1 // SPDX-License-Identifier: GPL-2.0
3 * Memory subsystem support
5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
6 * Dave Hansen <haveblue@us.ibm.com>
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.
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>
22 #include <linux/stat.h>
23 #include <linux/slab.h>
25 #include <linux/atomic.h>
26 #include <linux/uaccess.h>
28 #define MEMORY_CLASS_NAME "memory"
30 #define to_memory_block(dev) container_of(dev, struct memory_block, dev)
32 static int sections_per_block
;
34 static inline unsigned long base_memory_block_id(unsigned long section_nr
)
36 return section_nr
/ sections_per_block
;
39 static inline unsigned long pfn_to_block_id(unsigned long pfn
)
41 return base_memory_block_id(pfn_to_section_nr(pfn
));
44 static inline unsigned long phys_to_block_id(unsigned long phys
)
46 return pfn_to_block_id(PFN_DOWN(phys
));
49 static int memory_subsys_online(struct device
*dev
);
50 static int memory_subsys_offline(struct device
*dev
);
52 static struct bus_type memory_subsys
= {
53 .name
= MEMORY_CLASS_NAME
,
54 .dev_name
= MEMORY_CLASS_NAME
,
55 .online
= memory_subsys_online
,
56 .offline
= memory_subsys_offline
,
59 static BLOCKING_NOTIFIER_HEAD(memory_chain
);
61 int register_memory_notifier(struct notifier_block
*nb
)
63 return blocking_notifier_chain_register(&memory_chain
, nb
);
65 EXPORT_SYMBOL(register_memory_notifier
);
67 void unregister_memory_notifier(struct notifier_block
*nb
)
69 blocking_notifier_chain_unregister(&memory_chain
, nb
);
71 EXPORT_SYMBOL(unregister_memory_notifier
);
73 static void memory_block_release(struct device
*dev
)
75 struct memory_block
*mem
= to_memory_block(dev
);
80 unsigned long __weak
memory_block_size_bytes(void)
82 return MIN_MEMORY_BLOCK_SIZE
;
84 EXPORT_SYMBOL_GPL(memory_block_size_bytes
);
87 * Show the first physical section index (number) of this memory block.
89 static ssize_t
phys_index_show(struct device
*dev
,
90 struct device_attribute
*attr
, char *buf
)
92 struct memory_block
*mem
= to_memory_block(dev
);
93 unsigned long phys_index
;
95 phys_index
= mem
->start_section_nr
/ sections_per_block
;
96 return sprintf(buf
, "%08lx\n", phys_index
);
100 * Show whether the memory block is likely to be offlineable (or is already
101 * offline). Once offline, the memory block could be removed. The return
102 * value does, however, not indicate that there is a way to remove the
105 static ssize_t
removable_show(struct device
*dev
, struct device_attribute
*attr
,
108 struct memory_block
*mem
= to_memory_block(dev
);
112 if (mem
->state
!= MEM_ONLINE
)
115 for (i
= 0; i
< sections_per_block
; i
++) {
116 if (!present_section_nr(mem
->start_section_nr
+ i
))
118 pfn
= section_nr_to_pfn(mem
->start_section_nr
+ i
);
119 ret
&= is_mem_section_removable(pfn
, PAGES_PER_SECTION
);
123 return sprintf(buf
, "%d\n", ret
);
127 * online, offline, going offline, etc.
