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(*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         /*
 201          * We only change the mapping of the vmemmap virtual address range
 202          * [@start + PAGE_SIZE, end), so we only need to flush the TLB which
 203          * belongs to the range.
 204          */
 205         flush_tlb_kernel_range(start + PAGE_SIZE, end);
 206
 207         return 0;
 208 }
 209
 210 /*
 211  * Free a vmemmap page. A vmemmap page can be allocated from the memblock
 212  * allocator or buddy allocator. If the PG_reserved flag is set, it means
 213  * that it allocated from the memblock allocator, just free it via the
 214  * free_bootmem_page(). Otherwise, use __free_page().
 215  */
 216 static inline void free_vmemmap_page(struct page *page)
 217 {
 218         if (PageReserved(page))
 219                 free_bootmem_page(page);
 220         else
 221                 __free_page(page);
 222 }
 223
 224 /* Free a list of the vmemmap pages */
 225 static void free_vmemmap_page_list(struct list_head *list)
 226 {
 227         struct page *page, *next;
 228
 229         list_for_each_entry_safe(page, next, list, lru) {
 230                 list_del(&page->lru);
 231                 free_vmemmap_page(page);
 232         }
 233 }
 234
 235 static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
 236                               struct vmemmap_remap_walk *walk)
 237 {
 238         /*
 239          * Remap the tail pages as read-only to catch illegal write operation
 240          * to the tail pages.
 241          */
 242         pgprot_t pgprot = PAGE_KERNEL_RO;
 243         pte_t entry = mk_pte(walk->reuse_page, pgprot);
 244         struct page *page = pte_page(*pte);
 245
 246         list_add_tail(&page->lru, walk->vmemmap_pages);
 247         set_pte_at(&init_mm, addr, pte, entry);
 248 }
 249
 250 /*
 251  * How many struct page structs need to be reset. When we reuse the head
 252  * struct page, the special metadata (e.g. page->flags or page->mapping)
 253  * cannot copy to the tail struct page structs. The invalid value will be
 254  * checked in the free_tail_pages_check(). In order to avoid the message
 255  * of "corrupted mapping in tail page". We need to reset at least 3 (one
 256  * head struct page struct and two tail struct page structs) struct page
 257  * structs.
 258  */
 259 #define NR_RESET_STRUCT_PAGE            3
 260
 261 static inline void reset_struct_pages(struct page *start)
 262 {
 263         struct page *from = start + NR_RESET_STRUCT_PAGE;
 264
 265         BUILD_BUG_ON(NR_RESET_STRUCT_PAGE * 2 > PAGE_SIZE / sizeof(struct page));
 266         memcpy(start, from, sizeof(*from) * NR_RESET_STRUCT_PAGE);
 267 }
 268
 269 static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
 270                                 struct vmemmap_remap_walk *walk)
 271 {
 272         pgprot_t pgprot = PAGE_KERNEL;
 273         struct page *page;
 274         void *to;
 275
 276         BUG_ON(pte_page(*pte) != walk->reuse_page);
 277
 278         page = list_first_entry(walk->vmemmap_pages, struct page, lru);
 279         list_del(&page->lru);
 280         to = page_to_virt(page);
 281         copy_page(to, (void *)walk->reuse_addr);
 282         reset_struct_pages(to);
 283
 284         /*
 285          * Makes sure that preceding stores to the page contents become visible
 286          * before the set_pte_at() write.
 287          */
 288         smp_wmb();
 289         set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
 290 }
 291
 292 /**
 293  * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end)
 294  *                      to the page which @reuse is mapped to, then free vmemmap
 295  *                      which the range are mapped to.
 296  * @start:      start address of the vmemmap virtual address range that we want
 297  *              to remap.
 298  * @end:        end address of the vmemmap virtual address range that we want to
 299  *              remap.
 300  * @reuse:      reuse address.
 301  *
 302  * Return: %0 on success, negative error code otherwise.
