Linux Audio

Check our new training course

Loading...
v4.17
  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/mutex.h>
 23#include <linux/stat.h>
 24#include <linux/slab.h>
 25
 26#include <linux/atomic.h>
 27#include <linux/uaccess.h>
 28
 29static DEFINE_MUTEX(mem_sysfs_mutex);
 30
 31#define MEMORY_CLASS_NAME	"memory"
 32
 33#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
 34
 35static int sections_per_block;
 36
 37static inline int base_memory_block_id(int section_nr)
 38{
 39	return section_nr / sections_per_block;
 40}
 41
 
 
 
 
 
 
 
 
 
 
 42static int memory_subsys_online(struct device *dev);
 43static int memory_subsys_offline(struct device *dev);
 44
 45static struct bus_type memory_subsys = {
 46	.name = MEMORY_CLASS_NAME,
 47	.dev_name = MEMORY_CLASS_NAME,
 48	.online = memory_subsys_online,
 49	.offline = memory_subsys_offline,
 50};
 51
 52static BLOCKING_NOTIFIER_HEAD(memory_chain);
 53
 54int register_memory_notifier(struct notifier_block *nb)
 55{
 56	return blocking_notifier_chain_register(&memory_chain, nb);
 57}
 58EXPORT_SYMBOL(register_memory_notifier);
 59
 60void unregister_memory_notifier(struct notifier_block *nb)
 61{
 62	blocking_notifier_chain_unregister(&memory_chain, nb);
 63}
 64EXPORT_SYMBOL(unregister_memory_notifier);
 65
 66static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
 67
 68int register_memory_isolate_notifier(struct notifier_block *nb)
 69{
 70	return atomic_notifier_chain_register(&memory_isolate_chain, nb);
 71}
 72EXPORT_SYMBOL(register_memory_isolate_notifier);
 73
 74void unregister_memory_isolate_notifier(struct notifier_block *nb)
 75{
 76	atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
 77}
 78EXPORT_SYMBOL(unregister_memory_isolate_notifier);
 79
 80static void memory_block_release(struct device *dev)
 81{
 82	struct memory_block *mem = to_memory_block(dev);
 83
 84	kfree(mem);
 85}
 86
 87unsigned long __weak memory_block_size_bytes(void)
 88{
 89	return MIN_MEMORY_BLOCK_SIZE;
 90}
 91
 92static unsigned long get_memory_block_size(void)
 93{
 94	unsigned long block_sz;
 95
 96	block_sz = memory_block_size_bytes();
 97
 98	/* Validate blk_sz is a power of 2 and not less than section size */
 99	if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
100		WARN_ON(1);
101		block_sz = MIN_MEMORY_BLOCK_SIZE;
102	}
103
104	return block_sz;
105}
106
107/*
108 * use this as the physical section index that this memsection
109 * uses.
110 */
111
112static ssize_t show_mem_start_phys_index(struct device *dev,
113			struct device_attribute *attr, char *buf)
114{
115	struct memory_block *mem = to_memory_block(dev);
116	unsigned long phys_index;
117
118	phys_index = mem->start_section_nr / sections_per_block;
119	return sprintf(buf, "%08lx\n", phys_index);
120}
121
122/*
123 * Show whether the section of memory is likely to be hot-removable
 
 
 
124 */
125static ssize_t show_mem_removable(struct device *dev,
126			struct device_attribute *attr, char *buf)
127{
128	unsigned long i, pfn;
129	int ret = 1;
130	struct memory_block *mem = to_memory_block(dev);
 
 
131
132	if (mem->state != MEM_ONLINE)
133		goto out;
134
135	for (i = 0; i < sections_per_block; i++) {
136		if (!present_section_nr(mem->start_section_nr + i))
137			continue;
138		pfn = section_nr_to_pfn(mem->start_section_nr + i);
139		ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
140	}
141
142out:
143	return sprintf(buf, "%d\n", ret);
144}
145
146/*
147 * online, offline, going offline, etc.
148 */
149static ssize_t show_mem_state(struct device *dev,
150			struct device_attribute *attr, char *buf)
151{
152	struct memory_block *mem = to_memory_block(dev);
153	ssize_t len = 0;
154
155	/*
156	 * We can probably put these states in a nice little array
157	 * so that they're not open-coded
158	 */
159	switch (mem->state) {
160	case MEM_ONLINE:
161		len = sprintf(buf, "online\n");
162		break;
163	case MEM_OFFLINE:
164		len = sprintf(buf, "offline\n");
165		break;
166	case MEM_GOING_OFFLINE:
167		len = sprintf(buf, "going-offline\n");
168		break;
169	default:
170		len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
171				mem->state);
172		WARN_ON(1);
173		break;
174	}
175
176	return len;
177}
178
179int memory_notify(unsigned long val, void *v)
180{
181	return blocking_notifier_call_chain(&memory_chain, val, v);
182}
183
184int memory_isolate_notify(unsigned long val, void *v)
185{
186	return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
187}
188
189/*
190 * The probe routines leave the pages uninitialized, just as the bootmem code
191 * does. Make sure we do not access them, but instead use only information from
192 * within sections.
193 */
194static bool pages_correctly_probed(unsigned long start_pfn)
195{
196	unsigned long section_nr = pfn_to_section_nr(start_pfn);
197	unsigned long section_nr_end = section_nr + sections_per_block;
198	unsigned long pfn = start_pfn;
199
200	/*
201	 * memmap between sections is not contiguous except with
202	 * SPARSEMEM_VMEMMAP. We lookup the page once per section
203	 * and assume memmap is contiguous within each section
204	 */
205	for (; section_nr < section_nr_end; section_nr++) {
206		if (WARN_ON_ONCE(!pfn_valid(pfn)))
207			return false;
208
209		if (!present_section_nr(section_nr)) {
210			pr_warn("section %ld pfn[%lx, %lx) not present",
211				section_nr, pfn, pfn + PAGES_PER_SECTION);
212			return false;
213		} else if (!valid_section_nr(section_nr)) {
214			pr_warn("section %ld pfn[%lx, %lx) no valid memmap",
215				section_nr, pfn, pfn + PAGES_PER_SECTION);
216			return false;
217		} else if (online_section_nr(section_nr)) {
218			pr_warn("section %ld pfn[%lx, %lx) is already online",
219				section_nr, pfn, pfn + PAGES_PER_SECTION);
220			return false;
221		}
222		pfn += PAGES_PER_SECTION;
223	}
224
225	return true;
226}
227
228/*
229 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
230 * OK to have direct references to sparsemem variables in here.
231 * Must already be protected by mem_hotplug_begin().
232 */
233static int
234memory_block_action(unsigned long phys_index, unsigned long action, int online_type)
 
