projects/linux·blob uaccess-cleanups · refs

1254 lines · 32.7 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Memory subsystem support
4 *
5 * Written by Matt Tolentino <[email protected]>
6 * Dave Hansen <[email protected]>
7 *
8 * This file provides the necessary infrastructure to represent
9 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
10 * All arch-independent code that assumes MEMORY_HOTPLUG requires
11 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
12 */
13
14#include <linux/module.h>
15#include <linux/init.h>
16#include <linux/topology.h>
17#include <linux/capability.h>
18#include <linux/device.h>
19#include <linux/memory.h>
20#include <linux/memory_hotplug.h>
21#include <linux/mm.h>
22#include <linux/stat.h>
23#include <linux/slab.h>
24#include <linux/xarray.h>
25#include <linux/export.h>
26
27#include <linux/atomic.h>
28#include <linux/uaccess.h>
29
30#define MEMORY_CLASS_NAME "memory"
31
32static const char *const online_type_to_str[] = {
33 [MMOP_OFFLINE] = "offline",
34 [MMOP_ONLINE] = "online",
35 [MMOP_ONLINE_KERNEL] = "online_kernel",
36 [MMOP_ONLINE_MOVABLE] = "online_movable",
37};
38
39int mhp_online_type_from_str(const char *str)
40{
41 int i;
42
43 for (i = 0; i < ARRAY_SIZE(online_type_to_str); i++) {
44 if (sysfs_streq(str, online_type_to_str[i]))
45 return i;
46 }
47 return -EINVAL;
48}
49EXPORT_SYMBOL_GPL(mhp_online_type_from_str);
50
51const char *mhp_online_type_to_str(int online_type)
52{
53 if (online_type < 0 || online_type >= (int)ARRAY_SIZE(online_type_to_str))
54 return NULL;
55 return online_type_to_str[online_type];
56}
57EXPORT_SYMBOL_GPL(mhp_online_type_to_str);
58
59#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
60
61int sections_per_block;
62EXPORT_SYMBOL(sections_per_block);
63
64static int memory_subsys_online(struct device *dev);
65static int memory_subsys_offline(struct device *dev);
66
67static const struct bus_type memory_subsys = {
68 .name = MEMORY_CLASS_NAME,
69 .dev_name = MEMORY_CLASS_NAME,
70 .online = memory_subsys_online,
71 .offline = memory_subsys_offline,
72};
73
74/*
75 * Memory blocks are cached in a local radix tree to avoid
76 * a costly linear search for the corresponding device on
77 * the subsystem bus.
78 */
79static DEFINE_XARRAY(memory_blocks);
80
81/*
82 * Memory groups, indexed by memory group id (mgid).
83 */
84static DEFINE_XARRAY_FLAGS(memory_groups, XA_FLAGS_ALLOC);
85#define MEMORY_GROUP_MARK_DYNAMIC XA_MARK_1
86
87static BLOCKING_NOTIFIER_HEAD(memory_chain);
88
89int register_memory_notifier(struct notifier_block *nb)
90{
91 return blocking_notifier_chain_register(&memory_chain, nb);
92}
93EXPORT_SYMBOL(register_memory_notifier);
94
95void unregister_memory_notifier(struct notifier_block *nb)
96{
97 blocking_notifier_chain_unregister(&memory_chain, nb);
98}
99EXPORT_SYMBOL(unregister_memory_notifier);
100
101static void memory_block_release(struct device *dev)
102{
103 struct memory_block *mem = to_memory_block(dev);
104 /* Verify that the altmap is freed */
105 WARN_ON(mem->altmap);
106 kfree(mem);
107}
108
109
110/* Max block size to be set by memory_block_advise_max_size */
111static unsigned long memory_block_advised_size;
112static bool memory_block_advised_size_queried;
113
114/**
115 * memory_block_advise_max_size() - advise memory hotplug on the max suggested
116 * block size, usually for alignment.
117 * @size: suggestion for maximum block size. must be aligned on power of 2.
118 *
119 * Early boot software (pre-allocator init) may advise archs on the max block
120 * size. This value can only decrease after initialization, as the intent is
121 * to identify the largest supported alignment for all sources.
122 *
123 * Use of this value is arch-defined, as is min/max block size.
124 *
125 * Return: 0 on success
126 * -EINVAL if size is 0 or not pow2 aligned
127 * -EBUSY if value has already been probed
128 */
129int __init memory_block_advise_max_size(unsigned long size)
130{
131 if (!size || !is_power_of_2(size))
132 return -EINVAL;
133
134 if (memory_block_advised_size_queried)
135 return -EBUSY;
136
137 if (memory_block_advised_size)
138 memory_block_advised_size = min(memory_block_advised_size, size);
139 else
140 memory_block_advised_size = size;
141
142 return 0;
143}
144
145/**
146 * memory_block_advised_max_size() - query advised max hotplug block size.
