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
);
180 else if (chunk
->npages
== MTHCA_ICM_CHUNK_LEN
) {
181 chunk
->nsg
= pci_map_sg(dev
->pdev
, chunk
->mem
,
183 PCI_DMA_BIDIRECTIONAL
);
189 if (chunk
->npages
== MTHCA_ICM_CHUNK_LEN
)
192 npages
-= 1 << cur_order
;
200 if (!coherent
&& chunk
) {
201 chunk
->nsg
= pci_map_sg(dev
->pdev
, chunk
->mem
,
203 PCI_DMA_BIDIRECTIONAL
);
212 mthca_free_icm(dev
, icm
, coherent
);
216 int mthca_table_get(struct mthca_dev
*dev
, struct mthca_icm_table
*table
, int obj
)
218 int i
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
/ MTHCA_TABLE_CHUNK_SIZE
;
222 mutex_lock(&table
->mutex
);
225 ++table
->icm
[i
]->refcount
;
229 table
->icm
[i
] = mthca_alloc_icm(dev
, MTHCA_TABLE_CHUNK_SIZE
>> PAGE_SHIFT
,
230 (table
->lowmem
? GFP_KERNEL
: GFP_HIGHUSER
) |
231 __GFP_NOWARN
, table
->coherent
);
232 if (!table
->icm
[i
]) {
237 if (mthca_MAP_ICM(dev
, table
->icm
[i
], table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
238 &status
) || status
) {
239 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
240 table
->icm
[i
] = NULL
;
245 ++table
->icm
[i
]->refcount
;
248 mutex_unlock(&table
->mutex
);
252 void mthca_table_put(struct mthca_dev
*dev
, struct mthca_icm_table
*table
, int obj
)
257 if (!mthca_is_memfree(dev
))
260 i
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
/ MTHCA_TABLE_CHUNK_SIZE
;
262 mutex_lock(&table
->mutex
);
264 if (--table
->icm
[i
]->refcount
== 0) {
265 mthca_UNMAP_ICM(dev
, table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
266 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
,
268 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
269 table
->icm
[i
] = NULL
;
272 mutex_unlock(&table
->mutex
);
275 void *mthca_table_find(struct mthca_icm_table
*table
, int obj
, dma_addr_t
*dma_handle
)
277 int idx
, offset
, dma_offset
, i
;
278 struct mthca_icm_chunk
*chunk
;
279 struct mthca_icm
*icm
;
280 struct page
*page
= NULL
;
285 mutex_lock(&table
->mutex
);
287 idx
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
;
288 icm
= table
->icm
[idx
/ MTHCA_TABLE_CHUNK_SIZE
];
289 dma_offset
= offset
= idx
% MTHCA_TABLE_CHUNK_SIZE
;
294 list_for_each_entry(chunk
, &icm
->chunk_list
, list
) {
295 for (i
= 0; i
< chunk
->npages
; ++i
) {
296 if (dma_handle
&& dma_offset
>= 0) {
297 if (sg_dma_len(&chunk
->mem
[i
]) > dma_offset
)
298 *dma_handle
= sg_dma_address(&chunk
->mem
[i
]) +
300 dma_offset
-= sg_dma_len(&chunk
->mem
[i
]);
302 /* DMA mapping can merge pages but not split them,
303 * so if we found the page, dma_handle has already
304 * been assigned to. */
305 if (chunk
->mem
[i
].length
> offset
) {
306 page
= chunk
->mem
[i
].page
;
309 offset
-= chunk
->mem
[i
].length
;
314 mutex_unlock(&table
->mutex
);
315 return page
? lowmem_page_address(page
) + offset
: NULL
;
318 int mthca_table_get_range(struct mthca_dev
*dev
, struct mthca_icm_table
*table
,
321 int inc
= MTHCA_TABLE_CHUNK_SIZE
/ table
->obj_size
;
324 for (i
= start
; i
<= end
; i
+= inc
) {
325 err
= mthca_table_get(dev
, table
, i
);
335 mthca_table_put(dev
, table
, i
);
341 void mthca_table_put_range(struct mthca_dev
*dev
, struct mthca_icm_table
*table
,
346 if (!mthca_is_memfree(dev
))
349 for (i
= start
; i
<= end
; i
+= MTHCA_TABLE_CHUNK_SIZE
/ table
->obj_size
)
350 mthca_table_put(dev
, table
, i
);
353 struct mthca_icm_table
*mthca_alloc_icm_table(struct mthca_dev
*dev
,
354 u64 virt
, int obj_size
,
355 int nobj
, int reserved
,
356 int use_lowmem
, int use_coherent
)
358 struct mthca_icm_table
*table
;
364 num_icm
= (obj_size
* nobj
+ MTHCA_TABLE_CHUNK_SIZE
- 1) / MTHCA_TABLE_CHUNK_SIZE
;
366 table
= kmalloc(sizeof *table
+ num_icm
