xref: /illumos-gate/usr/src/uts/sun4v/vm/mach_vm_dep.c (revision 8119dad84d6416f13557b0ba8e2aaf9064cbcfd3)
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/vm.h>
39 #include <sys/exec.h>
40 #include <sys/cmn_err.h>
41 #include <sys/cpu_module.h>
42 #include <sys/cpu.h>
43 #include <sys/elf_SPARC.h>
44 #include <sys/archsystm.h>
45 #include <vm/hat_sfmmu.h>
46 #include <sys/memnode.h>
47 #include <sys/mem_cage.h>
48 #include <vm/vm_dep.h>
49 #include <sys/error.h>
50 #include <sys/machsystm.h>
51 #include <vm/seg_kmem.h>
52 #include <sys/stack.h>
53 #include <sys/atomic.h>
54 #include <sys/promif.h>
55 #include <sys/random.h>
56 
57 uint_t page_colors = 0;
58 uint_t page_colors_mask = 0;
59 uint_t page_coloring_shift = 0;
60 int consistent_coloring;
61 int update_proc_pgcolorbase_after_fork = 1;
62 
63 uint_t mmu_page_sizes = MMU_PAGE_SIZES;
64 uint_t max_mmu_page_sizes = MMU_PAGE_SIZES;
65 uint_t mmu_hashcnt = MAX_HASHCNT;
66 uint_t max_mmu_hashcnt = MAX_HASHCNT;
67 size_t mmu_ism_pagesize = DEFAULT_ISM_PAGESIZE;
68 
69 /*
70  * A bitmask of the page sizes supported by hardware based upon szc.
71  * The base pagesize (p_szc == 0) must always be supported by the hardware.
72  */
73 int mmu_exported_pagesize_mask;
74 uint_t mmu_exported_page_sizes;
75 
76 uint_t szc_2_userszc[MMU_PAGE_SIZES];
77 uint_t userszc_2_szc[MMU_PAGE_SIZES];
78 
79 extern uint_t vac_colors_mask;
80 extern int vac_shift;
81 
82 hw_pagesize_t hw_page_array[] = {
83 	{MMU_PAGESIZE, MMU_PAGESHIFT, 0, MMU_PAGESIZE >> MMU_PAGESHIFT},
84 	{MMU_PAGESIZE64K, MMU_PAGESHIFT64K, 0,
85 	    MMU_PAGESIZE64K >> MMU_PAGESHIFT},
86 	{MMU_PAGESIZE512K, MMU_PAGESHIFT512K, 0,
87 	    MMU_PAGESIZE512K >> MMU_PAGESHIFT},
88 	{MMU_PAGESIZE4M, MMU_PAGESHIFT4M, 0, MMU_PAGESIZE4M >> MMU_PAGESHIFT},
89 	{MMU_PAGESIZE32M, MMU_PAGESHIFT32M, 0,
90 	    MMU_PAGESIZE32M >> MMU_PAGESHIFT},
91 	{MMU_PAGESIZE256M, MMU_PAGESHIFT256M, 0,
92 	    MMU_PAGESIZE256M >> MMU_PAGESHIFT},
93 	{0, 0, 0, 0}
94 };
95 
96 /*
97  * Maximum page size used to map 64-bit memory segment kmem64_base..kmem64_end
98  */
99 int	max_bootlp_tteszc = TTE256M;
100 
101 /*
102  * Maximum and default segment size tunables for user heap, stack, private
103  * and shared anonymous memory, and user text and initialized data.
104  */
105 size_t max_uheap_lpsize = MMU_PAGESIZE64K;
106 size_t default_uheap_lpsize = MMU_PAGESIZE64K;
107 size_t max_ustack_lpsize = MMU_PAGESIZE64K;
108 size_t default_ustack_lpsize = MMU_PAGESIZE64K;
109 size_t max_privmap_lpsize = MMU_PAGESIZE64K;
110 size_t max_uidata_lpsize = MMU_PAGESIZE64K;
111 size_t max_utext_lpsize = MMU_PAGESIZE4M;
112 size_t max_shm_lpsize = MMU_PAGESIZE4M;
113 
114 /*
115  * Contiguous memory allocator data structures and variables.
