drivers/net/mlx4/icm.c

   1 /*
   2  * Copyright (c) 2005 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/init.h>
  35 #include <linux/errno.h>
  36 #include <linux/mm.h>
  37 #include <linux/scatterlist.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(chunk->mem[i].page,
  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(chunk->mem[i].page),
  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         mem->page = alloc_pages(gfp_mask, order);
  99         if (!mem->page)
 100                 return -ENOMEM;
 101
 102         mem->length = PAGE_SIZE << order;
 103         mem->offset = 0;
 104         return 0;
 105 }
 106
 107 static int mlx4_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
 108                                     int order, gfp_t gfp_mask)
 109 {
 110         void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order,
 111                                        &sg_dma_address(mem), gfp_mask);
 112         if (!buf)
 113                 return -ENOMEM;
 114
 115         sg_set_buf(mem, buf, PAGE_SIZE << order);
 116         BUG_ON(mem->offset);
 117         sg_dma_len(mem) = PAGE_SIZE << order;
 118         return 0;
 119 }
 120
 121 struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
 122                                 gfp_t gfp_mask, int coherent)
 123 {
 124         struct mlx4_icm *icm;
 125         struct mlx4_icm_chunk *chunk = NULL;
 126         int cur_order;
 127         int ret;
 128
 129         /* We use sg_set_buf for coherent allocs, which assumes low memory */
 130         BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
 131
 132         icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 133         if (!icm)
 134                 return NULL;
 135
 136         icm->refcount = 0;
 137         INIT_LIST_HEAD(&icm->chunk_list);
 138
 139         cur_order = get_order(MLX4_ICM_ALLOC_SIZE);
 140
 141         while (npages > 0) {
 142                 if (!chunk) {
 143                         chunk = kmalloc(sizeof *chunk,
 144                                         gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 145                         if (!chunk)
 146                                 goto fail;
 147
 148                         chunk->npages = 0;
 149                         chunk->nsg    = 0;
 150                         list_add_tail(&chunk->list, &icm->chunk_list);
 151                 }
 152
 153                 while (1 << cur_order > npages)
 154                         --cur_order;
 155
 156                 if (coherent)
 157                         ret = mlx4_alloc_icm_coherent(&dev->pdev->dev,
 158                                                       &chunk->mem[chunk->npages],
 159                                                       cur_order, gfp_mask);
 160                 else
 161                         ret = mlx4_alloc_icm_pages(&chunk->mem[chunk->npages],
 162                                                    cur_order, gfp_mask);
 163
 164                 if (!ret) {
 165                         ++chunk->npages;
 166
 167                         if (coherent)
 168                                 ++chunk->nsg;
 169                         else if (chunk->npages == MLX4_ICM_CHUNK_LEN) {
 170                                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 171                                                         chunk->npages,
 172                                                         PCI_DMA_BIDIRECTIONAL);
 173
 174                                 if (chunk->nsg <= 0)
 175                                         goto fail;
 176
 177                                 chunk = NULL;
 178                         }
 179
 180                         npages -= 1 << cur_order;
 181                 } else {
 182                         --cur_order;
 183                         if (cur_order < 0)
 184                                 goto fail;
 185                 }
 186         }
 187
 188         if (!coherent && chunk) {
 189                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 190                                         chunk->npages,
 191                                         PCI_DMA_BIDIRECTIONAL);
 192
 193                 if (chunk->nsg <= 0)
 194                         goto fail;
 195         }
 196
 197         return icm;
 198
 199 fail:
 200         mlx4_free_icm(dev, icm, coherent);
 201         return NULL;
 202 }
 203
 204 static int mlx4_MAP_ICM(struct mlx4_dev *dev, struct mlx4_icm *icm, u64 virt)
 205 {
 206         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM, icm, virt);
 207 }
 208
 209 int mlx4_UNMAP_ICM(struct mlx4_dev *dev, u64 virt, u32 page_count)
 210 {
 211         return mlx4_cmd(dev, virt, page_count, 0, MLX4_CMD_UNMAP_ICM,
 212                         MLX4_CMD_TIME_CLASS_B);
 213 }
 214
 215 int mlx4_MAP_ICM_page(struct mlx4_dev *dev, u64 dma_addr, u64 virt)
 216 {
 217         struct mlx4_cmd_mailbox *mailbox;
 218         __be64 *inbox;
 219         int err;
 220
 221         mailbox = mlx4_alloc_cmd_mailbox(dev);
 222         if (IS_ERR(mailbox))
 223                 return PTR_ERR(mailbox);
 224         inbox = mailbox->buf;
 225
 226         inbox[0] = cpu_to_be64(virt);
 227         inbox[1] = cpu_to_be64(dma_addr);
 228
 229         err = mlx4_cmd(dev, mailbox->dma, 1, 0, MLX4_CMD_MAP_ICM,
 230                        MLX4_CMD_TIME_CLASS_B);
 231
 232         mlx4_free_cmd_mailbox(dev, mailbox);
 233
 234         if (!err)
 235                 mlx4_dbg(dev, "Mapped page at %llx to %llx for ICM.\n",
 236                           (unsigned long long) dma_addr, (unsigned long long) virt);
 237
 238         return err;
 239 }
 240
 241 int mlx4_MAP_ICM_AUX(struct mlx4_dev *dev, struct mlx4_icm *icm)
 242 {
 243         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM_AUX, icm, -1);
 244 }
 245
 246 int mlx4_UNMAP_ICM_AUX(struct mlx4_dev *dev)
 247 {
 248         return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_UNMAP_ICM_AUX, MLX4_CMD_TIME_CLASS_B);
 249 }
 250
 251 int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj)
 252 {
