xref: /illumos-gate/usr/src/uts/sun4v/vm/mach_vm_dep.c (revision 8b80e8cb6855118d46f605e91b5ed4ce83417395)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /* 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/hsvc.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  * map_addr_proc() is the routine called when the system is to
173  * choose an address for the user.  We will pick an address
174  * range which is just below the current stack limit.  The
175  * algorithm used for cache consistency on machines with virtual
176  * address caches is such that offset 0 in the vnode is always
177  * on a shm_alignment'ed aligned address.  Unfortunately, this
178  * means that vnodes which are demand paged will not be mapped
179  * cache consistently with the executable images.  When the
180  * cache alignment for a given object is inconsistent, the
181  * lower level code must manage the translations so that this
182  * is not seen here (at the cost of efficiency, of course).
183  *
184  * Every mapping will have a redzone of a single page on either side of
185  * the request. This is done to leave one page unmapped between segments.
186  * This is not required, but it's useful for the user because if their
187  * program strays across a segment boundary, it will catch a fault
188  * immediately making debugging a little easier.  Currently the redzone
189  * is mandatory.
190  *
191  * addrp is a value/result parameter.
192  *	On input it is a hint from the user to be used in a completely
193  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
194  *	minimal alignment, which must be some "power of two" multiple of
195  *	pagesize.
196  *
197  *	On output it is NULL if no address can be found in the current
198  *	processes address space or else an address that is currently
199  *	not mapped for len bytes with a page of red zone on either side.
200  *	If vacalign is true, then the selected address will obey the alignment
201  *	constraints of a vac machine based on the given off value.
202  */
203 /*ARGSUSED3*/
204 void
205 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
206     caddr_t userlimit, struct proc *p, uint_t flags)
207 {
208 	struct as *as = p->p_as;
209 	caddr_t addr;
210 	caddr_t base;
211 	size_t slen;
212 	uintptr_t align_amount;
213 	int allow_largepage_alignment = 1;
214 
215 	base = p->p_brkbase;
216 	if (userlimit < as->a_userlimit) {
217 		/*
218 		 * This happens when a program wants to map something in
219 		 * a range that's accessible to a program in a smaller
220 		 * address space.  For example, a 64-bit program might
221 		 * be calling mmap32(2) to guarantee that the returned
222 		 * address is below 4Gbytes.
223 		 */
224 		ASSERT(userlimit > base);
225 		slen = userlimit - base;
226 	} else {
227 		slen = p->p_usrstack - base - (((size_t)rctl_enforced_value(
228 		    rctlproc_legacy[RLIMIT_STACK], p->p_rctls, p) + PAGEOFFSET)
229 		    & PAGEMASK);
230 	}
231 	/* Make len be a multiple of PAGESIZE */
232 	len = (len + PAGEOFFSET) & PAGEMASK;
233 
234 	/*
235 	 *  If the request is larger than the size of a particular
236 	 *  mmu level, then we use that level to map the request.
237 	 *  But this requires that both the virtual and the physical
238 	 *  addresses be aligned with respect to that level, so we
239 	 *  do the virtual bit of nastiness here.
240 	 *
241 	 *  For 32-bit processes, only those which have specified
242 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
243 	 *  we can potentially waste up to 256MB of the 4G process address
244 	 *  space just for alignment.
245 	 *
246 	 * XXXQ Should iterate trough hw_page_array here to catch
247 	 * all supported pagesizes
248 	 */
249 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
250 	    ((uintptr_t)*addrp) != 0)) {
251 		allow_largepage_alignment = 0;
252 	}
253 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
254 	    allow_largepage_alignment &&
255 	    (len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
256 		align_amount = MMU_PAGESIZE256M;
257 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
258 	    allow_largepage_alignment &&
259 	    (len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
260 		align_amount = MMU_PAGESIZE32M;
261 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
262 		align_amount = MMU_PAGESIZE4M;
263 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
264 		align_amount = MMU_PAGESIZE512K;
265 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
266 		align_amount = MMU_PAGESIZE64K;
267 	} else  {
268 		/*
269 		 * Align virtual addresses on a 64K boundary to ensure
270 		 * that ELF shared libraries are mapped with the appropriate
271 		 * alignment constraints by the run-time linker.
