Linux Audio

Check our new training course

Loading...
v3.15
 
  1/*
  2 * Memory subsystem support
  3 *
  4 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
  5 *            Dave Hansen <haveblue@us.ibm.com>
  6 *
  7 * This file provides the necessary infrastructure to represent
  8 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
  9 * All arch-independent code that assumes MEMORY_HOTPLUG requires
 10 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
 11 */
 12
 13#include <linux/module.h>
 14#include <linux/init.h>
 15#include <linux/topology.h>
 16#include <linux/capability.h>
 17#include <linux/device.h>
 18#include <linux/memory.h>
 19#include <linux/memory_hotplug.h>
 20#include <linux/mm.h>
 21#include <linux/mutex.h>
 22#include <linux/stat.h>
 23#include <linux/slab.h>
 
 24
 25#include <linux/atomic.h>
 26#include <asm/uaccess.h>
 27
 28static DEFINE_MUTEX(mem_sysfs_mutex);
 29
 30#define MEMORY_CLASS_NAME	"memory"
 31
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 32#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
 33
 34static int sections_per_block;
 35
 36static inline int base_memory_block_id(int section_nr)
 37{
 38	return section_nr / sections_per_block;
 39}
 40
 
 
 
 
 
 
 
 
 
 
 41static int memory_subsys_online(struct device *dev);
 42static int memory_subsys_offline(struct device *dev);
 43
 44static struct bus_type memory_subsys = {
 45	.name = MEMORY_CLASS_NAME,
 46	.dev_name = MEMORY_CLASS_NAME,
 47	.online = memory_subsys_online,
 48	.offline = memory_subsys_offline,
 49};
 50
 
 
 
 
 
 
 
 
 
 
 
 
 
 51static BLOCKING_NOTIFIER_HEAD(memory_chain);
 52
 53int register_memory_notifier(struct notifier_block *nb)
 54{
 55        return blocking_notifier_chain_register(&memory_chain, nb);
 56}
 57EXPORT_SYMBOL(register_memory_notifier);
 58
 59void unregister_memory_notifier(struct notifier_block *nb)
 60{
 61        blocking_notifier_chain_unregister(&memory_chain, nb);
 62}
 63EXPORT_SYMBOL(unregister_memory_notifier);
 64
 65static ATOMIC_NOTIFIER_HEAD(memory_isolate_chain);
 66
 67int register_memory_isolate_notifier(struct notifier_block *nb)
 68{
 69	return atomic_notifier_chain_register(&memory_isolate_chain, nb);
 70}
 71EXPORT_SYMBOL(register_memory_isolate_notifier);
 72
 73void unregister_memory_isolate_notifier(struct notifier_block *nb)
 74{
 75	atomic_notifier_chain_unregister(&memory_isolate_chain, nb);
 76}
 77EXPORT_SYMBOL(unregister_memory_isolate_notifier);
 78
 79static void memory_block_release(struct device *dev)
 80{
 81	struct memory_block *mem = to_memory_block(dev);
 82
 83	kfree(mem);
 84}
 85
 86unsigned long __weak memory_block_size_bytes(void)
 87{
 88	return MIN_MEMORY_BLOCK_SIZE;
 89}
 90
 91static unsigned long get_memory_block_size(void)
 92{
 93	unsigned long block_sz;
 94
 95	block_sz = memory_block_size_bytes();
 96
 97	/* Validate blk_sz is a power of 2 and not less than section size */
 98	if ((block_sz & (block_sz - 1)) || (block_sz < MIN_MEMORY_BLOCK_SIZE)) {
 99		WARN_ON(1);
100		block_sz = MIN_MEMORY_BLOCK_SIZE;
101	}
102
103	return block_sz;
104}
105
106/*
107 * use this as the physical section index that this memsection
108 * uses.
109 */
110
111static ssize_t show_mem_start_phys_index(struct device *dev,
112			struct device_attribute *attr, char *buf)
113{
114	struct memory_block *mem = to_memory_block(dev);
115	unsigned long phys_index;
116
117	phys_index = mem->start_section_nr / sections_per_block;
118	return sprintf(buf, "%08lx\n", phys_index);
119}
120
121static ssize_t show_mem_end_phys_index(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->end_section_nr / sections_per_block;
128	return sprintf(buf, "%08lx\n", phys_index);
129}
130
131/*
132 * Show whether the section of memory is likely to be hot-removable
 
133 */
134static ssize_t show_mem_removable(struct device *dev,
135			struct device_attribute *attr, char *buf)
136{
137	unsigned long i, pfn;
138	int ret = 1;
139	struct memory_block *mem = to_memory_block(dev);
140
141	for (i = 0; i < sections_per_block; i++) {
142		if (!present_section_nr(mem->start_section_nr + i))
143			continue;
144		pfn = section_nr_to_pfn(mem->start_section_nr + i);
145		ret &= is_mem_section_removable(pfn, PAGES_PER_SECTION);
146	}
147
148	return sprintf(buf, "%d\n", ret);
149}
150
151/*
152 * online, offline, going offline, etc.
153 */
154static ssize_t show_mem_state(struct device *dev,
155			struct device_attribute *attr, char *buf)
156{
157	struct memory_block *mem = to_memory_block(dev);
158	ssize_t len = 0;
159
160	/*
161	 * We can probably put these states in a nice little array
162	 * so that they're not open-coded
163	 */
164	switch (mem->state) {
165		case MEM_ONLINE:
166			len = sprintf(buf, "online\n");
167			break;
168		case MEM_OFFLINE:
169			len = sprintf(buf, "offline\n");
170			break;
171		case MEM_GOING_OFFLINE:
172			len = sprintf(buf, "going-offline\n");
173			break;
174		default:
175			len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
176					mem->state);
177			WARN_ON(1);
178			break;
179	}
180
181	return len;
182}
183
184int memory_notify(unsigned long val, void *v)
185{
186	return blocking_notifier_call_chain(&memory_chain, val, v);
187}
188
189int memory_isolate_notify(unsigned long val, void *v)
 
 
 
190{
191	return atomic_notifier_call_chain(&memory_isolate_chain, val, v);
192}
 
193
194/*
195 * The probe routines leave the pages reserved, just as the bootmem code does.
196 * Make sure they're still that way.
197 */
198static bool pages_correctly_reserved(unsigned long start_pfn)
199{
200	int i, j;
201	struct page *page;
202	unsigned long pfn = start_pfn;
 
 
 
 
 
 
 
 
203
204	/*
205	 * memmap between sections is not contiguous except with
206	 * SPARSEMEM_VMEMMAP. We lookup the page once per section
207	 * and assume memmap is contiguous within each section
 
 
208	 */
209	for (i = 0; i < sections_per_block; i++, pfn += PAGES_PER_SECTION) {
210		if (WARN_ON_ONCE(!pfn_valid(pfn)))
211			return false;
212		page = pfn_to_page(pfn);
 
213
214		for (j = 0; j < PAGES_PER_SECTION; j++) {
215			if (PageReserved(page + j))
216				continue;
 
 
 
 
217
218			printk(KERN_WARNING "section number %ld page number %d "
219				"not reserved, was it already online?\n",
220				pfn_to_section_nr(pfn), j);
 
 
 
 
221
222			return false;
223		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
224	}
225
226	return true;
 
 
 