147 *
148 * After the first call, the value can never change. Callers looking for the
149 * actual block size should use memory_block_size_bytes. This interface is
150 * intended for use by arch-init when initializing the hotplug block size.
151 *
152 * Return: advised size in bytes, or 0 if never set.
153 */
154unsigned long memory_block_advised_max_size(void)
155{
156 memory_block_advised_size_queried = true;
157 return memory_block_advised_size;
158}
159
160unsigned long __weak memory_block_size_bytes(void)
161{
162 return MIN_MEMORY_BLOCK_SIZE;
163}
164EXPORT_SYMBOL_GPL(memory_block_size_bytes);
165
166/* Show the memory block ID, relative to the memory block size */
167static ssize_t phys_index_show(struct device *dev,
168 struct device_attribute *attr, char *buf)
169{
170 struct memory_block *mem = to_memory_block(dev);
171
172 return sysfs_emit(buf, "%08lx\n", memory_block_id(mem->start_section_nr));
173}
174
175/*
176 * Legacy interface that we cannot remove. Always indicate "removable"
177 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
178 */
179static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
180 char *buf)
181{
182 return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
183}
184
185/*
186 * online, offline, going offline, etc.
187 */
188static ssize_t state_show(struct device *dev, struct device_attribute *attr,
189 char *buf)
190{
191 struct memory_block *mem = to_memory_block(dev);
192 const char *output;
193
194 /*
195 * We can probably put these states in a nice little array
196 * so that they're not open-coded
197 */
198 switch (mem->state) {
199 case MEM_ONLINE:
200 output = "online";
201 break;
202 case MEM_OFFLINE:
203 output = "offline";
204 break;
205 case MEM_GOING_OFFLINE:
206 output = "going-offline";
207 break;
208 default:
209 WARN_ON(1);
210 return sysfs_emit(buf, "ERROR-UNKNOWN-%d\n", mem->state);
211 }
212
213 return sysfs_emit(buf, "%s\n", output);
214}
215
216int memory_notify(enum memory_block_state state, void *v)
217{
218 return blocking_notifier_call_chain(&memory_chain, state, v);
219}
220
221#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
222static unsigned long memblk_nr_poison(struct memory_block *mem);
223#else
224static inline unsigned long memblk_nr_poison(struct memory_block *mem)
225{
226 return 0;
227}
228#endif
229
230/*
231 * Must acquire mem_hotplug_lock in write mode.
232 */
233static int memory_block_online(struct memory_block *mem)
234{
235 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
236 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
237 unsigned long nr_vmemmap_pages = 0;
238 struct zone *zone;
239 int ret;
240
241 if (memblk_nr_poison(mem))
242 return -EHWPOISON;
243
244 zone = zone_for_pfn_range(mem->online_type, mem->nid, mem->group,
245 start_pfn, nr_pages);
246
247 /*
248 * Although vmemmap pages have a different lifecycle than the pages
249 * they describe (they remain until the memory is unplugged), doing
250 * their initialization and accounting at memory onlining/offlining
251 * stage helps to keep accounting easier to follow - e.g vmemmaps
252 * belong to the same zone as the memory they backed.
253 */
254 if (mem->altmap)
255 nr_vmemmap_pages = mem->altmap->free;
256
257 mem_hotplug_begin();
258 if (nr_vmemmap_pages) {
259 ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
260 if (ret)
261 goto out;
262 }
263
264 ret = online_pages(start_pfn + nr_vmemmap_pages,
265 nr_pages - nr_vmemmap_pages, zone, mem->group);
266 if (ret) {
267 if (nr_vmemmap_pages)
268 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
269 goto out;
270 }
271
272 /*
273 * Account once onlining succeeded. If the zone was unpopulated, it is
274 * now already properly populated.
275 */
276 if (nr_vmemmap_pages)
277 adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
278 nr_vmemmap_pages);
279
280 mem->zone = zone;
281out:
282 mem_hotplug_done();
283 return ret;
284}
285
286/*
287 * Must acquire mem_hotplug_lock in write mode.
288 */
289static int memory_block_offline(struct memory_block *mem)
290{
291 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
292 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
293 unsigned long nr_vmemmap_pages = 0;
294 int ret;
295
296 if (!mem->zone)
297 return -EINVAL;
298
299 /*
300 * Unaccount before offlining, such that unpopulated zone and kthreads
301 * can properly be torn down in offline_pages().
302 */
303 if (mem->altmap)
304 nr_vmemmap_pages = mem->altmap->free;
305
306 mem_hotplug_begin();
307 if (nr_vmemmap_pages)
308 adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
309 -nr_vmemmap_pages);
310
311 ret = offline_pages(start_pfn + nr_vmemmap_pages,
312 nr_pages - nr_vmemmap_pages, mem->zone, mem->group);
313 if (ret) {
314 /* offline_pages() failed. Account back. */
315 if (nr_vmemmap_pages)
316 adjust_present_page_count(pfn_to_page(start_pfn),
317 mem->group, nr_vmemmap_pages);
318 goto out;
319 }
320
321 if (nr_vmemmap_pages)
322 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
323
324 mem->zone = NULL;
325out:
326 mem_hotplug_done();
327 return ret;
328}
329
330/*
331 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
332 * OK to have direct references to sparsemem variables in here.
