drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c

   1 /*
   2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <linux/highmem.h>
  34 #include <linux/kernel.h>
  35 #include <linux/module.h>
  36 #include <linux/mlx5/driver.h>
  37 #include <linux/mlx5/cmd.h>
  38 #include "mlx5_core.h"
  39
  40 enum {
  41         MLX5_PAGES_CANT_GIVE    = 0,
  42         MLX5_PAGES_GIVE         = 1,
  43         MLX5_PAGES_TAKE         = 2
  44 };
  45
  46 enum {
  47         MLX5_BOOT_PAGES         = 1,
  48         MLX5_INIT_PAGES         = 2,
  49         MLX5_POST_INIT_PAGES    = 3
  50 };
  51
  52 struct mlx5_pages_req {
  53         struct mlx5_core_dev *dev;
  54         u16     func_id;
  55         s32     npages;
  56         struct work_struct work;
  57 };
  58
  59 struct fw_page {
  60         struct rb_node          rb_node;
  61         u64                     addr;
  62         struct page            *page;
  63         u16                     func_id;
  64         unsigned long           bitmask;
  65         struct list_head        list;
  66         unsigned                free_count;
  67 };
  68
  69 struct mlx5_query_pages_inbox {
  70         struct mlx5_inbox_hdr   hdr;
  71         u8                      rsvd[8];
  72 };
  73
  74 struct mlx5_query_pages_outbox {
  75         struct mlx5_outbox_hdr  hdr;
  76         __be16                  rsvd;
  77         __be16                  func_id;
  78         __be32                  num_pages;
  79 };
  80
  81 struct mlx5_manage_pages_inbox {
  82         struct mlx5_inbox_hdr   hdr;
  83         __be16                  rsvd;
  84         __be16                  func_id;
  85         __be32                  num_entries;
  86         __be64                  pas[0];
  87 };
  88
  89 struct mlx5_manage_pages_outbox {
  90         struct mlx5_outbox_hdr  hdr;
  91         __be32                  num_entries;
  92         u8                      rsvd[4];
  93         __be64                  pas[0];
  94 };
  95
  96 enum {
  97         MAX_RECLAIM_TIME_MSECS  = 5000,
  98 };
  99
 100 enum {
 101         MLX5_MAX_RECLAIM_TIME_MILI      = 5000,
 102         MLX5_NUM_4K_IN_PAGE             = PAGE_SIZE / MLX5_ADAPTER_PAGE_SIZE,
 103 };
 104
 105 static int insert_page(struct mlx5_core_dev *dev, u64 addr, struct page *page, u16 func_id)
 106 {
 107         struct rb_root *root = &dev->priv.page_root;
 108         struct rb_node **new = &root->rb_node;
 109         struct rb_node *parent = NULL;
 110         struct fw_page *nfp;
 111         struct fw_page *tfp;
 112         int i;
 113
 114         while (*new) {
 115                 parent = *new;
 116                 tfp = rb_entry(parent, struct fw_page, rb_node);
 117                 if (tfp->addr < addr)
 118                         new = &parent->rb_left;
 119                 else if (tfp->addr > addr)
 120                         new = &parent->rb_right;
 121                 else
 122                         return -EEXIST;
 123         }
 124
 125         nfp = kzalloc(sizeof(*nfp), GFP_KERNEL);
 126         if (!nfp)
 127                 return -ENOMEM;
 128
 129         nfp->addr = addr;
 130         nfp->page = page;
 131         nfp->func_id = func_id;
 132         nfp->free_count = MLX5_NUM_4K_IN_PAGE;
 133         for (i = 0; i < MLX5_NUM_4K_IN_PAGE; i++)
 134                 set_bit(i, &nfp->bitmask);
 135
 136         rb_link_node(&nfp->rb_node, parent, new);
 137         rb_insert_color(&nfp->rb_node, root);
 138         list_add(&nfp->list, &dev->priv.free_list);
 139
 140         return 0;
 141 }
 142
 143 static struct fw_page *find_fw_page(struct mlx5_core_dev *dev, u64 addr)
