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 <linux/xarray.h>
  39 #include "mlx5_core.h"
  40 #include "lib/eq.h"
  41
  42 enum {
  43         MLX5_PAGES_CANT_GIVE    = 0,
  44         MLX5_PAGES_GIVE         = 1,
  45         MLX5_PAGES_TAKE         = 2
  46 };
  47
  48 struct mlx5_pages_req {
  49         struct mlx5_core_dev *dev;
  50         u16     func_id;
  51         u8      ec_function;
  52         s32     npages;
  53         struct work_struct work;
  54         u8      release_all;
  55 };
  56
  57 struct fw_page {
  58         struct rb_node          rb_node;
  59         u64                     addr;
  60         struct page            *page;
  61         u16                     func_id;
  62         unsigned long           bitmask;
  63         struct list_head        list;
  64         unsigned                free_count;
  65 };
  66
  67 enum {
  68         MAX_RECLAIM_TIME_MSECS  = 5000,
  69         MAX_RECLAIM_VFS_PAGES_TIME_MSECS = 2 * 1000 * 60,
  70 };
  71
  72 enum {
  73         MLX5_MAX_RECLAIM_TIME_MILI      = 5000,
  74         MLX5_NUM_4K_IN_PAGE             = PAGE_SIZE / MLX5_ADAPTER_PAGE_SIZE,
  75 };
  76
  77 static struct rb_root *page_root_per_func_id(struct mlx5_core_dev *dev, u16 func_id)
  78 {
  79         struct rb_root *root;
  80         int err;
  81
  82         root = xa_load(&dev->priv.page_root_xa, func_id);
  83         if (root)
  84                 return root;
  85
  86         root = kzalloc(sizeof(*root), GFP_KERNEL);
  87         if (!root)
  88                 return ERR_PTR(-ENOMEM);
  89
  90         err = xa_insert(&dev->priv.page_root_xa, func_id, root, GFP_KERNEL);
  91         if (err) {
  92                 kfree(root);
  93                 return ERR_PTR(err);
  94         }
  95
  96         *root = RB_ROOT;
  97
  98         return root;
  99 }
 100
 101 static int insert_page(struct mlx5_core_dev *dev, u64 addr, struct page *page, u16 func_id)
 102 {
 103         struct rb_node *parent = NULL;
 104         struct rb_root *root;
 105         struct rb_node **new;
 106         struct fw_page *nfp;
 107         struct fw_page *tfp;
 108         int i;
 109
 110         root = page_root_per_func_id(dev, func_id);
 111         if (IS_ERR(root))
 112                 return PTR_ERR(root);
 113
 114         new = &root->rb_node;
 115
 116         while (*new) {
 117                 parent = *new;
 118                 tfp = rb_entry(parent, struct fw_page, rb_node);
 119                 if (tfp->addr < addr)
 120                         new = &parent->rb_left;
 121                 else if (tfp->addr > addr)
 122                         new = &parent->rb_right;
 123                 else
 124                         return -EEXIST;
 125         }
 126
 127         nfp = kzalloc(sizeof(*nfp), GFP_KERNEL);
 128         if (!nfp)
 129                 return -ENOMEM;
 130
 131         nfp->addr = addr;
 132         nfp->page = page;
 133         nfp->func_id = func_id;
 134         nfp->free_count = MLX5_NUM_4K_IN_PAGE;
 135         for (i = 0; i < MLX5_NUM_4K_IN_PAGE; i++)
 136                 set_bit(i, &nfp->bitmask);
 137
 138         rb_link_node(&nfp->rb_node, parent, new);
 139         rb_insert_color(&nfp->rb_node, root);
 140         list_add(&nfp->list, &dev->priv.free_list);
 141
 142         return 0;
 143 }
 144
 145 static struct fw_page *find_fw_page(struct mlx5_core_dev *dev, u64 addr,
 146                                     u32 func_id)
 147 {
 148         struct fw_page *result = NULL;
 149         struct rb_root *root;
 150         struct rb_node *tmp;
