mm/sparse-vmemmap.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Virtual Memory Map support
   4  *
   5  * (C) 2007 sgi. Christoph Lameter.
   6  *
   7  * Virtual memory maps allow VM primitives pfn_to_page, page_to_pfn,
   8  * virt_to_page, page_address() to be implemented as a base offset
   9  * calculation without memory access.
  10  *
  11  * However, virtual mappings need a page table and TLBs. Many Linux
  12  * architectures already map their physical space using 1-1 mappings
  13  * via TLBs. For those arches the virtual memory map is essentially
  14  * for free if we use the same page size as the 1-1 mappings. In that
  15  * case the overhead consists of a few additional pages that are
  16  * allocated to create a view of memory for vmemmap.
  17  *
  18  * The architecture is expected to provide a vmemmap_populate() function
  19  * to instantiate the mapping.
  20  */
  21 #include <linux/mm.h>
  22 #include <linux/mmzone.h>
  23 #include <linux/memblock.h>
  24 #include <linux/memremap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/slab.h>
  27 #include <linux/spinlock.h>
  28 #include <linux/vmalloc.h>
  29 #include <linux/sched.h>
  30 #include <linux/pgtable.h>
  31 #include <linux/bootmem_info.h>
  32
  33 #include <asm/dma.h>
  34 #include <asm/pgalloc.h>
  35 #include <asm/tlbflush.h>
  36
  37 #ifdef CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP
  38 /**
  39  * struct vmemmap_remap_walk - walk vmemmap page table
  40  *
  41  * @remap_pte:          called for each lowest-level entry (PTE).
  42  * @nr_walked:          the number of walked pte.
  43  * @reuse_page:         the page which is reused for the tail vmemmap pages.
  44  * @reuse_addr:         the virtual address of the @reuse_page page.
  45  * @vmemmap_pages:      the list head of the vmemmap pages that can be freed
  46  *                      or is mapped from.
  47  */
  48 struct vmemmap_remap_walk {
  49         void (*remap_pte)(pte_t *pte, unsigned long addr,
  50                           struct vmemmap_remap_walk *walk);
  51         unsigned long nr_walked;
  52         struct page *reuse_page;
  53         unsigned long reuse_addr;
  54         struct list_head *vmemmap_pages;
  55 };
  56
  57 static int __split_vmemmap_huge_pmd(pmd_t *pmd, unsigned long start)
  58 {
  59         pmd_t __pmd;
  60         int i;
  61         unsigned long addr = start;
  62         struct page *page = pmd_page(*pmd);
  63         pte_t *pgtable = pte_alloc_one_kernel(&init_mm);
  64
  65         if (!pgtable)
  66                 return -ENOMEM;
  67
  68         pmd_populate_kernel(&init_mm, &__pmd, pgtable);
  69
  70         for (i = 0; i < PMD_SIZE / PAGE_SIZE; i++, addr += PAGE_SIZE) {
  71                 pte_t entry, *pte;
  72                 pgprot_t pgprot = PAGE_KERNEL;
  73
  74                 entry = mk_pte(page + i, pgprot);
  75                 pte = pte_offset_kernel(&__pmd, addr);
  76                 set_pte_at(&init_mm, addr, pte, entry);
  77         }
  78
  79         spin_lock(&init_mm.page_table_lock);
  80         if (likely(pmd_leaf(*pmd))) {
  81                 /*
  82                  * Higher order allocations from buddy allocator must be able to
  83                  * be treated as indepdenent small pages (as they can be freed
  84                  * individually).
  85                  */
  86                 if (!PageReserved(page))
  87                         split_page(page, get_order(PMD_SIZE));
  88
  89                 /* Make pte visible before pmd. See comment in pmd_install(). */
  90                 smp_wmb();
  91                 pmd_populate_kernel(&init_mm, pmd, pgtable);
  92                 flush_tlb_kernel_range(start, start + PMD_SIZE);
  93         } else {
  94                 pte_free_kernel(&init_mm, pgtable);
  95         }
  96         spin_unlock(&init_mm.page_table_lock);
  97
  98         return 0;
  99 }
 100
 101 static int split_vmemmap_huge_pmd(pmd_t *pmd, unsigned long start)
 102 {
 103         int leaf;
 104
 105         spin_lock(&init_mm.page_table_lock);
 106         leaf = pmd_leaf(*pmd);
 107         spin_unlock(&init_mm.page_table_lock);
 108
 109         if (!leaf)
 110                 return 0;
 111
 112         return __split_vmemmap_huge_pmd(pmd, start);
 113 }
 114
 115 static void vmemmap_pte_range(pmd_t *pmd, unsigned long addr,
 116                               unsigned long end,
 117                               struct vmemmap_remap_walk *walk)
 118 {
 119         pte_t *pte = pte_offset_kernel(pmd, addr);
 120
 121         /*
 122          * The reuse_page is found 'first' in table walk before we start
 123          * remapping (which is calling @walk->remap_pte).
