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 #include <linux/vmstat.h>
  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 static 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 = next_present_section_nr(section_nr))
 177
 178 static inline unsigned long first_present_section_nr(void)
 179 {
 180         return next_present_section_nr(-1);
 181 }
 182
 183 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 184 static void subsection_mask_set(unsigned long *map, unsigned long pfn,
 185                 unsigned long nr_pages)
 186 {
 187         int idx = subsection_map_index(pfn);
 188         int end = subsection_map_index(pfn + nr_pages - 1);
 189
 190         bitmap_set(map, idx, end - idx + 1);
 191 }
 192
 193 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 194 {
 195         int end_sec_nr = pfn_to_section_nr(pfn + nr_pages - 1);
 196         unsigned long nr, start_sec_nr = pfn_to_section_nr(pfn);
 197
 198         for (nr = start_sec_nr; nr <= end_sec_nr; nr++) {
 199                 struct mem_section *ms;
 200                 unsigned long pfns;
 201
 202                 pfns = min(nr_pages, PAGES_PER_SECTION
 203                                 - (pfn & ~PAGE_SECTION_MASK));
 204                 ms = __nr_to_section(nr);
 205                 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
 206
 207                 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
 208                                 pfns, subsection_map_index(pfn),
 209                                 subsection_map_index(pfn + pfns - 1));
 210
 211                 pfn += pfns;
 212                 nr_pages -= pfns;
 213         }
 214 }
 215 #else
 216 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 217 {
 218 }
 219 #endif
 220
 221 /* Record a memory area against a node. */
 222 static void __init memory_present(int nid, unsigned long start, unsigned long end)
 223 {
 224         unsigned long pfn;
 225
 226         start &= PAGE_SECTION_MASK;
 227         mminit_validate_memmodel_limits(&start, &end);
 228         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 229                 unsigned long section_nr = pfn_to_section_nr(pfn);
 230                 struct mem_section *ms;
 231
 232                 sparse_index_init(section_nr, nid);
 233                 set_section_nid(section_nr, nid);
 234
 235                 ms = __nr_to_section(section_nr);
 236                 if (!ms->section_mem_map) {
 237                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 238                                                         SECTION_IS_ONLINE;
 239                         __section_mark_present(ms, section_nr);
 240                 }
 241         }
 242 }
 243
 244 /*
 245  * Mark all memblocks as present using memory_present().
 246  * This is a convenience function that is useful to mark all of the systems
 247  * memory as present during initialization.
 248  */
 249 static void __init memblocks_present(void)
 250 {
 251         unsigned long start, end;
 252         int i, nid;
 253
 254 #ifdef CONFIG_SPARSEMEM_EXTREME
 255         if (unlikely(!mem_section)) {
 256                 unsigned long size, align;
 257
 258                 size = sizeof(struct mem_section *) * NR_SECTION_ROOTS;
 259                 align = 1 << (INTERNODE_CACHE_SHIFT);
 260                 mem_section = memblock_alloc(size, align);
 261                 if (!mem_section)
 262                         panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
 263                               __func__, size, align);
 264         }
 265 #endif
 266
 267         for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
 268                 memory_present(nid, start, end);
 269 }
 270
 271 /*
 272  * Subtle, we encode the real pfn into the mem_map such that
 273  * the identity pfn - section_mem_map will return the actual
 274  * physical page frame number.