235{
236	unsigned long start_pfn;
237	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
238	int ret;
239
240	start_pfn = section_nr_to_pfn(phys_index);
241
242	switch (action) {
243	case MEM_ONLINE:
244		if (!pages_correctly_probed(start_pfn))
245			return -EBUSY;
246
247		ret = online_pages(start_pfn, nr_pages, online_type);
248		break;
249	case MEM_OFFLINE:
250		ret = offline_pages(start_pfn, nr_pages);
251		break;
252	default:
253		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
254		     "%ld\n", __func__, phys_index, action, action);
255		ret = -EINVAL;
256	}
257
258	return ret;
259}
260
261static int memory_block_change_state(struct memory_block *mem,
262		unsigned long to_state, unsigned long from_state_req)
263{
264	int ret = 0;
265
266	if (mem->state != from_state_req)
267		return -EINVAL;
268
269	if (to_state == MEM_OFFLINE)
270		mem->state = MEM_GOING_OFFLINE;
271
272	ret = memory_block_action(mem->start_section_nr, to_state,
273				mem->online_type);
274
275	mem->state = ret ? from_state_req : to_state;
276
277	return ret;
278}
279
280/* The device lock serializes operations on memory_subsys_[online|offline] */
281static int memory_subsys_online(struct device *dev)
282{
283	struct memory_block *mem = to_memory_block(dev);
284	int ret;
285
286	if (mem->state == MEM_ONLINE)
287		return 0;
288
289	/*
290	 * If we are called from store_mem_state(), online_type will be
291	 * set >= 0 Otherwise we were called from the device online
292	 * attribute and need to set the online_type.
293	 */
294	if (mem->online_type < 0)
295		mem->online_type = MMOP_ONLINE_KEEP;
296
297	/* Already under protection of mem_hotplug_begin() */
298	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
299
300	/* clear online_type */
301	mem->online_type = -1;
302
303	return ret;
304}
305
306static int memory_subsys_offline(struct device *dev)
307{
308	struct memory_block *mem = to_memory_block(dev);
309
310	if (mem->state == MEM_OFFLINE)
311		return 0;
312
313	/* Can't offline block with non-present sections */
314	if (mem->section_count != sections_per_block)
315		return -EINVAL;
316
317	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
318}
319
320static ssize_t
321store_mem_state(struct device *dev,
322		struct device_attribute *attr, const char *buf, size_t count)
323{
324	struct memory_block *mem = to_memory_block(dev);
325	int ret, online_type;
326
327	ret = lock_device_hotplug_sysfs();
328	if (ret)
329		return ret;
330
331	if (sysfs_streq(buf, "online_kernel"))
332		online_type = MMOP_ONLINE_KERNEL;
333	else if (sysfs_streq(buf, "online_movable"))
334		online_type = MMOP_ONLINE_MOVABLE;
335	else if (sysfs_streq(buf, "online"))
336		online_type = MMOP_ONLINE_KEEP;
337	else if (sysfs_streq(buf, "offline"))
338		online_type = MMOP_OFFLINE;
339	else {
340		ret = -EINVAL;
341		goto err;
342	}
343
344	/*
345	 * Memory hotplug needs to hold mem_hotplug_begin() for probe to find
346	 * the correct memory block to online before doing device_online(dev),
347	 * which will take dev->mutex.  Take the lock early to prevent an
348	 * inversion, memory_subsys_online() callbacks will be implemented by
349	 * assuming it's already protected.
350	 */
351	mem_hotplug_begin();
352
353	switch (online_type) {
354	case MMOP_ONLINE_KERNEL:
355	case MMOP_ONLINE_MOVABLE:
356	case MMOP_ONLINE_KEEP:
 
357		mem->online_type = online_type;
358		ret = device_online(&mem->dev);
359		break;
360	case MMOP_OFFLINE:
361		ret = device_offline(&mem->dev);
362		break;
363	default:
364		ret = -EINVAL; /* should never happen */
365	}
366
367	mem_hotplug_done();
368err:
369	unlock_device_hotplug();
370
371	if (ret < 0)
372		return ret;
373	if (ret)
374		return -EINVAL;
375
376	return count;
377}
378
379/*
380 * phys_device is a bad name for this.  What I really want
381 * is a way to differentiate between memory ranges that
382 * are part of physical devices that constitute
383 * a complete removable unit or fru.
384 * i.e. do these ranges belong to the same physical device,
385 * s.t. if I offline all of these sections I can then
386 * remove the physical device?
387 */
388static ssize_t show_phys_device(struct device *dev,
389				struct device_attribute *attr, char *buf)
390{
391	struct memory_block *mem = to_memory_block(dev);
392	return sprintf(buf, "%d\n", mem->phys_device);
393}
394
395#ifdef CONFIG_MEMORY_HOTREMOVE
396static void print_allowed_zone(char *buf, int nid, unsigned long start_pfn,
397		unsigned long nr_pages, int online_type,
398		struct zone *default_zone)
399{
400	struct zone *zone;
401
402	zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
403	if (zone != default_zone) {
404		strcat(buf, " ");
405		strcat(buf, zone->name);
406	}
407}
408
409static ssize_t show_valid_zones(struct device *dev,
410				struct device_attribute *attr, char *buf)
411{
412	struct memory_block *mem = to_memory_block(dev);
413	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
414	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
415	unsigned long valid_start_pfn, valid_end_pfn;
416	struct zone *default_zone;
417	int nid;
418
419	/*
420	 * The block contains more than one zone can not be offlined.
421	 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
422	 */
423	if (!test_pages_in_a_zone(start_pfn, start_pfn + nr_pages, &valid_start_pfn, &valid_end_pfn))
424		return sprintf(buf, "none\n");
425
426	start_pfn = valid_start_pfn;
427	nr_pages = valid_end_pfn - start_pfn;
428
429	/*
430	 * Check the existing zone. Make sure that we do that only on the
431	 * online nodes otherwise the page_zone is not reliable
432	 */
433	if (mem->state == MEM_ONLINE) {
 
 
 
 
 
 
 