 124          */
 125         if (!walk->reuse_page) {
 126                 walk->reuse_page = pte_page(*pte);
 127                 /*
 128                  * Because the reuse address is part of the range that we are
 129                  * walking, skip the reuse address range.
 130                  */
 131                 addr += PAGE_SIZE;
 132                 pte++;
 133                 walk->nr_walked++;
 134         }
 135
 136         for (; addr != end; addr += PAGE_SIZE, pte++) {
 137                 walk->remap_pte(pte, addr, walk);
 138                 walk->nr_walked++;
 139         }
 140 }
 141
 142 static int vmemmap_pmd_range(pud_t *pud, unsigned long addr,
 143                              unsigned long end,
 144                              struct vmemmap_remap_walk *walk)
 145 {
 146         pmd_t *pmd;
 147         unsigned long next;
 148
 149         pmd = pmd_offset(pud, addr);
 150         do {
 151                 int ret;
 152
 153                 ret = split_vmemmap_huge_pmd(pmd, addr & PMD_MASK);
 154                 if (ret)
 155                         return ret;
 156
 157                 next = pmd_addr_end(addr, end);
 158                 vmemmap_pte_range(pmd, addr, next, walk);
 159         } while (pmd++, addr = next, addr != end);
 160
 161         return 0;
 162 }
 163
 164 static int vmemmap_pud_range(p4d_t *p4d, unsigned long addr,
 165                              unsigned long end,
 166                              struct vmemmap_remap_walk *walk)
 167 {
 168         pud_t *pud;
 169         unsigned long next;
 170
 171         pud = pud_offset(p4d, addr);
 172         do {
 173                 int ret;
 174
 175                 next = pud_addr_end(addr, end);
 176                 ret = vmemmap_pmd_range(pud, addr, next, walk);
 177                 if (ret)
 178                         return ret;
 179         } while (pud++, addr = next, addr != end);
 180
 181         return 0;
 182 }
 183
 184 static int vmemmap_p4d_range(pgd_t *pgd, unsigned long addr,
 185                              unsigned long end,
 186                              struct vmemmap_remap_walk *walk)
 187 {
 188         p4d_t *p4d;
 189         unsigned long next;
 190
 191         p4d = p4d_offset(pgd, addr);
 192         do {
 193                 int ret;
 194
 195                 next = p4d_addr_end(addr, end);
 196                 ret = vmemmap_pud_range(p4d, addr, next, walk);
 197                 if (ret)
 198                         return ret;
 199         } while (p4d++, addr = next, addr != end);
 200
 201         return 0;
 202 }
 203
 204 static int vmemmap_remap_range(unsigned long start, unsigned long end,
 205                                struct vmemmap_remap_walk *walk)
 206 {
 207         unsigned long addr = start;
 208         unsigned long next;
 209         pgd_t *pgd;
 210
 211         VM_BUG_ON(!IS_ALIGNED(start, PAGE_SIZE));
 212         VM_BUG_ON(!IS_ALIGNED(end, PAGE_SIZE));
 213
 214         pgd = pgd_offset_k(addr);
 215         do {
 216                 int ret;
 217
 218                 next = pgd_addr_end(addr, end);
 219                 ret = vmemmap_p4d_range(pgd, addr, next, walk);
 220                 if (ret)
 221                         return ret;
 222         } while (pgd++, addr = next, addr != end);
 223
 224         /*
 225          * We only change the mapping of the vmemmap virtual address range
 226          * [@start + PAGE_SIZE, end), so we only need to flush the TLB which
 227          * belongs to the range.
