xref: /titanic_41/usr/src/uts/i86pc/vm/vm_machdep.c (revision 8696d418011068e5cedf3a229f7a6613e7798e92)
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 2007 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  * Portions of this source code were derived from Berkeley 4.3 BSD
31  * under license from the Regents of the University of California.
32  */
33 
34 #pragma ident	"%Z%%M%	%I%	%E% SMI"
35 
36 /*
37  * UNIX machine dependent virtual memory support.
38  */
39 
40 #include <sys/types.h>
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/user.h>
44 #include <sys/proc.h>
45 #include <sys/kmem.h>
46 #include <sys/vmem.h>
47 #include <sys/buf.h>
48 #include <sys/cpuvar.h>
49 #include <sys/lgrp.h>
50 #include <sys/disp.h>
51 #include <sys/vm.h>
52 #include <sys/mman.h>
53 #include <sys/vnode.h>
54 #include <sys/cred.h>
55 #include <sys/exec.h>
56 #include <sys/exechdr.h>
57 #include <sys/debug.h>
58 #include <sys/vmsystm.h>
59 
60 #include <vm/hat.h>
61 #include <vm/as.h>
62 #include <vm/seg.h>
63 #include <vm/seg_kp.h>
64 #include <vm/seg_vn.h>
65 #include <vm/page.h>
66 #include <vm/seg_kmem.h>
67 #include <vm/seg_kpm.h>
68 #include <vm/vm_dep.h>
69 
70 #include <sys/cpu.h>
71 #include <sys/vm_machparam.h>
72 #include <sys/memlist.h>
73 #include <sys/bootconf.h> /* XXX the memlist stuff belongs in memlist_plat.h */
74 #include <vm/hat_i86.h>
75 #include <sys/x86_archext.h>
76 #include <sys/elf_386.h>
77 #include <sys/cmn_err.h>
78 #include <sys/archsystm.h>
79 #include <sys/machsystm.h>
80 
81 #include <sys/vtrace.h>
82 #include <sys/ddidmareq.h>
83 #include <sys/promif.h>
84 #include <sys/memnode.h>
85 #include <sys/stack.h>
86 #include <util/qsort.h>
87 #include <sys/taskq.h>
88 
89 #ifdef __xpv
90 
91 #include <sys/hypervisor.h>
92 #include <sys/xen_mmu.h>
93 #include <sys/balloon_impl.h>
94 
95 /*
96  * domain 0 pages usable for DMA are kept pre-allocated and kept in
97  * distinct lists, ordered by increasing mfn.
98  */
99 static kmutex_t io_pool_lock;
100 static page_t *io_pool_4g;	/* pool for 32 bit dma limited devices */
101 static page_t *io_pool_16m;	/* pool for 24 bit dma limited legacy devices */
102 static long io_pool_cnt;
103 static long io_pool_cnt_max = 0;
104 #define	DEFAULT_IO_POOL_MIN	128
105 static long io_pool_cnt_min = DEFAULT_IO_POOL_MIN;
106 static long io_pool_cnt_lowater = 0;
107 static long io_pool_shrink_attempts; /* how many times did we try to shrink */
108 static long io_pool_shrinks;	/* how many times did we really shrink */
109 static long io_pool_grows;	/* how many times did we grow */
110 static mfn_t start_mfn = 1;
111 static caddr_t io_pool_kva;	/* use to alloc pages when needed */
112 
113 static int create_contig_pfnlist(uint_t);
114 
115 /*
116  * percentage of phys mem to hold in the i/o pool
117  */
118 #define	DEFAULT_IO_POOL_PCT	2
119 static long io_pool_physmem_pct = DEFAULT_IO_POOL_PCT;
120 static void page_io_pool_sub(page_t **, page_t *, page_t *);
121 
122 #endif /* __xpv */
123 
124 uint_t vac_colors = 1;
125 
126 int largepagesupport = 0;
127 extern uint_t page_create_new;
128 extern uint_t page_create_exists;
129 extern uint_t page_create_putbacks;
130 extern uint_t page_create_putbacks;
131 /*
132  * Allow users to disable the kernel's use of SSE.
133  */
134 extern int use_sse_pagecopy, use_sse_pagezero;
135 
136 /*
137  * combined memory ranges from mnode and memranges[] to manage single
138  * mnode/mtype dimension in the page lists.
139  */
140 typedef struct {
141 	pfn_t	mnr_pfnlo;
142 	pfn_t	mnr_pfnhi;
143 	int	mnr_mnode;
144 	int	mnr_memrange;		/* index into memranges[] */
145 	/* maintain page list stats */
146 	pgcnt_t	mnr_mt_clpgcnt;		/* cache list cnt */
147 	pgcnt_t	mnr_mt_flpgcnt;		/* free list cnt - small pages */
148 	pgcnt_t	mnr_mt_lgpgcnt;		/* free list cnt - large pages */
149 #ifdef DEBUG
150 	struct mnr_mts {		/* mnode/mtype szc stats */
151 		pgcnt_t	mnr_mts_pgcnt;
152 		int	mnr_mts_colors;
153 		pgcnt_t *mnr_mtsc_pgcnt;
154 	} 	*mnr_mts;
155 #endif
156 } mnoderange_t;
157 
158 #define	MEMRANGEHI(mtype)						\
159 	((mtype > 0) ? memranges[mtype - 1] - 1: physmax)
160 #define	MEMRANGELO(mtype)	(memranges[mtype])
161 
162 #define	MTYPE_FREEMEM(mt)						\
163 	(mnoderanges[mt].mnr_mt_clpgcnt +				\
164 	    mnoderanges[mt].mnr_mt_flpgcnt +				\
165 	    mnoderanges[mt].mnr_mt_lgpgcnt)
166 
167 /*
168  * As the PC architecture evolved memory up was clumped into several
169  * ranges for various historical I/O devices to do DMA.
170  * < 16Meg - ISA bus
171  * < 2Gig - ???
172  * < 4Gig - PCI bus or drivers that don't understand PAE mode
173  *
174  * These are listed in reverse order, so that we can skip over unused
175  * ranges on machines with small memories.
176  *
177  * For now under the Hypervisor, we'll only ever have one memrange.
178  */
179 #define	PFN_4GIG	0x100000
180 #define	PFN_16MEG	0x1000
181 static pfn_t arch_memranges[NUM_MEM_RANGES] = {
182     PFN_4GIG,	/* pfn range for 4G and above */
183     0x80000,	/* pfn range for 2G-4G */
184     PFN_16MEG,	/* pfn range for 16M-2G */
185     0x00000,	/* pfn range for 0-16M */
186 };
187 pfn_t *memranges = &arch_memranges[0];
188 int nranges = NUM_MEM_RANGES;
189 
190 /*
191  * This combines mem_node_config and memranges into one data
192  * structure to be used for page list management.
193  */
194 mnoderange_t	*mnoderanges;
195 int		mnoderangecnt;
196 int		mtype4g;
197 
198 /*
199  * 4g memory management variables for systems with more than 4g of memory:
200  *
201  * physical memory below 4g is required for 32bit dma devices and, currently,
202  * for kmem memory. On systems with more than 4g of memory, the pool of memory
203  * below 4g can be depleted without any paging activity given that there is
204  * likely to be sufficient memory above 4g.
205  *
206  * physmax4g is set true if the largest pfn is over 4g. The rest of the
207  * 4g memory management code is enabled only when physmax4g is true.
208  *
209  * maxmem4g is the count of the maximum number of pages on the page lists
210  * with physical addresses below 4g. It can be a lot less then 4g given that
211  * BIOS may reserve large chunks of space below 4g for hot plug pci devices,
212  * agp aperture etc.
213  *
214  * freemem4g maintains the count of the number of available pages on the
215  * page lists with physical addresses below 4g.
216  *
217  * DESFREE4G specifies the desired amount of below 4g memory. It defaults to
218  * 6% (desfree4gshift = 4) of maxmem4g.
219  *
220  * RESTRICT4G_ALLOC returns true if freemem4g falls below DESFREE4G
221  * and the amount of physical memory above 4g is greater than freemem4g.
222  * In this case, page_get_* routines will restrict below 4g allocations
223  * for requests that don't specifically require it.
224  */
225 
226 #define	LOTSFREE4G	(maxmem4g >> lotsfree4gshift)
227 #define	DESFREE4G	(maxmem4g >> desfree4gshift)
228 
229 #define	RESTRICT4G_ALLOC					\
230 	(physmax4g && (freemem4g < DESFREE4G) && ((freemem4g << 1) < freemem))
231 
232 static pgcnt_t	maxmem4g;
233 static pgcnt_t	freemem4g;
234 static int	physmax4g;
235 static int	desfree4gshift = 4;	/* maxmem4g shift to derive DESFREE4G */
236 static int	lotsfree4gshift = 3;
237 
238 /*
239  * 16m memory management:
240  *
241  * reserve some amount of physical memory below 16m for legacy devices.
242  *
243  * RESTRICT16M_ALLOC returns true if an there are sufficient free pages above
244  * 16m or if the 16m pool drops below DESFREE16M.
245  *
246  * In this case, general page allocations via page_get_{free,cache}list
247  * routines will be restricted from allocating from the 16m pool. Allocations
248  * that require specific pfn ranges (page_get_anylist) and PG_PANIC allocations
249  * are not restricted.
250  */
251 
252 #define	FREEMEM16M	MTYPE_FREEMEM(0)
253 #define	DESFREE16M	desfree16m
254 #define	RESTRICT16M_ALLOC(freemem, pgcnt, flags)		\
255 	((freemem != 0) && ((flags & PG_PANIC) == 0) &&		\
256 	    ((freemem >= (FREEMEM16M)) ||			\
257 	    (FREEMEM16M  < (DESFREE16M + pgcnt))))
258 
259 static pgcnt_t	desfree16m = 0x380;
260 
261 /*
262  * This can be patched via /etc/system to allow old non-PAE aware device
263  * drivers to use kmem_alloc'd memory on 32 bit systems with > 4Gig RAM.
264  */
265 int restricted_kmemalloc = 0;
266 
267 #ifdef VM_STATS
268 struct {
269 	ulong_t	pga_alloc;
270 	ulong_t	pga_notfullrange;
271 	ulong_t	pga_nulldmaattr;
272 	ulong_t	pga_allocok;
273 	ulong_t	pga_allocfailed;
274 	ulong_t	pgma_alloc;
275 	ulong_t	pgma_allocok;
276 	ulong_t	pgma_allocfailed;
277 	ulong_t	pgma_allocempty;
278 } pga_vmstats;
279 #endif
280 
281 uint_t mmu_page_sizes;
282 
283 /* How many page sizes the users can see */
284 uint_t mmu_exported_page_sizes;
285 
286 /*
287  * Number of pages in 1 GB.  Don't enable automatic large pages if we have
288  * fewer than this many pages.
289  */
290 pgcnt_t shm_lpg_min_physmem = 1 << (30 - MMU_PAGESHIFT);
291 pgcnt_t privm_lpg_min_physmem = 1 << (30 - MMU_PAGESHIFT);
292 
293 /*
294  * Maximum and default segment size tunables for user private
295  * and shared anon memory, and user text and initialized data.
296  * These can be patched via /etc/system to allow large pages
297  * to be used for mapping application private and shared anon memory.
298  */
299 size_t mcntl0_lpsize = MMU_PAGESIZE;
300 size_t max_uheap_lpsize = MMU_PAGESIZE;
301 size_t default_uheap_lpsize = MMU_PAGESIZE;
302 size_t max_ustack_lpsize = MMU_PAGESIZE;
303 size_t default_ustack_lpsize = MMU_PAGESIZE;
304 size_t max_privmap_lpsize = MMU_PAGESIZE;
305 size_t max_uidata_lpsize = MMU_PAGESIZE;
306 size_t max_utext_lpsize = MMU_PAGESIZE;
307 size_t max_shm_lpsize = MMU_PAGESIZE;
308 
309 
310 /*
311  * initialized by page_coloring_init().
312  */
313 uint_t	page_colors;
314 uint_t	page_colors_mask;
315 uint_t	page_coloring_shift;
316 int	cpu_page_colors;
317 static uint_t	l2_colors;
318 
319 /*
320  * Page freelists and cachelists are dynamically allocated once mnoderangecnt
321  * and page_colors are calculated from the l2 cache n-way set size.  Within a
322  * mnode range, the page freelist and cachelist are hashed into bins based on
323  * color. This makes it easier to search for a page within a specific memory
324  * range.
325  */
326 #define	PAGE_COLORS_MIN	16
327 
328 page_t ****page_freelists;
329 page_t ***page_cachelists;
330 
331 
332 /*
333  * Used by page layer to know about page sizes
334  */
335 hw_pagesize_t hw_page_array[MAX_NUM_LEVEL + 1];
336 
337 kmutex_t	*fpc_mutex[NPC_MUTEX];
338 kmutex_t	*cpc_mutex[NPC_MUTEX];
339 
340 /*
341  * Only let one thread at a time try to coalesce large pages, to
342  * prevent them from working against each other.
343  */
344 static kmutex_t	contig_lock;
345 #define	CONTIG_LOCK()	mutex_enter(&contig_lock);
346 #define	CONTIG_UNLOCK()	mutex_exit(&contig_lock);
347 
348 #define	PFN_16M		(mmu_btop((uint64_t)0x1000000))
349 
350 /*
351  * Return the optimum page size for a given mapping
352  */
353 /*ARGSUSED*/
354 size_t
355 map_pgsz(int maptype, struct proc *p, caddr_t addr, size_t len, int memcntl)
356 {
357 	level_t l = 0;
358 	size_t pgsz = MMU_PAGESIZE;
359 	size_t max_lpsize;
360 	uint_t mszc;
361 
362 	ASSERT(maptype != MAPPGSZ_VA);
363 
364 	if (maptype != MAPPGSZ_ISM && physmem < privm_lpg_min_physmem) {
365 		return (MMU_PAGESIZE);
366 	}
367 
368 	switch (maptype) {
369 	case MAPPGSZ_HEAP:
370 	case MAPPGSZ_STK:
371 		max_lpsize = memcntl ? mcntl0_lpsize : (maptype ==
372 		    MAPPGSZ_HEAP ? max_uheap_lpsize : max_ustack_lpsize);
373 		if (max_lpsize == MMU_PAGESIZE) {
374 			return (MMU_PAGESIZE);
375 		}
376 		if (len == 0) {
377 			len = (maptype == MAPPGSZ_HEAP) ? p->p_brkbase +
378 			    p->p_brksize - p->p_bssbase : p->p_stksize;
379 		}
380 		len = (maptype == MAPPGSZ_HEAP) ? MAX(len,
381 		    default_uheap_lpsize) : MAX(len, default_ustack_lpsize);
382 
383 		/*
384 		 * use the pages size that best fits len
385 		 */
386 		for (l = mmu.max_page_level; l > 0; --l) {
387 			if (LEVEL_SIZE(l) > max_lpsize || len < LEVEL_SIZE(l)) {
388 				continue;
389 			} else {
390 				pgsz = LEVEL_SIZE(l);
391 			}
392 			break;
393 		}
394 
395 		mszc = (maptype == MAPPGSZ_HEAP ? p->p_brkpageszc :
396 		    p->p_stkpageszc);
397 		if (addr == 0 && (pgsz < hw_page_array[mszc].hp_size)) {
398 			pgsz = hw_page_array[mszc].hp_size;
399 		}
400 		return (pgsz);
401 
402 	/*
403 	 * for ISM use the 1st large page size.
404 	 */
405 	case MAPPGSZ_ISM:
406 		if (mmu.max_page_level == 0)
407 			return (MMU_PAGESIZE);
408 		return (LEVEL_SIZE(1));
409 	}
410 	return (pgsz);
411 }
412 
413 static uint_t
414 map_szcvec(caddr_t addr, size_t size, uintptr_t off, size_t max_lpsize,
415     size_t min_physmem)
416 {
417 	caddr_t eaddr = addr + size;
418 	uint_t szcvec = 0;
419 	caddr_t raddr;
420 	caddr_t readdr;
421 	size_t	pgsz;
422 	int i;
423 
424 	if (physmem < min_physmem || max_lpsize <= MMU_PAGESIZE) {
425 		return (0);
426 	}
427 
428 	for (i = mmu_page_sizes - 1; i > 0; i--) {
429 		pgsz = page_get_pagesize(i);
430 		if (pgsz > max_lpsize) {
431 			continue;
432 		}
433 		raddr = (caddr_t)P2ROUNDUP((uintptr_t)addr, pgsz);
434 		readdr = (caddr_t)P2ALIGN((uintptr_t)eaddr, pgsz);
435 		if (raddr < addr || raddr >= readdr) {
436 			continue;
437 		}
438 		if (P2PHASE((uintptr_t)addr ^ off, pgsz)) {
439 			continue;
440 		}
441 		/*
442 		 * Set szcvec to the remaining page sizes.
