xref: /titanic_50/usr/src/uts/common/os/mem_cage.c (revision 7c2fbfb345896881c631598ee3852ce9ce33fb07)
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 #include <sys/types.h>
27 #include <sys/param.h>
28 #include <sys/thread.h>
29 #include <sys/proc.h>
30 #include <sys/callb.h>
31 #include <sys/vnode.h>
32 #include <sys/debug.h>
33 #include <sys/systm.h>		/* for bzero */
34 #include <sys/memlist.h>
35 #include <sys/cmn_err.h>
36 #include <sys/sysmacros.h>
37 #include <sys/vmsystm.h>	/* for NOMEMWAIT() */
38 #include <sys/atomic.h>		/* used to update kcage_freemem */
39 #include <sys/kmem.h>		/* for kmem_reap */
40 #include <sys/errno.h>
41 #include <sys/mem_cage.h>
42 #include <vm/seg_kmem.h>
43 #include <vm/page.h>
44 #include <vm/hat.h>
45 #include <vm/vm_dep.h>
46 #include <sys/mem_config.h>
47 #include <sys/lgrp.h>
48 #include <sys/rwlock.h>
49 #include <sys/cpupart.h>
50 
51 extern pri_t maxclsyspri;
52 
53 #ifdef DEBUG
54 #define	KCAGE_STATS
55 #endif
56 
57 #ifdef KCAGE_STATS
58 
59 #define	KCAGE_STATS_VERSION 9	/* can help report generators */
60 #define	KCAGE_STATS_NSCANS 256	/* depth of scan statistics buffer */
61 
62 struct kcage_stats_scan {
63 	/* managed by KCAGE_STAT_* macros */
64 	clock_t	scan_lbolt;
65 	uint_t	scan_id;
66 
67 	/* set in kcage_cageout() */
68 	uint_t	kt_passes;
69 	clock_t	kt_ticks;
70 	pgcnt_t	kt_kcage_freemem_start;
71 	pgcnt_t	kt_kcage_freemem_end;
72 	pgcnt_t kt_freemem_start;
73 	pgcnt_t kt_freemem_end;
74 	uint_t	kt_examined;
75 	uint_t	kt_cantlock;
76 	uint_t	kt_gotone;
77 	uint_t	kt_gotonefree;
78 	uint_t	kt_skiplevel;
79 	uint_t	kt_skipshared;
80 	uint_t	kt_skiprefd;
81 	uint_t	kt_destroy;
82 
83 	/* set in kcage_invalidate_page() */
84 	uint_t	kip_reloclocked;
85 	uint_t	kip_relocmod;
86 	uint_t	kip_destroy;
87 	uint_t	kip_nomem;
88 	uint_t	kip_demotefailed;
89 
90 	/* set in kcage_expand() */
91 	uint_t	ke_wanted;
92 	uint_t	ke_examined;
93 	uint_t	ke_lefthole;
94 	uint_t	ke_gotone;
95 	uint_t	ke_gotonefree;
96 };
97 
98 struct kcage_stats {
99 	/* managed by KCAGE_STAT_* macros */
100 	uint_t	version;
101 	uint_t	size;
102 
103 	/* set in kcage_cageout */
104 	uint_t	kt_wakeups;
105 	uint_t	kt_scans;
106 	uint_t	kt_cageout_break;
107 
108 	/* set in kcage_expand */
109 	uint_t	ke_calls;
110 	uint_t	ke_nopfn;
111 	uint_t	ke_nopaget;
112 	uint_t	ke_isnoreloc;
113 	uint_t	ke_deleting;
114 	uint_t	ke_lowfreemem;
115 	uint_t	ke_terminate;
116 
117 	/* set in kcage_freemem_add() */
118 	uint_t	kfa_trottlewake;
119 
120 	/* set in kcage_freemem_sub() */
121 	uint_t	kfs_cagewake;
122 
123 	/* set in kcage_create_throttle */
124 	uint_t	kct_calls;
125 	uint_t	kct_cageout;
126 	uint_t	kct_critical;
127 	uint_t	kct_exempt;
128 	uint_t	kct_cagewake;
129 	uint_t	kct_wait;
130 	uint_t	kct_progress;
131 	uint_t	kct_noprogress;
132 	uint_t	kct_timeout;
133 
134 	/* set in kcage_cageout_wakeup */
135 	uint_t	kcw_expandearly;
136 
137 	/* managed by KCAGE_STAT_* macros */
138 	uint_t	scan_array_size;
139 	uint_t	scan_index;
140 	struct kcage_stats_scan scans[KCAGE_STATS_NSCANS];
141 };
142 
143 static struct kcage_stats kcage_stats;
144 static struct kcage_stats_scan kcage_stats_scan_zero;
145 
146 /*
147  * No real need for atomics here. For the most part the incs and sets are
148  * done by the kernel cage thread. There are a few that are done by any
149  * number of other threads. Those cases are noted by comments.
150  */
151 #define	KCAGE_STAT_INCR(m)	kcage_stats.m++
152 
153 #define	KCAGE_STAT_NINCR(m, v) kcage_stats.m += (v)
154 
155 #define	KCAGE_STAT_INCR_SCAN(m)	\
156 	KCAGE_STAT_INCR(scans[kcage_stats.scan_index].m)
157 
158 #define	KCAGE_STAT_NINCR_SCAN(m, v) \
159 	KCAGE_STAT_NINCR(scans[kcage_stats.scan_index].m, v)
160 
161 #define	KCAGE_STAT_SET(m, v)	kcage_stats.m = (v)
162 
163 #define	KCAGE_STAT_SETZ(m, v)	\
164 	if (kcage_stats.m == 0) kcage_stats.m = (v)
165 
166 #define	KCAGE_STAT_SET_SCAN(m, v)	\
167 	KCAGE_STAT_SET(scans[kcage_stats.scan_index].m, v)
168 
169 #define	KCAGE_STAT_SETZ_SCAN(m, v)	\
170 	KCAGE_STAT_SETZ(scans[kcage_stats.scan_index].m, v)
171 
172 #define	KCAGE_STAT_INC_SCAN_INDEX \
173 	KCAGE_STAT_SET_SCAN(scan_lbolt, lbolt); \
174 	KCAGE_STAT_SET_SCAN(scan_id, kcage_stats.scan_index); \
175 	kcage_stats.scan_index = \
176 	(kcage_stats.scan_index + 1) % KCAGE_STATS_NSCANS; \
177 	kcage_stats.scans[kcage_stats.scan_index] = kcage_stats_scan_zero
178 
179 #define	KCAGE_STAT_INIT_SCAN_INDEX \
180 	kcage_stats.version = KCAGE_STATS_VERSION; \
181 	kcage_stats.size = sizeof (kcage_stats); \
182 	kcage_stats.scan_array_size = KCAGE_STATS_NSCANS; \
183 	kcage_stats.scan_index = 0
184 
185 #else /* KCAGE_STATS */
186 
187 #define	KCAGE_STAT_INCR(v)
188 #define	KCAGE_STAT_NINCR(m, v)
189 #define	KCAGE_STAT_INCR_SCAN(v)
190 #define	KCAGE_STAT_NINCR_SCAN(m, v)
191 #define	KCAGE_STAT_SET(m, v)
192 #define	KCAGE_STAT_SETZ(m, v)
193 #define	KCAGE_STAT_SET_SCAN(m, v)
194 #define	KCAGE_STAT_SETZ_SCAN(m, v)
195 #define	KCAGE_STAT_INC_SCAN_INDEX
196 #define	KCAGE_STAT_INIT_SCAN_INDEX
197 
198 #endif /* KCAGE_STATS */
199 
200 static kmutex_t kcage_throttle_mutex;	/* protects kcage_throttle_cv */
201 static kcondvar_t kcage_throttle_cv;
202 
203 static kmutex_t kcage_cageout_mutex;	/* protects cv and ready flag */
204 static kcondvar_t kcage_cageout_cv;	/* cageout thread naps here */
205 static int kcage_cageout_ready;		/* nonzero when cageout thread ready */
206 kthread_id_t kcage_cageout_thread;	/* to aid debugging */
207 
208 static krwlock_t kcage_range_rwlock;	/* protects kcage_glist elements */
209 
210 /*
211  * Cage expansion happens within a range.
212  */
213 struct kcage_glist {
214 	struct kcage_glist	*next;
215 	pfn_t			base;
216 	pfn_t			lim;
217 	pfn_t			curr;
218 	int			decr;
219 };
220 
221 static struct kcage_glist *kcage_glist;
222 static struct kcage_glist *kcage_current_glist;
223 
224 /*
225  * The firstfree element is provided so that kmem_alloc can be avoided
226  * until that cage has somewhere to go. This is not currently a problem
227  * as early kmem_alloc's use BOP_ALLOC instead of page_create_va.
