xref: /titanic_41/usr/src/uts/common/os/vm_pageout.c (revision 35a5a3587fd94b666239c157d3722745250ccbd7)
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 2009 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 #include <sys/types.h>
40 #include <sys/t_lock.h>
41 #include <sys/param.h>
42 #include <sys/buf.h>
43 #include <sys/uio.h>
44 #include <sys/proc.h>
45 #include <sys/systm.h>
46 #include <sys/mman.h>
47 #include <sys/cred.h>
48 #include <sys/vnode.h>
49 #include <sys/vm.h>
50 #include <sys/vmparam.h>
51 #include <sys/vtrace.h>
52 #include <sys/cmn_err.h>
53 #include <sys/cpuvar.h>
54 #include <sys/user.h>
55 #include <sys/kmem.h>
56 #include <sys/debug.h>
57 #include <sys/callb.h>
58 #include <sys/tnf_probe.h>
59 #include <sys/mem_cage.h>
60 #include <sys/time.h>
61 
62 #include <vm/hat.h>
63 #include <vm/as.h>
64 #include <vm/seg.h>
65 #include <vm/page.h>
66 #include <vm/pvn.h>
67 #include <vm/seg_kmem.h>
68 
69 static int checkpage(page_t *, int);
70 
71 /*
72  * The following parameters control operation of the page replacement
73  * algorithm.  They are initialized to 0, and then computed at boot time
74  * based on the size of the system.  If they are patched non-zero in
75  * a loaded vmunix they are left alone and may thus be changed per system
76  * using adb on the loaded system.
77  */
78 pgcnt_t		slowscan = 0;
79 pgcnt_t		fastscan = 0;
80 
81 static pgcnt_t	handspreadpages = 0;
82 static int	loopfraction = 2;
83 static pgcnt_t	looppages;
84 static int	min_percent_cpu = 4;
85 static int	max_percent_cpu = 80;
86 static pgcnt_t	maxfastscan = 0;
87 static pgcnt_t	maxslowscan = 100;
88 
89 pgcnt_t	maxpgio = 0;
90 pgcnt_t	minfree = 0;
91 pgcnt_t	desfree = 0;
92 pgcnt_t	lotsfree = 0;
93 pgcnt_t	needfree = 0;
94 pgcnt_t	throttlefree = 0;
95 pgcnt_t	pageout_reserve = 0;
96 
97 pgcnt_t	deficit;
98 pgcnt_t	nscan;
99 pgcnt_t	desscan;
100 
101 /*
102  * Values for min_pageout_ticks, max_pageout_ticks and pageout_ticks
103  * are the number of ticks in each wakeup cycle that gives the
104  * equivalent of some underlying %CPU duty cycle.
105  * When RATETOSCHEDPAGING is 4,  and hz is 100, pageout_scanner is
106  * awakened every 25 clock ticks.  So, converting from %CPU to ticks
107  * per wakeup cycle would be x% of 25, that is (x * 100) / 25.
108  * So, for example, 4% == 1 tick and 80% == 20 ticks.
109  *
110  * min_pageout_ticks:
111  *     ticks/wakeup equivalent of min_percent_cpu.
112  *
113  * max_pageout_ticks:
114  *     ticks/wakeup equivalent of max_percent_cpu.
115  *
116  * pageout_ticks:
117  *     Number of clock ticks budgeted for each wakeup cycle.
118  *     Computed each time around by schedpaging().
119  *     Varies between min_pageout_ticks .. max_pageout_ticks,
120  *     depending on memory pressure.
121  *
122  * pageout_lbolt:
123  *     Timestamp of the last time pageout_scanner woke up and started
124  *     (or resumed) scanning for not recently referenced pages.
125  */
126 
127 static clock_t	min_pageout_ticks;
128 static clock_t	max_pageout_ticks;
129 static clock_t	pageout_ticks;
130 static clock_t	pageout_lbolt;
131 
132 static uint_t	reset_hands;
133 
134 #define	PAGES_POLL_MASK	1023
135 
136 /*
137  * pageout_sample_lim:
138  *     The limit on the number of samples needed to establish a value
139  *     for new pageout parameters, fastscan, slowscan, and handspreadpages.
140  *
141  * pageout_sample_cnt:
142  *     Current sample number.  Once the sample gets large enough,
143  *     set new values for handspreadpages, fastscan and slowscan.
144  *
145  * pageout_sample_pages:
146  *     The accumulated number of pages scanned during sampling.
147  *
148  * pageout_sample_ticks:
149  *     The accumulated clock ticks for the sample.
150  *
151  * pageout_rate:
152  *     Rate in pages/nanosecond, computed at the end of sampling.
153  *
154  * pageout_new_spread:
155  *     The new value to use for fastscan and handspreadpages.
156  *     Calculated after enough samples have been taken.
157  */
158 
159 typedef hrtime_t hrrate_t;
160 
161 static uint64_t	pageout_sample_lim = 4;
162 static uint64_t	pageout_sample_cnt = 0;
163 static pgcnt_t	pageout_sample_pages = 0;
164 static hrrate_t	pageout_rate = 0;
165 static pgcnt_t	pageout_new_spread = 0;
166 
167 static clock_t	pageout_cycle_ticks;
168 static hrtime_t	sample_start, sample_end;
169 static hrtime_t	pageout_sample_etime = 0;
170 
171 /*
172  * Record number of times a pageout_scanner wakeup cycle finished because it
173  * timed out (exceeded its CPU budget), rather than because it visited
174  * its budgeted number of pages.
