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         u32     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 / 4096,
 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                 mlx5_core_warn(dev, "\n");
 198         }
 199
 200         fp = list_entry(dev->priv.free_list.next, struct fw_page, list);
 201         n = find_first_bit(&fp->bitmask, 8 * sizeof(fp->bitmask));
 202         if (n >= MLX5_NUM_4K_IN_PAGE) {
 203                 mlx5_core_warn(dev, "alloc 4k bug\n");
 204                 return -ENOENT;
 205         }
 206         clear_bit(n, &fp->bitmask);
 207         fp->free_count--;
 208         if (!fp->free_count)
 209                 list_del(&fp->list);
 210
 211         *addr = fp->addr + n * 4096;
 212
 213         return 0;
 214 }
 215
 216 static void free_4k(struct mlx5_core_dev *dev, u64 addr)
 217 {
 218         struct fw_page *fwp;
 219         int n;
 220
 221         fwp = find_fw_page(dev, addr & PAGE_MASK);
 222         if (!fwp) {
 223                 mlx5_core_warn(dev, "page not found\n");
 224                 return;
 225         }
 226
 227         n = (addr & ~PAGE_MASK) % 4096;
 228         fwp->free_count++;
 229         set_bit(n, &fwp->bitmask);
 230         if (fwp->free_count == MLX5_NUM_4K_IN_PAGE) {
 231                 rb_erase(&fwp->rb_node, &dev->priv.page_root);
 232                 list_del(&fwp->list);
 233                 dma_unmap_page(&dev->pdev->dev, addr, PAGE_SIZE, 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
 247         page = alloc_page(GFP_HIGHUSER);
 248         if (!page) {
 249                 mlx5_core_warn(dev, "failed to allocate page\n");
 250                 return -ENOMEM;
 251         }
 252         addr = dma_map_page(&dev->pdev->dev, page, 0,
 253                             PAGE_SIZE, DMA_BIDIRECTIONAL);
 254         if (dma_mapping_error(&dev->pdev->dev, addr)) {
 255                 mlx5_core_warn(dev, "failed dma mapping page\n");
 256                 err = -ENOMEM;
 257                 goto out_alloc;
 258         }
 259         err = insert_page(dev, addr, page, func_id);
 260         if (err) {
 261                 mlx5_core_err(dev, "failed to track allocated page\n");
 262                 goto out_mapping;
 263         }
 264
 265         return 0;
 266
 267 out_mapping:
 268         dma_unmap_page(&dev->pdev->dev, addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
 269
 270 out_alloc:
 271         __free_page(page);
 272
 273         return err;
 274 }
 275 static int give_pages(struct mlx5_core_dev *dev, u16 func_id, int npages,
 276                       int notify_fail)
 277 {
 278         struct mlx5_manage_pages_inbox *in;
 279         struct mlx5_manage_pages_outbox out;
 280         struct mlx5_manage_pages_inbox *nin;
 281         int inlen;
 282         u64 addr;
 283         int err;
 284         int i;
 285
 286         inlen = sizeof(*in) + npages * sizeof(in->pas[0]);
 287         in = mlx5_vzalloc(inlen);
 288         if (!in) {
 289                 mlx5_core_warn(dev, "vzalloc failed %d\n", inlen);
 290                 return -ENOMEM;
 291         }
 292         memset(&out, 0, sizeof(out));
 293
 294         for (i = 0; i < npages; i++) {
 295 retry:
 296                 err = alloc_4k(dev, &addr);
 297                 if (err) {
 298                         if (err == -ENOMEM)
 299                                 err = alloc_system_page(dev, func_id);
 300                         if (err)
 301                                 goto out_4k;
 302
 303                         goto retry;
 304                 }
 305                 in->pas[i] = cpu_to_be64(addr);
 306         }
 307
 308         in->hdr.opcode = cpu_to_be16(MLX5_CMD_OP_MANAGE_PAGES);
 309         in->hdr.opmod = cpu_to_be16(MLX5_PAGES_GIVE);
 310         in->func_id = cpu_to_be16(func_id);
 311         in->num_entries = cpu_to_be32(npages);