129 static ssize_t
state_show(struct device
*dev
, struct device_attribute
*attr
,
132 struct memory_block
*mem
= to_memory_block(dev
);
136 * We can probably put these states in a nice little array
137 * so that they're not open-coded
139 switch (mem
->state
) {
141 len
= sprintf(buf
, "online\n");
144 len
= sprintf(buf
, "offline\n");
146 case MEM_GOING_OFFLINE
:
147 len
= sprintf(buf
, "going-offline\n");
150 len
= sprintf(buf
, "ERROR-UNKNOWN-%ld\n",
159 int memory_notify(unsigned long val
, void *v
)
161 return blocking_notifier_call_chain(&memory_chain
, val
, v
);
165 * The probe routines leave the pages uninitialized, just as the bootmem code
166 * does. Make sure we do not access them, but instead use only information from
169 static bool pages_correctly_probed(unsigned long start_pfn
)
171 unsigned long section_nr
= pfn_to_section_nr(start_pfn
);
172 unsigned long section_nr_end
= section_nr
+ sections_per_block
;
173 unsigned long pfn
= start_pfn
;
176 * memmap between sections is not contiguous except with
177 * SPARSEMEM_VMEMMAP. We lookup the page once per section
178 * and assume memmap is contiguous within each section
180 for (; section_nr
< section_nr_end
; section_nr
++) {
181 if (WARN_ON_ONCE(!pfn_valid(pfn
)))
184 if (!present_section_nr(section_nr
)) {
185 pr_warn("section %ld pfn[%lx, %lx) not present\n",
186 section_nr
, pfn
, pfn
+ PAGES_PER_SECTION
);
188 } else if (!valid_section_nr(section_nr
)) {
189 pr_warn("section %ld pfn[%lx, %lx) no valid memmap\n",
190 section_nr
, pfn
, pfn
+ PAGES_PER_SECTION
);
192 } else if (online_section_nr(section_nr
)) {
193 pr_warn("section %ld pfn[%lx, %lx) is already online\n",
194 section_nr
, pfn
, pfn
+ PAGES_PER_SECTION
);
197 pfn
+= PAGES_PER_SECTION
;
204 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
205 * OK to have direct references to sparsemem variables in here.
208 memory_block_action(unsigned long start_section_nr
, unsigned long action
,
209 int online_type
, int nid
)
211 unsigned long start_pfn
;
212 unsigned long nr_pages
= PAGES_PER_SECTION
* sections_per_block
;
215 start_pfn
= section_nr_to_pfn(start_section_nr
);
219 if (!pages_correctly_probed(start_pfn
))
222 ret
= online_pages(start_pfn
, nr_pages
, online_type
, nid
);
225 ret
= offline_pages(start_pfn
, nr_pages
);
228 WARN(1, KERN_WARNING
"%s(%ld, %ld) unknown action: "
229 "%ld\n", __func__
, start_section_nr
, action
, action
);
236 static int memory_block_change_state(struct memory_block
*mem
,
237 unsigned long to_state
, unsigned long from_state_req
)
241 if (mem
->state
!= from_state_req
)
244 if (to_state
== MEM_OFFLINE
)
245 mem
->state
= MEM_GOING_OFFLINE
;
247 ret
= memory_block_action(mem
->start_section_nr
, to_state
,
248 mem
->online_type
, mem
->nid
);
250 mem
->state
= ret
? from_state_req
: to_state
;
255 /* The device lock serializes operations on memory_subsys_[online|offline] */
256 static int memory_subsys_online(struct device
*dev
)
258 struct memory_block
*mem
= to_memory_block(dev
);
261 if (mem
->state
== MEM_ONLINE
)
265 * If we are called from state_store(), online_type will be
266 * set >= 0 Otherwise we were called from the device online
267 * attribute and need to set the online_type.
269 if (mem
->online_type
< 0)
270 mem
->online_type
= MMOP_ONLINE_KEEP
;
272 ret
= memory_block_change_state(mem
, MEM_ONLINE
, MEM_OFFLINE
);
274 /* clear online_type */
275 mem
->online_type
= -1;
280 static int memory_subsys_offline(struct device
*dev
)
282 struct memory_block
*mem
= to_memory_block(dev
);
284 if (mem
->state
== MEM_OFFLINE
)
287 /* Can't offline block with non-present sections */
288 if (mem
->section_count
!= sections_per_block
)
291 return memory_block_change_state(mem
, MEM_OFFLINE
, MEM_ONLINE
);
294 static ssize_t
state_store(struct device
*dev
, struct device_attribute
*attr
,
295 const char *buf
, size_t count
)
297 struct memory_block
*mem
= to_memory_block(dev
);
298 int ret
, online_type
;
300 ret
= lock_device_hotplug_sysfs();
304 if (sysfs_streq(buf
, "online_kernel"))
305 online_type
= MMOP_ONLINE_KERNEL
;
306 else if (sysfs_streq(buf
, "online_movable"))
307 online_type
= MMOP_ONLINE_MOVABLE
;
308 else if (sysfs_streq(buf
, "online"))
309 online_type
= MMOP_ONLINE_KEEP
;
310 else if (sysfs_streq(buf
, "offline"))
311 online_type
= MMOP_OFFLINE
;
317 switch (online_type
) {
318 case MMOP_ONLINE_KERNEL
:
319 case MMOP_ONLINE_MOVABLE
:
320 case MMOP_ONLINE_KEEP
:
321 /* mem->online_type is protected by device_hotplug_lock */
322 mem
->online_type
= online_type
;
323 ret
= device_online(&mem
->dev
);
326 ret
= device_offline(&mem
->dev
);
329 ret
= -EINVAL
; /* should never happen */
333 unlock_device_hotplug();
344 * phys_device is a bad name for this. What I really want
345 * is a way to differentiate between memory ranges that
346 * are part of physical devices that constitute
347 * a complete removable unit or fru.
