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/delay.h>
  37 #include <linux/mlx5/driver.h>
  38 #include "mlx5_core.h"
  39 #include "lib/eq.h"
  40
  41 enum {
  42         MLX5_PAGES_CANT_GIVE    = 0,
  43         MLX5_PAGES_GIVE         = 1,
  44         MLX5_PAGES_TAKE         = 2
  45 };
  46
  47 struct mlx5_pages_req {
  48         struct mlx5_core_dev *dev;
  49         u16     func_id;
  50         u8      ec_function;
  51         s32     npages;
  52         struct work_struct work;
  53         u8      release_all;
  54 };
  55
  56 struct fw_page {
  57         struct rb_node          rb_node;
  58         u64                     addr;
  59         struct page            *page;
  60         u16                     func_id;
  61         unsigned long           bitmask;
  62         struct list_head        list;
  63         unsigned                free_count;
  64 };
  65
  66 enum {
  67         MAX_RECLAIM_TIME_MSECS  = 5000,
  68         MAX_RECLAIM_VFS_PAGES_TIME_MSECS = 2 * 1000 * 60,
  69 };
  70
  71 enum {
  72         MLX5_MAX_RECLAIM_TIME_MILI      = 5000,
  73         MLX5_NUM_4K_IN_PAGE             = PAGE_SIZE / MLX5_ADAPTER_PAGE_SIZE,
  74 };
  75
  76 static int insert_page(struct mlx5_core_dev *dev, u64 addr, struct page *page, u16 func_id)
  77 {
  78         struct rb_root *root = &dev->priv.page_root;
  79         struct rb_node **new = &root->rb_node;
  80         struct rb_node *parent = NULL;
  81         struct fw_page *nfp;
  82         struct fw_page *tfp;
  83         int i;
  84
  85         while (*new) {
  86                 parent = *new;
  87                 tfp = rb_entry(parent, struct fw_page, rb_node);
  88                 if (tfp->addr < addr)
  89                         new = &parent->rb_left;
  90                 else if (tfp->addr > addr)
  91                         new = &parent->rb_right;
  92                 else
  93                         return -EEXIST;
  94         }
  95
  96         nfp = kzalloc(sizeof(*nfp), GFP_KERNEL);
  97         if (!nfp)
  98                 return -ENOMEM;
  99
 100         nfp->addr = addr;
 101         nfp->page = page;
 102         nfp->func_id = func_id;
 103         nfp->free_count = MLX5_NUM_4K_IN_PAGE;
 104         for (i = 0; i < MLX5_NUM_4K_IN_PAGE; i++)
 105                 set_bit(i, &nfp->bitmask);
 106
 107         rb_link_node(&nfp->rb_node, parent, new);
 108         rb_insert_color(&nfp->rb_node, root);
 109         list_add(&nfp->list, &dev->priv.free_list);
 110
 111         return 0;
 112 }
 113
 114 static struct fw_page *find_fw_page(struct mlx5_core_dev *dev, u64 addr)
 115 {
 116         struct rb_root *root = &dev->priv.page_root;
 117         struct rb_node *tmp = root->rb_node;
 118         struct fw_page *result = NULL;
 119         struct fw_page *tfp;
 120
 121         while (tmp) {
 122                 tfp = rb_entry(tmp, struct fw_page, rb_node);
 123                 if (tfp->addr < addr) {
 124                         tmp = tmp->rb_left;
 125                 } else if (tfp->addr > addr) {
 126                         tmp = tmp->rb_right;
 127                 } else {
 128                         result = tfp;
 129                         break;
 130                 }
 131         }
 132
 133         return result;
 134 }
 135
 136 static int mlx5_cmd_query_pages(struct mlx5_core_dev *dev, u16 *func_id,
 137                                 s32 *npages, int boot)
 138 {
 139         u32 out[MLX5_ST_SZ_DW(query_pages_out)] = {};
 140         u32 in[MLX5_ST_SZ_DW(query_pages_in)] = {};
 141         int err;
