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         u32                     function;
  62         unsigned long           bitmask;
  63         struct list_head        list;
  64         unsigned int 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 u32 get_function(u16 func_id, bool ec_function)
  78 {
  79         return (u32)func_id | (ec_function << 16);
  80 }
  81
  82 static struct rb_root *page_root_per_function(struct mlx5_core_dev *dev, u32 function)
  83 {
  84         struct rb_root *root;
  85         int err;
  86
  87         root = xa_load(&dev->priv.page_root_xa, function);
  88         if (root)
  89                 return root;
  90
  91         root = kzalloc(sizeof(*root), GFP_KERNEL);
  92         if (!root)
  93                 return ERR_PTR(-ENOMEM);
  94
  95         err = xa_insert(&dev->priv.page_root_xa, function, root, GFP_KERNEL);
  96         if (err) {
  97                 kfree(root);
  98                 return ERR_PTR(err);
  99         }
 100
 101         *root = RB_ROOT;
 102
 103         return root;
 104 }
 105
 106 static int insert_page(struct mlx5_core_dev *dev, u64 addr, struct page *page, u32 function)
 107 {
 108         struct rb_node *parent = NULL;
 109         struct rb_root *root;
 110         struct rb_node **new;
 111         struct fw_page *nfp;
 112         struct fw_page *tfp;
 113         int i;
 114
 115         root = page_root_per_function(dev, function);
 116         if (IS_ERR(root))
 117                 return PTR_ERR(root);
 118
 119         new = &root->rb_node;
 120
 121         while (*new) {
 122                 parent = *new;
 123                 tfp = rb_entry(parent, struct fw_page, rb_node);
 124                 if (tfp->addr < addr)
 125                         new = &parent->rb_left;
 126                 else if (tfp->addr > addr)
 127                         new = &parent->rb_right;
 128                 else
 129                         return -EEXIST;
 130         }
 131
 132         nfp = kzalloc(sizeof(*nfp), GFP_KERNEL);
 133         if (!nfp)
 134                 return -ENOMEM;
 135
 136         nfp->addr = addr;
 137         nfp->page = page;
 138         nfp->function = function;
 139         nfp->free_count = MLX5_NUM_4K_IN_PAGE;
 140         for (i = 0; i < MLX5_NUM_4K_IN_PAGE; i++)
 141                 set_bit(i, &nfp->bitmask);
 142
 143         rb_link_node(&nfp->rb_node, parent, new);
 144         rb_insert_color(&nfp->rb_node, root);
 145         list_add(&nfp->list, &dev->priv.free_list);
 146
 147         return 0;
 148 }
 149
 150 static struct fw_page *find_fw_page(struct mlx5_core_dev *dev, u64 addr,
 151                                     u32 function)
 152 {
 153         struct fw_page *result = NULL;
 154         struct rb_root *root;
 155         struct rb_node *tmp;
 156         struct fw_page *tfp;
 157
 158         root = xa_load(&dev->priv.page_root_xa, function);
 159         if (WARN_ON_ONCE(!root))
 160                 return NULL;
 161
 162         tmp = root->rb_node;
 163
 164         while (tmp) {
 165                 tfp = rb_entry(tmp, struct fw_page, rb_node);
 166                 if (tfp->addr < addr) {
 167                         tmp = tmp->rb_left;
 168                 } else if (tfp->addr > addr) {
 169                         tmp = tmp->rb_right;
 170                 } else {
 171                         result = tfp;
 172                         break;
 173                 }
 174         }
 175
 176         return result;
 177 }
 178
 179 static int mlx5_cmd_query_pages(struct mlx5_core_dev *dev, u16 *func_id,
 180                                 s32 *npages, int boot)
 181 {
 182         u32 out[MLX5_ST_SZ_DW(query_pages_out)] = {};
 183         u32 in[MLX5_ST_SZ_DW(query_pages_in)] = {};
 184         int err;
