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 *head;
  45         pte_t *pgtable;
  46
  47         spin_lock(&init_mm.page_table_lock);
  48         head = pmd_leaf(*pmd) ? pmd_page(*pmd) : NULL;
  49         spin_unlock(&init_mm.page_table_lock);
  50
  51         if (!head)
  52                 return 0;
  53
  54         pgtable = pte_alloc_one_kernel(&init_mm);
  55         if (!pgtable)
  56                 return -ENOMEM;
  57
  58         pmd_populate_kernel(&init_mm, &__pmd, pgtable);
  59
  60         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  61                 pte_t entry, *pte;
  62                 pgprot_t pgprot = PAGE_KERNEL;
  63
  64                 entry = mk_pte(head + i, pgprot);
  65                 pte = pte_offset_kernel(&__pmd, addr);
  66                 set_pte_at(&init_mm, addr, pte, entry);
  67         }
  68
  69         spin_lock(&init_mm.page_table_lock);
  70         if (likely(pmd_leaf(*pmd))) {
  71                 /*
  72                  * Higher order allocations from buddy allocator must be able to
  73                  * be treated as indepdenent small pages (as they can be freed
  74                  * individually).
  75                  */
  76                 if (!PageReserved(head))
  77                         split_page(head, get_order(PMD_SIZE));
  78
  79                 /* Make pte visible before pmd. See comment in pmd_install(). */
  80                 smp_wmb();
  81                 pmd_populate_kernel(&init_mm, pmd, pgtable);
  82                 flush_tlb_kernel_range(start, start + PMD_SIZE);
  83         } else {
  84                 pte_free_kernel(&init_mm, pgtable);
  85         }
  86         spin_unlock(&init_mm.page_table_lock);
  87
  88         return 0;
  89 }
  90
  91 static void vmemmap_pte_range(pmd_t *pmd, unsigned long addr,
  92                               unsigned long end,
  93                               struct vmemmap_remap_walk *walk)
  94 {
  95         pte_t *pte = pte_offset_kernel(pmd, addr);
  96
  97         /*
  98          * The reuse_page is found 'first' in table walk before we start
  99          * remapping (which is calling @walk->remap_pte).
 100          */
 101         if (!walk->reuse_page) {
 102                 walk->reuse_page = pte_page(ptep_get(pte));
 103                 /*
 104                  * Because the reuse address is part of the range that we are
 105                  * walking, skip the reuse address range.
 106                  */
 107                 addr += PAGE_SIZE;
 108                 pte++;
 109                 walk->nr_walked++;
 110         }
 111
 112         for (; addr != end; addr += PAGE_SIZE, pte++) {
 113                 walk->remap_pte(pte, addr, walk);
 114                 walk->nr_walked++;
 115         }
 116 }
 117
 118 static int vmemmap_pmd_range(pud_t *pud, unsigned long addr,
 119                              unsigned long end,
 120                              struct vmemmap_remap_walk *walk)
 121 {
 122         pmd_t *pmd;
 123         unsigned long next;
 124
 125         pmd = pmd_offset(pud, addr);
 126         do {
 127                 int ret;
 128
 129                 ret = split_vmemmap_huge_pmd(pmd, addr & PMD_MASK);
 130                 if (ret)
 131                         return ret;
 132
 133                 next = pmd_addr_end(addr, end);
 134                 vmemmap_pte_range(pmd, addr, next, walk);
 135         } while (pmd++, addr = next, addr != end);
 136
 137         return 0;
 138 }
 139
 140 static int vmemmap_pud_range(p4d_t *p4d, unsigned long addr,
 141                              unsigned long end,
 142                              struct vmemmap_remap_walk *walk)
 143 {
 144         pud_t *pud;
 145         unsigned long next;
 146
 147         pud = pud_offset(p4d, addr);
 148         do {
 149                 int ret;
 150
 151                 next = pud_addr_end(addr, end);
 152                 ret = vmemmap_pmd_range(pud, addr, next, walk);
 153                 if (ret)
 154                         return ret;
 155         } while (pud++, addr = next, addr != end);
 156
 157         return 0;
 158 }
 159
 160 static int vmemmap_p4d_range(pgd_t *pgd, unsigned long addr,
 161                              unsigned long end,
 162                              struct vmemmap_remap_walk *walk)
 163 {
 164         p4d_t *p4d;
 165         unsigned long next;
 166
 167         p4d = p4d_offset(pgd, addr);
 168         do {
 169                 int ret;
 170
 171                 next = p4d_addr_end(addr, end);
