xref: /titanic_41/usr/src/uts/intel/ia32/os/sundep.c (revision 5494fa53f9c55de7fb31e3bfb7e0da50d08159a5)
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) 1990, 1991 UNIX System Laboratories, Inc. */
27 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989, 1990 AT&T   */
28 /*	All Rights Reserved   */
29 
30 #pragma ident	"%Z%%M%	%I%	%E% SMI"
31 
32 #include <sys/types.h>
33 #include <sys/param.h>
34 #include <sys/sysmacros.h>
35 #include <sys/signal.h>
36 #include <sys/systm.h>
37 #include <sys/user.h>
38 #include <sys/mman.h>
39 #include <sys/class.h>
40 #include <sys/proc.h>
41 #include <sys/procfs.h>
42 #include <sys/buf.h>
43 #include <sys/kmem.h>
44 #include <sys/cred.h>
45 #include <sys/archsystm.h>
46 #include <sys/vmparam.h>
47 #include <sys/prsystm.h>
48 #include <sys/reboot.h>
49 #include <sys/uadmin.h>
50 #include <sys/vfs.h>
51 #include <sys/vnode.h>
52 #include <sys/file.h>
53 #include <sys/session.h>
54 #include <sys/ucontext.h>
55 #include <sys/dnlc.h>
56 #include <sys/var.h>
57 #include <sys/cmn_err.h>
58 #include <sys/debugreg.h>
59 #include <sys/thread.h>
60 #include <sys/vtrace.h>
61 #include <sys/consdev.h>
62 #include <sys/psw.h>
63 #include <sys/regset.h>
64 #include <sys/privregs.h>
65 #include <sys/cpu.h>
66 #include <sys/stack.h>
67 #include <sys/swap.h>
68 #include <vm/hat.h>
69 #include <vm/anon.h>
70 #include <vm/as.h>
71 #include <vm/page.h>
72 #include <vm/seg.h>
73 #include <vm/seg_kmem.h>
74 #include <vm/seg_map.h>
75 #include <vm/seg_vn.h>
76 #include <sys/exec.h>
77 #include <sys/acct.h>
78 #include <sys/core.h>
79 #include <sys/corectl.h>
80 #include <sys/modctl.h>
81 #include <sys/tuneable.h>
82 #include <c2/audit.h>
83 #include <sys/bootconf.h>
84 #include <sys/brand.h>
85 #include <sys/dumphdr.h>
86 #include <sys/promif.h>
87 #include <sys/systeminfo.h>
88 #include <sys/kdi.h>
89 #include <sys/contract_impl.h>
90 #include <sys/x86_archext.h>
91 #include <sys/segments.h>
92 #include <sys/ontrap.h>
93 
94 /*
95  * Compare the version of boot that boot says it is against
96  * the version of boot the kernel expects.
97  */
98 int
99 check_boot_version(int boots_version)
100 {
101 	if (boots_version == BO_VERSION)
102 		return (0);
103 
104 	prom_printf("Wrong boot interface - kernel needs v%d found v%d\n",
105 	    BO_VERSION, boots_version);
106 	prom_panic("halting");
107 	/*NOTREACHED*/
108 }
109 
110 /*
111  * Process the physical installed list for boot.
112  * Finds:
113  * 1) the pfn of the highest installed physical page,
114  * 2) the number of pages installed
115  * 3) the number of distinct contiguous regions these pages fall into.
116  */
117 void
118 installed_top_size(
119 	struct memlist *list,	/* pointer to start of installed list */
120 	pfn_t *high_pfn,	/* return ptr for top value */
121 	pgcnt_t *pgcnt,		/* return ptr for sum of installed pages */
122 	int	*ranges)	/* return ptr for the count of contig. ranges */
123 {
124 	pfn_t top = 0;
125 	pgcnt_t sumpages = 0;
126 	pfn_t highp;		/* high page in a chunk */
127 	int cnt = 0;
128 
129 	for (; list; list = list->next) {
130 		++cnt;
131 		highp = (list->address + list->size - 1) >> PAGESHIFT;
132 		if (top < highp)
133 			top = highp;
134 		sumpages += btop(list->size);
135 	}
136 
137 	*high_pfn = top;
138 	*pgcnt = sumpages;
139 	*ranges = cnt;
140 }
141 
142 /*
143  * Copy in a memory list from boot to kernel, with a filter function
144  * to remove pages. The filter function can increase the address and/or
145  * decrease the size to filter out pages.  It will also align addresses and
146  * sizes to PAGESIZE.