 
227}
228
229/*
230 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
231 * OK to have direct references to sparsemem variables in here.
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	struct page *first_page;
239	int ret;
240
241	first_page = pfn_to_page(phys_index << PFN_SECTION_SHIFT);
242	start_pfn = page_to_pfn(first_page);
243
244	switch (action) {
245		case MEM_ONLINE:
246			if (!pages_correctly_reserved(start_pfn))
247				return -EBUSY;
248
249			ret = online_pages(start_pfn, nr_pages, online_type);
250			break;
251		case MEM_OFFLINE:
252			ret = offline_pages(start_pfn, nr_pages);
253			break;
254		default:
255			WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
256			     "%ld\n", __func__, phys_index, action, action);
257			ret = -EINVAL;
258	}
259
260	return ret;
261}
262
263static int memory_block_change_state(struct memory_block *mem,
264		unsigned long to_state, unsigned long from_state_req)
265{
266	int ret = 0;
267
268	if (mem->state != from_state_req)
269		return -EINVAL;
270
271	if (to_state == MEM_OFFLINE)
272		mem->state = MEM_GOING_OFFLINE;
273
274	ret = memory_block_action(mem->start_section_nr, to_state,
275				mem->online_type);
276
277	mem->state = ret ? from_state_req : to_state;
278
279	return ret;
280}
281
282/* The device lock serializes operations on memory_subsys_[online|offline] */
283static int memory_subsys_online(struct device *dev)
284{
285	struct memory_block *mem = to_memory_block(dev);
286	int ret;
287
288	if (mem->state == MEM_ONLINE)
289		return 0;
290
291	/*
292	 * If we are called from store_mem_state(), online_type will be
293	 * set >= 0 Otherwise we were called from the device online
294	 * attribute and need to set the online_type.
295	 */
296	if (mem->online_type < 0)
297		mem->online_type = ONLINE_KEEP;
298
299	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
300
301	/* clear online_type */
302	mem->online_type = -1;
303
304	return ret;
305}
306
307static int memory_subsys_offline(struct device *dev)
308{
309	struct memory_block *mem = to_memory_block(dev);
310
311	if (mem->state == MEM_OFFLINE)
312		return 0;
313
314	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
315}
316
317static ssize_t
318store_mem_state(struct device *dev,
319		struct device_attribute *attr, const char *buf, size_t count)
320{
 
321	struct memory_block *mem = to_memory_block(dev);
322	int ret, online_type;
 
 
 
323
324	ret = lock_device_hotplug_sysfs();
325	if (ret)
326		return ret;
327
328	if (!strncmp(buf, "online_kernel", min_t(int, count, 13)))
329		online_type = ONLINE_KERNEL;
330	else if (!strncmp(buf, "online_movable", min_t(int, count, 14)))
331		online_type = ONLINE_MOVABLE;
332	else if (!strncmp(buf, "online", min_t(int, count, 6)))
333		online_type = ONLINE_KEEP;
334	else if (!strncmp(buf, "offline", min_t(int, count, 7)))
335		online_type = -1;
336	else {
337		ret = -EINVAL;
338		goto err;
339	}
340
341	switch (online_type) {
342	case ONLINE_KERNEL:
343	case ONLINE_MOVABLE:
344	case ONLINE_KEEP:
345		/*
346		 * mem->online_type is not protected so there can be a
347		 * race here.  However, when racing online, the first
348		 * will succeed and the second will just return as the
349		 * block will already be online.  The online type
350		 * could be either one, but that is expected.
351		 */
352		mem->online_type = online_type;
353		ret = device_online(&mem->dev);
354		break;
355	case -1:
356		ret = device_offline(&mem->dev);
357		break;
358	default:
359		ret = -EINVAL; /* should never happen */
360	}
361
362err:
363	unlock_device_hotplug();
364
365	if (ret)
366		return ret;
 
 
 
367	return count;
368}
369
370/*
371 * phys_device is a bad name for this.  What I really want
372 * is a way to differentiate between memory ranges that
373 * are part of physical devices that constitute
374 * a complete removable unit or fru.
375 * i.e. do these ranges belong to the same physical device,
376 * s.t. if I offline all of these sections I can then
377 * remove the physical device?
378 */
379static ssize_t show_phys_device(struct device *dev,
380				struct device_attribute *attr, char *buf)
381{
382	struct memory_block *mem = to_memory_block(dev);
383	return sprintf(buf, "%d\n", mem->phys_device);
 
 
 
384}
385
386static DEVICE_ATTR(phys_index, 0444, show_mem_start_phys_index, NULL);
387static DEVICE_ATTR(end_phys_index, 0444, show_mem_end_phys_index, NULL);
388static DEVICE_ATTR(state, 0644, show_mem_state, store_mem_state);
389static DEVICE_ATTR(phys_device, 0444, show_phys_device, NULL);
390static DEVICE_ATTR(removable, 0444, show_mem_removable, NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
391
392/*
393 * Block size attribute stuff
394 */
395static ssize_t
396print_block_size(struct device *dev, struct device_attribute *attr,
397		 char *buf)
398{
399	return sprintf(buf, "%lx\n", get_memory_block_size());
400}
401
402static DEVICE_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
403
404/*
405 * Some architectures will have custom drivers to do this, and
406 * will not need to do it from userspace.  The fake hot-add code
407 * as well as ppc64 will do all of their discovery in userspace
408 * and will require this interface.
409 */
410#ifdef CONFIG_ARCH_MEMORY_PROBE
411static ssize_t
412memory_probe_store(struct device *dev, struct device_attribute *attr,
413		   const char *buf, size_t count)
414{
415	u64 phys_addr;
416	int nid;
417	int i, ret;
418	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
419
420	phys_addr = simple_strtoull(buf, NULL, 0);
 
 
421
422	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
423		return -EINVAL;
424
425	for (i = 0; i < sections_per_block; i++) {
426		nid = memory_add_physaddr_to_nid(phys_addr);
427		ret = add_memory(nid, phys_addr,
428				 PAGES_PER_SECTION << PAGE_SHIFT);
429		if (ret)
430			goto out;
431
432		phys_addr += MIN_MEMORY_BLOCK_SIZE;
433	}
 
 
 
 
 
434
435	ret = count;
436out:
 
437	return ret;
438}
439
440static DEVICE_ATTR(probe, S_IWUSR, NULL, memory_probe_store);
441#endif
442
443#ifdef CONFIG_MEMORY_FAILURE
444/*
445 * Support for offlining pages of memory
446 */
447
448/* Soft offline a page */
449static ssize_t
450store_soft_offline_page(struct device *dev,
451			struct device_attribute *attr,
452			const char *buf, size_t count)
453{
454	int ret;
455	u64 pfn;
456	if (!capable(CAP_SYS_ADMIN))
457		return -EPERM;
458	if (kstrtoull(buf, 0, &pfn) < 0)
459		return -EINVAL;
460	pfn >>= PAGE_SHIFT;
461	if (!pfn_valid(pfn))
462		return -ENXIO;
463	ret = soft_offline_page(pfn_to_page(pfn), 0);
464	return ret == 0 ? count : ret;
465}
466
467/* Forcibly offline a page, including killing processes. */
468static ssize_t
469store_hard_offline_page(struct device *dev,
470			struct device_attribute *attr,
471			const char *buf, size_t count)
472{
473	int ret;
474	u64 pfn;
475	if (!capable(CAP_SYS_ADMIN))
476		return -EPERM;
477	if (kstrtoull(buf, 0, &pfn) < 0)
478		return -EINVAL;
479	pfn >>= PAGE_SHIFT;
480	ret = memory_failure(pfn, 0, 0);
 
 
481	return ret ? ret : count;
482}
483
484static DEVICE_ATTR(soft_offline_page, S_IWUSR, NULL, store_soft_offline_page);
485static DEVICE_ATTR(hard_offline_page, S_IWUSR, NULL, store_hard_offline_page);
486#endif
487
488/*
489 * Note that phys_device is optional.  It is here to allow for
490 * differentiation between which *physical* devices each
491 * section belongs to...
492 */
493int __weak arch_get_memory_phys_device(unsigned long start_pfn)
494{
495	return 0;
496}
497
498/*
499 * A reference for the returned object is held and the reference for the
500 * hinted object is released.
 