 142
 143         MLX5_SET(query_pages_in, in, opcode, MLX5_CMD_OP_QUERY_PAGES);
 144         MLX5_SET(query_pages_in, in, op_mod, boot ?
 145                  MLX5_QUERY_PAGES_IN_OP_MOD_BOOT_PAGES :
 146                  MLX5_QUERY_PAGES_IN_OP_MOD_INIT_PAGES);
 147         MLX5_SET(query_pages_in, in, embedded_cpu_function, mlx5_core_is_ecpf(dev));
 148
 149         err = mlx5_cmd_exec_inout(dev, query_pages, in, out);
 150         if (err)
 151                 return err;
 152
 153         *npages = MLX5_GET(query_pages_out, out, num_pages);
 154         *func_id = MLX5_GET(query_pages_out, out, function_id);
 155
 156         return err;
 157 }
 158
 159 static int alloc_4k(struct mlx5_core_dev *dev, u64 *addr)
 160 {
 161         struct fw_page *fp;
 162         unsigned n;
 163
 164         if (list_empty(&dev->priv.free_list))
 165                 return -ENOMEM;
 166
 167         fp = list_entry(dev->priv.free_list.next, struct fw_page, list);
 168         n = find_first_bit(&fp->bitmask, 8 * sizeof(fp->bitmask));
 169         if (n >= MLX5_NUM_4K_IN_PAGE) {
 170                 mlx5_core_warn(dev, "alloc 4k bug\n");
 171                 return -ENOENT;
 172         }
 173         clear_bit(n, &fp->bitmask);
 174         fp->free_count--;
 175         if (!fp->free_count)
 176                 list_del(&fp->list);
 177
 178         *addr = fp->addr + n * MLX5_ADAPTER_PAGE_SIZE;
 179
 180         return 0;
 181 }
 182
 183 #define MLX5_U64_4K_PAGE_MASK ((~(u64)0U) << PAGE_SHIFT)
 184
 185 static void free_fwp(struct mlx5_core_dev *dev, struct fw_page *fwp)
 186 {
 187         int n = (fwp->addr & ~MLX5_U64_4K_PAGE_MASK) >> MLX5_ADAPTER_PAGE_SHIFT;
 188
 189         fwp->free_count++;
 190         set_bit(n, &fwp->bitmask);
 191         if (fwp->free_count == MLX5_NUM_4K_IN_PAGE) {
 192                 rb_erase(&fwp->rb_node, &dev->priv.page_root);
 193                 if (fwp->free_count != 1)
 194                         list_del(&fwp->list);
 195                 dma_unmap_page(dev->device, fwp->addr & MLX5_U64_4K_PAGE_MASK,
 196                                PAGE_SIZE, DMA_BIDIRECTIONAL);
 197                 __free_page(fwp->page);
 198                 kfree(fwp);
 199         } else if (fwp->free_count == 1) {
 200                 list_add(&fwp->list, &dev->priv.free_list);
 201         }
 202 }
 203
 204 static void free_addr(struct mlx5_core_dev *dev, u64 addr)
 205 {
 206         struct fw_page *fwp;
 207
 208         fwp = find_fw_page(dev, addr & MLX5_U64_4K_PAGE_MASK);
 209         if (!fwp) {
 210                 mlx5_core_warn_rl(dev, "page not found\n");
 211                 return;
 212         }
 213         free_fwp(dev, fwp);
 214 }
 215
 216 static int alloc_system_page(struct mlx5_core_dev *dev, u16 func_id)
 217 {
 218         struct device *device = dev->device;
 219         int nid = dev_to_node(device);
 220         struct page *page;
 221         u64 zero_addr = 1;
 222         u64 addr;
 223         int err;
 224
 225         page = alloc_pages_node(nid, GFP_HIGHUSER, 0);
 226         if (!page) {
 227                 mlx5_core_warn(dev, "failed to allocate page\n");
 228                 return -ENOMEM;
 229         }
 230 map:
 231         addr = dma_map_page(device, page, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
 232         if (dma_mapping_error(device, addr)) {
 233                 mlx5_core_warn(dev, "failed dma mapping page\n");
 234                 err = -ENOMEM;
 235                 goto err_mapping;
 236         }
 237
 238         /* Firmware doesn't support page with physical address 0 */
 239         if (addr == 0) {
 240                 zero_addr = addr;
 241                 goto map;
 242         }
 243
 244         err = insert_page(dev, addr, page, func_id);
 245         if (err) {
 246                 mlx5_core_err(dev, "failed to track allocated page\n");
