mm/hugetlb_vmemmap.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * HugeTLB Vmemmap Optimization (HVO)
   4  *
   5  * Copyright (c) 2020, ByteDance. All rights reserved.
   6  *
   7  *     Author: Muchun Song <songmuchun@bytedance.com>
   8  *
   9  * See Documentation/mm/vmemmap_dedup.rst
  10  */
  11 #define pr_fmt(fmt)     "HugeTLB: " fmt
  12
  13 #include <linux/pgtable.h>
  14 #include <linux/moduleparam.h>
  15 #include <linux/bootmem_info.h>
  16 #include <asm/pgalloc.h>
  17 #include <asm/tlbflush.h>
  18 #include "hugetlb_vmemmap.h"
  19
  20 /**
  21  * struct vmemmap_remap_walk - walk vmemmap page table
  22  *
  23  * @remap_pte:          called for each lowest-level entry (PTE).
  24  * @nr_walked:          the number of walked pte.
  25  * @reuse_page:         the page which is reused for the tail vmemmap pages.
  26  * @reuse_addr:         the virtual address of the @reuse_page page.
  27  * @vmemmap_pages:      the list head of the vmemmap pages that can be freed
  28  *                      or is mapped from.
  29  */
  30 struct vmemmap_remap_walk {
  31         void                    (*remap_pte)(pte_t *pte, unsigned long addr,
  32                                              struct vmemmap_remap_walk *walk);
  33         unsigned long           nr_walked;
  34         struct page             *reuse_page;
  35         unsigned long           reuse_addr;
  36         struct list_head        *vmemmap_pages;
  37 };
  38
  39 static int __split_vmemmap_huge_pmd(pmd_t *pmd, unsigned long start)
  40 {
  41         pmd_t __pmd;
  42         int i;
  43         unsigned long addr = start;
  44         struct page *page = pmd_page(*pmd);
  45         pte_t *pgtable = pte_alloc_one_kernel(&init_mm);
  46
  47         if (!pgtable)
  48                 return -ENOMEM;
  49
  50         pmd_populate_kernel(&init_mm, &__pmd, pgtable);
  51
  52         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  53                 pte_t entry, *pte;
  54                 pgprot_t pgprot = PAGE_KERNEL;
  55
  56                 entry = mk_pte(page + i, pgprot);
  57                 pte = pte_offset_kernel(&__pmd, addr);
  58                 set_pte_at(&init_mm, addr, pte, entry);
  59         }
  60
  61         spin_lock(&init_mm.page_table_lock);
  62         if (likely(pmd_leaf(*pmd))) {
  63                 /*
  64                  * Higher order allocations from buddy allocator must be able to
  65                  * be treated as indepdenent small pages (as they can be freed
  66                  * individually).
  67                  */
  68                 if (!PageReserved(page))
  69                         split_page(page, get_order(PMD_SIZE));
  70
  71                 /* Make pte visible before pmd. See comment in pmd_install(). */
  72                 smp_wmb();
  73                 pmd_populate_kernel(&init_mm, pmd, pgtable);
  74                 flush_tlb_kernel_range(start, start + PMD_SIZE);
  75         } else {
  76                 pte_free_kernel(&init_mm, pgtable);
  77         }
  78         spin_unlock(&init_mm.page_table_lock);
  79
  80         return 0;
  81 }
  82
  83 static int split_vmemmap_huge_pmd(pmd_t *pmd, unsigned long start)
  84 {
  85         int leaf;
  86
  87         spin_lock(&init_mm.page_table_lock);
  88         leaf = pmd_leaf(*pmd);
  89         spin_unlock(&init_mm.page_table_lock);
  90
  91         if (!leaf)
  92                 return 0;
  93
  94         return __split_vmemmap_huge_pmd(pmd, start);
  95 }
  96
  97 static void vmemmap_pte_range(pmd_t *pmd, unsigned long addr,
  98                               unsigned long end,
  99                               struct vmemmap_remap_walk *walk)
 100 {
 101         pte_t *pte = pte_offset_kernel(pmd, addr);
 102
 103         /*
 104          * The reuse_page is found 'first' in table walk before we start
 105          * remapping (which is calling @walk->remap_pte).
 106          */
 107         if (!walk->reuse_page) {
 108                 walk->reuse_page = pte_page(*pte);
 109                 /*
 110                  * Because the reuse address is part of the range that we are
 111                  * walking, skip the reuse address range.
