| 1 | |
| 2 | |
| 3 | * linux/mm/memory_hotplug.c
|
| 4 | *
|
| 5 | * Copyright (C)
|
| 6 |
|
| 7 |
|
| 8 | #include <linux/stddef.h>
|
| 9 | #include <linux/mm.h>
|
| 10 | #include <linux/sched/signal.h>
|
| 11 | #include <linux/swap.h>
|
| 12 | #include <linux/interrupt.h>
|
| 13 | #include <linux/pagemap.h>
|
| 14 | #include <linux/compiler.h>
|
| 15 | #include <linux/export.h>
|
| 16 | #include <linux/pagevec.h>
|
| 17 | #include <linux/writeback.h>
|
| 18 | #include <linux/slab.h>
|
| 19 | #include <linux/sysctl.h>
|
| 20 | #include <linux/cpu.h>
|
| 21 | #include <linux/memory.h>
|
| 22 | #include <linux/memremap.h>
|
| 23 | #include <linux/memory_hotplug.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/compaction.h>
|
| 37 | #include <linux/rmap.h>
|
| 38 | #include <linux/module.h>
|
| 39 |
|
| 40 | #include <asm/tlbflush.h>
|
| 41 |
|
| 42 | #include "internal.h"
|
| 43 | #include "shuffle.h"
|
| 44 |
|
| 45 | #ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
|
| 46 | |
| 47 | * memory_hotplug.memmap_on_memory parameter
|
| 48 |
|
| 49 | static bool memmap_on_memory __ro_after_init;
|
| 50 | module_param(memmap_on_memory, bool, 0444);
|
| 51 | MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug");
|
| 52 |
|
| 53 | static inline bool mhp_memmap_on_memory(void)
|
| 54 | {
|
| 55 | return memmap_on_memory;
|
| 56 | }
|
| 57 | #else
|
| 58 | static inline bool mhp_memmap_on_memory(void)
|
| 59 | {
|
| 60 | return false;
|
| 61 | }
|
| 62 | #endif
|
| 63 |
|
| 64 | enum {
|
| 65 | ONLINE_POLICY_CONTIG_ZONES = 0,
|
| 66 | ONLINE_POLICY_AUTO_MOVABLE,
|
| 67 | };
|
| 68 |
|
| 69 | static const char * const online_policy_to_str[] = {
|
| 70 | [ONLINE_POLICY_CONTIG_ZONES] = "contig-zones",
|
| 71 | [ONLINE_POLICY_AUTO_MOVABLE] = "auto-movable",
|
| 72 | };
|
| 73 |
|
| 74 | static int set_online_policy(const char *val, const struct kernel_param *kp)
|
| 75 | {
|
| 76 | int ret = sysfs_match_string(online_policy_to_str, val);
|
| 77 |
|
| 78 | if (ret < 0)
|
| 79 | return ret;
|
| 80 | *((int *)kp->arg) = ret;
|
| 81 | return 0;
|
| 82 | }
|
| 83 |
|
| 84 | static int get_online_policy(char *buffer, const struct kernel_param *kp)
|
| 85 | {
|
| 86 | return sprintf(buffer, "%s\n", online_policy_to_str[*((int *)kp->arg)]);
|
| 87 | }
|
| 88 |
|
| 89 | |
| 90 | * memory_hotplug.online_policy: configure online behavior when onlining without
|
| 91 | * specifying a zone (MMOP_ONLINE)
|
| 92 | *
|
| 93 | * "contig-zones": keep zone contiguous
|
| 94 | * "auto-movable": online memory to ZONE_MOVABLE if the configuration
|
| 95 | * (auto_movable_ratio, auto_movable_numa_aware) allows for it
|
| 96 |
|
| 97 | static int online_policy __read_mostly = ONLINE_POLICY_CONTIG_ZONES;
|
| 98 | static const struct kernel_param_ops online_policy_ops = {
|
| 99 | .set = set_online_policy,
|
| 100 | .get = get_online_policy,
|
| 101 | };
|
| 102 | module_param_cb(online_policy, &online_policy_ops, &online_policy, 0644);
|
| 103 | MODULE_PARM_DESC(online_policy,
|
| 104 | "Set the online policy (\"contig-zones\", \"auto-movable\") "
|
| 105 | "Default: \"contig-zones\"");
|
| 106 |
|
| 107 | |
| 108 | * memory_hotplug.auto_movable_ratio: specify maximum MOVABLE:KERNEL ratio
|
| 109 | *
|
| 110 | * The ratio represent an upper limit and the kernel might decide to not
|
| 111 | * online some memory to ZONE_MOVABLE -- e.g., because hotplugged KERNEL memory
|
| 112 | * doesn't allow for more MOVABLE memory.
|
| 113 |
|
| 114 | static unsigned int auto_movable_ratio __read_mostly = 301;
|
| 115 | module_param(auto_movable_ratio, uint, 0644);
|
| 116 | MODULE_PARM_DESC(auto_movable_ratio,
|
| 117 | "Set the maximum ratio of MOVABLE:KERNEL memory in the system "
|
| 118 | "in percent for \"auto-movable\" online policy. Default: 301");
|
| 119 |
|
| 120 | |
| 121 | * memory_hotplug.auto_movable_numa_aware: consider numa node stats
|
| 122 |
|
| 123 | #ifdef CONFIG_NUMA
|
| 124 | static bool auto_movable_numa_aware __read_mostly = true;
|
| 125 | module_param(auto_movable_numa_aware, bool, 0644);
|
| 126 | MODULE_PARM_DESC(auto_movable_numa_aware,
|
| 127 | "Consider numa node stats in addition to global stats in "
|
| 128 | "\"auto-movable\" online policy. Default: true");
|
| 129 | #endif
|
| 130 |
|
| 131 | |
| 132 | * online_page_callback contains pointer to current page onlining function.
|
| 133 | * Initially it is generic_online_page(). If it is required it could be
|
| 134 | * changed by calling set_online_page_callback() for callback registration
|
| 135 | * and restore_online_page_callback() for generic callback restore.
|
| 136 |
|
| 137 |
|
| 138 | static online_page_callback_t online_page_callback = generic_online_page;
|
| 139 | static DEFINE_MUTEX(online_page_callback_lock);
|
| 140 |
|
| 141 | DEFINE_STATIC_PERCPU_RWSEM(mem_hotplug_lock);
|
| 142 |
|
| 143 | void get_online_mems(void)
|
| 144 | {
|
| 145 | percpu_down_read(&mem_hotplug_lock);
|
| 146 | }
|
| 147 |
|
| 148 | void put_online_mems(void)
|
| 149 | {
|
| 150 | percpu_up_read(&mem_hotplug_lock);
|
| 151 | }
|
| 152 |
|
| 153 | bool movable_node_enabled = false;
|
| 154 |
|
| 155 | #ifndef CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE
|
| 156 | int mhp_default_online_type = MMOP_OFFLINE;
|
| 157 | #else
|
| 158 | int mhp_default_online_type = MMOP_ONLINE;
|
| 159 | #endif
|
| 160 |
|
| 161 | static int __init setup_memhp_default_state(char *str)
|
| 162 | {
|
| 163 | const int online_type = mhp_online_type_from_str(str);
|
| 164 |
|
| 165 | if (online_type >= 0)
|
| 166 | mhp_default_online_type = online_type;
|
| 167 |
|
| 168 | return 1;
|
| 169 | }
|
| 170 | __setup("memhp_default_state=", setup_memhp_default_state);
|
| 171 |
|
| 172 | void mem_hotplug_begin(void)
|
| 173 | {
|
| 174 | cpus_read_lock();
|
| 175 | percpu_down_write(&mem_hotplug_lock);
|
| 176 | }
|
| 177 |
|
| 178 | void mem_hotplug_done(void)
|
| 179 | {
|
| 180 | percpu_up_write(&mem_hotplug_lock);
|
| 181 | cpus_read_unlock();
|
| 182 | }
|
| 183 |
|
| 184 | u64 max_mem_size = U64_MAX;
|
| 185 |
|
| 186 |
|
| 187 | static struct resource *register_memory_resource(u64 start, u64 size,
|
| 188 | const char *resource_name)
|
| 189 | {
|
| 190 | struct resource *res;
|
| 191 | unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
|
| 192 |
|
| 193 | if (strcmp(resource_name, "System RAM"))
|
| 194 | flags |= IORESOURCE_SYSRAM_DRIVER_MANAGED;
|
| 195 |
|
| 196 | if (!mhp_range_allowed(start, size, true))
|
| 197 | return ERR_PTR(-E2BIG);
|
| 198 |
|
| 199 | |
| 200 | * Make sure value parsed from 'mem=' only restricts memory adding
|
| 201 | * while booting, so that memory hotplug won't be impacted. Please
|
| 202 | * refer to document of 'mem=' in kernel-parameters.txt for more
|
| 203 | * details.
|
| 204 |
|
| 205 | if (start + size > max_mem_size && system_state < SYSTEM_RUNNING)
|
| 206 | return ERR_PTR(-E2BIG);
|
| 207 |
|
| 208 | |
| 209 | * Request ownership of the new memory range. This might be
|
| 210 | * a child of an existing resource that was present but
|
| 211 | * not marked as busy.
|
| 212 |
|
| 213 | res = __request_region(&iomem_resource, start, size,
|
| 214 | resource_name, flags);
|
| 215 |
|
| 216 | if (!res) {
|
| 217 | pr_debug("Unable to reserve System RAM region: %016llx->%016llx\n",
|
| 218 | start, start + size);
|
| 219 | return ERR_PTR(-EEXIST);
|
| 220 | }
|
| 221 | return res;
|
| 222 | }
|
| 223 |
|
| 224 | static void release_memory_resource(struct resource *res)
|
| 225 | {
|
| 226 | if (!res)
|
| 227 | return;
|
| 228 | release_resource(res);
|
| 229 | kfree(res);
|
| 230 | }
|
| 231 |
|
| 232 | static int check_pfn_span(unsigned long pfn, unsigned long nr_pages)
|
| 233 | {
|
| 234 | |
| 235 | * Disallow all operations smaller than a sub-section and only
|
| 236 | * allow operations smaller than a section for
|
| 237 | * SPARSEMEM_VMEMMAP. Note that check_hotplug_memory_range()
|
| 238 | * enforces a larger memory_block_size_bytes() granularity for
|
| 239 | * memory that will be marked online, so this check should only
|
| 240 | * fire for direct arch_{add,remove}_memory() users outside of
|
| 241 | * add_memory_resource().
|
| 242 |
|
| 243 | unsigned long min_align;
|
| 244 |
|
| 245 | if (IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
|
| 246 | min_align = PAGES_PER_SUBSECTION;
|
| 247 | else
|
| 248 | min_align = PAGES_PER_SECTION;
|
| 249 | if (!IS_ALIGNED(pfn | nr_pages, min_align))
|
| 250 | return -EINVAL;
|
| 251 | return 0;
|
| 252 | }
|
| 253 |
|
| 254 | |
| 255 | * Return page for the valid pfn only if the page is online. All pfn
|
| 256 | * walkers which rely on the fully initialized page->flags and others
|
| 257 | * should use this rather than pfn_valid && pfn_to_page
|
| 258 |
|
| 259 | struct page *pfn_to_online_page(unsigned long pfn)
|
| 260 | {
|
| 261 | unsigned long nr = pfn_to_section_nr(pfn);
|
| 262 | struct dev_pagemap *pgmap;
|
| 263 | struct mem_section *ms;
|
| 264 |
|
| 265 | if (nr >= NR_MEM_SECTIONS)
|
| 266 | return NULL;
|
| 267 |
|
| 268 | ms = __nr_to_section(nr);
|
| 269 | if (!online_section(ms))
|
| 270 | return NULL;
|
| 271 |
|
| 272 | |
| 273 | * Save some code text when online_section() +
|
| 274 | * pfn_section_valid() are sufficient.
|
| 275 |
|
| 276 | if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
|
| 277 | return NULL;
|
| 278 |
|
| 279 | if (!pfn_section_valid(ms, pfn))
|
| 280 | return NULL;
|
| 281 |
|
| 282 | if (!online_device_section(ms))
|
| 283 | return pfn_to_page(pfn);
|
| 284 |
|
| 285 | |
| 286 | * Slowpath: when ZONE_DEVICE collides with
|
| 287 | * ZONE_{NORMAL,MOVABLE} within the same section some pfns in
|
| 288 | * the section may be 'offline' but 'valid'. Only
|
| 289 | * get_dev_pagemap() can determine sub-section online status.