348 * i.e. do these ranges belong to the same physical device,
349 * s.t. if I offline all of these sections I can then
350 * remove the physical device?
352 static ssize_t
phys_device_show(struct device
*dev
,
353 struct device_attribute
*attr
, char *buf
)
355 struct memory_block
*mem
= to_memory_block(dev
);
356 return sprintf(buf
, "%d\n", mem
->phys_device
);
359 #ifdef CONFIG_MEMORY_HOTREMOVE
360 static void print_allowed_zone(char *buf
, int nid
, unsigned long start_pfn
,
361 unsigned long nr_pages
, int online_type
,
362 struct zone
*default_zone
)
366 zone
= zone_for_pfn_range(online_type
, nid
, start_pfn
, nr_pages
);
367 if (zone
!= default_zone
) {
369 strcat(buf
, zone
->name
);
373 static ssize_t
valid_zones_show(struct device
*dev
,
374 struct device_attribute
*attr
, char *buf
)
376 struct memory_block
*mem
= to_memory_block(dev
);
377 unsigned long start_pfn
= section_nr_to_pfn(mem
->start_section_nr
);
378 unsigned long nr_pages
= PAGES_PER_SECTION
* sections_per_block
;
379 struct zone
*default_zone
;
383 * Check the existing zone. Make sure that we do that only on the
384 * online nodes otherwise the page_zone is not reliable
386 if (mem
->state
== MEM_ONLINE
) {
388 * The block contains more than one zone can not be offlined.
389 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
391 default_zone
= test_pages_in_a_zone(start_pfn
,
392 start_pfn
+ nr_pages
);
394 return sprintf(buf
, "none\n");
395 strcat(buf
, default_zone
->name
);
400 default_zone
= zone_for_pfn_range(MMOP_ONLINE_KEEP
, nid
, start_pfn
, nr_pages
);
401 strcat(buf
, default_zone
->name
);
403 print_allowed_zone(buf
, nid
, start_pfn
, nr_pages
, MMOP_ONLINE_KERNEL
,
405 print_allowed_zone(buf
, nid
, start_pfn
, nr_pages
, MMOP_ONLINE_MOVABLE
,
412 static DEVICE_ATTR_RO(valid_zones
);
415 static DEVICE_ATTR_RO(phys_index
);
416 static DEVICE_ATTR_RW(state
);
417 static DEVICE_ATTR_RO(phys_device
);
418 static DEVICE_ATTR_RO(removable
);
421 * Show the memory block size (shared by all memory blocks).
423 static ssize_t
block_size_bytes_show(struct device
*dev
,
424 struct device_attribute
*attr
, char *buf
)
426 return sprintf(buf
, "%lx\n", memory_block_size_bytes());
429 static DEVICE_ATTR_RO(block_size_bytes
);
432 * Memory auto online policy.
