diff options
-rw-r--r-- | drivers/infiniband/hw/mthca/mthca_main.c | 36 | ||||
-rw-r--r-- | drivers/infiniband/hw/mthca/mthca_memfree.c | 127 | ||||
-rw-r--r-- | drivers/infiniband/hw/mthca/mthca_memfree.h | 9 | ||||
-rw-r--r-- | drivers/infiniband/hw/mthca/mthca_mr.c | 8 | ||||
-rw-r--r-- | drivers/infiniband/hw/mthca/mthca_provider.h | 1 |
5 files changed, 131 insertions, 50 deletions
diff --git a/drivers/infiniband/hw/mthca/mthca_main.c b/drivers/infiniband/hw/mthca/mthca_main.c index 9a9dd32885a0..0d9b7d06bbc2 100644 --- a/drivers/infiniband/hw/mthca/mthca_main.c +++ b/drivers/infiniband/hw/mthca/mthca_main.c | |||
@@ -379,7 +379,7 @@ static int mthca_load_fw(struct mthca_dev *mdev) | |||
379 | 379 | ||
380 | mdev->fw.arbel.fw_icm = | 380 | mdev->fw.arbel.fw_icm = |
381 | mthca_alloc_icm(mdev, mdev->fw.arbel.fw_pages, | 381 | mthca_alloc_icm(mdev, mdev->fw.arbel.fw_pages, |
382 | GFP_HIGHUSER | __GFP_NOWARN); | 382 | GFP_HIGHUSER | __GFP_NOWARN, 0); |
383 | if (!mdev->fw.arbel.fw_icm) { | 383 | if (!mdev->fw.arbel.fw_icm) { |
384 | mthca_err(mdev, "Couldn't allocate FW area, aborting.\n"); | 384 | mthca_err(mdev, "Couldn't allocate FW area, aborting.\n"); |
385 | return -ENOMEM; | 385 | return -ENOMEM; |
@@ -412,7 +412,7 @@ err_unmap_fa: | |||
412 | mthca_UNMAP_FA(mdev, &status); | 412 | mthca_UNMAP_FA(mdev, &status); |
413 | 413 | ||
414 | err_free: | 414 | err_free: |
415 | mthca_free_icm(mdev, mdev->fw.arbel.fw_icm); | 415 | mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0); |
416 | return err; | 416 | return err; |
417 | } | 417 | } |
418 | 418 | ||
@@ -441,7 +441,7 @@ static int mthca_init_icm(struct mthca_dev *mdev, | |||
441 | (unsigned long long) aux_pages << 2); | 441 | (unsigned long long) aux_pages << 2); |
442 | 442 | ||
443 | mdev->fw.arbel.aux_icm = mthca_alloc_icm(mdev, aux_pages, | 443 | mdev->fw.arbel.aux_icm = mthca_alloc_icm(mdev, aux_pages, |
444 | GFP_HIGHUSER | __GFP_NOWARN); | 444 | GFP_HIGHUSER | __GFP_NOWARN, 0); |
445 | if (!mdev->fw.arbel.aux_icm) { | 445 | if (!mdev->fw.arbel.aux_icm) { |
446 | mthca_err(mdev, "Couldn't allocate aux memory, aborting.\n"); | 446 | mthca_err(mdev, "Couldn't allocate aux memory, aborting.\n"); |
447 | return -ENOMEM; | 447 | return -ENOMEM; |
@@ -471,7 +471,8 @@ static int mthca_init_icm(struct mthca_dev *mdev, | |||
471 | mdev->mr_table.mtt_table = mthca_alloc_icm_table(mdev, init_hca->mtt_base, | 471 | mdev->mr_table.mtt_table = mthca_alloc_icm_table(mdev, init_hca->mtt_base, |
472 | MTHCA_MTT_SEG_SIZE, | 472 | MTHCA_MTT_SEG_SIZE, |
473 | mdev->limits.num_mtt_segs, | 473 | mdev->limits.num_mtt_segs, |
474 | mdev->limits.reserved_mtts, 1); | 474 | mdev->limits.reserved_mtts, |
475 | 1, 0); | ||
475 | if (!mdev->mr_table.mtt_table) { | 476 | if (!mdev->mr_table.mtt_table) { |
476 | mthca_err(mdev, "Failed to map MTT context memory, aborting.\n"); | 477 | mthca_err(mdev, "Failed to map MTT context memory, aborting.\n"); |
477 | err = -ENOMEM; | 478 | err = -ENOMEM; |
@@ -481,7 +482,8 @@ static int mthca_init_icm(struct mthca_dev *mdev, | |||
481 | mdev->mr_table.mpt_table = mthca_alloc_icm_table(mdev, init_hca->mpt_base, | 482 | mdev->mr_table.mpt_table = mthca_alloc_icm_table(mdev, init_hca->mpt_base, |
482 | dev_lim->mpt_entry_sz, | 483 | dev_lim->mpt_entry_sz, |
483 | mdev->limits.num_mpts, | 484 | mdev->limits.num_mpts, |
484 | mdev->limits.reserved_mrws, 1); | 485 | mdev->limits.reserved_mrws, |
486 | 1, 1); | ||
485 | if (!mdev->mr_table.mpt_table) { | 487 | if (!mdev->mr_table.mpt_table) { |