 172                 ret = vmemmap_pud_range(p4d, addr, next, walk);
 173                 if (ret)
 174                         return ret;
 175         } while (p4d++, addr = next, addr != end);
 176
 177         return 0;
 178 }
 179
 180 static int vmemmap_remap_range(unsigned long start, unsigned long end,
 181                                struct vmemmap_remap_walk *walk)
 182 {
 183         unsigned long addr = start;
 184         unsigned long next;
 185         pgd_t *pgd;
 186
 187         VM_BUG_ON(!PAGE_ALIGNED(start));
 188         VM_BUG_ON(!PAGE_ALIGNED(end));
 189
 190         pgd = pgd_offset_k(addr);
 191         do {
 192                 int ret;
 193
 194                 next = pgd_addr_end(addr, end);
 195                 ret = vmemmap_p4d_range(pgd, addr, next, walk);
 196                 if (ret)
 197                         return ret;
 198         } while (pgd++, addr = next, addr != end);
 199
 200         flush_tlb_kernel_range(start, end);
 201
 202         return 0;
 203 }
 204
 205 /*
 206  * Free a vmemmap page. A vmemmap page can be allocated from the memblock
 207  * allocator or buddy allocator. If the PG_reserved flag is set, it means
 208  * that it allocated from the memblock allocator, just free it via the
 209  * free_bootmem_page(). Otherwise, use __free_page().
 210  */
 211 static inline void free_vmemmap_page(struct page *page)
 212 {
 213         if (PageReserved(page))
 214                 free_bootmem_page(page);
 215         else
 216                 __free_page(page);
 217 }
 218
 219 /* Free a list of the vmemmap pages */
 220 static void free_vmemmap_page_list(struct list_head *list)
 221 {
 222         struct page *page, *next;
 223
 224         list_for_each_entry_safe(page, next, list, lru)
 225                 free_vmemmap_page(page);
 226 }
 227
 228 static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
 229                               struct vmemmap_remap_walk *walk)
 230 {
 231         /*
 232          * Remap the tail pages as read-only to catch illegal write operation
 233          * to the tail pages.
 234          */
 235         pgprot_t pgprot = PAGE_KERNEL_RO;
 236         struct page *page = pte_page(ptep_get(pte));
 237         pte_t entry;
 238
 239         /* Remapping the head page requires r/w */
 240         if (unlikely(addr == walk->reuse_addr)) {
 241                 pgprot = PAGE_KERNEL;
 242                 list_del(&walk->reuse_page->lru);
 243
 244                 /*
 245                  * Makes sure that preceding stores to the page contents from
 246                  * vmemmap_remap_free() become visible before the set_pte_at()
 247                  * write.
 248                  */
 249                 smp_wmb();
 250         }
 251
 252         entry = mk_pte(walk->reuse_page, pgprot);
 253         list_add_tail(&page->lru, walk->vmemmap_pages);
 254         set_pte_at(&init_mm, addr, pte, entry);
 255 }
 256
 257 /*
 258  * How many struct page structs need to be reset. When we reuse the head
 259  * struct page, the special metadata (e.g. page->flags or page->mapping)
 260  * cannot copy to the tail struct page structs. The invalid value will be
 261  * checked in the free_tail_page_prepare(). In order to avoid the message
 262  * of "corrupted mapping in tail page". We need to reset at least 3 (one
 263  * head struct page struct and two tail struct page structs) struct page
 264  * structs.
 265  */
 266 #define NR_RESET_STRUCT_PAGE            3
 267
 268 static inline void reset_struct_pages(struct page *start)
 269 {
 270         struct page *from = start + NR_RESET_STRUCT_PAGE;
 271
 272         BUILD_BUG_ON(NR_RESET_STRUCT_PAGE * 2 > PAGE_SIZE / sizeof(struct page));
 273         memcpy(start, from, sizeof(*from) * NR_RESET_STRUCT_PAGE);
 274 }
 275
 276 static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
 277                                 struct vmemmap_remap_walk *walk)
 278 {
 279         pgprot_t pgprot = PAGE_KERNEL;
 280         struct page *page;
 281         void *to;
 282
 283         BUG_ON(pte_page(ptep_get(pte)) != walk->reuse_page);
 284
 285         page = list_first_entry(walk->vmemmap_pages, struct page, lru);
 286         list_del(&page->lru);
 287         to = page_to_virt(page);
 288         copy_page(to, (void *)walk->reuse_addr);
 289         reset_struct_pages(to);
 290
 291         /*
 292          * Makes sure that preceding stores to the page contents become visible
 293          * before the set_pte_at() write.
