Linux Audio

Check our new training course

Embedded Linux training

Mar 10-20, 2025, special US time zones
Register
Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Memory subsystem support
   4 *
   5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
   6 *            Dave Hansen <haveblue@us.ibm.com>
   7 *
   8 * This file provides the necessary infrastructure to represent
   9 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
  10 * All arch-independent code that assumes MEMORY_HOTPLUG requires
  11 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
  12 */
  13
  14#include <linux/module.h>
  15#include <linux/init.h>
  16#include <linux/topology.h>
  17#include <linux/capability.h>
  18#include <linux/device.h>
  19#include <linux/memory.h>
  20#include <linux/memory_hotplug.h>
  21#include <linux/mm.h>
  22#include <linux/stat.h>
  23#include <linux/slab.h>
  24#include <linux/xarray.h>
  25
  26#include <linux/atomic.h>
  27#include <linux/uaccess.h>
  28
  29#define MEMORY_CLASS_NAME	"memory"
  30
  31static const char *const online_type_to_str[] = {
  32	[MMOP_OFFLINE] = "offline",
  33	[MMOP_ONLINE] = "online",
  34	[MMOP_ONLINE_KERNEL] = "online_kernel",
  35	[MMOP_ONLINE_MOVABLE] = "online_movable",
  36};
  37
  38int mhp_online_type_from_str(const char *str)
  39{
  40	int i;
  41
  42	for (i = 0; i < ARRAY_SIZE(online_type_to_str); i++) {
  43		if (sysfs_streq(str, online_type_to_str[i]))
  44			return i;
  45	}
  46	return -EINVAL;
  47}
  48
  49#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
  50
  51static int sections_per_block;
  52
  53static inline unsigned long memory_block_id(unsigned long section_nr)
  54{
  55	return section_nr / sections_per_block;
  56}
  57
  58static inline unsigned long pfn_to_block_id(unsigned long pfn)
  59{
  60	return memory_block_id(pfn_to_section_nr(pfn));
  61}
  62
  63static inline unsigned long phys_to_block_id(unsigned long phys)
  64{
  65	return pfn_to_block_id(PFN_DOWN(phys));
  66}
  67
  68static int memory_subsys_online(struct device *dev);
  69static int memory_subsys_offline(struct device *dev);
  70
  71static struct bus_type memory_subsys = {
  72	.name = MEMORY_CLASS_NAME,
  73	.dev_name = MEMORY_CLASS_NAME,
  74	.online = memory_subsys_online,
  75	.offline = memory_subsys_offline,
  76};
  77
  78/*
  79 * Memory blocks are cached in a local radix tree to avoid
  80 * a costly linear search for the corresponding device on
  81 * the subsystem bus.
  82 */
  83static DEFINE_XARRAY(memory_blocks);
  84
  85/*
  86 * Memory groups, indexed by memory group id (mgid).
  87 */
  88static DEFINE_XARRAY_FLAGS(memory_groups, XA_FLAGS_ALLOC);
  89#define MEMORY_GROUP_MARK_DYNAMIC	XA_MARK_1
  90
  91static BLOCKING_NOTIFIER_HEAD(memory_chain);
  92
  93int register_memory_notifier(struct notifier_block *nb)
  94{
  95	return blocking_notifier_chain_register(&memory_chain, nb);
  96}
  97EXPORT_SYMBOL(register_memory_notifier);
  98
  99void unregister_memory_notifier(struct notifier_block *nb)
 100{
 101	blocking_notifier_chain_unregister(&memory_chain, nb);
 102}
 103EXPORT_SYMBOL(unregister_memory_notifier);
 104
 105static void memory_block_release(struct device *dev)
 106{
 107	struct memory_block *mem = to_memory_block(dev);
 108
 
 109	kfree(mem);
 110}
 111
 112unsigned long __weak memory_block_size_bytes(void)
 113{
 114	return MIN_MEMORY_BLOCK_SIZE;
 115}
 116EXPORT_SYMBOL_GPL(memory_block_size_bytes);
 117
 118/*
 119 * Show the first physical section index (number) of this memory block.
 120 */
 121static ssize_t phys_index_show(struct device *dev,
 122			       struct device_attribute *attr, char *buf)
 123{
 124	struct memory_block *mem = to_memory_block(dev);
 125	unsigned long phys_index;
 126
 127	phys_index = mem->start_section_nr / sections_per_block;
 128
 129	return sysfs_emit(buf, "%08lx\n", phys_index);
 130}
 131
 132/*
 133 * Legacy interface that we cannot remove. Always indicate "removable"
 134 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
 135 */
 136static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
 137			      char *buf)
 138{
 139	return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
 140}
 141
 142/*
 143 * online, offline, going offline, etc.
 144 */
 145static ssize_t state_show(struct device *dev, struct device_attribute *attr,
 146			  char *buf)
 147{
 148	struct memory_block *mem = to_memory_block(dev);
 149	const char *output;
 150
 151	/*
 152	 * We can probably put these states in a nice little array
 153	 * so that they're not open-coded
 154	 */
 155	switch (mem->state) {
 156	case MEM_ONLINE:
 157		output = "online";
 158		break;
 159	case MEM_OFFLINE:
 160		output = "offline";
 161		break;
 162	case MEM_GOING_OFFLINE:
 163		output = "going-offline";
 164		break;
 165	default:
 166		WARN_ON(1);
 167		return sysfs_emit(buf, "ERROR-UNKNOWN-%ld\n", mem->state);
 168	}
 169
 170	return sysfs_emit(buf, "%s\n", output);
 171}
 172
 173int memory_notify(unsigned long val, void *v)
 174{
 175	return blocking_notifier_call_chain(&memory_chain, val, v);
 176}
 177
 178#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
 179static unsigned long memblk_nr_poison(struct memory_block *mem);
 180#else
 181static inline unsigned long memblk_nr_poison(struct memory_block *mem)
 182{
 183	return 0;
 184}
 185#endif
 186
 
 
 
 187static int memory_block_online(struct memory_block *mem)
 188{
 189	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 190	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 191	unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
 
 192	struct zone *zone;
 193	int ret;
 194
 195	if (memblk_nr_poison(mem))
 196		return -EHWPOISON;
 197
 198	zone = zone_for_pfn_range(mem->online_type, mem->nid, mem->group,
 199				  start_pfn, nr_pages);
 200
 201	/*
 202	 * Although vmemmap pages have a different lifecycle than the pages
 203	 * they describe (they remain until the memory is unplugged), doing
 204	 * their initialization and accounting at memory onlining/offlining
 205	 * stage helps to keep accounting easier to follow - e.g vmemmaps
 206	 * belong to the same zone as the memory they backed.
 207	 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 208	if (nr_vmemmap_pages) {
 209		ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
 
 210		if (ret)
 211			return ret;
 212	}
 213
 214	ret = online_pages(start_pfn + nr_vmemmap_pages,
 215			   nr_pages - nr_vmemmap_pages, zone, mem->group);
 216	if (ret) {
 217		if (nr_vmemmap_pages)
 218			mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
 219		return ret;
 220	}
 221
 222	/*
 223	 * Account once onlining succeeded. If the zone was unpopulated, it is
 224	 * now already properly populated.
 225	 */
 226	if (nr_vmemmap_pages)
 227		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
 228					  nr_vmemmap_pages);
 229
 230	mem->zone = zone;
 
 
 
 
 
 
 231	return ret;
 232}
 233
 
 
 
 234static int memory_block_offline(struct memory_block *mem)
 235{
 236	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 237	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 238	unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
 
 239	int ret;
 240
 241	if (!mem->zone)
 242		return -EINVAL;
 243
 244	/*
 245	 * Unaccount before offlining, such that unpopulated zone and kthreads
 246	 * can properly be torn down in offline_pages().
 247	 */
 
 
 
 
 248	if (nr_vmemmap_pages)
 249		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
 250					  -nr_vmemmap_pages);
 251
 252	ret = offline_pages(start_pfn + nr_vmemmap_pages,
 253			    nr_pages - nr_vmemmap_pages, mem->zone, mem->group);
 254	if (ret) {
 255		/* offline_pages() failed. Account back. */
 256		if (nr_vmemmap_pages)
 257			adjust_present_page_count(pfn_to_page(start_pfn),
 258						  mem->group, nr_vmemmap_pages);
 259		return ret;
 260	}
 261
 262	if (nr_vmemmap_pages)
 263		mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
 264
 265	mem->zone = NULL;
 
 
 
 
 
 
 