486 | mthca_err(mdev, "Failed to map MPT context memory, aborting.\n"); | 488 | mthca_err(mdev, "Failed to map MPT context memory, aborting.\n"); |
487 | err = -ENOMEM; | 489 | err = -ENOMEM; |
@@ -491,7 +493,8 @@ static int mthca_init_icm(struct mthca_dev *mdev, | |||
491 | mdev->qp_table.qp_table = mthca_alloc_icm_table(mdev, init_hca->qpc_base, | 493 | mdev->qp_table.qp_table = mthca_alloc_icm_table(mdev, init_hca->qpc_base, |
492 | dev_lim->qpc_entry_sz, | 494 | dev_lim->qpc_entry_sz, |
493 | mdev->limits.num_qps, | 495 | mdev->limits.num_qps, |
494 | mdev->limits.reserved_qps, 0); | 496 | mdev->limits.reserved_qps, |
497 | 0, 0); | ||
495 | if (!mdev->qp_table.qp_table) { | 498 | if (!mdev->qp_table.qp_table) { |
496 | mthca_err(mdev, "Failed to map QP context memory, aborting.\n"); | 499 | mthca_err(mdev, "Failed to map QP context memory, aborting.\n"); |
497 | err = -ENOMEM; | 500 | err = -ENOMEM; |
@@ -501,7 +504,8 @@ static int mthca_init_icm(struct mthca_dev *mdev, | |||
501 | mdev->qp_table.eqp_table = mthca_alloc_icm_table(mdev, init_hca->eqpc_base, | 504 | mdev->qp_table.eqp_table = mthca_alloc_icm_table(mdev, init_hca->eqpc_base, |
502 | dev_lim->eqpc_entry_sz, | 505 | dev_lim->eqpc_entry_sz, |
503 | mdev->limits.num_qps, | 506 | mdev->limits.num_qps, |
504 | mdev->limits.reserved_qps, 0); | 507 | mdev->limits.reserved_qps, |
508 | 0, 0); | ||
505 | if (!mdev->qp_table.eqp_table) { | 509 | if (!mdev->qp_table.eqp_table) { |
506 | mthca_err(mdev, "Failed to map EQP context memory, aborting.\n"); | 510 | mthca_err(mdev, "Failed to map EQP context memory, aborting.\n"); |
507 | err = -ENOMEM; | 511 | err = -ENOMEM; |
@@ -511,7 +515,7 @@ static int mthca_init_icm(struct mthca_dev *mdev, | |||
511 | mdev->qp_table.rdb_table = mthca_alloc_icm_table(mdev, init_hca->rdb_base, | 515 | mdev->qp_table.rdb_table = mthca_alloc_icm_table(mdev, init_hca->rdb_base, |
512 | MTHCA_RDB_ENTRY_SIZE, | 516 | MTHCA_RDB_ENTRY_SIZE, |
513 | mdev->limits.num_qps << | 517 | mdev->limits.num_qps << |
514 | mdev->qp_table.rdb_shift, | 518 | mdev->qp_table.rdb_shift, 0, |
515 | 0, 0); | 519 | 0, 0); |
516 | if (!mdev->qp_table.rdb_table) { | 520 | if (!mdev->qp_table.rdb_table) { |
517 | mthca_err(mdev, "Failed to map RDB context memory, aborting\n"); | 521 | mthca_err(mdev, "Failed to map RDB context memory, aborting\n"); |
@@ -522,7 +526,8 @@ static int mthca_init_icm(struct mthca_dev *mdev, | |||
522 | mdev->cq_table.table = mthca_alloc_icm_table(mdev, init_hca->cqc_base, | 526 | mdev->cq_table.table = mthca_alloc_icm_table(mdev, init_hca->cqc_base, |
523 | dev_lim->cqc_entry_sz, | 527 | dev_lim->cqc_entry_sz, |
524 | mdev->limits.num_cqs, | 528 | mdev->limits.num_cqs, |
525 | mdev->limits.reserved_cqs, 0); | 529 | mdev->limits.reserved_cqs, |
530 | 0, 0); | ||
526 | if (!mdev->cq_table.table) { | 531 | if (!mdev->cq_table.table) { |
527 | mthca_err(mdev, "Failed to map CQ context memory, aborting.\n"); | 532 | mthca_err(mdev, "Failed to map CQ context memory, aborting.\n"); |
528 | err = -ENOMEM; | 533 | err = -ENOMEM; |
@@ -534,7 +539,8 @@ static int mthca_init_icm(struct mthca_dev *mdev, | |||
534 | mthca_alloc_icm_table(mdev, init_hca->srqc_base, | 539 | mthca_alloc_icm_table(mdev, init_hca->srqc_base, |
535 | dev_lim->srq_entry_sz, | 540 | dev_lim->srq_entry_sz, |
536 | mdev->limits.num_srqs, | 541 | mdev->limits.num_srqs, |
537 | mdev->limits.reserved_srqs, 0); | 542 | mdev->limits.reserved_srqs, |
543 | 0, 0); | ||
538 | if (!mdev->srq_table.table) { | 544 | if (!mdev->srq_table.table) { |
