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
  38 #include <linux/mlx4/cmd.h>
  39
  40 #include "mlx4.h"
  41 #include "icm.h"
  42 #include "fw.h"
  43
  44 /*
  45  * We allocate in as big chunks as we can, up to a maximum of 256 KB
  46  * per chunk.
  47  */
  48 enum {
  49         MLX4_ICM_ALLOC_SIZE     = 1 << 18,
  50         MLX4_TABLE_CHUNK_SIZE   = 1 << 18
  51 };
  52
  53 static void mlx4_free_icm_pages(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk)
  54 {
  55         int i;
  56
  57         if (chunk->nsg > 0)
  58                 pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
  59                              PCI_DMA_BIDIRECTIONAL);
  60
  61         for (i = 0; i < chunk->npages; ++i)
  62                 __free_pages(sg_page(&chunk->mem[i]),
  63                              get_order(chunk->mem[i].length));
  64 }
  65
  66 static void mlx4_free_icm_coherent(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk)
  67 {
  68         int i;
  69
  70         for (i = 0; i < chunk->npages; ++i)
  71                 dma_free_coherent(&dev->pdev->dev, chunk->mem[i].length,
  72                                   lowmem_page_address(sg_page(&chunk->mem[i])),
  73                                   sg_dma_address(&chunk->mem[i]));
  74 }
  75
  76 void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm, int coherent)
  77 {
  78         struct mlx4_icm_chunk *chunk, *tmp;
  79
  80         if (!icm)
  81                 return;
  82
  83         list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
  84                 if (coherent)
  85                         mlx4_free_icm_coherent(dev, chunk);
  86                 else
  87                         mlx4_free_icm_pages(dev, chunk);
  88
  89                 kfree(chunk);
  90         }
  91
  92         kfree(icm);
  93 }
  94
  95 static int mlx4_alloc_icm_pages(struct scatterlist *mem, int order, gfp_t gfp_mask)
  96 {
  97         struct page *page;
  98
  99         page = alloc_pages(gfp_mask, order);
 100         if (!page)
 101                 return -ENOMEM;
 102
 103         sg_set_page(mem, page, PAGE_SIZE << order, 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                         sg_init_table(chunk->mem, MLX4_ICM_CHUNK_LEN);
 149                         chunk->npages = 0;
 150                         chunk->nsg    = 0;
 151                         list_add_tail(&chunk->list, &icm->chunk_list);
 152                 }
 153
 154                 while (1 << cur_order > npages)
 155                         --cur_order;
 156
 157                 if (coherent)
 158                         ret = mlx4_alloc_icm_coherent(&dev->pdev->dev,
 159                                                       &chunk->mem[chunk->npages],
 160                                                       cur_order, gfp_mask);
 161                 else
 162                         ret = mlx4_alloc_icm_pages(&chunk->mem[chunk->npages],
 163                                                    cur_order, gfp_mask);
 164
 165                 if (!ret) {
 166                         ++chunk->npages;
 167
 168                         if (coherent)
 169                                 ++chunk->nsg;
 170                         else if (chunk->npages == MLX4_ICM_CHUNK_LEN) {
 171                                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 172                                                         chunk->npages,
 173                                                         PCI_DMA_BIDIRECTIONAL);
 174
 175                                 if (chunk->nsg <= 0)
 176                                         goto fail;
 177
 178                                 chunk = NULL;
 179                         }
 180
 181                         npages -= 1 << cur_order;
 182                 } else {
 183                         --cur_order;
 184                         if (cur_order < 0)
 185                                 goto fail;
 186                 }
 187         }
 188
 189         if (!coherent && chunk) {
 190                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 191                                         chunk->npages,
 192                                         PCI_DMA_BIDIRECTIONAL);
 193
 194                 if (chunk->nsg <= 0)
 195                         goto fail;
 196         }
 197
 198         return icm;
 199
 200 fail:
 201         mlx4_free_icm(dev, icm, coherent);
 202         return NULL;
 203 }
 204
 205 static int mlx4_MAP_ICM(struct mlx4_dev *dev, struct mlx4_icm *icm, u64 virt)
 206 {
 207         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM, icm, virt);