272 		 */
273 		align_amount = ELF_SPARC_MAXPGSZ;
274 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
275 		    ((uintptr_t)*addrp < align_amount))
276 			align_amount = (uintptr_t)*addrp;
277 	}
278 
279 	/*
280 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
281 	 */
282 	if (p->p_model == DATAMODEL_LP64)
283 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
284 #ifdef VAC
285 	if (vac && vacalign && (align_amount < shm_alignment))
286 		align_amount = shm_alignment;
287 #endif
288 
289 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
290 		align_amount = (uintptr_t)*addrp;
291 	}
292 
293 	ASSERT(ISP2(align_amount));
294 	ASSERT(align_amount == 0 || align_amount >= PAGESIZE);
295 
296 	/*
297 	 * Look for a large enough hole starting below the stack limit.
298 	 * After finding it, use the upper part.
299 	 */
300 	as_purge(as);
301 	off = off & (align_amount - 1);
302 	if (as_gap_aligned(as, len, &base, &slen, AH_HI, NULL, align_amount,
303 	    PAGESIZE, off) == 0) {
304 		caddr_t as_addr;
305 
306 		/*
307 		 * addr is the highest possible address to use since we have
308 		 * a PAGESIZE redzone at the beginning and end.
309 		 */
310 		addr = base + slen - (PAGESIZE + len);
311 		as_addr = addr;
312 		/*
313 		 * Round address DOWN to the alignment amount and
314 		 * add the offset in.
315 		 * If addr is greater than as_addr, len would not be large
316 		 * enough to include the redzone, so we must adjust down
317 		 * by the alignment amount.
318 		 */
319 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
320 		addr += (long)off;
321 		if (addr > as_addr) {
322 			addr -= align_amount;
323 		}
324 
325 		ASSERT(addr > base);
326 		ASSERT(addr + len < base + slen);
327 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
328 		    ((uintptr_t)(off)));
329 		*addrp = addr;
330 
331 	} else {
332 		*addrp = NULL;	/* no more virtual space */
333 	}
334 }
335 
336 /*
337  * Platform-dependent page scrub call.
338  * We call hypervisor to scrub the page.
339  */
340 void
341 pagescrub(page_t *pp, uint_t off, uint_t len)
342 {
343 	uint64_t pa, length;
344 
345 	pa = (uint64_t)(pp->p_pagenum << MMU_PAGESHIFT + off);
346 	length = (uint64_t)len;
347 
348 	(void) mem_scrub(pa, length);
349 }
350 
351 void
352 sync_data_memory(caddr_t va, size_t len)
353 {
354 	/* Call memory sync function */
355 	(void) mem_sync(va, len);
356 }
357 
358 size_t
359 mmu_get_kernel_lpsize(size_t lpsize)
360 {
361 	extern int mmu_exported_pagesize_mask;
362 	uint_t tte;
363 
364 	if (lpsize == 0) {
365 		/* no setting for segkmem_lpsize in /etc/system: use default */
366 		if (mmu_exported_pagesize_mask & (1 << TTE256M)) {
367 			lpsize = MMU_PAGESIZE256M;
368 		} else if (mmu_exported_pagesize_mask & (1 << TTE4M)) {
369 			lpsize = MMU_PAGESIZE4M;
370 		} else if (mmu_exported_pagesize_mask & (1 << TTE64K)) {
371 			lpsize = MMU_PAGESIZE64K;
372 		} else {
373 			lpsize = MMU_PAGESIZE;
374 		}
375 
376 		return (lpsize);
377 	}
378 
379 	for (tte = TTE8K; tte <= TTE256M; tte++) {
380 
381 		if ((mmu_exported_pagesize_mask & (1 << tte)) == 0)
382 			continue;
383 
384 		if (lpsize == TTEBYTES(tte))
385 			return (lpsize);
386 	}
387 
388 	lpsize = TTEBYTES(TTE8K);
389 	return (lpsize);
390 }
391 
392 void
393 mmu_init_kcontext()
394 {
395 }
396 
397 /*ARGSUSED*/
398 void
