1 /* 2 * linux/mm/memory_hotplug.c 3 * 4 * Copyright (C) 5 */ 6 7 #include <linux/stddef.h> 8 #include <linux/mm.h> 9 #include <linux/sched/signal.h> 10 #include <linux/swap.h> 11 #include <linux/interrupt.h> 12 #include <linux/pagemap.h> 13 #include <linux/compiler.h> 14 #include <linux/export.h> 15 #include <linux/pagevec.h> 16 #include <linux/writeback.h> 17 #include <linux/slab.h> 18 #include <linux/sysctl.h> 19 #include <linux/cpu.h> 20 #include <linux/memory.h> 21 #include <linux/memremap.h> 22 #include <linux/memory_hotplug.h> 23 #include <linux/highmem.h> 24 #include <linux/vmalloc.h> 25 #include <linux/ioport.h> 26 #include <linux/delay.h> 27 #include <linux/migrate.h> 28 #include <linux/page-isolation.h> 29 #include <linux/pfn.h> 30 #include <linux/suspend.h> 31 #include <linux/mm_inline.h> 32 #include <linux/firmware-map.h> 33 #include <linux/stop_machine.h> 34 #include <linux/hugetlb.h> 35 #include <linux/memblock.h> 36 #include <linux/bootmem.h> 37 #include <linux/compaction.h> 38 39 #include <asm/tlbflush.h> 40 41 #include "internal.h" 42 43 /* 44 * online_page_callback contains pointer to current page onlining function. 45 * Initially it is generic_online_page(). If it is required it could be 46 * changed by calling set_online_page_callback() for callback registration 47 * and restore_online_page_callback() for generic callback restore. 48 */ 49 50 static void generic_online_page(struct page *page); 51 52 static online_page_callback_t online_page_callback = generic_online_page; 53 static DEFINE_MUTEX(online_page_callback_lock); 54 55 DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock); 56 57 void get_online_mems(void) 58 { 59 percpu_down_read(&mem_hotplug_lock); 60 } 61 62 void put_online_mems(void) 63 { 64 percpu_up_read(&mem_hotplug_lock); 65 } 66 67 bool movable_node_enabled = false; 68 69 #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE 70 bool memhp_auto_online; 71 #else 72 bool memhp_auto_online = true; 73 #endif 74 EXPORT_SYMBOL_GPL(memhp_auto_online); 75 76 static int __init setup_memhp_default_state(char *str) 77 { 78 if (!strcmp(str, "online")) 79 memhp_auto_online = true; 80 else if (!strcmp(str, "offline")) 81 memhp_auto_online = false; 82 83 return 1; 84 } 85 __setup("memhp_default_state=", setup_memhp_default_state); 86 87 void mem_hotplug_begin(void) 88 { 89 cpus_read_lock(); 90 percpu_down_write(&mem_hotplug_lock); 91 } 92 93 void mem_hotplug_done(void) 94 { 95 percpu_up_write(&mem_hotplug_lock); 96 cpus_read_unlock(); 97 } 98 99 /* add this memory to iomem resource */ 100 static struct resource *register_memory_resource(u64 start, u64 size) 101 { 102 struct resource *res, *conflict; 103 res = kzalloc(sizeof(struct resource), GFP_KERNEL); 104 if (!res) 105 return ERR_PTR(-ENOMEM); 106 107 res->name = "System RAM"; 108 res->start = start; 109 res->end = start + size - 1; 110 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY; 111 conflict = request_resource_conflict(&iomem_resource, res); 112 if (conflict) { 113 if (conflict->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY) { 114 pr_debug("Device unaddressable memory block " 115 "memory hotplug at %#010llx !\n", 116 (unsigned long long)start); 117 } 118 pr_debug("System RAM resource %pR cannot be added\n", res); 119 kfree(res); 120 return ERR_PTR(-EEXIST); 121 } 122 return res; 123 } 124 125 static void release_memory_resource(struct resource *res) 126 { 127 if (!res) 128 return; 129 release_resource(res); 130 kfree(res); 131 return; 132 } 133 134 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE 135 void get_page_bootmem(unsigned long info, struct page *page, 136 unsigned long type) 137 { 138 page->freelist = (void *)type; 139 SetPagePrivate(page); 140 set_page_private(page, info); 141 page_ref_inc(page); 142 } 143 144 void put_page_bootmem(struct page *page) 145 { 146 unsigned long type; 147 148 type = (unsigned long) page->freelist; 149 BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE || 150 type > MEMORY_HOTPLUG_MAX_BOOTMEM_TYPE); 151 152 if (page_ref_dec_return(page) == 1) { 153 page->freelist = NULL; 154 ClearPagePrivate(page); 155 set_page_private(page, 0); 156 INIT_LIST_HEAD(&page->lru); 157 free_reserved_page(page); 158 } 159 } 160 161 #ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE 162 #ifndef CONFIG_SPARSEMEM_VMEMMAP 163 static void register_page_bootmem_info_section(unsigned long start_pfn) 164 { 165 unsigned long *usemap, mapsize, section_nr, i; 166 struct mem_section *ms; 167 struct page *page, *memmap; 168 169 section_nr = pfn_to_section_nr(start_pfn); 170 ms = __nr_to_section(section_nr); 171 172 /* Get section's memmap address */ 173 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr); 174 175 /* 176 * Get page for the memmap's phys address 177 * XXX: need more consideration for sparse_vmemmap... 178 */ 179 page = virt_to_page(memmap); 180 mapsize = sizeof(struct page) * PAGES_PER_SECTION; 181 mapsize = PAGE_ALIGN(mapsize) >> PAGE_SHIFT; 182 183 /* remember memmap's page */ 184 for (i = 0; i < mapsize; i++, page++) 185 get_page_bootmem(section_nr, page, SECTION_INFO); 186 187 usemap = __nr_to_section(section_nr)->pageblock_flags; 188 page = virt_to_page(usemap); 189 190 mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT; 191 192 for (i = 0; i < mapsize; i++, page++) 193 get_page_bootmem(section_nr, page, MIX_SECTION_INFO); 194 195 } 196 #else /* CONFIG_SPARSEMEM_VMEMMAP */ 197 static void register_page_bootmem_info_section(unsigned long start_pfn) 198 { 199 unsigned long *usemap, mapsize, section_nr, i; 200 struct mem_section *ms; 201 struct page *page, *memmap; 202 203 if (!pfn_valid(start_pfn)) 204 return; 205 206 section_nr = pfn_to_section_nr(start_pfn); 207 ms = __nr_to_section(section_nr); 208 209 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr); 210 211 register_page_bootmem_memmap(section_nr, memmap, PAGES_PER_SECTION); 212 213 usemap = __nr_to_section(section_nr)->pageblock_flags; 214 page = virt_to_page(usemap); 215 216 mapsize = PAGE_ALIGN(usemap_size()) >> PAGE_SHIFT; 217 218 for (i = 0; i < mapsize; i++, page++) 219 get_page_bootmem(section_nr, page, MIX_SECTION_INFO); 220 } 221 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */ 222 223 void __init register_page_bootmem_info_node(struct pglist_data *pgdat) 224 { 225 unsigned long i, pfn, end_pfn, nr_pages; 226 int node = pgdat->node_id; 227 struct page *page; 228 229 nr_pages = PAGE_ALIGN(sizeof(struct pglist_data)) >> PAGE_SHIFT; 230 page = virt_to_page(pgdat); 231 232 for (i = 0; i < nr_pages; i++, page++) 233 get_page_bootmem(node, page, NODE_INFO); 234 235 pfn = pgdat->node_start_pfn; 236 end_pfn = pgdat_end_pfn(pgdat); 237 238 /* register section info */ 239 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 240 /* 241 * Some platforms can assign the same pfn to multiple nodes - on 242 * node0 as well as nodeN. To avoid registering a pfn against 243 * multiple nodes we check that this pfn does not already 244 * reside in some other nodes. 