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/mlx5/cmd.h>
  39 #include "mlx5_core.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         s32     npages;
  51         struct work_struct work;
  52 };
  53
  54 struct fw_page {
  55         struct rb_node          rb_node;
  56         u64                     addr;
  57         struct page            *page;
  58         u16                     func_id;
  59         unsigned long           bitmask;
  60         struct list_head        list;
  61         unsigned                free_count;
  62 };
  63
  64 enum {
  65         MAX_RECLAIM_TIME_MSECS  = 5000,
  66         MAX_RECLAIM_VFS_PAGES_TIME_MSECS = 2 * 1000 * 60,
  67 };
  68
  69 enum {
  70         MLX5_MAX_RECLAIM_TIME_MILI      = 5000,
  71         MLX5_NUM_4K_IN_PAGE             = PAGE_SIZE / MLX5_ADAPTER_PAGE_SIZE,
  72 };
  73
  74 static int insert_page(struct mlx5_core_dev *dev, u64 addr, struct page *page, u16 func_id)
  75 {
  76         struct rb_root *root = &dev->priv.page_root;
  77         struct rb_node **new = &root->rb_node;
  78         struct rb_node *parent = NULL;
  79         struct fw_page *nfp;
  80         struct fw_page *tfp;
  81         int i;
  82
  83         while (*new) {
  84                 parent = *new;
  85                 tfp = rb_entry(parent, struct fw_page, rb_node);
  86                 if (tfp->addr < addr)
  87                         new = &parent->rb_left;
  88                 else if (tfp->addr > addr)
  89                         new = &parent->rb_right;
  90                 else
  91                         return -EEXIST;
  92         }
  93
  94         nfp = kzalloc(sizeof(*nfp), GFP_KERNEL);
  95         if (!nfp)
  96                 return -ENOMEM;
  97
  98         nfp->addr = addr;
  99         nfp->page = page;
 100         nfp->func_id = func_id;
 101         nfp->free_count = MLX5_NUM_4K_IN_PAGE;
 102         for (i = 0; i < MLX5_NUM_4K_IN_PAGE; i++)
 103                 set_bit(i, &nfp->bitmask);
 104
 105         rb_link_node(&nfp->rb_node, parent, new);
 106         rb_insert_color(&nfp->rb_node, root);
 107         list_add(&nfp->list, &dev->priv.free_list);
 108
 109         return 0;
 110 }
 111
 112 static struct fw_page *find_fw_page(struct mlx5_core_dev *dev, u64 addr)
 113 {
 114         struct rb_root *root = &dev->priv.page_root;
 115         struct rb_node *tmp = root->rb_node;
 116         struct fw_page *result = NULL;
 117         struct fw_page *tfp;
 118
 119         while (tmp) {
 120                 tfp = rb_entry(tmp, struct fw_page, rb_node);
 121                 if (tfp->addr < addr) {
 122                         tmp = tmp->rb_left;
 123                 } else if (tfp->addr > addr) {
 124                         tmp = tmp->rb_right;
 125                 } else {
 126                         result = tfp;
 127                         break;
 128                 }
 129         }
 130
 131         return result;
 132 }
 133
 134 static int mlx5_cmd_query_pages(struct mlx5_core_dev *dev, u16 *func_id,
 135                                 s32 *npages, int boot)
 136 {
 137         u32 out[MLX5_ST_SZ_DW(query_pages_out)] = {0};
 138         u32 in[MLX5_ST_SZ_DW(query_pages_in)]   = {0};
 139         int err;
