memory.c

   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#include <linux/export.h>
  26
  27#include <linux/atomic.h>
  28#include <linux/uaccess.h>
  29
  30#define MEMORY_CLASS_NAME	"memory"
  31
  32static const char *const online_type_to_str[] = {
  33	[MMOP_OFFLINE] = "offline",
  34	[MMOP_ONLINE] = "online",
  35	[MMOP_ONLINE_KERNEL] = "online_kernel",
  36	[MMOP_ONLINE_MOVABLE] = "online_movable",
  37};
  38
  39int mhp_online_type_from_str(const char *str)
  40{
  41	int i;
  42
  43	for (i = 0; i < ARRAY_SIZE(online_type_to_str); i++) {
  44		if (sysfs_streq(str, online_type_to_str[i]))
  45			return i;
  46	}
  47	return -EINVAL;
  48}
  49
  50#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
  51
  52int sections_per_block;
  53EXPORT_SYMBOL(sections_per_block);
  54
  55static int memory_subsys_online(struct device *dev);
  56static int memory_subsys_offline(struct device *dev);
  57
  58static const struct bus_type memory_subsys = {
  59	.name = MEMORY_CLASS_NAME,
  60	.dev_name = MEMORY_CLASS_NAME,
  61	.online = memory_subsys_online,
  62	.offline = memory_subsys_offline,
  63};
  64
  65/*
  66 * Memory blocks are cached in a local radix tree to avoid
  67 * a costly linear search for the corresponding device on
  68 * the subsystem bus.
  69 */
  70static DEFINE_XARRAY(memory_blocks);
  71
  72/*
  73 * Memory groups, indexed by memory group id (mgid).
  74 */
  75static DEFINE_XARRAY_FLAGS(memory_groups, XA_FLAGS_ALLOC);
  76#define MEMORY_GROUP_MARK_DYNAMIC	XA_MARK_1
  77
  78static BLOCKING_NOTIFIER_HEAD(memory_chain);
  79
  80int register_memory_notifier(struct notifier_block *nb)
  81{
  82	return blocking_notifier_chain_register(&memory_chain, nb);
  83}
  84EXPORT_SYMBOL(register_memory_notifier);
  85
  86void unregister_memory_notifier(struct notifier_block *nb)
  87{
  88	blocking_notifier_chain_unregister(&memory_chain, nb);
  89}
  90EXPORT_SYMBOL(unregister_memory_notifier);
  91
  92static void memory_block_release(struct device *dev)
  93{
  94	struct memory_block *mem = to_memory_block(dev);
  95	/* Verify that the altmap is freed */
  96	WARN_ON(mem->altmap);
  97	kfree(mem);
  98}
  99
 100
 101/* Max block size to be set by memory_block_advise_max_size */
 102static unsigned long memory_block_advised_size;
 103static bool memory_block_advised_size_queried;
 104
 105/**
 106 * memory_block_advise_max_size() - advise memory hotplug on the max suggested
 107 *				    block size, usually for alignment.
 108 * @size: suggestion for maximum block size. must be aligned on power of 2.
 109 *
 110 * Early boot software (pre-allocator init) may advise archs on the max block
 111 * size. This value can only decrease after initialization, as the intent is
 112 * to identify the largest supported alignment for all sources.
 113 *
 114 * Use of this value is arch-defined, as is min/max block size.
 115 *
 116 * Return: 0 on success
 117 *	   -EINVAL if size is 0 or not pow2 aligned
 118 *	   -EBUSY if value has already been probed
 119 */
 120int __init memory_block_advise_max_size(unsigned long size)
 121{
 122	if (!size || !is_power_of_2(size))
 123		return -EINVAL;
 124
 125	if (memory_block_advised_size_queried)
 126		return -EBUSY;
 127
 128	if (memory_block_advised_size)
 129		memory_block_advised_size = min(memory_block_advised_size, size);
 130	else
 131		memory_block_advised_size = size;
 132
 133	return 0;
 134}
 135
 136/**
 137 * memory_block_advised_max_size() - query advised max hotplug block size.
 138 *
 139 * After the first call, the value can never change. Callers looking for the
 140 * actual block size should use memory_block_size_bytes. This interface is
 141 * intended for use by arch-init when initializing the hotplug block size.
 142 *
 143 * Return: advised size in bytes, or 0 if never set.
 144 */
 145unsigned long memory_block_advised_max_size(void)
 146{
 147	memory_block_advised_size_queried = true;
 148	return memory_block_advised_size;
 149}
 150
 151unsigned long __weak memory_block_size_bytes(void)
 152{
 153	return MIN_MEMORY_BLOCK_SIZE;
 154}
 155EXPORT_SYMBOL_GPL(memory_block_size_bytes);
 156
 157/* Show the memory block ID, relative to the memory block size */
 158static ssize_t phys_index_show(struct device *dev,
 159			       struct device_attribute *attr, char *buf)
 160{
 161	struct memory_block *mem = to_memory_block(dev);
 162
 163	return sysfs_emit(buf, "%08lx\n", memory_block_id(mem->start_section_nr));
 164}
 165
 166/*
 167 * Legacy interface that we cannot remove. Always indicate "removable"
 168 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
 169 */
 170static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
 171			      char *buf)
 172{
 173	return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
 174}
 175
 176/*
 177 * online, offline, going offline, etc.