399 mmu_init_kernel_pgsz(struct hat *hat)
400 {
401 }
402 
403 static void *
404 contig_mem_span_alloc(vmem_t *vmp, size_t size, int vmflag)
405 {
406 	page_t *ppl;
407 	page_t *rootpp;
408 	caddr_t addr = NULL;
409 	pgcnt_t npages = btopr(size);
410 	page_t **ppa;
411 	int pgflags;
412 	spgcnt_t i = 0;
413 
414 
415 	ASSERT(size <= contig_mem_import_size_max);
416 	ASSERT((size & (size - 1)) == 0);
417 
418 	if ((addr = vmem_xalloc(vmp, size, size, 0, 0,
419 	    NULL, NULL, vmflag)) == NULL) {
420 		return (NULL);
421 	}
422 
423 	/* The address should be slab-size aligned. */
424 	ASSERT(((uintptr_t)addr & (size - 1)) == 0);
425 
426 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
427 		vmem_xfree(vmp, addr, size);
428 		return (NULL);
429 	}
430 
431 	pgflags = PG_EXCL;
432 	if (vmflag & VM_NORELOC)
433 		pgflags |= PG_NORELOC;
434 
435 	ppl = page_create_va_large(&kvp, (u_offset_t)(uintptr_t)addr, size,
436 	    pgflags, &kvseg, addr, NULL);
437 
438 	if (ppl == NULL) {
439 		vmem_xfree(vmp, addr, size);
440 		page_unresv(npages);
441 		return (NULL);
442 	}
443 
444 	rootpp = ppl;
445 	ppa = kmem_zalloc(npages * sizeof (page_t *), KM_SLEEP);
446 	while (ppl != NULL) {
447 		page_t *pp = ppl;
448 		ppa[i++] = pp;
449 		page_sub(&ppl, pp);
450 		ASSERT(page_iolock_assert(pp));
451 		ASSERT(PAGE_EXCL(pp));
452 		page_io_unlock(pp);
453 	}
454 
455 	/*
456 	 * Load the locked entry.  It's OK to preload the entry into
457 	 * the TSB since we now support large mappings in the kernel TSB.
458 	 */
459 	hat_memload_array(kas.a_hat, (caddr_t)rootpp->p_offset, size,
460 	    ppa, (PROT_ALL & ~PROT_USER) | HAT_NOSYNC, HAT_LOAD_LOCK);
461 
462 	ASSERT(i == page_get_pagecnt(ppa[0]->p_szc));
463 	for (--i; i >= 0; --i) {
464 		ASSERT(ppa[i]->p_szc == ppa[0]->p_szc);
465 		ASSERT(page_pptonum(ppa[i]) == page_pptonum(ppa[0]) + i);
466 		(void) page_pp_lock(ppa[i], 0, 1);
467 		/*
468 		 * Leave the page share locked. For non-cage pages,
469 		 * this would prevent memory DR if it were supported
470 		 * on sun4v.
471 		 */
472 		page_downgrade(ppa[i]);
473 	}
474 
475 	kmem_free(ppa, npages * sizeof (page_t *));
476 	return (addr);
477 }
478 
479 /*
480  * Allocates a slab by first trying to use the largest slab size
481  * in contig_mem_import_sizes and then falling back to smaller slab
482  * sizes still large enough for the allocation. The sizep argument
483  * is a pointer to the requested size. When a slab is successfully
484  * allocated, the slab size, which must be >= *sizep and <=
485  * contig_mem_import_size_max, is returned in the *sizep argument.
486  * Returns the virtual address of the new slab.
487  */
488 static void *
489 span_alloc_downsize(vmem_t *vmp, size_t *sizep, size_t align, int vmflag)
490 {
491 	int i;
492 
493 	ASSERT(*sizep <= contig_mem_import_size_max);
494 
495 	for (i = 0; i < NUM_IMPORT_SIZES; i++) {
496 		size_t page_size = contig_mem_import_sizes[i];
497 
498 		/*
499 		 * Check that the alignment is also less than the
500 		 * import (large page) size. In the case where the
501 		 * alignment is larger than the size, a large page
502 		 * large enough for the allocation is not necessarily
503 		 * physical-address aligned to satisfy the requested
504 		 * alignment. Since alignment is required to be a
505 		 * power-of-2, any large page >= size && >= align will
506 		 * suffice.
