2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005, 2006, 2007, 2008 Mellanox Technologies. All rights reserved.
4 * Copyright (c) 2006, 2007 Cisco Systems, Inc. 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
35 #include <linux/errno.h>
36 #include <linux/export.h>
37 #include <linux/slab.h>
38 #include <linux/kernel.h>
39 #include <linux/vmalloc.h>
41 #include <linux/mlx4/cmd.h>
46 static u32 mlx4_buddy_alloc(struct mlx4_buddy *buddy, int order)
52 spin_lock(&buddy->lock);
54 for (o = order; o <= buddy->max_order; ++o)
55 if (buddy->num_free[o]) {
56 m = 1 << (buddy->max_order - o);
57 seg = find_first_bit(buddy->bits[o], m);
62 spin_unlock(&buddy->lock);
66 clear_bit(seg, buddy->bits[o]);
72 set_bit(seg ^ 1, buddy->bits[o]);
76 spin_unlock(&buddy->lock);
83 static void mlx4_buddy_free(struct mlx4_buddy *buddy, u32 seg, int order)
87 spin_lock(&buddy->lock);
89 while (test_bit(seg ^ 1, buddy->bits[order])) {
90 clear_bit(seg ^ 1, buddy->bits[order]);
91 --buddy->num_free[order];
96 set_bit(seg, buddy->bits[order]);
97 ++buddy->num_free[order];
99 spin_unlock(&buddy->lock);
102 static int mlx4_buddy_init(struct mlx4_buddy *buddy, int max_order)
106 buddy->max_order = max_order;
107 spin_lock_init(&buddy->lock);
109 buddy->bits = kcalloc(buddy->max_order + 1, sizeof(long *),
111 buddy->num_free = kcalloc(buddy->max_order + 1, sizeof(*buddy->num_free),
113 if (!buddy->bits || !buddy->num_free)
116 for (i = 0; i <= buddy->max_order; ++i) {
117 s = BITS_TO_LONGS(1 << (buddy->max_order - i));
118 buddy->bits[i] = kvmalloc_array(s, sizeof(long), GFP_KERNEL | __GFP_ZERO);
123 set_bit(0, buddy->bits[buddy->max_order]);
124 buddy->num_free[buddy->max_order] = 1;
129 for (i = 0; i <= buddy->max_order; ++i)
130 kvfree(buddy->bits[i]);
134 kfree(buddy->num_free);
139 static void mlx4_buddy_cleanup(struct mlx4_buddy *buddy)
143 for (i = 0; i <= buddy->max_order; ++i)
144 kvfree(buddy->bits[i]);
147 kfree(buddy->num_free);
150 u32 __mlx4_alloc_mtt_range(struct mlx4_dev *dev, int order)
152 struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
157 seg_order = max_t(int, order - log_mtts_per_seg, 0);
159 seg = mlx4_buddy_alloc(&mr_table->mtt_buddy, seg_order);
163 offset = seg * (1 << log_mtts_per_seg);
165 if (mlx4_table_get_range(dev, &mr_table->mtt_table, offset,
166 offset + (1 << order) - 1)) {
167 mlx4_buddy_free(&mr_table->mtt_buddy, seg, seg_order);
174 static u32 mlx4_alloc_mtt_range(struct mlx4_dev *dev, int order)
180 if (mlx4_is_mfunc(dev)) {
181 set_param_l(&in_param, order);
182 err = mlx4_cmd_imm(dev, in_param, &out_param, RES_MTT,
183 RES_OP_RESERVE_AND_MAP,
185 MLX4_CMD_TIME_CLASS_A,
189 return get_param_l(&out_param);
191 return __mlx4_alloc_mtt_range(dev, order);
194 int mlx4_mtt_init(struct mlx4_dev *dev, int npages, int page_shift,
195 struct mlx4_mtt *mtt)
