xref: /linux/mm/page_isolation.c (revision 1c4b5ecb7ea190fa3e9f9d6891e6c90b60e04f24)
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 
18 static 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 
67 static 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 (!is_migrate_isolate_page(buddy)) {
97 				isolated_page = !!__isolate_free_page(page, order);
98 				/*
99 				 * Isolating a free page in an isolated pageblock
100 				 * is expected to always work as watermarks don't
101 				 * apply here.
102 				 */
103 				VM_WARN_ON(!isolated_page);
104 			}
105 		}
106 	}
107 
108 	/*
109 	 * If we isolate freepage with more than pageblock_order, there
110 	 * should be no freepage in the range, so we could avoid costly
111 	 * pageblock scanning for freepage moving.
112 	 *
113 	 * We didn't actually touch any of the isolated pages, so place them
114 	 * to the tail of the freelist. This is an optimization for memory
115 	 * onlining - just onlined memory won't immediately be considered for
116 	 * allocation.
117 	 */
118 	if (!isolated_page) {
119 		nr_pages = move_freepages_block(zone, page, migratetype, NULL);
120 		__mod_zone_freepage_state(zone, nr_pages, migratetype);
121 	}
122 	set_pageblock_migratetype(page, migratetype);
123 	if (isolated_page)
124 		__putback_isolated_page(page, order, migratetype);
125 	zone->nr_isolate_pageblock--;
126 out:
127 	spin_unlock_irqrestore(&zone->lock, flags);
128 }
129 
130 static inline struct page *
131 __first_valid_page(unsigned long pfn, unsigned long nr_pages)
132 {
133 	int i;
134 
135 	for (i = 0; i < nr_pages; i++) {
136 		struct page *page;
137 
138 		page = pfn_to_online_page(pfn + i);
139 		if (!page)
140 			continue;
141 		return page;
142 	}
143 	return NULL;
144 }
145 
146 /**
147  * start_isolate_page_range() - make page-allocation-type of range of pages to
148  * be MIGRATE_ISOLATE.
149  * @start_pfn:		The lower PFN of the range to be isolated.
150  * @end_pfn:		The upper PFN of the range to be isolated.
151  *			start_pfn/end_pfn must be aligned to pageblock_order.
152  * @migratetype:	Migrate type to set in error recovery.
153  * @flags:		The following flags are allowed (they can be combined in
154  *			a bit mask)
155  *			MEMORY_OFFLINE - isolate to offline (!allocate) memory
156  *					 e.g., skip over PageHWPoison() pages
157  *					 and PageOffline() pages.
158  *			REPORT_FAILURE - report details about the failure to
159  *			isolate the range
160  *
161  * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
162  * the range will never be allocated. Any free pages and pages freed in the
163  * future will not be allocated again. If specified range includes migrate types
164  * other than MOVABLE or CMA, this will fail with -EBUSY. For isolating all
165  * pages in the range finally, the caller have to free all pages in the range.
166  * test_page_isolated() can be used for test it.
167  *
168  * There is no high level synchronization mechanism that prevents two threads
169  * from trying to isolate overlapping ranges. If this happens, one thread
170  * will notice pageblocks in the overlapping range already set to isolate.
171  * This happens in set_migratetype_isolate, and set_migratetype_isolate
172  * returns an error. We then clean up by restoring the migration type on
173  * pageblocks we may have modified and return -EBUSY to caller. This
174  * prevents two threads from simultaneously working on overlapping ranges.
175  *
176  * Please note that there is no strong synchronization with the page allocator
177  * either. Pages might be freed while their page blocks are marked ISOLATED.
178  * A call to drain_all_pages() after isolation can flush most of them. However
179  * in some cases pages might still end up on pcp lists and that would allow
180  * for their allocation even when they are in fact isolated already. Depending
181  * on how strong of a guarantee the caller needs, zone_pcp_disable/enable()
182  * might be used to flush and disable pcplist before isolation and enable after
183  * unisolation.