 294          */
 295         smp_wmb();
 296         set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
 297 }
 298
 299 /**
 300  * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end)
 301  *                      to the page which @reuse is mapped to, then free vmemmap
 302  *                      which the range are mapped to.
 303  * @start:      start address of the vmemmap virtual address range that we want
 304  *              to remap.
 305  * @end:        end address of the vmemmap virtual address range that we want to
 306  *              remap.
 307  * @reuse:      reuse address.
 308  *
 309  * Return: %0 on success, negative error code otherwise.
 310  */
 311 static int vmemmap_remap_free(unsigned long start, unsigned long end,
 312                               unsigned long reuse)
 313 {
 314         int ret;
 315         LIST_HEAD(vmemmap_pages);
 316         struct vmemmap_remap_walk walk = {
 317                 .remap_pte      = vmemmap_remap_pte,
 318                 .reuse_addr     = reuse,
 319                 .vmemmap_pages  = &vmemmap_pages,
 320         };
 321         int nid = page_to_nid((struct page *)start);
 322         gfp_t gfp_mask = GFP_KERNEL | __GFP_THISNODE | __GFP_NORETRY |
 323                         __GFP_NOWARN;
 324
 325         /*
 326          * Allocate a new head vmemmap page to avoid breaking a contiguous
 327          * block of struct page memory when freeing it back to page allocator
 328          * in free_vmemmap_page_list(). This will allow the likely contiguous
 329          * struct page backing memory to be kept contiguous and allowing for
 330          * more allocations of hugepages. Fallback to the currently
 331          * mapped head page in case should it fail to allocate.
 332          */
 333         walk.reuse_page = alloc_pages_node(nid, gfp_mask, 0);
 334         if (walk.reuse_page) {
 335                 copy_page(page_to_virt(walk.reuse_page),
 336                           (void *)walk.reuse_addr);
 337                 list_add(&walk.reuse_page->lru, &vmemmap_pages);
 338         }
 339
 340         /*
 341          * In order to make remapping routine most efficient for the huge pages,
 342          * the routine of vmemmap page table walking has the following rules
 343          * (see more details from the vmemmap_pte_range()):
 344          *
 345          * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE)
 346          *   should be continuous.
 347          * - The @reuse address is part of the range [@reuse, @end) that we are
 348          *   walking which is passed to vmemmap_remap_range().
 349          * - The @reuse address is the first in the complete range.
 350          *
 351          * So we need to make sure that @start and @reuse meet the above rules.
 352          */
 353         BUG_ON(start - reuse != PAGE_SIZE);
 354
 355         mmap_read_lock(&init_mm);
 356         ret = vmemmap_remap_range(reuse, end, &walk);
 357         if (ret && walk.nr_walked) {
 358                 end = reuse + walk.nr_walked * PAGE_SIZE;
 359                 /*
 360                  * vmemmap_pages contains pages from the previous
 361                  * vmemmap_remap_range call which failed.  These
 362                  * are pages which were removed from the vmemmap.
 363                  * They will be restored in the following call.
 364                  */
 365                 walk = (struct vmemmap_remap_walk) {
 366                         .remap_pte      = vmemmap_restore_pte,
 367                         .reuse_addr     = reuse,
 368                         .vmemmap_pages  = &vmemmap_pages,
 369                 };
 370
 371                 vmemmap_remap_range(reuse, end, &walk);
 372         }
 373         mmap_read_unlock(&init_mm);
 374
 375         free_vmemmap_page_list(&vmemmap_pages);
 376
 377         return ret;
 378 }
 379
 380 static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
 381                                    struct list_head *list)
 382 {
 383         gfp_t gfp_mask = GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_THISNODE;
 384         unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
 385         int nid = page_to_nid((struct page *)start);
 386         struct page *page, *next;
 387
 388         while (nr_pages--) {
 389                 page = alloc_pages_node(nid, gfp_mask, 0);
 390                 if (!page)
 391                         goto out;
 392                 list_add_tail(&page->lru, list);
 393         }
 394
 395         return 0;
 396 out:
 397         list_for_each_entry_safe(page, next, list, lru)
 398                 __free_page(page);
 399         return -ENOMEM;
 400 }
 401
 402 /**
 403  * vmemmap_remap_alloc - remap the vmemmap virtual address range [@start, end)
 404  *                       to the page which is from the @vmemmap_pages
 405  *                       respectively.
