mm/hugetlb_vmemmap.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * HugeTLB Vmemmap Optimization (HVO)
   4  *
   5  * Copyright (c) 2020, ByteDance. All rights reserved.
   6  *
   7  *     Author: Muchun Song <songmuchun@bytedance.com>
   8  *
   9  * See Documentation/mm/vmemmap_dedup.rst
  10  */
  11 #define pr_fmt(fmt)     "HugeTLB: " fmt
  12
  13 #include <linux/pgtable.h>
  14 #include <linux/moduleparam.h>
  15 #include <linux/bootmem_info.h>
  16 #include <asm/pgalloc.h>
  17 #include <asm/tlbflush.h>
  18 #include "hugetlb_vmemmap.h"
  19
  20 /**
  21  * struct vmemmap_remap_walk - walk vmemmap page table
  22  *
  23  * @remap_pte:          called for each lowest-level entry (PTE).
  24  * @nr_walked:          the number of walked pte.
  25  * @reuse_page:         the page which is reused for the tail vmemmap pages.
  26  * @reuse_addr:         the virtual address of the @reuse_page page.
  27  * @vmemmap_pages:      the list head of the vmemmap pages that can be freed
  28  *                      or is mapped from.
  29  */
  30 struct vmemmap_remap_walk {
  31         void                    (*remap_pte)(pte_t *pte, unsigned long addr,
  32                                              struct vmemmap_remap_walk *walk);
  33         unsigned long           nr_walked;
  34         struct page             *reuse_page;
  35         unsigned long           reuse_addr;
  36         struct list_head        *vmemmap_pages;
  37 };
  38
  39 static int __split_vmemmap_huge_pmd(pmd_t *pmd, unsigned long start)
  40 {
  41         pmd_t __pmd;
  42         int i;
  43         unsigned long addr = start;
  44         struct page *page = pmd_page(*pmd);
  45         pte_t *pgtable = pte_alloc_one_kernel(&init_mm);
  46
  47         if (!pgtable)
  48                 return -ENOMEM;
  49
  50         pmd_populate_kernel(&init_mm, &__pmd, pgtable);
  51
  52         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
  53                 pte_t entry, *pte;
  54                 pgprot_t pgprot = PAGE_KERNEL;
  55
  56                 entry = mk_pte(page + i, pgprot);
  57                 pte = pte_offset_kernel(&__pmd, addr);
  58                 set_pte_at(&init_mm, addr, pte, entry);
  59         }
  60
  61         spin_lock(&init_mm.page_table_lock);
  62         if (likely(pmd_leaf(*pmd))) {
  63                 /*
  64                  * Higher order allocations from buddy allocator must be able to
  65                  * be treated as indepdenent small pages (as they can be freed
  66                  * individually).
  67                  */
  68                 if (!PageReserved(page))
  69                         split_page(page, get_order(PMD_SIZE));
  70
  71                 /* Make pte visible before pmd. See comment in pmd_install(). */
  72                 smp_wmb();
  73                 pmd_populate_kernel(&init_mm, pmd, pgtable);
  74                 flush_tlb_kernel_range(start, start + PMD_SIZE);
  75         } else {
  76                 pte_free_kernel(&init_mm, pgtable);
  77         }
  78         spin_unlock(&init_mm.page_table_lock);
  79
  80         return 0;
  81 }
  82
  83 static int split_vmemmap_huge_pmd(pmd_t *pmd, unsigned long start)
  84 {
  85         int leaf;
  86
  87         spin_lock(&init_mm.page_table_lock);
  88         leaf = pmd_leaf(*pmd);
  89         spin_unlock(&init_mm.page_table_lock);
  90
  91         if (!leaf)
  92                 return 0;
  93
  94         return __split_vmemmap_huge_pmd(pmd, start);
  95 }
  96
  97 static void vmemmap_pte_range(pmd_t *pmd, unsigned long addr,
  98                               unsigned long end,
  99                               struct vmemmap_remap_walk *walk)
 100 {
 101         pte_t *pte = pte_offset_kernel(pmd, addr);
 102
 103         /*
 104          * The reuse_page is found 'first' in table walk before we start
 105          * remapping (which is calling @walk->remap_pte).