 112                  */
 113                 addr += PAGE_SIZE;
 114                 pte++;
 115                 walk->nr_walked++;
 116         }
 117
 118         for (; addr != end; addr += PAGE_SIZE, pte++) {
 119                 walk->remap_pte(pte, addr, walk);
 120                 walk->nr_walked++;
 121         }
 122 }
 123
 124 static int vmemmap_pmd_range(pud_t *pud, unsigned long addr,
 125                              unsigned long end,
 126                              struct vmemmap_remap_walk *walk)
 127 {
 128         pmd_t *pmd;
 129         unsigned long next;
 130
 131         pmd = pmd_offset(pud, addr);
 132         do {
 133                 int ret;
 134
 135                 ret = split_vmemmap_huge_pmd(pmd, addr & PMD_MASK);
 136                 if (ret)
 137                         return ret;
 138
 139                 next = pmd_addr_end(addr, end);
 140                 vmemmap_pte_range(pmd, addr, next, walk);
 141         } while (pmd++, addr = next, addr != end);
 142
 143         return 0;
 144 }
 145
 146 static int vmemmap_pud_range(p4d_t *p4d, unsigned long addr,
 147                              unsigned long end,
 148                              struct vmemmap_remap_walk *walk)
 149 {
 150         pud_t *pud;
 151         unsigned long next;
 152
 153         pud = pud_offset(p4d, addr);
 154         do {
 155                 int ret;
 156
 157                 next = pud_addr_end(addr, end);
 158                 ret = vmemmap_pmd_range(pud, addr, next, walk);
 159                 if (ret)
 160                         return ret;
 161         } while (pud++, addr = next, addr != end);
 162
 163         return 0;
 164 }
 165
 166 static int vmemmap_p4d_range(pgd_t *pgd, unsigned long addr,
 167                              unsigned long end,
 168                              struct vmemmap_remap_walk *walk)
 169 {
 170         p4d_t *p4d;
 171         unsigned long next;
 172
 173         p4d = p4d_offset(pgd, addr);
 174         do {
 175                 int ret;
 176
 177                 next = p4d_addr_end(addr, end);
 178                 ret = vmemmap_pud_range(p4d, addr, next, walk);
 179                 if (ret)
 180                         return ret;
 181         } while (p4d++, addr = next, addr != end);
 182
 183         return 0;
 184 }
 185
 186 static int vmemmap_remap_range(unsigned long start, unsigned long end,
 187                                struct vmemmap_remap_walk *walk)
 188 {
 189         unsigned long addr = start;
 190         unsigned long next;
 191         pgd_t *pgd;
 192
 193         VM_BUG_ON(!PAGE_ALIGNED(start));
 194         VM_BUG_ON(!PAGE_ALIGNED(end));
 195
 196         pgd = pgd_offset_k(addr);
 197         do {
 198                 int ret;
 199
 200                 next = pgd_addr_end(addr, end);
 201                 ret = vmemmap_p4d_range(pgd, addr, next, walk);
 202                 if (ret)
 203                         return ret;
 204         } while (pgd++, addr = next, addr != end);
 205
 206         /*
 207          * We only change the mapping of the vmemmap virtual address range
 208          * [@start + PAGE_SIZE, end), so we only need to flush the TLB which
 209          * belongs to the range.
 210          */
 211         flush_tlb_kernel_range(start + PAGE_SIZE, end);
 212
 213         return 0;
 214 }
 215
 216 /*
 217  * Free a vmemmap page. A vmemmap page can be allocated from the memblock
 218  * allocator or buddy allocator. If the PG_reserved flag is set, it means
 219  * that it allocated from the memblock allocator, just free it via the
 220  * free_bootmem_page(). Otherwise, use __free_page().
 221  */
 222 static inline void free_vmemmap_page(struct page *page)
 223 {
 224         if (PageReserved(page))
 225                 free_bootmem_page(page);
 226         else
 227                 __free_page(page);
 228 }
 229
 230 /* Free a list of the vmemmap pages */
 231 static void free_vmemmap_page_list(struct list_head *list)
 232 {
 233         struct page *page, *next;
 234
 235         list_for_each_entry_safe(page, next, list, lru) {
 236                 list_del(&page->lru);
 237                 free_vmemmap_page(page);
 238         }
 239 }
 240
 241 static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
 242                               struct vmemmap_remap_walk *walk)
 243 {
 244         /*
 245          * Remap the tail pages as read-only to catch illegal write operation
 246          * to the tail pages.