245 */ 246 if (pfn_valid(pfn) && (early_pfn_to_nid(pfn) == node)) 247 register_page_bootmem_info_section(pfn); 248 } 249 } 250 #endif /* CONFIG_HAVE_BOOTMEM_INFO_NODE */ 251 252 static int __meminit __add_section(int nid, unsigned long phys_start_pfn, 253 bool want_memblock) 254 { 255 int ret; 256 int i; 257 258 if (pfn_valid(phys_start_pfn)) 259 return -EEXIST; 260 261 ret = sparse_add_one_section(NODE_DATA(nid), phys_start_pfn); 262 if (ret < 0) 263 return ret; 264 265 /* 266 * Make all the pages reserved so that nobody will stumble over half 267 * initialized state. 268 * FIXME: We also have to associate it with a node because pfn_to_node 269 * relies on having page with the proper node. 270 */ 271 for (i = 0; i < PAGES_PER_SECTION; i++) { 272 unsigned long pfn = phys_start_pfn + i; 273 struct page *page; 274 if (!pfn_valid(pfn)) 275 continue; 276 277 page = pfn_to_page(pfn); 278 set_page_node(page, nid); 279 SetPageReserved(page); 280 } 281 282 if (!want_memblock) 283 return 0; 284 285 return register_new_memory(nid, __pfn_to_section(phys_start_pfn)); 286 } 287 288 /* 289 * Reasonably generic function for adding memory. It is 290 * expected that archs that support memory hotplug will 291 * call this function after deciding the zone to which to 292 * add the new pages. 293 */ 294 int __ref __add_pages(int nid, unsigned long phys_start_pfn, 295 unsigned long nr_pages, bool want_memblock) 296 { 297 unsigned long i; 298 int err = 0; 299 int start_sec, end_sec; 300 struct vmem_altmap *altmap; 301 302 /* during initialize mem_map, align hot-added range to section */ 303 start_sec = pfn_to_section_nr(phys_start_pfn); 304 end_sec = pfn_to_section_nr(phys_start_pfn + nr_pages - 1); 305 306 altmap = to_vmem_altmap((unsigned long) pfn_to_page(phys_start_pfn)); 307 if (altmap) { 308 /* 309 * Validate altmap is within bounds of the total request 310 */ 311 if (altmap->base_pfn != phys_start_pfn 312 || vmem_altmap_offset(altmap) > nr_pages) { 313 pr_warn_once("memory add fail, invalid altmap\n"); 314 err = -EINVAL; 315 goto out; 316 } 317 altmap->alloc = 0; 318 } 319 320 for (i = start_sec; i <= end_sec; i++) { 321 err = __add_section(nid, section_nr_to_pfn(i), want_memblock); 322 323 /* 324 * EEXIST is finally dealt with by ioresource collision 325 * check. see add_memory() => register_memory_resource() 326 * Warning will be printed if there is collision. 327 */ 328 if (err && (err != -EEXIST)) 329 break; 330 err = 0; 331 } 332 vmemmap_populate_print_last(); 333 out: 334 return err; 335 } 336 EXPORT_SYMBOL_GPL(__add_pages); 337 338 #ifdef CONFIG_MEMORY_HOTREMOVE 339 /* find the smallest valid pfn in the range [start_pfn, end_pfn) */ 340 static int find_smallest_section_pfn(int nid, struct zone *zone, 341 unsigned long start_pfn, 342 unsigned long end_pfn) 343 { 344 struct mem_section *ms; 345 346 for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SECTION) { 347 ms = __pfn_to_section(start_pfn); 348 349 if (unlikely(!valid_section(ms))) 350 continue; 351 352 if (unlikely(pfn_to_nid(start_pfn) != nid)) 353 continue; 354 355 if (zone && zone != page_zone(pfn_to_page(start_pfn))) 356 continue; 357 358 return start_pfn; 359 } 360 361 return 0; 362 } 363 364 /* find the biggest valid pfn in the range [start_pfn, end_pfn). */ 365 static int find_biggest_section_pfn(int nid, struct zone *zone, 366 unsigned long start_pfn, 367 unsigned long end_pfn) 368 { 369 struct mem_section *ms; 370 unsigned long pfn; 371 372 /* pfn is the end pfn of a memory section. */ 373 pfn = end_pfn - 1; 374 for (; pfn >= start_pfn; pfn -= PAGES_PER_SECTION) { 375 ms = __pfn_to_section(pfn); 376 377 if (unlikely(!valid_section(ms))) 378 continue; 379 380 if (unlikely(pfn_to_nid(pfn) != nid)) 381 continue; 382 383 if (zone && zone != page_zone(pfn_to_page(pfn))) 384 continue; 385 386 return pfn; 387 } 388 389 return 0; 390 } 391 392 static void shrink_zone_span(struct zone *zone, unsigned long start_pfn, 393 unsigned long end_pfn) 394 { 395 unsigned long zone_start_pfn = zone->zone_start_pfn; 396 unsigned long z = zone_end_pfn(zone); /* zone_end_pfn namespace clash */ 397 unsigned long zone_end_pfn = z; 398 unsigned long pfn; 399 struct mem_section *ms; 400 int nid = zone_to_nid(zone); 401 402 zone_span_writelock(zone); 403 if (zone_start_pfn == start_pfn) { 404 /* 405 * If the section is smallest section in the zone, it need 406 * shrink zone->zone_start_pfn and zone->zone_spanned_pages. 407 * In this case, we find second smallest valid mem_section 408 * for shrinking zone. 409 */ 410 pfn = find_smallest_section_pfn(nid, zone, end_pfn, 411 zone_end_pfn); 412 if (pfn) { 413 zone->zone_start_pfn = pfn; 414 zone->spanned_pages = zone_end_pfn - pfn; 415 } 416 } else if (zone_end_pfn == end_pfn) { 417 /* 418 * If the section is biggest section in the zone, it need 419 * shrink zone->spanned_pages. 420 * In this case, we find second biggest valid mem_section for 421 * shrinking zone. 422 */ 423 pfn = find_biggest_section_pfn(nid, zone, zone_start_pfn, 424 start_pfn); 425 if (pfn) 426 zone->spanned_pages = pfn - zone_start_pfn + 1; 427 } 428 429 /* 430 * The section is not biggest or smallest mem_section in the zone, it 431 * only creates a hole in the zone. So in this case, we need not 432 * change the zone. But perhaps, the zone has only hole data. Thus 433 * it check the zone has only hole or not. 434 */ 435 pfn = zone_start_pfn; 436 for (; pfn < zone_end_pfn; pfn += PAGES_PER_SECTION) { 437 ms = __pfn_to_section(pfn); 438 439 if (unlikely(!valid_section(ms))) 440 continue; 441 442 if (page_zone(pfn_to_page(pfn)) != zone) 443 continue; 444 445 /* If the section is current section, it continues the loop */ 446 if (start_pfn == pfn) 447 continue; 448 449 /* If we find valid section, we have nothing to do */ 450 zone_span_writeunlock(zone); 451 return; 452 } 453 454 /* The zone has no valid section */ 455 zone->zone_start_pfn = 0; 456 zone->spanned_pages = 0; 457 zone_span_writeunlock(zone); 458 } 459 460 static void shrink_pgdat_span(struct pglist_data *pgdat, 461 unsigned long start_pfn, unsigned long end_pfn) 462 { 463 unsigned long pgdat_start_pfn = pgdat->node_start_pfn; 464 unsigned long p = pgdat_end_pfn(pgdat); /* pgdat_end_pfn namespace clash */ 465 unsigned long pgdat_end_pfn = p; 466 unsigned long pfn; 467 struct mem_section *ms; 468 int nid = pgdat->node_id; 469 470 if (pgdat_start_pfn == start_pfn) { 471 /* 472 * If the section is smallest section in the pgdat, it need 473 * shrink pgdat->node_start_pfn and pgdat->node_spanned_pages. 474 * In this case, we find second smallest valid mem_section 475 * for shrinking zone. 476 */ 477 pfn = find_smallest_section_pfn(nid, NULL, end_pfn, 478 pgdat_end_pfn); 479 if (pfn) { 480 pgdat->node_start_pfn = pfn; 481 pgdat->node_spanned_pages = pgdat_end_pfn - pfn; 482 } 483 } else if (pgdat_end_pfn == end_pfn) { 484 /* 485 * If the section is biggest section in the pgdat, it need 486 * shrink pgdat->node_spanned_pages. 