xref: /titanic_52/usr/src/uts/intel/ia32/os/sundep.c (revision ae115bc77f6fcde83175c75b4206dc2e50747966)
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/stack.h>
66 #include <sys/swap.h>
67 #include <vm/hat.h>
68 #include <vm/anon.h>
69 #include <vm/as.h>
70 #include <vm/page.h>
71 #include <vm/seg.h>
72 #include <vm/seg_kmem.h>
73 #include <vm/seg_map.h>
74 #include <vm/seg_vn.h>
75 #include <sys/exec.h>
76 #include <sys/acct.h>
77 #include <sys/core.h>
78 #include <sys/corectl.h>
79 #include <sys/modctl.h>
80 #include <sys/tuneable.h>
81 #include <c2/audit.h>
82 #include <sys/bootconf.h>
83 #include <sys/brand.h>
84 #include <sys/dumphdr.h>
85 #include <sys/promif.h>
86 #include <sys/systeminfo.h>
87 #include <sys/kdi.h>
88 #include <sys/contract_impl.h>
89 #include <sys/x86_archext.h>
90 #include <sys/segments.h>
91 #include <sys/ontrap.h>
92 
93 /*
94  * Compare the version of boot that boot says it is against
95  * the version of boot the kernel expects.
96  */
97 int
98 check_boot_version(int boots_version)
99 {
100 	if (boots_version == BO_VERSION)
101 		return (0);
102 
103 	prom_printf("Wrong boot interface - kernel needs v%d found v%d\n",
104 	    BO_VERSION, boots_version);
105 	prom_panic("halting");
106 	/*NOTREACHED*/
107 }
108 
109 /*
110  * Process the physical installed list for boot.
111  * Finds:
112  * 1) the pfn of the highest installed physical page,
113  * 2) the number of pages installed
114  * 3) the number of distinct contiguous regions these pages fall into.
115  */
116 void
117 installed_top_size(
118 	struct memlist *list,	/* pointer to start of installed list */
119 	pfn_t *high_pfn,	/* return ptr for top value */
120 	pgcnt_t *pgcnt,		/* return ptr for sum of installed pages */
121 	int	*ranges)	/* return ptr for the count of contig. ranges */
122 {
123 	pfn_t top = 0;
124 	pgcnt_t sumpages = 0;
125 	pfn_t highp;		/* high page in a chunk */
126 	int cnt = 0;
127 
128 	for (; list; list = list->next) {
129 		++cnt;
130 		highp = (list->address + list->size - 1) >> PAGESHIFT;
131 		if (top < highp)
132 			top = highp;
133 		sumpages += btop(list->size);
134 	}
135 
136 	*high_pfn = top;
137 	*pgcnt = sumpages;
138 	*ranges = cnt;
139 }
140 
141 /*
142  * Copy in a memory list from boot to kernel, with a filter function
143  * to remove pages. The filter function can increase the address and/or
144  * decrease the size to filter out pages.  It will also align addresses and
145  * sizes to PAGESIZE.
146  */
147 void
148 copy_memlist_filter(
149 	struct memlist *src,
150 	struct memlist **dstp,
151 	void (*filter)(uint64_t *, uint64_t *))
152 {
153 	struct memlist *dst, *prev;
154 	uint64_t addr;
155 	uint64_t size;
156 	uint64_t eaddr;
157 
158 	dst = *dstp;
159 	prev = dst;
160 
161 	/*
162 	 * Move through the memlist applying a filter against
163 	 * each range of memory. Note that we may apply the
164 	 * filter multiple times against each memlist entry.
165 	 */
166 	for (; src; src = src->next) {
167 		addr = P2ROUNDUP(src->address, PAGESIZE);
168 		eaddr = P2ALIGN(src->address + src->size, PAGESIZE);
169 		while (addr < eaddr) {
170 			size = eaddr - addr;
171 			if (filter != NULL)
172 				filter(&addr, &size);
173 			if (size == 0)
174 				break;
175 			dst->address = addr;
176 			dst->size = size;
177 			dst->next = 0;
178 			if (prev == dst) {
179 				dst->prev = 0;
180 				dst++;
181 			} else {
182 				dst->prev = prev;
183 				prev->next = dst;
184 				dst++;
185 				prev++;
186 			}
187 			addr += size;
188 		}
189 	}
190 
191 	*dstp = dst;
192 }
193 
194 /*
195  * Kernel setup code, called from startup().
