drivers/net/ethernet/mellanox/mlx5/core/pagealloc.c

   1 /*
   2  * Copyright (c) 2013, Mellanox Technologies inc.  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 <asm-generic/kmap_types.h>
  34 #include <linux/kernel.h>
  35 #include <linux/module.h>
  36 #include <linux/mlx5/driver.h>
  37 #include <linux/mlx5/cmd.h>
  38 #include "mlx5_core.h"
  39
  40 enum {
  41         MLX5_PAGES_CANT_GIVE    = 0,
  42         MLX5_PAGES_GIVE         = 1,
  43         MLX5_PAGES_TAKE         = 2
  44 };
  45
  46 enum {
  47         MLX5_BOOT_PAGES         = 1,
  48         MLX5_INIT_PAGES         = 2,
  49         MLX5_POST_INIT_PAGES    = 3
  50 };
  51
  52 struct mlx5_pages_req {
  53         struct mlx5_core_dev *dev;
  54         u16     func_id;
  55         s32     npages;
  56         struct work_struct work;
  57 };
  58
  59 struct fw_page {
  60         struct rb_node          rb_node;
  61         u64                     addr;
  62         struct page            *page;
  63         u16                     func_id;
  64         unsigned long           bitmask;
  65         struct list_head        list;
  66         unsigned                free_count;
  67 };
  68
  69 struct mlx5_query_pages_inbox {
  70         struct mlx5_inbox_hdr   hdr;
  71         u8                      rsvd[8];
  72 };
  73
  74 struct mlx5_query_pages_outbox {
  75         struct mlx5_outbox_hdr  hdr;
  76         __be16                  rsvd;
  77         __be16                  func_id;
  78         __be32                  num_pages;
  79 };
  80
  81 struct mlx5_manage_pages_inbox {
  82         struct mlx5_inbox_hdr   hdr;
  83         __be16                  rsvd;
  84         __be16                  func_id;
  85         __be32                  num_entries;
  86         __be64                  pas[0];
  87 };
  88
  89 struct mlx5_manage_pages_outbox {
  90         struct mlx5_outbox_hdr  hdr;
  91         __be32                  num_entries;
  92         u8                      rsvd[4];
  93         __be64                  pas[0];
  94 };
  95
  96 enum {
  97         MAX_RECLAIM_TIME_MSECS  = 5000,
  98 };
  99
 100 enum {
 101         MLX5_MAX_RECLAIM_TIME_MILI      = 5000,
 102         MLX5_NUM_4K_IN_PAGE             = PAGE_SIZE / MLX5_ADAPTER_PAGE_SIZE,
 103 };
 104
 105 static int insert_page(struct mlx5_core_dev *dev, u64 addr, struct page *page, u16 func_id)
 106 {
 107         struct rb_root *root = &dev->priv.page_root;
 108         struct rb_node **new = &root->rb_node;
 109         struct rb_node *parent = NULL;
 110         struct fw_page *nfp;
 111         struct fw_page *tfp;
 112         int i;
 113
 114         while (*new) {
 115                 parent = *new;
 116                 tfp = rb_entry(parent, struct fw_page, rb_node);
 117                 if (tfp->addr < addr)
 118                         new = &parent->rb_left;
 119                 else if (tfp->addr > addr)
 120                         new = &parent->rb_right;
 121                 else
 122                         return -EEXIST;
 123         }
 124
 125         nfp = kzalloc(sizeof(*nfp), GFP_KERNEL);
 126         if (!nfp)
 127                 return -ENOMEM;
 128
 129         nfp->addr = addr;
 130         nfp->page = page;
 131         nfp->func_id = func_id;
 132         nfp->free_count = MLX5_NUM_4K_IN_PAGE;
 133         for (i = 0; i < MLX5_NUM_4K_IN_PAGE; i++)
 134                 set_bit(i, &nfp->bitmask);
 135
 136         rb_link_node(&nfp->rb_node, parent, new);
