page_idle.c

  1// SPDX-License-Identifier: GPL-2.0
  2#include <linux/init.h>
  3#include <linux/memblock.h>
  4#include <linux/fs.h>
  5#include <linux/sysfs.h>
  6#include <linux/kobject.h>
  7#include <linux/memory_hotplug.h>
  8#include <linux/mm.h>
  9#include <linux/mmzone.h>
 10#include <linux/pagemap.h>
 11#include <linux/rmap.h>
 12#include <linux/mmu_notifier.h>
 13#include <linux/page_ext.h>
 14#include <linux/page_idle.h>
 15
 16#include "internal.h"
 17
 18#define BITMAP_CHUNK_SIZE	sizeof(u64)
 19#define BITMAP_CHUNK_BITS	(BITMAP_CHUNK_SIZE * BITS_PER_BYTE)
 20
 21/*
 22 * Idle page tracking only considers user memory pages, for other types of
 23 * pages the idle flag is always unset and an attempt to set it is silently
 24 * ignored.
 25 *
 26 * We treat a page as a user memory page if it is on an LRU list, because it is
 27 * always safe to pass such a page to rmap_walk(), which is essential for idle
 28 * page tracking. With such an indicator of user pages we can skip isolated
 29 * pages, but since there are not usually many of them, it will hardly affect
 30 * the overall result.
 31 *
 32 * This function tries to get a user memory page by pfn as described above.
 33 */
 34static struct folio *page_idle_get_folio(unsigned long pfn)
 35{
 36	struct page *page = pfn_to_online_page(pfn);
 37	struct folio *folio;
 38
 39	if (!page || PageTail(page))
 40		return NULL;
 41
 42	folio = page_folio(page);
 43	if (!folio_test_lru(folio) || !folio_try_get(folio))
 44		return NULL;
 45	if (unlikely(page_folio(page) != folio || !folio_test_lru(folio))) {
 46		folio_put(folio);
 47		folio = NULL;
 48	}
 49	return folio;
 50}
 51
 52static bool page_idle_clear_pte_refs_one(struct folio *folio,
 53					struct vm_area_struct *vma,
 54					unsigned long addr, void *arg)
 55{
 56	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, addr, 0);
 57	bool referenced = false;
 58
 59	while (page_vma_mapped_walk(&pvmw)) {
 60		addr = pvmw.address;
 61		if (pvmw.pte) {
 62			/*
 63			 * For PTE-mapped THP, one sub page is referenced,
 64			 * the whole THP is referenced.
 65			 *
 66			 * PFN swap PTEs, such as device-exclusive ones, that
 67			 * actually map pages are "old" from a CPU perspective.
 68			 * The MMU notifier takes care of any device aspects.
 69			 */
 70			if (likely(pte_present(ptep_get(pvmw.pte))))
 71				referenced |= ptep_test_and_clear_young(vma, addr, pvmw.pte);
 72			referenced |= mmu_notifier_clear_young(vma->vm_mm, addr, addr + PAGE_SIZE);
 73		} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
 74			pmd_t pmdval = pmdp_get(pvmw.pmd);
 75
 76			if (likely(pmd_present(pmdval)))
 77				referenced |= pmdp_clear_young_notify(vma, addr, pvmw.pmd);
 78			referenced |= mmu_notifier_clear_young(vma->vm_mm, addr, addr + PMD_SIZE);
 79		} else {
 80			/* unexpected pmd-mapped page? */
 81			WARN_ON_ONCE(1);
 82		}
 83	}
 84
 85	if (referenced) {
 86		folio_clear_idle(folio);
 87		/*
 88		 * We cleared the referenced bit in a mapping to this page. To
 89		 * avoid interference with page reclaim, mark it young so that
 90		 * folio_referenced() will return > 0.
 91		 */
 92		folio_set_young(folio);
 93	}
 94	return true;
 95}
 96
 97static void page_idle_clear_pte_refs(struct folio *folio)
 98{
 99	/*
100	 * Since rwc.try_lock is unused, rwc is effectively immutable, so we
101	 * can make it static to save some cycles and stack.
