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>
39 #include <linux/sched.h>
43 #include "mthca_memfree.h"
44 #include "mthca_dev.h"
45 #include "mthca_cmd.h"
48 * We allocate in as big chunks as we can, up to a maximum of 256 KB
52 MTHCA_ICM_ALLOC_SIZE
= 1 << 18,
53 MTHCA_TABLE_CHUNK_SIZE
= 1 << 18
56 struct mthca_user_db_table
{
60 struct scatterlist mem
;
65 static void mthca_free_icm_pages(struct mthca_dev
*dev
, struct mthca_icm_chunk
*chunk
)
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(sg_page(&chunk
->mem
[i
]),
75 get_order(chunk
->mem
[i
].length
));
78 static void mthca_free_icm_coherent(struct mthca_dev
*dev
, struct mthca_icm_chunk
*chunk
)
82 for (i
= 0; i
< chunk
->npages
; ++i
) {
83 dma_free_coherent(&dev
->pdev
->dev
, chunk
->mem
[i
].length
,
84 lowmem_page_address(sg_page(&chunk
->mem
[i
])),
85 sg_dma_address(&chunk
->mem
[i
]));
89 void mthca_free_icm(struct mthca_dev
*dev
, struct mthca_icm
*icm
, int coherent
)
91 struct mthca_icm_chunk
*chunk
, *tmp
;
96 list_for_each_entry_safe(chunk
, tmp
, &icm
->chunk_list
, list
) {
98 mthca_free_icm_coherent(dev
, chunk
);
100 mthca_free_icm_pages(dev
, chunk
);
108 static int mthca_alloc_icm_pages(struct scatterlist
*mem
, int order
, gfp_t gfp_mask
)
112 page
= alloc_pages(gfp_mask
, order
);
116 sg_set_page(mem
, page
, PAGE_SIZE
<< order
, 0);
120 static int mthca_alloc_icm_coherent(struct device
*dev
, struct scatterlist
*mem
,
121 int order
, gfp_t gfp_mask
)
123 void *buf
= dma_alloc_coherent(dev
, PAGE_SIZE
<< order
, &sg_dma_address(mem
),
128 sg_set_buf(mem
, buf
, PAGE_SIZE
<< order
);
130 sg_dma_len(mem
) = PAGE_SIZE
<< order
;
134 struct mthca_icm
*mthca_alloc_icm(struct mthca_dev
*dev
, int npages
,
135 gfp_t gfp_mask
, int coherent
)
137 struct mthca_icm
*icm
;
138 struct mthca_icm_chunk
*chunk
= NULL
;
142 /* We use sg_set_buf for coherent allocs, which assumes low memory */
143 BUG_ON(coherent
&& (gfp_mask
& __GFP_HIGHMEM
));
145 icm
= kmalloc(sizeof *icm
, gfp_mask
& ~(__GFP_HIGHMEM
| __GFP_NOWARN
));
150 INIT_LIST_HEAD(&icm
->chunk_list
);
152 cur_order
= get_order(MTHCA_ICM_ALLOC_SIZE
);
156 chunk
= kmalloc(sizeof *chunk
,
157 gfp_mask
& ~(__GFP_HIGHMEM
| __GFP_NOWARN
));
161 sg_init_table(chunk
->mem
, MTHCA_ICM_CHUNK_LEN
);
164 list_add_tail(&chunk
->list
, &icm
->chunk_list
);
167 while (1 << cur_order
> npages
)
171 ret
= mthca_alloc_icm_coherent(&dev
->pdev
->dev
,
172 &chunk
->mem
[chunk
->npages
],
173 cur_order
, gfp_mask
);
175 ret
= mthca_alloc_icm_pages(&chunk
->mem
[chunk
->npages
],
176 cur_order
, gfp_mask
);
183 else if (chunk
->npages
== MTHCA_ICM_CHUNK_LEN
) {
184 chunk
->nsg
= pci_map_sg(dev
