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