xref: /illumos-gate/usr/src/uts/common/os/vm_pageout.c (revision 5bbb4db2c3f208d12bf0fd11769728f9e5ba66a2)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
27 /*	  All Rights Reserved  	*/
28 
29 /*
30  * University Copyright- Copyright (c) 1982, 1986, 1988
31  * The Regents of the University of California
32  * All Rights Reserved
33  *
34  * University Acknowledgment- Portions of this document are derived from
35  * software developed by the University of California, Berkeley, and its
36  * contributors.
37  */
38 
39 #pragma ident	"%Z%%M%	%I%	%E% SMI"
40 
41 #include <sys/types.h>
42 #include <sys/t_lock.h>
43 #include <sys/param.h>
44 #include <sys/buf.h>
45 #include <sys/uio.h>
46 #include <sys/proc.h>
47 #include <sys/systm.h>
48 #include <sys/mman.h>
49 #include <sys/cred.h>
50 #include <sys/vnode.h>
51 #include <sys/vm.h>
52 #include <sys/vmparam.h>
53 #include <sys/vtrace.h>
54 #include <sys/cmn_err.h>
55 #include <sys/cpuvar.h>
56 #include <sys/user.h>
57 #include <sys/kmem.h>
58 #include <sys/debug.h>
59 #include <sys/callb.h>
60 #include <sys/tnf_probe.h>
61 #include <sys/mem_cage.h>
62 #include <sys/time.h>
63 
64 #include <vm/hat.h>
65 #include <vm/as.h>
66 #include <vm/seg.h>
67 #include <vm/page.h>
68 #include <vm/pvn.h>
69 #include <vm/seg_kmem.h>
70 
71 static int checkpage(page_t *, int);
72 
73 /*
74  * The following parameters control operation of the page replacement
75  * algorithm.  They are initialized to 0, and then computed at boot time
76  * based on the size of the system.  If they are patched non-zero in
77  * a loaded vmunix they are left alone and may thus be changed per system
78  * using adb on the loaded system.
79  */
80 pgcnt_t		slowscan = 0;
81 pgcnt_t		fastscan = 0;
82 
83 static pgcnt_t	handspreadpages = 0;
84 static int	loopfraction = 2;
85 static pgcnt_t	looppages;
86 static int	min_percent_cpu = 4;
87 static int	max_percent_cpu = 80;
88 static pgcnt_t	maxfastscan = 0;
89 static pgcnt_t	maxslowscan = 100;
90 
91 pgcnt_t	maxpgio = 0;
92 pgcnt_t	minfree = 0;
93 pgcnt_t	desfree = 0;
94 pgcnt_t	lotsfree = 0;
95 pgcnt_t	needfree = 0;
96 pgcnt_t	throttlefree = 0;
97 pgcnt_t	pageout_reserve = 0;
98 
99 pgcnt_t	deficit;
100 pgcnt_t	nscan;
101 pgcnt_t	desscan;
102 
103 /*
104  * Values for min_pageout_ticks, max_pageout_ticks and pageout_ticks
105  * are the number of ticks in each wakeup cycle that gives the
106  * equivalent of some underlying %CPU duty cycle.
107  * When RATETOSCHEDPAGING is 4,  and hz is 100, pageout_scanner is
108  * awakened every 25 clock ticks.  So, converting from %CPU to ticks
109  * per wakeup cycle would be x% of 25, that is (x * 100) / 25.
110  * So, for example, 4% == 1 tick and 80% == 20 ticks.
111  *
112  * min_pageout_ticks:
113  *     ticks/wakeup equivalent of min_percent_cpu.
114  *
115  * max_pageout_ticks:
116  *     ticks/wakeup equivalent of max_percent_cpu.
117  *
118  * pageout_ticks:
119  *     Number of clock ticks budgeted for each wakeup cycle.
120  *     Computed each time around by schedpaging().
121  *     Varies between min_pageout_ticks .. max_pageout_ticks,
122  *     depending on memory pressure.
123  *
124  * pageout_lbolt:
125  *     Timestamp of the last time pageout_scanner woke up and started
126  *     (or resumed) scanning for not recently referenced pages.
127  */
128 
129 static clock_t	min_pageout_ticks;
130 static clock_t	max_pageout_ticks;
131 static clock_t	pageout_ticks;
132 static clock_t	pageout_lbolt;
133 
134 static uint_t	reset_hands;
135 
136 #define	PAGES_POLL_MASK	1023
137 
138 /*
139  * pageout_sample_lim:
140  *     The limit on the number of samples needed to establish a value
141  *     for new pageout parameters, fastscan, slowscan, and handspreadpages.
142  *
143  * pageout_sample_cnt:
144  *     Current sample number.  Once the sample gets large enough,
145  *     set new values for handspreadpages, fastscan and slowscan.
146  *
147  * pageout_sample_pages:
148  *     The accumulated number of pages scanned during sampling.
149  *
150  * pageout_sample_ticks:
151  *     The accumulated clock ticks for the sample.
152  *
153  * pageout_rate:
154  *     Rate in pages/nanosecond, computed at the end of sampling.
155  *
156  * pageout_new_spread:
157  *     The new value to use for fastscan and handspreadpages.
158  *     Calculated after enough samples have been taken.
159  */
160 
161 typedef hrtime_t hrrate_t;
162 
163 static uint64_t	pageout_sample_lim = 4;
164 static uint64_t	pageout_sample_cnt = 0;
165 static pgcnt_t	pageout_sample_pages = 0;
166 static hrrate_t	pageout_rate = 0;
167 static pgcnt_t	pageout_new_spread = 0;
168 
169 static clock_t	pageout_cycle_ticks;
170 static hrtime_t	sample_start, sample_end;
171 static hrtime_t	pageout_sample_etime = 0;
172 
173 /*
174  * Record number of times a pageout_scanner wakeup cycle finished because it
175  * timed out (exceeded its CPU budget), rather than because it visited
176  * its budgeted number of pages.