 178 */
 179static ssize_t state_show(struct device *dev, struct device_attribute *attr,
 180			  char *buf)
 181{
 182	struct memory_block *mem = to_memory_block(dev);
 183	const char *output;
 184
 185	/*
 186	 * We can probably put these states in a nice little array
 187	 * so that they're not open-coded
 188	 */
 189	switch (mem->state) {
 190	case MEM_ONLINE:
 191		output = "online";
 192		break;
 193	case MEM_OFFLINE:
 194		output = "offline";
 195		break;
 196	case MEM_GOING_OFFLINE:
 197		output = "going-offline";
 198		break;
 199	default:
 200		WARN_ON(1);
 201		return sysfs_emit(buf, "ERROR-UNKNOWN-%d\n", mem->state);
 202	}
 203
 204	return sysfs_emit(buf, "%s\n", output);
 205}
 206
 207int memory_notify(enum memory_block_state state, void *v)
 208{
 209	return blocking_notifier_call_chain(&memory_chain, state, v);
 210}
 211
 212#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
 213static unsigned long memblk_nr_poison(struct memory_block *mem);
 214#else
 215static inline unsigned long memblk_nr_poison(struct memory_block *mem)
 216{
 217	return 0;
 218}
 219#endif
 220
 221/*
 222 * Must acquire mem_hotplug_lock in write mode.
 223 */
 224static int memory_block_online(struct memory_block *mem)
 225{
 226	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 227	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 228	unsigned long nr_vmemmap_pages = 0;
 229	struct zone *zone;
 230	int ret;
 231
 232	if (memblk_nr_poison(mem))
 233		return -EHWPOISON;
 234
 235	zone = zone_for_pfn_range(mem->online_type, mem->nid, mem->group,
 236				  start_pfn, nr_pages);
 237
 238	/*
 239	 * Although vmemmap pages have a different lifecycle than the pages
 240	 * they describe (they remain until the memory is unplugged), doing
 241	 * their initialization and accounting at memory onlining/offlining
 242	 * stage helps to keep accounting easier to follow - e.g vmemmaps
 243	 * belong to the same zone as the memory they backed.
 244	 */
 245	if (mem->altmap)
 246		nr_vmemmap_pages = mem->altmap->free;
 247
 248	mem_hotplug_begin();
 249	if (nr_vmemmap_pages) {
 250		ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
 251		if (ret)
 252			goto out;
 253	}
 254
 255	ret = online_pages(start_pfn + nr_vmemmap_pages,
 256			   nr_pages - nr_vmemmap_pages, zone, mem->group);
 257	if (ret) {
 258		if (nr_vmemmap_pages)
 259			mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
 260		goto out;
 261	}
 262
 263	/*
 264	 * Account once onlining succeeded. If the zone was unpopulated, it is
 265	 * now already properly populated.
 266	 */
 267	if (nr_vmemmap_pages)
 268		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
 269					  nr_vmemmap_pages);
 270
 271	mem->zone = zone;
 272out:
 273	mem_hotplug_done();
 274	return ret;
 275}
 276
 277/*
 278 * Must acquire mem_hotplug_lock in write mode.
 279 */
 280static int memory_block_offline(struct memory_block *mem)
 281{
 282	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 283	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 284	unsigned long nr_vmemmap_pages = 0;
 285	int ret;
 286
 287	if (!mem->zone)
 288		return -EINVAL;
 289
 290	/*
 291	 * Unaccount before offlining, such that unpopulated zone and kthreads
 292	 * can properly be torn down in offline_pages().
 293	 */
 294	if (mem->altmap)
 295		nr_vmemmap_pages = mem->altmap->free;
 296
 297	mem_hotplug_begin();
 298	if (nr_vmemmap_pages)
 299		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
 300					  -nr_vmemmap_pages);
 301
 302	ret = offline_pages(start_pfn + nr_vmemmap_pages,
 303			    nr_pages - nr_vmemmap_pages, mem->zone, mem->group);
 304	if (ret) {
 305		/* offline_pages() failed. Account back. */
 306		if (nr_vmemmap_pages)
 307			adjust_present_page_count(pfn_to_page(start_pfn),
 308						  mem->group, nr_vmemmap_pages);
 309		goto out;
 310	}
 311
 312	if (nr_vmemmap_pages)
 313		mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
 314
 315	mem->zone = NULL;
 316out:
 317	mem_hotplug_done();
 318	return ret;
 319}
 320
 321/*
 322 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 323 * OK to have direct references to sparsemem variables in here.
