Linux Audio

Check our new training course

Loading...
v6.8
   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 zone *zone;
 192	int ret;
 193
 194	if (memblk_nr_poison(mem))
 195		return -EHWPOISON;
 196
 197	zone = zone_for_pfn_range(mem->online_type, mem->nid, mem->group,
 198				  start_pfn, nr_pages);
 199
 200	/*
 201	 * Although vmemmap pages have a different lifecycle than the pages
 202	 * they describe (they remain until the memory is unplugged), doing
 203	 * their initialization and accounting at memory onlining/offlining
 204	 * stage helps to keep accounting easier to follow - e.g vmemmaps
 205	 * belong to the same zone as the memory they backed.
 206	 */
 207	if (mem->altmap)
 208		nr_vmemmap_pages = mem->altmap->free;
 209
 210	mem_hotplug_begin();
 211	if (nr_vmemmap_pages) {
 212		ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
 213		if (ret)
 214			goto out;
 215	}
 216
 217	ret = online_pages(start_pfn + nr_vmemmap_pages,
 218			   nr_pages - nr_vmemmap_pages, zone, mem->group);
 219	if (ret) {
 220		if (nr_vmemmap_pages)
 221			mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
 222		goto out;
 223	}
 224
 225	/*
 226	 * Account once onlining succeeded. If the zone was unpopulated, it is
 227	 * now already properly populated.
 228	 */
 229	if (nr_vmemmap_pages)
 230		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
 231					  nr_vmemmap_pages);
 232
 233	mem->zone = zone;
 234out:
 235	mem_hotplug_done();
 236	return ret;
 237}
 238
 239/*
 240 * Must acquire mem_hotplug_lock in write mode.
 241 */
 242static int memory_block_offline(struct memory_block *mem)
 243{
 244	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 245	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 246	unsigned long nr_vmemmap_pages = 0;
 
 247	int ret;
 248
 249	if (!mem->zone)
 250		return -EINVAL;
 251
 252	/*
 253	 * Unaccount before offlining, such that unpopulated zone and kthreads
 254	 * can properly be torn down in offline_pages().
 255	 */
 256	if (mem->altmap)
 257		nr_vmemmap_pages = mem->altmap->free;
 258
 259	mem_hotplug_begin();
 260	if (nr_vmemmap_pages)
 261		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
 262					  -nr_vmemmap_pages);
 263
 264	ret = offline_pages(start_pfn + nr_vmemmap_pages,
 265			    nr_pages - nr_vmemmap_pages, mem->zone, mem->group);
 266	if (ret) {
 267		/* offline_pages() failed. Account back. */
 268		if (nr_vmemmap_pages)
 269			adjust_present_page_count(pfn_to_page(start_pfn),
 270						  mem->group, nr_vmemmap_pages);
 271		goto out;
 272	}
 273
 274	if (nr_vmemmap_pages)
 275		mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
 276
 277	mem->zone = NULL;
 278out:
 279	mem_hotplug_done();
 280	return ret;
 281}
 282
 283/*
 284 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 285 * OK to have direct references to sparsemem variables in here.
 286 */
 287static int
 288memory_block_action(struct memory_block *mem, unsigned long action)
 289{
 290	int ret;
 291
 292	switch (action) {
 293	case MEM_ONLINE:
 294		ret = memory_block_online(mem);
 295		break;
 296	case MEM_OFFLINE:
 297		ret = memory_block_offline(mem);
 298		break;
 299	default:
 300		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 301		     "%ld\n", __func__, mem->start_section_nr, action, action);
 302		ret = -EINVAL;
 303	}
 304
 305	return ret;
 306}
 307
 308static int memory_block_change_state(struct memory_block *mem,
 309		unsigned long to_state, unsigned long from_state_req)
 310{
 311	int ret = 0;
 312
 313	if (mem->state != from_state_req)
 314		return -EINVAL;
 315
 316	if (to_state == MEM_OFFLINE)
 317		mem->state = MEM_GOING_OFFLINE;
 318
 319	ret = memory_block_action(mem, to_state);
 320	mem->state = ret ? from_state_req : to_state;
 321
 322	return ret;
 323}
 324
 325/* The device lock serializes operations on memory_subsys_[online|offline] */
 326static int memory_subsys_online(struct device *dev)
 327{
 328	struct memory_block *mem = to_memory_block(dev);
 329	int ret;
 330
 331	if (mem->state == MEM_ONLINE)
 332		return 0;
 333
 334	/*
 335	 * When called via device_online() without configuring the online_type,
 336	 * we want to default to MMOP_ONLINE.
 337	 */
 338	if (mem->online_type == MMOP_OFFLINE)
 339		mem->online_type = MMOP_ONLINE;
 340
 341	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 342	mem->online_type = MMOP_OFFLINE;
 343
 344	return ret;
 345}
 346
 347static int memory_subsys_offline(struct device *dev)
 348{
 349	struct memory_block *mem = to_memory_block(dev);
 350
 351	if (mem->state == MEM_OFFLINE)
 352		return 0;
 353
 354	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 355}
 356
 357static ssize_t state_store(struct device *dev, struct device_attribute *attr,
 358			   const char *buf, size_t count)
 359{
 360	const int online_type = mhp_online_type_from_str(buf);
 361	struct memory_block *mem = to_memory_block(dev);
 362	int ret;
 363
 364	if (online_type < 0)
 365		return -EINVAL;
 366
 367	ret = lock_device_hotplug_sysfs();
 368	if (ret)
 369		return ret;
 370
 371	switch (online_type) {
 372	case MMOP_ONLINE_KERNEL:
 373	case MMOP_ONLINE_MOVABLE:
 374	case MMOP_ONLINE:
 375		/* mem->online_type is protected by device_hotplug_lock */
 376		mem->online_type = online_type;
 377		ret = device_online(&mem->dev);
 378		break;
 379	case MMOP_OFFLINE:
 380		ret = device_offline(&mem->dev);
 381		break;
 382	default:
 383		ret = -EINVAL; /* should never happen */
 384	}
 385
 386	unlock_device_hotplug();
 387
 388	if (ret < 0)
 389		return ret;
 390	if (ret)
 391		return -EINVAL;
 392
 393	return count;
 394}
 395
 396/*
 397 * Legacy interface that we cannot remove: s390x exposes the storage increment
 398 * covered by a memory block, allowing for identifying which memory blocks
 399 * comprise a storage increment. Since a memory block spans complete
 400 * storage increments nowadays, this interface is basically unused. Other
 401 * archs never exposed != 0.
 402 */
 403static ssize_t phys_device_show(struct device *dev,
 404				struct device_attribute *attr, char *buf)
 405{
 406	struct memory_block *mem = to_memory_block(dev);
 407	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 408
 409	return sysfs_emit(buf, "%d\n",
 410			  arch_get_memory_phys_device(start_pfn));
 411}
 412
 413#ifdef CONFIG_MEMORY_HOTREMOVE
 414static int print_allowed_zone(char *buf, int len, int nid,
 415			      struct memory_group *group,
 416			      unsigned long start_pfn, unsigned long nr_pages,
 417			      int online_type, struct zone *default_zone)
 418{
 419	struct zone *zone;
 420
 421	zone = zone_for_pfn_range(online_type, nid, group, start_pfn, nr_pages);
 422	if (zone == default_zone)
 423		return 0;
 424
 425	return sysfs_emit_at(buf, len, " %s", zone->name);
 426}
 427
 428static ssize_t valid_zones_show(struct device *dev,
 429				struct device_attribute *attr, char *buf)
 430{
 431	struct memory_block *mem = to_memory_block(dev);
 432	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 433	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 434	struct memory_group *group = mem->group;
 435	struct zone *default_zone;
 436	int nid = mem->nid;
 437	int len = 0;
 