 228          */
 229         flush_tlb_kernel_range(start + PAGE_SIZE, end);
 230
 231         return 0;
 232 }
 233
 234 /*
 235  * Free a vmemmap page. A vmemmap page can be allocated from the memblock
 236  * allocator or buddy allocator. If the PG_reserved flag is set, it means
 237  * that it allocated from the memblock allocator, just free it via the
 238  * free_bootmem_page(). Otherwise, use __free_page().
 239  */
 240 static inline void free_vmemmap_page(struct page *page)
 241 {
 242         if (PageReserved(page))
 243                 free_bootmem_page(page);
 244         else
 245                 __free_page(page);
 246 }
 247
 248 /* Free a list of the vmemmap pages */
 249 static void free_vmemmap_page_list(struct list_head *list)
 250 {
 251         struct page *page, *next;
 252
 253         list_for_each_entry_safe(page, next, list, lru) {
 254                 list_del(&page->lru);
 255                 free_vmemmap_page(page);
 256         }
 257 }
 258
 259 static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
 260                               struct vmemmap_remap_walk *walk)
 261 {
 262         /*
 263          * Remap the tail pages as read-only to catch illegal write operation
 264          * to the tail pages.
 265          */
 266         pgprot_t pgprot = PAGE_KERNEL_RO;
 267         pte_t entry = mk_pte(walk->reuse_page, pgprot);
 268         struct page *page = pte_page(*pte);
 269
 270         list_add_tail(&page->lru, walk->vmemmap_pages);
 271         set_pte_at(&init_mm, addr, pte, entry);
 272 }
 273
 274 /*
 275  * How many struct page structs need to be reset. When we reuse the head
 276  * struct page, the special metadata (e.g. page->flags or page->mapping)
 277  * cannot copy to the tail struct page structs. The invalid value will be
 278  * checked in the free_tail_pages_check(). In order to avoid the message
 279  * of "corrupted mapping in tail page". We need to reset at least 3 (one
 280  * head struct page struct and two tail struct page structs) struct page
 281  * structs.
 282  */
 283 #define NR_RESET_STRUCT_PAGE            3
 284
 285 static inline void reset_struct_pages(struct page *start)
 286 {
 287         int i;
 288         struct page *from = start + NR_RESET_STRUCT_PAGE;
 289
 290         for (i = 0; i < NR_RESET_STRUCT_PAGE; i++)
 291                 memcpy(start + i, from, sizeof(*from));
 292 }
 293
 294 static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
 295                                 struct vmemmap_remap_walk *walk)
 296 {
 297         pgprot_t pgprot = PAGE_KERNEL;
 298         struct page *page;
 299         void *to;
 300
 301         BUG_ON(pte_page(*pte) != walk->reuse_page);
 302
 303         page = list_first_entry(walk->vmemmap_pages, struct page, lru);
 304         list_del(&page->lru);
 305         to = page_to_virt(page);
 306         copy_page(to, (void *)walk->reuse_addr);
 307         reset_struct_pages(to);
 308
 309         set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
 310 }
 311
 312 /**
 313  * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end)
 314  *                      to the page which @reuse is mapped to, then free vmemmap
 315  *                      which the range are mapped to.
 316  * @start:      start address of the vmemmap virtual address range that we want
 317  *              to remap.
 318  * @end:        end address of the vmemmap virtual address range that we want to
 319  *              remap.
 320  * @reuse:      reuse address.
 321  *
 322  * Return: %0 on success, negative error code otherwise.
 323  */
 324 int vmemmap_remap_free(unsigned long start, unsigned long end,
 325                        unsigned long reuse)
 326 {
 327         int ret;
 328         LIST_HEAD(vmemmap_pages);
 329         struct vmemmap_remap_walk walk = {
 330                 .remap_pte      = vmemmap_remap_pte,
 331                 .reuse_addr     = reuse,
 332                 .vmemmap_pages  = &vmemmap_pages,
 333         };
 334
 335         /*
 336          * In order to make remapping routine most efficient for the huge pages,
 337          * the routine of vmemmap page table walking has the following rules
 338          * (see more details from the vmemmap_pte_range()):
 339          *
 340          * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE)
 341          *   should be continuous.
 342          * - The @reuse address is part of the range [@reuse, @end) that we are
 343          *   walking which is passed to vmemmap_remap_range().
 344          * - The @reuse address is the first in the complete range.
