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>
41 #include "mthca_memfree.h"
42 #include "mthca_dev.h"
43 #include "mthca_cmd.h"
46 * We allocate in as big chunks as we can, up to a maximum of 256 KB
50 MTHCA_ICM_ALLOC_SIZE
= 1 << 18,
51 MTHCA_TABLE_CHUNK_SIZE
= 1 << 18
54 struct mthca_user_db_table
{
58 struct scatterlist mem
;
63 static void mthca_free_icm_pages(struct mthca_dev
*dev
, struct mthca_icm_chunk
*chunk
)
68 pci_unmap_sg(dev
->pdev
, chunk
->mem
, chunk
->npages
,
69 PCI_DMA_BIDIRECTIONAL
);
71 for (i
= 0; i
< chunk
->npages
; ++i
)
72 __free_pages(sg_page(&chunk
->mem
[i
]),
73 get_order(chunk
->mem
[i
].length
));
76 static void mthca_free_icm_coherent(struct mthca_dev
*dev
, struct mthca_icm_chunk
*chunk
)
80 for (i
= 0; i
< chunk
->npages
; ++i
) {
81 dma_free_coherent(&dev
->pdev
->dev
, chunk
->mem
[i
].length
,
82 lowmem_page_address(sg_page(&chunk
->mem
[i
])),
83 sg_dma_address(&chunk
->mem
[i
]));
87 void mthca_free_icm(struct mthca_dev
*dev
, struct mthca_icm
*icm
, int coherent
)
89 struct mthca_icm_chunk
*chunk
, *tmp
;
94 list_for_each_entry_safe(chunk
, tmp
, &icm
->chunk_list
, list
) {
96 mthca_free_icm_coherent(dev
, chunk
);
98 mthca_free_icm_pages(dev
, chunk
);
106 static int mthca_alloc_icm_pages(struct scatterlist
*mem
, int order
, gfp_t gfp_mask
)
111 * Use __GFP_ZERO because buggy firmware assumes ICM pages are
112 * cleared, and subtle failures are seen if they aren't.
114 page
= alloc_pages(gfp_mask
| __GFP_ZERO
, order
);
118 sg_set_page(mem
, page
, PAGE_SIZE
<< order
, 0);
122 static int mthca_alloc_icm_coherent(struct device
*dev
, struct scatterlist
*mem
,
123 int order
, gfp_t gfp_mask
)
125 void *buf
= dma_alloc_coherent(dev
, PAGE_SIZE
<< order
, &sg_dma_address(mem
),
130 sg_set_buf(mem
, buf
, PAGE_SIZE
<< order
);
132 sg_dma_len(mem
) = PAGE_SIZE
<< order
;
136 struct mthca_icm
*mthca_alloc_icm(struct mthca_dev
*dev
, int npages
,
137 gfp_t gfp_mask
, int coherent
)
139 struct mthca_icm
*icm
;
140 struct mthca_icm_chunk
*chunk
= NULL
;
144 /* We use sg_set_buf for coherent allocs, which assumes low memory */
145 BUG_ON(coherent
&& (gfp_mask
& __GFP_HIGHMEM
));
147 icm
= kmalloc(sizeof *icm
, gfp_mask
& ~(__GFP_HIGHMEM
| __GFP_NOWARN
));
152 INIT_LIST_HEAD(&icm
->chunk_list
);
154 cur_order
= get_order(MTHCA_ICM_ALLOC_SIZE
);
158 chunk
= kmalloc(sizeof *chunk
,
159 gfp_mask
& ~(__GFP_HIGHMEM
| __GFP_NOWARN
));
163 sg_init_table(chunk
->mem
, MTHCA_ICM_CHUNK_LEN
);
166 list_add_tail(&chunk
->list
, &icm
->chunk_list
);
169 while (1 << cur_order
> npages
)
173 ret
= mthca_alloc_icm_coherent(&dev
->pdev
->dev
,
174 &chunk
->mem
[chunk
->npages
],
175 cur_order
, gfp_mask
);
177 ret
= mthca_alloc_icm_pages(&chunk
->mem
[chunk
->npages
],
178 cur_order
, gfp_mask
);
185 else if (chunk
->npages
== MTHCA_ICM_CHUNK_LEN
) {
186 chunk
->nsg
= pci_map_sg(dev
->pdev
, chunk
->mem
,
188 PCI_DMA_BIDIRECTIONAL
);
194 if (chunk
->npages
== MTHCA_ICM_CHUNK_LEN
)
197 npages
-= 1 << cur_order
;
205 if (!coherent
&& chunk
) {
206 chunk
->nsg
= pci_map_sg(dev
->pdev
, chunk
->mem
,
208 PCI_DMA_BIDIRECTIONAL
);
217 mthca_free_icm(dev
, icm
, coherent
);
221 int mthca_table_get(struct mthca_dev
*dev
, struct mthca_icm_table
*table
, int obj
)
223 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
], table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
243 &status
) || status
) {
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
)
262 if (!mthca_is_memfree(dev
