xref: /linux/drivers/base/memory.c (revision dec1c62e91ba268ab2a6e339d4d7a59287d5eba1)
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 
31 static 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 
38 int 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 
51 static int sections_per_block;
52 
53 static inline unsigned long memory_block_id(unsigned long section_nr)
54 {
55 	return section_nr / sections_per_block;
56 }
57 
58 static inline unsigned long pfn_to_block_id(unsigned long pfn)
59 {
60 	return memory_block_id(pfn_to_section_nr(pfn));
61 }
62 
63 static inline unsigned long phys_to_block_id(unsigned long phys)
64 {
65 	return pfn_to_block_id(PFN_DOWN(phys));
66 }
67 
68 static int memory_subsys_online(struct device *dev);
69 static int memory_subsys_offline(struct device *dev);
70 
71 static 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  */
83 static DEFINE_XARRAY(memory_blocks);
84 
85 /*
86  * Memory groups, indexed by memory group id (mgid).
87  */
88 static DEFINE_XARRAY_FLAGS(memory_groups, XA_FLAGS_ALLOC);
89 #define MEMORY_GROUP_MARK_DYNAMIC	XA_MARK_1
90 
91 static BLOCKING_NOTIFIER_HEAD(memory_chain);
92 
93 int register_memory_notifier(struct notifier_block *nb)
94 {
95 	return blocking_notifier_chain_register(&memory_chain, nb);
96 }
97 EXPORT_SYMBOL(register_memory_notifier);
98 
99 void unregister_memory_notifier(struct notifier_block *nb)
100 {
101 	blocking_notifier_chain_unregister(&memory_chain, nb);
102 }
103 EXPORT_SYMBOL(unregister_memory_notifier);
104 
105 static void memory_block_release(struct device *dev)
106 {
107 	struct memory_block *mem = to_memory_block(dev);
108 
109 	kfree(mem);
110 }
111 
112 unsigned long __weak memory_block_size_bytes(void)
113 {
114 	return MIN_MEMORY_BLOCK_SIZE;
115 }
116 EXPORT_SYMBOL_GPL(memory_block_size_bytes);
117 
118 /*
119  * Show the first physical section index (number) of this memory block.
120  */
121 static ssize_t phys_index_show(struct device *dev,
122 			       struct device_attribute *attr, char *buf)
123 {
124 	struct memory_block *mem = to_memory_block(dev);
125 	unsigned long phys_index;
126 
127 	phys_index = mem->start_section_nr / sections_per_block;
128 
129 	return sysfs_emit(buf, "%08lx\n", phys_index);
130 }
131 
132 /*
133  * Legacy interface that we cannot remove. Always indicate "removable"
134  * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
135  */
136 static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
137 			      char *buf)
138 {
139 	return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
140 }
141 
142 /*
143  * online, offline, going offline, etc.
144  */
145 static ssize_t state_show(struct device *dev, struct device_attribute *attr,
146 			  char *buf)
147 {
148 	struct memory_block *mem = to_memory_block(dev);
149 	const char *output;
150 
151 	/*
152 	 * We can probably put these states in a nice little array
153 	 * so that they're not open-coded
154 	 */
155 	switch (mem->state) {
156 	case MEM_ONLINE:
157 		output = "online";
158 		break;
159 	case MEM_OFFLINE:
160 		output = "offline";
161 		break;
162 	case MEM_GOING_OFFLINE:
163 		output = "going-offline";
164 		break;
165 	default:
166 		WARN_ON(1);
167 		return sysfs_emit(buf, "ERROR-UNKNOWN-%ld\n", mem->state);
168 	}
169 
170 	return sysfs_emit(buf, "%s\n", output);
171 }
172 
173 int memory_notify(unsigned long val, void *v)
174 {
175 	return blocking_notifier_call_chain(&memory_chain, val, v);
176 }
177 
178 static int memory_block_online(struct memory_block *mem)
179 {
180 	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
181 	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
182 	unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
183 	struct zone *zone;
184 	int ret;
185 
186 	zone = zone_for_pfn_range(mem->online_type, mem->nid, mem->group,
187 				  start_pfn, nr_pages);
188 
189 	/*
190 	 * Although vmemmap pages have a different lifecycle than the pages
191 	 * they describe (they remain until the memory is unplugged), doing
192 	 * their initialization and accounting at memory onlining/offlining
193 	 * stage helps to keep accounting easier to follow - e.g vmemmaps
194 	 * belong to the same zone as the memory they backed.
