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(chunk
->mem
[i
].page
,
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(chunk
->mem
[i
].page
),
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
)
110 mem
->page
= alloc_pages(gfp_mask
, order
);
114 mem
->length
= PAGE_SIZE
<< order
;
119 static int mthca_alloc_icm_coherent(struct device
*dev
, struct scatterlist
*mem
,
120 int order
, gfp_t gfp_mask
)
122 void *buf
= dma_alloc_coherent(dev
, PAGE_SIZE
<< order
, &sg_dma_address(mem
),
127 sg_set_buf(mem
, buf
, PAGE_SIZE
<< order
);
129 sg_dma_len(mem
) = PAGE_SIZE
<< order
;
133 struct mthca_icm
*mthca_alloc_icm(struct mthca_dev
*dev
, int npages
,
134 gfp_t gfp_mask
, int coherent
)
136 struct mthca_icm
*icm
;
137 struct mthca_icm_chunk
*chunk
= NULL
;
141 /* We use sg_set_buf for coherent allocs, which assumes low memory */
142 BUG_ON(coherent
&& (gfp_mask
& __GFP_HIGHMEM
));
144 icm
= kmalloc(sizeof *icm
, gfp_mask
& ~(__GFP_HIGHMEM
| __GFP_NOWARN
));
149 INIT_LIST_HEAD(&icm
->chunk_list
);
151 cur_order
= get_order(MTHCA_ICM_ALLOC_SIZE
);
155 chunk
= kmalloc(sizeof *chunk
,
156 gfp_mask
& ~(__GFP_HIGHMEM
| __GFP_NOWARN
));
162 list_add_tail(&chunk
->list
, &icm
->chunk_list
);
165 while (1 << cur_order
> npages
)
169 ret
= mthca_alloc_icm_coherent(&dev
->pdev
->dev
,
170 &chunk
->mem
[chunk
->npages
],
171 cur_order
, gfp_mask
);
173 ret
= mthca_alloc_icm_pages(&chunk
->mem
[chunk
->npages
],
174 cur_order
, gfp_mask
);
181 else if (chunk
->npages
== MTHCA_ICM_CHUNK_LEN
) {
182 chunk
->nsg
= pci_map_sg(dev
->pdev
, chunk
->mem
,
184 PCI_DMA_BIDIRECTIONAL
);
190 if (chunk
->npages
== MTHCA_ICM_CHUNK_LEN
)
193 npages
-= 1 << cur_order
;
201 if (!coherent
&& chunk
) {
202 chunk
->nsg
= pci_map_sg(dev
->pdev
, chunk
->mem
,
204 PCI_DMA_BIDIRECTIONAL
);
213 mthca_free_icm(dev
, icm
, coherent
);
217 int mthca_table_get(struct mthca_dev
*dev
, struct mthca_icm_table
*table
, int obj
)
219 int i
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
/ MTHCA_TABLE_CHUNK_SIZE
;
223 mutex_lock(&table
->mutex
);
226 ++table
->icm
[i
]->refcount
;
230 table
->icm
[i
] = mthca_alloc_icm(dev
, MTHCA_TABLE_CHUNK_SIZE
>> PAGE_SHIFT
,
231 (table
->lowmem
? GFP_KERNEL
: GFP_HIGHUSER
) |
232 __GFP_NOWARN
, table
->coherent
);
233 if (!table
->icm
[i
]) {
238 if (mthca_MAP_ICM(dev
, table
->icm
[i
], table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
239 &status
) || status
) {
240 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
241 table
->icm
[i
] = NULL
;
246 ++table
->icm
[i
]->refcount
;
249 mutex_unlock(&table
->mutex
);
253 void mthca_table_put(struct mthca_dev
*dev
, struct mthca_icm_table
*table
, int obj
)
258 if (!mthca_is_memfree(dev
))
261 i
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
/ MTHCA_TABLE_CHUNK_SIZE
;
263 mutex_lock(&table
->mutex
);
265 if (--table
->icm
[i
]->refcount
== 0) {
266 mthca_UNMAP_ICM(dev
, table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
267 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
,
269 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
