xref: /titanic_41/usr/src/uts/i86pc/vm/vm_machdep.c (revision 24fe0b3bf671e123467ce1df0b67cadd3614c8e4)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
27 /*	All Rights Reserved   */
28 
29 /*
30  * 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 /*
35  * UNIX machine dependent virtual memory support.
36  */
37 
38 #include <sys/types.h>
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/user.h>
42 #include <sys/proc.h>
43 #include <sys/kmem.h>
44 #include <sys/vmem.h>
45 #include <sys/buf.h>
46 #include <sys/cpuvar.h>
47 #include <sys/lgrp.h>
48 #include <sys/disp.h>
49 #include <sys/vm.h>
50 #include <sys/mman.h>
51 #include <sys/vnode.h>
52 #include <sys/cred.h>
53 #include <sys/exec.h>
54 #include <sys/exechdr.h>
55 #include <sys/debug.h>
56 #include <sys/vmsystm.h>
57 
58 #include <vm/hat.h>
59 #include <vm/as.h>
60 #include <vm/seg.h>
61 #include <vm/seg_kp.h>
62 #include <vm/seg_vn.h>
63 #include <vm/page.h>
64 #include <vm/seg_kmem.h>
65 #include <vm/seg_kpm.h>
66 #include <vm/vm_dep.h>
67 
68 #include <sys/cpu.h>
69 #include <sys/vm_machparam.h>
70 #include <sys/memlist.h>
71 #include <sys/bootconf.h> /* XXX the memlist stuff belongs in memlist_plat.h */
72 #include <vm/hat_i86.h>
73 #include <sys/x86_archext.h>
74 #include <sys/elf_386.h>
75 #include <sys/cmn_err.h>
76 #include <sys/archsystm.h>
77 #include <sys/machsystm.h>
78 
79 #include <sys/vtrace.h>
80 #include <sys/ddidmareq.h>
81 #include <sys/promif.h>
82 #include <sys/memnode.h>
83 #include <sys/stack.h>
84 #include <util/qsort.h>
85 #include <sys/taskq.h>
86 
87 #ifdef __xpv
88 
89 #include <sys/hypervisor.h>
90 #include <sys/xen_mmu.h>
91 #include <sys/balloon_impl.h>
92 
93 /*
94  * domain 0 pages usable for DMA are kept pre-allocated and kept in
95  * distinct lists, ordered by increasing mfn.
96  */
97 static kmutex_t io_pool_lock;
98 static kmutex_t contig_list_lock;
99 static page_t *io_pool_4g;	/* pool for 32 bit dma limited devices */
100 static page_t *io_pool_16m;	/* pool for 24 bit dma limited legacy devices */
101 static long io_pool_cnt;
102 static long io_pool_cnt_max = 0;
103 #define	DEFAULT_IO_POOL_MIN	128
104 static long io_pool_cnt_min = DEFAULT_IO_POOL_MIN;
105 static long io_pool_cnt_lowater = 0;
106 static long io_pool_shrink_attempts; /* how many times did we try to shrink */
107 static long io_pool_shrinks;	/* how many times did we really shrink */
108 static long io_pool_grows;	/* how many times did we grow */
109 static mfn_t start_mfn = 1;
110 static caddr_t io_pool_kva;	/* use to alloc pages when needed */
111 
112 static int create_contig_pfnlist(uint_t);
113 
114 /*
115  * percentage of phys mem to hold in the i/o pool
116  */
117 #define	DEFAULT_IO_POOL_PCT	2
118 static long io_pool_physmem_pct = DEFAULT_IO_POOL_PCT;
119 static void page_io_pool_sub(page_t **, page_t *, page_t *);
120 int ioalloc_dbg = 0;
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[MMU_PAGE_SIZES];	/* free list cnt per szc */
148 	pgcnt_t	mnr_mt_totcnt;		/* sum of cache and free lists */
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)	(mnoderanges[mt].mnr_mt_totcnt)
163 
164 /*
165  * As the PC architecture evolved memory up was clumped into several
166  * ranges for various historical I/O devices to do DMA.
167  * < 16Meg - ISA bus
168  * < 2Gig - ???
169  * < 4Gig - PCI bus or drivers that don't understand PAE mode
170  *
171  * These are listed in reverse order, so that we can skip over unused
172  * ranges on machines with small memories.
173  *
174  * For now under the Hypervisor, we'll only ever have one memrange.
175  */
176 #define	PFN_4GIG	0x100000
177 #define	PFN_16MEG	0x1000
178 static pfn_t arch_memranges[NUM_MEM_RANGES] = {
179     PFN_4GIG,	/* pfn range for 4G and above */
180     0x80000,	/* pfn range for 2G-4G */
181     PFN_16MEG,	/* pfn range for 16M-2G */
182     0x00000,	/* pfn range for 0-16M */
183 };
184 pfn_t *memranges = &arch_memranges[0];
185 int nranges = NUM_MEM_RANGES;
186 
187 /*
188  * This combines mem_node_config and memranges into one data
189  * structure to be used for page list management.
190  */
191 mnoderange_t	*mnoderanges;
192 int		mnoderangecnt;
193 int		mtype4g;
194 
195 /*
196  * 4g memory management variables for systems with more than 4g of memory:
197  *
198  * physical memory below 4g is required for 32bit dma devices and, currently,
199  * for kmem memory. On systems with more than 4g of memory, the pool of memory
200  * below 4g can be depleted without any paging activity given that there is
201  * likely to be sufficient memory above 4g.
202  *
203  * physmax4g is set true if the largest pfn is over 4g. The rest of the
204  * 4g memory management code is enabled only when physmax4g is true.
205  *
206  * maxmem4g is the count of the maximum number of pages on the page lists
207  * with physical addresses below 4g. It can be a lot less then 4g given that
208  * BIOS may reserve large chunks of space below 4g for hot plug pci devices,
209  * agp aperture etc.
210  *
211  * freemem4g maintains the count of the number of available pages on the
212  * page lists with physical addresses below 4g.
213  *
214  * DESFREE4G specifies the desired amount of below 4g memory. It defaults to
215  * 6% (desfree4gshift = 4) of maxmem4g.
216  *
217  * RESTRICT4G_ALLOC returns true if freemem4g falls below DESFREE4G
218  * and the amount of physical memory above 4g is greater than freemem4g.
219  * In this case, page_get_* routines will restrict below 4g allocations
220  * for requests that don't specifically require it.
221  */
222 
223 #define	LOTSFREE4G	(maxmem4g >> lotsfree4gshift)
224 #define	DESFREE4G	(maxmem4g >> desfree4gshift)
225 
226 #define	RESTRICT4G_ALLOC					\
227 	(physmax4g && (freemem4g < DESFREE4G) && ((freemem4g << 1) < freemem))
228 
229 static pgcnt_t	maxmem4g;
230 static pgcnt_t	freemem4g;
231 static int	physmax4g;
232 static int	desfree4gshift = 4;	/* maxmem4g shift to derive DESFREE4G */
233 static int	lotsfree4gshift = 3;
234 
235 /*
236  * 16m memory management:
237  *
238  * reserve some amount of physical memory below 16m for legacy devices.
239  *
240  * RESTRICT16M_ALLOC returns true if an there are sufficient free pages above
241  * 16m or if the 16m pool drops below DESFREE16M.
242  *
243  * In this case, general page allocations via page_get_{free,cache}list
244  * routines will be restricted from allocating from the 16m pool. Allocations
245  * that require specific pfn ranges (page_get_anylist) and PG_PANIC allocations
246  * are not restricted.
247  */
248 
249 #define	FREEMEM16M	MTYPE_FREEMEM(0)
250 #define	DESFREE16M	desfree16m
251 #define	RESTRICT16M_ALLOC(freemem, pgcnt, flags)		\
252 	((freemem != 0) && ((flags & PG_PANIC) == 0) &&		\
253 	    ((freemem >= (FREEMEM16M)) ||			\
254 	    (FREEMEM16M  < (DESFREE16M + pgcnt))))
255 
256 static pgcnt_t	desfree16m = 0x380;
257 
258 /*
259  * This can be patched via /etc/system to allow old non-PAE aware device
260  * drivers to use kmem_alloc'd memory on 32 bit systems with > 4Gig RAM.
261  */
262 int restricted_kmemalloc = 0;
263 
264 #ifdef VM_STATS
265 struct {
266 	ulong_t	pga_alloc;
267 	ulong_t	pga_notfullrange;
268 	ulong_t	pga_nulldmaattr;
269 	ulong_t	pga_allocok;
270 	ulong_t	pga_allocfailed;
271 	ulong_t	pgma_alloc;
272 	ulong_t	pgma_allocok;
273 	ulong_t	pgma_allocfailed;
274 	ulong_t	pgma_allocempty;
275 } pga_vmstats;
276 #endif
277 
278 uint_t mmu_page_sizes;
279 
280 /* How many page sizes the users can see */
281 uint_t mmu_exported_page_sizes;
282 
283 /* page sizes that legacy applications can see */
284 uint_t mmu_legacy_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.umax_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 	case MAPPGSZ_ISM:
403 		for (l = mmu.umax_page_level; l > 0; --l) {
404 			if (len >= LEVEL_SIZE(l))
405 				return (LEVEL_SIZE(l));
406 		}
407 		return (LEVEL_SIZE(0));
408 	}
409 	return (pgsz);
410 }
411 
412 static uint_t
413 map_szcvec(caddr_t addr, size_t size, uintptr_t off, size_t max_lpsize,
414     size_t min_physmem)
415 {
416 	caddr_t eaddr = addr + size;
417 	uint_t szcvec = 0;
418 	caddr_t raddr;
419 	caddr_t readdr;
420 	size_t	pgsz;
421 	int i;
422 
423 	if (physmem < min_physmem || max_lpsize <= MMU_PAGESIZE) {
424 		return (0);
425 	}
426 
427 	for (i = mmu_exported_page_sizes - 1; i > 0; i--) {
428 		pgsz = page_get_pagesize(i);
429 		if (pgsz > max_lpsize) {
430 			continue;
431 		}
432 		raddr = (caddr_t)P2ROUNDUP((uintptr_t)addr, pgsz);
433 		readdr = (caddr_t)P2ALIGN((uintptr_t)eaddr, pgsz);
434 		if (raddr < addr || raddr >= readdr) {
435 			continue;
436 		}
437 		if (P2PHASE((uintptr_t)addr ^ off, pgsz)) {
438 			continue;
439 		}
440 		/*
441 		 * Set szcvec to the remaining page sizes.
442 		 */
443 		szcvec = ((1 << (i + 1)) - 1) & ~1;
444 		break;
445 	}
446 	return (szcvec);
447 }
448 
449 /*
450  * Return a bit vector of large page size codes that
451  * can be used to map [addr, addr + len) region.
452  */
453 /*ARGSUSED*/
454 uint_t
455 map_pgszcvec(caddr_t addr, size_t size, uintptr_t off, int flags, int type,
456     int memcntl)
457 {
458 	size_t max_lpsize = mcntl0_lpsize;
459 
460 	if (mmu.max_page_level == 0)
461 		return (0);
462 
463 	if (flags & MAP_TEXT) {
464 		if (!memcntl)
465 			max_lpsize = max_utext_lpsize;
466 		return (map_szcvec(addr, size, off, max_lpsize,
467 		    shm_lpg_min_physmem));
468 
469 	} else if (flags & MAP_INITDATA) {
470 		if (!memcntl)
471 			max_lpsize = max_uidata_lpsize;
472 		return (map_szcvec(addr, size, off, max_lpsize,
473 		    privm_lpg_min_physmem));
474 
475 	} else if (type == MAPPGSZC_SHM) {
476 		if (!memcntl)
477 			max_lpsize = max_shm_lpsize;
478 		return (map_szcvec(addr, size, off, max_lpsize,
479 		    shm_lpg_min_physmem));
480 
481 	} else if (type == MAPPGSZC_HEAP) {
482 		if (!memcntl)
483 			max_lpsize = max_uheap_lpsize;
484 		return (map_szcvec(addr, size, off, max_lpsize,
485 		    privm_lpg_min_physmem));
486 
487 	} else if (type == MAPPGSZC_STACK) {
488 		if (!memcntl)
489 			max_lpsize = max_ustack_lpsize;
490 		return (map_szcvec(addr, size, off, max_lpsize,
491 		    privm_lpg_min_physmem));
492 
493 	} else {
494 		if (!memcntl)
495 			max_lpsize = max_privmap_lpsize;
496 		return (map_szcvec(addr, size, off, max_lpsize,
497 		    privm_lpg_min_physmem));
498 	}
499 }
500 
501 /*
502  * Handle a pagefault.
503  */
504 faultcode_t
505 pagefault(
506 	caddr_t addr,
507 	enum fault_type type,
508 	enum seg_rw rw,
509 	int iskernel)
510 {
511 	struct as *as;
512 	struct hat *hat;
513 	struct proc *p;
514 	kthread_t *t;
515 	faultcode_t res;
516 	caddr_t base;
517 	size_t len;
518 	int err;
519 	int mapped_red;
520 	uintptr_t ea;
521 
522 	ASSERT_STACK_ALIGNED();
523 
524 	if (INVALID_VADDR(addr))
525 		return (FC_NOMAP);
526 
527 	mapped_red = segkp_map_red();
528 
529 	if (iskernel) {
530 		as = &kas;
531 		hat = as->a_hat;
532 	} else {
533 		t = curthread;
534 		p = ttoproc(t);
535 		as = p->p_as;
536 		hat = as->a_hat;
537 	}
538 
539 	/*
540 	 * Dispatch pagefault.
541 	 */
542 	res = as_fault(hat, as, addr, 1, type, rw);
543 
544 	/*
545 	 * If this isn't a potential unmapped hole in the user's
546 	 * UNIX data or stack segments, just return status info.
547 	 */
548 	if (res != FC_NOMAP || iskernel)
549 		goto out;
550 
551 	/*
552 	 * Check to see if we happened to faulted on a currently unmapped
553 	 * part of the UNIX data or stack segments.  If so, create a zfod
554 	 * mapping there and then try calling the fault routine again.
555 	 */
556 	base = p->p_brkbase;
557 	len = p->p_brksize;
558 
559 	if (addr < base || addr >= base + len) {		/* data seg? */
560 		base = (caddr_t)p->p_usrstack - p->p_stksize;
561 		len = p->p_stksize;
562 		if (addr < base || addr >= p->p_usrstack) {	/* stack seg? */
563 			/* not in either UNIX data or stack segments */
564 			res = FC_NOMAP;
565 			goto out;
566 		}
567 	}
568 
569 	/*
570 	 * the rest of this function implements a 3.X 4.X 5.X compatibility
571 	 * This code is probably not needed anymore
572 	 */
573 	if (p->p_model == DATAMODEL_ILP32) {
574 
575 		/* expand the gap to the page boundaries on each side */
576 		ea = P2ROUNDUP((uintptr_t)base + len, MMU_PAGESIZE);
577 		base = (caddr_t)P2ALIGN((uintptr_t)base, MMU_PAGESIZE);
578 		len = ea - (uintptr_t)base;
579 
580 		as_rangelock(as);
581 		if (as_gap(as, MMU_PAGESIZE, &base, &len, AH_CONTAIN, addr) ==
582 		    0) {
583 			err = as_map(as, base, len, segvn_create, zfod_argsp);
584 			as_rangeunlock(as);
585 			if (err) {
586 				res = FC_MAKE_ERR(err);
587 				goto out;
588 			}
589 		} else {
590 			/*
591 			 * This page is already mapped by another thread after
592 			 * we returned from as_fault() above.  We just fall
593 			 * through as_fault() below.
594 			 */
595 			as_rangeunlock(as);
596 		}
597 
598 		res = as_fault(hat, as, addr, 1, F_INVAL, rw);
599 	}
600 
601 out:
602 	if (mapped_red)
603 		segkp_unmap_red();
604 
605 	return (res);
606 }
607 
608 void
609 map_addr(caddr_t *addrp, size_t len, offset_t off, int vacalign, uint_t flags)
610 {
611 	struct proc *p = curproc;
612 	caddr_t userlimit = (flags & _MAP_LOW32) ?
613 	    (caddr_t)_userlimit32 : p->p_as->a_userlimit;
614 
615 	map_addr_proc(addrp, len, off, vacalign, userlimit, curproc, flags);
616 }
617 
618 /*ARGSUSED*/
619 int
620 map_addr_vacalign_check(caddr_t addr, u_offset_t off)
621 {
622 	return (0);
623 }
624 
625 /*
626  * map_addr_proc() is the routine called when the system is to
627  * choose an address for the user.  We will pick an address
628  * range which is the highest available below userlimit.
629  *
630  * Every mapping will have a redzone of a single page on either side of
631  * the request. This is done to leave one page unmapped between segments.
632  * This is not required, but it's useful for the user because if their
633  * program strays across a segment boundary, it will catch a fault
634  * immediately making debugging a little easier.  Currently the redzone
635  * is mandatory.
636  *
637  * addrp is a value/result parameter.
638  *	On input it is a hint from the user to be used in a completely
639  *	machine dependent fashion.  We decide to completely ignore this hint.
640  *	If MAP_ALIGN was specified, addrp contains the minimal alignment, which
641  *	must be some "power of two" multiple of pagesize.