 185
 186         MLX5_SET(query_pages_in, in, opcode, MLX5_CMD_OP_QUERY_PAGES);
 187         MLX5_SET(query_pages_in, in, op_mod, boot ?
 188                  MLX5_QUERY_PAGES_IN_OP_MOD_BOOT_PAGES :
 189                  MLX5_QUERY_PAGES_IN_OP_MOD_INIT_PAGES);
 190         MLX5_SET(query_pages_in, in, embedded_cpu_function, mlx5_core_is_ecpf(dev));
 191
 192         err = mlx5_cmd_exec_inout(dev, query_pages, in, out);
 193         if (err)
 194                 return err;
 195
 196         *npages = MLX5_GET(query_pages_out, out, num_pages);
 197         *func_id = MLX5_GET(query_pages_out, out, function_id);
 198
 199         return err;
 200 }
 201
 202 static int alloc_4k(struct mlx5_core_dev *dev, u64 *addr, u32 function)
 203 {
 204         struct fw_page *fp = NULL;
 205         struct fw_page *iter;
 206         unsigned n;
 207
 208         list_for_each_entry(iter, &dev->priv.free_list, list) {
 209                 if (iter->function != function)
 210                         continue;
 211                 fp = iter;
 212         }
 213
 214         if (list_empty(&dev->priv.free_list) || !fp)
 215                 return -ENOMEM;
 216
 217         n = find_first_bit(&fp->bitmask, 8 * sizeof(fp->bitmask));
 218         if (n >= MLX5_NUM_4K_IN_PAGE) {
 219                 mlx5_core_warn(dev, "alloc 4k bug\n");
 220                 return -ENOENT;
 221         }
 222         clear_bit(n, &fp->bitmask);
 223         fp->free_count--;
 224         if (!fp->free_count)
 225                 list_del(&fp->list);
 226
 227         *addr = fp->addr + n * MLX5_ADAPTER_PAGE_SIZE;
 228
 229         return 0;
 230 }
 231
 232 #define MLX5_U64_4K_PAGE_MASK ((~(u64)0U) << PAGE_SHIFT)
 233
 234 static void free_fwp(struct mlx5_core_dev *dev, struct fw_page *fwp,
 235                      bool in_free_list)
 236 {
 237         struct rb_root *root;
 238
 239         root = xa_load(&dev->priv.page_root_xa, fwp->function);
 240         if (WARN_ON_ONCE(!root))
 241                 return;
 242
 243         rb_erase(&fwp->rb_node, root);
 244         if (in_free_list)
 245                 list_del(&fwp->list);
 246         dma_unmap_page(mlx5_core_dma_dev(dev), fwp->addr & MLX5_U64_4K_PAGE_MASK,
 247                        PAGE_SIZE, DMA_BIDIRECTIONAL);
 248         __free_page(fwp->page);
 249         kfree(fwp);
 250 }
 251
 252 static void free_4k(struct mlx5_core_dev *dev, u64 addr, u32 function)
 253 {
 254         struct fw_page *fwp;
 255         int n;
 256
 257         fwp = find_fw_page(dev, addr & MLX5_U64_4K_PAGE_MASK, function);
 258         if (!fwp) {
 259                 mlx5_core_warn_rl(dev, "page not found\n");
 260                 return;
 261         }
 262         n = (addr & ~MLX5_U64_4K_PAGE_MASK) >> MLX5_ADAPTER_PAGE_SHIFT;
 263         fwp->free_count++;
 264         set_bit(n, &fwp->bitmask);
 265         if (fwp->free_count == MLX5_NUM_4K_IN_PAGE)
 266                 free_fwp(dev, fwp, fwp->free_count != 1);
 267         else if (fwp->free_count == 1)
 268                 list_add(&fwp->list, &dev->priv.free_list);
 269 }
 270
 271 static int alloc_system_page(struct mlx5_core_dev *dev, u32 function)
 272 {
 273         struct device *device = mlx5_core_dma_dev(dev);
 274         int nid = dev_to_node(device);
 275         struct page *page;
 276         u64 zero_addr = 1;
 277         u64 addr;
 278         int err;
 279
 280         page = alloc_pages_node(nid, GFP_HIGHUSER, 0);
 281         if (!page) {
 282                 mlx5_core_warn(dev, "failed to allocate page\n");
 283                 return -ENOMEM;
 284         }
 285 map:
 286         addr = dma_map_page(device, page, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
 287         if (dma_mapping_error(device, addr)) {
 288                 mlx5_core_warn(dev, "failed dma mapping page\n");
 289                 err = -ENOMEM;
 290                 goto err_mapping;
 291         }
 292
 293         /* Firmware doesn't support page with physical address 0 */
 294         if (addr == 0) {
 295                 zero_addr = addr;
 296                 goto map;
 297         }
 298
 299         err = insert_page(dev, addr, page, function);
 300         if (err) {
 301                 mlx5_core_err(dev, "failed to track allocated page\n");
