Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/mm/page_isolation.c
4 */
5
6#include <linux/mm.h>
7#include <linux/page-isolation.h>
8#include <linux/pageblock-flags.h>
9#include <linux/memory.h>
10#include <linux/hugetlb.h>
11#include <linux/page_owner.h>
12#include <linux/migrate.h>
13#include "internal.h"
14
15#define CREATE_TRACE_POINTS
16#include <trace/events/page_isolation.h>
17
18static int set_migratetype_isolate(struct page *page, int migratetype,
19 bool skip_hwpoisoned_pages)
20{
21 struct zone *zone;
22 unsigned long flags, pfn;
23 struct memory_isolate_notify arg;
24 int notifier_ret;
25 int ret = -EBUSY;
26
27 zone = page_zone(page);
28
29 spin_lock_irqsave(&zone->lock, flags);
30
31 /*
32 * We assume the caller intended to SET migrate type to isolate.
33 * If it is already set, then someone else must have raced and
34 * set it before us. Return -EBUSY
35 */
36 if (is_migrate_isolate_page(page))
37 goto out;
38
39 pfn = page_to_pfn(page);
40 arg.start_pfn = pfn;
41 arg.nr_pages = pageblock_nr_pages;
42 arg.pages_found = 0;
43
44 /*
45 * It may be possible to isolate a pageblock even if the
46 * migratetype is not MIGRATE_MOVABLE. The memory isolation
47 * notifier chain is used by balloon drivers to return the
48 * number of pages in a range that are held by the balloon
49 * driver to shrink memory. If all the pages are accounted for
50 * by balloons, are free, or on the LRU, isolation can continue.
51 * Later, for example, when memory hotplug notifier runs, these
52 * pages reported as "can be isolated" should be isolated(freed)
53 * by the balloon driver through the memory notifier chain.
54 */
55 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
56 notifier_ret = notifier_to_errno(notifier_ret);
57 if (notifier_ret)
58 goto out;
59 /*
60 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
61 * We just check MOVABLE pages.
62 */
63 if (!has_unmovable_pages(zone, page, arg.pages_found, migratetype,
64 skip_hwpoisoned_pages))
65 ret = 0;
66
67 /*
68 * immobile means "not-on-lru" pages. If immobile is larger than
69 * removable-by-driver pages reported by notifier, we'll fail.
70 */
71
72out:
73 if (!ret) {
74 unsigned long nr_pages;
75 int mt = get_pageblock_migratetype(page);
76
77 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
78 zone->nr_isolate_pageblock++;
79 nr_pages = move_freepages_block(zone, page, MIGRATE_ISOLATE,
80 NULL);
81
82 __mod_zone_freepage_state(zone, -nr_pages, mt);
83 }
84
85 spin_unlock_irqrestore(&zone->lock, flags);
86 if (!ret)
87 drain_all_pages(zone);
88 return ret;
89}
90
91static void unset_migratetype_isolate(struct page *page, unsigned migratetype)
92{
93 struct zone *zone;
94 unsigned long flags, nr_pages;
95 bool isolated_page = false;
96 unsigned int order;
97 unsigned long pfn, buddy_pfn;
98 struct page *buddy;
99
100 zone = page_zone(page);
101 spin_lock_irqsave(&zone->lock, flags);
102 if (!is_migrate_isolate_page(page))
103 goto out;
104
105 /*
106 * Because freepage with more than pageblock_order on isolated
107 * pageblock is restricted to merge due to freepage counting problem,
108 * it is possible that there is free buddy page.
109 * move_freepages_block() doesn't care of merge so we need other
110 * approach in order to merge them. Isolation and free will make
111 * these pages to be merged.
