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
;
227 mutex_lock(&table
->mutex
);
230 ++table
->icm
[i
]->refcount
;
234 table
->icm
[i
] = mthca_alloc_icm(dev
, MTHCA_TABLE_CHUNK_SIZE
>> PAGE_SHIFT
,
235 (table
->lowmem
? GFP_KERNEL
: GFP_HIGHUSER
) |
236 __GFP_NOWARN
, table
->coherent
);
237 if (!table
->icm
[i
]) {
242 if (mthca_MAP_ICM(dev
, table
->icm
[i
],
243 table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
)) {
244 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
245 table
->icm
[i
] = NULL
;
250 ++table
->icm
[i
]->refcount
;
253 mutex_unlock(&table
->mutex
);
257 void mthca_table_put(struct mthca_dev
*dev
, struct mthca_icm_table
*table
, int obj
)
261 if (!mthca_is_memfree(dev
))
264 i
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
/ MTHCA_TABLE_CHUNK_SIZE
;
266 mutex_lock(&table
->mutex
);
268 if (--table
->icm
[i
]->refcount
== 0) {
269 mthca_UNMAP_ICM(dev
, table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
270 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
);
271 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
272 table
->icm
[i
] = NULL
;
275 mutex_unlock(&table
->mutex
);
278 void *mthca_table_find(struct mthca_icm_table
*table
, int obj
, dma_addr_t
*dma_handle
)
280 int idx
, offset
, dma_offset
, i
;
281 struct mthca_icm_chunk
*chunk
;
282 struct mthca_icm
*icm
;
283 struct page
*page
= NULL
;
288 mutex_lock(&table
->mutex
);
290 idx
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
;
291 icm
= table
->icm
[idx
/ MTHCA_TABLE_CHUNK_SIZE
];
292 dma_offset
= offset
= idx
% MTHCA_TABLE_CHUNK_SIZE
;
297 list_for_each_entry(chunk
, &icm
->chunk_list
, list
) {
298 for (i
= 0; i
< chunk
->npages
; ++i
) {
299 if (dma_handle
&& dma_offset
>= 0) {
300 if (sg_dma_len(&chunk
->mem
[i
]) > dma_offset
)
301 *dma_handle
= sg_dma_address(&chunk
->mem
[i
]) +
303 dma_offset
-= sg_dma_len(&chunk
->mem
[i
]);
305 /* DMA mapping can merge pages but not split them,
306 * so if we found the page, dma_handle has already
307 * been assigned to. */
308 if (chunk
->mem
[i
].length
> offset
) {
309 page
= sg_page(&chunk
->mem
[i
]);
312 offset
-= chunk
->mem
[i
].length
;
317 mutex_unlock(&table
->mutex
);
318 return page
? lowmem_page_address(page
) + offset
: NULL
;
321 int mthca_table_get_range(struct mthca_dev
*dev
, struct mthca_icm_table
*table
,
324 int inc
= MTHCA_TABLE_CHUNK_SIZE
/ table
->obj_size
;
327 for (i
= start
; i
<= end
; i
+= inc
) {
328 err
= mthca_table_get(dev
, table
, i
);
338 mthca_table_put(dev
, table
, i
);
344 void mthca_table_put_range(struct mthca_dev
*dev
, struct mthca_icm_table
*table
,
349 if (!mthca_is_memfree(dev
))
352 for (i
= start
; i
<= end
; i
+= MTHCA_TABLE_CHUNK_SIZE
/ table
->obj_size
)
353 mthca_table_put(dev
, table
, i
);
356 struct mthca_icm_table
*mthca_alloc_icm_table(struct mthca_dev
*dev
,
357 u64 virt
, int obj_size
,
358 int nobj
, int reserved
,
359 int use_lowmem
, int use_coherent
)
361 struct mthca_icm_table
*table
;
367 obj_per_chunk
= MTHCA_TABLE_CHUNK_SIZE
/ obj_size
;
368 num_icm
= DIV_ROUND_UP(nobj
, obj_per_chunk
);
370 table
= kmalloc(struct_size(table
, icm
, num_icm
), GFP_KERNEL
);
375 table
->num_icm
= num_icm
;
376 table
->num_obj
= nobj
;
377 table
->obj_size
= obj_size
;
378 table
->lowmem
= use_lowmem
;
379 table
->coherent
= use_coherent
;
380 mutex_init(&table
->mutex
);
382 for (i
= 0; i
< num_icm
; ++i
)
383 table
->icm
[i
] = NULL
;
385 for (i
= 0; i
* MTHCA_TABLE_CHUNK_SIZE
< reserved
* obj_size
; ++i
) {
386 chunk_size
= MTHCA_TABLE_CHUNK_SIZE
;
387 if ((i
+ 1) * MTHCA_TABLE_CHUNK_SIZE
> nobj
* obj_size
)
388 chunk_size
= nobj
* obj_size
- i
* MTHCA_TABLE_CHUNK_SIZE
;
390 table
->icm
[i
] = mthca_alloc_icm(dev
, chunk_size
>> PAGE_SHIFT
,
391 (use_lowmem
? GFP_KERNEL
: GFP_HIGHUSER
) |
392 __GFP_NOWARN
, use_coherent
);
395 if (mthca_MAP_ICM(dev
, table
->icm
[i
],
396 virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
)) {
397 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
398 table
->icm
[i
] = NULL
;
403 * Add a reference to this ICM chunk so that it never
404 * gets freed (since it contains reserved firmware objects).