 106          */
 107         if (!walk->reuse_page) {
 108                 walk->reuse_page = pte_page(*pte);
 109                 /*
 110                  * Because the reuse address is part of the range that we are
 111                  * walking, skip the reuse address range.
 112                  */
 113                 addr += PAGE_SIZE;
 114                 pte++;
 115                 walk->nr_walked++;
 116         }
 117
 118         for (; addr != end; addr += PAGE_SIZE, pte++) {
 119                 walk->remap_pte(pte, addr, walk);
 120                 walk->nr_walked++;
 121         }
 122 }
 123
 124 static int vmemmap_pmd_range(pud_t *pud, unsigned long addr,
 125                              unsigned long end,
 126                              struct vmemmap_remap_walk *walk)
 127 {
 128         pmd_t *pmd;
 129         unsigned long next;
 130
 131         pmd = pmd_offset(pud, addr);
 132         do {
 133                 int ret;
 134
 135                 ret = split_vmemmap_huge_pmd(pmd, addr & PMD_MASK);
 136                 if (ret)
 137                         return ret;
 138
 139                 next = pmd_addr_end(addr, end);
 140                 vmemmap_pte_range(pmd, addr, next, walk);
 141         } while (pmd++, addr = next, addr != end);
 142
 143         return 0;
 144 }
 145
 146 static int vmemmap_pud_range(p4d_t *p4d, unsigned long addr,
 147                              unsigned long end,
 148                              struct vmemmap_remap_walk *walk)
 149 {
 150         pud_t *pud;
 151         unsigned long next;
 152
 153         pud = pud_offset(p4d, addr);
 154         do {
 155                 int ret;
 156
 157                 next = pud_addr_end(addr, end);
 158                 ret = vmemmap_pmd_range(pud, addr, next, walk);
 159                 if (ret)
 160                         return ret;
 161         } while (pud++, addr = next, addr != end);
 162
 163         return 0;
 164 }
 165
 166 static int vmemmap_p4d_range(pgd_t *pgd, unsigned long addr,
 167                              unsigned long end,
 168                              struct vmemmap_remap_walk *walk)
 169 {
 170         p4d_t *p4d;
 171         unsigned long next;
 172
 173         p4d = p4d_offset(pgd, addr);
 174         do {
 175                 int ret;
 176
 177                 next = p4d_addr_end(addr, end);
 178                 ret = vmemmap_pud_range(p4d, addr, next, walk);
 179                 if (ret)
 180                         return ret;
 181         } while (p4d++, addr = next, addr != end);
 182
 183         return 0;
 184 }
 185
 186 static int vmemmap_remap_range(unsigned long start, unsigned long end,
 187                                struct vmemmap_remap_walk *walk)
 188 {
 189         unsigned long addr = start;
 190         unsigned long next;
 191         pgd_t *pgd;
 192
 193         VM_BUG_ON(!PAGE_ALIGNED(start));
 194         VM_BUG_ON(!PAGE_ALIGNED(end));
 195
 196         pgd = pgd_offset_k(addr);
 197         do {
 198                 int ret;
 199
 200                 next = pgd_addr_end(addr, end);
 201                 ret = vmemmap_p4d_range(pgd, addr, next, walk);
 202                 if (ret)
 203                         return ret;
 204         } while (pgd++, addr = next, addr != end);
 205
 206         flush_tlb_kernel_range(start, end);
 207
 208         return 0;
 209 }
 210
 211 /*
 212  * Free a vmemmap page. A vmemmap page can be allocated from the memblock
 213  * allocator or buddy allocator. If the PG_reserved flag is set, it means
 214  * that it allocated from the memblock allocator, just free it via the
 215  * free_bootmem_page(). Otherwise, use __free_page().
