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 /**
  38  * struct vmemmap_remap_walk - walk vmemmap page table
  39  *
  40  * @remap_pte:          called for each lowest-level entry (PTE).
  41  * @nr_walked:          the number of walked pte.
  42  * @reuse_page:         the page which is reused for the tail vmemmap pages.
  43  * @reuse_addr:         the virtual address of the @reuse_page page.
  44  * @vmemmap_pages:      the list head of the vmemmap pages that can be freed
  45  *                      or is mapped from.
  46  */
  47 struct vmemmap_remap_walk {
  48         void (*remap_pte)(pte_t *pte, unsigned long addr,
  49                           struct vmemmap_remap_walk *walk);
  50         unsigned long nr_walked;
  51         struct page *reuse_page;
  52         unsigned long reuse_addr;
  53         struct list_head *vmemmap_pages;
  54 };
  55
  56 static int split_vmemmap_huge_pmd(pmd_t *pmd, unsigned long start,
  57                                   struct vmemmap_remap_walk *walk)
  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         /* Make pte visible before pmd. See comment in pmd_install(). */
  80         smp_wmb();
  81         pmd_populate_kernel(&init_mm, pmd, pgtable);
  82
  83         flush_tlb_kernel_range(start, start + PMD_SIZE);
  84
  85         return 0;
  86 }
  87
  88 static void vmemmap_pte_range(pmd_t *pmd, unsigned long addr,
  89                               unsigned long end,
  90                               struct vmemmap_remap_walk *walk)
  91 {
  92         pte_t *pte = pte_offset_kernel(pmd, addr);
  93
  94         /*
  95          * The reuse_page is found 'first' in table walk before we start
  96          * remapping (which is calling @walk->remap_pte).
  97          */
  98         if (!walk->reuse_page) {
  99                 walk->reuse_page = pte_page(*pte);
 100                 /*
 101                  * Because the reuse address is part of the range that we are
 102                  * walking, skip the reuse address range.
 103                  */
 104                 addr += PAGE_SIZE;
 105                 pte++;
 106                 walk->nr_walked++;
 107         }
 108
 109         for (; addr != end; addr += PAGE_SIZE, pte++) {
 110                 walk->remap_pte(pte, addr, walk);
 111                 walk->nr_walked++;
 112         }
 113 }
 114
 115 static int vmemmap_pmd_range(pud_t *pud, unsigned long addr,
 116                              unsigned long end,
 117                              struct vmemmap_remap_walk *walk)
 118 {
 119         pmd_t *pmd;
 120         unsigned long next;
 121
 122         pmd = pmd_offset(pud, addr);
 123         do {
 124                 if (pmd_leaf(*pmd)) {
 125                         int ret;
 126
 127                         ret = split_vmemmap_huge_pmd(pmd, addr & PMD_MASK, walk);
 128                         if (ret)
 129                                 return ret;
 130                 }
 131                 next = pmd_addr_end(addr, end);
 132                 vmemmap_pte_range(pmd, addr, next, walk);
 133         } while (pmd++, addr = next, addr != end);
 134
 135         return 0;
 136 }
 137
 138 static int vmemmap_pud_range(p4d_t *p4d, unsigned long addr,
 139                              unsigned long end,
 140                              struct vmemmap_remap_walk *walk)
 141 {
 142         pud_t *pud;
 143         unsigned long next;
 144
 145         pud = pud_offset(p4d, addr);
 146         do {
 147                 int ret;
 148
 149                 next = pud_addr_end(addr, end);
 150                 ret = vmemmap_pmd_range(pud, addr, next, walk);
 151                 if (ret)
 152                         return ret;
 153         } while (pud++, addr = next, addr != end);
 154
 155         return 0;
 156 }
 157
 158 static int vmemmap_p4d_range(pgd_t *pgd, unsigned long addr,
 159                              unsigned long end,
 160                              struct vmemmap_remap_walk *walk)
 161 {
 162         p4d_t *p4d;
 163         unsigned long next;
 164
 165         p4d = p4d_offset(pgd, addr);
 166         do {
 167                 int ret;
 168
 169                 next = p4d_addr_end(addr, end);
 170                 ret = vmemmap_pud_range(p4d, addr, next, walk);
 171                 if (ret)
 172                         return ret;
 173         } while (p4d++, addr = next, addr != end);
 174
 175         return 0;
 176 }
 177
 178 static int vmemmap_remap_range(unsigned long start, unsigned long end,
 179                                struct vmemmap_remap_walk *walk)
 180 {
 181         unsigned long addr = start;
 182         unsigned long next;
 183         pgd_t *pgd;
 184
 185         VM_BUG_ON(!IS_ALIGNED(start, PAGE_SIZE));
 186         VM_BUG_ON(!IS_ALIGNED(end, PAGE_SIZE));
 187
 188         pgd = pgd_offset_k(addr);
 189         do {
 190                 int ret;
 191
 192                 next = pgd_addr_end(addr, end);
 193                 ret = vmemmap_p4d_range(pgd, addr, next, walk);
 194                 if (ret)
 195                         return ret;
 196         } while (pgd++, addr = next, addr != end);
 197
 198         /*
 199          * We only change the mapping of the vmemmap virtual address range
 200          * [@start + PAGE_SIZE, end), so we only need to flush the TLB which
 201          * belongs to the range.
