xref: /titanic_51/usr/src/uts/intel/ia32/os/sundep.c (revision 25c28e83beb90e7c80452a7c818c5e6f73a07dc8)
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 (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*	Copyright (c) 1990, 1991 UNIX System Laboratories, Inc. */
26 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989, 1990 AT&T   */
27 /*	All Rights Reserved   */
28 
29 #include <sys/types.h>
30 #include <sys/param.h>
31 #include <sys/sysmacros.h>
32 #include <sys/signal.h>
33 #include <sys/systm.h>
34 #include <sys/user.h>
35 #include <sys/mman.h>
36 #include <sys/class.h>
37 #include <sys/proc.h>
38 #include <sys/procfs.h>
39 #include <sys/buf.h>
40 #include <sys/kmem.h>
41 #include <sys/cred.h>
42 #include <sys/archsystm.h>
43 #include <sys/vmparam.h>
44 #include <sys/prsystm.h>
45 #include <sys/reboot.h>
46 #include <sys/uadmin.h>
47 #include <sys/vfs.h>
48 #include <sys/vnode.h>
49 #include <sys/file.h>
50 #include <sys/session.h>
51 #include <sys/ucontext.h>
52 #include <sys/dnlc.h>
53 #include <sys/var.h>
54 #include <sys/cmn_err.h>
55 #include <sys/debugreg.h>
56 #include <sys/thread.h>
57 #include <sys/vtrace.h>
58 #include <sys/consdev.h>
59 #include <sys/psw.h>
60 #include <sys/regset.h>
61 #include <sys/privregs.h>
62 #include <sys/cpu.h>
63 #include <sys/stack.h>
64 #include <sys/swap.h>
65 #include <vm/hat.h>
66 #include <vm/anon.h>
67 #include <vm/as.h>
68 #include <vm/page.h>
69 #include <vm/seg.h>
70 #include <vm/seg_kmem.h>
71 #include <vm/seg_map.h>
72 #include <vm/seg_vn.h>
73 #include <sys/exec.h>
74 #include <sys/acct.h>
75 #include <sys/core.h>
76 #include <sys/corectl.h>
77 #include <sys/modctl.h>
78 #include <sys/tuneable.h>
79 #include <c2/audit.h>
80 #include <sys/bootconf.h>
81 #include <sys/brand.h>
82 #include <sys/dumphdr.h>
83 #include <sys/promif.h>
84 #include <sys/systeminfo.h>
85 #include <sys/kdi.h>
86 #include <sys/contract_impl.h>
87 #include <sys/x86_archext.h>
88 #include <sys/segments.h>
89 #include <sys/ontrap.h>
90 #include <sys/cpu.h>
91 #ifdef __xpv
92 #include <sys/hypervisor.h>
93 #endif
94 
95 /*
96  * Compare the version of boot that boot says it is against
97  * the version of boot the kernel expects.
98  */
99 int
100 check_boot_version(int boots_version)
101 {
102 	if (boots_version == BO_VERSION)
103 		return (0);
104 
105 	prom_printf("Wrong boot interface - kernel needs v%d found v%d\n",
106 	    BO_VERSION, boots_version);
107 	prom_panic("halting");
108 	/*NOTREACHED*/
109 }
110 
111 /*
112  * Process the physical installed list for boot.
113  * Finds:
114  * 1) the pfn of the highest installed physical page,
115  * 2) the number of pages installed
116  * 3) the number of distinct contiguous regions these pages fall into.