201 mtt->page_shift = MLX4_ICM_PAGE_SHIFT;
204 mtt->page_shift = page_shift;
206 for (mtt->order = 0, i = 1; i < npages; i <<= 1)
209 mtt->offset = mlx4_alloc_mtt_range(dev, mtt->order);
210 if (mtt->offset == -1)
215 EXPORT_SYMBOL_GPL(mlx4_mtt_init);
217 void __mlx4_free_mtt_range(struct mlx4_dev *dev, u32 offset, int order)
221 struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
223 seg_order = max_t(int, order - log_mtts_per_seg, 0);
224 first_seg = offset / (1 << log_mtts_per_seg);
226 mlx4_buddy_free(&mr_table->mtt_buddy, first_seg, seg_order);
227 mlx4_table_put_range(dev, &mr_table->mtt_table, offset,
228 offset + (1 << order) - 1);
231 static void mlx4_free_mtt_range(struct mlx4_dev *dev, u32 offset, int order)
236 if (mlx4_is_mfunc(dev)) {
237 set_param_l(&in_param, offset);
238 set_param_h(&in_param, order);
239 err = mlx4_cmd(dev, in_param, RES_MTT, RES_OP_RESERVE_AND_MAP,
241 MLX4_CMD_TIME_CLASS_A,
244 mlx4_warn(dev, "Failed to free mtt range at:%d order:%d\n",
248 __mlx4_free_mtt_range(dev, offset, order);
251 void mlx4_mtt_cleanup(struct mlx4_dev *dev, struct mlx4_mtt *mtt)
256 mlx4_free_mtt_range(dev, mtt->offset, mtt->order);
258 EXPORT_SYMBOL_GPL(mlx4_mtt_cleanup);
260 u64 mlx4_mtt_addr(struct mlx4_dev *dev, struct mlx4_mtt *mtt)
262 return (u64) mtt->offset * dev->caps.mtt_entry_sz;
264 EXPORT_SYMBOL_GPL(mlx4_mtt_addr);
266 static u32 hw_index_to_key(u32 ind)
268 return (ind >> 24) | (ind << 8);
271 static u32 key_to_hw_index(u32 key)
273 return (key << 24) | (key >> 8);
276 static int mlx4_SW2HW_MPT(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
279 return mlx4_cmd(dev, mailbox->dma, mpt_index,
280 0, MLX4_CMD_SW2HW_MPT, MLX4_CMD_TIME_CLASS_B,
284 static int mlx4_HW2SW_MPT(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
287 return mlx4_cmd_box(dev, 0, mailbox ? mailbox->dma : 0, mpt_index,
288 !mailbox, MLX4_CMD_HW2SW_MPT,
289 MLX4_CMD_TIME_CLASS_B, MLX4_CMD_WRAPPED);
292 /* Must protect against concurrent access */
293 int mlx4_mr_hw_get_mpt(struct mlx4_dev *dev, struct mlx4_mr *mmr,
294 struct mlx4_mpt_entry ***mpt_entry)
297 int key = key_to_hw_index(mmr->key) & (dev->caps.num_mpts - 1);
298 struct mlx4_cmd_mailbox *mailbox = NULL;
300 if (mmr->enabled != MLX4_MPT_EN_HW)
303 err = mlx4_HW2SW_MPT(dev, NULL, key);
305 mlx4_warn(dev, "HW2SW_MPT failed (%d).", err);
306 mlx4_warn(dev, "Most likely the MR has MWs bound to it.\n");
310 mmr->enabled = MLX4_MPT_EN_SW;
312 if (!mlx4_is_mfunc(dev)) {
313 **mpt_entry = mlx4_table_find(
314 &mlx4_priv(dev)->mr_table.dmpt_table,
317 mailbox = mlx4_alloc_cmd_mailbox(dev);
319 return PTR_ERR(mailbox);
321 err = mlx4_cmd_box(dev, 0, mailbox->dma, key,
322 0, MLX4_CMD_QUERY_MPT,
323 MLX4_CMD_TIME_CLASS_B,
328 *mpt_entry = (struct mlx4_mpt_entry **)&mailbox->buf;
331 if (!(*mpt_entry) || !(**mpt_entry)) {