175  */
176 uint64_t pageout_timeouts = 0;
177 
178 #ifdef VM_STATS
179 static struct pageoutvmstats_str {
180 	ulong_t	checkpage[3];
181 } pageoutvmstats;
182 #endif /* VM_STATS */
183 
184 /*
185  * Threads waiting for free memory use this condition variable and lock until
186  * memory becomes available.
187  */
188 kmutex_t	memavail_lock;
189 kcondvar_t	memavail_cv;
190 
191 /*
192  * The size of the clock loop.
193  */
194 #define	LOOPPAGES	total_pages
195 
196 /*
197  * Set up the paging constants for the clock algorithm.
198  * Called after the system is initialized and the amount of memory
199  * and number of paging devices is known.
200  *
201  * lotsfree is 1/64 of memory, but at least 512K.
202  * desfree is 1/2 of lotsfree.
203  * minfree is 1/2 of desfree.
204  *
205  * Note: to revert to the paging algorithm of Solaris 2.4/2.5, set:
206  *
207  *	lotsfree = btop(512K)
208  *	desfree = btop(200K)
209  *	minfree = btop(100K)
210  *	throttlefree = INT_MIN
211  *	max_percent_cpu = 4
212  */
213 void
setupclock(int recalc)214 setupclock(int recalc)
215 {
216 
217 	static spgcnt_t init_lfree, init_dfree, init_mfree;
218 	static spgcnt_t init_tfree, init_preserve, init_mpgio;
219 	static spgcnt_t init_mfscan, init_fscan, init_sscan, init_hspages;
220 
221 	looppages = LOOPPAGES;
222 
223 	/*
224 	 * setupclock can now be called to recalculate the paging
225 	 * parameters in the case of dynamic addition of memory.
226 	 * So to make sure we make the proper calculations, if such a
227 	 * situation should arise, we save away the initial values
228 	 * of each parameter so we can recall them when needed. This
229 	 * way we don't lose the settings an admin might have made
230 	 * through the /etc/system file.
231 	 */
232 
233 	if (!recalc) {
234 		init_lfree = lotsfree;
235 		init_dfree = desfree;
236 		init_mfree = minfree;
237 		init_tfree = throttlefree;
238 		init_preserve = pageout_reserve;
239 		init_mpgio = maxpgio;
240 		init_mfscan = maxfastscan;
241 		init_fscan = fastscan;
242 		init_sscan = slowscan;
243 		init_hspages = handspreadpages;
244 	}
245 
246 	/*
247 	 * Set up thresholds for paging:
248 	 */
249 
250 	/*
251 	 * Lotsfree is threshold where paging daemon turns on.
252 	 */
253 	if (init_lfree == 0 || init_lfree >= looppages)
254 		lotsfree = MAX(looppages / 64, btop(512 * 1024));
255 	else
256 		lotsfree = init_lfree;
257 
258 	/*
259 	 * Desfree is amount of memory desired free.
260 	 * If less than this for extended period, start swapping.
261 	 */
262 	if (init_dfree == 0 || init_dfree >= lotsfree)
263 		desfree = lotsfree / 2;
264 	else
265 		desfree = init_dfree;
266 
267 	/*
268 	 * Minfree is minimal amount of free memory which is tolerable.
269 	 */
270 	if (init_mfree == 0 || init_mfree >= desfree)
271 		minfree = desfree / 2;
272 	else
273 		minfree = init_mfree;
274 
275 	/*
276 	 * Throttlefree is the point at which we start throttling
277 	 * PG_WAIT requests until enough memory becomes available.
278 	 */
279 	if (init_tfree == 0 || init_tfree >= desfree)
280 		throttlefree = minfree;
281 	else
282 		throttlefree = init_tfree;
283 
284 	/*
285 	 * Pageout_reserve is the number of pages that we keep in
286 	 * stock for pageout's own use.  Having a few such pages
287 	 * provides insurance against system deadlock due to
288 	 * pageout needing pages.  When freemem < pageout_reserve,
289 	 * non-blocking allocations are denied to any threads
290 	 * other than pageout and sched.  (At some point we might
291 	 * want to consider a per-thread flag like T_PUSHING_PAGES
292 	 * to indicate that a thread is part of the page-pushing
293 	 * dance (e.g. an interrupt thread) and thus is entitled
294 	 * to the same special dispensation we accord pageout.)
295 	 */
296 	if (init_preserve == 0 || init_preserve >= throttlefree)
297 		pageout_reserve = throttlefree / 2;
298 	else
299 		pageout_reserve = init_preserve;
300 
301 	/*
302 	 * Maxpgio thresholds how much paging is acceptable.
303 	 * This figures that 2/3 busy on an arm is all that is
304 	 * tolerable for paging.  We assume one operation per disk rev.
305 	 *
306 	 * XXX - Does not account for multiple swap devices.
307 	 */
308 	if (init_mpgio == 0)
309 		maxpgio = (DISKRPM * 2) / 3;
310 	else
311 		maxpgio = init_mpgio;
312 
313 	/*
314 	 * The clock scan rate varies between fastscan and slowscan
315 	 * based on the amount of free memory available.  Fastscan
316 	 * rate should be set based on the number pages that can be
317 	 * scanned per sec using ~10% of processor time.  Since this
318 	 * value depends on the processor, MMU, Mhz etc., it is
319 	 * difficult to determine it in a generic manner for all
320 	 * architectures.
321 	 *
322 	 * Instead of trying to determine the number of pages scanned
323 	 * per sec for every processor, fastscan is set to be the smaller
324 	 * of 1/2 of memory or MAXHANDSPREADPAGES and the sampling
325 	 * time is limited to ~4% of processor time.