 
434		strcat(buf, page_zone(pfn_to_page(start_pfn))->name);
435		goto out;
436	}
437
438	nid = pfn_to_nid(start_pfn);
439	default_zone = zone_for_pfn_range(MMOP_ONLINE_KEEP, nid, start_pfn, nr_pages);
440	strcat(buf, default_zone->name);
441
442	print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_KERNEL,
443			default_zone);
444	print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_MOVABLE,
445			default_zone);
446out:
447	strcat(buf, "\n");
448
449	return strlen(buf);
450}
451static DEVICE_ATTR(valid_zones, 0444, show_valid_zones, NULL);
452#endif
453
454static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
455static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
456static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
457static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
458
459/*
460 * Block size attribute stuff
461 */
462static ssize_t
463print_block_size(struct device *dev, struct device_attribute *attr,
464		 char *buf)
465{
466	return sprintf(buf, "%lx\n", get_memory_block_size());
467}
468
469static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
470
471/*
472 * Memory auto online policy.
473 */
474
475static ssize_t
476show_auto_online_blocks(struct device *dev, struct device_attribute *attr,
477			char *buf)
478{
479	if (memhp_auto_online)
480		return sprintf(buf, "online\n");
481	else
482		return sprintf(buf, "offline\n");
483}
484
485static ssize_t
486store_auto_online_blocks(struct device *dev, struct device_attribute *attr,
487			 const char *buf, size_t count)
488{
489	if (sysfs_streq(buf, "online"))
490		memhp_auto_online = true;
491	else if (sysfs_streq(buf, "offline"))
492		memhp_auto_online = false;
493	else
494		return -EINVAL;
495
496	return count;
497}
498
499static DEVICE_ATTR(auto_online_blocks, 0644, show_auto_online_blocks,
500		   store_auto_online_blocks);
501
502/*
503 * Some architectures will have custom drivers to do this, and
504 * will not need to do it from userspace.  The fake hot-add code
505 * as well as ppc64 will do all of their discovery in userspace
506 * and will require this interface.
507 */
508#ifdef CONFIG_ARCH_MEMORY_PROBE
509static ssize_t
510memory_probe_store(struct device *dev, struct device_attribute *attr,
511		   const char *buf, size_t count)
512{
513	u64 phys_addr;
514	int nid, ret;
515	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
516
517	ret = kstrtoull(buf, 0, &phys_addr);
518	if (ret)
519		return ret;
520
521	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
522		return -EINVAL;
523
 
 
 
 
524	nid = memory_add_physaddr_to_nid(phys_addr);
525	ret = add_memory(nid, phys_addr,
526			 MIN_MEMORY_BLOCK_SIZE * sections_per_block);
527
528	if (ret)
529		goto out;
530
531	ret = count;
532out:
 
533	return ret;
534}
535
536static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
537#endif
538
539#ifdef CONFIG_MEMORY_FAILURE
540/*
541 * Support for offlining pages of memory
542 */
543
544/* Soft offline a page */
545static ssize_t
546store_soft_offline_page(struct device *dev,
547			struct device_attribute *attr,
548			const char *buf, size_t count)
549{
550	int ret;
551	u64 pfn;
552	if (!capable(CAP_SYS_ADMIN))
553		return -EPERM;
554	if (kstrtoull(buf, 0, &pfn) < 0)
555		return -EINVAL;
556	pfn >>= PAGE_SHIFT;
557	if (!pfn_valid(pfn))
558		return -ENXIO;
 
 
 
559	ret = soft_offline_page(pfn_to_page(pfn), 0);
560	return ret == 0 ? count : ret;
561}
562
563/* Forcibly offline a page, including killing processes. */
564static ssize_t
565store_hard_offline_page(struct device *dev,
566			struct device_attribute *attr,
567			const char *buf, size_t count)
568{
569	int ret;
570	u64 pfn;
571	if (!capable(CAP_SYS_ADMIN))
572		return -EPERM;
573	if (kstrtoull(buf, 0, &pfn) < 0)
574		return -EINVAL;
575	pfn >>= PAGE_SHIFT;
576	ret = memory_failure(pfn, 0);
577	return ret ? ret : count;
578}
579
580static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
581static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
582#endif
583
584/*
585 * Note that phys_device is optional.  It is here to allow for
586 * differentiation between which *physical* devices each
587 * section belongs to...
588 */
589int __weak arch_get_memory_phys_device(unsigned long start_pfn)
590{
591	return 0;
592}
593
594/*
595 * A reference for the returned object is held and the reference for the
596 * hinted object is released.
597 */
598struct memory_block *find_memory_block_hinted(struct mem_section *section,
599					      struct memory_block *hint)
600{
601	int block_id = base_memory_block_id(__section_nr(section));
602	struct device *hintdev = hint ? &hint->dev : NULL;
603	struct device *dev;
604
605	dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
606	if (hint)
607		put_device(&hint->dev);
608	if (!dev)
609		return NULL;
610	return to_memory_block(dev);
611}
612
613/*
614 * For now, we have a linear search to go find the appropriate
615 * memory_block corresponding to a particular phys_index. If
616 * this gets to be a real problem, we can always use a radix
617 * tree or something here.
618 *
619 * This could be made generic for all device subsystems.
620 */
621struct memory_block *find_memory_block(struct mem_section *section)
622{
623	return find_memory_block_hinted(section, NULL);
 
 
624}
625
626static struct attribute *memory_memblk_attrs[] = {
627	&dev_attr_phys_index.attr,
628	&dev_attr_state.attr,
629	&dev_attr_phys_device.attr,
630	&dev_attr_removable.attr,
631#ifdef CONFIG_MEMORY_HOTREMOVE
632	&dev_attr_valid_zones.attr,
633#endif
634	NULL
635};
636
637static struct attribute_group memory_memblk_attr_group = {
638	.attrs = memory_memblk_attrs,
639};
640
641static const struct attribute_group *memory_memblk_attr_groups[] = {
642	&memory_memblk_attr_group,
643	NULL,
644};
645
646/*
647 * register_memory - Setup a sysfs device for a memory block
648 */
649static
650int register_memory(struct memory_block *memory)
651{
 
 
652	memory->dev.bus = &memory_subsys;
653	memory->dev.id = memory->start_section_nr / sections_per_block;
654	memory->dev.release = memory_block_release;
655	memory->dev.groups = memory_memblk_attr_groups;
656	memory->dev.offline = memory->state == MEM_OFFLINE;
657
658	return device_register(&memory->dev);
 
 
 
 
659}
660
661static int init_memory_block(struct memory_block **memory,
662			     struct mem_section *section, unsigned long state)
663{
664	struct memory_block *mem;
665	unsigned long start_pfn;
666	int scn_nr;
667	int ret = 0;
668
 
 
 