339 mlx4_free_cmd_mailbox(dev, mailbox);
342 EXPORT_SYMBOL_GPL(mlx4_mr_hw_get_mpt);
344 int mlx4_mr_hw_write_mpt(struct mlx4_dev *dev, struct mlx4_mr *mmr,
345 struct mlx4_mpt_entry **mpt_entry)
349 if (!mlx4_is_mfunc(dev)) {
350 /* Make sure any changes to this entry are flushed */
353 *(u8 *)(*mpt_entry) = MLX4_MPT_STATUS_HW;
355 /* Make sure the new status is written */
358 err = mlx4_SYNC_TPT(dev);
360 int key = key_to_hw_index(mmr->key) & (dev->caps.num_mpts - 1);
362 struct mlx4_cmd_mailbox *mailbox =
363 container_of((void *)mpt_entry, struct mlx4_cmd_mailbox,
366 err = mlx4_SW2HW_MPT(dev, mailbox, key);
370 mmr->pd = be32_to_cpu((*mpt_entry)->pd_flags) & MLX4_MPT_PD_MASK;
371 mmr->enabled = MLX4_MPT_EN_HW;
375 EXPORT_SYMBOL_GPL(mlx4_mr_hw_write_mpt);
377 void mlx4_mr_hw_put_mpt(struct mlx4_dev *dev,
378 struct mlx4_mpt_entry **mpt_entry)
380 if (mlx4_is_mfunc(dev)) {
381 struct mlx4_cmd_mailbox *mailbox =
382 container_of((void *)mpt_entry, struct mlx4_cmd_mailbox,
384 mlx4_free_cmd_mailbox(dev, mailbox);
387 EXPORT_SYMBOL_GPL(mlx4_mr_hw_put_mpt);
389 int mlx4_mr_hw_change_pd(struct mlx4_dev *dev, struct mlx4_mpt_entry *mpt_entry,
392 u32 pd_flags = be32_to_cpu(mpt_entry->pd_flags) & ~MLX4_MPT_PD_MASK;
393 /* The wrapper function will put the slave's id here */
394 if (mlx4_is_mfunc(dev))
395 pd_flags &= ~MLX4_MPT_PD_VF_MASK;
397 mpt_entry->pd_flags = cpu_to_be32(pd_flags |
398 (pdn & MLX4_MPT_PD_MASK)
399 | MLX4_MPT_PD_FLAG_EN_INV);
402 EXPORT_SYMBOL_GPL(mlx4_mr_hw_change_pd);
404 int mlx4_mr_hw_change_access(struct mlx4_dev *dev,
405 struct mlx4_mpt_entry *mpt_entry,
408 u32 flags = (be32_to_cpu(mpt_entry->flags) & ~MLX4_PERM_MASK) |
409 (access & MLX4_PERM_MASK);
411 mpt_entry->flags = cpu_to_be32(flags);
414 EXPORT_SYMBOL_GPL(mlx4_mr_hw_change_access);
416 static int mlx4_mr_alloc_reserved(struct mlx4_dev *dev, u32 mridx, u32 pd,
417 u64 iova, u64 size, u32 access, int npages,
418 int page_shift, struct mlx4_mr *mr)
424 mr->enabled = MLX4_MPT_DISABLED;
425 mr->key = hw_index_to_key(mridx);
427 return mlx4_mtt_init(dev, npages, page_shift, &mr->mtt);
430 static int mlx4_WRITE_MTT(struct mlx4_dev *dev,
431 struct mlx4_cmd_mailbox *mailbox,
434 return mlx4_cmd(dev, mailbox->dma, num_entries, 0, MLX4_CMD_WRITE_MTT,
435 MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
438 int __mlx4_mpt_reserve(struct mlx4_dev *dev)
440 struct mlx4_priv *priv = mlx4_priv(dev);
442 return mlx4_bitmap_alloc(&priv->mr_table.mpt_bitmap);
445 static int mlx4_mpt_reserve(struct mlx4_dev *dev)
449 if (mlx4_is_mfunc(dev)) {
450 if (mlx4_cmd_imm(dev, 0, &out_param, RES_MPT, RES_OP_RESERVE,
452 MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED))
454 return get_param_l(&out_param);
456 return __mlx4_mpt_reserve(dev);
459 void __mlx4_mpt_release(struct mlx4_dev *dev, u32 index)
