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. Note that the chunks are not necessarily in contiguous 48 * physical memory. 49 */ 50 enum { 51 MLX4_ICM_ALLOC_SIZE = 1 << 18, 52 MLX4_TABLE_CHUNK_SIZE = 1 << 18, 53 }; 54 55 static void mlx4_free_icm_pages(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk) 56 { 57 int i; 58 59 if (chunk->nsg > 0) 60 pci_unmap_sg(dev->persist->pdev, chunk->mem, chunk->npages, 61 PCI_DMA_BIDIRECTIONAL); 62 63 for (i = 0; i < chunk->npages; ++i) 64 __free_pages(sg_page(&chunk->mem[i]), 65 get_order(chunk->mem[i].length)); 66 } 67 68 static void mlx4_free_icm_coherent(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk) 69 { 70 int i; 71 72 for (i = 0; i < chunk->npages; ++i) 73 dma_free_coherent(&dev->persist->pdev->dev, 74 chunk->mem[i].length, 75 lowmem_page_address(sg_page(&chunk->mem[i])), 76 sg_dma_address(&chunk->mem[i])); 77 } 78 79 void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm, int coherent) 80 { 81 struct mlx4_icm_chunk *chunk, *tmp; 82 83 if (!icm) 84 return; 85 86 list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) { 87 if (coherent) 88 mlx4_free_icm_coherent(dev, chunk); 89 else 90 mlx4_free_icm_pages(dev, chunk); 91 92 kfree(chunk); 93 } 94 95 kfree(icm); 96 } 97 98 static int mlx4_alloc_icm_pages(struct scatterlist *mem, int order, 99 gfp_t gfp_mask, int node) 100 { 101 struct page *page; 102 103 page = alloc_pages_node(node, gfp_mask, order); 104 if (!page) { 105 page = alloc_pages(gfp_mask, order); 106 if (!page) 107 return -ENOMEM; 108 } 109 110 sg_set_page(mem, page, PAGE_SIZE << order, 0); 111 return 0; 112 } 113 114 static int mlx4_alloc_icm_coherent(struct device *dev, struct scatterlist *mem, 115 int order, gfp_t gfp_mask) 116 { 117 void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order, 118 &sg_dma_address(mem), gfp_mask); 119 if (!buf) 120 return -ENOMEM; 121 122 if (offset_in_page(buf)) { 123 dma_free_coherent(dev, PAGE_SIZE << order, 124 buf, sg_dma_address(mem)); 125 return -ENOMEM; 126 } 127 128 sg_set_buf(mem, buf, PAGE_SIZE << order); 129 sg_dma_len(mem) = PAGE_SIZE << order; 130 return 0; 131 } 132 133 struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages, 134 gfp_t gfp_mask, int coherent) 135 { 136 struct mlx4_icm *icm; 137 struct mlx4_icm_chunk *chunk = NULL; 138 int cur_order; 139 gfp_t mask; 140 int ret; 141 142 /* We use sg_set_buf for coherent allocs, which assumes low memory */ 143 BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM)); 144 145 icm = kmalloc_node(sizeof(*icm), 146 gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN), 147 dev->numa_node); 148 if (!icm) { 149 icm = kmalloc(sizeof(*icm), 150 gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN)); 151 if (!icm) 152 return NULL; 153 } 154 155 icm->refcount = 0; 156 INIT_LIST_HEAD(&icm->chunk_list); 157 158 cur_order = get_order(MLX4_ICM_ALLOC_SIZE); 159 160 while (npages > 0) { 161 if (!chunk) { 162 chunk = kmalloc_node(sizeof(*chunk), 163 gfp_mask & ~(__GFP_HIGHMEM | 164 __GFP_NOWARN), 165 dev->numa_node); 166 if (!chunk) { 167 chunk = kmalloc(sizeof(*chunk), 168 gfp_mask & ~(__GFP_HIGHMEM | 169 __GFP_NOWARN)); 170 if (!chunk) 171 goto fail; 172 } 173 174 sg_init_table(chunk->mem, MLX4_ICM_CHUNK_LEN); 175 chunk->npages = 0; 176 chunk->nsg = 0; 177 list_add_tail(&chunk->list, &icm->chunk_list); 178 } 179 180 while (1 << cur_order > npages) 181 --cur_order; 182 183 mask = gfp_mask; 184 if (cur_order) 185 mask &= ~__GFP_DIRECT_RECLAIM; 186 187 if (coherent) 188 ret = mlx4_alloc_icm_coherent(&dev->persist->pdev->dev, 189 &chunk->mem[chunk->npages], 190 cur_order, mask); 191 else 192 ret = mlx4_alloc_icm_pages(&chunk->mem[chunk->npages], 193 cur_order, mask, 194 dev->numa_node); 195 196 if (ret) { 197 if (--cur_order < 0) 198 goto fail; 199 else 200 continue; 201 } 202 203 ++chunk->npages; 204 205 if (coherent) 206 ++chunk->nsg; 207 else if (chunk->npages == MLX4_ICM_CHUNK_LEN) { 208 chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->mem, 209 chunk->npages, 210 PCI_DMA_BIDIRECTIONAL); 211 212 if (chunk->nsg <= 0) 213 goto fail; 214 } 215 216 if (chunk->npages == MLX4_ICM_CHUNK_LEN) 217 chunk = NULL; 218 219 npages -= 1 << cur_order; 220 } 221 222 if (!coherent && chunk) { 223 chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->mem, 224 chunk->npages, 225 PCI_DMA_BIDIRECTIONAL); 226 227 if (chunk->nsg <= 0) 228 goto fail; 229 } 230 231 return icm; 232 233 fail: 234 mlx4_free_icm(dev, icm, coherent); 235 return NULL; 236 } 237 238 static int mlx4_MAP_ICM(struct mlx4_dev *dev, struct mlx4_icm *icm, u64 virt) 239 { 240 return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM, icm, virt); 241 } 242 243 static int mlx4_UNMAP_ICM(struct mlx4_dev *dev, u64 virt, u32 page_count) 244 { 245 return mlx4_cmd(dev, virt, page_count, 0, MLX4_CMD_UNMAP_ICM, 246 MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE); 247 } 248 249 int mlx4_MAP_ICM_AUX(struct mlx4_dev *dev, struct mlx4_icm *icm) 250 { 251 return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM_AUX, icm, -1); 252 } 253 254 int mlx4_UNMAP_ICM_AUX(struct mlx4_dev *dev) 255 { 256 return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_UNMAP_ICM_AUX, 257 MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE); 258 } 259 260 int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, u32 obj) 261 { 262 u32 i = (obj & (table->num_obj - 1)) / 263 (MLX4_TABLE_CHUNK_SIZE / table->obj_size); 264 int ret = 0; 265 266 mutex_lock(&table->mutex); 267 268 if (table->icm[i]) { 269 ++table->icm[i]->refcount; 270 goto out; 271 } 272 273 table->icm[i] = mlx4_alloc_icm(dev, MLX4_TABLE_CHUNK_SIZE >> PAGE_SHIFT, 274 (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) | 275 __GFP_NOWARN, table->coherent); 276 if (!table->icm[i]) { 277 ret = -ENOMEM; 278 goto out; 279 } 280 281 if (mlx4_MAP_ICM(dev, table->icm[i], table->virt + 282 (u64) i * MLX4_TABLE_CHUNK_SIZE)) { 283 mlx4_free_icm(dev, table->icm[i], table->coherent); 284 table->icm[i] = NULL; 285 ret = -ENOMEM; 286 goto out; 287 } 288 289 ++table->icm[i]->refcount; 290 291 out: 292 mutex_unlock(&table->mutex); 293 return ret; 294 } 295 296 void mlx4_table_put(struct mlx4_dev *dev, struct mlx4_icm_table *table, u32 obj) 297 { 298 u32 i; 299 u64 offset; 300 301 i = (obj & (table->num_obj - 1)) / (MLX4_TABLE_CHUNK_SIZE / table->obj_size); 302 303 mutex_lock(&table->mutex); 304 305 if (--table->icm[i]->refcount == 0) { 306 offset = (u64) i * MLX4_TABLE_CHUNK_SIZE; 307 mlx4_UNMAP_ICM(dev, table->virt + offset, 308 MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE); 309 mlx4_free_icm(dev, table->icm[i], table->coherent); 310 table->icm[i] = NULL; 311 } 312 313 mutex_unlock(&table->mutex); 314 } 315 316 void *mlx4_table_find(struct mlx4_icm_table *table, u32 obj, 317 dma_addr_t *dma_handle) 318 { 319 int offset, dma_offset, i; 320 u64 idx; 321 struct mlx4_icm_chunk *chunk; 322 struct mlx4_icm *icm; 323 struct page *page = NULL; 324 325 if (!table->lowmem) 326 return NULL; 327 328 mutex_lock(&table->mutex); 329 330 idx = (u64) (obj & (table->num_obj - 1)) * table->obj_size; 331 icm = table->icm[idx / MLX4_TABLE_CHUNK_SIZE]; 332 dma_offset = offset = idx % MLX4_TABLE_CHUNK_SIZE; 333 334 if (!icm) 335 goto out; 336 337 list_for_each_entry(chunk, &icm->chunk_list, list) { 