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