_uar *uar, int page)
365 return dev->uar_table.uarc_base +
366 uar->index * dev->uar_table.uarc_size +
370 int mthca_map_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
371 struct mthca_user_db_table *db_tab, int index, u64 uaddr)
377 if (!mthca_is_memfree(dev))
380 if (index < 0 || index > dev->uar_table.uarc_size / 8)
383 down(&db_tab->mutex);
385 i = index / MTHCA_DB_REC_PER_PAGE;
387 if ((db_tab->page[i].refcount >= MTHCA_DB_REC_PER_PAGE) ||
388 (db_tab->page[i].uvirt && db_tab->page[i].uvirt != uaddr) ||
394 if (db_tab->page[i].refcount) {
395 ++db_tab->page[i].refcount;
399 ret = get_user_pages(current, current->mm, uaddr & PAGE_MASK, 1, 1, 0,
400 &db_tab->page[i].mem.page, NULL);
404 db_tab->page[i].mem.length = 4096;
405 db_tab->page[i].mem.offset = uaddr & ~PAGE_MASK;
407 ret = pci_map_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
409 put_page(db_tab->page[i].mem.page);
413 ret = mthca_MAP_ICM_page(dev, sg_dma_address(&db_tab->page[i].mem),
414 mthca_uarc_virt(dev, uar, i), &status);
418 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
419 put_page(db_tab->page[i].mem.page);
423 db_tab->page[i].uvirt = uaddr;
424 db_tab->page[i].refcount = 1;
431 void mthca_unmap_user_db(struct mthca_dev *dev, struct mthca_uar *uar,
432 struct mthca_user_db_table *db_tab, int index)
434 if (!mthca_is_memfree(dev))
438 * To make our bookkeeping simpler, we don't unmap DB
439 * pages until we clean up the whole db table.
442 down(&db_tab->mutex);
444 --db_tab->page[index / MTHCA_DB_REC_PER_PAGE].refcount;
449 struct mthca_user_db_table *mthca_init_user_db_tab(struct mthca_dev *dev)
451 struct mthca_user_db_table *db_tab;
455 if (!mthca_is_memfree(dev))
458 npages = dev->uar_table.uarc_size / 4096;
459 db_tab = kmalloc(sizeof *db_tab + npages * sizeof *db_tab->page, GFP_KERNEL);
461 return ERR_PTR(-ENOMEM);
463 init_MUTEX(&db_tab->mutex);
464 for (i = 0; i < npages; ++i) {
465 db_tab->page[i].refcount = 0;
466 db_tab->page[i].uvirt = 0;
472 void mthca_cleanup_user_db_tab(struct mthca_dev *dev, struct mthca_uar *uar,
473 struct mthca_user_db_table *db_tab)
478 if (!mthca_is_memfree(dev))
481 for (i = 0; i < dev->uar_table.uarc_size / 4096; ++i) {
482 if (db_tab->page[i].uvirt) {
483 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, uar, i), 1, &status);
484 pci_unmap_sg(dev->pdev, &db_tab->page[i].mem, 1, PCI_DMA_TODEVICE);
485 put_page(db_tab->page[i].mem.page);
490 int mthca_alloc_db(struct mthca_dev *dev, enum mthca_db_type type,
496 struct mthca_db_page *page;
500 down(&dev->db_tab->mutex);
503 case MTHCA_DB_TYPE_CQ_ARM:
504 case MTHCA_DB_TYPE_SQ:
507 end = dev->db_tab->max_group1;
511 case MTHCA_DB_TYPE_CQ_SET_CI:
512 case MTHCA_DB_TYPE_RQ:
513 case MTHCA_DB_TYPE_SRQ:
515 start = dev->db_tab->npages - 1;
516 end = dev->db_tab->min_group2;
525 for (i = start; i != end; i += dir)
526 if (dev->db_tab->page[i].db_rec &&
527 !bitmap_full(dev->db_tab->page[i].used,
528 MTHCA_DB_REC_PER_PAGE)) {
529 page = dev->db_tab->page + i;
533 for (i = start; i != end; i += dir)