406 ++table
->icm
[i
]->refcount
;
412 for (i
= 0; i
< num_icm
; ++i
)
414 mthca_UNMAP_ICM(dev
, virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
415 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
);
416 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
424 void mthca_free_icm_table(struct mthca_dev
*dev
, struct mthca_icm_table
*table
)
428 for (i
= 0; i
< table
->num_icm
; ++i
)
431 table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
432 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
);
433 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
439 static u64
mthca_uarc_virt(struct mthca_dev
*dev
, struct mthca_uar
*uar
, int page
)
441 return dev
->uar_table
.uarc_base
+
442 uar
->index
* dev
->uar_table
.uarc_size
+
443 page
* MTHCA_ICM_PAGE_SIZE
;
446 int mthca_map_user_db(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
447 struct mthca_user_db_table
*db_tab
, int index
, u64 uaddr
)
449 struct page
*pages
[1];
453 if (!mthca_is_memfree(dev
))
456 if (index
< 0 || index
> dev
->uar_table
.uarc_size
/ 8)
459 mutex_lock(&db_tab
->mutex
);
461 i
= index
/ MTHCA_DB_REC_PER_PAGE
;
463 if ((db_tab
->page
[i
].refcount
>= MTHCA_DB_REC_PER_PAGE
) ||
464 (db_tab
->page
[i
].uvirt
&& db_tab
->page
[i
].uvirt
!= uaddr
) ||
470 if (db_tab
->page
[i
].refcount
) {
471 ++db_tab
->page
[i
].refcount
;
475 ret
= pin_user_pages_fast(uaddr
& PAGE_MASK
, 1,
476 FOLL_WRITE
| FOLL_LONGTERM
, pages
);
480 sg_set_page(&db_tab
->page
[i
].mem
, pages
[0], MTHCA_ICM_PAGE_SIZE
,
483 ret
= pci_map_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
485 unpin_user_page(pages
[0]);
489 ret
= mthca_MAP_ICM_page(dev
, sg_dma_address(&db_tab
->page
[i
].mem
),
490 mthca_uarc_virt(dev
, uar
, i
));
492 pci_unmap_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
493 unpin_user_page(sg_page(&db_tab
->page
[i
].mem
));
497 db_tab
->page
[i
].uvirt
= uaddr
;
498 db_tab
->page
[i
].refcount
= 1;
501 mutex_unlock(&db_tab
->mutex
);
505 void mthca_unmap_user_db(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
506 struct mthca_user_db_table
*db_tab
, int index
)
508 if (!mthca_is_memfree(dev
))
512 * To make our bookkeeping simpler, we don't unmap DB
513 * pages until we clean up the whole db table.
516 mutex_lock(&db_tab
->mutex
);
518 --db_tab
->page
[index
/ MTHCA_DB_REC_PER_PAGE
].refcount
;
520 mutex_unlock(&db_tab
->mutex
);
523 struct mthca_user_db_table
*mthca_init_user_db_tab(struct mthca_dev
*dev
)
525 struct mthca_user_db_table
*db_tab
;
529 if (!mthca_is_memfree(dev
))
532 npages
= dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
;
533 db_tab
= kmalloc(struct_size(db_tab
, page
, npages
), GFP_KERNEL
);
535 return ERR_PTR(-ENOMEM
);
537 mutex_init(&db_tab
->mutex
);
538 for (i
= 0; i
< npages
; ++i
) {
539 db_tab
->page
[i
].refcount
= 0;
540 db_tab
->page
[i
].uvirt
= 0;
541 sg_init_table(&db_tab
->page
[i
].mem
, 1);
547 void mthca_cleanup_user_db_tab(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
548 struct mthca_user_db_table
*db_tab
)
552 if (!mthca_is_memfree(dev
))
555 for (i
= 0; i
< dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
; ++i
) {
556 if (db_tab
->page
[i
].uvirt
) {
557 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, uar
, i
), 1);
558 pci_unmap_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
559 unpin_user_page(sg_page(&db_tab
->page
[i
].mem
));
566 int mthca_alloc_db(struct mthca_dev
*dev
, enum mthca_db_type type
,
572 struct mthca_db_page
*page
;
575 mutex_lock(&dev
->db_tab
->mutex
);
578 case MTHCA_DB_TYPE_CQ_ARM
:
579 case MTHCA_DB_TYPE_SQ
:
582 end
= dev
->db_tab
