xref: /titanic_41/usr/src/uts/sun4u/vm/mach_vm_dep.c (revision 18c2aff776a775d34a4c9893a4c72e0434d68e36)
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 
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 
61 uint_t mmu_page_sizes = DEFAULT_MMU_PAGE_SIZES;
62 uint_t max_mmu_page_sizes = MMU_PAGE_SIZES;
63 uint_t mmu_hashcnt = DEFAULT_MAX_HASHCNT;
64 uint_t max_mmu_hashcnt = MAX_HASHCNT;
65 size_t mmu_ism_pagesize = DEFAULT_ISM_PAGESIZE;
66 
67 /*
68  * The sun4u hardware mapping sizes which will always be supported are
69  * 8K, 64K, 512K and 4M.  If sun4u based machines need to support other
70  * page sizes, platform or cpu specific routines need to modify the value.
71  * The base pagesize (p_szc == 0) must always be supported by the hardware.
72  */
73 int mmu_exported_pagesize_mask = (1 << TTE8K) | (1 << TTE64K) |
74 	(1 << TTE512K) | (1 << TTE4M);
75 uint_t mmu_exported_page_sizes;
76 
77 uint_t szc_2_userszc[MMU_PAGE_SIZES];
78 uint_t userszc_2_szc[MMU_PAGE_SIZES];
79 
80 extern uint_t vac_colors_mask;
81 extern int vac_shift;
82 
83 hw_pagesize_t hw_page_array[] = {
84 	{MMU_PAGESIZE, MMU_PAGESHIFT, MMU_PAGESIZE >> MMU_PAGESHIFT},
85 	{MMU_PAGESIZE64K, MMU_PAGESHIFT64K, MMU_PAGESIZE64K >> MMU_PAGESHIFT},
86 	{MMU_PAGESIZE512K, MMU_PAGESHIFT512K,
87 	    MMU_PAGESIZE512K >> MMU_PAGESHIFT},
88 	{MMU_PAGESIZE4M, MMU_PAGESHIFT4M, MMU_PAGESIZE4M >> MMU_PAGESHIFT},
89 	{MMU_PAGESIZE32M, MMU_PAGESHIFT32M, MMU_PAGESIZE32M >> MMU_PAGESHIFT},
90 	{MMU_PAGESIZE256M, MMU_PAGESHIFT256M,
91 	    MMU_PAGESIZE256M >> MMU_PAGESHIFT},
92 	{0, 0, 0}
93 };
94 
95 /*
96  * use_text_pgsz64k, use_initdata_pgsz64k and use_text_pgsz4m
97  * can be set in platform or CPU specific code but user can change the
98  * default values via /etc/system.
99  */
100 
101 int	use_text_pgsz64k = 0;
102 int	use_text_pgsz4m = 0;
103 int	use_initdata_pgsz64k = 0;
104 
105 /*
106  * disable_text_largepages and disable_initdata_largepages bitmaks are set in
107  * platform or CPU specific code to disable page sizes that should not be
108  * used. These variables normally shouldn't be changed via /etc/system. A
109  * particular page size for text or inititialized data will be used by default
110  * if both one of use_* variables is set to 1 AND this page size is not
111  * disabled in the corresponding disable_* bitmask variable.
112  */
113 
114 int disable_text_largepages = (1 << TTE4M) | (1 << TTE64K);
115 int disable_initdata_largepages = (1 << TTE64K);
116 
117 /*
118  * Minimum segment size tunables before 64K or 4M large pages
119  * should be used to map it.
120  */
121 size_t text_pgsz64k_minsize = MMU_PAGESIZE64K;
122 size_t text_pgsz4m_minsize = MMU_PAGESIZE4M;
123 size_t initdata_pgsz64k_minsize = MMU_PAGESIZE64K;
124 
125 size_t max_shm_lpsize = ULONG_MAX;
126 
127 /*
128  * Platforms with smaller or larger TLBs may wish to change this.  Most
129  * sun4u platforms can hold 1024 8K entries by default and most processes
130  * are observed to be < 6MB on these machines, so we decide to move up
131  * here to give ourselves some wiggle room for other, smaller segments.
132  */
133 int auto_lpg_tlb_threshold = 768;
134 int auto_lpg_minszc = TTE4M;
135 int auto_lpg_maxszc = TTE4M;
136 size_t auto_lpg_heap_default = MMU_PAGESIZE;
137 size_t auto_lpg_stack_default = MMU_PAGESIZE;
138 size_t auto_lpg_va_default = MMU_PAGESIZE;
139 size_t auto_lpg_remap_threshold = 0;
140 /*
141  * Number of pages in 1 GB.  Don't enable automatic large pages if we have
142  * fewer than this many pages.
143  */
144 pgcnt_t auto_lpg_min_physmem = 1 << (30 - MMU_PAGESHIFT);
145 
146 /*
147  * map_addr_proc() is the routine called when the system is to
148  * choose an address for the user.  We will pick an address
149  * range which is just below the current stack limit.  The
150  * algorithm used for cache consistency on machines with virtual
151  * address caches is such that offset 0 in the vnode is always
152  * on a shm_alignment'ed aligned address.  Unfortunately, this
153  * means that vnodes which are demand paged will not be mapped
154  * cache consistently with the executable images.  When the
155  * cache alignment for a given object is inconsistent, the
156  * lower level code must manage the translations so that this
157  * is not seen here (at the cost of efficiency, of course).
158  *
159  * addrp is a value/result parameter.
160  *	On input it is a hint from the user to be used in a completely
161  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
162  *	minimal alignment.
163  *
164  *	On output it is NULL if no address can be found in the current
165  *	processes address space or else an address that is currently
166  *	not mapped for len bytes with a page of red zone on either side.