333 */
334static int
335memory_block_action(struct memory_block *mem, unsigned long action)
336{
337 int ret;
338
339 switch (action) {
340 case MEM_ONLINE:
341 ret = memory_block_online(mem);
342 break;
343 case MEM_OFFLINE:
344 ret = memory_block_offline(mem);
345 break;
346 default:
347 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
348 "%ld\n", __func__, mem->start_section_nr, action, action);
349 ret = -EINVAL;
350 }
351
352 return ret;
353}
354
355static int memory_block_change_state(struct memory_block *mem,
356 unsigned long to_state, unsigned long from_state_req)
357{
358 int ret = 0;
359
360 if (mem->state != from_state_req)
361 return -EINVAL;
362
363 if (to_state == MEM_OFFLINE)
364 mem->state = MEM_GOING_OFFLINE;
365
366 ret = memory_block_action(mem, to_state);
367 mem->state = ret ? from_state_req : to_state;
368
369 return ret;
370}
371
372/* The device lock serializes operations on memory_subsys_[online|offline] */
373static int memory_subsys_online(struct device *dev)
374{
375 struct memory_block *mem = to_memory_block(dev);
376 int ret;
377
378 if (mem->state == MEM_ONLINE)
379 return 0;
380
381 /*
382 * When called via device_online() without configuring the online_type,
383 * we want to default to MMOP_ONLINE.
384 */
385 if (mem->online_type == MMOP_OFFLINE)
386 mem->online_type = MMOP_ONLINE;
387
388 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
389 mem->online_type = MMOP_OFFLINE;
390
391 return ret;
392}
393
394static int memory_subsys_offline(struct device *dev)
395{
396 struct memory_block *mem = to_memory_block(dev);
397
398 if (mem->state == MEM_OFFLINE)
399 return 0;
400
401 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
402}
403
404static ssize_t state_store(struct device *dev, struct device_attribute *attr,
405 const char *buf, size_t count)
406{
407 const int online_type = mhp_online_type_from_str(buf);
408 struct memory_block *mem = to_memory_block(dev);
409 int ret;
410
411 if (online_type < 0)
412 return -EINVAL;
413
414 ret = lock_device_hotplug_sysfs();
415 if (ret)
416 return ret;
417
418 switch (online_type) {
419 case MMOP_ONLINE_KERNEL:
420 case MMOP_ONLINE_MOVABLE:
421 case MMOP_ONLINE:
422 /* mem->online_type is protected by device_hotplug_lock */
423 mem->online_type = online_type;
424 ret = device_online(&mem->dev);
425 break;
426 case MMOP_OFFLINE:
427 ret = device_offline(&mem->dev);
428 break;
429 default:
430 ret = -EINVAL; /* should never happen */
431 }
432
433 unlock_device_hotplug();
434
435 if (ret < 0)
436 return ret;
437 if (ret)
438 return -EINVAL;
439
440 return count;
441}
442
443/*
444 * Legacy interface that we cannot remove: s390x exposes the storage increment
445 * covered by a memory block, allowing for identifying which memory blocks
446 * comprise a storage increment. Since a memory block spans complete
447 * storage increments nowadays, this interface is basically unused. Other
448 * archs never exposed != 0.
449 */
450static ssize_t phys_device_show(struct device *dev,
451 struct device_attribute *attr, char *buf)
452{
453 struct memory_block *mem = to_memory_block(dev);
454 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
455
456 return sysfs_emit(buf, "%d\n",
457 arch_get_memory_phys_device(start_pfn));
458}
459
460#ifdef CONFIG_MEMORY_HOTREMOVE
461static int print_allowed_zone(char *buf, int len, int nid,
462 struct memory_group *group,
463 unsigned long start_pfn, unsigned long nr_pages,
464 enum mmop online_type, struct zone *default_zone)
465{
466 struct zone *zone;
467
468 zone = zone_for_pfn_range(online_type, nid, group, start_pfn, nr_pages);
469 if (zone == default_zone)
470 return 0;
471
472 return sysfs_emit_at(buf, len, " %s", zone->name);
473}
474
475static ssize_t valid_zones_show(struct device *dev,
476 struct device_attribute *attr, char *buf)
477{
478 struct memory_block *mem = to_memory_block(dev);
479 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
480 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
481 struct memory_group *group = mem->group;
482 struct zone *default_zone;
483 int nid = mem->nid;
484 int len;
485
486 /*
487 * Check the existing zone. Make sure that we do that only on the
488 * online nodes otherwise the page_zone is not reliable
489 */
490 if (mem->state == MEM_ONLINE) {
491 /*
492 * If !mem->zone, the memory block spans multiple zones and
493 * cannot get offlined.