443 		 */
444 		szcvec = ((1 << (i + 1)) - 1) & ~1;
445 		break;
446 	}
447 	return (szcvec);
448 }
449 
450 /*
451  * Return a bit vector of large page size codes that
452  * can be used to map [addr, addr + len) region.
453  */
454 /*ARGSUSED*/
455 uint_t
456 map_pgszcvec(caddr_t addr, size_t size, uintptr_t off, int flags, int type,
457     int memcntl)
458 {
459 	size_t max_lpsize = mcntl0_lpsize;
460 
461 	if (mmu.max_page_level == 0)
462 		return (0);
463 
464 	if (flags & MAP_TEXT) {
465 		if (!memcntl)
466 			max_lpsize = max_utext_lpsize;
467 		return (map_szcvec(addr, size, off, max_lpsize,
468 		    shm_lpg_min_physmem));
469 
470 	} else if (flags & MAP_INITDATA) {
471 		if (!memcntl)
472 			max_lpsize = max_uidata_lpsize;
473 		return (map_szcvec(addr, size, off, max_lpsize,
474 		    privm_lpg_min_physmem));
475 
476 	} else if (type == MAPPGSZC_SHM) {
477 		if (!memcntl)
478 			max_lpsize = max_shm_lpsize;
479 		return (map_szcvec(addr, size, off, max_lpsize,
480 		    shm_lpg_min_physmem));
481 
482 	} else if (type == MAPPGSZC_HEAP) {
483 		if (!memcntl)
484 			max_lpsize = max_uheap_lpsize;
485 		return (map_szcvec(addr, size, off, max_lpsize,
486 		    privm_lpg_min_physmem));
487 
488 	} else if (type == MAPPGSZC_STACK) {
489 		if (!memcntl)
490 			max_lpsize = max_ustack_lpsize;
491 		return (map_szcvec(addr, size, off, max_lpsize,
492 		    privm_lpg_min_physmem));
493 
494 	} else {
495 		if (!memcntl)
496 			max_lpsize = max_privmap_lpsize;
497 		return (map_szcvec(addr, size, off, max_lpsize,
498 		    privm_lpg_min_physmem));
499 	}
500 }
501 
502 /*
503  * Handle a pagefault.
504  */
505 faultcode_t
506 pagefault(
507 	caddr_t addr,
508 	enum fault_type type,
509 	enum seg_rw rw,
510 	int iskernel)
511 {
512 	struct as *as;
513 	struct hat *hat;
514 	struct proc *p;
515 	kthread_t *t;
516 	faultcode_t res;
517 	caddr_t base;
518 	size_t len;
519 	int err;
520 	int mapped_red;
521 	uintptr_t ea;
522 
523 	ASSERT_STACK_ALIGNED();
524 
525 	if (INVALID_VADDR(addr))
526 		return (FC_NOMAP);
527 
528 	mapped_red = segkp_map_red();
529 
530 	if (iskernel) {
531 		as = &kas;
532 		hat = as->a_hat;
533 	} else {
534 		t = curthread;
535 		p = ttoproc(t);
536 		as = p->p_as;
537 		hat = as->a_hat;
538 	}
539 
540 	/*
541 	 * Dispatch pagefault.
542 	 */
543 	res = as_fault(hat, as, addr, 1, type, rw);
544 
545 	/*
546 	 * If this isn't a potential unmapped hole in the user's
547 	 * UNIX data or stack segments, just return status info.
548 	 */
549 	if (res != FC_NOMAP || iskernel)
550 		goto out;
551 
552 	/*
553 	 * Check to see if we happened to faulted on a currently unmapped
554 	 * part of the UNIX data or stack segments.  If so, create a zfod
555 	 * mapping there and then try calling the fault routine again.
556 	 */
557 	base = p->p_brkbase;
558 	len = p->p_brksize;
559 
560 	if (addr < base || addr >= base + len) {		/* data seg? */
561 		base = (caddr_t)p->p_usrstack - p->p_stksize;
562 		len = p->p_stksize;
563 		if (addr < base || addr >= p->p_usrstack) {	/* stack seg? */
564 			/* not in either UNIX data or stack segments */
565 			res = FC_NOMAP;
566 			goto out;
567 		}
568 	}
569 
570 	/*
571 	 * the rest of this function implements a 3.X 4.X 5.X compatibility
572 	 * This code is probably not needed anymore
573 	 */
574 	if (p->p_model == DATAMODEL_ILP32) {
575 
576 		/* expand the gap to the page boundaries on each side */
577 		ea = P2ROUNDUP((uintptr_t)base + len, MMU_PAGESIZE);
578 		base = (caddr_t)P2ALIGN((uintptr_t)base, MMU_PAGESIZE);
579 		len = ea - (uintptr_t)base;
580 
581 		as_rangelock(as);
582 		if (as_gap(as, MMU_PAGESIZE, &base, &len, AH_CONTAIN, addr) ==
583 		    0) {
584 			err = as_map(as, base, len, segvn_create, zfod_argsp);
585 			as_rangeunlock(as);
586 			if (err) {
587 				res = FC_MAKE_ERR(err);
588 				goto out;
589 			}
590 		} else {
591 			/*
592 			 * This page is already mapped by another thread after
593 			 * we returned from as_fault() above.  We just fall
594 			 * through as_fault() below.
595 			 */
596 			as_rangeunlock(as);
597 		}
598 
599 		res = as_fault(hat, as, addr, 1, F_INVAL, rw);
600 	}
601 
602 out:
603 	if (mapped_red)
604 		segkp_unmap_red();
605 
606 	return (res);
607 }
608 
609 void
610 map_addr(caddr_t *addrp, size_t len, offset_t off, int vacalign, uint_t flags)
611 {
612 	struct proc *p = curproc;
613 	caddr_t userlimit = (flags & _MAP_LOW32) ?
614 	    (caddr_t)_userlimit32 : p->p_as->a_userlimit;
615 
616 	map_addr_proc(addrp, len, off, vacalign, userlimit, curproc, flags);
617 }
618 
619 /*ARGSUSED*/
620 int
621 map_addr_vacalign_check(caddr_t addr, u_offset_t off)
622 {
623 	return (0);
624 }
625 
626 /*
627  * map_addr_proc() is the routine called when the system is to
628  * choose an address for the user.  We will pick an address
629  * range which is the highest available below userlimit.
630  *
631  * addrp is a value/result parameter.
632  *	On input it is a hint from the user to be used in a completely
633  *	machine dependent fashion.  We decide to completely ignore this hint.
634  *
635  *	On output it is NULL if no address can be found in the current
636  *	processes address space or else an address that is currently
637  *	not mapped for len bytes with a page of red zone on either side.
638  *
639  *	align is not needed on x86 (it's for viturally addressed caches)
640  */
641 /*ARGSUSED*/
642 void
643 map_addr_proc(
644 	caddr_t *addrp,
645 	size_t len,
646 	offset_t off,
647 	int vacalign,
648 	caddr_t userlimit,
649 	struct proc *p,
650 	uint_t flags)
651 {
652 	struct as *as = p->p_as;
653 	caddr_t addr;
654 	caddr_t base;
655 	size_t slen;
656 	size_t align_amount;
657 
658 	ASSERT32(userlimit == as->a_userlimit);
659 
660 	base = p->p_brkbase;
661 #if defined(__amd64)
662 	/*
663 	 * XX64 Yes, this needs more work.
664 	 */
665 	if (p->p_model == DATAMODEL_NATIVE) {
666 		if (userlimit < as->a_userlimit) {
667 			/*
668 			 * This happens when a program wants to map
669 			 * something in a range that's accessible to a
670 			 * program in a smaller address space.  For example,
671 			 * a 64-bit program calling mmap32(2) to guarantee
672 			 * that the returned address is below 4Gbytes.
673 			 */
674 			ASSERT((uintptr_t)userlimit < ADDRESS_C(0xffffffff));
675 
676 			if (userlimit > base)
677 				slen = userlimit - base;
678 			else {
679 				*addrp = NULL;
680 				return;
681 			}
682 		} else {
683 			/*
684 			 * XX64 This layout is probably wrong .. but in
685 			 * the event we make the amd64 address space look
686 			 * like sparcv9 i.e. with the stack -above- the
687 			 * heap, this bit of code might even be correct.
688 			 */
689 			slen = p->p_usrstack - base -
690 			    (((size_t)rctl_enforced_value(
691 			    rctlproc_legacy[RLIMIT_STACK],
692 			    p->p_rctls, p) + PAGEOFFSET) & PAGEMASK);
693 		}
694 	} else
695 #endif
696 		slen = userlimit - base;
697 
698 	len = (len + PAGEOFFSET) & PAGEMASK;
699 
700 	/*
701 	 * Redzone for each side of the request. This is done to leave
702 	 * one page unmapped between segments. This is not required, but
703 	 * it's useful for the user because if their program strays across
704 	 * a segment boundary, it will catch a fault immediately making
705 	 * debugging a little easier.
706 	 */
707 	len += 2 * MMU_PAGESIZE;
708 
709 	/*
710 	 * figure out what the alignment should be
711 	 *
712 	 * XX64 -- is there an ELF_AMD64_MAXPGSZ or is it the same????
713 	 */
714 	if (len <= ELF_386_MAXPGSZ) {
715 		/*
716 		 * Align virtual addresses to ensure that ELF shared libraries
717 		 * are mapped with the appropriate alignment constraints by
718 		 * the run-time linker.
719 		 */
720 		align_amount = ELF_386_MAXPGSZ;
721 	} else {
722 		int l = mmu.max_page_level;
723 
724 		while (l && len < LEVEL_SIZE(l))
725 			--l;
726 
727 		align_amount = LEVEL_SIZE(l);
728 	}
729 
730 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount))
731 		align_amount = (uintptr_t)*addrp;
732 
733 	len += align_amount;
734 
735 	/*
736 	 * Look for a large enough hole starting below userlimit.
737 	 * After finding it, use the upper part.  Addition of PAGESIZE
738 	 * is for the redzone as described above.
739 	 */
740 	if (as_gap(as, len, &base, &slen, AH_HI, NULL) == 0) {
741 		caddr_t as_addr;
742 
743 		addr = base + slen - len + MMU_PAGESIZE;
744 		as_addr = addr;
745 		/*
746 		 * Round address DOWN to the alignment amount,
747 		 * add the offset, and if this address is less
748 		 * than the original address, add alignment amount.
749 		 */
750 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1)));
751 		addr += (uintptr_t)(off & (align_amount - 1));
752 		if (addr < as_addr)
753 			addr += align_amount;
754 
755 		ASSERT(addr <= (as_addr + align_amount));
756 		ASSERT(((uintptr_t)addr & (align_amount - 1)) ==
757 		    ((uintptr_t)(off & (align_amount - 1))));
758 		*addrp = addr;
759 	} else {
760 		*addrp = NULL;	/* no more virtual space */
761 	}
762 }
763 
764 /*
765  * Determine whether [base, base+len] contains a valid range of
766  * addresses at least minlen long. base and len are adjusted if
767  * required to provide a valid range.
768  */
769 /*ARGSUSED3*/
770 int
771 valid_va_range(caddr_t *basep, size_t *lenp, size_t minlen, int dir)
772 {
773 	uintptr_t hi, lo;
774 
775 	lo = (uintptr_t)*basep;
776 	hi = lo + *lenp;
777 
778 	/*
779 	 * If hi rolled over the top, try cutting back.
780 	 */
781 	if (hi < lo) {
782 		if (0 - lo + hi < minlen)
783 			return (0);
784 		if (0 - lo < minlen)
785 			return (0);
786 		*lenp = 0 - lo;
787 	} else if (hi - lo < minlen) {
788 		return (0);
789 	}
790 #if defined(__amd64)
791 	/*
792 	 * Deal with a possible hole in the address range between
793 	 * hole_start and hole_end that should never be mapped.
794 	 */
795 	if (lo < hole_start) {
796 		if (hi > hole_start) {
797 			if (hi < hole_end) {
798 				hi = hole_start;
799 			} else {
800 				/* lo < hole_start && hi >= hole_end */
801 				if (dir == AH_LO) {
802 					/*
803 					 * prefer lowest range
804 					 */
805 					if (hole_start - lo >= minlen)
806 						hi = hole_start;
807 					else if (hi - hole_end >= minlen)
808 						lo = hole_end;
809 					else
810 						return (0);
811 				} else {
812 					/*
813 					 * prefer highest range
814 					 */
815 					if (hi - hole_end >= minlen)
816 						lo = hole_end;
817 					else if (hole_start - lo >= minlen)
818 						hi = hole_start;
819 					else
820 						return (0);
821 				}
822 			}
823 		}
824 	} else {
825 		/* lo >= hole_start */
826 		if (hi < hole_end)
827 			return (0);
828 		if (lo < hole_end)
829 			lo = hole_end;
830 	}
831 
832 	if (hi - lo < minlen)
833 		return (0);
834 
835 	*basep = (caddr_t)lo;
836 	*lenp = hi - lo;
837 #endif
838 	return (1);
839 }
840 
841 /*
842  * Determine whether [addr, addr+len] are valid user addresses.
843  */
844 /*ARGSUSED*/
845 int
846 valid_usr_range(caddr_t addr, size_t len, uint_t prot, struct as *as,
847     caddr_t userlimit)
848 {
849 	caddr_t eaddr = addr + len;
850 
851 	if (eaddr <= addr || addr >= userlimit || eaddr > userlimit)
852 		return (RANGE_BADADDR);
853 
854 #if defined(__amd64)
855 	/*
856 	 * Check for the VA hole
857 	 */
858 	if (eaddr > (caddr_t)hole_start && addr < (caddr_t)hole_end)
859 		return (RANGE_BADADDR);
860 #endif
861 
862 	return (RANGE_OKAY);
863 }
864 
865 /*
866  * Return 1 if the page frame is onboard memory, else 0.
867  */
868 int
869 pf_is_memory(pfn_t pf)
870 {
871 	if (pfn_is_foreign(pf))
872 		return (0);
873 	return (address_in_memlist(phys_install, pfn_to_pa(pf), 1));
874 }
875 
876 /*
877  * return the memrange containing pfn
878  */
879 int
880 memrange_num(pfn_t pfn)
881 {
882 	int n;
883 
884 	for (n = 0; n < nranges - 1; ++n) {
885 		if (pfn >= memranges[n])
886 			break;
887 	}
888 	return (n);
889 }
890 
891 /*
892  * return the mnoderange containing pfn
893  */
894 /*ARGSUSED*/
895 int
896 pfn_2_mtype(pfn_t pfn)
897 {
898 #if defined(__xpv)
899 	return (0);
900 #else
901 	int	n;
902 
903 	for (n = mnoderangecnt - 1; n >= 0; n--) {
904 		if (pfn >= mnoderanges[n].mnr_pfnlo) {
905 			break;
906 		}
907 	}
908 	return (n);
909 #endif
910 }
911 
912 #if !defined(__xpv)
913 /*
914  * is_contigpage_free:
915  *	returns a page list of contiguous pages. It minimally has to return
916  *	minctg pages. Caller determines minctg based on the scatter-gather
917  *	list length.
918  *
919  *	pfnp is set to the next page frame to search on return.
920  */
921 static page_t *
922 is_contigpage_free(
923 	pfn_t *pfnp,
924 	pgcnt_t *pgcnt,
925 	pgcnt_t minctg,
926 	uint64_t pfnseg,
927 	int iolock)
928 {
929 	int	i = 0;
930 	pfn_t	pfn = *pfnp;
931 	page_t	*pp;
932 	page_t	*plist = NULL;
933 
934 	/*
935 	 * fail if pfn + minctg crosses a segment boundary.
936 	 * Adjust for next starting pfn to begin at segment boundary.
937 	 */
938 
939 	if (((*pfnp + minctg - 1) & pfnseg) < (*pfnp & pfnseg)) {
940 		*pfnp = roundup(*pfnp, pfnseg + 1);
941 		return (NULL);
942 	}
943 
944 	do {
945 retry:
946 		pp = page_numtopp_nolock(pfn + i);
947 		if ((pp == NULL) ||
948 		    (page_trylock(pp, SE_EXCL) == 0)) {
949 			(*pfnp)++;
950 			break;
951 		}
952 		if (page_pptonum(pp) != pfn + i) {
953 			page_unlock(pp);
954 			goto retry;
955 		}
956 
957 		if (!(PP_ISFREE(pp))) {
958 			page_unlock(pp);
959 			(*pfnp)++;
960 			break;
961 		}
962 
963 		if (!PP_ISAGED(pp)) {
964 			page_list_sub(pp, PG_CACHE_LIST);
965 			page_hashout(pp, (kmutex_t *)NULL);
966 		} else {
967 			page_list_sub(pp, PG_FREE_LIST);
968 		}
969 
970 		if (iolock)
971 			page_io_lock(pp);
972 		page_list_concat(&plist, &pp);
973 
974 		/*
975 		 * exit loop when pgcnt satisfied or segment boundary reached.