195 	 */
196 	if (nr_vmemmap_pages) {
197 		ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
198 		if (ret)
199 			return ret;
200 	}
201 
202 	ret = online_pages(start_pfn + nr_vmemmap_pages,
203 			   nr_pages - nr_vmemmap_pages, zone, mem->group);
204 	if (ret) {
205 		if (nr_vmemmap_pages)
206 			mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
207 		return ret;
208 	}
209 
210 	/*
211 	 * Account once onlining succeeded. If the zone was unpopulated, it is
212 	 * now already properly populated.
213 	 */
214 	if (nr_vmemmap_pages)
215 		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
216 					  nr_vmemmap_pages);
217 
218 	mem->zone = zone;
219 	return ret;
220 }
221 
222 static int memory_block_offline(struct memory_block *mem)
223 {
224 	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
225 	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
226 	unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
227 	int ret;
228 
229 	if (!mem->zone)
230 		return -EINVAL;
231 
232 	/*
233 	 * Unaccount before offlining, such that unpopulated zone and kthreads
234 	 * can properly be torn down in offline_pages().
235 	 */
236 	if (nr_vmemmap_pages)
237 		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
238 					  -nr_vmemmap_pages);
239 
240 	ret = offline_pages(start_pfn + nr_vmemmap_pages,
241 			    nr_pages - nr_vmemmap_pages, mem->zone, mem->group);
242 	if (ret) {
243 		/* offline_pages() failed. Account back. */
244 		if (nr_vmemmap_pages)
245 			adjust_present_page_count(pfn_to_page(start_pfn),
246 						  mem->group, nr_vmemmap_pages);
247 		return ret;
248 	}
249 
250 	if (nr_vmemmap_pages)
251 		mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
252 
253 	mem->zone = NULL;
254 	return ret;
255 }
256 
257 /*
258  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
259  * OK to have direct references to sparsemem variables in here.
260  */
261 static int
262 memory_block_action(struct memory_block *mem, unsigned long action)
263 {
264 	int ret;
265 
266 	switch (action) {
267 	case MEM_ONLINE:
268 		ret = memory_block_online(mem);
269 		break;
270 	case MEM_OFFLINE:
271 		ret = memory_block_offline(mem);
272 		break;
273 	default:
274 		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
275 		     "%ld\n", __func__, mem->start_section_nr, action, action);
276 		ret = -EINVAL;
277 	}
278 
279 	return ret;
280 }
281 
282 static int memory_block_change_state(struct memory_block *mem,
283 		unsigned long to_state, unsigned long from_state_req)
284 {
285 	int ret = 0;
286 
287 	if (mem->state != from_state_req)
288 		return -EINVAL;
289 
290 	if (to_state == MEM_OFFLINE)
291 		mem->state = MEM_GOING_OFFLINE;
292 
293 	ret = memory_block_action(mem, to_state);
294 	mem->state = ret ? from_state_req : to_state;
295 
296 	return ret;
297 }
298 
299 /* The device lock serializes operations on memory_subsys_[online|offline] */
300 static int memory_subsys_online(struct device *dev)
301 {
302 	struct memory_block *mem = to_memory_block(dev);
303 	int ret;
304 
305 	if (mem->state == MEM_ONLINE)
306 		return 0;
307 
308 	/*
309 	 * When called via device_online() without configuring the online_type,
310 	 * we want to default to MMOP_ONLINE.
311 	 */
312 	if (mem->online_type == MMOP_OFFLINE)
313 		mem->online_type = MMOP_ONLINE;
314 
315 	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
316 	mem->online_type = MMOP_OFFLINE;
317 
318 	return ret;
319 }
320 
321 static int memory_subsys_offline(struct device *dev)
322 {
323 	struct memory_block *mem = to_memory_block(dev);
324 
325 	if (mem->state == MEM_OFFLINE)
326 		return 0;
327 
328 	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
329 }
330 
331 static ssize_t state_store(struct device *dev, struct device_attribute *attr,
332 			   const char *buf, size_t count)
333 {
334 	const int online_type = mhp_online_type_from_str(buf);
335 	struct memory_block *mem = to_memory_block(dev);
336 	int ret;
337 
338 	if (online_type < 0)
339 		return -EINVAL;
340 
341 	ret = lock_device_hotplug_sysfs();
342 	if (ret)
343 		return ret;
344 
345 	switch (online_type) {
346 	case MMOP_ONLINE_KERNEL:
347 	case MMOP_ONLINE_MOVABLE:
348 	case MMOP_ONLINE:
349 		/* mem->online_type is protected by device_hotplug_lock */
350 		mem->online_type = online_type;
351 		ret = device_online(&mem->dev);
352 		break;
353 	case MMOP_OFFLINE:
354 		ret = device_offline(&mem->dev);
355 		break;
356 	default:
357 		ret = -EINVAL; /* should never happen */
358 	}
359 
360 	unlock_device_hotplug();
361 
362 	if (ret < 0)
363 		return ret;
364 	if (ret)
365 		return -EINVAL;
366 
367 	return count;
368 }
369 
370 /*
371  * Legacy interface that we cannot remove: s390x exposes the storage increment
372  * covered by a memory block, allowing for identifying which memory blocks
373  * comprise a storage increment. Since a memory block spans complete
374  * storage increments nowadays, this interface is basically unused. Other
375  * archs never exposed != 0.
