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 <linux/mmdebug.h>
  17 #include <asm/pgalloc.h>
  18 #include <asm/tlbflush.h>
  19 #include "hugetlb_vmemmap.h"
  20
  21 /**
  22  * struct vmemmap_remap_walk - walk vmemmap page table
  23  *
  24  * @remap_pte:          called for each lowest-level entry (PTE).
  25  * @nr_walked:          the number of walked pte.
  26  * @reuse_page:         the page which is reused for the tail vmemmap pages.
  27  * @reuse_addr:         the virtual address of the @reuse_page page.
  28  * @vmemmap_pages:      the list head of the vmemmap pages that can be freed
  29  *                      or is mapped from.
  30  */
  31 struct vmemmap_remap_walk {
  32         void                    (*remap_pte)(pte_t *pte, unsigned long addr,
  33                                              struct vmemmap_remap_walk *walk);
  34         unsigned long           nr_walked;
  35         struct page             *reuse_page;
  36         unsigned long           reuse_addr;
  37         struct list_head        *vmemmap_pages;
  38 };
  39
  40 static int split_vmemmap_huge_pmd(pmd_t *pmd, unsigned long start)
  41 {
  42         pmd_t __pmd;
  43         int i;
  44         unsigned long addr = start;
  45         struct page *head;
  46         pte_t *pgtable;
  47
  48         spin_lock(&init_mm.page_table_lock);
  49         head = pmd_leaf(*pmd) ? pmd_page(*pmd) : NULL;
  50         spin_unlock(&init_mm.page_table_lock);
  51
  52         if (!head)
  53                 return 0;
  54
  55         pgtable = pte_alloc_one_kernel(&init_mm);
  56         if (!pgtable)
  57                 return -ENOMEM;
  58
  59         pmd_populate_kernel(&init_mm, &__pmd, pgtable);
  60
  61         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  62                 pte_t entry, *pte;
  63                 pgprot_t pgprot = PAGE_KERNEL;
  64
  65                 entry = mk_pte(head + i, pgprot);
  66                 pte = pte_offset_kernel(&__pmd, addr);
  67                 set_pte_at(&init_mm, addr, pte, entry);
  68         }
  69
  70         spin_lock(&init_mm.page_table_lock);
  71         if (likely(pmd_leaf(*pmd))) {
  72                 /*
  73                  * Higher order allocations from buddy allocator must be able to
  74                  * be treated as indepdenent small pages (as they can be freed
  75                  * individually).
  76                  */
  77                 if (!PageReserved(head))
  78                         split_page(head, get_order(PMD_SIZE));
  79
  80                 /* Make pte visible before pmd. See comment in pmd_install(). */
  81                 smp_wmb();
  82                 pmd_populate_kernel(&init_mm, pmd, pgtable);
  83                 flush_tlb_kernel_range(start, start + PMD_SIZE);
  84         } else {
  85                 pte_free_kernel(&init_mm, pgtable);
  86         }
  87         spin_unlock(&init_mm.page_table_lock);
  88
  89         return 0;
  90 }
  91
  92 static void vmemmap_pte_range(pmd_t *pmd, unsigned long addr,
  93                               unsigned long end,
  94                               struct vmemmap_remap_walk *walk)
  95 {
  96         pte_t *pte = pte_offset_kernel(pmd, addr);
  97
  98         /*
  99          * The reuse_page is found 'first' in table walk before we start
 100          * remapping (which is calling @walk->remap_pte).
 101          */
 102         if (!walk->reuse_page) {
 103                 walk->reuse_page = pte_page(ptep_get(pte));
 104                 /*
 105                  * Because the reuse address is part of the range that we are
 106                  * walking, skip the reuse address range.