 144 {
 145         struct rb_root *root = &dev->priv.page_root;
 146         struct rb_node *tmp = root->rb_node;
 147         struct fw_page *result = NULL;
 148         struct fw_page *tfp;
 149
 150         while (tmp) {
 151                 tfp = rb_entry(tmp, struct fw_page, rb_node);
 152                 if (tfp->addr < addr) {
 153                         tmp = tmp->rb_left;
 154                 } else if (tfp->addr > addr) {
 155                         tmp = tmp->rb_right;
 156                 } else {
 157                         result = tfp;
 158                         break;
 159                 }
 160         }
 161
 162         return result;
 163 }
 164
 165 static int mlx5_cmd_query_pages(struct mlx5_core_dev *dev, u16 *func_id,
 166                                 s32 *npages, int boot)
 167 {
 168         struct mlx5_query_pages_inbox   in;
 169         struct mlx5_query_pages_outbox  out;
 170         int err;
 171
 172         memset(&in, 0, sizeof(in));
 173         memset(&out, 0, sizeof(out));
 174         in.hdr.opcode = cpu_to_be16(MLX5_CMD_OP_QUERY_PAGES);
 175         in.hdr.opmod = boot ? cpu_to_be16(MLX5_BOOT_PAGES) : cpu_to_be16(MLX5_INIT_PAGES);
 176
 177         err = mlx5_cmd_exec(dev, &in, sizeof(in), &out, sizeof(out));
 178         if (err)
 179                 return err;
 180
 181         if (out.hdr.status)
 182                 return mlx5_cmd_status_to_err(&out.hdr);
 183
 184         *npages = be32_to_cpu(out.num_pages);
 185         *func_id = be16_to_cpu(out.func_id);
 186
 187         return err;
 188 }
 189
 190 static int alloc_4k(struct mlx5_core_dev *dev, u64 *addr)
 191 {
 192         struct fw_page *fp;
 193         unsigned n;
 194
 195         if (list_empty(&dev->priv.free_list))
 196                 return -ENOMEM;
 197
 198         fp = list_entry(dev->priv.free_list.next, struct fw_page, list);
 199         n = find_first_bit(&fp->bitmask, 8 * sizeof(fp->bitmask));
 200         if (n >= MLX5_NUM_4K_IN_PAGE) {
 201                 mlx5_core_warn(dev, "alloc 4k bug\n");
 202                 return -ENOENT;
 203         }
 204         clear_bit(n, &fp->bitmask);
 205         fp->free_count--;
 206         if (!fp->free_count)
 207                 list_del(&fp->list);
 208
 209         *addr = fp->addr + n * MLX5_ADAPTER_PAGE_SIZE;
 210
 211         return 0;
 212 }
 213
 214 #define MLX5_U64_4K_PAGE_MASK ((~(u64)0U) << PAGE_SHIFT)
 215
 216 static void free_4k(struct mlx5_core_dev *dev, u64 addr)
 217 {
 218         struct fw_page *fwp;
 219         int n;
 220
 221         fwp = find_fw_page(dev, addr & MLX5_U64_4K_PAGE_MASK);
 222         if (!fwp) {
 223                 mlx5_core_warn(dev, "page not found\n");
 224                 return;
 225         }
 226
 227         n = (addr & ~MLX5_U64_4K_PAGE_MASK) >> MLX5_ADAPTER_PAGE_SHIFT;
 228         fwp->free_count++;
 229         set_bit(n, &fwp->bitmask);
 230         if (fwp->free_count == MLX5_NUM_4K_IN_PAGE) {
 231                 rb_erase(&fwp->rb_node, &dev->priv.page_root);
 232                 if (fwp->free_count != 1)
 233                         list_del(&fwp->list);
 234                 dma_unmap_page(&dev->pdev->dev, addr & MLX5_U64_4K_PAGE_MASK,
 235                                PAGE_SIZE, DMA_BIDIRECTIONAL);
 236                 __free_page(fwp->page);
 237                 kfree(fwp);
 238         } else if (fwp->free_count == 1) {
 239                 list_add(&fwp->list, &dev->priv.free_list);
 240         }
 241 }
 242
 243 static int alloc_system_page(struct mlx5_core_dev *dev, u16 func_id)
 244 {
 245         struct page *page;
 246         u64 zero_addr = 1;
 247         u64 addr;