 247                 dma_unmap_page(device, addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
 248         }
 249
 250 err_mapping:
 251         if (err)
 252                 __free_page(page);
 253
 254         if (zero_addr == 0)
 255                 dma_unmap_page(device, zero_addr, PAGE_SIZE,
 256                                DMA_BIDIRECTIONAL);
 257
 258         return err;
 259 }
 260
 261 static void page_notify_fail(struct mlx5_core_dev *dev, u16 func_id,
 262                              bool ec_function)
 263 {
 264         u32 in[MLX5_ST_SZ_DW(manage_pages_in)] = {};
 265         int err;
 266
 267         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 268         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_CANT_GIVE);
 269         MLX5_SET(manage_pages_in, in, function_id, func_id);
 270         MLX5_SET(manage_pages_in, in, embedded_cpu_function, ec_function);
 271
 272         err = mlx5_cmd_exec_in(dev, manage_pages, in);
 273         if (err)
 274                 mlx5_core_warn(dev, "page notify failed func_id(%d) err(%d)\n",
 275                                func_id, err);
 276 }
 277
 278 static int give_pages(struct mlx5_core_dev *dev, u16 func_id, int npages,
 279                       int notify_fail, bool ec_function)
 280 {
 281         u32 out[MLX5_ST_SZ_DW(manage_pages_out)] = {0};
 282         int inlen = MLX5_ST_SZ_BYTES(manage_pages_in);
 283         u64 addr;
 284         int err;
 285         u32 *in;
 286         int i;
 287
 288         inlen += npages * MLX5_FLD_SZ_BYTES(manage_pages_in, pas[0]);
 289         in = kvzalloc(inlen, GFP_KERNEL);
 290         if (!in) {
 291                 err = -ENOMEM;
 292                 mlx5_core_warn(dev, "vzalloc failed %d\n", inlen);
 293                 goto out_free;
 294         }
 295
 296         for (i = 0; i < npages; i++) {
 297 retry:
 298                 err = alloc_4k(dev, &addr);
 299                 if (err) {
 300                         if (err == -ENOMEM)
 301                                 err = alloc_system_page(dev, func_id);
 302                         if (err)
 303                                 goto out_4k;
 304
 305                         goto retry;
 306                 }
 307                 MLX5_ARRAY_SET64(manage_pages_in, in, pas, i, addr);
 308         }
 309
 310         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 311         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_GIVE);
 312         MLX5_SET(manage_pages_in, in, function_id, func_id);
 313         MLX5_SET(manage_pages_in, in, input_num_entries, npages);
 314         MLX5_SET(manage_pages_in, in, embedded_cpu_function, ec_function);
 315
 316         err = mlx5_cmd_exec(dev, in, inlen, out, sizeof(out));
 317         if (err) {
 318                 mlx5_core_warn(dev, "func_id 0x%x, npages %d, err %d\n",
 319                                func_id, npages, err);
 320                 goto out_4k;
 321         }
 322
 323         dev->priv.fw_pages += npages;
 324         if (func_id)
 325                 dev->priv.vfs_pages += npages;
 326         else if (mlx5_core_is_ecpf(dev) && !ec_function)
 327                 dev->priv.peer_pf_pages += npages;
 328
 329         mlx5_core_dbg(dev, "npages %d, ec_function %d, func_id 0x%x, err %d\n",
 330                       npages, ec_function, func_id, err);
 331
 332         kvfree(in);
 333         return 0;
 334
 335 out_4k:
 336         for (i--; i >= 0; i--)
 337                 free_addr(dev, MLX5_GET64(manage_pages_in, in, pas[i]));
 338 out_free:
 339         kvfree(in);
 340         if (notify_fail)
 341                 page_notify_fail(dev, func_id, ec_function);
 342         return err;
 343 }
 344
 345 static void release_all_pages(struct mlx5_core_dev *dev, u32 func_id,
 346                               bool ec_function)
 347 {
 348         struct rb_node *p;
 349         int npages = 0;
 350