642  *
643  *	On output it is NULL if no address can be found in the current
644  *	processes address space or else an address that is currently
645  *	not mapped for len bytes with a page of red zone on either side.
646  *
647  *	vacalign is not needed on x86 (it's for viturally addressed caches)
648  */
649 /*ARGSUSED*/
650 void
651 map_addr_proc(
652 	caddr_t *addrp,
653 	size_t len,
654 	offset_t off,
655 	int vacalign,
656 	caddr_t userlimit,
657 	struct proc *p,
658 	uint_t flags)
659 {
660 	struct as *as = p->p_as;
661 	caddr_t addr;
662 	caddr_t base;
663 	size_t slen;
664 	size_t align_amount;
665 
666 	ASSERT32(userlimit == as->a_userlimit);
667 
668 	base = p->p_brkbase;
669 #if defined(__amd64)
670 	/*
671 	 * XX64 Yes, this needs more work.
672 	 */
673 	if (p->p_model == DATAMODEL_NATIVE) {
674 		if (userlimit < as->a_userlimit) {
675 			/*
676 			 * This happens when a program wants to map
677 			 * something in a range that's accessible to a
678 			 * program in a smaller address space.  For example,
679 			 * a 64-bit program calling mmap32(2) to guarantee
680 			 * that the returned address is below 4Gbytes.
681 			 */
682 			ASSERT((uintptr_t)userlimit < ADDRESS_C(0xffffffff));
683 
684 			if (userlimit > base)
685 				slen = userlimit - base;
686 			else {
687 				*addrp = NULL;
688 				return;
689 			}
690 		} else {
691 			/*
692 			 * XX64 This layout is probably wrong .. but in
693 			 * the event we make the amd64 address space look
694 			 * like sparcv9 i.e. with the stack -above- the
695 			 * heap, this bit of code might even be correct.
696 			 */
697 			slen = p->p_usrstack - base -
698 			    ((p->p_stk_ctl + PAGEOFFSET) & PAGEMASK);
699 		}
700 	} else
701 #endif
702 		slen = userlimit - base;
703 
704 	/* Make len be a multiple of PAGESIZE */
705 	len = (len + PAGEOFFSET) & PAGEMASK;
706 
707 	/*
708 	 * figure out what the alignment should be
709 	 *
710 	 * XX64 -- is there an ELF_AMD64_MAXPGSZ or is it the same????
711 	 */
712 	if (len <= ELF_386_MAXPGSZ) {
713 		/*
714 		 * Align virtual addresses to ensure that ELF shared libraries
715 		 * are mapped with the appropriate alignment constraints by
716 		 * the run-time linker.
717 		 */
718 		align_amount = ELF_386_MAXPGSZ;
719 	} else {
720 		int l = mmu.umax_page_level;
721 
722 		while (l && len < LEVEL_SIZE(l))
723 			--l;
724 
725 		align_amount = LEVEL_SIZE(l);
726 	}
727 
728 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount))
729 		align_amount = (uintptr_t)*addrp;
730 
731 	ASSERT(ISP2(align_amount));
732 	ASSERT(align_amount == 0 || align_amount >= PAGESIZE);
733 
734 	off = off & (align_amount - 1);
735 	/*
736 	 * Look for a large enough hole starting below userlimit.
737 	 * After finding it, use the upper part.
738 	 */
739 	if (as_gap_aligned(as, len, &base, &slen, AH_HI, NULL, align_amount,
740 	    PAGESIZE, off) == 0) {
741 		caddr_t as_addr;
742 
743 		/*
744 		 * addr is the highest possible address to use since we have
745 		 * a PAGESIZE redzone at the beginning and end.
746 		 */
747 		addr = base + slen - (PAGESIZE + len);
748 		as_addr = addr;
749 		/*
750 		 * Round address DOWN to the alignment amount and
751 		 * add the offset in.
752 		 * If addr is greater than as_addr, len would not be large
753 		 * enough to include the redzone, so we must adjust down
754 		 * by the alignment amount.
755 		 */
756 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1)));
757 		addr += (uintptr_t)off;
758 		if (addr > as_addr) {
759 			addr -= align_amount;
760 		}
761 
762 		ASSERT(addr > base);
763 		ASSERT(addr + len < base + slen);
764 		ASSERT(((uintptr_t)addr & (align_amount - 1)) ==
765 		    ((uintptr_t)(off)));
766 		*addrp = addr;
767 	} else {
768 		*addrp = NULL;	/* no more virtual space */
769 	}
770 }
771 
772 int valid_va_range_aligned_wraparound;
773 
774 /*
775  * Determine whether [*basep, *basep + *lenp) contains a mappable range of
776  * addresses at least "minlen" long, where the base of the range is at "off"
777  * phase from an "align" boundary and there is space for a "redzone"-sized
778  * redzone on either side of the range.  On success, 1 is returned and *basep
779  * and *lenp are adjusted to describe the acceptable range (including
780  * the redzone).  On failure, 0 is returned.
781  */
782 /*ARGSUSED3*/
783 int
784 valid_va_range_aligned(caddr_t *basep, size_t *lenp, size_t minlen, int dir,
785     size_t align, size_t redzone, size_t off)
786 {
787 	uintptr_t hi, lo;
788 	size_t tot_len;
789 
790 	ASSERT(align == 0 ? off == 0 : off < align);
791 	ASSERT(ISP2(align));
792 	ASSERT(align == 0 || align >= PAGESIZE);
793 
794 	lo = (uintptr_t)*basep;
795 	hi = lo + *lenp;
796 	tot_len = minlen + 2 * redzone; /* need at least this much space */
797 
798 	/*
799 	 * If hi rolled over the top, try cutting back.
800 	 */
801 	if (hi < lo) {
802 		*lenp = 0UL - lo - 1UL;
803 		/* See if this really happens. If so, then we figure out why */
804 		valid_va_range_aligned_wraparound++;
805 		hi = lo + *lenp;
806 	}
807 	if (*lenp < tot_len) {
808 		return (0);
809 	}
810 
811 #if defined(__amd64)
812 	/*
813 	 * Deal with a possible hole in the address range between
814 	 * hole_start and hole_end that should never be mapped.
815 	 */
816 	if (lo < hole_start) {
817 		if (hi > hole_start) {
818 			if (hi < hole_end) {
819 				hi = hole_start;
820 			} else {
821 				/* lo < hole_start && hi >= hole_end */
822 				if (dir == AH_LO) {
823 					/*
824 					 * prefer lowest range
825 					 */
826 					if (hole_start - lo >= tot_len)
827 						hi = hole_start;
828 					else if (hi - hole_end >= tot_len)
829 						lo = hole_end;
830 					else
831 						return (0);
832 				} else {
833 					/*
834 					 * prefer highest range
835 					 */
836 					if (hi - hole_end >= tot_len)
837 						lo = hole_end;
838 					else if (hole_start - lo >= tot_len)
839 						hi = hole_start;
840 					else
841 						return (0);
842 				}
843 			}
844 		}
845 	} else {
846 		/* lo >= hole_start */
847 		if (hi < hole_end)
848 			return (0);
849 		if (lo < hole_end)
850 			lo = hole_end;
851 	}
852 #endif
853 
854 	if (hi - lo < tot_len)
855 		return (0);
856 
857 	if (align > 1) {
858 		uintptr_t tlo = lo + redzone;
859 		uintptr_t thi = hi - redzone;
860 		tlo = (uintptr_t)P2PHASEUP(tlo, align, off);
861 		if (tlo < lo + redzone) {
862 			return (0);
863 		}
864 		if (thi < tlo || thi - tlo < minlen) {
865 			return (0);
866 		}
867 	}
868 
869 	*basep = (caddr_t)lo;
870 	*lenp = hi - lo;
871 	return (1);
872 }
873 
874 /*
875  * Determine whether [*basep, *basep + *lenp) contains a mappable range of
876  * addresses at least "minlen" long.  On success, 1 is returned and *basep
877  * and *lenp are adjusted to describe the acceptable range.  On failure, 0
878  * is returned.
879  */
880 int
881 valid_va_range(caddr_t *basep, size_t *lenp, size_t minlen, int dir)
882 {
883 	return (valid_va_range_aligned(basep, lenp, minlen, dir, 0, 0, 0));
884 }
885 
886 /*
887  * Determine whether [addr, addr+len] are valid user addresses.
888  */
889 /*ARGSUSED*/
890 int
891 valid_usr_range(caddr_t addr, size_t len, uint_t prot, struct as *as,
892     caddr_t userlimit)
893 {
894 	caddr_t eaddr = addr + len;
895 
896 	if (eaddr <= addr || addr >= userlimit || eaddr > userlimit)
897 		return (RANGE_BADADDR);
898 
899 #if defined(__amd64)
900 	/*
901 	 * Check for the VA hole
902 	 */
903 	if (eaddr > (caddr_t)hole_start && addr < (caddr_t)hole_end)
904 		return (RANGE_BADADDR);
905 #endif
906 
907 	return (RANGE_OKAY);
908 }
909 
910 /*
911  * Return 1 if the page frame is onboard memory, else 0.
912  */
913 int
914 pf_is_memory(pfn_t pf)
915 {
916 	if (pfn_is_foreign(pf))
917 		return (0);
918 	return (address_in_memlist(phys_install, pfn_to_pa(pf), 1));
919 }
920 
921 /*
922  * return the memrange containing pfn
923  */
924 int
925 memrange_num(pfn_t pfn)
926 {
927 	int n;
928 
929 	for (n = 0; n < nranges - 1; ++n) {
930 		if (pfn >= memranges[n])
931 			break;
932 	}
933 	return (n);
934 }
935 
936 /*
937  * return the mnoderange containing pfn
938  */
939 /*ARGSUSED*/
940 int
941 pfn_2_mtype(pfn_t pfn)
942 {
943 #if defined(__xpv)
944 	return (0);
945 #else
946 	int	n;
947 
948 	for (n = mnoderangecnt - 1; n >= 0; n--) {
949 		if (pfn >= mnoderanges[n].mnr_pfnlo) {
950 			break;
951 		}
952 	}
953 	return (n);
954 #endif
955 }
956 
957 #if !defined(__xpv)
958 /*
959  * is_contigpage_free:
960  *	returns a page list of contiguous pages. It minimally has to return
961  *	minctg pages. Caller determines minctg based on the scatter-gather
962  *	list length.
963  *
964  *	pfnp is set to the next page frame to search on return.
965  */
966 static page_t *
967 is_contigpage_free(
968 	pfn_t *pfnp,
969 	pgcnt_t *pgcnt,
970 	pgcnt_t minctg,
971 	uint64_t pfnseg,
972 	int iolock)
973 {
974 	int	i = 0;
975 	pfn_t	pfn = *pfnp;
976 	page_t	*pp;
977 	page_t	*plist = NULL;
978 
979 	/*
980 	 * fail if pfn + minctg crosses a segment boundary.
981 	 * Adjust for next starting pfn to begin at segment boundary.
982 	 */
983 
984 	if (((*pfnp + minctg - 1) & pfnseg) < (*pfnp & pfnseg)) {
985 		*pfnp = roundup(*pfnp, pfnseg + 1);
986 		return (NULL);
987 	}
988 
989 	do {
990 retry:
991 		pp = page_numtopp_nolock(pfn + i);
992 		if ((pp == NULL) ||
993 		    (page_trylock(pp, SE_EXCL) == 0)) {
994 			(*pfnp)++;
995 			break;
996 		}
997 		if (page_pptonum(pp) != pfn + i) {
998 			page_unlock(pp);
999 			goto retry;
1000 		}
1001 
1002 		if (!(PP_ISFREE(pp))) {
1003 			page_unlock(pp);
1004 			(*pfnp)++;
1005 			break;
1006 		}
1007 
1008 		if (!PP_ISAGED(pp)) {
1009 			page_list_sub(pp, PG_CACHE_LIST);
1010 			page_hashout(pp, (kmutex_t *)NULL);
1011 		} else {
1012 			page_list_sub(pp, PG_FREE_LIST);
1013 		}
1014 
1015 		if (iolock)
1016 			page_io_lock(pp);
1017 		page_list_concat(&plist, &pp);
1018 
1019 		/*
1020 		 * exit loop when pgcnt satisfied or segment boundary reached.
1021 		 */
1022 
1023 	} while ((++i < *pgcnt) && ((pfn + i) & pfnseg));
1024 
1025 	*pfnp += i;		/* set to next pfn to search */
1026 
1027 	if (i >= minctg) {
1028 		*pgcnt -= i;
1029 		return (plist);
1030 	}
1031 
1032 	/*
1033 	 * failure: minctg not satisfied.
1034 	 *
1035 	 * if next request crosses segment boundary, set next pfn
1036 	 * to search from the segment boundary.
1037 	 */
1038 	if (((*pfnp + minctg - 1) & pfnseg) < (*pfnp & pfnseg))
1039 		*pfnp = roundup(*pfnp, pfnseg + 1);
1040 
1041 	/* clean up any pages already allocated */
1042 
1043 	while (plist) {
1044 		pp = plist;
1045 		page_sub(&plist, pp);
1046 		page_list_add(pp, PG_FREE_LIST | PG_LIST_TAIL);
1047 		if (iolock)
1048 			page_io_unlock(pp);
1049 		page_unlock(pp);
1050 	}
1051 
1052 	return (NULL);
1053 }
1054 #endif	/* !__xpv */
1055 
1056 /*
1057  * verify that pages being returned from allocator have correct DMA attribute
1058  */
1059 #ifndef DEBUG
1060 #define	check_dma(a, b, c) (0)
1061 #else
1062 static void
1063 check_dma(ddi_dma_attr_t *dma_attr, page_t *pp, int cnt)
1064 {
1065 	if (dma_attr == NULL)
1066 		return;
1067 
1068 	while (cnt-- > 0) {
1069 		if (pa_to_ma(pfn_to_pa(pp->p_pagenum)) <
1070 		    dma_attr->dma_attr_addr_lo)
1071 			panic("PFN (pp=%p) below dma_attr_addr_lo", (void *)pp);
1072 		if (pa_to_ma(pfn_to_pa(pp->p_pagenum)) >=
1073 		    dma_attr->dma_attr_addr_hi)
1074 			panic("PFN (pp=%p) above dma_attr_addr_hi", (void *)pp);
1075 		pp = pp->p_next;
1076 	}
1077 }
1078 #endif
1079 
1080 #if !defined(__xpv)
1081 static page_t *
1082 page_get_contigpage(pgcnt_t *pgcnt, ddi_dma_attr_t *mattr, int iolock)
1083 {
1084 	pfn_t		pfn;
1085 	int		sgllen;
1086 	uint64_t	pfnseg;
1087 	pgcnt_t		minctg;
1088 	page_t		*pplist = NULL, *plist;
1089 	uint64_t	lo, hi;
1090 	pgcnt_t		pfnalign = 0;
1091 	static pfn_t	startpfn;
1092 	static pgcnt_t	lastctgcnt;
1093 	uintptr_t	align;
1094 
1095 	CONTIG_LOCK();
1096 
1097 	if (mattr) {
1098 		lo = mmu_btop((mattr->dma_attr_addr_lo + MMU_PAGEOFFSET));
1099 		hi = mmu_btop(mattr->dma_attr_addr_hi);
1100 		if (hi >= physmax)
1101 			hi = physmax - 1;
1102 		sgllen = mattr->dma_attr_sgllen;
1103 		pfnseg = mmu_btop(mattr->dma_attr_seg);
1104 
1105 		align = maxbit(mattr->dma_attr_align, mattr->dma_attr_minxfer);
1106 		if (align > MMU_PAGESIZE)
1107 			pfnalign = mmu_btop(align);
1108 
1109 		/*
1110 		 * in order to satisfy the request, must minimally
1111 		 * acquire minctg contiguous pages
1112 		 */
1113 		minctg = howmany(*pgcnt, sgllen);
1114 
1115 		ASSERT(hi >= lo);
1116 
1117 		/*
1118 		 * start from where last searched if the minctg >= lastctgcnt
1119 		 */
1120 		if (minctg < lastctgcnt || startpfn < lo || startpfn > hi)
1121 			startpfn = lo;
1122 	} else {
1123 		hi = physmax - 1;
1124 		lo = 0;
1125 		sgllen = 1;
1126 		pfnseg = mmu.highest_pfn;
1127 		minctg = *pgcnt;
1128 
1129 		if (minctg < lastctgcnt)
1130 			startpfn = lo;
1131 	}
1132 	lastctgcnt = minctg;
1133 
1134 	ASSERT(pfnseg + 1 >= (uint64_t)minctg);
1135 
1136 	/* conserve 16m memory - start search above 16m when possible */
1137 	if (hi > PFN_16M && startpfn < PFN_16M)
1138 		startpfn = PFN_16M;
1139 
1140 	pfn = startpfn;
1141 	if (pfnalign)
1142 		pfn = P2ROUNDUP(pfn, pfnalign);
1143 
1144 	while (pfn + minctg - 1 <= hi) {
1145 
1146 		plist = is_contigpage_free(&pfn, pgcnt, minctg, pfnseg, iolock);
1147 		if (plist) {
1148 			page_list_concat(&pplist, &plist);
1149 			sgllen--;
1150 			/*
1151 			 * return when contig pages no longer needed
1152 			 */
1153 			if (!*pgcnt || ((*pgcnt <= sgllen) && !pfnalign)) {
1154 				startpfn = pfn;
1155 				CONTIG_UNLOCK();
1156 				check_dma(mattr, pplist, *pgcnt);
1157 				return (pplist);
1158 			}
1159 			minctg = howmany(*pgcnt, sgllen);
1160 		}
1161 		if (pfnalign)
1162 			pfn = P2ROUNDUP(pfn, pfnalign);
1163 	}
1164 
1165 	/* cannot find contig pages in specified range */
1166 	if (startpfn == lo) {
1167 		CONTIG_UNLOCK();
1168 		return (NULL);
1169 	}
1170 
1171 	/* did not start with lo previously */
1172 	pfn = lo;
1173 	if (pfnalign)
1174 		pfn = P2ROUNDUP(pfn, pfnalign);
1175 
1176 	/* allow search to go above startpfn */
1177 	while (pfn < startpfn) {
1178 
1179 		plist = is_contigpage_free(&pfn, pgcnt, minctg, pfnseg, iolock);
1180 		if (plist != NULL) {
1181 
1182 			page_list_concat(&pplist, &plist);
1183 			sgllen--;
1184 
1185 			/*
1186 			 * return when contig pages no longer needed
1187 			 */
1188 			if (!*pgcnt || ((*pgcnt <= sgllen) && !pfnalign)) {
1189 				startpfn = pfn;
1190 				CONTIG_UNLOCK();
1191 				check_dma(mattr, pplist, *pgcnt);
1192 				return (pplist);
1193 			}
1194 			minctg = howmany(*pgcnt, sgllen);
1195 		}
1196 		if (pfnalign)
1197 			pfn = P2ROUNDUP(pfn, pfnalign);
1198 	}
1199 	CONTIG_UNLOCK();
1200 	return (NULL);
1201 }
1202 #endif	/* !__xpv */
1203 
1204 /*
1205  * mnode_range_cnt() calculates the number of memory ranges for mnode and
1206  * memranges[]. Used to determine the size of page lists and mnoderanges.