 438
 439	/*
 440	 * Check the existing zone. Make sure that we do that only on the
 441	 * online nodes otherwise the page_zone is not reliable
 442	 */
 443	if (mem->state == MEM_ONLINE) {
 444		/*
 445		 * If !mem->zone, the memory block spans multiple zones and
 446		 * cannot get offlined.
 447		 */
 448		default_zone = mem->zone;
 
 449		if (!default_zone)
 450			return sysfs_emit(buf, "%s\n", "none");
 451		len += sysfs_emit_at(buf, len, "%s", default_zone->name);
 452		goto out;
 453	}
 454
 455	default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, group,
 456					  start_pfn, nr_pages);
 
 457
 458	len += sysfs_emit_at(buf, len, "%s", default_zone->name);
 459	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
 460				  MMOP_ONLINE_KERNEL, default_zone);
 461	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
 462				  MMOP_ONLINE_MOVABLE, default_zone);
 463out:
 464	len += sysfs_emit_at(buf, len, "\n");
 465	return len;
 466}
 467static DEVICE_ATTR_RO(valid_zones);
 468#endif
 469
 470static DEVICE_ATTR_RO(phys_index);
 471static DEVICE_ATTR_RW(state);
 472static DEVICE_ATTR_RO(phys_device);
 473static DEVICE_ATTR_RO(removable);
 474
 475/*
 476 * Show the memory block size (shared by all memory blocks).
 477 */
 478static ssize_t block_size_bytes_show(struct device *dev,
 479				     struct device_attribute *attr, char *buf)
 480{
 481	return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
 482}
 483
 484static DEVICE_ATTR_RO(block_size_bytes);
 485
 486/*
 487 * Memory auto online policy.
 488 */
 489
 490static ssize_t auto_online_blocks_show(struct device *dev,
 491				       struct device_attribute *attr, char *buf)
 492{
 493	return sysfs_emit(buf, "%s\n",
 494			  online_type_to_str[mhp_default_online_type]);
 495}
 496
 497static ssize_t auto_online_blocks_store(struct device *dev,
 498					struct device_attribute *attr,
 499					const char *buf, size_t count)
 500{
 501	const int online_type = mhp_online_type_from_str(buf);
 502
 503	if (online_type < 0)
 504		return -EINVAL;
 505
 506	mhp_default_online_type = online_type;
 507	return count;
 508}
 509
 510static DEVICE_ATTR_RW(auto_online_blocks);
 511
 512#ifdef CONFIG_CRASH_HOTPLUG
 513#include <linux/kexec.h>
 514static ssize_t crash_hotplug_show(struct device *dev,
 515				       struct device_attribute *attr, char *buf)
 516{
 517	return sysfs_emit(buf, "%d\n", crash_hotplug_memory_support());
 518}
 519static DEVICE_ATTR_RO(crash_hotplug);
 520#endif
 521
 522/*
 523 * Some architectures will have custom drivers to do this, and
 524 * will not need to do it from userspace.  The fake hot-add code
 525 * as well as ppc64 will do all of their discovery in userspace
 526 * and will require this interface.
 527 */
 528#ifdef CONFIG_ARCH_MEMORY_PROBE
 529static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
 530			   const char *buf, size_t count)
 531{
 532	u64 phys_addr;
 533	int nid, ret;
 534	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
 535
 536	ret = kstrtoull(buf, 0, &phys_addr);
 537	if (ret)
 538		return ret;
 539
 540	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
 541		return -EINVAL;
 542
 543	ret = lock_device_hotplug_sysfs();
 544	if (ret)
 545		return ret;
 546
 547	nid = memory_add_physaddr_to_nid(phys_addr);
 548	ret = __add_memory(nid, phys_addr,
 549			   MIN_MEMORY_BLOCK_SIZE * sections_per_block,
 550			   MHP_NONE);
 551
 552	if (ret)
 553		goto out;
 554
 555	ret = count;
 556out:
 557	unlock_device_hotplug();
 558	return ret;
 559}
 560
 561static DEVICE_ATTR_WO(probe);
 562#endif
 563
 564#ifdef CONFIG_MEMORY_FAILURE
 565/*
 566 * Support for offlining pages of memory
 567 */
 568
 569/* Soft offline a page */
 570static ssize_t soft_offline_page_store(struct device *dev,
 571				       struct device_attribute *attr,
 572				       const char *buf, size_t count)
 573{
 574	int ret;
 575	u64 pfn;
 576	if (!capable(CAP_SYS_ADMIN))
 577		return -EPERM;
 578	if (kstrtoull(buf, 0, &pfn) < 0)
 579		return -EINVAL;
 580	pfn >>= PAGE_SHIFT;
 581	ret = soft_offline_page(pfn, 0);
 582	return ret == 0 ? count : ret;
 583}
 584
 585/* Forcibly offline a page, including killing processes. */
 586static ssize_t hard_offline_page_store(struct device *dev,
 587				       struct device_attribute *attr,
 588				       const char *buf, size_t count)
 589{
 590	int ret;
 591	u64 pfn;
 592	if (!capable(CAP_SYS_ADMIN))
 593		return -EPERM;
 594	if (kstrtoull(buf, 0, &pfn) < 0)
 595		return -EINVAL;
 596	pfn >>= PAGE_SHIFT;
 597	ret = memory_failure(pfn, MF_SW_SIMULATED);
 598	if (ret == -EOPNOTSUPP)
 599		ret = 0;
 600	return ret ? ret : count;
 601}
 602
 603static DEVICE_ATTR_WO(soft_offline_page);
 604static DEVICE_ATTR_WO(hard_offline_page);
 605#endif
 606
 607/* See phys_device_show(). */
 608int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 609{
 610	return 0;
 611}
 612
 613/*
 614 * A reference for the returned memory block device is acquired.
 615 *
 616 * Called under device_hotplug_lock.
 617 */
 618static struct memory_block *find_memory_block_by_id(unsigned long block_id)
 619{
 620	struct memory_block *mem;
 621
 622	mem = xa_load(&memory_blocks, block_id);
 623	if (mem)
 624		get_device(&mem->dev);
 625	return mem;
 626}
 627
 628/*
 629 * Called under device_hotplug_lock.
 630 */
 631struct memory_block *find_memory_block(unsigned long section_nr)
 632{
 633	unsigned long block_id = memory_block_id(section_nr);
 634
 635	return find_memory_block_by_id(block_id);
 636}
 637
 638static struct attribute *memory_memblk_attrs[] = {
 639	&dev_attr_phys_index.attr,
 640	&dev_attr_state.attr,
 641	&dev_attr_phys_device.attr,
 642	&dev_attr_removable.attr,
 643#ifdef CONFIG_MEMORY_HOTREMOVE
 644	&dev_attr_valid_zones.attr,
 645#endif
 646	NULL
 647};
 648
 649static const struct attribute_group memory_memblk_attr_group = {
 650	.attrs = memory_memblk_attrs,
 651};
 652
 653static const struct attribute_group *memory_memblk_attr_groups[] = {
 654	&memory_memblk_attr_group,
 655	NULL,
 656};
 657
 658static int __add_memory_block(struct memory_block *memory)
 