435 static ssize_t
auto_online_blocks_show(struct device
*dev
,
436 struct device_attribute
*attr
, char *buf
)
438 if (memhp_auto_online
)
439 return sprintf(buf
, "online\n");
441 return sprintf(buf
, "offline\n");
444 static ssize_t
auto_online_blocks_store(struct device
*dev
,
445 struct device_attribute
*attr
,
446 const char *buf
, size_t count
)
448 if (sysfs_streq(buf
, "online"))
449 memhp_auto_online
= true;
450 else if (sysfs_streq(buf
, "offline"))
451 memhp_auto_online
= false;
458 static DEVICE_ATTR_RW(auto_online_blocks
);
461 * Some architectures will have custom drivers to do this, and
462 * will not need to do it from userspace. The fake hot-add code
463 * as well as ppc64 will do all of their discovery in userspace
464 * and will require this interface.
466 #ifdef CONFIG_ARCH_MEMORY_PROBE
467 static ssize_t
probe_store(struct device
*dev
, struct device_attribute
*attr
,
468 const char *buf
, size_t count
)
472 unsigned long pages_per_block
= PAGES_PER_SECTION
* sections_per_block
;
474 ret
= kstrtoull(buf
, 0, &phys_addr
);
478 if (phys_addr
& ((pages_per_block
<< PAGE_SHIFT
) - 1))
481 ret
= lock_device_hotplug_sysfs();
485 nid
= memory_add_physaddr_to_nid(phys_addr
);
486 ret
= __add_memory(nid
, phys_addr
,
487 MIN_MEMORY_BLOCK_SIZE
* sections_per_block
);
494 unlock_device_hotplug();
498 static DEVICE_ATTR_WO(probe
);
501 #ifdef CONFIG_MEMORY_FAILURE
503 * Support for offlining pages of memory
506 /* Soft offline a page */
507 static ssize_t
soft_offline_page_store(struct device
*dev
,
508 struct device_attribute
*attr
,
509 const char *buf
, size_t count
)
513 if (!capable(CAP_SYS_ADMIN
))
515 if (kstrtoull(buf
, 0, &pfn
) < 0)
518 ret
= soft_offline_page(pfn
, 0);
519 return ret
== 0 ? count
: ret
;
522 /* Forcibly offline a page, including killing processes. */
523 static ssize_t
hard_offline_page_store(struct device
*dev
,
524 struct device_attribute
*attr
,
525 const char *buf
, size_t count
)
529 if (!capable(CAP_SYS_ADMIN
))
531 if (kstrtoull(buf
, 0, &pfn
) < 0)
534 ret
= memory_failure(pfn
, 0);
535 return ret
? ret
: count
;
538 static DEVICE_ATTR_WO(soft_offline_page
);
539 static DEVICE_ATTR_WO(hard_offline_page
);
543 * Note that phys_device is optional. It is here to allow for
544 * differentiation between which *physical* devices each
545 * section belongs to...
547 int __weak
arch_get_memory_phys_device(unsigned long start_pfn
)
552 /* A reference for the returned memory block device is acquired. */
553 static struct memory_block
*find_memory_block_by_id(unsigned long block_id
)
557 dev
= subsys_find_device_by_id(&memory_subsys
, block_id
, NULL
);
558 return dev
? to_memory_block(dev
) : NULL
;
562 * For now, we have a linear search to go find the appropriate
563 * memory_block corresponding to a particular phys_index. If
564 * this gets to be a real problem, we can always use a radix
565 * tree or something here.
567 * This could be made generic for all device subsystems.