507 		 */
508 		if (*sizep <= page_size && align <= page_size) {
509 			void *addr;
510 			addr = contig_mem_span_alloc(vmp, page_size, vmflag);
511 			if (addr == NULL)
512 				continue;
513 			*sizep = page_size;
514 			return (addr);
515 		}
516 		return (NULL);
517 	}
518 
519 	return (NULL);
520 }
521 
522 static void *
523 contig_mem_span_xalloc(vmem_t *vmp, size_t *sizep, size_t align, int vmflag)
524 {
525 	return (span_alloc_downsize(vmp, sizep, align, vmflag | VM_NORELOC));
526 }
527 
528 static void *
529 contig_mem_reloc_span_xalloc(vmem_t *vmp, size_t *sizep, size_t align,
530     int vmflag)
531 {
532 	ASSERT((vmflag & VM_NORELOC) == 0);
533 	return (span_alloc_downsize(vmp, sizep, align, vmflag));
534 }
535 
536 /*
537  * Free a span, which is always exactly one large page.
538  */
539 static void
540 contig_mem_span_free(vmem_t *vmp, void *inaddr, size_t size)
541 {
542 	page_t *pp;
543 	caddr_t addr = inaddr;
544 	caddr_t eaddr;
545 	pgcnt_t npages = btopr(size);
546 	page_t *rootpp = NULL;
547 
548 	ASSERT(size <= contig_mem_import_size_max);
549 	/* All slabs should be size aligned */
550 	ASSERT(((uintptr_t)addr & (size - 1)) == 0);
551 
552 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
553 
554 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
555 		pp = page_find(&kvp, (u_offset_t)(uintptr_t)addr);
556 		if (pp == NULL) {
557 			panic("contig_mem_span_free: page not found");
558 		}
559 		if (!page_tryupgrade(pp)) {
560 			page_unlock(pp);
561 			pp = page_lookup(&kvp,
562 			    (u_offset_t)(uintptr_t)addr, SE_EXCL);
563 			if (pp == NULL)
564 				panic("contig_mem_span_free: page not found");
565 		}
566 
567 		ASSERT(PAGE_EXCL(pp));
568 		ASSERT(size == page_get_pagesize(pp->p_szc));
569 		ASSERT(rootpp == NULL || rootpp->p_szc == pp->p_szc);
570 		ASSERT(rootpp == NULL || (page_pptonum(rootpp) +
571 		    (pgcnt_t)btop(addr - (caddr_t)inaddr) == page_pptonum(pp)));
572 
573 		page_pp_unlock(pp, 0, 1);
574 
575 		if (rootpp == NULL)
576 			rootpp = pp;
577 	}
578 	page_destroy_pages(rootpp);
579 	page_unresv(npages);
580 
581 	if (vmp != NULL)
582 		vmem_xfree(vmp, inaddr, size);
583 }
584 
585 static void *
586 contig_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t *sizep, size_t align,
587     int vmflag)
588 {
589 	ASSERT((align & (align - 1)) == 0);
590 	return (vmem_xalloc(vmp, *sizep, align, 0, 0, NULL, NULL, vmflag));
591 }
592 
593 /*
594  * contig_mem_alloc, contig_mem_alloc_align
595  *
596  * Caution: contig_mem_alloc and contig_mem_alloc_align should be
597  * used only when physically contiguous non-relocatable memory is
598  * required. Furthermore, use of these allocation routines should be
599  * minimized as well as should the allocation size. As described in the
600  * contig_mem_arena comment block above, slab allocations fall back to
601  * being outside of the cage. Therefore, overuse of these allocation
602  * routines can lead to non-relocatable large pages being allocated
603  * outside the cage. Such pages prevent the allocation of a larger page
604  * occupying overlapping pages. This can impact performance for
605  * applications that utilize e.g. 256M large pages.
