xref: /titanic_41/usr/src/uts/sun4u/vm/mach_vm_dep.c (revision 8ad60789b9913d5b2118fe76d8cd7a661d8ae8b6)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
28 /*	All Rights Reserved   */
29 
30 /*
31  * Portions of this source code were derived from Berkeley 4.3 BSD
32  * under license from the Regents of the University of California.
33  */
34 
35 #pragma ident	"%Z%%M%	%I%	%E% SMI"
36 
37 /*
38  * UNIX machine dependent virtual memory support.
39  */
40 
41 #include <sys/vm.h>
42 #include <sys/exec.h>
43 #include <sys/cmn_err.h>
44 #include <sys/cpu_module.h>
45 #include <sys/cpu.h>
46 #include <sys/elf_SPARC.h>
47 #include <sys/archsystm.h>
48 #include <vm/hat_sfmmu.h>
49 #include <sys/memnode.h>
50 #include <sys/mem_cage.h>
51 #include <vm/vm_dep.h>
52 
53 #if defined(__sparcv9) && defined(SF_ERRATA_57)
54 caddr_t errata57_limit;
55 #endif
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 
62 uint_t mmu_page_sizes = DEFAULT_MMU_PAGE_SIZES;
63 uint_t max_mmu_page_sizes = MMU_PAGE_SIZES;
64 uint_t mmu_hashcnt = DEFAULT_MAX_HASHCNT;
65 uint_t max_mmu_hashcnt = MAX_HASHCNT;
66 size_t mmu_ism_pagesize = DEFAULT_ISM_PAGESIZE;
67 
68 /*
69  * The sun4u hardware mapping sizes which will always be supported are
70  * 8K, 64K, 512K and 4M.  If sun4u based machines need to support other
71  * page sizes, platform or cpu specific routines need to modify the value.
72  * The base pagesize (p_szc == 0) must always be supported by the hardware.
73  */
74 int mmu_exported_pagesize_mask = (1 << TTE8K) | (1 << TTE64K) |
75 	(1 << TTE512K) | (1 << TTE4M);
76 uint_t mmu_exported_page_sizes;
77 
78 uint_t szc_2_userszc[MMU_PAGE_SIZES];
79 uint_t userszc_2_szc[MMU_PAGE_SIZES];
80 
81 extern uint_t vac_colors_mask;
82 extern int vac_shift;
83 
84 hw_pagesize_t hw_page_array[] = {
85 	{MMU_PAGESIZE, MMU_PAGESHIFT, MMU_PAGESIZE >> MMU_PAGESHIFT},
86 	{MMU_PAGESIZE64K, MMU_PAGESHIFT64K, MMU_PAGESIZE64K >> MMU_PAGESHIFT},
87 	{MMU_PAGESIZE512K, MMU_PAGESHIFT512K,
88 	    MMU_PAGESIZE512K >> MMU_PAGESHIFT},
89 	{MMU_PAGESIZE4M, MMU_PAGESHIFT4M, MMU_PAGESIZE4M >> MMU_PAGESHIFT},
90 	{MMU_PAGESIZE32M, MMU_PAGESHIFT32M, MMU_PAGESIZE32M >> MMU_PAGESHIFT},
91 	{MMU_PAGESIZE256M, MMU_PAGESHIFT256M,
92 	    MMU_PAGESIZE256M >> MMU_PAGESHIFT},
93 	{0, 0, 0}
94 };
95 
96 /*
97  * use_text_pgsz64k, use_initdata_pgsz64k and use_text_pgsz4m
98  * can be set in platform or CPU specific code but user can change the
99  * default values via /etc/system.
100  */
101 
102 int	use_text_pgsz64k = 0;
103 int	use_text_pgsz4m = 0;
104 int	use_initdata_pgsz64k = 0;
105 
106 /*
107  * disable_text_largepages and disable_initdata_largepages bitmaks are set in
108  * platform or CPU specific code to disable page sizes that should not be
109  * used. These variables normally shouldn't be changed via /etc/system. A
110  * particular page size for text or inititialized data will be used by default
111  * if both one of use_* variables is set to 1 AND this page size is not
112  * disabled in the corresponding disable_* bitmask variable.
113  */
114 
115 int disable_text_largepages = (1 << TTE4M) | (1 << TTE64K);
116 int disable_initdata_largepages = (1 << TTE64K);
117 
118 /*
119  * Minimum segment size tunables before 64K or 4M large pages
120  * should be used to map it.