 266	return ret;
 267}
 268
 269/*
 270 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 271 * OK to have direct references to sparsemem variables in here.
 272 */
 273static int
 274memory_block_action(struct memory_block *mem, unsigned long action)
 275{
 276	int ret;
 277
 278	switch (action) {
 279	case MEM_ONLINE:
 280		ret = memory_block_online(mem);
 281		break;
 282	case MEM_OFFLINE:
 283		ret = memory_block_offline(mem);
 284		break;
 285	default:
 286		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 287		     "%ld\n", __func__, mem->start_section_nr, action, action);
 288		ret = -EINVAL;
 289	}
 290
 291	return ret;
 292}
 293
 294static int memory_block_change_state(struct memory_block *mem,
 295		unsigned long to_state, unsigned long from_state_req)
 296{
 297	int ret = 0;
 298
 299	if (mem->state != from_state_req)
 300		return -EINVAL;
 301
 302	if (to_state == MEM_OFFLINE)
 303		mem->state = MEM_GOING_OFFLINE;
 304
 305	ret = memory_block_action(mem, to_state);
 306	mem->state = ret ? from_state_req : to_state;
 307
 308	return ret;
 309}
 310
 311/* The device lock serializes operations on memory_subsys_[online|offline] */
 312static int memory_subsys_online(struct device *dev)
 313{
 314	struct memory_block *mem = to_memory_block(dev);
 315	int ret;
 316
 317	if (mem->state == MEM_ONLINE)
 318		return 0;
 319
 320	/*
 321	 * When called via device_online() without configuring the online_type,
 322	 * we want to default to MMOP_ONLINE.
 323	 */
 324	if (mem->online_type == MMOP_OFFLINE)
 325		mem->online_type = MMOP_ONLINE;
 326
 327	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 328	mem->online_type = MMOP_OFFLINE;
 329
 330	return ret;
 331}
 332
 333static int memory_subsys_offline(struct device *dev)
 334{
 335	struct memory_block *mem = to_memory_block(dev);
 336
 337	if (mem->state == MEM_OFFLINE)
 338		return 0;
 339
 340	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 341}
 342
 343static ssize_t state_store(struct device *dev, struct device_attribute *attr,
 344			   const char *buf, size_t count)
 345{
 346	const int online_type = mhp_online_type_from_str(buf);
 347	struct memory_block *mem = to_memory_block(dev);
 348	int ret;
 349
 350	if (online_type < 0)
 351		return -EINVAL;
 352
 353	ret = lock_device_hotplug_sysfs();
 354	if (ret)
 355		return ret;
 356
 357	switch (online_type) {
 358	case MMOP_ONLINE_KERNEL:
 359	case MMOP_ONLINE_MOVABLE:
 360	case MMOP_ONLINE:
 361		/* mem->online_type is protected by device_hotplug_lock */
 362		mem->online_type = online_type;
 363		ret = device_online(&mem->dev);
 364		break;
 365	case MMOP_OFFLINE:
 366		ret = device_offline(&mem->dev);
 367		break;
 368	default:
 369		ret = -EINVAL; /* should never happen */
 370	}
 371
 372	unlock_device_hotplug();
 373
 374	if (ret < 0)
 375		return ret;
 376	if (ret)
 377		return -EINVAL;
 378
 379	return count;
 380}
 381
 382/*
 383 * Legacy interface that we cannot remove: s390x exposes the storage increment
 384 * covered by a memory block, allowing for identifying which memory blocks
 385 * comprise a storage increment. Since a memory block spans complete
 386 * storage increments nowadays, this interface is basically unused. Other
 387 * archs never exposed != 0.
 388 */
 389static ssize_t phys_device_show(struct device *dev,
 390				struct device_attribute *attr, char *buf)
 391{
 392	struct memory_block *mem = to_memory_block(dev);
 393	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 394
 395	return sysfs_emit(buf, "%d\n",
 396			  arch_get_memory_phys_device(start_pfn));
 397}
 398
 399#ifdef CONFIG_MEMORY_HOTREMOVE
 400static int print_allowed_zone(char *buf, int len, int nid,
 401			      struct memory_group *group,
 402			      unsigned long start_pfn, unsigned long nr_pages,
 403			      int online_type, struct zone *default_zone)
 404{
 405	struct zone *zone;
 406
 407	zone = zone_for_pfn_range(online_type, nid, group, start_pfn, nr_pages);
 408	if (zone == default_zone)
 409		return 0;
 410
 411	return sysfs_emit_at(buf, len, " %s", zone->name);
 412}
 413
 414static ssize_t valid_zones_show(struct device *dev,
 415				struct device_attribute *attr, char *buf)
 416{
 417	struct memory_block *mem = to_memory_block(dev);
 418	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 419	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 420	struct memory_group *group = mem->group;
 421	struct zone *default_zone;
 422	int nid = mem->nid;
 423	int len = 0;
 424
 425	/*
 426	 * Check the existing zone. Make sure that we do that only on the
 427	 * online nodes otherwise the page_zone is not reliable
 428	 */
 429	if (mem->state == MEM_ONLINE) {
 430		/*
 431		 * If !mem->zone, the memory block spans multiple zones and
 432		 * cannot get offlined.
 433		 */
 434		default_zone = mem->zone;
 435		if (!default_zone)
 436			return sysfs_emit(buf, "%s\n", "none");
 437		len += sysfs_emit_at(buf, len, "%s", default_zone->name);
 438		goto out;
 439	}
 440
 441	default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, group,
 442					  start_pfn, nr_pages);
 443
 444	len += sysfs_emit_at(buf, len, "%s", default_zone->name);
 445	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
 446				  MMOP_ONLINE_KERNEL, default_zone);
 447	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
 448				  MMOP_ONLINE_MOVABLE, default_zone);
 449out:
 450	len += sysfs_emit_at(buf, len, "\n");
 451	return len;
 452}
 453static DEVICE_ATTR_RO(valid_zones);
 454#endif
 455
 456static DEVICE_ATTR_RO(phys_index);
 457static DEVICE_ATTR_RW(state);
 458static DEVICE_ATTR_RO(phys_device);
 459static DEVICE_ATTR_RO(removable);
 460
 461/*
 462 * Show the memory block size (shared by all memory blocks).
 463 */
 464static ssize_t block_size_bytes_show(struct device *dev,
 465				     struct device_attribute *attr, char *buf)
 466{
 467	return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
 468}
 469
 470static DEVICE_ATTR_RO(block_size_bytes);
 471
 472/*
 473 * Memory auto online policy.
 474 */
 475
 476static ssize_t auto_online_blocks_show(struct device *dev,
 477				       struct device_attribute *attr, char *buf)
 478{
 479	return sysfs_emit(buf, "%s\n",
 480			  online_type_to_str[mhp_default_online_type]);
 481}
 482
 483static ssize_t auto_online_blocks_store(struct device *dev,
 484					struct device_attribute *attr,
 485					const char *buf, size_t count)
 486{
 487	const int online_type = mhp_online_type_from_str(buf);
 488
 489	if (online_type < 0)
 490		return -EINVAL;
 491
 492	mhp_default_online_type = online_type;
 493	return count;
 494}
 495
 496static DEVICE_ATTR_RW(auto_online_blocks);
 497
 
 
 
 
 
 
 
 
 