|
| 290 |
|
| 291 | pgmap = get_dev_pagemap(pfn, NULL);
|
| 292 | put_dev_pagemap(pgmap);
|
| 293 |
|
| 294 |
|
| 295 | if (pgmap)
|
| 296 | return NULL;
|
| 297 |
|
| 298 | return pfn_to_page(pfn);
|
| 299 | }
|
| 300 | EXPORT_SYMBOL_GPL(pfn_to_online_page);
|
| 301 |
|
| 302 | int __ref __add_pages(int nid, unsigned long pfn, unsigned long nr_pages,
|
| 303 | struct mhp_params *params)
|
| 304 | {
|
| 305 | const unsigned long end_pfn = pfn + nr_pages;
|
| 306 | unsigned long cur_nr_pages;
|
| 307 | int err;
|
| 308 | struct vmem_altmap *altmap = params->altmap;
|
| 309 |
|
| 310 | if (WARN_ON_ONCE(!pgprot_val(params->pgprot)))
|
| 311 | return -EINVAL;
|
| 312 |
|
| 313 | VM_BUG_ON(!mhp_range_allowed(PFN_PHYS(pfn), nr_pages * PAGE_SIZE, false));
|
| 314 |
|
| 315 | if (altmap) {
|
| 316 | |
| 317 | * Validate altmap is within bounds of the total request
|
| 318 |
|
| 319 | if (altmap->base_pfn != pfn
|
| 320 | || vmem_altmap_offset(altmap) > nr_pages) {
|
| 321 | pr_warn_once("memory add fail, invalid altmap\n");
|
| 322 | return -EINVAL;
|
| 323 | }
|
| 324 | altmap->alloc = 0;
|
| 325 | }
|
| 326 |
|
| 327 | if (check_pfn_span(pfn, nr_pages)) {
|
| 328 | WARN(1, "Misaligned %s start: %#lx end: #%lx\n", __func__, pfn, pfn + nr_pages - 1);
|
| 329 | return -EINVAL;
|
| 330 | }
|
| 331 |
|
| 332 | for (; pfn < end_pfn; pfn += cur_nr_pages) {
|
| 333 |
|
| 334 | cur_nr_pages = min(end_pfn - pfn,
|
| 335 | SECTION_ALIGN_UP(pfn + 1) - pfn);
|
| 336 | err = sparse_add_section(nid, pfn, cur_nr_pages, altmap,
|
| 337 | params->pgmap);
|
| 338 | if (err)
|
| 339 | break;
|
| 340 | cond_resched();
|
| 341 | }
|
| 342 | vmemmap_populate_print_last();
|
| 343 | return err;
|
| 344 | }
|
| 345 |
|
| 346 |
|
| 347 | static unsigned long find_smallest_section_pfn(int nid, struct zone *zone,
|
| 348 | unsigned long start_pfn,
|
| 349 | unsigned long end_pfn)
|
| 350 | {
|
| 351 | for (; start_pfn < end_pfn; start_pfn += PAGES_PER_SUBSECTION) {
|
| 352 | if (unlikely(!pfn_to_online_page(start_pfn)))
|
| 353 | continue;
|
| 354 |
|
| 355 | if (unlikely(pfn_to_nid(start_pfn) != nid))
|
| 356 | continue;
|
| 357 |
|
| 358 | if (zone != page_zone(pfn_to_page(start_pfn)))
|
| 359 | continue;
|
| 360 |
|
| 361 | return start_pfn;
|
| 362 | }
|
| 363 |
|
| 364 | return 0;
|
| 365 | }
|
| 366 |
|
| 367 |
|
| 368 | static unsigned long find_biggest_section_pfn(int nid, struct zone *zone,
|
| 369 | unsigned long start_pfn,
|
| 370 | unsigned long end_pfn)
|
| 371 | {
|
| 372 | unsigned long pfn;
|
| 373 |
|
| 374 |
|
| 375 | pfn = end_pfn - 1;
|
| 376 | for (; pfn >= start_pfn; pfn -= PAGES_PER_SUBSECTION) {
|
| 377 | if (unlikely(!pfn_to_online_page(pfn)))
|
| 378 | continue;
|
| 379 |
|
| 380 | if (unlikely(pfn_to_nid(pfn) != nid))
|
| 381 | continue;
|
| 382 |
|
| 383 | if (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 pfn;
|
| 396 | int nid = zone_to_nid(zone);
|
| 397 |
|
| 398 | if (zone->zone_start_pfn == start_pfn) {
|
| 399 | |
| 400 | * If the section is smallest section in the zone, it need
|
| 401 | * shrink zone->zone_start_pfn and zone->zone_spanned_pages.
|
| 402 | * In this case, we find second smallest valid mem_section
|
| 403 | * for shrinking zone.
|
| 404 |
|
| 405 | pfn = find_smallest_section_pfn(nid, zone, end_pfn,
|
| 406 | zone_end_pfn(zone));
|
| 407 | if (pfn) {
|
| 408 | zone->spanned_pages = zone_end_pfn(zone) - pfn;
|
| 409 | zone->zone_start_pfn = pfn;
|
| 410 | } else {
|
| 411 | zone->zone_start_pfn = 0;
|
| 412 | zone->spanned_pages = 0;
|
| 413 | }
|
| 414 | } else if (zone_end_pfn(zone) == end_pfn) {
|
| 415 | |
| 416 | * If the section is biggest section in the zone, it need
|
| 417 | * shrink zone->spanned_pages.
|
| 418 | * In this case, we find second biggest valid mem_section for
|
| 419 | * shrinking zone.
|
| 420 |
|
| 421 | pfn = find_biggest_section_pfn(nid, zone, zone->zone_start_pfn,
|
| 422 | start_pfn);
|
| 423 | if (pfn)
|
| 424 | zone->spanned_pages = pfn - zone->zone_start_pfn + 1;
|
| 425 | else {
|
| 426 | zone->zone_start_pfn = 0;
|
| 427 | zone->spanned_pages = 0;
|
| 428 | }
|
| 429 | }
|
| 430 | }
|
| 431 |
|
| 432 | static void update_pgdat_span(struct pglist_data *pgdat)
|
| 433 | {
|
| 434 | unsigned long node_start_pfn = 0, node_end_pfn = 0;
|
| 435 | struct zone *zone;
|
| 436 |
|
| 437 | for (zone = pgdat->node_zones;
|
| 438 | zone < pgdat->node_zones + MAX_NR_ZONES; zone++) {
|
| 439 | unsigned long end_pfn = zone_end_pfn(zone);
|
| 440 |
|
| 441 |
|
| 442 | if (!zone->spanned_pages)
|
| 443 | continue;
|
| 444 | if (!node_end_pfn) {
|
| 445 | node_start_pfn = zone->zone_start_pfn;
|
| 446 | node_end_pfn = end_pfn;
|
| 447 | continue;
|
| 448 | }
|
| 449 |
|
| 450 | if (end_pfn > node_end_pfn)
|
| 451 | node_end_pfn = end_pfn;
|
| 452 | if (zone->zone_start_pfn < node_start_pfn)
|
| 453 | node_start_pfn = zone->zone_start_pfn;
|
| 454 | }
|
| 455 |
|
| 456 | pgdat->node_start_pfn = node_start_pfn;
|
| 457 | pgdat->node_spanned_pages = node_end_pfn - node_start_pfn;
|
| 458 | }
|
| 459 |
|
| 460 | void __ref remove_pfn_range_from_zone(struct zone *zone,
|
| 461 | unsigned long start_pfn,
|
| 462 | unsigned long nr_pages)
|
| 463 | {
|
| 464 | const unsigned long end_pfn = start_pfn + nr_pages;
|
| 465 | struct pglist_data *pgdat = zone->zone_pgdat;
|
| 466 | unsigned long pfn, cur_nr_pages;
|
| 467 |
|
| 468 |
|
| 469 | for (pfn = start_pfn; pfn < end_pfn; pfn += cur_nr_pages) {
|
| 470 | cond_resched();
|
| 471 |
|
| 472 |
|
| 473 | cur_nr_pages =
|
| 474 | min(end_pfn - pfn, SECTION_ALIGN_UP(pfn + 1) - pfn);
|
| 475 | page_init_poison(pfn_to_page(pfn),
|
| 476 | sizeof(struct page) * cur_nr_pages);
|
| 477 | }
|
| 478 |
|
| 479 | |
| 480 | * Zone shrinking code cannot properly deal with ZONE_DEVICE. So
|
| 481 | * we will not try to shrink the zones - which is okay as
|
| 482 | * set_zone_contiguous() cannot deal with ZONE_DEVICE either way.
|
| 483 |
|
| 484 | if (zone_is_zone_device(zone))
|
| 485 | return;
|
| 486 |
|
| 487 | clear_zone_contiguous(zone);
|
| 488 |
|
| 489 | shrink_zone_span(zone, start_pfn, start_pfn + nr_pages);
|
| 490 | update_pgdat_span(pgdat);
|
| 491 |
|
| 492 | set_zone_contiguous(zone);
|
| 493 | }
|
| 494 |
|
| 495 | static void __remove_section(unsigned long pfn, unsigned long nr_pages,
|
| 496 | unsigned long map_offset,
|
| 497 | struct vmem_altmap *altmap)
|
| 498 | {
|
| 499 | struct mem_section *ms = __pfn_to_section(pfn);
|
| 500 |
|
| 501 | if (WARN_ON_ONCE(!valid_section(ms)))
|
| 502 | return;
|
| 503 |
|
| 504 | sparse_remove_section(ms, pfn, nr_pages, map_offset, altmap);
|
| 505 | }
|
| 506 |
|
| 507 | |
| 508 | * __remove_pages() - remove sections of pages
|
| 509 | * @pfn: starting pageframe (must be aligned to start of a section)
|
| 510 | * @nr_pages: number of pages to remove (must be multiple of section size)
|
| 511 | * @altmap: alternative device page map or %NULL if default memmap is used
|
| 512 | *
|
| 513 | * Generic helper function to remove section mappings and sysfs entries
|
| 514 | * for the section of the memory we are removing. Caller needs to make
|
| 515 | * sure that pages are marked reserved and zones are adjust properly by
|
| 516 | * calling offline_pages().
|
| 517 |
|
| 518 | void __remove_pages(unsigned long pfn, unsigned long nr_pages,
|
| 519 | struct vmem_altmap *altmap)
|
| 520 | {
|
| 521 | const unsigned long end_pfn = pfn + nr_pages;
|
| 522 | unsigned long cur_nr_pages;
|
| 523 | unsigned long map_offset = 0;
|
| 524 |
|
| 525 | map_offset = vmem_altmap_offset(altmap);
|
| 526 |
|
| 527 | if (check_pfn_span(pfn, nr_pages)) {
|
| 528 | WARN(1, "Misaligned %s start: %#lx end: #%lx\n", __func__, pfn, pfn + nr_pages - 1);
|
| 529 | return;
|
| 530 | }
|
| 531 |
|
| 532 | for (; pfn < end_pfn; pfn += cur_nr_pages) {
|
| 533 | cond_resched();
|
| 534 |
|
| 535 | cur_nr_pages = min(end_pfn - pfn,
|
| 536 | SECTION_ALIGN_UP(pfn + 1) - pfn);
|
| 537 | __remove_section(pfn, cur_nr_pages, map_offset, altmap);
|
| 538 | map_offset = 0;
|
| 539 | }
|
| 540 | }
|
| 541 |
|
| 542 | int set_online_page_callback(online_page_callback_t callback)
|
| 543 | {
|
| 544 | int rc = -EINVAL;
|
| 545 |
|
| 546 | get_online_mems();
|
| 547 | mutex_lock(&online_page_callback_lock);
|
| 548 |
|
| 549 | if (online_page_callback == generic_online_page) {
|
| 550 | online_page_callback = callback;
|
| 551 | rc = 0;
|
| 552 | }
|
| 553 |
|
| 554 | mutex_unlock(&online_page_callback_lock);
|
| 555 | put_online_mems();
|
| 556 |
|
| 557 | return rc;
|
| 558 | }
|
| 559 | EXPORT_SYMBOL_GPL(set_online_page_callback);
|
| 560 |
|
| 561 | int restore_online_page_callback(online_page_callback_t callback)
|
| 562 | {
|
| 563 | int rc = -EINVAL;
|
| 564 |
|
| 565 | get_online_mems();
|
| 566 | mutex_lock(&online_page_callback_lock);
|
| 567 |
|
| 568 | if (online_page_callback == callback) {
|
| 569 | online_page_callback = generic_online_page;
|
| 570 | rc = 0;
|
| 571 | }
|
| 572 |
|
| 573 | mutex_unlock(&online_page_callback_lock);
|
| 574 | put_online_mems();
|
| 575 |
|
| 576 | return rc;
|
| 577 | }
|
| 578 | EXPORT_SYMBOL_GPL(restore_online_page_callback);
|
| 579 |
|
| 580 | void generic_online_page(struct page *page, unsigned int order)
|
| 581 | {
|
| 582 | |
| 583 | * Freeing the page with debug_pagealloc enabled will try to unmap it,
|
| 584 | * so we should map it first. This is better than introducing a special
|
| 585 | * case in page freeing fast path.
|
| 586 |
|
| 587 | debug_pagealloc_map_pages(page, 1 << order);
|
| 588 | __free_pages_core(page, order);
|
| 589 | totalram_pages_add(1UL << order);
|
| 590 | }
|
| 591 | EXPORT_SYMBOL_GPL(generic_online_page);
|
| 592 |
|
| 593 | static void online_pages_range(unsigned long start_pfn, unsigned long nr_pages)
|
| 594 | {
|
| 595 | const unsigned long end_pfn = start_pfn + nr_pages;
|
| 596 | unsigned long pfn;
|
| 597 |
|
| 598 | |
| 599 | * Online the pages in MAX_ORDER aligned chunks. The callback might
|
| 600 | * decide to not expose all pages to the buddy (e.g., expose them
|
| 601 | * later). We account all pages as being online and belonging to this
|
| 602 | * zone ("present").
|
| 603 | * When using memmap_on_memory, the range might not be aligned to
|
| 604 | * MAX_ORDER_NR_PAGES - 1, but pageblock aligned. __ffs() will detect
|
| 605 | * this and the first chunk to online will be pageblock_nr_pages.