121  */
122 size_t text_pgsz64k_minsize = MMU_PAGESIZE64K;
123 size_t text_pgsz4m_minsize = MMU_PAGESIZE4M;
124 size_t initdata_pgsz64k_minsize = MMU_PAGESIZE64K;
125 
126 /*
127  * map_addr_proc() is the routine called when the system is to
128  * choose an address for the user.  We will pick an address
129  * range which is just below the current stack limit.  The
130  * algorithm used for cache consistency on machines with virtual
131  * address caches is such that offset 0 in the vnode is always
132  * on a shm_alignment'ed aligned address.  Unfortunately, this
133  * means that vnodes which are demand paged will not be mapped
134  * cache consistently with the executable images.  When the
135  * cache alignment for a given object is inconsistent, the
136  * lower level code must manage the translations so that this
137  * is not seen here (at the cost of efficiency, of course).
138  *
139  * addrp is a value/result parameter.
140  *	On input it is a hint from the user to be used in a completely
141  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
142  *	minimal alignment.
143  *
144  *	On output it is NULL if no address can be found in the current
145  *	processes address space or else an address that is currently
146  *	not mapped for len bytes with a page of red zone on either side.
147  *	If vacalign is true, then the selected address will obey the alignment
148  *	constraints of a vac machine based on the given off value.
149  */
150 /*ARGSUSED4*/
151 void
152 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
153     caddr_t userlimit, struct proc *p, uint_t flags)
154 {
155 	struct as *as = p->p_as;
156 	caddr_t addr;
157 	caddr_t base;
158 	size_t slen;
159 	uintptr_t align_amount;
160 	int allow_largepage_alignment = 1;
161 
162 	base = p->p_brkbase;
163 	if (userlimit < as->a_userlimit) {
164 		/*
165 		 * This happens when a program wants to map something in
166 		 * a range that's accessible to a program in a smaller
167 		 * address space.  For example, a 64-bit program might
168 		 * be calling mmap32(2) to guarantee that the returned
169 		 * address is below 4Gbytes.
170 		 */
171 		ASSERT(userlimit > base);
172 		slen = userlimit - base;
173 	} else {
174 		slen = p->p_usrstack - base - (((size_t)rctl_enforced_value(
175 		    rctlproc_legacy[RLIMIT_STACK], p->p_rctls, p) + PAGEOFFSET)
176 		    & PAGEMASK);
177 	}
178 	len = (len + PAGEOFFSET) & PAGEMASK;
179 
180 	/*
181 	 * Redzone for each side of the request. This is done to leave
182 	 * one page unmapped between segments. This is not required, but
183 	 * it's useful for the user because if their program strays across
184 	 * a segment boundary, it will catch a fault immediately making
185 	 * debugging a little easier.
186 	 */
187 	len += (2 * PAGESIZE);
188 
189 	/*
190 	 *  If the request is larger than the size of a particular
191 	 *  mmu level, then we use that level to map the request.
192 	 *  But this requires that both the virtual and the physical
193 	 *  addresses be aligned with respect to that level, so we
194 	 *  do the virtual bit of nastiness here.
195 	 *
196 	 *  For 32-bit processes, only those which have specified
197 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
198 	 *  we can potentially waste up to 256MB of the 4G process address
199 	 *  space just for alignment.
200 	 */
201 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
202 	    ((uintptr_t)*addrp) != 0)) {
203 		allow_largepage_alignment = 0;
204 	}
205 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
206 	    allow_largepage_alignment &&
207 		(len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
208 		align_amount = MMU_PAGESIZE256M;
209 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
210 	    allow_largepage_alignment &&
211 		(len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
212 		align_amount = MMU_PAGESIZE32M;
213 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
214 		align_amount = MMU_PAGESIZE4M;
215 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
216 		align_amount = MMU_PAGESIZE512K;
217 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
218 		align_amount = MMU_PAGESIZE64K;
219 	} else  {
220 		/*
221 		 * Align virtual addresses on a 64K boundary to ensure
222 		 * that ELF shared libraries are mapped with the appropriate
223 		 * alignment constraints by the run-time linker.
224 		 */
225 		align_amount = ELF_SPARC_MAXPGSZ;
226 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
227 			((uintptr_t)*addrp < align_amount))
228 			align_amount = (uintptr_t)*addrp;
229 	}
230 
231 	/*
232 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
233 	 */
234 	if (p->p_model == DATAMODEL_LP64)
235 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
236 #ifdef VAC
237 	if (vac && vacalign && (align_amount < shm_alignment))
238 		align_amount = shm_alignment;
239 #endif
240 
241 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
242 		align_amount = (uintptr_t)*addrp;
243 	}
244 	len += align_amount;
245 
246 	/*
247 	 * Look for a large enough hole starting below the stack limit.
248 	 * After finding it, use the upper part.  Addition of PAGESIZE is
249 	 * for the redzone as described above.