 137         rb_insert_color(&nfp->rb_node, root);
 138         list_add(&nfp->list, &dev->priv.free_list);
 139
 140         return 0;
 141 }
 142
 143 static struct fw_page *find_fw_page(struct mlx5_core_dev *dev, u64 addr)
 144 {
 145         struct rb_root *root = &dev->priv.page_root;
 146         struct rb_node *tmp = root->rb_node;
 147         struct fw_page *result = NULL;
 148         struct fw_page *tfp;
 149
 150         while (tmp) {
 151                 tfp = rb_entry(tmp, struct fw_page, rb_node);
 152                 if (tfp->addr < addr) {
 153                         tmp = tmp->rb_left;
 154                 } else if (tfp->addr > addr) {
 155                         tmp = tmp->rb_right;
 156                 } else {
 157                         result = tfp;
 158                         break;
 159                 }
 160         }
 161
 162         return result;
 163 }
 164
 165 static int mlx5_cmd_query_pages(struct mlx5_core_dev *dev, u16 *func_id,
 166                                 s32 *npages, int boot)
 167 {
 168         struct mlx5_query_pages_inbox   in;
 169         struct mlx5_query_pages_outbox  out;
 170         int err;
 171
 172         memset(&in, 0, sizeof(in));
 173         memset(&out, 0, sizeof(out));
 174         in.hdr.opcode = cpu_to_be16(MLX5_CMD_OP_QUERY_PAGES);
 175         in.hdr.opmod = boot ? cpu_to_be16(MLX5_BOOT_PAGES) : cpu_to_be16(MLX5_INIT_PAGES);
 176
 177         err = mlx5_cmd_exec(dev, &in, sizeof(in), &out, sizeof(out));
 178         if (err)
 179                 return err;
 180
 181         if (out.hdr.status)
 182                 return mlx5_cmd_status_to_err(&out.hdr);
 183
 184         *npages = be32_to_cpu(out.num_pages);
 185         *func_id = be16_to_cpu(out.func_id);
 186
 187         return err;
 188 }
 189
 190 static int alloc_4k(struct mlx5_core_dev *dev, u64 *addr)
 191 {
 192         struct fw_page *fp;
 193         unsigned n;
 194
 195         if (list_empty(&dev->priv.free_list))
 196                 return -ENOMEM;
 197
 198         fp = list_entry(dev->priv.free_list.next, struct fw_page, list);
 199         n = find_first_bit(&fp->bitmask, 8 * sizeof(fp->bitmask));
 200         if (n >= MLX5_NUM_4K_IN_PAGE) {
 201                 mlx5_core_warn(dev, "alloc 4k bug\n");
 202                 return -ENOENT;
 203         }
 204         clear_bit(n, &fp->bitmask);
 205         fp->free_count--;
 206         if (!fp->free_count)
 207                 list_del(&fp->list);
 208
 209         *addr = fp->addr + n * MLX5_ADAPTER_PAGE_SIZE;
 210
 211         return 0;
 212 }
 213
 214 static void free_4k(struct mlx5_core_dev *dev, u64 addr)
 215 {
 216         struct fw_page *fwp;
 217         int n;
 218
 219         fwp = find_fw_page(dev, addr & PAGE_MASK);
 220         if (!fwp) {
 221                 mlx5_core_warn(dev, "page not found\n");
 222                 return;
 223         }
 224
 225         n = (addr & ~PAGE_MASK) >> MLX5_ADAPTER_PAGE_SHIFT;
 226         fwp->free_count++;
 227         set_bit(n, &fwp->bitmask);
 228         if (fwp->free_count == MLX5_NUM_4K_IN_PAGE) {
 229                 rb_erase(&fwp->rb_node, &dev->priv.page_root);
 230                 if (fwp->free_count != 1)
 231                         list_del(&fwp->list);
 232                 dma_unmap_page(&dev->pdev->dev, addr & PAGE_MASK, PAGE_SIZE,
 233                                DMA_BIDIRECTIONAL);
 234                 __free_page(fwp->page);
 235                 kfree(fwp);
