xref: /titanic_41/usr/src/uts/sun4v/vm/mach_vm_dep.c (revision 8c74a1f9477c04aa8539a84a49aa2bf629c7a14d)
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 2006 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 
55 uint_t page_colors = 0;
56 uint_t page_colors_mask = 0;
57 uint_t page_coloring_shift = 0;
58 int consistent_coloring;
59 
60 uint_t mmu_page_sizes = MMU_PAGE_SIZES;
61 uint_t max_mmu_page_sizes = MMU_PAGE_SIZES;
62 uint_t mmu_hashcnt = MAX_HASHCNT;
63 uint_t max_mmu_hashcnt = MAX_HASHCNT;
64 size_t mmu_ism_pagesize = DEFAULT_ISM_PAGESIZE;
65 
66 /*
67  * A bitmask of the page sizes supported by hardware based upon szc.
68  * The base pagesize (p_szc == 0) must always be supported by the hardware.
69  */
70 int mmu_exported_pagesize_mask;
71 uint_t mmu_exported_page_sizes;
72 
73 uint_t szc_2_userszc[MMU_PAGE_SIZES];
74 uint_t userszc_2_szc[MMU_PAGE_SIZES];
75 
76 extern uint_t vac_colors_mask;
77 extern int vac_shift;
78 
79 hw_pagesize_t hw_page_array[] = {
80 	{MMU_PAGESIZE, MMU_PAGESHIFT, MMU_PAGESIZE >> MMU_PAGESHIFT},
81 	{MMU_PAGESIZE64K, MMU_PAGESHIFT64K, MMU_PAGESIZE64K >> MMU_PAGESHIFT},
82 	{MMU_PAGESIZE512K, MMU_PAGESHIFT512K,
83 	    MMU_PAGESIZE512K >> MMU_PAGESHIFT},
84 	{MMU_PAGESIZE4M, MMU_PAGESHIFT4M, MMU_PAGESIZE4M >> MMU_PAGESHIFT},
85 	{MMU_PAGESIZE32M, MMU_PAGESHIFT32M, MMU_PAGESIZE32M >> MMU_PAGESHIFT},
86 	{MMU_PAGESIZE256M, MMU_PAGESHIFT256M,
87 	    MMU_PAGESIZE256M >> MMU_PAGESHIFT},
88 	{0, 0, 0}
89 };
90 
91 /*
92  * Enable usage of 64k/4M pages for text and 64k pages for initdata for
93  * all sun4v platforms. These variables can be overwritten by the platmod
94  * or the CPU module. User can also change the setting via /etc/system.
95  */
96 
97 int	use_text_pgsz64k = 1;
98 int	use_text_pgsz4m = 1;
99 int	use_initdata_pgsz64k = 1;
100 
101 /*
102  * disable_text_largepages and disable_initdata_largepages bitmaks reflect
103  * both unconfigured and undesirable page sizes. Current implementation
104  * supports 64K and 4M page sizes for text and only 64K for data. Rest of
105  * the page sizes are not currently supported, hence disabled below. In
106  * future, when support is added for any other page size, it should be
107  * reflected below.
108  *
109  * Note that these bitmask can be set in platform or CPU specific code to
110  * disable page sizes that should not be used. These variables normally
111  * shouldn't be changed via /etc/system.
112  *
113  * These bitmasks are also updated within hat_init to reflect unsupported
114  * page sizes on a sun4v processor per mmu_exported_pagesize_mask global
115  * variable.
116  */
117 
118 int disable_text_largepages =
119 	(1 << TTE512K) | (1 << TTE32M) | (1 << TTE256M) | (1 << TTE2G) |
120 	(1 << TTE16G);
121 int disable_initdata_largepages =
122 	(1 << TTE512K) | (1 << TTE4M) | (1 << TTE32M) | (1 << TTE256M) |
123 	(1 << TTE2G) | (1 << TTE16G);
124 
125 /*
126  * Minimum segment size tunables before 64K or 4M large pages
127  * should be used to map it.
