xref: /illumos-gate/usr/src/uts/sun4u/vm/mach_vm_dep.c (revision eda50310abb3984bab11856a2aca8936d26881cb)
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, 0, MMU_PAGESIZE >> MMU_PAGESHIFT},
85 	{MMU_PAGESIZE64K, MMU_PAGESHIFT64K, 0,
86 	    MMU_PAGESIZE64K >> MMU_PAGESHIFT},
87 	{MMU_PAGESIZE512K, MMU_PAGESHIFT512K, 0,
88 	    MMU_PAGESIZE512K >> MMU_PAGESHIFT},
89 	{MMU_PAGESIZE4M, MMU_PAGESHIFT4M, 0, MMU_PAGESIZE4M >> MMU_PAGESHIFT},
90 	{MMU_PAGESIZE32M, MMU_PAGESHIFT32M, 0,
91 	    MMU_PAGESIZE32M >> MMU_PAGESHIFT},
92 	{MMU_PAGESIZE256M, MMU_PAGESHIFT256M, 0,
93 	    MMU_PAGESIZE256M >> MMU_PAGESHIFT},
94 	{0, 0, 0, 0}
95 };
96 
97 /*
98  * use_text_pgsz64k, use_initdata_pgsz64k and use_text_pgsz4m
99  * can be set in platform or CPU specific code but user can change the
100  * default values via /etc/system.
101  */
102 
103 int	use_text_pgsz64k = 0;
104 int	use_text_pgsz4m = 0;
105 int	use_initdata_pgsz64k = 0;
106 
107 /*
108  * disable_text_largepages and disable_initdata_largepages bitmaks are set in
109  * platform or CPU specific code to disable page sizes that should not be
110  * used. These variables normally shouldn't be changed via /etc/system. A
111  * particular page size for text or inititialized data will be used by default
112  * if both one of use_* variables is set to 1 AND this page size is not
113  * disabled in the corresponding disable_* bitmask variable.
114  */
115 
116 int disable_text_largepages = (1 << TTE4M) | (1 << TTE64K);
117 int disable_initdata_largepages = (1 << TTE64K);
118 
119 /*
120  * Minimum segment size tunables before 64K or 4M large pages
121  * should be used to map it.
122  */
123 size_t text_pgsz64k_minsize = MMU_PAGESIZE64K;
124 size_t text_pgsz4m_minsize = MMU_PAGESIZE4M;
125 size_t initdata_pgsz64k_minsize = MMU_PAGESIZE64K;
126 
127 size_t max_shm_lpsize = ULONG_MAX;
128 
129 /*
130  * Platforms with smaller or larger TLBs may wish to change this.  Most
131  * sun4u platforms can hold 1024 8K entries by default and most processes
132  * are observed to be < 6MB on these machines, so we decide to move up
133  * here to give ourselves some wiggle room for other, smaller segments.
134  */
135 int auto_lpg_tlb_threshold = 768;
136 int auto_lpg_minszc = TTE4M;
137 int auto_lpg_maxszc = TTE4M;
138 size_t auto_lpg_heap_default = MMU_PAGESIZE;
139 size_t auto_lpg_stack_default = MMU_PAGESIZE;
140 size_t auto_lpg_va_default = MMU_PAGESIZE;
141 size_t auto_lpg_remap_threshold = 0;
142 /*
143  * Number of pages in 1 GB.  Don't enable automatic large pages if we have
144  * fewer than this many pages.
145  */
146 pgcnt_t auto_lpg_min_physmem = 1 << (30 - MMU_PAGESHIFT);
147 
148 /*
149  * map_addr_proc() is the routine called when the system is to
150  * choose an address for the user.  We will pick an address
151  * range which is just below the current stack limit.  The
152  * algorithm used for cache consistency on machines with virtual
153  * address caches is such that offset 0 in the vnode is always
154  * on a shm_alignment'ed aligned address.  Unfortunately, this
155  * means that vnodes which are demand paged will not be mapped
156  * cache consistently with the executable images.  When the
157  * cache alignment for a given object is inconsistent, the
158  * lower level code must manage the translations so that this
159  * is not seen here (at the cost of efficiency, of course).
160  *
161  * addrp is a value/result parameter.
162  *	On input it is a hint from the user to be used in a completely
163  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
164  *	minimal alignment.
165  *
166  *	On output it is NULL if no address can be found in the current
167  *	processes address space or else an address that is currently
168  *	not mapped for len bytes with a page of red zone on either side.
169  *	If vacalign is true, then the selected address will obey the alignment
170  *	constraints of a vac machine based on the given off value.
171  */
172 /*ARGSUSED4*/
173 void
174 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
175     caddr_t userlimit, struct proc *p, uint_t flags)
176 {
177 	struct as *as = p->p_as;
178 	caddr_t addr;
179 	caddr_t base;
180 	size_t slen;
181 	uintptr_t align_amount;
182 	int allow_largepage_alignment = 1;
183 
184 	base = p->p_brkbase;
185 	if (userlimit < as->a_userlimit) {
186 		/*
187 		 * This happens when a program wants to map something in
188 		 * a range that's accessible to a program in a smaller
189 		 * address space.  For example, a 64-bit program might
190 		 * be calling mmap32(2) to guarantee that the returned
191 		 * address is below 4Gbytes.