494 */
495 return sysfs_emit(buf, "%s\n",
496 mem->zone ? mem->zone->name : "none");
497 }
498
499 default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, group,
500 start_pfn, nr_pages);
501
502 len = sysfs_emit(buf, "%s", default_zone->name);
503 len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
504 MMOP_ONLINE_KERNEL, default_zone);
505 len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
506 MMOP_ONLINE_MOVABLE, default_zone);
507 len += sysfs_emit_at(buf, len, "\n");
508 return len;
509}
510static DEVICE_ATTR_RO(valid_zones);
511#endif
512
513static DEVICE_ATTR_RO(phys_index);
514static DEVICE_ATTR_RW(state);
515static DEVICE_ATTR_RO(phys_device);
516static DEVICE_ATTR_RO(removable);
517
518/*
519 * Show the memory block size (shared by all memory blocks).
520 */
521static ssize_t block_size_bytes_show(struct device *dev,
522 struct device_attribute *attr, char *buf)
523{
524 return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
525}
526
527static DEVICE_ATTR_RO(block_size_bytes);
528
529/*
530 * Memory auto online policy.
531 */
532
533static ssize_t auto_online_blocks_show(struct device *dev,
534 struct device_attribute *attr, char *buf)
535{
536 return sysfs_emit(buf, "%s\n",
537 online_type_to_str[mhp_get_default_online_type()]);
538}
539
540static ssize_t auto_online_blocks_store(struct device *dev,
541 struct device_attribute *attr,
542 const char *buf, size_t count)
543{
544 const int online_type = mhp_online_type_from_str(buf);
545
546 if (online_type < 0)
547 return -EINVAL;
548
549 mhp_set_default_online_type(online_type);
550 return count;
551}
552
553static DEVICE_ATTR_RW(auto_online_blocks);
554
555#ifdef CONFIG_CRASH_HOTPLUG
556#include <linux/kexec.h>
557static ssize_t crash_hotplug_show(struct device *dev,
558 struct device_attribute *attr, char *buf)
559{
560 return sysfs_emit(buf, "%d\n", crash_check_hotplug_support());
561}
562static DEVICE_ATTR_RO(crash_hotplug);
563#endif
564
565/*
566 * Some architectures will have custom drivers to do this, and
567 * will not need to do it from userspace. The fake hot-add code
568 * as well as ppc64 will do all of their discovery in userspace
569 * and will require this interface.
570 */
571#ifdef CONFIG_ARCH_MEMORY_PROBE
572static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
573 const char *buf, size_t count)
574{
575 u64 phys_addr;
576 int nid, ret;
577 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
578
579 ret = kstrtoull(buf, 0, &phys_addr);
580 if (ret)
581 return ret;
582
583 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
584 return -EINVAL;
585
586 ret = lock_device_hotplug_sysfs();
587 if (ret)
588 return ret;
589
590 nid = memory_add_physaddr_to_nid(phys_addr);
591 ret = __add_memory(nid, phys_addr,
592 MIN_MEMORY_BLOCK_SIZE * sections_per_block,
593 MHP_NONE);
594
595 if (ret)
596 goto out;
597
598 ret = count;
599out:
600 unlock_device_hotplug();
601 return ret;
602}
603
604static DEVICE_ATTR_WO(probe);
605#endif
606
607#ifdef CONFIG_MEMORY_FAILURE
608/*
609 * Support for offlining pages of memory
610 */
611
612/* Soft offline a page */
613static ssize_t soft_offline_page_store(struct device *dev,
614 struct device_attribute *attr,
615 const char *buf, size_t count)
616{
617 int ret;
618 u64 pfn;
619 if (!capable(CAP_SYS_ADMIN))
620 return -EPERM;
621 if (kstrtoull(buf, 0, &pfn) < 0)
622 return -EINVAL;
623 pfn >>= PAGE_SHIFT;
624 ret = soft_offline_page(pfn, 0);
625 return ret == 0 ? count : ret;
626}
627
628/* Forcibly offline a page, including killing processes. */
629static ssize_t hard_offline_page_store(struct device *dev,
630 struct device_attribute *attr,
631 const char *buf, size_t count)
632{
633 int ret;
634 u64 pfn;
635 if (!capable(CAP_SYS_ADMIN))
636 return -EPERM;
637 if (kstrtoull(buf, 0, &pfn) < 0)
638 return -EINVAL;
639 pfn >>= PAGE_SHIFT;
640 ret = memory_failure(pfn, MF_SW_SIMULATED);
641 if (ret == -EOPNOTSUPP)
642 ret = 0;
643 return ret ? ret : count;
644}
645
646static DEVICE_ATTR_WO(soft_offline_page);
647static DEVICE_ATTR_WO(hard_offline_page);
648#endif
649
650/* See phys_device_show(). */
651int __weak arch_get_memory_phys_device(unsigned long start_pfn)
652{
653 return 0;
654}
655
656/*
657 * A reference for the returned memory block device is acquired.