976 		 */
977 
978 	} while ((++i < *pgcnt) && ((pfn + i) & pfnseg));
979 
980 	*pfnp += i;		/* set to next pfn to search */
981 
982 	if (i >= minctg) {
983 		*pgcnt -= i;
984 		return (plist);
985 	}
986 
987 	/*
988 	 * failure: minctg not satisfied.
989 	 *
990 	 * if next request crosses segment boundary, set next pfn
991 	 * to search from the segment boundary.
992 	 */
993 	if (((*pfnp + minctg - 1) & pfnseg) < (*pfnp & pfnseg))
994 		*pfnp = roundup(*pfnp, pfnseg + 1);
995 
996 	/* clean up any pages already allocated */
997 
998 	while (plist) {
999 		pp = plist;
1000 		page_sub(&plist, pp);
1001 		page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL);
1002 		if (iolock)
1003 			page_io_unlock(pp);
1004 		page_unlock(pp);
1005 	}
1006 
1007 	return (NULL);
1008 }
1009 #endif	/* !__xpv */
1010 
1011 /*
1012  * verify that pages being returned from allocator have correct DMA attribute
1013  */
1014 #ifndef DEBUG
1015 #define	check_dma(a, b, c) (0)
1016 #else
1017 static void
1018 check_dma(ddi_dma_attr_t *dma_attr, page_t *pp, int cnt)
1019 {
1020 	if (dma_attr == NULL)
1021 		return;
1022 
1023 	while (cnt-- > 0) {
1024 		if (pa_to_ma(pfn_to_pa(pp->p_pagenum)) <
1025 		    dma_attr->dma_attr_addr_lo)
1026 			panic("PFN (pp=%p) below dma_attr_addr_lo", pp);
1027 		if (pa_to_ma(pfn_to_pa(pp->p_pagenum)) >=
1028 		    dma_attr->dma_attr_addr_hi)
1029 			panic("PFN (pp=%p) above dma_attr_addr_hi", pp);
1030 		pp = pp->p_next;
1031 	}
1032 }
1033 #endif
1034 
1035 #if !defined(__xpv)
1036 static page_t *
1037 page_get_contigpage(pgcnt_t *pgcnt, ddi_dma_attr_t *mattr, int iolock)
1038 {
1039 	pfn_t		pfn;
1040 	int		sgllen;
1041 	uint64_t	pfnseg;
1042 	pgcnt_t		minctg;
1043 	page_t		*pplist = NULL, *plist;
1044 	uint64_t	lo, hi;
1045 	pgcnt_t		pfnalign = 0;
1046 	static pfn_t	startpfn;
1047 	static pgcnt_t	lastctgcnt;
1048 	uintptr_t	align;
1049 
1050 	CONTIG_LOCK();
1051 
1052 	if (mattr) {
1053 		lo = mmu_btop((mattr->dma_attr_addr_lo + MMU_PAGEOFFSET));
1054 		hi = mmu_btop(mattr->dma_attr_addr_hi);
1055 		if (hi >= physmax)
1056 			hi = physmax - 1;
1057 		sgllen = mattr->dma_attr_sgllen;
1058 		pfnseg = mmu_btop(mattr->dma_attr_seg);
1059 
1060 		align = maxbit(mattr->dma_attr_align, mattr->dma_attr_minxfer);
1061 		if (align > MMU_PAGESIZE)
1062 			pfnalign = mmu_btop(align);
1063 
1064 		/*
1065 		 * in order to satisfy the request, must minimally
1066 		 * acquire minctg contiguous pages
1067 		 */
1068 		minctg = howmany(*pgcnt, sgllen);
1069 
1070 		ASSERT(hi >= lo);
1071 
1072 		/*
1073 		 * start from where last searched if the minctg >= lastctgcnt
1074 		 */
1075 		if (minctg < lastctgcnt || startpfn < lo || startpfn > hi)
1076 			startpfn = lo;
1077 	} else {
1078 		hi = physmax - 1;
1079 		lo = 0;
1080 		sgllen = 1;
1081 		pfnseg = mmu.highest_pfn;
1082 		minctg = *pgcnt;
1083 
1084 		if (minctg < lastctgcnt)
1085 			startpfn = lo;
1086 	}
1087 	lastctgcnt = minctg;
1088 
1089 	ASSERT(pfnseg + 1 >= (uint64_t)minctg);
1090 
1091 	/* conserve 16m memory - start search above 16m when possible */
1092 	if (hi > PFN_16M && startpfn < PFN_16M)
1093 		startpfn = PFN_16M;
1094 
1095 	pfn = startpfn;
1096 	if (pfnalign)
1097 		pfn = P2ROUNDUP(pfn, pfnalign);
1098 
1099 	while (pfn + minctg - 1 <= hi) {
1100 
1101 		plist = is_contigpage_free(&pfn, pgcnt, minctg, pfnseg, iolock);
1102 		if (plist) {
1103 			page_list_concat(&pplist, &plist);
1104 			sgllen--;
1105 			/*
1106 			 * return when contig pages no longer needed
1107 			 */
1108 			if (!*pgcnt || ((*pgcnt <= sgllen) && !pfnalign)) {
1109 				startpfn = pfn;
1110 				CONTIG_UNLOCK();
1111 				check_dma(mattr, pplist, *pgcnt);
1112 				return (pplist);
1113 			}
1114 			minctg = howmany(*pgcnt, sgllen);
1115 		}
1116 		if (pfnalign)
1117 			pfn = P2ROUNDUP(pfn, pfnalign);
1118 	}
1119 
1120 	/* cannot find contig pages in specified range */
1121 	if (startpfn == lo) {
1122 		CONTIG_UNLOCK();
1123 		return (NULL);
1124 	}
1125 
1126 	/* did not start with lo previously */
1127 	pfn = lo;
1128 	if (pfnalign)
1129 		pfn = P2ROUNDUP(pfn, pfnalign);
1130 
1131 	/* allow search to go above startpfn */
1132 	while (pfn < startpfn) {
1133 
1134 		plist = is_contigpage_free(&pfn, pgcnt, minctg, pfnseg, iolock);
1135 		if (plist != NULL) {
1136 
1137 			page_list_concat(&pplist, &plist);
1138 			sgllen--;
1139 
1140 			/*
1141 			 * return when contig pages no longer needed
1142 			 */
1143 			if (!*pgcnt || ((*pgcnt <= sgllen) && !pfnalign)) {
1144 				startpfn = pfn;
1145 				CONTIG_UNLOCK();
1146 				check_dma(mattr, pplist, *pgcnt);
1147 				return (pplist);
1148 			}
1149 			minctg = howmany(*pgcnt, sgllen);
1150 		}
1151 		if (pfnalign)
1152 			pfn = P2ROUNDUP(pfn, pfnalign);
1153 	}
1154 	CONTIG_UNLOCK();
1155 	return (NULL);
1156 }
1157 #endif	/* !__xpv */
1158 
1159 /*
1160  * mnode_range_cnt() calculates the number of memory ranges for mnode and
1161  * memranges[]. Used to determine the size of page lists and mnoderanges.
1162  */
1163 int
1164 mnode_range_cnt(int mnode)
1165 {
1166 #if defined(__xpv)
1167 	ASSERT(mnode == 0);
1168 	return (1);
1169 #else	/* __xpv */
1170 	int	mri;
1171 	int	mnrcnt = 0;
1172 
1173 	if (mem_node_config[mnode].exists != 0) {
1174 		mri = nranges - 1;
1175 
1176 		/* find the memranges index below contained in mnode range */
1177 
1178 		while (MEMRANGEHI(mri) < mem_node_config[mnode].physbase)
1179 			mri--;
1180 
1181 		/*
1182 		 * increment mnode range counter when memranges or mnode
1183 		 * boundary is reached.
1184 		 */
1185 		while (mri >= 0 &&
1186 		    mem_node_config[mnode].physmax >= MEMRANGELO(mri)) {
1187 			mnrcnt++;
1188 			if (mem_node_config[mnode].physmax > MEMRANGEHI(mri))
1189 				mri--;
1190 			else
1191 				break;
1192 		}
1193 	}
1194 	ASSERT(mnrcnt <= MAX_MNODE_MRANGES);
1195 	return (mnrcnt);
1196 #endif	/* __xpv */
1197 }
1198 
1199 /*
1200  * mnode_range_setup() initializes mnoderanges.
1201  */
1202 void
1203 mnode_range_setup(mnoderange_t *mnoderanges)
1204 {
1205 	int	mnode, mri;
1206 
1207 	for (mnode = 0; mnode < max_mem_nodes; mnode++) {
1208 		if (mem_node_config[mnode].exists == 0)
1209 			continue;
1210 
1211 		mri = nranges - 1;
1212 
1213 		while (MEMRANGEHI(mri) < mem_node_config[mnode].physbase)
1214 			mri--;
1215 
1216 		while (mri >= 0 && mem_node_config[mnode].physmax >=
1217 		    MEMRANGELO(mri)) {
1218 			mnoderanges->mnr_pfnlo = MAX(MEMRANGELO(mri),
1219 			    mem_node_config[mnode].physbase);
1220 			mnoderanges->mnr_pfnhi = MIN(MEMRANGEHI(mri),
1221 			    mem_node_config[mnode].physmax);
1222 			mnoderanges->mnr_mnode = mnode;
1223 			mnoderanges->mnr_memrange = mri;
1224 			mnoderanges++;
1225 			if (mem_node_config[mnode].physmax > MEMRANGEHI(mri))
1226 				mri--;
1227 			else
1228 				break;
1229 		}
1230 	}
1231 }
1232 
1233 /*ARGSUSED*/
1234 int
1235 mtype_init(vnode_t *vp, caddr_t vaddr, uint_t *flags, size_t pgsz)
1236 {
1237 	int mtype = mnoderangecnt - 1;
1238 
1239 #if !defined(__xpv)
1240 #if defined(__i386)
1241 	/*
1242 	 * set the mtype range
1243 	 * - kmem requests needs to be below 4g if restricted_kmemalloc is set.
1244 	 * - for non kmem requests, set range to above 4g if memory below 4g
1245 	 * runs low.
1246 	 */
1247 	if (restricted_kmemalloc && VN_ISKAS(vp) &&
1248 	    (caddr_t)(vaddr) >= kernelheap &&
1249 	    (caddr_t)(vaddr) < ekernelheap) {
1250 		ASSERT(physmax4g);
1251 		mtype = mtype4g;
1252 		if (RESTRICT16M_ALLOC(freemem4g - btop(pgsz),
1253 		    btop(pgsz), *flags)) {
1254 			*flags |= PGI_MT_RANGE16M;
1255 		} else {
1256 			VM_STAT_ADD(vmm_vmstats.unrestrict16mcnt);
1257 			VM_STAT_COND_ADD((*flags & PG_PANIC),
1258 			    vmm_vmstats.pgpanicalloc);
1259 			*flags |= PGI_MT_RANGE0;
1260 		}
1261 		return (mtype);
1262 	}
1263 #endif	/* __i386 */
1264 
1265 	if (RESTRICT4G_ALLOC) {
1266 		VM_STAT_ADD(vmm_vmstats.restrict4gcnt);
1267 		/* here only for > 4g systems */
1268 		*flags |= PGI_MT_RANGE4G;
1269 	} else if (RESTRICT16M_ALLOC(freemem, btop(pgsz), *flags)) {
1270 		*flags |= PGI_MT_RANGE16M;
1271 	} else {
1272 		VM_STAT_ADD(vmm_vmstats.unrestrict16mcnt);
1273 		VM_STAT_COND_ADD((*flags & PG_PANIC), vmm_vmstats.pgpanicalloc);
1274 		*flags |= PGI_MT_RANGE0;
1275 	}
1276 #endif /* !__xpv */
1277 	return (mtype);
1278 }
1279 
1280 
1281 /* mtype init for page_get_replacement_page */
1282 /*ARGSUSED*/
1283 int
1284 mtype_pgr_init(int *flags, page_t *pp, int mnode, pgcnt_t pgcnt)
1285 {
1286 	int mtype = mnoderangecnt - 1;
1287 #if !defined(__ixpv)
1288 	if (RESTRICT16M_ALLOC(freemem, pgcnt, *flags)) {
1289 		*flags |= PGI_MT_RANGE16M;
1290 	} else {
1291 		VM_STAT_ADD(vmm_vmstats.unrestrict16mcnt);
1292 		*flags |= PGI_MT_RANGE0;
1293 	}
1294 #endif
1295 	return (mtype);
1296 }
1297 
1298 /*
1299  * Determine if the mnode range specified in mtype contains memory belonging
1300  * to memory node mnode.  If flags & PGI_MT_RANGE is set then mtype contains
1301  * the range of indices from high pfn to 0, 16m or 4g.
1302  *
1303  * Return first mnode range type index found otherwise return -1 if none found.
1304  */
1305 int
1306 mtype_func(int mnode, int mtype, uint_t flags)
1307 {
1308 	if (flags & PGI_MT_RANGE) {
1309 		int	mtlim = 0;
1310 
1311 		if (flags & PGI_MT_NEXT)
1312 			mtype--;
1313 		if (flags & PGI_MT_RANGE4G)
1314 			mtlim = mtype4g + 1;	/* exclude 0-4g range */
1315 		else if (flags & PGI_MT_RANGE16M)
1316 			mtlim = 1;		/* exclude 0-16m range */
1317 		while (mtype >= mtlim) {
1318 			if (mnoderanges[mtype].mnr_mnode == mnode)
1319 				return (mtype);
1320 			mtype--;
1321 		}
1322 	} else if (mnoderanges[mtype].mnr_mnode == mnode) {
1323 		return (mtype);
1324 	}
1325 	return (-1);
1326 }
1327 
1328 /*
1329  * Update the page list max counts with the pfn range specified by the
1330  * input parameters.  Called from add_physmem() when physical memory with
1331  * page_t's are initially added to the page lists.