326 	 *
327 	 * Setting fastscan to be 1/2 of memory allows pageout to scan
328 	 * all of memory in ~2 secs.  This implies that user pages not
329 	 * accessed within 1 sec (assuming, handspreadpages == fastscan)
330 	 * can be reclaimed when free memory is very low.  Stealing pages
331 	 * not accessed within 1 sec seems reasonable and ensures that
332 	 * active user processes don't thrash.
333 	 *
334 	 * Smaller values of fastscan result in scanning fewer pages
335 	 * every second and consequently pageout may not be able to free
336 	 * sufficient memory to maintain the minimum threshold.  Larger
337 	 * values of fastscan result in scanning a lot more pages which
338 	 * could lead to thrashing and higher CPU usage.
339 	 *
340 	 * Fastscan needs to be limited to a maximum value and should not
341 	 * scale with memory to prevent pageout from consuming too much
342 	 * time for scanning on slow CPU's and avoid thrashing, as a
343 	 * result of scanning too many pages, on faster CPU's.
344 	 * The value of 64 Meg was chosen for MAXHANDSPREADPAGES
345 	 * (the upper bound for fastscan) based on the average number
346 	 * of pages that can potentially be scanned in ~1 sec (using ~4%
347 	 * of the CPU) on some of the following machines that currently
348 	 * run Solaris 2.x:
349 	 *
350 	 *			average memory scanned in ~1 sec
351 	 *
352 	 *	25 Mhz SS1+:		23 Meg
353 	 *	LX:			37 Meg
354 	 *	50 Mhz SC2000:		68 Meg
355 	 *
356 	 *	40 Mhz 486:		26 Meg
357 	 *	66 Mhz 486:		42 Meg
358 	 *
359 	 * When free memory falls just below lotsfree, the scan rate
360 	 * goes from 0 to slowscan (i.e., pageout starts running).  This
361 	 * transition needs to be smooth and is achieved by ensuring that
362 	 * pageout scans a small number of pages to satisfy the transient
363 	 * memory demand.  This is set to not exceed 100 pages/sec (25 per
364 	 * wakeup) since scanning that many pages has no noticible impact
365 	 * on system performance.
366 	 *
367 	 * In addition to setting fastscan and slowscan, pageout is
368 	 * limited to using ~4% of the CPU.  This results in increasing
369 	 * the time taken to scan all of memory, which in turn means that
370 	 * user processes have a better opportunity of preventing their
371 	 * pages from being stolen.  This has a positive effect on
372 	 * interactive and overall system performance when memory demand
373 	 * is high.
374 	 *
375 	 * Thus, the rate at which pages are scanned for replacement will
376 	 * vary linearly between slowscan and the number of pages that
377 	 * can be scanned using ~4% of processor time instead of varying
378 	 * linearly between slowscan and fastscan.
379 	 *
380 	 * Also, the processor time used by pageout will vary from ~1%
381 	 * at slowscan to ~4% at fastscan instead of varying between
382 	 * ~1% at slowscan and ~10% at fastscan.
383 	 *
384 	 * The values chosen for the various VM parameters (fastscan,
385 	 * handspreadpages, etc) are not universally true for all machines,
386 	 * but appear to be a good rule of thumb for the machines we've
387 	 * tested.  They have the following ranges:
388 	 *
389 	 *	cpu speed:	20 to 70 Mhz
390 	 *	page size:	4K to 8K
391 	 *	memory size:	16M to 5G
392 	 *	page scan rate:	4000 - 17400 4K pages per sec
393 	 *
394 	 * The values need to be re-examined for machines which don't
395 	 * fall into the various ranges (e.g., slower or faster CPUs,
396 	 * smaller or larger pagesizes etc) shown above.
397 	 *
398 	 * On an MP machine, pageout is often unable to maintain the
399 	 * minimum paging thresholds under heavy load.  This is due to
400 	 * the fact that user processes running on other CPU's can be
401 	 * dirtying memory at a much faster pace than pageout can find
402 	 * pages to free.  The memory demands could be met by enabling
403 	 * more than one CPU to run the clock algorithm in such a manner
404 	 * that the various clock hands don't overlap.  This also makes
405 	 * it more difficult to determine the values for fastscan, slowscan
406 	 * and handspreadpages.
407 	 *
408 	 * The swapper is currently used to free up memory when pageout
409 	 * is unable to meet memory demands by swapping out processes.
410 	 * In addition to freeing up memory, swapping also reduces the
411 	 * demand for memory by preventing user processes from running
412 	 * and thereby consuming memory.
413 	 */
414 	if (init_mfscan == 0) {
415 		if (pageout_new_spread != 0)
416 			maxfastscan = pageout_new_spread;
417 		else
418 			maxfastscan = MAXHANDSPREADPAGES;
419 	} else {
420 		maxfastscan = init_mfscan;
421 	}
422 	if (init_fscan == 0)
423 		fastscan = MIN(looppages / loopfraction, maxfastscan);
424 	else
425 		fastscan = init_fscan;
426 	if (fastscan > looppages / loopfraction)
427 		fastscan = looppages / loopfraction;
428 
429 	/*
430 	 * Set slow scan time to 1/10 the fast scan time, but
431 	 * not to exceed maxslowscan.
432 	 */
433 	if (init_sscan == 0)
434 		slowscan = MIN(fastscan / 10, maxslowscan);
435 	else
436 		slowscan = init_sscan;
437 	if (slowscan > fastscan / 2)
438 		slowscan = fastscan / 2;
439 
440 	/*
441 	 * Handspreadpages is distance (in pages) between front and back
442 	 * pageout daemon hands.  The amount of time to reclaim a page
443 	 * once pageout examines it increases with this distance and
444 	 * decreases as the scan rate rises. It must be < the amount
445 	 * of pageable memory.
