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 u32 fwp_fill_manage_pages_out(struct fw_page *fwp, u32 *out, u32 index,
 426                                      u32 npages)
 427 {
 428         u32 pages_set = 0;
 429         unsigned int n;
 430
 431         for_each_clear_bit(n, &fwp->bitmask, MLX5_NUM_4K_IN_PAGE) {
 432                 MLX5_ARRAY_SET64(manage_pages_out, out, pas, index + pages_set,
 433                                  fwp->addr + (n * MLX5_ADAPTER_PAGE_SIZE));
 434                 pages_set++;
 435
 436                 if (!--npages)
 437                         break;
 438         }
 439
 440         return pages_set;
 441 }
 442
 443 static int reclaim_pages_cmd(struct mlx5_core_dev *dev,
 444                              u32 *in, int in_size, u32 *out, int out_size)
 445 {
 446         struct rb_root *root;
 447         struct fw_page *fwp;
 448         struct rb_node *p;
 449         u32 func_id;
 450         u32 npages;
 451         u32 i = 0;
 452
 453         if (!mlx5_cmd_is_down(dev))
 454                 return mlx5_cmd_exec(dev, in, in_size, out, out_size);
 455
 456         /* No hard feelings, we want our pages back! */
 457         npages = MLX5_GET(manage_pages_in, in, input_num_entries);
 458         func_id = MLX5_GET(manage_pages_in, in, function_id);
 459
 460         root = xa_load(&dev->priv.page_root_xa, func_id);
 461         if (WARN_ON_ONCE(!root))
 462                 return -EEXIST;
 463
 464         p = rb_first(root);
 465         while (p && i < npages) {
 466                 fwp = rb_entry(p, struct fw_page, rb_node);
 467                 p = rb_next(p);
 468
 469                 i += fwp_fill_manage_pages_out(fwp, out, i, npages - i);
 470         }
 471
 472         MLX5_SET(manage_pages_out, out, output_num_entries, i);
 473         return 0;
 474 }
 475
 476 static int reclaim_pages(struct mlx5_core_dev *dev, u32 func_id, int npages,
 477                          int *nclaimed, bool ec_function)
 478 {
 479         int outlen = MLX5_ST_SZ_BYTES(manage_pages_out);
 480         u32 in[MLX5_ST_SZ_DW(manage_pages_in)] = {};
 481         int num_claimed;
 482         u32 *out;
 483         int err;
 484         int i;
 485
 486         if (nclaimed)
 487                 *nclaimed = 0;
 488
 489         outlen += npages * MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
 490         out = kvzalloc(outlen, GFP_KERNEL);
 491         if (!out)
 492                 return -ENOMEM;
 493
 494         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 495         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_TAKE);
 496         MLX5_SET(manage_pages_in, in, function_id, func_id);
 497         MLX5_SET(manage_pages_in, in, input_num_entries, npages);
 498         MLX5_SET(manage_pages_in, in, embedded_cpu_function, ec_function);
 499
 500         mlx5_core_dbg(dev, "func 0x%x, npages %d, outlen %d\n",
 501                       func_id, npages, outlen);
 502         err = reclaim_pages_cmd(dev, in, sizeof(in), out, outlen);
 503         if (err) {
 504                 mlx5_core_err(dev, "failed reclaiming pages: err %d\n", err);
 505                 goto out_free;
 506         }
 507
 508         num_claimed = MLX5_GET(manage_pages_out, out, output_num_entries);
 509         if (num_claimed > npages) {
 510                 mlx5_core_warn(dev, "fw returned %d, driver asked %d => corruption\n",
 511                                num_claimed, npages);
 512                 err = -EINVAL;
 513                 goto out_free;
 514         }
 515
 516         for (i = 0; i < num_claimed; i++)
 517                 free_4k(dev, MLX5_GET64(manage_pages_out, out, pas[i]), func_id);
 518
 519         if (nclaimed)
 520                 *nclaimed = num_claimed;
 521
 522         dev->priv.fw_pages -= num_claimed;
 523         if (func_id)
 524                 dev->priv.vfs_pages -= num_claimed;
 525         else if (mlx5_core_is_ecpf(dev) && !ec_function)
 526                 dev->priv.peer_pf_pages -= num_claimed;