 247          */
 248         pgprot_t pgprot = PAGE_KERNEL_RO;
 249         pte_t entry = mk_pte(walk->reuse_page, pgprot);
 250         struct page *page = pte_page(*pte);
 251
 252         list_add_tail(&page->lru, walk->vmemmap_pages);
 253         set_pte_at(&init_mm, addr, pte, entry);
 254 }
 255
 256 /*
 257  * How many struct page structs need to be reset. When we reuse the head
 258  * struct page, the special metadata (e.g. page->flags or page->mapping)
 259  * cannot copy to the tail struct page structs. The invalid value will be
 260  * checked in the free_tail_pages_check(). In order to avoid the message
 261  * of "corrupted mapping in tail page". We need to reset at least 3 (one
 262  * head struct page struct and two tail struct page structs) struct page
 263  * structs.
 264  */
 265 #define NR_RESET_STRUCT_PAGE            3
 266
 267 static inline void reset_struct_pages(struct page *start)
 268 {
 269         struct page *from = start + NR_RESET_STRUCT_PAGE;
 270
 271         BUILD_BUG_ON(NR_RESET_STRUCT_PAGE * 2 > PAGE_SIZE / sizeof(struct page));
 272         memcpy(start, from, sizeof(*from) * NR_RESET_STRUCT_PAGE);
 273 }
 274
 275 static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
 276                                 struct vmemmap_remap_walk *walk)
 277 {
 278         pgprot_t pgprot = PAGE_KERNEL;
 279         struct page *page;
 280         void *to;
 281
 282         BUG_ON(pte_page(*pte) != walk->reuse_page);
 283
 284         page = list_first_entry(walk->vmemmap_pages, struct page, lru);
 285         list_del(&page->lru);
 286         to = page_to_virt(page);
 287         copy_page(to, (void *)walk->reuse_addr);
 288         reset_struct_pages(to);
 289
 290         /*
 291          * Makes sure that preceding stores to the page contents become visible
 292          * before the set_pte_at() write.
 293          */
 294         smp_wmb();
 295         set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
 296 }
 297
 298 /**
 299  * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end)
 300  *                      to the page which @reuse is mapped to, then free vmemmap
 301  *                      which the range are mapped to.
 302  * @start:      start address of the vmemmap virtual address range that we want
 303  *              to remap.
 304  * @end:        end address of the vmemmap virtual address range that we want to
 305  *              remap.
 306  * @reuse:      reuse address.
 307  *
 308  * Return: %0 on success, negative error code otherwise.
 309  */
 310 static int vmemmap_remap_free(unsigned long start, unsigned long end,
 311                               unsigned long reuse)
 312 {
 313         int ret;
 314         LIST_HEAD(vmemmap_pages);
 315         struct vmemmap_remap_walk walk = {
 316                 .remap_pte      = vmemmap_remap_pte,
 317                 .reuse_addr     = reuse,
 318                 .vmemmap_pages  = &vmemmap_pages,
 319         };
 320
 321         /*
 322          * In order to make remapping routine most efficient for the huge pages,
 323          * the routine of vmemmap page table walking has the following rules
 324          * (see more details from the vmemmap_pte_range()):
 325          *
 326          * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE)
 327          *   should be continuous.
 328          * - The @reuse address is part of the range [@reuse, @end) that we are
 329          *   walking which is passed to vmemmap_remap_range().
 330          * - The @reuse address is the first in the complete range.
 331          *
 332          * So we need to make sure that @start and @reuse meet the above rules.
 333          */
 334         BUG_ON(start - reuse != PAGE_SIZE);
 335
 336         mmap_read_lock(&init_mm);
 337         ret = vmemmap_remap_range(reuse, end, &walk);
 338         if (ret && walk.nr_walked) {
 339                 end = reuse + walk.nr_walked * PAGE_SIZE;
 340                 /*
 341                  * vmemmap_pages contains pages from the previous
 342                  * vmemmap_remap_range call which failed.  These
 343                  * are pages which were removed from the vmemmap.
 344                  * They will be restored in the following call.
 345                  */
 346                 walk = (struct vmemmap_remap_walk) {
 347                         .remap_pte      = vmemmap_restore_pte,
 348                         .reuse_addr     = reuse,
 349                         .vmemmap_pages  = &vmemmap_pages,
 350                 };
 351
 352                 vmemmap_remap_range(reuse, end, &walk);
 353         }
 354         mmap_read_unlock(&init_mm);
 355
 356         free_vmemmap_page_list(&vmemmap_pages);
 357
 358         return ret;
 359 }
 360
 361 static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
 362                                    gfp_t gfp_mask, struct list_head *list)
 363 {
 364         unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
 365         int nid = page_to_nid((struct page *)start);
 366         struct page *page, *next;
 367
 368         while (nr_pages--) {
 369                 page = alloc_pages_node(nid, gfp_mask, 0);
 370                 if (!page)
 371                         goto out;
 372                 list_add_tail(&page->lru, list);
 373         }
 374
 375         return 0;
 376 out:
 377         list_for_each_entry_safe(page, next, list, lru)
 378                 __free_pages(page, 0);
 379         return -ENOMEM;
 380 }
 381
 382 /**
 383  * vmemmap_remap_alloc - remap the vmemmap virtual address range [@start, end)
 384  *                       to the page which is from the @vmemmap_pages
 385  *                       respectively.