539 | mthca_err(mdev, "Failed to map SRQ context memory, " | 545 | mthca_err(mdev, "Failed to map SRQ context memory, " |
540 | "aborting.\n"); | 546 | "aborting.\n"); |
@@ -554,7 +560,7 @@ static int mthca_init_icm(struct mthca_dev *mdev, | |||
554 | mdev->limits.num_amgms, | 560 | mdev->limits.num_amgms, |
555 | mdev->limits.num_mgms + | 561 | mdev->limits.num_mgms + |
556 | mdev->limits.num_amgms, | 562 | mdev->limits.num_amgms, |
557 | 0); | 563 | 0, 0); |
558 | if (!mdev->mcg_table.table) { | 564 | if (!mdev->mcg_table.table) { |
559 | mthca_err(mdev, "Failed to map MCG context memory, aborting.\n"); | 565 | mthca_err(mdev, "Failed to map MCG context memory, aborting.\n"); |
560 | err = -ENOMEM; | 566 | err = -ENOMEM; |
@@ -592,7 +598,7 @@ err_unmap_aux: | |||
592 | mthca_UNMAP_ICM_AUX(mdev, &status); | 598 | mthca_UNMAP_ICM_AUX(mdev, &status); |
593 | 599 | ||
594 | err_free_aux: | 600 | err_free_aux: |
595 | mthca_free_icm(mdev, mdev->fw.arbel.aux_icm); | 601 | mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0); |
596 | 602 | ||
597 | return err; | 603 | return err; |
598 | } | 604 | } |
@@ -613,7 +619,7 @@ static void mthca_free_icms(struct mthca_dev *mdev) | |||
613 | mthca_unmap_eq_icm(mdev); | 619 | mthca_unmap_eq_icm(mdev); |
614 | 620 | ||
615 | mthca_UNMAP_ICM_AUX(mdev, &status); | 621 | mthca_UNMAP_ICM_AUX(mdev, &status); |
616 | mthca_free_icm(mdev, mdev->fw.arbel.aux_icm); | 622 | mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0); |
617 | } | 623 | } |
618 | 624 | ||
619 | static int mthca_init_arbel(struct mthca_dev *mdev) | 625 | static int mthca_init_arbel(struct mthca_dev *mdev) |
@@ -697,7 +703,7 @@ err_free_icm: | |||
697 | 703 | ||
698 | err_stop_fw: | 704 | err_stop_fw: |
699 | mthca_UNMAP_FA(mdev, &status); | 705 | mthca_UNMAP_FA(mdev, &status); |
700 | mthca_free_icm(mdev, mdev->fw.arbel.fw_icm); | 706 | mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0); |
701 | 707 | ||
702 | err_disable: | 708 | err_disable: |
703 | if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM)) | 709 | if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM)) |
@@ -716,7 +722,7 @@ static void mthca_close_hca(struct mthca_dev *mdev) | |||
716 | mthca_free_icms(mdev); | 722 | mthca_free_icms(mdev); |
717 | 723 | ||
718 | mthca_UNMAP_FA(mdev, &status); | 724 | mthca_UNMAP_FA(mdev, &status); |
719 | mthca_free_icm(mdev, mdev->fw.arbel.fw_icm); | 725 | mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0); |
720 | 726 | ||
721 | if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM)) | 727 | if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM)) |
722 | mthca_DISABLE_LAM(mdev, &status); | 728 | mthca_DISABLE_LAM(mdev, &status); |
diff --git a/drivers/infiniband/hw/mthca/mthca_memfree.c b/drivers/infiniband/hw/mthca/mthca_memfree.c index 6b19645d946c..0b9d053a599d 100644 --- a/drivers/infiniband/hw/mthca/mthca_memfree.c +++ b/drivers/infiniband/hw/mthca/mthca_memfree.c | |||
@@ -35,6 +35,9 @@ | |||
35 | */ | 35 | */ |
36 | 36 | ||
37 | #include <linux/mm.h> | 37 | #include <linux/mm.h> |
38 | #include <linux/scatterlist.h> | ||
39 | |||
40 | #include <asm/page.h> | ||
38 | 41 | ||
39 | #include "mthca_memfree.h" | 42 | #include "mthca_memfree.h" |
40 | #include "mthca_dev.h" | 43 | #include "mthca_dev.h" |
@@ -58,22 +61,42 @@ struct mthca_user_db_table { | |||
58 | } page[0]; | 61 | } page[0]; |
59 | }; | 62 | }; |
60 | 63 | ||