|
| 606 |
|
| 607 | for (pfn = start_pfn; pfn < end_pfn;) {
|
| 608 | int order;
|
| 609 |
|
| 610 | |
| 611 | * Free to online pages in the largest chunks alignment allows.
|
| 612 | *
|
| 613 | * __ffs() behaviour is undefined for 0. start == 0 is
|
| 614 | * MAX_ORDER-aligned, Set order to MAX_ORDER for the case.
|
| 615 |
|
| 616 | if (pfn)
|
| 617 | order = min_t(int, MAX_ORDER, __ffs(pfn));
|
| 618 | else
|
| 619 | order = MAX_ORDER;
|
| 620 |
|
| 621 | (*online_page_callback)(pfn_to_page(pfn), order);
|
| 622 | pfn += (1UL << order);
|
| 623 | }
|
| 624 |
|
| 625 |
|
| 626 | online_mem_sections(start_pfn, end_pfn);
|
| 627 | }
|
| 628 |
|
| 629 |
|
| 630 | static void node_states_check_changes_online(unsigned long nr_pages,
|
| 631 | struct zone *zone, struct memory_notify *arg)
|
| 632 | {
|
| 633 | int nid = zone_to_nid(zone);
|
| 634 |
|
| 635 | arg->status_change_nid = NUMA_NO_NODE;
|
| 636 | arg->status_change_nid_normal = NUMA_NO_NODE;
|
| 637 |
|
| 638 | if (!node_state(nid, N_MEMORY))
|
| 639 | arg->status_change_nid = nid;
|
| 640 | if (zone_idx(zone) <= ZONE_NORMAL && !node_state(nid, N_NORMAL_MEMORY))
|
| 641 | arg->status_change_nid_normal = nid;
|
| 642 | }
|
| 643 |
|
| 644 | static void node_states_set_node(int node, struct memory_notify *arg)
|
| 645 | {
|
| 646 | if (arg->status_change_nid_normal >= 0)
|
| 647 | node_set_state(node, N_NORMAL_MEMORY);
|
| 648 |
|
| 649 | if (arg->status_change_nid >= 0)
|
| 650 | node_set_state(node, N_MEMORY);
|
| 651 | }
|
| 652 |
|
| 653 | static void __meminit resize_zone_range(struct zone *zone, unsigned long start_pfn,
|
| 654 | unsigned long nr_pages)
|
| 655 | {
|
| 656 | unsigned long old_end_pfn = zone_end_pfn(zone);
|
| 657 |
|
| 658 | if (zone_is_empty(zone) || start_pfn < zone->zone_start_pfn)
|
| 659 | zone->zone_start_pfn = start_pfn;
|
| 660 |
|
| 661 | zone->spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - zone->zone_start_pfn;
|
| 662 | }
|
| 663 |
|
| 664 | static void __meminit resize_pgdat_range(struct pglist_data *pgdat, unsigned long start_pfn,
|
| 665 | unsigned long nr_pages)
|
| 666 | {
|
| 667 | unsigned long old_end_pfn = pgdat_end_pfn(pgdat);
|
| 668 |
|
| 669 | if (!pgdat->node_spanned_pages || start_pfn < pgdat->node_start_pfn)
|
| 670 | pgdat->node_start_pfn = start_pfn;
|
| 671 |
|
| 672 | pgdat->node_spanned_pages = max(start_pfn + nr_pages, old_end_pfn) - pgdat->node_start_pfn;
|
| 673 |
|
| 674 | }
|
| 675 |
|
| 676 | #ifdef CONFIG_ZONE_DEVICE
|
| 677 | static void section_taint_zone_device(unsigned long pfn)
|
| 678 | {
|
| 679 | struct mem_section *ms = __pfn_to_section(pfn);
|
| 680 |
|
| 681 | ms->section_mem_map |= SECTION_TAINT_ZONE_DEVICE;
|
| 682 | }
|
| 683 | #else
|
| 684 | static inline void section_taint_zone_device(unsigned long pfn)
|
| 685 | {
|
| 686 | }
|
| 687 | #endif
|
| 688 |
|
| 689 | |
| 690 | * Associate the pfn range with the given zone, initializing the memmaps
|
| 691 | * and resizing the pgdat/zone data to span the added pages. After this
|
| 692 | * call, all affected pages are PG_reserved.
|
| 693 | *
|
| 694 | * All aligned pageblocks are initialized to the specified migratetype
|
| 695 | * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
|
| 696 | * zone stats (e.g., nr_isolate_pageblock) are touched.
|
| 697 |
|
| 698 | void __ref move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
|
| 699 | unsigned long nr_pages,
|
| 700 | struct vmem_altmap *altmap, int migratetype)
|
| 701 | {
|
| 702 | struct pglist_data *pgdat = zone->zone_pgdat;
|
| 703 | int nid = pgdat->node_id;
|
| 704 |
|
| 705 | clear_zone_contiguous(zone);
|
| 706 |
|
| 707 | if (zone_is_empty(zone))
|
| 708 | init_currently_empty_zone(zone, start_pfn, nr_pages);
|
| 709 | resize_zone_range(zone, start_pfn, nr_pages);
|
| 710 | resize_pgdat_range(pgdat, start_pfn, nr_pages);
|
| 711 |
|
| 712 | |
| 713 | * Subsection population requires care in pfn_to_online_page().
|
| 714 | * Set the taint to enable the slow path detection of
|
| 715 | * ZONE_DEVICE pages in an otherwise ZONE_{NORMAL,MOVABLE}
|
| 716 | * section.
|
| 717 |
|
| 718 | if (zone_is_zone_device(zone)) {
|
| 719 | if (!IS_ALIGNED(start_pfn, PAGES_PER_SECTION))
|
| 720 | section_taint_zone_device(start_pfn);
|
| 721 | if (!IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION))
|
| 722 | section_taint_zone_device(start_pfn + nr_pages);
|
| 723 | }
|
| 724 |
|
| 725 | |
| 726 | * TODO now we have a visible range of pages which are not associated
|
| 727 | * with their zone properly. Not nice but set_pfnblock_flags_mask
|
| 728 | * expects the zone spans the pfn range. All the pages in the range
|
| 729 | * are reserved so nobody should be touching them so we should be safe
|
| 730 |
|
| 731 | memmap_init_range(nr_pages, nid, zone_idx(zone), start_pfn, 0,
|
| 732 | MEMINIT_HOTPLUG, altmap, migratetype);
|
| 733 |
|
| 734 | set_zone_contiguous(zone);
|
| 735 | }
|
| 736 |
|
| 737 | struct auto_movable_stats {
|
| 738 | unsigned long kernel_early_pages;
|
| 739 | unsigned long movable_pages;
|
| 740 | };
|
| 741 |
|
| 742 | static void auto_movable_stats_account_zone(struct auto_movable_stats *stats,
|
| 743 | struct zone *zone)
|
| 744 | {
|
| 745 | if (zone_idx(zone) == ZONE_MOVABLE) {
|
| 746 | stats->movable_pages += zone->present_pages;
|
| 747 | } else {
|
| 748 | stats->kernel_early_pages += zone->present_early_pages;
|
| 749 | #ifdef CONFIG_CMA
|
| 750 | |
| 751 | * CMA pages (never on hotplugged memory) behave like
|
| 752 | * ZONE_MOVABLE.
|
| 753 |
|
| 754 | stats->movable_pages += zone->cma_pages;
|
| 755 | stats->kernel_early_pages -= zone->cma_pages;
|
| 756 | #endif
|
| 757 | }
|
| 758 | }
|
| 759 | struct auto_movable_group_stats {
|
| 760 | unsigned long movable_pages;
|
| 761 | unsigned long req_kernel_early_pages;
|
| 762 | };
|
| 763 |
|
| 764 | static int auto_movable_stats_account_group(struct memory_group *group,
|
| 765 | void *arg)
|
| 766 | {
|
| 767 | const int ratio = READ_ONCE(auto_movable_ratio);
|
| 768 | struct auto_movable_group_stats *stats = arg;
|
| 769 | long pages;
|
| 770 |
|
| 771 | |
| 772 | * We don't support modifying the config while the auto-movable online
|
| 773 | * policy is already enabled. Just avoid the division by zero below.
|
| 774 |
|
| 775 | if (!ratio)
|
| 776 | return 0;
|
| 777 |
|
| 778 | |
| 779 | * Calculate how many early kernel pages this group requires to
|
| 780 | * satisfy the configured zone ratio.
|
| 781 |
|
| 782 | pages = group->present_movable_pages * 100 / ratio;
|
| 783 | pages -= group->present_kernel_pages;
|
| 784 |
|
| 785 | if (pages > 0)
|
| 786 | stats->req_kernel_early_pages += pages;
|
| 787 | stats->movable_pages += group->present_movable_pages;
|
| 788 | return 0;
|
| 789 | }
|
| 790 |
|
| 791 | static bool auto_movable_can_online_movable(int nid, struct memory_group *group,
|
| 792 | unsigned long nr_pages)
|
| 793 | {
|
| 794 | unsigned long kernel_early_pages, movable_pages;
|
| 795 | struct auto_movable_group_stats group_stats = {};
|
| 796 | struct auto_movable_stats stats = {};
|
| 797 | pg_data_t *pgdat = NODE_DATA(nid);
|
| 798 | struct zone *zone;
|
| 799 | int i;
|
| 800 |
|
| 801 |
|
| 802 | if (nid == NUMA_NO_NODE) {
|
| 803 |
|
| 804 | for_each_populated_zone(zone)
|
| 805 | auto_movable_stats_account_zone(&stats, zone);
|
| 806 | } else {
|
| 807 | for (i = 0; i < MAX_NR_ZONES; i++) {
|
| 808 | zone = pgdat->node_zones + i;
|
| 809 | if (populated_zone(zone))
|
| 810 | auto_movable_stats_account_zone(&stats, zone);
|
| 811 | }
|
| 812 | }
|
| 813 |
|
| 814 | kernel_early_pages = stats.kernel_early_pages;
|
| 815 | movable_pages = stats.movable_pages;
|
| 816 |
|
| 817 | |
| 818 | * Kernel memory inside dynamic memory group allows for more MOVABLE
|
| 819 | * memory within the same group. Remove the effect of all but the
|
| 820 | * current group from the stats.
|
| 821 |
|
| 822 | walk_dynamic_memory_groups(nid, auto_movable_stats_account_group,
|
| 823 | group, &group_stats);
|
| 824 | if (kernel_early_pages <= group_stats.req_kernel_early_pages)
|
| 825 | return false;
|
| 826 | kernel_early_pages -= group_stats.req_kernel_early_pages;
|
| 827 | movable_pages -= group_stats.movable_pages;
|
| 828 |
|
| 829 | if (group && group->is_dynamic)
|
| 830 | kernel_early_pages += group->present_kernel_pages;
|
| 831 |
|
| 832 | |
| 833 | * Test if we could online the given number of pages to ZONE_MOVABLE
|
| 834 | * and still stay in the configured ratio.
|
| 835 |
|
| 836 | movable_pages += nr_pages;
|
| 837 | return movable_pages <= (auto_movable_ratio * kernel_early_pages) / 100;
|
| 838 | }
|
| 839 |
|
| 840 | |
| 841 | * Returns a default kernel memory zone for the given pfn range.
|
| 842 | * If no kernel zone covers this pfn range it will automatically go
|
| 843 | * to the ZONE_NORMAL.
|
| 844 |
|
| 845 | static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn,
|
| 846 | unsigned long nr_pages)
|
| 847 | {
|
| 848 | struct pglist_data *pgdat = NODE_DATA(nid);
|
| 849 | int zid;
|
| 850 |
|
| 851 | for (zid = 0; zid < ZONE_NORMAL; zid++) {
|
| 852 | struct zone *zone = &pgdat->node_zones[zid];
|
| 853 |
|
| 854 | if (zone_intersects(zone, start_pfn, nr_pages))
|
| 855 | return zone;
|
| 856 | }
|
| 857 |
|
| 858 | return &pgdat->node_zones[ZONE_NORMAL];
|
| 859 | }
|
| 860 |
|
| 861 | |
| 862 | * Determine to which zone to online memory dynamically based on user
|
| 863 | * configuration and system stats. We care about the following ratio:
|
| 864 | *
|
| 865 | * MOVABLE : KERNEL
|
| 866 | *
|
| 867 | * Whereby MOVABLE is memory in ZONE_MOVABLE and KERNEL is memory in
|
| 868 | * one of the kernel zones. CMA pages inside one of the kernel zones really
|
| 869 | * behaves like ZONE_MOVABLE, so we treat them accordingly.
|
| 870 | *
|
| 871 | * We don't allow for hotplugged memory in a KERNEL zone to increase the
|
| 872 | * amount of MOVABLE memory we can have, so we end up with:
|
| 873 | *
|
| 874 | * MOVABLE : KERNEL_EARLY
|
| 875 | *
|
| 876 | * Whereby KERNEL_EARLY is memory in one of the kernel zones, available sinze
|
| 877 | * boot. We base our calculation on KERNEL_EARLY internally, because:
|
| 878 | *
|
| 879 | * a) Hotplugged memory in one of the kernel zones can sometimes still get
|
| 880 | * hotunplugged, especially when hot(un)plugging individual memory blocks.
|
| 881 | * There is no coordination across memory devices, therefore "automatic"
|
| 882 | * hotunplugging, as implemented in hypervisors, could result in zone
|
| 883 | * imbalances.