487 * In this case, we find second biggest valid mem_section for 488 * shrinking zone. 489 */ 490 pfn = find_biggest_section_pfn(nid, NULL, pgdat_start_pfn, 491 start_pfn); 492 if (pfn) 493 pgdat->node_spanned_pages = pfn - pgdat_start_pfn + 1; 494 } 495 496 /* 497 * If the section is not biggest or smallest mem_section in the pgdat, 498 * it only creates a hole in the pgdat. So in this case, we need not 499 * change the pgdat. 500 * But perhaps, the pgdat has only hole data. Thus it check the pgdat 501 * has only hole or not. 502 */ 503 pfn = pgdat_start_pfn; 504 for (; pfn < pgdat_end_pfn; pfn += PAGES_PER_SECTION) { 505 ms = __pfn_to_section(pfn); 506 507 if (unlikely(!valid_section(ms))) 508 continue; 509 510 if (pfn_to_nid(pfn) != nid) 511 continue; 512 513 /* If the section is current section, it continues the loop */ 514 if (start_pfn == pfn) 515 continue; 516 517 /* If we find valid section, we have nothing to do */ 518 return; 519 } 520 521 /* The pgdat has no valid section */ 522 pgdat->node_start_pfn = 0; 523 pgdat->node_spanned_pages = 0; 524 } 525 526 static void __remove_zone(struct zone *zone, unsigned long start_pfn) 527 { 528 struct pglist_data *pgdat = zone->zone_pgdat; 529 int nr_pages = PAGES_PER_SECTION; 530 unsigned long flags; 531 532 pgdat_resize_lock(zone->zone_pgdat, &flags); 533 shrink_zone_span(zone, start_pfn, start_pfn + nr_pages); 534 shrink_pgdat_span(pgdat, start_pfn, start_pfn + nr_pages); 535 pgdat_resize_unlock(zone->zone_pgdat, &flags); 536 } 537 538 static int __remove_section(struct zone *zone, struct mem_section *ms, 539 unsigned long map_offset) 540 { 541 unsigned long start_pfn; 542 int scn_nr; 543 int ret = -EINVAL; 544 545 if (!valid_section(ms)) 546 return ret; 547 548 ret = unregister_memory_section(ms); 549 if (ret) 550 return ret; 551 552 scn_nr = __section_nr(ms); 553 start_pfn = section_nr_to_pfn(scn_nr); 554 __remove_zone(zone, start_pfn); 555 556 sparse_remove_one_section(zone, ms, map_offset); 557 return 0; 558 } 559 560 /** 561 * __remove_pages() - remove sections of pages from a zone 562 * @zone: zone from which pages need to be removed 563 * @phys_start_pfn: starting pageframe (must be aligned to start of a section) 564 * @nr_pages: number of pages to remove (must be multiple of section size) 565 * 566 * Generic helper function to remove section mappings and sysfs entries 567 * for the section of the memory we are removing. Caller needs to make 568 * sure that pages are marked reserved and zones are adjust properly by 569 * calling offline_pages(). 570 */ 571 int __remove_pages(struct zone *zone, unsigned long phys_start_pfn, 572 unsigned long nr_pages) 573 { 574 unsigned long i; 575 unsigned long map_offset = 0; 576 int sections_to_remove, ret = 0; 577 578 /* In the ZONE_DEVICE case device driver owns the memory region */ 579 if (is_dev_zone(zone)) { 580 struct page *page = pfn_to_page(phys_start_pfn); 581 struct vmem_altmap *altmap; 582 583 altmap = to_vmem_altmap((unsigned long) page); 584 if (altmap) 585 map_offset = vmem_altmap_offset(altmap); 586 } else { 587 resource_size_t start, size; 588 589 start = phys_start_pfn << PAGE_SHIFT; 590 size = nr_pages * PAGE_SIZE; 591 592 ret = release_mem_region_adjustable(&iomem_resource, start, 593 size); 594 if (ret) { 595 resource_size_t endres = start + size - 1; 596 597 pr_warn("Unable to release resource <%pa-%pa> (%d)\n", 598 &start, &endres, ret); 599 } 600 } 601 602 clear_zone_contiguous(zone); 603 604 /* 605 * We can only remove entire sections 606 */ 607 BUG_ON(phys_start_pfn & ~PAGE_SECTION_MASK); 608 BUG_ON(nr_pages % PAGES_PER_SECTION); 609 610 sections_to_remove = nr_pages / PAGES_PER_SECTION; 611 for (i = 0; i < sections_to_remove; i++) { 612 unsigned long pfn = phys_start_pfn + i*PAGES_PER_SECTION; 613 614 ret = __remove_section(zone, __pfn_to_section(pfn), map_offset); 615 map_offset = 0; 616 if (ret) 617 break; 618 } 619 620 set_zone_contiguous(zone); 621 622 return ret; 623 } 624 #endif /* CONFIG_MEMORY_HOTREMOVE */ 625 626 int set_online_page_callback(online_page_callback_t callback) 627 { 628 int rc = -EINVAL; 629 630 get_online_mems(); 631 mutex_lock(&online_page_callback_lock); 632 633 if (online_page_callback == generic_online_page) { 634 online_page_callback = callback; 635 rc = 0; 636 } 637 638 mutex_unlock(&online_page_callback_lock); 639 put_online_mems(); 640 641 return rc; 642 } 643 EXPORT_SYMBOL_GPL(set_online_page_callback); 644 645 int restore_online_page_callback(online_page_callback_t callback) 646 { 647 int rc = -EINVAL; 648 649 get_online_mems(); 650 mutex_lock(&online_page_callback_lock); 651 652 if (online_page_callback == callback) { 653 online_page_callback = generic_online_page; 654 rc = 0; 655 } 656 657 mutex_unlock(&online_page_callback_lock); 658 put_online_mems(); 659 660 return rc; 661 } 662 EXPORT_SYMBOL_GPL(restore_online_page_callback); 663 664 void __online_page_set_limits(struct page *page) 665 { 666 } 667 EXPORT_SYMBOL_GPL(__online_page_set_limits); 668 669 void __online_page_increment_counters(struct page *page) 670 { 671 adjust_managed_page_count(page, 1); 672 } 673 EXPORT_SYMBOL_GPL(__online_page_increment_counters); 674 675 void __online_page_free(struct page *page) 676 { 677 __free_reserved_page(page); 678 } 679 EXPORT_SYMBOL_GPL(__online_page_free); 680 681 static void generic_online_page(struct page *page) 682 { 683 __online_page_set_limits(page); 684 __online_page_increment_counters(page); 685 __online_page_free(page); 686 } 687 688 static int online_pages_range(unsigned long start_pfn, unsigned long nr_pages, 689 void *arg) 690 { 691 unsigned long i; 692 unsigned long onlined_pages = *(unsigned long *)arg; 693 struct page *page; 694 695 if (PageReserved(pfn_to_page(start_pfn))) 696 for (i = 0; i < nr_pages; i++) { 697 page = pfn_to_page(start_pfn + i); 698 (*online_page_callback)(page); 699 onlined_pages++; 700 } 701 702 online_mem_sections(start_pfn, start_pfn + nr_pages); 703 704 *(unsigned long *)arg = onlined_pages; 705 return 0; 706 } 707 708 /* check which state of node_states will be changed when online memory */ 709 static void node_states_check_changes_online(unsigned long nr_pages, 710 struct zone *zone, struct memory_notify *arg) 711 { 712 int nid = zone_to_nid(zone); 713 enum zone_type zone_last = ZONE_NORMAL; 714 715 /* 716 * If we have HIGHMEM or movable node, node_states[N_NORMAL_MEMORY] 717 * contains nodes which have zones of 0...ZONE_NORMAL, 718 * set zone_last to ZONE_NORMAL. 719 * 720 * If we don't have HIGHMEM nor movable node, 721 * node_states[N_NORMAL_MEMORY] contains nodes which have zones of 722 * 0...ZONE_MOVABLE, set zone_last to ZONE_MOVABLE. 723 */ 724 if (N_MEMORY == N_NORMAL_MEMORY) 725 zone_last = ZONE_MOVABLE; 726 727 /* 728 * if the memory to be online is in a zone of 0...zone_last, and 729 * the zones of 0...zone_last don't have memory before online, we will 730 * need to set the node to node_states[N_NORMAL_MEMORY] after 731 * the memory is online. 