 
 
 
 659{
 660	int ret;
 661
 662	memory->dev.bus = &memory_subsys;
 663	memory->dev.id = memory->start_section_nr / sections_per_block;
 664	memory->dev.release = memory_block_release;
 665	memory->dev.groups = memory_memblk_attr_groups;
 666	memory->dev.offline = memory->state == MEM_OFFLINE;
 667
 668	ret = device_register(&memory->dev);
 669	if (ret) {
 670		put_device(&memory->dev);
 671		return ret;
 672	}
 673	ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
 674			      GFP_KERNEL));
 675	if (ret)
 
 676		device_unregister(&memory->dev);
 677
 678	return ret;
 679}
 680
 681static struct zone *early_node_zone_for_memory_block(struct memory_block *mem,
 682						     int nid)
 683{
 684	const unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 685	const unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 686	struct zone *zone, *matching_zone = NULL;
 687	pg_data_t *pgdat = NODE_DATA(nid);
 688	int i;
 689
 690	/*
 691	 * This logic only works for early memory, when the applicable zones
 692	 * already span the memory block. We don't expect overlapping zones on
 693	 * a single node for early memory. So if we're told that some PFNs
 694	 * of a node fall into this memory block, we can assume that all node
 695	 * zones that intersect with the memory block are actually applicable.
 696	 * No need to look at the memmap.
 697	 */
 698	for (i = 0; i < MAX_NR_ZONES; i++) {
 699		zone = pgdat->node_zones + i;
 700		if (!populated_zone(zone))
 701			continue;
 702		if (!zone_intersects(zone, start_pfn, nr_pages))
 703			continue;
 704		if (!matching_zone) {
 705			matching_zone = zone;
 706			continue;
 707		}
 708		/* Spans multiple zones ... */
 709		matching_zone = NULL;
 710		break;
 711	}
 712	return matching_zone;
 713}
 714
 715#ifdef CONFIG_NUMA
 716/**
 717 * memory_block_add_nid() - Indicate that system RAM falling into this memory
 718 *			    block device (partially) belongs to the given node.
 719 * @mem: The memory block device.
 720 * @nid: The node id.
 721 * @context: The memory initialization context.
 722 *
 723 * Indicate that system RAM falling into this memory block (partially) belongs
 724 * to the given node. If the context indicates ("early") that we are adding the
 725 * node during node device subsystem initialization, this will also properly
 726 * set/adjust mem->zone based on the zone ranges of the given node.
 727 */
 728void memory_block_add_nid(struct memory_block *mem, int nid,
 729			  enum meminit_context context)
 730{
 731	if (context == MEMINIT_EARLY && mem->nid != nid) {
 732		/*
 733		 * For early memory we have to determine the zone when setting
 734		 * the node id and handle multiple nodes spanning a single
 735		 * memory block by indicate via zone == NULL that we're not
 736		 * dealing with a single zone. So if we're setting the node id
 737		 * the first time, determine if there is a single zone. If we're
 738		 * setting the node id a second time to a different node,
 739		 * invalidate the single detected zone.
 740		 */
 741		if (mem->nid == NUMA_NO_NODE)
 742			mem->zone = early_node_zone_for_memory_block(mem, nid);
 743		else
 744			mem->zone = NULL;
 745	}
 746
 747	/*
 748	 * If this memory block spans multiple nodes, we only indicate
 749	 * the last processed node. If we span multiple nodes (not applicable
 750	 * to hotplugged memory), zone == NULL will prohibit memory offlining
 751	 * and consequently unplug.
 752	 */
 753	mem->nid = nid;
 754}
 755#endif
 756
 757static int add_memory_block(unsigned long block_id, unsigned long state,
 758			    struct vmem_altmap *altmap,
 759			    struct memory_group *group)
 760{
 761	struct memory_block *mem;
 762	int ret = 0;
 763
 764	mem = find_memory_block_by_id(block_id);
 765	if (mem) {
 766		put_device(&mem->dev);
 767		return -EEXIST;
 768	}
 769	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 770	if (!mem)
 771		return -ENOMEM;
 772
 773	mem->start_section_nr = block_id * sections_per_block;
 774	mem->state = state;
 775	mem->nid = NUMA_NO_NODE;
 776	mem->altmap = altmap;
 777	INIT_LIST_HEAD(&mem->group_next);
 778
 779#ifndef CONFIG_NUMA
 780	if (state == MEM_ONLINE)
 781		/*
 782		 * MEM_ONLINE at this point implies early memory. With NUMA,
 783		 * we'll determine the zone when setting the node id via
 784		 * memory_block_add_nid(). Memory hotplug updated the zone
 785		 * manually when memory onlining/offlining succeeds.
 786		 */
 787		mem->zone = early_node_zone_for_memory_block(mem, NUMA_NO_NODE);
 788#endif /* CONFIG_NUMA */
 789
 790	ret = __add_memory_block(mem);
 791	if (ret)
 792		return ret;
 793
 794	if (group) {
 795		mem->group = group;
 796		list_add(&mem->group_next, &group->memory_blocks);
 797	}
 798
 799	return 0;
 800}
 801
 802static int __init add_boot_memory_block(unsigned long base_section_nr)
 803{
 804	int section_count = 0;
 805	unsigned long nr;
 806
 807	for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
 808	     nr++)
 809		if (present_section_nr(nr))
 810			section_count++;
 811
 812	if (section_count == 0)
 813		return 0;
 814	return add_memory_block(memory_block_id(base_section_nr),
 815				MEM_ONLINE, NULL,  NULL);
 816}
 817
 818static int add_hotplug_memory_block(unsigned long block_id,
 819				    struct vmem_altmap *altmap,
 820				    struct memory_group *group)
 821{
 822	return add_memory_block(block_id, MEM_OFFLINE, altmap, group);
 823}
 824
 825static void remove_memory_block(struct memory_block *memory)
 826{
 827	if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
 828		return;
 829
 830	WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
 831
 832	if (memory->group) {
 833		list_del(&memory->group_next);
 834		memory->group = NULL;
 835	}
 836
 837	/* drop the ref. we got via find_memory_block() */
 838	put_device(&memory->dev);
 839	device_unregister(&memory->dev);
 840}
 841
 842/*
 843 * Create memory block devices for the given memory area. Start and size
 844 * have to be aligned to memory block granularity. Memory block devices
 845 * will be initialized as offline.
 846 *
 847 * Called under device_hotplug_lock.
 848 */
 849int create_memory_block_devices(unsigned long start, unsigned long size,
 850				struct vmem_altmap *altmap,
 851				struct memory_group *group)
 852{
 853	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
 854	unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
 855	struct memory_block *mem;
 856	unsigned long block_id;
 857	int ret = 0;
 858
 859	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
 860			 !IS_ALIGNED(size, memory_block_size_bytes())))
 861		return -EINVAL;
 862
 863	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
 864		ret = add_hotplug_memory_block(block_id, altmap, group);
 865		if (ret)
 866			break;
 867	}
 868	if (ret) {
 869		end_block_id = block_id;
 870		for (block_id = start_block_id; block_id != end_block_id;
 871		     block_id++) {
 872			mem = find_memory_block_by_id(block_id);
 873			if (WARN_ON_ONCE(!mem))
 874				continue;
 875			remove_memory_block(mem);
 876		}
 877	}
 878	return ret;
 879}
 880
 881/*
 882 * Remove memory block devices for the given memory area. Start and size
 883 * have to be aligned to memory block granularity. Memory block devices
 884 * have to be offline.
 885 *
 886 * Called under device_hotplug_lock.
 887 */
 888void remove_memory_block_devices(unsigned long start, unsigned long size)
 889{
 890	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
 891	const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
 892	struct memory_block *mem;
 893	unsigned long block_id;
 894
 895	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
 896			 !IS_ALIGNED(size, memory_block_size_bytes())))
 897		return;
 898
 899	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
 900		mem = find_memory_block_by_id(block_id);
 901		if (WARN_ON_ONCE(!mem))
 902			continue;
 903		num_poisoned_pages_sub(-1UL, memblk_nr_poison(mem));
 904		unregister_memory_block_under_nodes(mem);
 905		remove_memory_block(mem);
 906	}
 907}
 908
 
