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