732 */ 733 if (zone_idx(zone) <= zone_last && !node_state(nid, N_NORMAL_MEMORY)) 734 arg->status_change_nid_normal = nid; 735 else 736 arg->status_change_nid_normal = -1; 737 738 #ifdef CONFIG_HIGHMEM 739 /* 740 * If we have movable node, node_states[N_HIGH_MEMORY] 741 * contains nodes which have zones of 0...ZONE_HIGHMEM, 742 * set zone_last to ZONE_HIGHMEM. 743 * 744 * If we don't have movable node, node_states[N_NORMAL_MEMORY] 745 * contains nodes which have zones of 0...ZONE_MOVABLE, 746 * set zone_last to ZONE_MOVABLE. 747 */ 748 zone_last = ZONE_HIGHMEM; 749 if (N_MEMORY == N_HIGH_MEMORY) 750 zone_last = ZONE_MOVABLE; 751 752 if (zone_idx(zone) <= zone_last && !node_state(nid, N_HIGH_MEMORY)) 753 arg->status_change_nid_high = nid; 754 else 755 arg->status_change_nid_high = -1; 756 #else 757 arg->status_change_nid_high = arg->status_change_nid_normal; 758 #endif 759 760 /* 761 * if the node don't have memory befor online, we will need to 762 * set the node to node_states[N_MEMORY] after the memory 763 * is online. 764 */ 765 if (!node_state(nid, N_MEMORY)) 766 arg->status_change_nid = nid; 767 else 768 arg->status_change_nid = -1; 769 } 770 771 static void node_states_set_node(int node, struct memory_notify *arg) 772 { 773 if (arg->status_change_nid_normal >= 0) 774 node_set_state(node, N_NORMAL_MEMORY); 775 776 if (arg->status_change_nid_high >= 0) 777 node_set_state(node, N_HIGH_MEMORY); 778 779 node_set_state(node, N_MEMORY); 780 } 781 782 static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn, 783 unsigned long nr_pages) 784 { 785 unsigned long old_end_pfn = zone_end_pfn(zone); 786 787 if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn) 788 zone->zone_start_pfn = start_pfn; 789 790 zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn; 791 } 792 793 static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn, 794 unsigned long nr_pages) 795 { 796 unsigned long old_end_pfn = pgdat_end_pfn(pgdat); 797 798 if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn) 799 pgdat->node_start_pfn = start_pfn; 800 801 pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn; 802 } 803 804 void __ref move_pfn_range_to_zone(struct zone *zone, 805 unsigned long start_pfn, unsigned long nr_pages) 806 { 807 struct pglist_data *pgdat = zone->zone_pgdat; 808 int nid = pgdat->node_id; 809 unsigned long flags; 810 811 if (zone_is_empty(zone)) 812 init_currently_empty_zone(zone, start_pfn, nr_pages); 813 814 clear_zone_contiguous(zone); 815 816 /* TODO Huh pgdat is irqsave while zone is not. It used to be like that before */ 817 pgdat_resize_lock(pgdat, &flags); 818 zone_span_writelock(zone); 819 resize_zone_range(zone, start_pfn, nr_pages); 820 zone_span_writeunlock(zone); 821 resize_pgdat_range(pgdat, start_pfn, nr_pages); 822 pgdat_resize_unlock(pgdat, &flags); 823 824 /* 825 * TODO now we have a visible range of pages which are not associated 826 * with their zone properly. Not nice but set_pfnblock_flags_mask 827 * expects the zone spans the pfn range. All the pages in the range 828 * are reserved so nobody should be touching them so we should be safe 829 */ 830 memmap_init_zone(nr_pages, nid, zone_idx(zone), start_pfn, MEMMAP_HOTPLUG); 831 832 set_zone_contiguous(zone); 833 } 834 835 /* 836 * Returns a default kernel memory zone for the given pfn range. 837 * If no kernel zone covers this pfn range it will automatically go 838 * to the ZONE_NORMAL. 839 */ 840 static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn, 841 unsigned long nr_pages) 842 { 843 struct pglist_data *pgdat = NODE_DATA(nid); 844 int zid; 845 846 for (zid = 0; zid <= ZONE_NORMAL; zid++) { 847 struct zone *zone = &pgdat->node_zones[zid]; 848 849 if (zone_intersects(zone, start_pfn, nr_pages)) 850 return zone; 851 } 852 853 return &pgdat->node_zones[ZONE_NORMAL]; 854 } 855 856 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn, 857 unsigned long nr_pages) 858 { 859 struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn, 860 nr_pages); 861 struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 862 bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages); 863 bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages); 864 865 /* 866 * We inherit the existing zone in a simple case where zones do not 867 * overlap in the given range 868 */ 869 if (in_kernel ^ in_movable) 870 return (in_kernel) ? kernel_zone : movable_zone; 871 872 /* 873 * If the range doesn't belong to any zone or two zones overlap in the 874 * given range then we use movable zone only if movable_node is 875 * enabled because we always online to a kernel zone by default. 876 */ 877 return movable_node_enabled ? movable_zone : kernel_zone; 878 } 879 880 struct zone * zone_for_pfn_range(int online_type, int nid, unsigned start_pfn, 881 unsigned long nr_pages) 882 { 883 if (online_type == MMOP_ONLINE_KERNEL) 884 return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages); 885 886 if (online_type == MMOP_ONLINE_MOVABLE) 887 return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE]; 888 889 return default_zone_for_pfn(nid, start_pfn, nr_pages); 890 } 891 892 /* 893 * Associates the given pfn range with the given node and the zone appropriate 894 * for the given online type. 895 */ 896 static struct zone * __meminit move_pfn_range(int online_type, int nid, 897 unsigned long start_pfn, unsigned long nr_pages) 898 { 899 struct zone *zone; 900 901 zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages); 902 move_pfn_range_to_zone(zone, start_pfn, nr_pages); 903 return zone; 904 } 905 906 /* Must be protected by mem_hotplug_begin() or a device_lock */ 907 int __ref online_pages(unsigned long pfn, unsigned long nr_pages, int online_type) 908 { 909 unsigned long flags; 910 unsigned long onlined_pages = 0; 911 struct zone *zone; 912 int need_zonelists_rebuild = 0; 913 int nid; 914 int ret; 915 struct memory_notify arg; 916 917 nid = pfn_to_nid(pfn); 918 /* associate pfn range with the zone */ 919 zone = move_pfn_range(online_type, nid, pfn, nr_pages); 920 921 arg.start_pfn = pfn; 922 arg.nr_pages = nr_pages; 923 node_states_check_changes_online(nr_pages, zone, &arg); 924 925 ret = memory_notify(MEM_GOING_ONLINE, &arg); 926 ret = notifier_to_errno(ret); 927 if (ret) 928 goto failed_addition; 929 930 /* 931 * If this zone is not populated, then it is not in zonelist. 932 * This means the page allocator ignores this zone. 933 * So, zonelist must be updated after online. 