 140
 141         MLX5_SET(query_pages_in, in, opcode, MLX5_CMD_OP_QUERY_PAGES);
 142         MLX5_SET(query_pages_in, in, op_mod, boot ?
 143                  MLX5_QUERY_PAGES_IN_OP_MOD_BOOT_PAGES :
 144                  MLX5_QUERY_PAGES_IN_OP_MOD_INIT_PAGES);
 145
 146         err = mlx5_cmd_exec(dev, in, sizeof(in), out, sizeof(out));
 147         if (err)
 148                 return err;
 149
 150         *npages = MLX5_GET(query_pages_out, out, num_pages);
 151         *func_id = MLX5_GET(query_pages_out, out, function_id);
 152
 153         return err;
 154 }
 155
 156 static int alloc_4k(struct mlx5_core_dev *dev, u64 *addr)
 157 {
 158         struct fw_page *fp;
 159         unsigned n;
 160
 161         if (list_empty(&dev->priv.free_list))
 162                 return -ENOMEM;
 163
 164         fp = list_entry(dev->priv.free_list.next, struct fw_page, list);
 165         n = find_first_bit(&fp->bitmask, 8 * sizeof(fp->bitmask));
 166         if (n >= MLX5_NUM_4K_IN_PAGE) {
 167                 mlx5_core_warn(dev, "alloc 4k bug\n");
 168                 return -ENOENT;
 169         }
 170         clear_bit(n, &fp->bitmask);
 171         fp->free_count--;
 172         if (!fp->free_count)
 173                 list_del(&fp->list);
 174
 175         *addr = fp->addr + n * MLX5_ADAPTER_PAGE_SIZE;
 176
 177         return 0;
 178 }
 179
 180 #define MLX5_U64_4K_PAGE_MASK ((~(u64)0U) << PAGE_SHIFT)
 181
 182 static void free_4k(struct mlx5_core_dev *dev, u64 addr)
 183 {
 184         struct fw_page *fwp;
 185         int n;
 186
 187         fwp = find_fw_page(dev, addr & MLX5_U64_4K_PAGE_MASK);
 188         if (!fwp) {
 189                 mlx5_core_warn(dev, "page not found\n");
 190                 return;
 191         }
 192
 193         n = (addr & ~MLX5_U64_4K_PAGE_MASK) >> MLX5_ADAPTER_PAGE_SHIFT;
 194         fwp->free_count++;
 195         set_bit(n, &fwp->bitmask);
 196         if (fwp->free_count == MLX5_NUM_4K_IN_PAGE) {
 197                 rb_erase(&fwp->rb_node, &dev->priv.page_root);
 198                 if (fwp->free_count != 1)
 199                         list_del(&fwp->list);
 200                 dma_unmap_page(&dev->pdev->dev, addr & MLX5_U64_4K_PAGE_MASK,
 201                                PAGE_SIZE, DMA_BIDIRECTIONAL);
 202                 __free_page(fwp->page);
 203                 kfree(fwp);
 204         } else if (fwp->free_count == 1) {
 205                 list_add(&fwp->list, &dev->priv.free_list);
 206         }
 207 }
 208
 209 static int alloc_system_page(struct mlx5_core_dev *dev, u16 func_id)
 210 {
 211         struct page *page;
 212         u64 zero_addr = 1;
 213         u64 addr;
 214         int err;
 215         int nid = dev_to_node(&dev->pdev->dev);
 216
 217         page = alloc_pages_node(nid, GFP_HIGHUSER, 0);
 218         if (!page) {
 219                 mlx5_core_warn(dev, "failed to allocate page\n");
 220                 return -ENOMEM;
 221         }
 222 map:
 223         addr = dma_map_page(&dev->pdev->dev, page, 0,
 224                             PAGE_SIZE, DMA_BIDIRECTIONAL);
 225         if (dma_mapping_error(&dev->pdev->dev, addr)) {
 226                 mlx5_core_warn(dev, "failed dma mapping page\n");
 227                 err = -ENOMEM;
 228                 goto err_mapping;
 229         }
 230
 231         /* Firmware doesn't support page with physical address 0 */
 232         if (addr == 0) {
 233                 zero_addr = addr;
 234                 goto map;
 235         }
 236
 237         err = insert_page(dev, addr, page, func_id);
 238         if (err) {
 239                 mlx5_core_err(dev, "failed to track allocated page\n");
 240                 dma_unmap_page(&dev->pdev->dev, addr, PAGE_SIZE,
 241                                DMA_BIDIRECTIONAL);
 242         }
 243
 244 err_mapping:
 245         if (err)
 246                 __free_page(page);
 247
 248         if (zero_addr == 0)
 249                 dma_unmap_page(&dev->pdev->dev, zero_addr, PAGE_SIZE,
 250                                DMA_BIDIRECTIONAL);