147  */
148 void
149 copy_memlist_filter(
150 	struct memlist *src,
151 	struct memlist **dstp,
152 	void (*filter)(uint64_t *, uint64_t *))
153 {
154 	struct memlist *dst, *prev;
155 	uint64_t addr;
156 	uint64_t size;
157 	uint64_t eaddr;
158 
159 	dst = *dstp;
160 	prev = dst;
161 
162 	/*
163 	 * Move through the memlist applying a filter against
164 	 * each range of memory. Note that we may apply the
165 	 * filter multiple times against each memlist entry.
166 	 */
167 	for (; src; src = src->next) {
168 		addr = P2ROUNDUP(src->address, PAGESIZE);
169 		eaddr = P2ALIGN(src->address + src->size, PAGESIZE);
170 		while (addr < eaddr) {
171 			size = eaddr - addr;
172 			if (filter != NULL)
173 				filter(&addr, &size);
174 			if (size == 0)
175 				break;
176 			dst->address = addr;
177 			dst->size = size;
178 			dst->next = 0;
179 			if (prev == dst) {
180 				dst->prev = 0;
181 				dst++;
182 			} else {
183 				dst->prev = prev;
184 				prev->next = dst;
185 				dst++;
186 				prev++;
187 			}
188 			addr += size;
189 		}
190 	}
191 
192 	*dstp = dst;
193 }
194 
195 /*
196  * Kernel setup code, called from startup().
197  */
198 void
199 kern_setup1(void)
200 {
201 	proc_t *pp;
202 
203 	pp = &p0;
204 
205 	proc_sched = pp;
206 
207 	/*
208 	 * Initialize process 0 data structures
209 	 */
210 	pp->p_stat = SRUN;
211 	pp->p_flag = SSYS;
212 
213 	pp->p_pidp = &pid0;
214 	pp->p_pgidp = &pid0;
215 	pp->p_sessp = &session0;
216 	pp->p_tlist = &t0;
217 	pid0.pid_pglink = pp;
218 	pid0.pid_pgtail = pp;
219 
220 	/*
221 	 * XXX - we asssume that the u-area is zeroed out except for
222 	 * ttolwp(curthread)->lwp_regs.
223 	 */
224 	PTOU(curproc)->u_cmask = (mode_t)CMASK;
225 
226 	thread_init();		/* init thread_free list */
227 	pid_init();		/* initialize pid (proc) table */
228 	contract_init();	/* initialize contracts */
229 
230 	init_pages_pp_maximum();
231 }
232 
233 /*
234  * Load a procedure into a thread.
235  */
236 void
237 thread_load(kthread_t *t, void (*start)(), caddr_t arg, size_t len)
238 {
239 	caddr_t sp;
240 	size_t framesz;
241 	caddr_t argp;
242 	long *p;
243 	extern void thread_start();
244 
245 	/*
246 	 * Push a "c" call frame onto the stack to represent
247 	 * the caller of "start".
248 	 */
249 	sp = t->t_stk;
250 	ASSERT(((uintptr_t)t->t_stk & (STACK_ENTRY_ALIGN - 1)) == 0);
251 	if (len != 0) {
252 		/*
253 		 * the object that arg points at is copied into the
254 		 * caller's frame.
255 		 */
256 		framesz = SA(len);
257 		sp -= framesz;
258 		ASSERT(sp > t->t_stkbase);
259 		argp = sp + SA(MINFRAME);
260 		bcopy(arg, argp, len);
261 		arg = argp;
262 	}
263 	/*
264 	 * Set up arguments (arg and len) on the caller's stack frame.
265 	 */
266 	p = (long *)sp;
267 
268 	*--p = 0;		/* fake call */
269 	*--p = 0;		/* null frame pointer terminates stack trace */
270 	*--p = (long)len;
271 	*--p = (intptr_t)arg;
272 	*--p = (intptr_t)start;
273 
274 	/*
275 	 * initialize thread to resume at thread_start() which will
276 	 * turn around and invoke (*start)(arg, len).