1332  */
1333 void
1334 mtype_modify_max(pfn_t startpfn, long cnt)
1335 {
1336 	int	mtype = 0;
1337 	pfn_t	endpfn = startpfn + cnt, pfn;
1338 	pgcnt_t	inc;
1339 
1340 	ASSERT(cnt > 0);
1341 
1342 	if (!physmax4g)
1343 		return;
1344 
1345 	for (pfn = startpfn; pfn < endpfn; ) {
1346 		if (pfn <= mnoderanges[mtype].mnr_pfnhi) {
1347 			if (endpfn < mnoderanges[mtype].mnr_pfnhi) {
1348 				inc = endpfn - pfn;
1349 			} else {
1350 				inc = mnoderanges[mtype].mnr_pfnhi - pfn + 1;
1351 			}
1352 			if (mtype <= mtype4g)
1353 				maxmem4g += inc;
1354 			pfn += inc;
1355 		}
1356 		mtype++;
1357 		ASSERT(mtype < mnoderangecnt || pfn >= endpfn);
1358 	}
1359 }
1360 
1361 int
1362 mtype_2_mrange(int mtype)
1363 {
1364 	return (mnoderanges[mtype].mnr_memrange);
1365 }
1366 
1367 void
1368 mnodetype_2_pfn(int mnode, int mtype, pfn_t *pfnlo, pfn_t *pfnhi)
1369 {
1370 	ASSERT(mnoderanges[mtype].mnr_mnode == mnode);
1371 	*pfnlo = mnoderanges[mtype].mnr_pfnlo;
1372 	*pfnhi = mnoderanges[mtype].mnr_pfnhi;
1373 }
1374 
1375 size_t
1376 plcnt_sz(size_t ctrs_sz)
1377 {
1378 #ifdef DEBUG
1379 	int	szc, colors;
1380 
1381 	ctrs_sz += mnoderangecnt * sizeof (struct mnr_mts) * mmu_page_sizes;
1382 	for (szc = 0; szc < mmu_page_sizes; szc++) {
1383 		colors = page_get_pagecolors(szc);
1384 		ctrs_sz += mnoderangecnt * sizeof (pgcnt_t) * colors;
1385 	}
1386 #endif
1387 	return (ctrs_sz);
1388 }
1389 
1390 caddr_t
1391 plcnt_init(caddr_t addr)
1392 {
1393 #ifdef DEBUG
1394 	int	mt, szc, colors;
1395 
1396 	for (mt = 0; mt < mnoderangecnt; mt++) {
1397 		mnoderanges[mt].mnr_mts = (struct mnr_mts *)addr;
1398 		addr += (sizeof (struct mnr_mts) * mmu_page_sizes);
1399 		for (szc = 0; szc < mmu_page_sizes; szc++) {
1400 			colors = page_get_pagecolors(szc);
1401 			mnoderanges[mt].mnr_mts[szc].mnr_mts_colors = colors;
1402 			mnoderanges[mt].mnr_mts[szc].mnr_mtsc_pgcnt =
1403 			    (pgcnt_t *)addr;
1404 			addr += (sizeof (pgcnt_t) * colors);
1405 		}
1406 	}
1407 #endif
1408 	return (addr);
1409 }
1410 
1411 void
1412 plcnt_inc_dec(page_t *pp, int mtype, int szc, long cnt, int flags)
1413 {
1414 #ifdef DEBUG
1415 	int	bin = PP_2_BIN(pp);
1416 
1417 	atomic_add_long(&mnoderanges[mtype].mnr_mts[szc].mnr_mts_pgcnt, cnt);
1418 	atomic_add_long(&mnoderanges[mtype].mnr_mts[szc].mnr_mtsc_pgcnt[bin],
1419 	    cnt);
1420 #endif
1421 	ASSERT(mtype == PP_2_MTYPE(pp));
1422 	if (physmax4g && mtype <= mtype4g)
1423 		atomic_add_long(&freemem4g, cnt);
1424 	if (flags & PG_CACHE_LIST)
1425 		atomic_add_long(&mnoderanges[mtype].mnr_mt_clpgcnt, cnt);
1426 	else if (szc)
1427 		atomic_add_long(&mnoderanges[mtype].mnr_mt_lgpgcnt, cnt);
1428 	else
1429 		atomic_add_long(&mnoderanges[mtype].mnr_mt_flpgcnt, cnt);
1430 }
1431 
1432 /*
1433  * Returns the free page count for mnode
1434  */
1435 int
1436 mnode_pgcnt(int mnode)
1437 {
1438 	int	mtype = mnoderangecnt - 1;
1439 	int	flags = PGI_MT_RANGE0;
1440 	pgcnt_t	pgcnt = 0;
1441 
1442 	mtype = mtype_func(mnode, mtype, flags);
1443 
1444 	while (mtype != -1) {
1445 		pgcnt += MTYPE_FREEMEM(mtype);
1446 		mtype = mtype_func(mnode, mtype, flags | PGI_MT_NEXT);
1447 	}
1448 	return (pgcnt);
1449 }
1450 
1451 /*
1452  * Initialize page coloring variables based on the l2 cache parameters.
1453  * Calculate and return memory needed for page coloring data structures.
1454  */
1455 size_t
1456 page_coloring_init(uint_t l2_sz, int l2_linesz, int l2_assoc)
1457 {
1458 	size_t	colorsz = 0;
1459 	int	i;
1460 	int	colors;
1461 
1462 #if defined(__xpv)
1463 	/*
1464 	 * Hypervisor domains currently don't have any concept of NUMA.
1465 	 * Hence we'll act like there is only 1 memrange.
1466 	 */
1467 	i = memrange_num(1);
1468 #else /* !__xpv */
1469 	/*
1470 	 * Reduce the memory ranges lists if we don't have large amounts
1471 	 * of memory. This avoids searching known empty free lists.
1472 	 */
1473 	i = memrange_num(physmax);
1474 #if defined(__i386)
1475 	if (i > 0)
1476 		restricted_kmemalloc = 0;
1477 #endif
1478 	/* physmax greater than 4g */
1479 	if (i == 0)
1480 		physmax4g = 1;
1481 #endif /* !__xpv */
1482 	memranges += i;
1483 	nranges -= i;
1484 
1485 	ASSERT(ISP2(l2_sz));
1486 	ASSERT(ISP2(l2_linesz));
1487 	ASSERT(l2_sz > MMU_PAGESIZE);
1488 
1489 	/* l2_assoc is 0 for fully associative l2 cache */
1490 	if (l2_assoc)
1491 		l2_colors = MAX(1, l2_sz / (l2_assoc * MMU_PAGESIZE));
1492 	else
1493 		l2_colors = 1;
1494 
1495 	/* for scalability, configure at least PAGE_COLORS_MIN color bins */
1496 	page_colors = MAX(l2_colors, PAGE_COLORS_MIN);
1497 
1498 	/*
1499 	 * cpu_page_colors is non-zero when a page color may be spread across
1500 	 * multiple bins.
1501 	 */
1502 	if (l2_colors < page_colors)
1503 		cpu_page_colors = l2_colors;
1504 
1505 	ASSERT(ISP2(page_colors));
1506 
1507 	page_colors_mask = page_colors - 1;
1508 
1509 	ASSERT(ISP2(CPUSETSIZE()));
1510 	page_coloring_shift = lowbit(CPUSETSIZE());
1511 
1512 	/* initialize number of colors per page size */
1513 	for (i = 0; i <= mmu.max_page_level; i++) {
1514 		hw_page_array[i].hp_size = LEVEL_SIZE(i);
1515 		hw_page_array[i].hp_shift = LEVEL_SHIFT(i);
1516 		hw_page_array[i].hp_pgcnt = LEVEL_SIZE(i) >> LEVEL_SHIFT(0);
1517 		hw_page_array[i].hp_colors = (page_colors_mask >>
1518 		    (hw_page_array[i].hp_shift - hw_page_array[0].hp_shift))
1519 		    + 1;
1520 		colorequivszc[i] = 0;
1521 	}
1522 
1523 	/*
1524 	 * The value of cpu_page_colors determines if additional color bins
1525 	 * need to be checked for a particular color in the page_get routines.
1526 	 */
1527 	if (cpu_page_colors != 0) {
1528 
1529 		int a = lowbit(page_colors) - lowbit(cpu_page_colors);
1530 		ASSERT(a > 0);
1531 		ASSERT(a < 16);
1532 
1533 		for (i = 0; i <= mmu.max_page_level; i++) {
1534 			if ((colors = hw_page_array[i].hp_colors) <= 1) {
1535 				colorequivszc[i] = 0;
1536 				continue;
1537 			}
1538 			while ((colors >> a) == 0)
1539 				a--;
1540 			ASSERT(a >= 0);
1541 
1542 			/* higher 4 bits encodes color equiv mask */
1543 			colorequivszc[i] = (a << 4);
1544 		}
1545 	}
1546 
1547 	/* factor in colorequiv to check additional 'equivalent' bins. */
1548 	if (colorequiv > 1) {
1549 
1550 		int a = lowbit(colorequiv) - 1;
1551 		if (a > 15)
1552 			a = 15;
1553 
1554 		for (i = 0; i <= mmu.max_page_level; i++) {
1555 			if ((colors = hw_page_array[i].hp_colors) <= 1) {
1556 				continue;
1557 			}
1558 			while ((colors >> a) == 0)
1559 				a--;
1560 			if ((a << 4) > colorequivszc[i]) {
1561 				colorequivszc[i] = (a << 4);
1562 			}
1563 		}
1564 	}
1565 
1566 	/* size for mnoderanges */
1567 	for (mnoderangecnt = 0, i = 0; i < max_mem_nodes; i++)
1568 		mnoderangecnt += mnode_range_cnt(i);
1569 	colorsz = mnoderangecnt * sizeof (mnoderange_t);
1570 
1571 	/* size for fpc_mutex and cpc_mutex */
1572 	colorsz += (2 * max_mem_nodes * sizeof (kmutex_t) * NPC_MUTEX);
1573 
1574 	/* size of page_freelists */
1575 	colorsz += mnoderangecnt * sizeof (page_t ***);
1576 	colorsz += mnoderangecnt * mmu_page_sizes * sizeof (page_t **);
1577 
1578 	for (i = 0; i < mmu_page_sizes; i++) {
1579 		colors = page_get_pagecolors(i);
1580 		colorsz += mnoderangecnt * colors * sizeof (page_t *);
1581 	}
1582 
1583 	/* size of page_cachelists */
1584 	colorsz += mnoderangecnt * sizeof (page_t **);
1585 	colorsz += mnoderangecnt * page_colors * sizeof (page_t *);
1586 
1587 	return (colorsz);
1588 }
1589 
1590 /*
1591  * Called once at startup to configure page_coloring data structures and
1592  * does the 1st page_free()/page_freelist_add().
1593  */
1594 void
1595 page_coloring_setup(caddr_t pcmemaddr)
1596 {
1597 	int	i;
1598 	int	j;
1599 	int	k;
1600 	caddr_t	addr;
1601 	int	colors;
1602 
1603 	/*
1604 	 * do page coloring setup
1605 	 */
1606 	addr = pcmemaddr;
1607 
1608 	mnoderanges = (mnoderange_t *)addr;
1609 	addr += (mnoderangecnt * sizeof (mnoderange_t));
1610 
1611 	mnode_range_setup(mnoderanges);
1612 
1613 	if (physmax4g)
1614 		mtype4g = pfn_2_mtype(0xfffff);
1615 
1616 	for (k = 0; k < NPC_MUTEX; k++) {
1617 		fpc_mutex[k] = (kmutex_t *)addr;
1618 		addr += (max_mem_nodes * sizeof (kmutex_t));
1619 	}
1620 	for (k = 0; k < NPC_MUTEX; k++) {
1621 		cpc_mutex[k] = (kmutex_t *)addr;
1622 		addr += (max_mem_nodes * sizeof (kmutex_t));
1623 	}
1624 	page_freelists = (page_t ****)addr;
1625 	addr += (mnoderangecnt * sizeof (page_t ***));
1626 
1627 	page_cachelists = (page_t ***)addr;
1628 	addr += (mnoderangecnt * sizeof (page_t **));
1629 
1630 	for (i = 0; i < mnoderangecnt; i++) {
1631 		page_freelists[i] = (page_t ***)addr;
1632 		addr += (mmu_page_sizes * sizeof (page_t **));
1633 
1634 		for (j = 0; j < mmu_page_sizes; j++) {
1635 			colors = page_get_pagecolors(j);
1636 			page_freelists[i][j] = (page_t **)addr;
1637 			addr += (colors * sizeof (page_t *));
1638 		}
1639 		page_cachelists[i] = (page_t **)addr;
1640 		addr += (page_colors * sizeof (page_t *));
1641 	}
1642 }
1643 
1644 #if defined(__xpv)
1645 /*
1646  * Give back 10% of the io_pool pages to the free list.
1647  * Don't shrink the pool below some absolute minimum.
1648  */
1649 static void
1650 page_io_pool_shrink()
1651 {
1652 	int retcnt;
1653 	page_t *pp, *pp_first, *pp_last, **curpool;
1654 	mfn_t mfn;
1655 	int bothpools = 0;
1656 
1657 	mutex_enter(&io_pool_lock);
1658 	io_pool_shrink_attempts++;	/* should be a kstat? */
1659 	retcnt = io_pool_cnt / 10;
1660 	if (io_pool_cnt - retcnt < io_pool_cnt_min)
1661 		retcnt = io_pool_cnt - io_pool_cnt_min;
1662 	if (retcnt <= 0)
1663 		goto done;
1664 	io_pool_shrinks++;	/* should be a kstat? */
1665 	curpool = &io_pool_4g;
1666 domore:
1667 	/*
1668 	 * Loop through taking pages from the end of the list
1669 	 * (highest mfns) till amount to return reached.
1670 	 */
1671 	for (pp = *curpool; pp && retcnt > 0; ) {
1672 		pp_first = pp_last = pp->p_prev;
1673 		if (pp_first == *curpool)
1674 			break;
1675 		retcnt--;
1676 		io_pool_cnt--;
1677 		page_io_pool_sub(curpool, pp_first, pp_last);
1678 		if ((mfn = pfn_to_mfn(pp->p_pagenum)) < start_mfn)
1679 			start_mfn = mfn;
1680 		page_free(pp_first, 1);
1681 		pp = *curpool;
1682 	}
1683 	if (retcnt != 0 && !bothpools) {
1684 		/*
1685 		 * If not enough found in less constrained pool try the
1686 		 * more constrained one.
1687 		 */
1688 		curpool = &io_pool_16m;
1689 		bothpools = 1;
1690 		goto domore;
1691 	}
1692 done:
1693 	mutex_exit(&io_pool_lock);
1694 }
1695 
1696 #endif	/* __xpv */
1697 
1698 uint_t
1699 page_create_update_flags_x86(uint_t flags)
1700 {
1701 #if defined(__xpv)
1702 	/*
1703 	 * Check this is an urgent allocation and free pages are depleted.
1704 	 */
1705 	if (!(flags & PG_WAIT) && freemem < desfree)
1706 		page_io_pool_shrink();
1707 #else /* !__xpv */
1708 	/*
1709 	 * page_create_get_something may call this because 4g memory may be
1710 	 * depleted. Set flags to allow for relocation of base page below
1711 	 * 4g if necessary.
1712 	 */
1713 	if (physmax4g)
1714 		flags |= (PGI_PGCPSZC0 | PGI_PGCPHIPRI);
1715 #endif /* __xpv */
1716 	return (flags);
1717 }
1718 
1719 /*ARGSUSED*/
1720 int
1721 bp_color(struct buf *bp)
1722 {
1723 	return (0);
1724 }
1725 
1726 #if defined(__xpv)
1727 
1728 /*
1729  * Take pages out of an io_pool
1730  */
1731 static void
1732 page_io_pool_sub(page_t **poolp, page_t *pp_first, page_t *pp_last)
1733 {
1734 	if (*poolp == pp_first) {
1735 		*poolp = pp_last->p_next;
1736 		if (*poolp == pp_first)
1737 			*poolp = NULL;
1738 	}
1739 	pp_first->p_prev->p_next = pp_last->p_next;
1740 	pp_last->p_next->p_prev = pp_first->p_prev;
1741 	pp_first->p_prev = pp_last;
1742 	pp_last->p_next = pp_first;
1743 }
1744 
1745 /*
1746  * Put a page on the io_pool list. The list is ordered by increasing MFN.
1747  */
1748 static void
1749 page_io_pool_add(page_t **poolp, page_t *pp)
1750 {
1751 	page_t	*look;
1752 	mfn_t	mfn = mfn_list[pp->p_pagenum];
1753 
1754 	if (*poolp == NULL) {
1755 		*poolp = pp;
1756 		pp->p_next = pp;
1757 		pp->p_prev = pp;
1758 		return;
1759 	}
1760 
1761 	/*
1762 	 * Since we try to take pages from the high end of the pool
1763 	 * chances are good that the pages to be put on the list will
1764 	 * go at or near the end of the list. so start at the end and
1765 	 * work backwards.
1766 	 */
1767 	look = (*poolp)->p_prev;
1768 	while (mfn < mfn_list[look->p_pagenum]) {
1769 		look = look->p_prev;
1770 		if (look == (*poolp)->p_prev)
1771 			break; /* backed all the way to front of list */
1772 	}
1773 
1774 	/* insert after look */
1775 	pp->p_prev = look;
1776 	pp->p_next = look->p_next;
1777 	pp->p_next->p_prev = pp;
1778 	look->p_next = pp;
1779 	if (mfn < mfn_list[(*poolp)->p_pagenum]) {
1780 		/*
1781 		 * we inserted a new first list element
1782 		 * adjust pool pointer to newly inserted element
1783 		 */
1784 		*poolp = pp;
1785 	}
1786 }
1787 
1788 /*
1789  * Add a page to the io_pool.  Setting the force flag will force the page
1790  * into the io_pool no matter what.
1791  */
1792 static void
1793 add_page_to_pool(page_t *pp, int force)
1794 {
1795 	page_t *highest;
1796 	page_t *freep = NULL;
1797 
1798 	mutex_enter(&io_pool_lock);
1799 	/*
1800 	 * Always keep the scarce low memory pages
1801 	 */
1802 	if (mfn_list[pp->p_pagenum] < PFN_16MEG) {
1803 		++io_pool_cnt;
1804 		page_io_pool_add(&io_pool_16m, pp);
1805 		goto done;
1806 	}
1807 	if (io_pool_cnt < io_pool_cnt_max || force) {
1808 		++io_pool_cnt;
1809 		page_io_pool_add(&io_pool_4g, pp);
1810 	} else {
1811 		highest = io_pool_4g->p_prev;
1812 		if (mfn_list[pp->p_pagenum] < mfn_list[highest->p_pagenum]) {
1813 			page_io_pool_sub(&io_pool_4g, highest, highest);
1814 			page_io_pool_add(&io_pool_4g, pp);
1815 			freep = highest;
1816 		} else {
1817 			freep = pp;
1818 		}
1819 	}
1820 done:
1821 	mutex_exit(&io_pool_lock);
1822 	if (freep)
1823 		page_free(freep, 1);
1824 }
1825 
1826 
1827 int contig_pfn_cnt;	/* no of pfns in the contig pfn list */
1828 int contig_pfn_max;	/* capacity of the contig pfn list */
1829 int next_alloc_pfn;	/* next position in list to start a contig search */
1830 int contig_pfnlist_updates;	/* pfn list update count */
1831 int contig_pfnlist_locked;	/* contig pfn list locked against use */
1832 int contig_pfnlist_builds;	/* how many times have we (re)built list */
1833 int contig_pfnlist_buildfailed;	/* how many times has list build failed */
1834 int create_contig_pending;	/* nonzero means taskq creating contig list */
1835 pfn_t *contig_pfn_list = NULL;	/* list of contig pfns in ascending mfn order */
1836 
1837 /*
1838  * Function to use in sorting a list of pfns by their underlying mfns.
