mm/sparse.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * sparse memory mappings.
   4  */
   5 #include <linux/mm.h>
   6 #include <linux/slab.h>
   7 #include <linux/mmzone.h>
   8 #include <linux/memblock.h>
   9 #include <linux/compiler.h>
  10 #include <linux/highmem.h>
  11 #include <linux/export.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/vmalloc.h>
  14 #include <linux/swap.h>
  15 #include <linux/swapops.h>
  16 #include <linux/bootmem_info.h>
  17
  18 #include "internal.h"
  19 #include <asm/dma.h>
  20
  21 /*
  22  * Permanent SPARSEMEM data:
  23  *
  24  * 1) mem_section       - memory sections, mem_map's for valid memory
  25  */
  26 #ifdef CONFIG_SPARSEMEM_EXTREME
  27 struct mem_section **mem_section;
  28 #else
  29 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  30         ____cacheline_internodealigned_in_smp;
  31 #endif
  32 EXPORT_SYMBOL(mem_section);
  33
  34 #ifdef NODE_NOT_IN_PAGE_FLAGS
  35 /*
  36  * If we did not store the node number in the page then we have to
  37  * do a lookup in the section_to_node_table in order to find which
  38  * node the page belongs to.
  39  */
  40 #if MAX_NUMNODES <= 256
  41 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  42 #else
  43 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  44 #endif
  45
  46 int page_to_nid(const struct page *page)
  47 {
  48         return section_to_node_table[page_to_section(page)];
  49 }
  50 EXPORT_SYMBOL(page_to_nid);
  51
  52 static void set_section_nid(unsigned long section_nr, int nid)
  53 {
  54         section_to_node_table[section_nr] = nid;
  55 }
  56 #else /* !NODE_NOT_IN_PAGE_FLAGS */
  57 static inline void set_section_nid(unsigned long section_nr, int nid)
  58 {
  59 }
  60 #endif
  61
  62 #ifdef CONFIG_SPARSEMEM_EXTREME
  63 static noinline struct mem_section __ref *sparse_index_alloc(int nid)
  64 {
  65         struct mem_section *section = NULL;
  66         unsigned long array_size = SECTIONS_PER_ROOT *
  67                                    sizeof(struct mem_section);
  68
  69         if (slab_is_available()) {
  70                 section = kzalloc_node(array_size, GFP_KERNEL, nid);
  71         } else {
  72                 section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
  73                                               nid);
  74                 if (!section)
  75                         panic("%s: Failed to allocate %lu bytes nid=%d\n",
  76                               __func__, array_size, nid);
  77         }
  78
  79         return section;
  80 }
  81
  82 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  83 {
  84         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  85         struct mem_section *section;
  86
  87         /*
  88          * An existing section is possible in the sub-section hotplug
  89          * case. First hot-add instantiates, follow-on hot-add reuses
  90          * the existing section.
  91          *
  92          * The mem_hotplug_lock resolves the apparent race below.
  93          */
  94         if (mem_section[root])
  95                 return 0;
  96
  97         section = sparse_index_alloc(nid);
  98         if (!section)
  99                 return -ENOMEM;
 100
 101         mem_section[root] = section;
 102
 103         return 0;
 104 }
 105 #else /* !SPARSEMEM_EXTREME */
 106 static inline int sparse_index_init(unsigned long section_nr, int nid)
 107 {
 108         return 0;
 109 }
 110 #endif
 111
 112 /*
 113  * During early boot, before section_mem_map is used for an actual
 114  * mem_map, we use section_mem_map to store the section's NUMA
 115  * node.  This keeps us from having to use another data structure.  The
 116  * node information is cleared just before we store the real mem_map.