 151         struct fw_page *tfp;
 152
 153         root = xa_load(&dev->priv.page_root_xa, func_id);
 154         if (WARN_ON_ONCE(!root))
 155                 return NULL;
 156
 157         tmp = root->rb_node;
 158
 159         while (tmp) {
 160                 tfp = rb_entry(tmp, struct fw_page, rb_node);
 161                 if (tfp->addr < addr) {
 162                         tmp = tmp->rb_left;
 163                 } else if (tfp->addr > addr) {
 164                         tmp = tmp->rb_right;
 165                 } else {
 166                         result = tfp;
 167                         break;
 168                 }
 169         }
 170
 171         return result;
 172 }
 173
 174 static int mlx5_cmd_query_pages(struct mlx5_core_dev *dev, u16 *func_id,
 175                                 s32 *npages, int boot)
 176 {
 177         u32 out[MLX5_ST_SZ_DW(query_pages_out)] = {};
 178         u32 in[MLX5_ST_SZ_DW(query_pages_in)] = {};
 179         int err;
 180
 181         MLX5_SET(query_pages_in, in, opcode, MLX5_CMD_OP_QUERY_PAGES);
 182         MLX5_SET(query_pages_in, in, op_mod, boot ?
 183                  MLX5_QUERY_PAGES_IN_OP_MOD_BOOT_PAGES :
 184                  MLX5_QUERY_PAGES_IN_OP_MOD_INIT_PAGES);
 185         MLX5_SET(query_pages_in, in, embedded_cpu_function, mlx5_core_is_ecpf(dev));
 186
 187         err = mlx5_cmd_exec_inout(dev, query_pages, in, out);
 188         if (err)
 189                 return err;
 190
 191         *npages = MLX5_GET(query_pages_out, out, num_pages);
 192         *func_id = MLX5_GET(query_pages_out, out, function_id);
 193
 194         return err;
 195 }
 196
 197 static int alloc_4k(struct mlx5_core_dev *dev, u64 *addr, u16 func_id)
 198 {
 199         struct fw_page *fp = NULL;
 200         struct fw_page *iter;
 201         unsigned n;
 202
 203         list_for_each_entry(iter, &dev->priv.free_list, list) {
 204                 if (iter->func_id != func_id)
 205                         continue;
 206                 fp = iter;
 207         }
 208
 209         if (list_empty(&dev->priv.free_list) || !fp)
 210                 return -ENOMEM;
 211
 212         n = find_first_bit(&fp->bitmask, 8 * sizeof(fp->bitmask));
 213         if (n >= MLX5_NUM_4K_IN_PAGE) {
 214                 mlx5_core_warn(dev, "alloc 4k bug\n");
 215                 return -ENOENT;
 216         }
 217         clear_bit(n, &fp->bitmask);
 218         fp->free_count--;
 219         if (!fp->free_count)
 220                 list_del(&fp->list);
 221
 222         *addr = fp->addr + n * MLX5_ADAPTER_PAGE_SIZE;
 223
 224         return 0;
 225 }
 226
 227 #define MLX5_U64_4K_PAGE_MASK ((~(u64)0U) << PAGE_SHIFT)
 228
 229 static void free_fwp(struct mlx5_core_dev *dev, struct fw_page *fwp,
 230                      bool in_free_list)
 231 {
 232         struct rb_root *root;
 233
 234         root = xa_load(&dev->priv.page_root_xa, fwp->func_id);
 235         if (WARN_ON_ONCE(!root))
 236                 return;
 237
 238         rb_erase(&fwp->rb_node, root);
 239         if (in_free_list)
 240                 list_del(&fwp->list);
 241         dma_unmap_page(mlx5_core_dma_dev(dev), fwp->addr & MLX5_U64_4K_PAGE_MASK,
 242                        PAGE_SIZE, DMA_BIDIRECTIONAL);
 243         __free_page(fwp->page);
 244         kfree(fwp);
 245 }
 246
 247 static void free_4k(struct mlx5_core_dev *dev, u64 addr, u32 func_id)
 248 {
 249         struct fw_page *fwp;
 250         int n;
 251
 252         fwp = find_fw_page(dev, addr & MLX5_U64_4K_PAGE_MASK, func_id);
 253         if (!fwp) {