1207  */
1208 int
1209 mnode_range_cnt(int mnode)
1210 {
1211 #if defined(__xpv)
1212 	ASSERT(mnode == 0);
1213 	return (1);
1214 #else	/* __xpv */
1215 	int	mri;
1216 	int	mnrcnt = 0;
1217 
1218 	if (mem_node_config[mnode].exists != 0) {
1219 		mri = nranges - 1;
1220 
1221 		/* find the memranges index below contained in mnode range */
1222 
1223 		while (MEMRANGEHI(mri) < mem_node_config[mnode].physbase)
1224 			mri--;
1225 
1226 		/*
1227 		 * increment mnode range counter when memranges or mnode
1228 		 * boundary is reached.
1229 		 */
1230 		while (mri >= 0 &&
1231 		    mem_node_config[mnode].physmax >= MEMRANGELO(mri)) {
1232 			mnrcnt++;
1233 			if (mem_node_config[mnode].physmax > MEMRANGEHI(mri))
1234 				mri--;
1235 			else
1236 				break;
1237 		}
1238 	}
1239 	ASSERT(mnrcnt <= MAX_MNODE_MRANGES);
1240 	return (mnrcnt);
1241 #endif	/* __xpv */
1242 }
1243 
1244 /*
1245  * mnode_range_setup() initializes mnoderanges.
1246  */
1247 void
1248 mnode_range_setup(mnoderange_t *mnoderanges)
1249 {
1250 	int	mnode, mri;
1251 
1252 	for (mnode = 0; mnode < max_mem_nodes; mnode++) {
1253 		if (mem_node_config[mnode].exists == 0)
1254 			continue;
1255 
1256 		mri = nranges - 1;
1257 
1258 		while (MEMRANGEHI(mri) < mem_node_config[mnode].physbase)
1259 			mri--;
1260 
1261 		while (mri >= 0 && mem_node_config[mnode].physmax >=
1262 		    MEMRANGELO(mri)) {
1263 			mnoderanges->mnr_pfnlo = MAX(MEMRANGELO(mri),
1264 			    mem_node_config[mnode].physbase);
1265 			mnoderanges->mnr_pfnhi = MIN(MEMRANGEHI(mri),
1266 			    mem_node_config[mnode].physmax);
1267 			mnoderanges->mnr_mnode = mnode;
1268 			mnoderanges->mnr_memrange = mri;
1269 			mnoderanges++;
1270 			if (mem_node_config[mnode].physmax > MEMRANGEHI(mri))
1271 				mri--;
1272 			else
1273 				break;
1274 		}
1275 	}
1276 }
1277 
1278 /*ARGSUSED*/
1279 int
1280 mtype_init(vnode_t *vp, caddr_t vaddr, uint_t *flags, size_t pgsz)
1281 {
1282 	int mtype = mnoderangecnt - 1;
1283 
1284 #if !defined(__xpv)
1285 #if defined(__i386)
1286 	/*
1287 	 * set the mtype range
1288 	 * - kmem requests needs to be below 4g if restricted_kmemalloc is set.
1289 	 * - for non kmem requests, set range to above 4g if memory below 4g
1290 	 * runs low.
1291 	 */
1292 	if (restricted_kmemalloc && VN_ISKAS(vp) &&
1293 	    (caddr_t)(vaddr) >= kernelheap &&
1294 	    (caddr_t)(vaddr) < ekernelheap) {
1295 		ASSERT(physmax4g);
1296 		mtype = mtype4g;
1297 		if (RESTRICT16M_ALLOC(freemem4g - btop(pgsz),
1298 		    btop(pgsz), *flags)) {
1299 			*flags |= PGI_MT_RANGE16M;
1300 		} else {
1301 			VM_STAT_ADD(vmm_vmstats.unrestrict16mcnt);
1302 			VM_STAT_COND_ADD((*flags & PG_PANIC),
1303 			    vmm_vmstats.pgpanicalloc);
1304 			*flags |= PGI_MT_RANGE0;
1305 		}
1306 		return (mtype);
1307 	}
1308 #endif	/* __i386 */
1309 
1310 	if (RESTRICT4G_ALLOC) {
1311 		VM_STAT_ADD(vmm_vmstats.restrict4gcnt);
1312 		/* here only for > 4g systems */
1313 		*flags |= PGI_MT_RANGE4G;
1314 	} else if (RESTRICT16M_ALLOC(freemem, btop(pgsz), *flags)) {
1315 		*flags |= PGI_MT_RANGE16M;
1316 	} else {
1317 		VM_STAT_ADD(vmm_vmstats.unrestrict16mcnt);
1318 		VM_STAT_COND_ADD((*flags & PG_PANIC), vmm_vmstats.pgpanicalloc);
1319 		*flags |= PGI_MT_RANGE0;
1320 	}
1321 #endif /* !__xpv */
1322 	return (mtype);
1323 }
1324 
1325 
1326 /* mtype init for page_get_replacement_page */
1327 /*ARGSUSED*/
1328 int
1329 mtype_pgr_init(int *flags, page_t *pp, int mnode, pgcnt_t pgcnt)
1330 {
1331 	int mtype = mnoderangecnt - 1;
1332 #if !defined(__ixpv)
1333 	if (RESTRICT16M_ALLOC(freemem, pgcnt, *flags)) {
1334 		*flags |= PGI_MT_RANGE16M;
1335 	} else {
1336 		VM_STAT_ADD(vmm_vmstats.unrestrict16mcnt);
1337 		*flags |= PGI_MT_RANGE0;
1338 	}
1339 #endif
1340 	return (mtype);
1341 }
1342 
1343 /*
1344  * Determine if the mnode range specified in mtype contains memory belonging
1345  * to memory node mnode.  If flags & PGI_MT_RANGE is set then mtype contains
1346  * the range of indices from high pfn to 0, 16m or 4g.
1347  *
1348  * Return first mnode range type index found otherwise return -1 if none found.
1349  */
1350 int
1351 mtype_func(int mnode, int mtype, uint_t flags)
1352 {
1353 	if (flags & PGI_MT_RANGE) {
1354 		int	mtlim = 0;
1355 
1356 		if (flags & PGI_MT_NEXT)
1357 			mtype--;
1358 		if (flags & PGI_MT_RANGE4G)
1359 			mtlim = mtype4g + 1;	/* exclude 0-4g range */
1360 		else if (flags & PGI_MT_RANGE16M)
1361 			mtlim = 1;		/* exclude 0-16m range */
1362 		while (mtype >= mtlim) {
1363 			if (mnoderanges[mtype].mnr_mnode == mnode)
1364 				return (mtype);
1365 			mtype--;
1366 		}
1367 	} else if (mnoderanges[mtype].mnr_mnode == mnode) {
1368 		return (mtype);
1369 	}
1370 	return (-1);
1371 }
1372 
1373 /*
1374  * Update the page list max counts with the pfn range specified by the
1375  * input parameters.  Called from add_physmem() when physical memory with
1376  * page_t's are initially added to the page lists.
1377  */
1378 void
1379 mtype_modify_max(pfn_t startpfn, long cnt)
1380 {
1381 	int	mtype = 0;
1382 	pfn_t	endpfn = startpfn + cnt, pfn;
1383 	pgcnt_t	inc;
1384 
1385 	ASSERT(cnt > 0);
1386 
1387 	if (!physmax4g)
1388 		return;
1389 
1390 	for (pfn = startpfn; pfn < endpfn; ) {
1391 		if (pfn <= mnoderanges[mtype].mnr_pfnhi) {
1392 			if (endpfn < mnoderanges[mtype].mnr_pfnhi) {
1393 				inc = endpfn - pfn;
1394 			} else {
1395 				inc = mnoderanges[mtype].mnr_pfnhi - pfn + 1;
1396 			}
1397 			if (mtype <= mtype4g)
1398 				maxmem4g += inc;
1399 			pfn += inc;
1400 		}
1401 		mtype++;
1402 		ASSERT(mtype < mnoderangecnt || pfn >= endpfn);
1403 	}
1404 }
1405 
1406 int
1407 mtype_2_mrange(int mtype)
1408 {
1409 	return (mnoderanges[mtype].mnr_memrange);
1410 }
1411 
1412 void
1413 mnodetype_2_pfn(int mnode, int mtype, pfn_t *pfnlo, pfn_t *pfnhi)
1414 {
1415 	ASSERT(mnoderanges[mtype].mnr_mnode == mnode);
1416 	*pfnlo = mnoderanges[mtype].mnr_pfnlo;
1417 	*pfnhi = mnoderanges[mtype].mnr_pfnhi;
1418 }
1419 
1420 size_t
1421 plcnt_sz(size_t ctrs_sz)
1422 {
1423 #ifdef DEBUG
1424 	int	szc, colors;
1425 
1426 	ctrs_sz += mnoderangecnt * sizeof (struct mnr_mts) * mmu_page_sizes;
1427 	for (szc = 0; szc < mmu_page_sizes; szc++) {
1428 		colors = page_get_pagecolors(szc);
1429 		ctrs_sz += mnoderangecnt * sizeof (pgcnt_t) * colors;
1430 	}
1431 #endif
1432 	return (ctrs_sz);
1433 }
1434 
1435 caddr_t
1436 plcnt_init(caddr_t addr)
1437 {
1438 #ifdef DEBUG
1439 	int	mt, szc, colors;
1440 
1441 	for (mt = 0; mt < mnoderangecnt; mt++) {
1442 		mnoderanges[mt].mnr_mts = (struct mnr_mts *)addr;
1443 		addr += (sizeof (struct mnr_mts) * mmu_page_sizes);
1444 		for (szc = 0; szc < mmu_page_sizes; szc++) {
1445 			colors = page_get_pagecolors(szc);
1446 			mnoderanges[mt].mnr_mts[szc].mnr_mts_colors = colors;
1447 			mnoderanges[mt].mnr_mts[szc].mnr_mtsc_pgcnt =
1448 			    (pgcnt_t *)addr;
1449 			addr += (sizeof (pgcnt_t) * colors);
1450 		}
1451 	}
1452 #endif
1453 	return (addr);
1454 }
1455 
1456 void
1457 plcnt_inc_dec(page_t *pp, int mtype, int szc, long cnt, int flags)
1458 {
1459 #ifdef DEBUG
1460 	int	bin = PP_2_BIN(pp);
1461 
1462 	atomic_add_long(&mnoderanges[mtype].mnr_mts[szc].mnr_mts_pgcnt, cnt);
1463 	atomic_add_long(&mnoderanges[mtype].mnr_mts[szc].mnr_mtsc_pgcnt[bin],
1464 	    cnt);
1465 #endif
1466 	ASSERT(mtype == PP_2_MTYPE(pp));
1467 	if (physmax4g && mtype <= mtype4g)
1468 		atomic_add_long(&freemem4g, cnt);
1469 	if (flags & PG_CACHE_LIST)
1470 		atomic_add_long(&mnoderanges[mtype].mnr_mt_clpgcnt, cnt);
1471 	else
1472 		atomic_add_long(&mnoderanges[mtype].mnr_mt_flpgcnt[szc], cnt);
1473 	atomic_add_long(&mnoderanges[mtype].mnr_mt_totcnt, cnt);
1474 }
1475 
1476 /*
1477  * Returns the free page count for mnode
1478  */
1479 int
1480 mnode_pgcnt(int mnode)
1481 {
1482 	int	mtype = mnoderangecnt - 1;
1483 	int	flags = PGI_MT_RANGE0;
1484 	pgcnt_t	pgcnt = 0;
1485 
1486 	mtype = mtype_func(mnode, mtype, flags);
1487 
1488 	while (mtype != -1) {
1489 		pgcnt += MTYPE_FREEMEM(mtype);
1490 		mtype = mtype_func(mnode, mtype, flags | PGI_MT_NEXT);
1491 	}
1492 	return (pgcnt);
1493 }
1494 
1495 /*
1496  * Initialize page coloring variables based on the l2 cache parameters.
1497  * Calculate and return memory needed for page coloring data structures.
1498  */
1499 size_t
1500 page_coloring_init(uint_t l2_sz, int l2_linesz, int l2_assoc)
1501 {
1502 	size_t	colorsz = 0;
1503 	int	i;
1504 	int	colors;
1505 
1506 #if defined(__xpv)
1507 	/*
1508 	 * Hypervisor domains currently don't have any concept of NUMA.
1509 	 * Hence we'll act like there is only 1 memrange.
1510 	 */
1511 	i = memrange_num(1);
1512 #else /* !__xpv */
1513 	/*
1514 	 * Reduce the memory ranges lists if we don't have large amounts
1515 	 * of memory. This avoids searching known empty free lists.
1516 	 */
1517 	i = memrange_num(physmax);
1518 #if defined(__i386)
1519 	if (i > 0)
1520 		restricted_kmemalloc = 0;
1521 #endif
1522 	/* physmax greater than 4g */
1523 	if (i == 0)
1524 		physmax4g = 1;
1525 #endif /* !__xpv */
1526 	memranges += i;
1527 	nranges -= i;
1528 
1529 	ASSERT(mmu_page_sizes <= MMU_PAGE_SIZES);
1530 
1531 	ASSERT(ISP2(l2_linesz));
1532 	ASSERT(l2_sz > MMU_PAGESIZE);
1533 
1534 	/* l2_assoc is 0 for fully associative l2 cache */
1535 	if (l2_assoc)
1536 		l2_colors = MAX(1, l2_sz / (l2_assoc * MMU_PAGESIZE));
1537 	else
1538 		l2_colors = 1;
1539 
1540 	ASSERT(ISP2(l2_colors));
1541 
1542 	/* for scalability, configure at least PAGE_COLORS_MIN color bins */
1543 	page_colors = MAX(l2_colors, PAGE_COLORS_MIN);
1544 
1545 	/*
1546 	 * cpu_page_colors is non-zero when a page color may be spread across
1547 	 * multiple bins.
1548 	 */
1549 	if (l2_colors < page_colors)
1550 		cpu_page_colors = l2_colors;
1551 
1552 	ASSERT(ISP2(page_colors));
1553 
1554 	page_colors_mask = page_colors - 1;
1555 
1556 	ASSERT(ISP2(CPUSETSIZE()));
1557 	page_coloring_shift = lowbit(CPUSETSIZE());
1558 
1559 	/* initialize number of colors per page size */
1560 	for (i = 0; i <= mmu.max_page_level; i++) {
1561 		hw_page_array[i].hp_size = LEVEL_SIZE(i);
1562 		hw_page_array[i].hp_shift = LEVEL_SHIFT(i);
1563 		hw_page_array[i].hp_pgcnt = LEVEL_SIZE(i) >> LEVEL_SHIFT(0);
1564 		hw_page_array[i].hp_colors = (page_colors_mask >>
1565 		    (hw_page_array[i].hp_shift - hw_page_array[0].hp_shift))
1566 		    + 1;
1567 		colorequivszc[i] = 0;
1568 	}
1569 
1570 	/*
1571 	 * The value of cpu_page_colors determines if additional color bins
1572 	 * need to be checked for a particular color in the page_get routines.