196  */
197 void
198 kern_setup1(void)
199 {
200 	proc_t *pp;
201 
202 	pp = &p0;
203 
204 	proc_sched = pp;
205 
206 	/*
207 	 * Initialize process 0 data structures
208 	 */
209 	pp->p_stat = SRUN;
210 	pp->p_flag = SSYS;
211 
212 	pp->p_pidp = &pid0;
213 	pp->p_pgidp = &pid0;
214 	pp->p_sessp = &session0;
215 	pp->p_tlist = &t0;
216 	pid0.pid_pglink = pp;
217 	pid0.pid_pgtail = pp;
218 
219 	/*
220 	 * XXX - we asssume that the u-area is zeroed out except for
221 	 * ttolwp(curthread)->lwp_regs.
222 	 */
223 	PTOU(curproc)->u_cmask = (mode_t)CMASK;
224 
225 	thread_init();		/* init thread_free list */
226 	pid_init();		/* initialize pid (proc) table */
227 	contract_init();	/* initialize contracts */
228 
229 	init_pages_pp_maximum();
230 }
231 
232 /*
233  * Load a procedure into a thread.
234  */
235 void
236 thread_load(kthread_t *t, void (*start)(), caddr_t arg, size_t len)
237 {
238 	caddr_t sp;
239 	size_t framesz;
240 	caddr_t argp;
241 	long *p;
242 	extern void thread_start();
243 
244 	/*
245 	 * Push a "c" call frame onto the stack to represent
246 	 * the caller of "start".
247 	 */
248 	sp = t->t_stk;
249 	ASSERT(((uintptr_t)t->t_stk & (STACK_ENTRY_ALIGN - 1)) == 0);
250 	if (len != 0) {
251 		/*
252 		 * the object that arg points at is copied into the
253 		 * caller's frame.
254 		 */
255 		framesz = SA(len);
256 		sp -= framesz;
257 		ASSERT(sp > t->t_stkbase);
258 		argp = sp + SA(MINFRAME);
259 		bcopy(arg, argp, len);
260 		arg = argp;
261 	}
262 	/*
263 	 * Set up arguments (arg and len) on the caller's stack frame.
264 	 */
265 	p = (long *)sp;
266 
267 	*--p = 0;		/* fake call */
268 	*--p = 0;		/* null frame pointer terminates stack trace */
269 	*--p = (long)len;
270 	*--p = (intptr_t)arg;
271 	*--p = (intptr_t)start;
272 
273 	/*
274 	 * initialize thread to resume at thread_start() which will
275 	 * turn around and invoke (*start)(arg, len).
276 	 */
277 	t->t_pc = (uintptr_t)thread_start;
278 	t->t_sp = (uintptr_t)p;
279 
280 	ASSERT((t->t_sp & (STACK_ENTRY_ALIGN - 1)) == 0);
281 }
282 
283 /*
284  * load user registers into lwp.
285  */
286 /*ARGSUSED2*/
287 void
288 lwp_load(klwp_t *lwp, gregset_t grp, uintptr_t thrptr)
289 {
290 	struct regs *rp = lwptoregs(lwp);
291 
292 	setgregs(lwp, grp);
293 	rp->r_ps = PSL_USER;
294 
295 	/*
296 	 * For 64-bit lwps, we allow null %fs selector value, and null
297 	 * %gs selector to point anywhere in the address space using
298 	 * %fsbase and %gsbase behind the scenes.  libc uses %fs to point
299 	 * at the ulwp_t structure.