 
 498/*
 499 * Some architectures will have custom drivers to do this, and
 500 * will not need to do it from userspace.  The fake hot-add code
 501 * as well as ppc64 will do all of their discovery in userspace
 502 * and will require this interface.
 503 */
 504#ifdef CONFIG_ARCH_MEMORY_PROBE
 505static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
 506			   const char *buf, size_t count)
 507{
 508	u64 phys_addr;
 509	int nid, ret;
 510	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
 511
 512	ret = kstrtoull(buf, 0, &phys_addr);
 513	if (ret)
 514		return ret;
 515
 516	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
 517		return -EINVAL;
 518
 519	ret = lock_device_hotplug_sysfs();
 520	if (ret)
 521		return ret;
 522
 523	nid = memory_add_physaddr_to_nid(phys_addr);
 524	ret = __add_memory(nid, phys_addr,
 525			   MIN_MEMORY_BLOCK_SIZE * sections_per_block,
 526			   MHP_NONE);
 527
 528	if (ret)
 529		goto out;
 530
 531	ret = count;
 532out:
 533	unlock_device_hotplug();
 534	return ret;
 535}
 536
 537static DEVICE_ATTR_WO(probe);
 538#endif
 539
 540#ifdef CONFIG_MEMORY_FAILURE
 541/*
 542 * Support for offlining pages of memory
 543 */
 544
 545/* Soft offline a page */
 546static ssize_t soft_offline_page_store(struct device *dev,
 547				       struct device_attribute *attr,
 548				       const char *buf, size_t count)
 549{
 550	int ret;
 551	u64 pfn;
 552	if (!capable(CAP_SYS_ADMIN))
 553		return -EPERM;
 554	if (kstrtoull(buf, 0, &pfn) < 0)
 555		return -EINVAL;
 556	pfn >>= PAGE_SHIFT;
 557	ret = soft_offline_page(pfn, 0);
 558	return ret == 0 ? count : ret;
 559}
 560
 561/* Forcibly offline a page, including killing processes. */
 562static ssize_t hard_offline_page_store(struct device *dev,
 563				       struct device_attribute *attr,
 564				       const char *buf, size_t count)
 565{
 566	int ret;
 567	u64 pfn;
 568	if (!capable(CAP_SYS_ADMIN))
 569		return -EPERM;
 570	if (kstrtoull(buf, 0, &pfn) < 0)
 571		return -EINVAL;
 572	pfn >>= PAGE_SHIFT;
 573	ret = memory_failure(pfn, MF_SW_SIMULATED);
 574	if (ret == -EOPNOTSUPP)
 575		ret = 0;
 576	return ret ? ret : count;
 577}
 578
 579static DEVICE_ATTR_WO(soft_offline_page);
 580static DEVICE_ATTR_WO(hard_offline_page);
 581#endif
 582
 583/* See phys_device_show(). */
 584int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 585{
 586	return 0;
 587}
 588
 589/*
 590 * A reference for the returned memory block device is acquired.
 591 *
 592 * Called under device_hotplug_lock.
 593 */
 594static struct memory_block *find_memory_block_by_id(unsigned long block_id)
 595{
 596	struct memory_block *mem;
 597
 598	mem = xa_load(&memory_blocks, block_id);
 599	if (mem)
 600		get_device(&mem->dev);
 601	return mem;
 602}
 603
 604/*
 605 * Called under device_hotplug_lock.
 606 */
 607struct memory_block *find_memory_block(unsigned long section_nr)
 608{
 609	unsigned long block_id = memory_block_id(section_nr);
 610
 611	return find_memory_block_by_id(block_id);
 612}
 613
 614static struct attribute *memory_memblk_attrs[] = {
 615	&dev_attr_phys_index.attr,
 616	&dev_attr_state.attr,
 617	&dev_attr_phys_device.attr,
 618	&dev_attr_removable.attr,
 619#ifdef CONFIG_MEMORY_HOTREMOVE
 620	&dev_attr_valid_zones.attr,
 621#endif
 622	NULL
 623};
 624
 625static const struct attribute_group memory_memblk_attr_group = {
 626	.attrs = memory_memblk_attrs,
 627};
 628
 629static const struct attribute_group *memory_memblk_attr_groups[] = {
 630	&memory_memblk_attr_group,
 631	NULL,
 632};
 633
 634static int __add_memory_block(struct memory_block *memory)
 635{
 636	int ret;
 637
 638	memory->dev.bus = &memory_subsys;
 639	memory->dev.id = memory->start_section_nr / sections_per_block;
 640	memory->dev.release = memory_block_release;
 641	memory->dev.groups = memory_memblk_attr_groups;
 642	memory->dev.offline = memory->state == MEM_OFFLINE;
 643
 644	ret = device_register(&memory->dev);
 645	if (ret) {
 646		put_device(&memory->dev);
 647		return ret;
 648	}
 649	ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
 650			      GFP_KERNEL));
 651	if (ret)
 652		device_unregister(&memory->dev);
 653
 654	return ret;
 655}
 656
 657static struct zone *early_node_zone_for_memory_block(struct memory_block *mem,
 658						     int nid)
 659{
 660	const unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 661	const unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 662	struct zone *zone, *matching_zone = NULL;
 663	pg_data_t *pgdat = NODE_DATA(nid);
 664	int i;
 665
 666	/*
 667	 * This logic only works for early memory, when the applicable zones
 668	 * already span the memory block. We don't expect overlapping zones on
 669	 * a single node for early memory. So if we're told that some PFNs
 670	 * of a node fall into this memory block, we can assume that all node
 671	 * zones that intersect with the memory block are actually applicable.
 672	 * No need to look at the memmap.
 673	 */
 674	for (i = 0; i < MAX_NR_ZONES; i++) {
 675		zone = pgdat->node_zones + i;
 676		if (!populated_zone(zone))
 677			continue;
 678		if (!zone_intersects(zone, start_pfn, nr_pages))
 679			continue;
 680		if (!matching_zone) {
 681			matching_zone = zone;
 682			continue;
 683		}
 684		/* Spans multiple zones ... */
 685		matching_zone = NULL;
 686		break;
 687	}
 688	return matching_zone;
 689}
 690
 691#ifdef CONFIG_NUMA
 692/**
 693 * memory_block_add_nid() - Indicate that system RAM falling into this memory
 694 *			    block device (partially) belongs to the given node.
 695 * @mem: The memory block device.
 696 * @nid: The node id.
 697 * @context: The memory initialization context.
 698 *
 699 * Indicate that system RAM falling into this memory block (partially) belongs
 700 * to the given node. If the context indicates ("early") that we are adding the
 701 * node during node device subsystem initialization, this will also properly
 702 * set/adjust mem->zone based on the zone ranges of the given node.
 703 */
 704void memory_block_add_nid(struct memory_block *mem, int nid,
 705			  enum meminit_context context)
 706{
 707	if (context == MEMINIT_EARLY && mem->nid != nid) {
 708		/*
 709		 * For early memory we have to determine the zone when setting
 710		 * the node id and handle multiple nodes spanning a single
 711		 * memory block by indicate via zone == NULL that we're not
 712		 * dealing with a single zone. So if we're setting the node id
 713		 * the first time, determine if there is a single zone. If we're
 714		 * setting the node id a second time to a different node,
 715		 * invalidate the single detected zone.
 716		 */
 717		if (mem->nid == NUMA_NO_NODE)
 718			mem->zone = early_node_zone_for_memory_block(mem, nid);
 719		else
 720			mem->zone = NULL;
 721	}
 722
 723	/*
 724	 * If this memory block spans multiple nodes, we only indicate
 725	 * the last processed node. If we span multiple nodes (not applicable
 726	 * to hotplugged memory), zone == NULL will prohibit memory offlining
 727	 * and consequently unplug.
 728	 */
 729	mem->nid = nid;
 730}
 731#endif
 732
 733static int add_memory_block(unsigned long block_id, unsigned long state,
 734			    unsigned long nr_vmemmap_pages,
 735			    struct memory_group *group)
 736{
 737	struct memory_block *mem;
 738	int ret = 0;
 739
 740	mem = find_memory_block_by_id(block_id);
 741	if (mem) {
 742		put_device(&mem->dev);
 743		return -EEXIST;
 744	}
 745	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 746	if (!mem)
 747		return -ENOMEM;
 748
 749	mem->start_section_nr = block_id * sections_per_block;
 750	mem->state = state;
 751	mem->nid = NUMA_NO_NODE;
 752	mem->nr_vmemmap_pages = nr_vmemmap_pages;
 753	INIT_LIST_HEAD(&mem->group_next);
 754
 755#ifndef CONFIG_NUMA
 756	if (state == MEM_ONLINE)
 757		/*
 758		 * MEM_ONLINE at this point implies early memory. With NUMA,
 759		 * we'll determine the zone when setting the node id via
 760		 * memory_block_add_nid(). Memory hotplug updated the zone
 761		 * manually when memory onlining/offlining succeeds.
 762		 */
 763		mem->zone = early_node_zone_for_memory_block(mem, NUMA_NO_NODE);
 764#endif /* CONFIG_NUMA */
 765
 766	ret = __add_memory_block(mem);
 767	if (ret)
 768		return ret;
 769
 770	if (group) {
 771		mem->group = group;
 772		list_add(&mem->group_next, &group->memory_blocks);
 773	}
 774
 775	return 0;
 776}
 777
 778static int __init add_boot_memory_block(unsigned long base_section_nr)
 779{
 780	int section_count = 0;
 781	unsigned long nr;
 782
 783	for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
 784	     nr++)
 785		if (present_section_nr(nr))
 786			section_count++;
 787
 788	if (section_count == 0)
 789		return 0;
 790	return add_memory_block(memory_block_id(base_section_nr),
 791				MEM_ONLINE, 0,  NULL);
 792}
 793
 794static int add_hotplug_memory_block(unsigned long block_id,
 795				    unsigned long nr_vmemmap_pages,
 796				    struct memory_group *group)
 797{
 798	return add_memory_block(block_id, MEM_OFFLINE, nr_vmemmap_pages, group);
 799}
 800
 801static void remove_memory_block(struct memory_block *memory)
 802{
 803	if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
 804		return;
 805
 806	WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
 807
 808	if (memory->group) {
 809		list_del(&memory->group_next);
 810		memory->group = NULL;
 811	}
 812
 813	/* drop the ref. we got via find_memory_block() */
 814	put_device(&memory->dev);
 815	device_unregister(&memory->dev);
 816}
 817
 818/*
 819 * Create memory block devices for the given memory area. Start and size
 820 * have to be aligned to memory block granularity. Memory block devices
 821 * will be initialized as offline.
 822 *
 823 * Called under device_hotplug_lock.
 824 */
 825int create_memory_block_devices(unsigned long start, unsigned long size,
 826				unsigned long vmemmap_pages,
 827				struct memory_group *group)
 828{
 829	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
 830	unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
 831	struct memory_block *mem;
 832	unsigned long block_id;
 833	int ret = 0;
 834
 835	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
 836			 !IS_ALIGNED(size, memory_block_size_bytes())))
 837		return -EINVAL;
 838
 839	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
 840		ret = add_hotplug_memory_block(block_id, vmemmap_pages, group);
 841		if (ret)
 842			break;
 843	}
 844	if (ret) {
 845		end_block_id = block_id;
 846		for (block_id = start_block_id; block_id != end_block_id;
 847		     block_id++) {
 848			mem = find_memory_block_by_id(block_id);
 849			if (WARN_ON_ONCE(!mem))
 850				continue;
 851			remove_memory_block(mem);
 852		}
 853	}
 854	return ret;
 855}
 856
 857/*
 858 * Remove memory block devices for the given memory area. Start and size
 859 * have to be aligned to memory block granularity. Memory block devices
 860 * have to be offline.
 861 *
 862 * Called under device_hotplug_lock.
 863 */
 864void remove_memory_block_devices(unsigned long start, unsigned long size)
 865{
 866	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
 867	const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
 868	struct memory_block *mem;
 869	unsigned long block_id;
 870
 871	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
 872			 !IS_ALIGNED(size, memory_block_size_bytes())))
 873		return;
 874
 875	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
 876		mem = find_memory_block_by_id(block_id);
 877		if (WARN_ON_ONCE(!mem))
 878			continue;
 879		num_poisoned_pages_sub(-1UL, memblk_nr_poison(mem));
 880		unregister_memory_block_under_nodes(mem);
 881		remove_memory_block(mem);
 882	}
 883}
 884
 885static struct attribute *memory_root_attrs[] = {
 886#ifdef CONFIG_ARCH_MEMORY_PROBE
 887	&dev_attr_probe.attr,
 888#endif
 889
 890#ifdef CONFIG_MEMORY_FAILURE
 891	&dev_attr_soft_offline_page.attr,
 892	&dev_attr_hard_offline_page.attr,
 893#endif
 894
 895	&dev_attr_block_size_bytes.attr,
 896	&dev_attr_auto_online_blocks.attr,
 
 
 