534 if (!dev->db_tab->page[i].db_rec) {
535 page = dev->db_tab->page + i;
539 if (dev->db_tab->max_group1 >= dev->db_tab->min_group2 - 1) {
545 ++dev->db_tab->max_group1;
547 --dev->db_tab->min_group2;
549 page = dev->db_tab->page + end;
552 page->db_rec = dma_alloc_coherent(&dev->pdev->dev, 4096,
553 &page->mapping, GFP_KERNEL);
558 memset(page->db_rec, 0, 4096);
560 ret = mthca_MAP_ICM_page(dev, page->mapping,
561 mthca_uarc_virt(dev, &dev->driver_uar, i), &status);
565 dma_free_coherent(&dev->pdev->dev, 4096,
566 page->db_rec, page->mapping);
570 bitmap_zero(page->used, MTHCA_DB_REC_PER_PAGE);
573 j = find_first_zero_bit(page->used, MTHCA_DB_REC_PER_PAGE);
574 set_bit(j, page->used);
577 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
579 ret = i * MTHCA_DB_REC_PER_PAGE + j;
581 page->db_rec[j] = cpu_to_be64((qn << 8) | (type << 5));
583 *db = (__be32 *) &page->db_rec[j];
586 up(&dev->db_tab->mutex);
591 void mthca_free_db(struct mthca_dev *dev, int type, int db_index)
594 struct mthca_db_page *page;
597 i = db_index / MTHCA_DB_REC_PER_PAGE;
598 j = db_index % MTHCA_DB_REC_PER_PAGE;
600 page = dev->db_tab->page + i;
602 down(&dev->db_tab->mutex);
605 if (i >= dev->db_tab->min_group2)
606 j = MTHCA_DB_REC_PER_PAGE - 1 - j;
607 clear_bit(j, page->used);
609 if (bitmap_empty(page->used, MTHCA_DB_REC_PER_PAGE) &&
610 i >= dev->db_tab->max_group1 - 1) {
611 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
613 dma_free_coherent(&dev->pdev->dev, 4096,
614 page->db_rec, page->mapping);
617 if (i == dev->db_tab->max_group1) {
618 --dev->db_tab->max_group1;
619 /* XXX may be able to unmap more pages now */
621 if (i == dev->db_tab->min_group2)
622 ++dev->db_tab->min_group2;
625 up(&dev->db_tab->mutex);
628 int mthca_init_db_tab(struct mthca_dev *dev)
632 if (!mthca_is_memfree(dev))
635 dev->db_tab = kmalloc(sizeof *dev->db_tab, GFP_KERNEL);
639 init_MUTEX(&dev->db_tab->mutex);
641 dev->db_tab->npages = dev->uar_table.uarc_size / 4096;
642 dev->db_tab->max_group1 = 0;
643 dev->db_tab->min_group2 = dev->db_tab->npages - 1;
645 dev->db_tab->page = kmalloc(dev->db_tab->npages *
646 sizeof *dev->db_tab->page,
648 if (!dev->db_tab->page) {
653 for (i = 0; i < dev->db_tab->npages; ++i)
654 dev->db_tab->page[i].db_rec = NULL;
659 void mthca_cleanup_db_tab(struct mthca_dev *dev)
664 if (!mthca_is_memfree(dev))
668 * Because we don't always free our UARC pages when they
669 * become empty to make mthca_free_db() simpler we need to
670 * make a sweep through the doorbell pages and free any
671 * leftover pages now.
673 for (i = 0; i < dev->db_tab->npages; ++i) {
674 if (!dev->db_tab->page[i].db_rec)
677 if (!bitmap_empty(dev->db_tab->page[i].used, MTHCA_DB_REC_PER_PAGE))
678 mthca_warn(dev, "Kernel UARC page %d not empty\n", i);
680 mthca_UNMAP_ICM(dev, mthca_uarc_virt(dev, &dev->driver_uar, i), 1, &status);
682 dma_free_coherent(&dev->pdev->dev, 4096,
683 dev->db_tab->page[i].db_rec,
684 dev->db_tab->page[i].mapping);
687 kfree(dev->db_tab->page);