300 	 *
301 	 * For 32-bit lwps, libc wedges its lwp thread pointer into the
302 	 * ucontext ESP slot (which is otherwise irrelevant to setting a
303 	 * ucontext) and LWPGS_SEL value into gregs[REG_GS].  This is so
304 	 * syslwp_create() can atomically setup %gs.
305 	 *
306 	 * See setup_context() in libc.
307 	 */
308 #ifdef _SYSCALL32_IMPL
309 	if (lwp_getdatamodel(lwp) == DATAMODEL_ILP32) {
310 		if (grp[REG_GS] == LWPGS_SEL)
311 			(void) lwp_setprivate(lwp, _LWP_GSBASE, thrptr);
312 	} else {
313 		/*
314 		 * See lwp_setprivate in kernel and setup_context in libc.
315 		 *
316 		 * Currently libc constructs a ucontext from whole cloth for
317 		 * every new (not main) lwp created.  For 64 bit processes
318 		 * %fsbase is directly set to point to current thread pointer.
319 		 * In the past (solaris 10) %fs was also set LWPFS_SEL to
320 		 * indicate %fsbase. Now we use the null GDT selector for
321 		 * this purpose. LWP[FS|GS]_SEL are only intended for 32 bit
322 		 * processes. To ease transition we support older libcs in
323 		 * the newer kernel by forcing %fs or %gs selector to null
324 		 * by calling lwp_setprivate if LWP[FS|GS]_SEL is passed in
325 		 * the ucontext.  This is should be ripped out at some future
326 		 * date.  Another fix would be for libc to do a getcontext
327 		 * and inherit the null %fs/%gs from the current context but
328 		 * that means an extra system call and could hurt performance.
329 		 */
330 		if (grp[REG_FS] == 0x1bb) /* hard code legacy LWPFS_SEL */
331 		    (void) lwp_setprivate(lwp, _LWP_FSBASE,
332 		    (uintptr_t)grp[REG_FSBASE]);
333 
334 		if (grp[REG_GS] == 0x1c3) /* hard code legacy LWPGS_SEL */
335 		    (void) lwp_setprivate(lwp, _LWP_GSBASE,
336 		    (uintptr_t)grp[REG_GSBASE]);
337 	}
338 #else
339 	if (grp[GS] == LWPGS_SEL)
340 		(void) lwp_setprivate(lwp, _LWP_GSBASE, thrptr);
341 #endif
342 
343 	lwp->lwp_eosys = JUSTRETURN;
344 	lwptot(lwp)->t_post_sys = 1;
345 }
346 
347 /*
348  * set syscall()'s return values for a lwp.
349  */
350 void
351 lwp_setrval(klwp_t *lwp, int v1, int v2)
352 {
353 	lwptoregs(lwp)->r_ps &= ~PS_C;
354 	lwptoregs(lwp)->r_r0 = v1;
355 	lwptoregs(lwp)->r_r1 = v2;
356 }
357 
358 /*
359  * set syscall()'s return values for a lwp.
360  */
361 void
362 lwp_setsp(klwp_t *lwp, caddr_t sp)
363 {
364 	lwptoregs(lwp)->r_sp = (intptr_t)sp;
365 }
366 
367 /*
368  * Copy regs from parent to child.
369  */
370 void
371 lwp_forkregs(klwp_t *lwp, klwp_t *clwp)
372 {
373 #if defined(__amd64)
374 	struct pcb *pcb = &clwp->lwp_pcb;
375 	struct regs *rp = lwptoregs(lwp);
376 
377 	if ((pcb->pcb_flags & RUPDATE_PENDING) == 0) {
378 		pcb->pcb_ds = rp->r_ds;
379 		pcb->pcb_es = rp->r_es;
380 		pcb->pcb_fs = rp->r_fs;
381 		pcb->pcb_gs = rp->r_gs;
382 		pcb->pcb_flags |= RUPDATE_PENDING;
383 		lwptot(clwp)->t_post_sys = 1;
384 	}
385 	ASSERT(lwptot(clwp)->t_post_sys);
386 #endif
387 
388 	bcopy(lwp->lwp_regs, clwp->lwp_regs, sizeof (struct regs));
389 }
390 
391 /*
392  * This function is currently unused on x86.