277 	 */
278 	t->t_pc = (uintptr_t)thread_start;
279 	t->t_sp = (uintptr_t)p;
280 
281 	ASSERT((t->t_sp & (STACK_ENTRY_ALIGN - 1)) == 0);
282 }
283 
284 /*
285  * load user registers into lwp.
286  */
287 /*ARGSUSED2*/
288 void
289 lwp_load(klwp_t *lwp, gregset_t grp, uintptr_t thrptr)
290 {
291 	struct regs *rp = lwptoregs(lwp);
292 
293 	setgregs(lwp, grp);
294 	rp->r_ps = PSL_USER;
295 
296 	/*
297 	 * For 64-bit lwps, we allow null %fs selector value, and null
298 	 * %gs selector to point anywhere in the address space using
299 	 * %fsbase and %gsbase behind the scenes.  libc uses %fs to point
300 	 * at the ulwp_t structure.
301 	 *
302 	 * For 32-bit lwps, libc wedges its lwp thread pointer into the
303 	 * ucontext ESP slot (which is otherwise irrelevant to setting a
304 	 * ucontext) and LWPGS_SEL value into gregs[REG_GS].  This is so
305 	 * syslwp_create() can atomically setup %gs.
306 	 *
307 	 * See setup_context() in libc.
308 	 */
309 #ifdef _SYSCALL32_IMPL
310 	if (lwp_getdatamodel(lwp) == DATAMODEL_ILP32) {
311 		if (grp[REG_GS] == LWPGS_SEL)
312 			(void) lwp_setprivate(lwp, _LWP_GSBASE, thrptr);
313 	} else {
314 		/*
315 		 * See lwp_setprivate in kernel and setup_context in libc.
316 		 *
317 		 * Currently libc constructs a ucontext from whole cloth for
318 		 * every new (not main) lwp created.  For 64 bit processes
319 		 * %fsbase is directly set to point to current thread pointer.
320 		 * In the past (solaris 10) %fs was also set LWPFS_SEL to
321 		 * indicate %fsbase. Now we use the null GDT selector for
322 		 * this purpose. LWP[FS|GS]_SEL are only intended for 32 bit
323 		 * processes. To ease transition we support older libcs in
324 		 * the newer kernel by forcing %fs or %gs selector to null
325 		 * by calling lwp_setprivate if LWP[FS|GS]_SEL is passed in
326 		 * the ucontext.  This is should be ripped out at some future
327 		 * date.  Another fix would be for libc to do a getcontext
328 		 * and inherit the null %fs/%gs from the current context but
329 		 * that means an extra system call and could hurt performance.
330 		 */
331 		if (grp[REG_FS] == 0x1bb) /* hard code legacy LWPFS_SEL */
332 		    (void) lwp_setprivate(lwp, _LWP_FSBASE,
333 		    (uintptr_t)grp[REG_FSBASE]);
334 
335 		if (grp[REG_GS] == 0x1c3) /* hard code legacy LWPGS_SEL */
336 		    (void) lwp_setprivate(lwp, _LWP_GSBASE,
337 		    (uintptr_t)grp[REG_GSBASE]);
338 	}
339 #else
340 	if (grp[GS] == LWPGS_SEL)
341 		(void) lwp_setprivate(lwp, _LWP_GSBASE, thrptr);
342 #endif
343 
344 	lwp->lwp_eosys = JUSTRETURN;
345 	lwptot(lwp)->t_post_sys = 1;
346 }
347 
348 /*
349  * set syscall()'s return values for a lwp.
350  */
351 void
352 lwp_setrval(klwp_t *lwp, int v1, int v2)
353 {
354 	lwptoregs(lwp)->r_ps &= ~PS_C;
355 	lwptoregs(lwp)->r_r0 = v1;
356 	lwptoregs(lwp)->r_r1 = v2;
357 }
358 
359 /*
360  * set syscall()'s return values for a lwp.
361  */
362 void
363 lwp_setsp(klwp_t *lwp, caddr_t sp)
364 {
365 	lwptoregs(lwp)->r_sp = (intptr_t)sp;
366 }
367 
368 /*
369  * Copy regs from parent to child.