128  */
129 size_t text_pgsz64k_minsize = MMU_PAGESIZE64K;
130 size_t text_pgsz4m_minsize = MMU_PAGESIZE4M;
131 size_t initdata_pgsz64k_minsize = MMU_PAGESIZE64K;
132 
133 size_t max_shm_lpsize = MMU_PAGESIZE4M;
134 
135 /*
136  * map_addr_proc() is the routine called when the system is to
137  * choose an address for the user.  We will pick an address
138  * range which is just below the current stack limit.  The
139  * algorithm used for cache consistency on machines with virtual
140  * address caches is such that offset 0 in the vnode is always
141  * on a shm_alignment'ed aligned address.  Unfortunately, this
142  * means that vnodes which are demand paged will not be mapped
143  * cache consistently with the executable images.  When the
144  * cache alignment for a given object is inconsistent, the
145  * lower level code must manage the translations so that this
146  * is not seen here (at the cost of efficiency, of course).
147  *
148  * addrp is a value/result parameter.
149  *	On input it is a hint from the user to be used in a completely
150  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
151  *	minimal alignment.
152  *
153  *	On output it is NULL if no address can be found in the current
154  *	processes address space or else an address that is currently
155  *	not mapped for len bytes with a page of red zone on either side.
156  *	If vacalign is true, then the selected address will obey the alignment
157  *	constraints of a vac machine based on the given off value.
158  */
159 /*ARGSUSED3*/
160 void
161 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
162     caddr_t userlimit, struct proc *p, uint_t flags)
163 {
164 	struct as *as = p->p_as;
165 	caddr_t addr;
166 	caddr_t base;
167 	size_t slen;
168 	uintptr_t align_amount;
169 	int allow_largepage_alignment = 1;
170 
171 	base = p->p_brkbase;
172 	if (userlimit < as->a_userlimit) {
173 		/*
174 		 * This happens when a program wants to map something in
175 		 * a range that's accessible to a program in a smaller
176 		 * address space.  For example, a 64-bit program might
177 		 * be calling mmap32(2) to guarantee that the returned
178 		 * address is below 4Gbytes.
179 		 */
180 		ASSERT(userlimit > base);
181 		slen = userlimit - base;
182 	} else {
183 		slen = p->p_usrstack - base - (((size_t)rctl_enforced_value(
184 		    rctlproc_legacy[RLIMIT_STACK], p->p_rctls, p) + PAGEOFFSET)
185 		    & PAGEMASK);
186 	}
187 	len = (len + PAGEOFFSET) & PAGEMASK;
188 
189 	/*
190 	 * Redzone for each side of the request. This is done to leave
191 	 * one page unmapped between segments. This is not required, but
192 	 * it's useful for the user because if their program strays across
193 	 * a segment boundary, it will catch a fault immediately making
194 	 * debugging a little easier.
195 	 */
196 	len += (2 * PAGESIZE);
197 
198 	/*
199 	 *  If the request is larger than the size of a particular
200 	 *  mmu level, then we use that level to map the request.
201 	 *  But this requires that both the virtual and the physical
202 	 *  addresses be aligned with respect to that level, so we
203 	 *  do the virtual bit of nastiness here.
204 	 *
205 	 *  For 32-bit processes, only those which have specified
206 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
207 	 *  we can potentially waste up to 256MB of the 4G process address
208 	 *  space just for alignment.
209 	 *
210 	 * XXXQ Should iterate trough hw_page_array here to catch
211 	 * all supported pagesizes
212 	 */
213 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
214 	    ((uintptr_t)*addrp) != 0)) {
215 		allow_largepage_alignment = 0;
216 	}
217 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
218 	    allow_largepage_alignment &&
219 		(len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
220 		align_amount = MMU_PAGESIZE256M;
221 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
222 	    allow_largepage_alignment &&
223 		(len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
224 		align_amount = MMU_PAGESIZE32M;
225 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
226 		align_amount = MMU_PAGESIZE4M;
227 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
228 		align_amount = MMU_PAGESIZE512K;
229 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
230 		align_amount = MMU_PAGESIZE64K;
231 	} else  {
232 		/*
233 		 * Align virtual addresses on a 64K boundary to ensure
234 		 * that ELF shared libraries are mapped with the appropriate
235 		 * alignment constraints by the run-time linker.