 107                  */
 108                 addr += PAGE_SIZE;
 109                 pte++;
 110                 walk->nr_walked++;
 111         }
 112
 113         for (; addr != end; addr += PAGE_SIZE, pte++) {
 114                 walk->remap_pte(pte, addr, walk);
 115                 walk->nr_walked++;
 116         }
 117 }
 118
 119 static int vmemmap_pmd_range(pud_t *pud, unsigned long addr,
 120                              unsigned long end,
 121                              struct vmemmap_remap_walk *walk)
 122 {
 123         pmd_t *pmd;
 124         unsigned long next;
 125
 126         pmd = pmd_offset(pud, addr);
 127         do {
 128                 int ret;
 129
 130                 ret = split_vmemmap_huge_pmd(pmd, addr & PMD_MASK);
 131                 if (ret)
 132                         return ret;
 133
 134                 next = pmd_addr_end(addr, end);
 135                 vmemmap_pte_range(pmd, addr, next, walk);
 136         } while (pmd++, addr = next, addr != end);
 137
 138         return 0;
 139 }
 140
 141 static int vmemmap_pud_range(p4d_t *p4d, unsigned long addr,
 142                              unsigned long end,
 143                              struct vmemmap_remap_walk *walk)
 144 {
 145         pud_t *pud;
 146         unsigned long next;
 147
 148         pud = pud_offset(p4d, addr);
 149         do {
 150                 int ret;
 151
 152                 next = pud_addr_end(addr, end);
 153                 ret = vmemmap_pmd_range(pud, addr, next, walk);
 154                 if (ret)
 155                         return ret;
 156         } while (pud++, addr = next, addr != end);
 157
 158         return 0;
 159 }
 160
 161 static int vmemmap_p4d_range(pgd_t *pgd, unsigned long addr,
 162                              unsigned long end,
 163                              struct vmemmap_remap_walk *walk)
 164 {
 165         p4d_t *p4d;
 166         unsigned long next;
 167
 168         p4d = p4d_offset(pgd, addr);
 169         do {
 170                 int ret;
 171
 172                 next = p4d_addr_end(addr, end);
 173                 ret = vmemmap_pud_range(p4d, addr, next, walk);
 174                 if (ret)
 175                         return ret;
 176         } while (p4d++, addr = next, addr != end);
 177
 178         return 0;
 179 }
 180
 181 static int vmemmap_remap_range(unsigned long start, unsigned long end,
 182                                struct vmemmap_remap_walk *walk)
 183 {
 184         unsigned long addr = start;
 185         unsigned long next;
 186         pgd_t *pgd;
 187
 188         VM_BUG_ON(!PAGE_ALIGNED(start));
 189         VM_BUG_ON(!PAGE_ALIGNED(end));
 190
 191         pgd = pgd_offset_k(addr);
 192         do {
 193                 int ret;
 194
 195                 next = pgd_addr_end(addr, end);
 196                 ret = vmemmap_p4d_range(pgd, addr, next, walk);
 197                 if (ret)
 198                         return ret;
 199         } while (pgd++, addr = next, addr != end);
 200
 201         flush_tlb_kernel_range(start, end);
 202
 203         return 0;
 204 }
 205
 206 /*
 207  * Free a vmemmap page. A vmemmap page can be allocated from the memblock
 208  * allocator or buddy allocator. If the PG_reserved flag is set, it means
 209  * that it allocated from the memblock allocator, just free it via the
 210  * free_bootmem_page(). Otherwise, use __free_page().
 211  */
 212 static inline void free_vmemmap_page(struct page *page)
 213 {
 214         if (PageReserved(page))
 215                 free_bootmem_page(page);
 216         else
 217                 __free_page(page);
 218 }
 219
 220 /* Free a list of the vmemmap pages */
 221 static void free_vmemmap_page_list(struct list_head *list)
 222 {
 223         struct page *page, *next;
 224
 225         list_for_each_entry_safe(page, next, list, lru)
 226                 free_vmemmap_page(page);
 227 }
 228
 229 static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
 230                               struct vmemmap_remap_walk *walk)
 231 {
 232         /*
 233          * Remap the tail pages as read-only to catch illegal write operation
 234          * to the tail pages.