 248         int err;
 249         int nid = dev_to_node(&dev->pdev->dev);
 250
 251         page = alloc_pages_node(nid, GFP_HIGHUSER, 0);
 252         if (!page) {
 253                 mlx5_core_warn(dev, "failed to allocate page\n");
 254                 return -ENOMEM;
 255         }
 256 map:
 257         addr = dma_map_page(&dev->pdev->dev, page, 0,
 258                             PAGE_SIZE, DMA_BIDIRECTIONAL);
 259         if (dma_mapping_error(&dev->pdev->dev, addr)) {
 260                 mlx5_core_warn(dev, "failed dma mapping page\n");
 261                 err = -ENOMEM;
 262                 goto err_mapping;
 263         }
 264
 265         /* Firmware doesn't support page with physical address 0 */
 266         if (addr == 0) {
 267                 zero_addr = addr;
 268                 goto map;
 269         }
 270
 271         err = insert_page(dev, addr, page, func_id);
 272         if (err) {
 273                 mlx5_core_err(dev, "failed to track allocated page\n");
 274                 dma_unmap_page(&dev->pdev->dev, addr, PAGE_SIZE,
 275                                DMA_BIDIRECTIONAL);
 276         }
 277
 278 err_mapping:
 279         if (err)
 280                 __free_page(page);
 281
 282         if (zero_addr == 0)
 283                 dma_unmap_page(&dev->pdev->dev, zero_addr, PAGE_SIZE,
 284                                DMA_BIDIRECTIONAL);
 285
 286         return err;
 287 }
 288
 289 static void page_notify_fail(struct mlx5_core_dev *dev, u16 func_id)
 290 {
 291         struct mlx5_manage_pages_inbox *in;
 292         struct mlx5_manage_pages_outbox out;
 293         int err;
 294
 295         in = kzalloc(sizeof(*in), GFP_KERNEL);
 296         if (!in)
 297                 return;
 298
 299         memset(&out, 0, sizeof(out));
 300         in->hdr.opcode = cpu_to_be16(MLX5_CMD_OP_MANAGE_PAGES);
 301         in->hdr.opmod = cpu_to_be16(MLX5_PAGES_CANT_GIVE);
 302         in->func_id = cpu_to_be16(func_id);
 303         err = mlx5_cmd_exec(dev, in, sizeof(*in), &out, sizeof(out));
 304         if (!err)
 305                 err = mlx5_cmd_status_to_err(&out.hdr);
 306
 307         if (err)
 308                 mlx5_core_warn(dev, "page notify failed\n");
 309
 310         kfree(in);
 311 }
 312
 313 static int give_pages(struct mlx5_core_dev *dev, u16 func_id, int npages,
 314                       int notify_fail)
 315 {
 316         struct mlx5_manage_pages_inbox *in;
 317         struct mlx5_manage_pages_outbox out;
 318         int inlen;
 319         u64 addr;
 320         int err;
 321         int i;
 322
 323         inlen = sizeof(*in) + npages * sizeof(in->pas[0]);
 324         in = mlx5_vzalloc(inlen);
 325         if (!in) {
 326                 err = -ENOMEM;
 327                 mlx5_core_warn(dev, "vzalloc failed %d\n", inlen);
 328                 goto out_free;
 329         }
 330         memset(&out, 0, sizeof(out));
 331
 332         for (i = 0; i < npages; i++) {
 333 retry:
 334                 err = alloc_4k(dev, &addr);
 335                 if (err) {
 336                         if (err == -ENOMEM)
 337                                 err = alloc_system_page(dev, func_id);
 338                         if (err)
 339                                 goto out_4k;
 340
 341                         goto retry;
 342                 }
 343                 in->pas[i] = cpu_to_be64(addr);
 344         }
 345
 346         in->hdr.opcode = cpu_to_be16(MLX5_CMD_OP_MANAGE_PAGES);
 347         in->hdr.opmod = cpu_to_be16(MLX5_PAGES_GIVE);
 348         in->func_id = cpu_to_be16(func_id);
 349         in->num_entries = cpu_to_be32(npages);