177  */
178 uint64_t pageout_timeouts = 0;
179 
180 #ifdef VM_STATS
181 static struct pageoutvmstats_str {
182 	ulong_t	checkpage[3];
183 } pageoutvmstats;
184 #endif /* VM_STATS */
185 
186 /*
187  * Threads waiting for free memory use this condition variable and lock until
188  * memory becomes available.
189  */
190 kmutex_t	memavail_lock;
191 kcondvar_t	memavail_cv;
192 
193 /*
194  * The size of the clock loop.
195  */
196 #define	LOOPPAGES	total_pages
197 
198 /*
199  * Set up the paging constants for the clock algorithm.
200  * Called after the system is initialized and the amount of memory
201  * and number of paging devices is known.
202  *
203  * lotsfree is 1/64 of memory, but at least 512K.
204  * desfree is 1/2 of lotsfree.
205  * minfree is 1/2 of desfree.
206  *
207  * Note: to revert to the paging algorithm of Solaris 2.4/2.5, set:
208  *
209  *	lotsfree = btop(512K)
210  *	desfree = btop(200K)
211  *	minfree = btop(100K)
212  *	throttlefree = INT_MIN
213  *	max_percent_cpu = 4
214  */
215 void
216 setupclock(int recalc)
217 {
218 
219 	static spgcnt_t init_lfree, init_dfree, init_mfree;
220 	static spgcnt_t init_tfree, init_preserve, init_mpgio;
221 	static spgcnt_t init_mfscan, init_fscan, init_sscan, init_hspages;
222 
223 	looppages = LOOPPAGES;
224 
225 	/*
226 	 * setupclock can now be called to recalculate the paging
227 	 * parameters in the case of dynamic addition of memory.
228 	 * So to make sure we make the proper calculations, if such a
229 	 * situation should arise, we save away the initial values
230 	 * of each parameter so we can recall them when needed. This
231 	 * way we don't lose the settings an admin might have made
232 	 * through the /etc/system file.
233 	 */
234 
235 	if (!recalc) {
236 		init_lfree = lotsfree;
237 		init_dfree = desfree;
238 		init_mfree = minfree;
239 		init_tfree = throttlefree;
240 		init_preserve = pageout_reserve;
241 		init_mpgio = maxpgio;
242 		init_mfscan = maxfastscan;
243 		init_fscan = fastscan;
244 		init_sscan = slowscan;
245 		init_hspages = handspreadpages;
246 	}
247 
248 	/*
249 	 * Set up thresholds for paging:
250 	 */
251 
252 	/*
253 	 * Lotsfree is threshold where paging daemon turns on.
254 	 */
255 	if (init_lfree == 0 || init_lfree >= looppages)
256 		lotsfree = MAX(looppages / 64, btop(512 * 1024));
257 	else
258 		lotsfree = init_lfree;
259 
260 	/*
261 	 * Desfree is amount of memory desired free.
262 	 * If less than this for extended period, start swapping.
263 	 */
264 	if (init_dfree == 0 || init_dfree >= lotsfree)
265 		desfree = lotsfree / 2;
266 	else
267 		desfree = init_dfree;
268 
269 	/*
270 	 * Minfree is minimal amount of free memory which is tolerable.
271 	 */
272 	if (init_mfree == 0 || init_mfree >= desfree)
273 		minfree = desfree / 2;
274 	else
275 		minfree = init_mfree;
276 
277 	/*
278 	 * Throttlefree is the point at which we start throttling
279 	 * PG_WAIT requests until enough memory becomes available.
280 	 */
281 	if (init_tfree == 0 || init_tfree >= desfree)
282 		throttlefree = minfree;
283 	else
284 		throttlefree = init_tfree;
285 
286 	/*
287 	 * Pageout_reserve is the number of pages that we keep in
288 	 * stock for pageout's own use.  Having a few such pages
289 	 * provides insurance against system deadlock due to
290 	 * pageout needing pages.  When freemem < pageout_reserve,
291 	 * non-blocking allocations are denied to any threads
292 	 * other than pageout and sched.  (At some point we might
293 	 * want to consider a per-thread flag like T_PUSHING_PAGES
294 	 * to indicate that a thread is part of the page-pushing
295 	 * dance (e.g. an interrupt thread) and thus is entitled
296 	 * to the same special dispensation we accord pageout.)
297 	 */
298 	if (init_preserve == 0 || init_preserve >= throttlefree)
299 		pageout_reserve = throttlefree / 2;
300 	else
301 		pageout_reserve = init_preserve;
302 
303 	/*
304 	 * Maxpgio thresholds how much paging is acceptable.
305 	 * This figures that 2/3 busy on an arm is all that is
306 	 * tolerable for paging.  We assume one operation per disk rev.
307 	 *
308 	 * XXX - Does not account for multiple swap devices.
309 	 */
310 	if (init_mpgio == 0)
311 		maxpgio = (DISKRPM * 2) / 3;
312 	else
313 		maxpgio = init_mpgio;
314 
315 	/*
316 	 * The clock scan rate varies between fastscan and slowscan
317 	 * based on the amount of free memory available.  Fastscan
318 	 * rate should be set based on the number pages that can be
319 	 * scanned per sec using ~10% of processor time.  Since this
320 	 * value depends on the processor, MMU, Mhz etc., it is
321 	 * difficult to determine it in a generic manner for all
322 	 * architectures.