 235          */
 236         pgprot_t pgprot = PAGE_KERNEL_RO;
 237         struct page *page = pte_page(ptep_get(pte));
 238         pte_t entry;
 239
 240         /* Remapping the head page requires r/w */
 241         if (unlikely(addr == walk->reuse_addr)) {
 242                 pgprot = PAGE_KERNEL;
 243                 list_del(&walk->reuse_page->lru);
 244
 245                 /*
 246                  * Makes sure that preceding stores to the page contents from
 247                  * vmemmap_remap_free() become visible before the set_pte_at()
 248                  * write.
 249                  */
 250                 smp_wmb();
 251         }
 252
 253         entry = mk_pte(walk->reuse_page, pgprot);
 254         list_add_tail(&page->lru, walk->vmemmap_pages);
 255         set_pte_at(&init_mm, addr, pte, entry);
 256 }
 257
 258 /*
 259  * How many struct page structs need to be reset. When we reuse the head
 260  * struct page, the special metadata (e.g. page->flags or page->mapping)
 261  * cannot copy to the tail struct page structs. The invalid value will be
 262  * checked in the free_tail_page_prepare(). In order to avoid the message
 263  * of "corrupted mapping in tail page". We need to reset at least 3 (one
 264  * head struct page struct and two tail struct page structs) struct page
 265  * structs.
 266  */
 267 #define NR_RESET_STRUCT_PAGE            3
 268
 269 static inline void reset_struct_pages(struct page *start)
 270 {
 271         struct page *from = start + NR_RESET_STRUCT_PAGE;
 272
 273         BUILD_BUG_ON(NR_RESET_STRUCT_PAGE * 2 > PAGE_SIZE / sizeof(struct page));
 274         memcpy(start, from, sizeof(*from) * NR_RESET_STRUCT_PAGE);
 275 }
 276
 277 static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
 278                                 struct vmemmap_remap_walk *walk)
 279 {
 280         pgprot_t pgprot = PAGE_KERNEL;
 281         struct page *page;
 282         void *to;
 283
 284         BUG_ON(pte_page(ptep_get(pte)) != walk->reuse_page);
 285
 286         page = list_first_entry(walk->vmemmap_pages, struct page, lru);
 287         list_del(&page->lru);
 288         to = page_to_virt(page);
 289         copy_page(to, (void *)walk->reuse_addr);
 290         reset_struct_pages(to);
 291
 292         /*
 293          * Makes sure that preceding stores to the page contents become visible
 294          * before the set_pte_at() write.
 295          */
 296         smp_wmb();
 297         set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
 298 }
 299
 300 /**
 301  * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end)
 302  *                      to the page which @reuse is mapped to, then free vmemmap
 303  *                      which the range are mapped to.
 304  * @start:      start address of the vmemmap virtual address range that we want
 305  *              to remap.
 306  * @end:        end address of the vmemmap virtual address range that we want to
 307  *              remap.
 308  * @reuse:      reuse address.
 309  *
 310  * Return: %0 on success, negative error code otherwise.
 311  */
 312 static int vmemmap_remap_free(unsigned long start, unsigned long end,
 313                               unsigned long reuse)
 314 {
 315         int ret;
 316         LIST_HEAD(vmemmap_pages);
 317         struct vmemmap_remap_walk walk = {
 318                 .remap_pte      = vmemmap_remap_pte,
 319                 .reuse_addr     = reuse,
 320                 .vmemmap_pages  = &vmemmap_pages,
 321         };
 322         int nid = page_to_nid((struct page *)start);
 323         gfp_t gfp_mask = GFP_KERNEL | __GFP_THISNODE | __GFP_NORETRY |
 324                         __GFP_NOWARN;
 325
 326         /*
 327          * Allocate a new head vmemmap page to avoid breaking a contiguous
 328          * block of struct page memory when freeing it back to page allocator
 329          * in free_vmemmap_page_list(). This will allow the likely contiguous
 330          * struct page backing memory to be kept contiguous and allowing for
 331          * more allocations of hugepages. Fallback to the currently
 332          * mapped head page in case should it fail to allocate.
