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/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(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, gfp_t gfp_mask)
  97 {
  98         struct page *page;
  99
 100         page = alloc_pages(gfp_mask, order);
 101         if (!page)
 102                 return -ENOMEM;
 103
 104         sg_set_page(mem, page, PAGE_SIZE << order, 0);
 105         return 0;
 106 }
 107
 108 static int mlx4_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
 109                                     int order, gfp_t gfp_mask)
 110 {
 111         void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order,
 112                                        &sg_dma_address(mem), gfp_mask);
 113         if (!buf)
 114                 return -ENOMEM;
 115
 116         sg_set_buf(mem, buf, PAGE_SIZE << order);
 117         BUG_ON(mem->offset);
 118         sg_dma_len(mem) = PAGE_SIZE << order;
 119         return 0;
 120 }
 121
 122 struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
 123                                 gfp_t gfp_mask, int coherent)
 124 {
 125         struct mlx4_icm *icm;
 126         struct mlx4_icm_chunk *chunk = NULL;
 127         int cur_order;
 128         int ret;
 129
 130         /* We use sg_set_buf for coherent allocs, which assumes low memory */
 131         BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
 132
 133         icm = kmalloc(sizeof *icm, gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 134         if (!icm)
 135                 return NULL;
 136
 137         icm->refcount = 0;
 138         INIT_LIST_HEAD(&icm->chunk_list);
 139
 140         cur_order = get_order(MLX4_ICM_ALLOC_SIZE);
 141
 142         while (npages > 0) {
 143                 if (!chunk) {
 144                         chunk = kmalloc(sizeof *chunk,
 145                                         gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 146                         if (!chunk)
 147                                 goto fail;
 148
 149                         sg_init_table(chunk->mem, MLX4_ICM_CHUNK_LEN);
 150                         chunk->npages = 0;
 151                         chunk->nsg    = 0;
 152                         list_add_tail(&chunk->list, &icm->chunk_list);
 153                 }
 154
 155                 while (1 << cur_order > npages)
 156                         --cur_order;
 157
 158                 if (coherent)
 159                         ret = mlx4_alloc_icm_coherent(&dev->pdev->dev,
 160                                                       &chunk->mem[chunk->npages],
 161                                                       cur_order, gfp_mask);
 162                 else
 163                         ret = mlx4_alloc_icm_pages(&chunk->mem[chunk->npages],
 164                                                    cur_order, gfp_mask);
 165
 166                 if (!ret) {
 167                         ++chunk->npages;
 168
 169                         if (coherent)
 170                                 ++chunk->nsg;
 171                         else if (chunk->npages == MLX4_ICM_CHUNK_LEN) {
 172                                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 173                                                         chunk->npages,
 174                                                         PCI_DMA_BIDIRECTIONAL);
 175
 176                                 if (chunk->nsg <= 0)
 177                                         goto fail;
 178                         }
 179
 180                         if (chunk->npages == MLX4_ICM_CHUNK_LEN)
 181                                 chunk = NULL;
 182
 183                         npages -= 1 << cur_order;
 184                 } else {
 185                         --cur_order;
 186                         if (cur_order < 0)
 187                                 goto fail;
 188                 }
 189         }
 190
 191         if (!coherent && chunk) {
 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         return icm;
 201
 202 fail:
 203         mlx4_free_icm(dev, icm, coherent);
 204         return NULL;
 205 }
 206
 207 static int mlx4_MAP_ICM(struct mlx4_dev *dev, struct mlx4_icm *icm, u64 virt)
 208 {
 209         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM, icm, virt);
 210 }
 211
 212 int mlx4_UNMAP_ICM(struct mlx4_dev *dev, u64 virt, u32 page_count)
 213 {
 214         return mlx4_cmd(dev, virt, page_count, 0, MLX4_CMD_UNMAP_ICM,
 215                         MLX4_CMD_TIME_CLASS_B);
 216 }
 217