606  */
607 
608 /*
609  * Allocates size aligned contiguous memory up to contig_mem_import_size_max.
610  * Size must be a power of 2.
611  */
612 void *
613 contig_mem_alloc(size_t size)
614 {
615 	ASSERT((size & (size - 1)) == 0);
616 	return (contig_mem_alloc_align(size, size));
617 }
618 
619 /*
620  * contig_mem_alloc_align allocates real contiguous memory with the specified
621  * alignment up to contig_mem_import_size_max. The alignment must be a
622  * power of 2 and no greater than contig_mem_import_size_max. We assert
623  * the aligment is a power of 2. For non-debug, vmem_xalloc will panic
624  * for non power of 2 alignments.
625  */
626 void *
627 contig_mem_alloc_align(size_t size, size_t align)
628 {
629 	void *buf;
630 
631 	ASSERT(size <= contig_mem_import_size_max);
632 	ASSERT(align <= contig_mem_import_size_max);
633 	ASSERT((align & (align - 1)) == 0);
634 
635 	if (align < CONTIG_MEM_ARENA_QUANTUM)
636 		align = CONTIG_MEM_ARENA_QUANTUM;
637 
638 	/*
639 	 * We take the lock here to serialize span allocations.
640 	 * We do not lose concurrency for the common case, since
641 	 * allocations that don't require new span allocations
642 	 * are serialized by vmem_xalloc. Serializing span
643 	 * allocations also prevents us from trying to allocate
644 	 * more spans that necessary.
645 	 */
646 	mutex_enter(&contig_mem_lock);
647 
648 	buf = vmem_xalloc(contig_mem_arena, size, align, 0, 0,
649 	    NULL, NULL, VM_NOSLEEP | VM_NORELOC);
650 
651 	if ((buf == NULL) && (size <= MMU_PAGESIZE)) {
652 		mutex_exit(&contig_mem_lock);
653 		return (vmem_xalloc(static_alloc_arena, size, align, 0, 0,
654 		    NULL, NULL, VM_NOSLEEP));
655 	}
656 
657 	if (buf == NULL) {
658 		buf = vmem_xalloc(contig_mem_reloc_arena, size, align, 0, 0,
659 		    NULL, NULL, VM_NOSLEEP);
660 	}
661 
662 	mutex_exit(&contig_mem_lock);
663 
664 	return (buf);
665 }
666 
667 void
668 contig_mem_free(void *vaddr, size_t size)
669 {
670 	if (vmem_contains(contig_mem_arena, vaddr, size)) {
671 		vmem_xfree(contig_mem_arena, vaddr, size);
672 	} else if (size > MMU_PAGESIZE) {
673 		vmem_xfree(contig_mem_reloc_arena, vaddr, size);
674 	} else {
675 		vmem_xfree(static_alloc_arena, vaddr, size);
676 	}
677 }
678 
679 /*
680  * We create a set of stacked vmem arenas to enable us to
681  * allocate large >PAGESIZE chucks of contiguous Real Address space.
682  * The vmem_xcreate interface is used to create the contig_mem_arena
683  * allowing the import routine to downsize the requested slab size
684  * and return a smaller slab.