 350         err = mlx5_cmd_exec(dev, in, inlen, &out, sizeof(out));
 351         if (err) {
 352                 mlx5_core_warn(dev, "func_id 0x%x, npages %d, err %d\n",
 353                                func_id, npages, err);
 354                 goto out_4k;
 355         }
 356         dev->priv.fw_pages += npages;
 357
 358         err = mlx5_cmd_status_to_err(&out.hdr);
 359         if (err) {
 360                 mlx5_core_warn(dev, "func_id 0x%x, npages %d, status %d\n",
 361                                func_id, npages, out.hdr.status);
 362                 goto out_4k;
 363         }
 364
 365         mlx5_core_dbg(dev, "err %d\n", err);
 366
 367         kvfree(in);
 368         return 0;
 369
 370 out_4k:
 371         for (i--; i >= 0; i--)
 372                 free_4k(dev, be64_to_cpu(in->pas[i]));
 373 out_free:
 374         kvfree(in);
 375         if (notify_fail)
 376                 page_notify_fail(dev, func_id);
 377         return err;
 378 }
 379
 380 static int reclaim_pages(struct mlx5_core_dev *dev, u32 func_id, int npages,
 381                          int *nclaimed)
 382 {
 383         struct mlx5_manage_pages_inbox   in;
 384         struct mlx5_manage_pages_outbox *out;
 385         int num_claimed;
 386         int outlen;
 387         u64 addr;
 388         int err;
 389         int i;
 390
 391         if (nclaimed)
 392                 *nclaimed = 0;
 393
 394         memset(&in, 0, sizeof(in));
 395         outlen = sizeof(*out) + npages * sizeof(out->pas[0]);
 396         out = mlx5_vzalloc(outlen);
 397         if (!out)
 398                 return -ENOMEM;
 399
 400         in.hdr.opcode = cpu_to_be16(MLX5_CMD_OP_MANAGE_PAGES);
 401         in.hdr.opmod = cpu_to_be16(MLX5_PAGES_TAKE);
 402         in.func_id = cpu_to_be16(func_id);
 403         in.num_entries = cpu_to_be32(npages);
 404         mlx5_core_dbg(dev, "npages %d, outlen %d\n", npages, outlen);
 405         err = mlx5_cmd_exec(dev, &in, sizeof(in), out, outlen);
 406         if (err) {
 407                 mlx5_core_err(dev, "failed reclaiming pages\n");
 408                 goto out_free;
 409         }
 410         dev->priv.fw_pages -= npages;
 411
 412         if (out->hdr.status) {
 413                 err = mlx5_cmd_status_to_err(&out->hdr);
 414                 goto out_free;
 415         }
 416
 417         num_claimed = be32_to_cpu(out->num_entries);
 418         if (nclaimed)
 419                 *nclaimed = num_claimed;
 420
 421         for (i = 0; i < num_claimed; i++) {
 422                 addr = be64_to_cpu(out->pas[i]);
 423                 free_4k(dev, addr);
 424         }
 425
 426 out_free:
 427         kvfree(out);
 428         return err;
 429 }
 430
 431 static void pages_work_handler(struct work_struct *work)
 432 {
 433         struct mlx5_pages_req *req = container_of(work, struct mlx5_pages_req, work);
 434         struct mlx5_core_dev *dev = req->dev;
 435         int err = 0;
 436
 437         if (req->npages < 0)
 438                 err = reclaim_pages(dev, req->func_id, -1 * req->npages, NULL);
 439         else if (req->npages > 0)
 440                 err = give_pages(dev, req->func_id, req->npages, 1);
 441
 442         if (err)
 443                 mlx5_core_warn(dev, "%s fail %d\n",
 444                                req->npages < 0 ? "reclaim" : "give", err);
 445
 446         kfree(req);
 447 }
 448
 449 void mlx5_core_req_pages_handler(struct mlx5_core_dev *dev, u16 func_id,
 450                                  s32 npages)
 451 {
 452         struct mlx5_pages_req *req;
 453
 454         req = kzalloc(sizeof(*req), GFP_ATOMIC);
 455         if (!req) {