 303  */
 304 static int vmemmap_remap_free(unsigned long start, unsigned long end,
 305                               unsigned long reuse)
 306 {
 307         int ret;
 308         LIST_HEAD(vmemmap_pages);
 309         struct vmemmap_remap_walk walk = {
 310                 .remap_pte      = vmemmap_remap_pte,
 311                 .reuse_addr     = reuse,
 312                 .vmemmap_pages  = &vmemmap_pages,
 313         };
 314
 315         /*
 316          * In order to make remapping routine most efficient for the huge pages,
 317          * the routine of vmemmap page table walking has the following rules
 318          * (see more details from the vmemmap_pte_range()):
 319          *
 320          * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE)
 321          *   should be continuous.
 322          * - The @reuse address is part of the range [@reuse, @end) that we are
 323          *   walking which is passed to vmemmap_remap_range().
 324          * - The @reuse address is the first in the complete range.
 325          *
 326          * So we need to make sure that @start and @reuse meet the above rules.
 327          */
 328         BUG_ON(start - reuse != PAGE_SIZE);
 329
 330         mmap_read_lock(&init_mm);
 331         ret = vmemmap_remap_range(reuse, end, &walk);
 332         if (ret && walk.nr_walked) {
 333                 end = reuse + walk.nr_walked * PAGE_SIZE;
 334                 /*
 335                  * vmemmap_pages contains pages from the previous
 336                  * vmemmap_remap_range call which failed.  These
 337                  * are pages which were removed from the vmemmap.
 338                  * They will be restored in the following call.
 339                  */
 340                 walk = (struct vmemmap_remap_walk) {
 341                         .remap_pte      = vmemmap_restore_pte,
 342                         .reuse_addr     = reuse,
 343                         .vmemmap_pages  = &vmemmap_pages,
 344                 };
 345
 346                 vmemmap_remap_range(reuse, end, &walk);
 347         }
 348         mmap_read_unlock(&init_mm);
 349
 350         free_vmemmap_page_list(&vmemmap_pages);
 351
 352         return ret;
 353 }
 354
 355 static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
 356                                    gfp_t gfp_mask, struct list_head *list)
 357 {
 358         unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
 359         int nid = page_to_nid((struct page *)start);
 360         struct page *page, *next;
 361
 362         while (nr_pages--) {
 363                 page = alloc_pages_node(nid, gfp_mask, 0);
 364                 if (!page)
 365                         goto out;
 366                 list_add_tail(&page->lru, list);
 367         }
 368
 369         return 0;
 370 out:
 371         list_for_each_entry_safe(page, next, list, lru)
 372                 __free_pages(page, 0);
 373         return -ENOMEM;
 374 }
 375
 376 /**
 377  * vmemmap_remap_alloc - remap the vmemmap virtual address range [@start, end)
 378  *                       to the page which is from the @vmemmap_pages
 379  *                       respectively.
 380  * @start:      start address of the vmemmap virtual address range that we want
 381  *              to remap.
 382  * @end:        end address of the vmemmap virtual address range that we want to
 383  *              remap.
 384  * @reuse:      reuse address.
 385  * @gfp_mask:   GFP flag for allocating vmemmap pages.
 386  *
 387  * Return: %0 on success, negative error code otherwise.
