1 /*
2  * Copyright (c) 2005, 2006, 2007, 2008 Mellanox Technologies. All rights reserved.
3  * Copyright (c) 2006, 2007 Cisco Systems, Inc.  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/errno.h>
35 #include <linux/mm.h>
36 #include <linux/scatterlist.h>
37 #include <linux/slab.h>
38 
39 #include <linux/mlx4/cmd.h>
40 
41 #include "mlx4.h"
42 #include "icm.h"
43 #include "fw.h"
44 
45 /*
46  * We allocate in as big chunks as we can, up to a maximum of 256 KB
47  * per chunk.
48  */
49 enum {
50 	MLX4_ICM_ALLOC_SIZE	= 1 << 18,
51 	MLX4_TABLE_CHUNK_SIZE	= 1 << 18
52 };
53 
54 static void mlx4_free_icm_pages(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk)
55 {
56 	int i;
57 
58 	if (chunk->nsg > 0)
59 		pci_unmap_sg(dev->persist->pdev, chunk->mem, chunk->npages,
60 			     PCI_DMA_BIDIRECTIONAL);
61 
62 	for (i = 0; i < chunk->npages; ++i)
63 		__free_pages(sg_page(&chunk->mem[i]),
64 			     get_order(chunk->mem[i].length));
65 }
66 
67 static void mlx4_free_icm_coherent(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk)
68 {
69 	int i;
70 
71 	for (i = 0; i < chunk->npages; ++i)
72 		dma_free_coherent(&dev->persist->pdev->dev,
73 				  chunk->mem[i].length,
74 				  lowmem_page_address(sg_page(&chunk->mem[i])),
75 				  sg_dma_address(&chunk->mem[i]));
76 }
77 
78 void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm, int coherent)
79 {
80 	struct mlx4_icm_chunk *chunk, *tmp;
81 
82 	if (!icm)
83 		return;
84 
85 	list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
86 		if (coherent)
87 			mlx4_free_icm_coherent(dev, chunk);
88 		else
89 			mlx4_free_icm_pages(dev, chunk);
90 
91 		kfree(chunk);
92 	}
93 
94 	kfree(icm);
95 }
96 
97 static int mlx4_alloc_icm_pages(struct scatterlist *mem, int order,
98 				gfp_t gfp_mask, int node)
99 {
100 	struct page *page;
101 
102 	page = alloc_pages_node(node, gfp_mask, order);
103 	if (!page) {
104 		page = alloc_pages(gfp_mask, order);
105 		if (!page)
106 			return -ENOMEM;
107 	}
108 
109 	sg_set_page(mem, page, PAGE_SIZE << order, 0);
110 	return 0;
111 }
112 
113 static int mlx4_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
114 				    int order, gfp_t gfp_mask)
115 {
116 	void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order,
117 				       &sg_dma_address(mem), gfp_mask);
118 	if (!buf)
119 		return -ENOMEM;
120 
121 	if (offset_in_page(buf)) {
122 		dma_free_coherent(dev, PAGE_SIZE << order,
123 				  buf, sg_dma_address(mem));
124 		return -ENOMEM;
125 	}
126 
127 	sg_set_buf(mem, buf, PAGE_SIZE << order);
128 	sg_dma_len(mem) = PAGE_SIZE << order;
129 	return 0;
130 }
131 
132 struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
133 				gfp_t gfp_mask, int coherent)
134 {
135 	struct mlx4_icm *icm;
136 	struct mlx4_icm_chunk *chunk = NULL;
137 	int cur_order;
138 	int ret;
139 
140 	/* We use sg_set_buf for coherent allocs, which assumes low memory */
141 	BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
142 
143 	icm = kmalloc_node(sizeof(*icm),
144 			   gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN),
145 			   dev->numa_node);
146 	if (!icm) {
147 		icm = kmalloc(sizeof(*icm),
148 			      gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
149 		if (!icm)
150 			return NULL;
151 	}
152 
153 	icm->refcount = 0;
154 	INIT_LIST_HEAD(&icm->chunk_list);
155 
156 	cur_order = get_order(MLX4_ICM_ALLOC_SIZE);
157 
158 	while (npages > 0) {
159 		if (!chunk) {
160 			chunk = kmalloc_node(sizeof(*chunk),
161 					     gfp_mask & ~(__GFP_HIGHMEM |
162 							  __GFP_NOWARN),
163 					     dev->numa_node);
164 			if (!chunk) {
165 				chunk = kmalloc(sizeof(*chunk),
166 						gfp_mask & ~(__GFP_HIGHMEM |
167 							     __GFP_NOWARN));
168 				if (!chunk)
169 					goto fail;
170 			}
171 
172 			sg_init_table(chunk->mem, MLX4_ICM_CHUNK_LEN);
173 			chunk->npages = 0;
174 			chunk->nsg    = 0;
175 			list_add_tail(&chunk->list, &icm->chunk_list);
176 		}
177 
178 		while (1 << cur_order > npages)
179 			--cur_order;
180 
181 		if (coherent)
182 			ret = mlx4_alloc_icm_coherent(&dev->persist->pdev->dev,
183 						      &chunk->mem[chunk->npages],
184 						      cur_order, gfp_mask);
185 		else
186 			ret = mlx4_alloc_icm_pages(&chunk->mem[chunk->npages],
187 						   cur_order, gfp_mask,