 218 int mlx4_MAP_ICM_page(struct mlx4_dev *dev, u64 dma_addr, u64 virt)
 219 {
 220         struct mlx4_cmd_mailbox *mailbox;
 221         __be64 *inbox;
 222         int err;
 223
 224         mailbox = mlx4_alloc_cmd_mailbox(dev);
 225         if (IS_ERR(mailbox))
 226                 return PTR_ERR(mailbox);
 227         inbox = mailbox->buf;
 228
 229         inbox[0] = cpu_to_be64(virt);
 230         inbox[1] = cpu_to_be64(dma_addr);
 231
 232         err = mlx4_cmd(dev, mailbox->dma, 1, 0, MLX4_CMD_MAP_ICM,
 233                        MLX4_CMD_TIME_CLASS_B);
 234
 235         mlx4_free_cmd_mailbox(dev, mailbox);
 236
 237         if (!err)
 238                 mlx4_dbg(dev, "Mapped page at %llx to %llx for ICM.\n",
 239                           (unsigned long long) dma_addr, (unsigned long long) virt);
 240
 241         return err;
 242 }
 243
 244 int mlx4_MAP_ICM_AUX(struct mlx4_dev *dev, struct mlx4_icm *icm)
 245 {
 246         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM_AUX, icm, -1);
 247 }
 248
 249 int mlx4_UNMAP_ICM_AUX(struct mlx4_dev *dev)
 250 {
 251         return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_UNMAP_ICM_AUX, MLX4_CMD_TIME_CLASS_B);
 252 }
 253
 254 int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj)
 255 {
 256         int i = (obj & (table->num_obj - 1)) / (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
 257         int ret = 0;
 258
 259         mutex_lock(&table->mutex);
 260
 261         if (table->icm[i]) {
 262                 ++table->icm[i]->refcount;
 263                 goto out;
 264         }
 265
 266         table->icm[i] = mlx4_alloc_icm(dev, MLX4_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
 267                                        (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 268                                        __GFP_NOWARN, table->coherent);
 269         if (!table->icm[i]) {
 270                 ret = -ENOMEM;
 271                 goto out;
 272         }
 273
 274         if (mlx4_MAP_ICM(dev, table->icm[i], table->virt +
 275                          (u64) i * MLX4_TABLE_CHUNK_SIZE)) {
 276                 mlx4_free_icm(dev, table->icm[i], table->coherent);
 277                 table->icm[i] = NULL;
 278                 ret = -ENOMEM;
 279                 goto out;
 280         }
 281
 282         ++table->icm[i]->refcount;
 283
 284 out:
 285         mutex_unlock(&table->mutex);
 286         return ret;
 287 }
 288
 289 void mlx4_table_put(struct mlx4_dev *dev, struct mlx4_icm_table *table, int obj)
 290 {
 291         int i;
 292
 293         i = (obj & (table->num_obj - 1)) / (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
 294
 295         mutex_lock(&table->mutex);
 296
 297         if (--table->icm[i]->refcount == 0) {
 298                 mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
 299                                MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 300                 mlx4_free_icm(dev, table->icm[i], table->coherent);
 301                 table->icm[i] = NULL;
 302         }
 303
 304         mutex_unlock(&table->mutex);
 305 }
 306
 307 void *mlx4_table_find(struct mlx4_icm_table *table, int obj, dma_addr_t *dma_handle)
 308 {
 309         int idx, offset, dma_offset, i;
 310         struct mlx4_icm_chunk *chunk;
 311         struct mlx4_icm *icm;
 312         struct page *page = NULL;
 313
 314         if (!table->lowmem)
 315                 return NULL;
 316
 317         mutex_lock(&table->mutex);
 318
 319         idx = (obj & (table->num_obj - 1)) * table->obj_size;
 320         icm = table->icm[idx / MLX4_TABLE_CHUNK_SIZE];
 321         dma_offset = offset = idx % MLX4_TABLE_CHUNK_SIZE;
 322
 323         if (!icm)
 324                 goto out;
 325
 326         list_for_each_entry(chunk, &icm->chunk_list, list) {
 327                 for (i = 0; i < chunk->npages; ++i) {
 328                         if (dma_handle && dma_offset >= 0) {
 329                                 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
 330                                         *dma_handle = sg_dma_address(&chunk->mem[i]) +
 331                                                 dma_offset;
 332                                 dma_offset -= sg_dma_len(&chunk->mem[i]);
 333                         }
 334                         /*
 335                          * DMA mapping can merge pages but not split them,
 336                          * so if we found the page, dma_handle has already
 337                          * been assigned to.