1573 	 */
1574 	if (cpu_page_colors != 0) {
1575 
1576 		int a = lowbit(page_colors) - lowbit(cpu_page_colors);
1577 		ASSERT(a > 0);
1578 		ASSERT(a < 16);
1579 
1580 		for (i = 0; i <= mmu.max_page_level; i++) {
1581 			if ((colors = hw_page_array[i].hp_colors) <= 1) {
1582 				colorequivszc[i] = 0;
1583 				continue;
1584 			}
1585 			while ((colors >> a) == 0)
1586 				a--;
1587 			ASSERT(a >= 0);
1588 
1589 			/* higher 4 bits encodes color equiv mask */
1590 			colorequivszc[i] = (a << 4);
1591 		}
1592 	}
1593 
1594 	/* factor in colorequiv to check additional 'equivalent' bins. */
1595 	if (colorequiv > 1) {
1596 
1597 		int a = lowbit(colorequiv) - 1;
1598 		if (a > 15)
1599 			a = 15;
1600 
1601 		for (i = 0; i <= mmu.max_page_level; i++) {
1602 			if ((colors = hw_page_array[i].hp_colors) <= 1) {
1603 				continue;
1604 			}
1605 			while ((colors >> a) == 0)
1606 				a--;
1607 			if ((a << 4) > colorequivszc[i]) {
1608 				colorequivszc[i] = (a << 4);
1609 			}
1610 		}
1611 	}
1612 
1613 	/* size for mnoderanges */
1614 	for (mnoderangecnt = 0, i = 0; i < max_mem_nodes; i++)
1615 		mnoderangecnt += mnode_range_cnt(i);
1616 	colorsz = mnoderangecnt * sizeof (mnoderange_t);
1617 
1618 	/* size for fpc_mutex and cpc_mutex */
1619 	colorsz += (2 * max_mem_nodes * sizeof (kmutex_t) * NPC_MUTEX);
1620 
1621 	/* size of page_freelists */
1622 	colorsz += mnoderangecnt * sizeof (page_t ***);
1623 	colorsz += mnoderangecnt * mmu_page_sizes * sizeof (page_t **);
1624 
1625 	for (i = 0; i < mmu_page_sizes; i++) {
1626 		colors = page_get_pagecolors(i);
1627 		colorsz += mnoderangecnt * colors * sizeof (page_t *);
1628 	}
1629 
1630 	/* size of page_cachelists */
1631 	colorsz += mnoderangecnt * sizeof (page_t **);
1632 	colorsz += mnoderangecnt * page_colors * sizeof (page_t *);
1633 
1634 	return (colorsz);
1635 }
1636 
1637 /*
1638  * Called once at startup to configure page_coloring data structures and
1639  * does the 1st page_free()/page_freelist_add().
1640  */
1641 void
1642 page_coloring_setup(caddr_t pcmemaddr)
1643 {
1644 	int	i;
1645 	int	j;
1646 	int	k;
1647 	caddr_t	addr;
1648 	int	colors;
1649 
1650 	/*
1651 	 * do page coloring setup
1652 	 */
1653 	addr = pcmemaddr;
1654 
1655 	mnoderanges = (mnoderange_t *)addr;
1656 	addr += (mnoderangecnt * sizeof (mnoderange_t));
1657 
1658 	mnode_range_setup(mnoderanges);
1659 
1660 	if (physmax4g)
1661 		mtype4g = pfn_2_mtype(0xfffff);
1662 
1663 	for (k = 0; k < NPC_MUTEX; k++) {
1664 		fpc_mutex[k] = (kmutex_t *)addr;
1665 		addr += (max_mem_nodes * sizeof (kmutex_t));
1666 	}
1667 	for (k = 0; k < NPC_MUTEX; k++) {
1668 		cpc_mutex[k] = (kmutex_t *)addr;
1669 		addr += (max_mem_nodes * sizeof (kmutex_t));
1670 	}
1671 	page_freelists = (page_t ****)addr;
1672 	addr += (mnoderangecnt * sizeof (page_t ***));
1673 
1674 	page_cachelists = (page_t ***)addr;
1675 	addr += (mnoderangecnt * sizeof (page_t **));
1676 
1677 	for (i = 0; i < mnoderangecnt; i++) {
1678 		page_freelists[i] = (page_t ***)addr;
1679 		addr += (mmu_page_sizes * sizeof (page_t **));
1680 
1681 		for (j = 0; j < mmu_page_sizes; j++) {
1682 			colors = page_get_pagecolors(j);
1683 			page_freelists[i][j] = (page_t **)addr;
1684 			addr += (colors * sizeof (page_t *));
1685 		}
1686 		page_cachelists[i] = (page_t **)addr;
1687 		addr += (page_colors * sizeof (page_t *));
1688 	}
1689 }
1690 
1691 #if defined(__xpv)
1692 /*
1693  * Give back 10% of the io_pool pages to the free list.
1694  * Don't shrink the pool below some absolute minimum.
1695  */
1696 static void
1697 page_io_pool_shrink()
1698 {
1699 	int retcnt;
1700 	page_t *pp, *pp_first, *pp_last, **curpool;
1701 	mfn_t mfn;
1702 	int bothpools = 0;
1703 
1704 	mutex_enter(&io_pool_lock);
1705 	io_pool_shrink_attempts++;	/* should be a kstat? */
1706 	retcnt = io_pool_cnt / 10;
1707 	if (io_pool_cnt - retcnt < io_pool_cnt_min)
1708 		retcnt = io_pool_cnt - io_pool_cnt_min;
1709 	if (retcnt <= 0)
1710 		goto done;
1711 	io_pool_shrinks++;	/* should be a kstat? */
1712 	curpool = &io_pool_4g;
1713 domore:
1714 	/*
1715 	 * Loop through taking pages from the end of the list
1716 	 * (highest mfns) till amount to return reached.
1717 	 */
1718 	for (pp = *curpool; pp && retcnt > 0; ) {
1719 		pp_first = pp_last = pp->p_prev;
1720 		if (pp_first == *curpool)
1721 			break;
1722 		retcnt--;
1723 		io_pool_cnt--;
1724 		page_io_pool_sub(curpool, pp_first, pp_last);
1725 		if ((mfn = pfn_to_mfn(pp->p_pagenum)) < start_mfn)
1726 			start_mfn = mfn;
1727 		page_free(pp_first, 1);
1728 		pp = *curpool;
1729 	}
1730 	if (retcnt != 0 && !bothpools) {
1731 		/*
1732 		 * If not enough found in less constrained pool try the
1733 		 * more constrained one.
1734 		 */
1735 		curpool = &io_pool_16m;
1736 		bothpools = 1;
1737 		goto domore;
1738 	}
1739 done:
1740 	mutex_exit(&io_pool_lock);
1741 }
1742 
1743 #endif	/* __xpv */
1744 
1745 uint_t
1746 page_create_update_flags_x86(uint_t flags)
1747 {
1748 #if defined(__xpv)
1749 	/*
1750 	 * Check this is an urgent allocation and free pages are depleted.
1751 	 */
1752 	if (!(flags & PG_WAIT) && freemem < desfree)
1753 		page_io_pool_shrink();
1754 #else /* !__xpv */
1755 	/*
1756 	 * page_create_get_something may call this because 4g memory may be
1757 	 * depleted. Set flags to allow for relocation of base page below
1758 	 * 4g if necessary.
1759 	 */
1760 	if (physmax4g)
1761 		flags |= (PGI_PGCPSZC0 | PGI_PGCPHIPRI);
1762 #endif /* __xpv */
1763 	return (flags);
1764 }
1765 
1766 /*ARGSUSED*/
1767 int
1768 bp_color(struct buf *bp)
1769 {
1770 	return (0);
1771 }
1772 
1773 #if defined(__xpv)
1774 
1775 /*
1776  * Take pages out of an io_pool
1777  */
1778 static void
1779 page_io_pool_sub(page_t **poolp, page_t *pp_first, page_t *pp_last)
1780 {
1781 	if (*poolp == pp_first) {
1782 		*poolp = pp_last->p_next;
1783 		if (*poolp == pp_first)
1784 			*poolp = NULL;
1785 	}
1786 	pp_first->p_prev->p_next = pp_last->p_next;
1787 	pp_last->p_next->p_prev = pp_first->p_prev;
1788 	pp_first->p_prev = pp_last;
1789 	pp_last->p_next = pp_first;
1790 }
1791 
1792 /*
1793  * Put a page on the io_pool list. The list is ordered by increasing MFN.
1794  */
1795 static void
1796 page_io_pool_add(page_t **poolp, page_t *pp)
1797 {
1798 	page_t	*look;
1799 	mfn_t	mfn = mfn_list[pp->p_pagenum];
1800 
1801 	if (*poolp == NULL) {
1802 		*poolp = pp;
1803 		pp->p_next = pp;
1804 		pp->p_prev = pp;
1805 		return;
1806 	}
1807 
1808 	/*
1809 	 * Since we try to take pages from the high end of the pool
1810 	 * chances are good that the pages to be put on the list will
1811 	 * go at or near the end of the list. so start at the end and
1812 	 * work backwards.
1813 	 */
1814 	look = (*poolp)->p_prev;
1815 	while (mfn < mfn_list[look->p_pagenum]) {
1816 		look = look->p_prev;
1817 		if (look == (*poolp)->p_prev)
1818 			break; /* backed all the way to front of list */
1819 	}
1820 
1821 	/* insert after look */
1822 	pp->p_prev = look;
1823 	pp->p_next = look->p_next;
1824 	pp->p_next->p_prev = pp;
1825 	look->p_next = pp;
1826 	if (mfn < mfn_list[(*poolp)->p_pagenum]) {
1827 		/*
1828 		 * we inserted a new first list element
1829 		 * adjust pool pointer to newly inserted element
1830 		 */
1831 		*poolp = pp;
1832 	}
1833 }
1834 
1835 /*
1836  * Add a page to the io_pool.  Setting the force flag will force the page
1837  * into the io_pool no matter what.
1838  */
1839 static void
1840 add_page_to_pool(page_t *pp, int force)
1841 {
1842 	page_t *highest;
1843 	page_t *freep = NULL;
1844 
1845 	mutex_enter(&io_pool_lock);
1846 	/*
1847 	 * Always keep the scarce low memory pages
1848 	 */
1849 	if (mfn_list[pp->p_pagenum] < PFN_16MEG) {
1850 		++io_pool_cnt;
1851 		page_io_pool_add(&io_pool_16m, pp);
1852 		goto done;
1853 	}
1854 	if (io_pool_cnt < io_pool_cnt_max || force || io_pool_4g == NULL) {
1855 		++io_pool_cnt;
1856 		page_io_pool_add(&io_pool_4g, pp);
1857 	} else {
1858 		highest = io_pool_4g->p_prev;
1859 		if (mfn_list[pp->p_pagenum] < mfn_list[highest->p_pagenum]) {
1860 			page_io_pool_sub(&io_pool_4g, highest, highest);
1861 			page_io_pool_add(&io_pool_4g, pp);
1862 			freep = highest;
1863 		} else {
1864 			freep = pp;
1865 		}
1866 	}
1867 done:
1868 	mutex_exit(&io_pool_lock);
1869 	if (freep)
1870 		page_free(freep, 1);
1871 }
1872 
1873 
1874 int contig_pfn_cnt;	/* no of pfns in the contig pfn list */
1875 int contig_pfn_max;	/* capacity of the contig pfn list */
1876 int next_alloc_pfn;	/* next position in list to start a contig search */
1877 int contig_pfnlist_updates;	/* pfn list update count */
1878 int contig_pfnlist_builds;	/* how many times have we (re)built list */
1879 int contig_pfnlist_buildfailed;	/* how many times has list build failed */
1880 int create_contig_pending;	/* nonzero means taskq creating contig list */
1881 pfn_t *contig_pfn_list = NULL;	/* list of contig pfns in ascending mfn order */
1882 
1883 /*
1884  * Function to use in sorting a list of pfns by their underlying mfns.
1885  */
1886 static int
1887 mfn_compare(const void *pfnp1, const void *pfnp2)
1888 {
1889 	mfn_t mfn1 = mfn_list[*(pfn_t *)pfnp1];
1890 	mfn_t mfn2 = mfn_list[*(pfn_t *)pfnp2];
1891 
1892 	if (mfn1 > mfn2)
1893 		return (1);
1894 	if (mfn1 < mfn2)
1895 		return (-1);
1896 	return (0);
1897 }
1898 
1899 /*
1900  * Compact the contig_pfn_list by tossing all the non-contiguous
1901  * elements from the list.
1902  */
1903 static void
1904 compact_contig_pfn_list(void)
1905 {
1906 	pfn_t pfn, lapfn, prev_lapfn;
1907 	mfn_t mfn;
1908 	int i, newcnt = 0;
1909 
1910 	prev_lapfn = 0;
1911 	for (i = 0; i < contig_pfn_cnt - 1; i++) {
1912 		pfn = contig_pfn_list[i];
1913 		lapfn = contig_pfn_list[i + 1];
1914 		mfn = mfn_list[pfn];
1915 		/*
1916 		 * See if next pfn is for a contig mfn
1917 		 */
1918 		if (mfn_list[lapfn] != mfn + 1)
1919 			continue;
1920 		/*
1921 		 * pfn and lookahead are both put in list
1922 		 * unless pfn is the previous lookahead.
1923 		 */
1924 		if (pfn != prev_lapfn)
1925 			contig_pfn_list[newcnt++] = pfn;
1926 		contig_pfn_list[newcnt++] = lapfn;
1927 		prev_lapfn = lapfn;
1928 	}
1929 	for (i = newcnt; i < contig_pfn_cnt; i++)
1930 		contig_pfn_list[i] = 0;
1931 	contig_pfn_cnt = newcnt;
1932 }
1933 
1934 /*ARGSUSED*/
1935 static void
1936 call_create_contiglist(void *arg)
1937 {
1938 	(void) create_contig_pfnlist(PG_WAIT);
1939 }
1940 
1941 /*
1942  * Create list of freelist pfns that have underlying
1943  * contiguous mfns.  The list is kept in ascending mfn order.
1944  * returns 1 if list created else 0.
1945  */
1946 static int
1947 create_contig_pfnlist(uint_t flags)
1948 {
1949 	pfn_t pfn;
1950 	page_t *pp;
1951 	int ret = 1;
1952 
1953 	mutex_enter(&contig_list_lock);
1954 	if (contig_pfn_list != NULL)
1955 		goto out;
1956 	contig_pfn_max = freemem + (freemem / 10);
1957 	contig_pfn_list = kmem_zalloc(contig_pfn_max * sizeof (pfn_t),
1958 	    (flags & PG_WAIT) ? KM_SLEEP : KM_NOSLEEP);
1959 	if (contig_pfn_list == NULL) {
1960 		/*
1961 		 * If we could not create the contig list (because
1962 		 * we could not sleep for memory).  Dispatch a taskq that can
1963 		 * sleep to get the memory.
1964 		 */
1965 		if (!create_contig_pending) {
1966 			if (taskq_dispatch(system_taskq, call_create_contiglist,
1967 			    NULL, TQ_NOSLEEP) != NULL)
1968 				create_contig_pending = 1;
1969 		}
1970 		contig_pfnlist_buildfailed++;	/* count list build failures */
1971 		ret = 0;
1972 		goto out;
1973 	}
1974 	create_contig_pending = 0;
1975 	ASSERT(contig_pfn_cnt == 0);
1976 	for (pfn = 0; pfn < mfn_count; pfn++) {
1977 		pp = page_numtopp_nolock(pfn);
1978 		if (pp == NULL || !PP_ISFREE(pp))
1979 			continue;
1980 		contig_pfn_list[contig_pfn_cnt] = pfn;
1981 		if (++contig_pfn_cnt == contig_pfn_max)
1982 			break;
1983 	}
1984 	/*
1985 	 * Sanity check the new list.
1986 	 */
1987 	if (contig_pfn_cnt < 2) { /* no contig pfns */
1988 		contig_pfn_cnt = 0;
1989 		contig_pfnlist_buildfailed++;
1990 		kmem_free(contig_pfn_list, contig_pfn_max * sizeof (pfn_t));
1991 		contig_pfn_list = NULL;
1992 		contig_pfn_max = 0;
1993 		ret = 0;
1994 		goto out;
1995 	}
1996 	qsort(contig_pfn_list, contig_pfn_cnt, sizeof (pfn_t), mfn_compare);
1997 	compact_contig_pfn_list();
1998 	/*
1999 	 * Make sure next search of the newly created contiguous pfn
2000 	 * list starts at the beginning of the list.
2001 	 */
2002 	next_alloc_pfn = 0;
2003 	contig_pfnlist_builds++;	/* count list builds */
2004 out:
2005 	mutex_exit(&contig_list_lock);
2006 	return (ret);
2007 }
2008 
2009 
2010 /*
2011  * Toss the current contig pfnlist.  Someone is about to do a massive
2012  * update to pfn<->mfn mappings.  So we have them destroy the list and lock
2013  * it till they are done with their update.
2014  */
2015 void
2016 clear_and_lock_contig_pfnlist()
2017 {
2018 	pfn_t *listp = NULL;
2019 	size_t listsize;
2020 
2021 	mutex_enter(&contig_list_lock);
2022 	if (contig_pfn_list != NULL) {
2023 		listp = contig_pfn_list;
2024 		listsize = contig_pfn_max * sizeof (pfn_t);
2025 		contig_pfn_list = NULL;
2026 		contig_pfn_max = contig_pfn_cnt = 0;
2027 	}
2028 	if (listp != NULL)
2029 		kmem_free(listp, listsize);
2030 }
2031 
2032 /*
2033  * Unlock the contig_pfn_list.  The next attempted use of it will cause
2034  * it to be re-created.