236 		 */
237 		align_amount = ELF_SPARC_MAXPGSZ;
238 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
239 			((uintptr_t)*addrp < align_amount))
240 			align_amount = (uintptr_t)*addrp;
241 	}
242 
243 	/*
244 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
245 	 */
246 	if (p->p_model == DATAMODEL_LP64)
247 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
248 #ifdef VAC
249 	if (vac && vacalign && (align_amount < shm_alignment))
250 		align_amount = shm_alignment;
251 #endif
252 
253 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
254 		align_amount = (uintptr_t)*addrp;
255 	}
256 	len += align_amount;
257 
258 	/*
259 	 * Look for a large enough hole starting below the stack limit.
260 	 * After finding it, use the upper part.  Addition of PAGESIZE is
261 	 * for the redzone as described above.
262 	 */
263 	as_purge(as);
264 	if (as_gap(as, len, &base, &slen, AH_HI, NULL) == 0) {
265 		caddr_t as_addr;
266 
267 		addr = base + slen - len + PAGESIZE;
268 		as_addr = addr;
269 		/*
270 		 * Round address DOWN to the alignment amount,
271 		 * add the offset, and if this address is less
272 		 * than the original address, add alignment amount.
273 		 */
274 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
275 		addr += (long)(off & (align_amount - 1l));
276 		if (addr < as_addr) {
277 			addr += align_amount;
278 		}
279 
280 		ASSERT(addr <= (as_addr + align_amount));
281 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
282 		    ((uintptr_t)(off & (align_amount - 1l))));
283 		*addrp = addr;
284 
285 	} else {
286 		*addrp = NULL;	/* no more virtual space */
287 	}
288 }
289 
290 /* Auto large page tunables. */
291 int auto_lpg_tlb_threshold = 32;
292 int auto_lpg_minszc = TTE64K;
293 int auto_lpg_maxszc = TTE256M;
294 size_t auto_lpg_heap_default = MMU_PAGESIZE64K;
295 size_t auto_lpg_stack_default = MMU_PAGESIZE64K;
296 size_t auto_lpg_va_default = MMU_PAGESIZE64K;
297 size_t auto_lpg_remap_threshold = 0; /* always remap */
298 /*
299  * Number of pages in 1 GB.  Don't enable automatic large pages if we have
300  * fewer than this many pages.
301  */
302 pgcnt_t auto_lpg_min_physmem = 1 << (30 - MMU_PAGESHIFT);
303 
304 size_t
305 map_pgsz(int maptype, struct proc *p, caddr_t addr, size_t len, int *remap)
306 {
307 	uint_t	n;
308 	size_t	pgsz = 0;
309 
310 	if (remap)
311 		*remap = (len > auto_lpg_remap_threshold);
312 
313 	switch (maptype) {
314 	case MAPPGSZ_ISM:
315 		n = hat_preferred_pgsz(p->p_as->a_hat, addr, len, maptype);
316 		pgsz = hw_page_array[n].hp_size;
317 		break;
318 
319 	case MAPPGSZ_VA:
320 		n = hat_preferred_pgsz(p->p_as->a_hat, addr, len, maptype);
321 		pgsz = hw_page_array[n].hp_size;
322 		if ((pgsz <= MMU_PAGESIZE) ||
323 		    !IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz))
324 			pgsz = map_pgszva(p, addr, len);
325 		break;
326 
327 	case MAPPGSZ_STK:
328 		pgsz = map_pgszstk(p, addr, len);
329 		break;
330 
331 	case MAPPGSZ_HEAP:
332 		pgsz = map_pgszheap(p, addr, len);
333 		break;
334 	}
335 	return (pgsz);
336 }
337 
338 /*
339  * Platform-dependent page scrub call.
340  * We call hypervisor to scrub the page.