228  */
229 static vmem_t *kcage_arena;
230 static struct kcage_glist kcage_glist_firstfree;
231 static struct kcage_glist *kcage_glist_freelist = &kcage_glist_firstfree;
232 
233 /*
234  * Miscellaneous forward references
235  */
236 static struct kcage_glist *kcage_glist_alloc(void);
237 static int kcage_glist_delete(pfn_t, pfn_t, struct kcage_glist **);
238 static void kcage_cageout(void);
239 static int kcage_invalidate_page(page_t *, pgcnt_t *);
240 static int kcage_setnoreloc_pages(page_t *, se_t);
241 static int kcage_range_add_internal(pfn_t base, pgcnt_t npgs, kcage_dir_t);
242 static void kcage_init(pgcnt_t preferred_size);
243 static int kcage_range_delete_internal(pfn_t base, pgcnt_t npgs);
244 
245 /*
246  * Kernel Memory Cage counters and thresholds.
247  */
248 int kcage_on = 0;
249 pgcnt_t kcage_freemem;
250 pgcnt_t kcage_needfree;
251 pgcnt_t kcage_lotsfree;
252 pgcnt_t kcage_desfree;
253 pgcnt_t kcage_minfree;
254 pgcnt_t kcage_throttlefree;
255 pgcnt_t	kcage_reserve;
256 int kcage_maxwait = 10;	/* in seconds */
257 
258 /* when we use lp for kmem we start the cage at a higher initial value */
259 pgcnt_t kcage_kmemlp_mincage;
260 
261 #ifdef DEBUG
262 pgcnt_t	kcage_pagets;
263 #define	KCAGEPAGETS_INC()	kcage_pagets++
264 #else
265 #define	KCAGEPAGETS_INC()
266 #endif
267 
268 /* kstats to export what pages are currently caged */
269 kmutex_t kcage_kstat_lock;
270 static int kcage_kstat_update(kstat_t *ksp, int rw);
271 static int kcage_kstat_snapshot(kstat_t *ksp, void *buf, int rw);
272 
273 /*
274  * Startup and Dynamic Reconfiguration interfaces.
275  * kcage_range_add()
276  * kcage_range_del()
277  * kcage_range_delete_post_mem_del()
278  * kcage_range_init()
279  * kcage_set_thresholds()
280  */
281 
282 /*
283  * Called from page_get_contig_pages to get the approximate kcage pfn range
284  * for exclusion from search for contiguous pages. This routine is called
285  * without kcage_range lock (kcage routines can call page_get_contig_pages
286  * through page_relocate) and with the assumption, based on kcage_range_add,
287  * that kcage_current_glist always contain a valid pointer.
288  */
289 
290 int
291 kcage_current_pfn(pfn_t *pfncur)
292 {
293 	struct kcage_glist *lp = kcage_current_glist;
294 
295 	ASSERT(kcage_on);
296 
297 	ASSERT(lp != NULL);
298 
299 	*pfncur = lp->curr;
300 
301 	return (lp->decr);
302 }
303 
304 /*
305  * Called from vm_pagelist.c during coalesce to find kernel cage regions
306  * within an mnode. Looks for the lowest range between lo and hi.
307  *
308  * Kernel cage memory is defined between kcage_glist and kcage_current_glist.
309  * Non-cage memory is defined between kcage_current_glist and list end.
310  *
311  * If incage is set, returns the lowest kcage range. Otherwise returns lowest
312  * non-cage range.
313  *
314  * Returns zero on success and nlo, nhi:
315  * 	lo <= nlo < nhi <= hi
316  * Returns non-zero if no overlapping range is found.
317  */
318 int
319 kcage_next_range(int incage, pfn_t lo, pfn_t hi,
320     pfn_t *nlo, pfn_t *nhi)
321 {
322 	struct kcage_glist *lp;
323 	pfn_t tlo = hi;
324 	pfn_t thi = hi;
325 
326 	ASSERT(lo <= hi);
327 
328 	/*
329 	 * Reader lock protects the list, but kcage_get_pfn
330 	 * running concurrently may advance kcage_current_glist
331 	 * and also update kcage_current_glist->curr. Page
332 	 * coalesce can handle this race condition.
333 	 */
334 	rw_enter(&kcage_range_rwlock, RW_READER);
335 
336 	for (lp = incage ? kcage_glist : kcage_current_glist;
337 	    lp != NULL; lp = lp->next) {
338 
339 		pfn_t klo, khi;
340 
341 		/* find the range limits in this element */
342 		if ((incage && lp->decr) || (!incage && !lp->decr)) {
343 			klo = lp->curr;
344 			khi = lp->lim;
345 		} else {
346 			klo = lp->base;
347 			khi = lp->curr;
348 		}
349 
350 		/* handle overlap */
351 		if (klo < tlo && klo < khi && lo < khi && klo < hi) {
352 			tlo = MAX(lo, klo);
353 			thi = MIN(hi, khi);
354 			if (tlo == lo)
355 				break;
356 		}
357 
358 		/* check end of kcage */
359 		if (incage && lp == kcage_current_glist) {
360 			break;
361 		}
362 	}
363 
364 	rw_exit(&kcage_range_rwlock);
365 
366 	/* return non-zero if no overlapping range found */
367 	if (tlo == thi)
368 		return (1);
369 
370 	ASSERT(lo <= tlo && tlo < thi && thi <= hi);
371 
372 	/* return overlapping range */
373 	*nlo = tlo;
374 	*nhi = thi;
375 	return (0);
376 }
377 
378 void
379 kcage_range_init(struct memlist *ml, kcage_dir_t d, pgcnt_t preferred_size)
380 {
381 	int ret = 0;
382 
383 	ASSERT(kcage_arena == NULL);
384 	kcage_arena = vmem_create("kcage_arena", NULL, 0, sizeof (uint64_t),
385 	    segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
386 	ASSERT(kcage_arena != NULL);
387 
388 	if (d == KCAGE_DOWN) {
389 		while (ml->next != NULL)
390 			ml = ml->next;
391 	}
392 
393 	rw_enter(&kcage_range_rwlock, RW_WRITER);
394 
395 	while (ml != NULL) {
396 		ret = kcage_range_add_internal(btop(ml->address),
397 		    btop(ml->size), d);
398 		if (ret)
399 			panic("kcage_range_add_internal failed: "
400 			    "ml=%p, ret=0x%x\n", (void *)ml, ret);
401 
402 		ml = (d == KCAGE_DOWN ? ml->prev : ml->next);
403 	}
404 
405 	rw_exit(&kcage_range_rwlock);
406 
407 	if (ret == 0)
408 		kcage_init(preferred_size);
409 }
410 
411 /*
412  * Third arg controls direction of growth: 0: increasing pfns,
413  * 1: decreasing.
414  */
415 static int
416 kcage_range_add_internal(pfn_t base, pgcnt_t npgs, kcage_dir_t d)
417 {
418 	struct kcage_glist *new, **lpp;
419 	pfn_t lim;
420 
421 	ASSERT(rw_write_held(&kcage_range_rwlock));
422 
423 	ASSERT(npgs != 0);
424 	if (npgs == 0)
425 		return (EINVAL);
426 
427 	lim = base + npgs;
428 
429 	ASSERT(lim > base);
430 	if (lim <= base)
431 		return (EINVAL);
432 
433 	new = kcage_glist_alloc();
434 	if (new == NULL) {
435 		return (ENOMEM);
436 	}
437 
438 	new->base = base;
439 	new->lim = lim;
440 	new->decr = (d == KCAGE_DOWN);
441 	if (new->decr != 0)
442 		new->curr = new->lim;
443 	else
444 		new->curr = new->base;
445 	/*
446 	 * Any overlapping existing ranges are removed by deleting
447 	 * from the new list as we search for the tail.
448 	 */
449 	lpp = &kcage_glist;
450 	while (*lpp != NULL) {
451 		int ret;
452 		ret = kcage_glist_delete((*lpp)->base, (*lpp)->lim, &new);
453 		if (ret != 0)
454 			return (ret);
455 		lpp = &(*lpp)->next;
456 	}
457 
458 	*lpp = new;
459 
460 	if (kcage_current_glist == NULL) {
461 		kcage_current_glist = kcage_glist;
462 	}
463 
464 	return (0);
465 }
466 
467 int
468 kcage_range_add(pfn_t base, pgcnt_t npgs, kcage_dir_t d)
469 {
470 	int ret;
471 
472 	rw_enter(&kcage_range_rwlock, RW_WRITER);
473 	ret = kcage_range_add_internal(base, npgs, d);
474 	rw_exit(&kcage_range_rwlock);
475 	return (ret);
476 }
477 
478 /*
479  * Calls to add and delete must be protected by kcage_range_rwlock
480  */
481 static int
482 kcage_range_delete_internal(pfn_t base, pgcnt_t npgs)
483 {
484 	struct kcage_glist *lp;
485 	pfn_t lim;
486 
487 	ASSERT(rw_write_held(&kcage_range_rwlock));
488 
489 	ASSERT(npgs != 0);
490 	if (npgs == 0)
491 		return (EINVAL);
492 
493 	lim = base + npgs;
494 
495 	ASSERT(lim > base);
496 	if (lim <= base)
497 		return (EINVAL);
498 
499 	/*
500 	 * Check if the delete is OK first as a number of elements
501 	 * might be involved and it will be difficult to go
502 	 * back and undo (can't just add the range back in).
503 	 */
504 	for (lp = kcage_glist; lp != NULL; lp = lp->next) {
505 		/*
506 		 * If there have been no pages allocated from this
507 		 * element, we don't need to check it.
