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