 333          */
 334         walk.reuse_page = alloc_pages_node(nid, gfp_mask, 0);
 335         if (walk.reuse_page) {
 336                 copy_page(page_to_virt(walk.reuse_page),
 337                           (void *)walk.reuse_addr);
 338                 list_add(&walk.reuse_page->lru, &vmemmap_pages);
 339         }
 340
 341         /*
 342          * In order to make remapping routine most efficient for the huge pages,
 343          * the routine of vmemmap page table walking has the following rules
 344          * (see more details from the vmemmap_pte_range()):
 345          *
 346          * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE)
 347          *   should be continuous.
 348          * - The @reuse address is part of the range [@reuse, @end) that we are
 349          *   walking which is passed to vmemmap_remap_range().
 350          * - The @reuse address is the first in the complete range.
 351          *
 352          * So we need to make sure that @start and @reuse meet the above rules.
 353          */
 354         BUG_ON(start - reuse != PAGE_SIZE);
 355
 356         mmap_read_lock(&init_mm);
 357         ret = vmemmap_remap_range(reuse, end, &walk);
 358         if (ret && walk.nr_walked) {
 359                 end = reuse + walk.nr_walked * PAGE_SIZE;
 360                 /*
 361                  * vmemmap_pages contains pages from the previous
 362                  * vmemmap_remap_range call which failed.  These
 363                  * are pages which were removed from the vmemmap.
 364                  * They will be restored in the following call.
 365                  */
 366                 walk = (struct vmemmap_remap_walk) {
 367                         .remap_pte      = vmemmap_restore_pte,
 368                         .reuse_addr     = reuse,
 369                         .vmemmap_pages  = &vmemmap_pages,
 370                 };
 371
 372                 vmemmap_remap_range(reuse, end, &walk);
 373         }
 374         mmap_read_unlock(&init_mm);
 375
 376         free_vmemmap_page_list(&vmemmap_pages);
 377
 378         return ret;
 379 }
 380
 381 static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
 382                                    struct list_head *list)
 383 {
 384         gfp_t gfp_mask = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
 385         unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
 386         int nid = page_to_nid((struct page *)start);
 387         struct page *page, *next;
 388
 389         while (nr_pages--) {
 390                 page = alloc_pages_node(nid, gfp_mask, 0);
 391                 if (!page)
 392                         goto out;
 393                 list_add_tail(&page->lru, list);
 394         }
 395
 396         return 0;
 397 out:
 398         list_for_each_entry_safe(page, next, list, lru)
 399                 __free_page(page);
 400         return -ENOMEM;
 401 }
 402
 403 /**
 404  * vmemmap_remap_alloc - remap the vmemmap virtual address range [@start, end)
 405  *                       to the page which is from the @vmemmap_pages
 406  *                       respectively.
 407  * @start:      start address of the vmemmap virtual address range that we want
 408  *              to remap.
 409  * @end:        end address of the vmemmap virtual address range that we want to
 410  *              remap.
 411  * @reuse:      reuse address.
 412  *
 413  * Return: %0 on success, negative error code otherwise.