685  */
686 void
687 contig_mem_init(void)
688 {
689 	mutex_init(&contig_mem_lock, NULL, MUTEX_DEFAULT, NULL);
690 
691 	contig_mem_slab_arena = vmem_xcreate("contig_mem_slab_arena", NULL, 0,
692 	    CONTIG_MEM_SLAB_ARENA_QUANTUM, contig_vmem_xalloc_aligned_wrapper,
693 	    vmem_xfree, heap_arena, 0, VM_SLEEP | VMC_XALIGN);
694 
695 	contig_mem_arena = vmem_xcreate("contig_mem_arena", NULL, 0,
696 	    CONTIG_MEM_ARENA_QUANTUM, contig_mem_span_xalloc,
697 	    contig_mem_span_free, contig_mem_slab_arena, 0,
698 	    VM_SLEEP | VM_BESTFIT | VMC_XALIGN);
699 
700 	contig_mem_reloc_arena = vmem_xcreate("contig_mem_reloc_arena", NULL, 0,
701 	    CONTIG_MEM_ARENA_QUANTUM, contig_mem_reloc_span_xalloc,
702 	    contig_mem_span_free, contig_mem_slab_arena, 0,
703 	    VM_SLEEP | VM_BESTFIT | VMC_XALIGN);
704 
705 	if (contig_mem_prealloc_buf == NULL || vmem_add(contig_mem_arena,
706 	    contig_mem_prealloc_buf, contig_mem_prealloc_size, VM_SLEEP)
707 	    == NULL) {
708 		cmn_err(CE_WARN, "Failed to pre-populate contig_mem_arena");
709 	}
710 }
711 
712 /*
713  * In calculating how much memory to pre-allocate, we include a small
714  * amount per-CPU to account for per-CPU buffers in line with measured
715  * values for different size systems. contig_mem_prealloc_base_size is
716  * a cpu specific amount to be pre-allocated before considering per-CPU
717  * requirements and memory size. We always pre-allocate a minimum amount
718  * of memory determined by PREALLOC_MIN. Beyond that, we take the minimum
719  * of contig_mem_prealloc_base_size and a small percentage of physical
720  * memory to prevent allocating too much on smaller systems.
721  * contig_mem_prealloc_base_size is global, allowing for the CPU module
722  * to increase its value if necessary.
723  */
724 #define	PREALLOC_PER_CPU	(256 * 1024)		/* 256K */
725 #define	PREALLOC_PERCENT	(4)			/* 4% */
726 #define	PREALLOC_MIN		(16 * 1024 * 1024)	/* 16M */
727 size_t contig_mem_prealloc_base_size = 0;
728 
729 /*
730  * Called at boot-time allowing pre-allocation of contiguous memory.
731  * The argument 'alloc_base' is the requested base address for the
732  * allocation and originates in startup_memlist.
733  */
734 caddr_t
735 contig_mem_prealloc(caddr_t alloc_base, pgcnt_t npages)
736 {
737 	caddr_t	chunkp;
738 
739 	contig_mem_prealloc_size = MIN((PREALLOC_PER_CPU * ncpu_guest_max) +
740 	    contig_mem_prealloc_base_size,
741 	    (ptob(npages) * PREALLOC_PERCENT) / 100);
742 	contig_mem_prealloc_size = MAX(contig_mem_prealloc_size, PREALLOC_MIN);
743 	contig_mem_prealloc_size = P2ROUNDUP(contig_mem_prealloc_size,
744 	    MMU_PAGESIZE4M);
745 
746 	alloc_base = (caddr_t)roundup((uintptr_t)alloc_base, MMU_PAGESIZE4M);
747 	if (prom_alloc(alloc_base, contig_mem_prealloc_size,
748 	    MMU_PAGESIZE4M) != alloc_base) {
749 
750 		/*
751 		 * Failed.  This may mean the physical memory has holes in it
752 		 * and it will be more difficult to get large contiguous
753 		 * pieces of memory.  Since we only guarantee contiguous
754 		 * pieces of memory contig_mem_import_size_max or smaller,
755 		 * loop, getting contig_mem_import_size_max at a time, until
756 		 * failure or contig_mem_prealloc_size is reached.
757 		 */
758 		for (chunkp = alloc_base;
759 		    (chunkp - alloc_base) < contig_mem_prealloc_size;
760 		    chunkp += contig_mem_import_size_max) {
761 
762 			if (prom_alloc(chunkp, contig_mem_import_size_max,
763 			    MMU_PAGESIZE4M) != chunkp) {
764 				break;
765 			}
766 		}
767 		contig_mem_prealloc_size = chunkp - alloc_base;
768 		ASSERT(contig_mem_prealloc_size != 0);
769 	}
770 
771 	if (contig_mem_prealloc_size != 0) {
772 		contig_mem_prealloc_buf = alloc_base;
773 	} else {
774 		contig_mem_prealloc_buf = NULL;
775 	}
776 	alloc_base += contig_mem_prealloc_size;
777 
778 	return (alloc_base);
779 }
780 
781 static uint_t sp_color_stride = 16;
782 static uint_t sp_color_mask = 0x1f;
783 static uint_t sp_current_color = (uint_t)-1;
784 
785 size_t
786 exec_get_spslew(void)
787 {
788 	uint_t spcolor = atomic_inc_32_nv(&sp_current_color);
789 	return ((size_t)((spcolor & sp_color_mask) * SA(sp_color_stride)));
790 }
791 
792 /*
793  * This flag may be set via /etc/system to force the synchronization
794  * of I-cache with memory after every bcopy.  The default is 0, meaning
795  * that there is no need for an I-cache flush after each bcopy.  This
796  * flag is relevant only on platforms that have non-coherent I-caches.