 254                 mlx5_core_warn_rl(dev, "page not found\n");
 255                 return;
 256         }
 257         n = (addr & ~MLX5_U64_4K_PAGE_MASK) >> MLX5_ADAPTER_PAGE_SHIFT;
 258         fwp->free_count++;
 259         set_bit(n, &fwp->bitmask);
 260         if (fwp->free_count == MLX5_NUM_4K_IN_PAGE)
 261                 free_fwp(dev, fwp, fwp->free_count != 1);
 262         else if (fwp->free_count == 1)
 263                 list_add(&fwp->list, &dev->priv.free_list);
 264 }
 265
 266 static int alloc_system_page(struct mlx5_core_dev *dev, u16 func_id)
 267 {
 268         struct device *device = mlx5_core_dma_dev(dev);
 269         int nid = dev_to_node(device);
 270         struct page *page;
 271         u64 zero_addr = 1;
 272         u64 addr;
 273         int err;
 274
 275         page = alloc_pages_node(nid, GFP_HIGHUSER, 0);
 276         if (!page) {
 277                 mlx5_core_warn(dev, "failed to allocate page\n");
 278                 return -ENOMEM;
 279         }
 280 map:
 281         addr = dma_map_page(device, page, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
 282         if (dma_mapping_error(device, addr)) {
 283                 mlx5_core_warn(dev, "failed dma mapping page\n");
 284                 err = -ENOMEM;
 285                 goto err_mapping;
 286         }
 287
 288         /* Firmware doesn't support page with physical address 0 */
 289         if (addr == 0) {
 290                 zero_addr = addr;
 291                 goto map;
 292         }
 293
 294         err = insert_page(dev, addr, page, func_id);
 295         if (err) {
 296                 mlx5_core_err(dev, "failed to track allocated page\n");
 297                 dma_unmap_page(device, addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
 298         }
 299
 300 err_mapping:
 301         if (err)
 302                 __free_page(page);
 303
 304         if (zero_addr == 0)
 305                 dma_unmap_page(device, zero_addr, PAGE_SIZE,
 306                                DMA_BIDIRECTIONAL);
 307
 308         return err;
 309 }
 310
 311 static void page_notify_fail(struct mlx5_core_dev *dev, u16 func_id,
 312                              bool ec_function)
 313 {
 314         u32 in[MLX5_ST_SZ_DW(manage_pages_in)] = {};
 315         int err;
 316
 317         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 318         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_CANT_GIVE);
 319         MLX5_SET(manage_pages_in, in, function_id, func_id);
 320         MLX5_SET(manage_pages_in, in, embedded_cpu_function, ec_function);
 321
 322         err = mlx5_cmd_exec_in(dev, manage_pages, in);
 323         if (err)
 324                 mlx5_core_warn(dev, "page notify failed func_id(%d) err(%d)\n",
 325                                func_id, err);
 326 }
 327
 328 static int give_pages(struct mlx5_core_dev *dev, u16 func_id, int npages,
 329                       int notify_fail, bool ec_function)
 330 {
 331         u32 out[MLX5_ST_SZ_DW(manage_pages_out)] = {0};
 332         int inlen = MLX5_ST_SZ_BYTES(manage_pages_in);
 333         u64 addr;
 334         int err;
 335         u32 *in;
 336         int i;
 337
 338         inlen += npages * MLX5_FLD_SZ_BYTES(manage_pages_in, pas[0]);
 339         in = kvzalloc(inlen, GFP_KERNEL);
 340         if (!in) {
 341                 err = -ENOMEM;
 342                 mlx5_core_warn(dev, "vzalloc failed %d\n", inlen);
 343                 goto out_free;
 344         }
 345
 346         for (i = 0; i < npages; i++) {
 347 retry:
 348                 err = alloc_4k(dev, &addr, func_id);
 349                 if (err) {
 350                         if (err == -ENOMEM)
 351                                 err = alloc_system_page(dev, func_id);
 352                         if (err)
 353                                 goto out_4k;
 354
 355                         goto retry;
 356                 }
 357                 MLX5_ARRAY_SET64(manage_pages_in, in, pas, i, addr);
 358         }
 359
 360         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 361         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_GIVE);
 362         MLX5_SET(manage_pages_in, in, function_id, func_id);
 363         MLX5_SET(manage_pages_in, in, input_num_entries, npages);
 364         MLX5_SET(manage_pages_in, in, embedded_cpu_function, ec_function);
 365
 366         err = mlx5_cmd_exec(dev, in, inlen, out, sizeof(out));
 367         if (err) {
 368                 mlx5_core_warn(dev, "func_id 0x%x, npages %d, err %d\n",
 369                                func_id, npages, err);
 370                 goto out_4k;
 371         }
 372
 373         dev->priv.fw_pages += npages;
 374         if (func_id)
 375                 dev->priv.vfs_pages += npages;
 376         else if (mlx5_core_is_ecpf(dev) && !ec_function)
 377                 dev->priv.peer_pf_pages += npages;
 378
 379         mlx5_core_dbg(dev, "npages %d, ec_function %d, func_id 0x%x, err %d\n",
 380                       npages, ec_function, func_id, err);
 381
 382         kvfree(in);
 383         return 0;
 384
 385 out_4k:
 386         for (i--; i >= 0; i--)
 387                 free_4k(dev, MLX5_GET64(manage_pages_in, in, pas[i]), func_id);
 388 out_free:
 389         kvfree(in);
 390         if (notify_fail)
 391                 page_notify_fail(dev, func_id, ec_function);
 392         return err;
 393 }
 394
 395 static void release_all_pages(struct mlx5_core_dev *dev, u32 func_id,
 396                               bool ec_function)
 397 {
 398         struct rb_root *root;
 399         struct rb_node *p;
 400         int npages = 0;
 401
 402         root = xa_load(&dev->priv.page_root_xa, func_id);
 403         if (WARN_ON_ONCE(!root))
 404                 return;
 405
 406         p = rb_first(root);
 407         while (p) {
 408                 struct fw_page *fwp = rb_entry(p, struct fw_page, rb_node);
 409
 410                 p = rb_next(p);
 411                 npages += (MLX5_NUM_4K_IN_PAGE - fwp->free_count);
 412                 free_fwp(dev, fwp, fwp->free_count);
 413         }
 414
 415         dev->priv.fw_pages -= npages;
 416         if (func_id)
 417                 dev->priv.vfs_pages -= npages;
 418         else if (mlx5_core_is_ecpf(dev) && !ec_function)
 419                 dev->priv.peer_pf_pages -= npages;
 420
 421         mlx5_core_dbg(dev, "npages %d, ec_function %d, func_id 0x%x\n",
 422                       npages, ec_function, func_id);
 423 }
 424
 425 static int reclaim_pages_cmd(struct mlx5_core_dev *dev,
 426                              u32 *in, int in_size, u32 *out, int out_size)
 427 {
 428         struct rb_root *root;
 429         struct fw_page *fwp;
 430         struct rb_node *p;
 431         u32 func_id;
 432         u32 npages;
 433         u32 i = 0;
 434
 435         if (!mlx5_cmd_is_down(dev))
 436                 return mlx5_cmd_exec(dev, in, in_size, out, out_size);
 437
 438         /* No hard feelings, we want our pages back! */
 439         npages = MLX5_GET(manage_pages_in, in, input_num_entries);
 440         func_id = MLX5_GET(manage_pages_in, in, function_id);