934 */ 935 if (!populated_zone(zone)) { 936 need_zonelists_rebuild = 1; 937 setup_zone_pageset(zone); 938 } 939 940 ret = walk_system_ram_range(pfn, nr_pages, &onlined_pages, 941 online_pages_range); 942 if (ret) { 943 if (need_zonelists_rebuild) 944 zone_pcp_reset(zone); 945 goto failed_addition; 946 } 947 948 zone->present_pages += onlined_pages; 949 950 pgdat_resize_lock(zone->zone_pgdat, &flags); 951 zone->zone_pgdat->node_present_pages += onlined_pages; 952 pgdat_resize_unlock(zone->zone_pgdat, &flags); 953 954 if (onlined_pages) { 955 node_states_set_node(nid, &arg); 956 if (need_zonelists_rebuild) 957 build_all_zonelists(NULL); 958 else 959 zone_pcp_update(zone); 960 } 961 962 init_per_zone_wmark_min(); 963 964 if (onlined_pages) { 965 kswapd_run(nid); 966 kcompactd_run(nid); 967 } 968 969 vm_total_pages = nr_free_pagecache_pages(); 970 971 writeback_set_ratelimit(); 972 973 if (onlined_pages) 974 memory_notify(MEM_ONLINE, &arg); 975 return 0; 976 977 failed_addition: 978 pr_debug("online_pages [mem %#010llx-%#010llx] failed\n", 979 (unsigned long long) pfn << PAGE_SHIFT, 980 (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1); 981 memory_notify(MEM_CANCEL_ONLINE, &arg); 982 return ret; 983 } 984 #endif /* CONFIG_MEMORY_HOTPLUG_SPARSE */ 985 986 static void reset_node_present_pages(pg_data_t *pgdat) 987 { 988 struct zone *z; 989 990 for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++) 991 z->present_pages = 0; 992 993 pgdat->node_present_pages = 0; 994 } 995 996 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */ 997 static pg_data_t __ref *hotadd_new_pgdat(int nid, u64 start) 998 { 999 struct pglist_data *pgdat; 1000 unsigned long zones_size[MAX_NR_ZONES] = {0}; 1001 unsigned long zholes_size[MAX_NR_ZONES] = {0}; 1002 unsigned long start_pfn = PFN_DOWN(start); 1003 1004 pgdat = NODE_DATA(nid); 1005 if (!pgdat) { 1006 pgdat = arch_alloc_nodedata(nid); 1007 if (!pgdat) 1008 return NULL; 1009 1010 arch_refresh_nodedata(nid, pgdat); 1011 } else { 1012 /* 1013 * Reset the nr_zones, order and classzone_idx before reuse. 1014 * Note that kswapd will init kswapd_classzone_idx properly 1015 * when it starts in the near future. 1016 */ 1017 pgdat->nr_zones = 0; 1018 pgdat->kswapd_order = 0; 1019 pgdat->kswapd_classzone_idx = 0; 1020 } 1021 1022 /* we can use NODE_DATA(nid) from here */ 1023 1024 /* init node's zones as empty zones, we don't have any present pages.*/ 1025 free_area_init_node(nid, zones_size, start_pfn, zholes_size); 1026 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat); 1027 1028 /* 1029 * The node we allocated has no zone fallback lists. For avoiding 1030 * to access not-initialized zonelist, build here. 1031 */ 1032 build_all_zonelists(pgdat); 1033 1034 /* 1035 * zone->managed_pages is set to an approximate value in 1036 * free_area_init_core(), which will cause 1037 * /sys/device/system/node/nodeX/meminfo has wrong data. 1038 * So reset it to 0 before any memory is onlined. 1039 */ 1040 reset_node_managed_pages(pgdat); 1041 1042 /* 1043 * When memory is hot-added, all the memory is in offline state. So 1044 * clear all zones' present_pages because they will be updated in 1045 * online_pages() and offline_pages(). 1046 */ 1047 reset_node_present_pages(pgdat); 1048 1049 return pgdat; 1050 } 1051 1052 static void rollback_node_hotadd(int nid, pg_data_t *pgdat) 1053 { 1054 arch_refresh_nodedata(nid, NULL); 1055 free_percpu(pgdat->per_cpu_nodestats); 1056 arch_free_nodedata(pgdat); 1057 return; 1058 } 1059 1060 1061 /** 1062 * try_online_node - online a node if offlined 1063 * 1064 * called by cpu_up() to online a node without onlined memory. 1065 */ 1066 int try_online_node(int nid) 1067 { 1068 pg_data_t *pgdat; 1069 int ret; 1070 1071 if (node_online(nid)) 1072 return 0; 1073 1074 mem_hotplug_begin(); 1075 pgdat = hotadd_new_pgdat(nid, 0); 1076 if (!pgdat) { 1077 pr_err("Cannot online node %d due to NULL pgdat\n", nid); 1078 ret = -ENOMEM; 1079 goto out; 1080 } 1081 node_set_online(nid); 1082 ret = register_one_node(nid); 1083 BUG_ON(ret); 1084 out: 1085 mem_hotplug_done(); 1086 return ret; 1087 } 1088 1089 static int check_hotplug_memory_range(u64 start, u64 size) 1090 { 1091 u64 start_pfn = PFN_DOWN(start); 1092 u64 nr_pages = size >> PAGE_SHIFT; 1093 1094 /* Memory range must be aligned with section */ 1095 if ((start_pfn & ~PAGE_SECTION_MASK) || 1096 (nr_pages % PAGES_PER_SECTION) || (!nr_pages)) { 1097 pr_err("Section-unaligned hotplug range: start 0x%llx, size 0x%llx\n", 1098 (unsigned long long)start, 1099 (unsigned long long)size); 1100 return -EINVAL; 1101 } 1102 1103 return 0; 1104 } 1105 1106 static int online_memory_block(struct memory_block *mem, void *arg) 1107 { 1108 return device_online(&mem->dev); 1109 } 1110 1111 /* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */ 1112 int __ref add_memory_resource(int nid, struct resource *res, bool online) 1113 { 1114 u64 start, size; 1115 pg_data_t *pgdat = NULL; 1116 bool new_pgdat; 1117 bool new_node; 1118 int ret; 1119 1120 start = res->start; 1121 size = resource_size(res); 1122 1123 ret = check_hotplug_memory_range(start, size); 1124 if (ret) 1125 return ret; 1126 1127 { /* Stupid hack to suppress address-never-null warning */ 1128 void *p = NODE_DATA(nid); 1129 new_pgdat = !p; 1130 } 1131 1132 mem_hotplug_begin(); 1133 1134 /* 1135 * Add new range to memblock so that when hotadd_new_pgdat() is called 1136 * to allocate new pgdat, get_pfn_range_for_nid() will be able to find 1137 * this new range and calculate total pages correctly. The range will 1138 * be removed at hot-remove time. 1139 */ 1140 memblock_add_node(start, size, nid); 1141 1142 new_node = !node_online(nid); 1143 if (new_node) { 1144 pgdat = hotadd_new_pgdat(nid, start); 1145 ret = -ENOMEM; 1146 if (!pgdat) 1147 goto error; 1148 } 1149 1150 /* call arch's memory hotadd */ 1151 ret = arch_add_memory(nid, start, size, true); 1152 1153 if (ret < 0) 1154 goto error; 1155 1156 /* we online node here. we can't roll back from here. */ 1157 node_set_online(nid); 1158 1159 if (new_node) { 1160 unsigned long start_pfn = start >> PAGE_SHIFT; 1161 unsigned long nr_pages = size >> PAGE_SHIFT; 1162 1163 ret = __register_one_node(nid); 1164 if (ret) 1165 goto register_fail; 1166 1167 /* 1168 * link memory sections under this node. This is already 1169 * done when creatig memory section in register_new_memory 1170 * but that depends to have the node registered so offline 1171 * nodes have to go through register_node. 1172 * TODO clean up this mess. 1173 */ 1174 ret = link_mem_sections(nid, start_pfn, nr_pages); 1175 register_fail: 1176 /* 1177 * If sysfs file of new node can't create, cpu on the node 1178 * can't be hot-added. There is no rollback way now. 1179 * So, check by BUG_ON() to catch it reluctantly.. 1180 */ 1181 BUG_ON(ret); 1182 } 1183 1184 /* create new memmap entry */ 1185 firmware_map_add_hotplug(start, start + size, "System RAM"); 1186 1187 /* online pages if requested */ 1188 if (online) 1189 walk_memory_range(PFN_DOWN(start), PFN_UP(start + size - 1), 1190 NULL, online_memory_block); 1191 1192 goto out; 1193 1194 error: 1195 /* rollback pgdat allocation and others */ 1196 if (new_pgdat && pgdat) 1197 rollback_node_hotadd(nid, pgdat); 1198 memblock_remove(start, size); 1199 1200 out: 1201 mem_hotplug_done(); 1202 return ret; 1203 } 1204 EXPORT_SYMBOL_GPL(add_memory_resource); 1205 1206 int __ref add_memory(int nid, u64 start, u64 size) 1207 { 1208 struct resource *res; 1209 int ret; 1210 1211 res = register_memory_resource(start, size); 1212 if (IS_ERR(res)) 1213 return PTR_ERR(res); 1214 1215 ret = add_memory_resource(nid, res, memhp_auto_online); 1216 if (ret < 0) 1217 release_memory_resource(res); 1218 return ret; 1219 } 1220 EXPORT_SYMBOL_GPL(add_memory); 1221 1222 #ifdef CONFIG_MEMORY_HOTREMOVE 1223 /* 1224 * A free page on the buddy free lists (not the per-cpu lists) has PageBuddy 1225 * set and the size of the free page is given by page_order(). Using this, 1226 * the function determines if the pageblock contains only free pages. 