 324 */
 325static int
 326memory_block_action(struct memory_block *mem, unsigned long action)
 327{
 328	int ret;
 329
 330	switch (action) {
 331	case MEM_ONLINE:
 332		ret = memory_block_online(mem);
 333		break;
 334	case MEM_OFFLINE:
 335		ret = memory_block_offline(mem);
 336		break;
 337	default:
 338		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 339		     "%ld\n", __func__, mem->start_section_nr, action, action);
 340		ret = -EINVAL;
 341	}
 342
 343	return ret;
 344}
 345
 346static int memory_block_change_state(struct memory_block *mem,
 347		unsigned long to_state, unsigned long from_state_req)
 348{
 349	int ret = 0;
 350
 351	if (mem->state != from_state_req)
 352		return -EINVAL;
 353
 354	if (to_state == MEM_OFFLINE)
 355		mem->state = MEM_GOING_OFFLINE;
 356
 357	ret = memory_block_action(mem, to_state);
 358	mem->state = ret ? from_state_req : to_state;
 359
 360	return ret;
 361}
 362
 363/* The device lock serializes operations on memory_subsys_[online|offline] */
 364static int memory_subsys_online(struct device *dev)
 365{
 366	struct memory_block *mem = to_memory_block(dev);
 367	int ret;
 368
 369	if (mem->state == MEM_ONLINE)
 370		return 0;
 371
 372	/*
 373	 * When called via device_online() without configuring the online_type,
 374	 * we want to default to MMOP_ONLINE.
 375	 */
 376	if (mem->online_type == MMOP_OFFLINE)
 377		mem->online_type = MMOP_ONLINE;
 378
 379	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 380	mem->online_type = MMOP_OFFLINE;
 381
 382	return ret;
 383}
 384
 385static int memory_subsys_offline(struct device *dev)
 386{
 387	struct memory_block *mem = to_memory_block(dev);
 388
 389	if (mem->state == MEM_OFFLINE)
 390		return 0;
 391
 392	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 393}
 394
 395static ssize_t state_store(struct device *dev, struct device_attribute *attr,
 396			   const char *buf, size_t count)
 397{
 398	const int online_type = mhp_online_type_from_str(buf);
 399	struct memory_block *mem = to_memory_block(dev);
 400	int ret;
 401
 402	if (online_type < 0)
 403		return -EINVAL;
 404
 405	ret = lock_device_hotplug_sysfs();
 406	if (ret)
 407		return ret;
 408
 409	switch (online_type) {
 410	case MMOP_ONLINE_KERNEL:
 411	case MMOP_ONLINE_MOVABLE:
 412	case MMOP_ONLINE:
 413		/* mem->online_type is protected by device_hotplug_lock */
 414		mem->online_type = online_type;
 415		ret = device_online(&mem->dev);
 416		break;
 417	case MMOP_OFFLINE:
 418		ret = device_offline(&mem->dev);
 419		break;
 420	default:
 421		ret = -EINVAL; /* should never happen */
 422	}
 423
 424	unlock_device_hotplug();
 425
 426	if (ret < 0)
 427		return ret;
 428	if (ret)
 429		return -EINVAL;
 430
 431	return count;
 432}
 433
 434/*
 435 * Legacy interface that we cannot remove: s390x exposes the storage increment
 436 * covered by a memory block, allowing for identifying which memory blocks
 437 * comprise a storage increment. Since a memory block spans complete
 438 * storage increments nowadays, this interface is basically unused. Other
 439 * archs never exposed != 0.
 440 */
 441static ssize_t phys_device_show(struct device *dev,
 442				struct device_attribute *attr, char *buf)
 443{
 444	struct memory_block *mem = to_memory_block(dev);
 445	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 446
 447	return sysfs_emit(buf, "%d\n",
 448			  arch_get_memory_phys_device(start_pfn));
 449}
 450
 451#ifdef CONFIG_MEMORY_HOTREMOVE
 452static int print_allowed_zone(char *buf, int len, int nid,
 453			      struct memory_group *group,
 454			      unsigned long start_pfn, unsigned long nr_pages,
 455			      int online_type, struct zone *default_zone)
 456{
 457	struct zone *zone;
 458
 459	zone = zone_for_pfn_range(online_type, nid, group, start_pfn, nr_pages);
 460	if (zone == default_zone)
 461		return 0;
 462
 463	return sysfs_emit_at(buf, len, " %s", zone->name);
 464}
 465
 466static ssize_t valid_zones_show(struct device *dev,
 467				struct device_attribute *attr, char *buf)
 468{
 469	struct memory_block *mem = to_memory_block(dev);
 470	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 471	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 472	struct memory_group *group = mem->group;
 473	struct zone *default_zone;
 474	int nid = mem->nid;
 475	int len;
 476
 477	/*
 478	 * Check the existing zone. Make sure that we do that only on the
 479	 * online nodes otherwise the page_zone is not reliable
 480	 */
 481	if (mem->state == MEM_ONLINE) {
 482		/*
 483		 * If !mem->zone, the memory block spans multiple zones and
 484		 * cannot get offlined.