 345          *
 346          * So we need to make sure that @start and @reuse meet the above rules.
 347          */
 348         BUG_ON(start - reuse != PAGE_SIZE);
 349
 350         mmap_read_lock(&init_mm);
 351         ret = vmemmap_remap_range(reuse, end, &walk);
 352         if (ret && walk.nr_walked) {
 353                 end = reuse + walk.nr_walked * PAGE_SIZE;
 354                 /*
 355                  * vmemmap_pages contains pages from the previous
 356                  * vmemmap_remap_range call which failed.  These
 357                  * are pages which were removed from the vmemmap.
 358                  * They will be restored in the following call.
 359                  */
 360                 walk = (struct vmemmap_remap_walk) {
 361                         .remap_pte      = vmemmap_restore_pte,
 362                         .reuse_addr     = reuse,
 363                         .vmemmap_pages  = &vmemmap_pages,
 364                 };
 365
 366                 vmemmap_remap_range(reuse, end, &walk);
 367         }
 368         mmap_read_unlock(&init_mm);
 369
 370         free_vmemmap_page_list(&vmemmap_pages);
 371
 372         return ret;
 373 }
 374
 375 static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
 376                                    gfp_t gfp_mask, struct list_head *list)
 377 {
 378         unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
 379         int nid = page_to_nid((struct page *)start);
 380         struct page *page, *next;
 381
 382         while (nr_pages--) {
 383                 page = alloc_pages_node(nid, gfp_mask, 0);
 384                 if (!page)
 385                         goto out;
 386                 list_add_tail(&page->lru, list);
 387         }
 388
 389         return 0;
 390 out:
 391         list_for_each_entry_safe(page, next, list, lru)
 392                 __free_pages(page, 0);
 393         return -ENOMEM;
 394 }
 395
 396 /**
 397  * vmemmap_remap_alloc - remap the vmemmap virtual address range [@start, end)
 398  *                       to the page which is from the @vmemmap_pages
 399  *                       respectively.
 400  * @start:      start address of the vmemmap virtual address range that we want
 401  *              to remap.
 402  * @end:        end address of the vmemmap virtual address range that we want to
 403  *              remap.
 404  * @reuse:      reuse address.
 405  * @gfp_mask:   GFP flag for allocating vmemmap pages.
 406  *
 407  * Return: %0 on success, negative error code otherwise.
 408  */
 409 int vmemmap_remap_alloc(unsigned long start, unsigned long end,
 410                         unsigned long reuse, gfp_t gfp_mask)
 411 {
 412         LIST_HEAD(vmemmap_pages);
 413         struct vmemmap_remap_walk walk = {
 414                 .remap_pte      = vmemmap_restore_pte,
 415                 .reuse_addr     = reuse,
 416                 .vmemmap_pages  = &vmemmap_pages,
 417         };
 418
 419         /* See the comment in the vmemmap_remap_free(). */
 420         BUG_ON(start - reuse != PAGE_SIZE);
 421
 422         if (alloc_vmemmap_page_list(start, end, gfp_mask, &vmemmap_pages))
 423                 return -ENOMEM;
 424
 425         mmap_read_lock(&init_mm);
 426         vmemmap_remap_range(reuse, end, &walk);
 427         mmap_read_unlock(&init_mm);
 428
 429         return 0;
 430 }
 431 #endif /* CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP */
 432
 433 /*
 434  * Allocate a block of memory to be used to back the virtual memory map
 435  * or to back the page tables that are used to create the mapping.
 436  * Uses the main allocators if they are available, else bootmem.