 302                 dma_unmap_page(device, addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
 303         }
 304
 305 err_mapping:
 306         if (err)
 307                 __free_page(page);
 308
 309         if (zero_addr == 0)
 310                 dma_unmap_page(device, zero_addr, PAGE_SIZE,
 311                                DMA_BIDIRECTIONAL);
 312
 313         return err;
 314 }
 315
 316 static void page_notify_fail(struct mlx5_core_dev *dev, u16 func_id,
 317                              bool ec_function)
 318 {
 319         u32 in[MLX5_ST_SZ_DW(manage_pages_in)] = {};
 320         int err;
 321
 322         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 323         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_CANT_GIVE);
 324         MLX5_SET(manage_pages_in, in, function_id, func_id);
 325         MLX5_SET(manage_pages_in, in, embedded_cpu_function, ec_function);
 326
 327         err = mlx5_cmd_exec_in(dev, manage_pages, in);
 328         if (err)
 329                 mlx5_core_warn(dev, "page notify failed func_id(%d) err(%d)\n",
 330                                func_id, err);
 331 }
 332
 333 static int give_pages(struct mlx5_core_dev *dev, u16 func_id, int npages,
 334                       int notify_fail, bool ec_function)
 335 {
 336         u32 function = get_function(func_id, ec_function);
 337         u32 out[MLX5_ST_SZ_DW(manage_pages_out)] = {0};
 338         int inlen = MLX5_ST_SZ_BYTES(manage_pages_in);
 339         u64 addr;
 340         int err;
 341         u32 *in;
 342         int i;
 343
 344         inlen += npages * MLX5_FLD_SZ_BYTES(manage_pages_in, pas[0]);
 345         in = kvzalloc(inlen, GFP_KERNEL);
 346         if (!in) {
 347                 err = -ENOMEM;
 348                 mlx5_core_warn(dev, "vzalloc failed %d\n", inlen);
 349                 goto out_free;
 350         }
 351
 352         for (i = 0; i < npages; i++) {
 353 retry:
 354                 err = alloc_4k(dev, &addr, function);
 355                 if (err) {
 356                         if (err == -ENOMEM)
 357                                 err = alloc_system_page(dev, function);
 358                         if (err)
 359                                 goto out_4k;
 360
 361                         goto retry;
 362                 }
 363                 MLX5_ARRAY_SET64(manage_pages_in, in, pas, i, addr);
 364         }
 365
 366         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 367         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_GIVE);
 368         MLX5_SET(manage_pages_in, in, function_id, func_id);
 369         MLX5_SET(manage_pages_in, in, input_num_entries, npages);
 370         MLX5_SET(manage_pages_in, in, embedded_cpu_function, ec_function);
 371
 372         err = mlx5_cmd_exec(dev, in, inlen, out, sizeof(out));
 373         if (err) {
 374                 mlx5_core_warn(dev, "func_id 0x%x, npages %d, err %d\n",
 375                                func_id, npages, err);
 376                 goto out_4k;
 377         }
 378
 379         dev->priv.fw_pages += npages;
 380         if (func_id)
 381                 dev->priv.vfs_pages += npages;
 382         else if (mlx5_core_is_ecpf(dev) && !ec_function)
 383                 dev->priv.host_pf_pages += npages;
 384
 385         mlx5_core_dbg(dev, "npages %d, ec_function %d, func_id 0x%x, err %d\n",
 386                       npages, ec_function, func_id, err);
 387
 388         kvfree(in);
 389         return 0;
 390
 391 out_4k:
 392         for (i--; i >= 0; i--)
 393                 free_4k(dev, MLX5_GET64(manage_pages_in, in, pas[i]), function);
 394 out_free:
 395         kvfree(in);
 396         if (notify_fail)
 397                 page_notify_fail(dev, func_id, ec_function);
 398         return err;
 399 }
 400
 401 static void release_all_pages(struct mlx5_core_dev *dev, u16 func_id,
 402                               bool ec_function)
 403 {
 404         u32 function = get_function(func_id, ec_function);
 405         struct rb_root *root;
 406         struct rb_node *p;
 407         int npages = 0;
 408
 409         root = xa_load(&dev->priv.page_root_xa, function);
 410         if (WARN_ON_ONCE(!root))
 411                 return;
 412
 413         p = rb_first(root);
 414         while (p) {