501 */
502struct memory_block *find_memory_block_hinted(struct mem_section *section,
503					      struct memory_block *hint)
504{
505	int block_id = base_memory_block_id(__section_nr(section));
506	struct device *hintdev = hint ? &hint->dev : NULL;
507	struct device *dev;
508
509	dev = subsys_find_device_by_id(&memory_subsys, block_id, hintdev);
510	if (hint)
511		put_device(&hint->dev);
512	if (!dev)
513		return NULL;
514	return to_memory_block(dev);
515}
516
517/*
518 * For now, we have a linear search to go find the appropriate
519 * memory_block corresponding to a particular phys_index. If
520 * this gets to be a real problem, we can always use a radix
521 * tree or something here.
522 *
523 * This could be made generic for all device subsystems.
524 */
525struct memory_block *find_memory_block(struct mem_section *section)
526{
527	return find_memory_block_hinted(section, NULL);
 
 
528}
529
530static struct attribute *memory_memblk_attrs[] = {
531	&dev_attr_phys_index.attr,
532	&dev_attr_end_phys_index.attr,
533	&dev_attr_state.attr,
534	&dev_attr_phys_device.attr,
535	&dev_attr_removable.attr,
 
 
 
536	NULL
537};
538
539static struct attribute_group memory_memblk_attr_group = {
540	.attrs = memory_memblk_attrs,
541};
542
543static const struct attribute_group *memory_memblk_attr_groups[] = {
544	&memory_memblk_attr_group,
545	NULL,
546};
547
548/*
549 * register_memory - Setup a sysfs device for a memory block
550 */
551static
552int register_memory(struct memory_block *memory)
553{
 
 
554	memory->dev.bus = &memory_subsys;
555	memory->dev.id = memory->start_section_nr / sections_per_block;
556	memory->dev.release = memory_block_release;
557	memory->dev.groups = memory_memblk_attr_groups;
558	memory->dev.offline = memory->state == MEM_OFFLINE;
559
560	return device_register(&memory->dev);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
561}
 
562
563static int init_memory_block(struct memory_block **memory,
564			     struct mem_section *section, unsigned long state)
 
565{
566	struct memory_block *mem;
567	unsigned long start_pfn;
568	int scn_nr;
569	int ret = 0;
570
 
 
 
 
 
571	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
572	if (!mem)
573		return -ENOMEM;
574
575	scn_nr = __section_nr(section);
576	mem->start_section_nr =
577			base_memory_block_id(scn_nr) * sections_per_block;
578	mem->end_section_nr = mem->start_section_nr + sections_per_block - 1;
579	mem->state = state;
580	mem->section_count++;
581	start_pfn = section_nr_to_pfn(mem->start_section_nr);
582	mem->phys_device = arch_get_memory_phys_device(start_pfn);
583
584	ret = register_memory(mem);
 
 
 
 
 
 
 
 
 
585
586	*memory = mem;
587	return ret;
 
 
 
 
 
 
 
 
588}
589
590static int add_memory_block(int base_section_nr)
591{
592	struct memory_block *mem;
593	int i, ret, section_count = 0, section_nr;
594
595	for (i = base_section_nr;
596	     (i < base_section_nr + sections_per_block) && i < NR_MEM_SECTIONS;
597	     i++) {
598		if (!present_section_nr(i))
599			continue;
600		if (section_count == 0)
601			section_nr = i;
602		section_count++;
603	}
604
605	if (section_count == 0)
606		return 0;
607	ret = init_memory_block(&mem, __nr_to_section(section_nr), MEM_ONLINE);
608	if (ret)
609		return ret;
610	mem->section_count = section_count;
611	return 0;
612}
613
 
 
 
 
 
 
614
615/*
616 * need an interface for the VM to add new memory regions,
617 * but without onlining it.
618 */
619int register_new_memory(int nid, struct mem_section *section)
620{
621	int ret = 0;
622	struct memory_block *mem;
623
624	mutex_lock(&mem_sysfs_mutex);
625
626	mem = find_memory_block(section);
627	if (mem) {
628		mem->section_count++;
629		put_device(&mem->dev);
630	} else {
631		ret = init_memory_block(&mem, section, MEM_OFFLINE);
632		if (ret)
633			goto out;
634	}
635
636	if (mem->section_count == sections_per_block)
637		ret = register_mem_sect_under_node(mem, nid);
638out:
639	mutex_unlock(&mem_sysfs_mutex);
640	return ret;
641}
642
643#ifdef CONFIG_MEMORY_HOTREMOVE
644static void
645unregister_memory(struct memory_block *memory)
646{
647	BUG_ON(memory->dev.bus != &memory_subsys);
648
649	/* drop the ref. we got in remove_memory_block() */
650	put_device(&memory->dev);
651	device_unregister(&memory->dev);
652}
653
654static int remove_memory_block(unsigned long node_id,
655			       struct mem_section *section, int phys_device)
 
 
 
 
 
 
 
 
656{
 
 
657	struct memory_block *mem;
 
 
658
659	mutex_lock(&mem_sysfs_mutex);
660	mem = find_memory_block(section);
661	unregister_mem_sect_under_nodes(mem, __section_nr(section));
662
663	mem->section_count--;
664	if (mem->section_count == 0)
665		unregister_memory(mem);
666	else
667		put_device(&mem->dev);
668
669	mutex_unlock(&mem_sysfs_mutex);
670	return 0;
671}
672
673int unregister_memory_section(struct mem_section *section)
674{
675	if (!present_section(section))
676		return -EINVAL;
677
678	return remove_memory_block(0, section, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
679}
680#endif /* CONFIG_MEMORY_HOTREMOVE */
681
682/* return true if the memory block is offlined, otherwise, return false */
683bool is_memblock_offlined(struct memory_block *mem)
 
 
 
 
 
 
684{
685	return mem->state == MEM_OFFLINE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
686}
687
688static struct attribute *memory_root_attrs[] = {
689#ifdef CONFIG_ARCH_MEMORY_PROBE
690	&dev_attr_probe.attr,
691#endif
692
693#ifdef CONFIG_MEMORY_FAILURE
694	&dev_attr_soft_offline_page.attr,
695	&dev_attr_hard_offline_page.attr,
696#endif
697
698	&dev_attr_block_size_bytes.attr,
 
699	NULL
700};
701
702static struct attribute_group memory_root_attr_group = {
703	.attrs = memory_root_attrs,
704};
705
706static const struct attribute_group *memory_root_attr_groups[] = {
707	&memory_root_attr_group,
708	NULL,
709};
710
711/*
712 * Initialize the sysfs support for memory devices...
 
 
713 */
714int __init memory_dev_init(void)
715{
716	unsigned int i;
717	int ret;
718	int err;
719	unsigned long block_sz;
 
 
 
 
 
720
721	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
722	if (ret)
723		goto out;
724
725	block_sz = get_memory_block_size();
726	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
727
728	/*
729	 * Create entries for memory sections that were found
730	 * during boot and have been initialized
731	 */
732	mutex_lock(&mem_sysfs_mutex);
733	for (i = 0; i < NR_MEM_SECTIONS; i += sections_per_block) {
734		err = add_memory_block(i);
735		if (!ret)
736			ret = err;
 
737	}
738	mutex_unlock(&mem_sysfs_mutex);
739
740out:
741	if (ret)
742		printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
743	return ret;
744}
v6.2
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Memory subsystem support
   4 *
   5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
   6 *            Dave Hansen <haveblue@us.ibm.com>
   7 *
   8 * This file provides the necessary infrastructure to represent
   9 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
  10 * All arch-independent code that assumes MEMORY_HOTPLUG requires
  11 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
  12 */
  13
  14#include <linux/module.h>
  15#include <linux/init.h>
  16#include <linux/topology.h>
  17#include <linux/capability.h>
  18#include <linux/device.h>
  19#include <linux/memory.h>
  20#include <linux/memory_hotplug.h>
  21#include <linux/mm.h>
 