393  */
394 /*ARGSUSED*/
395 void
396 lwp_freeregs(klwp_t *lwp, int isexec)
397 {}
398 
399 /*
400  * This function is currently unused on x86.
401  */
402 void
403 lwp_pcb_exit(void)
404 {}
405 
406 /*
407  * Lwp context ops for segment registers.
408  */
409 
410 /*
411  * Every time we come into the kernel (syscall, interrupt or trap
412  * but not fast-traps) we capture the current values of the user's
413  * segment registers into the lwp's reg structure. This includes
414  * lcall for i386 generic system call support since it is handled
415  * as a segment-not-present trap.
416  *
417  * Here we save the current values from the lwp regs into the pcb
418  * and set the RUPDATE_PENDING bit to tell the rest of the kernel
419  * that the pcb copy of the segment registers is the current one.
420  * This ensures the lwp's next trip to user land via update_sregs.
421  * Finally we set t_post_sys to ensure that no system call fast-path's
422  * its way out of the kernel via sysret.
423  *
424  * (This means that we need to have interrupts disabled when we test
425  * t->t_post_sys in the syscall handlers; if the test fails, we need
426  * to keep interrupts disabled until we return to userland so we can't
427  * be switched away.)
428  *
429  * As a result of all this, we don't really have to do a whole lot if
430  * the thread is just mucking about in the kernel, switching on and
431  * off the cpu for whatever reason it feels like. And yet we still
432  * preserve fast syscalls, cause if we -don't- get descheduled,
433  * we never come here either.
434  */
435 
436 #define	VALID_LWP_DESC(udp) ((udp)->usd_type == SDT_MEMRWA && \
437 	    (udp)->usd_p == 1 && (udp)->usd_dpl == SEL_UPL)
438 
439 void
440 lwp_segregs_save(klwp_t *lwp)
441 {
442 #if defined(__amd64)
443 	pcb_t *pcb = &lwp->lwp_pcb;
444 	struct regs *rp;
445 
446 	ASSERT(VALID_LWP_DESC(&pcb->pcb_fsdesc));
447 	ASSERT(VALID_LWP_DESC(&pcb->pcb_gsdesc));
448 
449 	if ((pcb->pcb_flags & RUPDATE_PENDING) == 0) {
450 		rp = lwptoregs(lwp);
451 
452 		/*
453 		 * If there's no update already pending, capture the current
454 		 * %ds/%es/%fs/%gs values from lwp's regs in case the user
455 		 * changed them; %fsbase and %gsbase are privileged so the
456 		 * kernel versions of these registers in pcb_fsbase and
457 		 * pcb_gsbase are always up-to-date.
458 		 */
459 		pcb->pcb_ds = rp->r_ds;
460 		pcb->pcb_es = rp->r_es;
461 		pcb->pcb_fs = rp->r_fs;
462 		pcb->pcb_gs = rp->r_gs;
463 		pcb->pcb_flags |= RUPDATE_PENDING;
464 		lwp->lwp_thread->t_post_sys = 1;
465 	}
466 #endif	/* __amd64 */
467 
468 	ASSERT(bcmp(&CPU->cpu_gdt[GDT_LWPFS], &lwp->lwp_pcb.pcb_fsdesc,
469 	    sizeof (lwp->lwp_pcb.pcb_fsdesc)) == 0);
470 	ASSERT(bcmp(&CPU->cpu_gdt[GDT_LWPGS], &lwp->lwp_pcb.pcb_gsdesc,
471 	    sizeof (lwp->lwp_pcb.pcb_gsdesc)) == 0);
472 }
473 
474 #if defined(__amd64)
475 
476 /*
477  * Update the segment registers with new values from the pcb
478  *
479  * We have to do this carefully, and in the following order,
480  * in case any of the selectors points at a bogus descriptor.
481  * If they do, we'll catch trap with on_trap and return 1.
482  * returns 0 on success.
483  *
484  * This is particularly tricky for %gs.