 216  */
 217 static inline void free_vmemmap_page(struct page *page)
 218 {
 219         if (PageReserved(page))
 220                 free_bootmem_page(page);
 221         else
 222                 __free_page(page);
 223 }
 224
 225 /* Free a list of the vmemmap pages */
 226 static void free_vmemmap_page_list(struct list_head *list)
 227 {
 228         struct page *page, *next;
 229
 230         list_for_each_entry_safe(page, next, list, lru)
 231                 free_vmemmap_page(page);
 232 }
 233
 234 static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
 235                               struct vmemmap_remap_walk *walk)
 236 {
 237         /*
 238          * Remap the tail pages as read-only to catch illegal write operation
 239          * to the tail pages.
 240          */
 241         pgprot_t pgprot = PAGE_KERNEL_RO;
 242         struct page *page = pte_page(*pte);
 243         pte_t entry;
 244
 245         /* Remapping the head page requires r/w */
 246         if (unlikely(addr == walk->reuse_addr)) {
 247                 pgprot = PAGE_KERNEL;
 248                 list_del(&walk->reuse_page->lru);
 249
 250                 /*
 251                  * Makes sure that preceding stores to the page contents from
 252                  * vmemmap_remap_free() become visible before the set_pte_at()
 253                  * write.
 254                  */
 255                 smp_wmb();
 256         }
 257
 258         entry = mk_pte(walk->reuse_page, pgprot);
 259         list_add_tail(&page->lru, walk->vmemmap_pages);
 260         set_pte_at(&init_mm, addr, pte, entry);
 261 }
 262
 263 /*
 264  * How many struct page structs need to be reset. When we reuse the head
 265  * struct page, the special metadata (e.g. page->flags or page->mapping)
 266  * cannot copy to the tail struct page structs. The invalid value will be
 267  * checked in the free_tail_page_prepare(). In order to avoid the message
 268  * of "corrupted mapping in tail page". We need to reset at least 3 (one
 269  * head struct page struct and two tail struct page structs) struct page
 270  * structs.
 271  */
 272 #define NR_RESET_STRUCT_PAGE            3
 273
 274 static inline void reset_struct_pages(struct page *start)
 275 {
 276         struct page *from = start + NR_RESET_STRUCT_PAGE;
 277
 278         BUILD_BUG_ON(NR_RESET_STRUCT_PAGE * 2 > PAGE_SIZE / sizeof(struct page));
 279         memcpy(start, from, sizeof(*from) * NR_RESET_STRUCT_PAGE);
 280 }
 281
 282 static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
 283                                 struct vmemmap_remap_walk *walk)
 284 {
 285         pgprot_t pgprot = PAGE_KERNEL;
 286         struct page *page;
 287         void *to;
 288
 289         BUG_ON(pte_page(*pte) != walk->reuse_page);
 290
 291         page = list_first_entry(walk->vmemmap_pages, struct page, lru);
 292         list_del(&page->lru);
 293         to = page_to_virt(page);
 294         copy_page(to, (void *)walk->reuse_addr);
 295         reset_struct_pages(to);
 296
 297         /*
 298          * Makes sure that preceding stores to the page contents become visible
 299          * before the set_pte_at() write.
 300          */
 301         smp_wmb();
 302         set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
 303 }
 304
 305 /**
 306  * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end)
 307  *                      to the page which @reuse is mapped to, then free vmemmap
 308  *                      which the range are mapped to.
 309  * @start:      start address of the vmemmap virtual address range that we want
 310  *              to remap.
 311  * @end:        end address of the vmemmap virtual address range that we want to
 312  *              remap.
 313  * @reuse:      reuse address.
 314  *
 315  * Return: %0 on success, negative error code otherwise.