 527
 528 out_free:
 529         kvfree(out);
 530         return err;
 531 }
 532
 533 static void pages_work_handler(struct work_struct *work)
 534 {
 535         struct mlx5_pages_req *req = container_of(work, struct mlx5_pages_req, work);
 536         struct mlx5_core_dev *dev = req->dev;
 537         int err = 0;
 538
 539         if (req->release_all)
 540                 release_all_pages(dev, req->func_id, req->ec_function);
 541         else if (req->npages < 0)
 542                 err = reclaim_pages(dev, req->func_id, -1 * req->npages, NULL,
 543                                     req->ec_function);
 544         else if (req->npages > 0)
 545                 err = give_pages(dev, req->func_id, req->npages, 1, req->ec_function);
 546
 547         if (err)
 548                 mlx5_core_warn(dev, "%s fail %d\n",
 549                                req->npages < 0 ? "reclaim" : "give", err);
 550
 551         kfree(req);
 552 }
 553
 554 enum {
 555         EC_FUNCTION_MASK = 0x8000,
 556         RELEASE_ALL_PAGES_MASK = 0x4000,
 557 };
 558
 559 static int req_pages_handler(struct notifier_block *nb,
 560                              unsigned long type, void *data)
 561 {
 562         struct mlx5_pages_req *req;
 563         struct mlx5_core_dev *dev;
 564         struct mlx5_priv *priv;
 565         struct mlx5_eqe *eqe;
 566         bool ec_function;
 567         bool release_all;
 568         u16 func_id;
 569         s32 npages;
 570
 571         priv = mlx5_nb_cof(nb, struct mlx5_priv, pg_nb);
 572         dev  = container_of(priv, struct mlx5_core_dev, priv);
 573         eqe  = data;
 574
 575         func_id = be16_to_cpu(eqe->data.req_pages.func_id);
 576         npages  = be32_to_cpu(eqe->data.req_pages.num_pages);
 577         ec_function = be16_to_cpu(eqe->data.req_pages.ec_function) & EC_FUNCTION_MASK;
 578         release_all = be16_to_cpu(eqe->data.req_pages.ec_function) &
 579                       RELEASE_ALL_PAGES_MASK;
 580         mlx5_core_dbg(dev, "page request for func 0x%x, npages %d, release_all %d\n",
 581                       func_id, npages, release_all);
 582         req = kzalloc(sizeof(*req), GFP_ATOMIC);
 583         if (!req) {
 584                 mlx5_core_warn(dev, "failed to allocate pages request\n");
 585                 return NOTIFY_DONE;
 586         }
 587
 588         req->dev = dev;
 589         req->func_id = func_id;
 590         req->npages = npages;
 591         req->ec_function = ec_function;
 592         req->release_all = release_all;
 593         INIT_WORK(&req->work, pages_work_handler);
 594         queue_work(dev->priv.pg_wq, &req->work);
 595         return NOTIFY_OK;
 596 }
 597
 598 int mlx5_satisfy_startup_pages(struct mlx5_core_dev *dev, int boot)
 599 {
 600         u16 func_id;
 601         s32 npages;
 602         int err;
 603
 604         err = mlx5_cmd_query_pages(dev, &func_id, &npages, boot);
 605         if (err)
 606                 return err;
 607
 608         mlx5_core_dbg(dev, "requested %d %s pages for func_id 0x%x\n",
 609                       npages, boot ? "boot" : "init", func_id);
 610
 611         return give_pages(dev, func_id, npages, 0, mlx5_core_is_ecpf(dev));
 612 }
 613
 614 enum {
 615         MLX5_BLKS_FOR_RECLAIM_PAGES = 12
 616 };
 617
 618 static int optimal_reclaimed_pages(void)
 619 {
 620         struct mlx5_cmd_prot_block *block;
 621         struct mlx5_cmd_layout *lay;
 622         int ret;
 623
 624         ret = (sizeof(lay->out) + MLX5_BLKS_FOR_RECLAIM_PAGES * sizeof(block->data) -
 625                MLX5_ST_SZ_BYTES(manage_pages_out)) /
 626                MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
 627
 628         return ret;
 629 }
 630
 631 static int mlx5_reclaim_root_pages(struct mlx5_core_dev *dev,
 632                                    struct rb_root *root, u16 func_id)
 633 {
 634         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 635
 636         while (!RB_EMPTY_ROOT(root)) {