 485		 */
 486		return sysfs_emit(buf, "%s\n",
 487				  mem->zone ? mem->zone->name : "none");
 488	}
 489
 490	default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, group,
 491					  start_pfn, nr_pages);
 492
 493	len = sysfs_emit(buf, "%s", default_zone->name);
 494	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
 495				  MMOP_ONLINE_KERNEL, default_zone);
 496	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
 497				  MMOP_ONLINE_MOVABLE, default_zone);
 498	len += sysfs_emit_at(buf, len, "\n");
 499	return len;
 500}
 501static DEVICE_ATTR_RO(valid_zones);
 502#endif
 503
 504static DEVICE_ATTR_RO(phys_index);
 505static DEVICE_ATTR_RW(state);
 506static DEVICE_ATTR_RO(phys_device);
 507static DEVICE_ATTR_RO(removable);
 508
 509/*
 510 * Show the memory block size (shared by all memory blocks).
 511 */
 512static ssize_t block_size_bytes_show(struct device *dev,
 513				     struct device_attribute *attr, char *buf)
 514{
 515	return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
 516}
 517
 518static DEVICE_ATTR_RO(block_size_bytes);
 519
 520/*
 521 * Memory auto online policy.
 522 */
 523
 524static ssize_t auto_online_blocks_show(struct device *dev,
 525				       struct device_attribute *attr, char *buf)
 526{
 527	return sysfs_emit(buf, "%s\n",
 528			  online_type_to_str[mhp_get_default_online_type()]);
 529}
 530
 531static ssize_t auto_online_blocks_store(struct device *dev,
 532					struct device_attribute *attr,
 533					const char *buf, size_t count)
 534{
 535	const int online_type = mhp_online_type_from_str(buf);
 536
 537	if (online_type < 0)
 538		return -EINVAL;
 539
 540	mhp_set_default_online_type(online_type);
 541	return count;
 542}
 543
 544static DEVICE_ATTR_RW(auto_online_blocks);
 545
 546#ifdef CONFIG_CRASH_HOTPLUG
 547#include <linux/kexec.h>
 548static ssize_t crash_hotplug_show(struct device *dev,
 549				       struct device_attribute *attr, char *buf)
 550{
 551	return sysfs_emit(buf, "%d\n", crash_check_hotplug_support());
 552}
 553static DEVICE_ATTR_RO(crash_hotplug);
 554#endif
 555
 556/*
 557 * Some architectures will have custom drivers to do this, and
 558 * will not need to do it from userspace.  The fake hot-add code
 559 * as well as ppc64 will do all of their discovery in userspace
 560 * and will require this interface.
 561 */
 562#ifdef CONFIG_ARCH_MEMORY_PROBE
 563static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
 564			   const char *buf, size_t count)
 565{
 566	u64 phys_addr;
 567	int nid, ret;
 568	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
 569
 570	ret = kstrtoull(buf, 0, &phys_addr);
 571	if (ret)
 572		return ret;
 573
 574	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
 575		return -EINVAL;
 576
 577	ret = lock_device_hotplug_sysfs();
 578	if (ret)
 579		return ret;
 580
 581	nid = memory_add_physaddr_to_nid(phys_addr);
 582	ret = __add_memory(nid, phys_addr,
 583			   MIN_MEMORY_BLOCK_SIZE * sections_per_block,
 584			   MHP_NONE);
 585
 586	if (ret)
 587		goto out;
 588
 589	ret = count;
 590out:
 591	unlock_device_hotplug();
 592	return ret;
 593}
 594
 595static DEVICE_ATTR_WO(probe);
 596#endif
 597
 598#ifdef CONFIG_MEMORY_FAILURE
 599/*
 600 * Support for offlining pages of memory
 601 */
 602
 603/* Soft offline a page */
 604static ssize_t soft_offline_page_store(struct device *dev,
 605				       struct device_attribute *attr,
 606				       const char *buf, size_t count)
 607{
 608	int ret;
 609	u64 pfn;
 610	if (!capable(CAP_SYS_ADMIN))
 611		return -EPERM;
 612	if (kstrtoull(buf, 0, &pfn) < 0)
 613		return -EINVAL;
 614	pfn >>= PAGE_SHIFT;
 615	ret = soft_offline_page(pfn, 0);
 616	return ret == 0 ? count : ret;
 617}
 618
 619/* Forcibly offline a page, including killing processes. */
 620static ssize_t hard_offline_page_store(struct device *dev,
 621				       struct device_attribute *attr,
 622				       const char *buf, size_t count)
 623{
 624	int ret;
 625	u64 pfn;
 626	if (!capable(CAP_SYS_ADMIN))
 627		return -EPERM;
 628	if (kstrtoull(buf, 0, &pfn) < 0)
 629		return -EINVAL;
 630	pfn >>= PAGE_SHIFT;
 631	ret = memory_failure(pfn, MF_SW_SIMULATED);
 632	if (ret == -EOPNOTSUPP)
 633		ret = 0;
 634	return ret ? ret : count;
 635}
 636
 637static DEVICE_ATTR_WO(soft_offline_page);
 638static DEVICE_ATTR_WO(hard_offline_page);
 639#endif
 640
 641/* See phys_device_show(). */
 642int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 643{
 644	return 0;
 645}
 646
 647/*
 648 * A reference for the returned memory block device is acquired.
 649 *
 650 * Called under device_hotplug_lock.
