xref: /titanic_51/usr/src/uts/sun4v/vm/mach_vm_dep.c (revision 986fd29a0dc13f7608ef7f508f6e700bd7bc2720)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
27 /*	All Rights Reserved   */
28 
29 /*
30  * Portions of this source code were derived from Berkeley 4.3 BSD
31  * under license from the Regents of the University of California.
32  */
33 
34 #pragma ident	"%Z%%M%	%I%	%E% SMI"
35 
36 /*
37  * UNIX machine dependent virtual memory support.
38  */
39 
40 #include <sys/vm.h>
41 #include <sys/exec.h>
42 #include <sys/cmn_err.h>
43 #include <sys/cpu_module.h>
44 #include <sys/cpu.h>
45 #include <sys/elf_SPARC.h>
46 #include <sys/archsystm.h>
47 #include <vm/hat_sfmmu.h>
48 #include <sys/memnode.h>
49 #include <sys/mem_cage.h>
50 #include <vm/vm_dep.h>
51 #include <sys/error.h>
52 #include <sys/machsystm.h>
53 #include <vm/seg_kmem.h>
54 #include <sys/stack.h>
55 #include <sys/atomic.h>
56 #include <sys/promif.h>
57 
58 uint_t page_colors = 0;
59 uint_t page_colors_mask = 0;
60 uint_t page_coloring_shift = 0;
61 int consistent_coloring;
62 int update_proc_pgcolorbase_after_fork = 1;
63 
64 uint_t mmu_page_sizes = MMU_PAGE_SIZES;
65 uint_t max_mmu_page_sizes = MMU_PAGE_SIZES;
66 uint_t mmu_hashcnt = MAX_HASHCNT;
67 uint_t max_mmu_hashcnt = MAX_HASHCNT;
68 size_t mmu_ism_pagesize = DEFAULT_ISM_PAGESIZE;
69 
70 /*
71  * A bitmask of the page sizes supported by hardware based upon szc.
72  * The base pagesize (p_szc == 0) must always be supported by the hardware.
73  */
74 int mmu_exported_pagesize_mask;
75 uint_t mmu_exported_page_sizes;
76 
77 uint_t szc_2_userszc[MMU_PAGE_SIZES];
78 uint_t userszc_2_szc[MMU_PAGE_SIZES];
79 
80 extern uint_t vac_colors_mask;
81 extern int vac_shift;
82 
83 hw_pagesize_t hw_page_array[] = {
84 	{MMU_PAGESIZE, MMU_PAGESHIFT, 0, MMU_PAGESIZE >> MMU_PAGESHIFT},
85 	{MMU_PAGESIZE64K, MMU_PAGESHIFT64K, 0,
86 	    MMU_PAGESIZE64K >> MMU_PAGESHIFT},
87 	{MMU_PAGESIZE512K, MMU_PAGESHIFT512K, 0,
88 	    MMU_PAGESIZE512K >> MMU_PAGESHIFT},
89 	{MMU_PAGESIZE4M, MMU_PAGESHIFT4M, 0, MMU_PAGESIZE4M >> MMU_PAGESHIFT},
90 	{MMU_PAGESIZE32M, MMU_PAGESHIFT32M, 0,
91 	    MMU_PAGESIZE32M >> MMU_PAGESHIFT},
92 	{MMU_PAGESIZE256M, MMU_PAGESHIFT256M, 0,
93 	    MMU_PAGESIZE256M >> MMU_PAGESHIFT},
94 	{0, 0, 0, 0}
95 };
96 
97 /*
98  * Maximum page size used to map 64-bit memory segment kmem64_base..kmem64_end
99  */
100 int	max_bootlp_tteszc = TTE256M;
101 
102 /*
103  * Maximum and default segment size tunables for user heap, stack, private
104  * and shared anonymous memory, and user text and initialized data.
105  */
106 size_t max_uheap_lpsize = MMU_PAGESIZE64K;
107 size_t default_uheap_lpsize = MMU_PAGESIZE64K;
108 size_t max_ustack_lpsize = MMU_PAGESIZE64K;
109 size_t default_ustack_lpsize = MMU_PAGESIZE64K;
110 size_t max_privmap_lpsize = MMU_PAGESIZE64K;
111 size_t max_uidata_lpsize = MMU_PAGESIZE64K;
112 size_t max_utext_lpsize = MMU_PAGESIZE4M;
113 size_t max_shm_lpsize = MMU_PAGESIZE4M;
114 
115 /*
116  * Contiguous memory allocator data structures and variables.