 251
 252         return err;
 253 }
 254
 255 static void page_notify_fail(struct mlx5_core_dev *dev, u16 func_id)
 256 {
 257         u32 out[MLX5_ST_SZ_DW(manage_pages_out)] = {0};
 258         u32 in[MLX5_ST_SZ_DW(manage_pages_in)]   = {0};
 259         int err;
 260
 261         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 262         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_CANT_GIVE);
 263         MLX5_SET(manage_pages_in, in, function_id, func_id);
 264
 265         err = mlx5_cmd_exec(dev, in, sizeof(in), out, sizeof(out));
 266         if (err)
 267                 mlx5_core_warn(dev, "page notify failed func_id(%d) err(%d)\n",
 268                                func_id, err);
 269 }
 270
 271 static int give_pages(struct mlx5_core_dev *dev, u16 func_id, int npages,
 272                       int notify_fail)
 273 {
 274         u32 out[MLX5_ST_SZ_DW(manage_pages_out)] = {0};
 275         int inlen = MLX5_ST_SZ_BYTES(manage_pages_in);
 276         u64 addr;
 277         int err;
 278         u32 *in;
 279         int i;
 280
 281         inlen += npages * MLX5_FLD_SZ_BYTES(manage_pages_in, pas[0]);
 282         in = kvzalloc(inlen, GFP_KERNEL);
 283         if (!in) {
 284                 err = -ENOMEM;
 285                 mlx5_core_warn(dev, "vzalloc failed %d\n", inlen);
 286                 goto out_free;
 287         }
 288
 289         for (i = 0; i < npages; i++) {
 290 retry:
 291                 err = alloc_4k(dev, &addr);
 292                 if (err) {
 293                         if (err == -ENOMEM)
 294                                 err = alloc_system_page(dev, func_id);
 295                         if (err)
 296                                 goto out_4k;
 297
 298                         goto retry;
 299                 }
 300                 MLX5_ARRAY_SET64(manage_pages_in, in, pas, i, addr);
 301         }
 302
 303         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 304         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_GIVE);
 305         MLX5_SET(manage_pages_in, in, function_id, func_id);
 306         MLX5_SET(manage_pages_in, in, input_num_entries, npages);
 307
 308         err = mlx5_cmd_exec(dev, in, inlen, out, sizeof(out));
 309         if (err) {
 310                 mlx5_core_warn(dev, "func_id 0x%x, npages %d, err %d\n",
 311                                func_id, npages, err);
 312                 goto out_4k;
 313         }
 314
 315         dev->priv.fw_pages += npages;
 316         if (func_id)
 317                 dev->priv.vfs_pages += npages;
 318
 319         mlx5_core_dbg(dev, "err %d\n", err);
 320
 321         kvfree(in);
 322         return 0;
 323
 324 out_4k:
 325         for (i--; i >= 0; i--)
 326                 free_4k(dev, MLX5_GET64(manage_pages_in, in, pas[i]));
 327 out_free:
 328         kvfree(in);
 329         if (notify_fail)
 330                 page_notify_fail(dev, func_id);
 331         return err;
 332 }
 333
 334 static int reclaim_pages_cmd(struct mlx5_core_dev *dev,
 335                              u32 *in, int in_size, u32 *out, int out_size)
 336 {
 337         struct fw_page *fwp;
 338         struct rb_node *p;
 339         u32 func_id;
 340         u32 npages;
 341         u32 i = 0;
 342
 343         if (dev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR)
 344                 return mlx5_cmd_exec(dev, in, in_size, out, out_size);
 345
 346         /* No hard feelings, we want our pages back! */
 347         npages = MLX5_GET(manage_pages_in, in, input_num_entries);
 348         func_id = MLX5_GET(manage_pages_in, in, function_id);
 349
 350         p = rb_first(&dev->priv.page_root);
 351         while (p && i < npages) {
 352                 fwp = rb_entry(p, struct fw_page, rb_node);
 353                 p = rb_next(p);
 354                 if (fwp->func_id != func_id)
 355                         continue;
 356
 357                 MLX5_ARRAY_SET64(manage_pages_out, out, pas, i, fwp->addr);
 358                 i++;
 359         }