2035  */
2036 void
2037 unlock_contig_pfnlist()
2038 {
2039 	mutex_exit(&contig_list_lock);
2040 }
2041 
2042 /*
2043  * Update the contiguous pfn list in response to a pfn <-> mfn reassignment
2044  */
2045 void
2046 update_contig_pfnlist(pfn_t pfn, mfn_t oldmfn, mfn_t newmfn)
2047 {
2048 	int probe_hi, probe_lo, probe_pos, insert_after, insert_point;
2049 	pfn_t probe_pfn;
2050 	mfn_t probe_mfn;
2051 	int drop_lock = 0;
2052 
2053 	if (mutex_owner(&contig_list_lock) != curthread) {
2054 		drop_lock = 1;
2055 		mutex_enter(&contig_list_lock);
2056 	}
2057 	if (contig_pfn_list == NULL)
2058 		goto done;
2059 	contig_pfnlist_updates++;
2060 	/*
2061 	 * Find the pfn in the current list.  Use a binary chop to locate it.
2062 	 */
2063 	probe_hi = contig_pfn_cnt - 1;
2064 	probe_lo = 0;
2065 	probe_pos = (probe_hi + probe_lo) / 2;
2066 	while ((probe_pfn = contig_pfn_list[probe_pos]) != pfn) {
2067 		if (probe_pos == probe_lo) { /* pfn not in list */
2068 			probe_pos = -1;
2069 			break;
2070 		}
2071 		if (pfn_to_mfn(probe_pfn) <= oldmfn)
2072 			probe_lo = probe_pos;
2073 		else
2074 			probe_hi = probe_pos;
2075 		probe_pos = (probe_hi + probe_lo) / 2;
2076 	}
2077 	if (probe_pos >= 0) {
2078 		/*
2079 		 * Remove pfn from list and ensure next alloc
2080 		 * position stays in bounds.
2081 		 */
2082 		if (--contig_pfn_cnt <= next_alloc_pfn)
2083 			next_alloc_pfn = 0;
2084 		if (contig_pfn_cnt < 2) { /* no contig pfns */
2085 			contig_pfn_cnt = 0;
2086 			kmem_free(contig_pfn_list,
2087 			    contig_pfn_max * sizeof (pfn_t));
2088 			contig_pfn_list = NULL;
2089 			contig_pfn_max = 0;
2090 			goto done;
2091 		}
2092 		ovbcopy(&contig_pfn_list[probe_pos + 1],
2093 		    &contig_pfn_list[probe_pos],
2094 		    (contig_pfn_cnt - probe_pos) * sizeof (pfn_t));
2095 	}
2096 	if (newmfn == MFN_INVALID)
2097 		goto done;
2098 	/*
2099 	 * Check if new mfn has adjacent mfns in the list
2100 	 */
2101 	probe_hi = contig_pfn_cnt - 1;
2102 	probe_lo = 0;
2103 	insert_after = -2;
2104 	do {
2105 		probe_pos = (probe_hi + probe_lo) / 2;
2106 		probe_mfn = pfn_to_mfn(contig_pfn_list[probe_pos]);
2107 		if (newmfn == probe_mfn + 1)
2108 			insert_after = probe_pos;
2109 		else if (newmfn == probe_mfn - 1)
2110 			insert_after = probe_pos - 1;
2111 		if (probe_pos == probe_lo)
2112 			break;
2113 		if (probe_mfn <= newmfn)
2114 			probe_lo = probe_pos;
2115 		else
2116 			probe_hi = probe_pos;
2117 	} while (insert_after == -2);
2118 	/*
2119 	 * If there is space in the list and there are adjacent mfns
2120 	 * insert the pfn in to its proper place in the list.
2121 	 */
2122 	if (insert_after != -2 && contig_pfn_cnt + 1 <= contig_pfn_max) {
2123 		insert_point = insert_after + 1;
2124 		ovbcopy(&contig_pfn_list[insert_point],
2125 		    &contig_pfn_list[insert_point + 1],
2126 		    (contig_pfn_cnt - insert_point) * sizeof (pfn_t));
2127 		contig_pfn_list[insert_point] = pfn;
2128 		contig_pfn_cnt++;
2129 	}
2130 done:
2131 	if (drop_lock)
2132 		mutex_exit(&contig_list_lock);
2133 }
2134 
2135 /*
2136  * Called to (re-)populate the io_pool from the free page lists.
2137  */
2138 long
2139 populate_io_pool(void)
2140 {
2141 	pfn_t pfn;
2142 	mfn_t mfn, max_mfn;
2143 	page_t *pp;
2144 
2145 	/*
2146 	 * Figure out the bounds of the pool on first invocation.
2147 	 * We use a percentage of memory for the io pool size.
2148 	 * we allow that to shrink, but not to less than a fixed minimum
2149 	 */
2150 	if (io_pool_cnt_max == 0) {
2151 		io_pool_cnt_max = physmem / (100 / io_pool_physmem_pct);
2152 		io_pool_cnt_lowater = io_pool_cnt_max;
2153 		/*
2154 		 * This is the first time in populate_io_pool, grab a va to use
2155 		 * when we need to allocate pages.
2156 		 */
2157 		io_pool_kva = vmem_alloc(heap_arena, PAGESIZE, VM_SLEEP);
2158 	}
2159 	/*
2160 	 * If we are out of pages in the pool, then grow the size of the pool
2161 	 */
2162 	if (io_pool_cnt == 0) {
2163 		/*
2164 		 * Grow the max size of the io pool by 5%, but never more than
2165 		 * 25% of physical memory.
2166 		 */
2167 		if (io_pool_cnt_max < physmem / 4)
2168 			io_pool_cnt_max += io_pool_cnt_max / 20;
2169 	}
2170 	io_pool_grows++;	/* should be a kstat? */
2171 
2172 	/*
2173 	 * Get highest mfn on this platform, but limit to the 32 bit DMA max.
2174 	 */
2175 	(void) mfn_to_pfn(start_mfn);
2176 	max_mfn = MIN(cached_max_mfn, PFN_4GIG);
2177 	for (mfn = start_mfn; mfn < max_mfn; start_mfn = ++mfn) {
2178 		pfn = mfn_to_pfn(mfn);
2179 		if (pfn & PFN_IS_FOREIGN_MFN)
2180 			continue;
2181 		/*
2182 		 * try to allocate it from free pages
2183 		 */
2184 		pp = page_numtopp_alloc(pfn);
2185 		if (pp == NULL)
2186 			continue;
2187 		PP_CLRFREE(pp);
2188 		add_page_to_pool(pp, 1);
2189 		if (io_pool_cnt >= io_pool_cnt_max)
2190 			break;
2191 	}
2192 
2193 	return (io_pool_cnt);
2194 }
2195 
2196 /*
2197  * Destroy a page that was being used for DMA I/O. It may or
2198  * may not actually go back to the io_pool.
2199  */
2200 void
2201 page_destroy_io(page_t *pp)
2202 {
2203 	mfn_t mfn = mfn_list[pp->p_pagenum];
2204 
2205 	/*
2206 	 * When the page was alloc'd a reservation was made, release it now
2207 	 */
2208 	page_unresv(1);
2209 	/*
2210 	 * Unload translations, if any, then hash out the
2211 	 * page to erase its identity.
2212 	 */
2213 	(void) hat_pageunload(pp, HAT_FORCE_PGUNLOAD);
2214 	page_hashout(pp, NULL);
2215 
2216 	/*
2217 	 * If the page came from the free lists, just put it back to them.
2218 	 * DomU pages always go on the free lists as well.
2219 	 */
2220 	if (!DOMAIN_IS_INITDOMAIN(xen_info) || mfn >= PFN_4GIG) {
2221 		page_free(pp, 1);
2222 		return;
2223 	}
2224 
2225 	add_page_to_pool(pp, 0);
2226 }
2227 
2228 
2229 long contig_searches;		/* count of times contig pages requested */
2230 long contig_search_restarts;	/* count of contig ranges tried */
2231 long contig_search_failed;	/* count of contig alloc failures */
2232 
2233 /*
2234  * Free partial page list
2235  */
2236 static void
2237 free_partial_list(page_t **pplist)
2238 {
2239 	page_t *pp;
2240 
2241 	while (*pplist != NULL) {
2242 		pp = *pplist;
2243 		page_io_pool_sub(pplist, pp, pp);
2244 		page_free(pp, 1);
2245 	}
2246 }
2247 
2248 /*
2249  * Look thru the contiguous pfns that are not part of the io_pool for
2250  * contiguous free pages.  Return a list of the found pages or NULL.
2251  */
2252 page_t *
2253 find_contig_free(uint_t npages, uint_t flags, uint64_t pfnseg,
2254     pgcnt_t pfnalign)
2255 {
2256 	page_t *pp, *plist = NULL;
2257 	mfn_t mfn, prev_mfn, start_mfn;
2258 	pfn_t pfn;
2259 	int pages_needed, pages_requested;
2260 	int search_start;
2261 
2262 	/*
2263 	 * create the contig pfn list if not already done
2264 	 */
2265 retry:
2266 	mutex_enter(&contig_list_lock);
2267 	if (contig_pfn_list == NULL) {
2268 		mutex_exit(&contig_list_lock);
2269 		if (!create_contig_pfnlist(flags)) {
2270 			return (NULL);
2271 		}
2272 		goto retry;
2273 	}
2274 	contig_searches++;
2275 	/*
2276 	 * Search contiguous pfn list for physically contiguous pages not in
2277 	 * the io_pool.  Start the search where the last search left off.
2278 	 */
2279 	pages_requested = pages_needed = npages;
2280 	search_start = next_alloc_pfn;
2281 	start_mfn = prev_mfn = 0;
2282 	while (pages_needed) {
2283 		pfn = contig_pfn_list[next_alloc_pfn];
2284 		mfn = pfn_to_mfn(pfn);
2285 		/*
2286 		 * Check if mfn is first one or contig to previous one and
2287 		 * if page corresponding to mfn is free and that mfn
2288 		 * range is not crossing a segment boundary.
2289 		 */
2290 		if ((prev_mfn == 0 || mfn == prev_mfn + 1) &&
2291 		    (pp = page_numtopp_alloc(pfn)) != NULL &&
2292 		    !((mfn & pfnseg) < (start_mfn & pfnseg))) {
2293 			PP_CLRFREE(pp);
2294 			page_io_pool_add(&plist, pp);
2295 			pages_needed--;
2296 			if (prev_mfn == 0) {
2297 				if (pfnalign &&
2298 				    mfn != P2ROUNDUP(mfn, pfnalign)) {
2299 					/*
2300 					 * not properly aligned
2301 					 */
2302 					contig_search_restarts++;
2303 					free_partial_list(&plist);
2304 					pages_needed = pages_requested;
2305 					start_mfn = prev_mfn = 0;
2306 					goto skip;
2307 				}
2308 				start_mfn = mfn;
2309 			}
2310 			prev_mfn = mfn;
2311 		} else {
2312 			contig_search_restarts++;
2313 			free_partial_list(&plist);
2314 			pages_needed = pages_requested;
2315 			start_mfn = prev_mfn = 0;
2316 		}
2317 skip:
2318 		if (++next_alloc_pfn == contig_pfn_cnt)
2319 			next_alloc_pfn = 0;
2320 		if (next_alloc_pfn == search_start)
2321 			break; /* all pfns searched */
2322 	}
2323 	mutex_exit(&contig_list_lock);
2324 	if (pages_needed) {
2325 		contig_search_failed++;
2326 		/*
2327 		 * Failed to find enough contig pages.
2328 		 * free partial page list
2329 		 */
2330 		free_partial_list(&plist);
2331 	}
2332 	return (plist);
2333 }
2334 
2335 /*
2336  * Search the reserved io pool pages for a page range with the
2337  * desired characteristics.
2338  */
2339 page_t *
2340 page_io_pool_alloc(ddi_dma_attr_t *mattr, int contig, pgcnt_t minctg)
2341 {
2342 	page_t *pp_first, *pp_last;
2343 	page_t *pp, **poolp;
2344 	pgcnt_t nwanted, pfnalign;
2345 	uint64_t pfnseg;
2346 	mfn_t mfn, tmfn, hi_mfn, lo_mfn;
2347 	int align, attempt = 0;
2348 
2349 	if (minctg == 1)
2350 		contig = 0;
2351 	lo_mfn = mmu_btop(mattr->dma_attr_addr_lo);
2352 	hi_mfn = mmu_btop(mattr->dma_attr_addr_hi);
2353 	pfnseg = mmu_btop(mattr->dma_attr_seg);
2354 	align = maxbit(mattr->dma_attr_align, mattr->dma_attr_minxfer);
2355 	if (align > MMU_PAGESIZE)
2356 		pfnalign = mmu_btop(align);
2357 	else
2358 		pfnalign = 0;
2359 
2360 try_again:
2361 	/*
2362 	 * See if we want pages for a legacy device
2363 	 */
2364 	if (hi_mfn < PFN_16MEG)
2365 		poolp = &io_pool_16m;
2366 	else
2367 		poolp = &io_pool_4g;
2368 try_smaller:
2369 	/*
2370 	 * Take pages from I/O pool. We'll use pages from the highest
2371 	 * MFN range possible.
2372 	 */
2373 	pp_first = pp_last = NULL;
2374 	mutex_enter(&io_pool_lock);
2375 	nwanted = minctg;
2376 	for (pp = *poolp; pp && nwanted > 0; ) {
2377 		pp = pp->p_prev;
2378 
2379 		/*
2380 		 * skip pages above allowable range
2381 		 */
2382 		mfn = mfn_list[pp->p_pagenum];
2383 		if (hi_mfn < mfn)
2384 			goto skip;
2385 
2386 		/*
2387 		 * stop at pages below allowable range
2388 		 */
2389 		if (lo_mfn > mfn)
2390 			break;
2391 restart:
2392 		if (pp_last == NULL) {
2393 			/*
2394 			 * Check alignment
2395 			 */
2396 			tmfn = mfn - (minctg - 1);
2397 			if (pfnalign && tmfn != P2ROUNDUP(tmfn, pfnalign))
2398 				goto skip; /* not properly aligned */
2399 			/*
2400 			 * Check segment
2401 			 */
2402 			if ((mfn & pfnseg) < (tmfn & pfnseg))
2403 				goto skip; /* crosses seg boundary */
2404 			/*
2405 			 * Start building page list
2406 			 */
2407 			pp_first = pp_last = pp;
2408 			nwanted--;
2409 		} else {
2410 			/*
2411 			 * check physical contiguity if required
2412 			 */
2413 			if (contig &&
2414 			    mfn_list[pp_first->p_pagenum] != mfn + 1) {
2415 				/*
2416 				 * not a contiguous page, restart list.
2417 				 */
2418 				pp_last = NULL;
2419 				nwanted = minctg;
2420 				goto restart;
2421 			} else { /* add page to list */
2422 				pp_first = pp;
2423 				nwanted--;
2424 			}
2425 		}
2426 skip:
2427 		if (pp == *poolp)
2428 			break;
2429 	}
2430 
2431 	/*
2432 	 * If we didn't find memory. Try the more constrained pool, then
2433 	 * sweep free pages into the DMA pool and try again.
2434 	 */
2435 	if (nwanted != 0) {
2436 		mutex_exit(&io_pool_lock);
2437 		/*
2438 		 * If we were looking in the less constrained pool and
2439 		 * didn't find pages, try the more constrained pool.