->pdev
, chunk
->mem
,
186 PCI_DMA_BIDIRECTIONAL
);
192 if (chunk
->npages
== MTHCA_ICM_CHUNK_LEN
)
195 npages
-= 1 << cur_order
;
203 if (!coherent
&& chunk
) {
204 chunk
->nsg
= pci_map_sg(dev
->pdev
, chunk
->mem
,
206 PCI_DMA_BIDIRECTIONAL
);
215 mthca_free_icm(dev
, icm
, coherent
);
219 int mthca_table_get(struct mthca_dev
*dev
, struct mthca_icm_table
*table
, int obj
)
221 int i
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
/ MTHCA_TABLE_CHUNK_SIZE
;
225 mutex_lock(&table
->mutex
);
228 ++table
->icm
[i
]->refcount
;
232 table
->icm
[i
] = mthca_alloc_icm(dev
, MTHCA_TABLE_CHUNK_SIZE
>> PAGE_SHIFT
,
233 (table
->lowmem
? GFP_KERNEL
: GFP_HIGHUSER
) |
234 __GFP_NOWARN
, table
->coherent
);
235 if (!table
->icm
[i
]) {
240 if (mthca_MAP_ICM(dev
, table
->icm
[i
], table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
241 &status
) || status
) {
242 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
243 table
->icm
[i
] = NULL
;
248 ++table
->icm
[i
]->refcount
;
251 mutex_unlock(&table
->mutex
);
255 void mthca_table_put(struct mthca_dev
*dev
, struct mthca_icm_table
*table
, int obj
)
260 if (!mthca_is_memfree(dev
))
263 i
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
/ MTHCA_TABLE_CHUNK_SIZE
;
265 mutex_lock(&table
->mutex
);
267 if (--table
->icm
[i
]->refcount
== 0) {
268 mthca_UNMAP_ICM(dev
, table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
269 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
;
368 obj_per_chunk
= MTHCA_TABLE_CHUNK_SIZE
/ obj_size
;
369 num_icm
= DIV_ROUND_UP(nobj
, obj_per_chunk
);
371 table
= kmalloc(sizeof *table
+ num_icm
* sizeof *table
->icm
, GFP_KERNEL
);
376 table
->num_icm
= num_icm
;
377 table
->num_obj
= nobj
;
378 table
->obj_size
= obj_size
;
379 table
->lowmem
= use_lowmem
;
380 table
->coherent
= use_coherent
;
381 mutex_init(&table
->mutex
);
383 for (i
= 0; i
< num_icm
; ++i
)
384 table
->icm
[i
] = NULL
;
386 for (i
= 0; i
* MTHCA_TABLE_CHUNK_SIZE
< reserved
* obj_size
; ++i
) {
387 chunk_size
= MTHCA_TABLE_CHUNK_SIZE
;
388 if ((i
+ 1) * MTHCA_TABLE_CHUNK_SIZE
> nobj
* obj_size
)
389 chunk_size
= nobj
* obj_size
- i
* MTHCA_TABLE_CHUNK_SIZE
;
391 table
->icm
[i
] = mthca_alloc_icm(dev
, chunk_size
>> PAGE_SHIFT
,
392 (use_lowmem
? GFP_KERNEL
: GFP_HIGHUSER
) |
393 __GFP_NOWARN
, use_coherent
);
396 if (mthca_MAP_ICM(dev
, table
->icm
[i
], virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
397 &status
) || status
) {
398 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
399 table
->icm
[i
] = NULL
;
404 * Add a reference to this ICM chunk so that it never
405 * gets freed (since it contains reserved firmware objects).