 406  * @start:      start address of the vmemmap virtual address range that we want
 407  *              to remap.
 408  * @end:        end address of the vmemmap virtual address range that we want to
 409  *              remap.
 410  * @reuse:      reuse address.
 411  *
 412  * Return: %0 on success, negative error code otherwise.
 413  */
 414 static int vmemmap_remap_alloc(unsigned long start, unsigned long end,
 415                                unsigned long reuse)
 416 {
 417         LIST_HEAD(vmemmap_pages);
 418         struct vmemmap_remap_walk walk = {
 419                 .remap_pte      = vmemmap_restore_pte,
 420                 .reuse_addr     = reuse,
 421                 .vmemmap_pages  = &vmemmap_pages,
 422         };
 423
 424         /* See the comment in the vmemmap_remap_free(). */
 425         BUG_ON(start - reuse != PAGE_SIZE);
 426
 427         if (alloc_vmemmap_page_list(start, end, &vmemmap_pages))
 428                 return -ENOMEM;
 429
 430         mmap_read_lock(&init_mm);
 431         vmemmap_remap_range(reuse, end, &walk);
 432         mmap_read_unlock(&init_mm);
 433
 434         return 0;
 435 }
 436
 437 DEFINE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
 438 EXPORT_SYMBOL(hugetlb_optimize_vmemmap_key);
 439
 440 static bool vmemmap_optimize_enabled = IS_ENABLED(CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP_DEFAULT_ON);
 441 core_param(hugetlb_free_vmemmap, vmemmap_optimize_enabled, bool, 0);
 442
 443 /**
 444  * hugetlb_vmemmap_restore - restore previously optimized (by
 445  *                           hugetlb_vmemmap_optimize()) vmemmap pages which
 446  *                           will be reallocated and remapped.
 447  * @h:          struct hstate.
 448  * @head:       the head page whose vmemmap pages will be restored.
 449  *
 450  * Return: %0 if @head's vmemmap pages have been reallocated and remapped,
 451  * negative error code otherwise.
 452  */
 453 int hugetlb_vmemmap_restore(const struct hstate *h, struct page *head)
 454 {
 455         int ret;
 456         unsigned long vmemmap_start = (unsigned long)head, vmemmap_end;
 457         unsigned long vmemmap_reuse;
 458
 459         if (!HPageVmemmapOptimized(head))
 460                 return 0;
 461
 462         vmemmap_end     = vmemmap_start + hugetlb_vmemmap_size(h);
 463         vmemmap_reuse   = vmemmap_start;
 464         vmemmap_start   += HUGETLB_VMEMMAP_RESERVE_SIZE;
 465
 466         /*
 467          * The pages which the vmemmap virtual address range [@vmemmap_start,
 468          * @vmemmap_end) are mapped to are freed to the buddy allocator, and
 469          * the range is mapped to the page which @vmemmap_reuse is mapped to.
 470          * When a HugeTLB page is freed to the buddy allocator, previously
 471          * discarded vmemmap pages must be allocated and remapping.
 472          */
 473         ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, vmemmap_reuse);
 474         if (!ret) {
 475                 ClearHPageVmemmapOptimized(head);
 476                 static_branch_dec(&hugetlb_optimize_vmemmap_key);
 477         }
 478
 479         return ret;
 480 }
 481
 482 /* Return true iff a HugeTLB whose vmemmap should and can be optimized. */
 483 static bool vmemmap_should_optimize(const struct hstate *h, const struct page *head)
 484 {
 485         if (!READ_ONCE(vmemmap_optimize_enabled))
 486                 return false;
 487
 488         if (!hugetlb_vmemmap_optimizable(h))
 489                 return false;
 490
 491         if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG)) {
 492                 pmd_t *pmdp, pmd;
 493                 struct page *vmemmap_page;
 494                 unsigned long vaddr = (unsigned long)head;
 495
 496                 /*
 497                  * Only the vmemmap page's vmemmap page can be self-hosted.
 498                  * Walking the page tables to find the backing page of the
 499                  * vmemmap page.