 897	NULL
 898};
 899
 900static const struct attribute_group memory_root_attr_group = {
 901	.attrs = memory_root_attrs,
 902};
 903
 904static const struct attribute_group *memory_root_attr_groups[] = {
 905	&memory_root_attr_group,
 906	NULL,
 907};
 908
 909/*
 910 * Initialize the sysfs support for memory devices. At the time this function
 911 * is called, we cannot have concurrent creation/deletion of memory block
 912 * devices, the device_hotplug_lock is not needed.
 913 */
 914void __init memory_dev_init(void)
 915{
 916	int ret;
 917	unsigned long block_sz, nr;
 918
 919	/* Validate the configured memory block size */
 920	block_sz = memory_block_size_bytes();
 921	if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
 922		panic("Memory block size not suitable: 0x%lx\n", block_sz);
 923	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 924
 925	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
 926	if (ret)
 927		panic("%s() failed to register subsystem: %d\n", __func__, ret);
 928
 929	/*
 930	 * Create entries for memory sections that were found
 931	 * during boot and have been initialized
 932	 */
 933	for (nr = 0; nr <= __highest_present_section_nr;
 934	     nr += sections_per_block) {
 935		ret = add_boot_memory_block(nr);
 936		if (ret)
 937			panic("%s() failed to add memory block: %d\n", __func__,
 938			      ret);
 939	}
 940}
 941
 942/**
 943 * walk_memory_blocks - walk through all present memory blocks overlapped
 944 *			by the range [start, start + size)
 945 *
 946 * @start: start address of the memory range
 947 * @size: size of the memory range
 948 * @arg: argument passed to func
 949 * @func: callback for each memory section walked
 950 *
 951 * This function walks through all present memory blocks overlapped by the
 952 * range [start, start + size), calling func on each memory block.
 953 *
 954 * In case func() returns an error, walking is aborted and the error is
 955 * returned.
 956 *
 957 * Called under device_hotplug_lock.
 958 */
 959int walk_memory_blocks(unsigned long start, unsigned long size,
 960		       void *arg, walk_memory_blocks_func_t func)
 961{
 962	const unsigned long start_block_id = phys_to_block_id(start);
 963	const unsigned long end_block_id = phys_to_block_id(start + size - 1);
 964	struct memory_block *mem;
 965	unsigned long block_id;
 966	int ret = 0;
 967
 968	if (!size)
 969		return 0;
 970
 971	for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
 972		mem = find_memory_block_by_id(block_id);
 973		if (!mem)
 974			continue;
 975
 976		ret = func(mem, arg);
 977		put_device(&mem->dev);
 978		if (ret)
 979			break;
 980	}
 981	return ret;
 982}
 983
 984struct for_each_memory_block_cb_data {
 985	walk_memory_blocks_func_t func;
 986	void *arg;
 987};
 988
 989static int for_each_memory_block_cb(struct device *dev, void *data)
 990{
 991	struct memory_block *mem = to_memory_block(dev);
 992	struct for_each_memory_block_cb_data *cb_data = data;
 993
 994	return cb_data->func(mem, cb_data->arg);
 995}
 996
 997/**
 998 * for_each_memory_block - walk through all present memory blocks
 999 *
1000 * @arg: argument passed to func
1001 * @func: callback for each memory block walked
1002 *
1003 * This function walks through all present memory blocks, calling func on
1004 * each memory block.
1005 *
1006 * In case func() returns an error, walking is aborted and the error is
1007 * returned.
1008 */
1009int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
1010{
1011	struct for_each_memory_block_cb_data cb_data = {
1012		.func = func,
1013		.arg = arg,
1014	};
1015
1016	return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
1017				for_each_memory_block_cb);
1018}
1019
1020/*
1021 * This is an internal helper to unify allocation and initialization of
1022 * memory groups. Note that the passed memory group will be copied to a
1023 * dynamically allocated memory group. After this call, the passed
1024 * memory group should no longer be used.
1025 */
1026static int memory_group_register(struct memory_group group)
1027{
1028	struct memory_group *new_group;
1029	uint32_t mgid;
1030	int ret;
1031
1032	if (!node_possible(group.nid))
1033		return -EINVAL;
1034
1035	new_group = kzalloc(sizeof(group), GFP_KERNEL);
1036	if (!new_group)
1037		return -ENOMEM;
1038	*new_group = group;
1039	INIT_LIST_HEAD(&new_group->memory_blocks);
1040
1041	ret = xa_alloc(&memory_groups, &mgid, new_group, xa_limit_31b,
1042		       GFP_KERNEL);
1043	if (ret) {
1044		kfree(new_group);
1045		return ret;
1046	} else if (group.is_dynamic) {
1047		xa_set_mark(&memory_groups, mgid, MEMORY_GROUP_MARK_DYNAMIC);
1048	}
1049	return mgid;
1050}
1051
1052/**
1053 * memory_group_register_static() - Register a static memory group.
1054 * @nid: The node id.
1055 * @max_pages: The maximum number of pages we'll have in this static memory
1056 *	       group.
1057 *
1058 * Register a new static memory group and return the memory group id.
1059 * All memory in the group belongs to a single unit, such as a DIMM. All
1060 * memory belonging to a static memory group is added in one go to be removed
1061 * in one go -- it's static.
1062 *
1063 * Returns an error if out of memory, if the node id is invalid, if no new
1064 * memory groups can be registered, or if max_pages is invalid (0). Otherwise,
1065 * returns the new memory group id.
1066 */
1067int memory_group_register_static(int nid, unsigned long max_pages)
1068{
1069	struct memory_group group = {
1070		.nid = nid,
1071		.s = {
1072			.max_pages = max_pages,
1073		},
1074	};
1075
1076	if (!max_pages)
1077		return -EINVAL;
1078	return memory_group_register(group);
1079}
1080EXPORT_SYMBOL_GPL(memory_group_register_static);
1081
1082/**
1083 * memory_group_register_dynamic() - Register a dynamic memory group.
1084 * @nid: The node id.
1085 * @unit_pages: Unit in pages in which is memory added/removed in this dynamic
1086 *		memory group.
1087 *
1088 * Register a new dynamic memory group and return the memory group id.
1089 * Memory within a dynamic memory group is added/removed dynamically
1090 * in unit_pages.
1091 *
1092 * Returns an error if out of memory, if the node id is invalid, if no new
1093 * memory groups can be registered, or if unit_pages is invalid (0, not a
1094 * power of two, smaller than a single memory block). Otherwise, returns the
1095 * new memory group id.
1096 */
1097int memory_group_register_dynamic(int nid, unsigned long unit_pages)
1098{
1099	struct memory_group group = {
1100		.nid = nid,
1101		.is_dynamic = true,
1102		.d = {
1103			.unit_pages = unit_pages,
1104		},
1105	};
1106
1107	if (!unit_pages || !is_power_of_2(unit_pages) ||
1108	    unit_pages < PHYS_PFN(memory_block_size_bytes()))
1109		return -EINVAL;
1110	return memory_group_register(group);
1111}
1112EXPORT_SYMBOL_GPL(memory_group_register_dynamic);
1113
1114/**
1115 * memory_group_unregister() - Unregister a memory group.
1116 * @mgid: the memory group id
1117 *
1118 * Unregister a memory group. If any memory block still belongs to this
1119 * memory group, unregistering will fail.
1120 *
1121 * Returns -EINVAL if the memory group id is invalid, returns -EBUSY if some
1122 * memory blocks still belong to this memory group and returns 0 if
1123 * unregistering succeeded.
1124 */
1125int memory_group_unregister(int mgid)
1126{
1127	struct memory_group *group;
1128
1129	if (mgid < 0)
1130		return -EINVAL;
1131
1132	group = xa_load(&memory_groups, mgid);
1133	if (!group)
1134		return -EINVAL;
1135	if (!list_empty(&group->memory_blocks))
1136		return -EBUSY;
1137	xa_erase(&memory_groups, mgid);
1138	kfree(group);
1139	return 0;
1140}
1141EXPORT_SYMBOL_GPL(memory_group_unregister);
1142
1143/*
1144 * This is an internal helper only to be used in core memory hotplug code to
1145 * lookup a memory group. We don't care about locking, as we don't expect a
1146 * memory group to get unregistered while adding memory to it -- because
1147 * the group and the memory is managed by the same driver.
1148 */
1149struct memory_group *memory_group_find_by_id(int mgid)
1150{
1151	return xa_load(&memory_groups, mgid);
1152}
1153
1154/*
1155 * This is an internal helper only to be used in core memory hotplug code to
1156 * walk all dynamic memory groups excluding a given memory group, either
1157 * belonging to a specific node, or belonging to any node.
1158 */
1159int walk_dynamic_memory_groups(int nid, walk_memory_groups_func_t func,
1160			       struct memory_group *excluded, void *arg)
1161{
1162	struct memory_group *group;
1163	unsigned long index;
1164	int ret = 0;
1165
1166	xa_for_each_marked(&memory_groups, index, group,
1167			   MEMORY_GROUP_MARK_DYNAMIC) {
1168		if (group == excluded)
1169			continue;
1170#ifdef CONFIG_NUMA
1171		if (nid != NUMA_NO_NODE && group->nid != nid)
1172			continue;
1173#endif /* CONFIG_NUMA */
1174		ret = func(group, arg);
1175		if (ret)
1176			break;
1177	}
1178	return ret;
1179}
1180
1181#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
1182void memblk_nr_poison_inc(unsigned long pfn)
1183{
1184	const unsigned long block_id = pfn_to_block_id(pfn);
1185	struct memory_block *mem = find_memory_block_by_id(block_id);
1186
1187	if (mem)
1188		atomic_long_inc(&mem->nr_hwpoison);
1189}
1190
1191void memblk_nr_poison_sub(unsigned long pfn, long i)
1192{
1193	const unsigned long block_id = pfn_to_block_id(pfn);
1194	struct memory_block *mem = find_memory_block_by_id(block_id);
1195
1196	if (mem)
1197		atomic_long_sub(i, &mem->nr_hwpoison);
1198}
1199
1200static unsigned long memblk_nr_poison(struct memory_block *mem)
1201{
1202	return atomic_long_read(&mem->nr_hwpoison);
1203}
1204#endif
v6.9.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Memory subsystem support
   4 *
   5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
   6 *            Dave Hansen <haveblue@us.ibm.com>
   7 *
   8 * This file provides the necessary infrastructure to represent
   9 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
  10 * All arch-independent code that assumes MEMORY_HOTPLUG requires
  11 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
  12 */
  13
  14#include <linux/module.h>
  15#include <linux/init.h>
  16#include <linux/topology.h>
  17#include <linux/capability.h>
  18#include <linux/device.h>
  19#include <linux/memory.h>
  20#include <linux/memory_hotplug.h>
  21#include <linux/mm.h>
  22#include <linux/stat.h>
  23#include <linux/slab.h>
  24#include <linux/xarray.h>
  25
  26#include <linux/atomic.h>
  27#include <linux/uaccess.h>
  28
  29#define MEMORY_CLASS_NAME	"memory"
  30
  31static const char *const online_type_to_str[] = {
  32	[MMOP_OFFLINE] = "offline",
  33	[MMOP_ONLINE] = "online",
  34	[MMOP_ONLINE_KERNEL] = "online_kernel",
  35	[MMOP_ONLINE_MOVABLE] = "online_movable",
  36};
  37
  38int mhp_online_type_from_str(const char *str)
  39{
  40	int i;
  41
  42	for (i = 0; i < ARRAY_SIZE(online_type_to_str); i++) {
  43		if (sysfs_streq(str, online_type_to_str[i]))
  44			return i;
  45	}
  46	return -EINVAL;
  47}
  48
  49#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
  50
  51static int sections_per_block;
  52
  53static inline unsigned long memory_block_id(unsigned long section_nr)
  54{
  55	return section_nr / sections_per_block;
  56}
  57
  58static inline unsigned long pfn_to_block_id(unsigned long pfn)
  59{
  60	return memory_block_id(pfn_to_section_nr(pfn));
  61}
  62
  63static inline unsigned long phys_to_block_id(unsigned long phys)
  64{
  65	return pfn_to_block_id(PFN_DOWN(phys));
  66}
  67
  68static int memory_subsys_online(struct device *dev);
  69static int memory_subsys_offline(struct device *dev);
  70
  71static const struct bus_type memory_subsys = {
  72	.name = MEMORY_CLASS_NAME,
  73	.dev_name = MEMORY_CLASS_NAME,
  74	.online = memory_subsys_online,
  75	.offline = memory_subsys_offline,
  76};
  77
  78/*
  79 * Memory blocks are cached in a local radix tree to avoid
  80 * a costly linear search for the corresponding device on
  81 * the subsystem bus.
  82 */
  83static DEFINE_XARRAY(memory_blocks);
  84
  85/*
  86 * Memory groups, indexed by memory group id (mgid).
  87 */
  88static DEFINE_XARRAY_FLAGS(memory_groups, XA_FLAGS_ALLOC);
  89#define MEMORY_GROUP_MARK_DYNAMIC	XA_MARK_1
  90
  91static BLOCKING_NOTIFIER_HEAD(memory_chain);
  92
  93int register_memory_notifier(struct notifier_block *nb)
  94{
  95	return blocking_notifier_chain_register(&memory_chain, nb);
  96}
  97EXPORT_SYMBOL(register_memory_notifier);
  98
  99void unregister_memory_notifier(struct notifier_block *nb)
 100{
 101	blocking_notifier_chain_unregister(&memory_chain, nb);
 102}
 103EXPORT_SYMBOL(unregister_memory_notifier);
 104
 105static void memory_block_release(struct device *dev)
 106{
 107	struct memory_block *mem = to_memory_block(dev);
 108	/* Verify that the altmap is freed */
 109	WARN_ON(mem->altmap);
 110	kfree(mem);
 111}
 112
 113unsigned long __weak memory_block_size_bytes(void)
 114{
 115	return MIN_MEMORY_BLOCK_SIZE;
 116}
 117EXPORT_SYMBOL_GPL(memory_block_size_bytes);
 118
 119/* Show the memory block ID, relative to the memory block size */
 