376  */
377 static ssize_t phys_device_show(struct device *dev,
378 				struct device_attribute *attr, char *buf)
379 {
380 	struct memory_block *mem = to_memory_block(dev);
381 	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
382 
383 	return sysfs_emit(buf, "%d\n",
384 			  arch_get_memory_phys_device(start_pfn));
385 }
386 
387 #ifdef CONFIG_MEMORY_HOTREMOVE
388 static int print_allowed_zone(char *buf, int len, int nid,
389 			      struct memory_group *group,
390 			      unsigned long start_pfn, unsigned long nr_pages,
391 			      int online_type, struct zone *default_zone)
392 {
393 	struct zone *zone;
394 
395 	zone = zone_for_pfn_range(online_type, nid, group, start_pfn, nr_pages);
396 	if (zone == default_zone)
397 		return 0;
398 
399 	return sysfs_emit_at(buf, len, " %s", zone->name);
400 }
401 
402 static ssize_t valid_zones_show(struct device *dev,
403 				struct device_attribute *attr, char *buf)
404 {
405 	struct memory_block *mem = to_memory_block(dev);
406 	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
407 	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
408 	struct memory_group *group = mem->group;
409 	struct zone *default_zone;
410 	int nid = mem->nid;
411 	int len = 0;
412 
413 	/*
414 	 * Check the existing zone. Make sure that we do that only on the
415 	 * online nodes otherwise the page_zone is not reliable
416 	 */
417 	if (mem->state == MEM_ONLINE) {
418 		/*
419 		 * If !mem->zone, the memory block spans multiple zones and
420 		 * cannot get offlined.
421 		 */
422 		default_zone = mem->zone;
423 		if (!default_zone)
424 			return sysfs_emit(buf, "%s\n", "none");
425 		len += sysfs_emit_at(buf, len, "%s", default_zone->name);
426 		goto out;
427 	}
428 
429 	default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, group,
430 					  start_pfn, nr_pages);
431 
432 	len += sysfs_emit_at(buf, len, "%s", default_zone->name);
433 	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
434 				  MMOP_ONLINE_KERNEL, default_zone);
435 	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
436 				  MMOP_ONLINE_MOVABLE, default_zone);
437 out:
438 	len += sysfs_emit_at(buf, len, "\n");
439 	return len;
440 }
441 static DEVICE_ATTR_RO(valid_zones);
442 #endif
443 
444 static DEVICE_ATTR_RO(phys_index);
445 static DEVICE_ATTR_RW(state);
446 static DEVICE_ATTR_RO(phys_device);
447 static DEVICE_ATTR_RO(removable);
448 
449 /*
450  * Show the memory block size (shared by all memory blocks).
451  */
452 static ssize_t block_size_bytes_show(struct device *dev,
453 				     struct device_attribute *attr, char *buf)
454 {
455 	return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
456 }
457 
458 static DEVICE_ATTR_RO(block_size_bytes);
459 
460 /*
461  * Memory auto online policy.
462  */
463 
464 static ssize_t auto_online_blocks_show(struct device *dev,
465 				       struct device_attribute *attr, char *buf)
466 {
467 	return sysfs_emit(buf, "%s\n",
468 			  online_type_to_str[mhp_default_online_type]);
469 }
470 
471 static ssize_t auto_online_blocks_store(struct device *dev,
472 					struct device_attribute *attr,
473 					const char *buf, size_t count)
474 {
475 	const int online_type = mhp_online_type_from_str(buf);
476 
477 	if (online_type < 0)
478 		return -EINVAL;
479 
480 	mhp_default_online_type = online_type;
481 	return count;
482 }
483 
484 static DEVICE_ATTR_RW(auto_online_blocks);
485 
486 /*
487  * Some architectures will have custom drivers to do this, and
488  * will not need to do it from userspace.  The fake hot-add code
489  * as well as ppc64 will do all of their discovery in userspace
490  * and will require this interface.
