xref: /openbmc/linux/mm/page_reporting.c (revision ded1ffea)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/mm.h>
3 #include <linux/mmzone.h>
4 #include <linux/page_reporting.h>
5 #include <linux/gfp.h>
6 #include <linux/export.h>
7 #include <linux/module.h>
8 #include <linux/delay.h>
9 #include <linux/scatterlist.h>
10 
11 #include "page_reporting.h"
12 #include "internal.h"
13 
14 /* Initialize to an unsupported value */
15 unsigned int page_reporting_order = -1;
16 
page_order_update_notify(const char * val,const struct kernel_param * kp)17 static int page_order_update_notify(const char *val, const struct kernel_param *kp)
18 {
19 	/*
20 	 * If param is set beyond this limit, order is set to default
21 	 * pageblock_order value
22 	 */
23 	return  param_set_uint_minmax(val, kp, 0, MAX_ORDER);
24 }
25 
26 static const struct kernel_param_ops page_reporting_param_ops = {
27 	.set = &page_order_update_notify,
28 	/*
29 	 * For the get op, use param_get_int instead of param_get_uint.
30 	 * This is to make sure that when unset the initialized value of
31 	 * -1 is shown correctly
32 	 */
33 	.get = &param_get_int,
34 };
35 
36 module_param_cb(page_reporting_order, &page_reporting_param_ops,
37 			&page_reporting_order, 0644);
38 MODULE_PARM_DESC(page_reporting_order, "Set page reporting order");
39 
40 /*
41  * This symbol is also a kernel parameter. Export the page_reporting_order
42  * symbol so that other drivers can access it to control order values without
43  * having to introduce another configurable parameter. Only one driver can
44  * register with the page_reporting driver for the service, so we have just
45  * one control parameter for the use case(which can be accessed in both
46  * drivers)
47  */
48 EXPORT_SYMBOL_GPL(page_reporting_order);
49 
50 #define PAGE_REPORTING_DELAY	(2 * HZ)
51 static struct page_reporting_dev_info __rcu *pr_dev_info __read_mostly;
52 
53 enum {
54 	PAGE_REPORTING_IDLE = 0,
55 	PAGE_REPORTING_REQUESTED,
56 	PAGE_REPORTING_ACTIVE
57 };
58 
59 /* request page reporting */
60 static void
__page_reporting_request(struct page_reporting_dev_info * prdev)61 __page_reporting_request(struct page_reporting_dev_info *prdev)
62 {
63 	unsigned int state;
64 
65 	/* Check to see if we are in desired state */
66 	state = atomic_read(&prdev->state);
67 	if (state == PAGE_REPORTING_REQUESTED)
68 		return;
69 
70 	/*
71 	 * If reporting is already active there is nothing we need to do.
72 	 * Test against 0 as that represents PAGE_REPORTING_IDLE.
73 	 */
74 	state = atomic_xchg(&prdev->state, PAGE_REPORTING_REQUESTED);
75 	if (state != PAGE_REPORTING_IDLE)
76 		return;
77 
78 	/*
79 	 * Delay the start of work to allow a sizable queue to build. For
80 	 * now we are limiting this to running no more than once every
81 	 * couple of seconds.
82 	 */
83 	schedule_delayed_work(&prdev->work, PAGE_REPORTING_DELAY);
84 }
85 
86 /* notify prdev of free page reporting request */
__page_reporting_notify(void)87 void __page_reporting_notify(void)
88 {
89 	struct page_reporting_dev_info *prdev;
90 
91 	/*
92 	 * We use RCU to protect the pr_dev_info pointer. In almost all
93 	 * cases this should be present, however in the unlikely case of
94 	 * a shutdown this will be NULL and we should exit.
