xref: /titanic_50/usr/src/uts/sun4v/vm/mach_vm_dep.c (revision 144dfaa9a648eea321858b34d4941d2268130176)
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 /*
134  * map_addr_proc() is the routine called when the system is to
135  * choose an address for the user.  We will pick an address
136  * range which is just below the current stack limit.  The
137  * algorithm used for cache consistency on machines with virtual
138  * address caches is such that offset 0 in the vnode is always
139  * on a shm_alignment'ed aligned address.  Unfortunately, this
140  * means that vnodes which are demand paged will not be mapped
141  * cache consistently with the executable images.  When the
142  * cache alignment for a given object is inconsistent, the
143  * lower level code must manage the translations so that this
144  * is not seen here (at the cost of efficiency, of course).
145  *
146  * addrp is a value/result parameter.
147  *	On input it is a hint from the user to be used in a completely
148  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
149  *	minimal alignment.
150  *
151  *	On output it is NULL if no address can be found in the current
152  *	processes address space or else an address that is currently
153  *	not mapped for len bytes with a page of red zone on either side.
154  *	If vacalign is true, then the selected address will obey the alignment
155  *	constraints of a vac machine based on the given off value.
156  */
157 /*ARGSUSED3*/
158 void
159 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
160     caddr_t userlimit, struct proc *p, uint_t flags)
161 {
162 	struct as *as = p->p_as;
163 	caddr_t addr;
164 	caddr_t base;
165 	size_t slen;
166 	uintptr_t align_amount;
167 	int allow_largepage_alignment = 1;
168 
169 	base = p->p_brkbase;
170 	if (userlimit < as->a_userlimit) {
171 		/*
172 		 * This happens when a program wants to map something in
173 		 * a range that's accessible to a program in a smaller
174 		 * address space.  For example, a 64-bit program might
175 		 * be calling mmap32(2) to guarantee that the returned
176 		 * address is below 4Gbytes.
177 		 */
178 		ASSERT(userlimit > base);
179 		slen = userlimit - base;
180 	} else {
181 		slen = p->p_usrstack - base - (((size_t)rctl_enforced_value(
182 		    rctlproc_legacy[RLIMIT_STACK], p->p_rctls, p) + PAGEOFFSET)
183 		    & PAGEMASK);
184 	}
185 	len = (len + PAGEOFFSET) & PAGEMASK;
186 
187 	/*
188 	 * Redzone for each side of the request. This is done to leave
189 	 * one page unmapped between segments. This is not required, but
190 	 * it's useful for the user because if their program strays across
191 	 * a segment boundary, it will catch a fault immediately making
192 	 * debugging a little easier.
193 	 */
194 	len += (2 * PAGESIZE);
195 
196 	/*
197 	 *  If the request is larger than the size of a particular
198 	 *  mmu level, then we use that level to map the request.
199 	 *  But this requires that both the virtual and the physical
200 	 *  addresses be aligned with respect to that level, so we
201 	 *  do the virtual bit of nastiness here.
202 	 *
203 	 *  For 32-bit processes, only those which have specified
204 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
205 	 *  we can potentially waste up to 256MB of the 4G process address
206 	 *  space just for alignment.
207 	 *
208 	 * XXXQ Should iterate trough hw_page_array here to catch
209 	 * all supported pagesizes
210 	 */
211 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
212 	    ((uintptr_t)*addrp) != 0)) {
213 		allow_largepage_alignment = 0;
214 	}
215 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
216 	    allow_largepage_alignment &&
217 		(len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
218 		align_amount = MMU_PAGESIZE256M;
219 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
220 	    allow_largepage_alignment &&
221 		(len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
222 		align_amount = MMU_PAGESIZE32M;
223 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
224 		align_amount = MMU_PAGESIZE4M;
225 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
226 		align_amount = MMU_PAGESIZE512K;
227 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
228 		align_amount = MMU_PAGESIZE64K;
229 	} else  {
230 		/*
231 		 * Align virtual addresses on a 64K boundary to ensure
232 		 * that ELF shared libraries are mapped with the appropriate
233 		 * alignment constraints by the run-time linker.
234 		 */
235 		align_amount = ELF_SPARC_MAXPGSZ;
236 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
237 			((uintptr_t)*addrp < align_amount))
238 			align_amount = (uintptr_t)*addrp;
239 	}
240 
241 	/*
242 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
243 	 */
244 	if (p->p_model == DATAMODEL_LP64)
245 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
246 #ifdef VAC
247 	if (vac && vacalign && (align_amount < shm_alignment))
248 		align_amount = shm_alignment;
249 #endif
250 
251 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
252 		align_amount = (uintptr_t)*addrp;
253 	}
254 	len += align_amount;
255 
256 	/*
257 	 * Look for a large enough hole starting below the stack limit.