112 */
113 if (PageBuddy(page)) {
114 order = page_order(page);
115 if (order >= pageblock_order) {
116 pfn = page_to_pfn(page);
117 buddy_pfn = __find_buddy_pfn(pfn, order);
118 buddy = page + (buddy_pfn - pfn);
119
120 if (pfn_valid_within(buddy_pfn) &&
121 !is_migrate_isolate_page(buddy)) {
122 __isolate_free_page(page, order);
123 isolated_page = true;
124 }
125 }
126 }
127
128 /*
129 * If we isolate freepage with more than pageblock_order, there
130 * should be no freepage in the range, so we could avoid costly
131 * pageblock scanning for freepage moving.
132 */
133 if (!isolated_page) {
134 nr_pages = move_freepages_block(zone, page, migratetype, NULL);
135 __mod_zone_freepage_state(zone, nr_pages, migratetype);
136 }
137 set_pageblock_migratetype(page, migratetype);
138 zone->nr_isolate_pageblock--;
139out:
140 spin_unlock_irqrestore(&zone->lock, flags);
141 if (isolated_page) {
142 post_alloc_hook(page, order, __GFP_MOVABLE);
143 __free_pages(page, order);
144 }
145}
146
147static inline struct page *
148__first_valid_page(unsigned long pfn, unsigned long nr_pages)
149{
150 int i;
151
152 for (i = 0; i < nr_pages; i++) {
153 struct page *page;
154
155 if (!pfn_valid_within(pfn + i))
156 continue;
157 page = pfn_to_online_page(pfn + i);
158 if (!page)
159 continue;
160 return page;
161 }
162 return NULL;
163}
164
165/*
166 * start_isolate_page_range() -- make page-allocation-type of range of pages
167 * to be MIGRATE_ISOLATE.
168 * @start_pfn: The lower PFN of the range to be isolated.
169 * @end_pfn: The upper PFN of the range to be isolated.
170 * @migratetype: migrate type to set in error recovery.
171 *
172 * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
173 * the range will never be allocated. Any free pages and pages freed in the
174 * future will not be allocated again.
175 *
176 * start_pfn/end_pfn must be aligned to pageblock_order.
177 * Return 0 on success and -EBUSY if any part of range cannot be isolated.
178 *
179 * There is no high level synchronization mechanism that prevents two threads
180 * from trying to isolate overlapping ranges. If this happens, one thread
181 * will notice pageblocks in the overlapping range already set to isolate.
182 * This happens in set_migratetype_isolate, and set_migratetype_isolate
183 * returns an error. We then clean up by restoring the migration type on
184 * pageblocks we may have modified and return -EBUSY to caller. This
185 * prevents two threads from simultaneously working on overlapping ranges.
186 */
187int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
188 unsigned migratetype, bool skip_hwpoisoned_pages)
189{
190 unsigned long pfn;
191 unsigned long undo_pfn;
192 struct page *page;
193
194 BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
195 BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
196
197 for (pfn = start_pfn;
198 pfn < end_pfn;
199 pfn += pageblock_nr_pages) {
200 page = __first_valid_page(pfn, pageblock_nr_pages);
201 if (page &&
202 set_migratetype_isolate(page, migratetype, skip_hwpoisoned_pages)) {
203 undo_pfn = pfn;
204 goto undo;
205 }
206 }
207 return 0;
208undo:
209 for (pfn = start_pfn;
210 pfn < undo_pfn;
211 pfn += pageblock_nr_pages) {
212 struct page *page = pfn_to_online_page(pfn);
213 if (!page)
214 continue;
215 unset_migratetype_isolate(page, migratetype);
216 }
217
218 return -EBUSY;
219}
220
221/*
222 * Make isolated pages available again.
223 */
224int undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
225 unsigned migratetype)
226{
227 unsigned long pfn;
228 struct page *page;
229
230 BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
231 BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
232
233 for (pfn = start_pfn;
234 pfn < end_pfn;
235 pfn += pageblock_nr_pages) {
236 page = __first_valid_page(pfn, pageblock_nr_pages);
237 if (!page || !is_migrate_isolate_page(page))
238 continue;
239 unset_migratetype_isolate(page, migratetype);
240 }
241 return 0;
242}
243/*
244 * Test all pages in the range is free(means isolated) or not.