116  *
117  * The sun4v kernel must provide a means to allocate physically
118  * contiguous, non-relocatable memory. The contig_mem_arena
119  * and contig_mem_slab_arena exist for this purpose. Allocations
120  * that require physically contiguous non-relocatable memory should
121  * be made using contig_mem_alloc() or contig_mem_alloc_align()
122  * which return memory from contig_mem_arena or contig_mem_reloc_arena.
123  * These arenas import memory from the contig_mem_slab_arena one
124  * contiguous chunk at a time.
125  *
126  * When importing slabs, an attempt is made to allocate a large page
127  * to use as backing. As a result of the non-relocatable requirement,
128  * slabs are allocated from the kernel cage freelists. If the cage does
129  * not contain any free contiguous chunks large enough to satisfy the
130  * slab allocation, the slab size will be downsized and the operation
131  * retried. Large slab sizes are tried first to minimize cage
132  * fragmentation. If the slab allocation is unsuccessful still, the slab
133  * is allocated from outside the kernel cage. This is undesirable because,
134  * until slabs are freed, it results in non-relocatable chunks scattered
135  * throughout physical memory.
136  *
137  * Allocations from the contig_mem_arena are backed by slabs from the
138  * cage. Allocations from the contig_mem_reloc_arena are backed by
139  * slabs allocated outside the cage. Slabs are left share locked while
140  * in use to prevent non-cage slabs from being relocated.
141  *
142  * Since there is no guarantee that large pages will be available in
143  * the kernel cage, contiguous memory is reserved and added to the
144  * contig_mem_arena at boot time, making it available for later
145  * contiguous memory allocations. This reserve will be used to satisfy
146  * contig_mem allocations first and it is only when the reserve is
147  * completely allocated that new slabs will need to be imported.
148  */
149 static	vmem_t		*contig_mem_slab_arena;
150 static	vmem_t		*contig_mem_arena;
151 static	vmem_t		*contig_mem_reloc_arena;
152 static	kmutex_t	contig_mem_lock;
153 #define	CONTIG_MEM_ARENA_QUANTUM	64
154 #define	CONTIG_MEM_SLAB_ARENA_QUANTUM	MMU_PAGESIZE64K
155 
156 /* contig_mem_arena import slab sizes, in decreasing size order */
157 static size_t contig_mem_import_sizes[] = {
158 	MMU_PAGESIZE4M,
159 	MMU_PAGESIZE512K,
160 	MMU_PAGESIZE64K
161 };
162 #define	NUM_IMPORT_SIZES	\
163 	(sizeof (contig_mem_import_sizes) / sizeof (size_t))
164 static size_t contig_mem_import_size_max	= MMU_PAGESIZE4M;
165 size_t contig_mem_slab_size			= MMU_PAGESIZE4M;
166 
167 /* Boot-time allocated buffer to pre-populate the contig_mem_arena */
168 static size_t contig_mem_prealloc_size;
169 static void *contig_mem_prealloc_buf;
170 
171 /*
172  * The maximum amount a randomized mapping will be slewed.  We should perhaps
173  * arrange things so these tunables can be separate for mmap, mmapobj, and
174  * ld.so
175  */
176 size_t aslr_max_map_skew = 256 * 1024 * 1024; /* 256MB */
177 
178 /*
179  * map_addr_proc() is the routine called when the system is to
180  * choose an address for the user.  We will pick an address
181  * range which is just below the current stack limit.  The
182  * algorithm used for cache consistency on machines with virtual
183  * address caches is such that offset 0 in the vnode is always
184  * on a shm_alignment'ed aligned address.  Unfortunately, this
185  * means that vnodes which are demand paged will not be mapped
186  * cache consistently with the executable images.  When the
187  * cache alignment for a given object is inconsistent, the
188  * lower level code must manage the translations so that this
189  * is not seen here (at the cost of efficiency, of course).