 117  */
 118 static inline unsigned long sparse_encode_early_nid(int nid)
 119 {
 120         return ((unsigned long)nid << SECTION_NID_SHIFT);
 121 }
 122
 123 static inline int sparse_early_nid(struct mem_section *section)
 124 {
 125         return (section->section_mem_map >> SECTION_NID_SHIFT);
 126 }
 127
 128 /* Validate the physical addressing limitations of the model */
 129 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 130                                                 unsigned long *end_pfn)
 131 {
 132         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 133
 134         /*
 135          * Sanity checks - do not allow an architecture to pass
 136          * in larger pfns than the maximum scope of sparsemem:
 137          */
 138         if (*start_pfn > max_sparsemem_pfn) {
 139                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 140                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 141                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 142                 WARN_ON_ONCE(1);
 143                 *start_pfn = max_sparsemem_pfn;
 144                 *end_pfn = max_sparsemem_pfn;
 145         } else if (*end_pfn > max_sparsemem_pfn) {
 146                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 147                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 148                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 149                 WARN_ON_ONCE(1);
 150                 *end_pfn = max_sparsemem_pfn;
 151         }
 152 }
 153
 154 /*
 155  * There are a number of times that we loop over NR_MEM_SECTIONS,
 156  * looking for section_present() on each.  But, when we have very
 157  * large physical address spaces, NR_MEM_SECTIONS can also be
 158  * very large which makes the loops quite long.
 159  *
 160  * Keeping track of this gives us an easy way to break out of
 161  * those loops early.
 162  */
 163 unsigned long __highest_present_section_nr;
 164 static void __section_mark_present(struct mem_section *ms,
 165                 unsigned long section_nr)
 166 {
 167         if (section_nr > __highest_present_section_nr)
 168                 __highest_present_section_nr = section_nr;
 169
 170         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 171 }
 172
 173 #define for_each_present_section_nr(start, section_nr)          \
 174         for (section_nr = next_present_section_nr(start-1);     \
 175              ((section_nr != -1) &&                             \
 176               (section_nr <= __highest_present_section_nr));    \
 177              section_nr = next_present_section_nr(section_nr))
 178
 179 static inline unsigned long first_present_section_nr(void)
 180 {
 181         return next_present_section_nr(-1);
 182 }
 183
 184 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 185 static void subsection_mask_set(unsigned long *map, unsigned long pfn,
 186                 unsigned long nr_pages)
 187 {
 188         int idx = subsection_map_index(pfn);
 189         int end = subsection_map_index(pfn + nr_pages - 1);
 190
 191         bitmap_set(map, idx, end - idx + 1);
 192 }
 193
 194 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 195 {
 196         int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
 197         unsigned long nr, start_sec = pfn_to_section_nr(pfn);
 198
 199         if (!nr_pages)
 200                 return;
 201
 202         for (nr = start_sec; nr <= end_sec; nr++) {
 203                 struct mem_section *ms;
 204                 unsigned long pfns;
 205
 206                 pfns = min(nr_pages, PAGES_PER_SECTION
 207                                 - (pfn & ~PAGE_SECTION_MASK));
 208                 ms = __nr_to_section(nr);
 209                 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
 210
 211                 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
 212                                 pfns, subsection_map_index(pfn),
 213                                 subsection_map_index(pfn + pfns - 1));
 214
 215                 pfn += pfns;
 216                 nr_pages -= pfns;
 217         }
 218 }
 219 #else
 220 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 221 {
 222 }
 223 #endif
 224
 225 /* Record a memory area against a node. */
 226 static void __init memory_present(int nid, unsigned long start, unsigned long end)
 227 {
 228         unsigned long pfn;
 229
 230 #ifdef CONFIG_SPARSEMEM_EXTREME
 231         if (unlikely(!mem_section)) {
 232                 unsigned long size, align;
 233
 234                 size = sizeof(struct mem_section *) * NR_SECTION_ROOTS;
 235                 align = 1 << (INTERNODE_CACHE_SHIFT);
 236                 mem_section = memblock_alloc(size, align);
 237                 if (!mem_section)
 238                         panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
 239                               __func__, size, align);
 240         }
 241 #endif
 242
 243         start &= PAGE_SECTION_MASK;
 244         mminit_validate_memmodel_limits(&start, &end);
 245         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 246                 unsigned long section = pfn_to_section_nr(pfn);
 247                 struct mem_section *ms;
 248
 249                 sparse_index_init(section, nid);
 250                 set_section_nid(section, nid);
 251
 252                 ms = __nr_to_section(section);
 253                 if (!ms->section_mem_map) {
 254                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 255                                                         SECTION_IS_ONLINE;
 256                         __section_mark_present(ms, section);
 257                 }
 258         }
 259 }
 260
 261 /*
 262  * Mark all memblocks as present using memory_present().
 263  * This is a convenience function that is useful to mark all of the systems
 264  * memory as present during initialization.