 386  * @start:      start address of the vmemmap virtual address range that we want
 387  *              to remap.
 388  * @end:        end address of the vmemmap virtual address range that we want to
 389  *              remap.
 390  * @reuse:      reuse address.
 391  * @gfp_mask:   GFP flag for allocating vmemmap pages.
 392  *
 393  * Return: %0 on success, negative error code otherwise.
 394  */
 395 static int vmemmap_remap_alloc(unsigned long start, unsigned long end,
 396                                unsigned long reuse, gfp_t gfp_mask)
 397 {
 398         LIST_HEAD(vmemmap_pages);
 399         struct vmemmap_remap_walk walk = {
 400                 .remap_pte      = vmemmap_restore_pte,
 401                 .reuse_addr     = reuse,
 402                 .vmemmap_pages  = &vmemmap_pages,
 403         };
 404
 405         /* See the comment in the vmemmap_remap_free(). */
 406         BUG_ON(start - reuse != PAGE_SIZE);
 407
 408         if (alloc_vmemmap_page_list(start, end, gfp_mask, &vmemmap_pages))
 409                 return -ENOMEM;
 410
 411         mmap_read_lock(&init_mm);
 412         vmemmap_remap_range(reuse, end, &walk);
 413         mmap_read_unlock(&init_mm);
 414
 415         return 0;
 416 }
 417
 418 DEFINE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
 419 EXPORT_SYMBOL(hugetlb_optimize_vmemmap_key);
 420
 421 static bool vmemmap_optimize_enabled = IS_ENABLED(CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP_DEFAULT_ON);
 422 core_param(hugetlb_free_vmemmap, vmemmap_optimize_enabled, bool, 0);
 423
 424 /**
 425  * hugetlb_vmemmap_restore - restore previously optimized (by
 426  *                           hugetlb_vmemmap_optimize()) vmemmap pages which
 427  *                           will be reallocated and remapped.
 428  * @h:          struct hstate.
 429  * @head:       the head page whose vmemmap pages will be restored.
 430  *
 431  * Return: %0 if @head's vmemmap pages have been reallocated and remapped,
 432  * negative error code otherwise.
 433  */
 434 int hugetlb_vmemmap_restore(const struct hstate *h, struct page *head)
 435 {
 436         int ret;
 437         unsigned long vmemmap_start = (unsigned long)head, vmemmap_end;
 438         unsigned long vmemmap_reuse;
 439
 440         if (!HPageVmemmapOptimized(head))
 441                 return 0;
 442
 443         vmemmap_end     = vmemmap_start + hugetlb_vmemmap_size(h);
 444         vmemmap_reuse   = vmemmap_start;
 445         vmemmap_start   += HUGETLB_VMEMMAP_RESERVE_SIZE;
 446
 447         /*
 448          * The pages which the vmemmap virtual address range [@vmemmap_start,
 449          * @vmemmap_end) are mapped to are freed to the buddy allocator, and
 450          * the range is mapped to the page which @vmemmap_reuse is mapped to.
 451          * When a HugeTLB page is freed to the buddy allocator, previously
 452          * discarded vmemmap pages must be allocated and remapping.
 453          */
 454         ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, vmemmap_reuse,
 455                                   GFP_KERNEL | __GFP_NORETRY | __GFP_THISNODE);
 456         if (!ret) {
 457                 ClearHPageVmemmapOptimized(head);
 458                 static_branch_dec(&hugetlb_optimize_vmemmap_key);
 459         }
 460
 461         return ret;
 462 }
 463
 464 /* Return true iff a HugeTLB whose vmemmap should and can be optimized. */
 465 static bool vmemmap_should_optimize(const struct hstate *h, const struct page *head)
 466 {
 467         if (!READ_ONCE(vmemmap_optimize_enabled))
 468                 return false;
 469
 470         if (!hugetlb_vmemmap_optimizable(h))
 471                 return false;
 472
 473         if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG)) {
 474                 pmd_t *pmdp, pmd;
 475                 struct page *vmemmap_page;
 476                 unsigned long vaddr = (unsigned long)head;
 477
 478                 /*
 479                  * Only the vmemmap page's vmemmap page can be self-hosted.