491  */
492 #ifdef CONFIG_ARCH_MEMORY_PROBE
493 static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
494 			   const char *buf, size_t count)
495 {
496 	u64 phys_addr;
497 	int nid, ret;
498 	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
499 
500 	ret = kstrtoull(buf, 0, &phys_addr);
501 	if (ret)
502 		return ret;
503 
504 	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
505 		return -EINVAL;
506 
507 	ret = lock_device_hotplug_sysfs();
508 	if (ret)
509 		return ret;
510 
511 	nid = memory_add_physaddr_to_nid(phys_addr);
512 	ret = __add_memory(nid, phys_addr,
513 			   MIN_MEMORY_BLOCK_SIZE * sections_per_block,
514 			   MHP_NONE);
515 
516 	if (ret)
517 		goto out;
518 
519 	ret = count;
520 out:
521 	unlock_device_hotplug();
522 	return ret;
523 }
524 
525 static DEVICE_ATTR_WO(probe);
526 #endif
527 
528 #ifdef CONFIG_MEMORY_FAILURE
529 /*
530  * Support for offlining pages of memory
531  */
532 
533 /* Soft offline a page */
534 static ssize_t soft_offline_page_store(struct device *dev,
535 				       struct device_attribute *attr,
536 				       const char *buf, size_t count)
537 {
538 	int ret;
539 	u64 pfn;
540 	if (!capable(CAP_SYS_ADMIN))
541 		return -EPERM;
542 	if (kstrtoull(buf, 0, &pfn) < 0)
543 		return -EINVAL;
544 	pfn >>= PAGE_SHIFT;
545 	ret = soft_offline_page(pfn, 0);
546 	return ret == 0 ? count : ret;
547 }
548 
549 /* Forcibly offline a page, including killing processes. */
550 static ssize_t hard_offline_page_store(struct device *dev,
551 				       struct device_attribute *attr,
552 				       const char *buf, size_t count)
553 {
554 	int ret;
555 	u64 pfn;
556 	if (!capable(CAP_SYS_ADMIN))
557 		return -EPERM;
558 	if (kstrtoull(buf, 0, &pfn) < 0)
559 		return -EINVAL;
560 	pfn >>= PAGE_SHIFT;
561 	ret = memory_failure(pfn, 0);
562 	if (ret == -EOPNOTSUPP)
563 		ret = 0;
564 	return ret ? ret : count;
565 }
566 
567 static DEVICE_ATTR_WO(soft_offline_page);
568 static DEVICE_ATTR_WO(hard_offline_page);
569 #endif
570 
571 /* See phys_device_show(). */
572 int __weak arch_get_memory_phys_device(unsigned long start_pfn)
573 {
574 	return 0;
575 }
576 
577 /*
578  * A reference for the returned memory block device is acquired.
579  *
580  * Called under device_hotplug_lock.
581  */
582 static struct memory_block *find_memory_block_by_id(unsigned long block_id)
583 {
584 	struct memory_block *mem;
585 
586 	mem = xa_load(&memory_blocks, block_id);
587 	if (mem)
588 		get_device(&mem->dev);
589 	return mem;
590 }
591 
592 /*
593  * Called under device_hotplug_lock.
594  */
595 struct memory_block *find_memory_block(unsigned long section_nr)
596 {
597 	unsigned long block_id = memory_block_id(section_nr);
598 
599 	return find_memory_block_by_id(block_id);
600 }
601 
602 static struct attribute *memory_memblk_attrs[] = {
603 	&dev_attr_phys_index.attr,
604 	&dev_attr_state.attr,
605 	&dev_attr_phys_device.attr,
606 	&dev_attr_removable.attr,
607 #ifdef CONFIG_MEMORY_HOTREMOVE
608 	&dev_attr_valid_zones.attr,
609 #endif
610 	NULL
611 };
612 
613 static const struct attribute_group memory_memblk_attr_group = {
614 	.attrs = memory_memblk_attrs,
615 };
616 
617 static const struct attribute_group *memory_memblk_attr_groups[] = {
618 	&memory_memblk_attr_group,
619 	NULL,
620 };
621 
622 static int __add_memory_block(struct memory_block *memory)
623 {
624 	int ret;
625 
626 	memory->dev.bus = &memory_subsys;
627 	memory->dev.id = memory->start_section_nr / sections_per_block;
628 	memory->dev.release = memory_block_release;
629 	memory->dev.groups = memory_memblk_attr_groups;
630 	memory->dev.offline = memory->state == MEM_OFFLINE;
631 
632 	ret = device_register(&memory->dev);
633 	if (ret) {
634 		put_device(&memory->dev);
635 		return ret;
636 	}
637 	ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
638 			      GFP_KERNEL));
639 	if (ret)
640 		device_unregister(&memory->dev);
641 
642 	return ret;
643 }
644 
645 static struct zone *early_node_zone_for_memory_block(struct memory_block *mem,
646 						     int nid)
647 {
648 	const unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
649 	const unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
650 	struct zone *zone, *matching_zone = NULL;
651 	pg_data_t *pgdat = NODE_DATA(nid);
652 	int i;
653 
654 	/*
655 	 * This logic only works for early memory, when the applicable zones
656 	 * already span the memory block. We don't expect overlapping zones on
657 	 * a single node for early memory. So if we're told that some PFNs
658 	 * of a node fall into this memory block, we can assume that all node
659 	 * zones that intersect with the memory block are actually applicable.