 265  */
 266 static void __init memblocks_present(void)
 267 {
 268         unsigned long start, end;
 269         int i, nid;
 270
 271         for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
 272                 memory_present(nid, start, end);
 273 }
 274
 275 /*
 276  * Subtle, we encode the real pfn into the mem_map such that
 277  * the identity pfn - section_mem_map will return the actual
 278  * physical page frame number.
 279  */
 280 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 281 {
 282         unsigned long coded_mem_map =
 283                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 284         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
 285         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 286         return coded_mem_map;
 287 }
 288
 289 #ifdef CONFIG_MEMORY_HOTPLUG
 290 /*
 291  * Decode mem_map from the coded memmap
 292  */
 293 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 294 {
 295         /* mask off the extra low bits of information */
 296         coded_mem_map &= SECTION_MAP_MASK;
 297         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 298 }
 299 #endif /* CONFIG_MEMORY_HOTPLUG */
 300
 301 static void __meminit sparse_init_one_section(struct mem_section *ms,
 302                 unsigned long pnum, struct page *mem_map,
 303                 struct mem_section_usage *usage, unsigned long flags)
 304 {
 305         ms->section_mem_map &= ~SECTION_MAP_MASK;
 306         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
 307                 | SECTION_HAS_MEM_MAP | flags;
 308         ms->usage = usage;
 309 }
 310
 311 static unsigned long usemap_size(void)
 312 {
 313         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 314 }
 315
 316 size_t mem_section_usage_size(void)
 317 {
 318         return sizeof(struct mem_section_usage) + usemap_size();
 319 }
 320
 321 static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
 322 {
 323 #ifndef CONFIG_NUMA
 324         return __pa_symbol(pgdat);
 325 #else
 326         return __pa(pgdat);
 327 #endif
 328 }
 329
 330 #ifdef CONFIG_MEMORY_HOTREMOVE
 331 static struct mem_section_usage * __init
 332 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 333                                          unsigned long size)
 334 {
 335         struct mem_section_usage *usage;
 336         unsigned long goal, limit;
 337         int nid;
 338         /*
 339          * A page may contain usemaps for other sections preventing the
 340          * page being freed and making a section unremovable while
 341          * other sections referencing the usemap remain active. Similarly,
 342          * a pgdat can prevent a section being removed. If section A
 343          * contains a pgdat and section B contains the usemap, both
 344          * sections become inter-dependent. This allocates usemaps
 345          * from the same section as the pgdat where possible to avoid
 346          * this problem.
 347          */
 348         goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 349         limit = goal + (1UL << PA_SECTION_SHIFT);
 350         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 351 again:
 352         usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
 353         if (!usage && limit) {
 354                 limit = 0;
 355                 goto again;
 356         }
 357         return usage;
 358 }
 359
 360 static void __init check_usemap_section_nr(int nid,
 361                 struct mem_section_usage *usage)
 362 {
 363         unsigned long usemap_snr, pgdat_snr;
 364         static unsigned long old_usemap_snr;
 365         static unsigned long old_pgdat_snr;
 366         struct pglist_data *pgdat = NODE_DATA(nid);
 367         int usemap_nid;
 368
 369         /* First call */
 370         if (!old_usemap_snr) {
 371                 old_usemap_snr = NR_MEM_SECTIONS;
 372                 old_pgdat_snr = NR_MEM_SECTIONS;
 373         }
 374
 375         usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
 376         pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT);
 377         if (usemap_snr == pgdat_snr)
 378                 return;
 379
 380         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 381                 /* skip redundant message */
 382                 return;
 383
 384         old_usemap_snr = usemap_snr;
 385         old_pgdat_snr = pgdat_snr;
 386
 387         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 388         if (usemap_nid != nid) {
 389                 pr_info("node %d must be removed before remove section %ld\n",
 390                         nid, usemap_snr);
 391                 return;
 392         }
 393         /*
 394          * There is a circular dependency.
 395          * Some platforms allow un-removable section because they will just
 396          * gather other removable sections for dynamic partitioning.
 397          * Just notify un-removable section's number here.