 
 
 
 
 
 909static struct attribute *memory_root_attrs[] = {
 910#ifdef CONFIG_ARCH_MEMORY_PROBE
 911	&dev_attr_probe.attr,
 912#endif
 913
 914#ifdef CONFIG_MEMORY_FAILURE
 915	&dev_attr_soft_offline_page.attr,
 916	&dev_attr_hard_offline_page.attr,
 917#endif
 918
 919	&dev_attr_block_size_bytes.attr,
 920	&dev_attr_auto_online_blocks.attr,
 921#ifdef CONFIG_CRASH_HOTPLUG
 922	&dev_attr_crash_hotplug.attr,
 923#endif
 924	NULL
 925};
 926
 927static const struct attribute_group memory_root_attr_group = {
 928	.attrs = memory_root_attrs,
 929};
 930
 931static const struct attribute_group *memory_root_attr_groups[] = {
 932	&memory_root_attr_group,
 933	NULL,
 934};
 935
 936/*
 937 * Initialize the sysfs support for memory devices. At the time this function
 938 * is called, we cannot have concurrent creation/deletion of memory block
 939 * devices, the device_hotplug_lock is not needed.
 940 */
 941void __init memory_dev_init(void)
 942{
 943	int ret;
 944	unsigned long block_sz, nr;
 945
 946	/* Validate the configured memory block size */
 947	block_sz = memory_block_size_bytes();
 948	if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
 949		panic("Memory block size not suitable: 0x%lx\n", block_sz);
 950	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 951
 952	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
 953	if (ret)
 954		panic("%s() failed to register subsystem: %d\n", __func__, ret);
 955
 956	/*
 957	 * Create entries for memory sections that were found
 958	 * during boot and have been initialized
 959	 */
 960	for (nr = 0; nr <= __highest_present_section_nr;
 961	     nr += sections_per_block) {
 962		ret = add_boot_memory_block(nr);
 963		if (ret)
 964			panic("%s() failed to add memory block: %d\n", __func__,
 965			      ret);
 966	}
 967}
 968
 969/**
 970 * walk_memory_blocks - walk through all present memory blocks overlapped
 971 *			by the range [start, start + size)
 972 *
 973 * @start: start address of the memory range
 974 * @size: size of the memory range
 975 * @arg: argument passed to func
 976 * @func: callback for each memory section walked
 977 *
 978 * This function walks through all present memory blocks overlapped by the
 979 * range [start, start + size), calling func on each memory block.
 980 *
 981 * In case func() returns an error, walking is aborted and the error is
 982 * returned.
 983 *
 984 * Called under device_hotplug_lock.
 985 */
 986int walk_memory_blocks(unsigned long start, unsigned long size,
 987		       void *arg, walk_memory_blocks_func_t func)
 988{
 989	const unsigned long start_block_id = phys_to_block_id(start);
 990	const unsigned long end_block_id = phys_to_block_id(start + size - 1);
 991	struct memory_block *mem;
 992	unsigned long block_id;
 993	int ret = 0;
 994
 995	if (!size)
 996		return 0;
 997
 998	for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
 999		mem = find_memory_block_by_id(block_id);
1000		if (!mem)
1001			continue;
1002
1003		ret = func(mem, arg);
1004		put_device(&mem->dev);
1005		if (ret)
1006			break;
1007	}
1008	return ret;
1009}
1010
1011struct for_each_memory_block_cb_data {
1012	walk_memory_blocks_func_t func;
1013	void *arg;
1014};
1015
1016static int for_each_memory_block_cb(struct device *dev, void *data)
1017{
1018	struct memory_block *mem = to_memory_block(dev);
1019	struct for_each_memory_block_cb_data *cb_data = data;
1020
1021	return cb_data->func(mem, cb_data->arg);
1022}
1023
1024/**
1025 * for_each_memory_block - walk through all present memory blocks
1026 *
1027 * @arg: argument passed to func
1028 * @func: callback for each memory block walked
1029 *
1030 * This function walks through all present memory blocks, calling func on
1031 * each memory block.
1032 *
1033 * In case func() returns an error, walking is aborted and the error is
1034 * returned.
1035 */
1036int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
1037{
1038	struct for_each_memory_block_cb_data cb_data = {
1039		.func = func,
1040		.arg = arg,
1041	};
1042
1043	return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
1044				for_each_memory_block_cb);
1045}
1046
1047/*
1048 * This is an internal helper to unify allocation and initialization of
1049 * memory groups. Note that the passed memory group will be copied to a
1050 * dynamically allocated memory group. After this call, the passed
1051 * memory group should no longer be used.
1052 */
1053static int memory_group_register(struct memory_group group)
1054{
1055	struct memory_group *new_group;
1056	uint32_t mgid;
1057	int ret;
1058
1059	if (!node_possible(group.nid))
1060		return -EINVAL;
1061
1062	new_group = kzalloc(sizeof(group), GFP_KERNEL);
1063	if (!new_group)
1064		return -ENOMEM;
1065	*new_group = group;
1066	INIT_LIST_HEAD(&new_group->memory_blocks);
1067
1068	ret = xa_alloc(&memory_groups, &mgid, new_group, xa_limit_31b,
1069		       GFP_KERNEL);
1070	if (ret) {
1071		kfree(new_group);
1072		return ret;
1073	} else if (group.is_dynamic) {
1074		xa_set_mark(&memory_groups, mgid, MEMORY_GROUP_MARK_DYNAMIC);
1075	}
1076	return mgid;
1077}
1078
1079/**
1080 * memory_group_register_static() - Register a static memory group.
1081 * @nid: The node id.
1082 * @max_pages: The maximum number of pages we'll have in this static memory
1083 *	       group.
1084 *
1085 * Register a new static memory group and return the memory group id.
1086 * All memory in the group belongs to a single unit, such as a DIMM. All
1087 * memory belonging to a static memory group is added in one go to be removed
1088 * in one go -- it's static.
1089 *
1090 * Returns an error if out of memory, if the node id is invalid, if no new
1091 * memory groups can be registered, or if max_pages is invalid (0). Otherwise,
1092 * returns the new memory group id.
1093 */
1094int memory_group_register_static(int nid, unsigned long max_pages)
1095{
1096	struct memory_group group = {
1097		.nid = nid,
1098		.s = {
1099			.max_pages = max_pages,
1100		},
1101	};
1102
1103	if (!max_pages)
1104		return -EINVAL;
1105	return memory_group_register(group);
1106}
1107EXPORT_SYMBOL_GPL(memory_group_register_static);
1108
1109/**
1110 * memory_group_register_dynamic() - Register a dynamic memory group.
1111 * @nid: The node id.
1112 * @unit_pages: Unit in pages in which is memory added/removed in this dynamic
1113 *		memory group.
1114 *
1115 * Register a new dynamic memory group and return the memory group id.
1116 * Memory within a dynamic memory group is added/removed dynamically
1117 * in unit_pages.
1118 *
1119 * Returns an error if out of memory, if the node id is invalid, if no new
1120 * memory groups can be registered, or if unit_pages is invalid (0, not a
1121 * power of two, smaller than a single memory block). Otherwise, returns the
1122 * new memory group id.
1123 */
1124int memory_group_register_dynamic(int nid, unsigned long unit_pages)
1125{
1126	struct memory_group group = {
1127		.nid = nid,
1128		.is_dynamic = true,
1129		.d = {
1130			.unit_pages = unit_pages,
1131		},
1132	};
1133
1134	if (!unit_pages || !is_power_of_2(unit_pages) ||
1135	    unit_pages < PHYS_PFN(memory_block_size_bytes()))
1136		return -EINVAL;
1137	return memory_group_register(group);
1138}
1139EXPORT_SYMBOL_GPL(memory_group_register_dynamic);
1140
1141/**
1142 * memory_group_unregister() - Unregister a memory group.
1143 * @mgid: the memory group id
1144 *
1145 * Unregister a memory group. If any memory block still belongs to this
1146 * memory group, unregistering will fail.
1147 *
1148 * Returns -EINVAL if the memory group id is invalid, returns -EBUSY if some
1149 * memory blocks still belong to this memory group and returns 0 if
1150 * unregistering succeeded.
1151 */
1152int memory_group_unregister(int mgid)
1153{
1154	struct memory_group *group;
1155
1156	if (mgid < 0)
1157		return -EINVAL;
1158
1159	group = xa_load(&memory_groups, mgid);
1160	if (!group)
1161		return -EINVAL;
1162	if (!list_empty(&group->memory_blocks))
1163		return -EBUSY;
1164	xa_erase(&memory_groups, mgid);
1165	kfree(group);
1166	return 0;
1167}
1168EXPORT_SYMBOL_GPL(memory_group_unregister);
1169
1170/*
1171 * This is an internal helper only to be used in core memory hotplug code to
1172 * lookup a memory group. We don't care about locking, as we don't expect a
1173 * memory group to get unregistered while adding memory to it -- because
1174 * the group and the memory is managed by the same driver.
1175 */
1176struct memory_group *memory_group_find_by_id(int mgid)
1177{
1178	return xa_load(&memory_groups, mgid);
1179}
1180
1181/*
1182 * This is an internal helper only to be used in core memory hotplug code to
1183 * walk all dynamic memory groups excluding a given memory group, either
1184 * belonging to a specific node, or belonging to any node.
1185 */
1186int walk_dynamic_memory_groups(int nid, walk_memory_groups_func_t func,
1187			       struct memory_group *excluded, void *arg)
1188{
1189	struct memory_group *group;
1190	unsigned long index;
1191	int ret = 0;
1192
1193	xa_for_each_marked(&memory_groups, index, group,
1194			   MEMORY_GROUP_MARK_DYNAMIC) {
1195		if (group == excluded)
1196			continue;
1197#ifdef CONFIG_NUMA
1198		if (nid != NUMA_NO_NODE && group->nid != nid)
1199			continue;
1200#endif /* CONFIG_NUMA */
1201		ret = func(group, arg);
1202		if (ret)
1203			break;
1204	}
1205	return ret;
1206}
1207
1208#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
1209void memblk_nr_poison_inc(unsigned long pfn)
1210{
1211	const unsigned long block_id = pfn_to_block_id(pfn);
1212	struct memory_block *mem = find_memory_block_by_id(block_id);
1213
1214	if (mem)
1215		atomic_long_inc(&mem->nr_hwpoison);
1216}
1217
1218void memblk_nr_poison_sub(unsigned long pfn, long i)
1219{
1220	const unsigned long block_id = pfn_to_block_id(pfn);
1221	struct memory_block *mem = find_memory_block_by_id(block_id);
1222
1223	if (mem)
1224		atomic_long_sub(i, &mem->nr_hwpoison);
1225}
1226
1227static unsigned long memblk_nr_poison(struct memory_block *mem)
1228{
1229	return atomic_long_read(&mem->nr_hwpoison);
1230}
1231#endif
v5.14.15
  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
 