102	 */
103	static struct rmap_walk_control rwc = {
104		.rmap_one = page_idle_clear_pte_refs_one,
105		.anon_lock = folio_lock_anon_vma_read,
106	};
107
108	if (!folio_mapped(folio) || !folio_raw_mapping(folio))
109		return;
110
111	if (!folio_trylock(folio))
112		return;
113
114	rmap_walk(folio, &rwc);
115	folio_unlock(folio);
116}
117
118static ssize_t page_idle_bitmap_read(struct file *file, struct kobject *kobj,
119				     const struct bin_attribute *attr, char *buf,
120				     loff_t pos, size_t count)
121{
122	u64 *out = (u64 *)buf;
123	struct folio *folio;
124	unsigned long pfn, end_pfn;
125	int bit;
126
127	if (pos % BITMAP_CHUNK_SIZE || count % BITMAP_CHUNK_SIZE)
128		return -EINVAL;
129
130	pfn = pos * BITS_PER_BYTE;
131	if (pfn >= max_pfn)
132		return 0;
133
134	end_pfn = pfn + count * BITS_PER_BYTE;
135	if (end_pfn > max_pfn)
136		end_pfn = max_pfn;
137
138	for (; pfn < end_pfn; pfn++) {
139		bit = pfn % BITMAP_CHUNK_BITS;
140		if (!bit)
141			*out = 0ULL;
142		folio = page_idle_get_folio(pfn);
143		if (folio) {
144			if (folio_test_idle(folio)) {
145				/*
146				 * The page might have been referenced via a
147				 * pte, in which case it is not idle. Clear
148				 * refs and recheck.
149				 */
150				page_idle_clear_pte_refs(folio);
151				if (folio_test_idle(folio))
152					*out |= 1ULL << bit;
153			}
154			folio_put(folio);
155		}
156		if (bit == BITMAP_CHUNK_BITS - 1)
157			out++;
158		cond_resched();
159	}
160	return (char *)out - buf;
161}
162
163static ssize_t page_idle_bitmap_write(struct file *file, struct kobject *kobj,
164				      const struct bin_attribute *attr, char *buf,
165				      loff_t pos, size_t count)
166{
167	const u64 *in = (u64 *)buf;
168	struct folio *folio;
169	unsigned long pfn, end_pfn;
170	int bit;
171
172	if (pos % BITMAP_CHUNK_SIZE || count % BITMAP_CHUNK_SIZE)
173		return -EINVAL;
174
175	pfn = pos * BITS_PER_BYTE;
176	if (pfn >= max_pfn)
177		return -ENXIO;
178
179	end_pfn = pfn + count * BITS_PER_BYTE;
180	if (end_pfn > max_pfn)
181		end_pfn = max_pfn;
182
183	for (; pfn < end_pfn; pfn++) {
184		bit = pfn % BITMAP_CHUNK_BITS;
185		if ((*in >> bit) & 1) {
186			folio = page_idle_get_folio(pfn);
187			if (folio) {
188				page_idle_clear_pte_refs(folio);
189				folio_set_idle(folio);
190				folio_put(folio);
191			}
192		}
193		if (bit == BITMAP_CHUNK_BITS - 1)
194			in++;
195		cond_resched();
196	}
197	return (char *)in - buf;
198}
199
200static const struct bin_attribute page_idle_bitmap_attr =
201		__BIN_ATTR(bitmap, 0600,
202			   page_idle_bitmap_read, page_idle_bitmap_write, 0);
203
204static const struct bin_attribute *const page_idle_bin_attrs[] = {
205	&page_idle_bitmap_attr,
206	NULL,
207};
208
209static const struct attribute_group page_idle_attr_group = {
210	.bin_attrs = page_idle_bin_attrs,
211	.name = "page_idle",
212};
213
214static int __init page_idle_init(void)
215{
216	int err;
217
218	err = sysfs_create_group(mm_kobj, &page_idle_attr_group);
219	if (err) {
220		pr_err("page_idle: register sysfs failed\n");
221		return err;
222	}
223	return 0;
224}
225subsys_initcall(page_idle_init);