569 struct memory_block
*find_memory_block(struct mem_section
*section
)
571 unsigned long block_id
= base_memory_block_id(__section_nr(section
));
573 return find_memory_block_by_id(block_id
);
576 static struct attribute
*memory_memblk_attrs
[] = {
577 &dev_attr_phys_index
.attr
,
578 &dev_attr_state
.attr
,
579 &dev_attr_phys_device
.attr
,
580 &dev_attr_removable
.attr
,
581 #ifdef CONFIG_MEMORY_HOTREMOVE
582 &dev_attr_valid_zones
.attr
,
587 static struct attribute_group memory_memblk_attr_group
= {
588 .attrs
= memory_memblk_attrs
,
591 static const struct attribute_group
*memory_memblk_attr_groups
[] = {
592 &memory_memblk_attr_group
,
597 * register_memory - Setup a sysfs device for a memory block
600 int register_memory(struct memory_block
*memory
)
604 memory
->dev
.bus
= &memory_subsys
;
605 memory
->dev
.id
= memory
->start_section_nr
/ sections_per_block
;
606 memory
->dev
.release
= memory_block_release
;
607 memory
->dev
.groups
= memory_memblk_attr_groups
;
608 memory
->dev
.offline
= memory
->state
== MEM_OFFLINE
;
610 ret
= device_register(&memory
->dev
);
612 put_device(&memory
->dev
);
617 static int init_memory_block(struct memory_block
**memory
,
618 unsigned long block_id
, unsigned long state
)
620 struct memory_block
*mem
;
621 unsigned long start_pfn
;
624 mem
= find_memory_block_by_id(block_id
);
626 put_device(&mem
->dev
);
629 mem
= kzalloc(sizeof(*mem
), GFP_KERNEL
);
633 mem
->start_section_nr
= block_id
* sections_per_block
;
635 start_pfn
= section_nr_to_pfn(mem
->start_section_nr
);
636 mem
->phys_device
= arch_get_memory_phys_device(start_pfn
);
637 mem
->nid
= NUMA_NO_NODE
;
639 ret
= register_memory(mem
);
645 static int add_memory_block(unsigned long base_section_nr
)
647 int ret
, section_count
= 0;
648 struct memory_block
*mem
;
651 for (nr
= base_section_nr
; nr
< base_section_nr
+ sections_per_block
;
653 if (present_section_nr(nr
))
656 if (section_count
== 0)
658 ret
= init_memory_block(&mem
, base_memory_block_id(base_section_nr
),
662 mem
->section_count
= section_count
;
666 static void unregister_memory(struct memory_block
*memory
)
668 if (WARN_ON_ONCE(memory
->dev
.bus
!= &memory_subsys
))
671 /* drop the ref. we got via find_memory_block() */
672 put_device(&memory
->dev
);
673 device_unregister(&memory
->dev
);
677 * Create memory block devices for the given memory area. Start and size
678 * have to be aligned to memory block granularity. Memory block devices
679 * will be initialized as offline.
681 * Called under device_hotplug_lock.
683 int create_memory_block_devices(unsigned long start
, unsigned long size
)
685 const unsigned long start_block_id
= pfn_to_block_id(PFN_DOWN(start
));
686 unsigned long end_block_id
= pfn_to_block_id(PFN_DOWN(start
+ size
));
687 struct memory_block
*mem
;
688 unsigned long block_id
;
691 if (WARN_ON_ONCE(!IS_ALIGNED(start
, memory_block_size_bytes()) ||
692 !IS_ALIGNED(size
, memory_block_size_bytes())))
695 for (block_id
= start_block_id
; block_id
!= end_block_id
; block_id
++) {
696 ret
= init_memory_block(&mem
, block_id
, MEM_OFFLINE
);
699 mem
->section_count
= sections_per_block
;
702 end_block_id
= block_id
;
703 for (block_id
= start_block_id
; block_id
!= end_block_id
;
705 mem
= find_memory_block_by_id(block_id
);
706 if (WARN_ON_ONCE(!mem
))
708 mem
->section_count
= 0;
709 unregister_memory(mem
);
716 * Remove memory block devices for the given memory area. Start and size
717 * have to be aligned to memory block granularity. Memory block devices
718 * have to be offline.
720 * Called under device_hotplug_lock.
