1 /*
2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/errno.h>
34 #include <linux/slab.h>
35 #include <linux/mm.h>
36 #include <linux/export.h>
37 #include <linux/bitmap.h>
38 #include <linux/dma-mapping.h>
39 #include <linux/vmalloc.h>
40 #include <linux/mlx5/driver.h>
41 
42 #include "mlx5_core.h"
43 
44 /* Handling for queue buffers -- we allocate a bunch of memory and
45  * register it in a memory region at HCA virtual address 0.  If the
46  * requested size is > max_direct, we split the allocation into
47  * multiple pages, so we don't require too much contiguous memory.
48  */
49 
50 int mlx5_buf_alloc(struct mlx5_core_dev *dev, int size, int max_direct,
51 		   struct mlx5_buf *buf)
52 {
53 	dma_addr_t t;
54 
55 	buf->size = size;
56 	if (size <= max_direct) {
57 		buf->nbufs        = 1;
58 		buf->npages       = 1;
59 		buf->page_shift   = (u8)get_order(size) + PAGE_SHIFT;
60 		buf->direct.buf   = dma_zalloc_coherent(&dev->pdev->dev,
61 							size, &t, GFP_KERNEL);
62 		if (!buf->direct.buf)
63 			return -ENOMEM;
64 
65 		buf->direct.map = t;
66 
67 		while (t & ((1 << buf->page_shift) - 1)) {
68 			--buf->page_shift;
69 			buf->npages *= 2;
70 		}
71 	} else {
72 		int i;
73 
74 		buf->direct.buf  = NULL;
75 		buf->nbufs       = (size + PAGE_SIZE - 1) / PAGE_SIZE;
76 		buf->npages      = buf->nbufs;
77 		buf->page_shift  = PAGE_SHIFT;
78 		buf->page_list   = kcalloc(buf->nbufs, sizeof(*buf->page_list),
79 					   GFP_KERNEL);
80 		if (!buf->page_list)
81 			return -ENOMEM;
82 
83 		for (i = 0; i < buf->nbufs; i++) {
84 			buf->page_list[i].buf =
85 				dma_zalloc_coherent(&dev->pdev->dev, PAGE_SIZE,
86 						    &t, GFP_KERNEL);
87 			if (!buf->page_list[i].buf)
88 				goto err_free;
89 
90 			buf->page_list[i].map = t;
91 		}
92 
93 		if (BITS_PER_LONG == 64) {
94 			struct page **pages;
95 			pages = kmalloc(sizeof(*pages) * buf->nbufs, GFP_KERNEL);
96 			if (!pages)
97 				goto err_free;
98 			for (i = 0; i < buf->nbufs; i++)
99 				pages[i] = virt_to_page(buf->page_list[i].buf);
100 			buf->direct.buf = vmap(pages, buf->nbufs, VM_MAP, PAGE_KERNEL);
101 			kfree(pages);
102 			if (!buf->direct.buf)
103 				goto err_free;
104 		}
105 	}
106 
107 	return 0;
108 
109 err_free:
110 	mlx5_buf_free(dev, buf);
111 
112 	return -ENOMEM;
113 }
114 EXPORT_SYMBOL_GPL(mlx5_buf_alloc);
115 
116 void mlx5_buf_free(struct mlx5_core_dev *dev, struct mlx5_buf *buf)
117 {
118 	int i;
119 
120 	if (buf->nbufs == 1)
121 		dma_free_coherent(&dev->pdev->dev, buf->size, buf->direct.buf,
122 				  buf->direct.map);
123 	else {
124 		if (BITS_PER_LONG == 64)
125 			vunmap(buf->direct.buf);
126 
127 		for (i = 0; i < buf->nbufs; i++)
128 			if (buf->page_list[i].buf)
129 				dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
130 						  buf->page_list[i].buf,