1839  */
1840 static int
1841 mfn_compare(const void *pfnp1, const void *pfnp2)
1842 {
1843 	mfn_t mfn1 = mfn_list[*(pfn_t *)pfnp1];
1844 	mfn_t mfn2 = mfn_list[*(pfn_t *)pfnp2];
1845 
1846 	if (mfn1 > mfn2)
1847 		return (1);
1848 	if (mfn1 < mfn2)
1849 		return (-1);
1850 	return (0);
1851 }
1852 
1853 /*
1854  * Compact the contig_pfn_list by tossing all the non-contiguous
1855  * elements from the list.
1856  */
1857 static void
1858 compact_contig_pfn_list(void)
1859 {
1860 	pfn_t pfn, lapfn, prev_lapfn;
1861 	mfn_t mfn;
1862 	int i, newcnt = 0;
1863 
1864 	prev_lapfn = 0;
1865 	for (i = 0; i < contig_pfn_cnt - 1; i++) {
1866 		pfn = contig_pfn_list[i];
1867 		lapfn = contig_pfn_list[i + 1];
1868 		mfn = mfn_list[pfn];
1869 		/*
1870 		 * See if next pfn is for a contig mfn
1871 		 */
1872 		if (mfn_list[lapfn] != mfn + 1)
1873 			continue;
1874 		/*
1875 		 * pfn and lookahead are both put in list
1876 		 * unless pfn is the previous lookahead.
1877 		 */
1878 		if (pfn != prev_lapfn)
1879 			contig_pfn_list[newcnt++] = pfn;
1880 		contig_pfn_list[newcnt++] = lapfn;
1881 		prev_lapfn = lapfn;
1882 	}
1883 	for (i = newcnt; i < contig_pfn_cnt; i++)
1884 		contig_pfn_list[i] = 0;
1885 	contig_pfn_cnt = newcnt;
1886 }
1887 
1888 /*ARGSUSED*/
1889 static void
1890 call_create_contiglist(void *arg)
1891 {
1892 	mutex_enter(&io_pool_lock);
1893 	(void) create_contig_pfnlist(PG_WAIT);
1894 	create_contig_pending = 0;
1895 	mutex_exit(&io_pool_lock);
1896 }
1897 
1898 /*
1899  * Create list of freelist pfns that have underlying
1900  * contiguous mfns.  The list is kept in ascending mfn order.
1901  * returns 1 if list created else 0.
1902  */
1903 static int
1904 create_contig_pfnlist(uint_t flags)
1905 {
1906 	pfn_t pfn;
1907 	page_t *pp;
1908 
1909 	if (contig_pfn_list != NULL)
1910 		return (1);
1911 	ASSERT(!contig_pfnlist_locked);
1912 	contig_pfn_max = freemem + (freemem / 10);
1913 	contig_pfn_list = kmem_zalloc(contig_pfn_max * sizeof (pfn_t),
1914 	    (flags & PG_WAIT) ? KM_SLEEP : KM_NOSLEEP);
1915 	if (contig_pfn_list == NULL) {
1916 		/*
1917 		 * If we could not create the contig list (because
1918 		 * we could not sleep for memory).  Dispatch a taskq that can
1919 		 * sleep to get the memory.
1920 		 */
1921 		if (!create_contig_pending) {
1922 			if (taskq_dispatch(system_taskq, call_create_contiglist,
1923 			    NULL, TQ_NOSLEEP) != NULL)
1924 				create_contig_pending = 1;
1925 		}
1926 		contig_pfnlist_buildfailed++;	/* count list build failures */
1927 		return (0);
1928 	}
1929 	ASSERT(contig_pfn_cnt == 0);
1930 	for (pfn = 0; pfn < mfn_count; pfn++) {
1931 		pp = page_numtopp_nolock(pfn);
1932 		if (pp == NULL || !PP_ISFREE(pp))
1933 			continue;
1934 		contig_pfn_list[contig_pfn_cnt] = pfn;
1935 		if (++contig_pfn_cnt == contig_pfn_max)
1936 			break;
1937 	}
1938 	qsort(contig_pfn_list, contig_pfn_cnt, sizeof (pfn_t), mfn_compare);
1939 	compact_contig_pfn_list();
1940 	/*
1941 	 * Make sure next search of the newly created contiguous pfn
1942 	 * list starts at the beginning of the list.
1943 	 */
1944 	next_alloc_pfn = 0;
1945 	contig_pfnlist_builds++;	/* count list builds */
1946 	return (1);
1947 }
1948 
1949 
1950 /*
1951  * Toss the current contig pfnlist.  Someone is about to do a massive
1952  * update to pfn<->mfn mappings.  So we have them destroy the list and lock
1953  * it till they are done with their update.
1954  */
1955 void
1956 clear_and_lock_contig_pfnlist()
1957 {
1958 	pfn_t *listp = NULL;
1959 	size_t listsize;
1960 
1961 	mutex_enter(&io_pool_lock);
1962 	ASSERT(!contig_pfnlist_locked);
1963 	if (contig_pfn_list != NULL) {
1964 		listp = contig_pfn_list;
1965 		listsize = contig_pfn_max * sizeof (pfn_t);
1966 		contig_pfn_list = NULL;
1967 		contig_pfn_max = contig_pfn_cnt = 0;
1968 	}
1969 	contig_pfnlist_locked = 1;
1970 	mutex_exit(&io_pool_lock);
1971 	if (listp != NULL)
1972 		kmem_free(listp, listsize);
1973 }
1974 
1975 /*
1976  * Unlock the contig_pfn_list.  The next attempted use of it will cause
1977  * it to be re-created.
1978  */
1979 void
1980 unlock_contig_pfnlist()
1981 {
1982 	mutex_enter(&io_pool_lock);
1983 	ASSERT(contig_pfnlist_locked);
1984 	contig_pfnlist_locked = 0;
1985 	mutex_exit(&io_pool_lock);
1986 }
1987 
1988 /*
1989  * Update the contiguous pfn list in response to a pfn <-> mfn reassignment
1990  */
1991 void
1992 update_contig_pfnlist(pfn_t pfn, mfn_t oldmfn, mfn_t newmfn)
1993 {
1994 	int probe_hi, probe_lo, probe_pos, insert_after, insert_point;
1995 	pfn_t probe_pfn;
1996 	mfn_t probe_mfn;
1997 
1998 	if (contig_pfn_list == NULL)
1999 		return;
2000 	mutex_enter(&io_pool_lock);
2001 	contig_pfnlist_updates++;
2002 	/*
2003 	 * Find the pfn in the current list.  Use a binary chop to locate it.
2004 	 */
2005 	probe_hi = contig_pfn_cnt - 1;
2006 	probe_lo = 0;
2007 	probe_pos = (probe_hi + probe_lo) / 2;
2008 	while ((probe_pfn = contig_pfn_list[probe_pos]) != pfn) {
2009 		if (probe_pos == probe_lo) { /* pfn not in list */
2010 			probe_pos = -1;
2011 			break;
2012 		}
2013 		if (pfn_to_mfn(probe_pfn) <= oldmfn)
2014 			probe_lo = probe_pos;
2015 		else
2016 			probe_hi = probe_pos;
2017 		probe_pos = (probe_hi + probe_lo) / 2;
2018 	}
2019 	if (probe_pos >= 0)  { /* remove pfn fom list */
2020 		contig_pfn_cnt--;
2021 		ovbcopy(&contig_pfn_list[probe_pos + 1],
2022 		    &contig_pfn_list[probe_pos],
2023 		    (contig_pfn_cnt - probe_pos) * sizeof (pfn_t));
2024 	}
2025 	if (newmfn == MFN_INVALID)
2026 		goto done;
2027 	/*
2028 	 * Check if new mfn has adjacent mfns in the list
2029 	 */
2030 	probe_hi = contig_pfn_cnt - 1;
2031 	probe_lo = 0;
2032 	insert_after = -2;
2033 	do {
2034 		probe_pos = (probe_hi + probe_lo) / 2;
2035 		probe_mfn = pfn_to_mfn(contig_pfn_list[probe_pos]);
2036 		if (newmfn == probe_mfn + 1)
2037 			insert_after = probe_pos;
2038 		else if (newmfn == probe_mfn - 1)
2039 			insert_after = probe_pos - 1;
2040 		if (probe_pos == probe_lo)
2041 			break;
2042 		if (probe_mfn <= newmfn)
2043 			probe_lo = probe_pos;
2044 		else
2045 			probe_hi = probe_pos;
2046 	} while (insert_after == -2);
2047 	/*
2048 	 * If there is space in the list and there are adjacent mfns
2049 	 * insert the pfn in to its proper place in the list.
2050 	 */
2051 	if (insert_after != -2 && contig_pfn_cnt + 1 <= contig_pfn_max) {
2052 		insert_point = insert_after + 1;
2053 		ovbcopy(&contig_pfn_list[insert_point],
2054 		    &contig_pfn_list[insert_point + 1],
2055 		    (contig_pfn_cnt - insert_point) * sizeof (pfn_t));
2056 		contig_pfn_list[insert_point] = pfn;
2057 		contig_pfn_cnt++;
2058 	}
2059 done:
2060 	mutex_exit(&io_pool_lock);
2061 }
2062 
2063 /*
2064  * Called to (re-)populate the io_pool from the free page lists.
2065  */
2066 long
2067 populate_io_pool(void)
2068 {
2069 	pfn_t pfn;
2070 	mfn_t mfn, max_mfn;
2071 	page_t *pp;
2072 
2073 	/*
2074 	 * Figure out the bounds of the pool on first invocation.
2075 	 * We use a percentage of memory for the io pool size.
2076 	 * we allow that to shrink, but not to less than a fixed minimum
2077 	 */
2078 	if (io_pool_cnt_max == 0) {
2079 		io_pool_cnt_max = physmem / (100 / io_pool_physmem_pct);
2080 		io_pool_cnt_lowater = io_pool_cnt_max;
2081 		/*
2082 		 * This is the first time in populate_io_pool, grab a va to use
2083 		 * when we need to allocate pages.
2084 		 */
2085 		io_pool_kva = vmem_alloc(heap_arena, PAGESIZE, VM_SLEEP);
2086 	}
2087 	/*
2088 	 * If we are out of pages in the pool, then grow the size of the pool
2089 	 */
2090 	if (io_pool_cnt == 0)
2091 		io_pool_cnt_max += io_pool_cnt_max / 20; /* grow by 5% */
2092 	io_pool_grows++;	/* should be a kstat? */
2093 
2094 	/*
2095 	 * Get highest mfn on this platform, but limit to the 32 bit DMA max.
2096 	 */
2097 	(void) mfn_to_pfn(start_mfn);
2098 	max_mfn = MIN(cached_max_mfn, PFN_4GIG);
2099 	for (mfn = start_mfn; mfn < max_mfn; start_mfn = ++mfn) {
2100 		pfn = mfn_to_pfn(mfn);
2101 		if (pfn & PFN_IS_FOREIGN_MFN)
2102 			continue;
2103 		/*
2104 		 * try to allocate it from free pages
2105 		 */
2106 		pp = page_numtopp_alloc(pfn);
2107 		if (pp == NULL)
2108 			continue;
2109 		PP_CLRFREE(pp);
2110 		add_page_to_pool(pp, 1);
2111 		if (io_pool_cnt >= io_pool_cnt_max)
2112 			break;
2113 	}
2114 
2115 	return (io_pool_cnt);
2116 }
2117 
2118 /*
2119  * Destroy a page that was being used for DMA I/O. It may or
2120  * may not actually go back to the io_pool.
2121  */
2122 void
2123 page_destroy_io(page_t *pp)
2124 {
2125 	mfn_t mfn = mfn_list[pp->p_pagenum];
2126 
2127 	/*
2128 	 * When the page was alloc'd a reservation was made, release it now
2129 	 */
2130 	page_unresv(1);
2131 	/*
2132 	 * Unload translations, if any, then hash out the
2133 	 * page to erase its identity.
2134 	 */
2135 	(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
2136 	page_hashout(pp, NULL);
2137 
2138 	/*
2139 	 * If the page came from the free lists, just put it back to them.
2140 	 * DomU pages always go on the free lists as well.
2141 	 */
2142 	if (!DOMAIN_IS_INITDOMAIN(xen_info) || mfn >= PFN_4GIG) {
2143 		page_free(pp, 1);
2144 		return;
2145 	}
2146 
2147 	add_page_to_pool(pp, 0);
2148 }
2149 
2150 
2151 long contig_searches;		/* count of times contig pages requested */
2152 long contig_search_restarts;	/* count of contig ranges tried */
2153 long contig_search_failed;	/* count of contig alloc failures */
2154 
2155 /*
2156  * Look thru the contiguous pfns that are not part of the io_pool for
2157  * contiguous free pages.  Return a list of the found pages or NULL.
2158  */
2159 page_t *
2160 find_contig_free(uint_t bytes, uint_t flags)
2161 {
2162 	page_t *pp, *plist = NULL;
2163 	mfn_t mfn, prev_mfn;
2164 	pfn_t pfn;
2165 	int pages_needed, pages_requested;
2166 	int search_start;
2167 
2168 	/*
2169 	 * create the contig pfn list if not already done
2170 	 */
2171 	if (contig_pfn_list == NULL) {
2172 		if (contig_pfnlist_locked) {
2173 			return (NULL);
2174 		} else {
2175 			if (!create_contig_pfnlist(flags))
2176 				return (NULL);
2177 		}
2178 	}
2179 	contig_searches++;
2180 	/*
2181 	 * Search contiguous pfn list for physically contiguous pages not in
2182 	 * the io_pool.  Start the search where the last search left off.
2183 	 */
2184 	pages_requested = pages_needed = mmu_btop(bytes);
2185 	search_start = next_alloc_pfn;
2186 	prev_mfn = 0;
2187 	while (pages_needed) {
2188 		pfn = contig_pfn_list[next_alloc_pfn];
2189 		mfn = pfn_to_mfn(pfn);
2190 		if ((prev_mfn == 0 || mfn == prev_mfn + 1) &&
2191 		    (pp = page_numtopp_alloc(pfn)) != NULL) {
2192 			PP_CLRFREE(pp);
2193 			page_io_pool_add(&plist, pp);
2194 			pages_needed--;
2195 			prev_mfn = mfn;
2196 		} else {
2197 			contig_search_restarts++;
2198 			/*
2199 			 * free partial page list
2200 			 */
2201 			while (plist != NULL) {
2202 				pp = plist;
2203 				page_io_pool_sub(&plist, pp, pp);
2204 				page_free(pp, 1);
2205 			}
2206 			pages_needed = pages_requested;
2207 			prev_mfn = 0;
2208 		}
2209 		if (++next_alloc_pfn == contig_pfn_cnt)
2210 			next_alloc_pfn = 0;
2211 		if (next_alloc_pfn == search_start)
2212 			break; /* all pfns searched */
2213 	}
2214 	if (pages_needed) {
2215 		contig_search_failed++;
2216 		/*
2217 		 * Failed to find enough contig pages.
2218 		 * free partial page list
2219 		 */
2220 		while (plist != NULL) {
2221 			pp = plist;
2222 			page_io_pool_sub(&plist, pp, pp);
2223 			page_free(pp, 1);
2224 		}
2225 	}
2226 	return (plist);
2227 }
2228 
2229 /*
2230  * Allocator for domain 0 I/O pages. We match the required
2231  * DMA attributes and contiguity constraints.