 338                          */
 339                         if (chunk->mem[i].length > offset) {
 340                                 page = sg_page(&chunk->mem[i]);
 341                                 goto out;
 342                         }
 343                         offset -= chunk->mem[i].length;
 344                 }
 345         }
 346
 347 out:
 348         mutex_unlock(&table->mutex);
 349         return page ? lowmem_page_address(page) + offset : NULL;
 350 }
 351
 352 int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 353                          int start, int end)
 354 {
 355         int inc = MLX4_TABLE_CHUNK_SIZE / table->obj_size;
 356         int i, err;
 357
 358         for (i = start; i <= end; i += inc) {
 359                 err = mlx4_table_get(dev, table, i);
 360                 if (err)
 361                         goto fail;
 362         }
 363
 364         return 0;
 365
 366 fail:
 367         while (i > start) {
 368                 i -= inc;
 369                 mlx4_table_put(dev, table, i);
 370         }
 371
 372         return err;
 373 }
 374
 375 void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 376                           int start, int end)
 377 {
 378         int i;
 379
 380         for (i = start; i <= end; i += MLX4_TABLE_CHUNK_SIZE / table->obj_size)
 381                 mlx4_table_put(dev, table, i);
 382 }
 383
 384 int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 385                         u64 virt, int obj_size, int nobj, int reserved,
 386                         int use_lowmem, int use_coherent)
 387 {
 388         int obj_per_chunk;
 389         int num_icm;
 390         unsigned chunk_size;
 391         int i;
 392
 393         obj_per_chunk = MLX4_TABLE_CHUNK_SIZE / obj_size;
 394         num_icm = (nobj + obj_per_chunk - 1) / obj_per_chunk;
 395
 396         table->icm      = kcalloc(num_icm, sizeof *table->icm, GFP_KERNEL);
 397         if (!table->icm)
 398                 return -ENOMEM;
 399         table->virt     = virt;
 400         table->num_icm  = num_icm;
 401         table->num_obj  = nobj;
 402         table->obj_size = obj_size;
 403         table->lowmem   = use_lowmem;
 404         table->coherent = use_coherent;
 405         mutex_init(&table->mutex);
 406
 407         for (i = 0; i * MLX4_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
 408                 chunk_size = MLX4_TABLE_CHUNK_SIZE;
 409                 if ((i + 1) * MLX4_TABLE_CHUNK_SIZE > nobj * obj_size)
 410                         chunk_size = PAGE_ALIGN(nobj * obj_size - i * MLX4_TABLE_CHUNK_SIZE);
 411
 412                 table->icm[i] = mlx4_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
 413                                                (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 414                                                __GFP_NOWARN, use_coherent);
 415                 if (!table->icm[i])
 416                         goto err;
 417                 if (mlx4_MAP_ICM(dev, table->icm[i], virt + i * MLX4_TABLE_CHUNK_SIZE)) {
 418                         mlx4_free_icm(dev, table->icm[i], use_coherent);
 419                         table->icm[i] = NULL;
 420                         goto err;
 421                 }
 422
 423                 /*
 424                  * Add a reference to this ICM chunk so that it never
 425                  * gets freed (since it contains reserved firmware objects).
 426                  */
 427                 ++table->icm[i]->refcount;
 428         }
 429
 430         return 0;
 431
 432 err:
 433         for (i = 0; i < num_icm; ++i)
 434                 if (table->icm[i]) {
 435                         mlx4_UNMAP_ICM(dev, virt + i * MLX4_TABLE_CHUNK_SIZE,
 436                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 437                         mlx4_free_icm(dev, table->icm[i], use_coherent);
 438                 }
 439
 440         return -ENOMEM;
 441 }
 442
 443 void mlx4_cleanup_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table)
 444 {
 445         int i;
 446
 447         for (i = 0; i < table->num_icm; ++i)
 448                 if (table->icm[i]) {
 449                         mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
 450                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 451                         mlx4_free_icm(dev, table->icm[i], table->coherent);
 452                 }
 453
 454         kfree(table->icm);
 455 }