 208 }
 209
 210 int mlx4_UNMAP_ICM(struct mlx4_dev *dev, u64 virt, u32 page_count)
 211 {
 212         return mlx4_cmd(dev, virt, page_count, 0, MLX4_CMD_UNMAP_ICM,
 213                         MLX4_CMD_TIME_CLASS_B);
 214 }
 215
 216 int mlx4_MAP_ICM_page(struct mlx4_dev *dev, u64 dma_addr, u64 virt)
 217 {
 218         struct mlx4_cmd_mailbox *mailbox;
 219         __be64 *inbox;
 220         int err;
 221
 222         mailbox = mlx4_alloc_cmd_mailbox(dev);
 223         if (IS_ERR(mailbox))
 224                 return PTR_ERR(mailbox);
 225         inbox = mailbox->buf;
 226
 227         inbox[0] = cpu_to_be64(virt);
 228         inbox[1] = cpu_to_be64(dma_addr);
 229
 230         err = mlx4_cmd(dev, mailbox->dma, 1, 0, MLX4_CMD_MAP_ICM,
 231                        MLX4_CMD_TIME_CLASS_B);
 232
 233         mlx4_free_cmd_mailbox(dev, mailbox);
 234
 235         if (!err)
 236                 mlx4_dbg(dev, "Mapped page at %llx to %llx for ICM.\n",
 237                           (unsigned long long) dma_addr, (unsigned long long) virt);
 238
 239         return err;
 240 }
 241
 242 int mlx4_MAP_ICM_AUX(struct mlx4_dev *dev, struct mlx4_icm *icm)
 243 {
 244         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM_AUX, icm, -1);
 245 }
 246
 247 int mlx4_UNMAP_ICM_AUX(struct mlx4_dev *dev)
 248 {
 249         return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_UNMAP_ICM_AUX, MLX4_CMD_TIME_CLASS_B);
 250 }
 251
 252 int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj)
 253 {
 254         int i = (obj & (table->num_obj - 1)) / (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
 255         int ret = 0;
 256
 257         mutex_lock(&table->mutex);
 258
 259         if (table->icm[i]) {
 260                 ++table->icm[i]->refcount;
 261                 goto out;
 262         }
 263
 264         table->icm[i] = mlx4_alloc_icm(dev, MLX4_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
 265                                        (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 266                                        __GFP_NOWARN, table->coherent);
 267         if (!table->icm[i]) {
 268                 ret = -ENOMEM;
 269                 goto out;
 270         }
 271
 272         if (mlx4_MAP_ICM(dev, table->icm[i], table->virt +
 273                          (u64) i * MLX4_TABLE_CHUNK_SIZE)) {
 274                 mlx4_free_icm(dev, table->icm[i], table->coherent);
 275                 table->icm[i] = NULL;
 276                 ret = -ENOMEM;
 277                 goto out;
 278         }
 279
 280         ++table->icm[i]->refcount;
 281
 282 out:
 283         mutex_unlock(&table->mutex);
 284         return ret;
 285 }
 286
 287 void mlx4_table_put(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj)
 288 {
 289         int i;
 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                 mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
 297                                MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 298                 mlx4_free_icm(dev, table->icm[i], table->coherent);
 299                 table->icm[i] = NULL;
 300         }
 301
 302         mutex_unlock(&table->mutex);
 303 }
 304
 305 void *mlx4_table_find(struct mlx4_icm_table *table, int obj, dma_addr_t *dma_handle)
 306 {
 307         int idx, offset, dma_offset, i;
 308         struct mlx4_icm_chunk *chunk;
 309         struct mlx4_icm *icm;
 310         struct page *page = NULL;
 311
 312         if (!table->lowmem)
 313                 return NULL;
 314
 315         mutex_lock(&table->mutex);
 316
 317         idx = (obj & (table->num_obj - 1)) * table->obj_size;
 318         icm = table->icm[idx / MLX4_TABLE_CHUNK_SIZE];
 319         dma_offset = offset = idx % MLX4_TABLE_CHUNK_SIZE;
 320
 321         if (!icm)
 322                 goto out;
 323
 324         list_for_each_entry(chunk, &icm->chunk_list, list) {
 325                 for (i = 0; i < chunk->npages; ++i) {
 326                         if (dma_handle && dma_offset >= 0) {
 327                                 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
 328                                         *dma_handle = sg_dma_address(&chunk->mem[i]) +
 329                                                 dma_offset;
 330                                 dma_offset -= sg_dma_len(&chunk->mem[i]);
 331                         }
 332                         /*
 333                          * DMA mapping can merge pages but not split them,
 334                          * so if we found the page, dma_handle has already
 335                          * been assigned to.