117  * 4) the number of contiguous memory ranges
118  */
119 void
120 installed_top_size_ex(
121 	struct memlist *list,	/* pointer to start of installed list */
122 	pfn_t *high_pfn,	/* return ptr for top value */
123 	pgcnt_t *pgcnt,		/* return ptr for sum of installed pages */
124 	int	*ranges)	/* return ptr for the count of contig. ranges */
125 {
126 	pfn_t top = 0;
127 	pgcnt_t sumpages = 0;
128 	pfn_t highp;		/* high page in a chunk */
129 	int cnt = 0;
130 
131 	for (; list; list = list->ml_next) {
132 		++cnt;
133 		highp = (list->ml_address + list->ml_size - 1) >> PAGESHIFT;
134 		if (top < highp)
135 			top = highp;
136 		sumpages += btop(list->ml_size);
137 	}
138 
139 	*high_pfn = top;
140 	*pgcnt = sumpages;
141 	*ranges = cnt;
142 }
143 
144 void
145 installed_top_size(
146 	struct memlist *list,	/* pointer to start of installed list */
147 	pfn_t *high_pfn,	/* return ptr for top value */
148 	pgcnt_t *pgcnt)		/* return ptr for sum of installed pages */
149 {
150 	int ranges;
151 
152 	installed_top_size_ex(list, high_pfn, pgcnt, &ranges);
153 }
154 
155 void
156 phys_install_has_changed(void)
157 {}
158 
159 /*
160  * Copy in a memory list from boot to kernel, with a filter function
161  * to remove pages. The filter function can increase the address and/or
162  * decrease the size to filter out pages.  It will also align addresses and
163  * sizes to PAGESIZE.
164  */
165 void
166 copy_memlist_filter(
167 	struct memlist *src,
168 	struct memlist **dstp,
169 	void (*filter)(uint64_t *, uint64_t *))
170 {
171 	struct memlist *dst, *prev;
172 	uint64_t addr;
173 	uint64_t size;
174 	uint64_t eaddr;
175 
176 	dst = *dstp;
177 	prev = dst;
178 
179 	/*
180 	 * Move through the memlist applying a filter against
181 	 * each range of memory. Note that we may apply the
182 	 * filter multiple times against each memlist entry.
183 	 */
184 	for (; src; src = src->ml_next) {
185 		addr = P2ROUNDUP(src->ml_address, PAGESIZE);
186 		eaddr = P2ALIGN(src->ml_address + src->ml_size, PAGESIZE);
187 		while (addr < eaddr) {
188 			size = eaddr - addr;
189 			if (filter != NULL)
190 				filter(&addr, &size);
191 			if (size == 0)
192 				break;
193 			dst->ml_address = addr;
194 			dst->ml_size = size;
195 			dst->ml_next = 0;
196 			if (prev == dst) {
197 				dst->ml_prev = 0;
198 				dst++;
199 			} else {
200 				dst->ml_prev = prev;
201 				prev->ml_next = dst;
202 				dst++;
203 				prev++;
204 			}
205 			addr += size;
206 		}
207 	}
208 
209 	*dstp = dst;
210 }
211 
212 /*
213  * Kernel setup code, called from startup().
214  */
215 void
216 kern_setup1(void)
217 {
218 	proc_t *pp;
219 
220 	pp = &p0;
221 
222 	proc_sched = pp;
223 
224 	/*
225 	 * Initialize process 0 data structures
226 	 */
227 	pp->p_stat = SRUN;
228 	pp->p_flag = SSYS;
229 
230 	pp->p_pidp = &pid0;
231 	pp->p_pgidp = &pid0;
232 	pp->p_sessp = &session0;
233 	pp->p_tlist = &t0;
234 	pid0.pid_pglink = pp;
235 	pid0.pid_pgtail = pp;
236 
237 	/*
238 	 * XXX - we asssume that the u-area is zeroed out except for
239 	 * ttolwp(curthread)->lwp_regs.
240 	 */
241 	PTOU(curproc)->u_cmask = (mode_t)CMASK;
242 
243 	thread_init();		/* init thread_free list */
244 	pid_init();		/* initialize pid (proc) table */
245 	contract_init();	/* initialize contracts */
246 
247 	init_pages_pp_maximum();
248 }
249 
250 /*
251  * Load a procedure into a thread.