446 	 *
447 	 * Since pageout is limited to ~4% of the CPU, setting handspreadpages
448 	 * to be "fastscan" results in the front hand being a few secs
449 	 * (varies based on the processor speed) ahead of the back hand
450 	 * at fastscan rates.  This distance can be further reduced, if
451 	 * necessary, by increasing the processor time used by pageout
452 	 * to be more than ~4% and preferrably not more than ~10%.
453 	 *
454 	 * As a result, user processes have a much better chance of
455 	 * referencing their pages before the back hand examines them.
456 	 * This also significantly lowers the number of reclaims from
457 	 * the freelist since pageout does not end up freeing pages which
458 	 * may be referenced a sec later.
459 	 */
460 	if (init_hspages == 0)
461 		handspreadpages = fastscan;
462 	else
463 		handspreadpages = init_hspages;
464 
465 	/*
466 	 * Make sure that back hand follows front hand by at least
467 	 * 1/RATETOSCHEDPAGING seconds.  Without this test, it is possible
468 	 * for the back hand to look at a page during the same wakeup of
469 	 * the pageout daemon in which the front hand cleared its ref bit.
470 	 */
471 	if (handspreadpages >= looppages)
472 		handspreadpages = looppages - 1;
473 
474 	/*
475 	 * If we have been called to recalculate the parameters,
476 	 * set a flag to re-evaluate the clock hand pointers.
477 	 */
478 	if (recalc)
479 		reset_hands = 1;
480 }
481 
482 /*
483  * Pageout scheduling.
484  *
485  * Schedpaging controls the rate at which the page out daemon runs by
486  * setting the global variables nscan and desscan RATETOSCHEDPAGING
487  * times a second.  Nscan records the number of pages pageout has examined
488  * in its current pass; schedpaging resets this value to zero each time
489  * it runs.  Desscan records the number of pages pageout should examine
490  * in its next pass; schedpaging sets this value based on the amount of
491  * currently available memory.
492  */
493 
494 #define	RATETOSCHEDPAGING	4		/* hz that is */
495 
496 static kmutex_t	pageout_mutex;	/* held while pageout or schedpaging running */
497 
498 /*
499  * Pool of available async pageout putpage requests.
500  */
501 static struct async_reqs *push_req;
502 static struct async_reqs *req_freelist;	/* available req structs */
503 static struct async_reqs *push_list;	/* pending reqs */
504 static kmutex_t push_lock;		/* protects req pool */
505 static kcondvar_t push_cv;
506 
507 static int async_list_size = 256;	/* number of async request structs */
508 
509 static void pageout_scanner(void);
510 
511 /*
512  * If a page is being shared more than "po_share" times
513  * then leave it alone- don't page it out.
514  */
515 #define	MIN_PO_SHARE	(8)
516 #define	MAX_PO_SHARE	((MIN_PO_SHARE) << 24)
517 ulong_t	po_share = MIN_PO_SHARE;
518 
519 /*
520  * Schedule rate for paging.
521  * Rate is linear interpolation between
522  * slowscan with lotsfree and fastscan when out of memory.
523  */
524 static void
schedpaging(void * arg)525 schedpaging(void *arg)
526 {
527 	spgcnt_t vavail;
528 
529 	if (freemem < lotsfree + needfree + kmem_reapahead)
530 		kmem_reap();
531 
532 	if (freemem < lotsfree + needfree)
533 		seg_preap();
534 
535 	if (kcage_on && (kcage_freemem < kcage_desfree || kcage_needfree))
536 		kcage_cageout_wakeup();
537 
538 	if (mutex_tryenter(&pageout_mutex)) {
539 		/* pageout() not running */
540 		nscan = 0;
541 		vavail = freemem - deficit;
542 		if (pageout_new_spread != 0)
543 			vavail -= needfree;
544 		if (vavail < 0)
545 			vavail = 0;
546 		if (vavail > lotsfree)
547 			vavail = lotsfree;
548 
549 		/*
550 		 * Fix for 1161438 (CRS SPR# 73922).  All variables
551 		 * in the original calculation for desscan were 32 bit signed
552 		 * ints.  As freemem approaches 0x0 on a system with 1 Gig or
553 		 * more of memory, the calculation can overflow.  When this
554 		 * happens, desscan becomes negative and pageout_scanner()
555 		 * stops paging out.
556 		 */
557 		if ((needfree) && (pageout_new_spread == 0)) {
558 			/*
559 			 * If we've not yet collected enough samples to
560 			 * calculate a spread, use the old logic of kicking
561 			 * into high gear anytime needfree is non-zero.
562 			 */
563 			desscan = fastscan / RATETOSCHEDPAGING;
564 		} else {
565 			/*
566 			 * Once we've calculated a spread based on system
567 			 * memory and usage, just treat needfree as another
568 			 * form of deficit.
569 			 */
570 			spgcnt_t faststmp, slowstmp, result;
571 
572 			slowstmp = slowscan * vavail;
573 			faststmp = fastscan * (lotsfree - vavail);
574 			result = (slowstmp + faststmp) /
575 			    nz(lotsfree) / RATETOSCHEDPAGING;
576 			desscan = (pgcnt_t)result;
577 		}
578 
579 		pageout_ticks = min_pageout_ticks + (lotsfree - vavail) *
580 		    (max_pageout_ticks - min_pageout_ticks) / nz(lotsfree);
581 
582 		if (freemem < lotsfree + needfree ||
583 		    pageout_sample_cnt < pageout_sample_lim) {
584 			TRACE_1(TR_FAC_VM, TR_PAGEOUT_CV_SIGNAL,
585 			    "pageout_cv_signal:freemem %ld", freemem);
586 			cv_signal(&proc_pageout->p_cv);
587 		} else {
588 			/*
589 			 * There are enough free pages, no need to
590 			 * kick the scanner thread.  And next time
591 			 * around, keep more of the `highly shared'
592 			 * pages.