660 	 * No need to look at the memmap.
661 	 */
662 	for (i = 0; i < MAX_NR_ZONES; i++) {
663 		zone = pgdat->node_zones + i;
664 		if (!populated_zone(zone))
665 			continue;
666 		if (!zone_intersects(zone, start_pfn, nr_pages))
667 			continue;
668 		if (!matching_zone) {
669 			matching_zone = zone;
670 			continue;
671 		}
672 		/* Spans multiple zones ... */
673 		matching_zone = NULL;
674 		break;
675 	}
676 	return matching_zone;
677 }
678 
679 #ifdef CONFIG_NUMA
680 /**
681  * memory_block_add_nid() - Indicate that system RAM falling into this memory
682  *			    block device (partially) belongs to the given node.
683  * @mem: The memory block device.
684  * @nid: The node id.
685  * @context: The memory initialization context.
686  *
687  * Indicate that system RAM falling into this memory block (partially) belongs
688  * to the given node. If the context indicates ("early") that we are adding the
689  * node during node device subsystem initialization, this will also properly
690  * set/adjust mem->zone based on the zone ranges of the given node.
691  */
692 void memory_block_add_nid(struct memory_block *mem, int nid,
693 			  enum meminit_context context)
694 {
695 	if (context == MEMINIT_EARLY && mem->nid != nid) {
696 		/*
697 		 * For early memory we have to determine the zone when setting
698 		 * the node id and handle multiple nodes spanning a single
699 		 * memory block by indicate via zone == NULL that we're not
700 		 * dealing with a single zone. So if we're setting the node id
701 		 * the first time, determine if there is a single zone. If we're
702 		 * setting the node id a second time to a different node,
703 		 * invalidate the single detected zone.
704 		 */
705 		if (mem->nid == NUMA_NO_NODE)
706 			mem->zone = early_node_zone_for_memory_block(mem, nid);
707 		else
708 			mem->zone = NULL;
709 	}
710 
711 	/*
712 	 * If this memory block spans multiple nodes, we only indicate
713 	 * the last processed node. If we span multiple nodes (not applicable
714 	 * to hotplugged memory), zone == NULL will prohibit memory offlining
715 	 * and consequently unplug.
716 	 */
717 	mem->nid = nid;
718 }
719 #endif
720 
721 static int add_memory_block(unsigned long block_id, unsigned long state,
722 			    unsigned long nr_vmemmap_pages,
723 			    struct memory_group *group)
724 {
725 	struct memory_block *mem;
726 	int ret = 0;
727 
728 	mem = find_memory_block_by_id(block_id);
729 	if (mem) {
730 		put_device(&mem->dev);
731 		return -EEXIST;
732 	}
733 	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
734 	if (!mem)
735 		return -ENOMEM;
736 
737 	mem->start_section_nr = block_id * sections_per_block;
738 	mem->state = state;
739 	mem->nid = NUMA_NO_NODE;
740 	mem->nr_vmemmap_pages = nr_vmemmap_pages;
741 	INIT_LIST_HEAD(&mem->group_next);
742 
743 #ifndef CONFIG_NUMA
744 	if (state == MEM_ONLINE)
745 		/*
746 		 * MEM_ONLINE at this point implies early memory. With NUMA,
747 		 * we'll determine the zone when setting the node id via
748 		 * memory_block_add_nid(). Memory hotplug updated the zone
749 		 * manually when memory onlining/offlining succeeds.