270 table
->icm
[i
] = NULL
;
273 mutex_unlock(&table
->mutex
);
276 void *mthca_table_find(struct mthca_icm_table
*table
, int obj
, dma_addr_t
*dma_handle
)
278 int idx
, offset
, dma_offset
, i
;
279 struct mthca_icm_chunk
*chunk
;
280 struct mthca_icm
*icm
;
281 struct page
*page
= NULL
;
286 mutex_lock(&table
->mutex
);
288 idx
= (obj
& (table
->num_obj
- 1)) * table
->obj_size
;
289 icm
= table
->icm
[idx
/ MTHCA_TABLE_CHUNK_SIZE
];
290 dma_offset
= offset
= idx
% MTHCA_TABLE_CHUNK_SIZE
;
295 list_for_each_entry(chunk
, &icm
->chunk_list
, list
) {
296 for (i
= 0; i
< chunk
->npages
; ++i
) {
297 if (dma_handle
&& dma_offset
>= 0) {
298 if (sg_dma_len(&chunk
->mem
[i
]) > dma_offset
)
299 *dma_handle
= sg_dma_address(&chunk
->mem
[i
]) +
301 dma_offset
-= sg_dma_len(&chunk
->mem
[i
]);
303 /* DMA mapping can merge pages but not split them,
304 * so if we found the page, dma_handle has already
305 * been assigned to. */
306 if (chunk
->mem
[i
].length
> offset
) {
307 page
= chunk
->mem
[i
].page
;
310 offset
-= chunk
->mem
[i
].length
;
315 mutex_unlock(&table
->mutex
);
316 return page
? lowmem_page_address(page
) + offset
: NULL
;
319 int mthca_table_get_range(struct mthca_dev
*dev
, struct mthca_icm_table
*table
,
322 int inc
= MTHCA_TABLE_CHUNK_SIZE
/ table
->obj_size
;
325 for (i
= start
; i
<= end
; i
+= inc
) {
326 err
= mthca_table_get(dev
, table
, i
);
336 mthca_table_put(dev
, table
, i
);
342 void mthca_table_put_range(struct mthca_dev
*dev
, struct mthca_icm_table
*table
,
347 if (!mthca_is_memfree(dev
))
350 for (i
= start
; i
<= end
; i
+= MTHCA_TABLE_CHUNK_SIZE
/ table
->obj_size
)
351 mthca_table_put(dev
, table
, i
);
354 struct mthca_icm_table
*mthca_alloc_icm_table(struct mthca_dev
*dev
,
355 u64 virt
, int obj_size
,
356 int nobj
, int reserved
,
357 int use_lowmem
, int use_coherent
)
359 struct mthca_icm_table
*table
;
365 num_icm
= (obj_size
* nobj
+ MTHCA_TABLE_CHUNK_SIZE
- 1) / MTHCA_TABLE_CHUNK_SIZE
;
367 table
= kmalloc(sizeof *table
+ num_icm
* sizeof *table
->icm
, GFP_KERNEL
);
372 table
->num_icm
= num_icm
;
373 table
->num_obj
= nobj
;
374 table
->obj_size
= obj_size
;
375 table
->lowmem
= use_lowmem
;
376 table
->coherent
= use_coherent
;
377 mutex_init(&table
->mutex
);
379 for (i
= 0; i
< num_icm
; ++i
)
380 table
->icm
[i
] = NULL
;
382 for (i
= 0; i
* MTHCA_TABLE_CHUNK_SIZE
< reserved
* obj_size
; ++i
) {
383 chunk_size
= MTHCA_TABLE_CHUNK_SIZE
;
384 if ((i
+ 1) * MTHCA_TABLE_CHUNK_SIZE
> nobj
* obj_size
)
385 chunk_size
= nobj
* obj_size
- i
* MTHCA_TABLE_CHUNK_SIZE
;
387 table
->icm
[i
] = mthca_alloc_icm(dev
, chunk_size
>> PAGE_SHIFT
,
388 (use_lowmem
? GFP_KERNEL
: GFP_HIGHUSER
) |
389 __GFP_NOWARN
, use_coherent
);
392 if (mthca_MAP_ICM(dev
, table
->icm
[i
], virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
393 &status
) || status
) {
394 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
395 table
->icm
[i
] = NULL
;
400 * Add a reference to this ICM chunk so that it never
401 * gets freed (since it contains reserved firmware objects).