258 	 * After finding it, use the upper part.  Addition of PAGESIZE is
259 	 * for the redzone as described above.
260 	 */
261 	as_purge(as);
262 	if (as_gap(as, len, &base, &slen, AH_HI, NULL) == 0) {
263 		caddr_t as_addr;
264 
265 		addr = base + slen - len + PAGESIZE;
266 		as_addr = addr;
267 		/*
268 		 * Round address DOWN to the alignment amount,
269 		 * add the offset, and if this address is less
270 		 * than the original address, add alignment amount.
271 		 */
272 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
273 		addr += (long)(off & (align_amount - 1l));
274 		if (addr < as_addr) {
275 			addr += align_amount;
276 		}
277 
278 		ASSERT(addr <= (as_addr + align_amount));
279 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
280 		    ((uintptr_t)(off & (align_amount - 1l))));
281 		*addrp = addr;
282 
283 	} else {
284 		*addrp = NULL;	/* no more virtual space */
285 	}
286 }
287 
288 /* Auto large page tunables. */
289 int auto_lpg_tlb_threshold = 32;
290 int auto_lpg_minszc = TTE64K;
291 int auto_lpg_maxszc = TTE256M;
292 size_t auto_lpg_heap_default = MMU_PAGESIZE64K;
293 size_t auto_lpg_stack_default = MMU_PAGESIZE64K;
294 size_t auto_lpg_va_default = MMU_PAGESIZE64K;
295 size_t auto_lpg_remap_threshold = 0; /* always remap */
296 /*
297  * Number of pages in 1 GB.  Don't enable automatic large pages if we have
298  * fewer than this many pages.
299  */
300 pgcnt_t auto_lpg_min_physmem = 1 << (30 - MMU_PAGESHIFT);
301 
302 size_t
303 map_pgsz(int maptype, struct proc *p, caddr_t addr, size_t len, int *remap)
304 {
305 	uint_t	n;
306 	size_t	pgsz = 0;
307 
308 	if (remap)
309 		*remap = (len > auto_lpg_remap_threshold);
310 
311 	switch (maptype) {
312 	case MAPPGSZ_ISM:
313 		n = hat_preferred_pgsz(p->p_as->a_hat, addr, len, maptype);
314 		pgsz = hw_page_array[n].hp_size;
315 		break;
316 
317 	case MAPPGSZ_VA:
318 		n = hat_preferred_pgsz(p->p_as->a_hat, addr, len, maptype);
319 		pgsz = hw_page_array[n].hp_size;
320 		if ((pgsz <= MMU_PAGESIZE) ||
321 		    !IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz))
322 			pgsz = map_pgszva(p, addr, len);
323 		break;
324 
325 	case MAPPGSZ_STK:
326 		pgsz = map_pgszstk(p, addr, len);
327 		break;
328 
329 	case MAPPGSZ_HEAP:
330 		pgsz = map_pgszheap(p, addr, len);
331 		break;
332 	}
333 	return (pgsz);
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 	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 #define	QUANTUM_SIZE	64
404 
405 static	vmem_t	*contig_mem_slab_arena;
406 static	vmem_t	*contig_mem_arena;
407 
408 uint_t contig_mem_slab_size = MMU_PAGESIZE4M;
409 
410 static void *
411 contig_mem_span_alloc(vmem_t *vmp, size_t size, int vmflag)
412 {
413 	page_t *ppl;
414 	page_t *rootpp;
415 	caddr_t addr = NULL;
416 	pgcnt_t npages = btopr(size);
417 	page_t **ppa;
418 	int pgflags;
419 	int i = 0;
420 
421 
422 	/*
423 	 * The import request should be at least
424 	 * contig_mem_slab_size because that is the
425 	 * slab arena's quantum. The size can be
426 	 * further restricted since contiguous
427 	 * allocations larger than contig_mem_slab_size
428 	 * are not supported here.
429 	 */
430 	ASSERT(size == contig_mem_slab_size);
431 
432 	if ((addr = vmem_xalloc(vmp, size, size, 0, 0,
433 	    NULL, NULL, vmflag)) == NULL) {
434 		return (NULL);
435 	}
436 
437 	/* The address should be slab-size aligned. */
438 	ASSERT(((uintptr_t)addr & (contig_mem_slab_size - 1)) == 0);
439 
440 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
441 		vmem_xfree(vmp, addr, size);
442 		return (NULL);
443 	}
444 
445 	pgflags = PG_EXCL;
446 	if ((vmflag & VM_NOSLEEP) == 0)
447 		pgflags |= PG_WAIT;
448 	if (vmflag & VM_PANIC)
449 		pgflags |= PG_PANIC;
450 	if (vmflag & VM_PUSHPAGE)
451 		pgflags |= PG_PUSHPAGE;
452 
453 	ppl = page_create_va_large(&kvp, (u_offset_t)(uintptr_t)addr, size,
454 	    pgflags, &kvseg, addr, NULL);
455 
456 	if (ppl == NULL) {
457 		vmem_xfree(vmp, addr, size);
458 		page_unresv(npages);
459 		return (NULL);
460 	}
461 
462 	rootpp = ppl;
463 	ppa = kmem_zalloc(npages * sizeof (page_t *), KM_SLEEP);
464 	while (ppl != NULL) {
465 		page_t *pp = ppl;
466 		ppa[i++] = pp;
467 		page_sub(&ppl, pp);
468 		ASSERT(page_iolock_assert(pp));
469 		page_io_unlock(pp);
470 	}
471 
472 	/*
473 	 * Load the locked entry.  It's OK to preload the entry into
474 	 * the TSB since we now support large mappings in the kernel TSB.