 651 */
 652struct memory_block *find_memory_block_by_id(unsigned long block_id)
 653{
 654	struct memory_block *mem;
 655
 656	mem = xa_load(&memory_blocks, block_id);
 657	if (mem)
 658		get_device(&mem->dev);
 659	return mem;
 660}
 661
 662/*
 663 * Called under device_hotplug_lock.
 664 */
 665struct memory_block *find_memory_block(unsigned long section_nr)
 666{
 667	unsigned long block_id = memory_block_id(section_nr);
 668
 669	return find_memory_block_by_id(block_id);
 670}
 671
 672static struct attribute *memory_memblk_attrs[] = {
 673	&dev_attr_phys_index.attr,
 674	&dev_attr_state.attr,
 675	&dev_attr_phys_device.attr,
 676	&dev_attr_removable.attr,
 677#ifdef CONFIG_MEMORY_HOTREMOVE
 678	&dev_attr_valid_zones.attr,
 679#endif
 680	NULL
 681};
 682
 683static const struct attribute_group memory_memblk_attr_group = {
 684	.attrs = memory_memblk_attrs,
 685};
 686
 687static const struct attribute_group *memory_memblk_attr_groups[] = {
 688	&memory_memblk_attr_group,
 689	NULL,
 690};
 691
 692static int __add_memory_block(struct memory_block *memory)
 693{
 694	int ret;
 695
 696	memory->dev.bus = &memory_subsys;
 697	memory->dev.id = memory->start_section_nr / sections_per_block;
 698	memory->dev.release = memory_block_release;
 699	memory->dev.groups = memory_memblk_attr_groups;
 700	memory->dev.offline = memory->state == MEM_OFFLINE;
 701
 702	ret = device_register(&memory->dev);
 703	if (ret) {
 704		put_device(&memory->dev);
 705		return ret;
 706	}
 707	ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
 708			      GFP_KERNEL));
 709	if (ret)
 710		device_unregister(&memory->dev);
 711
 712	return ret;
 713}
 714
 715static struct zone *early_node_zone_for_memory_block(struct memory_block *mem,
 716						     int nid)
 717{
 718	const unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 719	const unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 720	struct zone *zone, *matching_zone = NULL;
 721	pg_data_t *pgdat = NODE_DATA(nid);
 722	int i;
 723
 724	/*
 725	 * This logic only works for early memory, when the applicable zones
 726	 * already span the memory block. We don't expect overlapping zones on
 727	 * a single node for early memory. So if we're told that some PFNs
 728	 * of a node fall into this memory block, we can assume that all node
 729	 * zones that intersect with the memory block are actually applicable.
 730	 * No need to look at the memmap.
 731	 */
 732	for (i = 0; i < MAX_NR_ZONES; i++) {
 733		zone = pgdat->node_zones + i;
 734		if (!populated_zone(zone))
 735			continue;
 736		if (!zone_intersects(zone, start_pfn, nr_pages))
 737			continue;
 738		if (!matching_zone) {
 739			matching_zone = zone;
 740			continue;
 741		}
 742		/* Spans multiple zones ... */
 743		matching_zone = NULL;
 744		break;
 745	}
 746	return matching_zone;
 747}
 748
 749#ifdef CONFIG_NUMA
 750/**
 751 * memory_block_add_nid_early() - Indicate that early system RAM falling into
 752 *				  this memory block device (partially) belongs
 753 *				  to the given node.
 754 * @mem: The memory block device.
 755 * @nid: The node id.
 756 *
 757 * Indicate that early system RAM falling into this memory block (partially)
 758 * belongs to the given node. This will also properly set/adjust mem->zone based
 759 * on the zone ranges of the given node.
 760 *
 761 * Memory hotplug handles this on memory block creation, where we can only have
 762 * a single nid span a memory block.
 763 */
 764void memory_block_add_nid_early(struct memory_block *mem, int nid)
 765{
 766	if (mem->nid != nid) {
 767		/*
 768		 * For early memory we have to determine the zone when setting
 769		 * the node id and handle multiple nodes spanning a single
 770		 * memory block by indicate via zone == NULL that we're not
 771		 * dealing with a single zone. So if we're setting the node id
 772		 * the first time, determine if there is a single zone. If we're
 773		 * setting the node id a second time to a different node,
 774		 * invalidate the single detected zone.
 775		 */
 776		if (mem->nid == NUMA_NO_NODE)
 777			mem->zone = early_node_zone_for_memory_block(mem, nid);
 778		else
 779			mem->zone = NULL;
 780		/*
 781		 * If this memory block spans multiple nodes, we only indicate
 782		 * the last processed node. If we span multiple nodes (not applicable
 783		 * to hotplugged memory), zone == NULL will prohibit memory offlining
 784		 * and consequently unplug.