 437  */
 438
 439 static void * __ref __earlyonly_bootmem_alloc(int node,
 440                                 unsigned long size,
 441                                 unsigned long align,
 442                                 unsigned long goal)
 443 {
 444         return memblock_alloc_try_nid_raw(size, align, goal,
 445                                                MEMBLOCK_ALLOC_ACCESSIBLE, node);
 446 }
 447
 448 void * __meminit vmemmap_alloc_block(unsigned long size, int node)
 449 {
 450         /* If the main allocator is up use that, fallback to bootmem. */
 451         if (slab_is_available()) {
 452                 gfp_t gfp_mask = GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
 453                 int order = get_order(size);
 454                 static bool warned;
 455                 struct page *page;
 456
 457                 page = alloc_pages_node(node, gfp_mask, order);
 458                 if (page)
 459                         return page_address(page);
 460
 461                 if (!warned) {
 462                         warn_alloc(gfp_mask & ~__GFP_NOWARN, NULL,
 463                                    "vmemmap alloc failure: order:%u", order);
 464                         warned = true;
 465                 }
 466                 return NULL;
 467         } else
 468                 return __earlyonly_bootmem_alloc(node, size, size,
 469                                 __pa(MAX_DMA_ADDRESS));
 470 }
 471
 472 static void * __meminit altmap_alloc_block_buf(unsigned long size,
 473                                                struct vmem_altmap *altmap);
 474
 475 /* need to make sure size is all the same during early stage */
 476 void * __meminit vmemmap_alloc_block_buf(unsigned long size, int node,
 477                                          struct vmem_altmap *altmap)
 478 {
 479         void *ptr;
 480
 481         if (altmap)
 482                 return altmap_alloc_block_buf(size, altmap);
 483
 484         ptr = sparse_buffer_alloc(size);
 485         if (!ptr)
 486                 ptr = vmemmap_alloc_block(size, node);
 487         return ptr;
 488 }
 489
 490 static unsigned long __meminit vmem_altmap_next_pfn(struct vmem_altmap *altmap)
 491 {
 492         return altmap->base_pfn + altmap->reserve + altmap->alloc
 493                 + altmap->align;
 494 }
 495
 496 static unsigned long __meminit vmem_altmap_nr_free(struct vmem_altmap *altmap)
 497 {
 498         unsigned long allocated = altmap->alloc + altmap->align;
 499
 500         if (altmap->free > allocated)
 501                 return altmap->free - allocated;
 502         return 0;
 503 }
 504
 505 static void * __meminit altmap_alloc_block_buf(unsigned long size,
 506                                                struct vmem_altmap *altmap)
 507 {
 508         unsigned long pfn, nr_pfns, nr_align;
 509
 510         if (size & ~PAGE_MASK) {
 511                 pr_warn_once("%s: allocations must be multiple of PAGE_SIZE (%ld)\n",
 512                                 __func__, size);
 513                 return NULL;
 514         }
 515
 516         pfn = vmem_altmap_next_pfn(altmap);
 517         nr_pfns = size >> PAGE_SHIFT;
 518         nr_align = 1UL << find_first_bit(&nr_pfns, BITS_PER_LONG);
 519         nr_align = ALIGN(pfn, nr_align) - pfn;
 520         if (nr_pfns + nr_align > vmem_altmap_nr_free(altmap))
 521                 return NULL;
 522
 523         altmap->alloc += nr_pfns;
 524         altmap->align += nr_align;
 525         pfn += nr_align;
 526
 527         pr_debug("%s: pfn: %#lx alloc: %ld align: %ld nr: %#lx\n",
 528                         __func__, pfn, altmap->alloc, altmap->align, nr_pfns);
 529         return __va(__pfn_to_phys(pfn));
 530 }
 531
 532 void __meminit vmemmap_verify(pte_t *pte, int node,
 533                                 unsigned long start, unsigned long end)
 534 {
 535         unsigned long pfn = pte_pfn(*pte);
 536         int actual_node = early_pfn_to_nid(pfn);
 537
 538         if (node_distance(actual_node, node) > LOCAL_DISTANCE)
 539                 pr_warn_once("[%lx-%lx] potential offnode page_structs\n",
 540                         start, end - 1);
 541 }
 542
 543 pte_t * __meminit vmemmap_pte_populate(pmd_t *pmd, unsigned long addr, int node,
 544                                        struct vmem_altmap *altmap,
 545                                        struct page *reuse)
 546 {
 547         pte_t *pte = pte_offset_kernel(pmd, addr);
 548         if (pte_none(*pte)) {
 549                 pte_t entry;
 550                 void *p;
 551
 552                 if (!reuse) {
 553                         p = vmemmap_alloc_block_buf(PAGE_SIZE, node, altmap);
 554                         if (!p)
 555                                 return NULL;
 556                 } else {
 557                         /*
 558                          * When a PTE/PMD entry is freed from the init_mm
 559                          * there's a a free_pages() call to this page allocated
 560                          * above. Thus this get_page() is paired with the
 561                          * put_page_testzero() on the freeing path.
