1 /*
2  * Copyright (c) 2004 Topspin Communications.  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/bitmap.h>
36 
37 #include "mthca_dev.h"
38 
39 /* Trivial bitmap-based allocator */
mthca_alloc(struct mthca_alloc * alloc)40 u32 mthca_alloc(struct mthca_alloc *alloc)
41 {
42 	unsigned long flags;
43 	u32 obj;
44 
45 	spin_lock_irqsave(&alloc->lock, flags);
46 
47 	obj = find_next_zero_bit(alloc->table, alloc->max, alloc->last);
48 	if (obj >= alloc->max) {
49 		alloc->top = (alloc->top + alloc->max) & alloc->mask;
50 		obj = find_first_zero_bit(alloc->table, alloc->max);
51 	}
52 
53 	if (obj < alloc->max) {
54 		__set_bit(obj, alloc->table);
55 		obj |= alloc->top;
56 	} else
57 		obj = -1;
58 
59 	spin_unlock_irqrestore(&alloc->lock, flags);
60 
61 	return obj;
62 }
63 
mthca_free(struct mthca_alloc * alloc,u32 obj)64 void mthca_free(struct mthca_alloc *alloc, u32 obj)
65 {
66 	unsigned long flags;
67 
68 	obj &= alloc->max - 1;
69 
70 	spin_lock_irqsave(&alloc->lock, flags);
71 
72 	__clear_bit(obj, alloc->table);
73 	alloc->last = min(alloc->last, obj);
74 	alloc->top = (alloc->top + alloc->max) & alloc->mask;
75 
76 	spin_unlock_irqrestore(&alloc->lock, flags);
77 }
78 
mthca_alloc_init(struct mthca_alloc * alloc,u32 num,u32 mask,u32 reserved)79 int mthca_alloc_init(struct mthca_alloc *alloc, u32 num, u32 mask,
80 		     u32 reserved)
81 {
82 	/* num must be a power of 2 */
83 	if (num != 1 << (ffs(num) - 1))
84 		return -EINVAL;
85 
86 	alloc->last = 0;
87 	alloc->top  = 0;
88 	alloc->max  = num;
89 	alloc->mask = mask;
90 	spin_lock_init(&alloc->lock);
91 	alloc->table = bitmap_zalloc(num, GFP_KERNEL);
92 	if (!alloc->table)
93 		return -ENOMEM;
94 
95 	bitmap_set(alloc->table, 0, reserved);
96 
97 	return 0;
98 }
99 
mthca_alloc_cleanup(struct mthca_alloc * alloc)100 void mthca_alloc_cleanup(struct mthca_alloc *alloc)
101 {
102 	bitmap_free(alloc->table);
103 }
104 
105 /*
106  * Array of pointers with lazy allocation of leaf pages.  Callers of
107  * _get, _set and _clear methods must use a lock or otherwise
108  * serialize access to the array.
109  */
110 
111 #define MTHCA_ARRAY_MASK (PAGE_SIZE / sizeof (void *) - 1)
112 
mthca_array_get(struct mthca_array * array,int index)113 void *mthca_array_get(struct mthca_array *array, int index)
114 {
115 	int p = (index * sizeof (void *)) >> PAGE_SHIFT;
116 
117 	if (array->page_list[p].page)
118 		return array->page_list[p].page[index & MTHCA_ARRAY_MASK];
119 	else
120 		return NULL;
121 }
122 
mthca_array_set(struct mthca_array * array,int index,void * value)123 int mthca_array_set(struct mthca_array *array, int index, void *value)
124 {
125 	int p = (index * sizeof (void *)) >> PAGE_SHIFT;
126 
127 	/* Allocate with GFP_ATOMIC because we'll be called with locks held. */
128 	if (!array->page_list[p].page)
129 		array->page_list[p].page = (void **) get_zeroed_page(GFP_ATOMIC);
130 
131 	if (!array->page_list[p].page)
132 		return -ENOMEM;
133 
134 	array->page_list[p].page[index & MTHCA_ARRAY_MASK] = value;
135 	++array->page_list[p].used;
136 
137 	return 0;
138 }
139 
mthca_array_clear(struct mthca_array * array,int index)140 void mthca_array_clear(struct mthca_array *array, int index)
141 {
142 	int p = (index * sizeof (void *)) >> PAGE_SHIFT;
143 
144 	if (--array->page_list[p].used == 0) {
145 		free_page((unsigned long) array->page_list[p].page);
146 		array->page_list[p].page = NULL;
147 	} else
148 		array->page_list[p].page[index & MTHCA_ARRAY_MASK] = NULL;
149 
150 	if (array->page_list[p].used < 0)
151 		pr_debug("Array %p index %d page %d with ref count %d < 0\n",
152 			 array, index, p, array->page_list[p].used);
153 }
154 
mthca_array_init(struct mthca_array * array,int nent)155 int mthca_array_init(struct mthca_array *array, int nent)
156 {
157 	int npage = (nent * sizeof (void *) + PAGE_SIZE - 1) / PAGE_SIZE;
158 	int i;
159 
160 	array->page_list = kmalloc_array(npage, sizeof(*array->page_list),
161 					 GFP_KERNEL);
162 	if (!array->page_list)
163 		return -ENOMEM;
164 
165 	for (i = 0; i < npage; ++i) {
166 		array->page_list[i].page = NULL;
167 		array->page_list[i].used = 0;
168 	}
169 
170 	return 0;
171 }
172 
mthca_array_cleanup(struct mthca_array * array,int nent)173 void mthca_array_cleanup(struct mthca_array *array, int nent)
174 {
175 	int i;
176 
177 	for (i = 0; i < (nent * sizeof (void *) + PAGE_SIZE - 1) / PAGE_SIZE; ++i)
178 		free_page((unsigned long) array->page_list[i].page);
179 
180 	kfree(array->page_list);
181 }
182 
183 /*
184  * Handling for queue buffers -- we allocate a bunch of memory and
185  * register it in a memory region at HCA virtual address 0.  If the
186  * requested size is > max_direct, we split the allocation into
187  * multiple pages, so we don't require too much contiguous memory.