  22#include <linux/stat.h>
  23#include <linux/slab.h>
  24#include <linux/xarray.h>
  25
  26#include <linux/atomic.h>
  27#include <linux/uaccess.h>
 
 
  28
  29#define MEMORY_CLASS_NAME	"memory"
  30
  31static const char *const online_type_to_str[] = {
  32	[MMOP_OFFLINE] = "offline",
  33	[MMOP_ONLINE] = "online",
  34	[MMOP_ONLINE_KERNEL] = "online_kernel",
  35	[MMOP_ONLINE_MOVABLE] = "online_movable",
  36};
  37
  38int mhp_online_type_from_str(const char *str)
  39{
  40	int i;
  41
  42	for (i = 0; i < ARRAY_SIZE(online_type_to_str); i++) {
  43		if (sysfs_streq(str, online_type_to_str[i]))
  44			return i;
  45	}
  46	return -EINVAL;
  47}
  48
  49#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
  50
  51static int sections_per_block;
  52
  53static inline unsigned long memory_block_id(unsigned long section_nr)
  54{
  55	return section_nr / sections_per_block;
  56}
  57
  58static inline unsigned long pfn_to_block_id(unsigned long pfn)
  59{
  60	return memory_block_id(pfn_to_section_nr(pfn));
  61}
  62
  63static inline unsigned long phys_to_block_id(unsigned long phys)
  64{
  65	return pfn_to_block_id(PFN_DOWN(phys));
  66}
  67
  68static int memory_subsys_online(struct device *dev);
  69static int memory_subsys_offline(struct device *dev);
  70
  71static struct bus_type memory_subsys = {
  72	.name = MEMORY_CLASS_NAME,
  73	.dev_name = MEMORY_CLASS_NAME,
  74	.online = memory_subsys_online,
  75	.offline = memory_subsys_offline,
  76};
  77
  78/*
  79 * Memory blocks are cached in a local radix tree to avoid
  80 * a costly linear search for the corresponding device on
  81 * the subsystem bus.
  82 */
  83static DEFINE_XARRAY(memory_blocks);
  84
  85/*
  86 * Memory groups, indexed by memory group id (mgid).
  87 */
  88static DEFINE_XARRAY_FLAGS(memory_groups, XA_FLAGS_ALLOC);
  89#define MEMORY_GROUP_MARK_DYNAMIC	XA_MARK_1
  90
  91static BLOCKING_NOTIFIER_HEAD(memory_chain);
  92
  93int register_memory_notifier(struct notifier_block *nb)
  94{
  95	return blocking_notifier_chain_register(&memory_chain, nb);
  96}
  97EXPORT_SYMBOL(register_memory_notifier);
  98
  99void unregister_memory_notifier(struct notifier_block *nb)
 100{
 101	blocking_notifier_chain_unregister(&memory_chain, nb);
 102}
 103EXPORT_SYMBOL(unregister_memory_notifier);
 104
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 105static void memory_block_release(struct device *dev)
 106{
 107	struct memory_block *mem = to_memory_block(dev);
 108
 109	kfree(mem);
 110}
 111
 112unsigned long __weak memory_block_size_bytes(void)
 113{
 114	return MIN_MEMORY_BLOCK_SIZE;
 115}
 116EXPORT_SYMBOL_GPL(memory_block_size_bytes);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 117
 118/*
 119 * Show the first physical section index (number) of this memory block.
 
 120 */
 121static ssize_t phys_index_show(struct device *dev,
 122			       struct device_attribute *attr, char *buf)
 
 123{
 124	struct memory_block *mem = to_memory_block(dev);
 125	unsigned long phys_index;
 126
 127	phys_index = mem->start_section_nr / sections_per_block;
 
 
 
 
 
 
 
 
 128
 129	return sysfs_emit(buf, "%08lx\n", phys_index);
 
 130}
 131
 132/*
 133 * Legacy interface that we cannot remove. Always indicate "removable"
 134 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
 135 */
 136static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
 137			      char *buf)
 138{
 139	return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
 
 
 
 
 
 
 
 
 
 
 
 140}
 141
 142/*
 143 * online, offline, going offline, etc.
 144 */
 145static ssize_t state_show(struct device *dev, struct device_attribute *attr,
 146			  char *buf)
 147{
 148	struct memory_block *mem = to_memory_block(dev);
 149	const char *output;
 150
 151	/*
 152	 * We can probably put these states in a nice little array
 153	 * so that they're not open-coded
 154	 */
 155	switch (mem->state) {
 156	case MEM_ONLINE:
 157		output = "online";
 158		break;
 159	case MEM_OFFLINE:
 160		output = "offline";
 161		break;
 162	case MEM_GOING_OFFLINE:
 163		output = "going-offline";
 164		break;
 165	default:
 166		WARN_ON(1);
 167		return sysfs_emit(buf, "ERROR-UNKNOWN-%ld\n", mem->state);
 
 
 168	}
 169
 170	return sysfs_emit(buf, "%s\n", output);
 171}
 172
 173int memory_notify(unsigned long val, void *v)
 174{
 175	return blocking_notifier_call_chain(&memory_chain, val, v);
 176}
 177
 178#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
 179static unsigned long memblk_nr_poison(struct memory_block *mem);
 180#else
 181static inline unsigned long memblk_nr_poison(struct memory_block *mem)
 182{
 183	return 0;
 184}
 185#endif
 186
 187static int memory_block_online(struct memory_block *mem)
 
 
 
 
 188{
 189	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 190	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 191	unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
 192	struct zone *zone;
 193	int ret;
 194
 195	if (memblk_nr_poison(mem))
 196		return -EHWPOISON;
 197
 198	zone = zone_for_pfn_range(mem->online_type, mem->nid, mem->group,
 199				  start_pfn, nr_pages);
 200
 201	/*
 202	 * Although vmemmap pages have a different lifecycle than the pages
 203	 * they describe (they remain until the memory is unplugged), doing
 204	 * their initialization and accounting at memory onlining/offlining
 205	 * stage helps to keep accounting easier to follow - e.g vmemmaps
 206	 * belong to the same zone as the memory they backed.
 207	 */
 208	if (nr_vmemmap_pages) {
 209		ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
 210		if (ret)
 211			return ret;
 212	}
 213
 214	ret = online_pages(start_pfn + nr_vmemmap_pages,
 215			   nr_pages - nr_vmemmap_pages, zone, mem->group);
 216	if (ret) {
 217		if (nr_vmemmap_pages)
 218			mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
 219		return ret;
 220	}
 221
 222	/*
 223	 * Account once onlining succeeded. If the zone was unpopulated, it is
 224	 * now already properly populated.
 225	 */
 226	if (nr_vmemmap_pages)
 227		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
 228					  nr_vmemmap_pages);
 229
 230	mem->zone = zone;
 231	return ret;
 232}
 233
 234static int memory_block_offline(struct memory_block *mem)
 235{
 236	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 237	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 238	unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
 239	int ret;
 240
 241	if (!mem->zone)
 242		return -EINVAL;
 243
 244	/*
 245	 * Unaccount before offlining, such that unpopulated zone and kthreads
 246	 * can properly be torn down in offline_pages().
 247	 */
 248	if (nr_vmemmap_pages)
 249		adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
 250					  -nr_vmemmap_pages);
 251
 252	ret = offline_pages(start_pfn + nr_vmemmap_pages,
 253			    nr_pages - nr_vmemmap_pages, mem->zone, mem->group);
 254	if (ret) {
 255		/* offline_pages() failed. Account back. */
 256		if (nr_vmemmap_pages)
 257			adjust_present_page_count(pfn_to_page(start_pfn),
 258						  mem->group, nr_vmemmap_pages);
 259		return ret;
 260	}
 261
 262	if (nr_vmemmap_pages)
 263		mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
 264
 265	mem->zone = NULL;
 266	return ret;
 267}
 268
 269/*
 270 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 271 * OK to have direct references to sparsemem variables in here.
 272 */
 273static int
 274memory_block_action(struct memory_block *mem, unsigned long action)
 275{
 