190  *
191  * Every mapping will have a redzone of a single page on either side of
192  * the request. This is done to leave one page unmapped between segments.
193  * This is not required, but it's useful for the user because if their
194  * program strays across a segment boundary, it will catch a fault
195  * immediately making debugging a little easier.  Currently the redzone
196  * is mandatory.
197  *
198  * addrp is a value/result parameter.
199  *	On input it is a hint from the user to be used in a completely
200  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
201  *	minimal alignment, which must be some "power of two" multiple of
202  *	pagesize.
203  *
204  *	On output it is NULL if no address can be found in the current
205  *	processes address space or else an address that is currently
206  *	not mapped for len bytes with a page of red zone on either side.
207  *	If vacalign is true, then the selected address will obey the alignment
208  *	constraints of a vac machine based on the given off value.
209  */
210 /*ARGSUSED3*/
211 void
212 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
213     caddr_t userlimit, struct proc *p, uint_t flags)
214 {
215 	struct as *as = p->p_as;
216 	caddr_t addr;
217 	caddr_t base;
218 	size_t slen;
219 	uintptr_t align_amount;
220 	int allow_largepage_alignment = 1;
221 
222 	base = p->p_brkbase;
223 	if (userlimit < as->a_userlimit) {
224 		/*
225 		 * This happens when a program wants to map something in
226 		 * a range that's accessible to a program in a smaller
227 		 * address space.  For example, a 64-bit program might
228 		 * be calling mmap32(2) to guarantee that the returned
229 		 * address is below 4Gbytes.
230 		 */
231 		ASSERT(userlimit > base);
232 		slen = userlimit - base;
233 	} else {
234 		slen = p->p_usrstack - base -
235 		    ((p->p_stk_ctl + PAGEOFFSET) & PAGEMASK);
236 	}
237 	/* Make len be a multiple of PAGESIZE */
238 	len = (len + PAGEOFFSET) & PAGEMASK;
239 
240 	/*
241 	 *  If the request is larger than the size of a particular
242 	 *  mmu level, then we use that level to map the request.
243 	 *  But this requires that both the virtual and the physical
244 	 *  addresses be aligned with respect to that level, so we
245 	 *  do the virtual bit of nastiness here.
246 	 *
247 	 *  For 32-bit processes, only those which have specified
248 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
249 	 *  we can potentially waste up to 256MB of the 4G process address
250 	 *  space just for alignment.
251 	 *
252 	 * XXXQ Should iterate trough hw_page_array here to catch
253 	 * all supported pagesizes
254 	 */
255 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
256 	    ((uintptr_t)*addrp) != 0)) {
257 		allow_largepage_alignment = 0;
258 	}
259 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
260 	    allow_largepage_alignment &&
261 	    (len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
262 		align_amount = MMU_PAGESIZE256M;
263 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
264 	    allow_largepage_alignment &&
265 	    (len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
266 		align_amount = MMU_PAGESIZE32M;
267 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
268 		align_amount = MMU_PAGESIZE4M;
269 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
270 		align_amount = MMU_PAGESIZE512K;
271 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
272 		align_amount = MMU_PAGESIZE64K;
273 	} else  {
274 		/*
275 		 * Align virtual addresses on a 64K boundary to ensure
276 		 * that ELF shared libraries are mapped with the appropriate
277 		 * alignment constraints by the run-time linker.
278 		 */
279 		align_amount = ELF_SPARC_MAXPGSZ;
280 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
281 		    ((uintptr_t)*addrp < align_amount))
282 			align_amount = (uintptr_t)*addrp;
283 	}
284 
285 	/*
286 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
287 	 */
288 	if (p->p_model == DATAMODEL_LP64)
289 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
290 #ifdef VAC
291 	if (vac && vacalign && (align_amount < shm_alignment))
292 		align_amount = shm_alignment;
293 #endif
294 
295 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
296 		align_amount = (uintptr_t)*addrp;
297 	}
298 
299 	ASSERT(ISP2(align_amount));
300 	ASSERT(align_amount == 0 || align_amount >= PAGESIZE);
301 
302 	/*
303 	 * Look for a large enough hole starting below the stack limit.
