xref: /titanic_41/usr/src/uts/sun4v/vm/mach_vm_dep.c (revision 7010c12ad3ac2cada55cf126121a8c46957d3632)
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 = TTE64K;
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 		pgsz = map_pgszva(p, addr, len);
321 		break;
322 
323 	case MAPPGSZ_STK:
324 		pgsz = map_pgszstk(p, addr, len);
325 		break;
326 
327 	case MAPPGSZ_HEAP:
328 		pgsz = map_pgszheap(p, addr, len);
329 		break;
330 	}
331 	return (pgsz);
332 }
333 
334 /*
335  * Platform-dependent page scrub call.
336  * We call hypervisor to scrub the page.
337  */
338 void
339 pagescrub(page_t *pp, uint_t off, uint_t len)
340 {
341 	uint64_t pa, length;
342 
343 	pa = (uint64_t)(pp->p_pagenum << MMU_PAGESHIFT + off);
344 	length = (uint64_t)len;
345 
346 	(void) mem_scrub(pa, length);
347 }
348 
349 void
350 sync_data_memory(caddr_t va, size_t len)
351 {
352 	/* Call memory sync function */
353 	mem_sync(va, len);
354 }
355 
356 size_t
357 mmu_get_kernel_lpsize(size_t lpsize)
358 {
359 	extern int mmu_exported_pagesize_mask;
360 	uint_t tte;
361 
362 	if (lpsize == 0) {
363 		/* no setting for segkmem_lpsize in /etc/system: use default */
364 		if (mmu_exported_pagesize_mask & (1 << TTE256M)) {
365 			lpsize = MMU_PAGESIZE256M;
366 		} else if (mmu_exported_pagesize_mask & (1 << TTE4M)) {
367 			lpsize = MMU_PAGESIZE4M;
368 		} else if (mmu_exported_pagesize_mask & (1 << TTE64K)) {
369 			lpsize = MMU_PAGESIZE64K;
370 		} else {
371 			lpsize = MMU_PAGESIZE;
372 		}
373 
374 		return (lpsize);
375 	}
376 
377 	for (tte = TTE8K; tte <= TTE256M; tte++) {
378 
379 		if ((mmu_exported_pagesize_mask & (1 << tte)) == 0)
380 			continue;
381 
382 		if (lpsize == TTEBYTES(tte))
383 			return (lpsize);
384 	}
385 
386 	lpsize = TTEBYTES(TTE8K);
387 	return (lpsize);
388 }
389 
390 void
391 mmu_init_kcontext()
392 {
393 }
394 
395 /*ARGSUSED*/
396 void
397 mmu_init_kernel_pgsz(struct hat *hat)
398 {
399 }
400 
401 #define	QUANTUM_SIZE	64
402 
403 static	vmem_t	*contig_mem_slab_arena;
404 static	vmem_t	*contig_mem_arena;
405 
406 uint_t contig_mem_slab_size = MMU_PAGESIZE4M;
407 
408 static void *
409 contig_mem_span_alloc(vmem_t *vmp, size_t size, int vmflag)
410 {
411 	page_t *ppl;
412 	page_t *rootpp;
413 	caddr_t addr = NULL;
414 	pgcnt_t npages = btopr(size);
415 	page_t **ppa;
416 	int pgflags;
417 	int i = 0;
418 
419 
420 	/*
421 	 * The import request should be at least
422 	 * contig_mem_slab_size because that is the
423 	 * slab arena's quantum. The size can be
424 	 * further restricted since contiguous
425 	 * allocations larger than contig_mem_slab_size
426 	 * are not supported here.
427 	 */
428 	ASSERT(size == contig_mem_slab_size);
429 
430 	if ((addr = vmem_xalloc(vmp, size, size, 0, 0,
431 	    NULL, NULL, vmflag)) == NULL) {
432 		return (NULL);
433 	}
434 
435 	/* The address should be slab-size aligned. */
436 	ASSERT(((uintptr_t)addr & (contig_mem_slab_size - 1)) == 0);
437 
438 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
439 		vmem_xfree(vmp, addr, size);
440 		return (NULL);
441 	}
442 
443 	pgflags = PG_EXCL;
444 	if ((vmflag & VM_NOSLEEP) == 0)
445 		pgflags |= PG_WAIT;
446 	if (vmflag & VM_PANIC)
447 		pgflags |= PG_PANIC;
448 	if (vmflag & VM_PUSHPAGE)
449 		pgflags |= PG_PUSHPAGE;
450 
451 	ppl = page_create_va_large(&kvp, (u_offset_t)(uintptr_t)addr, size,
452 	    pgflags, &kvseg, addr, NULL);
453 
454 	if (ppl == NULL) {
455 		vmem_xfree(vmp, addr, size);
456 		page_unresv(npages);
457 		return (NULL);
458 	}
459 
460 	rootpp = ppl;
461 	ppa = kmem_zalloc(npages * sizeof (page_t *), KM_SLEEP);
462 	while (ppl != NULL) {
463 		page_t *pp = ppl;
464 		ppa[i++] = pp;
465 		page_sub(&ppl, pp);
466 		ASSERT(page_iolock_assert(pp));
467 		page_io_unlock(pp);
468 	}
469 
470 	/*
471 	 * Load the locked entry.  It's OK to preload the entry into
472 	 * the TSB since we now support large mappings in the kernel TSB.