61 | void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm) | 64 | static void mthca_free_icm_pages(struct mthca_dev *dev, struct mthca_icm_chunk *chunk) |
65 | { | ||
66 | int i; | ||
67 | |||
68 | if (chunk->nsg > 0) | ||
69 | pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages, | ||
70 | PCI_DMA_BIDIRECTIONAL); | ||
71 | |||
72 | for (i = 0; i < chunk->npages; ++i) | ||
73 | __free_pages(chunk->mem[i].page, | ||
74 | get_order(chunk->mem[i].length)); | ||
75 | } | ||
76 | |||
77 | static void mthca_free_icm_coherent(struct mthca_dev *dev, struct mthca_icm_chunk *chunk) | ||
62 | { | 78 | { |
63 | struct mthca_icm_chunk *chunk, *tmp; | ||
64 | int i; | 79 | int i; |
65 | 80 | ||
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])); | ||
85 | } | ||
86 | } | ||
87 | |||
88 | void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent) | ||
89 | { | ||
90 | struct mthca_icm_chunk *chunk, *tmp; | ||
91 | |||
66 | if (!icm) | 92 | if (!icm) |
67 | return; | 93 | return; |
68 | 94 | ||
69 | list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) { | 95 | list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) { |
70 | if (chunk->nsg > 0) | 96 | if (coherent) |
71 | pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages, | 97 | mthca_free_icm_coherent(dev, chunk); |
72 | PCI_DMA_BIDIRECTIONAL); | 98 | else |
73 | 99 | mthca_free_icm_pages(dev, chunk); | |
74 | for (i = 0; i < chunk->npages; ++i) | ||
75 | __free_pages(chunk->mem[i].page, | ||
76 | get_order(chunk->mem[i].length)); | ||
77 | 100 | ||
78 | kfree(chunk); | 101 | kfree(chunk); |
79 | } | 102 | } |
@@ -81,12 +104,41 @@ void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm) | |||
81 | kfree(icm); | 104 | kfree(icm); |
82 | } | 105 | } |
83 | 106 | ||
107 | static int mthca_alloc_icm_pages(struct scatterlist *mem, int order, gfp_t gfp_mask) | ||
108 | { | ||
109 | mem->page = alloc_pages(gfp_mask, order); | ||
110 | if (!mem->page) | ||
111 | return -ENOMEM; | ||
112 | |||
113 | mem->length = PAGE_SIZE << order; | ||
114 | mem->offset = 0; | ||
115 | return 0; | ||
116 | } | ||
117 | |||
118 | static int mthca_alloc_icm_coherent(struct device *dev, struct scatterlist *mem, | ||
119 | int order, gfp_t gfp_mask) | ||
120 | { | ||
121 | void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order, &sg_dma_address(mem), | ||
122 | gfp_mask); | ||
123 | if (!buf) | ||
124 | return -ENOMEM; | ||
125 | |||
126 | sg_set_buf(mem, buf, PAGE_SIZE << order); | ||
127 | BUG_ON(mem->offset); | ||
128 | sg_dma_len(mem) = PAGE_SIZE << order; | ||
129 | return 0; | ||
130 | } | ||
131 | |||
84 | struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages, | 132 | struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages, |
85 | gfp_t gfp_mask) | 133 | gfp_t gfp_mask, int coherent) |
86 | { | 134 | { |
87 | struct mthca_icm *icm; | 135 | struct mthca_icm *icm; |
88 | struct mthca_icm_chunk *chunk = NULL; | 136 | struct mthca_icm_chunk *chunk = NULL; |
89 | int cur_order; | 137 | int cur_order; |
138 | int ret; | ||
139 | |||
140 | /* We use sg_set_buf for coherent allocs, which assumes low memory */ | ||
141 | BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM)); | ||
90 | 142 | ||
91 | icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN)); | 143 | icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN)); |
92 | if (!icm) | 144 | if (!icm) |
@@ -112,21 +164,28 @@ struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages, | |||
112 | while (1 << cur_order > npages) | 164 | while (1 << cur_order > npages) |