 312         err = mlx5_cmd_exec(dev, in, inlen, &out, sizeof(out));
 313         if (err) {
 314                 mlx5_core_warn(dev, "func_id 0x%x, npages %d, err %d\n", func_id, npages, err);
 315                 goto out_alloc;
 316         }
 317         dev->priv.fw_pages += npages;
 318
 319         if (out.hdr.status) {
 320                 err = mlx5_cmd_status_to_err(&out.hdr);
 321                 if (err) {
 322                         mlx5_core_warn(dev, "func_id 0x%x, npages %d, status %d\n", func_id, npages, out.hdr.status);
 323                         goto out_alloc;
 324                 }
 325         }
 326
 327         mlx5_core_dbg(dev, "err %d\n", err);
 328
 329         goto out_free;
 330
 331 out_alloc:
 332         if (notify_fail) {
 333                 nin = kzalloc(sizeof(*nin), GFP_KERNEL);
 334                 if (!nin) {
 335                         mlx5_core_warn(dev, "allocation failed\n");
 336                         goto out_4k;
 337                 }
 338                 memset(&out, 0, sizeof(out));
 339                 nin->hdr.opcode = cpu_to_be16(MLX5_CMD_OP_MANAGE_PAGES);
 340                 nin->hdr.opmod = cpu_to_be16(MLX5_PAGES_CANT_GIVE);
 341                 if (mlx5_cmd_exec(dev, nin, sizeof(*nin), &out, sizeof(out)))
 342                         mlx5_core_warn(dev, "page notify failed\n");
 343                 kfree(nin);
 344         }
 345
 346 out_4k:
 347         for (i--; i >= 0; i--)
 348                 free_4k(dev, be64_to_cpu(in->pas[i]));
 349 out_free:
 350         mlx5_vfree(in);
 351         return err;
 352 }
 353
 354 static int reclaim_pages(struct mlx5_core_dev *dev, u32 func_id, int npages,
 355                          int *nclaimed)
 356 {
 357         struct mlx5_manage_pages_inbox   in;
 358         struct mlx5_manage_pages_outbox *out;
 359         int num_claimed;
 360         int outlen;
 361         u64 addr;
 362         int err;
 363         int i;
 364
 365         if (nclaimed)
 366                 *nclaimed = 0;
 367
 368         memset(&in, 0, sizeof(in));
 369         outlen = sizeof(*out) + npages * sizeof(out->pas[0]);
 370         out = mlx5_vzalloc(outlen);
 371         if (!out)
 372                 return -ENOMEM;
 373
 374         in.hdr.opcode = cpu_to_be16(MLX5_CMD_OP_MANAGE_PAGES);
 375         in.hdr.opmod = cpu_to_be16(MLX5_PAGES_TAKE);
 376         in.func_id = cpu_to_be16(func_id);
 377         in.num_entries = cpu_to_be32(npages);
 378         mlx5_core_dbg(dev, "npages %d, outlen %d\n", npages, outlen);
 379         err = mlx5_cmd_exec(dev, &in, sizeof(in), out, outlen);
 380         if (err) {
 381                 mlx5_core_err(dev, "failed recliaming pages\n");
 382                 goto out_free;
 383         }
 384         dev->priv.fw_pages -= npages;
 385
 386         if (out->hdr.status) {
 387                 err = mlx5_cmd_status_to_err(&out->hdr);
 388                 goto out_free;
 389         }
 390
 391         num_claimed = be32_to_cpu(out->num_entries);
 392         if (nclaimed)
 393                 *nclaimed = num_claimed;
 394
 395         for (i = 0; i < num_claimed; i++) {
 396                 addr = be64_to_cpu(out->pas[i]);
 397                 free_4k(dev, addr);
 398         }
 399
 400 out_free:
 401         mlx5_vfree(out);
 402         return err;
 403 }
 404
 405 static void pages_work_handler(struct work_struct *work)
 406 {
 407         struct mlx5_pages_req *req = container_of(work, struct mlx5_pages_req, work);
 408         struct mlx5_core_dev *dev = req->dev;
 409         int err = 0;
 410
 411         if (req->npages < 0)
 412                 err = reclaim_pages(dev, req->func_id, -1 * req->npages, NULL);
 413         else if (req->npages > 0)
 414                 err = give_pages(dev, req->func_id, req->npages, 1);
 415
 416         if (err)
 417                 mlx5_core_warn(dev, "%s fail %d\n", req->npages < 0 ?
