xref: /titanic_44/usr/src/uts/sun4u/vm/mach_vm_dep.c (revision ee5416c9d7e449233197d5d20bc6b81e4ff091b2)
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 2007 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 
52 #if defined(__sparcv9) && defined(SF_ERRATA_57)
53 caddr_t errata57_limit;
54 #endif
55 
56 uint_t page_colors = 0;
57 uint_t page_colors_mask = 0;
58 uint_t page_coloring_shift = 0;
59 int consistent_coloring;
60 int update_proc_pgcolorbase_after_fork = 0;
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, 0, MMU_PAGESIZE >> MMU_PAGESHIFT},
86 	{MMU_PAGESIZE64K, MMU_PAGESHIFT64K, 0,
87 	    MMU_PAGESIZE64K >> MMU_PAGESHIFT},
88 	{MMU_PAGESIZE512K, MMU_PAGESHIFT512K, 0,
89 	    MMU_PAGESIZE512K >> MMU_PAGESHIFT},
90 	{MMU_PAGESIZE4M, MMU_PAGESHIFT4M, 0, MMU_PAGESIZE4M >> MMU_PAGESHIFT},
91 	{MMU_PAGESIZE32M, MMU_PAGESHIFT32M, 0,
92 	    MMU_PAGESIZE32M >> MMU_PAGESHIFT},
93 	{MMU_PAGESIZE256M, MMU_PAGESHIFT256M, 0,
94 	    MMU_PAGESIZE256M >> MMU_PAGESHIFT},
95 	{0, 0, 0, 0}
96 };
97 
98 /*
99  * Maximum page size used to map 64-bit memory segment kmem64_base..kmem64_end
100  */
101 int	max_bootlp_tteszc = TTE4M;
102 
103 /*
104  * use_text_pgsz64k and use_text_pgsz512k allow the user to turn on these
105  * additional text page sizes for USIII-IV+ and OPL by changing the default
106  * values via /etc/system.
107  */
108 int	use_text_pgsz64K = 0;
109 int	use_text_pgsz512K = 0;
110 
111 /*
112  * Maximum and default segment size tunables for user heap, stack, private
113  * and shared anonymous memory, and user text and initialized data.
114  */
115 size_t max_uheap_lpsize = MMU_PAGESIZE4M;
116 size_t default_uheap_lpsize = MMU_PAGESIZE;
117 size_t max_ustack_lpsize = MMU_PAGESIZE4M;
118 size_t default_ustack_lpsize = MMU_PAGESIZE;
119 size_t max_privmap_lpsize = MMU_PAGESIZE4M;
120 size_t max_uidata_lpsize = MMU_PAGESIZE;
121 size_t max_utext_lpsize = MMU_PAGESIZE4M;
122 size_t max_shm_lpsize = MMU_PAGESIZE4M;
123 
124 void
125 adjust_data_maxlpsize(size_t ismpagesize)
126 {
127 	if (max_uheap_lpsize == MMU_PAGESIZE4M) {
128 		max_uheap_lpsize = ismpagesize;
129 	}
130 	if (max_ustack_lpsize == MMU_PAGESIZE4M) {
131 		max_ustack_lpsize = ismpagesize;
132 	}
133 	if (max_privmap_lpsize == MMU_PAGESIZE4M) {
134 		max_privmap_lpsize = ismpagesize;
135 	}
136 	if (max_shm_lpsize == MMU_PAGESIZE4M) {
137 		max_shm_lpsize = ismpagesize;
138 	}
139 }
140 
141 /*
142  * map_addr_proc() is the routine called when the system is to
143  * choose an address for the user.  We will pick an address
144  * range which is just below the current stack limit.  The
145  * algorithm used for cache consistency on machines with virtual
146  * address caches is such that offset 0 in the vnode is always
147  * on a shm_alignment'ed aligned address.  Unfortunately, this
148  * means that vnodes which are demand paged will not be mapped
149  * cache consistently with the executable images.  When the
150  * cache alignment for a given object is inconsistent, the
151  * lower level code must manage the translations so that this
152  * is not seen here (at the cost of efficiency, of course).
153  *
154  * addrp is a value/result parameter.
155  *	On input it is a hint from the user to be used in a completely
156  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
157  *	minimal alignment.
158  *
159  *	On output it is NULL if no address can be found in the current
160  *	processes address space or else an address that is currently
161  *	not mapped for len bytes with a page of red zone on either side.
162  *	If vacalign is true, then the selected address will obey the alignment
163  *	constraints of a vac machine based on the given off value.
164  */
165 /*ARGSUSED4*/
166 void
167 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
168     caddr_t userlimit, struct proc *p, uint_t flags)
169 {
170 	struct as *as = p->p_as;
171 	caddr_t addr;
172 	caddr_t base;
173 	size_t slen;
174 	uintptr_t align_amount;
175 	int allow_largepage_alignment = 1;
176 
177 	base = p->p_brkbase;
178 	if (userlimit < as->a_userlimit) {
179 		/*
180 		 * This happens when a program wants to map something in
181 		 * a range that's accessible to a program in a smaller
182 		 * address space.  For example, a 64-bit program might
183 		 * be calling mmap32(2) to guarantee that the returned
184 		 * address is below 4Gbytes.
