xref: /openbmc/linux/mm/page_isolation.c (revision 239480ab)
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 <linux/page_owner.h>
11 #include "internal.h"
12 
13 #define CREATE_TRACE_POINTS
14 #include <trace/events/page_isolation.h>
15 
16 static int set_migratetype_isolate(struct page *page,
17 				bool skip_hwpoisoned_pages)
18 {
19 	struct zone *zone;
20 	unsigned long flags, pfn;
21 	struct memory_isolate_notify arg;
22 	int notifier_ret;
23 	int ret = -EBUSY;
24 
25 	zone = page_zone(page);
26 
27 	spin_lock_irqsave(&zone->lock, flags);
28 
29 	pfn = page_to_pfn(page);
30 	arg.start_pfn = pfn;
31 	arg.nr_pages = pageblock_nr_pages;
32 	arg.pages_found = 0;
33 
34 	/*
35 	 * It may be possible to isolate a pageblock even if the
36 	 * migratetype is not MIGRATE_MOVABLE. The memory isolation
37 	 * notifier chain is used by balloon drivers to return the
38 	 * number of pages in a range that are held by the balloon
39 	 * driver to shrink memory. If all the pages are accounted for
40 	 * by balloons, are free, or on the LRU, isolation can continue.
41 	 * Later, for example, when memory hotplug notifier runs, these
42 	 * pages reported as "can be isolated" should be isolated(freed)
43 	 * by the balloon driver through the memory notifier chain.
44 	 */
45 	notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
46 	notifier_ret = notifier_to_errno(notifier_ret);
47 	if (notifier_ret)
48 		goto out;
49 	/*
50 	 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
51 	 * We just check MOVABLE pages.
52 	 */
53 	if (!has_unmovable_pages(zone, page, arg.pages_found,
54 				 skip_hwpoisoned_pages))
55 		ret = 0;
56 
57 	/*
58 	 * immobile means "not-on-lru" pages. If immobile is larger than
59 	 * removable-by-driver pages reported by notifier, we'll fail.
60 	 */
61 
62 out:
63 	if (!ret) {
64 		unsigned long nr_pages;
65 		int migratetype = get_pageblock_migratetype(page);
66 
67 		set_pageblock_migratetype(page, MIGRATE_ISOLATE);
68 		zone->nr_isolate_pageblock++;
69 		nr_pages = move_freepages_block(zone, page, MIGRATE_ISOLATE,
70 									NULL);
71 
72 		__mod_zone_freepage_state(zone, -nr_pages, migratetype);
73 	}
74 
75 	spin_unlock_irqrestore(&zone->lock, flags);
76 	if (!ret)
77 		drain_all_pages(zone);
78 	return ret;
79 }
80 
81 static void unset_migratetype_isolate(struct page *page, unsigned migratetype)
82 {
83 	struct zone *zone;
84 	unsigned long flags, nr_pages;
85 	bool isolated_page = false;
86 	unsigned int order;
87 	unsigned long pfn, buddy_pfn;
88 	struct page *buddy;
89 
90 	zone = page_zone(page);
91 	spin_lock_irqsave(&zone->lock, flags);
92 	if (!is_migrate_isolate_page(page))
93 		goto out;
94 
95 	/*
96 	 * Because freepage with more than pageblock_order on isolated
97 	 * pageblock is restricted to merge due to freepage counting problem,
98 	 * it is possible that there is free buddy page.
99 	 * move_freepages_block() doesn't care of merge so we need other
100 	 * approach in order to merge them. Isolation and free will make
101 	 * these pages to be merged.
102 	 */
103 	if (PageBuddy(page)) {
104 		order = page_order(page);
105 		if (order >= pageblock_order) {
106 			pfn = page_to_pfn(page);
107 			buddy_pfn = __find_buddy_pfn(pfn, order);
108 			buddy = page + (buddy_pfn - pfn);
109 
110 			if (pfn_valid_within(buddy_pfn) &&
111 			    !is_migrate_isolate_page(buddy)) {
112 				__isolate_free_page(page, order);
113 				isolated_page = true;
114 			}
115 		}
116 	}
117 
118 	/*
119 	 * If we isolate freepage with more than pageblock_order, there
120 	 * should be no freepage in the range, so we could avoid costly
121 	 * pageblock scanning for freepage moving.
122 	 */
123 	if (!isolated_page) {
124 		nr_pages = move_freepages_block(zone, page, migratetype, NULL);
125 		__mod_zone_freepage_state(zone, nr_pages, migratetype);
126 	}
127 	set_pageblock_migratetype(page, migratetype);
128 	zone->nr_isolate_pageblock--;
129 out:
130 	spin_unlock_irqrestore(&zone->lock, flags);
131 	if (isolated_page) {
132 		post_alloc_hook(page, order, __GFP_MOVABLE);
133 		__free_pages(page, order);
134 	}
135 }
136 
137 static inline struct page *
138 __first_valid_page(unsigned long pfn, unsigned long nr_pages)
139 {
140 	int i;
141 	for (i = 0; i < nr_pages; i++)
142 		if (pfn_valid_within(pfn + i))
143 			break;
144 	if (unlikely(i == nr_pages))
145 		return NULL;
146 	return pfn_to_page(pfn + i);
147 }
148 
149 /*
150  * start_isolate_page_range() -- make page-allocation-type of range of pages
151  * to be MIGRATE_ISOLATE.
152  * @start_pfn: The lower PFN of the range to be isolated.
153  * @end_pfn: The upper PFN of the range to be isolated.