722 void remove_memory_block_devices(unsigned long start
, unsigned long size
)
724 const unsigned long start_block_id
= pfn_to_block_id(PFN_DOWN(start
));
725 const unsigned long end_block_id
= pfn_to_block_id(PFN_DOWN(start
+ size
));
726 struct memory_block
*mem
;
727 unsigned long block_id
;
729 if (WARN_ON_ONCE(!IS_ALIGNED(start
, memory_block_size_bytes()) ||
730 !IS_ALIGNED(size
, memory_block_size_bytes())))
733 for (block_id
= start_block_id
; block_id
!= end_block_id
; block_id
++) {
734 mem
= find_memory_block_by_id(block_id
);
735 if (WARN_ON_ONCE(!mem
))
737 mem
->section_count
= 0;
738 unregister_memory_block_under_nodes(mem
);
739 unregister_memory(mem
);
743 /* return true if the memory block is offlined, otherwise, return false */
744 bool is_memblock_offlined(struct memory_block
*mem
)
746 return mem
->state
== MEM_OFFLINE
;
749 static struct attribute
*memory_root_attrs
[] = {
750 #ifdef CONFIG_ARCH_MEMORY_PROBE
751 &dev_attr_probe
.attr
,
754 #ifdef CONFIG_MEMORY_FAILURE
755 &dev_attr_soft_offline_page
.attr
,
756 &dev_attr_hard_offline_page
.attr
,
759 &dev_attr_block_size_bytes
.attr
,
760 &dev_attr_auto_online_blocks
.attr
,
764 static struct attribute_group memory_root_attr_group
= {
765 .attrs
= memory_root_attrs
,
768 static const struct attribute_group
*memory_root_attr_groups
[] = {
769 &memory_root_attr_group
,
774 * Initialize the sysfs support for memory devices. At the time this function
775 * is called, we cannot have concurrent creation/deletion of memory block
776 * devices, the device_hotplug_lock is not needed.
778 void __init
memory_dev_init(void)
781 unsigned long block_sz
, nr
;
783 /* Validate the configured memory block size */
784 block_sz
= memory_block_size_bytes();
785 if (!is_power_of_2(block_sz
) || block_sz
< MIN_MEMORY_BLOCK_SIZE
)
786 panic("Memory block size not suitable: 0x%lx\n", block_sz
);
787 sections_per_block
= block_sz
/ MIN_MEMORY_BLOCK_SIZE
;
789 ret
= subsys_system_register(&memory_subsys
, memory_root_attr_groups
);
791 panic("%s() failed to register subsystem: %d\n", __func__
, ret
);
794 * Create entries for memory sections that were found
795 * during boot and have been initialized
797 for (nr
= 0; nr
<= __highest_present_section_nr
;
798 nr
+= sections_per_block
) {
799 ret
= add_memory_block(nr
);
801 panic("%s() failed to add memory block: %d\n", __func__
,
807 * walk_memory_blocks - walk through all present memory blocks overlapped
808 * by the range [start, start + size)
810 * @start: start address of the memory range
811 * @size: size of the memory range
812 * @arg: argument passed to func
813 * @func: callback for each memory section walked
815 * This function walks through all present memory blocks overlapped by the
816 * range [start, start + size), calling func on each memory block.
818 * In case func() returns an error, walking is aborted and the error is
821 int walk_memory_blocks(unsigned long start
, unsigned long size
,
822 void *arg
, walk_memory_blocks_func_t func
)
824 const unsigned long start_block_id
= phys_to_block_id(start
);
825 const unsigned long end_block_id
= phys_to_block_id(start
+ size
- 1);
826 struct memory_block
*mem
;
827 unsigned long block_id
;
833 for (block_id
= start_block_id
; block_id
<= end_block_id
; block_id
++) {
834 mem
= find_memory_block_by_id(block_id
);
838 ret
= func(mem
, arg
);
839 put_device(&mem
->dev
);
846 struct for_each_memory_block_cb_data
{
847 walk_memory_blocks_func_t func
;
851 static int for_each_memory_block_cb(struct device
*dev
, void *data
)
853 struct memory_block
*mem
= to_memory_block(dev
);
854 struct for_each_memory_block_cb_data
*cb_data
= data
;
856 return cb_data
->func(mem
, cb_data
->arg
);
860 * for_each_memory_block - walk through all present memory blocks
862 * @arg: argument passed to func
863 * @func: callback for each memory block walked
865 * This function walks through all present memory blocks, calling func on
868 * In case func() returns an error, walking is aborted and the error is
871 int for_each_memory_block(void *arg
, walk_memory_blocks_func_t func
)
873 struct for_each_memory_block_cb_data cb_data
= {
878 return bus_for_each_dev(&memory_subsys
, NULL
, &cb_data
,
879 for_each_memory_block_cb
);