95 	 */
96 	rcu_read_lock();
97 	prdev = rcu_dereference(pr_dev_info);
98 	if (likely(prdev))
99 		__page_reporting_request(prdev);
100 
101 	rcu_read_unlock();
102 }
103 
104 static void
page_reporting_drain(struct page_reporting_dev_info * prdev,struct scatterlist * sgl,unsigned int nents,bool reported)105 page_reporting_drain(struct page_reporting_dev_info *prdev,
106 		     struct scatterlist *sgl, unsigned int nents, bool reported)
107 {
108 	struct scatterlist *sg = sgl;
109 
110 	/*
111 	 * Drain the now reported pages back into their respective
112 	 * free lists/areas. We assume at least one page is populated.
113 	 */
114 	do {
115 		struct page *page = sg_page(sg);
116 		int mt = get_pageblock_migratetype(page);
117 		unsigned int order = get_order(sg->length);
118 
119 		__putback_isolated_page(page, order, mt);
120 
121 		/* If the pages were not reported due to error skip flagging */
122 		if (!reported)
123 			continue;
124 
125 		/*
126 		 * If page was not comingled with another page we can
127 		 * consider the result to be "reported" since the page
128 		 * hasn't been modified, otherwise we will need to
129 		 * report on the new larger page when we make our way
130 		 * up to that higher order.
131 		 */
132 		if (PageBuddy(page) && buddy_order(page) == order)
133 			__SetPageReported(page);
134 	} while ((sg = sg_next(sg)));
135 
136 	/* reinitialize scatterlist now that it is empty */
137 	sg_init_table(sgl, nents);
138 }
139 
140 /*
141  * The page reporting cycle consists of 4 stages, fill, report, drain, and
142  * idle. We will cycle through the first 3 stages until we cannot obtain a
143  * full scatterlist of pages, in that case we will switch to idle.
144  */
145 static int
page_reporting_cycle(struct page_reporting_dev_info * prdev,struct zone * zone,unsigned int order,unsigned int mt,struct scatterlist * sgl,unsigned int * offset)146 page_reporting_cycle(struct page_reporting_dev_info *prdev, struct zone *zone,
147 		     unsigned int order, unsigned int mt,
148 		     struct scatterlist *sgl, unsigned int *offset)
149 {
150 	struct free_area *area = &zone->free_area[order];
151 	struct list_head *list = &area->free_list[mt];
152 	unsigned int page_len = PAGE_SIZE << order;
153 	struct page *page, *next;
154 	long budget;
155 	int err = 0;
156 
157 	/*
158 	 * Perform early check, if free area is empty there is
159 	 * nothing to process so we can skip this free_list.
160 	 */
161 	if (list_empty(list))
162 		return err;
163 
164 	spin_lock_irq(&zone->lock);
165 
166 	/*
167 	 * Limit how many calls we will be making to the page reporting
168 	 * device for this list. By doing this we avoid processing any
169 	 * given list for too long.
170 	 *
171 	 * The current value used allows us enough calls to process over a
172 	 * sixteenth of the current list plus one additional call to handle
173 	 * any pages that may have already been present from the previous
174 	 * list processed. This should result in us reporting all pages on
175 	 * an idle system in about 30 seconds.
176 	 *
177 	 * The division here should be cheap since PAGE_REPORTING_CAPACITY
178 	 * should always be a power of 2.
179 	 */
180 	budget = DIV_ROUND_UP(area->nr_free, PAGE_REPORTING_CAPACITY * 16);
181 
182 	/* loop through free list adding unreported pages to sg list */
183 	list_for_each_entry_safe(page, next, list, lru) {
184 		/* We are going to skip over the reported pages. */
185 		if (PageReported(page))
186 			continue;
187 
188 		/*
189 		 * If we fully consumed our budget then update our
190 		 * state to indicate that we are requesting additional
191 		 * processing and exit this list.
192 		 */
193 		if (budget < 0) {
194 			atomic_set(&prdev->state, PAGE_REPORTING_REQUESTED);
195 			next = page;
196 			break;
197 		}
198 
199 		/* Attempt to pull page from list and place in scatterlist */
200 		if (*offset) {
201 			if (!__isolate_free_page(page, order)) {
202 				next = page;
203 				break;
204 			}
205 
206 			/* Add page to scatter list */
207 			--(*offset);
208 			sg_set_page(&sgl[*offset], page, page_len, 0);
209 
210 			continue;
211 		}
212 
213 		/*
214 		 * Make the first non-reported page in the free list
215 		 * the new head of the free list before we release the
216 		 * zone lock.