113 | --cur_order; | 165 | --cur_order; |
114 | 166 | ||
115 | chunk->mem[chunk->npages].page = alloc_pages(gfp_mask, cur_order); | 167 | if (coherent) |
116 | if (chunk->mem[chunk->npages].page) { | 168 | ret = mthca_alloc_icm_coherent(&dev->pdev->dev, |
117 | chunk->mem[chunk->npages].length = PAGE_SIZE << cur_order; | 169 | &chunk->mem[chunk->npages], |
118 | chunk->mem[chunk->npages].offset = 0; | 170 | cur_order, gfp_mask); |
171 | else | ||
172 | ret = mthca_alloc_icm_pages(&chunk->mem[chunk->npages], | ||
173 | cur_order, gfp_mask); | ||
119 | 174 | ||
120 | if (++chunk->npages == MTHCA_ICM_CHUNK_LEN) { | 175 | if (!ret) { |
176 | ++chunk->npages; | ||
177 | |||
178 | if (!coherent && chunk->npages == MTHCA_ICM_CHUNK_LEN) { | ||
121 | chunk->nsg = pci_map_sg(dev->pdev, chunk->mem, | 179 | chunk->nsg = pci_map_sg(dev->pdev, chunk->mem, |
122 | chunk->npages, | 180 | chunk->npages, |
123 | PCI_DMA_BIDIRECTIONAL); | 181 | PCI_DMA_BIDIRECTIONAL); |
124 | 182 | ||
125 | if (chunk->nsg <= 0) | 183 | if (chunk->nsg <= 0) |
126 | goto fail; | 184 | goto fail; |
185 | } | ||
127 | 186 | ||
187 | if (chunk->npages == MTHCA_ICM_CHUNK_LEN) | ||
128 | chunk = NULL; | 188 | chunk = NULL; |
129 | } | ||
130 | 189 | ||
131 | npages -= 1 << cur_order; | 190 | npages -= 1 << cur_order; |
132 | } else { | 191 | } else { |
@@ -136,7 +195,7 @@ struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages, | |||
136 | } | 195 | } |
137 | } | 196 | } |
138 | 197 | ||
139 | if (chunk) { | 198 | if (!coherent && chunk) { |
140 | chunk->nsg = pci_map_sg(dev->pdev, chunk->mem, | 199 | chunk->nsg = pci_map_sg(dev->pdev, chunk->mem, |
141 | chunk->npages, | 200 | chunk->npages, |
142 | PCI_DMA_BIDIRECTIONAL); | 201 | PCI_DMA_BIDIRECTIONAL); |
@@ -148,7 +207,7 @@ struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages, | |||
148 | return icm; | 207 | return icm; |
149 | 208 | ||
150 | fail: | 209 | fail: |
151 | mthca_free_icm(dev, icm); | 210 | mthca_free_icm(dev, icm, coherent); |
152 | return NULL; | 211 | return NULL; |
153 | } | 212 | } |
154 | 213 | ||
@@ -167,7 +226,7 @@ int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int ob | |||
167 | 226 | ||
168 | table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT, | 227 | table->icm[i] = mthca_alloc_icm(dev, MTHCA_TABLE_CHUNK_SIZE >> PAGE_SHIFT, |
169 | (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) | | 228 | (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) | |
170 | __GFP_NOWARN); | 229 | __GFP_NOWARN, table->coherent); |
171 | if (!table->icm[i]) { | 230 | if (!table->icm[i]) { |
172 | ret = -ENOMEM; | 231 | ret = -ENOMEM; |
173 | goto out; | 232 | goto out; |
@@ -175,7 +234,7 @@ int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int ob | |||
175 | 234 | ||
176 | if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE, | 235 | if (mthca_MAP_ICM(dev, table->icm[i], table->virt + i * MTHCA_TABLE_CHUNK_SIZE, |
177 | &status) || status) { | 236 | &status) || status) { |
178 | mthca_free_icm(dev, table->icm[i]); | 237 | mthca_free_icm(dev, table->icm[i], table->coherent); |
179 | table->icm[i] = NULL; | 238 | table->icm[i] = NULL; |
180 | ret = -ENOMEM; | 239 | ret = -ENOMEM; |
181 | goto out; | 240 | goto out; |
@@ -204,16 +263,16 @@ void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int o | |||
204 | mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE, | 263 | mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE, |
205 | MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE, | 264 | MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE, |