154  * @migratetype: migrate type to set in error recovery.
155  *
156  * Making page-allocation-type to be MIGRATE_ISOLATE means free pages in
157  * the range will never be allocated. Any free pages and pages freed in the
158  * future will not be allocated again.
159  *
160  * start_pfn/end_pfn must be aligned to pageblock_order.
161  * Returns 0 on success and -EBUSY if any part of range cannot be isolated.
162  */
163 int start_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
164 			     unsigned migratetype, bool skip_hwpoisoned_pages)
165 {
166 	unsigned long pfn;
167 	unsigned long undo_pfn;
168 	struct page *page;
169 
170 	BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
171 	BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
172 
173 	for (pfn = start_pfn;
174 	     pfn < end_pfn;
175 	     pfn += pageblock_nr_pages) {
176 		page = __first_valid_page(pfn, pageblock_nr_pages);
177 		if (page &&
178 		    set_migratetype_isolate(page, skip_hwpoisoned_pages)) {
179 			undo_pfn = pfn;
180 			goto undo;
181 		}
182 	}
183 	return 0;
184 undo:
185 	for (pfn = start_pfn;
186 	     pfn < undo_pfn;
187 	     pfn += pageblock_nr_pages)
188 		unset_migratetype_isolate(pfn_to_page(pfn), migratetype);
189 
190 	return -EBUSY;
191 }
192 
193 /*
194  * Make isolated pages available again.
195  */
196 int undo_isolate_page_range(unsigned long start_pfn, unsigned long end_pfn,
197 			    unsigned migratetype)
198 {
199 	unsigned long pfn;
200 	struct page *page;
201 
202 	BUG_ON(!IS_ALIGNED(start_pfn, pageblock_nr_pages));
203 	BUG_ON(!IS_ALIGNED(end_pfn, pageblock_nr_pages));
204 
205 	for (pfn = start_pfn;
206 	     pfn < end_pfn;
207 	     pfn += pageblock_nr_pages) {
208 		page = __first_valid_page(pfn, pageblock_nr_pages);
209 		if (!page || !is_migrate_isolate_page(page))
210 			continue;
211 		unset_migratetype_isolate(page, migratetype);
212 	}
213 	return 0;
214 }
215 /*
216  * Test all pages in the range is free(means isolated) or not.
217  * all pages in [start_pfn...end_pfn) must be in the same zone.
218  * zone->lock must be held before call this.
219  *
220  * Returns the last tested pfn.
221  */
222 static unsigned long
223 __test_page_isolated_in_pageblock(unsigned long pfn, unsigned long end_pfn,
224 				  bool skip_hwpoisoned_pages)
225 {
226 	struct page *page;
227 
228 	while (pfn < end_pfn) {
229 		if (!pfn_valid_within(pfn)) {
230 			pfn++;
231 			continue;
232 		}
233 		page = pfn_to_page(pfn);
234 		if (PageBuddy(page))
235 			/*
236 			 * If the page is on a free list, it has to be on
237 			 * the correct MIGRATE_ISOLATE freelist. There is no
238 			 * simple way to verify that as VM_BUG_ON(), though.
239 			 */
240 			pfn += 1 << page_order(page);
241 		else if (skip_hwpoisoned_pages && PageHWPoison(page))
242 			/* A HWPoisoned page cannot be also PageBuddy */
243 			pfn++;
244 		else
245 			break;
246 	}
247 
248 	return pfn;
249 }
250 
251 /* Caller should ensure that requested range is in a single zone */
252 int test_pages_isolated(unsigned long start_pfn, unsigned long end_pfn,
253 			bool skip_hwpoisoned_pages)
254 {
255 	unsigned long pfn, flags;
256 	struct page *page;
257 	struct zone *zone;
258 
259 	/*
260 	 * Note: pageblock_nr_pages != MAX_ORDER. Then, chunks of free pages
261 	 * are not aligned to pageblock_nr_pages.
262 	 * Then we just check migratetype first.
263 	 */
264 	for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
265 		page = __first_valid_page(pfn, pageblock_nr_pages);
266 		if (page && !is_migrate_isolate_page(page))
267 			break;
268 	}
269 	page = __first_valid_page(start_pfn, end_pfn - start_pfn);
270 	if ((pfn < end_pfn) || !page)
271 		return -EBUSY;
272 	/* Check all pages are free or marked as ISOLATED */
273 	zone = page_zone(page);
274 	spin_lock_irqsave(&zone->lock, flags);
275 	pfn = __test_page_isolated_in_pageblock(start_pfn, end_pfn,
276 						skip_hwpoisoned_pages);
277 	spin_unlock_irqrestore(&zone->lock, flags);
278 
279 	trace_test_pages_isolated(start_pfn, end_pfn, pfn);
280 
281 	return pfn < end_pfn ? -EBUSY : 0;
282 }
283 
284 struct page *alloc_migrate_target(struct page *page, unsigned long private,
285 				  int **resultp)
286 {
287 	gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE;
288 
289 	/*
290 	 * TODO: allocate a destination hugepage from a nearest neighbor node,
291 	 * accordance with memory policy of the user process if possible. For
292 	 * now as a simple work-around, we use the next node for destination.
293 	 */
294 	if (PageHuge(page))
295 		return alloc_huge_page_node(page_hstate(compound_head(page)),
296 					    next_node_in(page_to_nid(page),
297 							 node_online_map));
298 
299 	if (PageHighMem(page))
300 		gfp_mask |= __GFP_HIGHMEM;
301 
302 	return alloc_page(gfp_mask);
303 }
304