 351         p = rb_first(&dev->priv.page_root);
 352         while (p) {
 353                 struct fw_page *fwp = rb_entry(p, struct fw_page, rb_node);
 354
 355                 p = rb_next(p);
 356                 if (fwp->func_id != func_id)
 357                         continue;
 358                 free_fwp(dev, fwp);
 359                 npages++;
 360         }
 361
 362         dev->priv.fw_pages -= npages;
 363         if (func_id)
 364                 dev->priv.vfs_pages -= npages;
 365         else if (mlx5_core_is_ecpf(dev) && !ec_function)
 366                 dev->priv.peer_pf_pages -= npages;
 367
 368         mlx5_core_dbg(dev, "npages %d, ec_function %d, func_id 0x%x\n",
 369                       npages, ec_function, func_id);
 370 }
 371
 372 static int reclaim_pages_cmd(struct mlx5_core_dev *dev,
 373                              u32 *in, int in_size, u32 *out, int out_size)
 374 {
 375         struct fw_page *fwp;
 376         struct rb_node *p;
 377         u32 func_id;
 378         u32 npages;
 379         u32 i = 0;
 380
 381         if (dev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR)
 382                 return mlx5_cmd_exec(dev, in, in_size, out, out_size);
 383
 384         /* No hard feelings, we want our pages back! */
 385         npages = MLX5_GET(manage_pages_in, in, input_num_entries);
 386         func_id = MLX5_GET(manage_pages_in, in, function_id);
 387
 388         p = rb_first(&dev->priv.page_root);
 389         while (p && i < npages) {
 390                 fwp = rb_entry(p, struct fw_page, rb_node);
 391                 p = rb_next(p);
 392                 if (fwp->func_id != func_id)
 393                         continue;
 394
 395                 MLX5_ARRAY_SET64(manage_pages_out, out, pas, i, fwp->addr);
 396                 i++;
 397         }
 398
 399         MLX5_SET(manage_pages_out, out, output_num_entries, i);
 400         return 0;
 401 }
 402
 403 static int reclaim_pages(struct mlx5_core_dev *dev, u32 func_id, int npages,
 404                          int *nclaimed, bool ec_function)
 405 {
 406         int outlen = MLX5_ST_SZ_BYTES(manage_pages_out);
 407         u32 in[MLX5_ST_SZ_DW(manage_pages_in)] = {};
 408         int num_claimed;
 409         u32 *out;
 410         int err;
 411         int i;
 412
 413         if (nclaimed)
 414                 *nclaimed = 0;
 415
 416         outlen += npages * MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
 417         out = kvzalloc(outlen, GFP_KERNEL);
 418         if (!out)
 419                 return -ENOMEM;
 420
 421         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 422         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_TAKE);
 423         MLX5_SET(manage_pages_in, in, function_id, func_id);
 424         MLX5_SET(manage_pages_in, in, input_num_entries, npages);
 425         MLX5_SET(manage_pages_in, in, embedded_cpu_function, ec_function);
 426
 427         mlx5_core_dbg(dev, "npages %d, outlen %d\n", npages, outlen);
 428         err = reclaim_pages_cmd(dev, in, sizeof(in), out, outlen);
 429         if (err) {
 430                 mlx5_core_err(dev, "failed reclaiming pages: err %d\n", err);
 431                 goto out_free;
 432         }
 433
 434         num_claimed = MLX5_GET(manage_pages_out, out, output_num_entries);
 435         if (num_claimed > npages) {
 436                 mlx5_core_warn(dev, "fw returned %d, driver asked %d => corruption\n",
 437                                num_claimed, npages);
 438                 err = -EINVAL;
 439                 goto out_free;
 440         }
 441
 442         for (i = 0; i < num_claimed; i++)
 443                 free_addr(dev, MLX5_GET64(manage_pages_out, out, pas[i]));
 444
 445         if (nclaimed)
 446                 *nclaimed = num_claimed;
 447
 448         dev->priv.fw_pages -= num_claimed;
 449         if (func_id)
 450                 dev->priv.vfs_pages -= num_claimed;