304 	 * After finding it, use the upper part.
305 	 */
306 	as_purge(as);
307 	off = off & (align_amount - 1);
308 	if (as_gap_aligned(as, len, &base, &slen, AH_HI, NULL, align_amount,
309 	    PAGESIZE, off) == 0) {
310 		caddr_t as_addr;
311 
312 		/*
313 		 * addr is the highest possible address to use since we have
314 		 * a PAGESIZE redzone at the beginning and end.
315 		 */
316 		addr = base + slen - (PAGESIZE + len);
317 		as_addr = addr;
318 		/*
319 		 * Round address DOWN to the alignment amount and
320 		 * add the offset in.
321 		 * If addr is greater than as_addr, len would not be large
322 		 * enough to include the redzone, so we must adjust down
323 		 * by the alignment amount.
324 		 */
325 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
326 		addr += (long)off;
327 		if (addr > as_addr) {
328 			addr -= align_amount;
329 		}
330 
331 		/*
332 		 * If randomization is requested, slew the allocation
333 		 * backwards, within the same gap, by a random amount.
334 		 */
335 		if (flags & _MAP_RANDOMIZE) {
336 			uint32_t slew;
337 
338 			(void) random_get_pseudo_bytes((uint8_t *)&slew,
339 			    sizeof (slew));
340 
341 			slew = slew % MIN(aslr_max_map_skew, (addr - base));
342 			addr -= P2ALIGN(slew, align_amount);
343 		}
344 
345 		ASSERT(addr > base);
346 		ASSERT(addr + len < base + slen);
347 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
348 		    ((uintptr_t)(off)));
349 		*addrp = addr;
350 
351 	} else {
352 		*addrp = NULL;	/* no more virtual space */
353 	}
354 }
355 
356 /*
357  * Platform-dependent page scrub call.
358  * We call hypervisor to scrub the page.
359  */
360 void
361 pagescrub(page_t *pp, uint_t off, uint_t len)
362 {
363 	uint64_t pa, length;
364 
365 	pa = (uint64_t)(pp->p_pagenum << MMU_PAGESHIFT + off);
366 	length = (uint64_t)len;
367 
368 	(void) mem_scrub(pa, length);
369 }
370 
371 void
372 sync_data_memory(caddr_t va, size_t len)
373 {
374 	/* Call memory sync function */
375 	(void) mem_sync(va, len);
376 }
377 
378 size_t
379 mmu_get_kernel_lpsize(size_t lpsize)
380 {
381 	extern int mmu_exported_pagesize_mask;
382 	uint_t tte;
383 
384 	if (lpsize == 0) {
385 		/* no setting for segkmem_lpsize in /etc/system: use default */
386 		if (mmu_exported_pagesize_mask & (1 << TTE256M)) {
387 			lpsize = MMU_PAGESIZE256M;
388 		} else if (mmu_exported_pagesize_mask & (1 << TTE4M)) {
389 			lpsize = MMU_PAGESIZE4M;
390 		} else if (mmu_exported_pagesize_mask & (1 << TTE64K)) {
391 			lpsize = MMU_PAGESIZE64K;
392 		} else {
393 			lpsize = MMU_PAGESIZE;
394 		}
395 
396 		return (lpsize);
397 	}
398 
399 	for (tte = TTE8K; tte <= TTE256M; tte++) {
400 
401 		if ((mmu_exported_pagesize_mask & (1 << tte)) == 0)
402 			continue;
403 
404 		if (lpsize == TTEBYTES(tte))
405 			return (lpsize);
406 	}
407 
408 	lpsize = TTEBYTES(TTE8K);
409 	return (lpsize);
410 }
411 
412 void
413 mmu_init_kcontext()
414 {
415 }
416 
417 /*ARGSUSED*/
418 void
419 mmu_init_kernel_pgsz(struct hat *hat)
420 {
421 }
422 
423 static void *
424 contig_mem_span_alloc(vmem_t *vmp, size_t size, int vmflag)
425 {
426 	page_t *ppl;
427 	page_t *rootpp;
428 	caddr_t addr = NULL;