1227 * Due to buddy contraints, a free page at least the size of a pageblock will 1228 * be located at the start of the pageblock 1229 */ 1230 static inline int pageblock_free(struct page *page) 1231 { 1232 return PageBuddy(page) && page_order(page) >= pageblock_order; 1233 } 1234 1235 /* Return the start of the next active pageblock after a given page */ 1236 static struct page *next_active_pageblock(struct page *page) 1237 { 1238 /* Ensure the starting page is pageblock-aligned */ 1239 BUG_ON(page_to_pfn(page) & (pageblock_nr_pages - 1)); 1240 1241 /* If the entire pageblock is free, move to the end of free page */ 1242 if (pageblock_free(page)) { 1243 int order; 1244 /* be careful. we don't have locks, page_order can be changed.*/ 1245 order = page_order(page); 1246 if ((order < MAX_ORDER) && (order >= pageblock_order)) 1247 return page + (1 << order); 1248 } 1249 1250 return page + pageblock_nr_pages; 1251 } 1252 1253 /* Checks if this range of memory is likely to be hot-removable. */ 1254 bool is_mem_section_removable(unsigned long start_pfn, unsigned long nr_pages) 1255 { 1256 struct page *page = pfn_to_page(start_pfn); 1257 struct page *end_page = page + nr_pages; 1258 1259 /* Check the starting page of each pageblock within the range */ 1260 for (; page < end_page; page = next_active_pageblock(page)) { 1261 if (!is_pageblock_removable_nolock(page)) 1262 return false; 1263 cond_resched(); 1264 } 1265 1266 /* All pageblocks in the memory block are likely to be hot-removable */ 1267 return true; 1268 } 1269 1270 /* 1271 * Confirm all pages in a range [start, end) belong to the same zone. 1272 * When true, return its valid [start, end). 1273 */ 1274 int test_pages_in_a_zone(unsigned long start_pfn, unsigned long end_pfn, 1275 unsigned long *valid_start, unsigned long *valid_end) 1276 { 1277 unsigned long pfn, sec_end_pfn; 1278 unsigned long start, end; 1279 struct zone *zone = NULL; 1280 struct page *page; 1281 int i; 1282 for (pfn = start_pfn, sec_end_pfn = SECTION_ALIGN_UP(start_pfn + 1); 1283 pfn < end_pfn; 1284 pfn = sec_end_pfn, sec_end_pfn += PAGES_PER_SECTION) { 1285 /* Make sure the memory section is present first */ 1286 if (!present_section_nr(pfn_to_section_nr(pfn))) 1287 continue; 1288 for (; pfn < sec_end_pfn && pfn < end_pfn; 1289 pfn += MAX_ORDER_NR_PAGES) { 1290 i = 0; 1291 /* This is just a CONFIG_HOLES_IN_ZONE check.*/ 1292 while ((i < MAX_ORDER_NR_PAGES) && 1293 !pfn_valid_within(pfn + i)) 1294 i++; 1295 if (i == MAX_ORDER_NR_PAGES || pfn + i >= end_pfn) 1296 continue; 1297 page = pfn_to_page(pfn + i); 1298 if (zone && page_zone(page) != zone) 1299 return 0; 1300 if (!zone) 1301 start = pfn + i; 1302 zone = page_zone(page); 1303 end = pfn + MAX_ORDER_NR_PAGES; 1304 } 1305 } 1306 1307 if (zone) { 1308 *valid_start = start; 1309 *valid_end = min(end, end_pfn); 1310 return 1; 1311 } else { 1312 return 0; 1313 } 1314 } 1315 1316 /* 1317 * Scan pfn range [start,end) to find movable/migratable pages (LRU pages, 1318 * non-lru movable pages and hugepages). We scan pfn because it's much 1319 * easier than scanning over linked list. This function returns the pfn 1320 * of the first found movable page if it's found, otherwise 0. 1321 */ 1322 static unsigned long scan_movable_pages(unsigned long start, unsigned long end) 1323 { 1324 unsigned long pfn; 1325 struct page *page; 1326 for (pfn = start; pfn < end; pfn++) { 1327 if (pfn_valid(pfn)) { 1328 page = pfn_to_page(pfn); 1329 if (PageLRU(page)) 1330 return pfn; 1331 if (__PageMovable(page)) 1332 return pfn; 1333 if (PageHuge(page)) { 1334 if (page_huge_active(page)) 1335 return pfn; 1336 else 1337 pfn = round_up(pfn + 1, 1338 1 << compound_order(page)) - 1; 1339 } 1340 } 1341 } 1342 return 0; 1343 } 1344 1345 static struct page *new_node_page(struct page *page, unsigned long private, 1346 int **result) 1347 { 1348 int nid = page_to_nid(page); 1349 nodemask_t nmask = node_states[N_MEMORY]; 1350 1351 /* 1352 * try to allocate from a different node but reuse this node if there 1353 * are no other online nodes to be used (e.g. we are offlining a part 1354 * of the only existing node) 1355 */ 1356 node_clear(nid, nmask); 1357 if (nodes_empty(nmask)) 1358 node_set(nid, nmask); 1359 1360 return new_page_nodemask(page, nid, &nmask); 1361 } 1362 1363 #define NR_OFFLINE_AT_ONCE_PAGES (256) 1364 static int 1365 do_migrate_range(unsigned long start_pfn, unsigned long end_pfn) 1366 { 1367 unsigned long pfn; 1368 struct page *page; 1369 int move_pages = NR_OFFLINE_AT_ONCE_PAGES; 1370 int not_managed = 0; 1371 int ret = 0; 1372 LIST_HEAD(source); 1373 1374 for (pfn = start_pfn; pfn < end_pfn && move_pages > 0; pfn++) { 1375 if (!pfn_valid(pfn)) 1376 continue; 1377 page = pfn_to_page(pfn); 1378 1379 if (PageHuge(page)) { 1380 struct page *head = compound_head(page); 1381 pfn = page_to_pfn(head) + (1<<compound_order(head)) - 1; 1382 if (compound_order(head) > PFN_SECTION_SHIFT) { 1383 ret = -EBUSY; 1384 break; 1385 } 1386 if (isolate_huge_page(page, &source)) 1387 move_pages -= 1 << compound_order(head); 1388 continue; 1389 } else if (thp_migration_supported() && PageTransHuge(page)) 1390 pfn = page_to_pfn(compound_head(page)) 1391 + hpage_nr_pages(page) - 1; 1392 1393 if (!get_page_unless_zero(page)) 1394 continue; 1395 /* 1396 * We can skip free pages. And we can deal with pages on 1397 * LRU and non-lru movable pages. 1398 */ 1399 if (PageLRU(page)) 1400 ret = isolate_lru_page(page); 1401 else 1402 ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE); 1403 if (!ret) { /* Success */ 1404 put_page(page); 1405 list_add_tail(&page->lru, &source); 1406 move_pages--; 1407 if (!__PageMovable(page)) 1408 inc_node_page_state(page, NR_ISOLATED_ANON + 1409 page_is_file_cache(page)); 1410 1411 } else { 1412 #ifdef CONFIG_DEBUG_VM 1413 pr_alert("failed to isolate pfn %lx\n", pfn); 1414 dump_page(page, "isolation failed"); 1415 #endif 1416 put_page(page); 1417 /* Because we don't have big zone->lock. we should 1418 check this again here. */ 1419 if (page_count(page)) { 1420 not_managed++; 1421 ret = -EBUSY; 1422 break; 1423 } 1424 } 1425 } 1426 if (!list_empty(&source)) { 1427 if (not_managed) { 1428 putback_movable_pages(&source); 1429 goto out; 1430 } 1431 1432 /* Allocate a new page from the nearest neighbor node */ 1433 ret = migrate_pages(&source, new_node_page, NULL, 0, 1434 MIGRATE_SYNC, MR_MEMORY_HOTPLUG); 1435 if (ret) 1436 putback_movable_pages(&source); 1437 } 1438 out: 1439 return ret; 1440 } 1441 1442 /* 1443 * remove from free_area[] and mark all as Reserved. 