 202          */
 203         flush_tlb_kernel_range(start + PAGE_SIZE, end);
 204
 205         return 0;
 206 }
 207
 208 /*
 209  * Free a vmemmap page. A vmemmap page can be allocated from the memblock
 210  * allocator or buddy allocator. If the PG_reserved flag is set, it means
 211  * that it allocated from the memblock allocator, just free it via the
 212  * free_bootmem_page(). Otherwise, use __free_page().
 213  */
 214 static inline void free_vmemmap_page(struct page *page)
 215 {
 216         if (PageReserved(page))
 217                 free_bootmem_page(page);
 218         else
 219                 __free_page(page);
 220 }
 221
 222 /* Free a list of the vmemmap pages */
 223 static void free_vmemmap_page_list(struct list_head *list)
 224 {
 225         struct page *page, *next;
 226
 227         list_for_each_entry_safe(page, next, list, lru) {
 228                 list_del(&page->lru);
 229                 free_vmemmap_page(page);
 230         }
 231 }
 232
 233 static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
 234                               struct vmemmap_remap_walk *walk)
 235 {
 236         /*
 237          * Remap the tail pages as read-only to catch illegal write operation
 238          * to the tail pages.
 239          */
 240         pgprot_t pgprot = PAGE_KERNEL_RO;
 241         pte_t entry = mk_pte(walk->reuse_page, pgprot);
 242         struct page *page = pte_page(*pte);
 243
 244         list_add_tail(&page->lru, walk->vmemmap_pages);
 245         set_pte_at(&init_mm, addr, pte, entry);
 246 }
 247
 248 /*
 249  * How many struct page structs need to be reset. When we reuse the head
 250  * struct page, the special metadata (e.g. page->flags or page->mapping)
 251  * cannot copy to the tail struct page structs. The invalid value will be
 252  * checked in the free_tail_pages_check(). In order to avoid the message
 253  * of "corrupted mapping in tail page". We need to reset at least 3 (one
 254  * head struct page struct and two tail struct page structs) struct page
 255  * structs.
 256  */
 257 #define NR_RESET_STRUCT_PAGE            3
 258
 259 static inline void reset_struct_pages(struct page *start)
 260 {
 261         int i;
 262         struct page *from = start + NR_RESET_STRUCT_PAGE;
 263
 264         for (i = 0; i < NR_RESET_STRUCT_PAGE; i++)
 265                 memcpy(start + i, from, sizeof(*from));
 266 }
 267
 268 static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
 269                                 struct vmemmap_remap_walk *walk)
 270 {
 271         pgprot_t pgprot = PAGE_KERNEL;
 272         struct page *page;
 273         void *to;
 274
 275         BUG_ON(pte_page(*pte) != walk->reuse_page);
 276
 277         page = list_first_entry(walk->vmemmap_pages, struct page, lru);
 278         list_del(&page->lru);
 279         to = page_to_virt(page);
 280         copy_page(to, (void *)walk->reuse_addr);
 281         reset_struct_pages(to);
 282
 283         set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
 284 }
 285
 286 /**
 287  * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end)
 288  *                      to the page which @reuse is mapped to, then free vmemmap
 289  *                      which the range are mapped to.
 290  * @start:      start address of the vmemmap virtual address range that we want
 291  *              to remap.
 292  * @end:        end address of the vmemmap virtual address range that we want to
 293  *              remap.
 294  * @reuse:      reuse address.
 295  *
 296  * Return: %0 on success, negative error code otherwise.