461 struct mlx4_priv *priv = mlx4_priv(dev);
463 mlx4_bitmap_free(&priv->mr_table.mpt_bitmap, index, MLX4_NO_RR);
466 static void mlx4_mpt_release(struct mlx4_dev *dev, u32 index)
470 if (mlx4_is_mfunc(dev)) {
471 set_param_l(&in_param, index);
472 if (mlx4_cmd(dev, in_param, RES_MPT, RES_OP_RESERVE,
474 MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED))
475 mlx4_warn(dev, "Failed to release mr index:%d\n",
479 __mlx4_mpt_release(dev, index);
482 int __mlx4_mpt_alloc_icm(struct mlx4_dev *dev, u32 index)
484 struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
486 return mlx4_table_get(dev, &mr_table->dmpt_table, index);
489 static int mlx4_mpt_alloc_icm(struct mlx4_dev *dev, u32 index)
493 if (mlx4_is_mfunc(dev)) {
494 set_param_l(¶m, index);
495 return mlx4_cmd_imm(dev, param, ¶m, RES_MPT, RES_OP_MAP_ICM,
497 MLX4_CMD_TIME_CLASS_A,
500 return __mlx4_mpt_alloc_icm(dev, index);
503 void __mlx4_mpt_free_icm(struct mlx4_dev *dev, u32 index)
505 struct mlx4_mr_table *mr_table = &mlx4_priv(dev)->mr_table;
507 mlx4_table_put(dev, &mr_table->dmpt_table, index);
510 static void mlx4_mpt_free_icm(struct mlx4_dev *dev, u32 index)
514 if (mlx4_is_mfunc(dev)) {
515 set_param_l(&in_param, index);
516 if (mlx4_cmd(dev, in_param, RES_MPT, RES_OP_MAP_ICM,
517 MLX4_CMD_FREE_RES, MLX4_CMD_TIME_CLASS_A,
519 mlx4_warn(dev, "Failed to free icm of mr index:%d\n",
523 return __mlx4_mpt_free_icm(dev, index);
526 int mlx4_mr_alloc(struct mlx4_dev *dev, u32 pd, u64 iova, u64 size, u32 access,
527 int npages, int page_shift, struct mlx4_mr *mr)
532 index = mlx4_mpt_reserve(dev);
536 err = mlx4_mr_alloc_reserved(dev, index, pd, iova, size,
537 access, npages, page_shift, mr);
539 mlx4_mpt_release(dev, index);
543 EXPORT_SYMBOL_GPL(mlx4_mr_alloc);
545 static int mlx4_mr_free_reserved(struct mlx4_dev *dev, struct mlx4_mr *mr)
549 if (mr->enabled == MLX4_MPT_EN_HW) {
550 err = mlx4_HW2SW_MPT(dev, NULL,
551 key_to_hw_index(mr->key) &
552 (dev->caps.num_mpts - 1));
554 mlx4_warn(dev, "HW2SW_MPT failed (%d), MR has MWs bound to it\n",
559 mr->enabled = MLX4_MPT_EN_SW;
561 mlx4_mtt_cleanup(dev, &mr->mtt);
566 int mlx4_mr_free(struct mlx4_dev *dev, struct mlx4_mr *mr)
570 ret = mlx4_mr_free_reserved(dev, mr);
574 mlx4_mpt_free_icm(dev, key_to_hw_index(mr->key));
575 mlx4_mpt_release(dev, key_to_hw_index(mr->key));
579 EXPORT_SYMBOL_GPL(mlx4_mr_free);
581 void mlx4_mr_rereg_mem_cleanup(struct mlx4_dev *dev, struct mlx4_mr *mr)
583 mlx4_mtt_cleanup(dev, &mr->mtt);
586 EXPORT_SYMBOL_GPL(mlx4_mr_rereg_mem_cleanup);
588 int mlx4_mr_rereg_mem_write(struct mlx4_dev *dev, struct mlx4_mr *mr,
589 u64 iova, u64 size, int npages,
590 int page_shift, struct mlx4_mpt_entry *mpt_entry)
594 err = mlx4_mtt_init(dev, npages, page_shift, &mr->mtt);
598 mpt_entry->start = cpu_to_be64(iova);
599 mpt_entry->length = cpu_to_be64(size);