|
| 884 | * b) Early/boot memory in one of the kernel zones can usually not get
|
| 885 | * hotunplugged again (e.g., no firmware interface to unplug, fragmented
|
| 886 | * with unmovable allocations). While there are corner cases where it might
|
| 887 | * still work, it is barely relevant in practice.
|
| 888 | *
|
| 889 | * Exceptions are dynamic memory groups, which allow for more MOVABLE
|
| 890 | * memory within the same memory group -- because in that case, there is
|
| 891 | * coordination within the single memory device managed by a single driver.
|
| 892 | *
|
| 893 | * We rely on "present pages" instead of "managed pages", as the latter is
|
| 894 | * highly unreliable and dynamic in virtualized environments, and does not
|
| 895 | * consider boot time allocations. For example, memory ballooning adjusts the
|
| 896 | * managed pages when inflating/deflating the balloon, and balloon compaction
|
| 897 | * can even migrate inflated pages between zones.
|
| 898 | *
|
| 899 | * Using "present pages" is better but some things to keep in mind are:
|
| 900 | *
|
| 901 | * a) Some memblock allocations, such as for the crashkernel area, are
|
| 902 | * effectively unused by the kernel, yet they account to "present pages".
|
| 903 | * Fortunately, these allocations are comparatively small in relevant setups
|
| 904 | * (e.g., fraction of system memory).
|
| 905 | * b) Some hotplugged memory blocks in virtualized environments, esecially
|
| 906 | * hotplugged by virtio-mem, look like they are completely present, however,
|
| 907 | * only parts of the memory block are actually currently usable.
|
| 908 | * "present pages" is an upper limit that can get reached at runtime. As
|
| 909 | * we base our calculations on KERNEL_EARLY, this is not an issue.
|
| 910 |
|
| 911 | static struct zone *auto_movable_zone_for_pfn(int nid,
|
| 912 | struct memory_group *group,
|
| 913 | unsigned long pfn,
|
| 914 | unsigned long nr_pages)
|
| 915 | {
|
| 916 | unsigned long online_pages = 0, max_pages, end_pfn;
|
| 917 | struct page *page;
|
| 918 |
|
| 919 | if (!auto_movable_ratio)
|
| 920 | goto kernel_zone;
|
| 921 |
|
| 922 | if (group && !group->is_dynamic) {
|
| 923 | max_pages = group->s.max_pages;
|
| 924 | online_pages = group->present_movable_pages;
|
| 925 |
|
| 926 |
|
| 927 | if (group->present_kernel_pages)
|
| 928 | goto kernel_zone;
|
| 929 | } else if (!group || group->d.unit_pages == nr_pages) {
|
| 930 | max_pages = nr_pages;
|
| 931 | } else {
|
| 932 | max_pages = group->d.unit_pages;
|
| 933 | |
| 934 | * Take a look at all online sections in the current unit.
|
| 935 | * We can safely assume that all pages within a section belong
|
| 936 | * to the same zone, because dynamic memory groups only deal
|
| 937 | * with hotplugged memory.
|
| 938 |
|
| 939 | pfn = ALIGN_DOWN(pfn, group->d.unit_pages);
|
| 940 | end_pfn = pfn + group->d.unit_pages;
|
| 941 | for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
|
| 942 | page = pfn_to_online_page(pfn);
|
| 943 | if (!page)
|
| 944 | continue;
|
| 945 |
|
| 946 | if (!is_zone_movable_page(page))
|
| 947 | goto kernel_zone;
|
| 948 | online_pages += PAGES_PER_SECTION;
|
| 949 | }
|
| 950 | }
|
| 951 |
|
| 952 | |
| 953 | * Online MOVABLE if we could *currently* online all remaining parts
|
| 954 | * MOVABLE. We expect to (add+) online them immediately next, so if
|
| 955 | * nobody interferes, all will be MOVABLE if possible.
|
| 956 |
|
| 957 | nr_pages = max_pages - online_pages;
|
| 958 | if (!auto_movable_can_online_movable(NUMA_NO_NODE, group, nr_pages))
|
| 959 | goto kernel_zone;
|
| 960 |
|
| 961 | #ifdef CONFIG_NUMA
|
| 962 | if (auto_movable_numa_aware &&
|
| 963 | !auto_movable_can_online_movable(nid, group, nr_pages))
|
| 964 | goto kernel_zone;
|
| 965 | #endif
|
| 966 |
|
| 967 | return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
|
| 968 | kernel_zone:
|
| 969 | return default_kernel_zone_for_pfn(nid, pfn, nr_pages);
|
| 970 | }
|
| 971 |
|
| 972 | static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
|
| 973 | unsigned long nr_pages)
|
| 974 | {
|
| 975 | struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
|
| 976 | nr_pages);
|
| 977 | struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
|
| 978 | bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
|
| 979 | bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
|
| 980 |
|
| 981 | |
| 982 | * We inherit the existing zone in a simple case where zones do not
|
| 983 | * overlap in the given range
|
| 984 |
|
| 985 | if (in_kernel ^ in_movable)
|
| 986 | return (in_kernel) ? kernel_zone : movable_zone;
|
| 987 |
|
| 988 | |
| 989 | * If the range doesn't belong to any zone or two zones overlap in the
|
| 990 | * given range then we use movable zone only if movable_node is
|
| 991 | * enabled because we always online to a kernel zone by default.
|
| 992 |
|
| 993 | return movable_node_enabled ? movable_zone : kernel_zone;
|
| 994 | }
|
| 995 |
|
| 996 | struct zone *zone_for_pfn_range(int online_type, int nid,
|
| 997 | struct memory_group *group, unsigned long start_pfn,
|
| 998 | unsigned long nr_pages)
|
| 999 | {
|
| 1000 | if (online_type == MMOP_ONLINE_KERNEL)
|
| 1001 | return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
|
| 1002 |
|
| 1003 | if (online_type == MMOP_ONLINE_MOVABLE)
|
| 1004 | return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
|
| 1005 |
|
| 1006 | if (online_policy == ONLINE_POLICY_AUTO_MOVABLE)
|
| 1007 | return auto_movable_zone_for_pfn(nid, group, start_pfn, nr_pages);
|
| 1008 |
|
| 1009 | return default_zone_for_pfn(nid, start_pfn, nr_pages);
|
| 1010 | }
|
| 1011 |
|
| 1012 | |
| 1013 | * This function should only be called by memory_block_{online,offline},
|
| 1014 | * and {online,offline}_pages.
|
| 1015 |
|
| 1016 | void adjust_present_page_count(struct page *page, struct memory_group *group,
|
| 1017 | long nr_pages)
|
| 1018 | {
|
| 1019 | struct zone *zone = page_zone(page);
|
| 1020 | const bool movable = zone_idx(zone) == ZONE_MOVABLE;
|
| 1021 |
|
| 1022 | |
| 1023 | * We only support onlining/offlining/adding/removing of complete
|
| 1024 | * memory blocks; therefore, either all is either early or hotplugged.
|
| 1025 |
|
| 1026 | if (early_section(__pfn_to_section(page_to_pfn(page))))
|
| 1027 | zone->present_early_pages += nr_pages;
|
| 1028 | zone->present_pages += nr_pages;
|
| 1029 | zone->zone_pgdat->node_present_pages += nr_pages;
|
| 1030 |
|
| 1031 | if (group && movable)
|
| 1032 | group->present_movable_pages += nr_pages;
|
| 1033 | else if (group && !movable)
|
| 1034 | group->present_kernel_pages += nr_pages;
|
| 1035 | }
|
| 1036 |
|
| 1037 | int mhp_init_memmap_on_memory(unsigned long pfn, unsigned long nr_pages,
|
| 1038 | struct zone *zone)
|
| 1039 | {
|
| 1040 | unsigned long end_pfn = pfn + nr_pages;
|
| 1041 | int ret, i;
|
| 1042 |
|
| 1043 | ret = kasan_add_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
|
| 1044 | if (ret)
|
| 1045 | return ret;
|
| 1046 |
|
| 1047 | move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_UNMOVABLE);
|
| 1048 |
|
| 1049 | for (i = 0; i < nr_pages; i++)
|
| 1050 | SetPageVmemmapSelfHosted(pfn_to_page(pfn + i));
|
| 1051 |
|
| 1052 | |
| 1053 | * It might be that the vmemmap_pages fully span sections. If that is
|
| 1054 | * the case, mark those sections online here as otherwise they will be
|
| 1055 | * left offline.
|
| 1056 |
|
| 1057 | if (nr_pages >= PAGES_PER_SECTION)
|
| 1058 | online_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
|
| 1059 |
|
| 1060 | return ret;
|
| 1061 | }
|
| 1062 |
|
| 1063 | void mhp_deinit_memmap_on_memory(unsigned long pfn, unsigned long nr_pages)
|
| 1064 | {
|
| 1065 | unsigned long end_pfn = pfn + nr_pages;
|
| 1066 |
|
| 1067 | |
| 1068 | * It might be that the vmemmap_pages fully span sections. If that is
|
| 1069 | * the case, mark those sections offline here as otherwise they will be
|
| 1070 | * left online.
|
| 1071 |
|
| 1072 | if (nr_pages >= PAGES_PER_SECTION)
|
| 1073 | offline_mem_sections(pfn, ALIGN_DOWN(end_pfn, PAGES_PER_SECTION));
|
| 1074 |
|
| 1075 | |
| 1076 | * The pages associated with this vmemmap have been offlined, so
|
| 1077 | * we can reset its state here.
|
| 1078 |
|
| 1079 | remove_pfn_range_from_zone(page_zone(pfn_to_page(pfn)), pfn, nr_pages);
|
| 1080 | kasan_remove_zero_shadow(__va(PFN_PHYS(pfn)), PFN_PHYS(nr_pages));
|
| 1081 | }
|
| 1082 |
|
| 1083 | int __ref online_pages(unsigned long pfn, unsigned long nr_pages,
|
| 1084 | struct zone *zone, struct memory_group *group)
|
| 1085 | {
|
| 1086 | unsigned long flags;
|
| 1087 | int need_zonelists_rebuild = 0;
|
| 1088 | const int nid = zone_to_nid(zone);
|
| 1089 | int ret;
|
| 1090 | struct memory_notify arg;
|
| 1091 |
|
| 1092 | |
| 1093 | * {on,off}lining is constrained to full memory sections (or more
|
| 1094 | * precisely to memory blocks from the user space POV).
|
| 1095 | * memmap_on_memory is an exception because it reserves initial part
|
| 1096 | * of the physical memory space for vmemmaps. That space is pageblock
|
| 1097 | * aligned.
|
| 1098 |
|
| 1099 | if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(pfn) ||
|
| 1100 | !IS_ALIGNED(pfn + nr_pages, PAGES_PER_SECTION)))
|
| 1101 | return -EINVAL;
|
| 1102 |
|
| 1103 | mem_hotplug_begin();
|
| 1104 |
|
| 1105 |
|
| 1106 | move_pfn_range_to_zone(zone, pfn, nr_pages, NULL, MIGRATE_ISOLATE);
|
| 1107 |
|
| 1108 | arg.start_pfn = pfn;
|
| 1109 | arg.nr_pages = nr_pages;
|
| 1110 | node_states_check_changes_online(nr_pages, zone, &arg);
|
| 1111 |
|
| 1112 | ret = memory_notify(MEM_GOING_ONLINE, &arg);
|
| 1113 | ret = notifier_to_errno(ret);
|
| 1114 | if (ret)
|
| 1115 | goto failed_addition;
|
| 1116 |
|
| 1117 | |
| 1118 | * Fixup the number of isolated pageblocks before marking the sections
|
| 1119 | * onlining, such that undo_isolate_page_range() works correctly.
|
| 1120 |
|
| 1121 | spin_lock_irqsave(&zone->lock, flags);
|
| 1122 | zone->nr_isolate_pageblock += nr_pages / pageblock_nr_pages;
|
| 1123 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 1124 |
|
| 1125 | |
| 1126 | * If this zone is not populated, then it is not in zonelist.
|
| 1127 | * This means the page allocator ignores this zone.
|
| 1128 | * So, zonelist must be updated after online.
|
| 1129 |
|
| 1130 | if (!populated_zone(zone)) {
|
| 1131 | need_zonelists_rebuild = 1;
|
| 1132 | setup_zone_pageset(zone);
|
| 1133 | }
|
| 1134 |
|
| 1135 | online_pages_range(pfn, nr_pages);
|
| 1136 | adjust_present_page_count(pfn_to_page(pfn), group, nr_pages);
|
| 1137 |
|
| 1138 | node_states_set_node(nid, &arg);
|
| 1139 | if (need_zonelists_rebuild)
|
| 1140 | build_all_zonelists(NULL);
|
| 1141 |
|
| 1142 |
|
| 1143 | undo_isolate_page_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE);
|
| 1144 |
|
| 1145 | |
| 1146 | * Freshly onlined pages aren't shuffled (e.g., all pages are placed to
|
| 1147 | * the tail of the freelist when undoing isolation). Shuffle the whole
|
| 1148 | * zone to make sure the just onlined pages are properly distributed
|
| 1149 | * across the whole freelist - to create an initial shuffle.