 785		 */
 786		mem->nid = nid;
 787	}
 788}
 789#endif
 790
 791static int add_memory_block(unsigned long block_id, int nid, unsigned long state,
 792			    struct vmem_altmap *altmap,
 793			    struct memory_group *group)
 794{
 795	struct memory_block *mem;
 796	int ret = 0;
 797
 798	mem = find_memory_block_by_id(block_id);
 799	if (mem) {
 800		put_device(&mem->dev);
 801		return -EEXIST;
 802	}
 803	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 804	if (!mem)
 805		return -ENOMEM;
 806
 807	mem->start_section_nr = block_id * sections_per_block;
 808	mem->state = state;
 809	mem->nid = nid;
 810	mem->altmap = altmap;
 811	INIT_LIST_HEAD(&mem->group_next);
 812
 813#ifndef CONFIG_NUMA
 814	if (state == MEM_ONLINE)
 815		/*
 816		 * MEM_ONLINE at this point implies early memory. With NUMA,
 817		 * we'll determine the zone when setting the node id via
 818		 * memory_block_add_nid(). Memory hotplug updated the zone
 819		 * manually when memory onlining/offlining succeeds.
 820		 */
 821		mem->zone = early_node_zone_for_memory_block(mem, NUMA_NO_NODE);
 822#endif /* CONFIG_NUMA */
 823
 824	ret = __add_memory_block(mem);
 825	if (ret)
 826		return ret;
 827
 828	if (group) {
 829		mem->group = group;
 830		list_add(&mem->group_next, &group->memory_blocks);
 831	}
 832
 833	return 0;
 834}
 835
 836static void remove_memory_block(struct memory_block *memory)
 837{
 838	if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
 839		return;
 840
 841	WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
 842
 843	if (memory->group) {
 844		list_del(&memory->group_next);
 845		memory->group = NULL;
 846	}
 847
 848	/* drop the ref. we got via find_memory_block() */
 849	put_device(&memory->dev);
 850	device_unregister(&memory->dev);
 851}
 852
 853/*
 854 * Create memory block devices for the given memory area. Start and size
 855 * have to be aligned to memory block granularity. Memory block devices
 856 * will be initialized as offline.
 857 *
 858 * Called under device_hotplug_lock.
 859 */
 860int create_memory_block_devices(unsigned long start, unsigned long size,
 861				int nid, struct vmem_altmap *altmap,
 862				struct memory_group *group)
 863{
 864	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
 865	unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
 866	struct memory_block *mem;
 867	unsigned long block_id;
 868	int ret = 0;
 869
 870	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
 871			 !IS_ALIGNED(size, memory_block_size_bytes())))
 872		return -EINVAL;
 873
 874	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
 875		ret = add_memory_block(block_id, nid, MEM_OFFLINE, altmap, group);
 876		if (ret)
 877			break;
 878	}
 879	if (ret) {
 880		end_block_id = block_id;
 881		for (block_id = start_block_id; block_id != end_block_id;
 882		     block_id++) {
 883			mem = find_memory_block_by_id(block_id);
 884			if (WARN_ON_ONCE(!mem))
 885				continue;
 886			remove_memory_block(mem);
 887		}
 888	}
 889	return ret;
 890}
 891
 892/*
 893 * Remove memory block devices for the given memory area. Start and size
 894 * have to be aligned to memory block granularity. Memory block devices
 895 * have to be offline.
 896 *
 897 * Called under device_hotplug_lock.
 898 */
 899void remove_memory_block_devices(unsigned long start, unsigned long size)
 900{
 901	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
 902	const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
 903	struct memory_block *mem;
 904	unsigned long block_id;
 905
 906	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
 907			 !IS_ALIGNED(size, memory_block_size_bytes())))
 908		return;
 909
 910	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
 911		mem = find_memory_block_by_id(block_id);
 912		if (WARN_ON_ONCE(!mem))
 913			continue;
 914		num_poisoned_pages_sub(-1UL, memblk_nr_poison(mem));
 915		unregister_memory_block_under_nodes(mem);
 916		remove_memory_block(mem);
 917	}
 918}
 919
 920static struct attribute *memory_root_attrs[] = {
 921#ifdef CONFIG_ARCH_MEMORY_PROBE
 922	&dev_attr_probe.attr,
 923#endif
 924
 925#ifdef CONFIG_MEMORY_FAILURE
 926	&dev_attr_soft_offline_page.attr,
 927	&dev_attr_hard_offline_page.attr,
 928#endif
 929
 930	&dev_attr_block_size_bytes.attr,
 931	&dev_attr_auto_online_blocks.attr,
 932#ifdef CONFIG_CRASH_HOTPLUG
 933	&dev_attr_crash_hotplug.attr,
 934#endif
 935	NULL
 936};
 937
 938static const struct attribute_group memory_root_attr_group = {
 939	.attrs = memory_root_attrs,
 940};
 941
 942static const struct attribute_group *memory_root_attr_groups[] = {
 943	&memory_root_attr_group,
 944	NULL,
 945};
 946
 947/*
 948 * Initialize the sysfs support for memory devices. At the time this function
 949 * is called, we cannot have concurrent creation/deletion of memory block
 950 * devices, the device_hotplug_lock is not needed.