600 mpt_entry->entity_size = cpu_to_be32(page_shift);
601 mpt_entry->flags &= ~(cpu_to_be32(MLX4_MPT_FLAG_FREE |
602 MLX4_MPT_FLAG_SW_OWNS));
603 if (mr->mtt.order < 0) {
604 mpt_entry->flags |= cpu_to_be32(MLX4_MPT_FLAG_PHYSICAL);
605 mpt_entry->mtt_addr = 0;
607 mpt_entry->mtt_addr = cpu_to_be64(mlx4_mtt_addr(dev,
609 if (mr->mtt.page_shift == 0)
610 mpt_entry->mtt_sz = cpu_to_be32(1 << mr->mtt.order);
612 if (mr->mtt.order >= 0 && mr->mtt.page_shift == 0) {
613 /* fast register MR in free state */
614 mpt_entry->flags |= cpu_to_be32(MLX4_MPT_FLAG_FREE);
615 mpt_entry->pd_flags |= cpu_to_be32(MLX4_MPT_PD_FLAG_FAST_REG |
616 MLX4_MPT_PD_FLAG_RAE);
618 mpt_entry->flags |= cpu_to_be32(MLX4_MPT_FLAG_SW_OWNS);
620 mr->enabled = MLX4_MPT_EN_SW;
624 EXPORT_SYMBOL_GPL(mlx4_mr_rereg_mem_write);
626 int mlx4_mr_enable(struct mlx4_dev *dev, struct mlx4_mr *mr)
628 struct mlx4_cmd_mailbox *mailbox;
629 struct mlx4_mpt_entry *mpt_entry;
632 err = mlx4_mpt_alloc_icm(dev, key_to_hw_index(mr->key));
636 mailbox = mlx4_alloc_cmd_mailbox(dev);
637 if (IS_ERR(mailbox)) {
638 err = PTR_ERR(mailbox);
641 mpt_entry = mailbox->buf;
642 mpt_entry->flags = cpu_to_be32(MLX4_MPT_FLAG_MIO |
643 MLX4_MPT_FLAG_REGION |
646 mpt_entry->key = cpu_to_be32(key_to_hw_index(mr->key));
647 mpt_entry->pd_flags = cpu_to_be32(mr->pd | MLX4_MPT_PD_FLAG_EN_INV);
648 mpt_entry->start = cpu_to_be64(mr->iova);
649 mpt_entry->length = cpu_to_be64(mr->size);
650 mpt_entry->entity_size = cpu_to_be32(mr->mtt.page_shift);
652 if (mr->mtt.order < 0) {
653 mpt_entry->flags |= cpu_to_be32(MLX4_MPT_FLAG_PHYSICAL);
654 mpt_entry->mtt_addr = 0;
656 mpt_entry->mtt_addr = cpu_to_be64(mlx4_mtt_addr(dev,
660 if (mr->mtt.order >= 0 && mr->mtt.page_shift == 0) {
661 /* fast register MR in free state */
662 mpt_entry->flags |= cpu_to_be32(MLX4_MPT_FLAG_FREE);
663 mpt_entry->pd_flags |= cpu_to_be32(MLX4_MPT_PD_FLAG_FAST_REG |
664 MLX4_MPT_PD_FLAG_RAE);
665 mpt_entry->mtt_sz = cpu_to_be32(1 << mr->mtt.order);
667 mpt_entry->flags |= cpu_to_be32(MLX4_MPT_FLAG_SW_OWNS);
670 err = mlx4_SW2HW_MPT(dev, mailbox,
671 key_to_hw_index(mr->key) & (dev->caps.num_mpts - 1));
673 mlx4_warn(dev, "SW2HW_MPT failed (%d)\n", err);
676 mr->enabled = MLX4_MPT_EN_HW;
678 mlx4_free_cmd_mailbox(dev, mailbox);
683 mlx4_free_cmd_mailbox(dev, mailbox);
686 mlx4_mpt_free_icm(dev, key_to_hw_index(mr->key));
689 EXPORT_SYMBOL_GPL(mlx4_mr_enable);
691 static int mlx4_write_mtt_chunk(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
692 int start_index, int npages, u64 *page_list)
694 struct mlx4_priv *priv = mlx4_priv(dev);
696 dma_addr_t dma_handle;
699 mtts = mlx4_table_find(&priv->mr_table.mtt_table, mtt->offset +
700 start_index, &dma_handle);
705 dma_sync_single_for_cpu(&dev->persist->pdev->dev, dma_handle,
706 npages * sizeof(u64), DMA_TO_DEVICE);