184  *
185  * Return: 0 on success and -EBUSY if any part of range cannot be isolated.
186  */
187 int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
188 			     unsigned migratetype, int flags)
189 {
190 	unsigned long pfn;
191 	struct page *page;
192 
193 	BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
194 	BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
195 
196 	for (pfn = start_pfn;
197 	     pfn < end_pfn;
198 	     pfn += pageblock_nr_pages) {
199 		page = __first_valid_page(pfn, pageblock_nr_pages);
200 		if (page && set_migratetype_isolate(page, migratetype, flags)) {
201 			undo_isolate_page_range(start_pfn, pfn, migratetype);
202 			return -EBUSY;
203 		}
204 	}
205 	return 0;
206 }
207 
208 /*
209  * Make isolated pages available again.
210  */
211 void undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
212 			    unsigned migratetype)
213 {
214 	unsigned long pfn;
215 	struct page *page;
216 
217 	BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
218 	BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
219 
220 	for (pfn = start_pfn;
221 	     pfn < end_pfn;
222 	     pfn += pageblock_nr_pages) {
223 		page = __first_valid_page(pfn, pageblock_nr_pages);
224 		if (!page || !is_migrate_isolate_page(page))
225 			continue;
226 		unset_migratetype_isolate(page, migratetype);
227 	}
228 }
229 /*
230  * Test all pages in the range is free(means isolated) or not.
231  * all pages in [start_pfn...end_pfn) must be in the same zone.
232  * zone->lock must be held before call this.
233  *
234  * Returns the last tested pfn.
235  */
236 static unsigned long
237 __test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn,
238 				  int flags)
239 {
240 	struct page *page;
241 
242 	while (pfn < end_pfn) {
243 		page = pfn_to_page(pfn);
244 		if (PageBuddy(page))
245 			/*
246 			 * If the page is on a free list, it has to be on
247 			 * the correct MIGRATE_ISOLATE freelist. There is no
248 			 * simple way to verify that as VM_BUG_ON(), though.
249 			 */
250 			pfn += 1 << buddy_order(page);
251 		else if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
252 			/* A HWPoisoned page cannot be also PageBuddy */
253 			pfn++;
254 		else if ((flags & MEMORY_OFFLINE) && PageOffline(page) &&
255 			 !page_count(page))
256 			/*
257 			 * The responsible driver agreed to skip PageOffline()
258 			 * pages when offlining memory by dropping its
259 			 * reference in MEM_GOING_OFFLINE.
260 			 */
261 			pfn++;
262 		else
263 			break;
264 	}
265 
266 	return pfn;
267 }
268 
269 /* Caller should ensure that requested range is in a single zone */
270 int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
271 			int isol_flags)
272 {
273 	unsigned long pfn, flags;
274 	struct page *page;
275 	struct zone *zone;
276 	int ret;
277 
278 	/*
279 	 * Note: pageblock_nr_pages != MAX_ORDER. Then, chunks of free pages
280 	 * are not aligned to pageblock_nr_pages.
281 	 * Then we just check migratetype first.
282 	 */
283 	for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
284 		page = __first_valid_page(pfn, pageblock_nr_pages);
285 		if (page && !is_migrate_isolate_page(page))
286 			break;
287 	}
288 	page = __first_valid_page(start_pfn, end_pfn - start_pfn);
289 	if ((pfn < end_pfn) || !page) {
290 		ret = -EBUSY;
291 		goto out;
292 	}
293 
294 	/* Check all pages are free or marked as ISOLATED */
295 	zone = page_zone(page);
296 	spin_lock_irqsave(&zone->lock, flags);
297 	pfn = __test_page_isolated_in_pageblock(start_pfn, end_pfn, isol_flags);
298 	spin_unlock_irqrestore(&zone->lock, flags);
299 
300 	ret = pfn < end_pfn ? -EBUSY : 0;
301 
302 out:
303 	trace_test_pages_isolated(start_pfn, end_pfn, pfn);
304 
305 	return ret;
306 }
307