429 	pgcnt_t npages = btopr(size);
430 	page_t **ppa;
431 	int pgflags;
432 	spgcnt_t i = 0;
433 
434 
435 	ASSERT(size <= contig_mem_import_size_max);
436 	ASSERT((size & (size - 1)) == 0);
437 
438 	if ((addr = vmem_xalloc(vmp, size, size, 0, 0,
439 	    NULL, NULL, vmflag)) == NULL) {
440 		return (NULL);
441 	}
442 
443 	/* The address should be slab-size aligned. */
444 	ASSERT(((uintptr_t)addr & (size - 1)) == 0);
445 
446 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
447 		vmem_xfree(vmp, addr, size);
448 		return (NULL);
449 	}
450 
451 	pgflags = PG_EXCL;
452 	if (vmflag & VM_NORELOC)
453 		pgflags |= PG_NORELOC;
454 
455 	ppl = page_create_va_large(&kvp, (u_offset_t)(uintptr_t)addr, size,
456 	    pgflags, &kvseg, addr, NULL);
457 
458 	if (ppl == NULL) {
459 		vmem_xfree(vmp, addr, size);
460 		page_unresv(npages);
461 		return (NULL);
462 	}
463 
464 	rootpp = ppl;
465 	ppa = kmem_zalloc(npages * sizeof (page_t *), KM_SLEEP);
466 	while (ppl != NULL) {
467 		page_t *pp = ppl;
468 		ppa[i++] = pp;
469 		page_sub(&ppl, pp);
470 		ASSERT(page_iolock_assert(pp));
471 		ASSERT(PAGE_EXCL(pp));
472 		page_io_unlock(pp);
473 	}
474 
475 	/*
476 	 * Load the locked entry.  It's OK to preload the entry into
477 	 * the TSB since we now support large mappings in the kernel TSB.
478 	 */
479 	hat_memload_array(kas.a_hat, (caddr_t)rootpp->p_offset, size,
480 	    ppa, (PROT_ALL & ~PROT_USER) | HAT_NOSYNC, HAT_LOAD_LOCK);
481 
482 	ASSERT(i == page_get_pagecnt(ppa[0]->p_szc));
483 	for (--i; i >= 0; --i) {
484 		ASSERT(ppa[i]->p_szc == ppa[0]->p_szc);
485 		ASSERT(page_pptonum(ppa[i]) == page_pptonum(ppa[0]) + i);
486 		(void) page_pp_lock(ppa[i], 0, 1);
487 		/*
488 		 * Leave the page share locked. For non-cage pages,
489 		 * this would prevent memory DR if it were supported
490 		 * on sun4v.
491 		 */
492 		page_downgrade(ppa[i]);
493 	}
494 
495 	kmem_free(ppa, npages * sizeof (page_t *));
496 	return (addr);
497 }
498 
499 /*
500  * Allocates a slab by first trying to use the largest slab size
501  * in contig_mem_import_sizes and then falling back to smaller slab
502  * sizes still large enough for the allocation. The sizep argument
503  * is a pointer to the requested size. When a slab is successfully
504  * allocated, the slab size, which must be >= *sizep and <=
505  * contig_mem_import_size_max, is returned in the *sizep argument.
506  * Returns the virtual address of the new slab.
507  */
508 static void *
509 span_alloc_downsize(vmem_t *vmp, size_t *sizep, size_t align, int vmflag)
510 {
511 	int i;
512 
513 	ASSERT(*sizep <= contig_mem_import_size_max);
514 
515 	for (i = 0; i < NUM_IMPORT_SIZES; i++) {
516 		size_t page_size = contig_mem_import_sizes[i];
517 
518 		/*
519 		 * Check that the alignment is also less than the
520 		 * import (large page) size. In the case where the
521 		 * alignment is larger than the size, a large page
522 		 * large enough for the allocation is not necessarily
523 		 * physical-address aligned to satisfy the requested
524 		 * alignment. Since alignment is required to be a
525 		 * power-of-2, any large page >= size && >= align will
526 		 * suffice.