252  */
253 void
254 thread_load(kthread_t *t, void (*start)(), caddr_t arg, size_t len)
255 {
256 	caddr_t sp;
257 	size_t framesz;
258 	caddr_t argp;
259 	long *p;
260 	extern void thread_start();
261 
262 	/*
263 	 * Push a "c" call frame onto the stack to represent
264 	 * the caller of "start".
265 	 */
266 	sp = t->t_stk;
267 	ASSERT(((uintptr_t)t->t_stk & (STACK_ENTRY_ALIGN - 1)) == 0);
268 	if (len != 0) {
269 		/*
270 		 * the object that arg points at is copied into the
271 		 * caller's frame.
272 		 */
273 		framesz = SA(len);
274 		sp -= framesz;
275 		ASSERT(sp > t->t_stkbase);
276 		argp = sp + SA(MINFRAME);
277 		bcopy(arg, argp, len);
278 		arg = argp;
279 	}
280 	/*
281 	 * Set up arguments (arg and len) on the caller's stack frame.
282 	 */
283 	p = (long *)sp;
284 
285 	*--p = 0;		/* fake call */
286 	*--p = 0;		/* null frame pointer terminates stack trace */
287 	*--p = (long)len;
288 	*--p = (intptr_t)arg;
289 	*--p = (intptr_t)start;
290 
291 	/*
292 	 * initialize thread to resume at thread_start() which will
293 	 * turn around and invoke (*start)(arg, len).
294 	 */
295 	t->t_pc = (uintptr_t)thread_start;
296 	t->t_sp = (uintptr_t)p;
297 
298 	ASSERT((t->t_sp & (STACK_ENTRY_ALIGN - 1)) == 0);
299 }
300 
301 /*
302  * load user registers into lwp.
303  */
304 /*ARGSUSED2*/
305 void
306 lwp_load(klwp_t *lwp, gregset_t grp, uintptr_t thrptr)
307 {
308 	struct regs *rp = lwptoregs(lwp);
309 
310 	setgregs(lwp, grp);
311 	rp->r_ps = PSL_USER;
312 
313 	/*
314 	 * For 64-bit lwps, we allow one magic %fs selector value, and one
315 	 * magic %gs selector to point anywhere in the address space using
316 	 * %fsbase and %gsbase behind the scenes.  libc uses %fs to point
317 	 * at the ulwp_t structure.
318 	 *
319 	 * For 32-bit lwps, libc wedges its lwp thread pointer into the
320 	 * ucontext ESP slot (which is otherwise irrelevant to setting a
321 	 * ucontext) and LWPGS_SEL value into gregs[REG_GS].  This is so
322 	 * syslwp_create() can atomically setup %gs.
323 	 *
324 	 * See setup_context() in libc.
325 	 */
326 #ifdef _SYSCALL32_IMPL
327 	if (lwp_getdatamodel(lwp) == DATAMODEL_ILP32) {
328 		if (grp[REG_GS] == LWPGS_SEL)
329 			(void) lwp_setprivate(lwp, _LWP_GSBASE, thrptr);
330 	} else {
331 		/*
332 		 * See lwp_setprivate in kernel and setup_context in libc.
333 		 *
334 		 * Currently libc constructs a ucontext from whole cloth for
335 		 * every new (not main) lwp created.  For 64 bit processes
336 		 * %fsbase is directly set to point to current thread pointer.
337 		 * In the past (solaris 10) %fs was also set LWPFS_SEL to
338 		 * indicate %fsbase. Now we use the null GDT selector for
339 		 * this purpose. LWP[FS|GS]_SEL are only intended for 32 bit
340 		 * processes. To ease transition we support older libcs in
341 		 * the newer kernel by forcing %fs or %gs selector to null
342 		 * by calling lwp_setprivate if LWP[FS|GS]_SEL is passed in
343 		 * the ucontext.  This is should be ripped out at some future
344 		 * date.  Another fix would be for libc to do a getcontext
345 		 * and inherit the null %fs/%gs from the current context but
346 		 * that means an extra system call and could hurt performance.