 415                 struct fw_page *fwp = rb_entry(p, struct fw_page, rb_node);
 416
 417                 p = rb_next(p);
 418                 npages += (MLX5_NUM_4K_IN_PAGE - fwp->free_count);
 419                 free_fwp(dev, fwp, fwp->free_count);
 420         }
 421
 422         dev->priv.fw_pages -= npages;
 423         if (func_id)
 424                 dev->priv.vfs_pages -= npages;
 425         else if (mlx5_core_is_ecpf(dev) && !ec_function)
 426                 dev->priv.host_pf_pages -= npages;
 427
 428         mlx5_core_dbg(dev, "npages %d, ec_function %d, func_id 0x%x\n",
 429                       npages, ec_function, func_id);
 430 }
 431
 432 static u32 fwp_fill_manage_pages_out(struct fw_page *fwp, u32 *out, u32 index,
 433                                      u32 npages)
 434 {
 435         u32 pages_set = 0;
 436         unsigned int n;
 437
 438         for_each_clear_bit(n, &fwp->bitmask, MLX5_NUM_4K_IN_PAGE) {
 439                 MLX5_ARRAY_SET64(manage_pages_out, out, pas, index + pages_set,
 440                                  fwp->addr + (n * MLX5_ADAPTER_PAGE_SIZE));
 441                 pages_set++;
 442
 443                 if (!--npages)
 444                         break;
 445         }
 446
 447         return pages_set;
 448 }
 449
 450 static int reclaim_pages_cmd(struct mlx5_core_dev *dev,
 451                              u32 *in, int in_size, u32 *out, int out_size)
 452 {
 453         struct rb_root *root;
 454         struct fw_page *fwp;
 455         struct rb_node *p;
 456         bool ec_function;
 457         u32 func_id;
 458         u32 npages;
 459         u32 i = 0;
 460
 461         if (!mlx5_cmd_is_down(dev))
 462                 return mlx5_cmd_exec(dev, in, in_size, out, out_size);
 463
 464         /* No hard feelings, we want our pages back! */
 465         npages = MLX5_GET(manage_pages_in, in, input_num_entries);
 466         func_id = MLX5_GET(manage_pages_in, in, function_id);
 467         ec_function = MLX5_GET(manage_pages_in, in, embedded_cpu_function);
 468
 469         root = xa_load(&dev->priv.page_root_xa, get_function(func_id, ec_function));
 470         if (WARN_ON_ONCE(!root))
 471                 return -EEXIST;
 472
 473         p = rb_first(root);
 474         while (p && i < npages) {
 475                 fwp = rb_entry(p, struct fw_page, rb_node);
 476                 p = rb_next(p);
 477
 478                 i += fwp_fill_manage_pages_out(fwp, out, i, npages - i);
 479         }
 480
 481         MLX5_SET(manage_pages_out, out, output_num_entries, i);
 482         return 0;
 483 }
 484
 485 static int reclaim_pages(struct mlx5_core_dev *dev, u16 func_id, int npages,
 486                          int *nclaimed, bool ec_function)
 487 {
 488         u32 function = get_function(func_id, ec_function);
 489         int outlen = MLX5_ST_SZ_BYTES(manage_pages_out);
 490         u32 in[MLX5_ST_SZ_DW(manage_pages_in)] = {};
 491         int num_claimed;
 492         u32 *out;
 493         int err;
 494         int i;
 495
 496         if (nclaimed)
 497                 *nclaimed = 0;
 498
 499         outlen += npages * MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
 500         out = kvzalloc(outlen, GFP_KERNEL);
 501         if (!out)
 502                 return -ENOMEM;
 503
 504         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 505         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_TAKE);
 506         MLX5_SET(manage_pages_in, in, function_id, func_id);
 507         MLX5_SET(manage_pages_in, in, input_num_entries, npages);
 508         MLX5_SET(manage_pages_in, in, embedded_cpu_function, ec_function);
 509
 510         mlx5_core_dbg(dev, "func 0x%x, npages %d, outlen %d\n",
 511                       func_id, npages, outlen);
 512         err = reclaim_pages_cmd(dev, in, sizeof(in), out, outlen);
 513         if (err) {
 514                 mlx5_core_err(dev, "failed reclaiming pages: err %d\n", err);
 515                 goto out_free;
 516         }
 517
 518         num_claimed = MLX5_GET(manage_pages_out, out, output_num_entries);
 519         if (num_claimed > npages) {
 520                 mlx5_core_warn(dev, "fw returned %d, driver asked %d => corruption\n",
 521                                num_claimed, npages);
