xref: /openbmc/linux/mm/page_isolation.c (revision 05bcf503)
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 "internal.h"
10 
11 /* called while holding zone->lock */
12 static void set_pageblock_isolate(struct page *page)
13 {
14 	if (get_pageblock_migratetype(page) == MIGRATE_ISOLATE)
15 		return;
16 
17 	set_pageblock_migratetype(page, MIGRATE_ISOLATE);
18 	page_zone(page)->nr_pageblock_isolate++;
19 }
20 
21 /* called while holding zone->lock */
22 static void restore_pageblock_isolate(struct page *page, int migratetype)
23 {
24 	struct zone *zone = page_zone(page);
25 	if (WARN_ON(get_pageblock_migratetype(page) != MIGRATE_ISOLATE))
26 		return;
27 
28 	BUG_ON(zone->nr_pageblock_isolate <= 0);
29 	set_pageblock_migratetype(page, migratetype);
30 	zone->nr_pageblock_isolate--;
31 }
32 
33 int set_migratetype_isolate(struct page *page)
34 {
35 	struct zone *zone;
36 	unsigned long flags, pfn;
37 	struct memory_isolate_notify arg;
38 	int notifier_ret;
39 	int ret = -EBUSY;
40 
41 	zone = page_zone(page);
42 
43 	spin_lock_irqsave(&zone->lock, flags);
44 
45 	pfn = page_to_pfn(page);
46 	arg.start_pfn = pfn;
47 	arg.nr_pages = pageblock_nr_pages;
48 	arg.pages_found = 0;
49 
50 	/*
51 	 * It may be possible to isolate a pageblock even if the
52 	 * migratetype is not MIGRATE_MOVABLE. The memory isolation
53 	 * notifier chain is used by balloon drivers to return the
54 	 * number of pages in a range that are held by the balloon
55 	 * driver to shrink memory. If all the pages are accounted for
56 	 * by balloons, are free, or on the LRU, isolation can continue.
57 	 * Later, for example, when memory hotplug notifier runs, these
58 	 * pages reported as "can be isolated" should be isolated(freed)
59 	 * by the balloon driver through the memory notifier chain.
60 	 */
61 	notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
62 	notifier_ret = notifier_to_errno(notifier_ret);
63 	if (notifier_ret)
64 		goto out;
65 	/*
66 	 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
67 	 * We just check MOVABLE pages.
68 	 */
69 	if (!has_unmovable_pages(zone, page, arg.pages_found))
70 		ret = 0;
71 
72 	/*
73 	 * immobile means "not-on-lru" paes. If immobile is larger than
74 	 * removable-by-driver pages reported by notifier, we'll fail.
75 	 */
76 
77 out:
78 	if (!ret) {
79 		unsigned long nr_pages;
80 		int migratetype = get_pageblock_migratetype(page);
81 
82 		set_pageblock_isolate(page);
83 		nr_pages = move_freepages_block(zone, page, MIGRATE_ISOLATE);
84 
85 		__mod_zone_freepage_state(zone, -nr_pages, migratetype);
86 	}
87 
88 	spin_unlock_irqrestore(&zone->lock, flags);
89 	if (!ret)
90 		drain_all_pages();
91 	return ret;
92 }
93 
94 void unset_migratetype_isolate(struct page *page, unsigned migratetype)
95 {
96 	struct zone *zone;
97 	unsigned long flags, nr_pages;
98 
99 	zone = page_zone(page);
100 	spin_lock_irqsave(&zone->lock, flags);
101 	if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
102 		goto out;
103 	nr_pages = move_freepages_block(zone, page, migratetype);
104 	__mod_zone_freepage_state(zone, nr_pages, migratetype);
105 	restore_pageblock_isolate(page, migratetype);
106 out:
107 	spin_unlock_irqrestore(&zone->lock, flags);
108 }
109 
110 static inline struct page *
111 __first_valid_page(unsigned long pfn, unsigned long nr_pages)
112 {
113 	int i;
114 	for (i = 0; i < nr_pages; i++)
115 		if (pfn_valid_within(pfn + i))
116 			break;
117 	if (unlikely(i == nr_pages))
118 		return NULL;
119 	return pfn_to_page(pfn + i);
120 }
121 
122 /*
123  * start_isolate_page_range() -- make page-allocation-type of range of pages
124  * to be MIGRATE_ISOLATE.
125  * @start_pfn: The lower PFN of the range to be isolated.
126  * @end_pfn: The upper PFN of the range to be isolated.
127  * @migratetype: migrate type to set in error recovery.
128  *
129  * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
130  * the range will never be allocated. Any free pages and pages freed in the
131  * future will not be allocated again.
132  *
133  * start_pfn/end_pfn must be aligned to pageblock_order.
