xref: /openbmc/linux/mm/page_isolation.c (revision 7af6fbdd)
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 		drain_all_pages(zone);
53 		return 0;
54 	}
55 
56 	spin_unlock_irqrestore(&zone->lock, flags);
57 	if (isol_flags & REPORT_FAILURE) {
58 		/*
59 		 * printk() with zone->lock held will likely trigger a
60 		 * lockdep splat, so defer it here.
61 		 */
62 		dump_page(unmovable, "unmovable page");
63 	}
64 
65 	return -EBUSY;
66 }
67 
68 static void unset_migratetype_isolate(struct page *page, unsigned migratetype)
69 {
70 	struct zone *zone;
71 	unsigned long flags, nr_pages;
72 	bool isolated_page = false;
73 	unsigned int order;
74 	unsigned long pfn, buddy_pfn;
75 	struct page *buddy;
76 
77 	zone = page_zone(page);
78 	spin_lock_irqsave(&zone->lock, flags);
79 	if (!is_migrate_isolate_page(page))
80 		goto out;
81 
82 	/*
83 	 * Because freepage with more than pageblock_order on isolated
84 	 * pageblock is restricted to merge due to freepage counting problem,
85 	 * it is possible that there is free buddy page.
86 	 * move_freepages_block() doesn't care of merge so we need other
87 	 * approach in order to merge them. Isolation and free will make
88 	 * these pages to be merged.
89 	 */
90 	if (PageBuddy(page)) {
91 		order = page_order(page);
92 		if (order >= pageblock_order) {
93 			pfn = page_to_pfn(page);
94 			buddy_pfn = __find_buddy_pfn(pfn, order);
95 			buddy = page + (buddy_pfn - pfn);
96 
97 			if (pfn_valid_within(buddy_pfn) &&
98 			    !is_migrate_isolate_page(buddy)) {
99 				__isolate_free_page(page, order);
100 				isolated_page = true;
101 			}
102 		}
103 	}
104 
105 	/*
106 	 * If we isolate freepage with more than pageblock_order, there
107 	 * should be no freepage in the range, so we could avoid costly
108 	 * pageblock scanning for freepage moving.
109 	 */
110 	if (!isolated_page) {
111 		nr_pages = move_freepages_block(zone, page, migratetype, NULL);
112 		__mod_zone_freepage_state(zone, nr_pages, migratetype);
113 	}
114 	set_pageblock_migratetype(page, migratetype);
115 	if (isolated_page)
116 		__putback_isolated_page(page, order, migratetype);
117 	zone->nr_isolate_pageblock--;
118 out:
119 	spin_unlock_irqrestore(&zone->lock, flags);
120 }
121 
122 static inline struct page *
123 __first_valid_page(unsigned long pfn, unsigned long nr_pages)
124 {
125 	int i;
126 
127 	for (i = 0; i < nr_pages; i++) {
128 		struct page *page;
129 
130 		page = pfn_to_online_page(pfn + i);
131 		if (!page)
132 			continue;
133 		return page;
134 	}
135 	return NULL;
136 }
137 
138 /**
139  * start_isolate_page_range() - make page-allocation-type of range of pages to
140  * be MIGRATE_ISOLATE.
141  * @start_pfn:		The lower PFN of the range to be isolated.
142  * @end_pfn:		The upper PFN of the range to be isolated.
143  *			start_pfn/end_pfn must be aligned to pageblock_order.
144  * @migratetype:	Migrate type to set in error recovery.
145  * @flags:		The following flags are allowed (they can be combined in
146  *			a bit mask)
147  *			MEMORY_OFFLINE - isolate to offline (!allocate) memory
148  *					 e.g., skip over PageHWPoison() pages
149  *					 and PageOffline() pages.
150  *			REPORT_FAILURE - report details about the failure to
151  *			isolate the range
152  *
153  * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
154  * the range will never be allocated. Any free pages and pages freed in the
155  * future will not be allocated again. If specified range includes migrate types
156  * other than MOVABLE or CMA, this will fail with -EBUSY. For isolating all
157  * pages in the range finally, the caller have to free all pages in the range.
158  * test_page_isolated() can be used for test it.
159  *
160  * There is no high level synchronization mechanism that prevents two threads
161  * from trying to isolate overlapping ranges. If this happens, one thread
162  * will notice pageblocks in the overlapping range already set to isolate.
163  * This happens in set_migratetype_isolate, and set_migratetype_isolate
164  * returns an error. We then clean up by restoring the migration type on
165  * pageblocks we may have modified and return -EBUSY to caller. This
166  * prevents two threads from simultaneously working on overlapping ranges.
167  *
168  * Please note that there is no strong synchronization with the page allocator
169  * either. Pages might be freed while their page blocks are marked ISOLATED.