1444 */ 1445 static int 1446 offline_isolated_pages_cb(unsigned long start, unsigned long nr_pages, 1447 void *data) 1448 { 1449 __offline_isolated_pages(start, start + nr_pages); 1450 return 0; 1451 } 1452 1453 static void 1454 offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 1455 { 1456 walk_system_ram_range(start_pfn, end_pfn - start_pfn, NULL, 1457 offline_isolated_pages_cb); 1458 } 1459 1460 /* 1461 * Check all pages in range, recoreded as memory resource, are isolated. 1462 */ 1463 static int 1464 check_pages_isolated_cb(unsigned long start_pfn, unsigned long nr_pages, 1465 void *data) 1466 { 1467 int ret; 1468 long offlined = *(long *)data; 1469 ret = test_pages_isolated(start_pfn, start_pfn + nr_pages, true); 1470 offlined = nr_pages; 1471 if (!ret) 1472 *(long *)data += offlined; 1473 return ret; 1474 } 1475 1476 static long 1477 check_pages_isolated(unsigned long start_pfn, unsigned long end_pfn) 1478 { 1479 long offlined = 0; 1480 int ret; 1481 1482 ret = walk_system_ram_range(start_pfn, end_pfn - start_pfn, &offlined, 1483 check_pages_isolated_cb); 1484 if (ret < 0) 1485 offlined = (long)ret; 1486 return offlined; 1487 } 1488 1489 static int __init cmdline_parse_movable_node(char *p) 1490 { 1491 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 1492 movable_node_enabled = true; 1493 #else 1494 pr_warn("movable_node parameter depends on CONFIG_HAVE_MEMBLOCK_NODE_MAP to work properly\n"); 1495 #endif 1496 return 0; 1497 } 1498 early_param("movable_node", cmdline_parse_movable_node); 1499 1500 /* check which state of node_states will be changed when offline memory */ 1501 static void node_states_check_changes_offline(unsigned long nr_pages, 1502 struct zone *zone, struct memory_notify *arg) 1503 { 1504 struct pglist_data *pgdat = zone->zone_pgdat; 1505 unsigned long present_pages = 0; 1506 enum zone_type zt, zone_last = ZONE_NORMAL; 1507 1508 /* 1509 * If we have HIGHMEM or movable node, node_states[N_NORMAL_MEMORY] 1510 * contains nodes which have zones of 0...ZONE_NORMAL, 1511 * set zone_last to ZONE_NORMAL. 1512 * 1513 * If we don't have HIGHMEM nor movable node, 1514 * node_states[N_NORMAL_MEMORY] contains nodes which have zones of 1515 * 0...ZONE_MOVABLE, set zone_last to ZONE_MOVABLE. 1516 */ 1517 if (N_MEMORY == N_NORMAL_MEMORY) 1518 zone_last = ZONE_MOVABLE; 1519 1520 /* 1521 * check whether node_states[N_NORMAL_MEMORY] will be changed. 1522 * If the memory to be offline is in a zone of 0...zone_last, 1523 * and it is the last present memory, 0...zone_last will 1524 * become empty after offline , thus we can determind we will 1525 * need to clear the node from node_states[N_NORMAL_MEMORY]. 1526 */ 1527 for (zt = 0; zt <= zone_last; zt++) 1528 present_pages += pgdat->node_zones[zt].present_pages; 1529 if (zone_idx(zone) <= zone_last && nr_pages >= present_pages) 1530 arg->status_change_nid_normal = zone_to_nid(zone); 1531 else 1532 arg->status_change_nid_normal = -1; 1533 1534 #ifdef CONFIG_HIGHMEM 1535 /* 1536 * If we have movable node, node_states[N_HIGH_MEMORY] 1537 * contains nodes which have zones of 0...ZONE_HIGHMEM, 1538 * set zone_last to ZONE_HIGHMEM. 1539 * 1540 * If we don't have movable node, node_states[N_NORMAL_MEMORY] 1541 * contains nodes which have zones of 0...ZONE_MOVABLE, 1542 * set zone_last to ZONE_MOVABLE. 1543 */ 1544 zone_last = ZONE_HIGHMEM; 1545 if (N_MEMORY == N_HIGH_MEMORY) 1546 zone_last = ZONE_MOVABLE; 1547 1548 for (; zt <= zone_last; zt++) 1549 present_pages += pgdat->node_zones[zt].present_pages; 1550 if (zone_idx(zone) <= zone_last && nr_pages >= present_pages) 1551 arg->status_change_nid_high = zone_to_nid(zone); 1552 else 1553 arg->status_change_nid_high = -1; 1554 #else 1555 arg->status_change_nid_high = arg->status_change_nid_normal; 1556 #endif 1557 1558 /* 1559 * node_states[N_HIGH_MEMORY] contains nodes which have 0...ZONE_MOVABLE 1560 */ 1561 zone_last = ZONE_MOVABLE; 1562 1563 /* 1564 * check whether node_states[N_HIGH_MEMORY] will be changed 1565 * If we try to offline the last present @nr_pages from the node, 1566 * we can determind we will need to clear the node from 1567 * node_states[N_HIGH_MEMORY]. 1568 */ 1569 for (; zt <= zone_last; zt++) 1570 present_pages += pgdat->node_zones[zt].present_pages; 1571 if (nr_pages >= present_pages) 1572 arg->status_change_nid = zone_to_nid(zone); 1573 else 1574 arg->status_change_nid = -1; 1575 } 1576 1577 static void node_states_clear_node(int node, struct memory_notify *arg) 1578 { 1579 if (arg->status_change_nid_normal >= 0) 1580 node_clear_state(node, N_NORMAL_MEMORY); 1581 1582 if ((N_MEMORY != N_NORMAL_MEMORY) && 1583 (arg->status_change_nid_high >= 0)) 1584 node_clear_state(node, N_HIGH_MEMORY); 1585 1586 if ((N_MEMORY != N_HIGH_MEMORY) && 1587 (arg->status_change_nid >= 0)) 1588 node_clear_state(node, N_MEMORY); 1589 } 1590 1591 static int __ref __offline_pages(unsigned long start_pfn, 1592 unsigned long end_pfn, unsigned long timeout) 1593 { 1594 unsigned long pfn, nr_pages, expire; 1595 long offlined_pages; 1596 int ret, drain, retry_max, node; 1597 unsigned long flags; 1598 unsigned long valid_start, valid_end; 1599 struct zone *zone; 1600 struct memory_notify arg; 1601 1602 /* at least, alignment against pageblock is necessary */ 1603 if (!IS_ALIGNED(start_pfn, pageblock_nr_pages)) 1604 return -EINVAL; 1605 if (!IS_ALIGNED(end_pfn, pageblock_nr_pages)) 1606 return -EINVAL; 1607 /* This makes hotplug much easier...and readable. 1608 we assume this for now. .*/ 1609 if (!test_pages_in_a_zone(start_pfn, end_pfn, &valid_start, &valid_end)) 1610 return -EINVAL; 1611 1612 zone = page_zone(pfn_to_page(valid_start)); 1613 node = zone_to_nid(zone); 1614 nr_pages = end_pfn - start_pfn; 1615 1616 /* set above range as isolated */ 1617 ret = start_isolate_page_range(start_pfn, end_pfn, 1618 MIGRATE_MOVABLE, true); 1619 if (ret) 1620 return ret; 1621 1622 arg.start_pfn = start_pfn; 1623 arg.nr_pages = nr_pages; 1624 node_states_check_changes_offline(nr_pages, zone, &arg); 1625 1626 ret = memory_notify(MEM_GOING_OFFLINE, &arg); 1627 ret = notifier_to_errno(ret); 1628 if (ret) 1629 goto failed_removal; 1630 1631 pfn = start_pfn; 1632 expire = jiffies + timeout; 1633 drain = 0; 1634 retry_max = 5; 1635 repeat: 1636 /* start memory hot removal */ 1637 ret = -EAGAIN; 1638 if (time_after(jiffies, expire)) 1639 goto failed_removal; 1640 ret = -EINTR; 1641 if (signal_pending(current)) 1642 goto failed_removal; 1643 ret = 0; 1644 if (drain) { 1645 lru_add_drain_all_cpuslocked(); 1646 cond_resched(); 1647 drain_all_pages(zone); 1648 } 1649 1650 pfn = scan_movable_pages(start_pfn, end_pfn); 1651 if (pfn) { /* We have movable pages */ 1652 ret = do_migrate_range(pfn, end_pfn); 1653 if (!ret) { 1654 drain = 1; 1655 goto repeat; 1656 } else { 1657 if (ret < 0) 1658 if (--retry_max == 0) 1659 goto failed_removal; 1660 yield(); 1661 drain = 1; 1662 goto repeat; 1663 } 1664 } 1665 /* drain all zone's lru pagevec, this is asynchronous... */ 1666 lru_add_drain_all_cpuslocked(); 1667 yield(); 1668 /* drain pcp pages, this is synchronous. */ 1669 drain_all_pages(zone); 1670 /* 1671 * dissolve free hugepages in the memory block before doing offlining 1672 * actually in order to make hugetlbfs's object counting consistent. 