 236         } else if (fwp->free_count == 1) {
 237                 list_add(&fwp->list, &dev->priv.free_list);
 238         }
 239 }
 240
 241 static int alloc_system_page(struct mlx5_core_dev *dev, u16 func_id)
 242 {
 243         struct page *page;
 244         u64 addr;
 245         int err;
 246         int nid = dev_to_node(&dev->pdev->dev);
 247
 248         page = alloc_pages_node(nid, GFP_HIGHUSER, 0);
 249         if (!page) {
 250                 mlx5_core_warn(dev, "failed to allocate page\n");
 251                 return -ENOMEM;
 252         }
 253         addr = dma_map_page(&dev->pdev->dev, page, 0,
 254                             PAGE_SIZE, DMA_BIDIRECTIONAL);
 255         if (dma_mapping_error(&dev->pdev->dev, addr)) {
 256                 mlx5_core_warn(dev, "failed dma mapping page\n");
 257                 err = -ENOMEM;
 258                 goto out_alloc;
 259         }
 260         err = insert_page(dev, addr, page, func_id);
 261         if (err) {
 262                 mlx5_core_err(dev, "failed to track allocated page\n");
 263                 goto out_mapping;
 264         }
 265
 266         return 0;
 267
 268 out_mapping:
 269         dma_unmap_page(&dev->pdev->dev, addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
 270
 271 out_alloc:
 272         __free_page(page);
 273
 274         return err;
 275 }
 276 static int give_pages(struct mlx5_core_dev *dev, u16 func_id, int npages,
 277                       int notify_fail)
 278 {
 279         struct mlx5_manage_pages_inbox *in;
 280         struct mlx5_manage_pages_outbox out;
 281         struct mlx5_manage_pages_inbox *nin;
 282         int inlen;
 283         u64 addr;
 284         int err;
 285         int i;
 286
 287         inlen = sizeof(*in) + npages * sizeof(in->pas[0]);
 288         in = mlx5_vzalloc(inlen);
 289         if (!in) {
 290                 mlx5_core_warn(dev, "vzalloc failed %d\n", inlen);
 291                 return -ENOMEM;
 292         }
 293         memset(&out, 0, sizeof(out));
 294
 295         for (i = 0; i < npages; i++) {
 296 retry:
 297                 err = alloc_4k(dev, &addr);
 298                 if (err) {
 299                         if (err == -ENOMEM)
 300                                 err = alloc_system_page(dev, func_id);
 301                         if (err)
 302                                 goto out_4k;
 303
 304                         goto retry;
 305                 }
 306                 in->pas[i] = cpu_to_be64(addr);
 307         }
 308
 309         in->hdr.opcode = cpu_to_be16(MLX5_CMD_OP_MANAGE_PAGES);
 310         in->hdr.opmod = cpu_to_be16(MLX5_PAGES_GIVE);
 311         in->func_id = cpu_to_be16(func_id);
 312         in->num_entries = cpu_to_be32(npages);
 313         err = mlx5_cmd_exec(dev, in, inlen, &out, sizeof(out));
 314         if (err) {
 315                 mlx5_core_warn(dev, "func_id 0x%x, npages %d, err %d\n",
 316                                func_id, npages, err);
 317                 goto out_alloc;
 318         }
 319         dev->priv.fw_pages += npages;
 320
 321         if (out.hdr.status) {
 322                 err = mlx5_cmd_status_to_err(&out.hdr);
 323                 if (err) {
 324                         mlx5_core_warn(dev, "func_id 0x%x, npages %d, status %d\n",
 325                                        func_id, npages, out.hdr.status);
 326                         goto out_alloc;
 327                 }
 328         }
 329
 330         mlx5_core_dbg(dev, "err %d\n", err);