|
| 1150 |
|
| 1151 | shuffle_zone(zone);
|
| 1152 |
|
| 1153 |
|
| 1154 | init_per_zone_wmark_min();
|
| 1155 |
|
| 1156 | kswapd_run(nid);
|
| 1157 | kcompactd_run(nid);
|
| 1158 |
|
| 1159 | writeback_set_ratelimit();
|
| 1160 |
|
| 1161 | memory_notify(MEM_ONLINE, &arg);
|
| 1162 | mem_hotplug_done();
|
| 1163 | return 0;
|
| 1164 |
|
| 1165 | failed_addition:
|
| 1166 | pr_debug("online_pages [mem %#010llx-%#010llx] failed\n",
|
| 1167 | (unsigned long long) pfn << PAGE_SHIFT,
|
| 1168 | (((unsigned long long) pfn + nr_pages) << PAGE_SHIFT) - 1);
|
| 1169 | memory_notify(MEM_CANCEL_ONLINE, &arg);
|
| 1170 | remove_pfn_range_from_zone(zone, pfn, nr_pages);
|
| 1171 | mem_hotplug_done();
|
| 1172 | return ret;
|
| 1173 | }
|
| 1174 |
|
| 1175 | static void reset_node_present_pages(pg_data_t *pgdat)
|
| 1176 | {
|
| 1177 | struct zone *z;
|
| 1178 |
|
| 1179 | for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
|
| 1180 | z->present_pages = 0;
|
| 1181 |
|
| 1182 | pgdat->node_present_pages = 0;
|
| 1183 | }
|
| 1184 |
|
| 1185 |
|
| 1186 | static pg_data_t __ref *hotadd_init_pgdat(int nid)
|
| 1187 | {
|
| 1188 | struct pglist_data *pgdat;
|
| 1189 |
|
| 1190 | |
| 1191 | * NODE_DATA is preallocated (free_area_init) but its internal
|
| 1192 | * state is not allocated completely. Add missing pieces.
|
| 1193 | * Completely offline nodes stay around and they just need
|
| 1194 | * reintialization.
|
| 1195 |
|
| 1196 | pgdat = NODE_DATA(nid);
|
| 1197 |
|
| 1198 |
|
| 1199 | free_area_init_core_hotplug(pgdat);
|
| 1200 |
|
| 1201 | |
| 1202 | * The node we allocated has no zone fallback lists. For avoiding
|
| 1203 | * to access not-initialized zonelist, build here.
|
| 1204 |
|
| 1205 | build_all_zonelists(pgdat);
|
| 1206 |
|
| 1207 | |
| 1208 | * When memory is hot-added, all the memory is in offline state. So
|
| 1209 | * clear all zones' present_pages because they will be updated in
|
| 1210 | * online_pages() and offline_pages().
|
| 1211 | * TODO: should be in free_area_init_core_hotplug?
|
| 1212 |
|
| 1213 | reset_node_managed_pages(pgdat);
|
| 1214 | reset_node_present_pages(pgdat);
|
| 1215 |
|
| 1216 | return pgdat;
|
| 1217 | }
|
| 1218 |
|
| 1219 | |
| 1220 | * __try_online_node - online a node if offlined
|
| 1221 | * @nid: the node ID
|
| 1222 | * @set_node_online: Whether we want to online the node
|
| 1223 | * called by cpu_up() to online a node without onlined memory.
|
| 1224 | *
|
| 1225 | * Returns:
|
| 1226 | * 1 -> a new node has been allocated
|
| 1227 | * 0 -> the node is already online
|
| 1228 | * -ENOMEM -> the node could not be allocated
|
| 1229 |
|
| 1230 | static int __try_online_node(int nid, bool set_node_online)
|
| 1231 | {
|
| 1232 | pg_data_t *pgdat;
|
| 1233 | int ret = 1;
|
| 1234 |
|
| 1235 | if (node_online(nid))
|
| 1236 | return 0;
|
| 1237 |
|
| 1238 | pgdat = hotadd_init_pgdat(nid);
|
| 1239 | if (!pgdat) {
|
| 1240 | pr_err("Cannot online node %d due to NULL pgdat\n", nid);
|
| 1241 | ret = -ENOMEM;
|
| 1242 | goto out;
|
| 1243 | }
|
| 1244 |
|
| 1245 | if (set_node_online) {
|
| 1246 | node_set_online(nid);
|
| 1247 | ret = register_one_node(nid);
|
| 1248 | BUG_ON(ret);
|
| 1249 | }
|
| 1250 | out:
|
| 1251 | return ret;
|
| 1252 | }
|
| 1253 |
|
| 1254 | |
| 1255 | * Users of this function always want to online/register the node
|
| 1256 |
|
| 1257 | int try_online_node(int nid)
|
| 1258 | {
|
| 1259 | int ret;
|
| 1260 |
|
| 1261 | mem_hotplug_begin();
|
| 1262 | ret = __try_online_node(nid, true);
|
| 1263 | mem_hotplug_done();
|
| 1264 | return ret;
|
| 1265 | }
|
| 1266 |
|
| 1267 | static int check_hotplug_memory_range(u64 start, u64 size)
|
| 1268 | {
|
| 1269 |
|
| 1270 | if (!size || !IS_ALIGNED(start, memory_block_size_bytes()) ||
|
| 1271 | !IS_ALIGNED(size, memory_block_size_bytes())) {
|
| 1272 | pr_err("Block size [%#lx] unaligned hotplug range: start %#llx, size %#llx",
|
| 1273 | memory_block_size_bytes(), start, size);
|
| 1274 | return -EINVAL;
|
| 1275 | }
|
| 1276 |
|
| 1277 | return 0;
|
| 1278 | }
|
| 1279 |
|
| 1280 | static int online_memory_block(struct memory_block *mem, void *arg)
|
| 1281 | {
|
| 1282 | mem->online_type = mhp_default_online_type;
|
| 1283 | return device_online(&mem->dev);
|
| 1284 | }
|
| 1285 |
|
| 1286 | bool mhp_supports_memmap_on_memory(unsigned long size)
|
| 1287 | {
|
| 1288 | unsigned long nr_vmemmap_pages = size / PAGE_SIZE;
|
| 1289 | unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
|
| 1290 | unsigned long remaining_size = size - vmemmap_size;
|
| 1291 |
|
| 1292 | |
| 1293 | * Besides having arch support and the feature enabled at runtime, we
|
| 1294 | * need a few more assumptions to hold true:
|
| 1295 | *
|
| 1296 | * a) We span a single memory block: memory onlining/offlinin;g happens
|
| 1297 | * in memory block granularity. We don't want the vmemmap of online
|
| 1298 | * memory blocks to reside on offline memory blocks. In the future,
|
| 1299 | * we might want to support variable-sized memory blocks to make the
|
| 1300 | * feature more versatile.
|
| 1301 | *
|
| 1302 | * b) The vmemmap pages span complete PMDs: We don't want vmemmap code
|
| 1303 | * to populate memory from the altmap for unrelated parts (i.e.,
|
| 1304 | * other memory blocks)
|
| 1305 | *
|
| 1306 | * c) The vmemmap pages (and thereby the pages that will be exposed to
|
| 1307 | * the buddy) have to cover full pageblocks: memory onlining/offlining
|
| 1308 | * code requires applicable ranges to be page-aligned, for example, to
|
| 1309 | * set the migratetypes properly.
|
| 1310 | *
|
| 1311 | * TODO: Although we have a check here to make sure that vmemmap pages
|
| 1312 | * fully populate a PMD, it is not the right place to check for
|
| 1313 | * this. A much better solution involves improving vmemmap code
|
| 1314 | * to fallback to base pages when trying to populate vmemmap using
|
| 1315 | * altmap as an alternative source of memory, and we do not exactly
|
| 1316 | * populate a single PMD.
|
| 1317 |
|
| 1318 | return mhp_memmap_on_memory() &&
|
| 1319 | size == memory_block_size_bytes() &&
|
| 1320 | IS_ALIGNED(vmemmap_size, PMD_SIZE) &&
|
| 1321 | IS_ALIGNED(remaining_size, (pageblock_nr_pages << PAGE_SHIFT));
|
| 1322 | }
|
| 1323 |
|
| 1324 | |
| 1325 | * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
|
| 1326 | * and online/offline operations (triggered e.g. by sysfs).
|
| 1327 | *
|
| 1328 | * we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG
|
| 1329 |
|
| 1330 | int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
|
| 1331 | {
|
| 1332 | struct mhp_params params = { .pgprot = pgprot_mhp(PAGE_KERNEL) };
|
| 1333 | enum memblock_flags memblock_flags = MEMBLOCK_NONE;
|
| 1334 | struct vmem_altmap mhp_altmap = {};
|
| 1335 | struct memory_group *group = NULL;
|
| 1336 | u64 start, size;
|
| 1337 | bool new_node = false;
|
| 1338 | int ret;
|
| 1339 |
|
| 1340 | start = res->start;
|
| 1341 | size = resource_size(res);
|
| 1342 |
|
| 1343 | ret = check_hotplug_memory_range(start, size);
|
| 1344 | if (ret)
|
| 1345 | return ret;
|
| 1346 |
|
| 1347 | if (mhp_flags & MHP_NID_IS_MGID) {
|
| 1348 | group = memory_group_find_by_id(nid);
|
| 1349 | if (!group)
|
| 1350 | return -EINVAL;
|
| 1351 | nid = group->nid;
|
| 1352 | }
|
| 1353 |
|
| 1354 | if (!node_possible(nid)) {
|
| 1355 | WARN(1, "node %d was absent from the node_possible_map\n", nid);
|
| 1356 | return -EINVAL;
|
| 1357 | }
|
| 1358 |
|
| 1359 | mem_hotplug_begin();
|
| 1360 |
|
| 1361 | if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
|
| 1362 | if (res->flags & IORESOURCE_SYSRAM_DRIVER_MANAGED)
|
| 1363 | memblock_flags = MEMBLOCK_DRIVER_MANAGED;
|
| 1364 | ret = memblock_add_node(start, size, nid, memblock_flags);
|
| 1365 | if (ret)
|
| 1366 | goto error_mem_hotplug_end;
|
| 1367 | }
|
| 1368 |
|
| 1369 | ret = __try_online_node(nid, false);
|
| 1370 | if (ret < 0)
|
| 1371 | goto error;
|
| 1372 | new_node = ret;
|
| 1373 |
|
| 1374 | |
| 1375 | * Self hosted memmap array
|
| 1376 |
|
| 1377 | if (mhp_flags & MHP_MEMMAP_ON_MEMORY) {
|
| 1378 | if (!mhp_supports_memmap_on_memory(size)) {
|
| 1379 | ret = -EINVAL;
|
| 1380 | goto error;
|
| 1381 | }
|
| 1382 | mhp_altmap.free = PHYS_PFN(size);
|
| 1383 | mhp_altmap.base_pfn = PHYS_PFN(start);
|
| 1384 | params.altmap = &mhp_altmap;
|
| 1385 | }
|
| 1386 |
|
| 1387 |
|
| 1388 | ret = arch_add_memory(nid, start, size, ¶ms);
|
| 1389 | if (ret < 0)
|
| 1390 | goto error;
|
| 1391 |
|
| 1392 |
|
| 1393 | ret = create_memory_block_devices(start, size, mhp_altmap.alloc,
|
| 1394 | group);
|
| 1395 | if (ret) {
|
| 1396 | arch_remove_memory(start, size, NULL);
|
| 1397 | goto error;
|
| 1398 | }
|
| 1399 |
|
| 1400 | if (new_node) {
|
| 1401 | |
| 1402 | * can't be hot-added. There is no rollback way now.
|
| 1403 | * So, check by BUG_ON() to catch it reluctantly..
|
| 1404 | * We online node here. We can't roll back from here.
|
| 1405 |
|
| 1406 | node_set_online(nid);
|
| 1407 | ret = __register_one_node(nid);
|
| 1408 | BUG_ON(ret);
|
| 1409 | }
|
| 1410 |
|
| 1411 | register_memory_blocks_under_node(nid, PFN_DOWN(start),
|
| 1412 | PFN_UP(start + size - 1),
|
| 1413 | MEMINIT_HOTPLUG);
|
| 1414 |
|
| 1415 |
|
| 1416 | if (!strcmp(res->name, "System RAM"))
|
| 1417 | firmware_map_add_hotplug(start, start + size, "System RAM");
|
| 1418 |
|
| 1419 |
|
| 1420 | mem_hotplug_done();
|
| 1421 |
|
| 1422 | |
| 1423 | * In case we're allowed to merge the resource, flag it and trigger
|
| 1424 | * merging now that adding succeeded.
|
| 1425 |
|
| 1426 | if (mhp_flags & MHP_MERGE_RESOURCE)
|
| 1427 | merge_system_ram_resource(res);
|
| 1428 |
|
| 1429 |
|
| 1430 | if (mhp_default_online_type != MMOP_OFFLINE)
|
| 1431 | walk_memory_blocks(start, size, NULL, online_memory_block);
|
| 1432 |
|
| 1433 | return ret;
|
| 1434 | error:
|
| 1435 | if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK))
|
| 1436 | memblock_remove(start, size);
|
| 1437 | error_mem_hotplug_end:
|
| 1438 | mem_hotplug_done();
|
| 1439 | return ret;
|
| 1440 | }
|
| 1441 |
|
| 1442 |
|
| 1443 | int __ref __add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
|
| 1444 | {
|
| 1445 | struct resource *res;
|
| 1446 | int ret;
|
| 1447 |
|
| 1448 | res = register_memory_resource(start, size, "System RAM");
|
| 1449 | if (IS_ERR(res))
|
| 1450 | return PTR_ERR(res);
|
| 1451 |
|
| 1452 | ret = add_memory_resource(nid, res, mhp_flags);
|
| 1453 | if (ret < 0)
|
| 1454 | release_memory_resource(res);
|
| 1455 | return ret;
|
| 1456 | }
|
| 1457 |
|
| 1458 | int add_memory(int nid, u64 start, u64 size, mhp_t mhp_flags)
|
| 1459 | {
|
| 1460 | int rc;
|
| 1461 |
|
| 1462 | lock_device_hotplug();
|
| 1463 | rc = __add_memory(nid, start, size, mhp_flags);
|
| 1464 | unlock_device_hotplug();
|
| 1465 |
|
| 1466 | return rc;
|
| 1467 | }
|
| 1468 | EXPORT_SYMBOL_GPL(add_memory);
|
| 1469 |
|
| 1470 | |
| 1471 | * Add special, driver-managed memory to the system as system RAM. Such
|
| 1472 | * memory is not exposed via the raw firmware-provided memmap as system
|
| 1473 | * RAM, instead, it is detected and added by a driver - during cold boot,
|
| 1474 | * after a reboot, and after kexec.