323 	 *
324 	 * Instead of trying to determine the number of pages scanned
325 	 * per sec for every processor, fastscan is set to be the smaller
326 	 * of 1/2 of memory or MAXHANDSPREADPAGES and the sampling
327 	 * time is limited to ~4% of processor time.
328 	 *
329 	 * Setting fastscan to be 1/2 of memory allows pageout to scan
330 	 * all of memory in ~2 secs.  This implies that user pages not
331 	 * accessed within 1 sec (assuming, handspreadpages == fastscan)
332 	 * can be reclaimed when free memory is very low.  Stealing pages
333 	 * not accessed within 1 sec seems reasonable and ensures that
334 	 * active user processes don't thrash.
335 	 *
336 	 * Smaller values of fastscan result in scanning fewer pages
337 	 * every second and consequently pageout may not be able to free
338 	 * sufficient memory to maintain the minimum threshold.  Larger
339 	 * values of fastscan result in scanning a lot more pages which
340 	 * could lead to thrashing and higher CPU usage.
341 	 *
342 	 * Fastscan needs to be limited to a maximum value and should not
343 	 * scale with memory to prevent pageout from consuming too much
344 	 * time for scanning on slow CPU's and avoid thrashing, as a
345 	 * result of scanning too many pages, on faster CPU's.
346 	 * The value of 64 Meg was chosen for MAXHANDSPREADPAGES
347 	 * (the upper bound for fastscan) based on the average number
348 	 * of pages that can potentially be scanned in ~1 sec (using ~4%
349 	 * of the CPU) on some of the following machines that currently
350 	 * run Solaris 2.x:
351 	 *
352 	 *			average memory scanned in ~1 sec
353 	 *
354 	 *	25 Mhz SS1+:		23 Meg
355 	 *	LX:			37 Meg
356 	 *	50 Mhz SC2000:		68 Meg
357 	 *
358 	 *	40 Mhz 486:		26 Meg
359 	 *	66 Mhz 486:		42 Meg
360 	 *
361 	 * When free memory falls just below lotsfree, the scan rate
362 	 * goes from 0 to slowscan (i.e., pageout starts running).  This
363 	 * transition needs to be smooth and is achieved by ensuring that
364 	 * pageout scans a small number of pages to satisfy the transient
365 	 * memory demand.  This is set to not exceed 100 pages/sec (25 per
366 	 * wakeup) since scanning that many pages has no noticible impact
367 	 * on system performance.
368 	 *
369 	 * In addition to setting fastscan and slowscan, pageout is
370 	 * limited to using ~4% of the CPU.  This results in increasing
371 	 * the time taken to scan all of memory, which in turn means that
372 	 * user processes have a better opportunity of preventing their
373 	 * pages from being stolen.  This has a positive effect on
374 	 * interactive and overall system performance when memory demand
375 	 * is high.
376 	 *
377 	 * Thus, the rate at which pages are scanned for replacement will
378 	 * vary linearly between slowscan and the number of pages that
379 	 * can be scanned using ~4% of processor time instead of varying
380 	 * linearly between slowscan and fastscan.
381 	 *
382 	 * Also, the processor time used by pageout will vary from ~1%
383 	 * at slowscan to ~4% at fastscan instead of varying between
384 	 * ~1% at slowscan and ~10% at fastscan.
385 	 *
386 	 * The values chosen for the various VM parameters (fastscan,
387 	 * handspreadpages, etc) are not universally true for all machines,
388 	 * but appear to be a good rule of thumb for the machines we've
389 	 * tested.  They have the following ranges:
390 	 *
391 	 *	cpu speed:	20 to 70 Mhz
392 	 *	page size:	4K to 8K
393 	 *	memory size:	16M to 5G
394 	 *	page scan rate:	4000 - 17400 4K pages per sec
395 	 *
396 	 * The values need to be re-examined for machines which don't
397 	 * fall into the various ranges (e.g., slower or faster CPUs,
398 	 * smaller or larger pagesizes etc) shown above.
399 	 *
400 	 * On an MP machine, pageout is often unable to maintain the
401 	 * minimum paging thresholds under heavy load.  This is due to
402 	 * the fact that user processes running on other CPU's can be
403 	 * dirtying memory at a much faster pace than pageout can find
404 	 * pages to free.  The memory demands could be met by enabling
405 	 * more than one CPU to run the clock algorithm in such a manner
406 	 * that the various clock hands don't overlap.  This also makes
407 	 * it more difficult to determine the values for fastscan, slowscan
408 	 * and handspreadpages.
409 	 *
410 	 * The swapper is currently used to free up memory when pageout
411 	 * is unable to meet memory demands by swapping out processes.
412 	 * In addition to freeing up memory, swapping also reduces the
413 	 * demand for memory by preventing user processes from running
414 	 * and thereby consuming memory.
415 	 */
416 	if (init_mfscan == 0) {
417 		if (pageout_new_spread != 0)
418 			maxfastscan = pageout_new_spread;
419 		else
420 			maxfastscan = MAXHANDSPREADPAGES;
421 	} else {
422 		maxfastscan = init_mfscan;
423 	}
424 	if (init_fscan == 0)
425 		fastscan = MIN(looppages / loopfraction, maxfastscan);
426 	else
427 		fastscan = init_fscan;
428 	if (fastscan > looppages / loopfraction)
429 		fastscan = looppages / loopfraction;
430 
431 	/*
432 	 * Set slow scan time to 1/10 the fast scan time, but
433 	 * not to exceed maxslowscan.