593 			 */
594 			cv_signal_pageout();
595 			if (po_share > MIN_PO_SHARE) {
596 				po_share >>= 1;
597 			}
598 		}
599 		mutex_exit(&pageout_mutex);
600 	}
601 
602 	/*
603 	 * Signal threads waiting for available memory.
604 	 * NOTE: usually we need to grab memavail_lock before cv_broadcast, but
605 	 * in this case it is not needed - the waiters will be waken up during
606 	 * the next invocation of this function.
607 	 */
608 	if (kmem_avail() > 0)
609 		cv_broadcast(&memavail_cv);
610 
611 	(void) timeout(schedpaging, arg, hz / RATETOSCHEDPAGING);
612 }
613 
614 pgcnt_t		pushes;
615 ulong_t		push_list_size;		/* # of requests on pageout queue */
616 
617 #define	FRONT	1
618 #define	BACK	2
619 
620 int dopageout = 1;	/* must be non-zero to turn page stealing on */
621 
622 /*
623  * The page out daemon, which runs as process 2.
624  *
625  * As long as there are at least lotsfree pages,
626  * this process is not run.  When the number of free
627  * pages stays in the range desfree to lotsfree,
628  * this daemon runs through the pages in the loop
629  * at a rate determined in schedpaging().  Pageout manages
630  * two hands on the clock.  The front hand moves through
631  * memory, clearing the reference bit,
632  * and stealing pages from procs that are over maxrss.
633  * The back hand travels a distance behind the front hand,
634  * freeing the pages that have not been referenced in the time
635  * since the front hand passed.  If modified, they are pushed to
636  * swap before being freed.
637  *
638  * There are 2 threads that act on behalf of the pageout process.
639  * One thread scans pages (pageout_scanner) and frees them up if
640  * they don't require any VOP_PUTPAGE operation. If a page must be
641  * written back to its backing store, the request is put on a list
642  * and the other (pageout) thread is signaled. The pageout thread
643  * grabs VOP_PUTPAGE requests from the list, and processes them.
644  * Some filesystems may require resources for the VOP_PUTPAGE
645  * operations (like memory) and hence can block the pageout
646  * thread, but the scanner thread can still operate. There is still
647  * no guarantee that memory deadlocks cannot occur.
648  *
649  * For now, this thing is in very rough form.
650  */
651 void
pageout()652 pageout()
653 {
654 	struct async_reqs *arg;
655 	pri_t pageout_pri;
656 	int i;
657 	pgcnt_t max_pushes;
658 	callb_cpr_t cprinfo;
659 
660 	proc_pageout = ttoproc(curthread);
661 	proc_pageout->p_cstime = 0;
662 	proc_pageout->p_stime =  0;
663 	proc_pageout->p_cutime =  0;
664 	proc_pageout->p_utime = 0;
665 	bcopy("pageout", PTOU(curproc)->u_psargs, 8);
666 	bcopy("pageout", PTOU(curproc)->u_comm, 7);
667 
668 	/*
669 	 * Create pageout scanner thread
670 	 */
671 	mutex_init(&pageout_mutex, NULL, MUTEX_DEFAULT, NULL);
672 	mutex_init(&push_lock, NULL, MUTEX_DEFAULT, NULL);
673 
674 	/*
675 	 * Allocate and initialize the async request structures
676 	 * for pageout.
677 	 */
678 	push_req = (struct async_reqs *)
679 	    kmem_zalloc(async_list_size * sizeof (struct async_reqs), KM_SLEEP);
680 
681 	req_freelist = push_req;
682 	for (i = 0; i < async_list_size - 1; i++)
683 		push_req[i].a_next = &push_req[i + 1];
684 
685 	pageout_pri = curthread->t_pri;
686 
687 	/* Create the pageout scanner thread. */
688 	(void) lwp_kernel_create(proc_pageout, pageout_scanner, NULL, TS_RUN,
689 	    pageout_pri - 1);
690 
691 	/*
692 	 * kick off pageout scheduler.
693 	 */
694 	schedpaging(NULL);
695 
696 	/*
697 	 * Create kernel cage thread.
698 	 * The kernel cage thread is started under the pageout process
699 	 * to take advantage of the less restricted page allocation
700 	 * in page_create_throttle().
701 	 */
702 	kcage_cageout_init();
703 
704 	/*
705 	 * Limit pushes to avoid saturating pageout devices.
706 	 */
707 	max_pushes = maxpgio / RATETOSCHEDPAGING;
708 	CALLB_CPR_INIT(&cprinfo, &push_lock, callb_generic_cpr, "pageout");
709 
710 	for (;;) {
711 		mutex_enter(&push_lock);
712 
713 		while ((arg = push_list) == NULL || pushes > max_pushes) {
714 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
715 			cv_wait(&push_cv, &push_lock);
716 			pushes = 0;
717 			CALLB_CPR_SAFE_END(&cprinfo, &push_lock);
718 		}
719 		push_list = arg->a_next;
720 		arg->a_next = NULL;
721 		mutex_exit(&push_lock);
722 
723 		if (VOP_PUTPAGE(arg->a_vp, (offset_t)arg->a_off,
724 		    arg->a_len, arg->a_flags, arg->a_cred, NULL) == 0) {
725 			pushes++;
726 		}
727 
728 		/* vp held by checkpage() */
729 		VN_RELE(arg->a_vp);
730 
731 		mutex_enter(&push_lock);
732 		arg->a_next = req_freelist;	/* back on freelist */
733 		req_freelist = arg;
734 		push_list_size--;
735 		mutex_exit(&push_lock);
736 	}
737 }
738 
739 /*
740  * Kernel thread that scans pages looking for ones to free
741  */
742 static void
pageout_scanner(void)743 pageout_scanner(void)
744 {
745 	struct page *fronthand, *backhand;
746 	uint_t count;
747 	callb_cpr_t cprinfo;
748 	pgcnt_t	nscan_limit;
749 	pgcnt_t	pcount;
750 
751 	CALLB_CPR_INIT(&cprinfo, &pageout_mutex, callb_generic_cpr, "poscan");
752 	mutex_enter(&pageout_mutex);
753 
754 	/*
755 	 * The restart case does not attempt to point the hands at roughly
756 	 * the right point on the assumption that after one circuit things
757 	 * will have settled down - and restarts shouldn't be that often.