170  * In some cases pages might still end up on pcp lists and that would allow
171  * for their allocation even when they are in fact isolated already. Depending
172  * on how strong of a guarantee the caller needs drain_all_pages might be needed
173  * (e.g. __offline_pages will need to call it after check for isolated range for
174  * a next retry).
175  *
176  * Return: the number of isolated pageblocks on success and -EBUSY if any part
177  * of range cannot be isolated.
178  */
179 int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
180 			     unsigned migratetype, int flags)
181 {
182 	unsigned long pfn;
183 	unsigned long undo_pfn;
184 	struct page *page;
185 	int nr_isolate_pageblock = 0;
186 
187 	BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
188 	BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
189 
190 	for (pfn = start_pfn;
191 	     pfn < end_pfn;
192 	     pfn += pageblock_nr_pages) {
193 		page = __first_valid_page(pfn, pageblock_nr_pages);
194 		if (page) {
195 			if (set_migratetype_isolate(page, migratetype, flags)) {
196 				undo_pfn = pfn;
197 				goto undo;
198 			}
199 			nr_isolate_pageblock++;
200 		}
201 	}
202 	return nr_isolate_pageblock;
203 undo:
204 	for (pfn = start_pfn;
205 	     pfn < undo_pfn;
206 	     pfn += pageblock_nr_pages) {
207 		struct page *page = pfn_to_online_page(pfn);
208 		if (!page)
209 			continue;
210 		unset_migratetype_isolate(page, migratetype);
211 	}
212 
213 	return -EBUSY;
214 }
215 
216 /*
217  * Make isolated pages available again.
218  */
219 void undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
220 			    unsigned migratetype)
221 {
222 	unsigned long pfn;
223 	struct page *page;
224 
225 	BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
226 	BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
227 
228 	for (pfn = start_pfn;
229 	     pfn < end_pfn;
230 	     pfn += pageblock_nr_pages) {
231 		page = __first_valid_page(pfn, pageblock_nr_pages);
232 		if (!page || !is_migrate_isolate_page(page))
233 			continue;
234 		unset_migratetype_isolate(page, migratetype);
235 	}
236 }
237 /*
238  * Test all pages in the range is free(means isolated) or not.
239  * all pages in [start_pfn...end_pfn) must be in the same zone.
240  * zone->lock must be held before call this.
241  *
242  * Returns the last tested pfn.
243  */
244 static unsigned long
245 __test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn,
246 				  int flags)
247 {
248 	struct page *page;
249 
250 	while (pfn < end_pfn) {
251 		if (!pfn_valid_within(pfn)) {
252 			pfn++;
253 			continue;
254 		}
255 		page = pfn_to_page(pfn);
256 		if (PageBuddy(page))
257 			/*
258 			 * If the page is on a free list, it has to be on
259 			 * the correct MIGRATE_ISOLATE freelist. There is no
260 			 * simple way to verify that as VM_BUG_ON(), though.
261 			 */
262 			pfn += 1 << page_order(page);
263 		else if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
264 			/* A HWPoisoned page cannot be also PageBuddy */
265 			pfn++;
266 		else if ((flags & MEMORY_OFFLINE) && PageOffline(page) &&
267 			 !page_count(page))
268 			/*
269 			 * The responsible driver agreed to skip PageOffline()
270 			 * pages when offlining memory by dropping its
271 			 * reference in MEM_GOING_OFFLINE.
272 			 */
273 			pfn++;
274 		else
275 			break;
276 	}
277 
278 	return pfn;
279 }
280 
281 /* Caller should ensure that requested range is in a single zone */
282 int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
283 			int isol_flags)
284 {
285 	unsigned long pfn, flags;
286 	struct page *page;
287 	struct zone *zone;
288 
289 	/*
290 	 * Note: pageblock_nr_pages != MAX_ORDER. Then, chunks of free pages
291 	 * are not aligned to pageblock_nr_pages.
292 	 * Then we just check migratetype first.
293 	 */
294 	for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
295 		page = __first_valid_page(pfn, pageblock_nr_pages);
296 		if (page && !is_migrate_isolate_page(page))
297 			break;
298 	}
299 	page = __first_valid_page(start_pfn, end_pfn - start_pfn);
300 	if ((pfn < end_pfn) || !page)
301 		return -EBUSY;
302 	/* Check all pages are free or marked as ISOLATED */
303 	zone = page_zone(page);
304 	spin_lock_irqsave(&zone->lock, flags);
305 	pfn = __test_page_isolated_in_pageblock(start_pfn, end_pfn, isol_flags);
306 	spin_unlock_irqrestore(&zone->lock, flags);
307 
308 	trace_test_pages_isolated(start_pfn, end_pfn, pfn);
309 
310 	return pfn < end_pfn ? -EBUSY : 0;
311 }
312