 451         else if (mlx5_core_is_ecpf(dev) && !ec_function)
 452                 dev->priv.peer_pf_pages -= num_claimed;
 453
 454 out_free:
 455         kvfree(out);
 456         return err;
 457 }
 458
 459 static void pages_work_handler(struct work_struct *work)
 460 {
 461         struct mlx5_pages_req *req = container_of(work, struct mlx5_pages_req, work);
 462         struct mlx5_core_dev *dev = req->dev;
 463         int err = 0;
 464
 465         if (req->release_all)
 466                 release_all_pages(dev, req->func_id, req->ec_function);
 467         else if (req->npages < 0)
 468                 err = reclaim_pages(dev, req->func_id, -1 * req->npages, NULL,
 469                                     req->ec_function);
 470         else if (req->npages > 0)
 471                 err = give_pages(dev, req->func_id, req->npages, 1, req->ec_function);
 472
 473         if (err)
 474                 mlx5_core_warn(dev, "%s fail %d\n",
 475                                req->npages < 0 ? "reclaim" : "give", err);
 476
 477         kfree(req);
 478 }
 479
 480 enum {
 481         EC_FUNCTION_MASK = 0x8000,
 482         RELEASE_ALL_PAGES_MASK = 0x4000,
 483 };
 484
 485 static int req_pages_handler(struct notifier_block *nb,
 486                              unsigned long type, void *data)
 487 {
 488         struct mlx5_pages_req *req;
 489         struct mlx5_core_dev *dev;
 490         struct mlx5_priv *priv;
 491         struct mlx5_eqe *eqe;
 492         bool ec_function;
 493         bool release_all;
 494         u16 func_id;
 495         s32 npages;
 496
 497         priv = mlx5_nb_cof(nb, struct mlx5_priv, pg_nb);
 498         dev  = container_of(priv, struct mlx5_core_dev, priv);
 499         eqe  = data;
 500
 501         func_id = be16_to_cpu(eqe->data.req_pages.func_id);
 502         npages  = be32_to_cpu(eqe->data.req_pages.num_pages);
 503         ec_function = be16_to_cpu(eqe->data.req_pages.ec_function) & EC_FUNCTION_MASK;
 504         release_all = be16_to_cpu(eqe->data.req_pages.ec_function) &
 505                       RELEASE_ALL_PAGES_MASK;
 506         mlx5_core_dbg(dev, "page request for func 0x%x, npages %d, release_all %d\n",
 507                       func_id, npages, release_all);
 508         req = kzalloc(sizeof(*req), GFP_ATOMIC);
 509         if (!req) {
 510                 mlx5_core_warn(dev, "failed to allocate pages request\n");
 511                 return NOTIFY_DONE;
 512         }
 513
 514         req->dev = dev;
 515         req->func_id = func_id;
 516         req->npages = npages;
 517         req->ec_function = ec_function;
 518         req->release_all = release_all;
 519         INIT_WORK(&req->work, pages_work_handler);
 520         queue_work(dev->priv.pg_wq, &req->work);
 521         return NOTIFY_OK;
 522 }
 523
 524 int mlx5_satisfy_startup_pages(struct mlx5_core_dev *dev, int boot)
 525 {
 526         u16 uninitialized_var(func_id);
 527         s32 uninitialized_var(npages);
 528         int err;
 529
 530         err = mlx5_cmd_query_pages(dev, &func_id, &npages, boot);
 531         if (err)
 532                 return err;
 533
 534         mlx5_core_dbg(dev, "requested %d %s pages for func_id 0x%x\n",
 535                       npages, boot ? "boot" : "init", func_id);
 536
 537         return give_pages(dev, func_id, npages, 0, mlx5_core_is_ecpf(dev));
 538 }
 539
 540 enum {
 541         MLX5_BLKS_FOR_RECLAIM_PAGES = 12
 542 };
 543
 544 static int optimal_reclaimed_pages(void)
 545 {
 546         struct mlx5_cmd_prot_block *block;
 547         struct mlx5_cmd_layout *lay;
 548         int ret;
 549
 550         ret = (sizeof(lay->out) + MLX5_BLKS_FOR_RECLAIM_PAGES * sizeof(block->data) -
 551                MLX5_ST_SZ_BYTES(manage_pages_out)) /
 552                MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
 553
 554         return ret;
 555 }
 556