131 						  buf->page_list[i].map);
132 		kfree(buf->page_list);
133 	}
134 }
135 EXPORT_SYMBOL_GPL(mlx5_buf_free);
136 
137 static struct mlx5_db_pgdir *mlx5_alloc_db_pgdir(struct device *dma_device)
138 {
139 	struct mlx5_db_pgdir *pgdir;
140 
141 	pgdir = kzalloc(sizeof(*pgdir), GFP_KERNEL);
142 	if (!pgdir)
143 		return NULL;
144 
145 	bitmap_fill(pgdir->bitmap, MLX5_DB_PER_PAGE);
146 	pgdir->db_page = dma_alloc_coherent(dma_device, PAGE_SIZE,
147 					    &pgdir->db_dma, GFP_KERNEL);
148 	if (!pgdir->db_page) {
149 		kfree(pgdir);
150 		return NULL;
151 	}
152 
153 	return pgdir;
154 }
155 
156 static int mlx5_alloc_db_from_pgdir(struct mlx5_db_pgdir *pgdir,
157 				    struct mlx5_db *db)
158 {
159 	int offset;
160 	int i;
161 
162 	i = find_first_bit(pgdir->bitmap, MLX5_DB_PER_PAGE);
163 	if (i >= MLX5_DB_PER_PAGE)
164 		return -ENOMEM;
165 
166 	__clear_bit(i, pgdir->bitmap);
167 
168 	db->u.pgdir = pgdir;
169 	db->index   = i;
170 	offset = db->index * L1_CACHE_BYTES;
171 	db->db      = pgdir->db_page + offset / sizeof(*pgdir->db_page);
172 	db->dma     = pgdir->db_dma  + offset;
173 
174 	db->db[0] = 0;
175 	db->db[1] = 0;
176 
177 	return 0;
178 }
179 
180 int mlx5_db_alloc(struct mlx5_core_dev *dev, struct mlx5_db *db)
181 {
182 	struct mlx5_db_pgdir *pgdir;
183 	int ret = 0;
184 
185 	mutex_lock(&dev->priv.pgdir_mutex);
186 
187 	list_for_each_entry(pgdir, &dev->priv.pgdir_list, list)
188 		if (!mlx5_alloc_db_from_pgdir(pgdir, db))
189 			goto out;
190 
191 	pgdir = mlx5_alloc_db_pgdir(&(dev->pdev->dev));
192 	if (!pgdir) {
193 		ret = -ENOMEM;
194 		goto out;
195 	}
196 
197 	list_add(&pgdir->list, &dev->priv.pgdir_list);
198 
199 	/* This should never fail -- we just allocated an empty page: */
200 	WARN_ON(mlx5_alloc_db_from_pgdir(pgdir, db));
201 
202 out:
203 	mutex_unlock(&dev->priv.pgdir_mutex);
204 
205 	return ret;
206 }
207 EXPORT_SYMBOL_GPL(mlx5_db_alloc);
208 
209 void mlx5_db_free(struct mlx5_core_dev *dev, struct mlx5_db *db)
210 {
211 	mutex_lock(&dev->priv.pgdir_mutex);
212 
213 	__set_bit(db->index, db->u.pgdir->bitmap);
214 
215 	if (bitmap_full(db->u.pgdir->bitmap, MLX5_DB_PER_PAGE)) {
216 		dma_free_coherent(&(dev->pdev->dev), PAGE_SIZE,
217 				  db->u.pgdir->db_page, db->u.pgdir->db_dma);
218 		list_del(&db->u.pgdir->list);
219 		kfree(db->u.pgdir);
220 	}
221 
222 	mutex_unlock(&dev->priv.pgdir_mutex);
223 }
224 EXPORT_SYMBOL_GPL(mlx5_db_free);
225 
226 
227 void mlx5_fill_page_array(struct mlx5_buf *buf, __be64 *pas)
228 {
229 	u64 addr;
230 	int i;
231 
232 	for (i = 0; i < buf->npages; i++) {
233 		if (buf->nbufs == 1)
234 			addr = buf->direct.map + (i << buf->page_shift);
235 		else
236 			addr = buf->page_list[i].map;
237 
238 		pas[i] = cpu_to_be64(addr);
239 	}
240 }
241 EXPORT_SYMBOL_GPL(mlx5_fill_page_array);
242