508 		 */
509 		if ((lp->decr == 0 && lp->curr == lp->base) ||
510 		    (lp->decr != 0 && lp->curr == lp->lim))
511 			continue;
512 		/*
513 		 * If the element does not overlap, its OK.
514 		 */
515 		if (base >= lp->lim || lim <= lp->base)
516 			continue;
517 		/*
518 		 * Overlapping element: Does the range to be deleted
519 		 * overlap the area already used? If so fail.
520 		 */
521 		if (lp->decr == 0 && base < lp->curr && lim >= lp->base) {
522 			return (EBUSY);
523 		}
524 		if (lp->decr != 0 && base < lp->lim && lim >= lp->curr) {
525 			return (EBUSY);
526 		}
527 	}
528 	return (kcage_glist_delete(base, lim, &kcage_glist));
529 }
530 
531 int
532 kcage_range_delete(pfn_t base, pgcnt_t npgs)
533 {
534 	int ret;
535 
536 	rw_enter(&kcage_range_rwlock, RW_WRITER);
537 	ret = kcage_range_delete_internal(base, npgs);
538 	rw_exit(&kcage_range_rwlock);
539 	return (ret);
540 }
541 
542 /*
543  * Calls to add and delete must be protected by kcage_range_rwlock.
544  * This routine gets called after successful Solaris memory
545  * delete operation from DR post memory delete routines.
546  */
547 static int
548 kcage_range_delete_post_mem_del_internal(pfn_t base, pgcnt_t npgs)
549 {
550 	pfn_t lim;
551 
552 	ASSERT(rw_write_held(&kcage_range_rwlock));
553 
554 	ASSERT(npgs != 0);
555 	if (npgs == 0)
556 		return (EINVAL);
557 
558 	lim = base + npgs;
559 
560 	ASSERT(lim > base);
561 	if (lim <= base)
562 		return (EINVAL);
563 
564 	return (kcage_glist_delete(base, lim, &kcage_glist));
565 }
566 
567 int
568 kcage_range_delete_post_mem_del(pfn_t base, pgcnt_t npgs)
569 {
570 	int ret;
571 
572 	rw_enter(&kcage_range_rwlock, RW_WRITER);
573 	ret = kcage_range_delete_post_mem_del_internal(base, npgs);
574 	rw_exit(&kcage_range_rwlock);
575 	return (ret);
576 }
577 
578 /*
579  * No locking is required here as the whole operation is covered
580  * by kcage_range_rwlock writer lock.
581  */
582 static struct kcage_glist *
583 kcage_glist_alloc(void)
584 {
585 	struct kcage_glist *new;
586 
587 	if ((new = kcage_glist_freelist) != NULL) {
588 		kcage_glist_freelist = new->next;
589 	} else {
590 		new = vmem_alloc(kcage_arena, sizeof (*new), VM_NOSLEEP);
591 	}
592 
593 	if (new != NULL)
594 		bzero(new, sizeof (*new));
595 
596 	return (new);
597 }
598 
599 static void
600 kcage_glist_free(struct kcage_glist *lp)
601 {
602 	lp->next = kcage_glist_freelist;
603 	kcage_glist_freelist = lp;
604 }
605 
606 static int
607 kcage_glist_delete(pfn_t base, pfn_t lim, struct kcage_glist **lpp)
608 {
609 	struct kcage_glist *lp, *prev = *lpp;
610 
611 	while ((lp = *lpp) != NULL) {
612 		if (lim > lp->base && base < lp->lim) {
613 			/* The delete range overlaps this element. */
614 			if (base <= lp->base && lim >= lp->lim) {
615 				/* Delete whole element. */
616 				*lpp = lp->next;
617 				if (lp == kcage_current_glist) {
618 					/* This can never happen. */
619 					ASSERT(kcage_current_glist != prev);
620 					kcage_current_glist = prev;
621 				}
622 				kcage_glist_free(lp);
623 				continue;
624 			}
625 
626 			/* Partial delete. */
627 			if (base > lp->base && lim < lp->lim) {
628 				struct kcage_glist *new;
629 
630 				/*
631 				 * Remove a section from the middle,
632 				 * need to allocate a new element.
633 				 */
634 				new = kcage_glist_alloc();
635 				if (new == NULL) {
636 					return (ENOMEM);
637 				}
638 
639 				/*
640 				 * Tranfser unused range to new.
641 				 * Edit lp in place to preserve
642 				 * kcage_current_glist.
643 				 */
644 				new->decr = lp->decr;
645 				if (new->decr != 0) {
646 					new->base = lp->base;
647 					new->lim = base;
648 					new->curr = base;
649 
650 					lp->base = lim;
651 				} else {
652 					new->base = lim;
653 					new->lim = lp->lim;
654 					new->curr = new->base;
655 
656 					lp->lim = base;
657 				}
658 
659 				/* Insert new. */
660 				new->next = lp->next;
661 				lp->next = new;
662 				lpp = &lp->next;
663 			} else {
664 				/* Delete part of current block. */
665 				if (base > lp->base) {
666 					ASSERT(lim >= lp->lim);
667 					ASSERT(base < lp->lim);
668 					if (lp->decr != 0 &&
669 					    lp->curr == lp->lim)
670 						lp->curr = base;
671 					lp->lim = base;
672 				} else {
673 					ASSERT(base <= lp->base);
674 					ASSERT(lim > lp->base);
675 					if (lp->decr == 0 &&
676 					    lp->curr == lp->base)
677 						lp->curr = lim;
678 					lp->base = lim;
679 				}
680 			}
681 		}
682 		prev = *lpp;
683 		lpp = &(*lpp)->next;
684 	}
685 
686 	return (0);
687 }
688 
689 /*
690  * If lockit is 1, kcage_get_pfn holds the
691  * reader lock for kcage_range_rwlock.
692  * Changes to lp->curr can cause race conditions, but
693  * they are handled by higher level code (see kcage_next_range.)
694  */
695 static pfn_t
696 kcage_get_pfn(int lockit)
697 {
698 	struct kcage_glist *lp;
699 	pfn_t pfn = PFN_INVALID;
700 
701 	if (lockit && !rw_tryenter(&kcage_range_rwlock, RW_READER))
702 		return (pfn);
703 
704 	lp = kcage_current_glist;
705 	while (lp != NULL) {
706 		if (lp->decr != 0) {
707 			if (lp->curr != lp->base) {
708 				pfn = --lp->curr;
709 				break;
710 			}
711 		} else {
712 			if (lp->curr != lp->lim) {
713 				pfn = lp->curr++;
714 				break;
715 			}
716 		}
717 
718 		lp = lp->next;
719 		if (lp)
720 			kcage_current_glist = lp;
721 	}
722 
723 	if (lockit)
724 		rw_exit(&kcage_range_rwlock);
725 	return (pfn);
726 }
727 
728 /*
729  * Walk the physical address space of the cage.
730  * This routine does not guarantee to return PFNs in the order
731  * in which they were allocated to the cage. Instead, it walks
732  * each range as they appear on the growth list returning the PFNs
733  * range in ascending order.
734  *
735  * To begin scanning at lower edge of cage, reset should be nonzero.
736  * To step through cage, reset should be zero.
737  *
738  * PFN_INVALID will be returned when the upper end of the cage is
739  * reached -- indicating a full scan of the cage has been completed since
740  * previous reset. PFN_INVALID will continue to be returned until
741  * kcage_walk_cage is reset.
742  *
743  * It is possible to receive a PFN_INVALID result on reset if a growth
744  * list is not installed or if none of the PFNs in the installed list have
745  * been allocated to the cage. In otherwords, there is no cage.
746  *
747  * Caller need not hold kcage_range_rwlock while calling this function
748  * as the front part of the list is static - pages never come out of
749  * the cage.
750  *
751  * The caller is expected to only be kcage_cageout().
752  */
753 static pfn_t
754 kcage_walk_cage(int reset)
755 {
756 	static struct kcage_glist *lp = NULL;
757 	static pfn_t pfn;
758 
759 	if (reset)
760 		lp = NULL;
761 	if (lp == NULL) {
762 		lp = kcage_glist;
763 		pfn = PFN_INVALID;
764 	}
765 again:
766 	if (pfn == PFN_INVALID) {
767 		if (lp == NULL)
768 			return (PFN_INVALID);
769 
770 		if (lp->decr != 0) {
771 			/*
772 			 * In this range the cage grows from the highest
773 			 * address towards the lowest.
774 			 * Arrange to return pfns from curr to lim-1,
775 			 * inclusive, in ascending order.
776 			 */
777 
778 			pfn = lp->curr;
779 		} else {
780 			/*
781 			 * In this range the cage grows from the lowest
782 			 * address towards the highest.
783 			 * Arrange to return pfns from base to curr,
784 			 * inclusive, in ascending order.