))
265 i
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
/ MTHCA_TABLE_CHUNK_SIZE
;
267 mutex_lock(&table
->mutex
);
269 if (--table
->icm
[i
]->refcount
== 0) {
270 mthca_UNMAP_ICM(dev
, table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
271 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
,
273 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
274 table
->icm
[i
] = NULL
;
277 mutex_unlock(&table
->mutex
);
280 void *mthca_table_find(struct mthca_icm_table
*table
, int obj
, dma_addr_t
*dma_handle
)
282 int idx
, offset
, dma_offset
, i
;
283 struct mthca_icm_chunk
*chunk
;
284 struct mthca_icm
*icm
;
285 struct page
*page
= NULL
;
290 mutex_lock(&table
->mutex
);
292 idx
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
;
293 icm
= table
->icm
[idx
/ MTHCA_TABLE_CHUNK_SIZE
];
294 dma_offset
= offset
= idx
% MTHCA_TABLE_CHUNK_SIZE
;
299 list_for_each_entry(chunk
, &icm
->chunk_list
, list
) {
300 for (i
= 0; i
< chunk
->npages
; ++i
) {
301 if (dma_handle
&& dma_offset
>= 0) {
302 if (sg_dma_len(&chunk
->mem
[i
]) > dma_offset
)
303 *dma_handle
= sg_dma_address(&chunk
->mem
[i
]) +
305 dma_offset
-= sg_dma_len(&chunk
->mem
[i
]);
307 /* DMA mapping can merge pages but not split them,
308 * so if we found the page, dma_handle has already
309 * been assigned to. */
310 if (chunk
->mem
[i
].length
> offset
) {
311 page
= sg_page(&chunk
->mem
[i
]);
314 offset
-= chunk
->mem
[i
].length
;
319 mutex_unlock(&table
->mutex
);
320 return page
? lowmem_page_address(page
) + offset
: NULL
;
323 int mthca_table_get_range(struct mthca_dev
*dev
, struct mthca_icm_table
*table
,
326 int inc
= MTHCA_TABLE_CHUNK_SIZE
/ table
->obj_size
;
329 for (i
= start
; i
<= end
; i
+= inc
) {
330 err
= mthca_table_get(dev
, table
, i
);
340 mthca_table_put(dev
, table
, i
);
346 void mthca_table_put_range(struct mthca_dev
*dev
, struct mthca_icm_table
*table
,
351 if (!mthca_is_memfree(dev
))
354 for (i
= start
; i
<= end
; i
+= MTHCA_TABLE_CHUNK_SIZE
/ table
->obj_size
)
355 mthca_table_put(dev
, table
, i
);
358 struct mthca_icm_table
*mthca_alloc_icm_table(struct mthca_dev
*dev
,
359 u64 virt
, int obj_size
,
360 int nobj
, int reserved
,
361 int use_lowmem
, int use_coherent
)
363 struct mthca_icm_table
*table
;
370 obj_per_chunk
= MTHCA_TABLE_CHUNK_SIZE
/ obj_size
;
371 num_icm
= DIV_ROUND_UP(nobj
, obj_per_chunk
);
373 table
= kmalloc(sizeof *table
+ num_icm
* sizeof *table
->icm
, GFP_KERNEL
);
378 table
->num_icm
= num_icm
;
379 table
->num_obj
= nobj
;
380 table
->obj_size
= obj_size
;
381 table
->lowmem
= use_lowmem
;
382 table
->coherent
= use_coherent
;
383 mutex_init(&table
->mutex
);
385 for (i
= 0; i
< num_icm
; ++i
)
386 table
->icm
[i
] = NULL
;
388 for (i
= 0; i
* MTHCA_TABLE_CHUNK_SIZE
< reserved
* obj_size
; ++i
) {
389 chunk_size
= MTHCA_TABLE_CHUNK_SIZE
;
390 if ((i
+ 1) * MTHCA_TABLE_CHUNK_SIZE
> nobj
* obj_size
)
391 chunk_size
= nobj
* obj_size
- i
* MTHCA_TABLE_CHUNK_SIZE
;
393 table
->icm
[i
] = mthca_alloc_icm(dev
, chunk_size
>> PAGE_SHIFT
,
394 (use_lowmem
? GFP_KERNEL
: GFP_HIGHUSER
) |
395 __GFP_NOWARN
, use_coherent
);
398 if (mthca_MAP_ICM(dev
, table
->icm
[i
], virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
399 &status
) || status
) {
400 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
401 table
->icm
[i
] = NULL
;
406 * Add a reference to this ICM chunk so that it never
407 * gets freed (since it contains reserved firmware objects).