134  * Returns 0 on success and -EBUSY if any part of range cannot be isolated.
135  */
136 int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
137 			     unsigned migratetype)
138 {
139 	unsigned long pfn;
140 	unsigned long undo_pfn;
141 	struct page *page;
142 
143 	BUG_ON((start_pfn) & (pageblock_nr_pages - 1));
144 	BUG_ON((end_pfn) & (pageblock_nr_pages - 1));
145 
146 	for (pfn = start_pfn;
147 	     pfn < end_pfn;
148 	     pfn += pageblock_nr_pages) {
149 		page = __first_valid_page(pfn, pageblock_nr_pages);
150 		if (page && set_migratetype_isolate(page)) {
151 			undo_pfn = pfn;
152 			goto undo;
153 		}
154 	}
155 	return 0;
156 undo:
157 	for (pfn = start_pfn;
158 	     pfn < undo_pfn;
159 	     pfn += pageblock_nr_pages)
160 		unset_migratetype_isolate(pfn_to_page(pfn), migratetype);
161 
162 	return -EBUSY;
163 }
164 
165 /*
166  * Make isolated pages available again.
167  */
168 int undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
169 			    unsigned migratetype)
170 {
171 	unsigned long pfn;
172 	struct page *page;
173 	BUG_ON((start_pfn) & (pageblock_nr_pages - 1));
174 	BUG_ON((end_pfn) & (pageblock_nr_pages - 1));
175 	for (pfn = start_pfn;
176 	     pfn < end_pfn;
177 	     pfn += pageblock_nr_pages) {
178 		page = __first_valid_page(pfn, pageblock_nr_pages);
179 		if (!page || get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
180 			continue;
181 		unset_migratetype_isolate(page, migratetype);
182 	}
183 	return 0;
184 }
185 /*
186  * Test all pages in the range is free(means isolated) or not.
187  * all pages in [start_pfn...end_pfn) must be in the same zone.
188  * zone->lock must be held before call this.
189  *
190  * Returns 1 if all pages in the range are isolated.
191  */
192 static int
193 __test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn)
194 {
195 	struct page *page;
196 
197 	while (pfn < end_pfn) {
198 		if (!pfn_valid_within(pfn)) {
199 			pfn++;
200 			continue;
201 		}
202 		page = pfn_to_page(pfn);
203 		if (PageBuddy(page)) {
204 			/*
205 			 * If race between isolatation and allocation happens,
206 			 * some free pages could be in MIGRATE_MOVABLE list
207 			 * although pageblock's migratation type of the page
208 			 * is MIGRATE_ISOLATE. Catch it and move the page into
209 			 * MIGRATE_ISOLATE list.
210 			 */
211 			if (get_freepage_migratetype(page) != MIGRATE_ISOLATE) {
212 				struct page *end_page;
213 
214 				end_page = page + (1 << page_order(page)) - 1;
215 				move_freepages(page_zone(page), page, end_page,
216 						MIGRATE_ISOLATE);
217 			}
218 			pfn += 1 << page_order(page);
219 		}
220 		else if (page_count(page) == 0 &&
221 			get_freepage_migratetype(page) == MIGRATE_ISOLATE)
222 			pfn += 1;
223 		else
224 			break;
225 	}
226 	if (pfn < end_pfn)
227 		return 0;
228 	return 1;
229 }
230 
231 int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn)
232 {
233 	unsigned long pfn, flags;
234 	struct page *page;
235 	struct zone *zone;
236 	int ret;
237 
238 	/*
239 	 * Note: pageblock_nr_page != MAX_ORDER. Then, chunks of free page
240 	 * is not aligned to pageblock_nr_pages.
241 	 * Then we just check pagetype fist.
242 	 */
243 	for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
244 		page = __first_valid_page(pfn, pageblock_nr_pages);
245 		if (page && get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
246 			break;
247 	}
248 	page = __first_valid_page(start_pfn, end_pfn - start_pfn);
249 	if ((pfn < end_pfn) || !page)
250 		return -EBUSY;
251 	/* Check all pages are free or Marked as ISOLATED */
252 	zone = page_zone(page);
253 	spin_lock_irqsave(&zone->lock, flags);
254 	ret = __test_page_isolated_in_pageblock(start_pfn, end_pfn);
255 	spin_unlock_irqrestore(&zone->lock, flags);
256 	return ret ? 0 : -EBUSY;
257 }
258 
259 struct page *alloc_migrate_target(struct page *page, unsigned long private,
260 				  int **resultp)
261 {
262 	gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
263 
264 	if (PageHighMem(page))
265 		gfp_mask |= __GFP_HIGHMEM;
266 
267 	return alloc_page(gfp_mask);
268 }
269