117  *
118  * The sun4v kernel must provide a means to allocate physically
119  * contiguous, non-relocatable memory. The contig_mem_arena
120  * and contig_mem_slab_arena exist for this purpose. Allocations
121  * that require physically contiguous non-relocatable memory should
122  * be made using contig_mem_alloc() or contig_mem_alloc_align()
123  * which return memory from contig_mem_arena or contig_mem_reloc_arena.
124  * These arenas import memory from the contig_mem_slab_arena one
125  * contiguous chunk at a time.
126  *
127  * When importing slabs, an attempt is made to allocate a large page
128  * to use as backing. As a result of the non-relocatable requirement,
129  * slabs are allocated from the kernel cage freelists. If the cage does
130  * not contain any free contiguous chunks large enough to satisfy the
131  * slab allocation, the slab size will be downsized and the operation
132  * retried. Large slab sizes are tried first to minimize cage
133  * fragmentation. If the slab allocation is unsuccessful still, the slab
134  * is allocated from outside the kernel cage. This is undesirable because,
135  * until slabs are freed, it results in non-relocatable chunks scattered
136  * throughout physical memory.
137  *
138  * Allocations from the contig_mem_arena are backed by slabs from the
139  * cage. Allocations from the contig_mem_reloc_arena are backed by
140  * slabs allocated outside the cage. Slabs are left share locked while
141  * in use to prevent non-cage slabs from being relocated.
142  *
143  * Since there is no guarantee that large pages will be available in
144  * the kernel cage, contiguous memory is reserved and added to the
145  * contig_mem_arena at boot time, making it available for later
146  * contiguous memory allocations. This reserve will be used to satisfy
147  * contig_mem allocations first and it is only when the reserve is
148  * completely allocated that new slabs will need to be imported.
149  */
150 static	vmem_t		*contig_mem_slab_arena;
151 static	vmem_t		*contig_mem_arena;
152 static	vmem_t		*contig_mem_reloc_arena;
153 static	kmutex_t	contig_mem_lock;
154 #define	CONTIG_MEM_ARENA_QUANTUM	64
155 #define	CONTIG_MEM_SLAB_ARENA_QUANTUM	MMU_PAGESIZE64K
156 
157 /* contig_mem_arena import slab sizes, in decreasing size order */
158 static size_t contig_mem_import_sizes[] = {
159 	MMU_PAGESIZE4M,
160 	MMU_PAGESIZE512K,
161 	MMU_PAGESIZE64K
162 };
163 #define	NUM_IMPORT_SIZES	\
164 	(sizeof (contig_mem_import_sizes) / sizeof (size_t))
165 static size_t contig_mem_import_size_max	= MMU_PAGESIZE4M;
166 size_t contig_mem_slab_size			= MMU_PAGESIZE4M;
167 
168 /* Boot-time allocated buffer to pre-populate the contig_mem_arena */
169 static size_t contig_mem_prealloc_size;
170 static void *contig_mem_prealloc_buf;
171 
172 /*
173  * map_addr_proc() is the routine called when the system is to
174  * choose an address for the user.  We will pick an address
175  * range which is just below the current stack limit.  The
176  * algorithm used for cache consistency on machines with virtual
177  * address caches is such that offset 0 in the vnode is always
178  * on a shm_alignment'ed aligned address.  Unfortunately, this
179  * means that vnodes which are demand paged will not be mapped
180  * cache consistently with the executable images.  When the
181  * cache alignment for a given object is inconsistent, the
182  * lower level code must manage the translations so that this
183  * is not seen here (at the cost of efficiency, of course).
184  *
185  * addrp is a value/result parameter.
186  *	On input it is a hint from the user to be used in a completely
187  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
188  *	minimal alignment.
189  *
190  *	On output it is NULL if no address can be found in the current
191  *	processes address space or else an address that is currently
192  *	not mapped for len bytes with a page of red zone on either side.
193  *	If vacalign is true, then the selected address will obey the alignment
194  *	constraints of a vac machine based on the given off value.