 398          */
 399         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 400                 usemap_snr, pgdat_snr, nid);
 401 }
 402 #else
 403 static struct mem_section_usage * __init
 404 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 405                                          unsigned long size)
 406 {
 407         return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
 408 }
 409
 410 static void __init check_usemap_section_nr(int nid,
 411                 struct mem_section_usage *usage)
 412 {
 413 }
 414 #endif /* CONFIG_MEMORY_HOTREMOVE */
 415
 416 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 417 static unsigned long __init section_map_size(void)
 418 {
 419         return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
 420 }
 421
 422 #else
 423 static unsigned long __init section_map_size(void)
 424 {
 425         return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 426 }
 427
 428 struct page __init *__populate_section_memmap(unsigned long pfn,
 429                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 430 {
 431         unsigned long size = section_map_size();
 432         struct page *map = sparse_buffer_alloc(size);
 433         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 434
 435         if (map)
 436                 return map;
 437
 438         map = memblock_alloc_try_nid_raw(size, size, addr,
 439                                           MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 440         if (!map)
 441                 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
 442                       __func__, size, PAGE_SIZE, nid, &addr);
 443
 444         return map;
 445 }
 446 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 447
 448 static void *sparsemap_buf __meminitdata;
 449 static void *sparsemap_buf_end __meminitdata;
 450
 451 static inline void __meminit sparse_buffer_free(unsigned long size)
 452 {
 453         WARN_ON(!sparsemap_buf || size == 0);
 454         memblock_free_early(__pa(sparsemap_buf), size);
 455 }
 456
 457 static void __init sparse_buffer_init(unsigned long size, int nid)
 458 {
 459         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 460         WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
 461         /*
 462          * Pre-allocated buffer is mainly used by __populate_section_memmap
 463          * and we want it to be properly aligned to the section size - this is
 464          * especially the case for VMEMMAP which maps memmap to PMDs
 465          */
 466         sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(),
 467                                         addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 468         sparsemap_buf_end = sparsemap_buf + size;
 469 }
 470
 471 static void __init sparse_buffer_fini(void)
 472 {
 473         unsigned long size = sparsemap_buf_end - sparsemap_buf;
 474
 475         if (sparsemap_buf && size > 0)
 476                 sparse_buffer_free(size);
 477         sparsemap_buf = NULL;
 478 }
 479
 480 void * __meminit sparse_buffer_alloc(unsigned long size)
 481 {
 482         void *ptr = NULL;
 483
 484         if (sparsemap_buf) {
 485                 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
 486                 if (ptr + size > sparsemap_buf_end)
 487                         ptr = NULL;
 488                 else {
 489                         /* Free redundant aligned space */
 490                         if ((unsigned long)(ptr - sparsemap_buf) > 0)
 491                                 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
 492                         sparsemap_buf = ptr + size;
 493                 }
 494         }
 495         return ptr;
 496 }
 497
 498 void __weak __meminit vmemmap_populate_print_last(void)
 499 {
 500 }
 501
 502 /*
 503  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
 504  * And number of present sections in this node is map_count.
 505  */
 506 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
 507                                    unsigned long pnum_end,
 508                                    unsigned long map_count)
 509 {
 510         struct mem_section_usage *usage;
 511         unsigned long pnum;
 512         struct page *map;
 513
 514         usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
 515                         mem_section_usage_size() * map_count);
 516         if (!usage) {
 517                 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
 518                 goto failed;
 519         }
 520         sparse_buffer_init(map_count * section_map_size(), nid);
 521         for_each_present_section_nr(pnum_begin, pnum) {
 522                 unsigned long pfn = section_nr_to_pfn(pnum);
 523
 524                 if (pnum >= pnum_end)
 525                         break;
 526
 527                 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
 528                                 nid, NULL);
 529                 if (!map) {
 530                         pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
 531                                __func__, nid);
 532                         pnum_begin = pnum;
 533                         sparse_buffer_fini();
 534                         goto failed;
 535                 }
 536                 check_usemap_section_nr(nid, usage);
 537                 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
 538                                 SECTION_IS_EARLY);
 539                 usage = (void *) usage + mem_section_usage_size();
 540         }
 541         sparse_buffer_fini();
 542         return;
 543 failed:
 544         /* We failed to allocate, mark all the following pnums as not present */
 545         for_each_present_section_nr(pnum_begin, pnum) {
 546                 struct mem_section *ms;
 547
 548                 if (pnum >= pnum_end)
 549                         break;
 550                 ms = __nr_to_section(pnum);
 551                 ms->section_mem_map = 0;
 552         }
 553 }
 554
 555 /*
 556  * Allocate the accumulated non-linear sections, allocate a mem_map
 557  * for each and record the physical to section mapping.