192 		 */
193 		ASSERT(userlimit > base);
194 		slen = userlimit - base;
195 	} else {
196 		slen = p->p_usrstack - base - (((size_t)rctl_enforced_value(
197 		    rctlproc_legacy[RLIMIT_STACK], p->p_rctls, p) + PAGEOFFSET)
198 		    & PAGEMASK);
199 	}
200 	len = (len + PAGEOFFSET) & PAGEMASK;
201 
202 	/*
203 	 * Redzone for each side of the request. This is done to leave
204 	 * one page unmapped between segments. This is not required, but
205 	 * it's useful for the user because if their program strays across
206 	 * a segment boundary, it will catch a fault immediately making
207 	 * debugging a little easier.
208 	 */
209 	len += (2 * PAGESIZE);
210 
211 	/*
212 	 *  If the request is larger than the size of a particular
213 	 *  mmu level, then we use that level to map the request.
214 	 *  But this requires that both the virtual and the physical
215 	 *  addresses be aligned with respect to that level, so we
216 	 *  do the virtual bit of nastiness here.
217 	 *
218 	 *  For 32-bit processes, only those which have specified
219 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
220 	 *  we can potentially waste up to 256MB of the 4G process address
221 	 *  space just for alignment.
222 	 */
223 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
224 	    ((uintptr_t)*addrp) != 0)) {
225 		allow_largepage_alignment = 0;
226 	}
227 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
228 	    allow_largepage_alignment &&
229 		(len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
230 		align_amount = MMU_PAGESIZE256M;
231 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
232 	    allow_largepage_alignment &&
233 		(len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
234 		align_amount = MMU_PAGESIZE32M;
235 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
236 		align_amount = MMU_PAGESIZE4M;
237 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
238 		align_amount = MMU_PAGESIZE512K;
239 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
240 		align_amount = MMU_PAGESIZE64K;
241 	} else  {
242 		/*
243 		 * Align virtual addresses on a 64K boundary to ensure
244 		 * that ELF shared libraries are mapped with the appropriate
245 		 * alignment constraints by the run-time linker.
246 		 */
247 		align_amount = ELF_SPARC_MAXPGSZ;
248 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
249 			((uintptr_t)*addrp < align_amount))
250 			align_amount = (uintptr_t)*addrp;
251 	}
252 
253 	/*
254 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
255 	 */
256 	if (p->p_model == DATAMODEL_LP64)
257 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
258 #ifdef VAC
259 	if (vac && vacalign && (align_amount < shm_alignment))
260 		align_amount = shm_alignment;
261 #endif
262 
263 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
264 		align_amount = (uintptr_t)*addrp;
265 	}
266 	len += align_amount;
267 
268 	/*
269 	 * Look for a large enough hole starting below the stack limit.
270 	 * After finding it, use the upper part.  Addition of PAGESIZE is
271 	 * for the redzone as described above.
272 	 */
273 	as_purge(as);
274 	if (as_gap(as, len, &base, &slen, AH_HI, NULL) == 0) {
275 		caddr_t as_addr;
276 
277 		addr = base + slen - len + PAGESIZE;
278 		as_addr = addr;
279 		/*
280 		 * Round address DOWN to the alignment amount,
281 		 * add the offset, and if this address is less
282 		 * than the original address, add alignment amount.
283 		 */
284 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
285 		addr += (long)(off & (align_amount - 1l));
286 		if (addr < as_addr) {
287 			addr += align_amount;
288 		}
289 
290 		ASSERT(addr <= (as_addr + align_amount));
291 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
292 		    ((uintptr_t)(off & (align_amount - 1l))));
293 		*addrp = addr;
294 
295 #if defined(SF_ERRATA_57)
296 		if (AS_TYPE_64BIT(as) && addr < errata57_limit) {
297 			*addrp = NULL;
298 		}
299 #endif
300 	} else {
301 		*addrp = NULL;	/* no more virtual space */
302 	}
303 }
304 
305 /*
306  * Platform-dependent page scrub call.
307  */
308 void
309 pagescrub(page_t *pp, uint_t off, uint_t len)
310 {
311 	/*
312 	 * For now, we rely on the fact that pagezero() will
313 	 * always clear UEs.
314 	 */
315 	pagezero(pp, off, len);
316 }
317 
318 /*ARGSUSED*/
319 void
320 sync_data_memory(caddr_t va, size_t len)
321 {
322 	cpu_flush_ecache();
323 }
324 
325 /*
326  * platform specific large pages for kernel heap support
327  */
328 void
329 mmu_init_kcontext()
330 {
331 	extern void set_kcontextreg();
332 
333 	if (kcontextreg)
334 		set_kcontextreg();
335 }
336 
337 void
338 contig_mem_init(void)
339 {
340 	/* not applicable to sun4u */
341 }
342