434 	 */
435 	if (init_sscan == 0)
436 		slowscan = MIN(fastscan / 10, maxslowscan);
437 	else
438 		slowscan = init_sscan;
439 	if (slowscan > fastscan / 2)
440 		slowscan = fastscan / 2;
441 
442 	/*
443 	 * Handspreadpages is distance (in pages) between front and back
444 	 * pageout daemon hands.  The amount of time to reclaim a page
445 	 * once pageout examines it increases with this distance and
446 	 * decreases as the scan rate rises. It must be < the amount
447 	 * of pageable memory.
448 	 *
449 	 * Since pageout is limited to ~4% of the CPU, setting handspreadpages
450 	 * to be "fastscan" results in the front hand being a few secs
451 	 * (varies based on the processor speed) ahead of the back hand
452 	 * at fastscan rates.  This distance can be further reduced, if
453 	 * necessary, by increasing the processor time used by pageout
454 	 * to be more than ~4% and preferrably not more than ~10%.
455 	 *
456 	 * As a result, user processes have a much better chance of
457 	 * referencing their pages before the back hand examines them.
458 	 * This also significantly lowers the number of reclaims from
459 	 * the freelist since pageout does not end up freeing pages which
460 	 * may be referenced a sec later.
461 	 */
462 	if (init_hspages == 0)
463 		handspreadpages = fastscan;
464 	else
465 		handspreadpages = init_hspages;
466 
467 	/*
468 	 * Make sure that back hand follows front hand by at least
469 	 * 1/RATETOSCHEDPAGING seconds.  Without this test, it is possible
470 	 * for the back hand to look at a page during the same wakeup of
471 	 * the pageout daemon in which the front hand cleared its ref bit.
472 	 */
473 	if (handspreadpages >= looppages)
474 		handspreadpages = looppages - 1;
475 
476 	/*
477 	 * If we have been called to recalculate the parameters,
478 	 * set a flag to re-evaluate the clock hand pointers.
479 	 */
480 	if (recalc)
481 		reset_hands = 1;
482 }
483 
484 /*
485  * Pageout scheduling.
486  *
487  * Schedpaging controls the rate at which the page out daemon runs by
488  * setting the global variables nscan and desscan RATETOSCHEDPAGING
489  * times a second.  Nscan records the number of pages pageout has examined
490  * in its current pass; schedpaging resets this value to zero each time
491  * it runs.  Desscan records the number of pages pageout should examine
492  * in its next pass; schedpaging sets this value based on the amount of
493  * currently available memory.
494  */
495 
496 #define	RATETOSCHEDPAGING	4		/* hz that is */
497 
498 static kmutex_t	pageout_mutex;	/* held while pageout or schedpaging running */
499 
500 /*
501  * Pool of available async pageout putpage requests.
502  */
503 static struct async_reqs *push_req;
504 static struct async_reqs *req_freelist;	/* available req structs */
505 static struct async_reqs *push_list;	/* pending reqs */
506 static kmutex_t push_lock;		/* protects req pool */
507 static kcondvar_t push_cv;
508 
509 static int async_list_size = 256;	/* number of async request structs */
510 
511 static void pageout_scanner(void);
512 
513 /*
514  * If a page is being shared more than "po_share" times
515  * then leave it alone- don't page it out.
516  */
517 #define	MIN_PO_SHARE	(8)
518 #define	MAX_PO_SHARE	((MIN_PO_SHARE) << 24)
519 ulong_t	po_share = MIN_PO_SHARE;
520 
521 /*
522  * Schedule rate for paging.
523  * Rate is linear interpolation between
524  * slowscan with lotsfree and fastscan when out of memory.
525  */
526 static void
527 schedpaging(void *arg)
528 {
529 	spgcnt_t vavail;
530 
531 	if (freemem < lotsfree + needfree + kmem_reapahead)
532 		kmem_reap();
533 
534 	if (freemem < lotsfree + needfree)
535 		seg_preap();
536 
537 	if (kcage_on && (kcage_freemem < kcage_desfree || kcage_needfree))
538 		kcage_cageout_wakeup();
539 
540 	if (mutex_tryenter(&pageout_mutex)) {
541 		/* pageout() not running */
542 		nscan = 0;
543 		vavail = freemem - deficit;
544 		if (pageout_new_spread != 0)
545 			vavail -= needfree;
546 		if (vavail < 0)
547 			vavail = 0;
548 		if (vavail > lotsfree)
549 			vavail = lotsfree;
550 
551 		/*
552 		 * Fix for 1161438 (CRS SPR# 73922).  All variables
553 		 * in the original calculation for desscan were 32 bit signed
554 		 * ints.  As freemem approaches 0x0 on a system with 1 Gig or
555 		 * more of memory, the calculation can overflow.  When this
556 		 * happens, desscan becomes negative and pageout_scanner()
557 		 * stops paging out.
558 		 */
559 		if ((needfree) && (pageout_new_spread == 0)) {
560 			/*
561 			 * If we've not yet collected enough samples to
562 			 * calculate a spread, use the old logic of kicking
563 			 * into high gear anytime needfree is non-zero.
564 			 */
565 			desscan = fastscan / RATETOSCHEDPAGING;
566 		} else {
567 			/*
568 			 * Once we've calculated a spread based on system
569 			 * memory and usage, just treat needfree as another
570 			 * form of deficit.
571 			 */
572 			spgcnt_t faststmp, slowstmp, result;
573 
574 			slowstmp = slowscan * vavail;
575 			faststmp = fastscan * (lotsfree - vavail);
576 			result = (slowstmp + faststmp) /
577 			    nz(lotsfree) / RATETOSCHEDPAGING;
578 			desscan = (pgcnt_t)result;
579 		}
580 
581 		pageout_ticks = min_pageout_ticks + (lotsfree - vavail) *
582 		    (max_pageout_ticks - min_pageout_ticks) / nz(lotsfree);
583 
584 		if (freemem < lotsfree + needfree ||
585 		    pageout_sample_cnt < pageout_sample_lim) {
586 			TRACE_1(TR_FAC_VM, TR_PAGEOUT_CV_SIGNAL,
587 			    "pageout_cv_signal:freemem %ld", freemem);
588 			cv_signal(&proc_pageout->p_cv);
589 		} else {
590 			/*
591 			 * There are enough free pages, no need to
592 			 * kick the scanner thread.  And next time
593 			 * around, keep more of the `highly shared'
594 			 * pages.