785 			 */
786 
787 			pfn = lp->base;
788 		}
789 	}
790 
791 	if (lp->decr != 0) {		/* decrementing pfn */
792 		if (pfn == lp->lim) {
793 			/* Don't go beyond the static part of the glist. */
794 			if (lp == kcage_current_glist)
795 				lp = NULL;
796 			else
797 				lp = lp->next;
798 			pfn = PFN_INVALID;
799 			goto again;
800 		}
801 
802 		ASSERT(pfn >= lp->curr && pfn < lp->lim);
803 	} else {			/* incrementing pfn */
804 		if (pfn == lp->curr) {
805 			/* Don't go beyond the static part of the glist. */
806 			if (lp == kcage_current_glist)
807 				lp = NULL;
808 			else
809 				lp = lp->next;
810 			pfn = PFN_INVALID;
811 			goto again;
812 		}
813 
814 		ASSERT(pfn >= lp->base && pfn < lp->curr);
815 	}
816 
817 	return (pfn++);
818 }
819 
820 /*
821  * Callback functions for to recalc cage thresholds after
822  * Kphysm memory add/delete operations.
823  */
824 /*ARGSUSED*/
825 static void
826 kcage_kphysm_postadd_cb(void *arg, pgcnt_t delta_pages)
827 {
828 	kcage_recalc_thresholds();
829 }
830 
831 /*ARGSUSED*/
832 static int
833 kcage_kphysm_predel_cb(void *arg, pgcnt_t delta_pages)
834 {
835 	/* TODO: when should cage refuse memory delete requests? */
836 	return (0);
837 }
838 
839 /*ARGSUSED*/
840 static  void
841 kcage_kphysm_postdel_cb(void *arg, pgcnt_t delta_pages, int cancelled)
842 {
843 	kcage_recalc_thresholds();
844 }
845 
846 static kphysm_setup_vector_t kcage_kphysm_vectors = {
847 	KPHYSM_SETUP_VECTOR_VERSION,
848 	kcage_kphysm_postadd_cb,
849 	kcage_kphysm_predel_cb,
850 	kcage_kphysm_postdel_cb
851 };
852 
853 /*
854  * This is called before a CPR suspend and after a CPR resume.  We have to
855  * turn off kcage_cageout_ready before a suspend, and turn it back on after a
856  * restart.
857  */
858 /*ARGSUSED*/
859 static boolean_t
860 kcage_cageout_cpr(void *arg, int code)
861 {
862 	if (code == CB_CODE_CPR_CHKPT) {
863 		ASSERT(kcage_cageout_ready);
864 		kcage_cageout_ready = 0;
865 		return (B_TRUE);
866 	} else if (code == CB_CODE_CPR_RESUME) {
867 		ASSERT(kcage_cageout_ready == 0);
868 		kcage_cageout_ready = 1;
869 		return (B_TRUE);
870 	}
871 	return (B_FALSE);
872 }
873 
874 /*
875  * kcage_recalc_preferred_size() increases initial cage size to improve large
876  * page availability when lp for kmem is enabled and kpr is disabled
877  */
878 static pgcnt_t
879 kcage_recalc_preferred_size(pgcnt_t preferred_size)
880 {
881 	if (SEGKMEM_USE_LARGEPAGES && segkmem_reloc == 0) {
882 		pgcnt_t lpmincage = kcage_kmemlp_mincage;
883 		if (lpmincage == 0) {
884 			lpmincage = MIN(P2ROUNDUP(((physmem * PAGESIZE) / 8),
885 			    segkmem_heaplp_quantum), 0x40000000UL) / PAGESIZE;
886 		}
887 		kcage_kmemlp_mincage = MIN(lpmincage,
888 		    (segkmem_kmemlp_max / PAGESIZE));
889 		preferred_size = MAX(kcage_kmemlp_mincage, preferred_size);
890 	}
891 	return (preferred_size);
892 }
893 
894 /*
895  * Kcage_init() builds the cage and initializes the cage thresholds.
896  * The size of the cage is determined by the argument preferred_size.
897  * or the actual amount of memory, whichever is smaller.
898  */
899 static void
900 kcage_init(pgcnt_t preferred_size)
901 {
902 	pgcnt_t wanted;
903 	pfn_t pfn;
904 	page_t *pp;
905 	kstat_t *ksp;
906 
907 	extern struct vnode kvp;
908 	extern void page_list_noreloc_startup(page_t *);
909 
910 	ASSERT(!kcage_on);
911 
912 	/* increase preferred cage size for lp for kmem */
913 	preferred_size = kcage_recalc_preferred_size(preferred_size);
914 
915 	/* Debug note: initialize this now so early expansions can stat */
916 	KCAGE_STAT_INIT_SCAN_INDEX;
917 
918 	/*
919 	 * Initialize cage thresholds and install kphysm callback.
920 	 * If we can't arrange to have the thresholds track with
921 	 * available physical memory, then the cage thresholds may
922 	 * end up over time at levels that adversly effect system
923 	 * performance; so, bail out.
924 	 */
925 	kcage_recalc_thresholds();
926 	if (kphysm_setup_func_register(&kcage_kphysm_vectors, NULL)) {
927 		ASSERT(0);		/* Catch this in DEBUG kernels. */
928 		return;
929 	}
930 
931 	/*
932 	 * Limit startup cage size within the range of kcage_minfree
933 	 * and availrmem, inclusively.
934 	 */
935 	wanted = MIN(MAX(preferred_size, kcage_minfree), availrmem);
936 
937 	/*
938 	 * Construct the cage. PFNs are allocated from the glist. It
939 	 * is assumed that the list has been properly ordered for the
940 	 * platform by the platform code. Typically, this is as simple
941 	 * as calling kcage_range_init(phys_avail, decr), where decr is
942 	 * 1 if the kernel has been loaded into upper end of physical
943 	 * memory, or 0 if the kernel has been loaded at the low end.
944 	 *
945 	 * Note: it is assumed that we are in the startup flow, so there
946 	 * is no reason to grab the page lock.
947 	 */
948 	kcage_freemem = 0;
949 	pfn = PFN_INVALID;			/* prime for alignment test */
950 	while (wanted != 0) {
951 		if ((pfn = kcage_get_pfn(0)) == PFN_INVALID)
952 			break;
953 
954 		if ((pp = page_numtopp_nolock(pfn)) != NULL) {
955 			KCAGEPAGETS_INC();
956 			/*
957 			 * Set the noreloc state on the page.
958 			 * If the page is free and not already
959 			 * on the noreloc list then move it.
960 			 */
961 			if (PP_ISFREE(pp)) {
962 				if (PP_ISNORELOC(pp) == 0)
963 					page_list_noreloc_startup(pp);
964 			} else {
965 				ASSERT(pp->p_szc == 0);
966 				PP_SETNORELOC(pp);
967 			}
968 		}
969 		PLCNT_XFER_NORELOC(pp);
970 		wanted -= 1;
971 	}
972 
973 	/*
974 	 * Need to go through and find kernel allocated pages
975 	 * and capture them into the Cage.  These will primarily
976 	 * be pages gotten through boot_alloc().
977 	 */
978 	if (kvp.v_pages) {
979 
980 		pp = kvp.v_pages;
981 		do {
982 			ASSERT(!PP_ISFREE(pp));
983 			ASSERT(pp->p_szc == 0);
984 			if (PP_ISNORELOC(pp) == 0) {
985 				PP_SETNORELOC(pp);
986 				PLCNT_XFER_NORELOC(pp);
987 			}
988 		} while ((pp = pp->p_vpnext) != kvp.v_pages);
989 
990 	}
991 
992 	kcage_on = 1;
993 
994 	/*
995 	 * CB_CL_CPR_POST_KERNEL is the class that executes from cpr_suspend()
996 	 * after the cageout thread is blocked, and executes from cpr_resume()
997 	 * before the cageout thread is restarted.  By executing in this class,
998 	 * we are assured that the kernel cage thread won't miss wakeup calls
999 	 * and also CPR's larger kmem_alloc requests will not fail after
1000 	 * CPR shuts down the cageout kernel thread.