370  */
371 void
372 lwp_forkregs(klwp_t *lwp, klwp_t *clwp)
373 {
374 #if defined(__amd64)
375 	struct pcb *pcb = &clwp->lwp_pcb;
376 	struct regs *rp = lwptoregs(lwp);
377 
378 	if (pcb->pcb_rupdate == 0) {
379 		pcb->pcb_ds = rp->r_ds;
380 		pcb->pcb_es = rp->r_es;
381 		pcb->pcb_fs = rp->r_fs;
382 		pcb->pcb_gs = rp->r_gs;
383 		pcb->pcb_rupdate = 1;
384 		lwptot(clwp)->t_post_sys = 1;
385 	}
386 	ASSERT(lwptot(clwp)->t_post_sys);
387 #endif
388 
389 	bcopy(lwp->lwp_regs, clwp->lwp_regs, sizeof (struct regs));
390 }
391 
392 /*
393  * This function is currently unused on x86.
394  */
395 /*ARGSUSED*/
396 void
397 lwp_freeregs(klwp_t *lwp, int isexec)
398 {}
399 
400 /*
401  * This function is currently unused on x86.
402  */
403 void
404 lwp_pcb_exit(void)
405 {}
406 
407 /*
408  * Lwp context ops for segment registers.
409  */
410 
411 /*
412  * Every time we come into the kernel (syscall, interrupt or trap
413  * but not fast-traps) we capture the current values of the user's
414  * segment registers into the lwp's reg structure. This includes
415  * lcall for i386 generic system call support since it is handled
416  * as a segment-not-present trap.
417  *
418  * Here we save the current values from the lwp regs into the pcb
419  * and set pcb->pcb_rupdate to 1 to tell the rest of the kernel
420  * that the pcb copy of the segment registers is the current one.
421  * This ensures the lwp's next trip to user land via update_sregs.
422  * Finally we set t_post_sys to ensure that no system call fast-path's
423  * its way out of the kernel via sysret.
424  *
425  * (This means that we need to have interrupts disabled when we test
426  * t->t_post_sys in the syscall handlers; if the test fails, we need
427  * to keep interrupts disabled until we return to userland so we can't
428  * be switched away.)
429  *
430  * As a result of all this, we don't really have to do a whole lot if
431  * the thread is just mucking about in the kernel, switching on and
432  * off the cpu for whatever reason it feels like. And yet we still
433  * preserve fast syscalls, cause if we -don't- get descheduled,
434  * we never come here either.
435  */
436 
437 #define	VALID_LWP_DESC(udp) ((udp)->usd_type == SDT_MEMRWA && \
438 	    (udp)->usd_p == 1 && (udp)->usd_dpl == SEL_UPL)
439 
440 void
441 lwp_segregs_save(klwp_t *lwp)
442 {
443 #if defined(__amd64)
444 	pcb_t *pcb = &lwp->lwp_pcb;
445 	struct regs *rp;
446 
447 	ASSERT(VALID_LWP_DESC(&pcb->pcb_fsdesc));
448 	ASSERT(VALID_LWP_DESC(&pcb->pcb_gsdesc));
449 
450 	if (pcb->pcb_rupdate == 0) {
451 		rp = lwptoregs(lwp);
452 
453 		/*
454 		 * If there's no update already pending, capture the current
455 		 * %ds/%es/%fs/%gs values from lwp's regs in case the user
456 		 * changed them; %fsbase and %gsbase are privileged so the
457 		 * kernel versions of these registers in pcb_fsbase and
458 		 * pcb_gsbase are always up-to-date.