409 ++table
->icm
[i
]->refcount
;
415 for (i
= 0; i
< num_icm
; ++i
)
417 mthca_UNMAP_ICM(dev
, virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
418 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
,
420 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
428 void mthca_free_icm_table(struct mthca_dev
*dev
, struct mthca_icm_table
*table
)
433 for (i
= 0; i
< table
->num_icm
; ++i
)
435 mthca_UNMAP_ICM(dev
, table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
436 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
,
438 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
444 static u64
mthca_uarc_virt(struct mthca_dev
*dev
, struct mthca_uar
*uar
, int page
)
446 return dev
->uar_table
.uarc_base
+
447 uar
->index
* dev
->uar_table
.uarc_size
+
448 page
* MTHCA_ICM_PAGE_SIZE
;
451 int mthca_map_user_db(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
452 struct mthca_user_db_table
*db_tab
, int index
, u64 uaddr
)
454 struct page
*pages
[1];
459 if (!mthca_is_memfree(dev
))
462 if (index
< 0 || index
> dev
->uar_table
.uarc_size
/ 8)
465 mutex_lock(&db_tab
->mutex
);
467 i
= index
/ MTHCA_DB_REC_PER_PAGE
;
469 if ((db_tab
->page
[i
].refcount
>= MTHCA_DB_REC_PER_PAGE
) ||
470 (db_tab
->page
[i
].uvirt
&& db_tab
->page
[i
].uvirt
!= uaddr
) ||
476 if (db_tab
->page
[i
].refcount
) {
477 ++db_tab
->page
[i
].refcount
;
481 ret
= get_user_pages(current
, current
->mm
, uaddr
& PAGE_MASK
, 1, 1, 0,
486 sg_set_page(&db_tab
->page
[i
].mem
, pages
[0], MTHCA_ICM_PAGE_SIZE
,
489 ret
= pci_map_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
495 ret
= mthca_MAP_ICM_page(dev
, sg_dma_address(&db_tab
->page
[i
].mem
),
496 mthca_uarc_virt(dev
, uar
, i
), &status
);
500 pci_unmap_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
501 put_page(sg_page(&db_tab
->page
[i
].mem
));
505 db_tab
->page
[i
].uvirt
= uaddr
;
506 db_tab
->page
[i
].refcount
= 1;
509 mutex_unlock(&db_tab
->mutex
);
513 void mthca_unmap_user_db(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
514 struct mthca_user_db_table
*db_tab
, int index
)
516 if (!mthca_is_memfree(dev
))
520 * To make our bookkeeping simpler, we don't unmap DB
521 * pages until we clean up the whole db table.
524 mutex_lock(&db_tab
->mutex
);
526 --db_tab
->page
[index
/ MTHCA_DB_REC_PER_PAGE
].refcount
;
528 mutex_unlock(&db_tab
->mutex
);
531 struct mthca_user_db_table
*mthca_init_user_db_tab(struct mthca_dev
*dev
)
533 struct mthca_user_db_table
*db_tab
;
537 if (!mthca_is_memfree(dev
))
540 npages
= dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
;
541 db_tab
= kmalloc(sizeof *db_tab
+ npages
* sizeof *db_tab
->page
, GFP_KERNEL
);
543 return ERR_PTR(-ENOMEM
);
545 mutex_init(&db_tab
->mutex
);
546 for (i
= 0; i
< npages
; ++i
) {
547 db_tab
->page
[i
].refcount
= 0;
548 db_tab
->page
[i
].uvirt
= 0;
549 sg_init_table(&db_tab
->page
[i
].mem
, 1);
555 void mthca_cleanup_user_db_tab(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
556 struct mthca_user_db_table
*db_tab
)
561 if (!mthca_is_memfree(dev
))
564 for (i
= 0; i
< dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
; ++i
) {
565 if (db_tab
->page
[i
].uvirt
) {
566 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, uar
, i
), 1, &status
);
567 pci_unmap_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
568 put_page(sg_page(&db_tab
->page
[i
].mem
));
575 int mthca_alloc_db(struct mthca_dev
*dev
, enum mthca_db_type type
,
581 struct mthca_db_page
*page
;
585 mutex_lock(&dev
->db_tab
->mutex
);
588 case MTHCA_DB_TYPE_CQ_ARM
:
589 case MTHCA_DB_TYPE_SQ
:
592 end
= dev
->db_tab
->max_group1
;
596 case MTHCA_DB_TYPE_CQ_SET_CI
:
597 case MTHCA_DB_TYPE_RQ
:
598 case MTHCA_DB_TYPE_SRQ
:
600 start
= dev
->db_tab
->npages
- 1;
601 end
= dev
->db_tab
->min_group2
;
610 for (i
= start
; i
!= end
; i
+= dir
)
611 if (dev
->db_tab
->page
[i
].db_rec
&&
612 !bitmap_full(dev
->db_tab
->page
[i
].used
,
613 MTHCA_DB_REC_PER_PAGE
)) {
614 page
= dev
->db_tab
->page
+ i
;
618 for (i
= start
; i
!= end
; i
+= dir
)
619 if (!dev
->db_tab
->page
[i
].db_rec
) {
620 page
= dev
->db_tab
->page
+ i
;
624 if (dev
->db_tab
->max_group1
>= dev
->db_tab
->min_group2
- 1) {
630 ++dev
->db_tab
->max_group1
;
632 --dev
->db_tab
->min_group2
;
634 page
= dev
->db_tab
->page
+ end
;
637 page
->db_rec
= dma_alloc_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
638 &page
->mapping
, GFP_KERNEL
);
643 memset(page
->db_rec
, 0, MTHCA_ICM_PAGE_SIZE
);
645 ret
= mthca_MAP_ICM_page(dev
, page
->mapping
,
646 mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), &status
);
650 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
651 page
->db_rec
, page
->mapping
);
655 bitmap_zero(page
->used
, MTHCA_DB_REC_PER_PAGE
);
658 j
= find_first_zero_bit(page
->used
, MTHCA_DB_REC_PER_PAGE
);
659 set_bit(j
, page
->used
);
662 j
= MTHCA_DB_REC_PER_PAGE
- 1 - j
;
664 ret
= i
* MTHCA_DB_REC_PER_PAGE
+ j
;
666 page
->db_rec
[j
] = cpu_to_be64((qn
<< 8) | (type
<< 5));
668 *db
= (__be32
*) &page
->db_rec
[j
];
671 mutex_unlock(&dev
->db_tab
->mutex
);
676 void mthca_free_db(struct mthca_dev
*dev
, int type
, int db_index
)
679 struct mthca_db_page
*page
;
682 i
= db_index
/ MTHCA_DB_REC_PER_PAGE
;
683 j
= db_index
% MTHCA_DB_REC_PER_PAGE
;
685 page
= dev
->db_tab
->page
+ i
;
687 mutex_lock(&dev
->db_tab
->mutex
);
690 if (i
>= dev
->db_tab
->min_group2
)
691 j
= MTHCA_DB_REC_PER_PAGE
- 1 - j
;
692 clear_bit(j
, page
->used
);
694 if (bitmap_empty(page
->used
, MTHCA_DB_REC_PER_PAGE
) &&
695 i
>= dev
->db_tab
->max_group1
- 1) {
696 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), 1, &status
);
698 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
699 page
->db_rec
, page
->mapping
);
702 if (i
== dev
->db_tab
->max_group1
) {
703 --dev
->db_tab
->max_group1
;
704 /* XXX may be able to unmap more pages now */
706 if (i
== dev
->db_tab
->min_group2
)
707 ++dev
->db_tab
->min_group2
;
710 mutex_unlock(&dev
->db_tab
->mutex
);
713 int mthca_init_db_tab(struct mthca_dev
*dev
)
717 if (!mthca_is_memfree(dev
))
720 dev
->db_tab
= kmalloc(sizeof *dev
->db_tab
, GFP_KERNEL
);
724 mutex_init(&dev
->db_tab
->mutex
);
726 dev
->db_tab
->npages
= dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
;
727 dev
->db_tab
->max_group1
= 0;
728 dev
->db_tab
->min_group2
= dev
->db_tab
->npages
- 1;
730 dev
->db_tab
->page
= kmalloc(dev
->db_tab
->npages
*
731 sizeof *dev
->db_tab
->page
,
733 if (!dev
->db_tab
->page
) {
738 for (i
= 0; i
< dev
->db_tab
->npages
; ++i
)
739 dev
->db_tab
->page
[i
].db_rec
= NULL
;
744 void mthca_cleanup_db_tab(struct mthca_dev
*dev
)
749 if (!mthca_is_memfree(dev
))
753 * Because we don't always free our UARC pages when they
754 * become empty to make mthca_free_db() simpler we need to
755 * make a sweep through the doorbell pages and free any
756 * leftover pages now.
758 for (i
= 0; i
< dev
->db_tab
->npages
; ++i
) {
759 if (!dev
->db_tab
->page
[i
].db_rec
)
762 if (!bitmap_empty(dev
->db_tab
->page
[i
].used
, MTHCA_DB_REC_PER_PAGE
))
763 mthca_warn(dev
, "Kernel UARC page %d not empty\n", i
);
765 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), 1, &status
);
767 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
768 dev
->db_tab
->page
[i
].db_rec
,
769 dev
->db_tab
->page
[i
].mapping
);
772 kfree(dev
->db_tab
->page
);