 558  */
 559 void __init sparse_init(void)
 560 {
 561         unsigned long pnum_end, pnum_begin, map_count = 1;
 562         int nid_begin;
 563
 564         memblocks_present();
 565
 566         pnum_begin = first_present_section_nr();
 567         nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
 568
 569         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 570         set_pageblock_order();
 571
 572         for_each_present_section_nr(pnum_begin + 1, pnum_end) {
 573                 int nid = sparse_early_nid(__nr_to_section(pnum_end));
 574
 575                 if (nid == nid_begin) {
 576                         map_count++;
 577                         continue;
 578                 }
 579                 /* Init node with sections in range [pnum_begin, pnum_end) */
 580                 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 581                 nid_begin = nid;
 582                 pnum_begin = pnum_end;
 583                 map_count = 1;
 584         }
 585         /* cover the last node */
 586         sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 587         vmemmap_populate_print_last();
 588 }
 589
 590 #ifdef CONFIG_MEMORY_HOTPLUG
 591
 592 /* Mark all memory sections within the pfn range as online */
 593 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 594 {
 595         unsigned long pfn;
 596
 597         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 598                 unsigned long section_nr = pfn_to_section_nr(pfn);
 599                 struct mem_section *ms;
 600
 601                 /* onlining code should never touch invalid ranges */
 602                 if (WARN_ON(!valid_section_nr(section_nr)))
 603                         continue;
 604
 605                 ms = __nr_to_section(section_nr);
 606                 ms->section_mem_map |= SECTION_IS_ONLINE;
 607         }
 608 }
 609
 610 /* Mark all memory sections within the pfn range as offline */
 611 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 612 {
 613         unsigned long pfn;
 614
 615         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 616                 unsigned long section_nr = pfn_to_section_nr(pfn);
 617                 struct mem_section *ms;
 618
 619                 /*
 620                  * TODO this needs some double checking. Offlining code makes
 621                  * sure to check pfn_valid but those checks might be just bogus
 622                  */
 623                 if (WARN_ON(!valid_section_nr(section_nr)))
 624                         continue;
 625
 626                 ms = __nr_to_section(section_nr);
 627                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 628         }
 629 }
 630
 631 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 632 static struct page * __meminit populate_section_memmap(unsigned long pfn,
 633                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 634 {
 635         return __populate_section_memmap(pfn, nr_pages, nid, altmap);
 636 }
 637
 638 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 639                 struct vmem_altmap *altmap)
 640 {
 641         unsigned long start = (unsigned long) pfn_to_page(pfn);
 642         unsigned long end = start + nr_pages * sizeof(struct page);
 643
 644         vmemmap_free(start, end, altmap);
 645 }
 646 static void free_map_bootmem(struct page *memmap)
 647 {
 648         unsigned long start = (unsigned long)memmap;
 649         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 650
 651         vmemmap_free(start, end, NULL);
 652 }
 653
 654 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
 655 {
 656         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 657         DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
 658         struct mem_section *ms = __pfn_to_section(pfn);
 659         unsigned long *subsection_map = ms->usage
 660                 ? &ms->usage->subsection_map[0] : NULL;
 661
 662         subsection_mask_set(map, pfn, nr_pages);
 663         if (subsection_map)
 664                 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
 665
 666         if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
 667                                 "section already deactivated (%#lx + %ld)\n",
 668                                 pfn, nr_pages))
 669                 return -EINVAL;
 670
 671         bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
 672         return 0;
 673 }
 674
 675 static bool is_subsection_map_empty(struct mem_section *ms)
 676 {
 677         return bitmap_empty(&ms->usage->subsection_map[0],
 678                             SUBSECTIONS_PER_SECTION);
 679 }
 680
 681 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
 682 {
 683         struct mem_section *ms = __pfn_to_section(pfn);
 684         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 685         unsigned long *subsection_map;
 686         int rc = 0;
 687
 688         subsection_mask_set(map, pfn, nr_pages);
 689