245 * all pages in [start_pfn...end_pfn) must be in the same zone.
246 * zone->lock must be held before call this.
247 *
248 * Returns the last tested pfn.
249 */
250static unsigned long
251__test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn,
252 bool skip_hwpoisoned_pages)
253{
254 struct page *page;
255
256 while (pfn < end_pfn) {
257 if (!pfn_valid_within(pfn)) {
258 pfn++;
259 continue;
260 }
261 page = pfn_to_page(pfn);
262 if (PageBuddy(page))
263 /*
264 * If the page is on a free list, it has to be on
265 * the correct MIGRATE_ISOLATE freelist. There is no
266 * simple way to verify that as VM_BUG_ON(), though.
267 */
268 pfn += 1 << page_order(page);
269 else if (skip_hwpoisoned_pages && PageHWPoison(page))
270 /* A HWPoisoned page cannot be also PageBuddy */
271 pfn++;
272 else
273 break;
274 }
275
276 return pfn;
277}
278
279/* Caller should ensure that requested range is in a single zone */
280int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
281 bool skip_hwpoisoned_pages)
282{
283 unsigned long pfn, flags;
284 struct page *page;
285 struct zone *zone;
286
287 /*
288 * Note: pageblock_nr_pages != MAX_ORDER. Then, chunks of free pages
289 * are not aligned to pageblock_nr_pages.
290 * Then we just check migratetype first.
291 */
292 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
293 page = __first_valid_page(pfn, pageblock_nr_pages);
294 if (page && !is_migrate_isolate_page(page))
295 break;
296 }
297 page = __first_valid_page(start_pfn, end_pfn - start_pfn);
298 if ((pfn < end_pfn) || !page)
299 return -EBUSY;
300 /* Check all pages are free or marked as ISOLATED */
301 zone = page_zone(page);
302 spin_lock_irqsave(&zone->lock, flags);
303 pfn = __test_page_isolated_in_pageblock(start_pfn, end_pfn,
304 skip_hwpoisoned_pages);
305 spin_unlock_irqrestore(&zone->lock, flags);
306
307 trace_test_pages_isolated(start_pfn, end_pfn, pfn);
308
309 return pfn < end_pfn ? -EBUSY : 0;
310}
311
312struct page *alloc_migrate_target(struct page *page, unsigned long private)
313{
314 return new_page_nodemask(page, numa_node_id(), &node_states[N_MEMORY]);
315}
1/*
2 * linux/mm/page_isolation.c
3 */
4
5#include <linux/mm.h>
6#include <linux/page-isolation.h>
7#include <linux/pageblock-flags.h>
8#include <linux/memory.h>
9#include <linux/hugetlb.h>
10#include "internal.h"
11
12int set_migratetype_isolate(struct page *page, bool skip_hwpoisoned_pages)
13{
14 struct zone *zone;
15 unsigned long flags, pfn;
16 struct memory_isolate_notify arg;
17 int notifier_ret;
18 int ret = -EBUSY;
19
20 zone = page_zone(page);
21
22 spin_lock_irqsave(&zone->lock, flags);
23
24 pfn = page_to_pfn(page);
25 arg.start_pfn = pfn;
26 arg.nr_pages = pageblock_nr_pages;
27 arg.pages_found = 0;
28
29 /*
30 * It may be possible to isolate a pageblock even if the
31 * migratetype is not MIGRATE_MOVABLE. The memory isolation
32 * notifier chain is used by balloon drivers to return the
33 * number of pages in a range that are held by the balloon
34 * driver to shrink memory. If all the pages are accounted for
35 * by balloons, are free, or on the LRU, isolation can continue.
36 * Later, for example, when memory hotplug notifier runs, these
37 * pages reported as "can be isolated" should be isolated(freed)
38 * by the balloon driver through the memory notifier chain.