 388  */
 389 static int vmemmap_remap_alloc(unsigned long start, unsigned long end,
 390                                unsigned long reuse, gfp_t gfp_mask)
 391 {
 392         LIST_HEAD(vmemmap_pages);
 393         struct vmemmap_remap_walk walk = {
 394                 .remap_pte      = vmemmap_restore_pte,
 395                 .reuse_addr     = reuse,
 396                 .vmemmap_pages  = &vmemmap_pages,
 397         };
 398
 399         /* See the comment in the vmemmap_remap_free(). */
 400         BUG_ON(start - reuse != PAGE_SIZE);
 401
 402         if (alloc_vmemmap_page_list(start, end, gfp_mask, &vmemmap_pages))
 403                 return -ENOMEM;
 404
 405         mmap_read_lock(&init_mm);
 406         vmemmap_remap_range(reuse, end, &walk);
 407         mmap_read_unlock(&init_mm);
 408
 409         return 0;
 410 }
 411
 412 DEFINE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
 413 EXPORT_SYMBOL(hugetlb_optimize_vmemmap_key);
 414
 415 static bool vmemmap_optimize_enabled = IS_ENABLED(CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP_DEFAULT_ON);
 416 core_param(hugetlb_free_vmemmap, vmemmap_optimize_enabled, bool, 0);
 417
 418 /**
 419  * hugetlb_vmemmap_restore - restore previously optimized (by
 420  *                           hugetlb_vmemmap_optimize()) vmemmap pages which
 421  *                           will be reallocated and remapped.
 422  * @h:          struct hstate.
 423  * @head:       the head page whose vmemmap pages will be restored.
 424  *
 425  * Return: %0 if @head's vmemmap pages have been reallocated and remapped,
 426  * negative error code otherwise.
 427  */
 428 int hugetlb_vmemmap_restore(const struct hstate *h, struct page *head)
 429 {
 430         int ret;
 431         unsigned long vmemmap_start = (unsigned long)head, vmemmap_end;
 432         unsigned long vmemmap_reuse;
 433
 434         if (!HPageVmemmapOptimized(head))
 435                 return 0;
 436
 437         vmemmap_end     = vmemmap_start + hugetlb_vmemmap_size(h);
 438         vmemmap_reuse   = vmemmap_start;
 439         vmemmap_start   += HUGETLB_VMEMMAP_RESERVE_SIZE;
 440
 441         /*
 442          * The pages which the vmemmap virtual address range [@vmemmap_start,
 443          * @vmemmap_end) are mapped to are freed to the buddy allocator, and
 444          * the range is mapped to the page which @vmemmap_reuse is mapped to.
 445          * When a HugeTLB page is freed to the buddy allocator, previously
 446          * discarded vmemmap pages must be allocated and remapping.
 447          */
 448         ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, vmemmap_reuse,
 449                                   GFP_KERNEL | __GFP_NORETRY | __GFP_THISNODE);
 450         if (!ret) {
 451                 ClearHPageVmemmapOptimized(head);
 452                 static_branch_dec(&hugetlb_optimize_vmemmap_key);
 453         }
 454
 455         return ret;
 456 }
 457
 458 /* Return true iff a HugeTLB whose vmemmap should and can be optimized. */
 459 static bool vmemmap_should_optimize(const struct hstate *h, const struct page *head)
 460 {
 461         if (!READ_ONCE(vmemmap_optimize_enabled))
 462                 return false;
 463
 464         if (!hugetlb_vmemmap_optimizable(h))
 465                 return false;
 466
 467         if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG)) {
 468                 pmd_t *pmdp, pmd;
 469                 struct page *vmemmap_page;
 470                 unsigned long vaddr = (unsigned long)head;
 471
 472                 /*
 473                  * Only the vmemmap page's vmemmap page can be self-hosted.
 474                  * Walking the page tables to find the backing page of the
 475                  * vmemmap page.
 476                  */
 477                 pmdp = pmd_off_k(vaddr);
 478                 /*
 479                  * The READ_ONCE() is used to stabilize *pmdp in a register or
 480                  * on the stack so that it will stop changing under the code.
 481                  * The only concurrent operation where it can be changed is
 482                  * split_vmemmap_huge_pmd() (*pmdp will be stable after this
 483                  * operation).
