xref: /openbmc/linux/mm/page_isolation.c (revision 0da85d1e)
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 		zone->nr_isolate_pageblock++;
64 		nr_pages = move_freepages_block(zone, page, MIGRATE_ISOLATE);
65 
66 		__mod_zone_freepage_state(zone, -nr_pages, migratetype);
67 	}
68 
69 	spin_unlock_irqrestore(&zone->lock, flags);
70 	if (!ret)
71 		drain_all_pages(zone);
72 	return ret;
73 }
74 
75 void unset_migratetype_isolate(struct page *page, unsigned migratetype)
76 {
77 	struct zone *zone;
78 	unsigned long flags, nr_pages;
79 	struct page *isolated_page = NULL;
80 	unsigned int order;
81 	unsigned long page_idx, buddy_idx;
82 	struct page *buddy;
83 
84 	zone = page_zone(page);
85 	spin_lock_irqsave(&zone->lock, flags);
86 	if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
87 		goto out;
88 
89 	/*
90 	 * Because freepage with more than pageblock_order on isolated
91 	 * pageblock is restricted to merge due to freepage counting problem,
92 	 * it is possible that there is free buddy page.
93 	 * move_freepages_block() doesn't care of merge so we need other
94 	 * approach in order to merge them. Isolation and free will make
95 	 * these pages to be merged.
96 	 */
97 	if (PageBuddy(page)) {
98 		order = page_order(page);
99 		if (order >= pageblock_order) {
100 			page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
101 			buddy_idx = __find_buddy_index(page_idx, order);
102 			buddy = page + (buddy_idx - page_idx);
103 
104 			if (!is_migrate_isolate_page(buddy)) {
105 				__isolate_free_page(page, order);
106 				kernel_map_pages(page, (1 << order), 1);
107 				set_page_refcounted(page);
108 				isolated_page = page;
109 			}
110 		}
111 	}
112 
113 	/*
114 	 * If we isolate freepage with more than pageblock_order, there
115 	 * should be no freepage in the range, so we could avoid costly
116 	 * pageblock scanning for freepage moving.
117 	 */
118 	if (!isolated_page) {
119 		nr_pages = move_freepages_block(zone, page, migratetype);
120 		__mod_zone_freepage_state(zone, nr_pages, migratetype);
121 	}
122 	set_pageblock_migratetype(page, migratetype);
123 	zone->nr_isolate_pageblock--;
124 out:
125 	spin_unlock_irqrestore(&zone->lock, flags);
126 	if (isolated_page)
127 		__free_pages(isolated_page, order);
128 }
129 
130 static inline struct page *
131 __first_valid_page(unsigned long pfn, unsigned long nr_pages)
132 {
133 	int i;
134 	for (i = 0; i < nr_pages; i++)
135 		if (pfn_valid_within(pfn + i))
136 			break;
137 	if (unlikely(i == nr_pages))
138 		return NULL;
139 	return pfn_to_page(pfn + i);
140 }
141 
142 /*
143  * start_isolate_page_range() -- make page-allocation-type of range of pages
144  * to 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  * @migratetype: migrate type to set in error recovery.
148  *
149  * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
150  * the range will never be allocated. Any free pages and pages freed in the
151  * future will not be allocated again.
152  *
153  * start_pfn/end_pfn must be aligned to pageblock_order.
154  * Returns 0 on success and -EBUSY if any part of range cannot be isolated.
155  */
156 int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
157 			     unsigned migratetype, bool skip_hwpoisoned_pages)
158 {
159 	unsigned long pfn;
160 	unsigned long undo_pfn;
161 	struct page *page;
162 
163 	BUG_ON((start_pfn) & (pageblock_nr_pages - 1));
164 	BUG_ON((end_pfn) & (pageblock_nr_pages - 1));
165 
166 	for (pfn = start_pfn;
167 	     pfn < end_pfn;
168 	     pfn += pageblock_nr_pages) {
169 		page = __first_valid_page(pfn, pageblock_nr_pages);
170 		if (page &&
171 		    set_migratetype_isolate(page, skip_hwpoisoned_pages)) {
172 			undo_pfn = pfn;
173 			goto undo;
174 		}
175 	}
176 	return 0;
177 undo:
178 	for (pfn = start_pfn;
179 	     pfn < undo_pfn;
180 	     pfn += pageblock_nr_pages)
181 		unset_migratetype_isolate(pfn_to_page(pfn), migratetype);
182 
183 	return -EBUSY;
184 }
185 
186 /*
187  * Make isolated pages available again.