758 	 */
759 
760 	/*
761 	 * Set the two clock hands to be separated by a reasonable amount,
762 	 * but no more than 360 degrees apart.
763 	 */
764 	backhand = page_first();
765 	if (handspreadpages >= total_pages)
766 		fronthand = page_nextn(backhand, total_pages - 1);
767 	else
768 		fronthand = page_nextn(backhand, handspreadpages);
769 
770 	min_pageout_ticks = MAX(1,
771 	    ((hz * min_percent_cpu) / 100) / RATETOSCHEDPAGING);
772 	max_pageout_ticks = MAX(min_pageout_ticks,
773 	    ((hz * max_percent_cpu) / 100) / RATETOSCHEDPAGING);
774 
775 loop:
776 	cv_signal_pageout();
777 
778 	CALLB_CPR_SAFE_BEGIN(&cprinfo);
779 	cv_wait(&proc_pageout->p_cv, &pageout_mutex);
780 	CALLB_CPR_SAFE_END(&cprinfo, &pageout_mutex);
781 
782 	if (!dopageout)
783 		goto loop;
784 
785 	if (reset_hands) {
786 		reset_hands = 0;
787 
788 		backhand = page_first();
789 		if (handspreadpages >= total_pages)
790 			fronthand = page_nextn(backhand, total_pages - 1);
791 		else
792 			fronthand = page_nextn(backhand, handspreadpages);
793 	}
794 
795 	CPU_STATS_ADDQ(CPU, vm, pgrrun, 1);
796 	count = 0;
797 
798 	TRACE_4(TR_FAC_VM, TR_PAGEOUT_START,
799 	    "pageout_start:freemem %ld lotsfree %ld nscan %ld desscan %ld",
800 	    freemem, lotsfree, nscan, desscan);
801 
802 	/* Kernel probe */
803 	TNF_PROBE_2(pageout_scan_start, "vm pagedaemon", /* CSTYLED */,
804 	    tnf_ulong, pages_free, freemem, tnf_ulong, pages_needed, needfree);
805 
806 	pcount = 0;
807 	if (pageout_sample_cnt < pageout_sample_lim) {
808 		nscan_limit = total_pages;
809 	} else {
810 		nscan_limit = desscan;
811 	}
812 	pageout_lbolt = ddi_get_lbolt();
813 	sample_start = gethrtime();
814 
815 	/*
816 	 * Scan the appropriate number of pages for a single duty cycle.
817 	 * However, stop scanning as soon as there is enough free memory.
818 	 * For a short while, we will be sampling the performance of the
819 	 * scanner and need to keep running just to get sample data, in
820 	 * which case we keep going and don't pay attention to whether
821 	 * or not there is enough free memory.
822 	 */
823 
824 	while (nscan < nscan_limit && (freemem < lotsfree + needfree ||
825 	    pageout_sample_cnt < pageout_sample_lim)) {
826 		int rvfront, rvback;
827 
828 		/*
829 		 * Check to see if we have exceeded our %CPU budget
830 		 * for this wakeup, but not on every single page visited,
831 		 * just every once in a while.
832 		 */
833 		if ((pcount & PAGES_POLL_MASK) == PAGES_POLL_MASK) {
834 			pageout_cycle_ticks = ddi_get_lbolt() - pageout_lbolt;
835 			if (pageout_cycle_ticks >= pageout_ticks) {
836 				++pageout_timeouts;
837 				break;
838 			}
839 		}
840 
841 		/*
842 		 * If checkpage manages to add a page to the free list,
843 		 * we give ourselves another couple of trips around the loop.
844 		 */
845 		if ((rvfront = checkpage(fronthand, FRONT)) == 1)
846 			count = 0;
847 		if ((rvback = checkpage(backhand, BACK)) == 1)
848 			count = 0;
849 
850 		++pcount;
851 
852 		/*
853 		 * protected by pageout_mutex instead of cpu_stat_lock
854 		 */
855 		CPU_STATS_ADDQ(CPU, vm, scan, 1);
856 
857 		/*
858 		 * Don't include ineligible pages in the number scanned.
859 		 */
860 		if (rvfront != -1 || rvback != -1)
861 			nscan++;
862 
863 		backhand = page_next(backhand);
864 
865 		/*
866 		 * backhand update and wraparound check are done separately
867 		 * because lint barks when it finds an empty "if" body
868 		 */
869 
870 		if ((fronthand = page_next(fronthand)) == page_first())	{
871 			TRACE_2(TR_FAC_VM, TR_PAGEOUT_HAND_WRAP,
872 			    "pageout_hand_wrap:freemem %ld whichhand %d",
873 			    freemem, FRONT);
874 
875 			/*
876 			 * protected by pageout_mutex instead of cpu_stat_lock
877 			 */
878 			CPU_STATS_ADDQ(CPU, vm, rev, 1);
879 			if (++count > 1) {
880 				/*
881 				 * Extremely unlikely, but it happens.