* sizeof *table
->icm
, GFP_KERNEL
);
371 table
->num_icm
= num_icm
;
372 table
->num_obj
= nobj
;
373 table
->obj_size
= obj_size
;
374 table
->lowmem
= use_lowmem
;
375 table
->coherent
= use_coherent
;
376 mutex_init(&table
->mutex
);
378 for (i
= 0; i
< num_icm
; ++i
)
379 table
->icm
[i
] = NULL
;
381 for (i
= 0; i
* MTHCA_TABLE_CHUNK_SIZE
< reserved
* obj_size
; ++i
) {
382 chunk_size
= MTHCA_TABLE_CHUNK_SIZE
;
383 if ((i
+ 1) * MTHCA_TABLE_CHUNK_SIZE
> nobj
* obj_size
)
384 chunk_size
= nobj
* obj_size
- i
* MTHCA_TABLE_CHUNK_SIZE
;
386 table
->icm
[i
] = mthca_alloc_icm(dev
, chunk_size
>> PAGE_SHIFT
,
387 (use_lowmem
? GFP_KERNEL
: GFP_HIGHUSER
) |
388 __GFP_NOWARN
, use_coherent
);
391 if (mthca_MAP_ICM(dev
, table
->icm
[i
], virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
392 &status
) || status
) {
393 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
394 table
->icm
[i
] = NULL
;
399 * Add a reference to this ICM chunk so that it never
400 * gets freed (since it contains reserved firmware objects).
402 ++table
->icm
[i
]->refcount
;
408 for (i
= 0; i
< num_icm
; ++i
)
410 mthca_UNMAP_ICM(dev
, virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
411 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
,
413 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
421 void mthca_free_icm_table(struct mthca_dev
*dev
, struct mthca_icm_table
*table
)
426 for (i
= 0; i
< table
->num_icm
; ++i
)
428 mthca_UNMAP_ICM(dev
, table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
429 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
,
431 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
437 static u64
mthca_uarc_virt(struct mthca_dev
*dev
, struct mthca_uar
*uar
, int page
)
439 return dev
->uar_table
.uarc_base
+
440 uar
->index
* dev
->uar_table
.uarc_size
+
441 page
* MTHCA_ICM_PAGE_SIZE
;
444 int mthca_map_user_db(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
445 struct mthca_user_db_table
*db_tab
, int index
, u64 uaddr
)
451 if (!mthca_is_memfree(dev
))
454 if (index
< 0 || index
> dev
->uar_table
.uarc_size
/ 8)
457 mutex_lock(&db_tab
->mutex
);
459 i
= index
/ MTHCA_DB_REC_PER_PAGE
;
461 if ((db_tab
->page
[i
].refcount
>= MTHCA_DB_REC_PER_PAGE
) ||
462 (db_tab
->page
[i
].uvirt
&& db_tab
->page
[i
].uvirt
!= uaddr
) ||
468 if (db_tab
->page
[i
].refcount
) {
469 ++db_tab
->page
[i
].refcount
;
473 ret
= get_user_pages(current
, current
->mm
, uaddr
& PAGE_MASK
, 1, 1, 0,
474 &db_tab
->page
[i
].mem
.page
, NULL
);
478 db_tab
->page
[i
].mem
.length
= MTHCA_ICM_PAGE_SIZE
;
479 db_tab
->page
[i
].mem
.offset
= uaddr
& ~PAGE_MASK
;
481 ret
= pci_map_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
483 put_page(db_tab
->page
[i
].mem
.page
);
487 ret
= mthca_MAP_ICM_page(dev
, sg_dma_address(&db_tab
->page
[i
].mem
),
488 mthca_uarc_virt(dev
, uar
, i
), &status
);
492 pci_unmap_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
493 put_page(db_tab
->page
[i
].mem
.page
);
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(sizeof *db_tab
+ npages
* sizeof *db_tab
->page
, 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;
546 void mthca_cleanup_user_db_tab(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
547 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, &status
);
558 pci_unmap_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
559 put_page(db_tab
->page
[i
].mem
.page
);
566 int mthca_alloc_db(struct mthca_dev
*dev
, enum mthca_db_type type
,
572 struct mthca_db_page
*page
;
576 mutex_lock(&dev
->db_tab
->mutex
);
579 case MTHCA_DB_TYPE_CQ_ARM
:
580 case MTHCA_DB_TYPE_SQ
:
583 end
= dev
->db_tab
->max_group1
;
587 case MTHCA_DB_TYPE_CQ_SET_CI