 331
 332         goto out_free;
 333
 334 out_alloc:
 335         if (notify_fail) {
 336                 nin = kzalloc(sizeof(*nin), GFP_KERNEL);
 337                 if (!nin) {
 338                         mlx5_core_warn(dev, "allocation failed\n");
 339                         goto out_4k;
 340                 }
 341                 memset(&out, 0, sizeof(out));
 342                 nin->hdr.opcode = cpu_to_be16(MLX5_CMD_OP_MANAGE_PAGES);
 343                 nin->hdr.opmod = cpu_to_be16(MLX5_PAGES_CANT_GIVE);
 344                 if (mlx5_cmd_exec(dev, nin, sizeof(*nin), &out, sizeof(out)))
 345                         mlx5_core_warn(dev, "page notify failed\n");
 346                 kfree(nin);
 347         }
 348
 349 out_4k:
 350         for (i--; i >= 0; i--)
 351                 free_4k(dev, be64_to_cpu(in->pas[i]));
 352 out_free:
 353         kvfree(in);
 354         return err;
 355 }
 356
 357 static int reclaim_pages(struct mlx5_core_dev *dev, u32 func_id, int npages,
 358                          int *nclaimed)
 359 {
 360         struct mlx5_manage_pages_inbox   in;
 361         struct mlx5_manage_pages_outbox *out;
 362         int num_claimed;
 363         int outlen;
 364         u64 addr;
 365         int err;
 366         int i;
 367
 368         if (nclaimed)
 369                 *nclaimed = 0;
 370
 371         memset(&in, 0, sizeof(in));
 372         outlen = sizeof(*out) + npages * sizeof(out->pas[0]);
 373         out = mlx5_vzalloc(outlen);
 374         if (!out)
 375                 return -ENOMEM;
 376
 377         in.hdr.opcode = cpu_to_be16(MLX5_CMD_OP_MANAGE_PAGES);
 378         in.hdr.opmod = cpu_to_be16(MLX5_PAGES_TAKE);
 379         in.func_id = cpu_to_be16(func_id);
 380         in.num_entries = cpu_to_be32(npages);
 381         mlx5_core_dbg(dev, "npages %d, outlen %d\n", npages, outlen);
 382         err = mlx5_cmd_exec(dev, &in, sizeof(in), out, outlen);
 383         if (err) {
 384                 mlx5_core_err(dev, "failed reclaiming pages\n");
 385                 goto out_free;
 386         }
 387         dev->priv.fw_pages -= npages;
 388
 389         if (out->hdr.status) {
 390                 err = mlx5_cmd_status_to_err(&out->hdr);
 391                 goto out_free;
 392         }
 393
 394         num_claimed = be32_to_cpu(out->num_entries);
 395         if (nclaimed)
 396                 *nclaimed = num_claimed;
 397
 398         for (i = 0; i < num_claimed; i++) {
 399                 addr = be64_to_cpu(out->pas[i]);
 400                 free_4k(dev, addr);
 401         }
 402
 403 out_free:
 404         kvfree(out);
 405         return err;
 406 }
 407
 408 static void pages_work_handler(struct work_struct *work)
 409 {
 410         struct mlx5_pages_req *req = container_of(work, struct mlx5_pages_req, work);
 411         struct mlx5_core_dev *dev = req->dev;
 412         int err = 0;
 413
 414         if (req->npages < 0)
 415                 err = reclaim_pages(dev, req->func_id, -1 * req->npages, NULL);
 416         else if (req->npages > 0)
 417                 err = give_pages(dev, req->func_id, req->npages, 1);
 418
 419         if (err)
 420                 mlx5_core_warn(dev, "%s fail %d\n",
 421                                req->npages < 0 ? "reclaim" : "give", err);
 422
 423         kfree(req);
 424 }
 425
 426 void mlx5_core_req_pages_handler(struct mlx5_core_dev *dev, u16 func_id,
 427                                  s32 npages)
 428 {
 429         struct mlx5_pages_req *req;