459 		 */
460 		pcb->pcb_ds = rp->r_ds;
461 		pcb->pcb_es = rp->r_es;
462 		pcb->pcb_fs = rp->r_fs;
463 		pcb->pcb_gs = rp->r_gs;
464 		pcb->pcb_rupdate = 1;
465 		lwp->lwp_thread->t_post_sys = 1;
466 	}
467 #endif	/* __amd64 */
468 
469 	ASSERT(bcmp(&CPU->cpu_gdt[GDT_LWPFS], &lwp->lwp_pcb.pcb_fsdesc,
470 	    sizeof (lwp->lwp_pcb.pcb_fsdesc)) == 0);
471 	ASSERT(bcmp(&CPU->cpu_gdt[GDT_LWPGS], &lwp->lwp_pcb.pcb_gsdesc,
472 	    sizeof (lwp->lwp_pcb.pcb_gsdesc)) == 0);
473 }
474 
475 #if defined(__amd64)
476 
477 /*
478  * Update the segment registers with new values from the pcb
479  *
480  * We have to do this carefully, and in the following order,
481  * in case any of the selectors points at a bogus descriptor.
482  * If they do, we'll catch trap with on_trap and return 1.
483  * returns 0 on success.
484  *
485  * This is particularly tricky for %gs.
486  * This routine must be executed under a cli.
487  */
488 int
489 update_sregs(struct regs *rp,  klwp_t *lwp)
490 {
491 	pcb_t *pcb = &lwp->lwp_pcb;
492 	ulong_t	kgsbase;
493 	on_trap_data_t	otd;
494 	int rc = 0;
495 
496 	if (!on_trap(&otd, OT_SEGMENT_ACCESS)) {
497 
498 		kgsbase = (ulong_t)CPU;
499 		__set_gs(pcb->pcb_gs);
500 
501 		/*
502 		 * If __set_gs fails it's because the new %gs is a bad %gs,
503 		 * we'll be taking a trap but with the original %gs and %gsbase
504 		 * undamaged (i.e. pointing at curcpu).
505 		 *
506 		 * We've just mucked up the kernel's gsbase.  Oops.  In
507 		 * particular we can't take any traps at all.  Make the newly
508 		 * computed gsbase be the hidden gs via __swapgs , and fix
509 		 * the kernel's gsbase back again. Later, when we return to
510 		 * userland we'll swapgs again restoring gsbase just loaded
511 		 * above.
512 		 */
513 		__swapgs();
514 		rp->r_gs = pcb->pcb_gs;
515 
516 		/*
517 		 * restore kernel's gsbase
518 		 */
519 		wrmsr(MSR_AMD_GSBASE, kgsbase);
520 
521 		/*
522 		 * Only override the descriptor base address if
523 		 * r_gs == LWPGS_SEL or if r_gs == NULL. A note on
524 		 * NULL descriptors -- 32-bit programs take faults
525 		 * if they deference NULL descriptors; however,
526 		 * when 64-bit programs load them into %fs or %gs,
527 		 * they DONT fault -- only the base address remains
528 		 * whatever it was from the last load.   Urk.
529 		 *
530 		 * XXX - note that lwp_setprivate now sets %fs/%gs to the
531 		 * null selector for 64 bit processes. Whereas before
532 		 * %fs/%gs were set to LWP(FS|GS)_SEL regardless of
533 		 * the process's data model. For now we check for both
534 		 * values so that the kernel can also support the older
535 		 * libc. This should be ripped out at some point in the
536 		 * future.
537 		 */
538 		if (pcb->pcb_gs == LWPGS_SEL || pcb->pcb_gs == 0)
539 			wrmsr(MSR_AMD_KGSBASE, pcb->pcb_gsbase);
540 
541 		__set_ds(pcb->pcb_ds);
542 		rp->r_ds = pcb->pcb_ds;
543 
544 		__set_es(pcb->pcb_es);
545 		rp->r_es = pcb->pcb_es;
546 
547 		__set_fs(pcb->pcb_fs);
548 		rp->r_fs = pcb->pcb_fs;
549 
550 		/*
551 		 * Same as for %gs
552 		 */
553 		if (pcb->pcb_fs == LWPFS_SEL || pcb->pcb_fs == 0)
554 			wrmsr(MSR_AMD_FSBASE, pcb->pcb_fsbase);
555 
556 	} else {
557 		cli();
558 		rc = 1;
559 	}
560 	no_trap();
561 	return (rc);
562 }
563 #endif	/* __amd64 */
564 
565 #ifdef _SYSCALL32_IMPL
566 
567 /*
568  * Make it impossible for a process to change its data model.