658 *
659 * Called under device_hotplug_lock.
660 */
661struct memory_block *memory_block_get(unsigned long block_id)
662{
663 struct memory_block *mem;
664
665 mem = xa_load(&memory_blocks, block_id);
666 if (mem)
667 get_device(&mem->dev);
668 return mem;
669}
670
671static struct attribute *memory_memblk_attrs[] = {
672 &dev_attr_phys_index.attr,
673 &dev_attr_state.attr,
674 &dev_attr_phys_device.attr,
675 &dev_attr_removable.attr,
676#ifdef CONFIG_MEMORY_HOTREMOVE
677 &dev_attr_valid_zones.attr,
678#endif
679 NULL
680};
681
682static const struct attribute_group memory_memblk_attr_group = {
683 .attrs = memory_memblk_attrs,
684};
685
686static const struct attribute_group *memory_memblk_attr_groups[] = {
687 &memory_memblk_attr_group,
688 NULL,
689};
690
691static int __add_memory_block(struct memory_block *memory)
692{
693 int ret;
694
695 memory->dev.bus = &memory_subsys;
696 memory->dev.id = memory->start_section_nr / sections_per_block;
697 memory->dev.release = memory_block_release;
698 memory->dev.groups = memory_memblk_attr_groups;
699 dev_assign_offline(&memory->dev, memory->state == MEM_OFFLINE);
700
701 ret = device_register(&memory->dev);
702 if (ret) {
703 memory_block_put(memory);
704 return ret;
705 }
706 ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
707 GFP_KERNEL));
708 if (ret)
709 device_unregister(&memory->dev);
710
711 return ret;
712}
713
714static struct zone *early_node_zone_for_memory_block(struct memory_block *mem,
715 int nid)
716{
717 const unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
718 const unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
719 struct zone *zone, *matching_zone = NULL;
720 pg_data_t *pgdat = NODE_DATA(nid);
721 int i;
722
723 /*
724 * This logic only works for early memory, when the applicable zones
725 * already span the memory block. We don't expect overlapping zones on
726 * a single node for early memory. So if we're told that some PFNs
727 * of a node fall into this memory block, we can assume that all node
728 * zones that intersect with the memory block are actually applicable.
729 * No need to look at the memmap.
730 */
731 for (i = 0; i < MAX_NR_ZONES; i++) {
732 zone = pgdat->node_zones + i;
733 if (!populated_zone(zone))
734 continue;
735 if (!zone_intersects(zone, start_pfn, nr_pages))
736 continue;
737 if (!matching_zone) {
738 matching_zone = zone;
739 continue;
740 }
741 /* Spans multiple zones ... */
742 matching_zone = NULL;
743 break;
744 }
745 return matching_zone;
746}
747
748#ifdef CONFIG_NUMA
749/**
750 * memory_block_add_nid_early() - Indicate that early system RAM falling into
751 * this memory block device (partially) belongs
752 * to the given node.
753 * @mem: The memory block device.
754 * @nid: The node id.
755 *
756 * Indicate that early system RAM falling into this memory block (partially)
757 * belongs to the given node. This will also properly set/adjust mem->zone based
758 * on the zone ranges of the given node.
759 *
760 * Memory hotplug handles this on memory block creation, where we can only have
761 * a single nid span a memory block.
762 */
763void memory_block_add_nid_early(struct memory_block *mem, int nid)
764{
765 if (mem->nid != nid) {
766 /*
767 * For early memory we have to determine the zone when setting
768 * the node id and handle multiple nodes spanning a single
769 * memory block by indicate via zone == NULL that we're not
770 * dealing with a single zone. So if we're setting the node id
771 * the first time, determine if there is a single zone. If we're
772 * setting the node id a second time to a different node,
773 * invalidate the single detected zone.
774 */
775 if (mem->nid == NUMA_NO_NODE)
776 mem->zone = early_node_zone_for_memory_block(mem, nid);
777 else
778 mem->zone = NULL;
779 /*
780 * If this memory block spans multiple nodes, we only indicate
781 * the last processed node. If we span multiple nodes (not applicable
782 * to hotplugged memory), zone == NULL will prohibit memory offlining
783 * and consequently unplug.