595 			 */
596 			cv_signal_pageout();
597 			if (po_share > MIN_PO_SHARE) {
598 				po_share >>= 1;
599 			}
600 		}
601 		mutex_exit(&pageout_mutex);
602 	}
603 
604 	/*
605 	 * Signal threads waiting for available memory.
606 	 * NOTE: usually we need to grab memavail_lock before cv_broadcast, but
607 	 * in this case it is not needed - the waiters will be waken up during
608 	 * the next invocation of this function.
609 	 */
610 	if (kmem_avail() > 0)
611 		cv_broadcast(&memavail_cv);
612 
613 	(void) timeout(schedpaging, arg, hz / RATETOSCHEDPAGING);
614 }
615 
616 pgcnt_t		pushes;
617 ulong_t		push_list_size;		/* # of requests on pageout queue */
618 
619 #define	FRONT	1
620 #define	BACK	2
621 
622 int dopageout = 1;	/* must be non-zero to turn page stealing on */
623 
624 /*
625  * The page out daemon, which runs as process 2.
626  *
627  * As long as there are at least lotsfree pages,
628  * this process is not run.  When the number of free
629  * pages stays in the range desfree to lotsfree,
630  * this daemon runs through the pages in the loop
631  * at a rate determined in schedpaging().  Pageout manages
632  * two hands on the clock.  The front hand moves through
633  * memory, clearing the reference bit,
634  * and stealing pages from procs that are over maxrss.
635  * The back hand travels a distance behind the front hand,
636  * freeing the pages that have not been referenced in the time
637  * since the front hand passed.  If modified, they are pushed to
638  * swap before being freed.
639  *
640  * There are 2 threads that act on behalf of the pageout process.
641  * One thread scans pages (pageout_scanner) and frees them up if
642  * they don't require any VOP_PUTPAGE operation. If a page must be
643  * written back to its backing store, the request is put on a list
644  * and the other (pageout) thread is signaled. The pageout thread
645  * grabs VOP_PUTPAGE requests from the list, and processes them.
646  * Some filesystems may require resources for the VOP_PUTPAGE
647  * operations (like memory) and hence can block the pageout
648  * thread, but the scanner thread can still operate. There is still
649  * no guarantee that memory deadlocks cannot occur.
650  *
651  * For now, this thing is in very rough form.
652  */
653 void
654 pageout()
655 {
656 	struct async_reqs *arg;
657 	pri_t pageout_pri;
658 	int i;
659 	pgcnt_t max_pushes;
660 	callb_cpr_t cprinfo;
661 
662 	proc_pageout = ttoproc(curthread);
663 	proc_pageout->p_cstime = 0;
664 	proc_pageout->p_stime =  0;
665 	proc_pageout->p_cutime =  0;
666 	proc_pageout->p_utime = 0;
667 	bcopy("pageout", PTOU(curproc)->u_psargs, 8);
668 	bcopy("pageout", PTOU(curproc)->u_comm, 7);
669 
670 	/*
671 	 * Create pageout scanner thread
672 	 */
673 	mutex_init(&pageout_mutex, NULL, MUTEX_DEFAULT, NULL);
674 	mutex_init(&push_lock, NULL, MUTEX_DEFAULT, NULL);
675 
676 	/*
677 	 * Allocate and initialize the async request structures
678 	 * for pageout.
679 	 */
680 	push_req = (struct async_reqs *)
681 	    kmem_zalloc(async_list_size * sizeof (struct async_reqs), KM_SLEEP);
682 
683 	req_freelist = push_req;
684 	for (i = 0; i < async_list_size - 1; i++)
685 		push_req[i].a_next = &push_req[i + 1];
686 
687 	pageout_pri = curthread->t_pri;
688 	pageout_init(pageout_scanner, proc_pageout, pageout_pri - 1);
689 
690 	/*
691 	 * kick off pageout scheduler.
692 	 */
693 	schedpaging(NULL);
694 
695 	/*
696 	 * Create kernel cage thread.
697 	 * The kernel cage thread is started under the pageout process
698 	 * to take advantage of the less restricted page allocation
699 	 * in page_create_throttle().
700 	 */
701 	kcage_cageout_init();
702 
703 	/*
704 	 * Limit pushes to avoid saturating pageout devices.