1673 */ 1674 ret = dissolve_free_huge_pages(start_pfn, end_pfn); 1675 if (ret) 1676 goto failed_removal; 1677 /* check again */ 1678 offlined_pages = check_pages_isolated(start_pfn, end_pfn); 1679 if (offlined_pages < 0) { 1680 ret = -EBUSY; 1681 goto failed_removal; 1682 } 1683 pr_info("Offlined Pages %ld\n", offlined_pages); 1684 /* Ok, all of our target is isolated. 1685 We cannot do rollback at this point. */ 1686 offline_isolated_pages(start_pfn, end_pfn); 1687 /* reset pagetype flags and makes migrate type to be MOVABLE */ 1688 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1689 /* removal success */ 1690 adjust_managed_page_count(pfn_to_page(start_pfn), -offlined_pages); 1691 zone->present_pages -= offlined_pages; 1692 1693 pgdat_resize_lock(zone->zone_pgdat, &flags); 1694 zone->zone_pgdat->node_present_pages -= offlined_pages; 1695 pgdat_resize_unlock(zone->zone_pgdat, &flags); 1696 1697 init_per_zone_wmark_min(); 1698 1699 if (!populated_zone(zone)) { 1700 zone_pcp_reset(zone); 1701 build_all_zonelists(NULL); 1702 } else 1703 zone_pcp_update(zone); 1704 1705 node_states_clear_node(node, &arg); 1706 if (arg.status_change_nid >= 0) { 1707 kswapd_stop(node); 1708 kcompactd_stop(node); 1709 } 1710 1711 vm_total_pages = nr_free_pagecache_pages(); 1712 writeback_set_ratelimit(); 1713 1714 memory_notify(MEM_OFFLINE, &arg); 1715 return 0; 1716 1717 failed_removal: 1718 pr_debug("memory offlining [mem %#010llx-%#010llx] failed\n", 1719 (unsigned long long) start_pfn << PAGE_SHIFT, 1720 ((unsigned long long) end_pfn << PAGE_SHIFT) - 1); 1721 memory_notify(MEM_CANCEL_OFFLINE, &arg); 1722 /* pushback to free area */ 1723 undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE); 1724 return ret; 1725 } 1726 1727 /* Must be protected by mem_hotplug_begin() or a device_lock */ 1728 int offline_pages(unsigned long start_pfn, unsigned long nr_pages) 1729 { 1730 return __offline_pages(start_pfn, start_pfn + nr_pages, 120 * HZ); 1731 } 1732 #endif /* CONFIG_MEMORY_HOTREMOVE */ 1733 1734 /** 1735 * walk_memory_range - walks through all mem sections in [start_pfn, end_pfn) 1736 * @start_pfn: start pfn of the memory range 1737 * @end_pfn: end pfn of the memory range 1738 * @arg: argument passed to func 1739 * @func: callback for each memory section walked 1740 * 1741 * This function walks through all present mem sections in range 1742 * [start_pfn, end_pfn) and call func on each mem section. 1743 * 1744 * Returns the return value of func. 1745 */ 1746 int walk_memory_range(unsigned long start_pfn, unsigned long end_pfn, 1747 void *arg, int (*func)(struct memory_block *, void *)) 1748 { 1749 struct memory_block *mem = NULL; 1750 struct mem_section *section; 1751 unsigned long pfn, section_nr; 1752 int ret; 1753 1754 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 1755 section_nr = pfn_to_section_nr(pfn); 1756 if (!present_section_nr(section_nr)) 1757 continue; 1758 1759 section = __nr_to_section(section_nr); 1760 /* same memblock? */ 1761 if (mem) 1762 if ((section_nr >= mem->start_section_nr) && 1763 (section_nr <= mem->end_section_nr)) 1764 continue; 1765 1766 mem = find_memory_block_hinted(section, mem); 1767 if (!mem) 1768 continue; 1769 1770 ret = func(mem, arg); 1771 if (ret) { 1772 kobject_put(&mem->dev.kobj); 1773 return ret; 1774 } 1775 } 1776 1777 if (mem) 1778 kobject_put(&mem->dev.kobj); 1779 1780 return 0; 1781 } 1782 1783 #ifdef CONFIG_MEMORY_HOTREMOVE 1784 static int check_memblock_offlined_cb(struct memory_block *mem, void *arg) 1785 { 1786 int ret = !is_memblock_offlined(mem); 1787 1788 if (unlikely(ret)) { 1789 phys_addr_t beginpa, endpa; 1790 1791 beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr)); 1792 endpa = PFN_PHYS(section_nr_to_pfn(mem->end_section_nr + 1))-1; 1793 pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n", 1794 &beginpa, &endpa); 1795 } 1796 1797 return ret; 1798 } 1799 1800 static int check_cpu_on_node(pg_data_t *pgdat) 1801 { 1802 int cpu; 1803 1804 for_each_present_cpu(cpu) { 1805 if (cpu_to_node(cpu) == pgdat->node_id) 1806 /* 1807 * the cpu on this node isn't removed, and we can't 1808 * offline this node. 1809 */ 1810 return -EBUSY; 1811 } 1812 1813 return 0; 1814 } 1815 1816 static void unmap_cpu_on_node(pg_data_t *pgdat) 1817 { 1818 #ifdef CONFIG_ACPI_NUMA 1819 int cpu; 1820 1821 for_each_possible_cpu(cpu) 1822 if (cpu_to_node(cpu) == pgdat->node_id) 1823 numa_clear_node(cpu); 1824 #endif 1825 } 1826 1827 static int check_and_unmap_cpu_on_node(pg_data_t *pgdat) 1828 { 1829 int ret; 1830 1831 ret = check_cpu_on_node(pgdat); 1832 if (ret) 1833 return ret; 1834 1835 /* 1836 * the node will be offlined when we come here, so we can clear 1837 * the cpu_to_node() now. 1838 */ 1839 1840 unmap_cpu_on_node(pgdat); 1841 return 0; 1842 } 1843 1844 /** 1845 * try_offline_node 1846 * 1847 * Offline a node if all memory sections and cpus of the node are removed. 1848 * 1849 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1850 * and online/offline operations before this call. 1851 */ 1852 void try_offline_node(int nid) 1853 { 1854 pg_data_t *pgdat = NODE_DATA(nid); 1855 unsigned long start_pfn = pgdat->node_start_pfn; 1856 unsigned long end_pfn = start_pfn + pgdat->node_spanned_pages; 1857 unsigned long pfn; 1858 1859 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 1860 unsigned long section_nr = pfn_to_section_nr(pfn); 1861 1862 if (!present_section_nr(section_nr)) 1863 continue; 1864 1865 if (pfn_to_nid(pfn) != nid) 1866 continue; 1867 1868 /* 1869 * some memory sections of this node are not removed, and we 1870 * can't offline node now. 1871 */ 1872 return; 1873 } 1874 1875 if (check_and_unmap_cpu_on_node(pgdat)) 1876 return; 1877 1878 /* 1879 * all memory/cpu of this node are removed, we can offline this 1880 * node now. 1881 */ 1882 node_set_offline(nid); 1883 unregister_one_node(nid); 1884 } 1885 EXPORT_SYMBOL(try_offline_node); 1886 1887 /** 1888 * remove_memory 1889 * 1890 * NOTE: The caller must call lock_device_hotplug() to serialize hotplug 1891 * and online/offline operations before this call, as required by 1892 * try_offline_node(). 1893 */ 1894 void __ref remove_memory(int nid, u64 start, u64 size) 1895 { 1896 int ret; 1897 1898 BUG_ON(check_hotplug_memory_range(start, size)); 1899 1900 mem_hotplug_begin(); 1901 1902 /* 1903 * All memory blocks must be offlined before removing memory. Check 1904 * whether all memory blocks in question are offline and trigger a BUG() 1905 * if this is not the case. 1906 */ 1907 ret = walk_memory_range(PFN_DOWN(start), PFN_UP(start + size - 1), NULL, 1908 check_memblock_offlined_cb); 1909 if (ret) 1910 BUG(); 1911 1912 /* remove memmap entry */ 1913 firmware_map_remove(start, start + size, "System RAM"); 1914 memblock_free(start, size); 1915 memblock_remove(start, size); 1916 1917 arch_remove_memory(start, size); 1918 1919 try_offline_node(nid); 1920 1921 mem_hotplug_done(); 1922 } 1923 EXPORT_SYMBOL_GPL(remove_memory); 1924 #endif /* CONFIG_MEMORY_HOTREMOVE */ 1925