708 for (i = 0; i < npages; ++i)
709 mtts[i] = cpu_to_be64(page_list[i] | MLX4_MTT_FLAG_PRESENT);
711 dma_sync_single_for_device(&dev->persist->pdev->dev, dma_handle,
712 npages * sizeof(u64), DMA_TO_DEVICE);
717 int __mlx4_write_mtt(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
718 int start_index, int npages, u64 *page_list)
723 int max_mtts_first_page;
725 /* compute how may mtts fit in the first page */
726 mtts_per_page = PAGE_SIZE / sizeof(u64);
727 max_mtts_first_page = mtts_per_page - (mtt->offset + start_index)
730 chunk = min_t(int, max_mtts_first_page, npages);
733 err = mlx4_write_mtt_chunk(dev, mtt, start_index, chunk, page_list);
737 start_index += chunk;
740 chunk = min_t(int, mtts_per_page, npages);
745 int mlx4_write_mtt(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
746 int start_index, int npages, u64 *page_list)
748 struct mlx4_cmd_mailbox *mailbox = NULL;
749 __be64 *inbox = NULL;
757 if (mlx4_is_mfunc(dev)) {
758 mailbox = mlx4_alloc_cmd_mailbox(dev);
760 return PTR_ERR(mailbox);
761 inbox = mailbox->buf;
764 chunk = min_t(int, MLX4_MAILBOX_SIZE / sizeof(u64) - 2,
766 inbox[0] = cpu_to_be64(mtt->offset + start_index);
768 for (i = 0; i < chunk; ++i)
769 inbox[i + 2] = cpu_to_be64(page_list[i] |
770 MLX4_MTT_FLAG_PRESENT);
771 err = mlx4_WRITE_MTT(dev, mailbox, chunk);
773 mlx4_free_cmd_mailbox(dev, mailbox);
778 start_index += chunk;
781 mlx4_free_cmd_mailbox(dev, mailbox);
785 return __mlx4_write_mtt(dev, mtt, start_index, npages, page_list);
787 EXPORT_SYMBOL_GPL(mlx4_write_mtt);
789 int mlx4_buf_write_mtt(struct mlx4_dev *dev, struct mlx4_mtt *mtt,
790 struct mlx4_buf *buf)
796 page_list = kcalloc(buf->npages, sizeof(*page_list), GFP_KERNEL);
800 for (i = 0; i < buf->npages; ++i)
802 page_list[i] = buf->direct.map + (i << buf->page_shift);
804 page_list[i] = buf->page_list[i].map;
806 err = mlx4_write_mtt(dev, mtt, 0, buf->npages, page_list);
811 EXPORT_SYMBOL_GPL(mlx4_buf_write_mtt);
813 int mlx4_mw_alloc(struct mlx4_dev *dev, u32 pd, enum mlx4_mw_type type,
818 if ((type == MLX4_MW_TYPE_1 &&
819 !(dev->caps.flags & MLX4_DEV_CAP_FLAG_MEM_WINDOW)) ||
820 (type == MLX4_MW_TYPE_2 &&
821 !(dev->caps.bmme_flags & MLX4_BMME_FLAG_TYPE_2_WIN)))
824 index = mlx4_mpt_reserve(dev);
828 mw->key = hw_index_to_key(index);
831 mw->enabled = MLX4_MPT_DISABLED;
835 EXPORT_SYMBOL_GPL(mlx4_mw_alloc);
837 int mlx4_mw_enable(struct mlx4_dev *dev, struct mlx4_mw *mw)
839 struct mlx4_cmd_mailbox *mailbox;
840 struct mlx4_mpt_entry *mpt_entry;
843 err = mlx4_mpt_alloc_icm(dev, key_to_hw_index(mw->key));
847 mailbox = mlx4_alloc_cmd_mailbox(dev);
848 if (IS_ERR(mailbox)) {
849 err = PTR_ERR(mailbox);
852 mpt_entry = mailbox->buf;
854 /* Note that the MLX4_MPT_FLAG_REGION bit in mpt_entry->flags is turned
855 * off, thus creating a memory window and not a memory region.