784 */
785 mem->nid = nid;
786 }
787}
788#endif
789
790static int add_memory_block(unsigned long block_id, int nid, unsigned long state,
791 struct vmem_altmap *altmap,
792 struct memory_group *group)
793{
794 struct memory_block *mem;
795 int ret = 0;
796
797 mem = memory_block_get(block_id);
798 if (mem) {
799 memory_block_put(mem);
800 return -EEXIST;
801 }
802 mem = kzalloc_obj(*mem);
803 if (!mem)
804 return -ENOMEM;
805
806 mem->start_section_nr = block_id * sections_per_block;
807 mem->state = state;
808 mem->nid = nid;
809 INIT_LIST_HEAD(&mem->group_next);
810
811#ifndef CONFIG_NUMA
812 if (state == MEM_ONLINE)
813 /*
814 * MEM_ONLINE at this point implies early memory. With NUMA,
815 * we'll determine the zone when setting the node id via
816 * memory_block_add_nid_early(). Memory hotplug updated the zone
817 * manually when memory onlining/offlining succeeds.
818 */
819 mem->zone = early_node_zone_for_memory_block(mem, NUMA_NO_NODE);
820#endif /* CONFIG_NUMA */
821
822 ret = __add_memory_block(mem);
823 if (ret)
824 return ret;
825
826 mem->altmap = altmap;
827
828 if (group) {
829 mem->group = group;
830 list_add(&mem->group_next, &group->memory_blocks);
831 }
832
833 return 0;
834}
835
836static void remove_memory_block(struct memory_block *memory)
837{
838 if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
839 return;
840
841 WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
842
843 if (memory->group) {
844 list_del(&memory->group_next);
845 memory->group = NULL;
846 }
847
848 /* drop the ref. we got via memory_block_get() */
849 memory_block_put(memory);
850 device_unregister(&memory->dev);
851}
852
853/*
854 * Create memory block devices for the given memory area. Start and size
855 * have to be aligned to memory block granularity. Memory block devices
856 * will be initialized as offline.
857 *
858 * Called under device_hotplug_lock.
859 */
860int create_memory_block_devices(unsigned long start, unsigned long size,
861 int nid, struct vmem_altmap *altmap,
862 struct memory_group *group)
863{
864 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
865 unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
866 struct memory_block *mem;
867 unsigned long block_id;
868 int ret = 0;
869
870 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
871 !IS_ALIGNED(size, memory_block_size_bytes())))
872 return -EINVAL;
873
874 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
875 ret = add_memory_block(block_id, nid, MEM_OFFLINE, altmap, group);
876 if (ret)
877 break;
878 }
879 if (ret) {
880 end_block_id = block_id;
881 for (block_id = start_block_id; block_id != end_block_id;
882 block_id++) {
883 mem = memory_block_get(block_id);
884 if (WARN_ON_ONCE(!mem))
885 continue;
886 remove_memory_block(mem);
887 }
888 }
889 return ret;
890}
891
892/*
893 * Remove memory block devices for the given memory area. Start and size
894 * have to be aligned to memory block granularity. Memory block devices
895 * have to be offline.
896 *
897 * Called under device_hotplug_lock.
898 */
899void remove_memory_block_devices(unsigned long start, unsigned long size)
900{
901 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
902 const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
903 struct memory_block *mem;
904 unsigned long block_id;
905
906 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
907 !IS_ALIGNED(size, memory_block_size_bytes())))
908 return;
909
910 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
911 mem = memory_block_get(block_id);
912 if (WARN_ON_ONCE(!mem))
913 continue;
914 num_poisoned_pages_sub(-1UL, memblk_nr_poison(mem));
915 unregister_memory_block_under_nodes(mem);
916 remove_memory_block(mem);
917 }
918}
919
920static struct attribute *memory_root_attrs[] = {
921#ifdef CONFIG_ARCH_MEMORY_PROBE
922 &dev_attr_probe.attr,
923#endif
924
925#ifdef CONFIG_MEMORY_FAILURE
926 &dev_attr_soft_offline_page.attr,
927 &dev_attr_hard_offline_page.attr,
928#endif
929
930 &dev_attr_block_size_bytes.attr,
931 &dev_attr_auto_online_blocks.attr,
932#ifdef CONFIG_CRASH_HOTPLUG
933 &dev_attr_crash_hotplug.attr,
934#endif
935 NULL
936};
937
938static const struct attribute_group memory_root_attr_group = {
939 .attrs = memory_root_attrs,
940};
941
942static const struct attribute_group *memory_root_attr_groups[] = {
943 &memory_root_attr_group,
944 NULL,
945};
946
947/*
948 * Initialize the sysfs support for memory devices. At the time this function
949 * is called, we cannot have concurrent creation/deletion of memory block
950 * devices, the device_hotplug_lock is not needed.
951 */
952void __init memory_dev_init(void)
953{
954 int ret;
955 unsigned long block_sz, block_id, nr;
956
957 /* Validate the configured memory block size */
958 block_sz = memory_block_size_bytes();
959 if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
960 panic("Memory block size not suitable: 0x%lx\n", block_sz);
961 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
962
963 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
964 if (ret)
965 panic("%s() failed to register subsystem: %d\n", __func__, ret);
966
967 /*
968 * Create entries for memory sections that were found during boot
969 * and have been initialized. Use @block_id to track the last
970 * handled block and initialize it to an invalid value (ULONG_MAX)
971 * to bypass the block ID matching check for the first present
972 * block so that it can be covered.