195  */
196 /*ARGSUSED3*/
197 void
198 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
199     caddr_t userlimit, struct proc *p, uint_t flags)
200 {
201 	struct as *as = p->p_as;
202 	caddr_t addr;
203 	caddr_t base;
204 	size_t slen;
205 	uintptr_t align_amount;
206 	int allow_largepage_alignment = 1;
207 
208 	base = p->p_brkbase;
209 	if (userlimit < as->a_userlimit) {
210 		/*
211 		 * This happens when a program wants to map something in
212 		 * a range that's accessible to a program in a smaller
213 		 * address space.  For example, a 64-bit program might
214 		 * be calling mmap32(2) to guarantee that the returned
215 		 * address is below 4Gbytes.
216 		 */
217 		ASSERT(userlimit > base);
218 		slen = userlimit - base;
219 	} else {
220 		slen = p->p_usrstack - base - (((size_t)rctl_enforced_value(
221 		    rctlproc_legacy[RLIMIT_STACK], p->p_rctls, p) + PAGEOFFSET)
222 		    & PAGEMASK);
223 	}
224 	len = (len + PAGEOFFSET) & PAGEMASK;
225 
226 	/*
227 	 * Redzone for each side of the request. This is done to leave
228 	 * one page unmapped between segments. This is not required, but
229 	 * it's useful for the user because if their program strays across
230 	 * a segment boundary, it will catch a fault immediately making
231 	 * debugging a little easier.
232 	 */
233 	len += (2 * PAGESIZE);
234 
235 	/*
236 	 *  If the request is larger than the size of a particular
237 	 *  mmu level, then we use that level to map the request.
238 	 *  But this requires that both the virtual and the physical
239 	 *  addresses be aligned with respect to that level, so we
240 	 *  do the virtual bit of nastiness here.
241 	 *
242 	 *  For 32-bit processes, only those which have specified
243 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
244 	 *  we can potentially waste up to 256MB of the 4G process address
245 	 *  space just for alignment.
246 	 *
247 	 * XXXQ Should iterate trough hw_page_array here to catch
248 	 * all supported pagesizes
249 	 */
250 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
251 	    ((uintptr_t)*addrp) != 0)) {
252 		allow_largepage_alignment = 0;
253 	}
254 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
255 	    allow_largepage_alignment &&
256 	    (len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
257 		align_amount = MMU_PAGESIZE256M;
258 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
259 	    allow_largepage_alignment &&
260 	    (len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
261 		align_amount = MMU_PAGESIZE32M;
262 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
263 		align_amount = MMU_PAGESIZE4M;
264 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
265 		align_amount = MMU_PAGESIZE512K;
266 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
267 		align_amount = MMU_PAGESIZE64K;
268 	} else  {
269 		/*
270 		 * Align virtual addresses on a 64K boundary to ensure
271 		 * that ELF shared libraries are mapped with the appropriate
272 		 * alignment constraints by the run-time linker.
273 		 */
274 		align_amount = ELF_SPARC_MAXPGSZ;
275 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
276 		    ((uintptr_t)*addrp < align_amount))
277 			align_amount = (uintptr_t)*addrp;
278 	}
279 
280 	/*
281 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
282 	 */
283 	if (p->p_model == DATAMODEL_LP64)
284 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
285 #ifdef VAC
286 	if (vac && vacalign && (align_amount < shm_alignment))
287 		align_amount = shm_alignment;
288 #endif
289 
290 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
291 		align_amount = (uintptr_t)*addrp;
292 	}
293 	len += align_amount;
294 
295 	/*
296 	 * Look for a large enough hole starting below the stack limit.
297 	 * After finding it, use the upper part.  Addition of PAGESIZE is
298 	 * for the redzone as described above.
299 	 */
300 	as_purge(as);
301 	if (as_gap(as, len, &base, &slen, AH_HI, NULL) == 0) {
302 		caddr_t as_addr;
303 
304 		addr = base + slen - len + PAGESIZE;
305 		as_addr = addr;
306 		/*
307 		 * Round address DOWN to the alignment amount,
308 		 * add the offset, and if this address is less
309 		 * than the original address, add alignment amount.
310 		 */
311 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
312 		addr += (long)(off & (align_amount - 1l));
313 		if (addr < as_addr) {
314 			addr += align_amount;
315 		}
316 
317 		ASSERT(addr <= (as_addr + align_amount));
318 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
319 		    ((uintptr_t)(off & (align_amount - 1l))));
320 		*addrp = addr;
321 
322 	} else {
323 		*addrp = NULL;	/* no more virtual space */
324 	}
325 }
326 
327 /*
328  * Platform-dependent page scrub call.