 
 
669	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
670	if (!mem)
671		return -ENOMEM;
672
673	scn_nr = __section_nr(section);
674	mem->start_section_nr =
675			base_memory_block_id(scn_nr) * sections_per_block;
676	mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
677	mem->state = state;
678	start_pfn = section_nr_to_pfn(mem->start_section_nr);
679	mem->phys_device = arch_get_memory_phys_device(start_pfn);
 
680
681	ret = register_memory(mem);
682
683	*memory = mem;
684	return ret;
685}
686
687static int add_memory_block(int base_section_nr)
688{
 
689	struct memory_block *mem;
690	int i, ret, section_count = 0, section_nr;
691
692	for (i = base_section_nr;
693	     (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
694	     i++) {
695		if (!present_section_nr(i))
696			continue;
697		if (section_count == 0)
698			section_nr = i;
699		section_count++;
700	}
701
702	if (section_count == 0)
703		return 0;
704	ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
 
705	if (ret)
706		return ret;
707	mem->section_count = section_count;
708	return 0;
709}
710
 
 
 
 
 
 
 
 
 
 
711/*
712 * need an interface for the VM to add new memory regions,
713 * but without onlining it.
 
714 */
715int hotplug_memory_register(int nid, struct mem_section *section)
716{
717	int ret = 0;
 
718	struct memory_block *mem;
 
 
719
720	mutex_lock(&mem_sysfs_mutex);
 
 
721
722	mem = find_memory_block(section);
723	if (mem) {
724		mem->section_count++;
725		put_device(&mem->dev);
726	} else {
727		ret = init_memory_block(&mem, section, MEM_OFFLINE);
728		if (ret)
729			goto out;
730		mem->section_count++;
 
 
 
 
 
 
 
 
 
731	}
732
733	if (mem->section_count == sections_per_block)
734		ret = register_mem_sect_under_node(mem, nid, false);
735out:
736	mutex_unlock(&mem_sysfs_mutex);
737	return ret;
738}
739
740#ifdef CONFIG_MEMORY_HOTREMOVE
741static void
742unregister_memory(struct memory_block *memory)
743{
744	BUG_ON(memory->dev.bus != &memory_subsys);
745
746	/* drop the ref. we got in remove_memory_block() */
747	put_device(&memory->dev);
748	device_unregister(&memory->dev);
749}
750
751static int remove_memory_section(unsigned long node_id,
752			       struct mem_section *section, int phys_device)
753{
 
 
754	struct memory_block *mem;
 
755
756	mutex_lock(&mem_sysfs_mutex);
757
758	/*
759	 * Some users of the memory hotplug do not want/need memblock to
760	 * track all sections. Skip over those.
761	 */
762	mem = find_memory_block(section);
763	if (!mem)
764		goto out_unlock;
765
766	unregister_mem_sect_under_nodes(mem, __section_nr(section));
767
768	mem->section_count--;
769	if (mem->section_count == 0)
 
 
 
 
 
770		unregister_memory(mem);
771	else
772		put_device(&mem->dev);
773
774out_unlock:
775	mutex_unlock(&mem_sysfs_mutex);
776	return 0;
777}
778
779int unregister_memory_section(struct mem_section *section)
780{
781	if (!present_section(section))
782		return -EINVAL;
783
784	return remove_memory_section(0, section, 0);
785}
786#endif /* CONFIG_MEMORY_HOTREMOVE */
787
788/* return true if the memory block is offlined, otherwise, return false */
789bool is_memblock_offlined(struct memory_block *mem)
790{
791	return mem->state == MEM_OFFLINE;
792}
793
794static struct attribute *memory_root_attrs[] = {
795#ifdef CONFIG_ARCH_MEMORY_PROBE
796	&dev_attr_probe.attr,
797#endif
798
799#ifdef CONFIG_MEMORY_FAILURE
800	&dev_attr_soft_offline_page.attr,
801	&dev_attr_hard_offline_page.attr,
802#endif
803
804	&dev_attr_block_size_bytes.attr,
805	&dev_attr_auto_online_blocks.attr,
806	NULL
807};
808
809static struct attribute_group memory_root_attr_group = {
810	.attrs = memory_root_attrs,
811};
812
813static const struct attribute_group *memory_root_attr_groups[] = {
814	&memory_root_attr_group,
815	NULL,
816};
817
818/*
819 * Initialize the sysfs support for memory devices...
820 */
821int __init memory_dev_init(void)
822{
823	unsigned int i;
824	int ret;
825	int err;
826	unsigned long block_sz;
 
 
 
 
 
 
827
828	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
829	if (ret)
830		goto out;
831
832	block_sz = get_memory_block_size();
833	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
834
835	/*
836	 * Create entries for memory sections that were found
837	 * during boot and have been initialized
838	 */
839	mutex_lock(&mem_sysfs_mutex);
840	for (i = 0; i <= __highest_present_section_nr;
841		i += sections_per_block) {
842		err = add_memory_block(i);
843		if (!ret)
844			ret = err;
845	}
846	mutex_unlock(&mem_sysfs_mutex);
847
848out:
849	if (ret)
850		printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
851	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
852}
v5.4
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Memory subsystem support
  4 *
  5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
  6 *            Dave Hansen <haveblue@us.ibm.com>
  7 *
  8 * This file provides the necessary infrastructure to represent
  9 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
 10 * All arch-independent code that assumes MEMORY_HOTPLUG requires
 11 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
 12 */
 13
 14#include <linux/module.h>
 15#include <linux/init.h>
 16#include <linux/topology.h>
 17#include <linux/capability.h>
 18#include <linux/device.h>
 19#include <linux/memory.h>
 20#include <linux/memory_hotplug.h>
 21#include <linux/mm.h>
 22#include <linux/mutex.h>
 23#include <linux/stat.h>
 24#include <linux/slab.h>
 25
 26#include <linux/atomic.h>
 27#include <linux/uaccess.h>
 28
 29static DEFINE_MUTEX(mem_sysfs_mutex);
 30
 31#define MEMORY_CLASS_NAME	"memory"
 32
 33#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
 34
 35static int sections_per_block;
 36
 37static inline unsigned long base_memory_block_id(unsigned long section_nr)
 38{
 39	return section_nr / sections_per_block;
 40}
 41
 42static inline unsigned long pfn_to_block_id(unsigned long pfn)
 43{
 44	return base_memory_block_id(pfn_to_section_nr(pfn));
 45}
 46
 47static inline unsigned long phys_to_block_id(unsigned long phys)
 48{
 49	return pfn_to_block_id(PFN_DOWN(phys));
 50}
 51
 52static int memory_subsys_online(struct device *dev);
 53static int memory_subsys_offline(struct device *dev);
 54
 55static struct bus_type memory_subsys = {
 56	.name = MEMORY_CLASS_NAME,
 57	.dev_name = MEMORY_CLASS_NAME,
 58	.online = memory_subsys_online,
 59	.offline = memory_subsys_offline,
 60};
 61
 62static BLOCKING_NOTIFIER_HEAD(memory_chain);
 63
 64int register_memory_notifier(struct notifier_block *nb)
 65{
 66	return blocking_notifier_chain_register(&memory_chain, nb);
 67}
 68EXPORT_SYMBOL(register_memory_notifier);
 69
 70void unregister_memory_notifier(struct notifier_block *nb)
 71{
 72	blocking_notifier_chain_unregister(&memory_chain, nb);
 73}
 74EXPORT_SYMBOL(unregister_memory_notifier);
 75
 76static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
 77
 78int register_memory_isolate_notifier(struct notifier_block *nb)
 79{
 80	return atomic_notifier_chain_register(&memory_isolate_chain, nb);
 81}
 82EXPORT_SYMBOL(register_memory_isolate_notifier);
 83
 84void unregister_memory_isolate_notifier(struct notifier_block *nb)
 85{
 86	atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
 87}
 88EXPORT_SYMBOL(unregister_memory_isolate_notifier);
 89
 90static void memory_block_release(struct device *dev)
 91{
 92	struct memory_block *mem = to_memory_block(dev);
 93
 94	kfree(mem);
 95}
 96
 97unsigned long __weak memory_block_size_bytes(void)
 98{
 99	return MIN_MEMORY_BLOCK_SIZE;
100}
101EXPORT_SYMBOL_GPL(memory_block_size_bytes);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
102
103/*
104 * Show the first physical section index (number) of this memory block.
 