 253         int i = (obj & (table->num_obj - 1)) / (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_KERNEL : 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, int obj)
 287 {
 288         int i;
 289
 290         i = (obj & (table->num_obj - 1)) / (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
 291
 292         mutex_lock(&table->mutex);
 293
 294         if (--table->icm[i]->refcount == 0) {
 295                 mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
 296                                MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 297                 mlx4_free_icm(dev, table->icm[i], table->coherent);
 298                 table->icm[i] = NULL;
 299         }
 300
 301         mutex_unlock(&table->mutex);
 302 }
 303
 304 void *mlx4_table_find(struct mlx4_icm_table *table, int obj, dma_addr_t *dma_handle)
 305 {
 306         int idx, offset, dma_offset, i;
 307         struct mlx4_icm_chunk *chunk;
 308         struct mlx4_icm *icm;
 309         struct page *page = NULL;
 310
 311         if (!table->lowmem)
 312                 return NULL;
 313
 314         mutex_lock(&table->mutex);
 315
 316         idx = (obj & (table->num_obj - 1)) * table->obj_size;
 317         icm = table->icm[idx / MLX4_TABLE_CHUNK_SIZE];
 318         dma_offset = offset = idx % MLX4_TABLE_CHUNK_SIZE;
 319
 320         if (!icm)
 321                 goto out;
 322
 323         list_for_each_entry(chunk, &icm->chunk_list, list) {
 324                 for (i = 0; i < chunk->npages; ++i) {
 325                         if (dma_handle && dma_offset >= 0) {
 326                                 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
 327                                         *dma_handle = sg_dma_address(&chunk->mem[i]) +
 328                                                 dma_offset;
 329                                 dma_offset -= sg_dma_len(&chunk->mem[i]);
 330                         }
 331                         /*
 332                          * DMA mapping can merge pages but not split them,
 333                          * so if we found the page, dma_handle has already
 334                          * been assigned to.
 335                          */
 336                         if (chunk->mem[i].length > offset) {
 337                                 page = chunk->mem[i].page;
 338                                 goto out;
 339                         }
 340                         offset -= chunk->mem[i].length;
 341                 }
 342         }
 343
 344 out:
 345         mutex_unlock(&table->mutex);
 346         return page ? lowmem_page_address(page) + offset : NULL;
 347 }
 348
 349 int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 350                          int start, int end)
 351 {
 352         int inc = MLX4_TABLE_CHUNK_SIZE / table->obj_size;
 353         int i, err;
 354
 355         for (i = start; i <= end; i += inc) {
 356                 err = mlx4_table_get(dev, table, i);
 357                 if (err)
 358                         goto fail;
 359         }
 360
 361         return 0;
 362
 363 fail:
 364         while (i > start) {
 365                 i -= inc;
 366                 mlx4_table_put(dev, table, i);
 367         }
 368
 369         return err;
 370 }
 371
 372 void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 373                           int start, int end)
 374 {
 375         int i;
 376
 377         for (i = start; i <= end; i += MLX4_TABLE_CHUNK_SIZE / table->obj_size)
 378                 mlx4_table_put(dev, table, i);
 379 }
 380
 381 int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 382                         u64 virt, int obj_size, int nobj, int reserved,
 383                         int use_lowmem, int use_coherent)
 384 {
 385         int obj_per_chunk;
 386         int num_icm;
 387         unsigned chunk_size;
 388         int i;
 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         for (i = 0; i * MLX4_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
 405                 chunk_size = MLX4_TABLE_CHUNK_SIZE;
 406                 if ((i + 1) * MLX4_TABLE_CHUNK_SIZE > nobj * obj_size)
 407                         chunk_size = PAGE_ALIGN(nobj * obj_size - i * MLX4_TABLE_CHUNK_SIZE);
 408
 409                 table->icm[i] = mlx4_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
 410                                                (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 411                                                __GFP_NOWARN, use_coherent);
 412                 if (!table->icm[i])
 413                         goto err;
 414                 if (mlx4_MAP_ICM(dev, table->icm[i], virt + i * MLX4_TABLE_CHUNK_SIZE)) {
 415                         mlx4_free_icm(dev, table->icm[i], use_coherent);
 416                         table->icm[i] = NULL;
 417                         goto err;
 418                 }
 419
 420                 /*
 421                  * Add a reference to this ICM chunk so that it never
 422                  * gets freed (since it contains reserved firmware objects).
 423                  */
 424                 ++table->icm[i]->refcount;
 425         }
 426
 427         return 0;
 428
 429 err:
 430         for (i = 0; i < num_icm; ++i)
 431                 if (table->icm[i]) {
 432                         mlx4_UNMAP_ICM(dev, virt + i * MLX4_TABLE_CHUNK_SIZE,
 433                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 434                         mlx4_free_icm(dev, table->icm[i], use_coherent);
 435                 }
 436
 437         return -ENOMEM;
 438 }
 439
 440 void mlx4_cleanup_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table)
 441 {
 442         int i;
 443
 444         for (i = 0; i < table->num_icm; ++i)
 445                 if (table->icm[i]) {
 446                         mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
 447                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 448                         mlx4_free_icm(dev, table->icm[i], table->coherent);
 449                 }
 450
 451         kfree(table->icm);
 452 }