882 				 * We went around the loop at least once
883 				 * and didn't get far enough.
884 				 * If we are still skipping `highly shared'
885 				 * pages, skip fewer of them.  Otherwise,
886 				 * give up till the next clock tick.
887 				 */
888 				if (po_share < MAX_PO_SHARE) {
889 					po_share <<= 1;
890 				} else {
891 					/*
892 					 * Really a "goto loop", but
893 					 * if someone is TRACing or
894 					 * TNF_PROBE_ing, at least
895 					 * make records to show
896 					 * where we are.
897 					 */
898 					break;
899 				}
900 			}
901 		}
902 	}
903 
904 	sample_end = gethrtime();
905 
906 	TRACE_5(TR_FAC_VM, TR_PAGEOUT_END,
907 	    "pageout_end:freemem %ld lots %ld nscan %ld des %ld count %u",
908 	    freemem, lotsfree, nscan, desscan, count);
909 
910 	/* Kernel probe */
911 	TNF_PROBE_2(pageout_scan_end, "vm pagedaemon", /* CSTYLED */,
912 	    tnf_ulong, pages_scanned, nscan, tnf_ulong, pages_free, freemem);
913 
914 	if (pageout_sample_cnt < pageout_sample_lim) {
915 		pageout_sample_pages += pcount;
916 		pageout_sample_etime += sample_end - sample_start;
917 		++pageout_sample_cnt;
918 	}
919 	if (pageout_sample_cnt >= pageout_sample_lim &&
920 	    pageout_new_spread == 0) {
921 		pageout_rate = (hrrate_t)pageout_sample_pages *
922 		    (hrrate_t)(NANOSEC) / pageout_sample_etime;
923 		pageout_new_spread = pageout_rate / 10;
924 		setupclock(1);
925 	}
926 
927 	goto loop;
928 }
929 
930 /*
931  * Look at the page at hand.  If it is locked (e.g., for physical i/o),
932  * system (u., page table) or free, then leave it alone.  Otherwise,
933  * if we are running the front hand, turn off the page's reference bit.
934  * If the proc is over maxrss, we take it.  If running the back hand,
935  * check whether the page has been reclaimed.  If not, free the page,
936  * pushing it to disk first if necessary.
937  *
938  * Return values:
939  *	-1 if the page is not a candidate at all,
940  *	 0 if not freed, or
941  *	 1 if we freed it.
942  */
943 static int
checkpage(struct page * pp,int whichhand)944 checkpage(struct page *pp, int whichhand)
945 {
946 	int ppattr;
947 	int isfs = 0;
948 	int isexec = 0;
949 	int pagesync_flag;
950 
951 	/*
952 	 * Skip pages:
953 	 * 	- associated with the kernel vnode since
954 	 *	    they are always "exclusively" locked.
955 	 *	- that are free
956 	 *	- that are shared more than po_share'd times
957 	 *	- its already locked
958 	 *
959 	 * NOTE:  These optimizations assume that reads are atomic.
960 	 */
961 
962 	if (PP_ISKAS(pp) || PAGE_LOCKED(pp) || PP_ISFREE(pp) ||
963 	    pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
964 	    hat_page_checkshare(pp, po_share)) {
965 		return (-1);
966 	}
967 
968 	if (!page_trylock(pp, SE_EXCL)) {
969 		/*
970 		 * Skip the page if we can't acquire the "exclusive" lock.
971 		 */
972 		return (-1);
973 	} else if (PP_ISFREE(pp)) {
974 		/*
975 		 * It became free between the above check and our actually
976 		 * locking the page.  Oh, well there will be other pages.
977 		 */
978 		page_unlock(pp);
979 		return (-1);
980 	}
981 
982 	/*
983 	 * Reject pages that cannot be freed. The page_struct_lock
984 	 * need not be acquired to examine these
985 	 * fields since the page has an "exclusive" lock.
986 	 */
987 	if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
988 		page_unlock(pp);
989 		return (-1);
990 	}
991 
992 	/*
993 	 * Maintain statistics for what we are freeing
994 	 */
995 
996 	if (pp->p_vnode != NULL) {
997 		if (pp->p_vnode->v_flag & VVMEXEC)
998 			isexec = 1;
999 
1000 		if (!IS_SWAPFSVP(pp->p_vnode))
1001 			isfs = 1;
1002 	}
1003 
1004 	/*
1005 	 * Turn off REF and MOD bits with the front hand.
1006 	 * The back hand examines the REF bit and always considers
1007 	 * SHARED pages as referenced.
1008 	 */
1009 	if (whichhand == FRONT)
1010 		pagesync_flag = HAT_SYNC_ZERORM;
1011 	else
1012 		pagesync_flag = HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_REF |
1013 		    HAT_SYNC_STOPON_SHARED;
1014 
1015 	ppattr = hat_pagesync(pp, pagesync_flag);
1016 
1017 recheck:
1018 	/*
1019 	 * If page is referenced; make unreferenced but reclaimable.
1020 	 * If this page is not referenced, then it must be reclaimable
1021 	 * and we can add it to the free list.
1022 	 */
1023 	if (ppattr & P_REF) {
1024 		TRACE_2(TR_FAC_VM, TR_PAGEOUT_ISREF,
1025 		    "pageout_isref:pp %p whichhand %d", pp, whichhand);
1026 		if (whichhand == FRONT) {
1027 			/*
1028 			 * Checking of rss or madvise flags needed here...
1029 			 *
1030 			 * If not "well-behaved", fall through into the code
1031 			 * for not referenced.