 
 
 276	int ret;
 277
 
 
 
 278	switch (action) {
 279	case MEM_ONLINE:
 280		ret = memory_block_online(mem);
 281		break;
 282	case MEM_OFFLINE:
 283		ret = memory_block_offline(mem);
 284		break;
 285	default:
 286		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
 287		     "%ld\n", __func__, mem->start_section_nr, action, action);
 288		ret = -EINVAL;
 
 
 
 289	}
 290
 291	return ret;
 292}
 293
 294static int memory_block_change_state(struct memory_block *mem,
 295		unsigned long to_state, unsigned long from_state_req)
 296{
 297	int ret = 0;
 298
 299	if (mem->state != from_state_req)
 300		return -EINVAL;
 301
 302	if (to_state == MEM_OFFLINE)
 303		mem->state = MEM_GOING_OFFLINE;
 304
 305	ret = memory_block_action(mem, to_state);
 
 
 306	mem->state = ret ? from_state_req : to_state;
 307
 308	return ret;
 309}
 310
 311/* The device lock serializes operations on memory_subsys_[online|offline] */
 312static int memory_subsys_online(struct device *dev)
 313{
 314	struct memory_block *mem = to_memory_block(dev);
 315	int ret;
 316
 317	if (mem->state == MEM_ONLINE)
 318		return 0;
 319
 320	/*
 321	 * When called via device_online() without configuring the online_type,
 322	 * we want to default to MMOP_ONLINE.
 
 323	 */
 324	if (mem->online_type == MMOP_OFFLINE)
 325		mem->online_type = MMOP_ONLINE;
 326
 327	ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 328	mem->online_type = MMOP_OFFLINE;
 
 
 329
 330	return ret;
 331}
 332
 333static int memory_subsys_offline(struct device *dev)
 334{
 335	struct memory_block *mem = to_memory_block(dev);
 336
 337	if (mem->state == MEM_OFFLINE)
 338		return 0;
 339
 340	return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 341}
 342
 343static ssize_t state_store(struct device *dev, struct device_attribute *attr,
 344			   const char *buf, size_t count)
 
 345{
 346	const int online_type = mhp_online_type_from_str(buf);
 347	struct memory_block *mem = to_memory_block(dev);
 348	int ret;
 349
 350	if (online_type < 0)
 351		return -EINVAL;
 352
 353	ret = lock_device_hotplug_sysfs();
 354	if (ret)
 355		return ret;
 356
 
 
 
 
 
 
 
 
 
 
 
 
 
 357	switch (online_type) {
 358	case MMOP_ONLINE_KERNEL:
 359	case MMOP_ONLINE_MOVABLE:
 360	case MMOP_ONLINE:
 361		/* mem->online_type is protected by device_hotplug_lock */
 
 
 
 
 
 
 362		mem->online_type = online_type;
 363		ret = device_online(&mem->dev);
 364		break;
 365	case MMOP_OFFLINE:
 366		ret = device_offline(&mem->dev);
 367		break;
 368	default:
 369		ret = -EINVAL; /* should never happen */
 370	}
 371
 
 372	unlock_device_hotplug();
 373
 374	if (ret < 0)
 375		return ret;
 376	if (ret)
 377		return -EINVAL;
 378
 379	return count;
 380}
 381
 382/*
 383 * Legacy interface that we cannot remove: s390x exposes the storage increment
 384 * covered by a memory block, allowing for identifying which memory blocks
 385 * comprise a storage increment. Since a memory block spans complete
 386 * storage increments nowadays, this interface is basically unused. Other
 387 * archs never exposed != 0.
 
 
 388 */
 389static ssize_t phys_device_show(struct device *dev,
 390				struct device_attribute *attr, char *buf)
 391{
 392	struct memory_block *mem = to_memory_block(dev);
 393	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 394
 395	return sysfs_emit(buf, "%d\n",
 396			  arch_get_memory_phys_device(start_pfn));
 397}
 398
 399#ifdef CONFIG_MEMORY_HOTREMOVE
 400static int print_allowed_zone(char *buf, int len, int nid,
 401			      struct memory_group *group,
 402			      unsigned long start_pfn, unsigned long nr_pages,
 403			      int online_type, struct zone *default_zone)
 404{
 405	struct zone *zone;
 406
 407	zone = zone_for_pfn_range(online_type, nid, group, start_pfn, nr_pages);
 408	if (zone == default_zone)
 409		return 0;
 410
 411	return sysfs_emit_at(buf, len, " %s", zone->name);
 412}
 413
 414static ssize_t valid_zones_show(struct device *dev,
 415				struct device_attribute *attr, char *buf)
 416{
 417	struct memory_block *mem = to_memory_block(dev);
 418	unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 419	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 420	struct memory_group *group = mem->group;
 421	struct zone *default_zone;
 422	int nid = mem->nid;
 423	int len = 0;
 424
 425	/*
 426	 * Check the existing zone. Make sure that we do that only on the
 427	 * online nodes otherwise the page_zone is not reliable
 428	 */
 429	if (mem->state == MEM_ONLINE) {
 430		/*
 431		 * If !mem->zone, the memory block spans multiple zones and
 432		 * cannot get offlined.
 433		 */
 434		default_zone = mem->zone;
 435		if (!default_zone)
 436			return sysfs_emit(buf, "%s\n", "none");
 437		len += sysfs_emit_at(buf, len, "%s", default_zone->name);
 438		goto out;
 439	}
 440
 441	default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, group,
 442					  start_pfn, nr_pages);
 443
 444	len += sysfs_emit_at(buf, len, "%s", default_zone->name);
 445	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
 446				  MMOP_ONLINE_KERNEL, default_zone);
 447	len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
 448				  MMOP_ONLINE_MOVABLE, default_zone);
 449out:
 450	len += sysfs_emit_at(buf, len, "\n");
 451	return len;
 452}
 453static DEVICE_ATTR_RO(valid_zones);
 454#endif
 455
 456static DEVICE_ATTR_RO(phys_index);
 457static DEVICE_ATTR_RW(state);
 458static DEVICE_ATTR_RO(phys_device);
 459static DEVICE_ATTR_RO(removable);
 460
 461/*
 462 * Show the memory block size (shared by all memory blocks).
 463 */
 464static ssize_t block_size_bytes_show(struct device *dev,
 465				     struct device_attribute *attr, char *buf)
 
 466{
 467	return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
 468}
 469
 470static DEVICE_ATTR_RO(block_size_bytes);
 471
 472/*
 473 * Memory auto online policy.
 474 */
 475
 476static ssize_t auto_online_blocks_show(struct device *dev,
 477				       struct device_attribute *attr, char *buf)
 478{
 479	return sysfs_emit(buf, "%s\n",
 480			  online_type_to_str[mhp_default_online_type]);
 481}
 482
 483static ssize_t auto_online_blocks_store(struct device *dev,
 484					struct device_attribute *attr,
 485					const char *buf, size_t count)
 486{
 487	const int online_type = mhp_online_type_from_str(buf);
 488
 489	if (online_type < 0)
 490		return -EINVAL;
 491
 492	mhp_default_online_type = online_type;
 493	return count;
 494}
 495
 496static DEVICE_ATTR_RW(auto_online_blocks);
 497
 498/*
 499 * Some architectures will have custom drivers to do this, and
 500 * will not need to do it from userspace.  The fake hot-add code
 501 * as well as ppc64 will do all of their discovery in userspace
 502 * and will require this interface.
 503 */
 504#ifdef CONFIG_ARCH_MEMORY_PROBE
 505static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
 506			   const char *buf, size_t count)
 