 297  */
 298 int vmemmap_remap_free(unsigned long start, unsigned long end,
 299                        unsigned long reuse)
 300 {
 301         int ret;
 302         LIST_HEAD(vmemmap_pages);
 303         struct vmemmap_remap_walk walk = {
 304                 .remap_pte      = vmemmap_remap_pte,
 305                 .reuse_addr     = reuse,
 306                 .vmemmap_pages  = &vmemmap_pages,
 307         };
 308
 309         /*
 310          * In order to make remapping routine most efficient for the huge pages,
 311          * the routine of vmemmap page table walking has the following rules
 312          * (see more details from the vmemmap_pte_range()):
 313          *
 314          * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE)
 315          *   should be continuous.
 316          * - The @reuse address is part of the range [@reuse, @end) that we are
 317          *   walking which is passed to vmemmap_remap_range().
 318          * - The @reuse address is the first in the complete range.
 319          *
 320          * So we need to make sure that @start and @reuse meet the above rules.
 321          */
 322         BUG_ON(start - reuse != PAGE_SIZE);
 323
 324         mmap_write_lock(&init_mm);
 325         ret = vmemmap_remap_range(reuse, end, &walk);
 326         mmap_write_downgrade(&init_mm);
 327
 328         if (ret && walk.nr_walked) {
 329                 end = reuse + walk.nr_walked * PAGE_SIZE;
 330                 /*
 331                  * vmemmap_pages contains pages from the previous
 332                  * vmemmap_remap_range call which failed.  These
 333                  * are pages which were removed from the vmemmap.
 334                  * They will be restored in the following call.
 335                  */
 336                 walk = (struct vmemmap_remap_walk) {
 337                         .remap_pte      = vmemmap_restore_pte,
 338                         .reuse_addr     = reuse,
 339                         .vmemmap_pages  = &vmemmap_pages,
 340                 };
 341
 342                 vmemmap_remap_range(reuse, end, &walk);
 343         }
 344         mmap_read_unlock(&init_mm);
 345
 346         free_vmemmap_page_list(&vmemmap_pages);
 347
 348         return ret;
 349 }
 350
 351 static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
 352                                    gfp_t gfp_mask, struct list_head *list)
 353 {
 354         unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
 355         int nid = page_to_nid((struct page *)start);
 356         struct page *page, *next;
 357
 358         while (nr_pages--) {
 359                 page = alloc_pages_node(nid, gfp_mask, 0);
 360                 if (!page)
 361                         goto out;
 362                 list_add_tail(&page->lru, list);
 363         }
 364
 365         return 0;
 366 out:
 367         list_for_each_entry_safe(page, next, list, lru)
 368                 __free_pages(page, 0);
 369         return -ENOMEM;
 370 }
 371
 372 /**
 373  * vmemmap_remap_alloc - remap the vmemmap virtual address range [@start, end)
 374  *                       to the page which is from the @vmemmap_pages
 375  *                       respectively.
 376  * @start:      start address of the vmemmap virtual address range that we want
 377  *              to remap.
 378  * @end:        end address of the vmemmap virtual address range that we want to
 379  *              remap.
 380  * @reuse:      reuse address.
 381  * @gfp_mask:   GFP flag for allocating vmemmap pages.
 382  *
 383  * Return: %0 on success, negative error code otherwise.
 384  */
 385 int vmemmap_remap_alloc(unsigned long start, unsigned long end,
 386                         unsigned long reuse, gfp_t gfp_mask)
 387 {
 388         LIST_HEAD(vmemmap_pages);
 389         struct vmemmap_remap_walk walk = {
 390                 .remap_pte      = vmemmap_restore_pte,
 391                 .reuse_addr     = reuse,
 392                 .vmemmap_pages  = &vmemmap_pages,
 393         };
 394
 395         /* See the comment in the vmemmap_remap_free(). */
 396         BUG_ON(start - reuse != PAGE_SIZE);
 397
 398         if (alloc_vmemmap_page_list(start, end, gfp_mask, &vmemmap_pages))
 399                 return -ENOMEM;
 400
 401         mmap_read_lock(&init_mm);
 402         vmemmap_remap_range(reuse, end, &walk);
 403         mmap_read_unlock(&init_mm);
 404
 405         return 0;
 406 }
 407
 408 /*
 409  * Allocate a block of memory to be used to back the virtual memory map
 410  * or to back the page tables that are used to create the mapping.
