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. 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->sg, chunk->npages,
  61                              PCI_DMA_BIDIRECTIONAL);
  62
  63         for (i = 0; i < chunk->npages; ++i)
  64                 __free_pages(sg_page(&chunk->sg[i]),
  65                              get_order(chunk->sg[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->buf[i].size,
  75                                   chunk->buf[i].addr,
  76                                   chunk->buf[i].dma_addr);
  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 mlx4_icm_buf *buf,
 115                                    int order, gfp_t gfp_mask)
 116 {
 117         buf->addr = dma_alloc_coherent(dev, PAGE_SIZE << order,
 118                                        &buf->dma_addr, gfp_mask);
 119         if (!buf->addr)
 120                 return -ENOMEM;
 121
 122         if (offset_in_page(buf->addr)) {
 123                 dma_free_coherent(dev, PAGE_SIZE << order, buf->addr,
 124                                   buf->dma_addr);
 125                 return -ENOMEM;
 126         }
 127
 128         buf->size = 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         gfp_t mask;
 139         int ret;
 140
 141         /* We use sg_set_buf for coherent allocs, which assumes low memory */
 142         BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
 143
 144         icm = kmalloc_node(sizeof(*icm),
 145                            gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN),
 146                            dev->numa_node);
 147         if (!icm) {
 148                 icm = kmalloc(sizeof(*icm),
 149                               gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 150                 if (!icm)
 151                         return NULL;
 152         }
 153
 154         icm->refcount = 0;
 155         INIT_LIST_HEAD(&icm->chunk_list);
 156
 157         cur_order = get_order(MLX4_ICM_ALLOC_SIZE);
 158
 159         while (npages > 0) {
 160                 if (!chunk) {
 161                         chunk = kzalloc_node(sizeof(*chunk),
 162                                              gfp_mask & ~(__GFP_HIGHMEM |
 163                                                           __GFP_NOWARN),
 164                                              dev->numa_node);
 165                         if (!chunk) {
 166                                 chunk = kzalloc(sizeof(*chunk),
 167                                                 gfp_mask & ~(__GFP_HIGHMEM |
 168                                                              __GFP_NOWARN));
 169                                 if (!chunk)
 170                                         goto fail;
 171                         }
 172                         chunk->coherent = coherent;
 173
 174                         if (!coherent)
 175                                 sg_init_table(chunk->sg, MLX4_ICM_CHUNK_LEN);
 176                         list_add_tail(&chunk->list, &icm->chunk_list);
 177                 }
 178
 179                 while (1 << cur_order > npages)
 180                         --cur_order;
 181
 182                 mask = gfp_mask;
 183                 if (cur_order)
 184                         mask &= ~__GFP_DIRECT_RECLAIM;
 185
 186                 if (coherent)
 187                         ret = mlx4_alloc_icm_coherent(&dev->persist->pdev->dev,
 188                                                 &chunk->buf[chunk->npages],
 189                                                 cur_order, mask);
 190                 else
 191                         ret = mlx4_alloc_icm_pages(&chunk->sg[chunk->npages],
 192                                                    cur_order, mask,
 193                                                    dev->numa_node);
 194
 195                 if (ret) {
 196                         if (--cur_order < 0)
 197                                 goto fail;
 198                         else
 199                                 continue;
 200                 }
 201
 202                 ++chunk->npages;
 203
 204                 if (coherent)
 205                         ++chunk->nsg;
 206                 else if (chunk->npages == MLX4_ICM_CHUNK_LEN) {
 207                         chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->sg,
 208                                                 chunk->npages,
 209                                                 PCI_DMA_BIDIRECTIONAL);
 210
 211                         if (chunk->nsg <= 0)
 212                                 goto fail;
 213                 }
 214
 215                 if (chunk->npages == MLX4_ICM_CHUNK_LEN)
 216                         chunk = NULL;
 217
 218                 npages -= 1 << cur_order;
 219         }
 220
 221         if (!coherent && chunk) {
 222                 chunk->nsg = pci_map_sg(dev->persist->pdev, chunk->sg,
 223                                         chunk->npages,