1001 	 */
1002 	(void) callb_add(kcage_cageout_cpr, NULL, CB_CL_CPR_POST_KERNEL,
1003 	    "cageout");
1004 
1005 	/*
1006 	 * Coalesce pages to improve large page availability. A better fix
1007 	 * would to coalesce pages as they are included in the cage
1008 	 */
1009 	if (SEGKMEM_USE_LARGEPAGES) {
1010 		extern void page_freelist_coalesce_all(int mnode);
1011 		page_freelist_coalesce_all(-1);	/* do all mnodes */
1012 	}
1013 
1014 	ksp = kstat_create("kcage", 0, "kcage_page_list", "misc",
1015 	    KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VAR_SIZE | KSTAT_FLAG_VIRTUAL);
1016 	if (ksp != NULL) {
1017 		ksp->ks_update = kcage_kstat_update;
1018 		ksp->ks_snapshot = kcage_kstat_snapshot;
1019 		ksp->ks_lock = &kcage_kstat_lock; /* XXX - not really needed */
1020 		kstat_install(ksp);
1021 	}
1022 }
1023 
1024 static int
1025 kcage_kstat_update(kstat_t *ksp, int rw)
1026 {
1027 	struct kcage_glist *lp;
1028 	uint_t count;
1029 
1030 	if (rw == KSTAT_WRITE)
1031 		return (EACCES);
1032 
1033 	count = 0;
1034 	rw_enter(&kcage_range_rwlock, RW_WRITER);
1035 	for (lp = kcage_glist; lp != NULL; lp = lp->next) {
1036 		if (lp->decr) {
1037 			if (lp->curr != lp->lim) {
1038 				count++;
1039 			}
1040 		} else {
1041 			if (lp->curr != lp->base) {
1042 				count++;
1043 			}
1044 		}
1045 	}
1046 	rw_exit(&kcage_range_rwlock);
1047 
1048 	ksp->ks_ndata = count;
1049 	ksp->ks_data_size = count * 2 * sizeof (uint64_t);
1050 
1051 	return (0);
1052 }
1053 
1054 static int
1055 kcage_kstat_snapshot(kstat_t *ksp, void *buf, int rw)
1056 {
1057 	struct kcage_glist *lp;
1058 	struct memunit {
1059 		uint64_t address;
1060 		uint64_t size;
1061 	} *kspmem;
1062 
1063 	if (rw == KSTAT_WRITE)
1064 		return (EACCES);
1065 
1066 	ksp->ks_snaptime = gethrtime();
1067 
1068 	kspmem = (struct memunit *)buf;
1069 	rw_enter(&kcage_range_rwlock, RW_WRITER);
1070 	for (lp = kcage_glist; lp != NULL; lp = lp->next, kspmem++) {
1071 		if ((caddr_t)kspmem >= (caddr_t)buf + ksp->ks_data_size)
1072 			break;
1073 
1074 		if (lp->decr) {
1075 			if (lp->curr != lp->lim) {
1076 				kspmem->address = ptob(lp->curr);
1077 				kspmem->size = ptob(lp->lim - lp->curr);
1078 			}
1079 		} else {
1080 			if (lp->curr != lp->base) {
1081 				kspmem->address = ptob(lp->base);
1082 				kspmem->size = ptob(lp->curr - lp->base);
1083 			}
1084 		}
1085 	}
1086 	rw_exit(&kcage_range_rwlock);
1087 
1088 	return (0);
1089 }
1090 
1091 void
1092 kcage_recalc_thresholds()
1093 {
1094 	static int first = 1;
1095 	static pgcnt_t init_lotsfree;
1096 	static pgcnt_t init_desfree;
1097 	static pgcnt_t init_minfree;
1098 	static pgcnt_t init_throttlefree;
1099 	static pgcnt_t init_reserve;
1100 
1101 	/* TODO: any reason to take more care than this with live editing? */
1102 	mutex_enter(&kcage_cageout_mutex);
1103 	mutex_enter(&freemem_lock);
1104 
1105 	if (first) {
1106 		first = 0;
1107 		init_lotsfree = kcage_lotsfree;
1108 		init_desfree = kcage_desfree;
1109 		init_minfree = kcage_minfree;
1110 		init_throttlefree = kcage_throttlefree;
1111 		init_reserve = kcage_reserve;
1112 	} else {
1113 		kcage_lotsfree = init_lotsfree;
1114 		kcage_desfree = init_desfree;
1115 		kcage_minfree = init_minfree;
1116 		kcage_throttlefree = init_throttlefree;
1117 		kcage_reserve = init_reserve;
1118 	}
1119 
1120 	if (kcage_lotsfree == 0)
1121 		kcage_lotsfree = MAX(32, total_pages / 256);
1122 
1123 	if (kcage_minfree == 0)
1124 		kcage_minfree = MAX(32, kcage_lotsfree / 2);
1125 
1126 	if (kcage_desfree == 0)
1127 		kcage_desfree = MAX(32, kcage_minfree);
1128 
1129 	if (kcage_throttlefree == 0)
1130 		kcage_throttlefree = MAX(32, kcage_minfree / 2);
1131 
1132 	if (kcage_reserve == 0)
1133 		kcage_reserve = MIN(32, kcage_throttlefree / 2);
1134 
1135 	mutex_exit(&freemem_lock);
1136 	mutex_exit(&kcage_cageout_mutex);
1137 
1138 	if (kcage_cageout_ready) {
1139 		if (kcage_freemem < kcage_desfree)
1140 			kcage_cageout_wakeup();
1141 
1142 		if (kcage_needfree) {
1143 			mutex_enter(&kcage_throttle_mutex);
1144 			cv_broadcast(&kcage_throttle_cv);
1145 			mutex_exit(&kcage_throttle_mutex);
1146 		}
1147 	}
1148 }
1149 
1150 /*
1151  * Pageout interface:
1152  * kcage_cageout_init()
1153  */
1154 void
1155 kcage_cageout_init()
1156 {
1157 	if (kcage_on) {
1158 
1159 		(void) thread_create(NULL, 0, kcage_cageout,
1160 		    NULL, 0, proc_pageout, TS_RUN, maxclsyspri - 1);
1161 	}
1162 }
1163 
1164 
1165 /*
1166  * VM Interfaces:
1167  * kcage_create_throttle()
1168  * kcage_freemem_add()
1169  * kcage_freemem_sub()
1170  */
1171 
1172 /*
1173  * Wakeup cageout thread and throttle waiting for the number of pages
1174  * requested to become available.  For non-critical requests, a
1175  * timeout is added, since freemem accounting is separate from cage
1176  * freemem accounting: it's possible for us to get stuck and not make
1177  * forward progress even though there was sufficient freemem before
1178  * arriving here.
1179  */
1180 int
1181 kcage_create_throttle(pgcnt_t npages, int flags)
1182 {
1183 	int niter = 0;
1184 	pgcnt_t lastfree;
1185 	int enough = kcage_freemem > kcage_throttlefree + npages;
1186 
1187 	KCAGE_STAT_INCR(kct_calls);		/* unprotected incr. */
1188 
1189 	kcage_cageout_wakeup();			/* just to be sure */
1190 	KCAGE_STAT_INCR(kct_cagewake);		/* unprotected incr. */
1191 
1192 	/*
1193 	 * Obviously, we can't throttle the cageout thread since
1194 	 * we depend on it.  We also can't throttle the panic thread.
1195 	 */
1196 	if (curthread == kcage_cageout_thread || panicstr) {
1197 		KCAGE_STAT_INCR(kct_cageout);	/* unprotected incr. */
1198 		return (KCT_CRIT);
1199 	}
1200 
1201 	/*
1202 	 * Don't throttle threads which are critical for proper
1203 	 * vm management if we're above kcage_throttlefree or
1204 	 * if freemem is very low.
1205 	 */
1206 	if (NOMEMWAIT()) {
1207 		if (enough) {
1208 			KCAGE_STAT_INCR(kct_exempt);	/* unprotected incr. */
1209 			return (KCT_CRIT);
1210 		} else if (freemem < minfree) {
1211 			KCAGE_STAT_INCR(kct_critical);  /* unprotected incr. */
1212 			return (KCT_CRIT);
1213 		}
1214 	}
1215 
1216 	/*
1217 	 * Don't throttle real-time threads if kcage_freemem > kcage_reserve.
1218 	 */
1219 	if (DISP_PRIO(curthread) > maxclsyspri &&
1220 	    kcage_freemem > kcage_reserve) {
1221 		KCAGE_STAT_INCR(kct_exempt);	/* unprotected incr. */
1222 		return (KCT_CRIT);
1223 	}
1224 
1225 	/*
1226 	 * Cause all other threads (which are assumed to not be
1227 	 * critical to cageout) to wait here until their request
1228 	 * can be satisfied. Be a little paranoid and wake the
1229 	 * kernel cage on each loop through this logic.
1230 	 */
1231 	while (kcage_freemem < kcage_throttlefree + npages) {
1232 		ASSERT(kcage_on);
1233 
1234 		lastfree = kcage_freemem;
1235 
1236 		if (kcage_cageout_ready) {
1237 			mutex_enter(&kcage_throttle_mutex);
1238 
1239 			kcage_needfree += npages;
1240 			KCAGE_STAT_INCR(kct_wait);
1241 
1242 			kcage_cageout_wakeup();
1243 			KCAGE_STAT_INCR(kct_cagewake);
1244 
1245 			cv_wait(&kcage_throttle_cv, &kcage_throttle_mutex);
1246 
1247 			kcage_needfree -= npages;
1248 
1249 			mutex_exit(&kcage_throttle_mutex);
1250 		} else {
1251 			/*
1252 			 * NOTE: atomics are used just in case we enter
1253 			 * mp operation before the cageout thread is ready.