569  * We do this by toggling the present bits for the 32 and
570  * 64-bit user code descriptors. That way if a user lwp attempts
571  * to change its data model (by using the wrong code descriptor in
572  * %cs) it will fault immediately. This also allows us to simplify
573  * assertions and checks in the kernel.
574  */
575 static void
576 gdt_ucode_model(model_t model)
577 {
578 	cpu_t *cpu;
579 
580 	kpreempt_disable();
581 	cpu = CPU;
582 	if (model == DATAMODEL_NATIVE) {
583 		cpu->cpu_gdt[GDT_UCODE].usd_p = 1;
584 		cpu->cpu_gdt[GDT_U32CODE].usd_p = 0;
585 	} else {
586 		cpu->cpu_gdt[GDT_U32CODE].usd_p = 1;
587 		cpu->cpu_gdt[GDT_UCODE].usd_p = 0;
588 	}
589 	kpreempt_enable();
590 }
591 
592 #endif	/* _SYSCALL32_IMPL */
593 
594 /*
595  * Restore lwp private fs and gs segment descriptors
596  * on current cpu's GDT.
597  */
598 static void
599 lwp_segregs_restore(klwp_t *lwp)
600 {
601 	pcb_t *pcb = &lwp->lwp_pcb;
602 	cpu_t *cpu = CPU;
603 
604 	ASSERT(VALID_LWP_DESC(&pcb->pcb_fsdesc));
605 	ASSERT(VALID_LWP_DESC(&pcb->pcb_gsdesc));
606 
607 #ifdef	_SYSCALL32_IMPL
608 	gdt_ucode_model(DATAMODEL_NATIVE);
609 #endif
610 
611 	cpu->cpu_gdt[GDT_LWPFS] = pcb->pcb_fsdesc;
612 	cpu->cpu_gdt[GDT_LWPGS] = pcb->pcb_gsdesc;
613 
614 }
615 
616 #ifdef _SYSCALL32_IMPL
617 
618 static void
619 lwp_segregs_restore32(klwp_t *lwp)
620 {
621 	/*LINTED*/
622 	cpu_t *cpu = CPU;
623 	pcb_t *pcb = &lwp->lwp_pcb;
624 
625 	ASSERT(VALID_LWP_DESC(&lwp->lwp_pcb.pcb_fsdesc));
626 	ASSERT(VALID_LWP_DESC(&lwp->lwp_pcb.pcb_gsdesc));
627 
628 	gdt_ucode_model(DATAMODEL_ILP32);
629 	cpu->cpu_gdt[GDT_LWPFS] = pcb->pcb_fsdesc;
630 	cpu->cpu_gdt[GDT_LWPGS] = pcb->pcb_gsdesc;
631 }
632 
633 #endif	/* _SYSCALL32_IMPL */
634 
635 /*
636  * If this is a process in a branded zone, then we want it to use the brand
637  * syscall entry points instead of the standard Solaris entry points.  This
638  * routine must be called when a new lwp is created within a branded zone
639  * or when an existing lwp moves into a branded zone via a zone_enter()
640  * operation.
641  */
642 void
643 lwp_attach_brand_hdlrs(klwp_t *lwp)
644 {
645 	kthread_t *t = lwptot(lwp);
646 
647 	ASSERT(PROC_IS_BRANDED(lwptoproc(lwp)));
648 	ASSERT(removectx(t, NULL, brand_interpositioning_disable,
649 	    brand_interpositioning_enable, NULL, NULL, NULL, NULL) == 0);
650 
651 	installctx(t, NULL, brand_interpositioning_disable,
652 	    brand_interpositioning_enable, NULL, NULL, NULL, NULL);
653 
654 	if (t == curthread) {
655 		kpreempt_disable();
656 		brand_interpositioning_enable();
657 		kpreempt_enable();
658 	}
659 }
660 
661 /*
662  * Add any lwp-associated context handlers to the lwp at the beginning
663  * of the lwp's useful life.
664  *
665  * All paths which create lwp's invoke lwp_create(); lwp_create()
666  * invokes lwp_stk_init() which initializes the stack, sets up
667  * lwp_regs, and invokes this routine.
