xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_mem.c (revision ebacd8013fe5f7fdf9f6a5b286f6680dd2891036)
1 /*-
2  * Copyright (c) 2013-2020, Mellanox Technologies, Ltd.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  * $FreeBSD$
26  */
27 
28 #include "opt_rss.h"
29 #include "opt_ratelimit.h"
30 
31 #include <linux/module.h>
32 #include <rdma/ib_umem.h>
33 #include <rdma/ib_umem_odp.h>
34 #include <dev/mlx5/mlx5_ib/mlx5_ib.h>
35 
36 /* @umem: umem object to scan
37  * @addr: ib virtual address requested by the user
38  * @max_page_shift: high limit for page_shift - 0 means no limit
39  * @count: number of PAGE_SIZE pages covered by umem
40  * @shift: page shift for the compound pages found in the region
41  * @ncont: number of compund pages
42  * @order: log2 of the number of compound pages
43  */
44 void mlx5_ib_cont_pages(struct ib_umem *umem, u64 addr,
45 			unsigned long max_page_shift,
46 			int *count, int *shift,
47 			int *ncont, int *order)
48 {
49 	unsigned long tmp;
50 	unsigned long m;
51 	u64 base = ~0, p = 0;
52 	u64 len, pfn;
53 	int i = 0;
54 	struct scatterlist *sg;
55 	int entry;
56 
57 	addr = addr >> PAGE_SHIFT;
58 	tmp = (unsigned long)addr;
59 	m = find_first_bit(&tmp, BITS_PER_LONG);
60 	if (max_page_shift)
61 		m = min_t(unsigned long, max_page_shift - PAGE_SHIFT, m);
62 
63 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
64 		len = sg_dma_len(sg) >> PAGE_SHIFT;
65 		pfn = sg_dma_address(sg) >> PAGE_SHIFT;
66 		if (base + p != pfn) {
67 			/* If either the offset or the new
68 			 * base are unaligned update m
69 			 */
70 			tmp = (unsigned long)(pfn | p);
71 			if (!IS_ALIGNED(tmp, 1 << m))
72 				m = find_first_bit(&tmp, BITS_PER_LONG);
73 
74 			base = pfn;
75 			p = 0;
76 		}
77 
78 		p += len;
79 		i += len;
80 	}
81 
82 	if (i) {
83 		m = min_t(unsigned long, ilog2(roundup_pow_of_two(i)), m);
84 
85 		if (order)
86 			*order = ilog2(roundup_pow_of_two(i) >> m);
87 
88 		*ncont = DIV_ROUND_UP(i, (1 << m));
89 	} else {
90 		m  = 0;
91 
92 		if (order)
93 			*order = 0;
94 
95 		*ncont = 0;
96 	}
97 	*shift = PAGE_SHIFT + m;
98 	*count = i;
99 }
100 
101 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
102 static u64 umem_dma_to_mtt(dma_addr_t umem_dma)
103 {
104 	u64 mtt_entry = umem_dma & ODP_DMA_ADDR_MASK;
105 
106 	if (umem_dma & ODP_READ_ALLOWED_BIT)
107 		mtt_entry |= MLX5_IB_MTT_READ;
108 	if (umem_dma & ODP_WRITE_ALLOWED_BIT)
109 		mtt_entry |= MLX5_IB_MTT_WRITE;
110 
111 	return mtt_entry;
112 }
113 #endif
114 
115 /*
116  * Populate the given array with bus addresses from the umem.
117  *
118  * dev - mlx5_ib device
119  * umem - umem to use to fill the pages
120  * page_shift - determines the page size used in the resulting array
121  * offset - offset into the umem to start from,
122  *          only implemented for ODP umems
123  * num_pages - total number of pages to fill
124  * pas - bus addresses array to fill
125  * access_flags - access flags to set on all present pages.
126 		  use enum mlx5_ib_mtt_access_flags for this.
127  */
128 void __mlx5_ib_populate_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem,
129 			    int page_shift, size_t offset, size_t num_pages,
130 			    __be64 *pas, int access_flags)
131 {
132 	unsigned long umem_page_shift = ilog2(umem->page_size);
133 	int shift = page_shift - umem_page_shift;
134 	int mask = (1 << shift) - 1;
135 	int i, k;
136 	u64 cur = 0;
137 	u64 base;
138 	int len;
139 	struct scatterlist *sg;
140 	int entry;
141 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
142 	const bool odp = umem->odp_data != NULL;
143 
144 	if (odp) {
145 		WARN_ON(shift != 0);
146 		WARN_ON(access_flags != (MLX5_IB_MTT_READ | MLX5_IB_MTT_WRITE));
147 
148 		for (i = 0; i < num_pages; ++i) {
149 			dma_addr_t pa = umem->odp_data->dma_list[offset + i];
150 
151 			pas[i] = cpu_to_be64(umem_dma_to_mtt(pa));
152 		}
153 		return;
154 	}
155 #endif
156 
157 	i = 0;
158 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
159 		len = sg_dma_len(sg) >> umem_page_shift;
160 		base = sg_dma_address(sg);
161 		for (k = 0; k < len; k++) {
162 			if (!(i & mask)) {
163 				cur = base + (k << umem_page_shift);
164 				cur |= access_flags;
165 
166 				pas[i >> shift] = cpu_to_be64(cur);
167 				mlx5_ib_dbg(dev, "pas[%d] 0x%llx\n",
168 					    i >> shift, (long long)be64_to_cpu(pas[i >> shift]));
169 			}  else
170 				mlx5_ib_dbg(dev, "=====> 0x%llx\n",
171 					    (long long)(base + (k << umem_page_shift)));
172 			i++;
173 		}
174 	}
175 }
176 
177 void mlx5_ib_populate_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem,
178 			  int page_shift, __be64 *pas, int access_flags)
179 {
180 	return __mlx5_ib_populate_pas(dev, umem, page_shift, 0,
181 				      ib_umem_num_pages(umem), pas,
182 				      access_flags);
183 }
184 int mlx5_ib_get_buf_offset(u64 addr, int page_shift, u32 *offset)
185 {
186 	u64 page_size;
187 	u64 page_mask;
188 	u64 off_size;
189 	u64 off_mask;
190 	u64 buf_off;
191 
192 	page_size = (u64)1 << page_shift;
193 	page_mask = page_size - 1;
194 	buf_off = addr & page_mask;
195 	off_size = page_size >> 6;
196 	off_mask = off_size - 1;
197 
198 	if (buf_off & off_mask)
199 		return -EINVAL;
200 
201 	*offset = buf_off >> ilog2(off_size);
202 	return 0;
203 }
204