1254 			 */
1255 			atomic_add_long(&kcage_needfree, npages);
1256 
1257 			kcage_cageout_wakeup();
1258 			KCAGE_STAT_INCR(kct_cagewake);	/* unprotected incr. */
1259 
1260 			atomic_add_long(&kcage_needfree, -npages);
1261 		}
1262 
1263 		if ((flags & PG_WAIT) == 0) {
1264 			if (kcage_freemem > lastfree) {
1265 				KCAGE_STAT_INCR(kct_progress);
1266 				niter = 0;
1267 			} else {
1268 				KCAGE_STAT_INCR(kct_noprogress);
1269 				if (++niter >= kcage_maxwait) {
1270 					KCAGE_STAT_INCR(kct_timeout);
1271 					return (KCT_FAILURE);
1272 				}
1273 			}
1274 		}
1275 
1276 		if (NOMEMWAIT() && freemem < minfree) {
1277 			return (KCT_CRIT);
1278 		}
1279 
1280 	}
1281 	return (KCT_NONCRIT);
1282 }
1283 
1284 void
1285 kcage_freemem_add(pgcnt_t npages)
1286 {
1287 	extern void wakeup_pcgs(void);
1288 
1289 	atomic_add_long(&kcage_freemem, npages);
1290 
1291 	wakeup_pcgs();  /* wakeup threads in pcgs() */
1292 
1293 	if (kcage_needfree != 0 &&
1294 	    kcage_freemem >= (kcage_throttlefree + kcage_needfree)) {
1295 
1296 		mutex_enter(&kcage_throttle_mutex);
1297 		cv_broadcast(&kcage_throttle_cv);
1298 		KCAGE_STAT_INCR(kfa_trottlewake);
1299 		mutex_exit(&kcage_throttle_mutex);
1300 	}
1301 }
1302 
1303 void
1304 kcage_freemem_sub(pgcnt_t npages)
1305 {
1306 	atomic_add_long(&kcage_freemem, -npages);
1307 
1308 	if (kcage_freemem < kcage_desfree) {
1309 		kcage_cageout_wakeup();
1310 		KCAGE_STAT_INCR(kfs_cagewake); /* unprotected incr. */
1311 	}
1312 }
1313 
1314 /*
1315  * return 0 on failure and 1 on success.
1316  */
1317 static int
1318 kcage_setnoreloc_pages(page_t *rootpp, se_t se)
1319 {
1320 	pgcnt_t npgs, i;
1321 	page_t *pp;
1322 	pfn_t rootpfn = page_pptonum(rootpp);
1323 	uint_t szc;
1324 
1325 	ASSERT(!PP_ISFREE(rootpp));
1326 	ASSERT(PAGE_LOCKED_SE(rootpp, se));
1327 	if (!group_page_trylock(rootpp, se)) {
1328 		return (0);
1329 	}
1330 	szc = rootpp->p_szc;
1331 	if (szc == 0) {
1332 		/*
1333 		 * The szc of a locked page can only change for pages that are
1334 		 * non-swapfs (i.e. anonymous memory) file system pages.
1335 		 */
1336 		ASSERT(rootpp->p_vnode != NULL &&
1337 		    !PP_ISKAS(rootpp) &&
1338 		    !IS_SWAPFSVP(rootpp->p_vnode));
1339 		PP_SETNORELOC(rootpp);
1340 		return (1);
1341 	}
1342 	npgs = page_get_pagecnt(szc);
1343 	ASSERT(IS_P2ALIGNED(rootpfn, npgs));
1344 	pp = rootpp;
1345 	for (i = 0; i < npgs; i++, pp++) {
1346 		ASSERT(PAGE_LOCKED_SE(pp, se));
1347 		ASSERT(!PP_ISFREE(pp));
1348 		ASSERT(pp->p_szc == szc);
1349 		PP_SETNORELOC(pp);
1350 	}
1351 	group_page_unlock(rootpp);
1352 	return (1);
1353 }
1354 
1355 /*
1356  * Attempt to convert page to a caged page (set the P_NORELOC flag).
1357  * If successful and pages is free, move page to the tail of whichever
1358  * list it is on.
1359  * Returns:
1360  *   EBUSY  page already locked, assimilated but not free.
1361  *   ENOMEM page assimilated, but memory too low to relocate. Page not free.
1362  *   EAGAIN page not assimilated. Page not free.
1363  *   ERANGE page assimilated. Page not root.
1364  *   0      page assimilated. Page free.
1365  *   *nfreedp number of pages freed.
1366  * NOTE: With error codes ENOMEM, EBUSY, and 0 (zero), there is no way
1367  * to distinguish between a page that was already a NORELOC page from
1368  * those newly converted to NORELOC pages by this invocation of
1369  * kcage_assimilate_page.
1370  */
1371 static int
1372 kcage_assimilate_page(page_t *pp, pgcnt_t *nfreedp)
1373 {
1374 	if (page_trylock(pp, SE_EXCL)) {
1375 		if (PP_ISNORELOC(pp)) {
1376 check_free_and_return:
1377 			if (PP_ISFREE(pp)) {
1378 				page_unlock(pp);
1379 				*nfreedp = 0;
1380 				return (0);
1381 			} else {
1382 				page_unlock(pp);
1383 				return (EBUSY);
1384 			}
1385 			/*NOTREACHED*/
1386 		}
1387 	} else {
1388 		if (page_trylock(pp, SE_SHARED)) {
1389 			if (PP_ISNORELOC(pp))
1390 				goto check_free_and_return;
1391 		} else
1392 			return (EAGAIN);
1393 
1394 		if (!PP_ISFREE(pp)) {
1395 			page_unlock(pp);
1396 			return (EAGAIN);
1397 		}
1398 
1399 		/*
1400 		 * Need to upgrade the lock on it and set the NORELOC
1401 		 * bit. If it is free then remove it from the free
1402 		 * list so that the platform free list code can keep
1403 		 * NORELOC pages where they should be.
1404 		 */
1405 		/*
1406 		 * Before doing anything, get the exclusive lock.
1407 		 * This may fail (eg ISM pages are left shared locked).
1408 		 * If the page is free this will leave a hole in the
1409 		 * cage. There is no solution yet to this.
1410 		 */
1411 		if (!page_tryupgrade(pp)) {
1412 			page_unlock(pp);
1413 			return (EAGAIN);
1414 		}
1415 	}
1416 
1417 	ASSERT(PAGE_EXCL(pp));
1418 
1419 	if (PP_ISFREE(pp)) {
1420 		int which = PP_ISAGED(pp) ? PG_FREE_LIST : PG_CACHE_LIST;
1421 
1422 		page_list_sub(pp, which);
1423 		ASSERT(pp->p_szc == 0);
1424 		PP_SETNORELOC(pp);
1425 		PLCNT_XFER_NORELOC(pp);
1426 		page_list_add(pp, which | PG_LIST_TAIL);
1427 
1428 		page_unlock(pp);
1429 		*nfreedp = 1;
1430 		return (0);
1431 	} else {
1432 		if (pp->p_szc != 0) {
1433 			if (!kcage_setnoreloc_pages(pp, SE_EXCL)) {
1434 				page_unlock(pp);
1435 				return (EAGAIN);
1436 			}
1437 			ASSERT(PP_ISNORELOC(pp));
1438 		} else {
1439 			PP_SETNORELOC(pp);
1440 		}
1441 		PLCNT_XFER_NORELOC(pp);
1442 		return (kcage_invalidate_page(pp, nfreedp));
1443 	}
1444 	/*NOTREACHED*/
1445 }
1446 
1447 static int
1448 kcage_expand()
1449 {
1450 	int did_something = 0;
1451 
1452 	spgcnt_t wanted;
1453 	pfn_t pfn;
1454 	page_t *pp;
1455 	/* TODO: we don't really need n any more? */
1456 	pgcnt_t n;
1457 	pgcnt_t nf, nfreed;
1458 
1459 	/*
1460 	 * Expand the cage if available cage memory is really low. Calculate
1461 	 * the amount required to return kcage_freemem to the level of
1462 	 * kcage_lotsfree, or to satisfy throttled requests, whichever is
1463 	 * more.  It is rare for their sum to create an artificial threshold
1464 	 * above kcage_lotsfree, but it is possible.
1465 	 *
1466 	 * Exit early if expansion amount is equal to or less than zero.
1467 	 * (<0 is possible if kcage_freemem rises suddenly.)
1468 	 *
1469 	 * Exit early when the global page pool (apparently) does not
1470 	 * have enough free pages to page_relocate() even a single page.
1471 	 */
1472 	wanted = MAX(kcage_lotsfree, kcage_throttlefree + kcage_needfree)
1473 	    - kcage_freemem;
1474 	if (wanted <= 0)
1475 		return (0);
1476 	else if (freemem < pageout_reserve + 1) {
1477 		KCAGE_STAT_INCR(ke_lowfreemem);
1478 		return (0);
1479 	}
1480 
1481 	KCAGE_STAT_INCR(ke_calls);
1482 	KCAGE_STAT_SET_SCAN(ke_wanted, (uint_t)wanted);
1483 
1484 	/*
1485 	 * Assimilate more pages from the global page pool into the cage.
1486 	 */
1487 	n = 0;				/* number of pages PP_SETNORELOC'd */
1488 	nf = 0;				/* number of those actually free */
1489 	while (kcage_on && nf < wanted) {
1490 		pfn = kcage_get_pfn(1);
1491 		if (pfn == PFN_INVALID) {	/* eek! no where to grow */
1492 			KCAGE_STAT_INCR(ke_nopfn);
1493 			goto terminate;
1494 		}
1495 
1496 		KCAGE_STAT_INCR_SCAN(ke_examined);
1497 
1498 		if ((pp = page_numtopp_nolock(pfn)) == NULL) {
1499 			KCAGE_STAT_INCR(ke_nopaget);
1500 			continue;
1501 		}
1502 		KCAGEPAGETS_INC();
1503 		/*
1504 		 * Sanity check. Skip this pfn if it is
1505 		 * being deleted.