527 		 */
528 		if (*sizep <= page_size && align <= page_size) {
529 			void *addr;
530 			addr = contig_mem_span_alloc(vmp, page_size, vmflag);
531 			if (addr == NULL)
532 				continue;
533 			*sizep = page_size;
534 			return (addr);
535 		}
536 		return (NULL);
537 	}
538 
539 	return (NULL);
540 }
541 
542 static void *
543 contig_mem_span_xalloc(vmem_t *vmp, size_t *sizep, size_t align, int vmflag)
544 {
545 	return (span_alloc_downsize(vmp, sizep, align, vmflag | VM_NORELOC));
546 }
547 
548 static void *
549 contig_mem_reloc_span_xalloc(vmem_t *vmp, size_t *sizep, size_t align,
550     int vmflag)
551 {
552 	ASSERT((vmflag & VM_NORELOC) == 0);
553 	return (span_alloc_downsize(vmp, sizep, align, vmflag));
554 }
555 
556 /*
557  * Free a span, which is always exactly one large page.
558  */
559 static void
560 contig_mem_span_free(vmem_t *vmp, void *inaddr, size_t size)
561 {
562 	page_t *pp;
563 	caddr_t addr = inaddr;
564 	caddr_t eaddr;
565 	pgcnt_t npages = btopr(size);
566 	page_t *rootpp = NULL;
567 
568 	ASSERT(size <= contig_mem_import_size_max);
569 	/* All slabs should be size aligned */
570 	ASSERT(((uintptr_t)addr & (size - 1)) == 0);
571 
572 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
573 
574 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
575 		pp = page_find(&kvp, (u_offset_t)(uintptr_t)addr);
576 		if (pp == NULL) {
577 			panic("contig_mem_span_free: page not found");
578 		}
579 		if (!page_tryupgrade(pp)) {
580 			page_unlock(pp);
581 			pp = page_lookup(&kvp,
582 			    (u_offset_t)(uintptr_t)addr, SE_EXCL);
583 			if (pp == NULL)
584 				panic("contig_mem_span_free: page not found");
585 		}
586 
587 		ASSERT(PAGE_EXCL(pp));
588 		ASSERT(size == page_get_pagesize(pp->p_szc));
589 		ASSERT(rootpp == NULL || rootpp->p_szc == pp->p_szc);
590 		ASSERT(rootpp == NULL || (page_pptonum(rootpp) +
591 		    (pgcnt_t)btop(addr - (caddr_t)inaddr) == page_pptonum(pp)));
592 
593 		page_pp_unlock(pp, 0, 1);
594 
595 		if (rootpp == NULL)
596 			rootpp = pp;
597 	}
598 	page_destroy_pages(rootpp);
599 	page_unresv(npages);
600 
601 	if (vmp != NULL)
602 		vmem_xfree(vmp, inaddr, size);
603 }
604 
605 static void *
606 contig_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t *sizep, size_t align,
607     int vmflag)
608 {
609 	ASSERT((align & (align - 1)) == 0);
610 	return (vmem_xalloc(vmp, *sizep, align, 0, 0, NULL, NULL, vmflag));
611 }
612 
613 /*
614  * contig_mem_alloc, contig_mem_alloc_align
615  *
616  * Caution: contig_mem_alloc and contig_mem_alloc_align should be
617  * used only when physically contiguous non-relocatable memory is
618  * required. Furthermore, use of these allocation routines should be
619  * minimized as well as should the allocation size. As described in the
620  * contig_mem_arena comment block above, slab allocations fall back to
621  * being outside of the cage. Therefore, overuse of these allocation
622  * routines can lead to non-relocatable large pages being allocated
623  * outside the cage. Such pages prevent the allocation of a larger page
624  * occupying overlapping pages. This can impact performance for
625  * applications that utilize e.g. 256M large pages.
626  */
627 
628 /*
629  * Allocates size aligned contiguous memory up to contig_mem_import_size_max.