 414  */
 415 static int vmemmap_remap_alloc(unsigned long start, unsigned long end,
 416                                unsigned long reuse)
 417 {
 418         LIST_HEAD(vmemmap_pages);
 419         struct vmemmap_remap_walk walk = {
 420                 .remap_pte      = vmemmap_restore_pte,
 421                 .reuse_addr     = reuse,
 422                 .vmemmap_pages  = &vmemmap_pages,
 423         };
 424
 425         /* See the comment in the vmemmap_remap_free(). */
 426         BUG_ON(start - reuse != PAGE_SIZE);
 427
 428         if (alloc_vmemmap_page_list(start, end, &vmemmap_pages))
 429                 return -ENOMEM;
 430
 431         mmap_read_lock(&init_mm);
 432         vmemmap_remap_range(reuse, end, &walk);
 433         mmap_read_unlock(&init_mm);
 434
 435         return 0;
 436 }
 437
 438 DEFINE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
 439 EXPORT_SYMBOL(hugetlb_optimize_vmemmap_key);
 440
 441 static bool vmemmap_optimize_enabled = IS_ENABLED(CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP_DEFAULT_ON);
 442 core_param(hugetlb_free_vmemmap, vmemmap_optimize_enabled, bool, 0);
 443
 444 /**
 445  * hugetlb_vmemmap_restore - restore previously optimized (by
 446  *                           hugetlb_vmemmap_optimize()) vmemmap pages which
 447  *                           will be reallocated and remapped.
 448  * @h:          struct hstate.
 449  * @head:       the head page whose vmemmap pages will be restored.
 450  *
 451  * Return: %0 if @head's vmemmap pages have been reallocated and remapped,
 452  * negative error code otherwise.
 453  */
 454 int hugetlb_vmemmap_restore(const struct hstate *h, struct page *head)
 455 {
 456         int ret;
 457         unsigned long vmemmap_start = (unsigned long)head, vmemmap_end;
 458         unsigned long vmemmap_reuse;
 459
 460         VM_WARN_ON_ONCE(!PageHuge(head));
 461         if (!HPageVmemmapOptimized(head))
 462                 return 0;
 463
 464         vmemmap_end     = vmemmap_start + hugetlb_vmemmap_size(h);
 465         vmemmap_reuse   = vmemmap_start;
 466         vmemmap_start   += HUGETLB_VMEMMAP_RESERVE_SIZE;
 467
 468         /*
 469          * The pages which the vmemmap virtual address range [@vmemmap_start,
 470          * @vmemmap_end) are mapped to are freed to the buddy allocator, and
 471          * the range is mapped to the page which @vmemmap_reuse is mapped to.
 472          * When a HugeTLB page is freed to the buddy allocator, previously
 473          * discarded vmemmap pages must be allocated and remapping.
 474          */
 475         ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, vmemmap_reuse);
 476         if (!ret) {
 477                 ClearHPageVmemmapOptimized(head);
 478                 static_branch_dec(&hugetlb_optimize_vmemmap_key);
 479         }
 480
 481         return ret;
 482 }
 483
 484 /* Return true iff a HugeTLB whose vmemmap should and can be optimized. */
 485 static bool vmemmap_should_optimize(const struct hstate *h, const struct page *head)
 486 {
 487         if (!READ_ONCE(vmemmap_optimize_enabled))
 488                 return false;
 489
 490         if (!hugetlb_vmemmap_optimizable(h))
 491                 return false;
 492
 493         if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG)) {
 494                 pmd_t *pmdp, pmd;
 495                 struct page *vmemmap_page;
 496                 unsigned long vaddr = (unsigned long)head;
 497
 498                 /*
 499                  * Only the vmemmap page's vmemmap page can be self-hosted.
 500                  * Walking the page tables to find the backing page of the
 501                  * vmemmap page.
 502                  */
 503                 pmdp = pmd_off_k(vaddr);
 504                 /*
 505                  * The READ_ONCE() is used to stabilize *pmdp in a register or
 506                  * on the stack so that it will stop changing under the code.
 507                  * The only concurrent operation where it can be changed is
 508                  * split_vmemmap_huge_pmd() (*pmdp will be stable after this
 509                  * operation).