857 mpt_entry->key = cpu_to_be32(key_to_hw_index(mw->key));
858 mpt_entry->pd_flags = cpu_to_be32(mw->pd);
859 if (mw->type == MLX4_MW_TYPE_2) {
860 mpt_entry->flags |= cpu_to_be32(MLX4_MPT_FLAG_FREE);
861 mpt_entry->qpn = cpu_to_be32(MLX4_MPT_QP_FLAG_BOUND_QP);
862 mpt_entry->pd_flags |= cpu_to_be32(MLX4_MPT_PD_FLAG_EN_INV);
865 err = mlx4_SW2HW_MPT(dev, mailbox,
866 key_to_hw_index(mw->key) &
867 (dev->caps.num_mpts - 1));
869 mlx4_warn(dev, "SW2HW_MPT failed (%d)\n", err);
872 mw->enabled = MLX4_MPT_EN_HW;
874 mlx4_free_cmd_mailbox(dev, mailbox);
879 mlx4_free_cmd_mailbox(dev, mailbox);
882 mlx4_mpt_free_icm(dev, key_to_hw_index(mw->key));
885 EXPORT_SYMBOL_GPL(mlx4_mw_enable);
887 void mlx4_mw_free(struct mlx4_dev *dev, struct mlx4_mw *mw)
891 if (mw->enabled == MLX4_MPT_EN_HW) {
892 err = mlx4_HW2SW_MPT(dev, NULL,
893 key_to_hw_index(mw->key) &
894 (dev->caps.num_mpts - 1));
896 mlx4_warn(dev, "xxx HW2SW_MPT failed (%d)\n", err);
898 mw->enabled = MLX4_MPT_EN_SW;
901 mlx4_mpt_free_icm(dev, key_to_hw_index(mw->key));
902 mlx4_mpt_release(dev, key_to_hw_index(mw->key));
904 EXPORT_SYMBOL_GPL(mlx4_mw_free);
906 int mlx4_init_mr_table(struct mlx4_dev *dev)
908 struct mlx4_priv *priv = mlx4_priv(dev);
909 struct mlx4_mr_table *mr_table = &priv->mr_table;
912 /* Nothing to do for slaves - all MR handling is forwarded
914 if (mlx4_is_slave(dev))
917 if (!is_power_of_2(dev->caps.num_mpts))
920 err = mlx4_bitmap_init(&mr_table->mpt_bitmap, dev->caps.num_mpts,
921 ~0, dev->caps.reserved_mrws, 0);
925 err = mlx4_buddy_init(&mr_table->mtt_buddy,
926 ilog2((u32)dev->caps.num_mtts /
927 (1 << log_mtts_per_seg)));
931 if (dev->caps.reserved_mtts) {
932 priv->reserved_mtts =
933 mlx4_alloc_mtt_range(dev,
934 fls(dev->caps.reserved_mtts - 1));
935 if (priv->reserved_mtts < 0) {
936 mlx4_warn(dev, "MTT table of order %u is too small\n",
937 mr_table->mtt_buddy.max_order);
939 goto err_reserve_mtts;
946 mlx4_buddy_cleanup(&mr_table->mtt_buddy);
949 mlx4_bitmap_cleanup(&mr_table->mpt_bitmap);
954 void mlx4_cleanup_mr_table(struct mlx4_dev *dev)
956 struct mlx4_priv *priv = mlx4_priv(dev);
957 struct mlx4_mr_table *mr_table = &priv->mr_table;
959 if (mlx4_is_slave(dev))
961 if (priv->reserved_mtts >= 0)
962 mlx4_free_mtt_range(dev, priv->reserved_mtts,
963 fls(dev->caps.reserved_mtts - 1));
964 mlx4_buddy_cleanup(&mr_table->mtt_buddy);
965 mlx4_bitmap_cleanup(&mr_table->mpt_bitmap);
968 static inline int mlx4_check_fmr(struct mlx4_fmr *fmr, u64 *page_list,
969 int npages, u64 iova)
973 if (npages > fmr->max_pages)
976 page_mask = (1 << fmr->page_shift) - 1;
978 /* We are getting page lists, so va must be page aligned. */
979 if (iova & page_mask)
982 /* Trust the user not to pass misaligned data in page_list */
984 for (i = 0; i < npages; ++i) {
985 if (page_list[i] & ~page_mask)
989 if (fmr->maps >= fmr->max_maps)
995 int mlx4_map_phys_fmr(struct mlx4_dev *dev, struct mlx4_fmr *fmr, u64 *page_list,
996 int npages, u64 iova, u32 *lkey, u32 *rkey)
1001 err = mlx4_check_fmr(fmr, page_list, npages, iova);
1007 key = key_to_hw_index(fmr->mr.key);
1008 key += dev->caps.num_mpts;