1506 		 */
1507 		if (pfn_is_being_deleted(pfn)) {
1508 			KCAGE_STAT_INCR(ke_deleting);
1509 			continue;
1510 		}
1511 
1512 		if (PP_ISNORELOC(pp)) {
1513 			KCAGE_STAT_INCR(ke_isnoreloc);
1514 			continue;
1515 		}
1516 
1517 		switch (kcage_assimilate_page(pp, &nfreed)) {
1518 			case 0:		/* assimilated, page is free */
1519 				KCAGE_STAT_NINCR_SCAN(ke_gotonefree, nfreed);
1520 				did_something = 1;
1521 				nf += nfreed;
1522 				n++;
1523 				break;
1524 
1525 			case EBUSY:	/* assimilated, page not free */
1526 			case ERANGE:	/* assimilated, page not root */
1527 				KCAGE_STAT_INCR_SCAN(ke_gotone);
1528 				did_something = 1;
1529 				n++;
1530 				break;
1531 
1532 			case ENOMEM:	/* assimilated, but no mem */
1533 				KCAGE_STAT_INCR(ke_terminate);
1534 				did_something = 1;
1535 				n++;
1536 				goto terminate;
1537 
1538 			case EAGAIN:	/* can't assimilate */
1539 				KCAGE_STAT_INCR_SCAN(ke_lefthole);
1540 				break;
1541 
1542 			default:	/* catch this with debug kernels */
1543 				ASSERT(0);
1544 				break;
1545 		}
1546 	}
1547 
1548 	/*
1549 	 * Realign cage edge with the nearest physical address
1550 	 * boundry for big pages. This is done to give us a
1551 	 * better chance of actually getting usable big pages
1552 	 * in the cage.
1553 	 */
1554 
1555 terminate:
1556 
1557 	return (did_something);
1558 }
1559 
1560 /*
1561  * Relocate page opp (Original Page Pointer) from cage pool to page rpp
1562  * (Replacement Page Pointer) in the global pool. Page opp will be freed
1563  * if relocation is successful, otherwise it is only unlocked.
1564  * On entry, page opp must be exclusively locked and not free.
1565  * *nfreedp: number of pages freed.
1566  */
1567 static int
1568 kcage_relocate_page(page_t *pp, pgcnt_t *nfreedp)
1569 {
1570 	page_t *opp = pp;
1571 	page_t *rpp = NULL;
1572 	spgcnt_t npgs;
1573 	int result;
1574 
1575 	ASSERT(!PP_ISFREE(opp));
1576 	ASSERT(PAGE_EXCL(opp));
1577 
1578 	result = page_relocate(&opp, &rpp, 1, 1, &npgs, NULL);
1579 	*nfreedp = npgs;
1580 	if (result == 0) {
1581 		while (npgs-- > 0) {
1582 			page_t *tpp;
1583 
1584 			ASSERT(rpp != NULL);
1585 			tpp = rpp;
1586 			page_sub(&rpp, tpp);
1587 			page_unlock(tpp);
1588 		}
1589 
1590 		ASSERT(rpp == NULL);
1591 
1592 		return (0);		/* success */
1593 	}
1594 
1595 	page_unlock(opp);
1596 	return (result);
1597 }
1598 
1599 /*
1600  * Based on page_invalidate_pages()
1601  *
1602  * Kcage_invalidate_page() uses page_relocate() twice. Both instances
1603  * of use must be updated to match the new page_relocate() when it
1604  * becomes available.
1605  *
1606  * Return result of kcage_relocate_page or zero if page was directly freed.
1607  * *nfreedp: number of pages freed.
1608  */
1609 static int
1610 kcage_invalidate_page(page_t *pp, pgcnt_t *nfreedp)
1611 {
1612 	int result;
1613 
1614 #if defined(__sparc)
1615 	extern struct vnode prom_ppages;
1616 	ASSERT(pp->p_vnode != &prom_ppages);
1617 #endif /* __sparc */
1618 
1619 	ASSERT(!PP_ISFREE(pp));
1620 	ASSERT(PAGE_EXCL(pp));
1621 
1622 	/*
1623 	 * Is this page involved in some I/O? shared?
1624 	 * The page_struct_lock need not be acquired to
1625 	 * examine these fields since the page has an
1626 	 * "exclusive" lock.
1627 	 */
1628 	if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
1629 		result = kcage_relocate_page(pp, nfreedp);
1630 #ifdef KCAGE_STATS
1631 		if (result == 0)
1632 			KCAGE_STAT_INCR_SCAN(kip_reloclocked);
1633 		else if (result == ENOMEM)
1634 			KCAGE_STAT_INCR_SCAN(kip_nomem);
1635 #endif
1636 		return (result);
1637 	}
1638 
1639 	ASSERT(pp->p_vnode->v_type != VCHR);
1640 
1641 	/*
1642 	 * Unload the mappings and check if mod bit is set.
1643 	 */
1644 	(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
1645 
1646 	if (hat_ismod(pp)) {
1647 		result = kcage_relocate_page(pp, nfreedp);
1648 #ifdef KCAGE_STATS
1649 		if (result == 0)
1650 			KCAGE_STAT_INCR_SCAN(kip_relocmod);
1651 		else if (result == ENOMEM)
1652 			KCAGE_STAT_INCR_SCAN(kip_nomem);
1653 #endif
1654 		return (result);
1655 	}
1656 
1657 	if (!page_try_demote_pages(pp)) {
1658 		KCAGE_STAT_INCR_SCAN(kip_demotefailed);
1659 		page_unlock(pp);
1660 		return (EAGAIN);
1661 	}
1662 
1663 	/* LINTED: constant in conditional context */
1664 	VN_DISPOSE(pp, B_INVAL, 0, kcred);
1665 	KCAGE_STAT_INCR_SCAN(kip_destroy);
1666 	*nfreedp = 1;
1667 	return (0);
1668 }
1669 
1670 static void
1671 kcage_cageout()
1672 {
1673 	pfn_t pfn;
1674 	page_t *pp;
1675 	callb_cpr_t cprinfo;
1676 	int did_something;
1677 	int scan_again;
1678 	pfn_t start_pfn;
1679 	int pass;
1680 	int last_pass;
1681 	int pages_skipped;
1682 	int shared_skipped;
1683 	ulong_t shared_level = 8;
1684 	pgcnt_t nfreed;
1685 #ifdef KCAGE_STATS
1686 	clock_t scan_start;
1687 #endif
1688 
1689 	CALLB_CPR_INIT(&cprinfo, &kcage_cageout_mutex,
1690 	    callb_generic_cpr, "cageout");
1691 
1692 	mutex_enter(&kcage_cageout_mutex);
1693 	kcage_cageout_thread = curthread;
1694 
1695 	pfn = PFN_INVALID;		/* force scan reset */
1696 	start_pfn = PFN_INVALID;	/* force init with 1st cage pfn */
1697 	kcage_cageout_ready = 1;	/* switch kcage_cageout_wakeup mode */
1698 
1699 loop:
1700 	/*
1701 	 * Wait here. Sooner or later, kcage_freemem_sub() will notice
1702 	 * that kcage_freemem is less than kcage_desfree. When it does
1703 	 * notice, kcage_freemem_sub() will wake us up via call to
1704 	 * kcage_cageout_wakeup().
1705 	 */
1706 	CALLB_CPR_SAFE_BEGIN(&cprinfo);
1707 	cv_wait(&kcage_cageout_cv, &kcage_cageout_mutex);
1708 	CALLB_CPR_SAFE_END(&cprinfo, &kcage_cageout_mutex);
1709 
1710 	KCAGE_STAT_INCR(kt_wakeups);
1711 	KCAGE_STAT_SET_SCAN(kt_freemem_start, freemem);
1712 	KCAGE_STAT_SET_SCAN(kt_kcage_freemem_start, kcage_freemem);
1713 	pass = 0;
1714 	last_pass = 0;
1715 
1716 #ifdef KCAGE_STATS
1717 	scan_start = lbolt;
1718 #endif
1719 
1720 again:
1721 	if (!kcage_on)
1722 		goto loop;
1723 
1724 	KCAGE_STAT_INCR(kt_scans);
1725 	KCAGE_STAT_INCR_SCAN(kt_passes);
1726 
1727 	did_something = 0;
1728 	pages_skipped = 0;
1729 	shared_skipped = 0;
1730 	while ((kcage_freemem < kcage_lotsfree || kcage_needfree) &&
1731 	    (pfn = kcage_walk_cage(pfn == PFN_INVALID)) != PFN_INVALID) {
1732 
1733 		if (start_pfn == PFN_INVALID)
1734 			start_pfn = pfn;
1735 		else if (start_pfn == pfn) {
1736 			last_pass = pass;
1737 			pass += 1;
1738 			/*
1739 			 * Did a complete walk of kernel cage, but didn't free
1740 			 * any pages.  If only one cpu is active then
1741 			 * stop kernel cage walk and try expanding.