 690         subsection_map = &ms->usage->subsection_map[0];
 691
 692         if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
 693                 rc = -EINVAL;
 694         else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
 695                 rc = -EEXIST;
 696         else
 697                 bitmap_or(subsection_map, map, subsection_map,
 698                                 SUBSECTIONS_PER_SECTION);
 699
 700         return rc;
 701 }
 702 #else
 703 struct page * __meminit populate_section_memmap(unsigned long pfn,
 704                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 705 {
 706         return kvmalloc_node(array_size(sizeof(struct page),
 707                                         PAGES_PER_SECTION), GFP_KERNEL, nid);
 708 }
 709
 710 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 711                 struct vmem_altmap *altmap)
 712 {
 713         kvfree(pfn_to_page(pfn));
 714 }
 715
 716 static void free_map_bootmem(struct page *memmap)
 717 {
 718         unsigned long maps_section_nr, removing_section_nr, i;
 719         unsigned long magic, nr_pages;
 720         struct page *page = virt_to_page(memmap);
 721
 722         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 723                 >> PAGE_SHIFT;
 724
 725         for (i = 0; i < nr_pages; i++, page++) {
 726                 magic = (unsigned long) page->freelist;
 727
 728                 BUG_ON(magic == NODE_INFO);
 729
 730                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 731                 removing_section_nr = page_private(page);
 732
 733                 /*
 734                  * When this function is called, the removing section is
 735                  * logical offlined state. This means all pages are isolated
 736                  * from page allocator. If removing section's memmap is placed
 737                  * on the same section, it must not be freed.
 738                  * If it is freed, page allocator may allocate it which will
 739                  * be removed physically soon.
 740                  */
 741                 if (maps_section_nr != removing_section_nr)
 742                         put_page_bootmem(page);
 743         }
 744 }
 745
 746 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
 747 {
 748         return 0;
 749 }
 750
 751 static bool is_subsection_map_empty(struct mem_section *ms)
 752 {
 753         return true;
 754 }
 755
 756 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
 757 {
 758         return 0;
 759 }
 760 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 761
 762 /*
 763  * To deactivate a memory region, there are 3 cases to handle across
 764  * two configurations (SPARSEMEM_VMEMMAP={y,n}):
 765  *
 766  * 1. deactivation of a partial hot-added section (only possible in
 767  *    the SPARSEMEM_VMEMMAP=y case).
 768  *      a) section was present at memory init.
 769  *      b) section was hot-added post memory init.
 770  * 2. deactivation of a complete hot-added section.
 771  * 3. deactivation of a complete section from memory init.
 772  *
 773  * For 1, when subsection_map does not empty we will not be freeing the
 774  * usage map, but still need to free the vmemmap range.
 775  *
 776  * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
 777  */
 778 static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
 779                 struct vmem_altmap *altmap)
 780 {
 781         struct mem_section *ms = __pfn_to_section(pfn);
 782         bool section_is_early = early_section(ms);
 783         struct page *memmap = NULL;
 784         bool empty;
 785
 786         if (clear_subsection_map(pfn, nr_pages))
 787                 return;
 788
 789         empty = is_subsection_map_empty(ms);
 790         if (empty) {
 791                 unsigned long section_nr = pfn_to_section_nr(pfn);
 792
 793                 /*
 794                  * When removing an early section, the usage map is kept (as the
 795                  * usage maps of other sections fall into the same page). It
 796                  * will be re-used when re-adding the section - which is then no
 797                  * longer an early section. If the usage map is PageReserved, it
 798                  * was allocated during boot.
 799                  */
 800                 if (!PageReserved(virt_to_page(ms->usage))) {
 801                         kfree(ms->usage);
 802                         ms->usage = NULL;
 803                 }
 804                 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
 805                 /*
 806                  * Mark the section invalid so that valid_section()
 807                  * return false. This prevents code from dereferencing
 808                  * ms->usage array.
 809                  */
 810                 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
 811         }
 812
 813         /*
 814          * The memmap of early sections is always fully populated. See
 815          * section_activate() and pfn_valid() .