 
 120static ssize_t phys_index_show(struct device *dev,
 121			       struct device_attribute *attr, char *buf)
 122{
 123	struct memory_block *mem = to_memory_block(dev);
 
 
 
 124
 125	return sysfs_emit(buf, "%08lx\n", memory_block_id(mem->start_section_nr));
 126}
 127
 128/*
 129 * Legacy interface that we cannot remove. Always indicate "removable"
 130 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
 131 */
 132static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
 133			      char *buf)
 134{
 135	return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
 136}
 137
 138/*
 139 * online, offline, going offline, etc.
 140 */
 141static ssize_t state_show(struct device *dev, struct device_attribute *attr,
 142			  char *buf)
 143{
 144	struct memory_block *mem = to_memory_block(dev);
 145	const char *output;
 146
 147	/*
 148	 * We can probably put these states in a nice little array
 149	 * so that they're not open-coded
 150	 */
 151	switch (mem->state) {
 152	case MEM_ONLINE:
 153		output = "online";
 154		break;
 155	case MEM_OFFLINE:
 156		output = "offline";
 157		break;
 158	case MEM_GOING_OFFLINE:
 159		output = "going-offline";
 160		break;
 161	default:
 162		WARN_ON(1);
 163		return sysfs_emit(buf, "ERROR-UNKNOWN-%ld\n", mem->state);
 164	}
 165
 166	return sysfs_emit(buf, "%s\n", output);
 167}
 168
 169int memory_notify(unsigned long val, void *v)
 170{
 171	return blocking_notifier_call_chain(&memory_chain, val, v);
 172}
 173
 174#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
 175static unsigned long memblk_nr_poison(struct memory_block *mem);
 176#else
 177static inline unsigned long memblk_nr_poison(struct memory_block *mem)
 178{
 179	return 0;
 180}
 181#endif
 182
 183/*
 184 * Must acquire mem_hotplug_lock in write mode.
 185 */
 186static int memory_block_online(struct memory_block *mem)
 187{
 188	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 189	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 190	unsigned long nr_vmemmap_pages = 0;
 191	struct memory_notify arg;
 192	struct zone *zone;
 193	int ret;
 194
 195	if (memblk_nr_poison(mem))
 196		return -EHWPOISON;
 197
 198	zone = zone_for_pfn_range(mem->online_type, mem->nid, mem->group,
 199				  start_pfn, nr_pages);
 200
 201	/*
 202	 * Although vmemmap pages have a different lifecycle than the pages
 203	 * they describe (they remain until the memory is unplugged), doing
 204	 * their initialization and accounting at memory onlining/offlining
 205	 * stage helps to keep accounting easier to follow - e.g vmemmaps
 206	 * belong to the same zone as the memory they backed.
 207	 */
 208	if (mem->altmap)
 209		nr_vmemmap_pages = mem->altmap->free;
 210
 211	arg.altmap_start_pfn = start_pfn;
 212	arg.altmap_nr_pages = nr_vmemmap_pages;
 213	arg.start_pfn = start_pfn + nr_vmemmap_pages;
 214	arg.nr_pages = nr_pages - nr_vmemmap_pages;
 215	mem_hotplug_begin();
 216	ret = memory_notify(MEM_PREPARE_ONLINE, &arg);
 217	ret = notifier_to_errno(ret);
 218	if (ret)
 219		goto out_notifier;
 220
 221	if (nr_vmemmap_pages) {
 222		ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages,
 223						zone, mem->altmap->inaccessible);
 224		if (ret)
 225			goto out;
 226	}
 227
 228	ret = online_pages(start_pfn + nr_vmemmap_pages,
 229			   nr_pages - nr_vmemmap_pages, zone, mem->group);
 230	if (ret) {
 231		if (nr_vmemmap_pages)
 232			mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
 233		goto out;
 234	}
 235
 236	/*
 237	 * Account once onlining succeeded. If the zone was unpopulated, it is
 238	 * now already properly populated.
 239	 */
 240	if (nr_vmemmap_pages)
 241		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
 242					  nr_vmemmap_pages);
 243
 244	mem->zone = zone;
 245	mem_hotplug_done();
 246	return ret;
 247out:
 248	memory_notify(MEM_FINISH_OFFLINE, &arg);
 249out_notifier:
 250	mem_hotplug_done();
 251	return ret;
 252}
 253
 254/*
 255 * Must acquire mem_hotplug_lock in write mode.
 256 */
 257static int memory_block_offline(struct memory_block *mem)
 258{
 259	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 260	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 261	unsigned long nr_vmemmap_pages = 0;
 262	struct memory_notify arg;
 263	int ret;
 264
 265	if (!mem->zone)
 266		return -EINVAL;
 267
 268	/*
 269	 * Unaccount before offlining, such that unpopulated zone and kthreads
 270	 * can properly be torn down in offline_pages().
 271	 */
 272	if (mem->altmap)
 273		nr_vmemmap_pages = mem->altmap->free;
 274
 275	mem_hotplug_begin();
 276	if (nr_vmemmap_pages)
 277		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
 278					  -nr_vmemmap_pages);
 279
 280	ret = offline_pages(start_pfn + nr_vmemmap_pages,
 281			    nr_pages - nr_vmemmap_pages, mem->zone, mem->group);
 282	if (ret) {
 283		/* offline_pages() failed. Account back. */
 284		if (nr_vmemmap_pages)
 285			adjust_present_page_count(pfn_to_page(start_pfn),
 286						  mem->group, nr_vmemmap_pages);
 287		goto out;
 288	}
 289
 290	if (nr_vmemmap_pages)
 291		mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
 292
 293	mem->zone = NULL;
 294	arg.altmap_start_pfn = start_pfn;
 295	arg.altmap_nr_pages = nr_vmemmap_pages;
 296	arg.start_pfn = start_pfn + nr_vmemmap_pages;
 297	arg.nr_pages = nr_pages - nr_vmemmap_pages;
 298	memory_notify(MEM_FINISH_OFFLINE, &arg);
 299out:
 300	mem_hotplug_done();
 301	return ret;
 302}
 303
 304/*
 305 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 306 * OK to have direct references to sparsemem variables in here.
 307 */
 308static int
 309memory_block_action(struct memory_block *mem, unsigned long action)
 310{
 311	int ret;
 312
 313	switch (action) {
 314	case MEM_ONLINE:
 315		ret = memory_block_online(mem);
 316		break;
 317	case MEM_OFFLINE:
 318		ret = memory_block_offline(mem);
 319		break;
 320	default:
 321		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 322		     "%ld\n", __func__, mem->start_section_nr, action, action);
 323		ret = -EINVAL;
 324	}
 325
 326	return ret;
 327}
 328
 329static int memory_block_change_state(struct memory_block *mem,
 330		unsigned long to_state, unsigned long from_state_req)
 331{
 332	int ret = 0;
 333
 334	if (mem->state != from_state_req)
 335		return -EINVAL;
 336
 337	if (to_state == MEM_OFFLINE)
 338		mem->state = MEM_GOING_OFFLINE;
 339
 340	ret = memory_block_action(mem, to_state);
 341	mem->state = ret ? from_state_req : to_state;
 342
 343	return ret;
 344}
 345
 346/* The device lock serializes operations on memory_subsys_[online|offline] */
 347static int memory_subsys_online(struct device *dev)
 348{
 349	struct memory_block *mem = to_memory_block(dev);
 350	int ret;
 351
 352	if (mem->state == MEM_ONLINE)
 353		return 0;
 354
 355	/*
 356	 * When called via device_online() without configuring the online_type,
 357	 * we want to default to MMOP_ONLINE.
 358	 */
 359	if (mem->online_type == MMOP_OFFLINE)
 360		mem->online_type = MMOP_ONLINE;
 361
 362	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 363	mem->online_type = MMOP_OFFLINE;
 364
 365	return ret;
 366}
 367
 368static int memory_subsys_offline(struct device *dev)
 369{
 370	struct memory_block *mem = to_memory_block(dev);
 371
 372	if (mem->state == MEM_OFFLINE)
 373		return 0;
 374
 375	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 376}
 377
 378static ssize_t state_store(struct device *dev, struct device_attribute *attr,
 379			   const char *buf, size_t count)
 380{
 381	const int online_type = mhp_online_type_from_str(buf);
 382	struct memory_block *mem = to_memory_block(dev);
 383	int ret;
 384
 385	if (online_type < 0)
 386		return -EINVAL;
 387
 388	ret = lock_device_hotplug_sysfs();
 389	if (ret)
 390		return ret;
 391
 392	switch (online_type) {
 393	case MMOP_ONLINE_KERNEL:
 394	case MMOP_ONLINE_MOVABLE:
 395	case MMOP_ONLINE:
 396		/* mem->online_type is protected by device_hotplug_lock */
 397		mem->online_type = online_type;
 398		ret = device_online(&mem->dev);
 399		break;
 400	case MMOP_OFFLINE:
 401		ret = device_offline(&mem->dev);
 402		break;
 403	default:
 404		ret = -EINVAL; /* should never happen */
 405	}
 406
 407	unlock_device_hotplug();
 408
 409	if (ret < 0)
 410		return ret;
 411	if (ret)
 412		return -EINVAL;
 413
 414	return count;
 415}
 416
 417/*
 418 * Legacy interface that we cannot remove: s390x exposes the storage increment
 419 * covered by a memory block, allowing for identifying which memory blocks
 420 * comprise a storage increment. Since a memory block spans complete