630  * Size must be a power of 2.
631  */
632 void *
633 contig_mem_alloc(size_t size)
634 {
635 	ASSERT((size & (size - 1)) == 0);
636 	return (contig_mem_alloc_align(size, size));
637 }
638 
639 /*
640  * contig_mem_alloc_align allocates real contiguous memory with the
641  * specified alignment up to contig_mem_import_size_max. The alignment must
642  * be a power of 2 and no greater than contig_mem_import_size_max. We assert
643  * the aligment is a power of 2. For non-debug, vmem_xalloc will panic
644  * for non power of 2 alignments.
645  */
646 void *
647 contig_mem_alloc_align(size_t size, size_t align)
648 {
649 	void *buf;
650 
651 	ASSERT(size <= contig_mem_import_size_max);
652 	ASSERT(align <= contig_mem_import_size_max);
653 	ASSERT((align & (align - 1)) == 0);
654 
655 	if (align < CONTIG_MEM_ARENA_QUANTUM)
656 		align = CONTIG_MEM_ARENA_QUANTUM;
657 
658 	/*
659 	 * We take the lock here to serialize span allocations.
660 	 * We do not lose concurrency for the common case, since
661 	 * allocations that don't require new span allocations
662 	 * are serialized by vmem_xalloc. Serializing span
663 	 * allocations also prevents us from trying to allocate
664 	 * more spans than necessary.
665 	 */
666 	mutex_enter(&contig_mem_lock);
667 
668 	buf = vmem_xalloc(contig_mem_arena, size, align, 0, 0,
669 	    NULL, NULL, VM_NOSLEEP | VM_NORELOC);
670 
671 	if ((buf == NULL) && (size <= MMU_PAGESIZE)) {
672 		mutex_exit(&contig_mem_lock);
673 		return (vmem_xalloc(static_alloc_arena, size, align, 0, 0,
674 		    NULL, NULL, VM_NOSLEEP));
675 	}
676 
677 	if (buf == NULL) {
678 		buf = vmem_xalloc(contig_mem_reloc_arena, size, align, 0, 0,
679 		    NULL, NULL, VM_NOSLEEP);
680 	}
681 
682 	mutex_exit(&contig_mem_lock);
683 
684 	return (buf);
685 }
686 
687 void
688 contig_mem_free(void *vaddr, size_t size)
689 {
690 	if (vmem_contains(contig_mem_arena, vaddr, size)) {
691 		vmem_xfree(contig_mem_arena, vaddr, size);
692 	} else if (size > MMU_PAGESIZE) {
693 		vmem_xfree(contig_mem_reloc_arena, vaddr, size);
694 	} else {
695 		vmem_xfree(static_alloc_arena, vaddr, size);
696 	}
697 }
698 
699 /*
700  * We create a set of stacked vmem arenas to enable us to
701  * allocate large >PAGESIZE chucks of contiguous Real Address space.
702  * The vmem_xcreate interface is used to create the contig_mem_arena
703  * allowing the import routine to downsize the requested slab size
704  * and return a smaller slab.
