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