 421 * storage increments nowadays, this interface is basically unused. Other
 422 * archs never exposed != 0.
 423 */
 424static ssize_t phys_device_show(struct device *dev,
 425				struct device_attribute *attr, char *buf)
 426{
 427	struct memory_block *mem = to_memory_block(dev);
 428	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 429
 430	return sysfs_emit(buf, "%d\n",
 431			  arch_get_memory_phys_device(start_pfn));
 432}
 433
 434#ifdef CONFIG_MEMORY_HOTREMOVE
 435static int print_allowed_zone(char *buf, int len, int nid,
 436			      struct memory_group *group,
 437			      unsigned long start_pfn, unsigned long nr_pages,
 438			      int online_type, struct zone *default_zone)
 439{
 440	struct zone *zone;
 441
 442	zone = zone_for_pfn_range(online_type, nid, group, start_pfn, nr_pages);
 443	if (zone == default_zone)
 444		return 0;
 445
 446	return sysfs_emit_at(buf, len, " %s", zone->name);
 447}
 448
 449static ssize_t valid_zones_show(struct device *dev,
 450				struct device_attribute *attr, char *buf)
 451{
 452	struct memory_block *mem = to_memory_block(dev);
 453	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 454	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 455	struct memory_group *group = mem->group;
 456	struct zone *default_zone;
 457	int nid = mem->nid;
 458	int len = 0;
 459
 460	/*
 461	 * Check the existing zone. Make sure that we do that only on the
 462	 * online nodes otherwise the page_zone is not reliable
 463	 */
 464	if (mem->state == MEM_ONLINE) {
 465		/*
 466		 * If !mem->zone, the memory block spans multiple zones and
 467		 * cannot get offlined.
 468		 */
 469		default_zone = mem->zone;
 470		if (!default_zone)
 471			return sysfs_emit(buf, "%s\n", "none");
 472		len += sysfs_emit_at(buf, len, "%s", default_zone->name);
 473		goto out;
 474	}
 475
 476	default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, group,
 477					  start_pfn, nr_pages);
 478
 479	len += sysfs_emit_at(buf, len, "%s", default_zone->name);
 480	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
 481				  MMOP_ONLINE_KERNEL, default_zone);
 482	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
 483				  MMOP_ONLINE_MOVABLE, default_zone);
 484out:
 485	len += sysfs_emit_at(buf, len, "\n");
 486	return len;
 487}
 488static DEVICE_ATTR_RO(valid_zones);
 489#endif
 490
 491static DEVICE_ATTR_RO(phys_index);
 492static DEVICE_ATTR_RW(state);
 493static DEVICE_ATTR_RO(phys_device);
 494static DEVICE_ATTR_RO(removable);
 495
 496/*
 497 * Show the memory block size (shared by all memory blocks).
 498 */
 499static ssize_t block_size_bytes_show(struct device *dev,
 500				     struct device_attribute *attr, char *buf)
 501{
 502	return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
 503}
 504
 505static DEVICE_ATTR_RO(block_size_bytes);
 506
 507/*
 508 * Memory auto online policy.
 509 */
 510
 511static ssize_t auto_online_blocks_show(struct device *dev,
 512				       struct device_attribute *attr, char *buf)
 513{
 514	return sysfs_emit(buf, "%s\n",
 515			  online_type_to_str[mhp_default_online_type]);
 516}
 517
 518static ssize_t auto_online_blocks_store(struct device *dev,
 519					struct device_attribute *attr,
 520					const char *buf, size_t count)
 521{
 522	const int online_type = mhp_online_type_from_str(buf);
 523
 524	if (online_type < 0)
 525		return -EINVAL;
 526
 527	mhp_default_online_type = online_type;
 528	return count;
 529}
 530
 531static DEVICE_ATTR_RW(auto_online_blocks);
 532
 533#ifdef CONFIG_CRASH_HOTPLUG
 534#include <linux/kexec.h>
 535static ssize_t crash_hotplug_show(struct device *dev,
 536				       struct device_attribute *attr, char *buf)
 537{
 538	return sysfs_emit(buf, "%d\n", crash_hotplug_memory_support());
 539}
 540static DEVICE_ATTR_RO(crash_hotplug);
 541#endif
 542
 543/*
 544 * Some architectures will have custom drivers to do this, and
 545 * will not need to do it from userspace.  The fake hot-add code
 546 * as well as ppc64 will do all of their discovery in userspace
 547 * and will require this interface.
 548 */
 549#ifdef CONFIG_ARCH_MEMORY_PROBE
 550static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
 551			   const char *buf, size_t count)
 552{
 553	u64 phys_addr;
 554	int nid, ret;
 555	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
 556
 557	ret = kstrtoull(buf, 0, &phys_addr);
 558	if (ret)
 559		return ret;
 560
 561	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
 562		return -EINVAL;
 563
 564	ret = lock_device_hotplug_sysfs();
 565	if (ret)
 566		return ret;
 567
 568	nid = memory_add_physaddr_to_nid(phys_addr);
 569	ret = __add_memory(nid, phys_addr,
 570			   MIN_MEMORY_BLOCK_SIZE * sections_per_block,
 571			   MHP_NONE);
 572
 573	if (ret)
 574		goto out;
 575
 576	ret = count;
 577out:
 578	unlock_device_hotplug();
 579	return ret;
 580}
 581
 582static DEVICE_ATTR_WO(probe);
 583#endif
 584
 585#ifdef CONFIG_MEMORY_FAILURE
 586/*
 587 * Support for offlining pages of memory
 588 */
 589
 590/* Soft offline a page */
 591static ssize_t soft_offline_page_store(struct device *dev,
 592				       struct device_attribute *attr,
 593				       const char *buf, size_t count)
 594{
 595	int ret;
 596	u64 pfn;
 597	if (!capable(CAP_SYS_ADMIN))
 598		return -EPERM;
 599	if (kstrtoull(buf, 0, &pfn) < 0)
 600		return -EINVAL;
 601	pfn >>= PAGE_SHIFT;
 602	ret = soft_offline_page(pfn, 0);
 603	return ret == 0 ? count : ret;
 604}
 605
 606/* Forcibly offline a page, including killing processes. */
 607static ssize_t hard_offline_page_store(struct device *dev,
 608				       struct device_attribute *attr,
 609				       const char *buf, size_t count)
 610{
 611	int ret;
 612	u64 pfn;
 613	if (!capable(CAP_SYS_ADMIN))
 614		return -EPERM;
 615	if (kstrtoull(buf, 0, &pfn) < 0)
 616		return -EINVAL;
 617	pfn >>= PAGE_SHIFT;
 618	ret = memory_failure(pfn, MF_SW_SIMULATED);
 619	if (ret == -EOPNOTSUPP)
 620		ret = 0;
 621	return ret ? ret : count;
 622}
 623
 624static DEVICE_ATTR_WO(soft_offline_page);
 625static DEVICE_ATTR_WO(hard_offline_page);
 626#endif
 627
 628/* See phys_device_show(). */
 629int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 630{
 631	return 0;
 632}
 633
 634/*
 635 * A reference for the returned memory block device is acquired.
 636 *
 637 * Called under device_hotplug_lock.
 638 */
 639static struct memory_block *find_memory_block_by_id(unsigned long block_id)
 640{
 641	struct memory_block *mem;
 642
 643	mem = xa_load(&memory_blocks, block_id);
 644	if (mem)
 645		get_device(&mem->dev);
 646	return mem;
 647}
 648
 649/*
 650 * Called under device_hotplug_lock.
 651 */
 652struct memory_block *find_memory_block(unsigned long section_nr)
 653{
 654	unsigned long block_id = memory_block_id(section_nr);
 655
 656	return find_memory_block_by_id(block_id);
 657}
 658
 659static struct attribute *memory_memblk_attrs[] = {
 660	&dev_attr_phys_index.attr,
 661	&dev_attr_state.attr,
 662	&dev_attr_phys_device.attr,
 663	&dev_attr_removable.attr,
 664#ifdef CONFIG_MEMORY_HOTREMOVE
 665	&dev_attr_valid_zones.attr,
 666#endif
 667	NULL
 668};
 669
 670static const struct attribute_group memory_memblk_attr_group = {
 671	.attrs = memory_memblk_attrs,
 672};
 673
 674static const struct attribute_group *memory_memblk_attr_groups[] = {
 675	&memory_memblk_attr_group,
 676	NULL,
 677};
 678
 679static int __add_memory_block(struct memory_block *memory)
 680{
 681	int ret;
 682
 683	memory->dev.bus = &memory_subsys;
 684	memory->dev.id = memory->start_section_nr / sections_per_block;
 685	memory->dev.release = memory_block_release;
 686	memory->dev.groups = memory_memblk_attr_groups;
 687	memory->dev.offline = memory->state == MEM_OFFLINE;
 688
 689	ret = device_register(&memory->dev);
 690	if (ret) {
 691		put_device(&memory->dev);
 692		return ret;
 693	}
 694	ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
 695			      GFP_KERNEL));
 696	if (ret)
 697		device_unregister(&memory->dev);
 698
 699	return ret;
 700}
 701
 702static struct zone *early_node_zone_for_memory_block(struct memory_block *mem,
 703						     int nid)
 704{
 705	const unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 706	const unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 707	struct zone *zone, *matching_zone = NULL;
 708	pg_data_t *pgdat = NODE_DATA(nid);
 709	int i;
 710
 711	/*
 712	 * This logic only works for early memory, when the applicable zones
 713	 * already span the memory block. We don't expect overlapping zones on