329  * We call hypervisor to scrub the page.
330  */
331 void
332 pagescrub(page_t *pp, uint_t off, uint_t len)
333 {
334 	uint64_t pa, length;
335 
336 	pa = (uint64_t)(pp->p_pagenum << MMU_PAGESHIFT + off);
337 	length = (uint64_t)len;
338 
339 	(void) mem_scrub(pa, length);
340 }
341 
342 void
343 sync_data_memory(caddr_t va, size_t len)
344 {
345 	/* Call memory sync function */
346 	(void) mem_sync(va, len);
347 }
348 
349 size_t
350 mmu_get_kernel_lpsize(size_t lpsize)
351 {
352 	extern int mmu_exported_pagesize_mask;
353 	uint_t tte;
354 
355 	if (lpsize == 0) {
356 		/* no setting for segkmem_lpsize in /etc/system: use default */
357 		if (mmu_exported_pagesize_mask & (1 << TTE256M)) {
358 			lpsize = MMU_PAGESIZE256M;
359 		} else if (mmu_exported_pagesize_mask & (1 << TTE4M)) {
360 			lpsize = MMU_PAGESIZE4M;
361 		} else if (mmu_exported_pagesize_mask & (1 << TTE64K)) {
362 			lpsize = MMU_PAGESIZE64K;
363 		} else {
364 			lpsize = MMU_PAGESIZE;
365 		}
366 
367 		return (lpsize);
368 	}
369 
370 	for (tte = TTE8K; tte <= TTE256M; tte++) {
371 
372 		if ((mmu_exported_pagesize_mask & (1 << tte)) == 0)
373 			continue;
374 
375 		if (lpsize == TTEBYTES(tte))
376 			return (lpsize);
377 	}
378 
379 	lpsize = TTEBYTES(TTE8K);
380 	return (lpsize);
381 }
382 
383 void
384 mmu_init_kcontext()
385 {
386 }
387 
388 /*ARGSUSED*/
389 void
390 mmu_init_kernel_pgsz(struct hat *hat)
391 {
392 }
393 
394 static void *
395 contig_mem_span_alloc(vmem_t *vmp, size_t size, int vmflag)
396 {
397 	page_t *ppl;
398 	page_t *rootpp;
399 	caddr_t addr = NULL;
400 	pgcnt_t npages = btopr(size);
401 	page_t **ppa;
402 	int pgflags;
403 	spgcnt_t i = 0;
404 
405 
406 	ASSERT(size <= contig_mem_import_size_max);
407 	ASSERT((size & (size - 1)) == 0);
408 
409 	if ((addr = vmem_xalloc(vmp, size, size, 0, 0,
410 	    NULL, NULL, vmflag)) == NULL) {
411 		return (NULL);
412 	}
413 
414 	/* The address should be slab-size aligned. */
415 	ASSERT(((uintptr_t)addr & (size - 1)) == 0);
416 
417 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
418 		vmem_xfree(vmp, addr, size);
419 		return (NULL);
420 	}
421 
422 	pgflags = PG_EXCL;
423 	if (vmflag & VM_NORELOC)
424 		pgflags |= PG_NORELOC;
425 
426 	ppl = page_create_va_large(&kvp, (u_offset_t)(uintptr_t)addr, size,
427 	    pgflags, &kvseg, addr, NULL);
428 
429 	if (ppl == NULL) {
430 		vmem_xfree(vmp, addr, size);
431 		page_unresv(npages);
432 		return (NULL);
433 	}
434 
435 	rootpp = ppl;
436 	ppa = kmem_zalloc(npages * sizeof (page_t *), KM_SLEEP);
437 	while (ppl != NULL) {
438 		page_t *pp = ppl;
439 		ppa[i++] = pp;
440 		page_sub(&ppl, pp);
441 		ASSERT(page_iolock_assert(pp));
442 		ASSERT(PAGE_EXCL(pp));
443 		page_io_unlock(pp);
444 	}
445 
446 	/*
447 	 * Load the locked entry.  It's OK to preload the entry into
448 	 * the TSB since we now support large mappings in the kernel TSB.