2232  */
2233 /*ARGSUSED*/
2234 page_t *
2235 page_create_io(
2236 	struct vnode	*vp,
2237 	u_offset_t	off,
2238 	uint_t		bytes,
2239 	uint_t		flags,
2240 	struct as	*as,
2241 	caddr_t		vaddr,
2242 	ddi_dma_attr_t	*mattr)
2243 {
2244 	mfn_t	max_mfn = HYPERVISOR_memory_op(XENMEM_maximum_ram_page, NULL);
2245 	page_t	*pp_first;	/* list to return */
2246 	page_t	*pp_last;	/* last in list to return */
2247 	page_t	*pp, **poolp, **pplist = NULL, *expp;
2248 	int	i, extpages = 0, npages = 0, contig, anyaddr, extra;
2249 	mfn_t	lo_mfn;
2250 	mfn_t	hi_mfn;
2251 	mfn_t	mfn, tmfn;
2252 	mfn_t	*mfnlist = 0;
2253 	pgcnt_t	pfnalign = 0;
2254 	int	align, order, nbits, extents;
2255 	uint64_t pfnseg;
2256 	int	attempt = 0, is_domu = 0;
2257 	int	asked_hypervisor = 0;
2258 	uint_t	kflags;
2259 
2260 	ASSERT(mattr != NULL);
2261 	lo_mfn = mmu_btop(mattr->dma_attr_addr_lo);
2262 	hi_mfn = mmu_btop(mattr->dma_attr_addr_hi);
2263 	align = maxbit(mattr->dma_attr_align, mattr->dma_attr_minxfer);
2264 	if (align > MMU_PAGESIZE)
2265 		pfnalign = mmu_btop(align);
2266 	pfnseg = mmu_btop(mattr->dma_attr_seg);
2267 
2268 	/*
2269 	 * Clear the contig flag if only one page is needed.
2270 	 */
2271 	contig = (flags & PG_PHYSCONTIG);
2272 	flags &= ~PG_PHYSCONTIG;
2273 	bytes = P2ROUNDUP(bytes, MMU_PAGESIZE);
2274 	if (bytes == MMU_PAGESIZE)
2275 		contig = 0;
2276 
2277 	/*
2278 	 * Check if any old page in the system is fine.
2279 	 * DomU should always go down this path.
2280 	 */
2281 	is_domu = !DOMAIN_IS_INITDOMAIN(xen_info);
2282 	anyaddr = lo_mfn == 0 && hi_mfn >= max_mfn && !pfnalign;
2283 	if ((!contig && anyaddr) || is_domu) {
2284 		pp = page_create_va(vp, off, bytes, flags, &kvseg, vaddr);
2285 		if (pp)
2286 			return (pp);
2287 		else if (is_domu)
2288 			return (NULL); /* no memory available */
2289 	}
2290 	/*
2291 	 * DomU should never reach here
2292 	 */
2293 try_again:
2294 	/*
2295 	 * We could just want unconstrained but contig pages.
2296 	 */
2297 	if (anyaddr && contig && pfnseg >= max_mfn) {
2298 		/*
2299 		 * Look for free contig pages to satisfy the request.
2300 		 */
2301 		mutex_enter(&io_pool_lock);
2302 		pp_first = find_contig_free(bytes, flags);
2303 		mutex_exit(&io_pool_lock);
2304 		if (pp_first != NULL)
2305 			goto done;
2306 	}
2307 	/*
2308 	 * See if we want pages for a legacy device
2309 	 */
2310 	if (hi_mfn < PFN_16MEG)
2311 		poolp = &io_pool_16m;
2312 	else
2313 		poolp = &io_pool_4g;
2314 try_smaller:
2315 	/*
2316 	 * Take pages from I/O pool. We'll use pages from the highest MFN
2317 	 * range possible.
2318 	 */
2319 	pp_first = pp_last = NULL;
2320 	npages = mmu_btop(bytes);
2321 	mutex_enter(&io_pool_lock);
2322 	for (pp = *poolp; pp && npages > 0; ) {
2323 		pp = pp->p_prev;
2324 
2325 		/*
2326 		 * skip pages above allowable range
2327 		 */
2328 		mfn = mfn_list[pp->p_pagenum];
2329 		if (hi_mfn < mfn)
2330 			goto skip;
2331 
2332 		/*
2333 		 * stop at pages below allowable range
2334 		 */
2335 		if (lo_mfn > mfn)
2336 			break;
2337 restart:
2338 		if (pp_last == NULL) {
2339 			/*
2340 			 * Check alignment
2341 			 */
2342 			tmfn = mfn - (npages - 1);
2343 			if (pfnalign) {
2344 				if (tmfn != P2ROUNDUP(tmfn, pfnalign))
2345 					goto skip; /* not properly aligned */
2346 			}
2347 			/*
2348 			 * Check segment
2349 			 */
2350 			if ((mfn & pfnseg) < (tmfn & pfnseg))
2351 				goto skip; /* crosses segment boundary */
2352 			/*
2353 			 * Start building page list
2354 			 */
2355 			pp_first = pp_last = pp;
2356 			npages--;
2357 		} else {
2358 			/*
2359 			 * check physical contiguity if required
2360 			 */
2361 			if (contig &&
2362 			    mfn_list[pp_first->p_pagenum] != mfn + 1) {
2363 				/*
2364 				 * not a contiguous page, restart list.
2365 				 */
2366 				pp_last = NULL;
2367 				npages = mmu_btop(bytes);
2368 				goto restart;
2369 			} else { /* add page to list */
2370 				pp_first = pp;
2371 				--npages;
2372 			}
2373 		}
2374 skip:
2375 		if (pp == *poolp)
2376 			break;
2377 	}
2378 
2379 	/*
2380 	 * If we didn't find memory. Try the more constrained pool, then
2381 	 * sweep free pages into the DMA pool and try again. If we fail
2382 	 * repeatedly, ask the Hypervisor for help.
2383 	 */
2384 	if (npages != 0) {
2385 		mutex_exit(&io_pool_lock);
2386 		/*
2387 		 * If we were looking in the less constrained pool and didn't
2388 		 * find pages, try the more constrained pool.
2389 		 */
2390 		if (poolp == &io_pool_4g) {
2391 			poolp = &io_pool_16m;
2392 			goto try_smaller;
2393 		}
2394 		kmem_reap();
2395 		if (++attempt < 4) {
2396 			/*
2397 			 * Grab some more io_pool pages
2398 			 */
2399 			(void) populate_io_pool();
2400 			goto try_again;
2401 		}
2402 
2403 		if (asked_hypervisor++)
2404 			return (NULL);	/* really out of luck */
2405 		/*
2406 		 * Hypervisor exchange doesn't handle segment or alignment
2407 		 * constraints
2408 		 */
2409 		if (mattr->dma_attr_seg < mattr->dma_attr_addr_hi || pfnalign)
2410 			return (NULL);
2411 		/*
2412 		 * Try exchanging pages with the hypervisor.
2413 		 */
2414 		npages = mmu_btop(bytes);
2415 		kflags = flags & PG_WAIT ? KM_SLEEP : KM_NOSLEEP;
2416 		/*
2417 		 * Hypervisor will allocate extents, if we want contig pages
2418 		 * extent must be >= npages
2419 		 */
2420 		if (contig) {
2421 			order = highbit(npages) - 1;
2422 			if (npages & ((1 << order) - 1))
2423 				order++;
2424 			extpages = 1 << order;
2425 		} else {
2426 			order = 0;
2427 			extpages = npages;
2428 		}
2429 		if (extpages > npages) {
2430 			extra = extpages - npages;
2431 			if (!page_resv(extra, kflags))
2432 				return (NULL);
2433 		}
2434 		pplist = kmem_alloc(extpages * sizeof (page_t *), kflags);
2435 		if (pplist == NULL)
2436 			goto fail;
2437 		mfnlist = kmem_alloc(extpages * sizeof (mfn_t), kflags);
2438 		if (mfnlist == NULL)
2439 			goto fail;
2440 		pp = page_create_va(vp, off, npages * PAGESIZE, flags,
2441 		    &kvseg, vaddr);
2442 		if (pp == NULL)
2443 			goto fail;
2444 		pp_first = pp;
2445 		if (extpages > npages) {
2446 			/*
2447 			 * fill out the rest of extent pages to swap with the
2448 			 * hypervisor
2449 			 */
2450 			for (i = 0; i < extra; i++) {
2451 				expp = page_create_va(vp,
2452 				    (u_offset_t)(uintptr_t)io_pool_kva,
2453 				    PAGESIZE, flags, &kvseg, io_pool_kva);
2454 				if (expp == NULL)
2455 					goto balloon_fail;
2456 				(void) hat_pageunload(expp, HAT_FORCE_PGUNLOAD);
2457 				page_io_unlock(expp);
2458 				page_hashout(expp, NULL);
2459 				page_io_lock(expp);
2460 				/*
2461 				 * add page to end of list
2462 				 */
2463 				expp->p_prev = pp_first->p_prev;
2464 				expp->p_next = pp_first;
2465 				expp->p_prev->p_next = expp;
2466 				pp_first->p_prev = expp;
2467 			}
2468 
2469 		}
2470 		for (i = 0; i < extpages; i++) {
2471 			pplist[i] = pp;
2472 			pp = pp->p_next;
2473 		}
2474 		nbits = highbit(mattr->dma_attr_addr_hi);
2475 		extents = contig ? 1 : npages;
2476 		if (balloon_replace_pages(extents, pplist, nbits, order,
2477 		    mfnlist) != extents)
2478 			goto balloon_fail;
2479 
2480 		kmem_free(pplist, extpages * sizeof (page_t *));
2481 		kmem_free(mfnlist, extpages * sizeof (mfn_t));
2482 		/*
2483 		 * Return any excess pages to free list
2484 		 */
2485 		if (extpages > npages) {
2486 			for (i = 0; i < extra; i++) {
2487 				pp = pp_first->p_prev;
2488 				page_sub(&pp_first, pp);
2489 				page_io_unlock(pp);
2490 				page_unresv(1);
2491 				page_free(pp, 1);
2492 			}
2493 		}
2494 		check_dma(mattr, pp_first, mmu_btop(bytes));
2495 		return (pp_first);
2496 	}
2497 
2498 	/*
2499 	 * Found the pages, now snip them from the list
2500 	 */
2501 	page_io_pool_sub(poolp, pp_first, pp_last);
2502 	io_pool_cnt -= mmu_btop(bytes);
2503 	if (io_pool_cnt < io_pool_cnt_lowater)
2504 		io_pool_cnt_lowater = io_pool_cnt; /* io pool low water mark */
2505 	mutex_exit(&io_pool_lock);
2506 done:
2507 	check_dma(mattr, pp_first, mmu_btop(bytes));
2508 	pp = pp_first;
2509 	do {
2510 		if (!page_hashin(pp, vp, off, NULL)) {
2511 			panic("pg_create_io: hashin failed pp %p, vp %p,"
2512 			    " off %llx",
2513 			    (void *)pp, (void *)vp, off);
2514 		}
2515 		off += MMU_PAGESIZE;
2516 		PP_CLRFREE(pp);
2517 		PP_CLRAGED(pp);
2518 		page_set_props(pp, P_REF);
2519 		page_io_lock(pp);
2520 		pp = pp->p_next;
2521 	} while (pp != pp_first);
2522 	return (pp_first);
2523 balloon_fail:
2524 	/*
2525 	 * Return pages to free list and return failure
2526 	 */
2527 	while (pp_first != NULL) {
2528 		pp = pp_first;
2529 		page_sub(&pp_first, pp);
2530 		page_io_unlock(pp);
2531 		if (pp->p_vnode != NULL)
2532 			page_hashout(pp, NULL);
2533 		page_free(pp, 1);
2534 	}
2535 fail:
2536 	if (pplist)
2537 		kmem_free(pplist, extpages * sizeof (page_t *));
2538 	if (mfnlist)
2539 		kmem_free(mfnlist, extpages * sizeof (mfn_t));
2540 	page_unresv(extpages - npages);
2541 	return (NULL);
2542 }
2543 
2544 /*
2545  * Lock and return the page with the highest mfn that we can find.  last_mfn
2546  * holds the last one found, so the next search can start from there.  We
2547  * also keep a counter so that we don't loop forever if the machine has no
2548  * free pages.
2549  *
2550  * This is called from the balloon thread to find pages to give away.  new_high
2551  * is used when new mfn's have been added to the system - we will reset our
2552  * search if the new mfn's are higher than our current search position.