|
| 1475 | *
|
| 1476 | * Reasons why this memory should not be used for the initial memmap of a
|
| 1477 | * kexec kernel or for placing kexec images:
|
| 1478 | * - The booting kernel is in charge of determining how this memory will be
|
| 1479 | * used (e.g., use persistent memory as system RAM)
|
| 1480 | * - Coordination with a hypervisor is required before this memory
|
| 1481 | * can be used (e.g., inaccessible parts).
|
| 1482 | *
|
| 1483 | * For this memory, no entries in /sys/firmware/memmap ("raw firmware-provided
|
| 1484 | * memory map") are created. Also, the created memory resource is flagged
|
| 1485 | * with IORESOURCE_SYSRAM_DRIVER_MANAGED, so in-kernel users can special-case
|
| 1486 | * this memory as well (esp., not place kexec images onto it).
|
| 1487 | *
|
| 1488 | * The resource_name (visible via /proc/iomem) has to have the format
|
| 1489 | * "System RAM ($DRIVER)".
|
| 1490 |
|
| 1491 | int add_memory_driver_managed(int nid, u64 start, u64 size,
|
| 1492 | const char *resource_name, mhp_t mhp_flags)
|
| 1493 | {
|
| 1494 | struct resource *res;
|
| 1495 | int rc;
|
| 1496 |
|
| 1497 | if (!resource_name ||
|
| 1498 | strstr(resource_name, "System RAM (") != resource_name ||
|
| 1499 | resource_name[strlen(resource_name) - 1] != ')')
|
| 1500 | return -EINVAL;
|
| 1501 |
|
| 1502 | lock_device_hotplug();
|
| 1503 |
|
| 1504 | res = register_memory_resource(start, size, resource_name);
|
| 1505 | if (IS_ERR(res)) {
|
| 1506 | rc = PTR_ERR(res);
|
| 1507 | goto out_unlock;
|
| 1508 | }
|
| 1509 |
|
| 1510 | rc = add_memory_resource(nid, res, mhp_flags);
|
| 1511 | if (rc < 0)
|
| 1512 | release_memory_resource(res);
|
| 1513 |
|
| 1514 | out_unlock:
|
| 1515 | unlock_device_hotplug();
|
| 1516 | return rc;
|
| 1517 | }
|
| 1518 | EXPORT_SYMBOL_GPL(add_memory_driver_managed);
|
| 1519 |
|
| 1520 | |
| 1521 | * Platforms should define arch_get_mappable_range() that provides
|
| 1522 | * maximum possible addressable physical memory range for which the
|
| 1523 | * linear mapping could be created. The platform returned address
|
| 1524 | * range must adhere to these following semantics.
|
| 1525 | *
|
| 1526 | * - range.start <= range.end
|
| 1527 | * - Range includes both end points [range.start..range.end]
|
| 1528 | *
|
| 1529 | * There is also a fallback definition provided here, allowing the
|
| 1530 | * entire possible physical address range in case any platform does
|
| 1531 | * not define arch_get_mappable_range().
|
| 1532 |
|
| 1533 | struct range __weak arch_get_mappable_range(void)
|
| 1534 | {
|
| 1535 | struct range mhp_range = {
|
| 1536 | .start = 0UL,
|
| 1537 | .end = -1ULL,
|
| 1538 | };
|
| 1539 | return mhp_range;
|
| 1540 | }
|
| 1541 |
|
| 1542 | struct range mhp_get_pluggable_range(bool need_mapping)
|
| 1543 | {
|
| 1544 | const u64 max_phys = (1ULL << MAX_PHYSMEM_BITS) - 1;
|
| 1545 | struct range mhp_range;
|
| 1546 |
|
| 1547 | if (need_mapping) {
|
| 1548 | mhp_range = arch_get_mappable_range();
|
| 1549 | if (mhp_range.start > max_phys) {
|
| 1550 | mhp_range.start = 0;
|
| 1551 | mhp_range.end = 0;
|
| 1552 | }
|
| 1553 | mhp_range.end = min_t(u64, mhp_range.end, max_phys);
|
| 1554 | } else {
|
| 1555 | mhp_range.start = 0;
|
| 1556 | mhp_range.end = max_phys;
|
| 1557 | }
|
| 1558 | return mhp_range;
|
| 1559 | }
|
| 1560 | EXPORT_SYMBOL_GPL(mhp_get_pluggable_range);
|
| 1561 |
|
| 1562 | bool mhp_range_allowed(u64 start, u64 size, bool need_mapping)
|
| 1563 | {
|
| 1564 | struct range mhp_range = mhp_get_pluggable_range(need_mapping);
|
| 1565 | u64 end = start + size;
|
| 1566 |
|
| 1567 | if (start < end && start >= mhp_range.start && (end - 1) <= mhp_range.end)
|
| 1568 | return true;
|
| 1569 |
|
| 1570 | pr_warn("Hotplug memory [%#llx-%#llx] exceeds maximum addressable range [%#llx-%#llx]\n",
|
| 1571 | start, end, mhp_range.start, mhp_range.end);
|
| 1572 | return false;
|
| 1573 | }
|
| 1574 |
|
| 1575 | #ifdef CONFIG_MEMORY_HOTREMOVE
|
| 1576 | |
| 1577 | * Scan pfn range [start,end) to find movable/migratable pages (LRU pages,
|
| 1578 | * non-lru movable pages and hugepages). Will skip over most unmovable
|
| 1579 | * pages (esp., pages that can be skipped when offlining), but bail out on
|
| 1580 | * definitely unmovable pages.
|
| 1581 | *
|
| 1582 | * Returns:
|
| 1583 | * 0 in case a movable page is found and movable_pfn was updated.
|
| 1584 | * -ENOENT in case no movable page was found.
|
| 1585 | * -EBUSY in case a definitely unmovable page was found.
|
| 1586 |
|
| 1587 | static int scan_movable_pages(unsigned long start, unsigned long end,
|
| 1588 | unsigned long *movable_pfn)
|
| 1589 | {
|
| 1590 | unsigned long pfn;
|
| 1591 |
|
| 1592 | for (pfn = start; pfn < end; pfn++) {
|
| 1593 | struct page *page, *head;
|
| 1594 | unsigned long skip;
|
| 1595 |
|
| 1596 | if (!pfn_valid(pfn))
|
| 1597 | continue;
|
| 1598 | page = pfn_to_page(pfn);
|
| 1599 | if (PageLRU(page))
|
| 1600 | goto found;
|
| 1601 | if (__PageMovable(page))
|
| 1602 | goto found;
|
| 1603 |
|
| 1604 | |
| 1605 | * PageOffline() pages that are not marked __PageMovable() and
|
| 1606 | * have a reference count > 0 (after MEM_GOING_OFFLINE) are
|
| 1607 | * definitely unmovable. If their reference count would be 0,
|
| 1608 | * they could at least be skipped when offlining memory.
|
| 1609 |
|
| 1610 | if (PageOffline(page) && page_count(page))
|
| 1611 | return -EBUSY;
|
| 1612 |
|
| 1613 | if (!PageHuge(page))
|
| 1614 | continue;
|
| 1615 | head = compound_head(page);
|
| 1616 | |
| 1617 | * This test is racy as we hold no reference or lock. The
|
| 1618 | * hugetlb page could have been free'ed and head is no longer
|
| 1619 | * a hugetlb page before the following check. In such unlikely
|
| 1620 | * cases false positives and negatives are possible. Calling
|
| 1621 | * code must deal with these scenarios.
|
| 1622 |
|
| 1623 | if (HPageMigratable(head))
|
| 1624 | goto found;
|
| 1625 | skip = compound_nr(head) - (page - head);
|
| 1626 | pfn += skip - 1;
|
| 1627 | }
|
| 1628 | return -ENOENT;
|
| 1629 | found:
|
| 1630 | *movable_pfn = pfn;
|
| 1631 | return 0;
|
| 1632 | }
|
| 1633 |
|
| 1634 | static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
|
| 1635 | {
|
| 1636 | unsigned long pfn;
|
| 1637 | struct page *page, *head;
|
| 1638 | LIST_HEAD(source);
|
| 1639 | static DEFINE_RATELIMIT_STATE(migrate_rs, DEFAULT_RATELIMIT_INTERVAL,
|
| 1640 | DEFAULT_RATELIMIT_BURST);
|
| 1641 |
|
| 1642 | for (pfn = start_pfn; pfn < end_pfn; pfn++) {
|
| 1643 | struct folio *folio;
|
| 1644 | bool isolated;
|
| 1645 |
|
| 1646 | if (!pfn_valid(pfn))
|
| 1647 | continue;
|
| 1648 | page = pfn_to_page(pfn);
|
| 1649 | folio = page_folio(page);
|
| 1650 | head = &folio->page;
|
| 1651 |
|
| 1652 | if (PageHuge(page)) {
|
| 1653 | pfn = page_to_pfn(head) + compound_nr(head) - 1;
|
| 1654 | isolate_hugetlb(folio, &source);
|
| 1655 | continue;
|
| 1656 | } else if (PageTransHuge(page))
|
| 1657 | pfn = page_to_pfn(head) + thp_nr_pages(page) - 1;
|
| 1658 |
|
| 1659 | |
| 1660 | * HWPoison pages have elevated reference counts so the migration would
|
| 1661 | * fail on them. It also doesn't make any sense to migrate them in the
|
| 1662 | * first place. Still try to unmap such a page in case it is still mapped
|
| 1663 | * (e.g. current hwpoison implementation doesn't unmap KSM pages but keep
|
| 1664 | * the unmap as the catch all safety net).
|
| 1665 |
|
| 1666 | if (PageHWPoison(page)) {
|
| 1667 | if (WARN_ON(folio_test_lru(folio)))
|
| 1668 | folio_isolate_lru(folio);
|
| 1669 | if (folio_mapped(folio))
|
| 1670 | try_to_unmap(folio, TTU_IGNORE_MLOCK);
|
| 1671 | continue;
|
| 1672 | }
|
| 1673 |
|
| 1674 | if (!get_page_unless_zero(page))
|
| 1675 | continue;
|
| 1676 | |
| 1677 | * We can skip free pages. And we can deal with pages on
|
| 1678 | * LRU and non-lru movable pages.
|
| 1679 |
|
| 1680 | if (PageLRU(page))
|
| 1681 | isolated = isolate_lru_page(page);
|
| 1682 | else
|
| 1683 | isolated = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
|
| 1684 | if (isolated) {
|
| 1685 | list_add_tail(&page->lru, &source);
|
| 1686 | if (!__PageMovable(page))
|
| 1687 | inc_node_page_state(page, NR_ISOLATED_ANON +
|
| 1688 | page_is_file_lru(page));
|
| 1689 |
|
| 1690 | } else {
|
| 1691 | if (__ratelimit(&migrate_rs)) {
|
| 1692 | pr_warn("failed to isolate pfn %lx\n", pfn);
|
| 1693 | dump_page(page, "isolation failed");
|
| 1694 | }
|
| 1695 | }
|
| 1696 | put_page(page);
|
| 1697 | }
|
| 1698 | if (!list_empty(&source)) {
|
| 1699 | nodemask_t nmask = node_states[N_MEMORY];
|
| 1700 | struct migration_target_control mtc = {
|
| 1701 | .nmask = &nmask,
|
| 1702 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
|
| 1703 | };
|
| 1704 | int ret;
|
| 1705 |
|
| 1706 | |
| 1707 | * We have checked that migration range is on a single zone so
|
| 1708 | * we can use the nid of the first page to all the others.