217 		 */
218 		if (!list_is_first(&page->lru, list))
219 			list_rotate_to_front(&page->lru, list);
220 
221 		/* release lock before waiting on report processing */
222 		spin_unlock_irq(&zone->lock);
223 
224 		/* begin processing pages in local list */
225 		err = prdev->report(prdev, sgl, PAGE_REPORTING_CAPACITY);
226 
227 		/* reset offset since the full list was reported */
228 		*offset = PAGE_REPORTING_CAPACITY;
229 
230 		/* update budget to reflect call to report function */
231 		budget--;
232 
233 		/* reacquire zone lock and resume processing */
234 		spin_lock_irq(&zone->lock);
235 
236 		/* flush reported pages from the sg list */
237 		page_reporting_drain(prdev, sgl, PAGE_REPORTING_CAPACITY, !err);
238 
239 		/*
240 		 * Reset next to first entry, the old next isn't valid
241 		 * since we dropped the lock to report the pages
242 		 */
243 		next = list_first_entry(list, struct page, lru);
244 
245 		/* exit on error */
246 		if (err)
247 			break;
248 	}
249 
250 	/* Rotate any leftover pages to the head of the freelist */
251 	if (!list_entry_is_head(next, list, lru) && !list_is_first(&next->lru, list))
252 		list_rotate_to_front(&next->lru, list);
253 
254 	spin_unlock_irq(&zone->lock);
255 
256 	return err;
257 }
258 
259 static int
page_reporting_process_zone(struct page_reporting_dev_info * prdev,struct scatterlist * sgl,struct zone * zone)260 page_reporting_process_zone(struct page_reporting_dev_info *prdev,
261 			    struct scatterlist *sgl, struct zone *zone)
262 {
263 	unsigned int order, mt, leftover, offset = PAGE_REPORTING_CAPACITY;
264 	unsigned long watermark;
265 	int err = 0;
266 
267 	/* Generate minimum watermark to be able to guarantee progress */
268 	watermark = low_wmark_pages(zone) +
269 		    (PAGE_REPORTING_CAPACITY << page_reporting_order);
270 
271 	/*
272 	 * Cancel request if insufficient free memory or if we failed
273 	 * to allocate page reporting statistics for the zone.
274 	 */
275 	if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
276 		return err;
277 
278 	/* Process each free list starting from lowest order/mt */
279 	for (order = page_reporting_order; order < NR_PAGE_ORDERS; order++) {
280 		for (mt = 0; mt < MIGRATE_TYPES; mt++) {
281 			/* We do not pull pages from the isolate free list */
282 			if (is_migrate_isolate(mt))
283 				continue;
284 
285 			err = page_reporting_cycle(prdev, zone, order, mt,
286 						   sgl, &offset);
287 			if (err)
288 				return err;
289 		}
290 	}
291 
292 	/* report the leftover pages before going idle */
293 	leftover = PAGE_REPORTING_CAPACITY - offset;
294 	if (leftover) {
295 		sgl = &sgl[offset];
296 		err = prdev->report(prdev, sgl, leftover);
297 
298 		/* flush any remaining pages out from the last report */
299 		spin_lock_irq(&zone->lock);
300 		page_reporting_drain(prdev, sgl, leftover, !err);
301 		spin_unlock_irq(&zone->lock);
302 	}
303 
304 	return err;
305 }
306 
page_reporting_process(struct work_struct * work)307 static void page_reporting_process(struct work_struct *work)
308 {
309 	struct delayed_work *d_work = to_delayed_work(work);
310 	struct page_reporting_dev_info *prdev =
311 		container_of(d_work, struct page_reporting_dev_info, work);
312 	int err = 0, state = PAGE_REPORTING_ACTIVE;
313 	struct scatterlist *sgl;
314 	struct zone *zone;
315 
316 	/*
317 	 * Change the state to "Active" so that we can track if there is
318 	 * anyone requests page reporting after we complete our pass. If
319 	 * the state is not altered by the end of the pass we will switch
320 	 * to idle and quit scheduling reporting runs.