705 	 */
706 	max_pushes = maxpgio / RATETOSCHEDPAGING;
707 	CALLB_CPR_INIT(&cprinfo, &push_lock, callb_generic_cpr, "pageout");
708 
709 	for (;;) {
710 		mutex_enter(&push_lock);
711 
712 		while ((arg = push_list) == NULL || pushes > max_pushes) {
713 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
714 			cv_wait(&push_cv, &push_lock);
715 			pushes = 0;
716 			CALLB_CPR_SAFE_END(&cprinfo, &push_lock);
717 		}
718 		push_list = arg->a_next;
719 		arg->a_next = NULL;
720 		mutex_exit(&push_lock);
721 
722 		if (VOP_PUTPAGE(arg->a_vp, (offset_t)arg->a_off,
723 		    arg->a_len, arg->a_flags, arg->a_cred, NULL) == 0) {
724 			pushes++;
725 		}
726 
727 		/* vp held by checkpage() */
728 		VN_RELE(arg->a_vp);
729 
730 		mutex_enter(&push_lock);
731 		arg->a_next = req_freelist;	/* back on freelist */
732 		req_freelist = arg;
733 		push_list_size--;
734 		mutex_exit(&push_lock);
735 	}
736 }
737 
738 /*
739  * Kernel thread that scans pages looking for ones to free
740  */
741 static void
742 pageout_scanner(void)
743 {
744 	struct page *fronthand, *backhand;
745 	uint_t count;
746 	callb_cpr_t cprinfo;
747 	pgcnt_t	nscan_limit;
748 	pgcnt_t	pcount;
749 
750 	CALLB_CPR_INIT(&cprinfo, &pageout_mutex, callb_generic_cpr, "poscan");
751 	mutex_enter(&pageout_mutex);
752 
753 	/*
754 	 * The restart case does not attempt to point the hands at roughly
755 	 * the right point on the assumption that after one circuit things
756 	 * will have settled down - and restarts shouldn't be that often.
757 	 */
758 
759 	/*
760 	 * Set the two clock hands to be separated by a reasonable amount,
761 	 * but no more than 360 degrees apart.
762 	 */
763 	backhand = page_first();
764 	if (handspreadpages >= total_pages)
765 		fronthand = page_nextn(backhand, total_pages - 1);
766 	else
767 		fronthand = page_nextn(backhand, handspreadpages);
768 
769 	min_pageout_ticks = MAX(1,
770 	    ((hz * min_percent_cpu) / 100) / RATETOSCHEDPAGING);
771 	max_pageout_ticks = MAX(min_pageout_ticks,
772 	    ((hz * max_percent_cpu) / 100) / RATETOSCHEDPAGING);
773 
774 loop:
775 	cv_signal_pageout();
776 
777 	CALLB_CPR_SAFE_BEGIN(&cprinfo);
778 	cv_wait(&proc_pageout->p_cv, &pageout_mutex);
779 	CALLB_CPR_SAFE_END(&cprinfo, &pageout_mutex);
780 
781 	if (!dopageout)
782 		goto loop;
783 
784 	if (reset_hands) {
785 		reset_hands = 0;
786 
787 		backhand = page_first();
788 		if (handspreadpages >= total_pages)
789 			fronthand = page_nextn(backhand, total_pages - 1);
790 		else
791 			fronthand = page_nextn(backhand, handspreadpages);
792 	}
793 
794 	CPU_STATS_ADDQ(CPU, vm, pgrrun, 1);
795 	count = 0;
796 
797 	TRACE_4(TR_FAC_VM, TR_PAGEOUT_START,
798 	    "pageout_start:freemem %ld lotsfree %ld nscan %ld desscan %ld",
799 	    freemem, lotsfree, nscan, desscan);
800 
801 	/* Kernel probe */
802 	TNF_PROBE_2(pageout_scan_start, "vm pagedaemon", /* CSTYLED */,
803 	    tnf_ulong, pages_free, freemem, tnf_ulong, pages_needed, needfree);
804 
805 	pcount = 0;
806 	if (pageout_sample_cnt < pageout_sample_lim) {
807 		nscan_limit = total_pages;
808 	} else {
809 		nscan_limit = desscan;
810 	}
811 	pageout_lbolt = lbolt;
812 	sample_start = gethrtime();
813 
814 	/*
815 	 * Scan the appropriate number of pages for a single duty cycle.
816 	 * However, stop scanning as soon as there is enough free memory.
817 	 * For a short while, we will be sampling the performance of the
818 	 * scanner and need to keep running just to get sample data, in
819 	 * which case we keep going and don't pay attention to whether
820 	 * or not there is enough free memory.
821 	 */
822 
823 	while (nscan < nscan_limit && (freemem < lotsfree + needfree ||
824 	    pageout_sample_cnt < pageout_sample_lim)) {
825 		int rvfront, rvback;
826 
827 		/*
828 		 * Check to see if we have exceeded our %CPU budget
829 		 * for this wakeup, but not on every single page visited,
830 		 * just every once in a while.
831 		 */
832 		if ((pcount & PAGES_POLL_MASK) == PAGES_POLL_MASK) {
833 			pageout_cycle_ticks = lbolt - pageout_lbolt;
834 			if (pageout_cycle_ticks >= pageout_ticks) {
835 				++pageout_timeouts;
836 				break;
837 			}
838 		}
839 
840 		/*
841 		 * If checkpage manages to add a page to the free list,
842 		 * we give ourselves another couple of trips around the loop.
843 		 */
844 		if ((rvfront = checkpage(fronthand, FRONT)) == 1)
845 			count = 0;
846 		if ((rvback = checkpage(backhand, BACK)) == 1)
847 			count = 0;
848 
849 		++pcount;
850 
851 		/*
852 		 * protected by pageout_mutex instead of cpu_stat_lock
853 		 */
854 		CPU_STATS_ADDQ(CPU, vm, scan, 1);
855 
856 		/*
857 		 * Don't include ineligible pages in the number scanned.
858 		 */
859 		if (rvfront != -1 || rvback != -1)
860 			nscan++;
861 
862 		backhand = page_next(backhand);
863 
864 		/*
865 		 * backhand update and wraparound check are done separately
866 		 * because lint barks when it finds an empty "if" body
867 		 */
868 
869 		if ((fronthand = page_next(fronthand)) == page_first())	{
870 			TRACE_2(TR_FAC_VM, TR_PAGEOUT_HAND_WRAP,
871 			    "pageout_hand_wrap:freemem %ld whichhand %d",
872 			    freemem, FRONT);
873 
874 			/*
875 			 * protected by pageout_mutex instead of cpu_stat_lock
876 			 */
877 			CPU_STATS_ADDQ(CPU, vm, rev, 1);
878 			if (++count > 1) {
879 				/*
880 				 * Extremely unlikely, but it happens.