973 */
974 block_id = ULONG_MAX;
975 for_each_present_section_nr(0, nr) {
976 if (block_id != ULONG_MAX && memory_block_id(nr) == block_id)
977 continue;
978
979 block_id = memory_block_id(nr);
980 ret = add_memory_block(block_id, NUMA_NO_NODE, MEM_ONLINE, NULL, NULL);
981 if (ret) {
982 panic("%s() failed to add memory block: %d\n",
983 __func__, ret);
984 }
985 }
986}
987
988/**
989 * walk_memory_blocks - walk through all present memory blocks overlapped
990 * by the range [start, start + size)
991 *
992 * @start: start address of the memory range
993 * @size: size of the memory range
994 * @arg: argument passed to func
995 * @func: callback for each memory section walked
996 *
997 * This function walks through all present memory blocks overlapped by the
998 * range [start, start + size), calling func on each memory block.
999 *
1000 * In case func() returns an error, walking is aborted and the error is
1001 * returned.
1002 *
1003 * Called under device_hotplug_lock.
1004 */
1005int walk_memory_blocks(unsigned long start, unsigned long size,
1006 void *arg, walk_memory_blocks_func_t func)
1007{
1008 const unsigned long start_block_id = phys_to_block_id(start);
1009 const unsigned long end_block_id = phys_to_block_id(start + size - 1);
1010 struct memory_block *mem;
1011 unsigned long block_id;
1012 int ret = 0;
1013
1014 if (!size)
1015 return 0;
1016
1017 for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
1018 mem = memory_block_get(block_id);
1019 if (!mem)
1020 continue;
1021
1022 ret = func(mem, arg);
1023 memory_block_put(mem);
1024 if (ret)
1025 break;
1026 }
1027 return ret;
1028}
1029
1030struct for_each_memory_block_cb_data {
1031 walk_memory_blocks_func_t func;
1032 void *arg;
1033};
1034
1035static int for_each_memory_block_cb(struct device *dev, void *data)
1036{
1037 struct memory_block *mem = to_memory_block(dev);
1038 struct for_each_memory_block_cb_data *cb_data = data;
1039
1040 return cb_data->func(mem, cb_data->arg);
1041}
1042
1043/**
1044 * for_each_memory_block - walk through all present memory blocks
1045 *
1046 * @arg: argument passed to func
1047 * @func: callback for each memory block walked
1048 *
1049 * This function walks through all present memory blocks, calling func on
1050 * each memory block.
1051 *
1052 * In case func() returns an error, walking is aborted and the error is
1053 * returned.
1054 */
1055int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
1056{
1057 struct for_each_memory_block_cb_data cb_data = {
1058 .func = func,
1059 .arg = arg,
1060 };
1061
1062 return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
1063 for_each_memory_block_cb);
1064}
1065
1066/*
1067 * This is an internal helper to unify allocation and initialization of
1068 * memory groups. Note that the passed memory group will be copied to a
1069 * dynamically allocated memory group. After this call, the passed
1070 * memory group should no longer be used.
1071 */
1072static int memory_group_register(struct memory_group group)
1073{
1074 struct memory_group *new_group;
1075 uint32_t mgid;
1076 int ret;
1077
1078 if (!node_possible(group.nid))
1079 return -EINVAL;
1080
1081 new_group = kzalloc_obj(group);
1082 if (!new_group)
1083 return -ENOMEM;
1084 *new_group = group;
1085 INIT_LIST_HEAD(&new_group->memory_blocks);
1086
1087 ret = xa_alloc(&memory_groups, &mgid, new_group, xa_limit_31b,
1088 GFP_KERNEL);
1089 if (ret) {
1090 kfree(new_group);
1091 return ret;
1092 } else if (group.is_dynamic) {
1093 xa_set_mark(&memory_groups, mgid, MEMORY_GROUP_MARK_DYNAMIC);
1094 }
1095 return mgid;
1096}
1097
1098/**
1099 * memory_group_register_static() - Register a static memory group.
1100 * @nid: The node id.
1101 * @max_pages: The maximum number of pages we'll have in this static memory
1102 * group.
1103 *
1104 * Register a new static memory group and return the memory group id.
1105 * All memory in the group belongs to a single unit, such as a DIMM. All
1106 * memory belonging to a static memory group is added in one go to be removed
1107 * in one go -- it's static.
1108 *
1109 * Returns an error if out of memory, if the node id is invalid, if no new
1110 * memory groups can be registered, or if max_pages is invalid (0). Otherwise,
1111 * returns the new memory group id.