39 */
40 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
41 notifier_ret = notifier_to_errno(notifier_ret);
42 if (notifier_ret)
43 goto out;
44 /*
45 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
46 * We just check MOVABLE pages.
47 */
48 if (!has_unmovable_pages(zone, page, arg.pages_found,
49 skip_hwpoisoned_pages))
50 ret = 0;
51
52 /*
53 * immobile means "not-on-lru" paes. If immobile is larger than
54 * removable-by-driver pages reported by notifier, we'll fail.
55 */
56
57out:
58 if (!ret) {
59 unsigned long nr_pages;
60 int migratetype = get_pageblock_migratetype(page);
61
62 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
63 nr_pages = move_freepages_block(zone, page, MIGRATE_ISOLATE);
64
65 __mod_zone_freepage_state(zone, -nr_pages, migratetype);
66 }
67
68 spin_unlock_irqrestore(&zone->lock, flags);
69 if (!ret)
70 drain_all_pages();
71 return ret;
72}
73
74void unset_migratetype_isolate(struct page *page, unsigned migratetype)
75{
76 struct zone *zone;
77 unsigned long flags, nr_pages;
78
79 zone = page_zone(page);
80 spin_lock_irqsave(&zone->lock, flags);
81 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
82 goto out;
83 nr_pages = move_freepages_block(zone, page, migratetype);
84 __mod_zone_freepage_state(zone, nr_pages, migratetype);
85 set_pageblock_migratetype(page, migratetype);
86out:
87 spin_unlock_irqrestore(&zone->lock, flags);
88}
89
90static inline struct page *
91__first_valid_page(unsigned long pfn, unsigned long nr_pages)
92{
93 int i;
94 for (i = 0; i < nr_pages; i++)
95 if (pfn_valid_within(pfn + i))
96 break;
97 if (unlikely(i == nr_pages))
98 return NULL;
99 return pfn_to_page(pfn + i);
100}
101
102/*
103 * start_isolate_page_range() -- make page-allocation-type of range of pages
104 * to be MIGRATE_ISOLATE.
105 * @start_pfn: The lower PFN of the range to be isolated.
106 * @end_pfn: The upper PFN of the range to be isolated.
107 * @migratetype: migrate type to set in error recovery.
108 *
109 * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
110 * the range will never be allocated. Any free pages and pages freed in the
111 * future will not be allocated again.
112 *
113 * start_pfn/end_pfn must be aligned to pageblock_order.
114 * Returns 0 on success and -EBUSY if any part of range cannot be isolated.
115 */
116int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
117 unsigned migratetype, bool skip_hwpoisoned_pages)
118{
119 unsigned long pfn;
120 unsigned long undo_pfn;
121 struct page *page;
122
123 BUG_ON((start_pfn) & (pageblock_nr_pages - 1));
124 BUG_ON((end_pfn) & (pageblock_nr_pages - 1));
125
126 for (pfn = start_pfn;
127 pfn < end_pfn;
128 pfn += pageblock_nr_pages) {
129 page = __first_valid_page(pfn, pageblock_nr_pages);
130 if (page &&
131 set_migratetype_isolate(page, skip_hwpoisoned_pages)) {
132 undo_pfn = pfn;
133 goto undo;
134 }
135 }
136 return 0;
137undo:
138 for (pfn = start_pfn;
139 pfn < undo_pfn;
140 pfn += pageblock_nr_pages)
141 unset_migratetype_isolate(pfn_to_page(pfn), migratetype);
142
143 return -EBUSY;
144}
145
146/*
147 * Make isolated pages available again.