750 		 */
751 		mem->zone = early_node_zone_for_memory_block(mem, NUMA_NO_NODE);
752 #endif /* CONFIG_NUMA */
753 
754 	ret = __add_memory_block(mem);
755 	if (ret)
756 		return ret;
757 
758 	if (group) {
759 		mem->group = group;
760 		list_add(&mem->group_next, &group->memory_blocks);
761 	}
762 
763 	return 0;
764 }
765 
766 static int __init add_boot_memory_block(unsigned long base_section_nr)
767 {
768 	int section_count = 0;
769 	unsigned long nr;
770 
771 	for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
772 	     nr++)
773 		if (present_section_nr(nr))
774 			section_count++;
775 
776 	if (section_count == 0)
777 		return 0;
778 	return add_memory_block(memory_block_id(base_section_nr),
779 				MEM_ONLINE, 0,  NULL);
780 }
781 
782 static int add_hotplug_memory_block(unsigned long block_id,
783 				    unsigned long nr_vmemmap_pages,
784 				    struct memory_group *group)
785 {
786 	return add_memory_block(block_id, MEM_OFFLINE, nr_vmemmap_pages, group);
787 }
788 
789 static void remove_memory_block(struct memory_block *memory)
790 {
791 	if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
792 		return;
793 
794 	WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
795 
796 	if (memory->group) {
797 		list_del(&memory->group_next);
798 		memory->group = NULL;
799 	}
800 
801 	/* drop the ref. we got via find_memory_block() */
802 	put_device(&memory->dev);
803 	device_unregister(&memory->dev);
804 }
805 
806 /*
807  * Create memory block devices for the given memory area. Start and size
808  * have to be aligned to memory block granularity. Memory block devices
809  * will be initialized as offline.
810  *
811  * Called under device_hotplug_lock.
812  */
813 int create_memory_block_devices(unsigned long start, unsigned long size,
814 				unsigned long vmemmap_pages,
815 				struct memory_group *group)
816 {
817 	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
818 	unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
819 	struct memory_block *mem;
820 	unsigned long block_id;
821 	int ret = 0;
822 
823 	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
824 			 !IS_ALIGNED(size, memory_block_size_bytes())))
825 		return -EINVAL;
826 
827 	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
828 		ret = add_hotplug_memory_block(block_id, vmemmap_pages, group);
829 		if (ret)
830 			break;
831 	}
832 	if (ret) {
833 		end_block_id = block_id;
834 		for (block_id = start_block_id; block_id != end_block_id;
835 		     block_id++) {
836 			mem = find_memory_block_by_id(block_id);
837 			if (WARN_ON_ONCE(!mem))
838 				continue;
839 			remove_memory_block(mem);
840 		}
841 	}
842 	return ret;
843 }
844 
845 /*
846  * Remove memory block devices for the given memory area. Start and size
847  * have to be aligned to memory block granularity. Memory block devices
848  * have to be offline.
849  *
850  * Called under device_hotplug_lock.
851  */
852 void remove_memory_block_devices(unsigned long start, unsigned long size)
853 {
854 	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
855 	const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
856 	struct memory_block *mem;
857 	unsigned long block_id;
858 
859 	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
860 			 !IS_ALIGNED(size, memory_block_size_bytes())))
861 		return;
862 
863 	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
864 		mem = find_memory_block_by_id(block_id);
865 		if (WARN_ON_ONCE(!mem))
866 			continue;
867 		unregister_memory_block_under_nodes(mem);
868 		remove_memory_block(mem);
869 	}
870 }
871 
872 /* return true if the memory block is offlined, otherwise, return false */
873 bool is_memblock_offlined(struct memory_block *mem)
874 {
875 	return mem->state == MEM_OFFLINE;
876 }
877 
878 static struct attribute *memory_root_attrs[] = {
879 #ifdef CONFIG_ARCH_MEMORY_PROBE
880 	&dev_attr_probe.attr,
881 #endif
882 
883 #ifdef CONFIG_MEMORY_FAILURE
884 	&dev_attr_soft_offline_page.attr,
885 	&dev_attr_hard_offline_page.attr,
886 #endif
887 
888 	&dev_attr_block_size_bytes.attr,
889 	&dev_attr_auto_online_blocks.attr,
890 	NULL
891 };
892 
893 static const struct attribute_group memory_root_attr_group = {
894 	.attrs = memory_root_attrs,
895 };
896 
897 static const struct attribute_group *memory_root_attr_groups[] = {
898 	&memory_root_attr_group,
899 	NULL,
900 };
901 
902 /*
903  * Initialize the sysfs support for memory devices. At the time this function
904  * is called, we cannot have concurrent creation/deletion of memory block
905  * devices, the device_hotplug_lock is not needed.
