xref: /titanic_51/usr/src/uts/sun4u/vm/mach_vm_dep.c (revision b9bd317cda1afb3a01f4812de73e8cec888cbbd7)
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  * Every mapping will have a redzone of a single page on either side of
155  * the request. This is done to leave one page unmapped between segments.
156  * This is not required, but it's useful for the user because if their
157  * program strays across a segment boundary, it will catch a fault
158  * immediately making debugging a little easier.  Currently the redzone
159  * is mandatory.
160  *
161  *
162  * addrp is a value/result parameter.
163  *	On input it is a hint from the user to be used in a completely
164  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
165  *	minimal alignment, which must be some "power of two" multiple of
166  *	pagesize.
167  *
168  *	On output it is NULL if no address can be found in the current
169  *	processes address space or else an address that is currently
170  *	not mapped for len bytes with a page of red zone on either side.
171  *	If vacalign is true, then the selected address will obey the alignment
172  *	constraints of a vac machine based on the given off value.
173  */
174 /*ARGSUSED4*/
175 void
176 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
177     caddr_t userlimit, struct proc *p, uint_t flags)
178 {
179 	struct as *as = p->p_as;
180 	caddr_t addr;
181 	caddr_t base;
182 	size_t slen;
183 	uintptr_t align_amount;
184 	int allow_largepage_alignment = 1;
185 
186 	base = p->p_brkbase;
187 	if (userlimit < as->a_userlimit) {
188 		/*
189 		 * This happens when a program wants to map something in
190 		 * a range that's accessible to a program in a smaller
191 		 * address space.  For example, a 64-bit program might
192 		 * be calling mmap32(2) to guarantee that the returned
193 		 * address is below 4Gbytes.
194 		 */
195 		ASSERT(userlimit > base);
196 		slen = userlimit - base;
197 	} else {
198 		slen = p->p_usrstack - base - (((size_t)rctl_enforced_value(
199 		    rctlproc_legacy[RLIMIT_STACK], p->p_rctls, p) + PAGEOFFSET)
200 		    & PAGEMASK);
201 	}
202 
203 	/* Make len be a multiple of PAGESIZE */
204 	len = (len + PAGEOFFSET) & PAGEMASK;
205 
206 	/*
207 	 *  If the request is larger than the size of a particular
208 	 *  mmu level, then we use that level to map the request.
209 	 *  But this requires that both the virtual and the physical
210 	 *  addresses be aligned with respect to that level, so we
211 	 *  do the virtual bit of nastiness here.
212 	 *
213 	 *  For 32-bit processes, only those which have specified
214 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
215 	 *  we can potentially waste up to 256MB of the 4G process address
216 	 *  space just for alignment.
217 	 */
218 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
219 	    ((uintptr_t)*addrp) != 0)) {
220 		allow_largepage_alignment = 0;
221 	}
222 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
223 	    allow_largepage_alignment &&
224 	    (len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
225 		align_amount = MMU_PAGESIZE256M;
226 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
227 	    allow_largepage_alignment &&
228 	    (len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
229 		align_amount = MMU_PAGESIZE32M;
230 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
231 		align_amount = MMU_PAGESIZE4M;
232 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
233 		align_amount = MMU_PAGESIZE512K;
234 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
235 		align_amount = MMU_PAGESIZE64K;
236 	} else  {
237 		/*
238 		 * Align virtual addresses on a 64K boundary to ensure
239 		 * that ELF shared libraries are mapped with the appropriate
240 		 * alignment constraints by the run-time linker.
241 		 */
242 		align_amount = ELF_SPARC_MAXPGSZ;
243 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
244 		    ((uintptr_t)*addrp < align_amount))
245 			align_amount = (uintptr_t)*addrp;
246 	}
247 
248 	/*
249 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
250 	 */
251 	if (p->p_model == DATAMODEL_LP64)
252 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
253 #ifdef VAC
254 	if (vac && vacalign && (align_amount < shm_alignment))
255 		align_amount = shm_alignment;
256 #endif
257 
258 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
259 		align_amount = (uintptr_t)*addrp;
260 	}
261 
262 	ASSERT(ISP2(align_amount));
263 	ASSERT(align_amount == 0 || align_amount >= PAGESIZE);
264 
265 	/*
266 	 * Look for a large enough hole starting below the stack limit.
267 	 * After finding it, use the upper part.
268 	 */
269 	as_purge(as);
270 	off = off & (align_amount - 1);
271 	if (as_gap_aligned(as, len, &base, &slen, AH_HI, NULL, align_amount,
272 	    PAGESIZE, off) == 0) {
273 		caddr_t as_addr;
274 
275 		/*
276 		 * addr is the highest possible address to use since we have
277 		 * a PAGESIZE redzone at the beginning and end.
278 		 */
279 		addr = base + slen - (PAGESIZE + len);
280 		as_addr = addr;
281 		/*
282 		 * Round address DOWN to the alignment amount and
283 		 * add the offset in.
284 		 * If addr is greater than as_addr, len would not be large
285 		 * enough to include the redzone, so we must adjust down
286 		 * by the alignment amount.
287 		 */
288 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
289 		addr += (long)off;
290 		if (addr > as_addr) {
291 			addr -= align_amount;
292 		}
293 
294 		ASSERT(addr > base);
295 		ASSERT(addr + len < base + slen);
296 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
297 		    ((uintptr_t)(off)));
298 		*addrp = addr;
299 
300 #if defined(SF_ERRATA_57)
301 		if (AS_TYPE_64BIT(as) && addr < errata57_limit) {
302 			*addrp = NULL;
303 		}
304 #endif
305 	} else {
306 		*addrp = NULL;	/* no more virtual space */
307 	}
308 }
309 
310 /*
311  * Platform-dependent page scrub call.
312  */
313 void
314 pagescrub(page_t *pp, uint_t off, uint_t len)
315 {
316 	/*
317 	 * For now, we rely on the fact that pagezero() will
318 	 * always clear UEs.
319 	 */
320 	pagezero(pp, off, len);
321 }
322 
323 /*ARGSUSED*/
324 void
325 sync_data_memory(caddr_t va, size_t len)
326 {
327 	cpu_flush_ecache();
328 }
329 
330 /*
331  * platform specific large pages for kernel heap support
332  */
333 void
334 mmu_init_kcontext()
335 {
336 	extern void set_kcontextreg();
337 
338 	if (kcontextreg)
339 		set_kcontextreg();
340 }
341 
342 void
343 contig_mem_init(void)
344 {
345 	/* not applicable to sun4u */
346 }
347 
348 /*ARGSUSED*/
349 caddr_t
350 contig_mem_prealloc(caddr_t alloc_base, pgcnt_t npages)
351 {
352 	/* not applicable to sun4u */
353 	return (alloc_base);
354 }
355 
356 size_t
357 exec_get_spslew(void)
358 {
359 	return (0);
360 }
361