321 	 */
322 	atomic_set(&prdev->state, state);
323 
324 	/* allocate scatterlist to store pages being reported on */
325 	sgl = kmalloc_array(PAGE_REPORTING_CAPACITY, sizeof(*sgl), GFP_KERNEL);
326 	if (!sgl)
327 		goto err_out;
328 
329 	sg_init_table(sgl, PAGE_REPORTING_CAPACITY);
330 
331 	for_each_zone(zone) {
332 		err = page_reporting_process_zone(prdev, sgl, zone);
333 		if (err)
334 			break;
335 	}
336 
337 	kfree(sgl);
338 err_out:
339 	/*
340 	 * If the state has reverted back to requested then there may be
341 	 * additional pages to be processed. We will defer for 2s to allow
342 	 * more pages to accumulate.
343 	 */
344 	state = atomic_cmpxchg(&prdev->state, state, PAGE_REPORTING_IDLE);
345 	if (state == PAGE_REPORTING_REQUESTED)
346 		schedule_delayed_work(&prdev->work, PAGE_REPORTING_DELAY);
347 }
348 
349 static DEFINE_MUTEX(page_reporting_mutex);
350 DEFINE_STATIC_KEY_FALSE(page_reporting_enabled);
351 
page_reporting_register(struct page_reporting_dev_info * prdev)352 int page_reporting_register(struct page_reporting_dev_info *prdev)
353 {
354 	int err = 0;
355 
356 	mutex_lock(&page_reporting_mutex);
357 
358 	/* nothing to do if already in use */
359 	if (rcu_dereference_protected(pr_dev_info,
360 				lockdep_is_held(&page_reporting_mutex))) {
361 		err = -EBUSY;
362 		goto err_out;
363 	}
364 
365 	/*
366 	 * If the page_reporting_order value is not set, we check if
367 	 * an order is provided from the driver that is performing the
368 	 * registration. If that is not provided either, we default to
369 	 * pageblock_order.
370 	 */
371 
372 	if (page_reporting_order == -1) {
373 		if (prdev->order > 0 && prdev->order <= MAX_ORDER)
374 			page_reporting_order = prdev->order;
375 		else
376 			page_reporting_order = pageblock_order;
377 	}
378 
379 	/* initialize state and work structures */
380 	atomic_set(&prdev->state, PAGE_REPORTING_IDLE);
381 	INIT_DELAYED_WORK(&prdev->work, &page_reporting_process);
382 
383 	/* Begin initial flush of zones */
384 	__page_reporting_request(prdev);
385 
386 	/* Assign device to allow notifications */
387 	rcu_assign_pointer(pr_dev_info, prdev);
388 
389 	/* enable page reporting notification */
390 	if (!static_key_enabled(&page_reporting_enabled)) {
391 		static_branch_enable(&page_reporting_enabled);
392 		pr_info("Free page reporting enabled\n");
393 	}
394 err_out:
395 	mutex_unlock(&page_reporting_mutex);
396 
397 	return err;
398 }
399 EXPORT_SYMBOL_GPL(page_reporting_register);
400 
page_reporting_unregister(struct page_reporting_dev_info * prdev)401 void page_reporting_unregister(struct page_reporting_dev_info *prdev)
402 {
403 	mutex_lock(&page_reporting_mutex);
404 
405 	if (prdev == rcu_dereference_protected(pr_dev_info,
406 				lockdep_is_held(&page_reporting_mutex))) {
407 		/* Disable page reporting notification */
408 		RCU_INIT_POINTER(pr_dev_info, NULL);
409 		synchronize_rcu();
410 
411 		/* Flush any existing work, and lock it out */
412 		cancel_delayed_work_sync(&prdev->work);
413 	}
414 
415 	mutex_unlock(&page_reporting_mutex);
416 }
417 EXPORT_SYMBOL_GPL(page_reporting_unregister);
418