668  *
669  * All paths which destroy lwp's invoke lwp_exit() to rip the lwp
670  * apart and put it on 'lwp_deathrow'; if the lwp is destroyed it
671  * ends up in thread_free() which invokes freectx(t, 0) before
672  * invoking lwp_stk_fini().  When the lwp is recycled from death
673  * row, lwp_stk_fini() is invoked, then thread_free(), and thus
674  * freectx(t, 0) as before.
675  *
676  * In the case of exec, the surviving lwp is thoroughly scrubbed
677  * clean; exec invokes freectx(t, 1) to destroy associated contexts.
678  * On the way back to the new image, it invokes setregs() which
679  * in turn invokes this routine.
680  */
681 void
682 lwp_installctx(klwp_t *lwp)
683 {
684 	kthread_t *t = lwptot(lwp);
685 	int thisthread = t == curthread;
686 #ifdef _SYSCALL32_IMPL
687 	void (*restop)(klwp_t *) = lwp_getdatamodel(lwp) == DATAMODEL_NATIVE ?
688 	    lwp_segregs_restore : lwp_segregs_restore32;
689 #else
690 	void (*restop)(klwp_t *) = lwp_segregs_restore;
691 #endif
692 
693 	/*
694 	 * Install the basic lwp context handlers on each lwp.
695 	 *
696 	 * On the amd64 kernel, the context handlers are responsible for
697 	 * virtualizing %ds, %es, %fs, and %gs to the lwp.  The register
698 	 * values are only ever changed via sys_rtt when the
699 	 * pcb->pcb_rupdate == 1.  Only sys_rtt gets to clear the bit.
700 	 *
701 	 * On the i386 kernel, the context handlers are responsible for
702 	 * virtualizing %gs/%fs to the lwp by updating the per-cpu GDTs
703 	 */
704 	ASSERT(removectx(t, lwp, lwp_segregs_save, restop,
705 	    NULL, NULL, NULL, NULL) == 0);
706 	if (thisthread)
707 		kpreempt_disable();
708 	installctx(t, lwp, lwp_segregs_save, restop,
709 	    NULL, NULL, NULL, NULL);
710 	if (thisthread) {
711 		/*
712 		 * Since we're the right thread, set the values in the GDT
713 		 */
714 		restop(lwp);
715 		kpreempt_enable();
716 	}
717 
718 	/*
719 	 * If we have sysenter/sysexit instructions enabled, we need
720 	 * to ensure that the hardware mechanism is kept up-to-date with the
721 	 * lwp's kernel stack pointer across context switches.
722 	 *
723 	 * sep_save zeros the sysenter stack pointer msr; sep_restore sets
724 	 * it to the lwp's kernel stack pointer (kstktop).
725 	 */
726 	if (x86_feature & X86_SEP) {
727 #if defined(__amd64)
728 		caddr_t kstktop = (caddr_t)lwp->lwp_regs;
729 #elif defined(__i386)
730 		caddr_t kstktop = ((caddr_t)lwp->lwp_regs - MINFRAME) +
731 		    SA(sizeof (struct regs) + MINFRAME);
732 #endif
733 		ASSERT(removectx(t, kstktop,
734 		    sep_save, sep_restore, NULL, NULL, NULL, NULL) == 0);
735 
736 		if (thisthread)
737 			kpreempt_disable();
738 		installctx(t, kstktop,
739 		    sep_save, sep_restore, NULL, NULL, NULL, NULL);
740 		if (thisthread) {
741 			/*
742 			 * We're the right thread, so set the stack pointer
743 			 * for the first sysenter instruction to use
744 			 */
745 			sep_restore(kstktop);
746 			kpreempt_enable();
747 		}
748 	}
749 
750 	if (PROC_IS_BRANDED(ttoproc(t)))
751 		lwp_attach_brand_hdlrs(lwp);
752 }
753 
754 /*
755  * Clear registers on exec(2).