2440 		 */
2441 		if (poolp == &io_pool_4g) {
2442 			poolp = &io_pool_16m;
2443 			goto try_smaller;
2444 		}
2445 		kmem_reap();
2446 		if (++attempt < 4) {
2447 			/*
2448 			 * Grab some more io_pool pages
2449 			 */
2450 			(void) populate_io_pool();
2451 			goto try_again; /* go around and retry */
2452 		}
2453 		return (NULL);
2454 	}
2455 	/*
2456 	 * Found the pages, now snip them from the list
2457 	 */
2458 	page_io_pool_sub(poolp, pp_first, pp_last);
2459 	io_pool_cnt -= minctg;
2460 	/*
2461 	 * reset low water mark
2462 	 */
2463 	if (io_pool_cnt < io_pool_cnt_lowater)
2464 		io_pool_cnt_lowater = io_pool_cnt;
2465 	mutex_exit(&io_pool_lock);
2466 	return (pp_first);
2467 }
2468 
2469 page_t *
2470 page_swap_with_hypervisor(struct vnode *vp, u_offset_t off, caddr_t vaddr,
2471     ddi_dma_attr_t *mattr, uint_t flags, pgcnt_t minctg)
2472 {
2473 	uint_t kflags;
2474 	int order, extra, extpages, i, contig, nbits, extents;
2475 	page_t *pp, *expp, *pp_first, **pplist = NULL;
2476 	mfn_t *mfnlist = NULL;
2477 
2478 	contig = flags & PG_PHYSCONTIG;
2479 	if (minctg == 1)
2480 		contig = 0;
2481 	flags &= ~PG_PHYSCONTIG;
2482 	kflags = flags & PG_WAIT ? KM_SLEEP : KM_NOSLEEP;
2483 	/*
2484 	 * Hypervisor will allocate extents, if we want contig
2485 	 * pages extent must be >= minctg
2486 	 */
2487 	if (contig) {
2488 		order = highbit(minctg) - 1;
2489 		if (minctg & ((1 << order) - 1))
2490 			order++;
2491 		extpages = 1 << order;
2492 	} else {
2493 		order = 0;
2494 		extpages = minctg;
2495 	}
2496 	if (extpages > minctg) {
2497 		extra = extpages - minctg;
2498 		if (!page_resv(extra, kflags))
2499 			return (NULL);
2500 	}
2501 	pp_first = NULL;
2502 	pplist = kmem_alloc(extpages * sizeof (page_t *), kflags);
2503 	if (pplist == NULL)
2504 		goto balloon_fail;
2505 	mfnlist = kmem_alloc(extpages * sizeof (mfn_t), kflags);
2506 	if (mfnlist == NULL)
2507 		goto balloon_fail;
2508 	pp = page_create_va(vp, off, minctg * PAGESIZE, flags, &kvseg, vaddr);
2509 	if (pp == NULL)
2510 		goto balloon_fail;
2511 	pp_first = pp;
2512 	if (extpages > minctg) {
2513 		/*
2514 		 * fill out the rest of extent pages to swap
2515 		 * with the hypervisor
2516 		 */
2517 		for (i = 0; i < extra; i++) {
2518 			expp = page_create_va(vp,
2519 			    (u_offset_t)(uintptr_t)io_pool_kva,
2520 			    PAGESIZE, flags, &kvseg, io_pool_kva);
2521 			if (expp == NULL)
2522 				goto balloon_fail;
2523 			(void) hat_pageunload(expp, HAT_FORCE_PGUNLOAD);
2524 			page_io_unlock(expp);
2525 			page_hashout(expp, NULL);
2526 			page_io_lock(expp);
2527 			/*
2528 			 * add page to end of list
2529 			 */
2530 			expp->p_prev = pp_first->p_prev;
2531 			expp->p_next = pp_first;
2532 			expp->p_prev->p_next = expp;
2533 			pp_first->p_prev = expp;
2534 		}
2535 
2536 	}
2537 	for (i = 0; i < extpages; i++) {
2538 		pplist[i] = pp;
2539 		pp = pp->p_next;
2540 	}
2541 	nbits = highbit(mattr->dma_attr_addr_hi);
2542 	extents = contig ? 1 : minctg;
2543 	if (balloon_replace_pages(extents, pplist, nbits, order,
2544 	    mfnlist) != extents) {
2545 		if (ioalloc_dbg)
2546 			cmn_err(CE_NOTE, "request to hypervisor"
2547 			    " for %d pages, maxaddr %" PRIx64 " failed",
2548 			    extpages, mattr->dma_attr_addr_hi);
2549 		goto balloon_fail;
2550 	}
2551 
2552 	kmem_free(pplist, extpages * sizeof (page_t *));
2553 	kmem_free(mfnlist, extpages * sizeof (mfn_t));
2554 	/*
2555 	 * Return any excess pages to free list
2556 	 */
2557 	if (extpages > minctg) {
2558 		for (i = 0; i < extra; i++) {
2559 			pp = pp_first->p_prev;
2560 			page_sub(&pp_first, pp);
2561 			page_io_unlock(pp);
2562 			page_unresv(1);
2563 			page_free(pp, 1);
2564 		}
2565 	}
2566 	return (pp_first);
2567 balloon_fail:
2568 	/*
2569 	 * Return pages to free list and return failure
2570 	 */
2571 	while (pp_first != NULL) {
2572 		pp = pp_first;
2573 		page_sub(&pp_first, pp);
2574 		page_io_unlock(pp);
2575 		if (pp->p_vnode != NULL)
2576 			page_hashout(pp, NULL);
2577 		page_free(pp, 1);
2578 	}
2579 	if (pplist)
2580 		kmem_free(pplist, extpages * sizeof (page_t *));
2581 	if (mfnlist)
2582 		kmem_free(mfnlist, extpages * sizeof (mfn_t));
2583 	page_unresv(extpages - minctg);
2584 	return (NULL);
2585 }
2586 
2587 static void
2588 return_partial_alloc(page_t *plist)
2589 {
2590 	page_t *pp;
2591 
2592 	while (plist != NULL) {
2593 		pp = plist;
2594 		page_sub(&plist, pp);
2595 		page_io_unlock(pp);
2596 		page_destroy_io(pp);
2597 	}
2598 }
2599 
2600 static page_t *
2601 page_get_contigpages(
2602 	struct vnode	*vp,
2603 	u_offset_t	off,
2604 	int		*npagesp,
2605 	uint_t		flags,
2606 	caddr_t		vaddr,
2607 	ddi_dma_attr_t	*mattr)
2608 {
2609 	mfn_t	max_mfn = HYPERVISOR_memory_op(XENMEM_maximum_ram_page, NULL);
2610 	page_t	*plist;	/* list to return */
2611 	page_t	*pp, *mcpl;
2612 	int	contig, anyaddr, npages, getone = 0;
2613 	mfn_t	lo_mfn;
2614 	mfn_t	hi_mfn;
2615 	pgcnt_t	pfnalign = 0;
2616 	int	align, sgllen;
2617 	uint64_t pfnseg;
2618 	pgcnt_t	minctg;
2619 
2620 	npages = *npagesp;
2621 	ASSERT(mattr != NULL);
2622 	lo_mfn = mmu_btop(mattr->dma_attr_addr_lo);
2623 	hi_mfn = mmu_btop(mattr->dma_attr_addr_hi);
2624 	sgllen = mattr->dma_attr_sgllen;
2625 	pfnseg = mmu_btop(mattr->dma_attr_seg);
2626 	align = maxbit(mattr->dma_attr_align, mattr->dma_attr_minxfer);
2627 	if (align > MMU_PAGESIZE)
2628 		pfnalign = mmu_btop(align);
2629 
2630 	contig = flags & PG_PHYSCONTIG;
2631 	if (npages == -1) {
2632 		npages = 1;
2633 		pfnalign = 0;
2634 	}
2635 	/*
2636 	 * Clear the contig flag if only one page is needed.
2637 	 */
2638 	if (npages == 1) {
2639 		getone = 1;
2640 		contig = 0;
2641 	}
2642 
2643 	/*
2644 	 * Check if any page in the system is fine.
2645 	 */
2646 	anyaddr = lo_mfn == 0 && hi_mfn >= max_mfn;
2647 	if (!contig && anyaddr && !pfnalign) {
2648 		flags &= ~PG_PHYSCONTIG;
2649 		plist = page_create_va(vp, off, npages * MMU_PAGESIZE,
2650 		    flags, &kvseg, vaddr);
2651 		if (plist != NULL) {
2652 			*npagesp = 0;
2653 			return (plist);
2654 		}
2655 	}
2656 	plist = NULL;
2657 	minctg = howmany(npages, sgllen);
2658 	while (npages > sgllen || getone) {
2659 		if (minctg > npages)
2660 			minctg = npages;
2661 		mcpl = NULL;
2662 		/*
2663 		 * We could want contig pages with no address range limits.
2664 		 */
2665 		if (anyaddr && contig) {
2666 			/*
2667 			 * Look for free contig pages to satisfy the request.
2668 			 */
2669 			mcpl = find_contig_free(minctg, flags, pfnseg,
2670 			    pfnalign);
2671 		}
2672 		/*
2673 		 * Try the reserved io pools next
2674 		 */
2675 		if (mcpl == NULL)
2676 			mcpl = page_io_pool_alloc(mattr, contig, minctg);
2677 		if (mcpl != NULL) {
2678 			pp = mcpl;
2679 			do {
2680 				if (!page_hashin(pp, vp, off, NULL)) {
2681 					panic("page_get_contigpages:"
2682 					    " hashin failed"
2683 					    " pp %p, vp %p, off %llx",
2684 					    (void *)pp, (void *)vp, off);
2685 				}
2686 				off += MMU_PAGESIZE;
2687 				PP_CLRFREE(pp);
2688 				PP_CLRAGED(pp);
2689 				page_set_props(pp, P_REF);
2690 				page_io_lock(pp);
2691 				pp = pp->p_next;
2692 			} while (pp != mcpl);
2693 		} else {
2694 			/*
2695 			 * Hypervisor exchange doesn't handle segment or
2696 			 * alignment constraints
2697 			 */
2698 			if (mattr->dma_attr_seg < mattr->dma_attr_addr_hi ||
2699 			    pfnalign)
2700 				goto fail;
2701 			/*
2702 			 * Try exchanging pages with the hypervisor
2703 			 */
2704 			mcpl = page_swap_with_hypervisor(vp, off, vaddr, mattr,
2705 			    flags, minctg);
2706 			if (mcpl == NULL)
2707 				goto fail;
2708 			off += minctg * MMU_PAGESIZE;
2709 		}
2710 		check_dma(mattr, mcpl, minctg);
2711 		/*
2712 		 * Here with a minctg run of contiguous pages, add them to the
2713 		 * list we will return for this request.
2714 		 */
2715 		page_list_concat(&plist, &mcpl);
2716 		npages -= minctg;
2717 		*npagesp = npages;
2718 		sgllen--;
2719 		if (getone)
2720 			break;
2721 	}
2722 	return (plist);
2723 fail:
2724 	return_partial_alloc(plist);
2725 	return (NULL);
2726 }
2727 
2728 /*
2729  * Allocator for domain 0 I/O pages. We match the required
2730  * DMA attributes and contiguity constraints.
2731  */
2732 /*ARGSUSED*/
2733 page_t *
2734 page_create_io(
2735 	struct vnode	*vp,
2736 	u_offset_t	off,
2737 	uint_t		bytes,
2738 	uint_t		flags,
2739 	struct as	*as,
2740 	caddr_t		vaddr,
2741 	ddi_dma_attr_t	*mattr)
2742 {
2743 	page_t	*plist = NULL, *pp;
2744 	int	npages = 0, contig, anyaddr, pages_req;
2745 	mfn_t	lo_mfn;
2746 	mfn_t	hi_mfn;
2747 	pgcnt_t	pfnalign = 0;
2748 	int	align;
2749 	int	is_domu = 0;
2750 	int	dummy, bytes_got;
2751 	mfn_t	max_mfn = HYPERVISOR_memory_op(XENMEM_maximum_ram_page, NULL);
2752 
2753 	ASSERT(mattr != NULL);
2754 	lo_mfn = mmu_btop(mattr->dma_attr_addr_lo);
2755 	hi_mfn = mmu_btop(mattr->dma_attr_addr_hi);
2756 	align = maxbit(mattr->dma_attr_align, mattr->dma_attr_minxfer);
2757 	if (align > MMU_PAGESIZE)
2758 		pfnalign = mmu_btop(align);
2759 
2760 	/*
2761 	 * Clear the contig flag if only one page is needed or the scatter
2762 	 * gather list length is >= npages.
2763 	 */
2764 	pages_req = npages = mmu_btopr(bytes);
2765 	contig = (flags & PG_PHYSCONTIG);
2766 	bytes = P2ROUNDUP(bytes, MMU_PAGESIZE);
2767 	if (bytes == MMU_PAGESIZE || mattr->dma_attr_sgllen >= npages)
2768 		contig = 0;
2769 
2770 	/*
2771 	 * Check if any old page in the system is fine.
2772 	 * DomU should always go down this path.
2773 	 */
2774 	is_domu = !DOMAIN_IS_INITDOMAIN(xen_info);
2775 	anyaddr = lo_mfn == 0 && hi_mfn >= max_mfn && !pfnalign;
2776 	if ((!contig && anyaddr) || is_domu) {
2777 		flags &= ~PG_PHYSCONTIG;
2778 		plist = page_create_va(vp, off, bytes, flags, &kvseg, vaddr);
2779 		if (plist != NULL)
2780 			return (plist);
2781 		else if (is_domu)
2782 			return (NULL); /* no memory available */
2783 	}
2784 	/*
2785 	 * DomU should never reach here
2786 	 */
2787 	if (contig) {
2788 		plist = page_get_contigpages(vp, off, &npages, flags, vaddr,
2789 		    mattr);
2790 		if (plist == NULL)
2791 			goto fail;
2792 		bytes_got = (pages_req - npages) << MMU_PAGESHIFT;
2793 		vaddr += bytes_got;
2794 		off += bytes_got;
2795 		/*
2796 		 * We now have all the contiguous pages we need, but
2797 		 * we may still need additional non-contiguous pages.
2798 		 */
2799 	}
2800 	/*
2801 	 * now loop collecting the requested number of pages, these do
2802 	 * not have to be contiguous pages but we will use the contig
2803 	 * page alloc code to get the pages since it will honor any
2804 	 * other constraints the pages may have.
2805 	 */
2806 	while (npages--) {
2807 		dummy = -1;
2808 		pp = page_get_contigpages(vp, off, &dummy, flags, vaddr, mattr);
2809 		if (pp == NULL)
2810 			goto fail;
2811 		page_add(&plist, pp);
2812 		vaddr += MMU_PAGESIZE;
2813 		off += MMU_PAGESIZE;
2814 	}
2815 	return (plist);
2816 fail:
2817 	/*
2818 	 * Failed to get enough pages, return ones we did get
2819 	 */
2820 	return_partial_alloc(plist);
2821 	return (NULL);
2822 }
2823 
2824 /*
2825  * Lock and return the page with the highest mfn that we can find.  last_mfn
2826  * holds the last one found, so the next search can start from there.  We
2827  * also keep a counter so that we don't loop forever if the machine has no
2828  * free pages.
2829  *
2830  * This is called from the balloon thread to find pages to give away.  new_high
2831  * is used when new mfn's have been added to the system - we will reset our
2832  * search if the new mfn's are higher than our current search position.