 816          */
 817         if (!section_is_early)
 818                 depopulate_section_memmap(pfn, nr_pages, altmap);
 819         else if (memmap)
 820                 free_map_bootmem(memmap);
 821
 822         if (empty)
 823                 ms->section_mem_map = (unsigned long)NULL;
 824 }
 825
 826 static struct page * __meminit section_activate(int nid, unsigned long pfn,
 827                 unsigned long nr_pages, struct vmem_altmap *altmap)
 828 {
 829         struct mem_section *ms = __pfn_to_section(pfn);
 830         struct mem_section_usage *usage = NULL;
 831         struct page *memmap;
 832         int rc = 0;
 833
 834         if (!ms->usage) {
 835                 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
 836                 if (!usage)
 837                         return ERR_PTR(-ENOMEM);
 838                 ms->usage = usage;
 839         }
 840
 841         rc = fill_subsection_map(pfn, nr_pages);
 842         if (rc) {
 843                 if (usage)
 844                         ms->usage = NULL;
 845                 kfree(usage);
 846                 return ERR_PTR(rc);
 847         }
 848
 849         /*
 850          * The early init code does not consider partially populated
 851          * initial sections, it simply assumes that memory will never be
 852          * referenced.  If we hot-add memory into such a section then we
 853          * do not need to populate the memmap and can simply reuse what
 854          * is already there.
 855          */
 856         if (nr_pages < PAGES_PER_SECTION && early_section(ms))
 857                 return pfn_to_page(pfn);
 858
 859         memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
 860         if (!memmap) {
 861                 section_deactivate(pfn, nr_pages, altmap);
 862                 return ERR_PTR(-ENOMEM);
 863         }
 864
 865         return memmap;
 866 }
 867
 868 /**
 869  * sparse_add_section - add a memory section, or populate an existing one
 870  * @nid: The node to add section on
 871  * @start_pfn: start pfn of the memory range
 872  * @nr_pages: number of pfns to add in the section
 873  * @altmap: device page map
 874  *
 875  * This is only intended for hotplug.
 876  *
 877  * Note that only VMEMMAP supports sub-section aligned hotplug,
 878  * the proper alignment and size are gated by check_pfn_span().
 879  *
 880  *
 881  * Return:
 882  * * 0          - On success.
 883  * * -EEXIST    - Section has been present.
 884  * * -ENOMEM    - Out of memory.
 885  */
 886 int __meminit sparse_add_section(int nid, unsigned long start_pfn,
 887                 unsigned long nr_pages, struct vmem_altmap *altmap)
 888 {
 889         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 890         struct mem_section *ms;
 891         struct page *memmap;
 892         int ret;
 893
 894         ret = sparse_index_init(section_nr, nid);
 895         if (ret < 0)
 896                 return ret;
 897
 898         memmap = section_activate(nid, start_pfn, nr_pages, altmap);
 899         if (IS_ERR(memmap))
 900                 return PTR_ERR(memmap);
 901
 902         /*
 903          * Poison uninitialized struct pages in order to catch invalid flags
 904          * combinations.
 905          */
 906         page_init_poison(memmap, sizeof(struct page) * nr_pages);
 907
 908         ms = __nr_to_section(section_nr);
 909         set_section_nid(section_nr, nid);
 910         __section_mark_present(ms, section_nr);
 911
 912         /* Align memmap to section boundary in the subsection case */
 913         if (section_nr_to_pfn(section_nr) != start_pfn)
 914                 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
 915         sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
 916
 917         return 0;
 918 }
 919
 920 #ifdef CONFIG_MEMORY_FAILURE
 921 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 922 {
 923         int i;
 924
 925         /*
 926          * A further optimization is to have per section refcounted
 927          * num_poisoned_pages.  But that would need more space per memmap, so
 928          * for now just do a quick global check to speed up this routine in the
 929          * absence of bad pages.
 930          */
 931         if (atomic_long_read(&num_poisoned_pages) == 0)
 932                 return;
 933
 934         for (i = 0; i < nr_pages; i++) {
 935                 if (PageHWPoison(&memmap[i])) {
 936                         num_poisoned_pages_dec();
 937                         ClearPageHWPoison(&memmap[i]);
 938                 }
 939         }
 940 }
 941 #else
 942 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 943 {
 944 }
 945 #endif
 946
 947 void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
 948                 unsigned long nr_pages, unsigned long map_offset,
 949                 struct vmem_altmap *altmap)
 950 {
 951         clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
 952                         nr_pages - map_offset);
 953         section_deactivate(pfn, nr_pages, altmap);
 954 }
 955 #endif /* CONFIG_MEMORY_HOTPLUG */