 484                  */
 485                 pmd = READ_ONCE(*pmdp);
 486                 if (pmd_leaf(pmd))
 487                         vmemmap_page = pmd_page(pmd) + pte_index(vaddr);
 488                 else
 489                         vmemmap_page = pte_page(*pte_offset_kernel(pmdp, vaddr));
 490                 /*
 491                  * Due to HugeTLB alignment requirements and the vmemmap pages
 492                  * being at the start of the hotplugged memory region in
 493                  * memory_hotplug.memmap_on_memory case. Checking any vmemmap
 494                  * page's vmemmap page if it is marked as VmemmapSelfHosted is
 495                  * sufficient.
 496                  *
 497                  * [                  hotplugged memory                  ]
 498                  * [        section        ][...][        section        ]
 499                  * [ vmemmap ][              usable memory               ]
 500                  *   ^   |     |                                        |
 501                  *   +---+     |                                        |
 502                  *     ^       |                                        |
 503                  *     +-------+                                        |
 504                  *          ^                                           |
 505                  *          +-------------------------------------------+
 506                  */
 507                 if (PageVmemmapSelfHosted(vmemmap_page))
 508                         return false;
 509         }
 510
 511         return true;
 512 }
 513
 514 /**
 515  * hugetlb_vmemmap_optimize - optimize @head page's vmemmap pages.
 516  * @h:          struct hstate.
 517  * @head:       the head page whose vmemmap pages will be optimized.
 518  *
 519  * This function only tries to optimize @head's vmemmap pages and does not
 520  * guarantee that the optimization will succeed after it returns. The caller
 521  * can use HPageVmemmapOptimized(@head) to detect if @head's vmemmap pages
 522  * have been optimized.
 523  */
 524 void hugetlb_vmemmap_optimize(const struct hstate *h, struct page *head)
 525 {
 526         unsigned long vmemmap_start = (unsigned long)head, vmemmap_end;
 527         unsigned long vmemmap_reuse;
 528
 529         if (!vmemmap_should_optimize(h, head))
 530                 return;
 531
 532         static_branch_inc(&hugetlb_optimize_vmemmap_key);
 533
 534         vmemmap_end     = vmemmap_start + hugetlb_vmemmap_size(h);
 535         vmemmap_reuse   = vmemmap_start;
 536         vmemmap_start   += HUGETLB_VMEMMAP_RESERVE_SIZE;
 537
 538         /*
 539          * Remap the vmemmap virtual address range [@vmemmap_start, @vmemmap_end)
 540          * to the page which @vmemmap_reuse is mapped to, then free the pages
 541          * which the range [@vmemmap_start, @vmemmap_end] is mapped to.
 542          */
 543         if (vmemmap_remap_free(vmemmap_start, vmemmap_end, vmemmap_reuse))
 544                 static_branch_dec(&hugetlb_optimize_vmemmap_key);
 545         else
 546                 SetHPageVmemmapOptimized(head);
 547 }
 548
 549 static struct ctl_table hugetlb_vmemmap_sysctls[] = {
 550         {
 551                 .procname       = "hugetlb_optimize_vmemmap",
 552                 .data           = &vmemmap_optimize_enabled,
 553                 .maxlen         = sizeof(int),
 554                 .mode           = 0644,
 555                 .proc_handler   = proc_dobool,
 556         },
 557         { }
 558 };
 559
 560 static int __init hugetlb_vmemmap_init(void)
 561 {
 562         /* HUGETLB_VMEMMAP_RESERVE_SIZE should cover all used struct pages */
 563         BUILD_BUG_ON(__NR_USED_SUBPAGE * sizeof(struct page) > HUGETLB_VMEMMAP_RESERVE_SIZE);
 564
 565         if (IS_ENABLED(CONFIG_PROC_SYSCTL)) {
 566                 const struct hstate *h;
 567
 568                 for_each_hstate(h) {
 569                         if (hugetlb_vmemmap_optimizable(h)) {
 570                                 register_sysctl_init("vm", hugetlb_vmemmap_sysctls);
 571                                 break;
 572                         }
 573                 }
 574         }
 575         return 0;
 576 }
 577 late_initcall(hugetlb_vmemmap_init);