 275  */
 276 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 277 {
 278         unsigned long coded_mem_map =
 279                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 280         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > PFN_SECTION_SHIFT);
 281         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 282         return coded_mem_map;
 283 }
 284
 285 #ifdef CONFIG_MEMORY_HOTPLUG
 286 /*
 287  * Decode mem_map from the coded memmap
 288  */
 289 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 290 {
 291         /* mask off the extra low bits of information */
 292         coded_mem_map &= SECTION_MAP_MASK;
 293         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 294 }
 295 #endif /* CONFIG_MEMORY_HOTPLUG */
 296
 297 static void __meminit sparse_init_one_section(struct mem_section *ms,
 298                 unsigned long pnum, struct page *mem_map,
 299                 struct mem_section_usage *usage, unsigned long flags)
 300 {
 301         ms->section_mem_map &= ~SECTION_MAP_MASK;
 302         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
 303                 | SECTION_HAS_MEM_MAP | flags;
 304         ms->usage = usage;
 305 }
 306
 307 static unsigned long usemap_size(void)
 308 {
 309         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 310 }
 311
 312 size_t mem_section_usage_size(void)
 313 {
 314         return sizeof(struct mem_section_usage) + usemap_size();
 315 }
 316
 317 #ifdef CONFIG_MEMORY_HOTREMOVE
 318 static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
 319 {
 320 #ifndef CONFIG_NUMA
 321         VM_BUG_ON(pgdat != &contig_page_data);
 322         return __pa_symbol(&contig_page_data);
 323 #else
 324         return __pa(pgdat);
 325 #endif
 326 }
 327
 328 static struct mem_section_usage * __init
 329 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 330                                          unsigned long size)
 331 {
 332         struct mem_section_usage *usage;
 333         unsigned long goal, limit;
 334         int nid;
 335         /*
 336          * A page may contain usemaps for other sections preventing the
 337          * page being freed and making a section unremovable while
 338          * other sections referencing the usemap remain active. Similarly,
 339          * a pgdat can prevent a section being removed. If section A
 340          * contains a pgdat and section B contains the usemap, both
 341          * sections become inter-dependent. This allocates usemaps
 342          * from the same section as the pgdat where possible to avoid
 343          * this problem.
 344          */
 345         goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 346         limit = goal + (1UL << PA_SECTION_SHIFT);
 347         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 348 again:
 349         usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
 350         if (!usage && limit) {
 351                 limit = MEMBLOCK_ALLOC_ACCESSIBLE;
 352                 goto again;
 353         }
 354         return usage;
 355 }
 356
 357 static void __init check_usemap_section_nr(int nid,
 358                 struct mem_section_usage *usage)
 359 {
 360         unsigned long usemap_snr, pgdat_snr;
 361         static unsigned long old_usemap_snr;
 362         static unsigned long old_pgdat_snr;
 363         struct pglist_data *pgdat = NODE_DATA(nid);
 364         int usemap_nid;
 365
 366         /* First call */
 367         if (!old_usemap_snr) {
 368                 old_usemap_snr = NR_MEM_SECTIONS;
 369                 old_pgdat_snr = NR_MEM_SECTIONS;
 370         }
 371
 372         usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
 373         pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT);
 374         if (usemap_snr == pgdat_snr)
 375                 return;
 376
 377         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 378                 /* skip redundant message */
 379                 return;
 380
 381         old_usemap_snr = usemap_snr;
 382         old_pgdat_snr = pgdat_snr;
 383
 384         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 385         if (usemap_nid != nid) {
 386                 pr_info("node %d must be removed before remove section %ld\n",
 387                         nid, usemap_snr);
 388                 return;
 389         }
 390         /*
 391          * There is a circular dependency.
 392          * Some platforms allow un-removable section because they will just
 393          * gather other removable sections for dynamic partitioning.
 394          * Just notify un-removable section's number here.