188  */
189 
mthca_buf_alloc(struct mthca_dev * dev,int size,int max_direct,union mthca_buf * buf,int * is_direct,struct mthca_pd * pd,int hca_write,struct mthca_mr * mr)190 int mthca_buf_alloc(struct mthca_dev *dev, int size, int max_direct,
191 		    union mthca_buf *buf, int *is_direct, struct mthca_pd *pd,
192 		    int hca_write, struct mthca_mr *mr)
193 {
194 	int err = -ENOMEM;
195 	int npages, shift;
196 	u64 *dma_list = NULL;
197 	dma_addr_t t;
198 	int i;
199 
200 	if (size <= max_direct) {
201 		*is_direct = 1;
202 		npages     = 1;
203 		shift      = get_order(size) + PAGE_SHIFT;
204 
205 		buf->direct.buf = dma_alloc_coherent(&dev->pdev->dev,
206 						     size, &t, GFP_KERNEL);
207 		if (!buf->direct.buf)
208 			return -ENOMEM;
209 
210 		dma_unmap_addr_set(&buf->direct, mapping, t);
211 
212 		while (t & ((1 << shift) - 1)) {
213 			--shift;
214 			npages *= 2;
215 		}
216 
217 		dma_list = kmalloc_array(npages, sizeof(*dma_list),
218 					 GFP_KERNEL);
219 		if (!dma_list)
220 			goto err_free;
221 
222 		for (i = 0; i < npages; ++i)
223 			dma_list[i] = t + i * (1 << shift);
224 	} else {
225 		*is_direct = 0;
226 		npages     = (size + PAGE_SIZE - 1) / PAGE_SIZE;
227 		shift      = PAGE_SHIFT;
228 
229 		dma_list = kmalloc_array(npages, sizeof(*dma_list),
230 					 GFP_KERNEL);
231 		if (!dma_list)
232 			return -ENOMEM;
233 
234 		buf->page_list = kmalloc_array(npages,
235 					       sizeof(*buf->page_list),
236 					       GFP_KERNEL);
237 		if (!buf->page_list)
238 			goto err_out;
239 
240 		for (i = 0; i < npages; ++i)
241 			buf->page_list[i].buf = NULL;
242 
243 		for (i = 0; i < npages; ++i) {
244 			buf->page_list[i].buf =
245 				dma_alloc_coherent(&dev->pdev->dev, PAGE_SIZE,
246 						   &t, GFP_KERNEL);
247 			if (!buf->page_list[i].buf)
248 				goto err_free;
249 
250 			dma_list[i] = t;
251 			dma_unmap_addr_set(&buf->page_list[i], mapping, t);
252 
253 			clear_page(buf->page_list[i].buf);
254 		}
255 	}
256 
257 	err = mthca_mr_alloc_phys(dev, pd->pd_num,
258 				  dma_list, shift, npages,
259 				  0, size,
260 				  MTHCA_MPT_FLAG_LOCAL_READ |
261 				  (hca_write ? MTHCA_MPT_FLAG_LOCAL_WRITE : 0),
262 				  mr);
263 	if (err)
264 		goto err_free;
265 
266 	kfree(dma_list);
267 
268 	return 0;
269 
270 err_free:
271 	mthca_buf_free(dev, size, buf, *is_direct, NULL);
272 
273 err_out:
274 	kfree(dma_list);
275 
276 	return err;
277 }
278 
mthca_buf_free(struct mthca_dev * dev,int size,union mthca_buf * buf,int is_direct,struct mthca_mr * mr)279 void mthca_buf_free(struct mthca_dev *dev, int size, union mthca_buf *buf,
280 		    int is_direct, struct mthca_mr *mr)
281 {
282 	int i;
283 
284 	if (mr)
285 		mthca_free_mr(dev, mr);
286 
287 	if (is_direct)
288 		dma_free_coherent(&dev->pdev->dev, size, buf->direct.buf,
289 				  dma_unmap_addr(&buf->direct, mapping));
290 	else {
291 		for (i = 0; i < (size + PAGE_SIZE - 1) / PAGE_SIZE; ++i)
292 			dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
293 					  buf->page_list[i].buf,
294 					  dma_unmap_addr(&buf->page_list[i],
295 							 mapping));
296 		kfree(buf->page_list);
297 	}
298 }
299