 336                          */
 337                         if (chunk->mem[i].length > offset) {
 338                                 page = sg_page(&chunk->mem[i]);
 339                                 goto out;
 340                         }
 341                         offset -= chunk->mem[i].length;
 342                 }
 343         }
 344
 345 out:
 346         mutex_unlock(&table->mutex);
 347         return page ? lowmem_page_address(page) + offset : NULL;
 348 }
 349
 350 int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 351                          int start, int end)
 352 {
 353         int inc = MLX4_TABLE_CHUNK_SIZE / table->obj_size;
 354         int i, err;
 355
 356         for (i = start; i <= end; i += inc) {
 357                 err = mlx4_table_get(dev, table, i);
 358                 if (err)
 359                         goto fail;
 360         }
 361
 362         return 0;
 363
 364 fail:
 365         while (i > start) {
 366                 i -= inc;
 367                 mlx4_table_put(dev, table, i);
 368         }
 369
 370         return err;
 371 }
 372
 373 void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 374                           int start, int end)
 375 {
 376         int i;
 377
 378         for (i = start; i <= end; i += MLX4_TABLE_CHUNK_SIZE / table->obj_size)
 379                 mlx4_table_put(dev, table, i);
 380 }
 381
 382 int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 383                         u64 virt, int obj_size, int nobj, int reserved,
 384                         int use_lowmem, int use_coherent)
 385 {
 386         int obj_per_chunk;
 387         int num_icm;
 388         unsigned chunk_size;
 389         int i;
 390
 391         obj_per_chunk = MLX4_TABLE_CHUNK_SIZE / obj_size;
 392         num_icm = (nobj + obj_per_chunk - 1) / obj_per_chunk;
 393
 394         table->icm      = kcalloc(num_icm, sizeof *table->icm, GFP_KERNEL);
 395         if (!table->icm)
 396                 return -ENOMEM;
 397         table->virt     = virt;
 398         table->num_icm  = num_icm;
 399         table->num_obj  = nobj;
 400         table->obj_size = obj_size;
 401         table->lowmem   = use_lowmem;
 402         table->coherent = use_coherent;
 403         mutex_init(&table->mutex);
 404
 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 > nobj * obj_size)
 408                         chunk_size = PAGE_ALIGN(nobj * obj_size - i * MLX4_TABLE_CHUNK_SIZE);
 409
 410                 table->icm[i] = mlx4_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
 411                                                (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 412                                                __GFP_NOWARN, use_coherent);
 413                 if (!table->icm[i])
 414                         goto err;
 415                 if (mlx4_MAP_ICM(dev, table->icm[i], virt + i * MLX4_TABLE_CHUNK_SIZE)) {
 416                         mlx4_free_icm(dev, table->icm[i], use_coherent);
 417                         table->icm[i] = NULL;
 418                         goto err;
 419                 }
 420
 421                 /*
 422                  * Add a reference to this ICM chunk so that it never
 423                  * gets freed (since it contains reserved firmware objects).
 424                  */
 425                 ++table->icm[i]->refcount;
 426         }
 427
 428         return 0;
 429
 430 err:
 431         for (i = 0; i < num_icm; ++i)
 432                 if (table->icm[i]) {
 433                         mlx4_UNMAP_ICM(dev, virt + i * MLX4_TABLE_CHUNK_SIZE,
 434                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 435                         mlx4_free_icm(dev, table->icm[i], use_coherent);
 436                 }
 437
 438         return -ENOMEM;
 439 }
 440
 441 void mlx4_cleanup_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table)
 442 {
 443         int i;
 444
 445         for (i = 0; i < table->num_icm; ++i)
 446                 if (table->icm[i]) {
 447                         mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
 448                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 449                         mlx4_free_icm(dev, table->icm[i], table->coherent);
 450                 }
 451
 452         kfree(table->icm);
 453 }