2833  */
2834 page_t *
2835 page_get_high_mfn(mfn_t new_high)
2836 {
2837 	static mfn_t last_mfn = 0;
2838 	pfn_t pfn;
2839 	page_t *pp;
2840 	ulong_t loop_count = 0;
2841 
2842 	if (new_high > last_mfn)
2843 		last_mfn = new_high;
2844 
2845 	for (; loop_count < mfn_count; loop_count++, last_mfn--) {
2846 		if (last_mfn == 0) {
2847 			last_mfn = cached_max_mfn;
2848 		}
2849 
2850 		pfn = mfn_to_pfn(last_mfn);
2851 		if (pfn & PFN_IS_FOREIGN_MFN)
2852 			continue;
2853 
2854 		/* See if the page is free.  If so, lock it. */
2855 		pp = page_numtopp_alloc(pfn);
2856 		if (pp == NULL)
2857 			continue;
2858 		PP_CLRFREE(pp);
2859 
2860 		ASSERT(PAGE_EXCL(pp));
2861 		ASSERT(pp->p_vnode == NULL);
2862 		ASSERT(!hat_page_is_mapped(pp));
2863 		last_mfn--;
2864 		return (pp);
2865 	}
2866 	return (NULL);
2867 }
2868 
2869 #else /* !__xpv */
2870 
2871 /*
2872  * get a page from any list with the given mnode
2873  */
2874 static page_t *
2875 page_get_mnode_anylist(ulong_t origbin, uchar_t szc, uint_t flags,
2876     int mnode, int mtype, ddi_dma_attr_t *dma_attr)
2877 {
2878 	kmutex_t		*pcm;
2879 	int			i;
2880 	page_t			*pp;
2881 	page_t			*first_pp;
2882 	uint64_t		pgaddr;
2883 	ulong_t			bin;
2884 	int			mtypestart;
2885 	int			plw_initialized;
2886 	page_list_walker_t	plw;
2887 
2888 	VM_STAT_ADD(pga_vmstats.pgma_alloc);
2889 
2890 	ASSERT((flags & PG_MATCH_COLOR) == 0);
2891 	ASSERT(szc == 0);
2892 	ASSERT(dma_attr != NULL);
2893 
2894 	MTYPE_START(mnode, mtype, flags);
2895 	if (mtype < 0) {
2896 		VM_STAT_ADD(pga_vmstats.pgma_allocempty);
2897 		return (NULL);
2898 	}
2899 
2900 	mtypestart = mtype;
2901 
2902 	bin = origbin;
2903 
2904 	/*
2905 	 * check up to page_colors + 1 bins - origbin may be checked twice
2906 	 * because of BIN_STEP skip
2907 	 */
2908 	do {
2909 		plw_initialized = 0;
2910 
2911 		for (plw.plw_count = 0;
2912 		    plw.plw_count < page_colors; plw.plw_count++) {
2913 
2914 			if (PAGE_FREELISTS(mnode, szc, bin, mtype) == NULL)
2915 				goto nextfreebin;
2916 
2917 			pcm = PC_BIN_MUTEX(mnode, bin, PG_FREE_LIST);
2918 			mutex_enter(pcm);
2919 			pp = PAGE_FREELISTS(mnode, szc, bin, mtype);
2920 			first_pp = pp;
2921 			while (pp != NULL) {
2922 				if (page_trylock(pp, SE_EXCL) == 0) {
2923 					pp = pp->p_next;
2924 					if (pp == first_pp) {
2925 						pp = NULL;
2926 					}
2927 					continue;
2928 				}
2929 
2930 				ASSERT(PP_ISFREE(pp));
2931 				ASSERT(PP_ISAGED(pp));
2932 				ASSERT(pp->p_vnode == NULL);
2933 				ASSERT(pp->p_hash == NULL);
2934 				ASSERT(pp->p_offset == (u_offset_t)-1);
2935 				ASSERT(pp->p_szc == szc);
2936 				ASSERT(PFN_2_MEM_NODE(pp->p_pagenum) == mnode);
2937 				/* check if page within DMA attributes */
2938 				pgaddr = pa_to_ma(pfn_to_pa(pp->p_pagenum));
2939 				if ((pgaddr >= dma_attr->dma_attr_addr_lo) &&
2940 				    (pgaddr + MMU_PAGESIZE - 1 <=
2941 				    dma_attr->dma_attr_addr_hi)) {
2942 					break;
2943 				}
2944 
2945 				/* continue looking */
2946 				page_unlock(pp);
2947 				pp = pp->p_next;
2948 				if (pp == first_pp)
2949 					pp = NULL;
2950 
2951 			}
2952 			if (pp != NULL) {
2953 				ASSERT(mtype == PP_2_MTYPE(pp));
2954 				ASSERT(pp->p_szc == 0);
2955 
2956 				/* found a page with specified DMA attributes */
2957 				page_sub(&PAGE_FREELISTS(mnode, szc, bin,
2958 				    mtype), pp);
2959 				page_ctr_sub(mnode, mtype, pp, PG_FREE_LIST);
2960 
2961 				if ((PP_ISFREE(pp) == 0) ||
2962 				    (PP_ISAGED(pp) == 0)) {
2963 					cmn_err(CE_PANIC, "page %p is not free",
2964 					    (void *)pp);
2965 				}
2966 
2967 				mutex_exit(pcm);
2968 				check_dma(dma_attr, pp, 1);
2969 				VM_STAT_ADD(pga_vmstats.pgma_allocok);
2970 				return (pp);
2971 			}
2972 			mutex_exit(pcm);
2973 nextfreebin:
2974 			if (plw_initialized == 0) {
2975 				page_list_walk_init(szc, 0, bin, 1, 0, &plw);
2976 				ASSERT(plw.plw_ceq_dif == page_colors);
2977 				plw_initialized = 1;
2978 			}
2979 
2980 			if (plw.plw_do_split) {
2981 				pp = page_freelist_split(szc, bin, mnode,
2982 				    mtype,
2983 				    mmu_btop(dma_attr->dma_attr_addr_lo),
2984 				    mmu_btop(dma_attr->dma_attr_addr_hi + 1),
2985 				    &plw);
2986 				if (pp != NULL) {
2987 					check_dma(dma_attr, pp, 1);
2988 					return (pp);
2989 				}
2990 			}
2991 
2992 			bin = page_list_walk_next_bin(szc, bin, &plw);
2993 		}
2994 
2995 		MTYPE_NEXT(mnode, mtype, flags);
2996 	} while (mtype >= 0);
2997 
2998 	/* failed to find a page in the freelist; try it in the cachelist */
2999 
3000 	/* reset mtype start for cachelist search */
3001 	mtype = mtypestart;
3002 	ASSERT(mtype >= 0);
3003 
3004 	/* start with the bin of matching color */
3005 	bin = origbin;
3006 
3007 	do {
3008 		for (i = 0; i <= page_colors; i++) {
3009 			if (PAGE_CACHELISTS(mnode, bin, mtype) == NULL)
3010 				goto nextcachebin;
3011 			pcm = PC_BIN_MUTEX(mnode, bin, PG_CACHE_LIST);
3012 			mutex_enter(pcm);
3013 			pp = PAGE_CACHELISTS(mnode, bin, mtype);
3014 			first_pp = pp;
3015 			while (pp != NULL) {
3016 				if (page_trylock(pp, SE_EXCL) == 0) {
3017 					pp = pp->p_next;
3018 					if (pp == first_pp)
3019 						pp = NULL;
3020 					continue;
3021 				}
3022 				ASSERT(pp->p_vnode);
3023 				ASSERT(PP_ISAGED(pp) == 0);
3024 				ASSERT(pp->p_szc == 0);
3025 				ASSERT(PFN_2_MEM_NODE(pp->p_pagenum) == mnode);
3026 
3027 				/* check if page within DMA attributes */
3028 
3029 				pgaddr = pa_to_ma(pfn_to_pa(pp->p_pagenum));
3030 				if ((pgaddr >= dma_attr->dma_attr_addr_lo) &&
3031 				    (pgaddr + MMU_PAGESIZE - 1 <=
3032 				    dma_attr->dma_attr_addr_hi)) {
3033 					break;
3034 				}
3035 
3036 				/* continue looking */
3037 				page_unlock(pp);
3038 				pp = pp->p_next;
3039 				if (pp == first_pp)
3040 					pp = NULL;
3041 			}
3042 
3043 			if (pp != NULL) {
3044 				ASSERT(mtype == PP_2_MTYPE(pp));
3045 				ASSERT(pp->p_szc == 0);
3046 
3047 				/* found a page with specified DMA attributes */
3048 				page_sub(&PAGE_CACHELISTS(mnode, bin,
3049 				    mtype), pp);
3050 				page_ctr_sub(mnode, mtype, pp, PG_CACHE_LIST);
3051 
3052 				mutex_exit(pcm);
3053 				ASSERT(pp->p_vnode);
3054 				ASSERT(PP_ISAGED(pp) == 0);
3055 				check_dma(dma_attr, pp, 1);
3056 				VM_STAT_ADD(pga_vmstats.pgma_allocok);
3057 				return (pp);
3058 			}
3059 			mutex_exit(pcm);
3060 nextcachebin:
3061 			bin += (i == 0) ? BIN_STEP : 1;
3062 			bin &= page_colors_mask;
3063 		}
3064 		MTYPE_NEXT(mnode, mtype, flags);
3065 	} while (mtype >= 0);
3066 
3067 	VM_STAT_ADD(pga_vmstats.pgma_allocfailed);
3068 	return (NULL);
3069 }
3070 
3071 /*
3072  * This function is similar to page_get_freelist()/page_get_cachelist()
3073  * but it searches both the lists to find a page with the specified
3074  * color (or no color) and DMA attributes. The search is done in the
3075  * freelist first and then in the cache list within the highest memory
3076  * range (based on DMA attributes) before searching in the lower
3077  * memory ranges.
3078  *
3079  * Note: This function is called only by page_create_io().
3080  */
3081 /*ARGSUSED*/
3082 static page_t *
3083 page_get_anylist(struct vnode *vp, u_offset_t off, struct as *as, caddr_t vaddr,
3084     size_t size, uint_t flags, ddi_dma_attr_t *dma_attr, lgrp_t	*lgrp)
3085 {
3086 	uint_t		bin;
3087 	int		mtype;
3088 	page_t		*pp;
3089 	int		n;
3090 	int		m;
3091 	int		szc;
3092 	int		fullrange;
3093 	int		mnode;
3094 	int		local_failed_stat = 0;
3095 	lgrp_mnode_cookie_t	lgrp_cookie;
3096 
3097 	VM_STAT_ADD(pga_vmstats.pga_alloc);
3098 
3099 	/* only base pagesize currently supported */
3100 	if (size != MMU_PAGESIZE)
3101 		return (NULL);
3102 
3103 	/*
3104 	 * If we're passed a specific lgroup, we use it.  Otherwise,
3105 	 * assume first-touch placement is desired.
3106 	 */
3107 	if (!LGRP_EXISTS(lgrp))
3108 		lgrp = lgrp_home_lgrp();
3109 
3110 	/* LINTED */
3111 	AS_2_BIN(as, seg, vp, vaddr, bin, 0);
3112 
3113 	/*
3114 	 * Only hold one freelist or cachelist lock at a time, that way we
3115 	 * can start anywhere and not have to worry about lock
3116 	 * ordering.
3117 	 */
3118 	if (dma_attr == NULL) {
3119 		n = 0;
3120 		m = mnoderangecnt - 1;
3121 		fullrange = 1;
3122 		VM_STAT_ADD(pga_vmstats.pga_nulldmaattr);
3123 	} else {
3124 		pfn_t pfnlo = mmu_btop(dma_attr->dma_attr_addr_lo);
3125 		pfn_t pfnhi = mmu_btop(dma_attr->dma_attr_addr_hi);
3126 
3127 		/*
3128 		 * We can guarantee alignment only for page boundary.
3129 		 */
3130 		if (dma_attr->dma_attr_align > MMU_PAGESIZE)
3131 			return (NULL);
3132 
3133 		n = pfn_2_mtype(pfnlo);
3134 		m = pfn_2_mtype(pfnhi);
3135 
3136 		fullrange = ((pfnlo == mnoderanges[n].mnr_pfnlo) &&
3137 		    (pfnhi >= mnoderanges[m].mnr_pfnhi));
3138 	}
3139 	VM_STAT_COND_ADD(fullrange == 0, pga_vmstats.pga_notfullrange);
3140 
3141 	if (n > m)
3142 		return (NULL);
3143 
3144 	szc = 0;
3145 
3146 	/* cylcing thru mtype handled by RANGE0 if n == 0 */
3147 	if (n == 0) {
3148 		flags |= PGI_MT_RANGE0;
3149 		n = m;
3150 	}
3151 
3152 	/*
3153 	 * Try local memory node first, but try remote if we can't
3154 	 * get a page of the right color.
3155 	 */
3156 	LGRP_MNODE_COOKIE_INIT(lgrp_cookie, lgrp, LGRP_SRCH_HIER);
3157 	while ((mnode = lgrp_memnode_choose(&lgrp_cookie)) >= 0) {
3158 		/*
3159 		 * allocate pages from high pfn to low.
3160 		 */
3161 		for (mtype = m; mtype >= n; mtype--) {
3162 			if (fullrange != 0) {
3163 				pp = page_get_mnode_freelist(mnode,
3164 				    bin, mtype, szc, flags);
3165 				if (pp == NULL) {
3166 					pp = page_get_mnode_cachelist(
3167 					    bin, flags, mnode, mtype);
3168 				}
3169 			} else {
3170 				pp = page_get_mnode_anylist(bin, szc,
3171 				    flags, mnode, mtype, dma_attr);
3172 			}
3173 			if (pp != NULL) {
3174 				VM_STAT_ADD(pga_vmstats.pga_allocok);
3175 				check_dma(dma_attr, pp, 1);
3176 				return (pp);
3177 			}
3178 		}
3179 		if (!local_failed_stat) {
3180 			lgrp_stat_add(lgrp->lgrp_id, LGRP_NUM_ALLOC_FAIL, 1);
3181 			local_failed_stat = 1;
3182 		}
3183 	}
3184 	VM_STAT_ADD(pga_vmstats.pga_allocfailed);
3185 
3186 	return (NULL);
3187 }
3188 
3189 /*
3190  * page_create_io()
3191  *
3192  * This function is a copy of page_create_va() with an additional
3193  * argument 'mattr' that specifies DMA memory requirements to
3194  * the page list functions. This function is used by the segkmem
3195  * allocator so it is only to create new pages (i.e PG_EXCL is
3196  * set).
3197  *
3198  * Note: This interface is currently used by x86 PSM only and is
3199  *	 not fully specified so the commitment level is only for
3200  *	 private interface specific to x86. This interface uses PSM
3201  *	 specific page_get_anylist() interface.
3202  */
3203 
3204 #define	PAGE_HASH_SEARCH(index, pp, vp, off) { \
3205 	for ((pp) = page_hash[(index)]; (pp); (pp) = (pp)->p_hash) { \
3206 		if ((pp)->p_vnode == (vp) && (pp)->p_offset == (off)) \
3207 			break; \
3208 	} \
3209 }
3210 
3211 
3212 page_t *
3213 page_create_io(
3214 	struct vnode	*vp,
3215 	u_offset_t	off,
3216 	uint_t		bytes,
3217 	uint_t		flags,
3218 	struct as	*as,
3219 	caddr_t		vaddr,
3220 	ddi_dma_attr_t	*mattr)	/* DMA memory attributes if any */
3221 {
3222 	page_t		*plist = NULL;
3223 	uint_t		plist_len = 0;
3224 	pgcnt_t		npages;
3225 	page_t		*npp = NULL;
3226 	uint_t		pages_req;
3227 	page_t		*pp;
3228 	kmutex_t	*phm = NULL;
3229 	uint_t		index;
3230 
3231 	TRACE_4(TR_FAC_VM, TR_PAGE_CREATE_START,
3232 	    "page_create_start:vp %p off %llx bytes %u flags %x",
3233 	    vp, off, bytes, flags);
3234 
3235 	ASSERT((flags & ~(PG_EXCL | PG_WAIT | PG_PHYSCONTIG)) == 0);
3236 
3237 	pages_req = npages = mmu_btopr(bytes);
3238 
3239 	/*
3240 	 * Do the freemem and pcf accounting.
3241 	 */
3242 	if (!page_create_wait(npages, flags)) {
3243 		return (NULL);
3244 	}
3245 
3246 	TRACE_2(TR_FAC_VM, TR_PAGE_CREATE_SUCCESS,
3247 	    "page_create_success:vp %p off %llx", vp, off);
3248 
3249 	/*
3250 	 * If satisfying this request has left us with too little
3251 	 * memory, start the wheels turning to get some back.  The
3252 	 * first clause of the test prevents waking up the pageout
3253 	 * daemon in situations where it would decide that there's
3254 	 * nothing to do.
3255 	 */
3256 	if (nscan < desscan && freemem < minfree) {
3257 		TRACE_1(TR_FAC_VM, TR_PAGEOUT_CV_SIGNAL,
3258 		    "pageout_cv_signal:freemem %ld", freemem);
3259 		cv_signal(&proc_pageout->p_cv);
3260 	}
3261 
3262 	if (flags & PG_PHYSCONTIG) {
3263 
3264 		plist = page_get_contigpage(&npages, mattr, 1);
3265 		if (plist == NULL) {
3266 			page_create_putback(npages);
3267 			return (NULL);
3268 		}
3269 
3270 		pp = plist;
3271 
3272 		do {
3273 			if (!page_hashin(pp, vp, off, NULL)) {
3274 				panic("pg_creat_io: hashin failed %p %p %llx",
3275 				    (void *)pp, (void *)vp, off);
3276 			}
3277 			VM_STAT_ADD(page_create_new);
3278 			off += MMU_PAGESIZE;
3279 			PP_CLRFREE(pp);
3280 			PP_CLRAGED(pp);
3281 			page_set_props(pp, P_REF);
3282 			pp = pp->p_next;
3283 		} while (pp != plist);
3284 
3285 		if (!npages) {
3286 			check_dma(mattr, plist, pages_req);
3287 			return (plist);
3288 		} else {
3289 			vaddr += (pages_req - npages) << MMU_PAGESHIFT;
3290 		}
3291 
3292 		/*
3293 		 * fall-thru:
3294 		 *
3295 		 * page_get_contigpage returns when npages <= sgllen.
3296 		 * Grab the rest of the non-contig pages below from anylist.
3297 		 */
3298 	}
3299 
3300 	/*
3301 	 * Loop around collecting the requested number of pages.
3302 	 * Most of the time, we have to `create' a new page. With
3303 	 * this in mind, pull the page off the free list before
3304 	 * getting the hash lock.  This will minimize the hash
3305 	 * lock hold time, nesting, and the like.  If it turns
3306 	 * out we don't need the page, we put it back at the end.
3307 	 */
3308 	while (npages--) {
3309 		phm = NULL;
3310 
3311 		index = PAGE_HASH_FUNC(vp, off);
3312 top:
3313 		ASSERT(phm == NULL);
3314 		ASSERT(index == PAGE_HASH_FUNC(vp, off));
3315 		ASSERT(MUTEX_NOT_HELD(page_vnode_mutex(vp)));
3316 
3317 		if (npp == NULL) {
3318 			/*
3319 			 * Try to get the page of any color either from
3320 			 * the freelist or from the cache list.
3321 			 */
3322 			npp = page_get_anylist(vp, off, as, vaddr, MMU_PAGESIZE,
3323 			    flags & ~PG_MATCH_COLOR, mattr, NULL);
3324 			if (npp == NULL) {
3325 				if (mattr == NULL) {
3326 					/*
3327 					 * Not looking for a special page;
3328 					 * panic!
3329 					 */
3330 					panic("no page found %d", (int)npages);
3331 				}
3332 				/*
3333 				 * No page found! This can happen
3334 				 * if we are looking for a page
3335 				 * within a specific memory range
3336 				 * for DMA purposes. If PG_WAIT is
3337 				 * specified then we wait for a
3338 				 * while and then try again. The
3339 				 * wait could be forever if we
3340 				 * don't get the page(s) we need.
3341 				 *
3342 				 * Note: XXX We really need a mechanism
3343 				 * to wait for pages in the desired
3344 				 * range. For now, we wait for any
3345 				 * pages and see if we can use it.
3346 				 */
3347 
3348 				if ((mattr != NULL) && (flags & PG_WAIT)) {
3349 					delay(10);
3350 					goto top;
3351 				}
3352 				goto fail; /* undo accounting stuff */
3353 			}
3354 
3355 			if (PP_ISAGED(npp) == 0) {
3356 				/*
3357 				 * Since this page came from the
3358 				 * cachelist, we must destroy the
3359 				 * old vnode association.
3360 				 */
3361 				page_hashout(npp, (kmutex_t *)NULL);
3362 			}
3363 		}
3364 
3365 		/*
3366 		 * We own this page!