 360
 361         MLX5_SET(manage_pages_out, out, output_num_entries, i);
 362         return 0;
 363 }
 364
 365 static int reclaim_pages(struct mlx5_core_dev *dev, u32 func_id, int npages,
 366                          int *nclaimed)
 367 {
 368         int outlen = MLX5_ST_SZ_BYTES(manage_pages_out);
 369         u32 in[MLX5_ST_SZ_DW(manage_pages_in)] = {0};
 370         int num_claimed;
 371         u32 *out;
 372         int err;
 373         int i;
 374
 375         if (nclaimed)
 376                 *nclaimed = 0;
 377
 378         outlen += npages * MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
 379         out = kvzalloc(outlen, GFP_KERNEL);
 380         if (!out)
 381                 return -ENOMEM;
 382
 383         MLX5_SET(manage_pages_in, in, opcode, MLX5_CMD_OP_MANAGE_PAGES);
 384         MLX5_SET(manage_pages_in, in, op_mod, MLX5_PAGES_TAKE);
 385         MLX5_SET(manage_pages_in, in, function_id, func_id);
 386         MLX5_SET(manage_pages_in, in, input_num_entries, npages);
 387
 388         mlx5_core_dbg(dev, "npages %d, outlen %d\n", npages, outlen);
 389         err = reclaim_pages_cmd(dev, in, sizeof(in), out, outlen);
 390         if (err) {
 391                 mlx5_core_err(dev, "failed reclaiming pages: err %d\n", err);
 392                 goto out_free;
 393         }
 394
 395         num_claimed = MLX5_GET(manage_pages_out, out, output_num_entries);
 396         if (num_claimed > npages) {
 397                 mlx5_core_warn(dev, "fw returned %d, driver asked %d => corruption\n",
 398                                num_claimed, npages);
 399                 err = -EINVAL;
 400                 goto out_free;
 401         }
 402
 403         for (i = 0; i < num_claimed; i++)
 404                 free_4k(dev, MLX5_GET64(manage_pages_out, out, pas[i]));
 405
 406         if (nclaimed)
 407                 *nclaimed = num_claimed;
 408
 409         dev->priv.fw_pages -= num_claimed;
 410         if (func_id)
 411                 dev->priv.vfs_pages -= num_claimed;
 412
 413 out_free:
 414         kvfree(out);
 415         return err;
 416 }
 417
 418 static void pages_work_handler(struct work_struct *work)
 419 {
 420         struct mlx5_pages_req *req = container_of(work, struct mlx5_pages_req, work);
 421         struct mlx5_core_dev *dev = req->dev;
 422         int err = 0;
 423
 424         if (req->npages < 0)
 425                 err = reclaim_pages(dev, req->func_id, -1 * req->npages, NULL);
 426         else if (req->npages > 0)
 427                 err = give_pages(dev, req->func_id, req->npages, 1);
 428
 429         if (err)
 430                 mlx5_core_warn(dev, "%s fail %d\n",
 431                                req->npages < 0 ? "reclaim" : "give", err);
 432
 433         kfree(req);
 434 }
 435
 436 void mlx5_core_req_pages_handler(struct mlx5_core_dev *dev, u16 func_id,
 437                                  s32 npages)
 438 {
 439         struct mlx5_pages_req *req;
 440
 441         req = kzalloc(sizeof(*req), GFP_ATOMIC);
 442         if (!req) {
 443                 mlx5_core_warn(dev, "failed to allocate pages request\n");
 444                 return;
 445         }
 446
 447         req->dev = dev;
 448         req->func_id = func_id;
 449         req->npages = npages;
 450         INIT_WORK(&req->work, pages_work_handler);
 451         queue_work(dev->priv.pg_wq, &req->work);
 452 }
 453
 454 int mlx5_satisfy_startup_pages(struct mlx5_core_dev *dev, int boot)
 455 {
 456         u16 uninitialized_var(func_id);
 457         s32 uninitialized_var(npages);
 458         int err;
 459
 460         err = mlx5_cmd_query_pages(dev, &func_id, &npages, boot);
 461         if (err)
 462                 return err;
 463
 464         mlx5_core_dbg(dev, "requested %d %s pages for func_id 0x%x\n",
 465                       npages, boot ? "boot" : "init", func_id);
 466
 467         return give_pages(dev, func_id, npages, 0);
 468 }
 469
 470 enum {
 471         MLX5_BLKS_FOR_RECLAIM_PAGES = 12
 472 };
 473
 474 static int optimal_reclaimed_pages(void)
