xref: /openbmc/linux/mm/page_isolation.c (revision 5bd8e16d)
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 int 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 
57 out:
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 
74 void 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);
86 out:
87 	spin_unlock_irqrestore(&zone->lock, flags);
88 }
89 
90 static 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  */
116 int 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;
137 undo:
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  */
149 int 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  */
173 static 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 
221 int 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 
251 struct 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 }
274