 411  * Uses the main allocators if they are available, else bootmem.
 412  */
 413
 414 static void * __ref __earlyonly_bootmem_alloc(int node,
 415                                 unsigned long size,
 416                                 unsigned long align,
 417                                 unsigned long goal)
 418 {
 419         return memblock_alloc_try_nid_raw(size, align, goal,
 420                                                MEMBLOCK_ALLOC_ACCESSIBLE, node);
 421 }
 422
 423 void * __meminit vmemmap_alloc_block(unsigned long size, int node)
 424 {
 425         /* If the main allocator is up use that, fallback to bootmem. */
 426         if (slab_is_available()) {
 427                 gfp_t gfp_mask = GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
 428                 int order = get_order(size);
 429                 static bool warned;
 430                 struct page *page;
 431
 432                 page = alloc_pages_node(node, gfp_mask, order);
 433                 if (page)
 434                         return page_address(page);
 435
 436                 if (!warned) {
 437                         warn_alloc(gfp_mask & ~__GFP_NOWARN, NULL,
 438                                    "vmemmap alloc failure: order:%u", order);
 439                         warned = true;
 440                 }
 441                 return NULL;
 442         } else
 443                 return __earlyonly_bootmem_alloc(node, size, size,
 444                                 __pa(MAX_DMA_ADDRESS));
 445 }
 446
 447 static void * __meminit altmap_alloc_block_buf(unsigned long size,
 448                                                struct vmem_altmap *altmap);
 449
 450 /* need to make sure size is all the same during early stage */
 451 void * __meminit vmemmap_alloc_block_buf(unsigned long size, int node,
 452                                          struct vmem_altmap *altmap)
 453 {
 454         void *ptr;
 455
 456         if (altmap)
 457                 return altmap_alloc_block_buf(size, altmap);
 458
 459         ptr = sparse_buffer_alloc(size);
 460         if (!ptr)
 461                 ptr = vmemmap_alloc_block(size, node);
 462         return ptr;
 463 }
 464
 465 static unsigned long __meminit vmem_altmap_next_pfn(struct vmem_altmap *altmap)
 466 {
 467         return altmap->base_pfn + altmap->reserve + altmap->alloc
 468                 + altmap->align;
 469 }
 470
 471 static unsigned long __meminit vmem_altmap_nr_free(struct vmem_altmap *altmap)
 472 {
 473         unsigned long allocated = altmap->alloc + altmap->align;
 474
 475         if (altmap->free > allocated)
 476                 return altmap->free - allocated;
 477         return 0;
 478 }
 479
 480 static void * __meminit altmap_alloc_block_buf(unsigned long size,
 481                                                struct vmem_altmap *altmap)
 482 {
 483         unsigned long pfn, nr_pfns, nr_align;
 484
 485         if (size & ~PAGE_MASK) {
 486                 pr_warn_once("%s: allocations must be multiple of PAGE_SIZE (%ld)\n",
 487                                 __func__, size);
 488                 return NULL;
 489         }
 490
 491         pfn = vmem_altmap_next_pfn(altmap);
 492         nr_pfns = size >> PAGE_SHIFT;
 493         nr_align = 1UL << find_first_bit(&nr_pfns, BITS_PER_LONG);
 494         nr_align = ALIGN(pfn, nr_align) - pfn;
 495         if (nr_pfns + nr_align > vmem_altmap_nr_free(altmap))
 496                 return NULL;
 497
 498         altmap->alloc += nr_pfns;
 499         altmap->align += nr_align;
 500         pfn += nr_align;
 501
 502         pr_debug("%s: pfn: %#lx alloc: %ld align: %ld nr: %#lx\n",
 503                         __func__, pfn, altmap->alloc, altmap->align, nr_pfns);
 504         return __va(__pfn_to_phys(pfn));
 505 }
 506
 507 void __meminit vmemmap_verify(pte_t *pte, int node,
 508                                 unsigned long start, unsigned long end)
 509 {
 510         unsigned long pfn = pte_pfn(*pte);
 511         int actual_node = early_pfn_to_nid(pfn);
 512
 513         if (node_distance(actual_node, node) > LOCAL_DISTANCE)
 514                 pr_warn("[%lx-%lx] potential offnode page_structs\n",
 515                         start, end - 1);
 516 }
 517
 518 pte_t * __meminit vmemmap_pte_populate(pmd_t *pmd, unsigned long addr, int node,
 519                                        struct vmem_altmap *altmap)