 
 
 
 
 
 85static BLOCKING_NOTIFIER_HEAD(memory_chain);
 86
 87int register_memory_notifier(struct notifier_block *nb)
 88{
 89	return blocking_notifier_chain_register(&memory_chain, nb);
 90}
 91EXPORT_SYMBOL(register_memory_notifier);
 92
 93void unregister_memory_notifier(struct notifier_block *nb)
 94{
 95	blocking_notifier_chain_unregister(&memory_chain, nb);
 96}
 97EXPORT_SYMBOL(unregister_memory_notifier);
 98
 99static void memory_block_release(struct device *dev)
100{
101	struct memory_block *mem = to_memory_block(dev);
102
 
103	kfree(mem);
104}
105
106unsigned long __weak memory_block_size_bytes(void)
107{
108	return MIN_MEMORY_BLOCK_SIZE;
109}
110EXPORT_SYMBOL_GPL(memory_block_size_bytes);
111
112/*
113 * Show the first physical section index (number) of this memory block.
114 */
115static ssize_t phys_index_show(struct device *dev,
116			       struct device_attribute *attr, char *buf)
117{
118	struct memory_block *mem = to_memory_block(dev);
119	unsigned long phys_index;
120
121	phys_index = mem->start_section_nr / sections_per_block;
122
123	return sysfs_emit(buf, "%08lx\n", phys_index);
124}
125
126/*
127 * Legacy interface that we cannot remove. Always indicate "removable"
128 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
129 */
130static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
131			      char *buf)
132{
133	return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
134}
135
136/*
137 * online, offline, going offline, etc.
138 */
139static ssize_t state_show(struct device *dev, struct device_attribute *attr,
140			  char *buf)
141{
142	struct memory_block *mem = to_memory_block(dev);
143	const char *output;
144
145	/*
146	 * We can probably put these states in a nice little array
147	 * so that they're not open-coded
148	 */
149	switch (mem->state) {
150	case MEM_ONLINE:
151		output = "online";
152		break;
153	case MEM_OFFLINE:
154		output = "offline";
155		break;
156	case MEM_GOING_OFFLINE:
157		output = "going-offline";
158		break;
159	default:
160		WARN_ON(1);
161		return sysfs_emit(buf, "ERROR-UNKNOWN-%ld\n", mem->state);
162	}
163
164	return sysfs_emit(buf, "%s\n", output);
165}
166
167int memory_notify(unsigned long val, void *v)
168{
169	return blocking_notifier_call_chain(&memory_chain, val, v);
170}
171
 