2553  */
2554 page_t *
2555 page_get_high_mfn(mfn_t new_high)
2556 {
2557 	static mfn_t last_mfn = 0;
2558 	pfn_t pfn;
2559 	page_t *pp;
2560 	ulong_t loop_count = 0;
2561 
2562 	if (new_high > last_mfn)
2563 		last_mfn = new_high;
2564 
2565 	for (; loop_count < mfn_count; loop_count++, last_mfn--) {
2566 		if (last_mfn == 0) {
2567 			last_mfn = cached_max_mfn;
2568 		}
2569 
2570 		pfn = mfn_to_pfn(last_mfn);
2571 		if (pfn & PFN_IS_FOREIGN_MFN)
2572 			continue;
2573 
2574 		/* See if the page is free.  If so, lock it. */
2575 		pp = page_numtopp_alloc(pfn);
2576 		if (pp == NULL)
2577 			continue;
2578 		PP_CLRFREE(pp);
2579 
2580 		ASSERT(PAGE_EXCL(pp));
2581 		ASSERT(pp->p_vnode == NULL);
2582 		ASSERT(!hat_page_is_mapped(pp));
2583 		last_mfn--;
2584 		return (pp);
2585 	}
2586 	return (NULL);
2587 }
2588 
2589 #else /* !__xpv */
2590 
2591 /*
2592  * get a page from any list with the given mnode
2593  */
2594 static page_t *
2595 page_get_mnode_anylist(ulong_t origbin, uchar_t szc, uint_t flags,
2596     int mnode, int mtype, ddi_dma_attr_t *dma_attr)
2597 {
2598 	kmutex_t		*pcm;
2599 	int			i;
2600 	page_t			*pp;
2601 	page_t			*first_pp;
2602 	uint64_t		pgaddr;
2603 	ulong_t			bin;
2604 	int			mtypestart;
2605 	int			plw_initialized;
2606 	page_list_walker_t	plw;
2607 
2608 	VM_STAT_ADD(pga_vmstats.pgma_alloc);
2609 
2610 	ASSERT((flags & PG_MATCH_COLOR) == 0);
2611 	ASSERT(szc == 0);
2612 	ASSERT(dma_attr != NULL);
2613 
2614 	MTYPE_START(mnode, mtype, flags);
2615 	if (mtype < 0) {
2616 		VM_STAT_ADD(pga_vmstats.pgma_allocempty);
2617 		return (NULL);
2618 	}
2619 
2620 	mtypestart = mtype;
2621 
2622 	bin = origbin;
2623 
2624 	/*
2625 	 * check up to page_colors + 1 bins - origbin may be checked twice
2626 	 * because of BIN_STEP skip
2627 	 */
2628 	do {
2629 		plw_initialized = 0;
2630 
2631 		for (plw.plw_count = 0;
2632 		    plw.plw_count < page_colors; plw.plw_count++) {
2633 
2634 			if (PAGE_FREELISTS(mnode, szc, bin, mtype) == NULL)
2635 				goto nextfreebin;
2636 
2637 			pcm = PC_BIN_MUTEX(mnode, bin, PG_FREE_LIST);
2638 			mutex_enter(pcm);
2639 			pp = PAGE_FREELISTS(mnode, szc, bin, mtype);
2640 			first_pp = pp;
2641 			while (pp != NULL) {
2642 				if (page_trylock(pp, SE_EXCL) == 0) {
2643 					pp = pp->p_next;
2644 					if (pp == first_pp) {
2645 						pp = NULL;
2646 					}
2647 					continue;
2648 				}
2649 
2650 				ASSERT(PP_ISFREE(pp));
2651 				ASSERT(PP_ISAGED(pp));
2652 				ASSERT(pp->p_vnode == NULL);
2653 				ASSERT(pp->p_hash == NULL);
2654 				ASSERT(pp->p_offset == (u_offset_t)-1);
2655 				ASSERT(pp->p_szc == szc);
2656 				ASSERT(PFN_2_MEM_NODE(pp->p_pagenum) == mnode);
2657 				/* check if page within DMA attributes */
2658 				pgaddr = pa_to_ma(pfn_to_pa(pp->p_pagenum));
2659 				if ((pgaddr >= dma_attr->dma_attr_addr_lo) &&
2660 				    (pgaddr + MMU_PAGESIZE - 1 <=
2661 				    dma_attr->dma_attr_addr_hi)) {
2662 					break;
2663 				}
2664 
2665 				/* continue looking */
2666 				page_unlock(pp);
2667 				pp = pp->p_next;
2668 				if (pp == first_pp)
2669 					pp = NULL;
2670 
2671 			}
2672 			if (pp != NULL) {
2673 				ASSERT(mtype == PP_2_MTYPE(pp));
2674 				ASSERT(pp->p_szc == 0);
2675 
2676 				/* found a page with specified DMA attributes */
2677 				page_sub(&PAGE_FREELISTS(mnode, szc, bin,
2678 				    mtype), pp);
2679 				page_ctr_sub(mnode, mtype, pp, PG_FREE_LIST);
2680 
2681 				if ((PP_ISFREE(pp) == 0) ||
2682 				    (PP_ISAGED(pp) == 0)) {
2683 					cmn_err(CE_PANIC, "page %p is not free",
2684 					    (void *)pp);
2685 				}
2686 
2687 				mutex_exit(pcm);
2688 				check_dma(dma_attr, pp, 1);
2689 				VM_STAT_ADD(pga_vmstats.pgma_allocok);
2690 				return (pp);
2691 			}
2692 			mutex_exit(pcm);
2693 nextfreebin:
2694 			if (plw_initialized == 0) {
2695 				page_list_walk_init(szc, 0, bin, 1, 0, &plw);
2696 				ASSERT(plw.plw_ceq_dif == page_colors);
2697 				plw_initialized = 1;
2698 			}
2699 
2700 			if (plw.plw_do_split) {
2701 				pp = page_freelist_split(szc, bin, mnode,
2702 				    mtype,
2703 				    mmu_btop(dma_attr->dma_attr_addr_hi + 1),
2704 				    &plw);
2705 				if (pp != NULL)
2706 					return (pp);
2707 			}
2708 
2709 			bin = page_list_walk_next_bin(szc, bin, &plw);
2710 		}
2711 
2712 		MTYPE_NEXT(mnode, mtype, flags);
2713 	} while (mtype >= 0);
2714 
2715 	/* failed to find a page in the freelist; try it in the cachelist */
2716 
2717 	/* reset mtype start for cachelist search */
2718 	mtype = mtypestart;
2719 	ASSERT(mtype >= 0);
2720 
2721 	/* start with the bin of matching color */
2722 	bin = origbin;
2723 
2724 	do {
2725 		for (i = 0; i <= page_colors; i++) {
2726 			if (PAGE_CACHELISTS(mnode, bin, mtype) == NULL)
2727 				goto nextcachebin;
2728 			pcm = PC_BIN_MUTEX(mnode, bin, PG_CACHE_LIST);
2729 			mutex_enter(pcm);
2730 			pp = PAGE_CACHELISTS(mnode, bin, mtype);
2731 			first_pp = pp;
2732 			while (pp != NULL) {
2733 				if (page_trylock(pp, SE_EXCL) == 0) {
2734 					pp = pp->p_next;
2735 					if (pp == first_pp)
2736 						break;
2737 					continue;
2738 				}
2739 				ASSERT(pp->p_vnode);
2740 				ASSERT(PP_ISAGED(pp) == 0);
2741 				ASSERT(pp->p_szc == 0);
2742 				ASSERT(PFN_2_MEM_NODE(pp->p_pagenum) == mnode);
2743 
2744 				/* check if page within DMA attributes */
2745 
2746 				pgaddr = pa_to_ma(pfn_to_pa(pp->p_pagenum));
2747 				if ((pgaddr >= dma_attr->dma_attr_addr_lo) &&
2748 				    (pgaddr + MMU_PAGESIZE - 1 <=
2749 				    dma_attr->dma_attr_addr_hi)) {
2750 					break;
2751 				}
2752 
2753 				/* continue looking */
2754 				page_unlock(pp);
2755 				pp = pp->p_next;
2756 				if (pp == first_pp)
2757 					pp = NULL;
2758 			}
2759 
2760 			if (pp != NULL) {
2761 				ASSERT(mtype == PP_2_MTYPE(pp));
2762 				ASSERT(pp->p_szc == 0);
2763 
2764 				/* found a page with specified DMA attributes */
2765 				page_sub(&PAGE_CACHELISTS(mnode, bin,
2766 				    mtype), pp);
2767 				page_ctr_sub(mnode, mtype, pp, PG_CACHE_LIST);
2768 
2769 				mutex_exit(pcm);
2770 				ASSERT(pp->p_vnode);
2771 				ASSERT(PP_ISAGED(pp) == 0);
2772 				check_dma(dma_attr, pp, 1);
2773 				VM_STAT_ADD(pga_vmstats.pgma_allocok);
2774 				return (pp);
2775 			}
2776 			mutex_exit(pcm);
2777 nextcachebin:
2778 			bin += (i == 0) ? BIN_STEP : 1;
2779 			bin &= page_colors_mask;
2780 		}
2781 		MTYPE_NEXT(mnode, mtype, flags);
2782 	} while (mtype >= 0);
2783 
2784 	VM_STAT_ADD(pga_vmstats.pgma_allocfailed);
2785 	return (NULL);
2786 }
2787 
2788 /*
2789  * This function is similar to page_get_freelist()/page_get_cachelist()
2790  * but it searches both the lists to find a page with the specified
2791  * color (or no color) and DMA attributes. The search is done in the
2792  * freelist first and then in the cache list within the highest memory
2793  * range (based on DMA attributes) before searching in the lower
2794  * memory ranges.
2795  *
2796  * Note: This function is called only by page_create_io().
2797  */
2798 /*ARGSUSED*/
2799 static page_t *
2800 page_get_anylist(struct vnode *vp, u_offset_t off, struct as *as, caddr_t vaddr,
2801     size_t size, uint_t flags, ddi_dma_attr_t *dma_attr, lgrp_t	*lgrp)
2802 {
2803 	uint_t		bin;
2804 	int		mtype;
2805 	page_t		*pp;
2806 	int		n;
2807 	int		m;
2808 	int		szc;
2809 	int		fullrange;
2810 	int		mnode;
2811 	int		local_failed_stat = 0;
2812 	lgrp_mnode_cookie_t	lgrp_cookie;
2813 
2814 	VM_STAT_ADD(pga_vmstats.pga_alloc);
2815 
2816 	/* only base pagesize currently supported */
2817 	if (size != MMU_PAGESIZE)
2818 		return (NULL);
2819 
2820 	/*
2821 	 * If we're passed a specific lgroup, we use it.  Otherwise,
2822 	 * assume first-touch placement is desired.
2823 	 */
2824 	if (!LGRP_EXISTS(lgrp))
2825 		lgrp = lgrp_home_lgrp();
2826 
2827 	/* LINTED */
2828 	AS_2_BIN(as, seg, vp, vaddr, bin, 0);
2829 
2830 	/*
2831 	 * Only hold one freelist or cachelist lock at a time, that way we
2832 	 * can start anywhere and not have to worry about lock
2833 	 * ordering.
2834 	 */
2835 	if (dma_attr == NULL) {
2836 		n = 0;
2837 		m = mnoderangecnt - 1;
2838 		fullrange = 1;
2839 		VM_STAT_ADD(pga_vmstats.pga_nulldmaattr);
2840 	} else {
2841 		pfn_t pfnlo = mmu_btop(dma_attr->dma_attr_addr_lo);
2842 		pfn_t pfnhi = mmu_btop(dma_attr->dma_attr_addr_hi);
2843 
2844 		/*
2845 		 * We can guarantee alignment only for page boundary.
2846 		 */
2847 		if (dma_attr->dma_attr_align > MMU_PAGESIZE)
2848 			return (NULL);
2849 
2850 		n = pfn_2_mtype(pfnlo);
2851 		m = pfn_2_mtype(pfnhi);
2852 
2853 		fullrange = ((pfnlo == mnoderanges[n].mnr_pfnlo) &&
2854 		    (pfnhi >= mnoderanges[m].mnr_pfnhi));
2855 	}
2856 	VM_STAT_COND_ADD(fullrange == 0, pga_vmstats.pga_notfullrange);
2857 
2858 	if (n > m)
2859 		return (NULL);
2860 
2861 	szc = 0;
2862 
2863 	/* cylcing thru mtype handled by RANGE0 if n == 0 */
2864 	if (n == 0) {
2865 		flags |= PGI_MT_RANGE0;
2866 		n = m;
2867 	}
2868 
2869 	/*
2870 	 * Try local memory node first, but try remote if we can't
2871 	 * get a page of the right color.
2872 	 */
2873 	LGRP_MNODE_COOKIE_INIT(lgrp_cookie, lgrp, LGRP_SRCH_HIER);
2874 	while ((mnode = lgrp_memnode_choose(&lgrp_cookie)) >= 0) {
2875 		/*
2876 		 * allocate pages from high pfn to low.
2877 		 */
2878 		for (mtype = m; mtype >= n; mtype--) {
2879 			if (fullrange != 0) {
2880 				pp = page_get_mnode_freelist(mnode,
2881 				    bin, mtype, szc, flags);
2882 				if (pp == NULL) {
2883 					pp = page_get_mnode_cachelist(
2884 					    bin, flags, mnode, mtype);
2885 				}
2886 			} else {
2887 				pp = page_get_mnode_anylist(bin, szc,
2888 				    flags, mnode, mtype, dma_attr);
2889 			}
2890 			if (pp != NULL) {
2891 				VM_STAT_ADD(pga_vmstats.pga_allocok);
2892 				check_dma(dma_attr, pp, 1);
2893 				return (pp);
2894 			}
2895 		}
2896 		if (!local_failed_stat) {
2897 			lgrp_stat_add(lgrp->lgrp_id, LGRP_NUM_ALLOC_FAIL, 1);
2898 			local_failed_stat = 1;
2899 		}
2900 	}
2901 	VM_STAT_ADD(pga_vmstats.pga_allocfailed);
2902 
2903 	return (NULL);
2904 }
2905 
2906 /*
2907  * page_create_io()
2908  *
2909  * This function is a copy of page_create_va() with an additional
2910  * argument 'mattr' that specifies DMA memory requirements to
2911  * the page list functions. This function is used by the segkmem
2912  * allocator so it is only to create new pages (i.e PG_EXCL is
2913  * set).
2914  *
2915  * Note: This interface is currently used by x86 PSM only and is
2916  *	 not fully specified so the commitment level is only for
2917  *	 private interface specific to x86. This interface uses PSM
2918  *	 specific page_get_anylist() interface.
2919  */
2920 
2921 #define	PAGE_HASH_SEARCH(index, pp, vp, off) { \
2922 	for ((pp) = page_hash[(index)]; (pp); (pp) = (pp)->p_hash) { \
2923 		if ((pp)->p_vnode == (vp) && (pp)->p_offset == (off)) \
2924 			break; \
2925 	} \
2926 }
2927 
2928 
2929 page_t *
2930 page_create_io(
2931 	struct vnode	*vp,
2932 	u_offset_t	off,
2933 	uint_t		bytes,
2934 	uint_t		flags,
2935 	struct as	*as,
2936 	caddr_t		vaddr,
2937 	ddi_dma_attr_t	*mattr)	/* DMA memory attributes if any */
2938 {
2939 	page_t		*plist = NULL;
2940 	uint_t		plist_len = 0;
2941 	pgcnt_t		npages;
2942 	page_t		*npp = NULL;
2943 	uint_t		pages_req;
2944 	page_t		*pp;
2945 	kmutex_t	*phm = NULL;
2946 	uint_t		index;
2947 
2948 	TRACE_4(TR_FAC_VM, TR_PAGE_CREATE_START,
2949 	    "page_create_start:vp %p off %llx bytes %u flags %x",
2950 	    vp, off, bytes, flags);
2951 
2952 	ASSERT((flags & ~(PG_EXCL | PG_WAIT | PG_PHYSCONTIG)) == 0);
2953 
2954 	pages_req = npages = mmu_btopr(bytes);
2955 
2956 	/*
2957 	 * Do the freemem and pcf accounting.
2958 	 */
2959 	if (!page_create_wait(npages, flags)) {
2960 		return (NULL);
2961 	}
2962 
2963 	TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SUCCESS,
2964 	    "page_create_success:vp %p off %llx", vp, off);
2965 
2966 	/*
2967 	 * If satisfying this request has left us with too little
2968 	 * memory, start the wheels turning to get some back.  The
2969 	 * first clause of the test prevents waking up the pageout
2970 	 * daemon in situations where it would decide that there's
2971 	 * nothing to do.
2972 	 */
2973 	if (nscan < desscan && freemem < minfree) {
2974 		TRACE_1(TR_FAC_VM, TR_PAGEOUT_CV_SIGNAL,
2975 		    "pageout_cv_signal:freemem %ld", freemem);
2976 		cv_signal(&proc_pageout->p_cv);
2977 	}
2978 
2979 	if (flags & PG_PHYSCONTIG) {
2980 
2981 		plist = page_get_contigpage(&npages, mattr, 1);
2982 		if (plist == NULL) {
2983 			page_create_putback(npages);
2984 			return (NULL);
2985 		}
2986 
2987 		pp = plist;
2988 
2989 		do {
2990 			if (!page_hashin(pp, vp, off, NULL)) {
2991 				panic("pg_creat_io: hashin failed %p %p %llx",
2992 				    (void *)pp, (void *)vp, off);
2993 			}
2994 			VM_STAT_ADD(page_create_new);
2995 			off += MMU_PAGESIZE;
2996 			PP_CLRFREE(pp);
2997 			PP_CLRAGED(pp);
2998 			page_set_props(pp, P_REF);
2999 			pp = pp->p_next;
3000 		} while (pp != plist);
3001 
3002 		if (!npages) {
3003 			check_dma(mattr, plist, pages_req);
3004 			return (plist);
3005 		} else {
3006 			vaddr += (pages_req - npages) << MMU_PAGESHIFT;
3007 		}
3008 
3009 		/*
3010 		 * fall-thru:
3011 		 *
3012 		 * page_get_contigpage returns when npages <= sgllen.
3013 		 * Grab the rest of the non-contig pages below from anylist.
3014 		 */
3015 	}
3016 
3017 	/*
3018 	 * Loop around collecting the requested number of pages.
3019 	 * Most of the time, we have to `create' a new page. With
3020 	 * this in mind, pull the page off the free list before
3021 	 * getting the hash lock.  This will minimize the hash
3022 	 * lock hold time, nesting, and the like.  If it turns
3023 	 * out we don't need the page, we put it back at the end.
3024 	 */
3025 	while (npages--) {
3026 		phm = NULL;
3027 
3028 		index = PAGE_HASH_FUNC(vp, off);
3029 top:
3030 		ASSERT(phm == NULL);
3031 		ASSERT(index == PAGE_HASH_FUNC(vp, off));
3032 		ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
3033 
3034 		if (npp == NULL) {
3035 			/*
3036 			 * Try to get the page of any color either from
3037 			 * the freelist or from the cache list.
3038 			 */
3039 			npp = page_get_anylist(vp, off, as, vaddr, MMU_PAGESIZE,
3040 			    flags & ~PG_MATCH_COLOR, mattr, NULL);
3041 			if (npp == NULL) {
3042 				if (mattr == NULL) {
3043 					/*
3044 					 * Not looking for a special page;
3045 					 * panic!
3046 					 */
3047 					panic("no page found %d", (int)npages);
3048 				}
3049 				/*
3050 				 * No page found! This can happen
3051 				 * if we are looking for a page
3052 				 * within a specific memory range
3053 				 * for DMA purposes. If PG_WAIT is
3054 				 * specified then we wait for a
3055 				 * while and then try again. The
3056 				 * wait could be forever if we
3057 				 * don't get the page(s) we need.
3058 				 *
3059 				 * Note: XXX We really need a mechanism
3060 				 * to wait for pages in the desired
3061 				 * range. For now, we wait for any
3062 				 * pages and see if we can use it.
3063 				 */
3064 
3065 				if ((mattr != NULL) && (flags & PG_WAIT)) {
3066 					delay(10);
3067 					goto top;
3068 				}
3069 				goto fail; /* undo accounting stuff */
3070 			}
3071 
3072 			if (PP_ISAGED(npp) == 0) {
3073 				/*
3074 				 * Since this page came from the
3075 				 * cachelist, we must destroy the
3076 				 * old vnode association.
3077 				 */
3078 				page_hashout(npp, (kmutex_t *)NULL);
3079 			}
3080 		}
3081 
3082 		/*
3083 		 * We own this page!
3084 		 */
3085 		ASSERT(PAGE_EXCL(npp));
3086 		ASSERT(npp->p_vnode == NULL);
3087 		ASSERT(!hat_page_is_mapped(npp));
3088 		PP_CLRFREE(npp);
3089 		PP_CLRAGED(npp);
3090 
3091 		/*
3092 		 * Here we have a page in our hot little mits and are
3093 		 * just waiting to stuff it on the appropriate lists.
3094 		 * Get the mutex and check to see if it really does
3095 		 * not exist.