1112 */
1113int memory_group_register_static(int nid, unsigned long max_pages)
1114{
1115 struct memory_group group = {
1116 .nid = nid,
1117 .s = {
1118 .max_pages = max_pages,
1119 },
1120 };
1121
1122 if (!max_pages)
1123 return -EINVAL;
1124 return memory_group_register(group);
1125}
1126EXPORT_SYMBOL_GPL(memory_group_register_static);
1127
1128/**
1129 * memory_group_register_dynamic() - Register a dynamic memory group.
1130 * @nid: The node id.
1131 * @unit_pages: Unit in pages in which is memory added/removed in this dynamic
1132 * memory group.
1133 *
1134 * Register a new dynamic memory group and return the memory group id.
1135 * Memory within a dynamic memory group is added/removed dynamically
1136 * in unit_pages.
1137 *
1138 * Returns an error if out of memory, if the node id is invalid, if no new
1139 * memory groups can be registered, or if unit_pages is invalid (0, not a
1140 * power of two, smaller than a single memory block). Otherwise, returns the
1141 * new memory group id.
1142 */
1143int memory_group_register_dynamic(int nid, unsigned long unit_pages)
1144{
1145 struct memory_group group = {
1146 .nid = nid,
1147 .is_dynamic = true,
1148 .d = {
1149 .unit_pages = unit_pages,
1150 },
1151 };
1152
1153 if (!unit_pages || !is_power_of_2(unit_pages) ||
1154 unit_pages < PHYS_PFN(memory_block_size_bytes()))
1155 return -EINVAL;
1156 return memory_group_register(group);
1157}
1158EXPORT_SYMBOL_GPL(memory_group_register_dynamic);
1159
1160/**
1161 * memory_group_unregister() - Unregister a memory group.
1162 * @mgid: the memory group id
1163 *
1164 * Unregister a memory group. If any memory block still belongs to this
1165 * memory group, unregistering will fail.
1166 *
1167 * Returns -EINVAL if the memory group id is invalid, returns -EBUSY if some
1168 * memory blocks still belong to this memory group and returns 0 if
1169 * unregistering succeeded.
1170 */
1171int memory_group_unregister(int mgid)
1172{
1173 struct memory_group *group;
1174
1175 if (mgid < 0)
1176 return -EINVAL;
1177
1178 group = xa_load(&memory_groups, mgid);
1179 if (!group)
1180 return -EINVAL;
1181 if (!list_empty(&group->memory_blocks))
1182 return -EBUSY;
1183 xa_erase(&memory_groups, mgid);
1184 kfree(group);
1185 return 0;
1186}
1187EXPORT_SYMBOL_GPL(memory_group_unregister);
1188
1189/*
1190 * This is an internal helper only to be used in core memory hotplug code to
1191 * lookup a memory group. We don't care about locking, as we don't expect a
1192 * memory group to get unregistered while adding memory to it -- because
1193 * the group and the memory is managed by the same driver.
1194 */
1195struct memory_group *memory_group_find_by_id(int mgid)
1196{
1197 return xa_load(&memory_groups, mgid);
1198}
1199
1200/*
1201 * This is an internal helper only to be used in core memory hotplug code to
1202 * walk all dynamic memory groups excluding a given memory group, either
1203 * belonging to a specific node, or belonging to any node.
1204 */
1205int walk_dynamic_memory_groups(int nid, walk_memory_groups_func_t func,
1206 struct memory_group *excluded, void *arg)
1207{
1208 struct memory_group *group;
1209 unsigned long index;
1210 int ret = 0;
1211
1212 xa_for_each_marked(&memory_groups, index, group,
1213 MEMORY_GROUP_MARK_DYNAMIC) {
1214 if (group == excluded)
1215 continue;
1216#ifdef CONFIG_NUMA
1217 if (nid != NUMA_NO_NODE && group->nid != nid)
1218 continue;
1219#endif /* CONFIG_NUMA */
1220 ret = func(group, arg);
1221 if (ret)
1222 break;
1223 }
1224 return ret;
1225}
1226
1227#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
1228void memblk_nr_poison_inc(unsigned long pfn)
1229{
1230 const unsigned long block_id = pfn_to_block_id(pfn);
1231 struct memory_block *mem = memory_block_get(block_id);
1232
1233 if (mem) {
1234 atomic_long_inc(&mem->nr_hwpoison);
1235 memory_block_put(mem);
1236 }
1237}
1238
1239void memblk_nr_poison_sub(unsigned long pfn, long i)
1240{
1241 const unsigned long block_id = pfn_to_block_id(pfn);
1242 struct memory_block *mem = memory_block_get(block_id);
1243
1244 if (mem) {
1245 atomic_long_sub(i, &mem->nr_hwpoison);
1246 memory_block_put(mem);
1247 }
1248}
1249
1250static unsigned long memblk_nr_poison(struct memory_block *mem)
1251{
1252 return atomic_long_read(&mem->nr_hwpoison);
1253}
1254#endif