407 ++table
->icm
[i
]->refcount
;
413 for (i
= 0; i
< num_icm
; ++i
)
415 mthca_UNMAP_ICM(dev
, virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
416 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
,
418 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
426 void mthca_free_icm_table(struct mthca_dev
*dev
, struct mthca_icm_table
*table
)
431 for (i
= 0; i
< table
->num_icm
; ++i
)
433 mthca_UNMAP_ICM(dev
, table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
434 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
,
436 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
442 static u64
mthca_uarc_virt(struct mthca_dev
*dev
, struct mthca_uar
*uar
, int page
)
444 return dev
->uar_table
.uarc_base
+
445 uar
->index
* dev
->uar_table
.uarc_size
+
446 page
* MTHCA_ICM_PAGE_SIZE
;
449 int mthca_map_user_db(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
450 struct mthca_user_db_table
*db_tab
, int index
, u64 uaddr
)
452 struct page
*pages
[1];
457 if (!mthca_is_memfree(dev
))
460 if (index
< 0 || index
> dev
->uar_table
.uarc_size
/ 8)
463 mutex_lock(&db_tab
->mutex
);
465 i
= index
/ MTHCA_DB_REC_PER_PAGE
;
467 if ((db_tab
->page
[i
].refcount
>= MTHCA_DB_REC_PER_PAGE
) ||
468 (db_tab
->page
[i
].uvirt
&& db_tab
->page
[i
].uvirt
!= uaddr
) ||
474 if (db_tab
->page
[i
].refcount
) {
475 ++db_tab
->page
[i
].refcount
;
479 ret
= get_user_pages(current
, current
->mm
, uaddr
& PAGE_MASK
, 1, 1, 0,
484 sg_set_page(&db_tab
->page
[i
].mem
, pages
[0], MTHCA_ICM_PAGE_SIZE
,
487 ret
= pci_map_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
493 ret
= mthca_MAP_ICM_page(dev
, sg_dma_address(&db_tab
->page
[i
].mem
),
494 mthca_uarc_virt(dev
, uar
, i
), &status
);
498 pci_unmap_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
499 put_page(sg_page(&db_tab
->page
[i
].mem
));
503 db_tab
->page
[i
].uvirt
= uaddr
;
504 db_tab
->page
[i
].refcount
= 1;
507 mutex_unlock(&db_tab
->mutex
);
511 void mthca_unmap_user_db(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
512 struct mthca_user_db_table
*db_tab
, int index
)
514 if (!mthca_is_memfree(dev
))
518 * To make our bookkeeping simpler, we don't unmap DB
519 * pages until we clean up the whole db table.
522 mutex_lock(&db_tab
->mutex
);
524 --db_tab
->page
[index
/ MTHCA_DB_REC_PER_PAGE
].refcount
;
526 mutex_unlock(&db_tab
->mutex
);
529 struct mthca_user_db_table
*mthca_init_user_db_tab(struct mthca_dev
*dev
)
531 struct mthca_user_db_table
*db_tab
;
535 if (!mthca_is_memfree(dev
))
538 npages
= dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
;
539 db_tab
= kmalloc(sizeof *db_tab
+ npages
* sizeof *db_tab
->page
, GFP_KERNEL
);
541 return ERR_PTR(-ENOMEM
);
543 mutex_init(&db_tab
->mutex
);
544 for (i
= 0; i
< npages
; ++i
) {
545 db_tab
->page
[i
].refcount
= 0;
546 db_tab
->page
[i
].uvirt
= 0;
547 sg_init_table(&db_tab
->page
[i
].mem
, 1);
553 void mthca_cleanup_user_db_tab(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
554 struct mthca_user_db_table
*db_tab
)
559 if (!mthca_is_memfree(dev
))
562 for (i
= 0; i
< dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
; ++i
) {
563 if (db_tab
->page
[i
].uvirt
) {
564 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, uar
, i
), 1, &status
);
565 pci_unmap_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
566 put_page(sg_page(&db_tab
->page
[i
].mem
));
573 int mthca_alloc_db(struct mthca_dev
*dev
, enum mthca_db_type type
,
579 struct mthca_db_page
*page
;
583 mutex_lock(&dev
->db_tab
->mutex
);
586 case MTHCA_DB_TYPE_CQ_ARM
:
587 case MTHCA_DB_TYPE_SQ
:
590 end
= dev
->db_tab
->max_group1
;
594 case MTHCA_DB_TYPE_CQ_SET_CI
:
595 case MTHCA_DB_TYPE_RQ
:
596 case MTHCA_DB_TYPE_SRQ
:
598 start
= dev
->db_tab
->npages
- 1;
599 end
= dev
->db_tab
->min_group2
;
608 for (i
= start
; i
!= end
; i
+= dir
)
609 if (dev
->db_tab
->page
[i
].db_rec
&&
610 !bitmap_full(dev
->db_tab
->page
[i
].used
,
611 MTHCA_DB_REC_PER_PAGE
)) {
612 page
= dev
->db_tab
->page
+ i
;
616 for (i
= start
; i
!= end
; i
+= dir
)
617 if (!dev
->db_tab
->page
[i
].db_rec
) {
618 page
= dev
->db_tab
->page
+ i
;
622 if (dev
->db_tab
->max_group1
>= dev
->db_tab
->min_group2
- 1) {
628 ++dev
->db_tab
->max_group1
;
630 --dev
->db_tab
->min_group2
;
632 page
= dev
->db_tab
->page
+ end
;
635 page
->db_rec
= dma_alloc_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
636 &page
->mapping
, GFP_KERNEL
);
641 memset(page
->db_rec
, 0, MTHCA_ICM_PAGE_SIZE
);
643 ret
= mthca_MAP_ICM_page(dev
, page
->mapping
,
644 mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), &status
);
648 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
649 page
->db_rec
, page
->mapping
);
653 bitmap_zero(page
->used
, MTHCA_DB_REC_PER_PAGE
);
656 j
= find_first_zero_bit(page
->used
, MTHCA_DB_REC_PER_PAGE
);
657 set_bit(j
, page
->used
);
660 j
= MTHCA_DB_REC_PER_PAGE
- 1 - j
;
662 ret
= i
* MTHCA_DB_REC_PER_PAGE
+ j
;
664 page
->db_rec
[j
] = cpu_to_be64((qn
<< 8) | (type
<< 5));
666 *db
= (__be32
*) &page
->db_rec
[j
];
669 mutex_unlock(&dev
->db_tab
->mutex
);
674 void mthca_free_db(struct mthca_dev
*dev
, int type
, int db_index
)
677 struct mthca_db_page
*page
;
680 i
= db_index
/ MTHCA_DB_REC_PER_PAGE
;
681 j
= db_index
% MTHCA_DB_REC_PER_PAGE
;
683 page
= dev
->db_tab
->page
+ i
;
685 mutex_lock(&dev
->db_tab
->mutex
);
688 if (i
>= dev
->db_tab
->min_group2
)
689 j
= MTHCA_DB_REC_PER_PAGE
- 1 - j
;
690 clear_bit(j
, page
->used
);
692 if (bitmap_empty(page
->used
, MTHCA_DB_REC_PER_PAGE
) &&
693 i
>= dev
->db_tab
->max_group1
- 1) {
694 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), 1, &status
);
696 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
697 page
->db_rec
, page
->mapping
);
700 if (i
== dev
->db_tab
->max_group1
) {
701 --dev
->db_tab
->max_group1
;
702 /* XXX may be able to unmap more pages now */
704 if (i
== dev
->db_tab
->min_group2
)
705 ++dev
->db_tab
->min_group2
;
708 mutex_unlock(&dev
->db_tab
->mutex
);
711 int mthca_init_db_tab(struct mthca_dev
*dev
)
715 if (!mthca_is_memfree(dev
))
718 dev
->db_tab
= kmalloc(sizeof *dev
->db_tab
, GFP_KERNEL
);
722 mutex_init(&dev
->db_tab
->mutex
);
724 dev
->db_tab
->npages
= dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
;
725 dev
->db_tab
->max_group1
= 0;
726 dev
->db_tab
->min_group2
= dev
->db_tab
->npages
- 1;
728 dev
->db_tab
->page
= kmalloc(dev
->db_tab
->npages
*
729 sizeof *dev
->db_tab
->page
,
731 if (!dev
->db_tab
->page
) {
736 for (i
= 0; i
< dev
->db_tab
->npages
; ++i
)
737 dev
->db_tab
->page
[i
].db_rec
= NULL
;
742 void mthca_cleanup_db_tab(struct mthca_dev
*dev
)
747 if (!mthca_is_memfree(dev
))
751 * Because we don't always free our UARC pages when they
752 * become empty to make mthca_free_db() simpler we need to
753 * make a sweep through the doorbell pages and free any
754 * leftover pages now.
756 for (i
= 0; i
< dev
->db_tab
->npages
; ++i
) {
757 if (!dev
->db_tab
->page
[i
].db_rec
)
760 if (!bitmap_empty(dev
->db_tab
->page
[i
].used
, MTHCA_DB_REC_PER_PAGE
))
761 mthca_warn(dev
, "Kernel UARC page %d not empty\n", i
);
763 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), 1, &status
);
765 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
766 dev
->db_tab
->page
[i
].db_rec
,
767 dev
->db_tab
->page
[i
].mapping
);
770 kfree(dev
->db_tab
->page
);