3367 		 */
3368 		ASSERT(PAGE_EXCL(npp));
3369 		ASSERT(npp->p_vnode == NULL);
3370 		ASSERT(!hat_page_is_mapped(npp));
3371 		PP_CLRFREE(npp);
3372 		PP_CLRAGED(npp);
3373 
3374 		/*
3375 		 * Here we have a page in our hot little mits and are
3376 		 * just waiting to stuff it on the appropriate lists.
3377 		 * Get the mutex and check to see if it really does
3378 		 * not exist.
3379 		 */
3380 		phm = PAGE_HASH_MUTEX(index);
3381 		mutex_enter(phm);
3382 		PAGE_HASH_SEARCH(index, pp, vp, off);
3383 		if (pp == NULL) {
3384 			VM_STAT_ADD(page_create_new);
3385 			pp = npp;
3386 			npp = NULL;
3387 			if (!page_hashin(pp, vp, off, phm)) {
3388 				/*
3389 				 * Since we hold the page hash mutex and
3390 				 * just searched for this page, page_hashin
3391 				 * had better not fail.  If it does, that
3392 				 * means somethread did not follow the
3393 				 * page hash mutex rules.  Panic now and
3394 				 * get it over with.  As usual, go down
3395 				 * holding all the locks.
3396 				 */
3397 				ASSERT(MUTEX_HELD(phm));
3398 				panic("page_create: hashin fail %p %p %llx %p",
3399 				    (void *)pp, (void *)vp, off, (void *)phm);
3400 
3401 			}
3402 			ASSERT(MUTEX_HELD(phm));
3403 			mutex_exit(phm);
3404 			phm = NULL;
3405 
3406 			/*
3407 			 * Hat layer locking need not be done to set
3408 			 * the following bits since the page is not hashed
3409 			 * and was on the free list (i.e., had no mappings).
3410 			 *
3411 			 * Set the reference bit to protect
3412 			 * against immediate pageout
3413 			 *
3414 			 * XXXmh modify freelist code to set reference
3415 			 * bit so we don't have to do it here.
3416 			 */
3417 			page_set_props(pp, P_REF);
3418 		} else {
3419 			ASSERT(MUTEX_HELD(phm));
3420 			mutex_exit(phm);
3421 			phm = NULL;
3422 			/*
3423 			 * NOTE: This should not happen for pages associated
3424 			 *	 with kernel vnode 'kvp'.
3425 			 */
3426 			/* XX64 - to debug why this happens! */
3427 			ASSERT(!VN_ISKAS(vp));
3428 			if (VN_ISKAS(vp))
3429 				cmn_err(CE_NOTE,
3430 				    "page_create: page not expected "
3431 				    "in hash list for kernel vnode - pp 0x%p",
3432 				    (void *)pp);
3433 			VM_STAT_ADD(page_create_exists);
3434 			goto fail;
3435 		}
3436 
3437 		/*
3438 		 * Got a page!  It is locked.  Acquire the i/o
3439 		 * lock since we are going to use the p_next and
3440 		 * p_prev fields to link the requested pages together.
3441 		 */
3442 		page_io_lock(pp);
3443 		page_add(&plist, pp);
3444 		plist = plist->p_next;
3445 		off += MMU_PAGESIZE;
3446 		vaddr += MMU_PAGESIZE;
3447 	}
3448 
3449 	check_dma(mattr, plist, pages_req);
3450 	return (plist);
3451 
3452 fail:
3453 	if (npp != NULL) {
3454 		/*
3455 		 * Did not need this page after all.
3456 		 * Put it back on the free list.
3457 		 */
3458 		VM_STAT_ADD(page_create_putbacks);
3459 		PP_SETFREE(npp);
3460 		PP_SETAGED(npp);
3461 		npp->p_offset = (u_offset_t)-1;
3462 		page_list_add(npp, PG_FREE_LIST | PG_LIST_TAIL);
3463 		page_unlock(npp);
3464 	}
3465 
3466 	/*
3467 	 * Give up the pages we already got.
3468 	 */
3469 	while (plist != NULL) {
3470 		pp = plist;
3471 		page_sub(&plist, pp);
3472 		page_io_unlock(pp);
3473 		plist_len++;
3474 		/*LINTED: constant in conditional ctx*/
3475 		VN_DISPOSE(pp, B_INVAL, 0, kcred);
3476 	}
3477 
3478 	/*
3479 	 * VN_DISPOSE does freemem accounting for the pages in plist
3480 	 * by calling page_free. So, we need to undo the pcf accounting
3481 	 * for only the remaining pages.
3482 	 */
3483 	VM_STAT_ADD(page_create_putbacks);
3484 	page_create_putback(pages_req - plist_len);
3485 
3486 	return (NULL);
3487 }
3488 #endif /* !__xpv */
3489 
3490 
3491 /*
3492  * Copy the data from the physical page represented by "frompp" to
3493  * that represented by "topp". ppcopy uses CPU->cpu_caddr1 and
3494  * CPU->cpu_caddr2.  It assumes that no one uses either map at interrupt
3495  * level and no one sleeps with an active mapping there.
3496  *
3497  * Note that the ref/mod bits in the page_t's are not affected by
3498  * this operation, hence it is up to the caller to update them appropriately.
3499  */
3500 int
3501 ppcopy(page_t *frompp, page_t *topp)
3502 {
3503 	caddr_t		pp_addr1;
3504 	caddr_t		pp_addr2;
3505 	hat_mempte_t	pte1;
3506 	hat_mempte_t	pte2;
3507 	kmutex_t	*ppaddr_mutex;
3508 	label_t		ljb;
3509 	int		ret = 1;
3510 
3511 	ASSERT_STACK_ALIGNED();
3512 	ASSERT(PAGE_LOCKED(frompp));
3513 	ASSERT(PAGE_LOCKED(topp));
3514 
3515 	if (kpm_enable) {
3516 		pp_addr1 = hat_kpm_page2va(frompp, 0);
3517 		pp_addr2 = hat_kpm_page2va(topp, 0);
3518 		kpreempt_disable();
3519 	} else {
3520 		/*
3521 		 * disable pre-emption so that CPU can't change
3522 		 */
3523 		kpreempt_disable();
3524 
3525 		pp_addr1 = CPU->cpu_caddr1;
3526 		pp_addr2 = CPU->cpu_caddr2;
3527 		pte1 = CPU->cpu_caddr1pte;
3528 		pte2 = CPU->cpu_caddr2pte;
3529 
3530 		ppaddr_mutex = &CPU->cpu_ppaddr_mutex;
3531 		mutex_enter(ppaddr_mutex);
3532 
3533 		hat_mempte_remap(page_pptonum(frompp), pp_addr1, pte1,
3534 		    PROT_READ | HAT_STORECACHING_OK, HAT_LOAD_NOCONSIST);
3535 		hat_mempte_remap(page_pptonum(topp), pp_addr2, pte2,
3536 		    PROT_READ | PROT_WRITE | HAT_STORECACHING_OK,
3537 		    HAT_LOAD_NOCONSIST);
3538 	}
3539 
3540 	if (on_fault(&ljb)) {
3541 		ret = 0;
3542 		goto faulted;
3543 	}
3544 	if (use_sse_pagecopy)
3545 #ifdef __xpv
3546 		page_copy_no_xmm(pp_addr2, pp_addr1);
3547 #else
3548 		hwblkpagecopy(pp_addr1, pp_addr2);
3549 #endif
3550 	else
3551 		bcopy(pp_addr1, pp_addr2, PAGESIZE);
3552 
3553 	no_fault();
3554 faulted:
3555 	if (!kpm_enable) {
3556 #ifdef __xpv
3557 		/*
3558 		 * We can't leave unused mappings laying about under the
3559 		 * hypervisor, so blow them away.
3560 		 */
3561 		if (HYPERVISOR_update_va_mapping((uintptr_t)pp_addr1, 0,
3562 		    UVMF_INVLPG | UVMF_LOCAL) < 0)
3563 			panic("HYPERVISOR_update_va_mapping() failed");
3564 		if (HYPERVISOR_update_va_mapping((uintptr_t)pp_addr2, 0,
3565 		    UVMF_INVLPG | UVMF_LOCAL) < 0)
3566 			panic("HYPERVISOR_update_va_mapping() failed");
3567 #endif
3568 		mutex_exit(ppaddr_mutex);
3569 	}
3570 	kpreempt_enable();
3571 	return (ret);
3572 }
3573 
3574 void
3575 pagezero(page_t *pp, uint_t off, uint_t len)
3576 {
3577 	ASSERT(PAGE_LOCKED(pp));
3578 	pfnzero(page_pptonum(pp), off, len);
3579 }
3580 
3581 /*
3582  * Zero the physical page from off to off + len given by pfn
3583  * without changing the reference and modified bits of page.
3584  *
3585  * We use this using CPU private page address #2, see ppcopy() for more info.
3586  * pfnzero() must not be called at interrupt level.
3587  */
3588 void
3589 pfnzero(pfn_t pfn, uint_t off, uint_t len)
3590 {
3591 	caddr_t		pp_addr2;
3592 	hat_mempte_t	pte2;
3593 	kmutex_t	*ppaddr_mutex = NULL;
3594 
3595 	ASSERT_STACK_ALIGNED();
3596 	ASSERT(len <= MMU_PAGESIZE);
3597 	ASSERT(off <= MMU_PAGESIZE);
3598 	ASSERT(off + len <= MMU_PAGESIZE);
3599 
3600 	if (kpm_enable && !pfn_is_foreign(pfn)) {
3601 		pp_addr2 = hat_kpm_pfn2va(pfn);
3602 		kpreempt_disable();
3603 	} else {
3604 		kpreempt_disable();
3605 
3606 		pp_addr2 = CPU->cpu_caddr2;
3607 		pte2 = CPU->cpu_caddr2pte;
3608 
3609 		ppaddr_mutex = &CPU->cpu_ppaddr_mutex;
3610 		mutex_enter(ppaddr_mutex);
3611 
3612 		hat_mempte_remap(pfn, pp_addr2, pte2,
3613 		    PROT_READ | PROT_WRITE | HAT_STORECACHING_OK,
3614 		    HAT_LOAD_NOCONSIST);
3615 	}
3616 
3617 	if (use_sse_pagezero) {
3618 #ifdef __xpv
3619 		uint_t rem;
3620 
3621 		/*
3622 		 * zero a byte at a time until properly aligned for
3623 		 * block_zero_no_xmm().
3624 		 */
3625 		while (!P2NPHASE(off, ((uint_t)BLOCKZEROALIGN)) && len-- > 0)
3626 			pp_addr2[off++] = 0;
3627 
3628 		/*
3629 		 * Now use faster block_zero_no_xmm() for any range
3630 		 * that is properly aligned and sized.
3631 		 */
3632 		rem = P2PHASE(len, ((uint_t)BLOCKZEROALIGN));
3633 		len -= rem;
3634 		if (len != 0) {
3635 			block_zero_no_xmm(pp_addr2 + off, len);
3636 			off += len;
3637 		}
3638 
3639 		/*
3640 		 * zero remainder with byte stores.
3641 		 */
3642 		while (rem-- > 0)
3643 			pp_addr2[off++] = 0;
3644 #else
3645 		hwblkclr(pp_addr2 + off, len);
3646 #endif
3647 	} else {
3648 		bzero(pp_addr2 + off, len);
3649 	}
3650 
3651 	if (!kpm_enable || pfn_is_foreign(pfn)) {
3652 #ifdef __xpv
3653 		/*
3654 		 * On the hypervisor this page might get used for a page
3655 		 * table before any intervening change to this mapping,
3656 		 * so blow it away.
3657 		 */
3658 		if (HYPERVISOR_update_va_mapping((uintptr_t)pp_addr2, 0,
3659 		    UVMF_INVLPG) < 0)
3660 			panic("HYPERVISOR_update_va_mapping() failed");
3661 #endif
3662 		mutex_exit(ppaddr_mutex);
3663 	}
3664 
3665 	kpreempt_enable();
3666 }
3667 
3668 /*
3669  * Platform-dependent page scrub call.
3670  */
3671 void
3672 pagescrub(page_t *pp, uint_t off, uint_t len)
3673 {
3674 	/*
3675 	 * For now, we rely on the fact that pagezero() will
3676 	 * always clear UEs.
3677 	 */
3678 	pagezero(pp, off, len);
3679 }
3680 
3681 /*
3682  * set up two private addresses for use on a given CPU for use in ppcopy()
3683  */
3684 void
3685 setup_vaddr_for_ppcopy(struct cpu *cpup)
3686 {
3687 	void *addr;
3688 	hat_mempte_t pte_pa;
3689 
3690 	addr = vmem_alloc(heap_arena, mmu_ptob(1), VM_SLEEP);
3691 	pte_pa = hat_mempte_setup(addr);
3692 	cpup->cpu_caddr1 = addr;
3693 	cpup->cpu_caddr1pte = pte_pa;
3694 
3695 	addr = vmem_alloc(heap_arena, mmu_ptob(1), VM_SLEEP);
3696 	pte_pa = hat_mempte_setup(addr);
3697 	cpup->cpu_caddr2 = addr;
3698 	cpup->cpu_caddr2pte = pte_pa;
3699 
3700 	mutex_init(&cpup->cpu_ppaddr_mutex, NULL, MUTEX_DEFAULT, NULL);
3701 }
3702 
3703 /*
3704  * Undo setup_vaddr_for_ppcopy
3705  */
3706 void
3707 teardown_vaddr_for_ppcopy(struct cpu *cpup)
3708 {
3709 	mutex_destroy(&cpup->cpu_ppaddr_mutex);
3710 
3711 	hat_mempte_release(cpup->cpu_caddr2, cpup->cpu_caddr2pte);
3712 	cpup->cpu_caddr2pte = 0;
3713 	vmem_free(heap_arena, cpup->cpu_caddr2, mmu_ptob(1));
3714 	cpup->cpu_caddr2 = 0;
3715 
3716 	hat_mempte_release(cpup->cpu_caddr1, cpup->cpu_caddr1pte);
3717 	cpup->cpu_caddr1pte = 0;
3718 	vmem_free(heap_arena, cpup->cpu_caddr1, mmu_ptob(1));
3719 	cpup->cpu_caddr1 = 0;
3720 }
3721 
3722 /*
3723  * Create the pageout scanner thread. The thread has to
3724  * start at procedure with process pp and priority pri.
3725  */
3726 void
3727 pageout_init(void (*procedure)(), proc_t *pp, pri_t pri)
3728 {
3729 	(void) thread_create(NULL, 0, procedure, NULL, 0, pp, TS_RUN, pri);
3730 }
3731 
3732 /*
3733  * Function for flushing D-cache when performing module relocations
3734  * to an alternate mapping.  Unnecessary on Intel / AMD platforms.
3735  */
3736 void
3737 dcache_flushall()
3738 {}
3739 
3740 size_t
3741 exec_get_spslew(void)
3742 {
3743 	return (0);
3744 }
3745 
3746 /*
3747  * Allocate a memory page.  The argument 'seed' can be any pseudo-random
3748  * number to vary where the pages come from.  This is quite a hacked up
3749  * method -- it works for now, but really needs to be fixed up a bit.
3750  *
3751  * We currently use page_create_va() on the kvp with fake offsets,
3752  * segments and virt address.  This is pretty bogus, but was copied from the
3753  * old hat_i86.c code.  A better approach would be to specify either mnode
3754  * random or mnode local and takes a page from whatever color has the MOST
3755  * available - this would have a minimal impact on page coloring.
3756  */
3757 page_t *
3758 page_get_physical(uintptr_t seed)
3759 {
3760 	page_t *pp;
3761 	u_offset_t offset;
3762 	static struct seg tmpseg;
3763 	static uintptr_t ctr = 0;
3764 
3765 	/*
3766 	 * This code is gross, we really need a simpler page allocator.
3767 	 *
3768 	 * We need to assign an offset for the page to call page_create_va()
3769 	 * To avoid conflicts with other pages, we get creative with the offset.
3770 	 * For 32 bits, we need an offset > 4Gig
3771 	 * For 64 bits, need an offset somewhere in the VA hole.
3772 	 */
3773 	offset = seed;
3774 	if (offset > kernelbase)
3775 		offset -= kernelbase;
3776 	offset <<= MMU_PAGESHIFT;
3777 #if defined(__amd64)
3778 	offset += mmu.hole_start;	/* something in VA hole */
3779 #else
3780 	offset += 1ULL << 40;	/* something > 4 Gig */
3781 #endif
3782 
3783 	if (page_resv(1, KM_NOSLEEP) == 0)
3784 		return (NULL);
3785 
3786 #ifdef	DEBUG
3787 	pp = page_exists(&kvp, offset);
3788 	if (pp != NULL)
3789 		panic("page already exists %p", (void *)pp);
3790 #endif
3791 
3792 	pp = page_create_va(&kvp, offset, MMU_PAGESIZE, PG_EXCL,
3793 	    &tmpseg, (caddr_t)(ctr += MMU_PAGESIZE));	/* changing VA usage */
3794 	if (pp != NULL) {
3795 		page_io_unlock(pp);
3796 		page_hashout(pp, NULL);
3797 		page_downgrade(pp);
3798 	}
3799 	return (pp);
3800 }
3801