|
| 1709 |
|
| 1710 | mtc.nid = page_to_nid(list_first_entry(&source, struct page, lru));
|
| 1711 |
|
| 1712 | |
| 1713 | * try to allocate from a different node but reuse this node
|
| 1714 | * if there are no other online nodes to be used (e.g. we are
|
| 1715 | * offlining a part of the only existing node)
|
| 1716 |
|
| 1717 | node_clear(mtc.nid, nmask);
|
| 1718 | if (nodes_empty(nmask))
|
| 1719 | node_set(mtc.nid, nmask);
|
| 1720 | ret = migrate_pages(&source, alloc_migration_target, NULL,
|
| 1721 | (unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG, NULL);
|
| 1722 | if (ret) {
|
| 1723 | list_for_each_entry(page, &source, lru) {
|
| 1724 | if (__ratelimit(&migrate_rs)) {
|
| 1725 | pr_warn("migrating pfn %lx failed ret:%d\n",
|
| 1726 | page_to_pfn(page), ret);
|
| 1727 | dump_page(page, "migration failure");
|
| 1728 | }
|
| 1729 | }
|
| 1730 | putback_movable_pages(&source);
|
| 1731 | }
|
| 1732 | }
|
| 1733 | }
|
| 1734 |
|
| 1735 | static int __init cmdline_parse_movable_node(char *p)
|
| 1736 | {
|
| 1737 | movable_node_enabled = true;
|
| 1738 | return 0;
|
| 1739 | }
|
| 1740 | early_param("movable_node", cmdline_parse_movable_node);
|
| 1741 |
|
| 1742 |
|
| 1743 | static void node_states_check_changes_offline(unsigned long nr_pages,
|
| 1744 | struct zone *zone, struct memory_notify *arg)
|
| 1745 | {
|
| 1746 | struct pglist_data *pgdat = zone->zone_pgdat;
|
| 1747 | unsigned long present_pages = 0;
|
| 1748 | enum zone_type zt;
|
| 1749 |
|
| 1750 | arg->status_change_nid = NUMA_NO_NODE;
|
| 1751 | arg->status_change_nid_normal = NUMA_NO_NODE;
|
| 1752 |
|
| 1753 | |
| 1754 | * Check whether node_states[N_NORMAL_MEMORY] will be changed.
|
| 1755 | * If the memory to be offline is within the range
|
| 1756 | * [0..ZONE_NORMAL], and it is the last present memory there,
|
| 1757 | * the zones in that range will become empty after the offlining,
|
| 1758 | * thus we can determine that we need to clear the node from
|
| 1759 | * node_states[N_NORMAL_MEMORY].
|
| 1760 |
|
| 1761 | for (zt = 0; zt <= ZONE_NORMAL; zt++)
|
| 1762 | present_pages += pgdat->node_zones[zt].present_pages;
|
| 1763 | if (zone_idx(zone) <= ZONE_NORMAL && nr_pages >= present_pages)
|
| 1764 | arg->status_change_nid_normal = zone_to_nid(zone);
|
| 1765 |
|
| 1766 | |
| 1767 | * We have accounted the pages from [0..ZONE_NORMAL); ZONE_HIGHMEM
|
| 1768 | * does not apply as we don't support 32bit.
|
| 1769 | * Here we count the possible pages from ZONE_MOVABLE.
|
| 1770 | * If after having accounted all the pages, we see that the nr_pages
|
| 1771 | * to be offlined is over or equal to the accounted pages,
|
| 1772 | * we know that the node will become empty, and so, we can clear
|
| 1773 | * it for N_MEMORY as well.
|
| 1774 |
|
| 1775 | present_pages += pgdat->node_zones[ZONE_MOVABLE].present_pages;
|
| 1776 |
|
| 1777 | if (nr_pages >= present_pages)
|
| 1778 | arg->status_change_nid = zone_to_nid(zone);
|
| 1779 | }
|
| 1780 |
|
| 1781 | static void node_states_clear_node(int node, struct memory_notify *arg)
|
| 1782 | {
|
| 1783 | if (arg->status_change_nid_normal >= 0)
|
| 1784 | node_clear_state(node, N_NORMAL_MEMORY);
|
| 1785 |
|
| 1786 | if (arg->status_change_nid >= 0)
|
| 1787 | node_clear_state(node, N_MEMORY);
|
| 1788 | }
|
| 1789 |
|
| 1790 | static int count_system_ram_pages_cb(unsigned long start_pfn,
|
| 1791 | unsigned long nr_pages, void *data)
|
| 1792 | {
|
| 1793 | unsigned long *nr_system_ram_pages = data;
|
| 1794 |
|
| 1795 | *nr_system_ram_pages += nr_pages;
|
| 1796 | return 0;
|
| 1797 | }
|
| 1798 |
|
| 1799 | int __ref offline_pages(unsigned long start_pfn, unsigned long nr_pages,
|
| 1800 | struct zone *zone, struct memory_group *group)
|
| 1801 | {
|
| 1802 | const unsigned long end_pfn = start_pfn + nr_pages;
|
| 1803 | unsigned long pfn, system_ram_pages = 0;
|
| 1804 | const int node = zone_to_nid(zone);
|
| 1805 | unsigned long flags;
|
| 1806 | struct memory_notify arg;
|
| 1807 | char *reason;
|
| 1808 | int ret;
|
| 1809 |
|
| 1810 | |
| 1811 | * {on,off}lining is constrained to full memory sections (or more
|
| 1812 | * precisely to memory blocks from the user space POV).
|
| 1813 | * memmap_on_memory is an exception because it reserves initial part
|
| 1814 | * of the physical memory space for vmemmaps. That space is pageblock
|
| 1815 | * aligned.
|
| 1816 |
|
| 1817 | if (WARN_ON_ONCE(!nr_pages || !pageblock_aligned(start_pfn) ||
|
| 1818 | !IS_ALIGNED(start_pfn + nr_pages, PAGES_PER_SECTION)))
|
| 1819 | return -EINVAL;
|
| 1820 |
|
| 1821 | mem_hotplug_begin();
|
| 1822 |
|
| 1823 | |
| 1824 | * Don't allow to offline memory blocks that contain holes.
|
| 1825 | * Consequently, memory blocks with holes can never get onlined
|
| 1826 | * via the hotplug path - online_pages() - as hotplugged memory has
|
| 1827 | * no holes. This way, we e.g., don't have to worry about marking
|
| 1828 | * memory holes PG_reserved, don't need pfn_valid() checks, and can
|
| 1829 | * avoid using walk_system_ram_range() later.
|
| 1830 |
|
| 1831 | walk_system_ram_range(start_pfn, nr_pages, &system_ram_pages,
|
| 1832 | count_system_ram_pages_cb);
|
| 1833 | if (system_ram_pages != nr_pages) {
|
| 1834 | ret = -EINVAL;
|
| 1835 | reason = "memory holes";
|
| 1836 | goto failed_removal;
|
| 1837 | }
|
| 1838 |
|
| 1839 | |
| 1840 | * We only support offlining of memory blocks managed by a single zone,
|
| 1841 | * checked by calling code. This is just a sanity check that we might
|
| 1842 | * want to remove in the future.
|
| 1843 |
|
| 1844 | if (WARN_ON_ONCE(page_zone(pfn_to_page(start_pfn)) != zone ||
|
| 1845 | page_zone(pfn_to_page(end_pfn - 1)) != zone)) {
|
| 1846 | ret = -EINVAL;
|
| 1847 | reason = "multizone range";
|
| 1848 | goto failed_removal;
|
| 1849 | }
|
| 1850 |
|
| 1851 | |
| 1852 | * Disable pcplists so that page isolation cannot race with freeing
|
| 1853 | * in a way that pages from isolated pageblock are left on pcplists.
|
| 1854 |
|
| 1855 | zone_pcp_disable(zone);
|
| 1856 | lru_cache_disable();
|
| 1857 |
|
| 1858 |
|
| 1859 | ret = start_isolate_page_range(start_pfn, end_pfn,
|
| 1860 | MIGRATE_MOVABLE,
|
| 1861 | MEMORY_OFFLINE | REPORT_FAILURE,
|
| 1862 | GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL);
|
| 1863 | if (ret) {
|
| 1864 | reason = "failure to isolate range";
|
| 1865 | goto failed_removal_pcplists_disabled;
|
| 1866 | }
|
| 1867 |
|
| 1868 | arg.start_pfn = start_pfn;
|
| 1869 | arg.nr_pages = nr_pages;
|
| 1870 | node_states_check_changes_offline(nr_pages, zone, &arg);
|
| 1871 |
|
| 1872 | ret = memory_notify(MEM_GOING_OFFLINE, &arg);
|
| 1873 | ret = notifier_to_errno(ret);
|
| 1874 | if (ret) {
|
| 1875 | reason = "notifier failure";
|
| 1876 | goto failed_removal_isolated;
|
| 1877 | }
|
| 1878 |
|
| 1879 | do {
|
| 1880 | pfn = start_pfn;
|
| 1881 | do {
|
| 1882 | if (signal_pending(current)) {
|
| 1883 | ret = -EINTR;
|
| 1884 | reason = "signal backoff";
|
| 1885 | goto failed_removal_isolated;
|
| 1886 | }
|
| 1887 |
|
| 1888 | cond_resched();
|
| 1889 |
|
| 1890 | ret = scan_movable_pages(pfn, end_pfn, &pfn);
|
| 1891 | if (!ret) {
|
| 1892 | |
| 1893 | * TODO: fatal migration failures should bail
|
| 1894 | * out
|
| 1895 |
|
| 1896 | do_migrate_range(pfn, end_pfn);
|
| 1897 | }
|
| 1898 | } while (!ret);
|
| 1899 |
|
| 1900 | if (ret != -ENOENT) {
|
| 1901 | reason = "unmovable page";
|
| 1902 | goto failed_removal_isolated;
|
| 1903 | }
|
| 1904 |
|
| 1905 | |
| 1906 | * Dissolve free hugepages in the memory block before doing
|
| 1907 | * offlining actually in order to make hugetlbfs's object
|
| 1908 | * counting consistent.
|
| 1909 |
|
| 1910 | ret = dissolve_free_huge_pages(start_pfn, end_pfn);
|
| 1911 | if (ret) {
|
| 1912 | reason = "failure to dissolve huge pages";
|
| 1913 | goto failed_removal_isolated;
|
| 1914 | }
|
| 1915 |
|
| 1916 | ret = test_pages_isolated(start_pfn, end_pfn, MEMORY_OFFLINE);
|
| 1917 |
|
| 1918 | } while (ret);
|
| 1919 |
|
| 1920 |
|
| 1921 | __offline_isolated_pages(start_pfn, end_pfn);
|
| 1922 | pr_debug("Offlined Pages %ld\n", nr_pages);
|
| 1923 |
|
| 1924 | |
| 1925 | * The memory sections are marked offline, and the pageblock flags
|
| 1926 | * effectively stale; nobody should be touching them. Fixup the number
|
| 1927 | * of isolated pageblocks, memory onlining will properly revert this.
|
| 1928 |
|
| 1929 | spin_lock_irqsave(&zone->lock, flags);
|
| 1930 | zone->nr_isolate_pageblock -= nr_pages / pageblock_nr_pages;
|
| 1931 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 1932 |
|
| 1933 | lru_cache_enable();
|
| 1934 | zone_pcp_enable(zone);
|
| 1935 |
|
| 1936 |
|
| 1937 | adjust_managed_page_count(pfn_to_page(start_pfn), -nr_pages);
|
| 1938 | adjust_present_page_count(pfn_to_page(start_pfn), group, -nr_pages);
|
| 1939 |
|
| 1940 |
|
| 1941 | init_per_zone_wmark_min();
|
| 1942 |
|
| 1943 | if (!populated_zone(zone)) {
|
| 1944 | zone_pcp_reset(zone);
|
| 1945 | build_all_zonelists(NULL);
|
| 1946 | }
|
| 1947 |
|
| 1948 | node_states_clear_node(node, &arg);
|
| 1949 | if (arg.status_change_nid >= 0) {
|
| 1950 | kcompactd_stop(node);
|
| 1951 | kswapd_stop(node);
|
| 1952 | }
|
| 1953 |
|
| 1954 | writeback_set_ratelimit();
|
| 1955 |
|
| 1956 | memory_notify(MEM_OFFLINE, &arg);
|
| 1957 | remove_pfn_range_from_zone(zone, start_pfn, nr_pages);
|
| 1958 | mem_hotplug_done();
|
| 1959 | return 0;
|
| 1960 |
|
| 1961 | failed_removal_isolated:
|
| 1962 |
|
| 1963 | undo_isolate_page_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
|
| 1964 | memory_notify(MEM_CANCEL_OFFLINE, &arg);
|
| 1965 | failed_removal_pcplists_disabled:
|
| 1966 | lru_cache_enable();
|
| 1967 | zone_pcp_enable(zone);
|
| 1968 | failed_removal:
|
| 1969 | pr_debug("memory offlining [mem %#010llx-%#010llx] failed due to %s\n",
|
| 1970 | (unsigned long long) start_pfn << PAGE_SHIFT,
|
| 1971 | ((unsigned long long) end_pfn << PAGE_SHIFT) - 1,
|
| 1972 | reason);
|
| 1973 | mem_hotplug_done();
|
| 1974 | return ret;
|
| 1975 | }
|
| 1976 |
|
| 1977 | static int check_memblock_offlined_cb(struct memory_block *mem, void *arg)
|
| 1978 | {
|
| 1979 | int *nid = arg;
|
| 1980 |
|
| 1981 | *nid = mem->nid;
|
| 1982 | if (unlikely(mem->state != MEM_OFFLINE)) {
|
| 1983 | phys_addr_t beginpa, endpa;
|
| 1984 |
|
| 1985 | beginpa = PFN_PHYS(section_nr_to_pfn(mem->start_section_nr));
|
| 1986 | endpa = beginpa + memory_block_size_bytes() - 1;
|
| 1987 | pr_warn("removing memory fails, because memory [%pa-%pa] is onlined\n",
|
| 1988 | &beginpa, &endpa);
|
| 1989 |
|
| 1990 | return -EBUSY;
|
| 1991 | }
|
| 1992 | return 0;
|
| 1993 | }
|
| 1994 |
|
| 1995 | static int get_nr_vmemmap_pages_cb(struct memory_block *mem, void *arg)
|
| 1996 | {
|
| 1997 | |
| 1998 | * If not set, continue with the next block.