 714	 * a single node for early memory. So if we're told that some PFNs
 715	 * of a node fall into this memory block, we can assume that all node
 716	 * zones that intersect with the memory block are actually applicable.
 717	 * No need to look at the memmap.
 718	 */
 719	for (i = 0; i < MAX_NR_ZONES; i++) {
 720		zone = pgdat->node_zones + i;
 721		if (!populated_zone(zone))
 722			continue;
 723		if (!zone_intersects(zone, start_pfn, nr_pages))
 724			continue;
 725		if (!matching_zone) {
 726			matching_zone = zone;
 727			continue;
 728		}
 729		/* Spans multiple zones ... */
 730		matching_zone = NULL;
 731		break;
 732	}
 733	return matching_zone;
 734}
 735
 736#ifdef CONFIG_NUMA
 737/**
 738 * memory_block_add_nid() - Indicate that system RAM falling into this memory
 739 *			    block device (partially) belongs to the given node.
 740 * @mem: The memory block device.
 741 * @nid: The node id.
 742 * @context: The memory initialization context.
 743 *
 744 * Indicate that system RAM falling into this memory block (partially) belongs
 745 * to the given node. If the context indicates ("early") that we are adding the
 746 * node during node device subsystem initialization, this will also properly
 747 * set/adjust mem->zone based on the zone ranges of the given node.
 748 */
 749void memory_block_add_nid(struct memory_block *mem, int nid,
 750			  enum meminit_context context)
 751{
 752	if (context == MEMINIT_EARLY && mem->nid != nid) {
 753		/*
 754		 * For early memory we have to determine the zone when setting
 755		 * the node id and handle multiple nodes spanning a single
 756		 * memory block by indicate via zone == NULL that we're not
 757		 * dealing with a single zone. So if we're setting the node id
 758		 * the first time, determine if there is a single zone. If we're
 759		 * setting the node id a second time to a different node,
 760		 * invalidate the single detected zone.
 761		 */
 762		if (mem->nid == NUMA_NO_NODE)
 763			mem->zone = early_node_zone_for_memory_block(mem, nid);
 764		else
 765			mem->zone = NULL;
 766	}
 767
 768	/*
 769	 * If this memory block spans multiple nodes, we only indicate
 770	 * the last processed node. If we span multiple nodes (not applicable
 771	 * to hotplugged memory), zone == NULL will prohibit memory offlining
 772	 * and consequently unplug.
 773	 */
 774	mem->nid = nid;
 775}
 776#endif
 777
 778static int add_memory_block(unsigned long block_id, unsigned long state,
 779			    struct vmem_altmap *altmap,
 780			    struct memory_group *group)
 781{
 782	struct memory_block *mem;
 783	int ret = 0;
 784
 785	mem = find_memory_block_by_id(block_id);
 786	if (mem) {
 787		put_device(&mem->dev);
 788		return -EEXIST;
 789	}
 790	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 791	if (!mem)
 792		return -ENOMEM;
 793
 794	mem->start_section_nr = block_id * sections_per_block;
 795	mem->state = state;
 796	mem->nid = NUMA_NO_NODE;
 797	mem->altmap = altmap;
 798	INIT_LIST_HEAD(&mem->group_next);
 799
 800#ifndef CONFIG_NUMA
 801	if (state == MEM_ONLINE)
 802		/*
 803		 * MEM_ONLINE at this point implies early memory. With NUMA,
 804		 * we'll determine the zone when setting the node id via
 805		 * memory_block_add_nid(). Memory hotplug updated the zone
 806		 * manually when memory onlining/offlining succeeds.
 807		 */
 808		mem->zone = early_node_zone_for_memory_block(mem, NUMA_NO_NODE);
 809#endif /* CONFIG_NUMA */
 810
 811	ret = __add_memory_block(mem);
 812	if (ret)
 813		return ret;
 814
 815	if (group) {
 816		mem->group = group;
 817		list_add(&mem->group_next, &group->memory_blocks);
 818	}
 819
 820	return 0;
 821}
 822
 823static int __init add_boot_memory_block(unsigned long base_section_nr)
 824{
 825	int section_count = 0;
 826	unsigned long nr;
 827
 828	for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
 829	     nr++)
 830		if (present_section_nr(nr))
 831			section_count++;
 832
 833	if (section_count == 0)
 834		return 0;
 835	return add_memory_block(memory_block_id(base_section_nr),
 836				MEM_ONLINE, NULL,  NULL);
 837}
 838
 839static int add_hotplug_memory_block(unsigned long block_id,
 840				    struct vmem_altmap *altmap,
 841				    struct memory_group *group)
 842{
 843	return add_memory_block(block_id, MEM_OFFLINE, altmap, group);
 844}
 845
 846static void remove_memory_block(struct memory_block *memory)
 847{
 848	if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
 849		return;
 850
 851	WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
 852
 853	if (memory->group) {
 854		list_del(&memory->group_next);
 855		memory->group = NULL;
 856	}
 857
 858	/* drop the ref. we got via find_memory_block() */
 859	put_device(&memory->dev);
 860	device_unregister(&memory->dev);
 861}
 862
 863/*
 864 * Create memory block devices for the given memory area. Start and size
 865 * have to be aligned to memory block granularity. Memory block devices
 866 * will be initialized as offline.
 867 *
 868 * Called under device_hotplug_lock.
 869 */
 870int create_memory_block_devices(unsigned long start, unsigned long size,
 871				struct vmem_altmap *altmap,
 872				struct memory_group *group)
 873{
 874	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
 875	unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
 876	struct memory_block *mem;
 877	unsigned long block_id;
 878	int ret = 0;
 879
 880	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
 881			 !IS_ALIGNED(size, memory_block_size_bytes())))
 882		return -EINVAL;
 883
 884	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
 885		ret = add_hotplug_memory_block(block_id, altmap, group);
 886		if (ret)
 887			break;
 888	}
 889	if (ret) {
 890		end_block_id = block_id;
 891		for (block_id = start_block_id; block_id != end_block_id;
 892		     block_id++) {
 893			mem = find_memory_block_by_id(block_id);
 894			if (WARN_ON_ONCE(!mem))
 895				continue;
 896			remove_memory_block(mem);
 897		}
 898	}
 899	return ret;
 900}
 901
 902/*
 903 * Remove memory block devices for the given memory area. Start and size
 904 * have to be aligned to memory block granularity. Memory block devices
 905 * have to be offline.
 906 *
 907 * Called under device_hotplug_lock.
 908 */
 909void remove_memory_block_devices(unsigned long start, unsigned long size)
 910{
 911	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
 912	const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
 913	struct memory_block *mem;
 914	unsigned long block_id;
 915
 916	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
 917			 !IS_ALIGNED(size, memory_block_size_bytes())))
 918		return;
 919
 920	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
 921		mem = find_memory_block_by_id(block_id);
 922		if (WARN_ON_ONCE(!mem))
 923			continue;
 924		num_poisoned_pages_sub(-1UL, memblk_nr_poison(mem));
 925		unregister_memory_block_under_nodes(mem);
 926		remove_memory_block(mem);
 927	}
 928}
 929
 930static struct attribute *memory_root_attrs[] = {
 931#ifdef CONFIG_ARCH_MEMORY_PROBE
 932	&dev_attr_probe.attr,
 933#endif
 934
 935#ifdef CONFIG_MEMORY_FAILURE
 936	&dev_attr_soft_offline_page.attr,
 937	&dev_attr_hard_offline_page.attr,
 938#endif
 939
 940	&dev_attr_block_size_bytes.attr,
 941	&dev_attr_auto_online_blocks.attr,
 942#ifdef CONFIG_CRASH_HOTPLUG
 943	&dev_attr_crash_hotplug.attr,
 944#endif
 945	NULL
 946};
 947
 948static const struct attribute_group memory_root_attr_group = {
 949	.attrs = memory_root_attrs,
 950};
 951
 952static const struct attribute_group *memory_root_attr_groups[] = {
 953	&memory_root_attr_group,
 954	NULL,
 955};
 956
 957/*
 958 * Initialize the sysfs support for memory devices. At the time this function
 959 * is called, we cannot have concurrent creation/deletion of memory block
 960 * devices, the device_hotplug_lock is not needed.
 961 */
 962void __init memory_dev_init(void)
 963{
 964	int ret;
 965	unsigned long block_sz, nr;
 966
 967	/* Validate the configured memory block size */
 968	block_sz = memory_block_size_bytes();
 969	if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
 970		panic("Memory block size not suitable: 0x%lx\n", block_sz);
 971	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 972
 973	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
 974	if (ret)
 975		panic("%s() failed to register subsystem: %d\n", __func__, ret);
 976
 977	/*
 978	 * Create entries for memory sections that were found
 979	 * during boot and have been initialized