148 */
149int undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
150 unsigned migratetype)
151{
152 unsigned long pfn;
153 struct page *page;
154 BUG_ON((start_pfn) & (pageblock_nr_pages - 1));
155 BUG_ON((end_pfn) & (pageblock_nr_pages - 1));
156 for (pfn = start_pfn;
157 pfn < end_pfn;
158 pfn += pageblock_nr_pages) {
159 page = __first_valid_page(pfn, pageblock_nr_pages);
160 if (!page || get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
161 continue;
162 unset_migratetype_isolate(page, migratetype);
163 }
164 return 0;
165}
166/*
167 * Test all pages in the range is free(means isolated) or not.
168 * all pages in [start_pfn...end_pfn) must be in the same zone.
169 * zone->lock must be held before call this.
170 *
171 * Returns 1 if all pages in the range are isolated.
172 */
173static int
174__test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn,
175 bool skip_hwpoisoned_pages)
176{
177 struct page *page;
178
179 while (pfn < end_pfn) {
180 if (!pfn_valid_within(pfn)) {
181 pfn++;
182 continue;
183 }
184 page = pfn_to_page(pfn);
185 if (PageBuddy(page)) {
186 /*
187 * If race between isolatation and allocation happens,
188 * some free pages could be in MIGRATE_MOVABLE list
189 * although pageblock's migratation type of the page
190 * is MIGRATE_ISOLATE. Catch it and move the page into
191 * MIGRATE_ISOLATE list.
192 */
193 if (get_freepage_migratetype(page) != MIGRATE_ISOLATE) {
194 struct page *end_page;
195
196 end_page = page + (1 << page_order(page)) - 1;
197 move_freepages(page_zone(page), page, end_page,
198 MIGRATE_ISOLATE);
199 }
200 pfn += 1 << page_order(page);
201 }
202 else if (page_count(page) == 0 &&
203 get_freepage_migratetype(page) == MIGRATE_ISOLATE)
204 pfn += 1;
205 else if (skip_hwpoisoned_pages && PageHWPoison(page)) {
206 /*
207 * The HWPoisoned page may be not in buddy
208 * system, and page_count() is not 0.
209 */
210 pfn++;
211 continue;
212 }
213 else
214 break;
215 }
216 if (pfn < end_pfn)
217 return 0;
218 return 1;
219}
220
221int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
222 bool skip_hwpoisoned_pages)
223{
224 unsigned long pfn, flags;
225 struct page *page;
226 struct zone *zone;
227 int ret;
228
229 /*
230 * Note: pageblock_nr_pages != MAX_ORDER. Then, chunks of free pages
231 * are not aligned to pageblock_nr_pages.
232 * Then we just check migratetype first.
233 */
234 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
235 page = __first_valid_page(pfn, pageblock_nr_pages);
236 if (page && get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
237 break;
238 }
239 page = __first_valid_page(start_pfn, end_pfn - start_pfn);
240 if ((pfn < end_pfn) || !page)
241 return -EBUSY;
242 /* Check all pages are free or marked as ISOLATED */
243 zone = page_zone(page);
244 spin_lock_irqsave(&zone->lock, flags);
245 ret = __test_page_isolated_in_pageblock(start_pfn, end_pfn,
246 skip_hwpoisoned_pages);
247 spin_unlock_irqrestore(&zone->lock, flags);
248 return ret ? 0 : -EBUSY;
249}
250
251struct page *alloc_migrate_target(struct page *page, unsigned long private,
252 int **resultp)
253{
254 gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
255
256 /*
257 * TODO: allocate a destination hugepage from a nearest neighbor node,
258 * accordance with memory policy of the user process if possible. For
259 * now as a simple work-around, we use the next node for destination.
260 */
261 if (PageHuge(page)) {
262 nodemask_t src = nodemask_of_node(page_to_nid(page));
263 nodemask_t dst;
264 nodes_complement(dst, src);
265 return alloc_huge_page_node(page_hstate(compound_head(page)),
266 next_node(page_to_nid(page), dst));
267 }
268
269 if (PageHighMem(page))
270 gfp_mask |= __GFP_HIGHMEM;
271
272 return alloc_page(gfp_mask);
273}