341  */
342 void
343 pagescrub(page_t *pp, uint_t off, uint_t len)
344 {
345 	uint64_t pa, length;
346 
347 	pa = (uint64_t)(pp->p_pagenum << MMU_PAGESHIFT + off);
348 	length = (uint64_t)len;
349 
350 	(void) mem_scrub(pa, length);
351 }
352 
353 void
354 sync_data_memory(caddr_t va, size_t len)
355 {
356 	/* Call memory sync function */
357 	mem_sync(va, len);
358 }
359 
360 size_t
361 mmu_get_kernel_lpsize(size_t lpsize)
362 {
363 	extern int mmu_exported_pagesize_mask;
364 	uint_t tte;
365 
366 	if (lpsize == 0) {
367 		/* no setting for segkmem_lpsize in /etc/system: use default */
368 		if (mmu_exported_pagesize_mask & (1 << TTE256M)) {
369 			lpsize = MMU_PAGESIZE256M;
370 		} else if (mmu_exported_pagesize_mask & (1 << TTE4M)) {
371 			lpsize = MMU_PAGESIZE4M;
372 		} else if (mmu_exported_pagesize_mask & (1 << TTE64K)) {
373 			lpsize = MMU_PAGESIZE64K;
374 		} else {
375 			lpsize = MMU_PAGESIZE;
376 		}
377 
378 		return (lpsize);
379 	}
380 
381 	for (tte = TTE8K; tte <= TTE256M; tte++) {
382 
383 		if ((mmu_exported_pagesize_mask & (1 << tte)) == 0)
384 			continue;
385 
386 		if (lpsize == TTEBYTES(tte))
387 			return (lpsize);
388 	}
389 
390 	lpsize = TTEBYTES(TTE8K);
391 	return (lpsize);
392 }
393 
394 void
395 mmu_init_kcontext()
396 {
397 }
398 
399 /*ARGSUSED*/
400 void
401 mmu_init_kernel_pgsz(struct hat *hat)
402 {
403 }
404 
405 #define	QUANTUM_SIZE	64
406 
407 static	vmem_t	*contig_mem_slab_arena;
408 static	vmem_t	*contig_mem_arena;
409 
410 uint_t contig_mem_slab_size = MMU_PAGESIZE4M;
411 
412 static void *
413 contig_mem_span_alloc(vmem_t *vmp, size_t size, int vmflag)
414 {
415 	page_t *ppl;
416 	page_t *rootpp;
417 	caddr_t addr = NULL;
418 	pgcnt_t npages = btopr(size);
419 	page_t **ppa;
420 	int pgflags;
421 	int i = 0;
422 
423 
424 	/*
425 	 * The import request should be at least
426 	 * contig_mem_slab_size because that is the
427 	 * slab arena's quantum. The size can be
428 	 * further restricted since contiguous
429 	 * allocations larger than contig_mem_slab_size
430 	 * are not supported here.
431 	 */
432 	ASSERT(size == contig_mem_slab_size);
433 
434 	if ((addr = vmem_xalloc(vmp, size, size, 0, 0,
435 	    NULL, NULL, vmflag)) == NULL) {
436 		return (NULL);
437 	}
438 
439 	/* The address should be slab-size aligned. */
440 	ASSERT(((uintptr_t)addr & (contig_mem_slab_size - 1)) == 0);
441 
442 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
443 		vmem_xfree(vmp, addr, size);
444 		return (NULL);
445 	}
446 
447 	pgflags = PG_EXCL;
448 	if ((vmflag & VM_NOSLEEP) == 0)
449 		pgflags |= PG_WAIT;
450 	if (vmflag & VM_PANIC)
451 		pgflags |= PG_PANIC;
452 	if (vmflag & VM_PUSHPAGE)
453 		pgflags |= PG_PUSHPAGE;
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 		page_io_unlock(pp);
472 	}
473 
474 	/*
475 	 * Load the locked entry.  It's OK to preload the entry into
476 	 * the TSB since we now support large mappings in the kernel TSB.