 980	 */
 981	for (nr = 0; nr <= __highest_present_section_nr;
 982	     nr += sections_per_block) {
 983		ret = add_boot_memory_block(nr);
 984		if (ret)
 985			panic("%s() failed to add memory block: %d\n", __func__,
 986			      ret);
 987	}
 988}
 989
 990/**
 991 * walk_memory_blocks - walk through all present memory blocks overlapped
 992 *			by the range [start, start + size)
 993 *
 994 * @start: start address of the memory range
 995 * @size: size of the memory range
 996 * @arg: argument passed to func
 997 * @func: callback for each memory section walked
 998 *
 999 * This function walks through all present memory blocks overlapped by the
1000 * range [start, start + size), calling func on each memory block.
1001 *
1002 * In case func() returns an error, walking is aborted and the error is
1003 * returned.
1004 *
1005 * Called under device_hotplug_lock.
1006 */
1007int walk_memory_blocks(unsigned long start, unsigned long size,
1008		       void *arg, walk_memory_blocks_func_t func)
1009{
1010	const unsigned long start_block_id = phys_to_block_id(start);
1011	const unsigned long end_block_id = phys_to_block_id(start + size - 1);
1012	struct memory_block *mem;
1013	unsigned long block_id;
1014	int ret = 0;
1015
1016	if (!size)
1017		return 0;
1018
1019	for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
1020		mem = find_memory_block_by_id(block_id);
1021		if (!mem)
1022			continue;
1023
1024		ret = func(mem, arg);
1025		put_device(&mem->dev);
1026		if (ret)
1027			break;
1028	}
1029	return ret;
1030}
1031
1032struct for_each_memory_block_cb_data {
1033	walk_memory_blocks_func_t func;
1034	void *arg;
1035};
1036
1037static int for_each_memory_block_cb(struct device *dev, void *data)
1038{
1039	struct memory_block *mem = to_memory_block(dev);
1040	struct for_each_memory_block_cb_data *cb_data = data;
1041
1042	return cb_data->func(mem, cb_data->arg);
1043}
1044
1045/**
1046 * for_each_memory_block - walk through all present memory blocks
1047 *
1048 * @arg: argument passed to func
1049 * @func: callback for each memory block walked
1050 *
1051 * This function walks through all present memory blocks, calling func on
1052 * each memory block.
1053 *
1054 * In case func() returns an error, walking is aborted and the error is
1055 * returned.
1056 */
1057int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
1058{
1059	struct for_each_memory_block_cb_data cb_data = {
1060		.func = func,
1061		.arg = arg,
1062	};
1063
1064	return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
1065				for_each_memory_block_cb);
1066}
1067
1068/*
1069 * This is an internal helper to unify allocation and initialization of
1070 * memory groups. Note that the passed memory group will be copied to a
1071 * dynamically allocated memory group. After this call, the passed
1072 * memory group should no longer be used.
1073 */
1074static int memory_group_register(struct memory_group group)
1075{
1076	struct memory_group *new_group;
1077	uint32_t mgid;
1078	int ret;
1079
1080	if (!node_possible(group.nid))
1081		return -EINVAL;
1082
1083	new_group = kzalloc(sizeof(group), GFP_KERNEL);
1084	if (!new_group)
1085		return -ENOMEM;
1086	*new_group = group;
1087	INIT_LIST_HEAD(&new_group->memory_blocks);
1088
1089	ret = xa_alloc(&memory_groups, &mgid, new_group, xa_limit_31b,
1090		       GFP_KERNEL);
1091	if (ret) {
1092		kfree(new_group);
1093		return ret;
1094	} else if (group.is_dynamic) {
1095		xa_set_mark(&memory_groups, mgid, MEMORY_GROUP_MARK_DYNAMIC);
1096	}
1097	return mgid;
1098}
1099
1100/**
1101 * memory_group_register_static() - Register a static memory group.
1102 * @nid: The node id.
1103 * @max_pages: The maximum number of pages we'll have in this static memory
1104 *	       group.
1105 *
1106 * Register a new static memory group and return the memory group id.
1107 * All memory in the group belongs to a single unit, such as a DIMM. All
1108 * memory belonging to a static memory group is added in one go to be removed
1109 * in one go -- it's static.
1110 *
1111 * Returns an error if out of memory, if the node id is invalid, if no new
1112 * memory groups can be registered, or if max_pages is invalid (0). Otherwise,
1113 * returns the new memory group id.
1114 */
1115int memory_group_register_static(int nid, unsigned long max_pages)
1116{
1117	struct memory_group group = {
1118		.nid = nid,
1119		.s = {
1120			.max_pages = max_pages,
1121		},
1122	};
1123
1124	if (!max_pages)
1125		return -EINVAL;
1126	return memory_group_register(group);
1127}
1128EXPORT_SYMBOL_GPL(memory_group_register_static);
1129
1130/**
1131 * memory_group_register_dynamic() - Register a dynamic memory group.
1132 * @nid: The node id.
1133 * @unit_pages: Unit in pages in which is memory added/removed in this dynamic
1134 *		memory group.
1135 *
1136 * Register a new dynamic memory group and return the memory group id.
1137 * Memory within a dynamic memory group is added/removed dynamically
1138 * in unit_pages.
1139 *
1140 * Returns an error if out of memory, if the node id is invalid, if no new
1141 * memory groups can be registered, or if unit_pages is invalid (0, not a
1142 * power of two, smaller than a single memory block). Otherwise, returns the
1143 * new memory group id.
1144 */
1145int memory_group_register_dynamic(int nid, unsigned long unit_pages)
1146{
1147	struct memory_group group = {
1148		.nid = nid,
1149		.is_dynamic = true,
1150		.d = {
1151			.unit_pages = unit_pages,
1152		},
1153	};
1154
1155	if (!unit_pages || !is_power_of_2(unit_pages) ||
1156	    unit_pages < PHYS_PFN(memory_block_size_bytes()))
1157		return -EINVAL;
1158	return memory_group_register(group);
1159}
1160EXPORT_SYMBOL_GPL(memory_group_register_dynamic);
1161
1162/**
1163 * memory_group_unregister() - Unregister a memory group.
1164 * @mgid: the memory group id
1165 *
1166 * Unregister a memory group. If any memory block still belongs to this
1167 * memory group, unregistering will fail.
1168 *
1169 * Returns -EINVAL if the memory group id is invalid, returns -EBUSY if some
1170 * memory blocks still belong to this memory group and returns 0 if
1171 * unregistering succeeded.
1172 */
1173int memory_group_unregister(int mgid)
1174{
1175	struct memory_group *group;
1176
1177	if (mgid < 0)
1178		return -EINVAL;
1179
1180	group = xa_load(&memory_groups, mgid);
1181	if (!group)
1182		return -EINVAL;
1183	if (!list_empty(&group->memory_blocks))
1184		return -EBUSY;
1185	xa_erase(&memory_groups, mgid);
1186	kfree(group);
1187	return 0;
1188}
1189EXPORT_SYMBOL_GPL(memory_group_unregister);
1190
1191/*
1192 * This is an internal helper only to be used in core memory hotplug code to
1193 * lookup a memory group. We don't care about locking, as we don't expect a
1194 * memory group to get unregistered while adding memory to it -- because
1195 * the group and the memory is managed by the same driver.
1196 */
1197struct memory_group *memory_group_find_by_id(int mgid)
1198{
1199	return xa_load(&memory_groups, mgid);
1200}
1201
1202/*
1203 * This is an internal helper only to be used in core memory hotplug code to
1204 * walk all dynamic memory groups excluding a given memory group, either
1205 * belonging to a specific node, or belonging to any node.
1206 */
1207int walk_dynamic_memory_groups(int nid, walk_memory_groups_func_t func,
1208			       struct memory_group *excluded, void *arg)
1209{
1210	struct memory_group *group;
1211	unsigned long index;
1212	int ret = 0;
1213
1214	xa_for_each_marked(&memory_groups, index, group,
1215			   MEMORY_GROUP_MARK_DYNAMIC) {
1216		if (group == excluded)
1217			continue;
1218#ifdef CONFIG_NUMA
1219		if (nid != NUMA_NO_NODE && group->nid != nid)
1220			continue;
1221#endif /* CONFIG_NUMA */
1222		ret = func(group, arg);
1223		if (ret)
1224			break;
1225	}
1226	return ret;
1227}
1228
1229#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
1230void memblk_nr_poison_inc(unsigned long pfn)
1231{
1232	const unsigned long block_id = pfn_to_block_id(pfn);
1233	struct memory_block *mem = find_memory_block_by_id(block_id);
1234
1235	if (mem)
1236		atomic_long_inc(&mem->nr_hwpoison);
1237}
1238
1239void memblk_nr_poison_sub(unsigned long pfn, long i)
1240{
1241	const unsigned long block_id = pfn_to_block_id(pfn);
1242	struct memory_block *mem = find_memory_block_by_id(block_id);
1243
1244	if (mem)
1245		atomic_long_sub(i, &mem->nr_hwpoison);
1246}
1247
1248static unsigned long memblk_nr_poison(struct memory_block *mem)
1249{
1250	return atomic_long_read(&mem->nr_hwpoison);
1251}
1252#endif