 557 int mlx5_reclaim_startup_pages(struct mlx5_core_dev *dev)
 558 {
 559         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 560         struct fw_page *fwp;
 561         struct rb_node *p;
 562         int nclaimed = 0;
 563         int err = 0;
 564
 565         do {
 566                 p = rb_first(&dev->priv.page_root);
 567                 if (p) {
 568                         fwp = rb_entry(p, struct fw_page, rb_node);
 569                         err = reclaim_pages(dev, fwp->func_id,
 570                                             optimal_reclaimed_pages(),
 571                                             &nclaimed, mlx5_core_is_ecpf(dev));
 572
 573                         if (err) {
 574                                 mlx5_core_warn(dev, "failed reclaiming pages (%d)\n",
 575                                                err);
 576                                 return err;
 577                         }
 578                         if (nclaimed)
 579                                 end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 580                 }
 581                 if (time_after(jiffies, end)) {
 582                         mlx5_core_warn(dev, "FW did not return all pages. giving up...\n");
 583                         break;
 584                 }
 585         } while (p);
 586
 587         WARN(dev->priv.fw_pages,
 588              "FW pages counter is %d after reclaiming all pages\n",
 589              dev->priv.fw_pages);
 590         WARN(dev->priv.vfs_pages,
 591              "VFs FW pages counter is %d after reclaiming all pages\n",
 592              dev->priv.vfs_pages);
 593         WARN(dev->priv.peer_pf_pages,
 594              "Peer PF FW pages counter is %d after reclaiming all pages\n",
 595              dev->priv.peer_pf_pages);
 596
 597         return 0;
 598 }
 599
 600 int mlx5_pagealloc_init(struct mlx5_core_dev *dev)
 601 {
 602         dev->priv.page_root = RB_ROOT;
 603         INIT_LIST_HEAD(&dev->priv.free_list);
 604         dev->priv.pg_wq = create_singlethread_workqueue("mlx5_page_allocator");
 605         if (!dev->priv.pg_wq)
 606                 return -ENOMEM;
 607
 608         return 0;
 609 }
 610
 611 void mlx5_pagealloc_cleanup(struct mlx5_core_dev *dev)
 612 {
 613         destroy_workqueue(dev->priv.pg_wq);
 614 }
 615
 616 void mlx5_pagealloc_start(struct mlx5_core_dev *dev)
 617 {
 618         MLX5_NB_INIT(&dev->priv.pg_nb, req_pages_handler, PAGE_REQUEST);
 619         mlx5_eq_notifier_register(dev, &dev->priv.pg_nb);
 620 }
 621
 622 void mlx5_pagealloc_stop(struct mlx5_core_dev *dev)
 623 {
 624         mlx5_eq_notifier_unregister(dev, &dev->priv.pg_nb);
 625         flush_workqueue(dev->priv.pg_wq);
 626 }
 627
 628 int mlx5_wait_for_pages(struct mlx5_core_dev *dev, int *pages)
 629 {
 630         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_VFS_PAGES_TIME_MSECS);
 631         int prev_pages = *pages;
 632
 633         /* In case of internal error we will free the pages manually later */
 634         if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
 635                 mlx5_core_warn(dev, "Skipping wait for vf pages stage");
 636                 return 0;
 637         }
 638
 639         mlx5_core_dbg(dev, "Waiting for %d pages\n", prev_pages);
 640         while (*pages) {
 641                 if (time_after(jiffies, end)) {
 642                         mlx5_core_warn(dev, "aborting while there are %d pending pages\n", *pages);
 643                         return -ETIMEDOUT;
 644                 }
 645                 if (*pages < prev_pages) {
 646                         end = jiffies + msecs_to_jiffies(MAX_RECLAIM_VFS_PAGES_TIME_MSECS);
 647                         prev_pages = *pages;
 648                 }
 649                 msleep(50);
 650         }
 651
 652         mlx5_core_dbg(dev, "All pages received\n");
 653         return 0;
 654 }