 395          */
 396         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 397                 usemap_snr, pgdat_snr, nid);
 398 }
 399 #else
 400 static struct mem_section_usage * __init
 401 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 402                                          unsigned long size)
 403 {
 404         return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
 405 }
 406
 407 static void __init check_usemap_section_nr(int nid,
 408                 struct mem_section_usage *usage)
 409 {
 410 }
 411 #endif /* CONFIG_MEMORY_HOTREMOVE */
 412
 413 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 414 static unsigned long __init section_map_size(void)
 415 {
 416         return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
 417 }
 418
 419 #else
 420 static unsigned long __init section_map_size(void)
 421 {
 422         return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 423 }
 424
 425 struct page __init *__populate_section_memmap(unsigned long pfn,
 426                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
 427                 struct dev_pagemap *pgmap)
 428 {
 429         unsigned long size = section_map_size();
 430         struct page *map = sparse_buffer_alloc(size);
 431         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 432
 433         if (map)
 434                 return map;
 435
 436         map = memmap_alloc(size, size, addr, nid, false);
 437         if (!map)
 438                 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
 439                       __func__, size, PAGE_SIZE, nid, &addr);
 440
 441         return map;
 442 }
 443 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 444
 445 static void *sparsemap_buf __meminitdata;
 446 static void *sparsemap_buf_end __meminitdata;
 447
 448 static inline void __meminit sparse_buffer_free(unsigned long size)
 449 {
 450         WARN_ON(!sparsemap_buf || size == 0);
 451         memblock_free(sparsemap_buf, size);
 452 }
 453
 454 static void __init sparse_buffer_init(unsigned long size, int nid)
 455 {
 456         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 457         WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
 458         /*
 459          * Pre-allocated buffer is mainly used by __populate_section_memmap
 460          * and we want it to be properly aligned to the section size - this is
 461          * especially the case for VMEMMAP which maps memmap to PMDs
 462          */
 463         sparsemap_buf = memmap_alloc(size, section_map_size(), addr, nid, true);
 464         sparsemap_buf_end = sparsemap_buf + size;
 465 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 466         mod_node_early_perpage_metadata(nid, DIV_ROUND_UP(size, PAGE_SIZE));
 467 #endif
 468 }
 469
 470 static void __init sparse_buffer_fini(void)
 471 {
 472         unsigned long size = sparsemap_buf_end - sparsemap_buf;
 473
 474         if (sparsemap_buf && size > 0)
 475                 sparse_buffer_free(size);
 476         sparsemap_buf = NULL;
 477 }
 478
 479 void * __meminit sparse_buffer_alloc(unsigned long size)
 480 {
 481         void *ptr = NULL;
 482
 483         if (sparsemap_buf) {
 484                 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
 485                 if (ptr + size > sparsemap_buf_end)
 486                         ptr = NULL;
 487                 else {
 488                         /* Free redundant aligned space */
 489                         if ((unsigned long)(ptr - sparsemap_buf) > 0)
 490                                 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
 491                         sparsemap_buf = ptr + size;
 492                 }
 493         }
 494         return ptr;
 495 }
 496
 497 void __weak __meminit vmemmap_populate_print_last(void)
 498 {
 499 }
 500
 501 /*
 502  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
 503  * And number of present sections in this node is map_count.
 504  */
 505 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
 506                                    unsigned long pnum_end,
 507                                    unsigned long map_count)
 508 {
 509         struct mem_section_usage *usage;
 510         unsigned long pnum;
 511         struct page *map;
 512
 513         usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
 514                         mem_section_usage_size() * map_count);
 515         if (!usage) {
 516                 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
 517                 goto failed;
 518         }
 519         sparse_buffer_init(map_count * section_map_size(), nid);
 520         for_each_present_section_nr(pnum_begin, pnum) {
 521                 unsigned long pfn = section_nr_to_pfn(pnum);
 522
 523                 if (pnum >= pnum_end)
 524                         break;
 525
 526                 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
 527                                 nid, NULL, NULL);
 528                 if (!map) {
 529                         pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
 530                                __func__, nid);
 531                         pnum_begin = pnum;
 532                         sparse_buffer_fini();
 533                         goto failed;
 534                 }
 535                 check_usemap_section_nr(nid, usage);
 536                 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
 537                                 SECTION_IS_EARLY);
 538                 usage = (void *) usage + mem_section_usage_size();
 539         }
 540         sparse_buffer_fini();
 541         return;
 542 failed:
 543         /* We failed to allocate, mark all the following pnums as not present */
 544         for_each_present_section_nr(pnum_begin, pnum) {
 545                 struct mem_section *ms;
 546
 547                 if (pnum >= pnum_end)
 548                         break;
 549                 ms = __nr_to_section(pnum);
 550                 ms->section_mem_map = 0;
 551         }
 552 }
 553
 554 /*
 555  * Allocate the accumulated non-linear sections, allocate a mem_map
 556  * for each and record the physical to section mapping.