 316  */
 317 static int vmemmap_remap_free(unsigned long start, unsigned long end,
 318                               unsigned long reuse)
 319 {
 320         int ret;
 321         LIST_HEAD(vmemmap_pages);
 322         struct vmemmap_remap_walk walk = {
 323                 .remap_pte      = vmemmap_remap_pte,
 324                 .reuse_addr     = reuse,
 325                 .vmemmap_pages  = &vmemmap_pages,
 326         };
 327         int nid = page_to_nid((struct page *)start);
 328         gfp_t gfp_mask = GFP_KERNEL | __GFP_THISNODE | __GFP_NORETRY |
 329                         __GFP_NOWARN;
 330
 331         /*
 332          * Allocate a new head vmemmap page to avoid breaking a contiguous
 333          * block of struct page memory when freeing it back to page allocator
 334          * in free_vmemmap_page_list(). This will allow the likely contiguous
 335          * struct page backing memory to be kept contiguous and allowing for
 336          * more allocations of hugepages. Fallback to the currently
 337          * mapped head page in case should it fail to allocate.
 338          */
 339         walk.reuse_page = alloc_pages_node(nid, gfp_mask, 0);
 340         if (walk.reuse_page) {
 341                 copy_page(page_to_virt(walk.reuse_page),
 342                           (void *)walk.reuse_addr);
 343                 list_add(&walk.reuse_page->lru, &vmemmap_pages);
 344         }
 345
 346         /*
 347          * In order to make remapping routine most efficient for the huge pages,
 348          * the routine of vmemmap page table walking has the following rules
 349          * (see more details from the vmemmap_pte_range()):
 350          *
 351          * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE)
 352          *   should be continuous.
 353          * - The @reuse address is part of the range [@reuse, @end) that we are
 354          *   walking which is passed to vmemmap_remap_range().
 355          * - The @reuse address is the first in the complete range.
 356          *
 357          * So we need to make sure that @start and @reuse meet the above rules.
 358          */
 359         BUG_ON(start - reuse != PAGE_SIZE);
 360
 361         mmap_read_lock(&init_mm);
 362         ret = vmemmap_remap_range(reuse, end, &walk);
 363         if (ret && walk.nr_walked) {
 364                 end = reuse + walk.nr_walked * PAGE_SIZE;
 365                 /*
 366                  * vmemmap_pages contains pages from the previous
 367                  * vmemmap_remap_range call which failed.  These
 368                  * are pages which were removed from the vmemmap.
 369                  * They will be restored in the following call.
 370                  */
 371                 walk = (struct vmemmap_remap_walk) {
 372                         .remap_pte      = vmemmap_restore_pte,
 373                         .reuse_addr     = reuse,
 374                         .vmemmap_pages  = &vmemmap_pages,
 375                 };
 376
 377                 vmemmap_remap_range(reuse, end, &walk);
 378         }
 379         mmap_read_unlock(&init_mm);
 380
 381         free_vmemmap_page_list(&vmemmap_pages);
 382
 383         return ret;
 384 }
 385
 386 static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
 387                                    gfp_t gfp_mask, struct list_head *list)
 388 {
 389         unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
 390         int nid = page_to_nid((struct page *)start);
 391         struct page *page, *next;
 392
 393         while (nr_pages--) {
 394                 page = alloc_pages_node(nid, gfp_mask, 0);
 395                 if (!page)
 396                         goto out;
 397                 list_add_tail(&page->lru, list);
 398         }
 399
 400         return 0;
 401 out:
 402         list_for_each_entry_safe(page, next, list, lru)
 403                 __free_page(page);
 404         return -ENOMEM;
 405 }
 406
 407 /**
 408  * vmemmap_remap_alloc - remap the vmemmap virtual address range [@start, end)
 409  *                       to the page which is from the @vmemmap_pages
 410  *                       respectively.
 411  * @start:      start address of the vmemmap virtual address range that we want
 412  *              to remap.
 413  * @end:        end address of the vmemmap virtual address range that we want to
 414  *              remap.
 415  * @reuse:      reuse address.
 416  * @gfp_mask:   GFP flag for allocating vmemmap pages.
 417  *
 418  * Return: %0 on success, negative error code otherwise.