 522                 err = -EINVAL;
 523                 goto out_free;
 524         }
 525
 526         for (i = 0; i < num_claimed; i++)
 527                 free_4k(dev, MLX5_GET64(manage_pages_out, out, pas[i]), function);
 528
 529         if (nclaimed)
 530                 *nclaimed = num_claimed;
 531
 532         dev->priv.fw_pages -= num_claimed;
 533         if (func_id)
 534                 dev->priv.vfs_pages -= num_claimed;
 535         else if (mlx5_core_is_ecpf(dev) && !ec_function)
 536                 dev->priv.host_pf_pages -= num_claimed;
 537
 538 out_free:
 539         kvfree(out);
 540         return err;
 541 }
 542
 543 static void pages_work_handler(struct work_struct *work)
 544 {
 545         struct mlx5_pages_req *req = container_of(work, struct mlx5_pages_req, work);
 546         struct mlx5_core_dev *dev = req->dev;
 547         int err = 0;
 548
 549         if (req->release_all)
 550                 release_all_pages(dev, req->func_id, req->ec_function);
 551         else if (req->npages < 0)
 552                 err = reclaim_pages(dev, req->func_id, -1 * req->npages, NULL,
 553                                     req->ec_function);
 554         else if (req->npages > 0)
 555                 err = give_pages(dev, req->func_id, req->npages, 1, req->ec_function);
 556
 557         if (err)
 558                 mlx5_core_warn(dev, "%s fail %d\n",
 559                                req->npages < 0 ? "reclaim" : "give", err);
 560
 561         kfree(req);
 562 }
 563
 564 enum {
 565         EC_FUNCTION_MASK = 0x8000,
 566         RELEASE_ALL_PAGES_MASK = 0x4000,
 567 };
 568
 569 static int req_pages_handler(struct notifier_block *nb,
 570                              unsigned long type, void *data)
 571 {
 572         struct mlx5_pages_req *req;
 573         struct mlx5_core_dev *dev;
 574         struct mlx5_priv *priv;
 575         struct mlx5_eqe *eqe;
 576         bool ec_function;
 577         bool release_all;
 578         u16 func_id;
 579         s32 npages;
 580
 581         priv = mlx5_nb_cof(nb, struct mlx5_priv, pg_nb);
 582         dev  = container_of(priv, struct mlx5_core_dev, priv);
 583         eqe  = data;
 584
 585         func_id = be16_to_cpu(eqe->data.req_pages.func_id);
 586         npages  = be32_to_cpu(eqe->data.req_pages.num_pages);
 587         ec_function = be16_to_cpu(eqe->data.req_pages.ec_function) & EC_FUNCTION_MASK;
 588         release_all = be16_to_cpu(eqe->data.req_pages.ec_function) &
 589                       RELEASE_ALL_PAGES_MASK;
 590         mlx5_core_dbg(dev, "page request for func 0x%x, npages %d, release_all %d\n",
 591                       func_id, npages, release_all);
 592         req = kzalloc(sizeof(*req), GFP_ATOMIC);
 593         if (!req) {
 594                 mlx5_core_warn(dev, "failed to allocate pages request\n");
 595                 return NOTIFY_DONE;
 596         }
 597
 598         req->dev = dev;
 599         req->func_id = func_id;
 600         req->npages = npages;
 601         req->ec_function = ec_function;
 602         req->release_all = release_all;
 603         INIT_WORK(&req->work, pages_work_handler);
 604         queue_work(dev->priv.pg_wq, &req->work);
 605         return NOTIFY_OK;
 606 }
 607
 608 int mlx5_satisfy_startup_pages(struct mlx5_core_dev *dev, int boot)
 609 {
 610         u16 func_id;
 611         s32 npages;
 612         int err;
 613
 614         err = mlx5_cmd_query_pages(dev, &func_id, &npages, boot);
 615         if (err)
 616                 return err;
 617
 618         mlx5_core_dbg(dev, "requested %d %s pages for func_id 0x%x\n",
 619                       npages, boot ? "boot" : "init", func_id);
 620
 621         return give_pages(dev, func_id, npages, 0, mlx5_core_is_ecpf(dev));
 622 }
 623
 624 enum {
 625         MLX5_BLKS_FOR_RECLAIM_PAGES = 12
 626 };
 627
 628 static int optimal_reclaimed_pages(void)
 629 {
 630         struct mlx5_cmd_prot_block *block;
 631         struct mlx5_cmd_layout *lay;
 632         int ret;
 633
 634         ret = (sizeof(lay->out) + MLX5_BLKS_FOR_RECLAIM_PAGES * sizeof(block->data) -
 635                MLX5_ST_SZ_BYTES(manage_pages_out)) /
 636                MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
 637
 638         return ret;
 639 }