250 	 */
251 	as_purge(as);
252 	if (as_gap(as, len, &base, &slen, AH_HI, NULL) == 0) {
253 		caddr_t as_addr;
254 
255 		addr = base + slen - len + PAGESIZE;
256 		as_addr = addr;
257 		/*
258 		 * Round address DOWN to the alignment amount,
259 		 * add the offset, and if this address is less
260 		 * than the original address, add alignment amount.
261 		 */
262 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
263 		addr += (long)(off & (align_amount - 1l));
264 		if (addr < as_addr) {
265 			addr += align_amount;
266 		}
267 
268 		ASSERT(addr <= (as_addr + align_amount));
269 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
270 		    ((uintptr_t)(off & (align_amount - 1l))));
271 		*addrp = addr;
272 
273 #if defined(SF_ERRATA_57)
274 		if (AS_TYPE_64BIT(as) && addr < errata57_limit) {
275 			*addrp = NULL;
276 		}
277 #endif
278 	} else {
279 		*addrp = NULL;	/* no more virtual space */
280 	}
281 }
282 
283 /*
284  * Platforms with smaller or larger TLBs may wish to change this.  Most
285  * sun4u platforms can hold 1024 8K entries by default and most processes
286  * are observed to be < 6MB on these machines, so we decide to move up
287  * here to give ourselves some wiggle room for other, smaller segments.
288  */
289 int auto_lpg_tlb_threshold = 768;
290 int auto_lpg_minszc = TTE4M;
291 int auto_lpg_maxszc = TTE4M;
292 size_t auto_lpg_heap_default = MMU_PAGESIZE;
293 size_t auto_lpg_stack_default = MMU_PAGESIZE;
294 size_t auto_lpg_va_default = MMU_PAGESIZE;
295 size_t auto_lpg_remap_threshold = 0;
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 /*
303  * Suggest a page size to be used to map a segment of type maptype and length
304  * len.  Returns a page size (not a size code).
305  * If remap is non-NULL, fill in a value suggesting whether or not to remap
306  * this segment.
307  */
308 size_t
309 map_pgsz(int maptype, struct proc *p, caddr_t addr, size_t len, int *remap)
310 {
311 	uint_t	n;
312 	size_t	pgsz = 0;
313 
314 	if (remap)
315 		*remap = (len > auto_lpg_remap_threshold);
316 
317 	switch (maptype) {
318 	case MAPPGSZ_ISM:
319 		n = hat_preferred_pgsz(p->p_as->a_hat, addr, len, maptype);
320 		pgsz = hw_page_array[n].hp_size;
321 
322 		/*
323 		 * For non-Panther systems, the following code sets the [D]ISM
324 		 * pagesize to 4M if either of the DTLBs happens to be
325 		 * programmed to a different large pagesize.
326 		 * The Panther code might hit this case as well,
327 		 * if and only if the addr is not aligned to >= 4M.
328 		 */
329 		if ((pgsz > 0) && (pgsz < MMU_PAGESIZE4M))
330 			pgsz = MMU_PAGESIZE4M;
331 		break;
332 
333 	case MAPPGSZ_VA:
334 		n = hat_preferred_pgsz(p->p_as->a_hat, addr, len, maptype);
335 		pgsz = hw_page_array[n].hp_size;
336 		if ((pgsz <= MMU_PAGESIZE) ||
337 		    !IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz))
338 			pgsz = map_pgszva(p, addr, len);
339 		break;
340 
341 	case MAPPGSZ_STK:
342 		pgsz = map_pgszstk(p, addr, len);
343 		break;
344 
345 	case MAPPGSZ_HEAP:
346 		pgsz = map_pgszheap(p, addr, len);
347 		break;
348 	}
349 	return (pgsz);
350 }
351 
352 /*
353  * Platform-dependent page scrub call.
354  */
355 void
356 pagescrub(page_t *pp, uint_t off, uint_t len)
357 {
358 	/*
359 	 * For now, we rely on the fact that pagezero() will
360 	 * always clear UEs.
361 	 */
362 	pagezero(pp, off, len);
363 }
364 
365 /*ARGSUSED*/
366 void
367 sync_data_memory(caddr_t va, size_t len)
368 {
369 	cpu_flush_ecache();
370 }
371 
372 /*
373  * platform specific large pages for kernel heap support
374  */
375 void
376 mmu_init_kcontext()
377 {
378 	extern void set_kcontextreg();
379 
380 	if (kcontextreg)
381 		set_kcontextreg();
382 }
383 
384 void
385 contig_mem_init(void)
386 {
387 	/* not applicable to sun4u */
388 }
389