 
 
 
 
 
 
 
 
 
 
 
172static int memory_block_online(struct memory_block *mem)
173{
174	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
175	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
176	unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
177	struct zone *zone;
178	int ret;
179
180	zone = zone_for_pfn_range(mem->online_type, mem->nid, start_pfn, nr_pages);
 
 
 
 
181
182	/*
183	 * Although vmemmap pages have a different lifecycle than the pages
184	 * they describe (they remain until the memory is unplugged), doing
185	 * their initialization and accounting at memory onlining/offlining
186	 * stage helps to keep accounting easier to follow - e.g vmemmaps
187	 * belong to the same zone as the memory they backed.
188	 */
 
 
 
 
189	if (nr_vmemmap_pages) {
190		ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
191		if (ret)
192			return ret;
193	}
194
195	ret = online_pages(start_pfn + nr_vmemmap_pages,
196			   nr_pages - nr_vmemmap_pages, zone);
197	if (ret) {
198		if (nr_vmemmap_pages)
199			mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
200		return ret;
201	}
202
203	/*
204	 * Account once onlining succeeded. If the zone was unpopulated, it is
205	 * now already properly populated.
206	 */
207	if (nr_vmemmap_pages)
208		adjust_present_page_count(zone, nr_vmemmap_pages);
 
209
 
 
 
210	return ret;
211}
212
 
 
 
213static int memory_block_offline(struct memory_block *mem)
214{
215	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
216	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
217	unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
218	struct zone *zone;
219	int ret;
220
 
 
 
221	/*
222	 * Unaccount before offlining, such that unpopulated zone and kthreads
223	 * can properly be torn down in offline_pages().
224	 */
225	if (nr_vmemmap_pages) {
226		zone = page_zone(pfn_to_page(start_pfn));
227		adjust_present_page_count(zone, -nr_vmemmap_pages);
228	}
 
 
 
229
230	ret = offline_pages(start_pfn + nr_vmemmap_pages,
231			    nr_pages - nr_vmemmap_pages);
232	if (ret) {
233		/* offline_pages() failed. Account back. */
234		if (nr_vmemmap_pages)
235			adjust_present_page_count(zone, nr_vmemmap_pages);
236		return ret;
 
237	}
238
239	if (nr_vmemmap_pages)
240		mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
241
 
 
 
242	return ret;
243}
244
245/*
246 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
247 * OK to have direct references to sparsemem variables in here.
248 */
249static int
250memory_block_action(struct memory_block *mem, unsigned long action)
251{
252	int ret;
253
254	switch (action) {
255	case MEM_ONLINE:
256		ret = memory_block_online(mem);
257		break;
258	case MEM_OFFLINE:
259		ret = memory_block_offline(mem);
260		break;
261	default:
262		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
263		     "%ld\n", __func__, mem->start_section_nr, action, action);
264		ret = -EINVAL;
265	}
266
267	return ret;
268}
269
270static int memory_block_change_state(struct memory_block *mem,
271		unsigned long to_state, unsigned long from_state_req)
272{
273	int ret = 0;
274
275	if (mem->state != from_state_req)
276		return -EINVAL;
277
278	if (to_state == MEM_OFFLINE)
279		mem->state = MEM_GOING_OFFLINE;
280
281	ret = memory_block_action(mem, to_state);
282	mem->state = ret ? from_state_req : to_state;
283
284	return ret;
285}
286
287/* The device lock serializes operations on memory_subsys_[online|offline] */
288static int memory_subsys_online(struct device *dev)
289{
290	struct memory_block *mem = to_memory_block(dev);
291	int ret;
292
293	if (mem->state == MEM_ONLINE)
294		return 0;
295
296	/*
297	 * When called via device_online() without configuring the online_type,
298	 * we want to default to MMOP_ONLINE.
299	 */
300	if (mem->online_type == MMOP_OFFLINE)
301		mem->online_type = MMOP_ONLINE;
302
303	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
304	mem->online_type = MMOP_OFFLINE;
305
306	return ret;
307}
308
309static int memory_subsys_offline(struct device *dev)
310{
311	struct memory_block *mem = to_memory_block(dev);
312
313	if (mem->state == MEM_OFFLINE)
314		return 0;
315
316	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
317}
318
319static ssize_t state_store(struct device *dev, struct device_attribute *attr,
320			   const char *buf, size_t count)
321{
322	const int online_type = mhp_online_type_from_str(buf);
323	struct memory_block *mem = to_memory_block(dev);
324	int ret;
325
326	if (online_type < 0)
327		return -EINVAL;
328
329	ret = lock_device_hotplug_sysfs();
330	if (ret)
331		return ret;
332
333	switch (online_type) {
334	case MMOP_ONLINE_KERNEL:
335	case MMOP_ONLINE_MOVABLE:
336	case MMOP_ONLINE:
337		/* mem->online_type is protected by device_hotplug_lock */
338		mem->online_type = online_type;
339		ret = device_online(&mem->dev);
340		break;
341	case MMOP_OFFLINE:
342		ret = device_offline(&mem->dev);
343		break;
344	default:
345		ret = -EINVAL; /* should never happen */
346	}
347
348	unlock_device_hotplug();
349
350	if (ret < 0)
351		return ret;
352	if (ret)
353		return -EINVAL;
354
355	return count;
356}
357
358/*
359 * Legacy interface that we cannot remove: s390x exposes the storage increment
360 * covered by a memory block, allowing for identifying which memory blocks
361 * comprise a storage increment. Since a memory block spans complete
362 * storage increments nowadays, this interface is basically unused. Other
363 * archs never exposed != 0.
364 */
365static ssize_t phys_device_show(struct device *dev,
366				struct device_attribute *attr, char *buf)
367{
368	struct memory_block *mem = to_memory_block(dev);
369	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
370
371	return sysfs_emit(buf, "%d\n",
372			  arch_get_memory_phys_device(start_pfn));
373}
374
375#ifdef CONFIG_MEMORY_HOTREMOVE
376static int print_allowed_zone(char *buf, int len, int nid,
 
377			      unsigned long start_pfn, unsigned long nr_pages,
378			      int online_type, struct zone *default_zone)
379{
380	struct zone *zone;
381
382	zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
383	if (zone == default_zone)
384		return 0;
385
386	return sysfs_emit_at(buf, len, " %s", zone->name);
387}
388
389static ssize_t valid_zones_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	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 
395	struct zone *default_zone;
 