 951 */
 952void __init memory_dev_init(void)
 953{
 954	int ret;
 955	unsigned long block_sz, block_id, nr;
 956
 957	/* Validate the configured memory block size */
 958	block_sz = memory_block_size_bytes();
 959	if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
 960		panic("Memory block size not suitable: 0x%lx\n", block_sz);
 961	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 962
 963	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
 964	if (ret)
 965		panic("%s() failed to register subsystem: %d\n", __func__, ret);
 966
 967	/*
 968	 * Create entries for memory sections that were found during boot
 969	 * and have been initialized. Use @block_id to track the last
 970	 * handled block and initialize it to an invalid value (ULONG_MAX)
 971	 * to bypass the block ID matching check for the first present
 972	 * block so that it can be covered.
 973	 */
 974	block_id = ULONG_MAX;
 975	for_each_present_section_nr(0, nr) {
 976		if (block_id != ULONG_MAX && memory_block_id(nr) == block_id)
 977			continue;
 978
 979		block_id = memory_block_id(nr);
 980		ret = add_memory_block(block_id, NUMA_NO_NODE, MEM_ONLINE, NULL, NULL);
 981		if (ret) {
 982			panic("%s() failed to add memory block: %d\n",
 983			      __func__, ret);
 984		}
 985	}
 986}
 987
 988/**
 989 * walk_memory_blocks - walk through all present memory blocks overlapped
 990 *			by the range [start, start + size)
 991 *
 992 * @start: start address of the memory range
 993 * @size: size of the memory range
 994 * @arg: argument passed to func
 995 * @func: callback for each memory section walked
 996 *
 997 * This function walks through all present memory blocks overlapped by the
 998 * range [start, start + size), calling func on each memory block.
 999 *
1000 * In case func() returns an error, walking is aborted and the error is
1001 * returned.
1002 *
1003 * Called under device_hotplug_lock.
1004 */
1005int walk_memory_blocks(unsigned long start, unsigned long size,
1006		       void *arg, walk_memory_blocks_func_t func)
1007{
1008	const unsigned long start_block_id = phys_to_block_id(start);
1009	const unsigned long end_block_id = phys_to_block_id(start + size - 1);
1010	struct memory_block *mem;
1011	unsigned long block_id;
1012	int ret = 0;
1013
1014	if (!size)
1015		return 0;
1016
1017	for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
1018		mem = find_memory_block_by_id(block_id);
1019		if (!mem)
1020			continue;
1021
1022		ret = func(mem, arg);
1023		put_device(&mem->dev);
1024		if (ret)
1025			break;
1026	}
1027	return ret;
1028}
1029
1030struct for_each_memory_block_cb_data {
1031	walk_memory_blocks_func_t func;
1032	void *arg;
1033};
1034
1035static int for_each_memory_block_cb(struct device *dev, void *data)
1036{
1037	struct memory_block *mem = to_memory_block(dev);
1038	struct for_each_memory_block_cb_data *cb_data = data;
1039
1040	return cb_data->func(mem, cb_data->arg);
1041}
1042
1043/**
1044 * for_each_memory_block - walk through all present memory blocks
1045 *
1046 * @arg: argument passed to func
1047 * @func: callback for each memory block walked
1048 *
1049 * This function walks through all present memory blocks, calling func on
1050 * each memory block.
1051 *
1052 * In case func() returns an error, walking is aborted and the error is
1053 * returned.
1054 */
1055int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
1056{
1057	struct for_each_memory_block_cb_data cb_data = {
1058		.func = func,
1059		.arg = arg,
1060	};
1061
1062	return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
1063				for_each_memory_block_cb);
1064}
1065
1066/*
1067 * This is an internal helper to unify allocation and initialization of
1068 * memory groups. Note that the passed memory group will be copied to a
1069 * dynamically allocated memory group. After this call, the passed
1070 * memory group should no longer be used.
1071 */
1072static int memory_group_register(struct memory_group group)
1073{
1074	struct memory_group *new_group;
1075	uint32_t mgid;
1076	int ret;
1077
1078	if (!node_possible(group.nid))
1079		return -EINVAL;
1080
1081	new_group = kzalloc(sizeof(group), GFP_KERNEL);
1082	if (!new_group)
1083		return -ENOMEM;
1084	*new_group = group;
1085	INIT_LIST_HEAD(&new_group->memory_blocks);
1086
1087	ret = xa_alloc(&memory_groups, &mgid, new_group, xa_limit_31b,
1088		       GFP_KERNEL);
1089	if (ret) {
1090		kfree(new_group);
1091		return ret;
1092	} else if (group.is_dynamic) {
1093		xa_set_mark(&memory_groups, mgid, MEMORY_GROUP_MARK_DYNAMIC);
1094	}
1095	return mgid;
1096}
1097
1098/**
1099 * memory_group_register_static() - Register a static memory group.
1100 * @nid: The node id.
1101 * @max_pages: The maximum number of pages we'll have in this static memory
1102 *	       group.
1103 *
1104 * Register a new static memory group and return the memory group id.
1105 * All memory in the group belongs to a single unit, such as a DIMM. All
1106 * memory belonging to a static memory group is added in one go to be removed
1107 * in one go -- it's static.