881 				 * We went around the loop at least once
882 				 * and didn't get far enough.
883 				 * If we are still skipping `highly shared'
884 				 * pages, skip fewer of them.  Otherwise,
885 				 * give up till the next clock tick.
886 				 */
887 				if (po_share < MAX_PO_SHARE) {
888 					po_share <<= 1;
889 				} else {
890 					/*
891 					 * Really a "goto loop", but
892 					 * if someone is TRACing or
893 					 * TNF_PROBE_ing, at least
894 					 * make records to show
895 					 * where we are.
896 					 */
897 					break;
898 				}
899 			}
900 		}
901 	}
902 
903 	sample_end = gethrtime();
904 
905 	TRACE_5(TR_FAC_VM, TR_PAGEOUT_END,
906 	    "pageout_end:freemem %ld lots %ld nscan %ld des %ld count %u",
907 	    freemem, lotsfree, nscan, desscan, count);
908 
909 	/* Kernel probe */
910 	TNF_PROBE_2(pageout_scan_end, "vm pagedaemon", /* CSTYLED */,
911 	    tnf_ulong, pages_scanned, nscan, tnf_ulong, pages_free, freemem);
912 
913 	if (pageout_sample_cnt < pageout_sample_lim) {
914 		pageout_sample_pages += pcount;
915 		pageout_sample_etime += sample_end - sample_start;
916 		++pageout_sample_cnt;
917 	}
918 	if (pageout_sample_cnt >= pageout_sample_lim &&
919 	    pageout_new_spread == 0) {
920 		pageout_rate = (hrrate_t)pageout_sample_pages *
921 		    (hrrate_t)(NANOSEC) / pageout_sample_etime;
922 		pageout_new_spread = pageout_rate / 10;
923 		setupclock(1);
924 	}
925 
926 	goto loop;
927 }
928 
929 /*
930  * Look at the page at hand.  If it is locked (e.g., for physical i/o),
931  * system (u., page table) or free, then leave it alone.  Otherwise,
932  * if we are running the front hand, turn off the page's reference bit.
933  * If the proc is over maxrss, we take it.  If running the back hand,
934  * check whether the page has been reclaimed.  If not, free the page,
935  * pushing it to disk first if necessary.
936  *
937  * Return values:
938  *	-1 if the page is not a candidate at all,
939  *	 0 if not freed, or
940  *	 1 if we freed it.
941  */
942 static int
943 checkpage(struct page *pp, int whichhand)
944 {
945 	int ppattr;
946 	int isfs = 0;
947 	int isexec = 0;
948 	int pagesync_flag;
949 
950 	/*
951 	 * Skip pages:
952 	 * 	- associated with the kernel vnode since
953 	 *	    they are always "exclusively" locked.
954 	 *	- that are free
955 	 *	- that are shared more than po_share'd times
956 	 *	- its already locked
957 	 *
958 	 * NOTE:  These optimizations assume that reads are atomic.
959 	 */
960 
961 	if (PP_ISKAS(pp) || PAGE_LOCKED(pp) || PP_ISFREE(pp) ||
962 	    pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
963 	    hat_page_checkshare(pp, po_share)) {
964 		return (-1);
965 	}
966 
967 	if (!page_trylock(pp, SE_EXCL)) {
968 		/*
969 		 * Skip the page if we can't acquire the "exclusive" lock.
970 		 */
971 		return (-1);
972 	} else if (PP_ISFREE(pp)) {
973 		/*
974 		 * It became free between the above check and our actually
975 		 * locking the page.  Oh, well there will be other pages.
976 		 */
977 		page_unlock(pp);
978 		return (-1);
979 	}
980 
981 	/*
982 	 * Reject pages that cannot be freed. The page_struct_lock
983 	 * need not be acquired to examine these
984 	 * fields since the page has an "exclusive" lock.
985 	 */
986 	if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
987 		page_unlock(pp);
988 		return (-1);
989 	}
990 
991 	/*
992 	 * Maintain statistics for what we are freeing
993 	 */
994 
995 	if (pp->p_vnode != NULL) {
996 		if (pp->p_vnode->v_flag & VVMEXEC)
997 			isexec = 1;
998 
999 		if (!IS_SWAPFSVP(pp->p_vnode))
1000 			isfs = 1;
1001 	}
1002 
1003 	/*
1004 	 * Turn off REF and MOD bits with the front hand.
1005 	 * The back hand examines the REF bit and always considers
1006 	 * SHARED pages as referenced.
1007 	 */
1008 	if (whichhand == FRONT)
1009 		pagesync_flag = HAT_SYNC_ZERORM;
1010 	else
1011 		pagesync_flag = HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_REF |
1012 		    HAT_SYNC_STOPON_SHARED;
1013 
1014 	ppattr = hat_pagesync(pp, pagesync_flag);
1015 
1016 recheck:
1017 	/*
1018 	 * If page is referenced; make unreferenced but reclaimable.
1019 	 * If this page is not referenced, then it must be reclaimable
1020 	 * and we can add it to the free list.
1021 	 */
1022 	if (ppattr & P_REF) {
1023 		TRACE_2(TR_FAC_VM, TR_PAGEOUT_ISREF,
1024 		    "pageout_isref:pp %p whichhand %d", pp, whichhand);
1025 		if (whichhand == FRONT) {
1026 			/*
1027 			 * Checking of rss or madvise flags needed here...
1028 			 *
1029 			 * If not "well-behaved", fall through into the code
1030 			 * for not referenced.