 419  */
 420 static int vmemmap_remap_alloc(unsigned long start, unsigned long end,
 421                                unsigned long reuse, gfp_t gfp_mask)
 422 {
 423         LIST_HEAD(vmemmap_pages);
 424         struct vmemmap_remap_walk walk = {
 425                 .remap_pte      = vmemmap_restore_pte,
 426                 .reuse_addr     = reuse,
 427                 .vmemmap_pages  = &vmemmap_pages,
 428         };
 429
 430         /* See the comment in the vmemmap_remap_free(). */
 431         BUG_ON(start - reuse != PAGE_SIZE);
 432
 433         if (alloc_vmemmap_page_list(start, end, gfp_mask, &vmemmap_pages))
 434                 return -ENOMEM;
 435
 436         mmap_read_lock(&init_mm);
 437         vmemmap_remap_range(reuse, end, &walk);
 438         mmap_read_unlock(&init_mm);
 439
 440         return 0;
 441 }
 442
 443 DEFINE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
 444 EXPORT_SYMBOL(hugetlb_optimize_vmemmap_key);
 445
 446 static bool vmemmap_optimize_enabled = IS_ENABLED(CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP_DEFAULT_ON);
 447 core_param(hugetlb_free_vmemmap, vmemmap_optimize_enabled, bool, 0);
 448
 449 /**
 450  * hugetlb_vmemmap_restore - restore previously optimized (by
 451  *                           hugetlb_vmemmap_optimize()) vmemmap pages which
 452  *                           will be reallocated and remapped.
 453  * @h:          struct hstate.
 454  * @head:       the head page whose vmemmap pages will be restored.
 455  *
 456  * Return: %0 if @head's vmemmap pages have been reallocated and remapped,
 457  * negative error code otherwise.
 458  */
 459 int hugetlb_vmemmap_restore(const struct hstate *h, struct page *head)
 460 {
 461         int ret;
 462         unsigned long vmemmap_start = (unsigned long)head, vmemmap_end;
 463         unsigned long vmemmap_reuse;
 464
 465         if (!HPageVmemmapOptimized(head))
 466                 return 0;
 467
 468         vmemmap_end     = vmemmap_start + hugetlb_vmemmap_size(h);
 469         vmemmap_reuse   = vmemmap_start;
 470         vmemmap_start   += HUGETLB_VMEMMAP_RESERVE_SIZE;
 471
 472         /*
 473          * The pages which the vmemmap virtual address range [@vmemmap_start,
 474          * @vmemmap_end) are mapped to are freed to the buddy allocator, and
 475          * the range is mapped to the page which @vmemmap_reuse is mapped to.
 476          * When a HugeTLB page is freed to the buddy allocator, previously
 477          * discarded vmemmap pages must be allocated and remapping.
 478          */
 479         ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, vmemmap_reuse,
 480                                   GFP_KERNEL | __GFP_NORETRY | __GFP_THISNODE);
 481         if (!ret) {
 482                 ClearHPageVmemmapOptimized(head);
 483                 static_branch_dec(&hugetlb_optimize_vmemmap_key);
 484         }
 485
 486         return ret;
 487 }
 488
 489 /* Return true iff a HugeTLB whose vmemmap should and can be optimized. */
 490 static bool vmemmap_should_optimize(const struct hstate *h, const struct page *head)
 491 {
 492         if (!READ_ONCE(vmemmap_optimize_enabled))
 493                 return false;
 494
 495         if (!hugetlb_vmemmap_optimizable(h))
 496                 return false;
 497
 498         if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG)) {
 499                 pmd_t *pmdp, pmd;
 500                 struct page *vmemmap_page;
 501                 unsigned long vaddr = (unsigned long)head;
 502
 503                 /*
 504                  * Only the vmemmap page's vmemmap page can be self-hosted.
 505                  * Walking the page tables to find the backing page of the
 506                  * vmemmap page.
 507                  */
 508                 pmdp = pmd_off_k(vaddr);
 509                 /*
 510                  * The READ_ONCE() is used to stabilize *pmdp in a register or
 511                  * on the stack so that it will stop changing under the code.
 512                  * The only concurrent operation where it can be changed is
 513                  * split_vmemmap_huge_pmd() (*pmdp will be stable after this
 514                  * operation).