188 						   dev->numa_node);
189 
190 		if (ret) {
191 			if (--cur_order < 0)
192 				goto fail;
193 			else
194 				continue;
195 		}
196 
197 		++chunk->npages;
198 
199 		if (coherent)
200 			++chunk->nsg;
201 		else if (chunk->npages == MLX4_ICM_CHUNK_LEN) {
202 			chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->mem,
203 						chunk->npages,
204 						PCI_DMA_BIDIRECTIONAL);
205 
206 			if (chunk->nsg <= 0)
207 				goto fail;
208 		}
209 
210 		if (chunk->npages == MLX4_ICM_CHUNK_LEN)
211 			chunk = NULL;
212 
213 		npages -= 1 << cur_order;
214 	}
215 
216 	if (!coherent && chunk) {
217 		chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->mem,
218 					chunk->npages,
219 					PCI_DMA_BIDIRECTIONAL);
220 
221 		if (chunk->nsg <= 0)
222 			goto fail;
223 	}
224 
225 	return icm;
226 
227 fail:
228 	mlx4_free_icm(dev, icm, coherent);
229 	return NULL;
230 }
231 
232 static int mlx4_MAP_ICM(struct mlx4_dev *dev, struct mlx4_icm *icm, u64 virt)
233 {
234 	return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM, icm, virt);
235 }
236 
237 static int mlx4_UNMAP_ICM(struct mlx4_dev *dev, u64 virt, u32 page_count)
238 {
239 	return mlx4_cmd(dev, virt, page_count, 0, MLX4_CMD_UNMAP_ICM,
240 			MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE);
241 }
242 
243 int mlx4_MAP_ICM_AUX(struct mlx4_dev *dev, struct mlx4_icm *icm)
244 {
245 	return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM_AUX, icm, -1);
246 }
247 
248 int mlx4_UNMAP_ICM_AUX(struct mlx4_dev *dev)
249 {
250 	return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_UNMAP_ICM_AUX,
251 			MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE);
252 }
253 
254 int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, u32 obj)
255 {
256 	u32 i = (obj & (table->num_obj - 1)) /
257 			(MLX4_TABLE_CHUNK_SIZE / table->obj_size);
258 	int ret = 0;
259 
260 	mutex_lock(&table->mutex);
261 
262 	if (table->icm[i]) {
263 		++table->icm[i]->refcount;
264 		goto out;
265 	}
266 
267 	table->icm[i] = mlx4_alloc_icm(dev, MLX4_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
268 				       (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
269 				       __GFP_NOWARN, table->coherent);
270 	if (!table->icm[i]) {
271 		ret = -ENOMEM;
272 		goto out;
273 	}
274 
275 	if (mlx4_MAP_ICM(dev, table->icm[i], table->virt +
276 			 (u64) i * MLX4_TABLE_CHUNK_SIZE)) {
277 		mlx4_free_icm(dev, table->icm[i], table->coherent);
278 		table->icm[i] = NULL;
279 		ret = -ENOMEM;
280 		goto out;
281 	}
282 
283 	++table->icm[i]->refcount;
284 
285 out:
286 	mutex_unlock(&table->mutex);
287 	return ret;
288 }
289 
290 void mlx4_table_put(struct mlx4_dev *dev, struct mlx4_icm_table *table, u32 obj)
291 {
292 	u32 i;
293 	u64 offset;
294 
295 	i = (obj & (table->num_obj - 1)) / (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
296 
297 	mutex_lock(&table->mutex);
298 
299 	if (--table->icm[i]->refcount == 0) {
300 		offset = (u64) i * MLX4_TABLE_CHUNK_SIZE;
301 		mlx4_UNMAP_ICM(dev, table->virt + offset,
302 			       MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
303 		mlx4_free_icm(dev, table->icm[i], table->coherent);
304 		table->icm[i] = NULL;
305 	}
306 
307 	mutex_unlock(&table->mutex);
308 }
309 
310 void *mlx4_table_find(struct mlx4_icm_table *table, u32 obj,
311 			dma_addr_t *dma_handle)
312 {
313 	int offset, dma_offset, i;
314 	u64 idx;
315 	struct mlx4_icm_chunk *chunk;
316 	struct mlx4_icm *icm;
317 	struct page *page = NULL;
318 
319 	if (!table->lowmem)
320 		return NULL;
321 
322 	mutex_lock(&table->mutex);
323 
324 	idx = (u64) (obj & (table->num_obj - 1)) * table->obj_size;
325 	icm = table->icm[idx / MLX4_TABLE_CHUNK_SIZE];
326 	dma_offset = offset = idx % MLX4_TABLE_CHUNK_SIZE;
327 
328 	if (!icm)
329 		goto out;
330 
331 	list_for_each_entry(chunk, &icm->chunk_list, list) {
332 		for (i = 0; i < chunk->npages; ++i) {
333 			if (dma_handle && dma_offset >= 0) {
334 				if (sg_dma_len(&chunk->mem[i]) > dma_offset)
335 					*dma_handle = sg_dma_address(&chunk->mem[i]) +
336 						dma_offset;
337 				dma_offset -= sg_dma_len(&chunk->mem[i]);
338 			}
339 			/*
340 			 * DMA mapping can merge pages but not split them,
341 			 * so if we found the page, dma_handle has already
342 			 * been assigned to.