 640
 641 static int mlx5_reclaim_root_pages(struct mlx5_core_dev *dev,
 642                                    struct rb_root *root, u16 func_id)
 643 {
 644         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 645
 646         while (!RB_EMPTY_ROOT(root)) {
 647                 int nclaimed;
 648                 int err;
 649
 650                 err = reclaim_pages(dev, func_id, optimal_reclaimed_pages(),
 651                                     &nclaimed, mlx5_core_is_ecpf(dev));
 652                 if (err) {
 653                         mlx5_core_warn(dev, "failed reclaiming pages (%d) for func id 0x%x\n",
 654                                        err, func_id);
 655                         return err;
 656                 }
 657
 658                 if (nclaimed)
 659                         end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 660
 661                 if (time_after(jiffies, end)) {
 662                         mlx5_core_warn(dev, "FW did not return all pages. giving up...\n");
 663                         break;
 664                 }
 665         }
 666
 667         return 0;
 668 }
 669
 670 int mlx5_reclaim_startup_pages(struct mlx5_core_dev *dev)
 671 {
 672         struct rb_root *root;
 673         unsigned long id;
 674         void *entry;
 675
 676         xa_for_each(&dev->priv.page_root_xa, id, entry) {
 677                 root = entry;
 678                 mlx5_reclaim_root_pages(dev, root, id);
 679                 xa_erase(&dev->priv.page_root_xa, id);
 680                 kfree(root);
 681         }
 682
 683         WARN_ON(!xa_empty(&dev->priv.page_root_xa));
 684
 685         WARN(dev->priv.fw_pages,
 686              "FW pages counter is %d after reclaiming all pages\n",
 687              dev->priv.fw_pages);
 688         WARN(dev->priv.vfs_pages,
 689              "VFs FW pages counter is %d after reclaiming all pages\n",
 690              dev->priv.vfs_pages);
 691         WARN(dev->priv.host_pf_pages,
 692              "External host PF FW pages counter is %d after reclaiming all pages\n",
 693              dev->priv.host_pf_pages);
 694
 695         return 0;
 696 }
 697
 698 int mlx5_pagealloc_init(struct mlx5_core_dev *dev)
 699 {
 700         INIT_LIST_HEAD(&dev->priv.free_list);
 701         dev->priv.pg_wq = create_singlethread_workqueue("mlx5_page_allocator");
 702         if (!dev->priv.pg_wq)
 703                 return -ENOMEM;
 704
 705         xa_init(&dev->priv.page_root_xa);
 706
 707         return 0;
 708 }
 709
 710 void mlx5_pagealloc_cleanup(struct mlx5_core_dev *dev)
 711 {
 712         xa_destroy(&dev->priv.page_root_xa);
 713         destroy_workqueue(dev->priv.pg_wq);
 714 }
 715
 716 void mlx5_pagealloc_start(struct mlx5_core_dev *dev)
 717 {
 718         MLX5_NB_INIT(&dev->priv.pg_nb, req_pages_handler, PAGE_REQUEST);
 719         mlx5_eq_notifier_register(dev, &dev->priv.pg_nb);
 720 }
 721
 722 void mlx5_pagealloc_stop(struct mlx5_core_dev *dev)
 723 {
 724         mlx5_eq_notifier_unregister(dev, &dev->priv.pg_nb);
 725         flush_workqueue(dev->priv.pg_wq);
 726 }
 727
 728 int mlx5_wait_for_pages(struct mlx5_core_dev *dev, int *pages)
 729 {
 730         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_VFS_PAGES_TIME_MSECS);
 731         int prev_pages = *pages;
 732
 733         /* In case of internal error we will free the pages manually later */
 734         if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
 735                 mlx5_core_warn(dev, "Skipping wait for vf pages stage");
 736                 return 0;
 737         }
 738
 739         mlx5_core_dbg(dev, "Waiting for %d pages\n", prev_pages);
 740         while (*pages) {
 741                 if (time_after(jiffies, end)) {
 742                         mlx5_core_warn(dev, "aborting while there are %d pending pages\n", *pages);
 743                         return -ETIMEDOUT;
 744                 }
 745                 if (*pages < prev_pages) {
 746                         end = jiffies + msecs_to_jiffies(MAX_RECLAIM_VFS_PAGES_TIME_MSECS);
 747                         prev_pages = *pages;
 748                 }
 749                 msleep(50);
 750         }
 751
 752         mlx5_core_dbg(dev, "All pages received\n");
 753         return 0;
 754 }