906  */
907 void __init memory_dev_init(void)
908 {
909 	int ret;
910 	unsigned long block_sz, nr;
911 
912 	/* Validate the configured memory block size */
913 	block_sz = memory_block_size_bytes();
914 	if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
915 		panic("Memory block size not suitable: 0x%lx\n", block_sz);
916 	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
917 
918 	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
919 	if (ret)
920 		panic("%s() failed to register subsystem: %d\n", __func__, ret);
921 
922 	/*
923 	 * Create entries for memory sections that were found
924 	 * during boot and have been initialized
925 	 */
926 	for (nr = 0; nr <= __highest_present_section_nr;
927 	     nr += sections_per_block) {
928 		ret = add_boot_memory_block(nr);
929 		if (ret)
930 			panic("%s() failed to add memory block: %d\n", __func__,
931 			      ret);
932 	}
933 }
934 
935 /**
936  * walk_memory_blocks - walk through all present memory blocks overlapped
937  *			by the range [start, start + size)
938  *
939  * @start: start address of the memory range
940  * @size: size of the memory range
941  * @arg: argument passed to func
942  * @func: callback for each memory section walked
943  *
944  * This function walks through all present memory blocks overlapped by the
945  * range [start, start + size), calling func on each memory block.
946  *
947  * In case func() returns an error, walking is aborted and the error is
948  * returned.
949  *
950  * Called under device_hotplug_lock.
951  */
952 int walk_memory_blocks(unsigned long start, unsigned long size,
953 		       void *arg, walk_memory_blocks_func_t func)
954 {
955 	const unsigned long start_block_id = phys_to_block_id(start);
956 	const unsigned long end_block_id = phys_to_block_id(start + size - 1);
957 	struct memory_block *mem;
958 	unsigned long block_id;
959 	int ret = 0;
960 
961 	if (!size)
962 		return 0;
963 
964 	for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
965 		mem = find_memory_block_by_id(block_id);
966 		if (!mem)
967 			continue;
968 
969 		ret = func(mem, arg);
970 		put_device(&mem->dev);
971 		if (ret)
972 			break;
973 	}
974 	return ret;
975 }
976 
977 struct for_each_memory_block_cb_data {
978 	walk_memory_blocks_func_t func;
979 	void *arg;
980 };
981 
982 static int for_each_memory_block_cb(struct device *dev, void *data)
983 {
984 	struct memory_block *mem = to_memory_block(dev);
985 	struct for_each_memory_block_cb_data *cb_data = data;
986 
987 	return cb_data->func(mem, cb_data->arg);
988 }
989 
990 /**
991  * for_each_memory_block - walk through all present memory blocks
992  *
993  * @arg: argument passed to func
994  * @func: callback for each memory block walked
995  *
996  * This function walks through all present memory blocks, calling func on
997  * each memory block.
998  *
999  * In case func() returns an error, walking is aborted and the error is
1000  * returned.
1001  */
1002 int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
1003 {
1004 	struct for_each_memory_block_cb_data cb_data = {
1005 		.func = func,
1006 		.arg = arg,
1007 	};
1008 
1009 	return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
1010 				for_each_memory_block_cb);
1011 }
1012 
1013 /*
1014  * This is an internal helper to unify allocation and initialization of
1015  * memory groups. Note that the passed memory group will be copied to a
1016  * dynamically allocated memory group. After this call, the passed
1017  * memory group should no longer be used.
1018  */
1019 static int memory_group_register(struct memory_group group)
1020 {
1021 	struct memory_group *new_group;
1022 	uint32_t mgid;
1023 	int ret;
1024 
1025 	if (!node_possible(group.nid))
1026 		return -EINVAL;
1027 
1028 	new_group = kzalloc(sizeof(group), GFP_KERNEL);
1029 	if (!new_group)
1030 		return -ENOMEM;
1031 	*new_group = group;
1032 	INIT_LIST_HEAD(&new_group->memory_blocks);
1033 
1034 	ret = xa_alloc(&memory_groups, &mgid, new_group, xa_limit_31b,
1035 		       GFP_KERNEL);
1036 	if (ret) {
1037 		kfree(new_group);
1038 		return ret;
1039 	} else if (group.is_dynamic) {
1040 		xa_set_mark(&memory_groups, mgid, MEMORY_GROUP_MARK_DYNAMIC);
1041 	}
1042 	return mgid;
1043 }
1044 
1045 /**
1046  * memory_group_register_static() - Register a static memory group.
1047  * @nid: The node id.
1048  * @max_pages: The maximum number of pages we'll have in this static memory
1049  *	       group.
1050  *
1051  * Register a new static memory group and return the memory group id.