->max_group1
;
586 case MTHCA_DB_TYPE_CQ_SET_CI
:
587 case MTHCA_DB_TYPE_RQ
:
588 case MTHCA_DB_TYPE_SRQ
:
590 start
= dev
->db_tab
->npages
- 1;
591 end
= dev
->db_tab
->min_group2
;
600 for (i
= start
; i
!= end
; i
+= dir
)
601 if (dev
->db_tab
->page
[i
].db_rec
&&
602 !bitmap_full(dev
->db_tab
->page
[i
].used
,
603 MTHCA_DB_REC_PER_PAGE
)) {
604 page
= dev
->db_tab
->page
+ i
;
608 for (i
= start
; i
!= end
; i
+= dir
)
609 if (!dev
->db_tab
->page
[i
].db_rec
) {
610 page
= dev
->db_tab
->page
+ i
;
614 if (dev
->db_tab
->max_group1
>= dev
->db_tab
->min_group2
- 1) {
620 ++dev
->db_tab
->max_group1
;
622 --dev
->db_tab
->min_group2
;
624 page
= dev
->db_tab
->page
+ end
;
627 page
->db_rec
= dma_alloc_coherent(&dev
->pdev
->dev
,
628 MTHCA_ICM_PAGE_SIZE
, &page
->mapping
,
635 ret
= mthca_MAP_ICM_page(dev
, page
->mapping
,
636 mthca_uarc_virt(dev
, &dev
->driver_uar
, i
));
638 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
639 page
->db_rec
, page
->mapping
);
643 bitmap_zero(page
->used
, MTHCA_DB_REC_PER_PAGE
);
646 j
= find_first_zero_bit(page
->used
, MTHCA_DB_REC_PER_PAGE
);
647 set_bit(j
, page
->used
);
650 j
= MTHCA_DB_REC_PER_PAGE
- 1 - j
;
652 ret
= i
* MTHCA_DB_REC_PER_PAGE
+ j
;
654 page
->db_rec
[j
] = cpu_to_be64((qn
<< 8) | (type
<< 5));
656 *db
= (__be32
*) &page
->db_rec
[j
];
659 mutex_unlock(&dev
->db_tab
->mutex
);
664 void mthca_free_db(struct mthca_dev
*dev
, int type
, int db_index
)
667 struct mthca_db_page
*page
;
669 i
= db_index
/ MTHCA_DB_REC_PER_PAGE
;
670 j
= db_index
% MTHCA_DB_REC_PER_PAGE
;
672 page
= dev
->db_tab
->page
+ i
;
674 mutex_lock(&dev
->db_tab
->mutex
);
677 if (i
>= dev
->db_tab
->min_group2
)
678 j
= MTHCA_DB_REC_PER_PAGE
- 1 - j
;
679 clear_bit(j
, page
->used
);
681 if (bitmap_empty(page
->used
, MTHCA_DB_REC_PER_PAGE
) &&
682 i
>= dev
->db_tab
->max_group1
- 1) {
683 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), 1);
685 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
686 page
->db_rec
, page
->mapping
);
689 if (i
== dev
->db_tab
->max_group1
) {
690 --dev
->db_tab
->max_group1
;
691 /* XXX may be able to unmap more pages now */
693 if (i
== dev
->db_tab
->min_group2
)
694 ++dev
->db_tab
->min_group2
;
697 mutex_unlock(&dev
->db_tab
->mutex
);
700 int mthca_init_db_tab(struct mthca_dev
*dev
)
704 if (!mthca_is_memfree(dev
))
707 dev
->db_tab
= kmalloc(sizeof *dev
->db_tab
, GFP_KERNEL
);
711 mutex_init(&dev
->db_tab
->mutex
);
713 dev
->db_tab
->npages
= dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
;
714 dev
->db_tab
->max_group1
= 0;
715 dev
->db_tab
->min_group2
= dev
->db_tab
->npages
- 1;
717 dev
->db_tab
->page
= kmalloc_array(dev
->db_tab
->npages
,
718 sizeof(*dev
->db_tab
->page
),
720 if (!dev
->db_tab
->page
) {
725 for (i
= 0; i
< dev
->db_tab
->npages
; ++i
)
726 dev
->db_tab
->page
[i
].db_rec
= NULL
;
731 void mthca_cleanup_db_tab(struct mthca_dev
*dev
)
735 if (!mthca_is_memfree(dev
))
739 * Because we don't always free our UARC pages when they
740 * become empty to make mthca_free_db() simpler we need to
741 * make a sweep through the doorbell pages and free any
742 * leftover pages now.
744 for (i
= 0; i
< dev
->db_tab
->npages
; ++i
) {
745 if (!dev
->db_tab
->page
[i
].db_rec
)
748 if (!bitmap_empty(dev
->db_tab
->page
[i
].used
, MTHCA_DB_REC_PER_PAGE
))
749 mthca_warn(dev
, "Kernel UARC page %d not empty\n", i
);
751 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), 1);
753 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
754 dev
->db_tab
->page
[i
].db_rec
,
755 dev
->db_tab
->page
[i
].mapping
);
758 kfree(dev
->db_tab
->page
);