 224                                         PCI_DMA_BIDIRECTIONAL);
 225
 226                 if (chunk->nsg <= 0)
 227                         goto fail;
 228         }
 229
 230         return icm;
 231
 232 fail:
 233         mlx4_free_icm(dev, icm, coherent);
 234         return NULL;
 235 }
 236
 237 static int mlx4_MAP_ICM(struct mlx4_dev *dev, struct mlx4_icm *icm, u64 virt)
 238 {
 239         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM, icm, virt);
 240 }
 241
 242 static int mlx4_UNMAP_ICM(struct mlx4_dev *dev, u64 virt, u32 page_count)
 243 {
 244         return mlx4_cmd(dev, virt, page_count, 0, MLX4_CMD_UNMAP_ICM,
 245                         MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE);
 246 }
 247
 248 int mlx4_MAP_ICM_AUX(struct mlx4_dev *dev, struct mlx4_icm *icm)
 249 {
 250         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM_AUX, icm, -1);
 251 }
 252
 253 int mlx4_UNMAP_ICM_AUX(struct mlx4_dev *dev)
 254 {
 255         return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_UNMAP_ICM_AUX,
 256                         MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE);
 257 }
 258
 259 int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, u32 obj)
 260 {
 261         u32 i = (obj & (table->num_obj - 1)) /
 262                         (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
 263         int ret = 0;
 264
 265         mutex_lock(&table->mutex);
 266
 267         if (table->icm[i]) {
 268                 ++table->icm[i]->refcount;
 269                 goto out;
 270         }
 271
 272         table->icm[i] = mlx4_alloc_icm(dev, MLX4_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
 273                                        (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 274                                        __GFP_NOWARN, table->coherent);
 275         if (!table->icm[i]) {
 276                 ret = -ENOMEM;
 277                 goto out;
 278         }
 279
 280         if (mlx4_MAP_ICM(dev, table->icm[i], table->virt +
 281                          (u64) i * MLX4_TABLE_CHUNK_SIZE)) {
 282                 mlx4_free_icm(dev, table->icm[i], table->coherent);
 283                 table->icm[i] = NULL;
 284                 ret = -ENOMEM;
 285                 goto out;
 286         }
 287
 288         ++table->icm[i]->refcount;
 289
 290 out:
 291         mutex_unlock(&table->mutex);
 292         return ret;
 293 }
 294
 295 void mlx4_table_put(struct mlx4_dev *dev, struct mlx4_icm_table *table, u32 obj)
 296 {
 297         u32 i;
 298         u64 offset;
 299
 300         i = (obj & (table->num_obj - 1)) / (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
 301
 302         mutex_lock(&table->mutex);
 303
 304         if (--table->icm[i]->refcount == 0) {
 305                 offset = (u64) i * MLX4_TABLE_CHUNK_SIZE;
 306                 mlx4_UNMAP_ICM(dev, table->virt + offset,
 307                                MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 308                 mlx4_free_icm(dev, table->icm[i], table->coherent);
 309                 table->icm[i] = NULL;
 310         }
 311
 312         mutex_unlock(&table->mutex);
 313 }
 314
 315 void *mlx4_table_find(struct mlx4_icm_table *table, u32 obj,
 316                         dma_addr_t *dma_handle)
 317 {
 318         int offset, dma_offset, i;
 319         u64 idx;
 320         struct mlx4_icm_chunk *chunk;
 321         struct mlx4_icm *icm;
 322         void *addr = NULL;
 323
 324         if (!table->lowmem)
 325                 return NULL;
 326
 327         mutex_lock(&table->mutex);
 328
 329         idx = (u64) (obj & (table->num_obj - 1)) * table->obj_size;
 330         icm = table->icm[idx / MLX4_TABLE_CHUNK_SIZE];
 331         dma_offset = offset = idx % MLX4_TABLE_CHUNK_SIZE;
 332
 333         if (!icm)
 334                 goto out;
 335
 336         list_for_each_entry(chunk, &icm->chunk_list, list) {
 337                 for (i = 0; i < chunk->npages; ++i) {
 338                         dma_addr_t dma_addr;
 339                         size_t len;
 340
 341                         if (table->coherent) {
 342                                 len = chunk->buf[i].size;
 343                                 dma_addr = chunk->buf[i].dma_addr;
 344                                 addr = chunk->buf[i].addr;
 345                         } else {
 346                                 struct page *page;
 347
 348                                 len = sg_dma_len(&chunk->sg[i]);
 349                                 dma_addr = sg_dma_address(&chunk->sg[i]);
 350
 351                                 /* XXX: we should never do this for highmem
 352                                  * allocation.  This function either needs
 353                                  * to be split, or the kernel virtual address
 354                                  * return needs to be made optional.