477 	 */
478 	hat_memload_array(kas.a_hat, (caddr_t)rootpp->p_offset, size,
479 	    ppa, (PROT_ALL & ~PROT_USER) | HAT_NOSYNC, HAT_LOAD_LOCK);
480 
481 	for (--i; i >= 0; --i) {
482 		(void) page_pp_lock(ppa[i], 0, 1);
483 		page_unlock(ppa[i]);
484 	}
485 
486 	kmem_free(ppa, npages * sizeof (page_t *));
487 	return (addr);
488 }
489 
490 void
491 contig_mem_span_free(vmem_t *vmp, void *inaddr, size_t size)
492 {
493 	page_t *pp;
494 	caddr_t addr = inaddr;
495 	caddr_t eaddr;
496 	pgcnt_t npages = btopr(size);
497 	pgcnt_t pgs_left = npages;
498 	page_t *rootpp = NULL;
499 
500 	ASSERT(((uintptr_t)addr & (contig_mem_slab_size - 1)) == 0);
501 
502 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
503 
504 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
505 		pp = page_lookup(&kvp, (u_offset_t)(uintptr_t)addr, SE_EXCL);
506 		if (pp == NULL)
507 			panic("contig_mem_span_free: page not found");
508 
509 		ASSERT(PAGE_EXCL(pp));
510 		page_pp_unlock(pp, 0, 1);
511 
512 		if (rootpp == NULL)
513 			rootpp = pp;
514 		if (--pgs_left == 0) {
515 			/*
516 			 * similar logic to segspt_free_pages, but we know we
517 			 * have one large page.
518 			 */
519 			page_destroy_pages(rootpp);
520 		}
521 	}
522 	page_unresv(npages);
523 
524 	if (vmp != NULL)
525 		vmem_xfree(vmp, inaddr, size);
526 }
527 
528 static void *
529 contig_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t size, int vmflag)
530 {
531 	return (vmem_xalloc(vmp, size, size, 0, 0, NULL, NULL, vmflag));
532 }
533 
534 /*
535  * conting_mem_alloc_align allocates real contiguous memory with the specified
536  * alignment upto contig_mem_slab_size. The alignment must be a power of 2.
537  */
538 void *
539 contig_mem_alloc_align(size_t size, size_t align)
540 {
541 	ASSERT(align <= contig_mem_slab_size);
542 
543 	if ((align & (align - 1)) != 0)
544 		return (NULL);
545 
546 	return (vmem_xalloc(contig_mem_arena, size, align, 0, 0,
547 	    NULL, NULL, VM_NOSLEEP));
548 }
549 
550 /*
551  * Allocates size aligned contiguous memory upto contig_mem_slab_size.
552  * Size must be a power of 2.
553  */
554 void *
555 contig_mem_alloc(size_t size)
556 {
557 	ASSERT((size & (size - 1)) == 0);
558 	return (contig_mem_alloc_align(size, size));
559 }
560 
561 void
562 contig_mem_free(void *vaddr, size_t size)
563 {
564 	vmem_xfree(contig_mem_arena, vaddr, size);
565 }
566 
567 /*
568  * We create a set of stacked vmem arenas to enable us to
569  * allocate large >PAGESIZE chucks of contiguous Real Address space
570  * This is  what the Dynamics TSB support does for TSBs.
571  * The contig_mem_arena import functions are exactly the same as the
572  * TSB kmem_default arena import functions.
573  */
574 void
575 contig_mem_init(void)
576 {
577 
578 	contig_mem_slab_arena = vmem_create("contig_mem_slab_arena", NULL, 0,
579 	    contig_mem_slab_size, contig_vmem_xalloc_aligned_wrapper,
580 	    vmem_xfree, heap_arena, 0, VM_SLEEP);
581 
582 	contig_mem_arena = vmem_create("contig_mem_arena", NULL, 0,
583 	    QUANTUM_SIZE, contig_mem_span_alloc, contig_mem_span_free,
584 	    contig_mem_slab_arena, 0, VM_SLEEP | VM_BESTFIT);
585 
586 }
587