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