185 		 */
186 		ASSERT(userlimit > base);
187 		slen = userlimit - base;
188 	} else {
189 		slen = p->p_usrstack - base - (((size_t)rctl_enforced_value(
190 		    rctlproc_legacy[RLIMIT_STACK], p->p_rctls, p) + PAGEOFFSET)
191 		    & PAGEMASK);
192 	}
193 	len = (len + PAGEOFFSET) & PAGEMASK;
194 
195 	/*
196 	 * Redzone for each side of the request. This is done to leave
197 	 * one page unmapped between segments. This is not required, but
198 	 * it's useful for the user because if their program strays across
199 	 * a segment boundary, it will catch a fault immediately making
200 	 * debugging a little easier.
201 	 */
202 	len += (2 * PAGESIZE);
203 
204 	/*
205 	 *  If the request is larger than the size of a particular
206 	 *  mmu level, then we use that level to map the request.
207 	 *  But this requires that both the virtual and the physical
208 	 *  addresses be aligned with respect to that level, so we
209 	 *  do the virtual bit of nastiness here.
210 	 *
211 	 *  For 32-bit processes, only those which have specified
212 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
213 	 *  we can potentially waste up to 256MB of the 4G process address
214 	 *  space just for alignment.
215 	 */
216 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
217 	    ((uintptr_t)*addrp) != 0)) {
218 		allow_largepage_alignment = 0;
219 	}
220 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
221 	    allow_largepage_alignment &&
222 		(len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
223 		align_amount = MMU_PAGESIZE256M;
224 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
225 	    allow_largepage_alignment &&
226 		(len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
227 		align_amount = MMU_PAGESIZE32M;
228 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
229 		align_amount = MMU_PAGESIZE4M;
230 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
231 		align_amount = MMU_PAGESIZE512K;
232 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
233 		align_amount = MMU_PAGESIZE64K;
234 	} else  {
235 		/*
236 		 * Align virtual addresses on a 64K boundary to ensure
237 		 * that ELF shared libraries are mapped with the appropriate
238 		 * alignment constraints by the run-time linker.
239 		 */
240 		align_amount = ELF_SPARC_MAXPGSZ;
241 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
242 			((uintptr_t)*addrp < align_amount))
243 			align_amount = (uintptr_t)*addrp;
244 	}
245 
246 	/*
247 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
248 	 */
249 	if (p->p_model == DATAMODEL_LP64)
250 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
251 #ifdef VAC
252 	if (vac && vacalign && (align_amount < shm_alignment))
253 		align_amount = shm_alignment;
254 #endif
255 
256 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
257 		align_amount = (uintptr_t)*addrp;
258 	}
259 	len += align_amount;
260 
261 	/*
262 	 * Look for a large enough hole starting below the stack limit.
263 	 * After finding it, use the upper part.  Addition of PAGESIZE is
264 	 * for the redzone as described above.
265 	 */
266 	as_purge(as);
267 	if (as_gap(as, len, &base, &slen, AH_HI, NULL) == 0) {
268 		caddr_t as_addr;
269 
270 		addr = base + slen - len + PAGESIZE;
271 		as_addr = addr;
272 		/*
273 		 * Round address DOWN to the alignment amount,
274 		 * add the offset, and if this address is less
275 		 * than the original address, add alignment amount.
276 		 */
277 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
278 		addr += (long)(off & (align_amount - 1l));
279 		if (addr < as_addr) {
280 			addr += align_amount;
281 		}
282 
283 		ASSERT(addr <= (as_addr + align_amount));
284 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
285 		    ((uintptr_t)(off & (align_amount - 1l))));
286 		*addrp = addr;
287 
288 #if defined(SF_ERRATA_57)
289 		if (AS_TYPE_64BIT(as) && addr < errata57_limit) {
290 			*addrp = NULL;
291 		}
292 #endif
293 	} else {
294 		*addrp = NULL;	/* no more virtual space */
295 	}
296 }
297 
298 /*
299  * Platform-dependent page scrub call.
300  */
301 void
302 pagescrub(page_t *pp, uint_t off, uint_t len)
303 {
304 	/*
305 	 * For now, we rely on the fact that pagezero() will
306 	 * always clear UEs.
307 	 */
308 	pagezero(pp, off, len);
309 }
310 
311 /*ARGSUSED*/
312 void
313 sync_data_memory(caddr_t va, size_t len)
314 {
315 	cpu_flush_ecache();
316 }
317 
318 /*
319  * platform specific large pages for kernel heap support
320  */
321 void
322 mmu_init_kcontext()
323 {
324 	extern void set_kcontextreg();
325 
326 	if (kcontextreg)
327 		set_kcontextreg();
328 }
329 
330 void
331 contig_mem_init(void)
332 {
333 	/* not applicable to sun4u */
334 }
335 
336 /*ARGSUSED*/
337 caddr_t
338 contig_mem_prealloc(caddr_t alloc_base, pgcnt_t npages)
339 {
340 	/* not applicable to sun4u */
341 	return (alloc_base);
342 }
343 
344 size_t
345 exec_get_spslew(void)
346 {
347 	return (0);
348 }
349