 557  */
 558 void __init sparse_init(void)
 559 {
 560         unsigned long pnum_end, pnum_begin, map_count = 1;
 561         int nid_begin;
 562
 563         /* see include/linux/mmzone.h 'struct mem_section' definition */
 564         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 565         memblocks_present();
 566
 567         pnum_begin = first_present_section_nr();
 568         nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
 569
 570         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 571         set_pageblock_order();
 572
 573         for_each_present_section_nr(pnum_begin + 1, pnum_end) {
 574                 int nid = sparse_early_nid(__nr_to_section(pnum_end));
 575
 576                 if (nid == nid_begin) {
 577                         map_count++;
 578                         continue;
 579                 }
 580                 /* Init node with sections in range [pnum_begin, pnum_end) */
 581                 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 582                 nid_begin = nid;
 583                 pnum_begin = pnum_end;
 584                 map_count = 1;
 585         }
 586         /* cover the last node */
 587         sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 588         vmemmap_populate_print_last();
 589 }
 590
 591 #ifdef CONFIG_MEMORY_HOTPLUG
 592
 593 /* Mark all memory sections within the pfn range as online */
 594 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 595 {
 596         unsigned long pfn;
 597
 598         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 599                 unsigned long section_nr = pfn_to_section_nr(pfn);
 600                 struct mem_section *ms;
 601
 602                 /* onlining code should never touch invalid ranges */
 603                 if (WARN_ON(!valid_section_nr(section_nr)))
 604                         continue;
 605
 606                 ms = __nr_to_section(section_nr);
 607                 ms->section_mem_map |= SECTION_IS_ONLINE;
 608         }
 609 }
 610
 611 /* Mark all memory sections within the pfn range as offline */
 612 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 613 {
 614         unsigned long pfn;
 615
 616         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 617                 unsigned long section_nr = pfn_to_section_nr(pfn);
 618                 struct mem_section *ms;
 619
 620                 /*
 621                  * TODO this needs some double checking. Offlining code makes
 622                  * sure to check pfn_valid but those checks might be just bogus
 623                  */
 624                 if (WARN_ON(!valid_section_nr(section_nr)))
 625                         continue;
 626
 627                 ms = __nr_to_section(section_nr);
 628                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 629         }
 630 }
 631
 632 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 633 static struct page * __meminit populate_section_memmap(unsigned long pfn,
 634                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
 635                 struct dev_pagemap *pgmap)
 636 {
 637         return __populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
 638 }
 639
 640 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 641                 struct vmem_altmap *altmap)
 642 {
 643         unsigned long start = (unsigned long) pfn_to_page(pfn);
 644         unsigned long end = start + nr_pages * sizeof(struct page);
 645
 646         mod_node_page_state(page_pgdat(pfn_to_page(pfn)), NR_MEMMAP,
 647                             -1L * (DIV_ROUND_UP(end - start, PAGE_SIZE)));
 648         vmemmap_free(start, end, altmap);
 649 }
 650 static void free_map_bootmem(struct page *memmap)
 651 {
 652         unsigned long start = (unsigned long)memmap;
 653         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 654
 655         vmemmap_free(start, end, NULL);
 656 }
 657
 658 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
 659 {
 660         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 661         DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
 662         struct mem_section *ms = __pfn_to_section(pfn);
 663         unsigned long *subsection_map = ms->usage
 664                 ? &ms->usage->subsection_map[0] : NULL;
 665
 666         subsection_mask_set(map, pfn, nr_pages);
 667         if (subsection_map)
 668                 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
 669
 670         if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
 671                                 "section already deactivated (%#lx + %ld)\n",
 672                                 pfn, nr_pages))
 673                 return -EINVAL;
 674
 675         bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
 676         return 0;
 677 }
 678
 679 static bool is_subsection_map_empty(struct mem_section *ms)
 680 {
 681         return bitmap_empty(&ms->usage->subsection_map[0],
 682                             SUBSECTIONS_PER_SECTION);
 683 }
 684
 685 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
 686 {
 687         struct mem_section *ms = __pfn_to_section(pfn);
 688         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 689         unsigned long *subsection_map;