:
588 case MTHCA_DB_TYPE_RQ
:
589 case MTHCA_DB_TYPE_SRQ
:
591 start
= dev
->db_tab
->npages
- 1;
592 end
= dev
->db_tab
->min_group2
;
601 for (i
= start
; i
!= end
; i
+= dir
)
602 if (dev
->db_tab
->page
[i
].db_rec
&&
603 !bitmap_full(dev
->db_tab
->page
[i
].used
,
604 MTHCA_DB_REC_PER_PAGE
)) {
605 page
= dev
->db_tab
->page
+ i
;
609 for (i
= start
; i
!= end
; i
+= dir
)
610 if (!dev
->db_tab
->page
[i
].db_rec
) {
611 page
= dev
->db_tab
->page
+ i
;
615 if (dev
->db_tab
->max_group1
>= dev
->db_tab
->min_group2
- 1) {
621 ++dev
->db_tab
->max_group1
;
623 --dev
->db_tab
->min_group2
;
625 page
= dev
->db_tab
->page
+ end
;
628 page
->db_rec
= dma_alloc_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
629 &page
->mapping
, GFP_KERNEL
);
634 memset(page
->db_rec
, 0, MTHCA_ICM_PAGE_SIZE
);
636 ret
= mthca_MAP_ICM_page(dev
, page
->mapping
,
637 mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), &status
);
641 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
642 page
->db_rec
, page
->mapping
);
646 bitmap_zero(page
->used
, MTHCA_DB_REC_PER_PAGE
);
649 j
= find_first_zero_bit(page
->used
, MTHCA_DB_REC_PER_PAGE
);
650 set_bit(j
, page
->used
);
653 j
= MTHCA_DB_REC_PER_PAGE
- 1 - j
;
655 ret
= i
* MTHCA_DB_REC_PER_PAGE
+ j
;
657 page
->db_rec
[j
] = cpu_to_be64((qn
<< 8) | (type
<< 5));
659 *db
= (__be32
*) &page
->db_rec
[j
];
662 mutex_unlock(&dev
->db_tab
->mutex
);
667 void mthca_free_db(struct mthca_dev
*dev
, int type
, int db_index
)
670 struct mthca_db_page
*page
;
673 i
= db_index
/ MTHCA_DB_REC_PER_PAGE
;
674 j
= db_index
% MTHCA_DB_REC_PER_PAGE
;
676 page
= dev
->db_tab
->page
+ i
;
678 mutex_lock(&dev
->db_tab
->mutex
);
681 if (i
>= dev
->db_tab
->min_group2
)
682 j
= MTHCA_DB_REC_PER_PAGE
- 1 - j
;
683 clear_bit(j
, page
->used
);
685 if (bitmap_empty(page
->used
, MTHCA_DB_REC_PER_PAGE
) &&
686 i
>= dev
->db_tab
->max_group1
- 1) {
687 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), 1, &status
);
689 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
690 page
->db_rec
, page
->mapping
);
693 if (i
== dev
->db_tab
->max_group1
) {
694 --dev
->db_tab
->max_group1
;
695 /* XXX may be able to unmap more pages now */
697 if (i
== dev
->db_tab
->min_group2
)
698 ++dev
->db_tab
->min_group2
;
701 mutex_unlock(&dev
->db_tab
->mutex
);
704 int mthca_init_db_tab(struct mthca_dev
*dev
)
708 if (!mthca_is_memfree(dev
))
711 dev
->db_tab
= kmalloc(sizeof *dev
->db_tab
, GFP_KERNEL
);
715 mutex_init(&dev
->db_tab
->mutex
);
717 dev
->db_tab
->npages
= dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
;
718 dev
->db_tab
->max_group1
= 0;
719 dev
->db_tab
->min_group2
= dev
->db_tab
->npages
- 1;
721 dev
->db_tab
->page
= kmalloc(dev
->db_tab
->npages
*
722 sizeof *dev
->db_tab
->page
,
724 if (!dev
->db_tab
->page
) {
729 for (i
= 0; i
< dev
->db_tab
->npages
; ++i
)
730 dev
->db_tab
->page
[i
].db_rec
= NULL
;
735 void mthca_cleanup_db_tab(struct mthca_dev
*dev
)
740 if (!mthca_is_memfree(dev
))
744 * Because we don't always free our UARC pages when they
745 * become empty to make mthca_free_db() simpler we need to
746 * make a sweep through the doorbell pages and free any
747 * leftover pages now.
749 for (i
= 0; i
< dev
->db_tab
->npages
; ++i
) {
750 if (!dev
->db_tab
->page
[i
].db_rec
)
753 if (!bitmap_empty(dev
->db_tab
->page
[i
].used
, MTHCA_DB_REC_PER_PAGE
))
754 mthca_warn(dev
, "Kernel UARC page %d not empty\n", i
);
756 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), 1, &status
);
758 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
759 dev
->db_tab
->page
[i
].db_rec
,
760 dev
->db_tab
->page
[i
].mapping
);
763 kfree(dev
->db_tab
->page
);