 515                  */
 516                 pmd = READ_ONCE(*pmdp);
 517                 if (pmd_leaf(pmd))
 518                         vmemmap_page = pmd_page(pmd) + pte_index(vaddr);
 519                 else
 520                         vmemmap_page = pte_page(*pte_offset_kernel(pmdp, vaddr));
 521                 /*
 522                  * Due to HugeTLB alignment requirements and the vmemmap pages
 523                  * being at the start of the hotplugged memory region in
 524                  * memory_hotplug.memmap_on_memory case. Checking any vmemmap
 525                  * page's vmemmap page if it is marked as VmemmapSelfHosted is
 526                  * sufficient.
 527                  *
 528                  * [                  hotplugged memory                  ]
 529                  * [        section        ][...][        section        ]
 530                  * [ vmemmap ][              usable memory               ]
 531                  *   ^   |     |                                        |
 532                  *   +---+     |                                        |
 533                  *     ^       |                                        |
 534                  *     +-------+                                        |
 535                  *          ^                                           |
 536                  *          +-------------------------------------------+
 537                  */
 538                 if (PageVmemmapSelfHosted(vmemmap_page))
 539                         return false;
 540         }
 541
 542         return true;
 543 }
 544
 545 /**
 546  * hugetlb_vmemmap_optimize - optimize @head page's vmemmap pages.
 547  * @h:          struct hstate.
 548  * @head:       the head page whose vmemmap pages will be optimized.
 549  *
 550  * This function only tries to optimize @head's vmemmap pages and does not
 551  * guarantee that the optimization will succeed after it returns. The caller
 552  * can use HPageVmemmapOptimized(@head) to detect if @head's vmemmap pages
 553  * have been optimized.
 554  */
 555 void hugetlb_vmemmap_optimize(const struct hstate *h, struct page *head)
 556 {
 557         unsigned long vmemmap_start = (unsigned long)head, vmemmap_end;
 558         unsigned long vmemmap_reuse;
 559
 560         if (!vmemmap_should_optimize(h, head))
 561                 return;
 562
 563         static_branch_inc(&hugetlb_optimize_vmemmap_key);
 564
 565         vmemmap_end     = vmemmap_start + hugetlb_vmemmap_size(h);
 566         vmemmap_reuse   = vmemmap_start;
 567         vmemmap_start   += HUGETLB_VMEMMAP_RESERVE_SIZE;
 568
 569         /*
 570          * Remap the vmemmap virtual address range [@vmemmap_start, @vmemmap_end)
 571          * to the page which @vmemmap_reuse is mapped to, then free the pages
 572          * which the range [@vmemmap_start, @vmemmap_end] is mapped to.
 573          */
 574         if (vmemmap_remap_free(vmemmap_start, vmemmap_end, vmemmap_reuse))
 575                 static_branch_dec(&hugetlb_optimize_vmemmap_key);
 576         else
 577                 SetHPageVmemmapOptimized(head);
 578 }
 579
 580 static struct ctl_table hugetlb_vmemmap_sysctls[] = {
 581         {
 582                 .procname       = "hugetlb_optimize_vmemmap",
 583                 .data           = &vmemmap_optimize_enabled,
 584                 .maxlen         = sizeof(vmemmap_optimize_enabled),
 585                 .mode           = 0644,
 586                 .proc_handler   = proc_dobool,
 587         },
 588         { }
 589 };
 590
 591 static int __init hugetlb_vmemmap_init(void)
 592 {
 593         const struct hstate *h;
 594
 595         /* HUGETLB_VMEMMAP_RESERVE_SIZE should cover all used struct pages */
 596         BUILD_BUG_ON(__NR_USED_SUBPAGE * sizeof(struct page) > HUGETLB_VMEMMAP_RESERVE_SIZE);
 597
 598         for_each_hstate(h) {
 599                 if (hugetlb_vmemmap_optimizable(h)) {
 600                         register_sysctl_init("vm", hugetlb_vmemmap_sysctls);
 601                         break;
 602                 }
 603         }
 604         return 0;
 605 }
 606 late_initcall(hugetlb_vmemmap_init);