 475 {
 476         struct mlx5_cmd_prot_block *block;
 477         struct mlx5_cmd_layout *lay;
 478         int ret;
 479
 480         ret = (sizeof(lay->out) + MLX5_BLKS_FOR_RECLAIM_PAGES * sizeof(block->data) -
 481                MLX5_ST_SZ_BYTES(manage_pages_out)) /
 482                MLX5_FLD_SZ_BYTES(manage_pages_out, pas[0]);
 483
 484         return ret;
 485 }
 486
 487 int mlx5_reclaim_startup_pages(struct mlx5_core_dev *dev)
 488 {
 489         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 490         struct fw_page *fwp;
 491         struct rb_node *p;
 492         int nclaimed = 0;
 493         int err = 0;
 494
 495         do {
 496                 p = rb_first(&dev->priv.page_root);
 497                 if (p) {
 498                         fwp = rb_entry(p, struct fw_page, rb_node);
 499                         err = reclaim_pages(dev, fwp->func_id,
 500                                             optimal_reclaimed_pages(),
 501                                             &nclaimed);
 502
 503                         if (err) {
 504                                 mlx5_core_warn(dev, "failed reclaiming pages (%d)\n",
 505                                                err);
 506                                 return err;
 507                         }
 508                         if (nclaimed)
 509                                 end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 510                 }
 511                 if (time_after(jiffies, end)) {
 512                         mlx5_core_warn(dev, "FW did not return all pages. giving up...\n");
 513                         break;
 514                 }
 515         } while (p);
 516
 517         WARN(dev->priv.fw_pages,
 518              "FW pages counter is %d after reclaiming all pages\n",
 519              dev->priv.fw_pages);
 520         WARN(dev->priv.vfs_pages,
 521              "VFs FW pages counter is %d after reclaiming all pages\n",
 522              dev->priv.vfs_pages);
 523
 524         return 0;
 525 }
 526
 527 void mlx5_pagealloc_init(struct mlx5_core_dev *dev)
 528 {
 529         dev->priv.page_root = RB_ROOT;
 530         INIT_LIST_HEAD(&dev->priv.free_list);
 531 }
 532
 533 void mlx5_pagealloc_cleanup(struct mlx5_core_dev *dev)
 534 {
 535         /* nothing */
 536 }
 537
 538 int mlx5_pagealloc_start(struct mlx5_core_dev *dev)
 539 {
 540         dev->priv.pg_wq = create_singlethread_workqueue("mlx5_page_allocator");
 541         if (!dev->priv.pg_wq)
 542                 return -ENOMEM;
 543
 544         return 0;
 545 }
 546
 547 void mlx5_pagealloc_stop(struct mlx5_core_dev *dev)
 548 {
 549         destroy_workqueue(dev->priv.pg_wq);
 550 }
 551
 552 int mlx5_wait_for_vf_pages(struct mlx5_core_dev *dev)
 553 {
 554         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_VFS_PAGES_TIME_MSECS);
 555         int prev_vfs_pages = dev->priv.vfs_pages;
 556
 557         /* In case of internal error we will free the pages manually later */
 558         if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
 559                 mlx5_core_warn(dev, "Skipping wait for vf pages stage");
 560                 return 0;
 561         }
 562
 563         mlx5_core_dbg(dev, "Waiting for %d pages from %s\n", prev_vfs_pages,
 564                       dev->priv.name);
 565         while (dev->priv.vfs_pages) {
 566                 if (time_after(jiffies, end)) {
 567                         mlx5_core_warn(dev, "aborting while there are %d pending pages\n", dev->priv.vfs_pages);
 568                         return -ETIMEDOUT;
 569                 }
 570                 if (dev->priv.vfs_pages < prev_vfs_pages) {
 571                         end = jiffies + msecs_to_jiffies(MAX_RECLAIM_VFS_PAGES_TIME_MSECS);
 572                         prev_vfs_pages = dev->priv.vfs_pages;
 573                 }
 574                 msleep(50);
 575         }
 576
 577         mlx5_core_dbg(dev, "All pages received from %s\n", dev->priv.name);
 578         return 0;
 579 }