3096 		 */
3097 		phm = PAGE_HASH_MUTEX(index);
3098 		mutex_enter(phm);
3099 		PAGE_HASH_SEARCH(index, pp, vp, off);
3100 		if (pp == NULL) {
3101 			VM_STAT_ADD(page_create_new);
3102 			pp = npp;
3103 			npp = NULL;
3104 			if (!page_hashin(pp, vp, off, phm)) {
3105 				/*
3106 				 * Since we hold the page hash mutex and
3107 				 * just searched for this page, page_hashin
3108 				 * had better not fail.  If it does, that
3109 				 * means somethread did not follow the
3110 				 * page hash mutex rules.  Panic now and
3111 				 * get it over with.  As usual, go down
3112 				 * holding all the locks.
3113 				 */
3114 				ASSERT(MUTEX_HELD(phm));
3115 				panic("page_create: hashin fail %p %p %llx %p",
3116 				    (void *)pp, (void *)vp, off, (void *)phm);
3117 
3118 			}
3119 			ASSERT(MUTEX_HELD(phm));
3120 			mutex_exit(phm);
3121 			phm = NULL;
3122 
3123 			/*
3124 			 * Hat layer locking need not be done to set
3125 			 * the following bits since the page is not hashed
3126 			 * and was on the free list (i.e., had no mappings).
3127 			 *
3128 			 * Set the reference bit to protect
3129 			 * against immediate pageout
3130 			 *
3131 			 * XXXmh modify freelist code to set reference
3132 			 * bit so we don't have to do it here.
3133 			 */
3134 			page_set_props(pp, P_REF);
3135 		} else {
3136 			ASSERT(MUTEX_HELD(phm));
3137 			mutex_exit(phm);
3138 			phm = NULL;
3139 			/*
3140 			 * NOTE: This should not happen for pages associated
3141 			 *	 with kernel vnode 'kvp'.
3142 			 */
3143 			/* XX64 - to debug why this happens! */
3144 			ASSERT(!VN_ISKAS(vp));
3145 			if (VN_ISKAS(vp))
3146 				cmn_err(CE_NOTE,
3147 				    "page_create: page not expected "
3148 				    "in hash list for kernel vnode - pp 0x%p",
3149 				    (void *)pp);
3150 			VM_STAT_ADD(page_create_exists);
3151 			goto fail;
3152 		}
3153 
3154 		/*
3155 		 * Got a page!  It is locked.  Acquire the i/o
3156 		 * lock since we are going to use the p_next and
3157 		 * p_prev fields to link the requested pages together.
3158 		 */
3159 		page_io_lock(pp);
3160 		page_add(&plist, pp);
3161 		plist = plist->p_next;
3162 		off += MMU_PAGESIZE;
3163 		vaddr += MMU_PAGESIZE;
3164 	}
3165 
3166 	check_dma(mattr, plist, pages_req);
3167 	return (plist);
3168 
3169 fail:
3170 	if (npp != NULL) {
3171 		/*
3172 		 * Did not need this page after all.
3173 		 * Put it back on the free list.
3174 		 */
3175 		VM_STAT_ADD(page_create_putbacks);
3176 		PP_SETFREE(npp);
3177 		PP_SETAGED(npp);
3178 		npp->p_offset = (u_offset_t)-1;
3179 		page_list_add(npp, PG_FREE_LIST | PG_LIST_TAIL);
3180 		page_unlock(npp);
3181 	}
3182 
3183 	/*
3184 	 * Give up the pages we already got.
3185 	 */
3186 	while (plist != NULL) {
3187 		pp = plist;
3188 		page_sub(&plist, pp);
3189 		page_io_unlock(pp);
3190 		plist_len++;
3191 		/*LINTED: constant in conditional ctx*/
3192 		VN_DISPOSE(pp, B_INVAL, 0, kcred);
3193 	}
3194 
3195 	/*
3196 	 * VN_DISPOSE does freemem accounting for the pages in plist
3197 	 * by calling page_free. So, we need to undo the pcf accounting
3198 	 * for only the remaining pages.
3199 	 */
3200 	VM_STAT_ADD(page_create_putbacks);
3201 	page_create_putback(pages_req - plist_len);
3202 
3203 	return (NULL);
3204 }
3205 #endif /* !__xpv */
3206 
3207 
3208 /*
3209  * Copy the data from the physical page represented by "frompp" to
3210  * that represented by "topp". ppcopy uses CPU->cpu_caddr1 and
3211  * CPU->cpu_caddr2.  It assumes that no one uses either map at interrupt
3212  * level and no one sleeps with an active mapping there.
3213  *
3214  * Note that the ref/mod bits in the page_t's are not affected by
3215  * this operation, hence it is up to the caller to update them appropriately.
3216  */
3217 int
3218 ppcopy(page_t *frompp, page_t *topp)
3219 {
3220 	caddr_t		pp_addr1;
3221 	caddr_t		pp_addr2;
3222 	hat_mempte_t	pte1;
3223 	hat_mempte_t	pte2;
3224 	kmutex_t	*ppaddr_mutex;
3225 	label_t		ljb;
3226 	int		ret = 1;
3227 
3228 	ASSERT_STACK_ALIGNED();
3229 	ASSERT(PAGE_LOCKED(frompp));
3230 	ASSERT(PAGE_LOCKED(topp));
3231 
3232 	if (kpm_enable) {
3233 		pp_addr1 = hat_kpm_page2va(frompp, 0);
3234 		pp_addr2 = hat_kpm_page2va(topp, 0);
3235 		kpreempt_disable();
3236 	} else {
3237 		/*
3238 		 * disable pre-emption so that CPU can't change
3239 		 */
3240 		kpreempt_disable();
3241 
3242 		pp_addr1 = CPU->cpu_caddr1;
3243 		pp_addr2 = CPU->cpu_caddr2;
3244 		pte1 = CPU->cpu_caddr1pte;
3245 		pte2 = CPU->cpu_caddr2pte;
3246 
3247 		ppaddr_mutex = &CPU->cpu_ppaddr_mutex;
3248 		mutex_enter(ppaddr_mutex);
3249 
3250 		hat_mempte_remap(page_pptonum(frompp), pp_addr1, pte1,
3251 		    PROT_READ | HAT_STORECACHING_OK, HAT_LOAD_NOCONSIST);
3252 		hat_mempte_remap(page_pptonum(topp), pp_addr2, pte2,
3253 		    PROT_READ | PROT_WRITE | HAT_STORECACHING_OK,
3254 		    HAT_LOAD_NOCONSIST);
3255 	}
3256 
3257 	if (on_fault(&ljb)) {
3258 		ret = 0;
3259 		goto faulted;
3260 	}
3261 	if (use_sse_pagecopy)
3262 #ifdef __xpv
3263 		page_copy_no_xmm(pp_addr2, pp_addr1);
3264 #else
3265 		hwblkpagecopy(pp_addr1, pp_addr2);
3266 #endif
3267 	else
3268 		bcopy(pp_addr1, pp_addr2, PAGESIZE);
3269 
3270 	no_fault();
3271 faulted:
3272 	if (!kpm_enable) {
3273 #ifdef __xpv
3274 		/*
3275 		 * We can't leave unused mappings laying about under the
3276 		 * hypervisor, so blow them away.
3277 		 */
3278 		if (HYPERVISOR_update_va_mapping((uintptr_t)pp_addr1, 0,
3279 		    UVMF_INVLPG | UVMF_LOCAL) < 0)
3280 			panic("HYPERVISOR_update_va_mapping() failed");
3281 		if (HYPERVISOR_update_va_mapping((uintptr_t)pp_addr2, 0,
3282 		    UVMF_INVLPG | UVMF_LOCAL) < 0)
3283 			panic("HYPERVISOR_update_va_mapping() failed");
3284 #endif
3285 		mutex_exit(ppaddr_mutex);
3286 	}
3287 	kpreempt_enable();
3288 	return (ret);
3289 }
3290 
3291 void
3292 pagezero(page_t *pp, uint_t off, uint_t len)
3293 {
3294 	ASSERT(PAGE_LOCKED(pp));
3295 	pfnzero(page_pptonum(pp), off, len);
3296 }
3297 
3298 /*
3299  * Zero the physical page from off to off + len given by pfn
3300  * without changing the reference and modified bits of page.
3301  *
3302  * We use this using CPU private page address #2, see ppcopy() for more info.
3303  * pfnzero() must not be called at interrupt level.
3304  */
3305 void
3306 pfnzero(pfn_t pfn, uint_t off, uint_t len)
3307 {
3308 	caddr_t		pp_addr2;
3309 	hat_mempte_t	pte2;
3310 	kmutex_t	*ppaddr_mutex = NULL;
3311 
3312 	ASSERT_STACK_ALIGNED();
3313 	ASSERT(len <= MMU_PAGESIZE);
3314 	ASSERT(off <= MMU_PAGESIZE);
3315 	ASSERT(off + len <= MMU_PAGESIZE);
3316 
3317 	if (kpm_enable && !pfn_is_foreign(pfn)) {
3318 		pp_addr2 = hat_kpm_pfn2va(pfn);
3319 		kpreempt_disable();
3320 	} else {
3321 		kpreempt_disable();
3322 
3323 		pp_addr2 = CPU->cpu_caddr2;
3324 		pte2 = CPU->cpu_caddr2pte;
3325 
3326 		ppaddr_mutex = &CPU->cpu_ppaddr_mutex;
3327 		mutex_enter(ppaddr_mutex);
3328 
3329 		hat_mempte_remap(pfn, pp_addr2, pte2,
3330 		    PROT_READ | PROT_WRITE | HAT_STORECACHING_OK,
3331 		    HAT_LOAD_NOCONSIST);
3332 	}
3333 
3334 	if (use_sse_pagezero) {
3335 #ifdef __xpv
3336 		uint_t rem;
3337 
3338 		/*
3339 		 * zero a byte at a time until properly aligned for
3340 		 * block_zero_no_xmm().
3341 		 */
3342 		while (!P2NPHASE(off, ((uint_t)BLOCKZEROALIGN)) && len-- > 0)
3343 			pp_addr2[off++] = 0;
3344 
3345 		/*
3346 		 * Now use faster block_zero_no_xmm() for any range
3347 		 * that is properly aligned and sized.
3348 		 */
3349 		rem = P2PHASE(len, ((uint_t)BLOCKZEROALIGN));
3350 		len -= rem;
3351 		if (len != 0) {
3352 			block_zero_no_xmm(pp_addr2 + off, len);
3353 			off += len;
3354 		}
3355 
3356 		/*
3357 		 * zero remainder with byte stores.
3358 		 */
3359 		while (rem-- > 0)
3360 			pp_addr2[off++] = 0;
3361 #else
3362 		hwblkclr(pp_addr2 + off, len);
3363 #endif
3364 	} else {
3365 		bzero(pp_addr2 + off, len);
3366 	}
3367 
3368 	if (!kpm_enable || pfn_is_foreign(pfn)) {
3369 #ifdef __xpv
3370 		/*
3371 		 * On the hypervisor this page might get used for a page
3372 		 * table before any intervening change to this mapping,
3373 		 * so blow it away.
3374 		 */
3375 		if (HYPERVISOR_update_va_mapping((uintptr_t)pp_addr2, 0,
3376 		    UVMF_INVLPG) < 0)
3377 			panic("HYPERVISOR_update_va_mapping() failed");
3378 #endif
3379 		mutex_exit(ppaddr_mutex);
3380 	}
3381 
3382 	kpreempt_enable();
3383 }
3384 
3385 /*
3386  * Platform-dependent page scrub call.
3387  */
3388 void
3389 pagescrub(page_t *pp, uint_t off, uint_t len)
3390 {
3391 	/*
3392 	 * For now, we rely on the fact that pagezero() will
3393 	 * always clear UEs.
3394 	 */
3395 	pagezero(pp, off, len);
3396 }
3397 
3398 /*
3399  * set up two private addresses for use on a given CPU for use in ppcopy()
3400  */
3401 void
3402 setup_vaddr_for_ppcopy(struct cpu *cpup)
3403 {
3404 	void *addr;
3405 	hat_mempte_t pte_pa;
3406 
3407 	addr = vmem_alloc(heap_arena, mmu_ptob(1), VM_SLEEP);
3408 	pte_pa = hat_mempte_setup(addr);
3409 	cpup->cpu_caddr1 = addr;
3410 	cpup->cpu_caddr1pte = pte_pa;
3411 
3412 	addr = vmem_alloc(heap_arena, mmu_ptob(1), VM_SLEEP);
3413 	pte_pa = hat_mempte_setup(addr);
3414 	cpup->cpu_caddr2 = addr;
3415 	cpup->cpu_caddr2pte = pte_pa;
3416 
3417 	mutex_init(&cpup->cpu_ppaddr_mutex, NULL, MUTEX_DEFAULT, NULL);
3418 }
3419 
3420 /*
3421  * Undo setup_vaddr_for_ppcopy
3422  */
3423 void
3424 teardown_vaddr_for_ppcopy(struct cpu *cpup)
3425 {
3426 	mutex_destroy(&cpup->cpu_ppaddr_mutex);
3427 
3428 	hat_mempte_release(cpup->cpu_caddr2, cpup->cpu_caddr2pte);
3429 	cpup->cpu_caddr2pte = 0;
3430 	vmem_free(heap_arena, cpup->cpu_caddr2, mmu_ptob(1));
3431 	cpup->cpu_caddr2 = 0;
3432 
3433 	hat_mempte_release(cpup->cpu_caddr1, cpup->cpu_caddr1pte);
3434 	cpup->cpu_caddr1pte = 0;
3435 	vmem_free(heap_arena, cpup->cpu_caddr1, mmu_ptob(1));
3436 	cpup->cpu_caddr1 = 0;
3437 }
3438 
3439 /*
3440  * Create the pageout scanner thread. The thread has to
3441  * start at procedure with process pp and priority pri.
3442  */
3443 void
3444 pageout_init(void (*procedure)(), proc_t *pp, pri_t pri)
3445 {
3446 	(void) thread_create(NULL, 0, procedure, NULL, 0, pp, TS_RUN, pri);
3447 }
3448 
3449 /*
3450  * Function for flushing D-cache when performing module relocations
3451  * to an alternate mapping.  Unnecessary on Intel / AMD platforms.
3452  */
3453 void
3454 dcache_flushall()
3455 {}
3456 
3457 size_t
3458 exec_get_spslew(void)
3459 {
3460 	return (0);
3461 }
3462 
3463 /*
3464  * Allocate a memory page.  The argument 'seed' can be any pseudo-random
3465  * number to vary where the pages come from.  This is quite a hacked up
3466  * method -- it works for now, but really needs to be fixed up a bit.
3467  *
3468  * We currently use page_create_va() on the kvp with fake offsets,
3469  * segments and virt address.  This is pretty bogus, but was copied from the
3470  * old hat_i86.c code.  A better approach would be to specify either mnode
3471  * random or mnode local and takes a page from whatever color has the MOST
3472  * available - this would have a minimal impact on page coloring.
3473  */
3474 page_t *
3475 page_get_physical(uintptr_t seed)
3476 {
3477 	page_t *pp;
3478 	u_offset_t offset;
3479 	static struct seg tmpseg;
3480 	static uintptr_t ctr = 0;
3481 
3482 	/*
3483 	 * This code is gross, we really need a simpler page allocator.
3484 	 *
3485 	 * We need assign an offset for the page to call page_create_va().
3486 	 * To avoid conflicts with other pages, we get creative with the offset.
3487 	 * For 32 bits, we pick an offset > 4Gig
3488 	 * For 64 bits, pick an offset somewhere in the VA hole.
3489 	 */
3490 	offset = seed;
3491 	if (offset > kernelbase)
3492 		offset -= kernelbase;
3493 	offset <<= MMU_PAGESHIFT;
3494 #if defined(__amd64)
3495 	offset += mmu.hole_start;	/* something in VA hole */
3496 #else
3497 	offset += 1ULL << 40;		/* something > 4 Gig */
3498 #endif
3499 
3500 	if (page_resv(1, KM_NOSLEEP) == 0)
3501 		return (NULL);
3502 
3503 #ifdef	DEBUG
3504 	pp = page_exists(&kvp, offset);
3505 	if (pp != NULL)
3506 		panic("page already exists %p", pp);
3507 #endif
3508 
3509 	pp = page_create_va(&kvp, offset, MMU_PAGESIZE, PG_EXCL,
3510 	    &tmpseg, (caddr_t)(ctr += MMU_PAGESIZE));	/* changing VA usage */
3511 	if (pp == NULL)
3512 		return (NULL);
3513 	page_io_unlock(pp);
3514 	page_hashout(pp, NULL);
3515 	return (pp);
3516 }
3517