105 */
106static ssize_t phys_index_show(struct device *dev,
107			       struct device_attribute *attr, char *buf)
 
108{
109	struct memory_block *mem = to_memory_block(dev);
110	unsigned long phys_index;
111
112	phys_index = mem->start_section_nr / sections_per_block;
113	return sprintf(buf, "%08lx\n", phys_index);
114}
115
116/*
117 * Show whether the memory block is likely to be offlineable (or is already
118 * offline). Once offline, the memory block could be removed. The return
119 * value does, however, not indicate that there is a way to remove the
120 * memory block.
121 */
122static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
123			      char *buf)
124{
 
 
125	struct memory_block *mem = to_memory_block(dev);
126	unsigned long pfn;
127	int ret = 1, i;
128
129	if (mem->state != MEM_ONLINE)
130		goto out;
131
132	for (i = 0; i < sections_per_block; i++) {
133		if (!present_section_nr(mem->start_section_nr + i))
134			continue;
135		pfn = section_nr_to_pfn(mem->start_section_nr + i);
136		ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
137	}
138
139out:
140	return sprintf(buf, "%d\n", ret);
141}
142
143/*
144 * online, offline, going offline, etc.
145 */
146static ssize_t state_show(struct device *dev, struct device_attribute *attr,
147			  char *buf)
148{
149	struct memory_block *mem = to_memory_block(dev);
150	ssize_t len = 0;
151
152	/*
153	 * We can probably put these states in a nice little array
154	 * so that they're not open-coded
155	 */
156	switch (mem->state) {
157	case MEM_ONLINE:
158		len = sprintf(buf, "online\n");
159		break;
160	case MEM_OFFLINE:
161		len = sprintf(buf, "offline\n");
162		break;
163	case MEM_GOING_OFFLINE:
164		len = sprintf(buf, "going-offline\n");
165		break;
166	default:
167		len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
168				mem->state);
169		WARN_ON(1);
170		break;
171	}
172
173	return len;
174}
175
176int memory_notify(unsigned long val, void *v)
177{
178	return blocking_notifier_call_chain(&memory_chain, val, v);
179}
180
181int memory_isolate_notify(unsigned long val, void *v)
182{
183	return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
184}
185
186/*
187 * The probe routines leave the pages uninitialized, just as the bootmem code
188 * does. Make sure we do not access them, but instead use only information from
189 * within sections.
190 */
191static bool pages_correctly_probed(unsigned long start_pfn)
192{
193	unsigned long section_nr = pfn_to_section_nr(start_pfn);
194	unsigned long section_nr_end = section_nr + sections_per_block;
195	unsigned long pfn = start_pfn;
196
197	/*
198	 * memmap between sections is not contiguous except with
199	 * SPARSEMEM_VMEMMAP. We lookup the page once per section
200	 * and assume memmap is contiguous within each section
201	 */
202	for (; section_nr < section_nr_end; section_nr++) {
203		if (WARN_ON_ONCE(!pfn_valid(pfn)))
204			return false;
205
206		if (!present_section_nr(section_nr)) {
207			pr_warn("section %ld pfn[%lx, %lx) not present\n",
208				section_nr, pfn, pfn + PAGES_PER_SECTION);
209			return false;
210		} else if (!valid_section_nr(section_nr)) {
211			pr_warn("section %ld pfn[%lx, %lx) no valid memmap\n",
212				section_nr, pfn, pfn + PAGES_PER_SECTION);
213			return false;
214		} else if (online_section_nr(section_nr)) {
215			pr_warn("section %ld pfn[%lx, %lx) is already online\n",
216				section_nr, pfn, pfn + PAGES_PER_SECTION);
217			return false;
218		}
219		pfn += PAGES_PER_SECTION;
220	}
221
222	return true;
223}
224
225/*
226 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
227 * OK to have direct references to sparsemem variables in here.
 