343 			 */
344 			if (chunk->mem[i].length > offset) {
345 				page = sg_page(&chunk->mem[i]);
346 				goto out;
347 			}
348 			offset -= chunk->mem[i].length;
349 		}
350 	}
351 
352 out:
353 	mutex_unlock(&table->mutex);
354 	return page ? lowmem_page_address(page) + offset : NULL;
355 }
356 
357 int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
358 			 u32 start, u32 end)
359 {
360 	int inc = MLX4_TABLE_CHUNK_SIZE / table->obj_size;
361 	int err;
362 	u32 i;
363 
364 	for (i = start; i <= end; i += inc) {
365 		err = mlx4_table_get(dev, table, i);
366 		if (err)
367 			goto fail;
368 	}
369 
370 	return 0;
371 
372 fail:
373 	while (i > start) {
374 		i -= inc;
375 		mlx4_table_put(dev, table, i);
376 	}
377 
378 	return err;
379 }
380 
381 void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
382 			  u32 start, u32 end)
383 {
384 	u32 i;
385 
386 	for (i = start; i <= end; i += MLX4_TABLE_CHUNK_SIZE / table->obj_size)
387 		mlx4_table_put(dev, table, i);
388 }
389 
390 int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
391 			u64 virt, int obj_size,	u32 nobj, int reserved,
392 			int use_lowmem, int use_coherent)
393 {
394 	int obj_per_chunk;
395 	int num_icm;
396 	unsigned chunk_size;
397 	int i;
398 	u64 size;
399 
400 	obj_per_chunk = MLX4_TABLE_CHUNK_SIZE / obj_size;
401 	num_icm = (nobj + obj_per_chunk - 1) / obj_per_chunk;
402 
403 	table->icm      = kcalloc(num_icm, sizeof *table->icm, GFP_KERNEL);
404 	if (!table->icm)
405 		return -ENOMEM;
406 	table->virt     = virt;
407 	table->num_icm  = num_icm;
408 	table->num_obj  = nobj;
409 	table->obj_size = obj_size;
410 	table->lowmem   = use_lowmem;
411 	table->coherent = use_coherent;
412 	mutex_init(&table->mutex);
413 
414 	size = (u64) nobj * obj_size;
415 	for (i = 0; i * MLX4_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
416 		chunk_size = MLX4_TABLE_CHUNK_SIZE;
417 		if ((i + 1) * MLX4_TABLE_CHUNK_SIZE > size)
418 			chunk_size = PAGE_ALIGN(size -
419 					i * MLX4_TABLE_CHUNK_SIZE);
420 
421 		table->icm[i] = mlx4_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
422 					       (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
423 					       __GFP_NOWARN, use_coherent);
424 		if (!table->icm[i])
425 			goto err;
426 		if (mlx4_MAP_ICM(dev, table->icm[i], virt + i * MLX4_TABLE_CHUNK_SIZE)) {
427 			mlx4_free_icm(dev, table->icm[i], use_coherent);
428 			table->icm[i] = NULL;
429 			goto err;
430 		}
431 
432 		/*
433 		 * Add a reference to this ICM chunk so that it never
434 		 * gets freed (since it contains reserved firmware objects).
435 		 */
436 		++table->icm[i]->refcount;
437 	}
438 
439 	return 0;
440 
441 err:
442 	for (i = 0; i < num_icm; ++i)
443 		if (table->icm[i]) {
444 			mlx4_UNMAP_ICM(dev, virt + i * MLX4_TABLE_CHUNK_SIZE,
445 				       MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
446 			mlx4_free_icm(dev, table->icm[i], use_coherent);
447 		}
448 
449 	kfree(table->icm);
450 
451 	return -ENOMEM;
452 }
453 
454 void mlx4_cleanup_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table)
455 {
456 	int i;
457 
458 	for (i = 0; i < table->num_icm; ++i)
459 		if (table->icm[i]) {
460 			mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
461 				       MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
462 			mlx4_free_icm(dev, table->icm[i], table->coherent);
463 		}
464 
465 	kfree(table->icm);
466 }
467