 430
 431         req = kzalloc(sizeof(*req), GFP_ATOMIC);
 432         if (!req) {
 433                 mlx5_core_warn(dev, "failed to allocate pages request\n");
 434                 return;
 435         }
 436
 437         req->dev = dev;
 438         req->func_id = func_id;
 439         req->npages = npages;
 440         INIT_WORK(&req->work, pages_work_handler);
 441         queue_work(dev->priv.pg_wq, &req->work);
 442 }
 443
 444 int mlx5_satisfy_startup_pages(struct mlx5_core_dev *dev, int boot)
 445 {
 446         u16 uninitialized_var(func_id);
 447         s32 uninitialized_var(npages);
 448         int err;
 449
 450         err = mlx5_cmd_query_pages(dev, &func_id, &npages, boot);
 451         if (err)
 452                 return err;
 453
 454         mlx5_core_dbg(dev, "requested %d %s pages for func_id 0x%x\n",
 455                       npages, boot ? "boot" : "init", func_id);
 456
 457         return give_pages(dev, func_id, npages, 0);
 458 }
 459
 460 enum {
 461         MLX5_BLKS_FOR_RECLAIM_PAGES = 12
 462 };
 463
 464 static int optimal_reclaimed_pages(void)
 465 {
 466         struct mlx5_cmd_prot_block *block;
 467         struct mlx5_cmd_layout *lay;
 468         int ret;
 469
 470         ret = (sizeof(lay->out) + MLX5_BLKS_FOR_RECLAIM_PAGES * sizeof(block->data) -
 471                sizeof(struct mlx5_manage_pages_outbox)) /
 472                FIELD_SIZEOF(struct mlx5_manage_pages_outbox, pas[0]);
 473
 474         return ret;
 475 }
 476
 477 int mlx5_reclaim_startup_pages(struct mlx5_core_dev *dev)
 478 {
 479         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 480         struct fw_page *fwp;
 481         struct rb_node *p;
 482         int nclaimed = 0;
 483         int err;
 484
 485         do {
 486                 p = rb_first(&dev->priv.page_root);
 487                 if (p) {
 488                         fwp = rb_entry(p, struct fw_page, rb_node);
 489                         err = reclaim_pages(dev, fwp->func_id,
 490                                             optimal_reclaimed_pages(),
 491                                             &nclaimed);
 492                         if (err) {
 493                                 mlx5_core_warn(dev, "failed reclaiming pages (%d)\n",
 494                                                err);
 495                                 return err;
 496                         }
 497                         if (nclaimed)
 498                                 end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 499                 }
 500                 if (time_after(jiffies, end)) {
 501                         mlx5_core_warn(dev, "FW did not return all pages. giving up...\n");
 502                         break;
 503                 }
 504         } while (p);
 505
 506         return 0;
 507 }
 508
 509 void mlx5_pagealloc_init(struct mlx5_core_dev *dev)
 510 {
 511         dev->priv.page_root = RB_ROOT;
 512         INIT_LIST_HEAD(&dev->priv.free_list);
 513 }
 514
 515 void mlx5_pagealloc_cleanup(struct mlx5_core_dev *dev)
 516 {
 517         /* nothing */
 518 }
 519
 520 int mlx5_pagealloc_start(struct mlx5_core_dev *dev)
 521 {
 522         dev->priv.pg_wq = create_singlethread_workqueue("mlx5_page_allocator");
 523         if (!dev->priv.pg_wq)
 524                 return -ENOMEM;
 525
 526         return 0;
 527 }
 528
 529 void mlx5_pagealloc_stop(struct mlx5_core_dev *dev)
 530 {
 531         destroy_workqueue(dev->priv.pg_wq);
 532 }