 510                  */
 511                 pmd = READ_ONCE(*pmdp);
 512                 if (pmd_leaf(pmd))
 513                         vmemmap_page = pmd_page(pmd) + pte_index(vaddr);
 514                 else
 515                         vmemmap_page = pte_page(*pte_offset_kernel(pmdp, vaddr));
 516                 /*
 517                  * Due to HugeTLB alignment requirements and the vmemmap pages
 518                  * being at the start of the hotplugged memory region in
 519                  * memory_hotplug.memmap_on_memory case. Checking any vmemmap
 520                  * page's vmemmap page if it is marked as VmemmapSelfHosted is
 521                  * sufficient.
 522                  *
 523                  * [                  hotplugged memory                  ]
 524                  * [        section        ][...][        section        ]
 525                  * [ vmemmap ][              usable memory               ]
 526                  *   ^   |     |                                        |
 527                  *   +---+     |                                        |
 528                  *     ^       |                                        |
 529                  *     +-------+                                        |
 530                  *          ^                                           |
 531                  *          +-------------------------------------------+
 532                  */
 533                 if (PageVmemmapSelfHosted(vmemmap_page))
 534                         return false;
 535         }
 536
 537         return true;
 538 }
 539
 540 /**
 541  * hugetlb_vmemmap_optimize - optimize @head page's vmemmap pages.
 542  * @h:          struct hstate.
 543  * @head:       the head page whose vmemmap pages will be optimized.
 544  *
 545  * This function only tries to optimize @head's vmemmap pages and does not
 546  * guarantee that the optimization will succeed after it returns. The caller
 547  * can use HPageVmemmapOptimized(@head) to detect if @head's vmemmap pages
 548  * have been optimized.
 549  */
 550 void hugetlb_vmemmap_optimize(const struct hstate *h, struct page *head)
 551 {
 552         unsigned long vmemmap_start = (unsigned long)head, vmemmap_end;
 553         unsigned long vmemmap_reuse;
 554
 555         VM_WARN_ON_ONCE(!PageHuge(head));
 556         if (!vmemmap_should_optimize(h, head))
 557                 return;
 558
 559         static_branch_inc(&hugetlb_optimize_vmemmap_key);
 560
 561         vmemmap_end     = vmemmap_start + hugetlb_vmemmap_size(h);
 562         vmemmap_reuse   = vmemmap_start;
 563         vmemmap_start   += HUGETLB_VMEMMAP_RESERVE_SIZE;
 564
 565         /*
 566          * Remap the vmemmap virtual address range [@vmemmap_start, @vmemmap_end)
 567          * to the page which @vmemmap_reuse is mapped to, then free the pages
 568          * which the range [@vmemmap_start, @vmemmap_end] is mapped to.
 569          */
 570         if (vmemmap_remap_free(vmemmap_start, vmemmap_end, vmemmap_reuse))
 571                 static_branch_dec(&hugetlb_optimize_vmemmap_key);
 572         else
 573                 SetHPageVmemmapOptimized(head);
 574 }
 575
 576 static struct ctl_table hugetlb_vmemmap_sysctls[] = {
 577         {
 578                 .procname       = "hugetlb_optimize_vmemmap",
 579                 .data           = &vmemmap_optimize_enabled,
 580                 .maxlen         = sizeof(vmemmap_optimize_enabled),
 581                 .mode           = 0644,
 582                 .proc_handler   = proc_dobool,
 583         },
 584         { }
 585 };
 586
 587 static int __init hugetlb_vmemmap_init(void)
 588 {
 589         const struct hstate *h;
 590
 591         /* HUGETLB_VMEMMAP_RESERVE_SIZE should cover all used struct pages */
 592         BUILD_BUG_ON(__NR_USED_SUBPAGE * sizeof(struct page) > HUGETLB_VMEMMAP_RESERVE_SIZE);
 593
 594         for_each_hstate(h) {
 595                 if (hugetlb_vmemmap_optimizable(h)) {
 596                         register_sysctl_init("vm", hugetlb_vmemmap_sysctls);
 597                         break;
 598                 }
 599         }
 600         return 0;
 601 }
 602 late_initcall(hugetlb_vmemmap_init);