206 | &status); | 265 | &status); |
207 | mthca_free_icm(dev, table->icm[i]); | 266 | mthca_free_icm(dev, table->icm[i], table->coherent); |
208 | table->icm[i] = NULL; | 267 | table->icm[i] = NULL; |
209 | } | 268 | } |
210 | 269 | ||
211 | mutex_unlock(&table->mutex); | 270 | mutex_unlock(&table->mutex); |
212 | } | 271 | } |
213 | 272 | ||
214 | void *mthca_table_find(struct mthca_icm_table *table, int obj) | 273 | void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle) |
215 | { | 274 | { |
216 | int idx, offset, i; | 275 | int idx, offset, dma_offset, i; |
217 | struct mthca_icm_chunk *chunk; | 276 | struct mthca_icm_chunk *chunk; |
218 | struct mthca_icm *icm; | 277 | struct mthca_icm *icm; |
219 | struct page *page = NULL; | 278 | struct page *page = NULL; |
@@ -225,13 +284,22 @@ void *mthca_table_find(struct mthca_icm_table *table, int obj) | |||
225 | 284 | ||
226 | idx = (obj & (table->num_obj - 1)) * table->obj_size; | 285 | idx = (obj & (table->num_obj - 1)) * table->obj_size; |
227 | icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE]; | 286 | icm = table->icm[idx / MTHCA_TABLE_CHUNK_SIZE]; |
228 | offset = idx % MTHCA_TABLE_CHUNK_SIZE; | 287 | dma_offset = offset = idx % MTHCA_TABLE_CHUNK_SIZE; |
229 | 288 | ||
230 | if (!icm) | 289 | if (!icm) |
231 | goto out; | 290 | goto out; |
232 | 291 | ||
233 | list_for_each_entry(chunk, &icm->chunk_list, list) { | 292 | list_for_each_entry(chunk, &icm->chunk_list, list) { |
234 | for (i = 0; i < chunk->npages; ++i) { | 293 | for (i = 0; i < chunk->npages; ++i) { |
294 | if (dma_handle && dma_offset >= 0) { | ||
295 | if (sg_dma_len(&chunk->mem[i]) > dma_offset) | ||
296 | *dma_handle = sg_dma_address(&chunk->mem[i]) + | ||
297 | dma_offset; | ||
298 | dma_offset -= sg_dma_len(&chunk->mem[i]); | ||
299 | } | ||
300 | /* DMA mapping can merge pages but not split them, | ||
301 | * so if we found the page, dma_handle has already | ||
302 | * been assigned to. */ | ||
235 | if (chunk->mem[i].length > offset) { | 303 | if (chunk->mem[i].length > offset) { |
236 | page = chunk->mem[i].page; | 304 | page = chunk->mem[i].page; |
237 | goto out; | 305 | goto out; |
@@ -283,7 +351,7 @@ void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table, | |||
283 | struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev, | 351 | struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev, |
284 | u64 virt, int obj_size, | 352 | u64 virt, int obj_size, |
285 | int nobj, int reserved, | 353 | int nobj, int reserved, |
286 | int use_lowmem) | 354 | int use_lowmem, int use_coherent) |
287 | { | 355 | { |
288 | struct mthca_icm_table *table; | 356 | struct mthca_icm_table *table; |
289 | int num_icm; | 357 | int num_icm; |
@@ -302,6 +370,7 @@ struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev, | |||
302 | table->num_obj = nobj; | 370 | table->num_obj = nobj; |
303 | table->obj_size = obj_size; | 371 | table->obj_size = obj_size; |
304 | table->lowmem = use_lowmem; | 372 | table->lowmem = use_lowmem; |
373 | table->coherent = use_coherent; | ||
305 | mutex_init(&table->mutex); | 374 | mutex_init(&table->mutex); |
306 | 375 | ||
307 | for (i = 0; i < num_icm; ++i) | 376 | for (i = 0; i < num_icm; ++i) |
@@ -314,12 +383,12 @@ struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev, | |||
314 | 383 | ||
315 | table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT, | 384 | table->icm[i] = mthca_alloc_icm(dev, chunk_size >> PAGE_SHIFT, |