705  */
706 void
707 contig_mem_init(void)
708 {
709 	mutex_init(&contig_mem_lock, NULL, MUTEX_DEFAULT, NULL);
710 
711 	contig_mem_slab_arena = vmem_xcreate("contig_mem_slab_arena", NULL, 0,
712 	    CONTIG_MEM_SLAB_ARENA_QUANTUM, contig_vmem_xalloc_aligned_wrapper,
713 	    vmem_xfree, heap_arena, 0, VM_SLEEP | VMC_XALIGN);
714 
715 	contig_mem_arena = vmem_xcreate("contig_mem_arena", NULL, 0,
716 	    CONTIG_MEM_ARENA_QUANTUM, contig_mem_span_xalloc,
717 	    contig_mem_span_free, contig_mem_slab_arena, 0,
718 	    VM_SLEEP | VM_BESTFIT | VMC_XALIGN);
719 
720 	contig_mem_reloc_arena = vmem_xcreate("contig_mem_reloc_arena", NULL, 0,
721 	    CONTIG_MEM_ARENA_QUANTUM, contig_mem_reloc_span_xalloc,
722 	    contig_mem_span_free, contig_mem_slab_arena, 0,
723 	    VM_SLEEP | VM_BESTFIT | VMC_XALIGN);
724 
725 	if (contig_mem_prealloc_buf == NULL || vmem_add(contig_mem_arena,
726 	    contig_mem_prealloc_buf, contig_mem_prealloc_size, VM_SLEEP)
727 	    == NULL) {
728 		cmn_err(CE_WARN, "Failed to pre-populate contig_mem_arena");
729 	}
730 }
731 
732 /*
733  * In calculating how much memory to pre-allocate, we include a small
734  * amount per-CPU to account for per-CPU buffers in line with measured
735  * values for different size systems. contig_mem_prealloc_base_size is
736  * a cpu specific amount to be pre-allocated before considering per-CPU
737  * requirements and memory size. We always pre-allocate a minimum amount
738  * of memory determined by PREALLOC_MIN. Beyond that, we take the minimum
739  * of contig_mem_prealloc_base_size and a small percentage of physical
740  * memory to prevent allocating too much on smaller systems.
741  * contig_mem_prealloc_base_size is global, allowing for the CPU module
742  * to increase its value if necessary.
743  */
744 #define	PREALLOC_PER_CPU	(256 * 1024)		/* 256K */
745 #define	PREALLOC_PERCENT	(4)			/* 4% */
746 #define	PREALLOC_MIN		(16 * 1024 * 1024)	/* 16M */
747 size_t contig_mem_prealloc_base_size = 0;
748 
749 /*
750  * Called at boot-time allowing pre-allocation of contiguous memory.
751  * The argument 'alloc_base' is the requested base address for the
752  * allocation and originates in startup_memlist.
753  */
754 caddr_t
755 contig_mem_prealloc(caddr_t alloc_base, pgcnt_t npages)
756 {
757 	caddr_t	chunkp;
758 
759 	contig_mem_prealloc_size = MIN((PREALLOC_PER_CPU * ncpu_guest_max) +
760 	    contig_mem_prealloc_base_size,
761 	    (ptob(npages) * PREALLOC_PERCENT) / 100);
762 	contig_mem_prealloc_size = MAX(contig_mem_prealloc_size, PREALLOC_MIN);
763 	contig_mem_prealloc_size = P2ROUNDUP(contig_mem_prealloc_size,
764 	    MMU_PAGESIZE4M);
765 
766 	alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, MMU_PAGESIZE4M);
767 	if (prom_alloc(alloc_base, contig_mem_prealloc_size,
768 	    MMU_PAGESIZE4M) != alloc_base) {
769 
770 		/*
771 		 * Failed.  This may mean the physical memory has holes in it
772 		 * and it will be more difficult to get large contiguous
773 		 * pieces of memory.  Since we only guarantee contiguous
774 		 * pieces of memory contig_mem_import_size_max or smaller,
775 		 * loop, getting contig_mem_import_size_max at a time, until
776 		 * failure or contig_mem_prealloc_size is reached.
777 		 */
778 		for (chunkp = alloc_base;
779 		    (chunkp - alloc_base) < contig_mem_prealloc_size;
780 		    chunkp += contig_mem_import_size_max) {
781 
782 			if (prom_alloc(chunkp, contig_mem_import_size_max,
783 			    MMU_PAGESIZE4M) != chunkp) {
784 				break;
785 			}
786 		}
787 		contig_mem_prealloc_size = chunkp - alloc_base;
788 		ASSERT(contig_mem_prealloc_size != 0);
789 	}
790 
791 	if (contig_mem_prealloc_size != 0) {
792 		contig_mem_prealloc_buf = alloc_base;
793 	} else {
794 		contig_mem_prealloc_buf = NULL;
795 	}
796 	alloc_base += contig_mem_prealloc_size;
797 
798 	return (alloc_base);
799 }
800