1009 *lkey = *rkey = fmr->mr.key = hw_index_to_key(key);
1011 *(u8 *) fmr->mpt = MLX4_MPT_STATUS_SW;
1013 /* Make sure MPT status is visible before writing MTT entries */
1016 dma_sync_single_for_cpu(&dev->persist->pdev->dev, fmr->dma_handle,
1017 npages * sizeof(u64), DMA_TO_DEVICE);
1019 for (i = 0; i < npages; ++i)
1020 fmr->mtts[i] = cpu_to_be64(page_list[i] | MLX4_MTT_FLAG_PRESENT);
1022 dma_sync_single_for_device(&dev->persist->pdev->dev, fmr->dma_handle,
1023 npages * sizeof(u64), DMA_TO_DEVICE);
1025 fmr->mpt->key = cpu_to_be32(key);
1026 fmr->mpt->lkey = cpu_to_be32(key);
1027 fmr->mpt->length = cpu_to_be64(npages * (1ull << fmr->page_shift));
1028 fmr->mpt->start = cpu_to_be64(iova);
1030 /* Make MTT entries are visible before setting MPT status */
1033 *(u8 *) fmr->mpt = MLX4_MPT_STATUS_HW;
1035 /* Make sure MPT status is visible before consumer can use FMR */
1040 EXPORT_SYMBOL_GPL(mlx4_map_phys_fmr);
1042 int mlx4_fmr_alloc(struct mlx4_dev *dev, u32 pd, u32 access, int max_pages,
1043 int max_maps, u8 page_shift, struct mlx4_fmr *fmr)
1045 struct mlx4_priv *priv = mlx4_priv(dev);
1048 if (max_maps > dev->caps.max_fmr_maps)
1051 if (page_shift < (ffs(dev->caps.page_size_cap) - 1) || page_shift >= 32)
1054 /* All MTTs must fit in the same page */
1055 if (max_pages * sizeof(*fmr->mtts) > PAGE_SIZE)
1058 fmr->page_shift = page_shift;
1059 fmr->max_pages = max_pages;
1060 fmr->max_maps = max_maps;
1063 err = mlx4_mr_alloc(dev, pd, 0, 0, access, max_pages,
1064 page_shift, &fmr->mr);
1068 fmr->mtts = mlx4_table_find(&priv->mr_table.mtt_table,
1080 (void) mlx4_mr_free(dev, &fmr->mr);
1083 EXPORT_SYMBOL_GPL(mlx4_fmr_alloc);
1085 int mlx4_fmr_enable(struct mlx4_dev *dev, struct mlx4_fmr *fmr)
1087 struct mlx4_priv *priv = mlx4_priv(dev);
1090 err = mlx4_mr_enable(dev, &fmr->mr);
1094 fmr->mpt = mlx4_table_find(&priv->mr_table.dmpt_table,
1095 key_to_hw_index(fmr->mr.key), NULL);
1101 EXPORT_SYMBOL_GPL(mlx4_fmr_enable);
1103 void mlx4_fmr_unmap(struct mlx4_dev *dev, struct mlx4_fmr *fmr,
1104 u32 *lkey, u32 *rkey)
1109 /* To unmap: it is sufficient to take back ownership from HW */
1110 *(u8 *)fmr->mpt = MLX4_MPT_STATUS_SW;
1112 /* Make sure MPT status is visible */
1117 EXPORT_SYMBOL_GPL(mlx4_fmr_unmap);
1119 int mlx4_fmr_free(struct mlx4_dev *dev, struct mlx4_fmr *fmr)
1125 if (fmr->mr.enabled == MLX4_MPT_EN_HW) {
1126 /* In case of FMR was enabled and unmapped
1127 * make sure to give ownership of MPT back to HW
1128 * so HW2SW_MPT command will success.
1130 *(u8 *)fmr->mpt = MLX4_MPT_STATUS_SW;
1131 /* Make sure MPT status is visible before changing MPT fields */
1133 fmr->mpt->length = 0;
1134 fmr->mpt->start = 0;
1135 /* Make sure MPT data is visible after changing MPT status */
1137 *(u8 *)fmr->mpt = MLX4_MPT_STATUS_HW;
1138 /* make sure MPT status is visible */
1142 ret = mlx4_mr_free(dev, &fmr->mr);
1145 fmr->mr.enabled = MLX4_MPT_DISABLED;
1149 EXPORT_SYMBOL_GPL(mlx4_fmr_free);
1151 int mlx4_SYNC_TPT(struct mlx4_dev *dev)
1153 return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_SYNC_TPT,
1154 MLX4_CMD_TIME_CLASS_A, MLX4_CMD_NATIVE);
1156 EXPORT_SYMBOL_GPL(mlx4_SYNC_TPT);