485  * This routine must be executed under a cli.
486  */
487 int
488 update_sregs(struct regs *rp,  klwp_t *lwp)
489 {
490 	pcb_t *pcb = &lwp->lwp_pcb;
491 	ulong_t	kgsbase;
492 	on_trap_data_t	otd;
493 	int rc = 0;
494 
495 	if (!on_trap(&otd, OT_SEGMENT_ACCESS)) {
496 
497 		kgsbase = (ulong_t)CPU;
498 		__set_gs(pcb->pcb_gs);
499 
500 		/*
501 		 * If __set_gs fails it's because the new %gs is a bad %gs,
502 		 * we'll be taking a trap but with the original %gs and %gsbase
503 		 * undamaged (i.e. pointing at curcpu).
504 		 *
505 		 * We've just mucked up the kernel's gsbase.  Oops.  In
506 		 * particular we can't take any traps at all.  Make the newly
507 		 * computed gsbase be the hidden gs via __swapgs , and fix
508 		 * the kernel's gsbase back again. Later, when we return to
509 		 * userland we'll swapgs again restoring gsbase just loaded
510 		 * above.
511 		 */
512 		__swapgs();
513 		rp->r_gs = pcb->pcb_gs;
514 
515 		/*
516 		 * restore kernel's gsbase
517 		 */
518 		wrmsr(MSR_AMD_GSBASE, kgsbase);
519 
520 		/*
521 		 * Only override the descriptor base address if
522 		 * r_gs == LWPGS_SEL or if r_gs == NULL. A note on
523 		 * NULL descriptors -- 32-bit programs take faults
524 		 * if they deference NULL descriptors; however,
525 		 * when 64-bit programs load them into %fs or %gs,
526 		 * they DONT fault -- only the base address remains
527 		 * whatever it was from the last load.   Urk.
528 		 *
529 		 * XXX - note that lwp_setprivate now sets %fs/%gs to the
530 		 * null selector for 64 bit processes. Whereas before
531 		 * %fs/%gs were set to LWP(FS|GS)_SEL regardless of
532 		 * the process's data model. For now we check for both
533 		 * values so that the kernel can also support the older
534 		 * libc. This should be ripped out at some point in the
535 		 * future.
536 		 */
537 		if (pcb->pcb_gs == LWPGS_SEL || pcb->pcb_gs == 0)
538 			wrmsr(MSR_AMD_KGSBASE, pcb->pcb_gsbase);
539 
540 		__set_ds(pcb->pcb_ds);
541 		rp->r_ds = pcb->pcb_ds;
542 
543 		__set_es(pcb->pcb_es);
544 		rp->r_es = pcb->pcb_es;
545 
546 		__set_fs(pcb->pcb_fs);
547 		rp->r_fs = pcb->pcb_fs;
548 
549 		/*
550 		 * Same as for %gs
551 		 */
552 		if (pcb->pcb_fs == LWPFS_SEL || pcb->pcb_fs == 0)
553 			wrmsr(MSR_AMD_FSBASE, pcb->pcb_fsbase);
554 
555 	} else {
556 		cli();
557 		rc = 1;
558 	}
559 	no_trap();
560 	return (rc);
561 }
562 #endif	/* __amd64 */
563 
564 #ifdef _SYSCALL32_IMPL
565 
566 /*
567  * Make it impossible for a process to change its data model.
568  * We do this by toggling the present bits for the 32 and
569  * 64-bit user code descriptors. That way if a user lwp attempts
570  * to change its data model (by using the wrong code descriptor in
571  * %cs) it will fault immediately. This also allows us to simplify
572  * assertions and checks in the kernel.