 562                          * This can only called by certain ZONE_DEVICE path,
 563                          * and through vmemmap_populate_compound_pages() when
 564                          * slab is available.
 565                          */
 566                         get_page(reuse);
 567                         p = page_to_virt(reuse);
 568                 }
 569                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT, PAGE_KERNEL);
 570                 set_pte_at(&init_mm, addr, pte, entry);
 571         }
 572         return pte;
 573 }
 574
 575 static void * __meminit vmemmap_alloc_block_zero(unsigned long size, int node)
 576 {
 577         void *p = vmemmap_alloc_block(size, node);
 578
 579         if (!p)
 580                 return NULL;
 581         memset(p, 0, size);
 582
 583         return p;
 584 }
 585
 586 pmd_t * __meminit vmemmap_pmd_populate(pud_t *pud, unsigned long addr, int node)
 587 {
 588         pmd_t *pmd = pmd_offset(pud, addr);
 589         if (pmd_none(*pmd)) {
 590                 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
 591                 if (!p)
 592                         return NULL;
 593                 pmd_populate_kernel(&init_mm, pmd, p);
 594         }
 595         return pmd;
 596 }
 597
 598 pud_t * __meminit vmemmap_pud_populate(p4d_t *p4d, unsigned long addr, int node)
 599 {
 600         pud_t *pud = pud_offset(p4d, addr);
 601         if (pud_none(*pud)) {
 602                 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
 603                 if (!p)
 604                         return NULL;
 605                 pud_populate(&init_mm, pud, p);
 606         }
 607         return pud;
 608 }
 609
 610 p4d_t * __meminit vmemmap_p4d_populate(pgd_t *pgd, unsigned long addr, int node)
 611 {
 612         p4d_t *p4d = p4d_offset(pgd, addr);
 613         if (p4d_none(*p4d)) {
 614                 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
 615                 if (!p)
 616                         return NULL;
 617                 p4d_populate(&init_mm, p4d, p);
 618         }
 619         return p4d;
 620 }
 621
 622 pgd_t * __meminit vmemmap_pgd_populate(unsigned long addr, int node)
 623 {
 624         pgd_t *pgd = pgd_offset_k(addr);
 625         if (pgd_none(*pgd)) {
 626                 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
 627                 if (!p)
 628                         return NULL;
 629                 pgd_populate(&init_mm, pgd, p);
 630         }
 631         return pgd;
 632 }
 633
 634 static pte_t * __meminit vmemmap_populate_address(unsigned long addr, int node,
 635                                               struct vmem_altmap *altmap,
 636                                               struct page *reuse)
 637 {
 638         pgd_t *pgd;
 639         p4d_t *p4d;
 640         pud_t *pud;
 641         pmd_t *pmd;
 642         pte_t *pte;
 643
 644         pgd = vmemmap_pgd_populate(addr, node);
 645         if (!pgd)
 646                 return NULL;
 647         p4d = vmemmap_p4d_populate(pgd, addr, node);
 648         if (!p4d)
 649                 return NULL;
 650         pud = vmemmap_pud_populate(p4d, addr, node);
 651         if (!pud)
 652                 return NULL;
 653         pmd = vmemmap_pmd_populate(pud, addr, node);
 654         if (!pmd)
 655                 return NULL;
 656         pte = vmemmap_pte_populate(pmd, addr, node, altmap, reuse);
 657         if (!pte)
 658                 return NULL;
 659         vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
 660
 661         return pte;
 662 }
 663
 664 static int __meminit vmemmap_populate_range(unsigned long start,
 665                                             unsigned long end, int node,
 666                                             struct vmem_altmap *altmap,
 667                                             struct page *reuse)
 668 {
 669         unsigned long addr = start;
 670         pte_t *pte;
 671
 672         for (; addr < end; addr += PAGE_SIZE) {
 673                 pte = vmemmap_populate_address(addr, node, altmap, reuse);
 674                 if (!pte)
 675                         return -ENOMEM;
 676         }
 677
 678         return 0;
 679 }
 680
 681 int __meminit vmemmap_populate_basepages(unsigned long start, unsigned long end,
 682                                          int node, struct vmem_altmap *altmap)
 683 {
 684         return vmemmap_populate_range(start, end, node, altmap, NULL);
 685 }
 686
 687 /*
 688  * For compound pages bigger than section size (e.g. x86 1G compound
 689  * pages with 2M subsection size) fill the rest of sections as tail
 690  * pages.