396	int len = 0;
397	int nid;
398
399	/*
400	 * Check the existing zone. Make sure that we do that only on the
401	 * online nodes otherwise the page_zone is not reliable
402	 */
403	if (mem->state == MEM_ONLINE) {
404		/*
405		 * The block contains more than one zone can not be offlined.
406		 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
407		 */
408		default_zone = test_pages_in_a_zone(start_pfn,
409						    start_pfn + nr_pages);
410		if (!default_zone)
411			return sysfs_emit(buf, "%s\n", "none");
412		len += sysfs_emit_at(buf, len, "%s", default_zone->name);
413		goto out;
414	}
415
416	nid = mem->nid;
417	default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, start_pfn,
418					  nr_pages);
419
420	len += sysfs_emit_at(buf, len, "%s", default_zone->name);
421	len += print_allowed_zone(buf, len, nid, start_pfn, nr_pages,
422				  MMOP_ONLINE_KERNEL, default_zone);
423	len += print_allowed_zone(buf, len, nid, start_pfn, nr_pages,
424				  MMOP_ONLINE_MOVABLE, default_zone);
425out:
426	len += sysfs_emit_at(buf, len, "\n");
427	return len;
428}
429static DEVICE_ATTR_RO(valid_zones);
430#endif
431
432static DEVICE_ATTR_RO(phys_index);
433static DEVICE_ATTR_RW(state);
434static DEVICE_ATTR_RO(phys_device);
435static DEVICE_ATTR_RO(removable);
436
437/*
438 * Show the memory block size (shared by all memory blocks).
439 */
440static ssize_t block_size_bytes_show(struct device *dev,
441				     struct device_attribute *attr, char *buf)
442{
443	return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
444}
445
446static DEVICE_ATTR_RO(block_size_bytes);
447
448/*
449 * Memory auto online policy.
450 */
451
452static ssize_t auto_online_blocks_show(struct device *dev,
453				       struct device_attribute *attr, char *buf)
454{
455	return sysfs_emit(buf, "%s\n",
456			  online_type_to_str[mhp_default_online_type]);
457}
458
459static ssize_t auto_online_blocks_store(struct device *dev,
460					struct device_attribute *attr,
461					const char *buf, size_t count)
462{
463	const int online_type = mhp_online_type_from_str(buf);
464
465	if (online_type < 0)
466		return -EINVAL;
467
468	mhp_default_online_type = online_type;
469	return count;
470}
471
472static DEVICE_ATTR_RW(auto_online_blocks);
473
 
 
 
 
 
 
 
 
 
 
474/*
475 * Some architectures will have custom drivers to do this, and
476 * will not need to do it from userspace.  The fake hot-add code
477 * as well as ppc64 will do all of their discovery in userspace
478 * and will require this interface.
479 */
480#ifdef CONFIG_ARCH_MEMORY_PROBE
481static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
482			   const char *buf, size_t count)
483{
484	u64 phys_addr;
485	int nid, ret;
486	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
487
488	ret = kstrtoull(buf, 0, &phys_addr);
489	if (ret)
490		return ret;
491
492	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
493		return -EINVAL;
494
495	ret = lock_device_hotplug_sysfs();
496	if (ret)
497		return ret;
498
499	nid = memory_add_physaddr_to_nid(phys_addr);
500	ret = __add_memory(nid, phys_addr,
501			   MIN_MEMORY_BLOCK_SIZE * sections_per_block,
502			   MHP_NONE);
503
504	if (ret)
505		goto out;
506
507	ret = count;
508out:
509	unlock_device_hotplug();
510	return ret;
511}
512
513static DEVICE_ATTR_WO(probe);
514#endif
515
516#ifdef CONFIG_MEMORY_FAILURE
517/*
518 * Support for offlining pages of memory
519 */
520
521/* Soft offline a page */
522static ssize_t soft_offline_page_store(struct device *dev,
523				       struct device_attribute *attr,
524				       const char *buf, size_t count)
525{
526	int ret;
527	u64 pfn;
528	if (!capable(CAP_SYS_ADMIN))
529		return -EPERM;
530	if (kstrtoull(buf, 0, &pfn) < 0)
531		return -EINVAL;
532	pfn >>= PAGE_SHIFT;
533	ret = soft_offline_page(pfn, 0);
534	return ret == 0 ? count : ret;
535}
536
537/* Forcibly offline a page, including killing processes. */
538static ssize_t hard_offline_page_store(struct device *dev,
539				       struct device_attribute *attr,
540				       const char *buf, size_t count)
541{
542	int ret;
543	u64 pfn;
544	if (!capable(CAP_SYS_ADMIN))
545		return -EPERM;
546	if (kstrtoull(buf, 0, &pfn) < 0)
547		return -EINVAL;
548	pfn >>= PAGE_SHIFT;
549	ret = memory_failure(pfn, 0);
 
 
550	return ret ? ret : count;
551}
552
553static DEVICE_ATTR_WO(soft_offline_page);
554static DEVICE_ATTR_WO(hard_offline_page);
555#endif
556
557/* See phys_device_show(). */
558int __weak arch_get_memory_phys_device(unsigned long start_pfn)
559{
560	return 0;
561}
562
563/*
564 * A reference for the returned memory block device is acquired.
565 *
566 * Called under device_hotplug_lock.
567 */
568static struct memory_block *find_memory_block_by_id(unsigned long block_id)
569{
570	struct memory_block *mem;
571
572	mem = xa_load(&memory_blocks, block_id);
573	if (mem)
574		get_device(&mem->dev);
575	return mem;
576}
577
578/*
579 * Called under device_hotplug_lock.
580 */
581struct memory_block *find_memory_block(struct mem_section *section)
582{
583	unsigned long block_id = memory_block_id(__section_nr(section));
584
585	return find_memory_block_by_id(block_id);
586}
587
588static struct attribute *memory_memblk_attrs[] = {
589	&dev_attr_phys_index.attr,
590	&dev_attr_state.attr,
591	&dev_attr_phys_device.attr,
592	&dev_attr_removable.attr,
593#ifdef CONFIG_MEMORY_HOTREMOVE
594	&dev_attr_valid_zones.attr,
595#endif
596	NULL
597};
598
599static const struct attribute_group memory_memblk_attr_group = {
600	.attrs = memory_memblk_attrs,
601};
602
603static const struct attribute_group *memory_memblk_attr_groups[] = {
604	&memory_memblk_attr_group,
605	NULL,
606};
607
608/*
609 * register_memory - Setup a sysfs device for a memory block
610 */
611static
612int register_memory(struct memory_block *memory)
613{
614	int ret;
615
616	memory->dev.bus = &memory_subsys;
617	memory->dev.id = memory->start_section_nr / sections_per_block;
618	memory->dev.release = memory_block_release;
619	memory->dev.groups = memory_memblk_attr_groups;
620	memory->dev.offline = memory->state == MEM_OFFLINE;
621
622	ret = device_register(&memory->dev);
623	if (ret) {
624		put_device(&memory->dev);
625		return ret;
626	}
627	ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
628			      GFP_KERNEL));
629	if (ret) {
630		put_device(&memory->dev);
631		device_unregister(&memory->dev);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
632	}
633	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
634}
 