797  */
798 uint_t	force_sync_icache_after_bcopy = 0;
799 
800 /*
801  * This flag may be set via /etc/system to force the synchronization
802  * of I-cache to memory after every DMA. The default is 0, meaning
803  * that there is no need for an I-cache flush after each dma write to
804  * memory. This flag is relevant only on platforms that have
805  * non-coherent I-caches.
806  */
807 uint_t	force_sync_icache_after_dma = 0;
808 
809 /*
810  * This internal flag enables mach_sync_icache_pa, which is always
811  * called from common code if it is defined. However, not all
812  * platforms support the hv_mem_iflush firmware call.
813  */
814 static uint_t	do_mach_sync_icache_pa = 0;
815 
816 int	hsvc_kdi_mem_iflush_negotiated = B_FALSE;
817 
818 #define	MEM_IFLUSH_MAJOR	1
819 #define	MEM_IFLUSH_MINOR	0
820 static hsvc_info_t kdi_mem_iflush_hsvc = {
821 	HSVC_REV_1,		/* HSVC rev num */
822 	NULL,			/* Private */
823 	HSVC_GROUP_MEM_IFLUSH,	/* Requested API Group */
824 	MEM_IFLUSH_MAJOR,	/* Requested Major */
825 	MEM_IFLUSH_MINOR,	/* Requested Minor */
826 	"kdi"			/* Module name */
827 };
828 
829 /*
830  * Setup soft exec mode.
831  * Since /etc/system is read later on init, it
832  * may be used to override these flags.
833  */
834 void
835 mach_setup_icache(uint_t coherency)
836 {
837 	int		status;
838 	uint64_t	sup_minor;
839 
840 	if (coherency == 0 && icache_is_coherent) {
841 		extern void kdi_flush_caches(void);
842 		status = hsvc_register(&kdi_mem_iflush_hsvc, &sup_minor);
843 		if (status != 0)
844 			cmn_err(CE_PANIC, "I$ flush not implemented on "
845 			    "I$ incoherent system");
846 		hsvc_kdi_mem_iflush_negotiated = B_TRUE;
847 		kdi_flush_caches();
848 		icache_is_coherent = 0;
849 		do_mach_sync_icache_pa = 1;
850 	}
851 }
852 
853 /*
854  * Flush specified physical address range from I$ via hv_mem_iflush interface
855  */
856 /*ARGSUSED*/
857 void
858 mach_sync_icache_pa(caddr_t paddr, size_t size)
859 {
860 	if (do_mach_sync_icache_pa) {
861 		uint64_t pa = (uint64_t)paddr;
862 		uint64_t sz = (uint64_t)size;
863 		uint64_t i, flushed;
864 
865 		for (i = 0; i < sz; i += flushed) {
866 			if (hv_mem_iflush(pa + i, sz - i, &flushed) != H_EOK) {
867 				cmn_err(CE_PANIC, "Flushing the Icache failed");
868 				break;
869 			}
870 		}
871 	}
872 }
873 
874 /*
875  * Flush the page if it has been marked as executed
876  */
877 /*ARGSUSED*/
878 void
879 mach_sync_icache_pp(page_t *pp)
880 {
881 	if (PP_ISEXEC(pp))
882 		mach_sync_icache_pa((caddr_t)ptob(pp->p_pagenum), PAGESIZE);
883 }
884