 691  *
 692  * Note that memremap_pages() resets @nr_range value and will increment
 693  * it after each range successful onlining. Thus the value or @nr_range
 694  * at section memmap populate corresponds to the in-progress range
 695  * being onlined here.
 696  */
 697 static bool __meminit reuse_compound_section(unsigned long start_pfn,
 698                                              struct dev_pagemap *pgmap)
 699 {
 700         unsigned long nr_pages = pgmap_vmemmap_nr(pgmap);
 701         unsigned long offset = start_pfn -
 702                 PHYS_PFN(pgmap->ranges[pgmap->nr_range].start);
 703
 704         return !IS_ALIGNED(offset, nr_pages) && nr_pages > PAGES_PER_SUBSECTION;
 705 }
 706
 707 static pte_t * __meminit compound_section_tail_page(unsigned long addr)
 708 {
 709         pte_t *pte;
 710
 711         addr -= PAGE_SIZE;
 712
 713         /*
 714          * Assuming sections are populated sequentially, the previous section's
 715          * page data can be reused.
 716          */
 717         pte = pte_offset_kernel(pmd_off_k(addr), addr);
 718         if (!pte)
 719                 return NULL;
 720
 721         return pte;
 722 }
 723
 724 static int __meminit vmemmap_populate_compound_pages(unsigned long start_pfn,
 725                                                      unsigned long start,
 726                                                      unsigned long end, int node,
 727                                                      struct dev_pagemap *pgmap)
 728 {
 729         unsigned long size, addr;
 730         pte_t *pte;
 731         int rc;
 732
 733         if (reuse_compound_section(start_pfn, pgmap)) {
 734                 pte = compound_section_tail_page(start);
 735                 if (!pte)
 736                         return -ENOMEM;
 737
 738                 /*
 739                  * Reuse the page that was populated in the prior iteration
 740                  * with just tail struct pages.
 741                  */
 742                 return vmemmap_populate_range(start, end, node, NULL,
 743                                               pte_page(*pte));
 744         }
 745
 746         size = min(end - start, pgmap_vmemmap_nr(pgmap) * sizeof(struct page));
 747         for (addr = start; addr < end; addr += size) {
 748                 unsigned long next = addr, last = addr + size;
 749
 750                 /* Populate the head page vmemmap page */
 751                 pte = vmemmap_populate_address(addr, node, NULL, NULL);
 752                 if (!pte)
 753                         return -ENOMEM;
 754
 755                 /* Populate the tail pages vmemmap page */
 756                 next = addr + PAGE_SIZE;
 757                 pte = vmemmap_populate_address(next, node, NULL, NULL);
 758                 if (!pte)
 759                         return -ENOMEM;
 760
 761                 /*
 762                  * Reuse the previous page for the rest of tail pages
 763                  * See layout diagram in Documentation/vm/vmemmap_dedup.rst
 764                  */
 765                 next += PAGE_SIZE;
 766                 rc = vmemmap_populate_range(next, last, node, NULL,
 767                                             pte_page(*pte));
 768                 if (rc)
 769                         return -ENOMEM;
 770         }
 771
 772         return 0;
 773 }
 774
 775 struct page * __meminit __populate_section_memmap(unsigned long pfn,
 776                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
 777                 struct dev_pagemap *pgmap)
 778 {
 779         unsigned long start = (unsigned long) pfn_to_page(pfn);
 780         unsigned long end = start + nr_pages * sizeof(struct page);
 781         int r;
 782
 783         if (WARN_ON_ONCE(!IS_ALIGNED(pfn, PAGES_PER_SUBSECTION) ||
 784                 !IS_ALIGNED(nr_pages, PAGES_PER_SUBSECTION)))
 785                 return NULL;
 786
 787         if (is_power_of_2(sizeof(struct page)) &&
 788             pgmap && pgmap_vmemmap_nr(pgmap) > 1 && !altmap)
 789                 r = vmemmap_populate_compound_pages(pfn, start, end, nid, pgmap);
 790         else
 791                 r = vmemmap_populate(start, end, nid, altmap);
 792
 793         if (r < 0)
 794                 return NULL;
 795
 796         return pfn_to_page(pfn);
 797 }