167  *	If vacalign is true, then the selected address will obey the alignment
168  *	constraints of a vac machine based on the given off value.
169  */
170 /*ARGSUSED4*/
171 void
172 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
173     caddr_t userlimit, struct proc *p, uint_t flags)
174 {
175 	struct as *as = p->p_as;
176 	caddr_t addr;
177 	caddr_t base;
178 	size_t slen;
179 	uintptr_t align_amount;
180 	int allow_largepage_alignment = 1;
181 
182 	base = p->p_brkbase;
183 	if (userlimit < as->a_userlimit) {
184 		/*
185 		 * This happens when a program wants to map something in
186 		 * a range that's accessible to a program in a smaller
187 		 * address space.  For example, a 64-bit program might
188 		 * be calling mmap32(2) to guarantee that the returned
189 		 * address is below 4Gbytes.
190 		 */
191 		ASSERT(userlimit > base);
192 		slen = userlimit - base;
193 	} else {
194 		slen = p->p_usrstack - base - (((size_t)rctl_enforced_value(
195 		    rctlproc_legacy[RLIMIT_STACK], p->p_rctls, p) + PAGEOFFSET)
196 		    & PAGEMASK);
197 	}
198 	len = (len + PAGEOFFSET) & PAGEMASK;
199 
200 	/*
201 	 * Redzone for each side of the request. This is done to leave
202 	 * one page unmapped between segments. This is not required, but
203 	 * it's useful for the user because if their program strays across
204 	 * a segment boundary, it will catch a fault immediately making
205 	 * debugging a little easier.
206 	 */
207 	len += (2 * PAGESIZE);
208 
209 	/*
210 	 *  If the request is larger than the size of a particular
211 	 *  mmu level, then we use that level to map the request.
212 	 *  But this requires that both the virtual and the physical
213 	 *  addresses be aligned with respect to that level, so we
214 	 *  do the virtual bit of nastiness here.
215 	 *
216 	 *  For 32-bit processes, only those which have specified
217 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
218 	 *  we can potentially waste up to 256MB of the 4G process address
219 	 *  space just for alignment.
220 	 */
221 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
222 	    ((uintptr_t)*addrp) != 0)) {
223 		allow_largepage_alignment = 0;
224 	}
225 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
226 	    allow_largepage_alignment &&
227 		(len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
228 		align_amount = MMU_PAGESIZE256M;
229 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
230 	    allow_largepage_alignment &&
231 		(len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
232 		align_amount = MMU_PAGESIZE32M;
233 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
234 		align_amount = MMU_PAGESIZE4M;
235 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
236 		align_amount = MMU_PAGESIZE512K;
237 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
238 		align_amount = MMU_PAGESIZE64K;
239 	} else  {
240 		/*
241 		 * Align virtual addresses on a 64K boundary to ensure
242 		 * that ELF shared libraries are mapped with the appropriate
243 		 * alignment constraints by the run-time linker.
244 		 */
245 		align_amount = ELF_SPARC_MAXPGSZ;
246 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
247 			((uintptr_t)*addrp < align_amount))
248 			align_amount = (uintptr_t)*addrp;
249 	}
250 
251 	/*
252 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
253 	 */
254 	if (p->p_model == DATAMODEL_LP64)
255 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
256 #ifdef VAC
257 	if (vac && vacalign && (align_amount < shm_alignment))
258 		align_amount = shm_alignment;
259 #endif
260 
261 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
262 		align_amount = (uintptr_t)*addrp;
263 	}
264 	len += align_amount;
265 
266 	/*
267 	 * Look for a large enough hole starting below the stack limit.
268 	 * After finding it, use the upper part.  Addition of PAGESIZE is
269 	 * for the redzone as described above.
270 	 */
271 	as_purge(as);
272 	if (as_gap(as, len, &base, &slen, AH_HI, NULL) == 0) {
273 		caddr_t as_addr;
274 
275 		addr = base + slen - len + PAGESIZE;
276 		as_addr = addr;
277 		/*
278 		 * Round address DOWN to the alignment amount,
279 		 * add the offset, and if this address is less
280 		 * than the original address, add alignment amount.
281 		 */
282 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
283 		addr += (long)(off & (align_amount - 1l));
284 		if (addr < as_addr) {
285 			addr += align_amount;
286 		}
287 
288 		ASSERT(addr <= (as_addr + align_amount));
289 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
290 		    ((uintptr_t)(off & (align_amount - 1l))));
291 		*addrp = addr;
292 
293 #if defined(SF_ERRATA_57)
294 		if (AS_TYPE_64BIT(as) && addr < errata57_limit) {
295 			*addrp = NULL;
296 		}
297 #endif
298 	} else {
299 		*addrp = NULL;	/* no more virtual space */
300 	}
301 }
302 
303 /*
304  * Platform-dependent page scrub call.
305  */
306 void
307 pagescrub(page_t *pp, uint_t off, uint_t len)
308 {
309 	/*
310 	 * For now, we rely on the fact that pagezero() will
311 	 * always clear UEs.
312 	 */
313 	pagezero(pp, off, len);
314 }
315 
316 /*ARGSUSED*/
317 void
318 sync_data_memory(caddr_t va, size_t len)
319 {
320 	cpu_flush_ecache();
321 }
322 
323 /*
324  * platform specific large pages for kernel heap support
325  */
326 void
327 mmu_init_kcontext()
328 {
329 	extern void set_kcontextreg();
330 
331 	if (kcontextreg)
332 		set_kcontextreg();
333 }
334 
335 void
336 contig_mem_init(void)
337 {
338 	/* not applicable to sun4u */
339 }
340