403 ++table
->icm
[i
]->refcount
;
409 for (i
= 0; i
< num_icm
; ++i
)
411 mthca_UNMAP_ICM(dev
, virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
412 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
,
414 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
422 void mthca_free_icm_table(struct mthca_dev
*dev
, struct mthca_icm_table
*table
)
427 for (i
= 0; i
< table
->num_icm
; ++i
)
429 mthca_UNMAP_ICM(dev
, table
->virt
+ i
* MTHCA_TABLE_CHUNK_SIZE
,
430 MTHCA_TABLE_CHUNK_SIZE
/ MTHCA_ICM_PAGE_SIZE
,
432 mthca_free_icm(dev
, table
->icm
[i
], table
->coherent
);
438 static u64
mthca_uarc_virt(struct mthca_dev
*dev
, struct mthca_uar
*uar
, int page
)
440 return dev
->uar_table
.uarc_base
+
441 uar
->index
* dev
->uar_table
.uarc_size
+
442 page
* MTHCA_ICM_PAGE_SIZE
;
445 int mthca_map_user_db(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
446 struct mthca_user_db_table
*db_tab
, int index
, u64 uaddr
)
452 if (!mthca_is_memfree(dev
))
455 if (index
< 0 || index
> dev
->uar_table
.uarc_size
/ 8)
458 mutex_lock(&db_tab
->mutex
);
460 i
= index
/ MTHCA_DB_REC_PER_PAGE
;
462 if ((db_tab
->page
[i
].refcount
>= MTHCA_DB_REC_PER_PAGE
) ||
463 (db_tab
->page
[i
].uvirt
&& db_tab
->page
[i
].uvirt
!= uaddr
) ||
469 if (db_tab
->page
[i
].refcount
) {
470 ++db_tab
->page
[i
].refcount
;
474 ret
= get_user_pages(current
, current
->mm
, uaddr
& PAGE_MASK
, 1, 1, 0,
475 &db_tab
->page
[i
].mem
.page
, NULL
);
479 db_tab
->page
[i
].mem
.length
= MTHCA_ICM_PAGE_SIZE
;
480 db_tab
->page
[i
].mem
.offset
= uaddr
& ~PAGE_MASK
;
482 ret
= pci_map_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
484 put_page(db_tab
->page
[i
].mem
.page
);
488 ret
= mthca_MAP_ICM_page(dev
, sg_dma_address(&db_tab
->page
[i
].mem
),
489 mthca_uarc_virt(dev
, uar
, i
), &status
);
493 pci_unmap_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
494 put_page(db_tab
->page
[i
].mem
.page
);
498 db_tab
->page
[i
].uvirt
= uaddr
;
499 db_tab
->page
[i
].refcount
= 1;
502 mutex_unlock(&db_tab
->mutex
);
506 void mthca_unmap_user_db(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
507 struct mthca_user_db_table
*db_tab
, int index
)
509 if (!mthca_is_memfree(dev
))
513 * To make our bookkeeping simpler, we don't unmap DB
514 * pages until we clean up the whole db table.
517 mutex_lock(&db_tab
->mutex
);
519 --db_tab
->page
[index
/ MTHCA_DB_REC_PER_PAGE
].refcount
;
521 mutex_unlock(&db_tab
->mutex
);
524 struct mthca_user_db_table
*mthca_init_user_db_tab(struct mthca_dev
*dev
)
526 struct mthca_user_db_table
*db_tab
;
530 if (!mthca_is_memfree(dev
))
533 npages
= dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
;
534 db_tab
= kmalloc(sizeof *db_tab
+ npages
* sizeof *db_tab
->page
, GFP_KERNEL
);
536 return ERR_PTR(-ENOMEM
);
538 mutex_init(&db_tab
->mutex
);
539 for (i
= 0; i
< npages
; ++i
) {
540 db_tab
->page
[i
].refcount
= 0;
541 db_tab
->page
[i
].uvirt
= 0;
547 void mthca_cleanup_user_db_tab(struct mthca_dev
*dev
, struct mthca_uar
*uar
,
548 struct mthca_user_db_table
*db_tab
)
553 if (!mthca_is_memfree(dev
))
556 for (i
= 0; i
< dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
; ++i
) {
557 if (db_tab
->page
[i
].uvirt
) {
558 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, uar
, i
), 1, &status
);
559 pci_unmap_sg(dev
->pdev
, &db_tab
->page
[i
].mem
, 1, PCI_DMA_TODEVICE
);
560 put_page(db_tab
->page
[i
].mem
.page
);
567 int mthca_alloc_db(struct mthca_dev
*dev
, enum mthca_db_type type
,
573 struct mthca_db_page
*page
;
577 mutex_lock(&dev
->db_tab
->mutex
);
580 case MTHCA_DB_TYPE_CQ_ARM
:
581 case MTHCA_DB_TYPE_SQ
:
584 end
= dev
->db_tab
->max_group1
;
588 case MTHCA_DB_TYPE_CQ_SET_CI
:
589 case MTHCA_DB_TYPE_RQ
:
590 case MTHCA_DB_TYPE_SRQ
:
592 start
= dev
->db_tab
->npages
- 1;
593 end
= dev
->db_tab
->min_group2
;
602 for (i
= start
; i
!= end
; i
+= dir
)
603 if (dev
->db_tab
->page
[i
].db_rec
&&
604 !bitmap_full(dev
->db_tab
->page
[i
].used
,
605 MTHCA_DB_REC_PER_PAGE
)) {
606 page
= dev
->db_tab
->page
+ i
;
610 for (i
= start
; i
!= end
; i
+= dir
)
611 if (!dev
->db_tab
->page
[i
].db_rec
) {
612 page
= dev
->db_tab
->page
+ i
;
616 if (dev
->db_tab
->max_group1
>= dev
->db_tab
->min_group2
- 1) {
622 ++dev
->db_tab
->max_group1
;
624 --dev
->db_tab
->min_group2
;
626 page
= dev
->db_tab
->page
+ end
;
629 page
->db_rec
= dma_alloc_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
630 &page
->mapping
, GFP_KERNEL
);
635 memset(page
->db_rec
, 0, MTHCA_ICM_PAGE_SIZE
);
637 ret
= mthca_MAP_ICM_page(dev
, page
->mapping
,
638 mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), &status
);
642 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
643 page
->db_rec
, page
->mapping
);
647 bitmap_zero(page
->used
, MTHCA_DB_REC_PER_PAGE
);
650 j
= find_first_zero_bit(page
->used
, MTHCA_DB_REC_PER_PAGE
);
651 set_bit(j
, page
->used
);
654 j
= MTHCA_DB_REC_PER_PAGE
- 1 - j
;
656 ret
= i
* MTHCA_DB_REC_PER_PAGE
+ j
;
658 page
->db_rec
[j
] = cpu_to_be64((qn
<< 8) | (type
<< 5));
660 *db
= (__be32
*) &page
->db_rec
[j
];
663 mutex_unlock(&dev
->db_tab
->mutex
);
668 void mthca_free_db(struct mthca_dev
*dev
, int type
, int db_index
)
671 struct mthca_db_page
*page
;
674 i
= db_index
/ MTHCA_DB_REC_PER_PAGE
;
675 j
= db_index
% MTHCA_DB_REC_PER_PAGE
;
677 page
= dev
->db_tab
->page
+ i
;
679 mutex_lock(&dev
->db_tab
->mutex
);
682 if (i
>= dev
->db_tab
->min_group2
)
683 j
= MTHCA_DB_REC_PER_PAGE
- 1 - j
;
684 clear_bit(j
, page
->used
);
686 if (bitmap_empty(page
->used
, MTHCA_DB_REC_PER_PAGE
) &&
687 i
>= dev
->db_tab
->max_group1
- 1) {
688 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), 1, &status
);
690 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
691 page
->db_rec
, page
->mapping
);
694 if (i
== dev
->db_tab
->max_group1
) {
695 --dev
->db_tab
->max_group1
;
696 /* XXX may be able to unmap more pages now */
698 if (i
== dev
->db_tab
->min_group2
)
699 ++dev
->db_tab
->min_group2
;
702 mutex_unlock(&dev
->db_tab
->mutex
);
705 int mthca_init_db_tab(struct mthca_dev
*dev
)
709 if (!mthca_is_memfree(dev
))
712 dev
->db_tab
= kmalloc(sizeof *dev
->db_tab
, GFP_KERNEL
);
716 mutex_init(&dev
->db_tab
->mutex
);
718 dev
->db_tab
->npages
= dev
->uar_table
.uarc_size
/ MTHCA_ICM_PAGE_SIZE
;
719 dev
->db_tab
->max_group1
= 0;
720 dev
->db_tab
->min_group2
= dev
->db_tab
->npages
- 1;
722 dev
->db_tab
->page
= kmalloc(dev
->db_tab
->npages
*
723 sizeof *dev
->db_tab
->page
,
725 if (!dev
->db_tab
->page
) {
730 for (i
= 0; i
< dev
->db_tab
->npages
; ++i
)
731 dev
->db_tab
->page
[i
].db_rec
= NULL
;
736 void mthca_cleanup_db_tab(struct mthca_dev
*dev
)
741 if (!mthca_is_memfree(dev
))
745 * Because we don't always free our UARC pages when they
746 * become empty to make mthca_free_db() simpler we need to
747 * make a sweep through the doorbell pages and free any
748 * leftover pages now.
750 for (i
= 0; i
< dev
->db_tab
->npages
; ++i
) {
751 if (!dev
->db_tab
->page
[i
].db_rec
)
754 if (!bitmap_empty(dev
->db_tab
->page
[i
].used
, MTHCA_DB_REC_PER_PAGE
))
755 mthca_warn(dev
, "Kernel UARC page %d not empty\n", i
);
757 mthca_UNMAP_ICM(dev
, mthca_uarc_virt(dev
, &dev
->driver_uar
, i
), 1, &status
);
759 dma_free_coherent(&dev
->pdev
->dev
, MTHCA_ICM_PAGE_SIZE
,
760 dev
->db_tab
->page
[i
].db_rec
,
761 dev
->db_tab
->page
[i
].mapping
);
764 kfree(dev
->db_tab
->page
);
This page took 0.051281 seconds and 5 git commands to generate.