1108 *
1109 * Returns an error if out of memory, if the node id is invalid, if no new
1110 * memory groups can be registered, or if max_pages is invalid (0). Otherwise,
1111 * returns the new memory group id.
1112 */
1113int memory_group_register_static(int nid, unsigned long max_pages)
1114{
1115	struct memory_group group = {
1116		.nid = nid,
1117		.s = {
1118			.max_pages = max_pages,
1119		},
1120	};
1121
1122	if (!max_pages)
1123		return -EINVAL;
1124	return memory_group_register(group);
1125}
1126EXPORT_SYMBOL_GPL(memory_group_register_static);
1127
1128/**
1129 * memory_group_register_dynamic() - Register a dynamic memory group.
1130 * @nid: The node id.
1131 * @unit_pages: Unit in pages in which is memory added/removed in this dynamic
1132 *		memory group.
1133 *
1134 * Register a new dynamic memory group and return the memory group id.
1135 * Memory within a dynamic memory group is added/removed dynamically
1136 * in unit_pages.
1137 *
1138 * Returns an error if out of memory, if the node id is invalid, if no new
1139 * memory groups can be registered, or if unit_pages is invalid (0, not a
1140 * power of two, smaller than a single memory block). Otherwise, returns the
1141 * new memory group id.
1142 */
1143int memory_group_register_dynamic(int nid, unsigned long unit_pages)
1144{
1145	struct memory_group group = {
1146		.nid = nid,
1147		.is_dynamic = true,
1148		.d = {
1149			.unit_pages = unit_pages,
1150		},
1151	};
1152
1153	if (!unit_pages || !is_power_of_2(unit_pages) ||
1154	    unit_pages < PHYS_PFN(memory_block_size_bytes()))
1155		return -EINVAL;
1156	return memory_group_register(group);
1157}
1158EXPORT_SYMBOL_GPL(memory_group_register_dynamic);
1159
1160/**
1161 * memory_group_unregister() - Unregister a memory group.
1162 * @mgid: the memory group id
1163 *
1164 * Unregister a memory group. If any memory block still belongs to this
1165 * memory group, unregistering will fail.
1166 *
1167 * Returns -EINVAL if the memory group id is invalid, returns -EBUSY if some
1168 * memory blocks still belong to this memory group and returns 0 if
1169 * unregistering succeeded.
1170 */
1171int memory_group_unregister(int mgid)
1172{
1173	struct memory_group *group;
1174
1175	if (mgid < 0)
1176		return -EINVAL;
1177
1178	group = xa_load(&memory_groups, mgid);
1179	if (!group)
1180		return -EINVAL;
1181	if (!list_empty(&group->memory_blocks))
1182		return -EBUSY;
1183	xa_erase(&memory_groups, mgid);
1184	kfree(group);
1185	return 0;
1186}
1187EXPORT_SYMBOL_GPL(memory_group_unregister);
1188
1189/*
1190 * This is an internal helper only to be used in core memory hotplug code to
1191 * lookup a memory group. We don't care about locking, as we don't expect a
1192 * memory group to get unregistered while adding memory to it -- because
1193 * the group and the memory is managed by the same driver.
1194 */
1195struct memory_group *memory_group_find_by_id(int mgid)
1196{
1197	return xa_load(&memory_groups, mgid);
1198}
1199
1200/*
1201 * This is an internal helper only to be used in core memory hotplug code to
1202 * walk all dynamic memory groups excluding a given memory group, either
1203 * belonging to a specific node, or belonging to any node.
1204 */
1205int walk_dynamic_memory_groups(int nid, walk_memory_groups_func_t func,
1206			       struct memory_group *excluded, void *arg)
1207{
1208	struct memory_group *group;
1209	unsigned long index;
1210	int ret = 0;
1211
1212	xa_for_each_marked(&memory_groups, index, group,
1213			   MEMORY_GROUP_MARK_DYNAMIC) {
1214		if (group == excluded)
1215			continue;
1216#ifdef CONFIG_NUMA
1217		if (nid != NUMA_NO_NODE && group->nid != nid)
1218			continue;
1219#endif /* CONFIG_NUMA */
1220		ret = func(group, arg);
1221		if (ret)
1222			break;
1223	}
1224	return ret;
1225}
1226
1227#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
1228void memblk_nr_poison_inc(unsigned long pfn)
1229{
1230	const unsigned long block_id = pfn_to_block_id(pfn);
1231	struct memory_block *mem = find_memory_block_by_id(block_id);
1232
1233	if (mem)
1234		atomic_long_inc(&mem->nr_hwpoison);
1235}
1236
1237void memblk_nr_poison_sub(unsigned long pfn, long i)
1238{
1239	const unsigned long block_id = pfn_to_block_id(pfn);
1240	struct memory_block *mem = find_memory_block_by_id(block_id);
1241
1242	if (mem)
1243		atomic_long_sub(i, &mem->nr_hwpoison);
1244}
1245
1246static unsigned long memblk_nr_poison(struct memory_block *mem)
1247{
1248	return atomic_long_read(&mem->nr_hwpoison);
1249}
1250#endif