449 	 */
450 	hat_memload_array(kas.a_hat, (caddr_t)rootpp->p_offset, size,
451 	    ppa, (PROT_ALL & ~PROT_USER) | HAT_NOSYNC, HAT_LOAD_LOCK);
452 
453 	ASSERT(i == page_get_pagecnt(ppa[0]->p_szc));
454 	for (--i; i >= 0; --i) {
455 		ASSERT(ppa[i]->p_szc == ppa[0]->p_szc);
456 		ASSERT(page_pptonum(ppa[i]) == page_pptonum(ppa[0]) + i);
457 		(void) page_pp_lock(ppa[i], 0, 1);
458 		/*
459 		 * Leave the page share locked. For non-cage pages,
460 		 * this would prevent memory DR if it were supported
461 		 * on sun4v.
462 		 */
463 		page_downgrade(ppa[i]);
464 	}
465 
466 	kmem_free(ppa, npages * sizeof (page_t *));
467 	return (addr);
468 }
469 
470 /*
471  * Allocates a slab by first trying to use the largest slab size
472  * in contig_mem_import_sizes and then falling back to smaller slab
473  * sizes still large enough for the allocation. The sizep argument
474  * is a pointer to the requested size. When a slab is successfully
475  * allocated, the slab size, which must be >= *sizep and <=
476  * contig_mem_import_size_max, is returned in the *sizep argument.
477  * Returns the virtual address of the new slab.
478  */
479 static void *
480 span_alloc_downsize(vmem_t *vmp, size_t *sizep, size_t align, int vmflag)
481 {
482 	int i;
483 
484 	ASSERT(*sizep <= contig_mem_import_size_max);
485 
486 	for (i = 0; i < NUM_IMPORT_SIZES; i++) {
487 		size_t page_size = contig_mem_import_sizes[i];
488 
489 		/*
490 		 * Check that the alignment is also less than the
491 		 * import (large page) size. In the case where the
492 		 * alignment is larger than the size, a large page
493 		 * large enough for the allocation is not necessarily
494 		 * physical-address aligned to satisfy the requested
495 		 * alignment. Since alignment is required to be a
496 		 * power-of-2, any large page >= size && >= align will
497 		 * suffice.
498 		 */
499 		if (*sizep <= page_size && align <= page_size) {
500 			void *addr;
501 			addr = contig_mem_span_alloc(vmp, page_size, vmflag);
502 			if (addr == NULL)
503 				continue;
504 			*sizep = page_size;
505 			return (addr);
506 		}
507 		return (NULL);
508 	}
509 
510 	return (NULL);
511 }
512 
513 static void *
514 contig_mem_span_xalloc(vmem_t *vmp, size_t *sizep, size_t align, int vmflag)
515 {
516 	return (span_alloc_downsize(vmp, sizep, align, vmflag | VM_NORELOC));
517 }
518 
519 static void *
520 contig_mem_reloc_span_xalloc(vmem_t *vmp, size_t *sizep, size_t align,
521     int vmflag)
522 {
523 	ASSERT((vmflag & VM_NORELOC) == 0);
524 	return (span_alloc_downsize(vmp, sizep, align, vmflag));
525 }
526 
527 /*
528  * Free a span, which is always exactly one large page.
529  */
530 static void
531 contig_mem_span_free(vmem_t *vmp, void *inaddr, size_t size)
532 {
533 	page_t *pp;
534 	caddr_t addr = inaddr;
535 	caddr_t eaddr;
536 	pgcnt_t npages = btopr(size);
537 	page_t *rootpp = NULL;
538 
539 	ASSERT(size <= contig_mem_import_size_max);
540 	/* All slabs should be size aligned */
541 	ASSERT(((uintptr_t)addr & (size - 1)) == 0);
542 
543 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
544 
545 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
546 		pp = page_find(&kvp, (u_offset_t)(uintptr_t)addr);
547 		if (pp == NULL) {
548 			panic("contig_mem_span_free: page not found");
549 		}
550 		if (!page_tryupgrade(pp)) {
551 			page_unlock(pp);
552 			pp = page_lookup(&kvp,
553 			    (u_offset_t)(uintptr_t)addr, SE_EXCL);
554 			if (pp == NULL)
555 				panic("contig_mem_span_free: page not found");
556 		}
557 
558 		ASSERT(PAGE_EXCL(pp));
559 		ASSERT(size == page_get_pagesize(pp->p_szc));
560 		ASSERT(rootpp == NULL || rootpp->p_szc == pp->p_szc);
561 		ASSERT(rootpp == NULL || (page_pptonum(rootpp) +