473 	 */
474 	hat_memload_array(kas.a_hat, (caddr_t)rootpp->p_offset, size,
475 	    ppa, (PROT_ALL & ~PROT_USER) | HAT_NOSYNC, HAT_LOAD_LOCK);
476 
477 	for (--i; i >= 0; --i) {
478 		(void) page_pp_lock(ppa[i], 0, 1);
479 		page_unlock(ppa[i]);
480 	}
481 
482 	kmem_free(ppa, npages * sizeof (page_t *));
483 	return (addr);
484 }
485 
486 void
487 contig_mem_span_free(vmem_t *vmp, void *inaddr, size_t size)
488 {
489 	page_t *pp;
490 	caddr_t addr = inaddr;
491 	caddr_t eaddr;
492 	pgcnt_t npages = btopr(size);
493 	pgcnt_t pgs_left = npages;
494 	page_t *rootpp = NULL;
495 
496 	ASSERT(((uintptr_t)addr & (contig_mem_slab_size - 1)) == 0);
497 
498 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
499 
500 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
501 		pp = page_lookup(&kvp, (u_offset_t)(uintptr_t)addr, SE_EXCL);
502 		if (pp == NULL)
503 			panic("contig_mem_span_free: page not found");
504 
505 		ASSERT(PAGE_EXCL(pp));
506 		page_pp_unlock(pp, 0, 1);
507 
508 		if (rootpp == NULL)
509 			rootpp = pp;
510 		if (--pgs_left == 0) {
511 			/*
512 			 * similar logic to segspt_free_pages, but we know we
513 			 * have one large page.
514 			 */
515 			page_destroy_pages(rootpp);
516 		}
517 	}
518 	page_unresv(npages);
519 
520 	if (vmp != NULL)
521 		vmem_xfree(vmp, inaddr, size);
522 }
523 
524 static void *
525 contig_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t size, int vmflag)
526 {
527 	return (vmem_xalloc(vmp, size, size, 0, 0, NULL, NULL, vmflag));
528 }
529 
530 /*
531  * conting_mem_alloc_align allocates real contiguous memory with the specified
532  * alignment upto contig_mem_slab_size. The alignment must be a power of 2.
533  */
534 void *
535 contig_mem_alloc_align(size_t size, size_t align)
536 {
537 	ASSERT(align <= contig_mem_slab_size);
538 
539 	if ((align & (align - 1)) != 0)
540 		return (NULL);
541 
542 	return (vmem_xalloc(contig_mem_arena, size, align, 0, 0,
543 	    NULL, NULL, VM_NOSLEEP));
544 }
545 
546 /*
547  * Allocates size aligned contiguous memory upto contig_mem_slab_size.
548  * Size must be a power of 2.
549  */
550 void *
551 contig_mem_alloc(size_t size)
552 {
553 	ASSERT((size & (size - 1)) == 0);
554 	return (contig_mem_alloc_align(size, size));
555 }
556 
557 void
558 contig_mem_free(void *vaddr, size_t size)
559 {
560 	vmem_xfree(contig_mem_arena, vaddr, size);
561 }
562 
563 /*
564  * We create a set of stacked vmem arenas to enable us to
565  * allocate large >PAGESIZE chucks of contiguous Real Address space
566  * This is  what the Dynamics TSB support does for TSBs.
567  * The contig_mem_arena import functions are exactly the same as the
568  * TSB kmem_default arena import functions.
569  */
570 void
571 contig_mem_init(void)
572 {
573 
574 	contig_mem_slab_arena = vmem_create("contig_mem_slab_arena", NULL, 0,
575 	    contig_mem_slab_size, contig_vmem_xalloc_aligned_wrapper,
576 	    vmem_xfree, heap_arena, 0, VM_SLEEP);
577 
578 	contig_mem_arena = vmem_create("contig_mem_arena", NULL, 0,
579 	    QUANTUM_SIZE, contig_mem_span_alloc, contig_mem_span_free,
580 	    contig_mem_slab_arena, 0, VM_SLEEP | VM_BESTFIT);
581 
582 }
583