1 /*
2  * Copyright (c) 2016 Hisilicon Limited.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/platform_device.h>
35 #include <linux/vmalloc.h>
36 #include "hns_roce_device.h"
37 #include <rdma/ib_umem.h>
38 
39 void hns_roce_buf_free(struct hns_roce_dev *hr_dev, struct hns_roce_buf *buf)
40 {
41 	struct hns_roce_buf_list *trunks;
42 	u32 i;
43 
44 	if (!buf)
45 		return;
46 
47 	trunks = buf->trunk_list;
48 	if (trunks) {
49 		buf->trunk_list = NULL;
50 		for (i = 0; i < buf->ntrunks; i++)
51 			dma_free_coherent(hr_dev->dev, 1 << buf->trunk_shift,
52 					  trunks[i].buf, trunks[i].map);
53 
54 		kfree(trunks);
55 	}
56 
57 	kfree(buf);
58 }
59 
60 /*
61  * Allocate the dma buffer for storing ROCEE table entries
62  *
63  * @size: required size
64  * @page_shift: the unit size in a continuous dma address range
65  * @flags: HNS_ROCE_BUF_ flags to control the allocation flow.
66  */
67 struct hns_roce_buf *hns_roce_buf_alloc(struct hns_roce_dev *hr_dev, u32 size,
68 					u32 page_shift, u32 flags)
69 {
70 	u32 trunk_size, page_size, alloced_size;
71 	struct hns_roce_buf_list *trunks;
72 	struct hns_roce_buf *buf;
73 	gfp_t gfp_flags;
74 	u32 ntrunk, i;
75 
76 	/* The minimum shift of the page accessed by hw is HNS_HW_PAGE_SHIFT */
77 	if (WARN_ON(page_shift < HNS_HW_PAGE_SHIFT))
78 		return ERR_PTR(-EINVAL);
79 
80 	gfp_flags = (flags & HNS_ROCE_BUF_NOSLEEP) ? GFP_ATOMIC : GFP_KERNEL;
81 	buf = kzalloc(sizeof(*buf), gfp_flags);
82 	if (!buf)
83 		return ERR_PTR(-ENOMEM);
84 
85 	buf->page_shift = page_shift;
86 	page_size = 1 << buf->page_shift;
87 
88 	/* Calc the trunk size and num by required size and page_shift */
89 	if (flags & HNS_ROCE_BUF_DIRECT) {
90 		buf->trunk_shift = order_base_2(ALIGN(size, PAGE_SIZE));
91 		ntrunk = 1;
92 	} else {
93 		buf->trunk_shift = order_base_2(ALIGN(page_size, PAGE_SIZE));
94 		ntrunk = DIV_ROUND_UP(size, 1 << buf->trunk_shift);
95 	}
96 
97 	trunks = kcalloc(ntrunk, sizeof(*trunks), gfp_flags);
98 	if (!trunks) {
99 		kfree(buf);
100 		return ERR_PTR(-ENOMEM);
101 	}
102 
103 	trunk_size = 1 << buf->trunk_shift;
104 	alloced_size = 0;
105 	for (i = 0; i < ntrunk; i++) {
106 		trunks[i].buf = dma_alloc_coherent(hr_dev->dev, trunk_size,
107 						   &trunks[i].map, gfp_flags);
108 		if (!trunks[i].buf)
109 			break;
110 
111 		alloced_size += trunk_size;
112 	}
113 
114 	buf->ntrunks = i;
115 
116 	/* In nofail mode, it's only failed when the alloced size is 0 */
117 	if ((flags & HNS_ROCE_BUF_NOFAIL) ? i == 0 : i != ntrunk) {
118 		for (i = 0; i < buf->ntrunks; i++)
119 			dma_free_coherent(hr_dev->dev, trunk_size,
120 					  trunks[i].buf, trunks[i].map);
121 
122 		kfree(trunks);
123 		kfree(buf);
124 		return ERR_PTR(-ENOMEM);
125 	}
126 
127 	buf->npages = DIV_ROUND_UP(alloced_size, page_size);
128 	buf->trunk_list = trunks;
129 
130 	return buf;
131 }
132 
133 int hns_roce_get_kmem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
134 			   int buf_cnt, struct hns_roce_buf *buf,
135 			   unsigned int page_shift)
136 {
137 	unsigned int offset, max_size;
138 	int total = 0;
139 	int i;
140 
141 	if (page_shift > buf->trunk_shift) {
142 		dev_err(hr_dev->dev, "failed to check kmem buf shift %u > %u\n",
143 			page_shift, buf->trunk_shift);
144 		return -EINVAL;
145 	}
146 
147 	offset = 0;
148 	max_size = buf->ntrunks << buf->trunk_shift;
149 	for (i = 0; i < buf_cnt && offset < max_size; i++) {
150 		bufs[total++] = hns_roce_buf_dma_addr(buf, offset);
151 		offset += (1 << page_shift);
152 	}
153 
154 	return total;
155 }
156 
157 int hns_roce_get_umem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
158 			   int buf_cnt, struct ib_umem *umem,
159 			   unsigned int page_shift)
160 {
161 	struct ib_block_iter biter;
162 	int total = 0;
163 
164 	/* convert system page cnt to hw page cnt */
165 	rdma_umem_for_each_dma_block(umem, &biter, 1 << page_shift) {
166 		bufs[total++] = rdma_block_iter_dma_address(&biter);
167 		if (total >= buf_cnt)
168 			goto done;
169 	}
170 
171 done:
172 	return total;
173 }
174 
175 void hns_roce_cleanup_bitmap(struct hns_roce_dev *hr_dev)
176 {
177 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_XRC)
178 		ida_destroy(&hr_dev->xrcd_ida.ida);
179 
180 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
181 		ida_destroy(&hr_dev->srq_table.srq_ida.ida);
182 	hns_roce_cleanup_qp_table(hr_dev);
183 	hns_roce_cleanup_cq_table(hr_dev);
184 	ida_destroy(&hr_dev->mr_table.mtpt_ida.ida);
185 	ida_destroy(&hr_dev->pd_ida.ida);
186 	ida_destroy(&hr_dev->uar_ida.ida);
187 }
188