 500                  */
 501                 pmdp = pmd_off_k(vaddr);
 502                 /*
 503                  * The READ_ONCE() is used to stabilize *pmdp in a register or
 504                  * on the stack so that it will stop changing under the code.
 505                  * The only concurrent operation where it can be changed is
 506                  * split_vmemmap_huge_pmd() (*pmdp will be stable after this
 507                  * operation).
 508                  */
 509                 pmd = READ_ONCE(*pmdp);
 510                 if (pmd_leaf(pmd))
 511                         vmemmap_page = pmd_page(pmd) + pte_index(vaddr);
 512                 else
 513                         vmemmap_page = pte_page(*pte_offset_kernel(pmdp, vaddr));
 514                 /*
 515                  * Due to HugeTLB alignment requirements and the vmemmap pages
 516                  * being at the start of the hotplugged memory region in
 517                  * memory_hotplug.memmap_on_memory case. Checking any vmemmap
 518                  * page's vmemmap page if it is marked as VmemmapSelfHosted is
 519                  * sufficient.
 520                  *
 521                  * [                  hotplugged memory                  ]
 522                  * [        section        ][...][        section        ]
 523                  * [ vmemmap ][              usable memory               ]
 524                  *   ^   |     |                                        |
 525                  *   +---+     |                                        |
 526                  *     ^       |                                        |
 527                  *     +-------+                                        |
 528                  *          ^                                           |
 529                  *          +-------------------------------------------+
 530                  */
 531                 if (PageVmemmapSelfHosted(vmemmap_page))
 532                         return false;
 533         }
 534
 535         return true;
 536 }
 537
 538 /**
 539  * hugetlb_vmemmap_optimize - optimize @head page's vmemmap pages.
 540  * @h:          struct hstate.
 541  * @head:       the head page whose vmemmap pages will be optimized.
 542  *
 543  * This function only tries to optimize @head's vmemmap pages and does not
 544  * guarantee that the optimization will succeed after it returns. The caller
 545  * can use HPageVmemmapOptimized(@head) to detect if @head's vmemmap pages
 546  * have been optimized.
 547  */
 548 void hugetlb_vmemmap_optimize(const struct hstate *h, struct page *head)
 549 {
 550         unsigned long vmemmap_start = (unsigned long)head, vmemmap_end;
 551         unsigned long vmemmap_reuse;
 552
 553         if (!vmemmap_should_optimize(h, head))
 554                 return;
 555
 556         static_branch_inc(&hugetlb_optimize_vmemmap_key);
 557
 558         vmemmap_end     = vmemmap_start + hugetlb_vmemmap_size(h);
 559         vmemmap_reuse   = vmemmap_start;
 560         vmemmap_start   += HUGETLB_VMEMMAP_RESERVE_SIZE;
 561
 562         /*
 563          * Remap the vmemmap virtual address range [@vmemmap_start, @vmemmap_end)
 564          * to the page which @vmemmap_reuse is mapped to, then free the pages
 565          * which the range [@vmemmap_start, @vmemmap_end] is mapped to.
 566          */
 567         if (vmemmap_remap_free(vmemmap_start, vmemmap_end, vmemmap_reuse))
 568                 static_branch_dec(&hugetlb_optimize_vmemmap_key);
 569         else
 570                 SetHPageVmemmapOptimized(head);
 571 }
 572
 573 static struct ctl_table hugetlb_vmemmap_sysctls[] = {
 574         {
 575                 .procname       = "hugetlb_optimize_vmemmap",
 576                 .data           = &vmemmap_optimize_enabled,
 577                 .maxlen         = sizeof(vmemmap_optimize_enabled),
 578                 .mode           = 0644,
 579                 .proc_handler   = proc_dobool,
 580         },
 581         { }
 582 };
 583
 584 static int __init hugetlb_vmemmap_init(void)
 585 {
 586         const struct hstate *h;
 587
 588         /* HUGETLB_VMEMMAP_RESERVE_SIZE should cover all used struct pages */
 589         BUILD_BUG_ON(__NR_USED_SUBPAGE * sizeof(struct page) > HUGETLB_VMEMMAP_RESERVE_SIZE);
 590
 591         for_each_hstate(h) {
 592                 if (hugetlb_vmemmap_optimizable(h)) {
 593                         register_sysctl_init("vm", hugetlb_vmemmap_sysctls);
 594                         break;
 595                 }
 596         }
 597         return 0;
 598 }
 599 late_initcall(hugetlb_vmemmap_init);