635
636static int init_memory_block(unsigned long block_id, unsigned long state,
637			     unsigned long nr_vmemmap_pages)
 
638{
639	struct memory_block *mem;
640	int ret = 0;
641
642	mem = find_memory_block_by_id(block_id);
643	if (mem) {
644		put_device(&mem->dev);
645		return -EEXIST;
646	}
647	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
648	if (!mem)
649		return -ENOMEM;
650
651	mem->start_section_nr = block_id * sections_per_block;
652	mem->state = state;
653	mem->nid = NUMA_NO_NODE;
654	mem->nr_vmemmap_pages = nr_vmemmap_pages;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
655
656	ret = register_memory(mem);
 
 
 
657
658	return ret;
659}
660
661static int add_memory_block(unsigned long base_section_nr)
662{
663	int section_count = 0;
664	unsigned long nr;
665
666	for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
667	     nr++)
668		if (present_section_nr(nr))
669			section_count++;
670
671	if (section_count == 0)
672		return 0;
673	return init_memory_block(memory_block_id(base_section_nr),
674				 MEM_ONLINE, 0);
675}
676
677static void unregister_memory(struct memory_block *memory)
 
 
 
 
 
 
 
678{
679	if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
680		return;
681
682	WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
683
 
 
 
 
 
684	/* drop the ref. we got via find_memory_block() */
685	put_device(&memory->dev);
686	device_unregister(&memory->dev);
687}
688
689/*
690 * Create memory block devices for the given memory area. Start and size
691 * have to be aligned to memory block granularity. Memory block devices
692 * will be initialized as offline.
693 *
694 * Called under device_hotplug_lock.
695 */
696int create_memory_block_devices(unsigned long start, unsigned long size,
697				unsigned long vmemmap_pages)
 
698{
699	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
700	unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
701	struct memory_block *mem;
702	unsigned long block_id;
703	int ret = 0;
704
705	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
706			 !IS_ALIGNED(size, memory_block_size_bytes())))
707		return -EINVAL;
708
709	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
710		ret = init_memory_block(block_id, MEM_OFFLINE, vmemmap_pages);
711		if (ret)
712			break;
713	}
714	if (ret) {
715		end_block_id = block_id;
716		for (block_id = start_block_id; block_id != end_block_id;
717		     block_id++) {
718			mem = find_memory_block_by_id(block_id);
719			if (WARN_ON_ONCE(!mem))
720				continue;
721			unregister_memory(mem);
722		}
723	}
724	return ret;
725}
726
727/*
728 * Remove memory block devices for the given memory area. Start and size
729 * have to be aligned to memory block granularity. Memory block devices
730 * have to be offline.
731 *
732 * Called under device_hotplug_lock.
733 */
734void remove_memory_block_devices(unsigned long start, unsigned long size)
735{
736	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
737	const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
738	struct memory_block *mem;
739	unsigned long block_id;
740
741	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
742			 !IS_ALIGNED(size, memory_block_size_bytes())))
743		return;
744
745	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
746		mem = find_memory_block_by_id(block_id);
747		if (WARN_ON_ONCE(!mem))
748			continue;
 
749		unregister_memory_block_under_nodes(mem);
750		unregister_memory(mem);
751	}
752}
753
754/* return true if the memory block is offlined, otherwise, return false */
755bool is_memblock_offlined(struct memory_block *mem)
756{
757	return mem->state == MEM_OFFLINE;
758}
759
760static struct attribute *memory_root_attrs[] = {
761#ifdef CONFIG_ARCH_MEMORY_PROBE
762	&dev_attr_probe.attr,
763#endif
764
765#ifdef CONFIG_MEMORY_FAILURE
766	&dev_attr_soft_offline_page.attr,
767	&dev_attr_hard_offline_page.attr,
768#endif
769
770	&dev_attr_block_size_bytes.attr,
771	&dev_attr_auto_online_blocks.attr,
 
 
 
772	NULL
773};
774
775static const struct attribute_group memory_root_attr_group = {
776	.attrs = memory_root_attrs,
777};
778
779static const struct attribute_group *memory_root_attr_groups[] = {
780	&memory_root_attr_group,
781	NULL,
782};
783
784/*
785 * Initialize the sysfs support for memory devices. At the time this function
786 * is called, we cannot have concurrent creation/deletion of memory block
787 * devices, the device_hotplug_lock is not needed.
788 */
789void __init memory_dev_init(void)
790{
791	int ret;
792	unsigned long block_sz, nr;
793
794	/* Validate the configured memory block size */
795	block_sz = memory_block_size_bytes();
796	if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
797		panic("Memory block size not suitable: 0x%lx\n", block_sz);
798	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
799
800	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
801	if (ret)
802		panic("%s() failed to register subsystem: %d\n", __func__, ret);
803
804	/*
805	 * Create entries for memory sections that were found
806	 * during boot and have been initialized
807	 */
808	for (nr = 0; nr <= __highest_present_section_nr;
809	     nr += sections_per_block) {
810		ret = add_memory_block(nr);
811		if (ret)
812			panic("%s() failed to add memory block: %d\n", __func__,
813			      ret);
814	}
815}
816
817/**
818 * walk_memory_blocks - walk through all present memory blocks overlapped
819 *			by the range [start, start + size)
820 *
821 * @start: start address of the memory range
822 * @size: size of the memory range
823 * @arg: argument passed to func
824 * @func: callback for each memory section walked
825 *
826 * This function walks through all present memory blocks overlapped by the
827 * range [start, start + size), calling func on each memory block.
828 *
829 * In case func() returns an error, walking is aborted and the error is
830 * returned.
831 *
832 * Called under device_hotplug_lock.
833 */
834int walk_memory_blocks(unsigned long start, unsigned long size,
835		       void *arg, walk_memory_blocks_func_t func)
836{
837	const unsigned long start_block_id = phys_to_block_id(start);
838	const unsigned long end_block_id = phys_to_block_id(start + size - 1);
839	struct memory_block *mem;
840	unsigned long block_id;
841	int ret = 0;
842
843	if (!size)
844		return 0;
845
846	for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
847		mem = find_memory_block_by_id(block_id);
848		if (!mem)
849			continue;
850
851		ret = func(mem, arg);
852		put_device(&mem->dev);
853		if (ret)
854			break;
855	}
856	return ret;
857}
858
859struct for_each_memory_block_cb_data {
860	walk_memory_blocks_func_t func;
861	void *arg;
862};
863
864static int for_each_memory_block_cb(struct device *dev, void *data)
865{
866	struct memory_block *mem = to_memory_block(dev);
867	struct for_each_memory_block_cb_data *cb_data = data;
868
869	return cb_data->func(mem, cb_data->arg);
870}
871
872/**
873 * for_each_memory_block - walk through all present memory blocks
874 *
875 * @arg: argument passed to func
876 * @func: callback for each memory block walked
877 *
878 * This function walks through all present memory blocks, calling func on
879 * each memory block.
880 *
881 * In case func() returns an error, walking is aborted and the error is
882 * returned.
883 */
884int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
885{
886	struct for_each_memory_block_cb_data cb_data = {
887		.func = func,
888		.arg = arg,
889	};
890
891	return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
892				for_each_memory_block_cb);
893}