562 		    (pgcnt_t)btop(addr - (caddr_t)inaddr) == page_pptonum(pp)));
563 
564 		page_pp_unlock(pp, 0, 1);
565 
566 		if (rootpp == NULL)
567 			rootpp = pp;
568 	}
569 	page_destroy_pages(rootpp);
570 	page_unresv(npages);
571 
572 	if (vmp != NULL)
573 		vmem_xfree(vmp, inaddr, size);
574 }
575 
576 static void *
577 contig_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t *sizep, size_t align,
578     int vmflag)
579 {
580 	ASSERT((align & (align - 1)) == 0);
581 	return (vmem_xalloc(vmp, *sizep, align, 0, 0, NULL, NULL, vmflag));
582 }
583 
584 /*
585  * contig_mem_alloc, contig_mem_alloc_align
586  *
587  * Caution: contig_mem_alloc and contig_mem_alloc_align should be
588  * used only when physically contiguous non-relocatable memory is
589  * required. Furthermore, use of these allocation routines should be
590  * minimized as well as should the allocation size. As described in the
591  * contig_mem_arena comment block above, slab allocations fall back to
592  * being outside of the cage. Therefore, overuse of these allocation
593  * routines can lead to non-relocatable large pages being allocated
594  * outside the cage. Such pages prevent the allocation of a larger page
595  * occupying overlapping pages. This can impact performance for
596  * applications that utilize e.g. 256M large pages.
597  */
598 
599 /*
600  * Allocates size aligned contiguous memory up to contig_mem_import_size_max.
601  * Size must be a power of 2.
602  */
603 void *
604 contig_mem_alloc(size_t size)
605 {
606 	ASSERT((size & (size - 1)) == 0);
607 	return (contig_mem_alloc_align(size, size));
608 }
609 
610 /*
611  * contig_mem_alloc_align allocates real contiguous memory with the specified
612  * alignment up to contig_mem_import_size_max. The alignment must be a
613  * power of 2 and no greater than contig_mem_import_size_max. We assert
614  * the aligment is a power of 2. For non-debug, vmem_xalloc will panic
615  * for non power of 2 alignments.
616  */
617 void *
618 contig_mem_alloc_align(size_t size, size_t align)
619 {
620 	void *buf;
621 
622 	ASSERT(size <= contig_mem_import_size_max);
623 	ASSERT(align <= contig_mem_import_size_max);
624 	ASSERT((align & (align - 1)) == 0);
625 
626 	if (align < CONTIG_MEM_ARENA_QUANTUM)
627 		align = CONTIG_MEM_ARENA_QUANTUM;
628 
629 	/*
630 	 * We take the lock here to serialize span allocations.
631 	 * We do not lose concurrency for the common case, since
632 	 * allocations that don't require new span allocations
633 	 * are serialized by vmem_xalloc. Serializing span
634 	 * allocations also prevents us from trying to allocate
635 	 * more spans that necessary.
636 	 */
637 	mutex_enter(&contig_mem_lock);
638 
639 	buf = vmem_xalloc(contig_mem_arena, size, align, 0, 0,
640 	    NULL, NULL, VM_NOSLEEP | VM_NORELOC);
641 
642 	if ((buf == NULL) && (size <= MMU_PAGESIZE)) {
643 		mutex_exit(&contig_mem_lock);
644 		return (vmem_xalloc(static_alloc_arena, size, align, 0, 0,
645 		    NULL, NULL, VM_NOSLEEP));
646 	}
647 
648 	if (buf == NULL) {
649 		buf = vmem_xalloc(contig_mem_reloc_arena, size, align, 0, 0,
650 		    NULL, NULL, VM_NOSLEEP);
651 	}
652 
653 	mutex_exit(&contig_mem_lock);
654 
655 	return (buf);
656 }
657 
658 void
659 contig_mem_free(void *vaddr, size_t size)
660 {
661 	if (vmem_contains(contig_mem_arena, vaddr, size)) {
662 		vmem_xfree(contig_mem_arena, vaddr, size);
663 	} else if (size > MMU_PAGESIZE) {
664 		vmem_xfree(contig_mem_reloc_arena, vaddr, size);
665 	} else {
666 		vmem_xfree(static_alloc_arena, vaddr, size);
667 	}
668 }
669 
670 /*
671  * We create a set of stacked vmem arenas to enable us to
672  * allocate large >PAGESIZE chucks of contiguous Real Address space.