 456                 mlx5_core_warn(dev, "failed to allocate pages request\n");
 457                 return;
 458         }
 459
 460         req->dev = dev;
 461         req->func_id = func_id;
 462         req->npages = npages;
 463         INIT_WORK(&req->work, pages_work_handler);
 464         queue_work(dev->priv.pg_wq, &req->work);
 465 }
 466
 467 int mlx5_satisfy_startup_pages(struct mlx5_core_dev *dev, int boot)
 468 {
 469         u16 uninitialized_var(func_id);
 470         s32 uninitialized_var(npages);
 471         int err;
 472
 473         err = mlx5_cmd_query_pages(dev, &func_id, &npages, boot);
 474         if (err)
 475                 return err;
 476
 477         mlx5_core_dbg(dev, "requested %d %s pages for func_id 0x%x\n",
 478                       npages, boot ? "boot" : "init", func_id);
 479
 480         return give_pages(dev, func_id, npages, 0);
 481 }
 482
 483 enum {
 484         MLX5_BLKS_FOR_RECLAIM_PAGES = 12
 485 };
 486
 487 static int optimal_reclaimed_pages(void)
 488 {
 489         struct mlx5_cmd_prot_block *block;
 490         struct mlx5_cmd_layout *lay;
 491         int ret;
 492
 493         ret = (sizeof(lay->out) + MLX5_BLKS_FOR_RECLAIM_PAGES * sizeof(block->data) -
 494                sizeof(struct mlx5_manage_pages_outbox)) /
 495                FIELD_SIZEOF(struct mlx5_manage_pages_outbox, pas[0]);
 496
 497         return ret;
 498 }
 499
 500 int mlx5_reclaim_startup_pages(struct mlx5_core_dev *dev)
 501 {
 502         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 503         struct fw_page *fwp;
 504         struct rb_node *p;
 505         int nclaimed = 0;
 506         int err = 0;
 507
 508         do {
 509                 p = rb_first(&dev->priv.page_root);
 510                 if (p) {
 511                         fwp = rb_entry(p, struct fw_page, rb_node);
 512                         if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
 513                                 free_4k(dev, fwp->addr);
 514                                 nclaimed = 1;
 515                         } else {
 516                                 err = reclaim_pages(dev, fwp->func_id,
 517                                                     optimal_reclaimed_pages(),
 518                                                     &nclaimed);
 519                         }
 520                         if (err) {
 521                                 mlx5_core_warn(dev, "failed reclaiming pages (%d)\n",
 522                                                err);
 523                                 return err;
 524                         }
 525                         if (nclaimed)
 526                                 end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 527                 }
 528                 if (time_after(jiffies, end)) {
 529                         mlx5_core_warn(dev, "FW did not return all pages. giving up...\n");
 530                         break;
 531                 }
 532         } while (p);
 533
 534         return 0;
 535 }
 536
 537 void mlx5_pagealloc_init(struct mlx5_core_dev *dev)
 538 {
 539         dev->priv.page_root = RB_ROOT;
 540         INIT_LIST_HEAD(&dev->priv.free_list);
 541 }
 542
 543 void mlx5_pagealloc_cleanup(struct mlx5_core_dev *dev)
 544 {
 545         /* nothing */
 546 }
 547
 548 int mlx5_pagealloc_start(struct mlx5_core_dev *dev)
 549 {
 550         dev->priv.pg_wq = create_singlethread_workqueue("mlx5_page_allocator");
 551         if (!dev->priv.pg_wq)
 552                 return -ENOMEM;
 553
 554         return 0;
 555 }
 556
 557 void mlx5_pagealloc_stop(struct mlx5_core_dev *dev)
 558 {
 559         destroy_workqueue(dev->priv.pg_wq);
 560 }