475 	 */
476 	hat_memload_array(kas.a_hat, (caddr_t)rootpp->p_offset, size,
477 	    ppa, (PROT_ALL & ~PROT_USER) | HAT_NOSYNC, HAT_LOAD_LOCK);
478 
479 	for (--i; i >= 0; --i) {
480 		(void) page_pp_lock(ppa[i], 0, 1);
481 		page_unlock(ppa[i]);
482 	}
483 
484 	kmem_free(ppa, npages * sizeof (page_t *));
485 	return (addr);
486 }
487 
488 void
489 contig_mem_span_free(vmem_t *vmp, void *inaddr, size_t size)
490 {
491 	page_t *pp;
492 	caddr_t addr = inaddr;
493 	caddr_t eaddr;
494 	pgcnt_t npages = btopr(size);
495 	pgcnt_t pgs_left = npages;
496 	page_t *rootpp = NULL;
497 
498 	ASSERT(((uintptr_t)addr & (contig_mem_slab_size - 1)) == 0);
499 
500 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
501 
502 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
503 		pp = page_lookup(&kvp, (u_offset_t)(uintptr_t)addr, SE_EXCL);
504 		if (pp == NULL)
505 			panic("contig_mem_span_free: page not found");
506 
507 		ASSERT(PAGE_EXCL(pp));
508 		page_pp_unlock(pp, 0, 1);
509 
510 		if (rootpp == NULL)
511 			rootpp = pp;
512 		if (--pgs_left == 0) {
513 			/*
514 			 * similar logic to segspt_free_pages, but we know we
515 			 * have one large page.
516 			 */
517 			page_destroy_pages(rootpp);
518 		}
519 	}
520 	page_unresv(npages);
521 
522 	if (vmp != NULL)
523 		vmem_xfree(vmp, inaddr, size);
524 }
525 
526 static void *
527 contig_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t size, int vmflag)
528 {
529 	return (vmem_xalloc(vmp, size, size, 0, 0, NULL, NULL, vmflag));
530 }
531 
532 /*
533  * conting_mem_alloc_align allocates real contiguous memory with the specified
534  * alignment upto contig_mem_slab_size. The alignment must be a power of 2.
535  */
536 void *
537 contig_mem_alloc_align(size_t size, size_t align)
538 {
539 	ASSERT(align <= contig_mem_slab_size);
540 
541 	if ((align & (align - 1)) != 0)
542 		return (NULL);
543 
544 	return (vmem_xalloc(contig_mem_arena, size, align, 0, 0,
545 	    NULL, NULL, VM_NOSLEEP));
546 }
547 
548 /*
549  * Allocates size aligned contiguous memory upto contig_mem_slab_size.
550  * Size must be a power of 2.
551  */
552 void *
553 contig_mem_alloc(size_t size)
554 {
555 	ASSERT((size & (size - 1)) == 0);
556 	return (contig_mem_alloc_align(size, size));
557 }
558 
559 void
560 contig_mem_free(void *vaddr, size_t size)
561 {
562 	vmem_xfree(contig_mem_arena, vaddr, size);
563 }
564 
565 /*
566  * We create a set of stacked vmem arenas to enable us to
567  * allocate large >PAGESIZE chucks of contiguous Real Address space
568  * This is  what the Dynamics TSB support does for TSBs.
569  * The contig_mem_arena import functions are exactly the same as the
570  * TSB kmem_default arena import functions.
571  */
572 void
573 contig_mem_init(void)
574 {
575 
576 	contig_mem_slab_arena = vmem_create("contig_mem_slab_arena", NULL, 0,
577 	    contig_mem_slab_size, contig_vmem_xalloc_aligned_wrapper,
578 	    vmem_xfree, heap_arena, 0, VM_SLEEP);
579 
580 	contig_mem_arena = vmem_create("contig_mem_arena", NULL, 0,
581 	    QUANTUM_SIZE, contig_mem_span_alloc, contig_mem_span_free,
582 	    contig_mem_slab_arena, 0, VM_SLEEP | VM_BESTFIT);
583 
584 }
585