188  */
189 int undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
190 			    unsigned migratetype)
191 {
192 	unsigned long pfn;
193 	struct page *page;
194 	BUG_ON((start_pfn) & (pageblock_nr_pages - 1));
195 	BUG_ON((end_pfn) & (pageblock_nr_pages - 1));
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 || get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
201 			continue;
202 		unset_migratetype_isolate(page, migratetype);
203 	}
204 	return 0;
205 }
206 /*
207  * Test all pages in the range is free(means isolated) or not.
208  * all pages in [start_pfn...end_pfn) must be in the same zone.
209  * zone->lock must be held before call this.
210  *
211  * Returns 1 if all pages in the range are isolated.
212  */
213 static int
214 __test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn,
215 				  bool skip_hwpoisoned_pages)
216 {
217 	struct page *page;
218 
219 	while (pfn < end_pfn) {
220 		if (!pfn_valid_within(pfn)) {
221 			pfn++;
222 			continue;
223 		}
224 		page = pfn_to_page(pfn);
225 		if (PageBuddy(page)) {
226 			/*
227 			 * If race between isolatation and allocation happens,
228 			 * some free pages could be in MIGRATE_MOVABLE list
229 			 * although pageblock's migratation type of the page
230 			 * is MIGRATE_ISOLATE. Catch it and move the page into
231 			 * MIGRATE_ISOLATE list.
232 			 */
233 			if (get_freepage_migratetype(page) != MIGRATE_ISOLATE) {
234 				struct page *end_page;
235 
236 				end_page = page + (1 << page_order(page)) - 1;
237 				move_freepages(page_zone(page), page, end_page,
238 						MIGRATE_ISOLATE);
239 			}
240 			pfn += 1 << page_order(page);
241 		}
242 		else if (page_count(page) == 0 &&
243 			get_freepage_migratetype(page) == MIGRATE_ISOLATE)
244 			pfn += 1;
245 		else if (skip_hwpoisoned_pages && PageHWPoison(page)) {
246 			/*
247 			 * The HWPoisoned page may be not in buddy
248 			 * system, and page_count() is not 0.
249 			 */
250 			pfn++;
251 			continue;
252 		}
253 		else
254 			break;
255 	}
256 	if (pfn < end_pfn)
257 		return 0;
258 	return 1;
259 }
260 
261 int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
262 			bool skip_hwpoisoned_pages)
263 {
264 	unsigned long pfn, flags;
265 	struct page *page;
266 	struct zone *zone;
267 	int ret;
268 
269 	/*
270 	 * Note: pageblock_nr_pages != MAX_ORDER. Then, chunks of free pages
271 	 * are not aligned to pageblock_nr_pages.
272 	 * Then we just check migratetype first.
273 	 */
274 	for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
275 		page = __first_valid_page(pfn, pageblock_nr_pages);
276 		if (page && get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
277 			break;
278 	}
279 	page = __first_valid_page(start_pfn, end_pfn - start_pfn);
280 	if ((pfn < end_pfn) || !page)
281 		return -EBUSY;
282 	/* Check all pages are free or marked as ISOLATED */
283 	zone = page_zone(page);
284 	spin_lock_irqsave(&zone->lock, flags);
285 	ret = __test_page_isolated_in_pageblock(start_pfn, end_pfn,
286 						skip_hwpoisoned_pages);
287 	spin_unlock_irqrestore(&zone->lock, flags);
288 	return ret ? 0 : -EBUSY;
289 }
290 
291 struct page *alloc_migrate_target(struct page *page, unsigned long private,
292 				  int **resultp)
293 {
294 	gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
295 
296 	/*
297 	 * TODO: allocate a destination hugepage from a nearest neighbor node,
298 	 * accordance with memory policy of the user process if possible. For
299 	 * now as a simple work-around, we use the next node for destination.
300 	 */
301 	if (PageHuge(page)) {
302 		nodemask_t src = nodemask_of_node(page_to_nid(page));
303 		nodemask_t dst;
304 		nodes_complement(dst, src);
305 		return alloc_huge_page_node(page_hstate(compound_head(page)),
306 					    next_node(page_to_nid(page), dst));
307 	}
308 
309 	if (PageHighMem(page))
310 		gfp_mask |= __GFP_HIGHMEM;
311 
312 	return alloc_page(gfp_mask);
313 }
314