338 for (i = 0; i < chunk->npages; ++i) { 339 if (dma_handle && dma_offset >= 0) { 340 if (sg_dma_len(&chunk->mem[i]) > dma_offset) 341 *dma_handle = sg_dma_address(&chunk->mem[i]) + 342 dma_offset; 343 dma_offset -= sg_dma_len(&chunk->mem[i]); 344 } 345 /* 346 * DMA mapping can merge pages but not split them, 347 * so if we found the page, dma_handle has already 348 * been assigned to. 349 */ 350 if (chunk->mem[i].length > offset) { 351 page = sg_page(&chunk->mem[i]); 352 goto out; 353 } 354 offset -= chunk->mem[i].length; 355 } 356 } 357 358 out: 359 mutex_unlock(&table->mutex); 360 return page ? lowmem_page_address(page) + offset : NULL; 361 } 362 363 int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table, 364 u32 start, u32 end) 365 { 366 int inc = MLX4_TABLE_CHUNK_SIZE / table->obj_size; 367 int err; 368 u32 i; 369 370 for (i = start; i <= end; i += inc) { 371 err = mlx4_table_get(dev, table, i); 372 if (err) 373 goto fail; 374 } 375 376 return 0; 377 378 fail: 379 while (i > start) { 380 i -= inc; 381 mlx4_table_put(dev, table, i); 382 } 383 384 return err; 385 } 386 387 void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table, 388 u32 start, u32 end) 389 { 390 u32 i; 391 392 for (i = start; i <= end; i += MLX4_TABLE_CHUNK_SIZE / table->obj_size) 393 mlx4_table_put(dev, table, i); 394 } 395 396 int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table, 397 u64 virt, int obj_size, u32 nobj, int reserved, 398 int use_lowmem, int use_coherent) 399 { 400 int obj_per_chunk; 401 int num_icm; 402 unsigned chunk_size; 403 int i; 404 u64 size; 405 406 obj_per_chunk = MLX4_TABLE_CHUNK_SIZE / obj_size; 407 if (WARN_ON(!obj_per_chunk)) 408 return -EINVAL; 409 num_icm = (nobj + obj_per_chunk - 1) / obj_per_chunk; 410 411 table->icm = kvcalloc(num_icm, sizeof(*table->icm), GFP_KERNEL); 412 if (!table->icm) 413 return -ENOMEM; 414 table->virt = virt; 415 table->num_icm = num_icm; 416 table->num_obj = nobj; 417 table->obj_size = obj_size; 418 table->lowmem = use_lowmem; 419 table->coherent = use_coherent; 420 mutex_init(&table->mutex); 421 422 size = (u64) nobj * obj_size; 423 for (i = 0; i * MLX4_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) { 424 chunk_size = MLX4_TABLE_CHUNK_SIZE; 425 if ((i + 1) * MLX4_TABLE_CHUNK_SIZE > size) 426 chunk_size = PAGE_ALIGN(size - 427 i * MLX4_TABLE_CHUNK_SIZE); 428 429 table->icm[i] = mlx4_alloc_icm(dev, chunk_size >> PAGE_SHIFT, 430 (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) | 431 __GFP_NOWARN, use_coherent); 432 if (!table->icm[i]) 433 goto err; 434 if (mlx4_MAP_ICM(dev, table->icm[i], virt + i * MLX4_TABLE_CHUNK_SIZE)) { 435 mlx4_free_icm(dev, table->icm[i], use_coherent); 436 table->icm[i] = NULL; 437 goto err; 438 } 439 440 /* 441 * Add a reference to this ICM chunk so that it never 442 * gets freed (since it contains reserved firmware objects). 443 */ 444 ++table->icm[i]->refcount; 445 } 446 447 return 0; 448 449 err: 450 for (i = 0; i < num_icm; ++i) 451 if (table->icm[i]) { 452 mlx4_UNMAP_ICM(dev, virt + i * MLX4_TABLE_CHUNK_SIZE, 453 MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE); 454 mlx4_free_icm(dev, table->icm[i], use_coherent); 455 } 456 457 kvfree(table->icm); 458 459 return -ENOMEM; 460 } 461 462 void mlx4_cleanup_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table) 463 { 464 int i; 465 466 for (i = 0; i < table->num_icm; ++i) 467 if (table->icm[i]) { 468 mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE, 469 MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE); 470 mlx4_free_icm(dev, table->icm[i], table->coherent); 471 } 472 473 kvfree(table->icm); 474 } 475