 418                                "reclaim" : "give", err);
 419
 420         kfree(req);
 421 }
 422
 423 void mlx5_core_req_pages_handler(struct mlx5_core_dev *dev, u16 func_id,
 424                                  s32 npages)
 425 {
 426         struct mlx5_pages_req *req;
 427
 428         req = kzalloc(sizeof(*req), GFP_ATOMIC);
 429         if (!req) {
 430                 mlx5_core_warn(dev, "failed to allocate pages request\n");
 431                 return;
 432         }
 433
 434         req->dev = dev;
 435         req->func_id = func_id;
 436         req->npages = npages;
 437         INIT_WORK(&req->work, pages_work_handler);
 438         queue_work(dev->priv.pg_wq, &req->work);
 439 }
 440
 441 int mlx5_satisfy_startup_pages(struct mlx5_core_dev *dev, int boot)
 442 {
 443         u16 uninitialized_var(func_id);
 444         s32 uninitialized_var(npages);
 445         int err;
 446
 447         err = mlx5_cmd_query_pages(dev, &func_id, &npages, boot);
 448         if (err)
 449                 return err;
 450
 451         mlx5_core_dbg(dev, "requested %d %s pages for func_id 0x%x\n",
 452                       npages, boot ? "boot" : "init", func_id);
 453
 454         return give_pages(dev, func_id, npages, 0);
 455 }
 456
 457 static int optimal_reclaimed_pages(void)
 458 {
 459         struct mlx5_cmd_prot_block *block;
 460         struct mlx5_cmd_layout *lay;
 461         int ret;
 462
 463         ret = (sizeof(lay->in) + sizeof(block->data) -
 464                sizeof(struct mlx5_manage_pages_outbox)) / 8;
 465
 466         return ret;
 467 }
 468
 469 int mlx5_reclaim_startup_pages(struct mlx5_core_dev *dev)
 470 {
 471         unsigned long end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 472         struct fw_page *fwp;
 473         struct rb_node *p;
 474         int nclaimed = 0;
 475         int err;
 476
 477         do {
 478                 p = rb_first(&dev->priv.page_root);
 479                 if (p) {
 480                         fwp = rb_entry(p, struct fw_page, rb_node);
 481                         err = reclaim_pages(dev, fwp->func_id,
 482                                             optimal_reclaimed_pages(),
 483                                             &nclaimed);
 484                         if (err) {
 485                                 mlx5_core_warn(dev, "failed reclaiming pages (%d)\n", err);
 486                                 return err;
 487                         }
 488                         if (nclaimed)
 489                                 end = jiffies + msecs_to_jiffies(MAX_RECLAIM_TIME_MSECS);
 490                 }
 491                 if (time_after(jiffies, end)) {
 492                         mlx5_core_warn(dev, "FW did not return all pages. giving up...\n");
 493                         break;
 494                 }
 495         } while (p);
 496
 497         return 0;
 498 }
 499
 500 void mlx5_pagealloc_init(struct mlx5_core_dev *dev)
 501 {
 502         dev->priv.page_root = RB_ROOT;
 503         INIT_LIST_HEAD(&dev->priv.free_list);
 504 }
 505
 506 void mlx5_pagealloc_cleanup(struct mlx5_core_dev *dev)
 507 {
 508         /* nothing */
 509 }
 510
 511 int mlx5_pagealloc_start(struct mlx5_core_dev *dev)
 512 {
 513         dev->priv.pg_wq = create_singlethread_workqueue("mlx5_page_allocator");
 514         if (!dev->priv.pg_wq)
 515                 return -ENOMEM;
 516
 517         return 0;
 518 }
 519
 520 void mlx5_pagealloc_stop(struct mlx5_core_dev *dev)
 521 {
 522         destroy_workqueue(dev->priv.pg_wq);
 523 }