1031 			 */
1032 			hat_clrref(pp);
1033 		}
1034 		/*
1035 		 * Somebody referenced the page since the front
1036 		 * hand went by, so it's not a candidate for
1037 		 * freeing up.
1038 		 */
1039 		page_unlock(pp);
1040 		return (0);
1041 	}
1042 
1043 	VM_STAT_ADD(pageoutvmstats.checkpage[0]);
1044 
1045 	/*
1046 	 * If large page, attempt to demote it. If successfully demoted,
1047 	 * retry the checkpage.
1048 	 */
1049 	if (pp->p_szc != 0) {
1050 		if (!page_try_demote_pages(pp)) {
1051 			VM_STAT_ADD(pageoutvmstats.checkpage[1]);
1052 			page_unlock(pp);
1053 			return (-1);
1054 		}
1055 		ASSERT(pp->p_szc == 0);
1056 		VM_STAT_ADD(pageoutvmstats.checkpage[2]);
1057 		/*
1058 		 * since page_try_demote_pages() could have unloaded some
1059 		 * mappings it makes sense to reload ppattr.
1060 		 */
1061 		ppattr = hat_page_getattr(pp, P_MOD | P_REF);
1062 	}
1063 
1064 	/*
1065 	 * If the page is currently dirty, we have to arrange
1066 	 * to have it cleaned before it can be freed.
1067 	 *
1068 	 * XXX - ASSERT(pp->p_vnode != NULL);
1069 	 */
1070 	if ((ppattr & P_MOD) && pp->p_vnode) {
1071 		struct vnode *vp = pp->p_vnode;
1072 		u_offset_t offset = pp->p_offset;
1073 
1074 		/*
1075 		 * XXX - Test for process being swapped out or about to exit?
1076 		 * [Can't get back to process(es) using the page.]
1077 		 */
1078 
1079 		/*
1080 		 * Hold the vnode before releasing the page lock to
1081 		 * prevent it from being freed and re-used by some
1082 		 * other thread.
1083 		 */
1084 		VN_HOLD(vp);
1085 		page_unlock(pp);
1086 
1087 		/*
1088 		 * Queue i/o request for the pageout thread.
1089 		 */
1090 		if (!queue_io_request(vp, offset)) {
1091 			VN_RELE(vp);
1092 			return (0);
1093 		}
1094 		return (1);
1095 	}
1096 
1097 	/*
1098 	 * Now we unload all the translations,
1099 	 * and put the page back on to the free list.
1100 	 * If the page was used (referenced or modified) after
1101 	 * the pagesync but before it was unloaded we catch it
1102 	 * and handle the page properly.
1103 	 */
1104 	TRACE_2(TR_FAC_VM, TR_PAGEOUT_FREE,
1105 	    "pageout_free:pp %p whichhand %d", pp, whichhand);
1106 	(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
1107 	ppattr = hat_page_getattr(pp, P_MOD | P_REF);
1108 	if ((ppattr & P_REF) || ((ppattr & P_MOD) && pp->p_vnode))
1109 		goto recheck;
1110 
1111 	/*LINTED: constant in conditional context*/
1112 	VN_DISPOSE(pp, B_FREE, 0, kcred);
1113 
1114 	CPU_STATS_ADD_K(vm, dfree, 1);
1115 
1116 	if (isfs) {
1117 		if (isexec) {
1118 			CPU_STATS_ADD_K(vm, execfree, 1);
1119 		} else {
1120 			CPU_STATS_ADD_K(vm, fsfree, 1);
1121 		}
1122 	} else {
1123 		CPU_STATS_ADD_K(vm, anonfree, 1);
1124 	}
1125 
1126 	return (1);		/* freed a page! */
1127 }
1128 
1129 /*
1130  * Queue async i/o request from pageout_scanner and segment swapout
1131  * routines on one common list.  This ensures that pageout devices (swap)
1132  * are not saturated by pageout_scanner or swapout requests.
1133  * The pageout thread empties this list by initiating i/o operations.
1134  */
1135 int
1136 queue_io_request(vnode_t *vp, u_offset_t off)
1137 {
1138 	struct async_reqs *arg;
1139 
1140 	/*
1141 	 * If we cannot allocate an async request struct,
1142 	 * skip this page.
1143 	 */
1144 	mutex_enter(&push_lock);
1145 	if ((arg = req_freelist) == NULL) {
1146 		mutex_exit(&push_lock);
1147 		return (0);
1148 	}
1149 	req_freelist = arg->a_next;		/* adjust freelist */
1150 	push_list_size++;
1151 
1152 	arg->a_vp = vp;
1153 	arg->a_off = off;
1154 	arg->a_len = PAGESIZE;
1155 	arg->a_flags = B_ASYNC | B_FREE;
1156 	arg->a_cred = kcred;		/* always held */
1157 
1158 	/*
1159 	 * Add to list of pending write requests.
1160 	 */
1161 	arg->a_next = push_list;
1162 	push_list = arg;
1163 
1164 	if (req_freelist == NULL) {
1165 		/*
1166 		 * No free async requests left. The lock is held so we
1167 		 * might as well signal the pusher thread now.
1168 		 */
1169 		cv_signal(&push_cv);
1170 	}
1171 	mutex_exit(&push_lock);
1172 	return (1);
1173 }
1174 
1175 /*
1176  * Wakeup pageout to initiate i/o if push_list is not empty.
1177  */
1178 void
1179 cv_signal_pageout()
1180 {
1181 	if (push_list != NULL) {
1182 		mutex_enter(&push_lock);
1183 		cv_signal(&push_cv);
1184 		mutex_exit(&push_lock);
1185 	}
1186 }
1187