316 | (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) | | 385 | (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) | |
317 | __GFP_NOWARN); | 386 | __GFP_NOWARN, use_coherent); |
318 | if (!table->icm[i]) | 387 | if (!table->icm[i]) |
319 | goto err; | 388 | goto err; |
320 | if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE, | 389 | if (mthca_MAP_ICM(dev, table->icm[i], virt + i * MTHCA_TABLE_CHUNK_SIZE, |
321 | &status) || status) { | 390 | &status) || status) { |
322 | mthca_free_icm(dev, table->icm[i]); | 391 | mthca_free_icm(dev, table->icm[i], table->coherent); |
323 | table->icm[i] = NULL; | 392 | table->icm[i] = NULL; |
324 | goto err; | 393 | goto err; |
325 | } | 394 | } |
@@ -339,7 +408,7 @@ err: | |||
339 | mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE, | 408 | mthca_UNMAP_ICM(dev, virt + i * MTHCA_TABLE_CHUNK_SIZE, |
340 | MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE, | 409 | MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE, |
341 | &status); | 410 | &status); |
342 | mthca_free_icm(dev, table->icm[i]); | 411 | mthca_free_icm(dev, table->icm[i], table->coherent); |
343 | } | 412 | } |
344 | 413 | ||
345 | kfree(table); | 414 | kfree(table); |
@@ -357,7 +426,7 @@ void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table) | |||
357 | mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE, | 426 | mthca_UNMAP_ICM(dev, table->virt + i * MTHCA_TABLE_CHUNK_SIZE, |
358 | MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE, | 427 | MTHCA_TABLE_CHUNK_SIZE / MTHCA_ICM_PAGE_SIZE, |
359 | &status); | 428 | &status); |
360 | mthca_free_icm(dev, table->icm[i]); | 429 | mthca_free_icm(dev, table->icm[i], table->coherent); |
361 | } | 430 | } |
362 | 431 | ||
363 | kfree(table); | 432 | kfree(table); |
diff --git a/drivers/infiniband/hw/mthca/mthca_memfree.h b/drivers/infiniband/hw/mthca/mthca_memfree.h index 6d42947e1dc4..594144145f45 100644 --- a/drivers/infiniband/hw/mthca/mthca_memfree.h +++ b/drivers/infiniband/hw/mthca/mthca_memfree.h | |||
@@ -69,6 +69,7 @@ struct mthca_icm_table { | |||
69 | int num_obj; | 69 | int num_obj; |
70 | int obj_size; | 70 | int obj_size; |
71 | int lowmem; | 71 | int lowmem; |
72 | int coherent; | ||
72 | struct mutex mutex; | 73 | struct mutex mutex; |
73 | struct mthca_icm *icm[0]; | 74 | struct mthca_icm *icm[0]; |
74 | }; | 75 | }; |
@@ -82,17 +83,17 @@ struct mthca_icm_iter { | |||
82 | struct mthca_dev; | 83 | struct mthca_dev; |
83 | 84 | ||
84 | struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages, | 85 | struct mthca_icm *mthca_alloc_icm(struct mthca_dev *dev, int npages, |
85 | gfp_t gfp_mask); | 86 | gfp_t gfp_mask, int coherent); |
86 | void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm); | 87 | void mthca_free_icm(struct mthca_dev *dev, struct mthca_icm *icm, int coherent); |
87 | 88 | ||
88 | struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev, | 89 | struct mthca_icm_table *mthca_alloc_icm_table(struct mthca_dev *dev, |
89 | u64 virt, int obj_size, | 90 | u64 virt, int obj_size, |
90 | int nobj, int reserved, | 91 | int nobj, int reserved, |
91 | int use_lowmem); | 92 | int use_lowmem, int use_coherent); |
92 | void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table); | 93 | void mthca_free_icm_table(struct mthca_dev *dev, struct mthca_icm_table *table); |
93 | int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj); | 94 | int mthca_table_get(struct mthca_dev *dev, struct mthca_icm_table *table, int obj); |