|
| 1999 |
|
| 2000 | return mem->nr_vmemmap_pages;
|
| 2001 | }
|
| 2002 |
|
| 2003 | static int check_cpu_on_node(int nid)
|
| 2004 | {
|
| 2005 | int cpu;
|
| 2006 |
|
| 2007 | for_each_present_cpu(cpu) {
|
| 2008 | if (cpu_to_node(cpu) == nid)
|
| 2009 | |
| 2010 | * the cpu on this node isn't removed, and we can't
|
| 2011 | * offline this node.
|
| 2012 |
|
| 2013 | return -EBUSY;
|
| 2014 | }
|
| 2015 |
|
| 2016 | return 0;
|
| 2017 | }
|
| 2018 |
|
| 2019 | static int check_no_memblock_for_node_cb(struct memory_block *mem, void *arg)
|
| 2020 | {
|
| 2021 | int nid = *(int *)arg;
|
| 2022 |
|
| 2023 | |
| 2024 | * If a memory block belongs to multiple nodes, the stored nid is not
|
| 2025 | * reliable. However, such blocks are always online (e.g., cannot get
|
| 2026 | * offlined) and, therefore, are still spanned by the node.
|
| 2027 |
|
| 2028 | return mem->nid == nid ? -EEXIST : 0;
|
| 2029 | }
|
| 2030 |
|
| 2031 | |
| 2032 | * try_offline_node
|
| 2033 | * @nid: the node ID
|
| 2034 | *
|
| 2035 | * Offline a node if all memory sections and cpus of the node are removed.
|
| 2036 | *
|
| 2037 | * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
|
| 2038 | * and online/offline operations before this call.
|
| 2039 |
|
| 2040 | void try_offline_node(int nid)
|
| 2041 | {
|
| 2042 | int rc;
|
| 2043 |
|
| 2044 | |
| 2045 | * If the node still spans pages (especially ZONE_DEVICE), don't
|
| 2046 | * offline it. A node spans memory after move_pfn_range_to_zone(),
|
| 2047 | * e.g., after the memory block was onlined.
|
| 2048 |
|
| 2049 | if (node_spanned_pages(nid))
|
| 2050 | return;
|
| 2051 |
|
| 2052 | |
| 2053 | * Especially offline memory blocks might not be spanned by the
|
| 2054 | * node. They will get spanned by the node once they get onlined.
|
| 2055 | * However, they link to the node in sysfs and can get onlined later.
|
| 2056 |
|
| 2057 | rc = for_each_memory_block(&nid, check_no_memblock_for_node_cb);
|
| 2058 | if (rc)
|
| 2059 | return;
|
| 2060 |
|
| 2061 | if (check_cpu_on_node(nid))
|
| 2062 | return;
|
| 2063 |
|
| 2064 | |
| 2065 | * all memory/cpu of this node are removed, we can offline this
|
| 2066 | * node now.
|
| 2067 |
|
| 2068 | node_set_offline(nid);
|
| 2069 | unregister_one_node(nid);
|
| 2070 | }
|
| 2071 | EXPORT_SYMBOL(try_offline_node);
|
| 2072 |
|
| 2073 | static int __ref try_remove_memory(u64 start, u64 size)
|
| 2074 | {
|
| 2075 | struct vmem_altmap mhp_altmap = {};
|
| 2076 | struct vmem_altmap *altmap = NULL;
|
| 2077 | unsigned long nr_vmemmap_pages;
|
| 2078 | int rc = 0, nid = NUMA_NO_NODE;
|
| 2079 |
|
| 2080 | BUG_ON(check_hotplug_memory_range(start, size));
|
| 2081 |
|
| 2082 | |
| 2083 | * All memory blocks must be offlined before removing memory. Check
|
| 2084 | * whether all memory blocks in question are offline and return error
|
| 2085 | * if this is not the case.
|
| 2086 | *
|
| 2087 | * While at it, determine the nid. Note that if we'd have mixed nodes,
|
| 2088 | * we'd only try to offline the last determined one -- which is good
|
| 2089 | * enough for the cases we care about.
|
| 2090 |
|
| 2091 | rc = walk_memory_blocks(start, size, &nid, check_memblock_offlined_cb);
|
| 2092 | if (rc)
|
| 2093 | return rc;
|
| 2094 |
|
| 2095 | |
| 2096 | * We only support removing memory added with MHP_MEMMAP_ON_MEMORY in
|
| 2097 | * the same granularity it was added - a single memory block.
|
| 2098 |
|
| 2099 | if (mhp_memmap_on_memory()) {
|
| 2100 | nr_vmemmap_pages = walk_memory_blocks(start, size, NULL,
|
| 2101 | get_nr_vmemmap_pages_cb);
|
| 2102 | if (nr_vmemmap_pages) {
|
| 2103 | if (size != memory_block_size_bytes()) {
|
| 2104 | pr_warn("Refuse to remove %#llx - %#llx,"
|
| 2105 | "wrong granularity\n",
|
| 2106 | start, start + size);
|
| 2107 | return -EINVAL;
|
| 2108 | }
|
| 2109 |
|
| 2110 | |
| 2111 | * Let remove_pmd_table->free_hugepage_table do the
|
| 2112 | * right thing if we used vmem_altmap when hot-adding
|
| 2113 | * the range.
|
| 2114 |
|
| 2115 | mhp_altmap.alloc = nr_vmemmap_pages;
|
| 2116 | altmap = &mhp_altmap;
|
| 2117 | }
|
| 2118 | }
|
| 2119 |
|
| 2120 |
|
| 2121 | firmware_map_remove(start, start + size, "System RAM");
|
| 2122 |
|
| 2123 | |
| 2124 | * Memory block device removal under the device_hotplug_lock is
|
| 2125 | * a barrier against racing online attempts.
|
| 2126 |
|
| 2127 | remove_memory_block_devices(start, size);
|
| 2128 |
|
| 2129 | mem_hotplug_begin();
|
| 2130 |
|
| 2131 | arch_remove_memory(start, size, altmap);
|
| 2132 |
|
| 2133 | if (IS_ENABLED(CONFIG_ARCH_KEEP_MEMBLOCK)) {
|
| 2134 | memblock_phys_free(start, size);
|
| 2135 | memblock_remove(start, size);
|
| 2136 | }
|
| 2137 |
|
| 2138 | release_mem_region_adjustable(start, size);
|
| 2139 |
|
| 2140 | if (nid != NUMA_NO_NODE)
|
| 2141 | try_offline_node(nid);
|
| 2142 |
|
| 2143 | mem_hotplug_done();
|
| 2144 | return 0;
|
| 2145 | }
|
| 2146 |
|
| 2147 | |
| 2148 | * __remove_memory - Remove memory if every memory block is offline
|
| 2149 | * @start: physical address of the region to remove
|
| 2150 | * @size: size of the region to remove
|
| 2151 | *
|
| 2152 | * NOTE: The caller must call lock_device_hotplug() to serialize hotplug
|
| 2153 | * and online/offline operations before this call, as required by
|
| 2154 | * try_offline_node().
|
| 2155 |
|
| 2156 | void __remove_memory(u64 start, u64 size)
|
| 2157 | {
|
| 2158 |
|
| 2159 | |
| 2160 | * trigger BUG() if some memory is not offlined prior to calling this
|
| 2161 | * function
|
| 2162 |
|
| 2163 | if (try_remove_memory(start, size))
|
| 2164 | BUG();
|
| 2165 | }
|
| 2166 |
|
| 2167 | |
| 2168 | * Remove memory if every memory block is offline, otherwise return -EBUSY is
|
| 2169 | * some memory is not offline
|
| 2170 |
|
| 2171 | int remove_memory(u64 start, u64 size)
|
| 2172 | {
|
| 2173 | int rc;
|
| 2174 |
|
| 2175 | lock_device_hotplug();
|
| 2176 | rc = try_remove_memory(start, size);
|
| 2177 | unlock_device_hotplug();
|
| 2178 |
|
| 2179 | return rc;
|
| 2180 | }
|
| 2181 | EXPORT_SYMBOL_GPL(remove_memory);
|
| 2182 |
|
| 2183 | static int try_offline_memory_block(struct memory_block *mem, void *arg)
|
| 2184 | {
|
| 2185 | uint8_t online_type = MMOP_ONLINE_KERNEL;
|
| 2186 | uint8_t **online_types = arg;
|
| 2187 | struct page *page;
|
| 2188 | int rc;
|
| 2189 |
|
| 2190 | |
| 2191 | * Sense the online_type via the zone of the memory block. Offlining
|
| 2192 | * with multiple zones within one memory block will be rejected
|
| 2193 | * by offlining code ... so we don't care about that.
|
| 2194 |
|
| 2195 | page = pfn_to_online_page(section_nr_to_pfn(mem->start_section_nr));
|
| 2196 | if (page && zone_idx(page_zone(page)) == ZONE_MOVABLE)
|
| 2197 | online_type = MMOP_ONLINE_MOVABLE;
|
| 2198 |
|
| 2199 | rc = device_offline(&mem->dev);
|
| 2200 | |
| 2201 | * Default is MMOP_OFFLINE - change it only if offlining succeeded,
|
| 2202 | * so try_reonline_memory_block() can do the right thing.
|
| 2203 |
|
| 2204 | if (!rc)
|
| 2205 | **online_types = online_type;
|
| 2206 |
|
| 2207 | (*online_types)++;
|
| 2208 |
|
| 2209 | return rc < 0 ? rc : 0;
|
| 2210 | }
|
| 2211 |
|
| 2212 | static int try_reonline_memory_block(struct memory_block *mem, void *arg)
|
| 2213 | {
|
| 2214 | uint8_t **online_types = arg;
|
| 2215 | int rc;
|
| 2216 |
|
| 2217 | if (**online_types != MMOP_OFFLINE) {
|
| 2218 | mem->online_type = **online_types;
|
| 2219 | rc = device_online(&mem->dev);
|
| 2220 | if (rc < 0)
|
| 2221 | pr_warn("%s: Failed to re-online memory: %d",
|
| 2222 | __func__, rc);
|
| 2223 | }
|
| 2224 |
|
| 2225 |
|
| 2226 | (*online_types)++;
|
| 2227 | return 0;
|
| 2228 | }
|
| 2229 |
|
| 2230 | |
| 2231 | * Try to offline and remove memory. Might take a long time to finish in case
|
| 2232 | * memory is still in use. Primarily useful for memory devices that logically
|
| 2233 | * unplugged all memory (so it's no longer in use) and want to offline + remove
|
| 2234 | * that memory.
|
| 2235 |
|
| 2236 | int offline_and_remove_memory(u64 start, u64 size)
|
| 2237 | {
|
| 2238 | const unsigned long mb_count = size / memory_block_size_bytes();
|
| 2239 | uint8_t *online_types, *tmp;
|
| 2240 | int rc;
|
| 2241 |
|
| 2242 | if (!IS_ALIGNED(start, memory_block_size_bytes()) ||
|
| 2243 | !IS_ALIGNED(size, memory_block_size_bytes()) || !size)
|
| 2244 | return -EINVAL;
|
| 2245 |
|
| 2246 | |
| 2247 | * We'll remember the old online type of each memory block, so we can
|
| 2248 | * try to revert whatever we did when offlining one memory block fails
|
| 2249 | * after offlining some others succeeded.
|
| 2250 |
|
| 2251 | online_types = kmalloc_array(mb_count, sizeof(*online_types),
|
| 2252 | GFP_KERNEL);
|
| 2253 | if (!online_types)
|
| 2254 | return -ENOMEM;
|
| 2255 | |
| 2256 | * Initialize all states to MMOP_OFFLINE, so when we abort processing in
|
| 2257 | * try_offline_memory_block(), we'll skip all unprocessed blocks in
|
| 2258 | * try_reonline_memory_block().
|
| 2259 |
|
| 2260 | memset(online_types, MMOP_OFFLINE, mb_count);
|
| 2261 |
|
| 2262 | lock_device_hotplug();
|
| 2263 |
|
| 2264 | tmp = online_types;
|
| 2265 | rc = walk_memory_blocks(start, size, &tmp, try_offline_memory_block);
|
| 2266 |
|
| 2267 | |
| 2268 | * In case we succeeded to offline all memory, remove it.
|
| 2269 | * This cannot fail as it cannot get onlined in the meantime.
|
| 2270 |
|
| 2271 | if (!rc) {
|
| 2272 | rc = try_remove_memory(start, size);
|
| 2273 | if (rc)
|
| 2274 | pr_err("%s: Failed to remove memory: %d", __func__, rc);
|
| 2275 | }
|
| 2276 |
|
| 2277 | |
| 2278 | * Rollback what we did. While memory onlining might theoretically fail
|
| 2279 | * (nacked by a notifier), it barely ever happens.
|
| 2280 |
|
| 2281 | if (rc) {
|
| 2282 | tmp = online_types;
|
| 2283 | walk_memory_blocks(start, size, &tmp,
|
| 2284 | try_reonline_memory_block);
|
| 2285 | }
|
| 2286 | unlock_device_hotplug();
|
| 2287 |
|
| 2288 | kfree(online_types);
|
| 2289 | return rc;
|
| 2290 | }
|
| 2291 | EXPORT_SYMBOL_GPL(offline_and_remove_memory);
|
| 2292 | #endif
|