673  * The vmem_xcreate interface is used to create the contig_mem_arena
674  * allowing the import routine to downsize the requested slab size
675  * and return a smaller slab.
676  */
677 void
678 contig_mem_init(void)
679 {
680 	mutex_init(&contig_mem_lock, NULL, MUTEX_DEFAULT, NULL);
681 
682 	contig_mem_slab_arena = vmem_xcreate("contig_mem_slab_arena", NULL, 0,
683 	    CONTIG_MEM_SLAB_ARENA_QUANTUM, contig_vmem_xalloc_aligned_wrapper,
684 	    vmem_xfree, heap_arena, 0, VM_SLEEP | VMC_XALIGN);
685 
686 	contig_mem_arena = vmem_xcreate("contig_mem_arena", NULL, 0,
687 	    CONTIG_MEM_ARENA_QUANTUM, contig_mem_span_xalloc,
688 	    contig_mem_span_free, contig_mem_slab_arena, 0,
689 	    VM_SLEEP | VM_BESTFIT | VMC_XALIGN);
690 
691 	contig_mem_reloc_arena = vmem_xcreate("contig_mem_reloc_arena", NULL, 0,
692 	    CONTIG_MEM_ARENA_QUANTUM, contig_mem_reloc_span_xalloc,
693 	    contig_mem_span_free, contig_mem_slab_arena, 0,
694 	    VM_SLEEP | VM_BESTFIT | VMC_XALIGN);
695 
696 	if (vmem_add(contig_mem_arena, contig_mem_prealloc_buf,
697 	    contig_mem_prealloc_size, VM_SLEEP) == NULL)
698 		cmn_err(CE_PANIC, "Failed to pre-populate contig_mem_arena");
699 }
700 
701 /*
702  * In calculating how much memory to pre-allocate, we include a small
703  * amount per-CPU to account for per-CPU buffers in line with measured
704  * values for different size systems. contig_mem_prealloc_base_size is
705  * a cpu specific amount to be pre-allocated before considering per-CPU
706  * requirements and memory size. We always pre-allocate a minimum amount
707  * of memory determined by PREALLOC_MIN. Beyond that, we take the minimum
708  * of contig_mem_prealloc_base_size and a small percentage of physical
709  * memory to prevent allocating too much on smaller systems.
710  * contig_mem_prealloc_base_size is global, allowing for the CPU module
711  * to increase its value if necessary.
712  */
713 #define	PREALLOC_PER_CPU	(256 * 1024)		/* 256K */
714 #define	PREALLOC_PERCENT	(4)			/* 4% */
715 #define	PREALLOC_MIN		(16 * 1024 * 1024)	/* 16M */
716 size_t contig_mem_prealloc_base_size = 0;
717 
718 /*
719  * Called at boot-time allowing pre-allocation of contiguous memory.
720  * The argument 'alloc_base' is the requested base address for the
721  * allocation and originates in startup_memlist.
722  */
723 caddr_t
724 contig_mem_prealloc(caddr_t alloc_base, pgcnt_t npages)
725 {
726 	contig_mem_prealloc_size = MIN((PREALLOC_PER_CPU * ncpu_guest_max) +
727 	    contig_mem_prealloc_base_size,
728 	    (ptob(npages) * PREALLOC_PERCENT) / 100);
729 	contig_mem_prealloc_size = MAX(contig_mem_prealloc_size, PREALLOC_MIN);
730 	contig_mem_prealloc_size = P2ROUNDUP(contig_mem_prealloc_size,
731 	    MMU_PAGESIZE4M);
732 
733 	alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, MMU_PAGESIZE4M);
734 	if (prom_alloc(alloc_base, contig_mem_prealloc_size,
735 	    MMU_PAGESIZE4M) != alloc_base)
736 		prom_panic("can't allocate contig mem");
737 
738 	contig_mem_prealloc_buf = alloc_base;
739 	alloc_base += contig_mem_prealloc_size;
740 
741 	return (alloc_base);
742 }
743 
744 static uint_t sp_color_stride = 16;
745 static uint_t sp_color_mask = 0x1f;
746 static uint_t sp_current_color = (uint_t)-1;
747 
748 size_t
749 exec_get_spslew(void)
750 {
751 	uint_t spcolor = atomic_inc_32_nv(&sp_current_color);
752 	return ((size_t)((spcolor & sp_color_mask) * SA(sp_color_stride)));
753 }
754