1052  * All memory in the group belongs to a single unit, such as a DIMM. All
1053  * memory belonging to a static memory group is added in one go to be removed
1054  * in one go -- it's static.
1055  *
1056  * Returns an error if out of memory, if the node id is invalid, if no new
1057  * memory groups can be registered, or if max_pages is invalid (0). Otherwise,
1058  * returns the new memory group id.
1059  */
1060 int memory_group_register_static(int nid, unsigned long max_pages)
1061 {
1062 	struct memory_group group = {
1063 		.nid = nid,
1064 		.s = {
1065 			.max_pages = max_pages,
1066 		},
1067 	};
1068 
1069 	if (!max_pages)
1070 		return -EINVAL;
1071 	return memory_group_register(group);
1072 }
1073 EXPORT_SYMBOL_GPL(memory_group_register_static);
1074 
1075 /**
1076  * memory_group_register_dynamic() - Register a dynamic memory group.
1077  * @nid: The node id.
1078  * @unit_pages: Unit in pages in which is memory added/removed in this dynamic
1079  *		memory group.
1080  *
1081  * Register a new dynamic memory group and return the memory group id.
1082  * Memory within a dynamic memory group is added/removed dynamically
1083  * in unit_pages.
1084  *
1085  * Returns an error if out of memory, if the node id is invalid, if no new
1086  * memory groups can be registered, or if unit_pages is invalid (0, not a
1087  * power of two, smaller than a single memory block). Otherwise, returns the
1088  * new memory group id.
1089  */
1090 int memory_group_register_dynamic(int nid, unsigned long unit_pages)
1091 {
1092 	struct memory_group group = {
1093 		.nid = nid,
1094 		.is_dynamic = true,
1095 		.d = {
1096 			.unit_pages = unit_pages,
1097 		},
1098 	};
1099 
1100 	if (!unit_pages || !is_power_of_2(unit_pages) ||
1101 	    unit_pages < PHYS_PFN(memory_block_size_bytes()))
1102 		return -EINVAL;
1103 	return memory_group_register(group);
1104 }
1105 EXPORT_SYMBOL_GPL(memory_group_register_dynamic);
1106 
1107 /**
1108  * memory_group_unregister() - Unregister a memory group.
1109  * @mgid: the memory group id
1110  *
1111  * Unregister a memory group. If any memory block still belongs to this
1112  * memory group, unregistering will fail.
1113  *
1114  * Returns -EINVAL if the memory group id is invalid, returns -EBUSY if some
1115  * memory blocks still belong to this memory group and returns 0 if
1116  * unregistering succeeded.
1117  */
1118 int memory_group_unregister(int mgid)
1119 {
1120 	struct memory_group *group;
1121 
1122 	if (mgid < 0)
1123 		return -EINVAL;
1124 
1125 	group = xa_load(&memory_groups, mgid);
1126 	if (!group)
1127 		return -EINVAL;
1128 	if (!list_empty(&group->memory_blocks))
1129 		return -EBUSY;
1130 	xa_erase(&memory_groups, mgid);
1131 	kfree(group);
1132 	return 0;
1133 }
1134 EXPORT_SYMBOL_GPL(memory_group_unregister);
1135 
1136 /*
1137  * This is an internal helper only to be used in core memory hotplug code to
1138  * lookup a memory group. We don't care about locking, as we don't expect a
1139  * memory group to get unregistered while adding memory to it -- because
1140  * the group and the memory is managed by the same driver.
1141  */
1142 struct memory_group *memory_group_find_by_id(int mgid)
1143 {
1144 	return xa_load(&memory_groups, mgid);
1145 }
1146 
1147 /*
1148  * This is an internal helper only to be used in core memory hotplug code to
1149  * walk all dynamic memory groups excluding a given memory group, either
1150  * belonging to a specific node, or belonging to any node.
1151  */
1152 int walk_dynamic_memory_groups(int nid, walk_memory_groups_func_t func,
1153 			       struct memory_group *excluded, void *arg)
1154 {
1155 	struct memory_group *group;
1156 	unsigned long index;
1157 	int ret = 0;
1158 
1159 	xa_for_each_marked(&memory_groups, index, group,
1160 			   MEMORY_GROUP_MARK_DYNAMIC) {
1161 		if (group == excluded)
1162 			continue;
1163 #ifdef CONFIG_NUMA
1164 		if (nid != NUMA_NO_NODE && group->nid != nid)
1165 			continue;
1166 #endif /* CONFIG_NUMA */
1167 		ret = func(group, arg);
1168 		if (ret)
1169 			break;
1170 	}
1171 	return ret;
1172 }
1173