 441
 442         root = xa_load(&dev->priv.page_root_xa, func_id);
 443         if (WARN_ON_ONCE(!root))
 444                 return -EEXIST;
 445
 446         p = rb_first(root);
 447         while (p && i < npages) {
 448                 fwp = rb_entry(p, struct fw_page, rb_node);
 449                 p = rb_next(p);
 450
 451                 MLX5_ARRAY_SET64(manage_pages_out, out, pas, i, fwp->addr);
 452                 i++;
 453         }
 454
 455         MLX5_SET(manage_pages_out, out, output_num_entries, i);
 456         return 0;
 457 }
 458
 459 static int reclaim_pages(struct mlx5_core_dev *dev, u32 func_id, int npages,
 460                          int *nclaimed, bool ec_function)
 461 {
 462         int outlen = MLX5_ST_SZ_BYTES(manage_pages_out);
 463         u32 in[MLX5_ST_SZ_DW(manage_pages_in)] = {};
 464         int num_claimed;
 465         u32 *out;
 466         int err;
 467         int i;
 468
 469         if (nclaimed)
 470                 *nclaimed = 0;
 471
 472         outlen += npages * MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
 473         out = kvzalloc(outlen, GFP_KERNEL);
 474         if (!out)
 475                 return -ENOMEM;
 476
 477         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 478         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_TAKE);
 479         MLX5_SET(manage_pages_in, in, function_id, func_id);
 480         MLX5_SET(manage_pages_in, in, input_num_entries, npages);
 481         MLX5_SET(manage_pages_in, in, embedded_cpu_function, ec_function);
 482
 483         mlx5_core_dbg(dev, "func 0x%x, npages %d, outlen %d\n",
 484                       func_id, npages, outlen);
 485         err = reclaim_pages_cmd(dev, in, sizeof(in), out, outlen);
 486         if (err) {
 487                 mlx5_core_err(dev, "failed reclaiming pages: err %d\n", err);
 488                 goto out_free;
 489         }
 490
 491         num_claimed = MLX5_GET(manage_pages_out, out, output_num_entries);
 492         if (num_claimed > npages) {
 493                 mlx5_core_warn(dev, "fw returned %d, driver asked %d => corruption\n",
 494                                num_claimed, npages);
 495                 err = -EINVAL;
 496                 goto out_free;
 497         }
 498
 499         for (i = 0; i < num_claimed; i++)
 500                 free_4k(dev, MLX5_GET64(manage_pages_out, out, pas[i]), func_id);
 501
 502         if (nclaimed)
 503                 *nclaimed = num_claimed;
 504
 505         dev->priv.fw_pages -= num_claimed;
 506         if (func_id)
 507                 dev->priv.vfs_pages -= num_claimed;
 508         else if (mlx5_core_is_ecpf(dev) && !ec_function)
 509                 dev->priv.peer_pf_pages -= num_claimed;
 510
 511 out_free:
 512         kvfree(out);
 513         return err;
 514 }
 515
 516 static void pages_work_handler(struct work_struct *work)
 517 {
 518         struct mlx5_pages_req *req = container_of(work, struct mlx5_pages_req, work);
 519         struct mlx5_core_dev *dev = req->dev;
 520         int err = 0;
 521
 522         if (req->release_all)
 523                 release_all_pages(dev, req->func_id, req->ec_function);
 524         else if (req->npages < 0)
 525                 err = reclaim_pages(dev, req->func_id, -1 * req->npages, NULL,
 526                                     req->ec_function);
 527         else if (req->npages > 0)
 528                 err = give_pages(dev, req->func_id, req->npages, 1, req->ec_function);
 529
 530         if (err)
 531                 mlx5_core_warn(dev, "%s fail %d\n",