94 | void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj); | 95 | void mthca_table_put(struct mthca_dev *dev, struct mthca_icm_table *table, int obj); |
95 | void *mthca_table_find(struct mthca_icm_table *table, int obj); | 96 | void *mthca_table_find(struct mthca_icm_table *table, int obj, dma_addr_t *dma_handle); |
96 | int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table, | 97 | int mthca_table_get_range(struct mthca_dev *dev, struct mthca_icm_table *table, |
97 | int start, int end); | 98 | int start, int end); |
98 | void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table, | 99 | void mthca_table_put_range(struct mthca_dev *dev, struct mthca_icm_table *table, |
diff --git a/drivers/infiniband/hw/mthca/mthca_mr.c b/drivers/infiniband/hw/mthca/mthca_mr.c index f71ffa88db3a..7d08f2038aff 100644 --- a/drivers/infiniband/hw/mthca/mthca_mr.c +++ b/drivers/infiniband/hw/mthca/mthca_mr.c | |||
@@ -524,7 +524,7 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd, | |||
524 | if (err) | 524 | if (err) |
525 | goto err_out_mpt_free; | 525 | goto err_out_mpt_free; |
526 | 526 | ||
527 | mr->mem.arbel.mpt = mthca_table_find(dev->mr_table.mpt_table, key); | 527 | mr->mem.arbel.mpt = mthca_table_find(dev->mr_table.mpt_table, key, NULL); |
528 | BUG_ON(!mr->mem.arbel.mpt); | 528 | BUG_ON(!mr->mem.arbel.mpt); |
529 | } else | 529 | } else |
530 | mr->mem.tavor.mpt = dev->mr_table.tavor_fmr.mpt_base + | 530 | mr->mem.tavor.mpt = dev->mr_table.tavor_fmr.mpt_base + |
@@ -538,7 +538,8 @@ int mthca_fmr_alloc(struct mthca_dev *dev, u32 pd, | |||
538 | 538 | ||
539 | if (mthca_is_memfree(dev)) { | 539 | if (mthca_is_memfree(dev)) { |
540 | mr->mem.arbel.mtts = mthca_table_find(dev->mr_table.mtt_table, | 540 | mr->mem.arbel.mtts = mthca_table_find(dev->mr_table.mtt_table, |
541 | mr->mtt->first_seg); | 541 | mr->mtt->first_seg, |
542 | &mr->mem.arbel.dma_handle); | ||
542 | BUG_ON(!mr->mem.arbel.mtts); | 543 | BUG_ON(!mr->mem.arbel.mtts); |
543 | } else | 544 | } else |
544 | mr->mem.tavor.mtts = dev->mr_table.tavor_fmr.mtt_base + mtt_seg; | 545 | mr->mem.tavor.mtts = dev->mr_table.tavor_fmr.mtt_base + mtt_seg; |
@@ -712,6 +713,9 @@ int mthca_arbel_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list, | |||
712 | fmr->mem.arbel.mtts[i] = cpu_to_be64(page_list[i] | | 713 | fmr->mem.arbel.mtts[i] = cpu_to_be64(page_list[i] | |
713 | MTHCA_MTT_FLAG_PRESENT); | 714 | MTHCA_MTT_FLAG_PRESENT); |
714 | 715 | ||
716 | dma_sync_single(&dev->pdev->dev, fmr->mem.arbel.dma_handle, | ||
717 | list_len * sizeof(u64), DMA_TO_DEVICE); | ||
718 | |||
715 | fmr->mem.arbel.mpt->key = cpu_to_be32(key); | 719 | fmr->mem.arbel.mpt->key = cpu_to_be32(key); |
716 | fmr->mem.arbel.mpt->lkey = cpu_to_be32(key); | 720 | fmr->mem.arbel.mpt->lkey = cpu_to_be32(key); |
717 | fmr->mem.arbel.mpt->length = cpu_to_be64(list_len * (1ull << fmr->attr.page_shift)); | 721 | fmr->mem.arbel.mpt->length = cpu_to_be64(list_len * (1ull << fmr->attr.page_shift)); |
diff --git a/drivers/infiniband/hw/mthca/mthca_provider.h b/drivers/infiniband/hw/mthca/mthca_provider.h index 9a5bece3fa5c..1d266ac2e094 100644 --- a/drivers/infiniband/hw/mthca/mthca_provider.h +++ b/drivers/infiniband/hw/mthca/mthca_provider.h | |||
@@ -89,6 +89,7 @@ struct mthca_fmr { | |||
89 | struct { | 89 | struct { |
90 | struct mthca_mpt_entry *mpt; | 90 | struct mthca_mpt_entry *mpt; |
91 | __be64 *mtts; | 91 | __be64 *mtts; |
92 | dma_addr_t dma_handle; | ||
92 | } arbel; | 93 | } arbel; |
93 | } mem; | 94 | } mem; |
94 | }; | 95 | }; |