228 */
229static int
230memory_block_action(unsigned long start_section_nr, unsigned long action,
231		    int online_type)
232{
233	unsigned long start_pfn;
234	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
235	int ret;
236
237	start_pfn = section_nr_to_pfn(start_section_nr);
238
239	switch (action) {
240	case MEM_ONLINE:
241		if (!pages_correctly_probed(start_pfn))
242			return -EBUSY;
243
244		ret = online_pages(start_pfn, nr_pages, online_type);
245		break;
246	case MEM_OFFLINE:
247		ret = offline_pages(start_pfn, nr_pages);
248		break;
249	default:
250		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
251		     "%ld\n", __func__, start_section_nr, action, action);
252		ret = -EINVAL;
253	}
254
255	return ret;
256}
257
258static int memory_block_change_state(struct memory_block *mem,
259		unsigned long to_state, unsigned long from_state_req)
260{
261	int ret = 0;
262
263	if (mem->state != from_state_req)
264		return -EINVAL;
265
266	if (to_state == MEM_OFFLINE)
267		mem->state = MEM_GOING_OFFLINE;
268
269	ret = memory_block_action(mem->start_section_nr, to_state,
270				mem->online_type);
271
272	mem->state = ret ? from_state_req : to_state;
273
274	return ret;
275}
276
277/* The device lock serializes operations on memory_subsys_[online|offline] */
278static int memory_subsys_online(struct device *dev)
279{
280	struct memory_block *mem = to_memory_block(dev);
281	int ret;
282
283	if (mem->state == MEM_ONLINE)
284		return 0;
285
286	/*
287	 * If we are called from state_store(), online_type will be
288	 * set >= 0 Otherwise we were called from the device online
289	 * attribute and need to set the online_type.
290	 */
291	if (mem->online_type < 0)
292		mem->online_type = MMOP_ONLINE_KEEP;
293
 
294	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
295
296	/* clear online_type */
297	mem->online_type = -1;
298
299	return ret;
300}
301
302static int memory_subsys_offline(struct device *dev)
303{
304	struct memory_block *mem = to_memory_block(dev);
305
306	if (mem->state == MEM_OFFLINE)
307		return 0;
308
309	/* Can't offline block with non-present sections */
310	if (mem->section_count != sections_per_block)
311		return -EINVAL;
312
313	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
314}
315
316static ssize_t state_store(struct device *dev, struct device_attribute *attr,
317			   const char *buf, size_t count)
 
318{
319	struct memory_block *mem = to_memory_block(dev);
320	int ret, online_type;
321
322	ret = lock_device_hotplug_sysfs();
323	if (ret)
324		return ret;
325
326	if (sysfs_streq(buf, "online_kernel"))
327		online_type = MMOP_ONLINE_KERNEL;
328	else if (sysfs_streq(buf, "online_movable"))
329		online_type = MMOP_ONLINE_MOVABLE;
330	else if (sysfs_streq(buf, "online"))
331		online_type = MMOP_ONLINE_KEEP;
332	else if (sysfs_streq(buf, "offline"))
333		online_type = MMOP_OFFLINE;
334	else {
335		ret = -EINVAL;
336		goto err;
337	}
338
 
 
 
 
 
 
 
 
 
339	switch (online_type) {
340	case MMOP_ONLINE_KERNEL:
341	case MMOP_ONLINE_MOVABLE:
342	case MMOP_ONLINE_KEEP:
343		/* mem->online_type is protected by device_hotplug_lock */
344		mem->online_type = online_type;
345		ret = device_online(&mem->dev);
346		break;
347	case MMOP_OFFLINE:
348		ret = device_offline(&mem->dev);
349		break;
350	default:
351		ret = -EINVAL; /* should never happen */
352	}
353
 
354err:
355	unlock_device_hotplug();
356
357	if (ret < 0)
358		return ret;
359	if (ret)
360		return -EINVAL;
361
362	return count;
363}
364
365/*
366 * phys_device is a bad name for this.  What I really want
367 * is a way to differentiate between memory ranges that
368 * are part of physical devices that constitute
369 * a complete removable unit or fru.
370 * i.e. do these ranges belong to the same physical device,
371 * s.t. if I offline all of these sections I can then
372 * remove the physical device?
373 */
374static ssize_t phys_device_show(struct device *dev,
375				struct device_attribute *attr, char *buf)
376{
377	struct memory_block *mem = to_memory_block(dev);
378	return sprintf(buf, "%d\n", mem->phys_device);
379}
380
381#ifdef CONFIG_MEMORY_HOTREMOVE
382static void print_allowed_zone(char *buf, int nid, unsigned long start_pfn,
383		unsigned long nr_pages, int online_type,
384		struct zone *default_zone)
385{
386	struct zone *zone;
387
388	zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
389	if (zone != default_zone) {
390		strcat(buf, " ");
391		strcat(buf, zone->name);
392	}
393}
394
395static ssize_t valid_zones_show(struct device *dev,
396				struct device_attribute *attr, char *buf)
397{
398	struct memory_block *mem = to_memory_block(dev);
399	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
400	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
401	unsigned long valid_start_pfn, valid_end_pfn;
402	struct zone *default_zone;
403	int nid;
404
405	/*
 
 
 
 
 
 
 
 
 
 
406	 * Check the existing zone. Make sure that we do that only on the
407	 * online nodes otherwise the page_zone is not reliable
408	 */
409	if (mem->state == MEM_ONLINE) {
410		/*
411		 * The block contains more than one zone can not be offlined.
412		 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
413		 */
414		if (!test_pages_in_a_zone(start_pfn, start_pfn + nr_pages,
415					  &valid_start_pfn, &valid_end_pfn))
416			return sprintf(buf, "none\n");
417		start_pfn = valid_start_pfn;
418		strcat(buf, page_zone(pfn_to_page(start_pfn))->name);
419		goto out;
420	}
421
422	nid = mem->nid;
423	default_zone = zone_for_pfn_range(MMOP_ONLINE_KEEP, nid, start_pfn, nr_pages);
424	strcat(buf, default_zone->name);
425
426	print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_KERNEL,
427			default_zone);
428	print_allowed_zone(buf, nid, start_pfn, nr_pages, MMOP_ONLINE_MOVABLE,
429			default_zone);
430out:
431	strcat(buf, "\n");
432
433	return strlen(buf);
434}
435static DEVICE_ATTR_RO(valid_zones);
436#endif
437
438static DEVICE_ATTR_RO(phys_index);
439static DEVICE_ATTR_RW(state);
440static DEVICE_ATTR_RO(phys_device);
441static DEVICE_ATTR_RO(removable);
442
443/*
444 * Show the memory block size (shared by all memory blocks).
445 */
446static ssize_t block_size_bytes_show(struct device *dev,
447				     struct device_attribute *attr, char *buf)
 
448{
449	return sprintf(buf, "%lx\n", memory_block_size_bytes());
450}
451
452static DEVICE_ATTR_RO(block_size_bytes);
453
454/*
455 * Memory auto online policy.
456 */
457
458static ssize_t auto_online_blocks_show(struct device *dev,
459				       struct device_attribute *attr, char *buf)
 