756  */
757 void
758 setregs(uarg_t *args)
759 {
760 	struct regs *rp;
761 	kthread_t *t = curthread;
762 	klwp_t *lwp = ttolwp(t);
763 	pcb_t *pcb = &lwp->lwp_pcb;
764 	greg_t sp;
765 
766 	/*
767 	 * Initialize user registers
768 	 */
769 	(void) save_syscall_args();	/* copy args from registers first */
770 	rp = lwptoregs(lwp);
771 	sp = rp->r_sp;
772 	bzero(rp, sizeof (*rp));
773 
774 	rp->r_ss = UDS_SEL;
775 	rp->r_sp = sp;
776 	rp->r_pc = args->entry;
777 	rp->r_ps = PSL_USER;
778 
779 #if defined(__amd64)
780 
781 	pcb->pcb_fs = pcb->pcb_gs = 0;
782 	pcb->pcb_fsbase = pcb->pcb_gsbase = 0;
783 
784 	if (ttoproc(t)->p_model == DATAMODEL_NATIVE) {
785 
786 		rp->r_cs = UCS_SEL;
787 
788 		/*
789 		 * Only allow 64-bit user code descriptor to be present.
790 		 */
791 		gdt_ucode_model(DATAMODEL_NATIVE);
792 
793 		/*
794 		 * Arrange that the virtualized %fs and %gs GDT descriptors
795 		 * have a well-defined initial state (present, ring 3
796 		 * and of type data).
797 		 */
798 		pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_udesc;
799 
800 		/*
801 		 * thrptr is either NULL or a value used by DTrace.
802 		 * 64-bit processes use %fs as their "thread" register.
803 		 */
804 		if (args->thrptr)
805 			(void) lwp_setprivate(lwp, _LWP_FSBASE, args->thrptr);
806 
807 	} else {
808 
809 		rp->r_cs = U32CS_SEL;
810 		rp->r_ds = rp->r_es = UDS_SEL;
811 
812 		/*
813 		 * only allow 32-bit user code selector to be present.
814 		 */
815 		gdt_ucode_model(DATAMODEL_ILP32);
816 
817 		pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_u32desc;
818 
819 		/*
820 		 * thrptr is either NULL or a value used by DTrace.
821 		 * 32-bit processes use %gs as their "thread" register.
822 		 */
823 		if (args->thrptr)
824 			(void) lwp_setprivate(lwp, _LWP_GSBASE, args->thrptr);
825 
826 	}
827 
828 	pcb->pcb_ds = rp->r_ds;
829 	pcb->pcb_es = rp->r_es;
830 	pcb->pcb_rupdate = 1;
831 
832 #elif defined(__i386)
833 
834 	rp->r_cs = UCS_SEL;
835 	rp->r_ds = rp->r_es = UDS_SEL;
836 
837 	/*
838 	 * Arrange that the virtualized %fs and %gs GDT descriptors
839 	 * have a well-defined initial state (present, ring 3
840 	 * and of type data).
841 	 */
842 	pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_udesc;
843 
844 	/*
845 	 * For %gs we need to reset LWP_GSBASE in pcb and the
846 	 * per-cpu GDT descriptor. thrptr is either NULL
847 	 * or a value used by DTrace.
848 	 */
849 	if (args->thrptr)
850 		(void) lwp_setprivate(lwp, _LWP_GSBASE, args->thrptr);
851 #endif
852 
853 	lwp->lwp_eosys = JUSTRETURN;
854 	t->t_post_sys = 1;
855 
856 	/*
857 	 * Here we initialize minimal fpu state.
858 	 * The rest is done at the first floating
859 	 * point instruction that a process executes.
860 	 */
861 	pcb->pcb_fpu.fpu_flags = 0;
862 
863 	/*
864 	 * Add the lwp context handlers that virtualize segment registers,
865 	 * and/or system call stacks etc.
866 	 */
867 	lwp_installctx(lwp);
868 }
869 
870 user_desc_t *
871 cpu_get_gdt(void)
872 {
873 	return (CPU->cpu_gdt);
874 }
875 
876 
877 #if !defined(lwp_getdatamodel)
878 
879 /*
880  * Return the datamodel of the given lwp.
881  */
882 /*ARGSUSED*/
883 model_t
884 lwp_getdatamodel(klwp_t *lwp)
885 {
886 	return (lwp->lwp_procp->p_model);
887 }
888 
889 #endif	/* !lwp_getdatamodel */
890 
891 #if !defined(get_udatamodel)
892 
893 model_t
894 get_udatamodel(void)
895 {
896 	return (curproc->p_model);
897 }
898 
899 #endif	/* !get_udatamodel */
900