 637                 int nclaimed;
 638                 int err;
 639
 640                 err = reclaim_pages(dev, func_id, optimal_reclaimed_pages(),
 641                                     &nclaimed, mlx5_core_is_ecpf(dev));
 642                 if (err) {
 643                         mlx5_core_warn(dev, "failed reclaiming pages (%d) for func id 0x%x\n",
 644                                        err, func_id);
 645                         return err;
 646                 }
 647
 648                 if (nclaimed)
 649                         end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 650
 651                 if (time_after(jiffies, end)) {
 652                         mlx5_core_warn(dev, "FW did not return all pages. giving up...\n");
 653                         break;
 654                 }
 655         }
 656
 657         return 0;
 658 }
 659
 660 int mlx5_reclaim_startup_pages(struct mlx5_core_dev *dev)
 661 {
 662         struct rb_root *root;
 663         unsigned long id;
 664         void *entry;
 665
 666         xa_for_each(&dev->priv.page_root_xa, id, entry) {
 667                 root = entry;
 668                 mlx5_reclaim_root_pages(dev, root, id);
 669                 xa_erase(&dev->priv.page_root_xa, id);
 670                 kfree(root);
 671         }
 672
 673         WARN_ON(!xa_empty(&dev->priv.page_root_xa));
 674
 675         WARN(dev->priv.fw_pages,
 676              "FW pages counter is %d after reclaiming all pages\n",
 677              dev->priv.fw_pages);
 678         WARN(dev->priv.vfs_pages,
 679              "VFs FW pages counter is %d after reclaiming all pages\n",
 680              dev->priv.vfs_pages);
 681         WARN(dev->priv.peer_pf_pages,
 682              "Peer PF FW pages counter is %d after reclaiming all pages\n",
 683              dev->priv.peer_pf_pages);
 684
 685         return 0;
 686 }
 687
 688 int mlx5_pagealloc_init(struct mlx5_core_dev *dev)
 689 {
 690         INIT_LIST_HEAD(&dev->priv.free_list);
 691         dev->priv.pg_wq = create_singlethread_workqueue("mlx5_page_allocator");
 692         if (!dev->priv.pg_wq)
 693                 return -ENOMEM;
 694
 695         xa_init(&dev->priv.page_root_xa);
 696
 697         return 0;
 698 }
 699
 700 void mlx5_pagealloc_cleanup(struct mlx5_core_dev *dev)
 701 {
 702         xa_destroy(&dev->priv.page_root_xa);
 703         destroy_workqueue(dev->priv.pg_wq);
 704 }
 705
 706 void mlx5_pagealloc_start(struct mlx5_core_dev *dev)
 707 {
 708         MLX5_NB_INIT(&dev->priv.pg_nb, req_pages_handler, PAGE_REQUEST);
 709         mlx5_eq_notifier_register(dev, &dev->priv.pg_nb);
 710 }
 711
 712 void mlx5_pagealloc_stop(struct mlx5_core_dev *dev)
 713 {
 714         mlx5_eq_notifier_unregister(dev, &dev->priv.pg_nb);
 715         flush_workqueue(dev->priv.pg_wq);
 716 }
 717
 718 int mlx5_wait_for_pages(struct mlx5_core_dev *dev, int *pages)
 719 {
 720         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_VFS_PAGES_TIME_MSECS);
 721         int prev_pages = *pages;
 722
 723         /* In case of internal error we will free the pages manually later */
 724         if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
 725                 mlx5_core_warn(dev, "Skipping wait for vf pages stage");
 726                 return 0;
 727         }
 728
 729         mlx5_core_dbg(dev, "Waiting for %d pages\n", prev_pages);
 730         while (*pages) {
 731                 if (time_after(jiffies, end)) {
 732                         mlx5_core_warn(dev, "aborting while there are %d pending pages\n", *pages);
 733                         return -ETIMEDOUT;
 734                 }
 735                 if (*pages < prev_pages) {
 736                         end = jiffies + msecs_to_jiffies(MAX_RECLAIM_VFS_PAGES_TIME_MSECS);
 737                         prev_pages = *pages;
 738                 }
 739                 msleep(50);
 740         }
 741
 742         mlx5_core_dbg(dev, "All pages received\n");
 743         return 0;
 744 }