 690         int rc = 0;
 691
 692         subsection_mask_set(map, pfn, nr_pages);
 693
 694         subsection_map = &ms->usage->subsection_map[0];
 695
 696         if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
 697                 rc = -EINVAL;
 698         else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
 699                 rc = -EEXIST;
 700         else
 701                 bitmap_or(subsection_map, map, subsection_map,
 702                                 SUBSECTIONS_PER_SECTION);
 703
 704         return rc;
 705 }
 706 #else
 707 static struct page * __meminit populate_section_memmap(unsigned long pfn,
 708                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
 709                 struct dev_pagemap *pgmap)
 710 {
 711         return kvmalloc_node(array_size(sizeof(struct page),
 712                                         PAGES_PER_SECTION), GFP_KERNEL, nid);
 713 }
 714
 715 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 716                 struct vmem_altmap *altmap)
 717 {
 718         kvfree(pfn_to_page(pfn));
 719 }
 720
 721 static void free_map_bootmem(struct page *memmap)
 722 {
 723         unsigned long maps_section_nr, removing_section_nr, i;
 724         unsigned long magic, nr_pages;
 725         struct page *page = virt_to_page(memmap);
 726
 727         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 728                 >> PAGE_SHIFT;
 729
 730         for (i = 0; i < nr_pages; i++, page++) {
 731                 magic = page->index;
 732
 733                 BUG_ON(magic == NODE_INFO);
 734
 735                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 736                 removing_section_nr = page_private(page);
 737
 738                 /*
 739                  * When this function is called, the removing section is
 740                  * logical offlined state. This means all pages are isolated
 741                  * from page allocator. If removing section's memmap is placed
 742                  * on the same section, it must not be freed.
 743                  * If it is freed, page allocator may allocate it which will
 744                  * be removed physically soon.
 745                  */
 746                 if (maps_section_nr != removing_section_nr)
 747                         put_page_bootmem(page);
 748         }
 749 }
 750
 751 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
 752 {
 753         return 0;
 754 }
 755
 756 static bool is_subsection_map_empty(struct mem_section *ms)
 757 {
 758         return true;
 759 }
 760
 761 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
 762 {
 763         return 0;
 764 }
 765 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 766
 767 /*
 768  * To deactivate a memory region, there are 3 cases to handle across
 769  * two configurations (SPARSEMEM_VMEMMAP={y,n}):
 770  *
 771  * 1. deactivation of a partial hot-added section (only possible in
 772  *    the SPARSEMEM_VMEMMAP=y case).
 773  *      a) section was present at memory init.
 774  *      b) section was hot-added post memory init.
 775  * 2. deactivation of a complete hot-added section.
 776  * 3. deactivation of a complete section from memory init.
 777  *
 778  * For 1, when subsection_map does not empty we will not be freeing the
 779  * usage map, but still need to free the vmemmap range.
 780  *
 781  * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
 782  */
 783 static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
 784                 struct vmem_altmap *altmap)
 785 {
 786         struct mem_section *ms = __pfn_to_section(pfn);
 787         bool section_is_early = early_section(ms);
 788         struct page *memmap = NULL;
 789         bool empty;
 790
 791         if (clear_subsection_map(pfn, nr_pages))
 792                 return;
 793
 794         empty = is_subsection_map_empty(ms);
 795         if (empty) {
 796                 unsigned long section_nr = pfn_to_section_nr(pfn);
 797
 798                 /*
 799                  * Mark the section invalid so that valid_section()
 800                  * return false. This prevents code from dereferencing
 801                  * ms->usage array.
 802                  */
 803                 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
 804
 805                 /*
 806                  * When removing an early section, the usage map is kept (as the
 807                  * usage maps of other sections fall into the same page). It
 808                  * will be re-used when re-adding the section - which is then no
 809                  * longer an early section. If the usage map is PageReserved, it
 810                  * was allocated during boot.
 811                  */
 812                 if (!PageReserved(virt_to_page(ms->usage))) {
 813                         kfree_rcu(ms->usage, rcu);
 814                         WRITE_ONCE(ms->usage, NULL);
 815                 }
 816                 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
 817         }
 818
 819         /*
 820          * The memmap of early sections is always fully populated. See
 821          * section_activate() and pfn_valid() .