573  */
574 static void
575 gdt_ucode_model(model_t model)
576 {
577 	cpu_t *cpu;
578 
579 	kpreempt_disable();
580 	cpu = CPU;
581 	if (model == DATAMODEL_NATIVE) {
582 		cpu->cpu_gdt[GDT_UCODE].usd_p = 1;
583 		cpu->cpu_gdt[GDT_U32CODE].usd_p = 0;
584 	} else {
585 		cpu->cpu_gdt[GDT_U32CODE].usd_p = 1;
586 		cpu->cpu_gdt[GDT_UCODE].usd_p = 0;
587 	}
588 	kpreempt_enable();
589 }
590 
591 #endif	/* _SYSCALL32_IMPL */
592 
593 /*
594  * Restore lwp private fs and gs segment descriptors
595  * on current cpu's GDT.
596  */
597 static void
598 lwp_segregs_restore(klwp_t *lwp)
599 {
600 	pcb_t *pcb = &lwp->lwp_pcb;
601 	cpu_t *cpu = CPU;
602 
603 	ASSERT(VALID_LWP_DESC(&pcb->pcb_fsdesc));
604 	ASSERT(VALID_LWP_DESC(&pcb->pcb_gsdesc));
605 
606 #ifdef	_SYSCALL32_IMPL
607 	gdt_ucode_model(DATAMODEL_NATIVE);
608 #endif
609 
610 	cpu->cpu_gdt[GDT_LWPFS] = pcb->pcb_fsdesc;
611 	cpu->cpu_gdt[GDT_LWPGS] = pcb->pcb_gsdesc;
612 
613 }
614 
615 #ifdef _SYSCALL32_IMPL
616 
617 static void
618 lwp_segregs_restore32(klwp_t *lwp)
619 {
620 	/*LINTED*/
621 	cpu_t *cpu = CPU;
622 	pcb_t *pcb = &lwp->lwp_pcb;
623 
624 	ASSERT(VALID_LWP_DESC(&lwp->lwp_pcb.pcb_fsdesc));
625 	ASSERT(VALID_LWP_DESC(&lwp->lwp_pcb.pcb_gsdesc));
626 
627 	gdt_ucode_model(DATAMODEL_ILP32);
628 	cpu->cpu_gdt[GDT_LWPFS] = pcb->pcb_fsdesc;
629 	cpu->cpu_gdt[GDT_LWPGS] = pcb->pcb_gsdesc;
630 }
631 
632 #endif	/* _SYSCALL32_IMPL */
633 
634 /*
635  * If this is a process in a branded zone, then we want it to use the brand
636  * syscall entry points instead of the standard Solaris entry points.  This
637  * routine must be called when a new lwp is created within a branded zone
638  * or when an existing lwp moves into a branded zone via a zone_enter()
639  * operation.
640  */
641 void
642 lwp_attach_brand_hdlrs(klwp_t *lwp)
643 {
644 	kthread_t *t = lwptot(lwp);
645 
646 	ASSERT(PROC_IS_BRANDED(lwptoproc(lwp)));
647 	ASSERT(removectx(t, NULL, brand_interpositioning_disable,
648 	    brand_interpositioning_enable, NULL, NULL, NULL, NULL) == 0);
649 
650 	installctx(t, NULL, brand_interpositioning_disable,
651 	    brand_interpositioning_enable, NULL, NULL, NULL, NULL);
652 
653 	if (t == curthread) {
654 		kpreempt_disable();
655 		brand_interpositioning_enable();
656 		kpreempt_enable();
657 	}
658 }
659 
660 /*
661  * Add any lwp-associated context handlers to the lwp at the beginning
662  * of the lwp's useful life.
663  *
664  * All paths which create lwp's invoke lwp_create(); lwp_create()
665  * invokes lwp_stk_init() which initializes the stack, sets up
666  * lwp_regs, and invokes this routine.
667  *
668  * All paths which destroy lwp's invoke lwp_exit() to rip the lwp
669  * apart and put it on 'lwp_deathrow'; if the lwp is destroyed it
670  * ends up in thread_free() which invokes freectx(t, 0) before
671  * invoking lwp_stk_fini().  When the lwp is recycled from death
672  * row, lwp_stk_fini() is invoked, then thread_free(), and thus
673  * freectx(t, 0) as before.
674  *
675  * In the case of exec, the surviving lwp is thoroughly scrubbed
676  * clean; exec invokes freectx(t, 1) to destroy associated contexts.