 532                                req->npages < 0 ? "reclaim" : "give", err);
 533
 534         kfree(req);
 535 }
 536
 537 enum {
 538         EC_FUNCTION_MASK = 0x8000,
 539         RELEASE_ALL_PAGES_MASK = 0x4000,
 540 };
 541
 542 static int req_pages_handler(struct notifier_block *nb,
 543                              unsigned long type, void *data)
 544 {
 545         struct mlx5_pages_req *req;
 546         struct mlx5_core_dev *dev;
 547         struct mlx5_priv *priv;
 548         struct mlx5_eqe *eqe;
 549         bool ec_function;
 550         bool release_all;
 551         u16 func_id;
 552         s32 npages;
 553
 554         priv = mlx5_nb_cof(nb, struct mlx5_priv, pg_nb);
 555         dev  = container_of(priv, struct mlx5_core_dev, priv);
 556         eqe  = data;
 557
 558         func_id = be16_to_cpu(eqe->data.req_pages.func_id);
 559         npages  = be32_to_cpu(eqe->data.req_pages.num_pages);
 560         ec_function = be16_to_cpu(eqe->data.req_pages.ec_function) & EC_FUNCTION_MASK;
 561         release_all = be16_to_cpu(eqe->data.req_pages.ec_function) &
 562                       RELEASE_ALL_PAGES_MASK;
 563         mlx5_core_dbg(dev, "page request for func 0x%x, npages %d, release_all %d\n",
 564                       func_id, npages, release_all);
 565         req = kzalloc(sizeof(*req), GFP_ATOMIC);
 566         if (!req) {
 567                 mlx5_core_warn(dev, "failed to allocate pages request\n");
 568                 return NOTIFY_DONE;
 569         }
 570
 571         req->dev = dev;
 572         req->func_id = func_id;
 573         req->npages = npages;
 574         req->ec_function = ec_function;
 575         req->release_all = release_all;
 576         INIT_WORK(&req->work, pages_work_handler);
 577         queue_work(dev->priv.pg_wq, &req->work);
 578         return NOTIFY_OK;
 579 }
 580
 581 int mlx5_satisfy_startup_pages(struct mlx5_core_dev *dev, int boot)
 582 {
 583         u16 func_id;
 584         s32 npages;
 585         int err;
 586
 587         err = mlx5_cmd_query_pages(dev, &func_id, &npages, boot);
 588         if (err)
 589                 return err;
 590
 591         mlx5_core_dbg(dev, "requested %d %s pages for func_id 0x%x\n",
 592                       npages, boot ? "boot" : "init", func_id);
 593
 594         return give_pages(dev, func_id, npages, 0, mlx5_core_is_ecpf(dev));
 595 }
 596
 597 enum {
 598         MLX5_BLKS_FOR_RECLAIM_PAGES = 12
 599 };
 600
 601 static int optimal_reclaimed_pages(void)
 602 {
 603         struct mlx5_cmd_prot_block *block;
 604         struct mlx5_cmd_layout *lay;
 605         int ret;
 606
 607         ret = (sizeof(lay->out) + MLX5_BLKS_FOR_RECLAIM_PAGES * sizeof(block->data) -
 608                MLX5_ST_SZ_BYTES(manage_pages_out)) /
 609                MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
 610
 611         return ret;
 612 }
 613
 614 static int mlx5_reclaim_root_pages(struct mlx5_core_dev *dev,
 615                                    struct rb_root *root, u16 func_id)
 616 {
 617         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 618
 619         while (!RB_EMPTY_ROOT(root)) {
 620                 int nclaimed;
 621                 int err;
 622
 623                 err = reclaim_pages(dev, func_id, optimal_reclaimed_pages(),
 624                                     &nclaimed, mlx5_core_is_ecpf(dev));
 625                 if (err) {
 626                         mlx5_core_warn(dev, "failed reclaiming pages (%d) for func id 0x%x\n",
 627                                        err, func_id);