 822          */
 823         if (!section_is_early)
 824                 depopulate_section_memmap(pfn, nr_pages, altmap);
 825         else if (memmap)
 826                 free_map_bootmem(memmap);
 827
 828         if (empty)
 829                 ms->section_mem_map = (unsigned long)NULL;
 830 }
 831
 832 static struct page * __meminit section_activate(int nid, unsigned long pfn,
 833                 unsigned long nr_pages, struct vmem_altmap *altmap,
 834                 struct dev_pagemap *pgmap)
 835 {
 836         struct mem_section *ms = __pfn_to_section(pfn);
 837         struct mem_section_usage *usage = NULL;
 838         struct page *memmap;
 839         int rc;
 840
 841         if (!ms->usage) {
 842                 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
 843                 if (!usage)
 844                         return ERR_PTR(-ENOMEM);
 845                 ms->usage = usage;
 846         }
 847
 848         rc = fill_subsection_map(pfn, nr_pages);
 849         if (rc) {
 850                 if (usage)
 851                         ms->usage = NULL;
 852                 kfree(usage);
 853                 return ERR_PTR(rc);
 854         }
 855
 856         /*
 857          * The early init code does not consider partially populated
 858          * initial sections, it simply assumes that memory will never be
 859          * referenced.  If we hot-add memory into such a section then we
 860          * do not need to populate the memmap and can simply reuse what
 861          * is already there.
 862          */
 863         if (nr_pages < PAGES_PER_SECTION && early_section(ms))
 864                 return pfn_to_page(pfn);
 865
 866         memmap = populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
 867         if (!memmap) {
 868                 section_deactivate(pfn, nr_pages, altmap);
 869                 return ERR_PTR(-ENOMEM);
 870         }
 871
 872         return memmap;
 873 }
 874
 875 /**
 876  * sparse_add_section - add a memory section, or populate an existing one
 877  * @nid: The node to add section on
 878  * @start_pfn: start pfn of the memory range
 879  * @nr_pages: number of pfns to add in the section
 880  * @altmap: alternate pfns to allocate the memmap backing store
 881  * @pgmap: alternate compound page geometry for devmap mappings
 882  *
 883  * This is only intended for hotplug.
 884  *
 885  * Note that only VMEMMAP supports sub-section aligned hotplug,
 886  * the proper alignment and size are gated by check_pfn_span().
 887  *
 888  *
 889  * Return:
 890  * * 0          - On success.
 891  * * -EEXIST    - Section has been present.
 892  * * -ENOMEM    - Out of memory.
 893  */
 894 int __meminit sparse_add_section(int nid, unsigned long start_pfn,
 895                 unsigned long nr_pages, struct vmem_altmap *altmap,
 896                 struct dev_pagemap *pgmap)
 897 {
 898         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 899         struct mem_section *ms;
 900         struct page *memmap;
 901         int ret;
 902
 903         ret = sparse_index_init(section_nr, nid);
 904         if (ret < 0)
 905                 return ret;
 906
 907         memmap = section_activate(nid, start_pfn, nr_pages, altmap, pgmap);
 908         if (IS_ERR(memmap))
 909                 return PTR_ERR(memmap);
 910
 911         /*
 912          * Poison uninitialized struct pages in order to catch invalid flags
 913          * combinations.
 914          */
 915         if (!altmap || !altmap->inaccessible)
 916                 page_init_poison(memmap, sizeof(struct page) * nr_pages);
 917
 918         ms = __nr_to_section(section_nr);
 919         set_section_nid(section_nr, nid);
 920         __section_mark_present(ms, section_nr);
 921
 922         /* Align memmap to section boundary in the subsection case */
 923         if (section_nr_to_pfn(section_nr) != start_pfn)
 924                 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
 925         sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
 926
 927         return 0;
 928 }
 929
 930 void sparse_remove_section(unsigned long pfn, unsigned long nr_pages,
 931                            struct vmem_altmap *altmap)
 932 {
 933         struct mem_section *ms = __pfn_to_section(pfn);
 934
 935         if (WARN_ON_ONCE(!valid_section(ms)))
 936                 return;
 937
 938         section_deactivate(pfn, nr_pages, altmap);
 939 }
 940 #endif /* CONFIG_MEMORY_HOTPLUG */