 520 {
 521         pte_t *pte = pte_offset_kernel(pmd, addr);
 522         if (pte_none(*pte)) {
 523                 pte_t entry;
 524                 void *p;
 525
 526                 p = vmemmap_alloc_block_buf(PAGE_SIZE, node, altmap);
 527                 if (!p)
 528                         return NULL;
 529                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT, PAGE_KERNEL);
 530                 set_pte_at(&init_mm, addr, pte, entry);
 531         }
 532         return pte;
 533 }
 534
 535 static void * __meminit vmemmap_alloc_block_zero(unsigned long size, int node)
 536 {
 537         void *p = vmemmap_alloc_block(size, node);
 538
 539         if (!p)
 540                 return NULL;
 541         memset(p, 0, size);
 542
 543         return p;
 544 }
 545
 546 pmd_t * __meminit vmemmap_pmd_populate(pud_t *pud, unsigned long addr, int node)
 547 {
 548         pmd_t *pmd = pmd_offset(pud, addr);
 549         if (pmd_none(*pmd)) {
 550                 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
 551                 if (!p)
 552                         return NULL;
 553                 pmd_populate_kernel(&init_mm, pmd, p);
 554         }
 555         return pmd;
 556 }
 557
 558 pud_t * __meminit vmemmap_pud_populate(p4d_t *p4d, unsigned long addr, int node)
 559 {
 560         pud_t *pud = pud_offset(p4d, addr);
 561         if (pud_none(*pud)) {
 562                 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
 563                 if (!p)
 564                         return NULL;
 565                 pud_populate(&init_mm, pud, p);
 566         }
 567         return pud;
 568 }
 569
 570 p4d_t * __meminit vmemmap_p4d_populate(pgd_t *pgd, unsigned long addr, int node)
 571 {
 572         p4d_t *p4d = p4d_offset(pgd, addr);
 573         if (p4d_none(*p4d)) {
 574                 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
 575                 if (!p)
 576                         return NULL;
 577                 p4d_populate(&init_mm, p4d, p);
 578         }
 579         return p4d;
 580 }
 581
 582 pgd_t * __meminit vmemmap_pgd_populate(unsigned long addr, int node)
 583 {
 584         pgd_t *pgd = pgd_offset_k(addr);
 585         if (pgd_none(*pgd)) {
 586                 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
 587                 if (!p)
 588                         return NULL;
 589                 pgd_populate(&init_mm, pgd, p);
 590         }
 591         return pgd;
 592 }
 593
 594 int __meminit vmemmap_populate_basepages(unsigned long start, unsigned long end,
 595                                          int node, struct vmem_altmap *altmap)
 596 {
 597         unsigned long addr = start;
 598         pgd_t *pgd;
 599         p4d_t *p4d;
 600         pud_t *pud;
 601         pmd_t *pmd;
 602         pte_t *pte;
 603
 604         for (; addr < end; addr += PAGE_SIZE) {
 605                 pgd = vmemmap_pgd_populate(addr, node);
 606                 if (!pgd)
 607                         return -ENOMEM;
 608                 p4d = vmemmap_p4d_populate(pgd, addr, node);
 609                 if (!p4d)
 610                         return -ENOMEM;
 611                 pud = vmemmap_pud_populate(p4d, addr, node);
 612                 if (!pud)
 613                         return -ENOMEM;
 614                 pmd = vmemmap_pmd_populate(pud, addr, node);
 615                 if (!pmd)
 616                         return -ENOMEM;
 617                 pte = vmemmap_pte_populate(pmd, addr, node, altmap);
 618                 if (!pte)
 619                         return -ENOMEM;
 620                 vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
 621         }
 622
 623         return 0;
 624 }
 625
 626 struct page * __meminit __populate_section_memmap(unsigned long pfn,
 627                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 628 {
 629         unsigned long start = (unsigned long) pfn_to_page(pfn);
 630         unsigned long end = start + nr_pages * sizeof(struct page);
 631
 632         if (WARN_ON_ONCE(!IS_ALIGNED(pfn, PAGES_PER_SUBSECTION) ||
 633                 !IS_ALIGNED(nr_pages, PAGES_PER_SUBSECTION)))
 634                 return NULL;
 635
 636         if (vmemmap_populate(start, end, nid, altmap))
 637                 return NULL;
 638
 639         return pfn_to_page(pfn);
 640 }