 507{
 508	u64 phys_addr;
 509	int nid, ret;
 
 510	unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
 511
 512	ret = kstrtoull(buf, 0, &phys_addr);
 513	if (ret)
 514		return ret;
 515
 516	if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
 517		return -EINVAL;
 518
 519	ret = lock_device_hotplug_sysfs();
 520	if (ret)
 521		return ret;
 
 
 
 522
 523	nid = memory_add_physaddr_to_nid(phys_addr);
 524	ret = __add_memory(nid, phys_addr,
 525			   MIN_MEMORY_BLOCK_SIZE * sections_per_block,
 526			   MHP_NONE);
 527
 528	if (ret)
 529		goto out;
 530
 531	ret = count;
 532out:
 533	unlock_device_hotplug();
 534	return ret;
 535}
 536
 537static DEVICE_ATTR_WO(probe);
 538#endif
 539
 540#ifdef CONFIG_MEMORY_FAILURE
 541/*
 542 * Support for offlining pages of memory
 543 */
 544
 545/* Soft offline a page */
 546static ssize_t soft_offline_page_store(struct device *dev,
 547				       struct device_attribute *attr,
 548				       const char *buf, size_t count)
 
 549{
 550	int ret;
 551	u64 pfn;
 552	if (!capable(CAP_SYS_ADMIN))
 553		return -EPERM;
 554	if (kstrtoull(buf, 0, &pfn) < 0)
 555		return -EINVAL;
 556	pfn >>= PAGE_SHIFT;
 557	ret = soft_offline_page(pfn, 0);
 
 
 558	return ret == 0 ? count : ret;
 559}
 560
 561/* Forcibly offline a page, including killing processes. */
 562static ssize_t hard_offline_page_store(struct device *dev,
 563				       struct device_attribute *attr,
 564				       const char *buf, size_t count)
 
 565{
 566	int ret;
 567	u64 pfn;
 568	if (!capable(CAP_SYS_ADMIN))
 569		return -EPERM;
 570	if (kstrtoull(buf, 0, &pfn) < 0)
 571		return -EINVAL;
 572	pfn >>= PAGE_SHIFT;
 573	ret = memory_failure(pfn, MF_SW_SIMULATED);
 574	if (ret == -EOPNOTSUPP)
 575		ret = 0;
 576	return ret ? ret : count;
 577}
 578
 579static DEVICE_ATTR_WO(soft_offline_page);
 580static DEVICE_ATTR_WO(hard_offline_page);
 581#endif
 582
 583/* See phys_device_show(). */
 
 
 
 
 584int __weak arch_get_memory_phys_device(unsigned long start_pfn)
 585{
 586	return 0;
 587}
 588
 589/*
 590 * A reference for the returned memory block device is acquired.
 591 *
 592 * Called under device_hotplug_lock.
 593 */
 594static struct memory_block *find_memory_block_by_id(unsigned long block_id)
 
 595{
 596	struct memory_block *mem;
 
 
 597
 598	mem = xa_load(&memory_blocks, block_id);
 599	if (mem)
 600		get_device(&mem->dev);
 601	return mem;
 
 
 602}
 603
 604/*
 605 * Called under device_hotplug_lock.
 
 
 
 
 
 606 */
 607struct memory_block *find_memory_block(unsigned long section_nr)
 608{
 609	unsigned long block_id = memory_block_id(section_nr);
 610
 611	return find_memory_block_by_id(block_id);
 612}
 613
 614static struct attribute *memory_memblk_attrs[] = {
 615	&dev_attr_phys_index.attr,
 
 616	&dev_attr_state.attr,
 617	&dev_attr_phys_device.attr,
 618	&dev_attr_removable.attr,
 619#ifdef CONFIG_MEMORY_HOTREMOVE
 620	&dev_attr_valid_zones.attr,
 621#endif
 622	NULL
 623};
 624
 625static const struct attribute_group memory_memblk_attr_group = {
 626	.attrs = memory_memblk_attrs,
 627};
 628
 629static const struct attribute_group *memory_memblk_attr_groups[] = {
 630	&memory_memblk_attr_group,
 631	NULL,
 632};
 633
 634static int __add_memory_block(struct memory_block *memory)
 
 
 
 
 635{
 636	int ret;
 637
 638	memory->dev.bus = &memory_subsys;
 639	memory->dev.id = memory->start_section_nr / sections_per_block;
 640	memory->dev.release = memory_block_release;
 641	memory->dev.groups = memory_memblk_attr_groups;
 642	memory->dev.offline = memory->state == MEM_OFFLINE;
 643
 644	ret = device_register(&memory->dev);
 645	if (ret) {
 646		put_device(&memory->dev);
 647		return ret;
 648	}
 649	ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
 650			      GFP_KERNEL));
 651	if (ret)
 652		device_unregister(&memory->dev);
 653
 654	return ret;
 655}
 656
 657static struct zone *early_node_zone_for_memory_block(struct memory_block *mem,
 658						     int nid)
 659{
 660	const unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
 661	const unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
 662	struct zone *zone, *matching_zone = NULL;
 663	pg_data_t *pgdat = NODE_DATA(nid);
 664	int i;
 665
 666	/*
 667	 * This logic only works for early memory, when the applicable zones
 668	 * already span the memory block. We don't expect overlapping zones on
 669	 * a single node for early memory. So if we're told that some PFNs
 670	 * of a node fall into this memory block, we can assume that all node
 671	 * zones that intersect with the memory block are actually applicable.
 672	 * No need to look at the memmap.
 673	 */
 674	for (i = 0; i < MAX_NR_ZONES; i++) {
 675		zone = pgdat->node_zones + i;
 676		if (!populated_zone(zone))
 677			continue;
 678		if (!zone_intersects(zone, start_pfn, nr_pages))
 679			continue;
 680		if (!matching_zone) {
 681			matching_zone = zone;
 682			continue;
 683		}
 684		/* Spans multiple zones ... */
 685		matching_zone = NULL;
 686		break;
 687	}
 688	return matching_zone;
 689}
 690
 691#ifdef CONFIG_NUMA
 692/**
 693 * memory_block_add_nid() - Indicate that system RAM falling into this memory
 694 *			    block device (partially) belongs to the given node.
 695 * @mem: The memory block device.
 696 * @nid: The node id.
 697 * @context: The memory initialization context.
 698 *
 699 * Indicate that system RAM falling into this memory block (partially) belongs
 700 * to the given node. If the context indicates ("early") that we are adding the
 701 * node during node device subsystem initialization, this will also properly
 702 * set/adjust mem->zone based on the zone ranges of the given node.
 703 */
 704void memory_block_add_nid(struct memory_block *mem, int nid,
 705			  enum meminit_context context)
 706{
 707	if (context == MEMINIT_EARLY && mem->nid != nid) {
 708		/*
 709		 * For early memory we have to determine the zone when setting
 710		 * the node id and handle multiple nodes spanning a single
 711		 * memory block by indicate via zone == NULL that we're not
 712		 * dealing with a single zone. So if we're setting the node id
 713		 * the first time, determine if there is a single zone. If we're
 714		 * setting the node id a second time to a different node,
 715		 * invalidate the single detected zone.
 716		 */
 717		if (mem->nid == NUMA_NO_NODE)
 718			mem->zone = early_node_zone_for_memory_block(mem, nid);
 719		else
 720			mem->zone = NULL;
 721	}
 722
 723	/*
 724	 * If this memory block spans multiple nodes, we only indicate
 725	 * the last processed node. If we span multiple nodes (not applicable
 726	 * to hotplugged memory), zone == NULL will prohibit memory offlining
 727	 * and consequently unplug.
 728	 */
 729	mem->nid = nid;
 730}
 731#endif
 732
 733static int add_memory_block(unsigned long block_id, unsigned long state,
 734			    unsigned long nr_vmemmap_pages,
 735			    struct memory_group *group)
 736{
 737	struct memory_block *mem;
 