 355                                  */
 356                                 page = sg_page(&chunk->sg[i]);
 357                                 addr = lowmem_page_address(page);
 358                         }
 359
 360                         if (dma_handle && dma_offset >= 0) {
 361                                 if (len > dma_offset)
 362                                         *dma_handle = dma_addr + dma_offset;
 363                                 dma_offset -= len;
 364                         }
 365
 366                         /*
 367                          * DMA mapping can merge pages but not split them,
 368                          * so if we found the page, dma_handle has already
 369                          * been assigned to.
 370                          */
 371                         if (len > offset)
 372                                 goto out;
 373                         offset -= len;
 374                 }
 375         }
 376
 377         addr = NULL;
 378 out:
 379         mutex_unlock(&table->mutex);
 380         return addr ? addr + offset : NULL;
 381 }
 382
 383 int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 384                          u32 start, u32 end)
 385 {
 386         int inc = MLX4_TABLE_CHUNK_SIZE / table->obj_size;
 387         int err;
 388         u32 i;
 389
 390         for (i = start; i <= end; i += inc) {
 391                 err = mlx4_table_get(dev, table, i);
 392                 if (err)
 393                         goto fail;
 394         }
 395
 396         return 0;
 397
 398 fail:
 399         while (i > start) {
 400                 i -= inc;
 401                 mlx4_table_put(dev, table, i);
 402         }
 403
 404         return err;
 405 }
 406
 407 void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 408                           u32 start, u32 end)
 409 {
 410         u32 i;
 411
 412         for (i = start; i <= end; i += MLX4_TABLE_CHUNK_SIZE / table->obj_size)
 413                 mlx4_table_put(dev, table, i);
 414 }
 415
 416 int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 417                         u64 virt, int obj_size, u32 nobj, int reserved,
 418                         int use_lowmem, int use_coherent)
 419 {
 420         int obj_per_chunk;
 421         int num_icm;
 422         unsigned chunk_size;
 423         int i;
 424         u64 size;
 425
 426         obj_per_chunk = MLX4_TABLE_CHUNK_SIZE / obj_size;
 427         if (WARN_ON(!obj_per_chunk))
 428                 return -EINVAL;
 429         num_icm = (nobj + obj_per_chunk - 1) / obj_per_chunk;
 430
 431         table->icm      = kvcalloc(num_icm, sizeof(*table->icm), GFP_KERNEL);
 432         if (!table->icm)
 433                 return -ENOMEM;
 434         table->virt     = virt;
 435         table->num_icm  = num_icm;
 436         table->num_obj  = nobj;
 437         table->obj_size = obj_size;
 438         table->lowmem   = use_lowmem;
 439         table->coherent = use_coherent;
 440         mutex_init(&table->mutex);
 441
 442         size = (u64) nobj * obj_size;
 443         for (i = 0; i * MLX4_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
 444                 chunk_size = MLX4_TABLE_CHUNK_SIZE;
 445                 if ((i + 1) * MLX4_TABLE_CHUNK_SIZE > size)
 446                         chunk_size = PAGE_ALIGN(size -
 447                                         i * MLX4_TABLE_CHUNK_SIZE);
 448
 449                 table->icm[i] = mlx4_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
 450                                                (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 451                                                __GFP_NOWARN, use_coherent);
 452                 if (!table->icm[i])
 453                         goto err;
 454                 if (mlx4_MAP_ICM(dev, table->icm[i], virt + i * MLX4_TABLE_CHUNK_SIZE)) {
 455                         mlx4_free_icm(dev, table->icm[i], use_coherent);
 456                         table->icm[i] = NULL;
 457                         goto err;
 458                 }
 459
 460                 /*
 461                  * Add a reference to this ICM chunk so that it never
 462                  * gets freed (since it contains reserved firmware objects).
 463                  */
 464                 ++table->icm[i]->refcount;
 465         }
 466
 467         return 0;
 468
 469 err:
 470         for (i = 0; i < num_icm; ++i)
 471                 if (table->icm[i]) {
 472                         mlx4_UNMAP_ICM(dev, virt + i * MLX4_TABLE_CHUNK_SIZE,
 473                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 474                         mlx4_free_icm(dev, table->icm[i], use_coherent);
 475                 }
 476
 477         kvfree(table->icm);
 478
 479         return -ENOMEM;
 480 }
 481
 482 void mlx4_cleanup_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table)
 483 {
 484         int i;
 485
 486         for (i = 0; i < table->num_icm; ++i)
 487                 if (table->icm[i]) {
 488                         mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
 489                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 490                         mlx4_free_icm(dev, table->icm[i], table->coherent);
 491                 }
 492
 493         kvfree(table->icm);
 494 }