drivers/net/ethernet/mellanox/mlx4/icm.c

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