1032 			 */
1033 			hat_clrref(pp);
1034 		}
1035 		/*
1036 		 * Somebody referenced the page since the front
1037 		 * hand went by, so it's not a candidate for
1038 		 * freeing up.
1039 		 */
1040 		page_unlock(pp);
1041 		return (0);
1042 	}
1043 
1044 	VM_STAT_ADD(pageoutvmstats.checkpage[0]);
1045 
1046 	/*
1047 	 * If large page, attempt to demote it. If successfully demoted,
1048 	 * retry the checkpage.
1049 	 */
1050 	if (pp->p_szc != 0) {
1051 		if (!page_try_demote_pages(pp)) {
1052 			VM_STAT_ADD(pageoutvmstats.checkpage[1]);
1053 			page_unlock(pp);
1054 			return (-1);
1055 		}
1056 		ASSERT(pp->p_szc == 0);
1057 		VM_STAT_ADD(pageoutvmstats.checkpage[2]);
1058 		/*
1059 		 * since page_try_demote_pages() could have unloaded some
1060 		 * mappings it makes sense to reload ppattr.
1061 		 */
1062 		ppattr = hat_page_getattr(pp, P_MOD | P_REF);
1063 	}
1064 
1065 	/*
1066 	 * If the page is currently dirty, we have to arrange
1067 	 * to have it cleaned before it can be freed.
1068 	 *
1069 	 * XXX - ASSERT(pp->p_vnode != NULL);
1070 	 */
1071 	if ((ppattr & P_MOD) && pp->p_vnode) {
1072 		struct vnode *vp = pp->p_vnode;
1073 		u_offset_t offset = pp->p_offset;
1074 
1075 		/*
1076 		 * XXX - Test for process being swapped out or about to exit?
1077 		 * [Can't get back to process(es) using the page.]
1078 		 */
1079 
1080 		/*
1081 		 * Hold the vnode before releasing the page lock to
1082 		 * prevent it from being freed and re-used by some
1083 		 * other thread.
1084 		 */
1085 		VN_HOLD(vp);
1086 		page_unlock(pp);
1087 
1088 		/*
1089 		 * Queue i/o request for the pageout thread.
1090 		 */
1091 		if (!queue_io_request(vp, offset)) {
1092 			VN_RELE(vp);
1093 			return (0);
1094 		}
1095 		return (1);
1096 	}
1097 
1098 	/*
1099 	 * Now we unload all the translations,
1100 	 * and put the page back on to the free list.
1101 	 * If the page was used (referenced or modified) after
1102 	 * the pagesync but before it was unloaded we catch it
1103 	 * and handle the page properly.
1104 	 */
1105 	TRACE_2(TR_FAC_VM, TR_PAGEOUT_FREE,
1106 	    "pageout_free:pp %p whichhand %d", pp, whichhand);
1107 	(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
1108 	ppattr = hat_page_getattr(pp, P_MOD | P_REF);
1109 	if ((ppattr & P_REF) || ((ppattr & P_MOD) && pp->p_vnode))
1110 		goto recheck;
1111 
1112 	/*LINTED: constant in conditional context*/
1113 	VN_DISPOSE(pp, B_FREE, 0, kcred);
1114 
1115 	CPU_STATS_ADD_K(vm, dfree, 1);
1116 
1117 	if (isfs) {
1118 		if (isexec) {
1119 			CPU_STATS_ADD_K(vm, execfree, 1);
1120 		} else {
1121 			CPU_STATS_ADD_K(vm, fsfree, 1);
1122 		}
1123 	} else {
1124 		CPU_STATS_ADD_K(vm, anonfree, 1);
1125 	}
1126 
1127 	return (1);		/* freed a page! */
1128 }
1129 
1130 /*
1131  * Queue async i/o request from pageout_scanner and segment swapout
1132  * routines on one common list.  This ensures that pageout devices (swap)
1133  * are not saturated by pageout_scanner or swapout requests.
1134  * The pageout thread empties this list by initiating i/o operations.
1135  */
1136 int
queue_io_request(vnode_t * vp,u_offset_t off)1137 queue_io_request(vnode_t *vp, u_offset_t off)
1138 {
1139 	struct async_reqs *arg;
1140 
1141 	/*
1142 	 * If we cannot allocate an async request struct,
1143 	 * skip this page.
1144 	 */
1145 	mutex_enter(&push_lock);
1146 	if ((arg = req_freelist) == NULL) {
1147 		mutex_exit(&push_lock);
1148 		return (0);
1149 	}
1150 	req_freelist = arg->a_next;		/* adjust freelist */
1151 	push_list_size++;
1152 
1153 	arg->a_vp = vp;
1154 	arg->a_off = off;
1155 	arg->a_len = PAGESIZE;
1156 	arg->a_flags = B_ASYNC | B_FREE;
1157 	arg->a_cred = kcred;		/* always held */
1158 
1159 	/*
1160 	 * Add to list of pending write requests.
1161 	 */
1162 	arg->a_next = push_list;
1163 	push_list = arg;
1164 
1165 	if (req_freelist == NULL) {
1166 		/*
1167 		 * No free async requests left. The lock is held so we
1168 		 * might as well signal the pusher thread now.
1169 		 */
1170 		cv_signal(&push_cv);
1171 	}
1172 	mutex_exit(&push_lock);
1173 	return (1);
1174 }
1175 
1176 /*
1177  * Wakeup pageout to initiate i/o if push_list is not empty.
1178  */
1179 void
cv_signal_pageout()1180 cv_signal_pageout()
1181 {
1182 	if (push_list != NULL) {
1183 		mutex_enter(&push_lock);
1184 		cv_signal(&push_cv);
1185 		mutex_exit(&push_lock);
1186 	}
1187 }
1188