460{
461	if (memhp_auto_online)
462		return sprintf(buf, "online\n");
463	else
464		return sprintf(buf, "offline\n");
465}
466
467static ssize_t auto_online_blocks_store(struct device *dev,
468					struct device_attribute *attr,
469					const char *buf, size_t count)
470{
471	if (sysfs_streq(buf, "online"))
472		memhp_auto_online = true;
473	else if (sysfs_streq(buf, "offline"))
474		memhp_auto_online = false;
475	else
476		return -EINVAL;
477
478	return count;
479}
480
481static DEVICE_ATTR_RW(auto_online_blocks);
 
482
483/*
484 * Some architectures will have custom drivers to do this, and
485 * will not need to do it from userspace.  The fake hot-add code
486 * as well as ppc64 will do all of their discovery in userspace
487 * and will require this interface.
488 */
489#ifdef CONFIG_ARCH_MEMORY_PROBE
490static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
491			   const char *buf, size_t count)
 
492{
493	u64 phys_addr;
494	int nid, ret;
495	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
496
497	ret = kstrtoull(buf, 0, &phys_addr);
498	if (ret)
499		return ret;
500
501	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
502		return -EINVAL;
503
504	ret = lock_device_hotplug_sysfs();
505	if (ret)
506		return ret;
507
508	nid = memory_add_physaddr_to_nid(phys_addr);
509	ret = __add_memory(nid, phys_addr,
510			   MIN_MEMORY_BLOCK_SIZE * sections_per_block);
511
512	if (ret)
513		goto out;
514
515	ret = count;
516out:
517	unlock_device_hotplug();
518	return ret;
519}
520
521static DEVICE_ATTR_WO(probe);
522#endif
523
524#ifdef CONFIG_MEMORY_FAILURE
525/*
526 * Support for offlining pages of memory
527 */
528
529/* Soft offline a page */
530static ssize_t soft_offline_page_store(struct device *dev,
531				       struct device_attribute *attr,
532				       const char *buf, size_t count)
 
533{
534	int ret;
535	u64 pfn;
536	if (!capable(CAP_SYS_ADMIN))
537		return -EPERM;
538	if (kstrtoull(buf, 0, &pfn) < 0)
539		return -EINVAL;
540	pfn >>= PAGE_SHIFT;
541	if (!pfn_valid(pfn))
542		return -ENXIO;
543	/* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
544	if (!pfn_to_online_page(pfn))
545		return -EIO;
546	ret = soft_offline_page(pfn_to_page(pfn), 0);
547	return ret == 0 ? count : ret;
548}
549
550/* Forcibly offline a page, including killing processes. */
551static ssize_t hard_offline_page_store(struct device *dev,
552				       struct device_attribute *attr,
553				       const char *buf, size_t count)
 
554{
555	int ret;
556	u64 pfn;
557	if (!capable(CAP_SYS_ADMIN))
558		return -EPERM;
559	if (kstrtoull(buf, 0, &pfn) < 0)
560		return -EINVAL;
561	pfn >>= PAGE_SHIFT;
562	ret = memory_failure(pfn, 0);
563	return ret ? ret : count;
564}
565
566static DEVICE_ATTR_WO(soft_offline_page);
567static DEVICE_ATTR_WO(hard_offline_page);
568#endif
569
570/*
571 * Note that phys_device is optional.  It is here to allow for
572 * differentiation between which *physical* devices each
573 * section belongs to...
574 */
575int __weak arch_get_memory_phys_device(unsigned long start_pfn)
576{
577	return 0;
578}
579
580/* A reference for the returned memory block device is acquired. */
581static struct memory_block *find_memory_block_by_id(unsigned long block_id)
 
 
 
 
582{
 
 
583	struct device *dev;
584
585	dev = subsys_find_device_by_id(&memory_subsys, block_id, NULL);
586	return dev ? to_memory_block(dev) : NULL;
 
 
 
 
587}
588
589/*
590 * For now, we have a linear search to go find the appropriate
591 * memory_block corresponding to a particular phys_index. If
592 * this gets to be a real problem, we can always use a radix
593 * tree or something here.
594 *
595 * This could be made generic for all device subsystems.
596 */
597struct memory_block *find_memory_block(struct mem_section *section)
598{
599	unsigned long block_id = base_memory_block_id(__section_nr(section));
600
601	return find_memory_block_by_id(block_id);
602}
603
604static struct attribute *memory_memblk_attrs[] = {
605	&dev_attr_phys_index.attr,
606	&dev_attr_state.attr,
607	&dev_attr_phys_device.attr,
608	&dev_attr_removable.attr,
609#ifdef CONFIG_MEMORY_HOTREMOVE
610	&dev_attr_valid_zones.attr,
611#endif
612	NULL
613};
614
615static struct attribute_group memory_memblk_attr_group = {
616	.attrs = memory_memblk_attrs,
617};
618
619static const struct attribute_group *memory_memblk_attr_groups[] = {
620	&memory_memblk_attr_group,
621	NULL,
622};
623
624/*
625 * register_memory - Setup a sysfs device for a memory block
626 */
627static
628int register_memory(struct memory_block *memory)
629{
630	int ret;
631
632	memory->dev.bus = &memory_subsys;
633	memory->dev.id = memory->start_section_nr / sections_per_block;
634	memory->dev.release = memory_block_release;
635	memory->dev.groups = memory_memblk_attr_groups;
636	memory->dev.offline = memory->state == MEM_OFFLINE;
637
638	ret = device_register(&memory->dev);
639	if (ret)
640		put_device(&memory->dev);
641
642	return ret;
643}
644
645static int init_memory_block(struct memory_block **memory,
646			     unsigned long block_id, unsigned long state)
647{
648	struct memory_block *mem;
649	unsigned long start_pfn;
 
650	int ret = 0;
651
652	mem = find_memory_block_by_id(block_id);
653	if (mem) {
654		put_device(&mem->dev);
655		return -EEXIST;
656	}
657	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
658	if (!mem)
659		return -ENOMEM;
660
661	mem->start_section_nr = block_id * sections_per_block;
 
 
 
662	mem->state = state;
663	start_pfn = section_nr_to_pfn(mem->start_section_nr);
664	mem->phys_device = arch_get_memory_phys_device(start_pfn);
665	mem->nid = NUMA_NO_NODE;
666
667	ret = register_memory(mem);
668
669	*memory = mem;
670	return ret;
671}
672
673static int add_memory_block(unsigned long base_section_nr)
674{
675	int ret, section_count = 0;
676	struct memory_block *mem;
677	unsigned long nr;
678
679	for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
680	     nr++)
681		if (present_section_nr(nr))
682			section_count++;
 