677  * On the way back to the new image, it invokes setregs() which
678  * in turn invokes this routine.
679  */
680 void
681 lwp_installctx(klwp_t *lwp)
682 {
683 	kthread_t *t = lwptot(lwp);
684 	int thisthread = t == curthread;
685 #ifdef _SYSCALL32_IMPL
686 	void (*restop)(klwp_t *) = lwp_getdatamodel(lwp) == DATAMODEL_NATIVE ?
687 	    lwp_segregs_restore : lwp_segregs_restore32;
688 #else
689 	void (*restop)(klwp_t *) = lwp_segregs_restore;
690 #endif
691 
692 	/*
693 	 * Install the basic lwp context handlers on each lwp.
694 	 *
695 	 * On the amd64 kernel, the context handlers are responsible for
696 	 * virtualizing %ds, %es, %fs, and %gs to the lwp.  The register
697 	 * values are only ever changed via sys_rtt when the
698 	 * RUPDATE_PENDING bit is set.  Only sys_rtt gets to clear the bit.
699 	 *
700 	 * On the i386 kernel, the context handlers are responsible for
701 	 * virtualizing %gs/%fs to the lwp by updating the per-cpu GDTs
702 	 */
703 	ASSERT(removectx(t, lwp, lwp_segregs_save, restop,
704 	    NULL, NULL, NULL, NULL) == 0);
705 	if (thisthread)
706 		kpreempt_disable();
707 	installctx(t, lwp, lwp_segregs_save, restop,
708 	    NULL, NULL, NULL, NULL);
709 	if (thisthread) {
710 		/*
711 		 * Since we're the right thread, set the values in the GDT
712 		 */
713 		restop(lwp);
714 		kpreempt_enable();
715 	}
716 
717 	/*
718 	 * If we have sysenter/sysexit instructions enabled, we need
719 	 * to ensure that the hardware mechanism is kept up-to-date with the
720 	 * lwp's kernel stack pointer across context switches.
721 	 *
722 	 * sep_save zeros the sysenter stack pointer msr; sep_restore sets
723 	 * it to the lwp's kernel stack pointer (kstktop).
724 	 */
725 	if (x86_feature & X86_SEP) {
726 #if defined(__amd64)
727 		caddr_t kstktop = (caddr_t)lwp->lwp_regs;
728 #elif defined(__i386)
729 		caddr_t kstktop = ((caddr_t)lwp->lwp_regs - MINFRAME) +
730 		    SA(sizeof (struct regs) + MINFRAME);
731 #endif
732 		ASSERT(removectx(t, kstktop,
733 		    sep_save, sep_restore, NULL, NULL, NULL, NULL) == 0);
734 
735 		if (thisthread)
736 			kpreempt_disable();
737 		installctx(t, kstktop,
738 		    sep_save, sep_restore, NULL, NULL, NULL, NULL);
739 		if (thisthread) {
740 			/*
741 			 * We're the right thread, so set the stack pointer
742 			 * for the first sysenter instruction to use
743 			 */
744 			sep_restore(kstktop);
745 			kpreempt_enable();
746 		}
747 	}
748 
749 	if (PROC_IS_BRANDED(ttoproc(t)))
750 		lwp_attach_brand_hdlrs(lwp);
751 }
752 
753 /*
754  * Clear registers on exec(2).
755  */
756 void
757 setregs(uarg_t *args)
758 {
759 	struct regs *rp;
760 	kthread_t *t = curthread;
761 	klwp_t *lwp = ttolwp(t);
762 	pcb_t *pcb = &lwp->lwp_pcb;
763 	greg_t sp;
764 
765 	/*
766 	 * Initialize user registers
767 	 */
768 	(void) save_syscall_args();	/* copy args from registers first */
769 	rp = lwptoregs(lwp);
770 	sp = rp->r_sp;
771 	bzero(rp, sizeof (*rp));
772 
773 	rp->r_ss = UDS_SEL;
774 	rp->r_sp = sp;
775 	rp->r_pc = args->entry;
776 	rp->r_ps = PSL_USER;
777 
778 #if defined(__amd64)
779 
780 	pcb->pcb_fs = pcb->pcb_gs = 0;
781 	pcb->pcb_fsbase = pcb->pcb_gsbase = 0;
782 
783 	if (ttoproc(t)->p_model == DATAMODEL_NATIVE) {
784 
785 		rp->r_cs = UCS_SEL;
786 
787 		/*
788 		 * Only allow 64-bit user code descriptor to be present.