 
 738	int ret = 0;
 739
 740	mem = find_memory_block_by_id(block_id);
 741	if (mem) {
 742		put_device(&mem->dev);
 743		return -EEXIST;
 744	}
 745	mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 746	if (!mem)
 747		return -ENOMEM;
 748
 749	mem->start_section_nr = block_id * sections_per_block;
 
 
 
 750	mem->state = state;
 751	mem->nid = NUMA_NO_NODE;
 752	mem->nr_vmemmap_pages = nr_vmemmap_pages;
 753	INIT_LIST_HEAD(&mem->group_next);
 754
 755#ifndef CONFIG_NUMA
 756	if (state == MEM_ONLINE)
 757		/*
 758		 * MEM_ONLINE at this point implies early memory. With NUMA,
 759		 * we'll determine the zone when setting the node id via
 760		 * memory_block_add_nid(). Memory hotplug updated the zone
 761		 * manually when memory onlining/offlining succeeds.
 762		 */
 763		mem->zone = early_node_zone_for_memory_block(mem, NUMA_NO_NODE);
 764#endif /* CONFIG_NUMA */
 765
 766	ret = __add_memory_block(mem);
 767	if (ret)
 768		return ret;
 769
 770	if (group) {
 771		mem->group = group;
 772		list_add(&mem->group_next, &group->memory_blocks);
 773	}
 774
 775	return 0;
 776}
 777
 778static int __init add_boot_memory_block(unsigned long base_section_nr)
 779{
 780	int section_count = 0;
 781	unsigned long nr;
 782
 783	for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
 784	     nr++)
 785		if (present_section_nr(nr))
 786			section_count++;
 
 
 
 
 
 787
 788	if (section_count == 0)
 789		return 0;
 790	return add_memory_block(memory_block_id(base_section_nr),
 791				MEM_ONLINE, 0,  NULL);
 
 
 
 792}
 793
 794static int add_hotplug_memory_block(unsigned long block_id,
 795				    unsigned long nr_vmemmap_pages,
 796				    struct memory_group *group)
 797{
 798	return add_memory_block(block_id, MEM_OFFLINE, nr_vmemmap_pages, group);
 799}
 800
 801static void remove_memory_block(struct memory_block *memory)
 
 
 
 
 802{
 803	if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
 804		return;
 805
 806	WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
 807
 808	if (memory->group) {
 809		list_del(&memory->group_next);
 810		memory->group = NULL;
 
 
 
 
 
 811	}
 812
 813	/* drop the ref. we got via find_memory_block() */
 
 
 
 
 
 
 
 
 
 
 
 
 
 814	put_device(&memory->dev);
 815	device_unregister(&memory->dev);
 816}
 817
 818/*
 819 * Create memory block devices for the given memory area. Start and size
 820 * have to be aligned to memory block granularity. Memory block devices
 821 * will be initialized as offline.
 822 *
 823 * Called under device_hotplug_lock.
 824 */
 825int create_memory_block_devices(unsigned long start, unsigned long size,
 826				unsigned long vmemmap_pages,
 827				struct memory_group *group)
 828{
 829	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
 830	unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
 831	struct memory_block *mem;
 832	unsigned long block_id;
 833	int ret = 0;
 834
 835	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
 836			 !IS_ALIGNED(size, memory_block_size_bytes())))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 837		return -EINVAL;
 838
 839	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
 840		ret = add_hotplug_memory_block(block_id, vmemmap_pages, group);
 841		if (ret)
 842			break;
 843	}
 844	if (ret) {
 845		end_block_id = block_id;
 846		for (block_id = start_block_id; block_id != end_block_id;
 847		     block_id++) {
 848			mem = find_memory_block_by_id(block_id);
 849			if (WARN_ON_ONCE(!mem))
 850				continue;
 851			remove_memory_block(mem);
 852		}
 853	}
 854	return ret;
 855}
 
 856
 857/*
 858 * Remove memory block devices for the given memory area. Start and size
 859 * have to be aligned to memory block granularity. Memory block devices
 860 * have to be offline.
 861 *
 862 * Called under device_hotplug_lock.
 863 */
 864void remove_memory_block_devices(unsigned long start, unsigned long size)
 865{
 866	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
 867	const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
 868	struct memory_block *mem;
 869	unsigned long block_id;
 870
 871	if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
 872			 !IS_ALIGNED(size, memory_block_size_bytes())))
 873		return;
 874
 875	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
 876		mem = find_memory_block_by_id(block_id);
 877		if (WARN_ON_ONCE(!mem))
 878			continue;
 879		num_poisoned_pages_sub(-1UL, memblk_nr_poison(mem));
 880		unregister_memory_block_under_nodes(mem);
 881		remove_memory_block(mem);
 882	}
 883}
 884
 885static struct attribute *memory_root_attrs[] = {
 886#ifdef CONFIG_ARCH_MEMORY_PROBE
 887	&dev_attr_probe.attr,
 888#endif
 889
 890#ifdef CONFIG_MEMORY_FAILURE
 891	&dev_attr_soft_offline_page.attr,
 892	&dev_attr_hard_offline_page.attr,
 893#endif
 894
 895	&dev_attr_block_size_bytes.attr,
 896	&dev_attr_auto_online_blocks.attr,
 897	NULL
 898};
 899
 900static const struct attribute_group memory_root_attr_group = {
 901	.attrs = memory_root_attrs,
 902};
 903
 904static const struct attribute_group *memory_root_attr_groups[] = {
 905	&memory_root_attr_group,
 906	NULL,
 907};
 908
 909/*
 910 * Initialize the sysfs support for memory devices. At the time this function
 911 * is called, we cannot have concurrent creation/deletion of memory block
 912 * devices, the device_hotplug_lock is not needed.
 913 */
 914void __init memory_dev_init(void)
 915{
 
 916	int ret;
 917	unsigned long block_sz, nr;
 918
 919	/* Validate the configured memory block size */
 920	block_sz = memory_block_size_bytes();
 921	if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
 922		panic("Memory block size not suitable: 0x%lx\n", block_sz);
 923	sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
 924
 925	ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
 926	if (ret)
 927		panic("%s() failed to register subsystem: %d\n", __func__, ret);
 
 
 