 
 
 
 
683
684	if (section_count == 0)
685		return 0;
686	ret = init_memory_block(&mem, base_memory_block_id(base_section_nr),
687				MEM_ONLINE);
688	if (ret)
689		return ret;
690	mem->section_count = section_count;
691	return 0;
692}
693
694static void unregister_memory(struct memory_block *memory)
695{
696	if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
697		return;
698
699	/* drop the ref. we got via find_memory_block() */
700	put_device(&memory->dev);
701	device_unregister(&memory->dev);
702}
703
704/*
705 * Create memory block devices for the given memory area. Start and size
706 * have to be aligned to memory block granularity. Memory block devices
707 * will be initialized as offline.
708 */
709int create_memory_block_devices(unsigned long start, unsigned long size)
710{
711	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
712	unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
713	struct memory_block *mem;
714	unsigned long block_id;
715	int ret = 0;
716
717	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
718			 !IS_ALIGNED(size, memory_block_size_bytes())))
719		return -EINVAL;
720
721	mutex_lock(&mem_sysfs_mutex);
722	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
723		ret = init_memory_block(&mem, block_id, MEM_OFFLINE);
 
 
 
724		if (ret)
725			break;
726		mem->section_count = sections_per_block;
727	}
728	if (ret) {
729		end_block_id = block_id;
730		for (block_id = start_block_id; block_id != end_block_id;
731		     block_id++) {
732			mem = find_memory_block_by_id(block_id);
733			mem->section_count = 0;
734			unregister_memory(mem);
735		}
736	}
 
 
 
 
737	mutex_unlock(&mem_sysfs_mutex);
738	return ret;
739}
740
741/*
742 * Remove memory block devices for the given memory area. Start and size
743 * have to be aligned to memory block granularity. Memory block devices
744 * have to be offline.
745 */
746void remove_memory_block_devices(unsigned long start, unsigned long size)
 
 
 
 
 
 
 
747{
748	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
749	const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
750	struct memory_block *mem;
751	unsigned long block_id;
752
753	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
754			 !IS_ALIGNED(size, memory_block_size_bytes())))
755		return;
 
 
 
 
 
 
 
 
756
757	mutex_lock(&mem_sysfs_mutex);
758	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
759		mem = find_memory_block_by_id(block_id);
760		if (WARN_ON_ONCE(!mem))
761			continue;
762		mem->section_count = 0;
763		unregister_memory_block_under_nodes(mem);
764		unregister_memory(mem);
765	}
 
 
 
766	mutex_unlock(&mem_sysfs_mutex);
 
 
 
 
 
 
 
 
 
767}
 
768
769/* return true if the memory block is offlined, otherwise, return false */
770bool is_memblock_offlined(struct memory_block *mem)
771{
772	return mem->state == MEM_OFFLINE;
773}
774
775static struct attribute *memory_root_attrs[] = {
776#ifdef CONFIG_ARCH_MEMORY_PROBE
777	&dev_attr_probe.attr,
778#endif
779
780#ifdef CONFIG_MEMORY_FAILURE
781	&dev_attr_soft_offline_page.attr,
782	&dev_attr_hard_offline_page.attr,
783#endif
784
785	&dev_attr_block_size_bytes.attr,
786	&dev_attr_auto_online_blocks.attr,
787	NULL
788};
789
790static struct attribute_group memory_root_attr_group = {
791	.attrs = memory_root_attrs,
792};
793
794static const struct attribute_group *memory_root_attr_groups[] = {
795	&memory_root_attr_group,
796	NULL,
797};
798
799/*
800 * Initialize the sysfs support for memory devices...
801 */
802void __init memory_dev_init(void)
803{
 
804	int ret;
805	int err;
806	unsigned long block_sz, nr;
807
808	/* Validate the configured memory block size */
809	block_sz = memory_block_size_bytes();
810	if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
811		panic("Memory block size not suitable: 0x%lx\n", block_sz);
812	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
813
814	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
815	if (ret)
816		goto out;
817
 
 
 
818	/*
819	 * Create entries for memory sections that were found
820	 * during boot and have been initialized
821	 */
822	mutex_lock(&mem_sysfs_mutex);
823	for (nr = 0; nr <= __highest_present_section_nr;
824	     nr += sections_per_block) {
825		err = add_memory_block(nr);
826		if (!ret)
827			ret = err;
828	}
829	mutex_unlock(&mem_sysfs_mutex);
830
831out:
832	if (ret)
833		panic("%s() failed: %d\n", __func__, ret);
834}
835
836/**
837 * walk_memory_blocks - walk through all present memory blocks overlapped
838 *			by the range [start, start + size)
839 *
840 * @start: start address of the memory range
841 * @size: size of the memory range
842 * @arg: argument passed to func
843 * @func: callback for each memory section walked
844 *
845 * This function walks through all present memory blocks overlapped by the
846 * range [start, start + size), calling func on each memory block.
847 *
848 * In case func() returns an error, walking is aborted and the error is
849 * returned.
850 */
851int walk_memory_blocks(unsigned long start, unsigned long size,
852		       void *arg, walk_memory_blocks_func_t func)
853{
854	const unsigned long start_block_id = phys_to_block_id(start);
855	const unsigned long end_block_id = phys_to_block_id(start + size - 1);
856	struct memory_block *mem;
857	unsigned long block_id;
858	int ret = 0;
859
860	if (!size)
861		return 0;
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 (!mem)
866			continue;
867
868		ret = func(mem, arg);
869		put_device(&mem->dev);
870		if (ret)
871			break;
872	}
873	return ret;
874}
875
876struct for_each_memory_block_cb_data {
877	walk_memory_blocks_func_t func;
878	void *arg;
879};
880
881static int for_each_memory_block_cb(struct device *dev, void *data)
882{
883	struct memory_block *mem = to_memory_block(dev);
884	struct for_each_memory_block_cb_data *cb_data = data;
885
886	return cb_data->func(mem, cb_data->arg);
887}
888
889/**
890 * for_each_memory_block - walk through all present memory blocks
891 *
892 * @arg: argument passed to func
893 * @func: callback for each memory block walked
894 *
895 * This function walks through all present memory blocks, calling func on
896 * each memory block.
897 *
898 * In case func() returns an error, walking is aborted and the error is
899 * returned.
900 */
901int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
902{
903	struct for_each_memory_block_cb_data cb_data = {
904		.func = func,
905		.arg = arg,
906	};
907
908	return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
909				for_each_memory_block_cb);
910}