1742 			 */
1743 			if (cp_default.cp_ncpus == 1 && did_something == 0) {
1744 				KCAGE_STAT_INCR(kt_cageout_break);
1745 				break;
1746 			}
1747 		}
1748 
1749 		pp = page_numtopp_nolock(pfn);
1750 		if (pp == NULL) {
1751 			continue;
1752 		}
1753 
1754 		KCAGE_STAT_INCR_SCAN(kt_examined);
1755 
1756 		/*
1757 		 * Do a quick PP_ISNORELOC() and PP_ISFREE test outside
1758 		 * of the lock. If one is missed it will be seen next
1759 		 * time through.
1760 		 *
1761 		 * Skip non-caged-pages. These pages can exist in the cage
1762 		 * because, if during cage expansion, a page is
1763 		 * encountered that is long-term locked the lock prevents the
1764 		 * expansion logic from setting the P_NORELOC flag. Hence,
1765 		 * non-caged-pages surrounded by caged-pages.
1766 		 */
1767 		if (!PP_ISNORELOC(pp)) {
1768 			switch (kcage_assimilate_page(pp, &nfreed)) {
1769 				case 0:
1770 					did_something = 1;
1771 					KCAGE_STAT_NINCR_SCAN(kt_gotonefree,
1772 					    nfreed);
1773 					break;
1774 
1775 				case EBUSY:
1776 				case ERANGE:
1777 					did_something = 1;
1778 					KCAGE_STAT_INCR_SCAN(kt_gotone);
1779 					break;
1780 
1781 				case EAGAIN:
1782 				case ENOMEM:
1783 					break;
1784 
1785 				default:
1786 					/* catch this with debug kernels */
1787 					ASSERT(0);
1788 					break;
1789 			}
1790 
1791 			continue;
1792 		} else {
1793 			int prm;
1794 
1795 			if (PP_ISFREE(pp)) {
1796 				continue;
1797 			}
1798 
1799 			if ((PP_ISKAS(pp) && pp->p_lckcnt > 0) ||
1800 			    !page_trylock(pp, SE_EXCL)) {
1801 				KCAGE_STAT_INCR_SCAN(kt_cantlock);
1802 				continue;
1803 			}
1804 
1805 			/* P_NORELOC bit should not have gone away. */
1806 			ASSERT(PP_ISNORELOC(pp));
1807 			if (PP_ISFREE(pp) || (PP_ISKAS(pp) &&
1808 			    pp->p_lckcnt > 0)) {
1809 				page_unlock(pp);
1810 				continue;
1811 			}
1812 
1813 			KCAGE_STAT_SET_SCAN(kt_skiplevel, shared_level);
1814 			if (hat_page_checkshare(pp, shared_level)) {
1815 				page_unlock(pp);
1816 				pages_skipped = 1;
1817 				shared_skipped = 1;
1818 				KCAGE_STAT_INCR_SCAN(kt_skipshared);
1819 				continue;
1820 			}
1821 
1822 			/*
1823 			 * In pass {0, 1}, skip page if ref bit is set.
1824 			 * In pass {0, 1, 2}, skip page if mod bit is set.
1825 			 */
1826 			prm = hat_pagesync(pp,
1827 			    HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
1828 
1829 			/* On first pass ignore ref'd pages */
1830 			if (pass <= 1 && (prm & P_REF)) {
1831 				KCAGE_STAT_INCR_SCAN(kt_skiprefd);
1832 				pages_skipped = 1;
1833 				page_unlock(pp);
1834 				continue;
1835 			}
1836 
1837 			/* On pass 2, VN_DISPOSE if mod bit is not set */
1838 			if (pass <= 2) {
1839 				if (pp->p_szc != 0 || (prm & P_MOD) ||
1840 				    pp->p_lckcnt || pp->p_cowcnt) {
1841 					pages_skipped = 1;
1842 					page_unlock(pp);
1843 				} else {
1844 
1845 					/*
1846 					 * unload the mappings before
1847 					 * checking if mod bit is set
1848 					 */
1849 					(void) hat_pageunload(pp,
1850 					    HAT_FORCE_PGUNLOAD);
1851 
1852 					/*
1853 					 * skip this page if modified
1854 					 */
1855 					if (hat_ismod(pp)) {
1856 						pages_skipped = 1;
1857 						page_unlock(pp);
1858 						continue;
1859 					}
1860 
1861 					KCAGE_STAT_INCR_SCAN(kt_destroy);
1862 					/* constant in conditional context */
1863 					/* LINTED */
1864 					VN_DISPOSE(pp, B_INVAL, 0, kcred);
1865 					did_something = 1;
1866 				}
1867 				continue;
1868 			}
1869 
1870 			if (kcage_invalidate_page(pp, &nfreed) == 0) {
1871 				did_something = 1;
1872 				KCAGE_STAT_NINCR_SCAN(kt_gotonefree, nfreed);
1873 			}
1874 
1875 			/*
1876 			 * No need to drop the page lock here.
1877 			 * Kcage_invalidate_page has done that for us
1878 			 * either explicitly or through a page_free.
1879 			 */
1880 		}
1881 	}
1882 
1883 	/*
1884 	 * Expand the cage only if available cage memory is really low.
1885 	 * This test is done only after a complete scan of the cage.
1886 	 * The reason for not checking and expanding more often is to
1887 	 * avoid rapid expansion of the cage. Naturally, scanning the
1888 	 * cage takes time. So by scanning first, we use that work as a
1889 	 * delay loop in between expand decisions.
1890 	 */
1891 
1892 	scan_again = 0;
1893 	if (kcage_freemem < kcage_minfree || kcage_needfree) {
1894 		/*
1895 		 * Kcage_expand() will return a non-zero value if it was
1896 		 * able to expand the cage -- whether or not the new
1897 		 * pages are free and immediately usable. If non-zero,
1898 		 * we do another scan of the cage. The pages might be
1899 		 * freed during that scan or by time we get back here.
1900 		 * If not, we will attempt another expansion.
1901 		 * However, if kcage_expand() returns zero, then it was
1902 		 * unable to expand the cage. This is the case when the
1903 		 * the growth list is exausted, therefore no work was done
1904 		 * and there is no reason to scan the cage again.
1905 		 * Note: Kernel cage scan is not repeated when only one
1906 		 * cpu is active to avoid kernel cage thread hogging cpu.
1907 		 */
1908 		if (pass <= 3 && pages_skipped && cp_default.cp_ncpus > 1)
1909 			scan_again = 1;
1910 		else
1911 			(void) kcage_expand(); /* don't scan again */
1912 	} else if (kcage_freemem < kcage_lotsfree) {
1913 		/*
1914 		 * If available cage memory is less than abundant
1915 		 * and a full scan of the cage has not yet been completed,
1916 		 * or a scan has completed and some work was performed,
1917 		 * or pages were skipped because of sharing,
1918 		 * or we simply have not yet completed two passes,
1919 		 * then do another scan.
1920 		 */
1921 		if (pass <= 2 && pages_skipped)
1922 			scan_again = 1;
1923 		if (pass == last_pass || did_something)
1924 			scan_again = 1;
1925 		else if (shared_skipped && shared_level < (8<<24)) {
1926 			shared_level <<= 1;
1927 			scan_again = 1;
1928 		}
1929 	}
1930 
1931 	if (scan_again && cp_default.cp_ncpus > 1)
1932 		goto again;
1933 	else {
1934 		if (shared_level > 8)
1935 			shared_level >>= 1;
1936 
1937 		KCAGE_STAT_SET_SCAN(kt_freemem_end, freemem);
1938 		KCAGE_STAT_SET_SCAN(kt_kcage_freemem_end, kcage_freemem);
1939 		KCAGE_STAT_SET_SCAN(kt_ticks, lbolt - scan_start);
1940 		KCAGE_STAT_INC_SCAN_INDEX;
1941 		goto loop;
1942 	}
1943 
1944 	/*NOTREACHED*/
1945 }
1946 
1947 void
1948 kcage_cageout_wakeup()
1949 {
1950 	if (mutex_tryenter(&kcage_cageout_mutex)) {
1951 		if (kcage_cageout_ready) {
1952 			cv_signal(&kcage_cageout_cv);
1953 		} else if (kcage_freemem < kcage_minfree || kcage_needfree) {
1954 			/*
1955 			 * Available cage memory is really low. Time to
1956 			 * start expanding the cage. However, the
1957 			 * kernel cage thread is not yet ready to
1958 			 * do the work. Use *this* thread, which is
1959 			 * most likely to be t0, to do the work.
1960 			 */
1961 			KCAGE_STAT_INCR(kcw_expandearly);
1962 			(void) kcage_expand();
1963 			KCAGE_STAT_INC_SCAN_INDEX;
1964 		}
1965 
1966 		mutex_exit(&kcage_cageout_mutex);
1967 	}
1968 	/* else, kernel cage thread is already running */
1969 }
1970 
1971 void
1972 kcage_tick()
1973 {
1974 	/*
1975 	 * Once per second we wake up all the threads throttled
1976 	 * waiting for cage memory, in case we've become stuck
1977 	 * and haven't made forward progress expanding the cage.
1978 	 */
1979 	if (kcage_on && kcage_cageout_ready)
1980 		cv_broadcast(&kcage_throttle_cv);
1981 }
1982