 928
 929	/*
 930	 * Create entries for memory sections that were found
 931	 * during boot and have been initialized
 932	 */
 933	for (nr = 0; nr <= __highest_present_section_nr;
 934	     nr += sections_per_block) {
 935		ret = add_boot_memory_block(nr);
 936		if (ret)
 937			panic("%s() failed to add memory block: %d\n", __func__,
 938			      ret);
 939	}
 940}
 941
 942/**
 943 * walk_memory_blocks - walk through all present memory blocks overlapped
 944 *			by the range [start, start + size)
 945 *
 946 * @start: start address of the memory range
 947 * @size: size of the memory range
 948 * @arg: argument passed to func
 949 * @func: callback for each memory section walked
 950 *
 951 * This function walks through all present memory blocks overlapped by the
 952 * range [start, start + size), calling func on each memory block.
 953 *
 954 * In case func() returns an error, walking is aborted and the error is
 955 * returned.
 956 *
 957 * Called under device_hotplug_lock.
 958 */
 959int walk_memory_blocks(unsigned long start, unsigned long size,
 960		       void *arg, walk_memory_blocks_func_t func)
 961{
 962	const unsigned long start_block_id = phys_to_block_id(start);
 963	const unsigned long end_block_id = phys_to_block_id(start + size - 1);
 964	struct memory_block *mem;
 965	unsigned long block_id;
 966	int ret = 0;
 967
 968	if (!size)
 969		return 0;
 970
 971	for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
 972		mem = find_memory_block_by_id(block_id);
 973		if (!mem)
 974			continue;
 975
 976		ret = func(mem, arg);
 977		put_device(&mem->dev);
 978		if (ret)
 979			break;
 980	}
 981	return ret;
 982}
 983
 984struct for_each_memory_block_cb_data {
 985	walk_memory_blocks_func_t func;
 986	void *arg;
 987};
 988
 989static int for_each_memory_block_cb(struct device *dev, void *data)
 990{
 991	struct memory_block *mem = to_memory_block(dev);
 992	struct for_each_memory_block_cb_data *cb_data = data;
 993
 994	return cb_data->func(mem, cb_data->arg);
 995}
 996
 997/**
 998 * for_each_memory_block - walk through all present memory blocks
 999 *
1000 * @arg: argument passed to func
1001 * @func: callback for each memory block walked
1002 *
1003 * This function walks through all present memory blocks, calling func on
1004 * each memory block.
1005 *
1006 * In case func() returns an error, walking is aborted and the error is
1007 * returned.
1008 */
1009int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
1010{
1011	struct for_each_memory_block_cb_data cb_data = {
1012		.func = func,
1013		.arg = arg,
1014	};
1015
1016	return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
1017				for_each_memory_block_cb);
1018}
1019
1020/*
1021 * This is an internal helper to unify allocation and initialization of
1022 * memory groups. Note that the passed memory group will be copied to a
1023 * dynamically allocated memory group. After this call, the passed
1024 * memory group should no longer be used.
1025 */
1026static int memory_group_register(struct memory_group group)
1027{
1028	struct memory_group *new_group;
1029	uint32_t mgid;
1030	int ret;
1031
1032	if (!node_possible(group.nid))
1033		return -EINVAL;
1034
1035	new_group = kzalloc(sizeof(group), GFP_KERNEL);
1036	if (!new_group)
1037		return -ENOMEM;
1038	*new_group = group;
1039	INIT_LIST_HEAD(&new_group->memory_blocks);
1040
1041	ret = xa_alloc(&memory_groups, &mgid, new_group, xa_limit_31b,
1042		       GFP_KERNEL);
1043	if (ret) {
1044		kfree(new_group);
1045		return ret;
1046	} else if (group.is_dynamic) {
1047		xa_set_mark(&memory_groups, mgid, MEMORY_GROUP_MARK_DYNAMIC);
1048	}
1049	return mgid;
1050}
1051
1052/**
1053 * memory_group_register_static() - Register a static memory group.
1054 * @nid: The node id.
1055 * @max_pages: The maximum number of pages we'll have in this static memory
1056 *	       group.
1057 *
1058 * Register a new static memory group and return the memory group id.
1059 * All memory in the group belongs to a single unit, such as a DIMM. All
1060 * memory belonging to a static memory group is added in one go to be removed
1061 * in one go -- it's static.
1062 *
1063 * Returns an error if out of memory, if the node id is invalid, if no new
1064 * memory groups can be registered, or if max_pages is invalid (0). Otherwise,
1065 * returns the new memory group id.
1066 */
1067int memory_group_register_static(int nid, unsigned long max_pages)
1068{
1069	struct memory_group group = {
1070		.nid = nid,
1071		.s = {
1072			.max_pages = max_pages,
1073		},
1074	};
1075
1076	if (!max_pages)
1077		return -EINVAL;
1078	return memory_group_register(group);
1079}
1080EXPORT_SYMBOL_GPL(memory_group_register_static);
1081
1082/**
1083 * memory_group_register_dynamic() - Register a dynamic memory group.
1084 * @nid: The node id.
1085 * @unit_pages: Unit in pages in which is memory added/removed in this dynamic
1086 *		memory group.
1087 *
1088 * Register a new dynamic memory group and return the memory group id.
1089 * Memory within a dynamic memory group is added/removed dynamically
1090 * in unit_pages.
1091 *
1092 * Returns an error if out of memory, if the node id is invalid, if no new
1093 * memory groups can be registered, or if unit_pages is invalid (0, not a
1094 * power of two, smaller than a single memory block). Otherwise, returns the
1095 * new memory group id.
1096 */
1097int memory_group_register_dynamic(int nid, unsigned long unit_pages)
1098{
1099	struct memory_group group = {
1100		.nid = nid,
1101		.is_dynamic = true,
1102		.d = {
1103			.unit_pages = unit_pages,
1104		},
1105	};
1106
1107	if (!unit_pages || !is_power_of_2(unit_pages) ||
1108	    unit_pages < PHYS_PFN(memory_block_size_bytes()))
1109		return -EINVAL;
1110	return memory_group_register(group);
1111}
1112EXPORT_SYMBOL_GPL(memory_group_register_dynamic);
1113
1114/**
1115 * memory_group_unregister() - Unregister a memory group.
1116 * @mgid: the memory group id
1117 *
1118 * Unregister a memory group. If any memory block still belongs to this
1119 * memory group, unregistering will fail.
1120 *
1121 * Returns -EINVAL if the memory group id is invalid, returns -EBUSY if some
1122 * memory blocks still belong to this memory group and returns 0 if
1123 * unregistering succeeded.
1124 */
1125int memory_group_unregister(int mgid)
1126{
1127	struct memory_group *group;
1128
1129	if (mgid < 0)
1130		return -EINVAL;
1131
1132	group = xa_load(&memory_groups, mgid);
1133	if (!group)
1134		return -EINVAL;
1135	if (!list_empty(&group->memory_blocks))
1136		return -EBUSY;
1137	xa_erase(&memory_groups, mgid);
1138	kfree(group);
1139	return 0;
1140}
1141EXPORT_SYMBOL_GPL(memory_group_unregister);
1142
1143/*
1144 * This is an internal helper only to be used in core memory hotplug code to
1145 * lookup a memory group. We don't care about locking, as we don't expect a
1146 * memory group to get unregistered while adding memory to it -- because
1147 * the group and the memory is managed by the same driver.
1148 */
1149struct memory_group *memory_group_find_by_id(int mgid)
1150{
1151	return xa_load(&memory_groups, mgid);
1152}
1153
1154/*
1155 * This is an internal helper only to be used in core memory hotplug code to
1156 * walk all dynamic memory groups excluding a given memory group, either
1157 * belonging to a specific node, or belonging to any node.
1158 */
1159int walk_dynamic_memory_groups(int nid, walk_memory_groups_func_t func,
1160			       struct memory_group *excluded, void *arg)
1161{
1162	struct memory_group *group;
1163	unsigned long index;
1164	int ret = 0;
1165
1166	xa_for_each_marked(&memory_groups, index, group,
1167			   MEMORY_GROUP_MARK_DYNAMIC) {
1168		if (group == excluded)
1169			continue;
1170#ifdef CONFIG_NUMA
1171		if (nid != NUMA_NO_NODE && group->nid != nid)
1172			continue;
1173#endif /* CONFIG_NUMA */
1174		ret = func(group, arg);
1175		if (ret)
1176			break;
1177	}
1178	return ret;
1179}
1180
1181#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
1182void memblk_nr_poison_inc(unsigned long pfn)
1183{
1184	const unsigned long block_id = pfn_to_block_id(pfn);
1185	struct memory_block *mem = find_memory_block_by_id(block_id);
1186
1187	if (mem)
1188		atomic_long_inc(&mem->nr_hwpoison);
1189}
1190
1191void memblk_nr_poison_sub(unsigned long pfn, long i)
1192{
1193	const unsigned long block_id = pfn_to_block_id(pfn);
1194	struct memory_block *mem = find_memory_block_by_id(block_id);
1195
1196	if (mem)
1197		atomic_long_sub(i, &mem->nr_hwpoison);
1198}
1199
1200static unsigned long memblk_nr_poison(struct memory_block *mem)
1201{
1202	return atomic_long_read(&mem->nr_hwpoison);
1203}
1204#endif