 628                         return err;
 629                 }
 630
 631                 if (nclaimed)
 632                         end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 633
 634                 if (time_after(jiffies, end)) {
 635                         mlx5_core_warn(dev, "FW did not return all pages. giving up...\n");
 636                         break;
 637                 }
 638         }
 639
 640         return 0;
 641 }
 642
 643 int mlx5_reclaim_startup_pages(struct mlx5_core_dev *dev)
 644 {
 645         struct rb_root *root;
 646         unsigned long id;
 647         void *entry;
 648
 649         xa_for_each(&dev->priv.page_root_xa, id, entry) {
 650                 root = entry;
 651                 mlx5_reclaim_root_pages(dev, root, id);
 652                 xa_erase(&dev->priv.page_root_xa, id);
 653                 kfree(root);
 654         }
 655
 656         WARN_ON(!xa_empty(&dev->priv.page_root_xa));
 657
 658         WARN(dev->priv.fw_pages,
 659              "FW pages counter is %d after reclaiming all pages\n",
 660              dev->priv.fw_pages);
 661         WARN(dev->priv.vfs_pages,
 662              "VFs FW pages counter is %d after reclaiming all pages\n",
 663              dev->priv.vfs_pages);
 664         WARN(dev->priv.peer_pf_pages,
 665              "Peer PF FW pages counter is %d after reclaiming all pages\n",
 666              dev->priv.peer_pf_pages);
 667
 668         return 0;
 669 }
 670
 671 int mlx5_pagealloc_init(struct mlx5_core_dev *dev)
 672 {
 673         INIT_LIST_HEAD(&dev->priv.free_list);
 674         dev->priv.pg_wq = create_singlethread_workqueue("mlx5_page_allocator");
 675         if (!dev->priv.pg_wq)
 676                 return -ENOMEM;
 677
 678         xa_init(&dev->priv.page_root_xa);
 679
 680         return 0;
 681 }
 682
 683 void mlx5_pagealloc_cleanup(struct mlx5_core_dev *dev)
 684 {
 685         xa_destroy(&dev->priv.page_root_xa);
 686         destroy_workqueue(dev->priv.pg_wq);
 687 }
 688
 689 void mlx5_pagealloc_start(struct mlx5_core_dev *dev)
 690 {
 691         MLX5_NB_INIT(&dev->priv.pg_nb, req_pages_handler, PAGE_REQUEST);
 692         mlx5_eq_notifier_register(dev, &dev->priv.pg_nb);
 693 }
 694
 695 void mlx5_pagealloc_stop(struct mlx5_core_dev *dev)
 696 {
 697         mlx5_eq_notifier_unregister(dev, &dev->priv.pg_nb);
 698         flush_workqueue(dev->priv.pg_wq);
 699 }
 700
 701 int mlx5_wait_for_pages(struct mlx5_core_dev *dev, int *pages)
 702 {
 703         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_VFS_PAGES_TIME_MSECS);
 704         int prev_pages = *pages;
 705
 706         /* In case of internal error we will free the pages manually later */
 707         if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
 708                 mlx5_core_warn(dev, "Skipping wait for vf pages stage");
 709                 return 0;
 710         }
 711
 712         mlx5_core_dbg(dev, "Waiting for %d pages\n", prev_pages);
 713         while (*pages) {
 714                 if (time_after(jiffies, end)) {
 715                         mlx5_core_warn(dev, "aborting while there are %d pending pages\n", *pages);
 716                         return -ETIMEDOUT;
 717                 }
 718                 if (*pages < prev_pages) {
 719                         end = jiffies + msecs_to_jiffies(MAX_RECLAIM_VFS_PAGES_TIME_MSECS);
 720                         prev_pages = *pages;
 721                 }
 722                 msleep(50);
 723         }
 724
 725         mlx5_core_dbg(dev, "All pages received\n");
 726         return 0;
 727 }