xref: /freebsd/sys/dev/mlx5/mlx5_ib/mlx5_ib_mem.c (revision 77013d11e6483b970af25e13c9b892075742f7e5)
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 <linux/module.h>
29 #include <rdma/ib_umem.h>
30 #include <rdma/ib_umem_odp.h>
31 #include "mlx5_ib.h"
32 
33 /* @umem: umem object to scan
34  * @addr: ib virtual address requested by the user
35  * @max_page_shift: high limit for page_shift - 0 means no limit
36  * @count: number of PAGE_SIZE pages covered by umem
37  * @shift: page shift for the compound pages found in the region
38  * @ncont: number of compund pages
39  * @order: log2 of the number of compound pages
40  */
41 void mlx5_ib_cont_pages(struct ib_umem *umem, u64 addr,
42 			unsigned long max_page_shift,
43 			int *count, int *shift,
44 			int *ncont, int *order)
45 {
46 	unsigned long tmp;
47 	unsigned long m;
48 	u64 base = ~0, p = 0;
49 	u64 len, pfn;
50 	int i = 0;
51 	struct scatterlist *sg;
52 	int entry;
53 
54 	addr = addr >> PAGE_SHIFT;
55 	tmp = (unsigned long)addr;
56 	m = find_first_bit(&tmp, BITS_PER_LONG);
57 	if (max_page_shift)
58 		m = min_t(unsigned long, max_page_shift - PAGE_SHIFT, m);
59 
60 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
61 		len = sg_dma_len(sg) >> PAGE_SHIFT;
62 		pfn = sg_dma_address(sg) >> PAGE_SHIFT;
63 		if (base + p != pfn) {
64 			/* If either the offset or the new
65 			 * base are unaligned update m
66 			 */
67 			tmp = (unsigned long)(pfn | p);
68 			if (!IS_ALIGNED(tmp, 1 << m))
69 				m = find_first_bit(&tmp, BITS_PER_LONG);
70 
71 			base = pfn;
72 			p = 0;
73 		}
74 
75 		p += len;
76 		i += len;
77 	}
78 
79 	if (i) {
80 		m = min_t(unsigned long, ilog2(roundup_pow_of_two(i)), m);
81 
82 		if (order)
83 			*order = ilog2(roundup_pow_of_two(i) >> m);
84 
85 		*ncont = DIV_ROUND_UP(i, (1 << m));
86 	} else {
87 		m  = 0;
88 
89 		if (order)
90 			*order = 0;
91 
92 		*ncont = 0;
93 	}
94 	*shift = PAGE_SHIFT + m;
95 	*count = i;
96 }
97 
98 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
99 static u64 umem_dma_to_mtt(dma_addr_t umem_dma)
100 {
101 	u64 mtt_entry = umem_dma & ODP_DMA_ADDR_MASK;
102 
103 	if (umem_dma & ODP_READ_ALLOWED_BIT)
104 		mtt_entry |= MLX5_IB_MTT_READ;
105 	if (umem_dma & ODP_WRITE_ALLOWED_BIT)
106 		mtt_entry |= MLX5_IB_MTT_WRITE;
107 
108 	return mtt_entry;
109 }
110 #endif
111 
112 /*
113  * Populate the given array with bus addresses from the umem.
114  *
115  * dev - mlx5_ib device
116  * umem - umem to use to fill the pages
117  * page_shift - determines the page size used in the resulting array
118  * offset - offset into the umem to start from,
119  *          only implemented for ODP umems
120  * num_pages - total number of pages to fill
121  * pas - bus addresses array to fill
122  * access_flags - access flags to set on all present pages.
123 		  use enum mlx5_ib_mtt_access_flags for this.
124  */
125 void __mlx5_ib_populate_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem,
126 			    int page_shift, size_t offset, size_t num_pages,
127 			    __be64 *pas, int access_flags)
128 {
129 	unsigned long umem_page_shift = ilog2(umem->page_size);
130 	int shift = page_shift - umem_page_shift;
131 	int mask = (1 << shift) - 1;
132 	int i, k;
133 	u64 cur = 0;
134 	u64 base;
135 	int len;
136 	struct scatterlist *sg;
137 	int entry;
138 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
139 	const bool odp = umem->odp_data != NULL;
140 
141 	if (odp) {
142 		WARN_ON(shift != 0);
143 		WARN_ON(access_flags != (MLX5_IB_MTT_READ | MLX5_IB_MTT_WRITE));
144 
145 		for (i = 0; i < num_pages; ++i) {
146 			dma_addr_t pa = umem->odp_data->dma_list[offset + i];
147 
148 			pas[i] = cpu_to_be64(umem_dma_to_mtt(pa));
149 		}
150 		return;
151 	}
152 #endif
153 
154 	i = 0;
155 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
156 		len = sg_dma_len(sg) >> umem_page_shift;
157 		base = sg_dma_address(sg);
158 		for (k = 0; k < len; k++) {
159 			if (!(i & mask)) {
160 				cur = base + (k << umem_page_shift);
161 				cur |= access_flags;
162 
163 				pas[i >> shift] = cpu_to_be64(cur);
164 				mlx5_ib_dbg(dev, "pas[%d] 0x%llx\n",
165 					    i >> shift, (long long)be64_to_cpu(pas[i >> shift]));
166 			}  else
167 				mlx5_ib_dbg(dev, "=====> 0x%llx\n",
168 					    (long long)(base + (k << umem_page_shift)));
169 			i++;
170 		}
171 	}
172 }
173 
174 void mlx5_ib_populate_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem,
175 			  int page_shift, __be64 *pas, int access_flags)
176 {
177 	return __mlx5_ib_populate_pas(dev, umem, page_shift, 0,
178 				      ib_umem_num_pages(umem), pas,
179 				      access_flags);
180 }
181 int mlx5_ib_get_buf_offset(u64 addr, int page_shift, u32 *offset)
182 {
183 	u64 page_size;
184 	u64 page_mask;
185 	u64 off_size;
186 	u64 off_mask;
187 	u64 buf_off;
188 
189 	page_size = (u64)1 << page_shift;
190 	page_mask = page_size - 1;
191 	buf_off = addr & page_mask;
192 	off_size = page_size >> 6;
193 	off_mask = off_size - 1;
194 
195 	if (buf_off & off_mask)
196 		return -EINVAL;
197 
198 	*offset = buf_off >> ilog2(off_size);
199 	return 0;
200 }
201