 480                  * Walking the page tables to find the backing page of the
 481                  * vmemmap page.
 482                  */
 483                 pmdp = pmd_off_k(vaddr);
 484                 /*
 485                  * The READ_ONCE() is used to stabilize *pmdp in a register or
 486                  * on the stack so that it will stop changing under the code.
 487                  * The only concurrent operation where it can be changed is
 488                  * split_vmemmap_huge_pmd() (*pmdp will be stable after this
 489                  * operation).
 490                  */
 491                 pmd = READ_ONCE(*pmdp);
 492                 if (pmd_leaf(pmd))
 493                         vmemmap_page = pmd_page(pmd) + pte_index(vaddr);
 494                 else
 495                         vmemmap_page = pte_page(*pte_offset_kernel(pmdp, vaddr));
 496                 /*
 497                  * Due to HugeTLB alignment requirements and the vmemmap pages
 498                  * being at the start of the hotplugged memory region in
 499                  * memory_hotplug.memmap_on_memory case. Checking any vmemmap
 500                  * page's vmemmap page if it is marked as VmemmapSelfHosted is
 501                  * sufficient.
 502                  *
 503                  * [                  hotplugged memory                  ]
 504                  * [        section        ][...][        section        ]
 505                  * [ vmemmap ][              usable memory               ]
 506                  *   ^   |     |                                        |
 507                  *   +---+     |                                        |
 508                  *     ^       |                                        |
 509                  *     +-------+                                        |
 510                  *          ^                                           |
 511                  *          +-------------------------------------------+
 512                  */
 513                 if (PageVmemmapSelfHosted(vmemmap_page))
 514                         return false;
 515         }
 516
 517         return true;
 518 }
 519
 520 /**
 521  * hugetlb_vmemmap_optimize - optimize @head page's vmemmap pages.
 522  * @h:          struct hstate.
 523  * @head:       the head page whose vmemmap pages will be optimized.
 524  *
 525  * This function only tries to optimize @head's vmemmap pages and does not
 526  * guarantee that the optimization will succeed after it returns. The caller
 527  * can use HPageVmemmapOptimized(@head) to detect if @head's vmemmap pages
 528  * have been optimized.
 529  */
 530 void hugetlb_vmemmap_optimize(const struct hstate *h, struct page *head)
 531 {
 532         unsigned long vmemmap_start = (unsigned long)head, vmemmap_end;
 533         unsigned long vmemmap_reuse;
 534
 535         if (!vmemmap_should_optimize(h, head))
 536                 return;
 537
 538         static_branch_inc(&hugetlb_optimize_vmemmap_key);
 539
 540         vmemmap_end     = vmemmap_start + hugetlb_vmemmap_size(h);
 541         vmemmap_reuse   = vmemmap_start;
 542         vmemmap_start   += HUGETLB_VMEMMAP_RESERVE_SIZE;
 543
 544         /*
 545          * Remap the vmemmap virtual address range [@vmemmap_start, @vmemmap_end)
 546          * to the page which @vmemmap_reuse is mapped to, then free the pages
 547          * which the range [@vmemmap_start, @vmemmap_end] is mapped to.
 548          */
 549         if (vmemmap_remap_free(vmemmap_start, vmemmap_end, vmemmap_reuse))
 550                 static_branch_dec(&hugetlb_optimize_vmemmap_key);
 551         else
 552                 SetHPageVmemmapOptimized(head);
 553 }
 554
 555 static struct ctl_table hugetlb_vmemmap_sysctls[] = {
 556         {
 557                 .procname       = "hugetlb_optimize_vmemmap",
 558                 .data           = &vmemmap_optimize_enabled,
 559                 .maxlen         = sizeof(int),
 560                 .mode           = 0644,
 561                 .proc_handler   = proc_dobool,
 562         },
 563         { }
 564 };
 565
 566 static int __init hugetlb_vmemmap_init(void)
 567 {
 568         /* HUGETLB_VMEMMAP_RESERVE_SIZE should cover all used struct pages */
 569         BUILD_BUG_ON(__NR_USED_SUBPAGE * sizeof(struct page) > HUGETLB_VMEMMAP_RESERVE_SIZE);
 570
 571         if (IS_ENABLED(CONFIG_PROC_SYSCTL)) {
 572                 const struct hstate *h;
 573
 574                 for_each_hstate(h) {
 575                         if (hugetlb_vmemmap_optimizable(h)) {
 576                                 register_sysctl_init("vm", hugetlb_vmemmap_sysctls);
 577                                 break;
 578                         }
 579                 }
 580         }
 581         return 0;
 582 }
 583 late_initcall(hugetlb_vmemmap_init);