789 		 */
790 		gdt_ucode_model(DATAMODEL_NATIVE);
791 
792 		/*
793 		 * Arrange that the virtualized %fs and %gs GDT descriptors
794 		 * have a well-defined initial state (present, ring 3
795 		 * and of type data).
796 		 */
797 		pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_udesc;
798 
799 		/*
800 		 * thrptr is either NULL or a value used by DTrace.
801 		 * 64-bit processes use %fs as their "thread" register.
802 		 */
803 		if (args->thrptr)
804 			(void) lwp_setprivate(lwp, _LWP_FSBASE, args->thrptr);
805 
806 	} else {
807 
808 		rp->r_cs = U32CS_SEL;
809 		rp->r_ds = rp->r_es = UDS_SEL;
810 
811 		/*
812 		 * only allow 32-bit user code selector to be present.
813 		 */
814 		gdt_ucode_model(DATAMODEL_ILP32);
815 
816 		pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_u32desc;
817 
818 		/*
819 		 * thrptr is either NULL or a value used by DTrace.
820 		 * 32-bit processes use %gs as their "thread" register.
821 		 */
822 		if (args->thrptr)
823 			(void) lwp_setprivate(lwp, _LWP_GSBASE, args->thrptr);
824 
825 	}
826 
827 	pcb->pcb_ds = rp->r_ds;
828 	pcb->pcb_es = rp->r_es;
829 	pcb->pcb_flags |= RUPDATE_PENDING;
830 
831 #elif defined(__i386)
832 
833 	rp->r_cs = UCS_SEL;
834 	rp->r_ds = rp->r_es = UDS_SEL;
835 
836 	/*
837 	 * Arrange that the virtualized %fs and %gs GDT descriptors
838 	 * have a well-defined initial state (present, ring 3
839 	 * and of type data).
840 	 */
841 	pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_udesc;
842 
843 	/*
844 	 * For %gs we need to reset LWP_GSBASE in pcb and the
845 	 * per-cpu GDT descriptor. thrptr is either NULL
846 	 * or a value used by DTrace.
847 	 */
848 	if (args->thrptr)
849 		(void) lwp_setprivate(lwp, _LWP_GSBASE, args->thrptr);
850 #endif
851 
852 	lwp->lwp_eosys = JUSTRETURN;
853 	t->t_post_sys = 1;
854 
855 	/*
856 	 * Here we initialize minimal fpu state.
857 	 * The rest is done at the first floating
858 	 * point instruction that a process executes.
859 	 */
860 	pcb->pcb_fpu.fpu_flags = 0;
861 
862 	/*
863 	 * Add the lwp context handlers that virtualize segment registers,
864 	 * and/or system call stacks etc.
865 	 */
866 	lwp_installctx(lwp);
867 }
868 
869 user_desc_t *
870 cpu_get_gdt(void)
871 {
872 	return (CPU->cpu_gdt);
873 }
874 
875 
876 #if !defined(lwp_getdatamodel)
877 
878 /*
879  * Return the datamodel of the given lwp.
880  */
881 /*ARGSUSED*/
882 model_t
883 lwp_getdatamodel(klwp_t *lwp)
884 {
885 	return (lwp->lwp_procp->p_model);
886 }
887 
888 #endif	/* !lwp_getdatamodel */
889 
890 #if !defined(get_udatamodel)
891 
892 model_t
893 get_udatamodel(void)
894 {
895 	return (curproc->p_model);
896 }
897 
898 #endif	/* !get_udatamodel */
899