347 		 */
348 		if (grp[REG_FS] == 0x1bb) /* hard code legacy LWPFS_SEL */
349 			(void) lwp_setprivate(lwp, _LWP_FSBASE,
350 			    (uintptr_t)grp[REG_FSBASE]);
351 
352 		if (grp[REG_GS] == 0x1c3) /* hard code legacy LWPGS_SEL */
353 			(void) lwp_setprivate(lwp, _LWP_GSBASE,
354 			    (uintptr_t)grp[REG_GSBASE]);
355 	}
356 #else
357 	if (grp[GS] == LWPGS_SEL)
358 		(void) lwp_setprivate(lwp, _LWP_GSBASE, thrptr);
359 #endif
360 
361 	lwp->lwp_eosys = JUSTRETURN;
362 	lwptot(lwp)->t_post_sys = 1;
363 }
364 
365 /*
366  * set syscall()'s return values for a lwp.
367  */
368 void
369 lwp_setrval(klwp_t *lwp, int v1, int v2)
370 {
371 	lwptoregs(lwp)->r_ps &= ~PS_C;
372 	lwptoregs(lwp)->r_r0 = v1;
373 	lwptoregs(lwp)->r_r1 = v2;
374 }
375 
376 /*
377  * set syscall()'s return values for a lwp.
378  */
379 void
380 lwp_setsp(klwp_t *lwp, caddr_t sp)
381 {
382 	lwptoregs(lwp)->r_sp = (intptr_t)sp;
383 }
384 
385 /*
386  * Copy regs from parent to child.
387  */
388 void
389 lwp_forkregs(klwp_t *lwp, klwp_t *clwp)
390 {
391 #if defined(__amd64)
392 	struct pcb *pcb = &clwp->lwp_pcb;
393 	struct regs *rp = lwptoregs(lwp);
394 
395 	if (pcb->pcb_rupdate == 0) {
396 		pcb->pcb_ds = rp->r_ds;
397 		pcb->pcb_es = rp->r_es;
398 		pcb->pcb_fs = rp->r_fs;
399 		pcb->pcb_gs = rp->r_gs;
400 		pcb->pcb_rupdate = 1;
401 		lwptot(clwp)->t_post_sys = 1;
402 	}
403 	ASSERT(lwptot(clwp)->t_post_sys);
404 #endif
405 
406 	bcopy(lwp->lwp_regs, clwp->lwp_regs, sizeof (struct regs));
407 }
408 
409 /*
410  * This function is currently unused on x86.
411  */
412 /*ARGSUSED*/
413 void
414 lwp_freeregs(klwp_t *lwp, int isexec)
415 {}
416 
417 /*
418  * This function is currently unused on x86.
419  */
420 void
421 lwp_pcb_exit(void)
422 {}
423 
424 /*
425  * Lwp context ops for segment registers.
426  */
427 
428 /*
429  * Every time we come into the kernel (syscall, interrupt or trap
430  * but not fast-traps) we capture the current values of the user's
431  * segment registers into the lwp's reg structure. This includes
432  * lcall for i386 generic system call support since it is handled
433  * as a segment-not-present trap.
434  *
435  * Here we save the current values from the lwp regs into the pcb
436  * and set pcb->pcb_rupdate to 1 to tell the rest of the kernel
437  * that the pcb copy of the segment registers is the current one.
438  * This ensures the lwp's next trip to user land via update_sregs.
439  * Finally we set t_post_sys to ensure that no system call fast-path's
440  * its way out of the kernel via sysret.
441  *
442  * (This means that we need to have interrupts disabled when we test
443  * t->t_post_sys in the syscall handlers; if the test fails, we need
444  * to keep interrupts disabled until we return to userland so we can't
445  * be switched away.)
446  *
447  * As a result of all this, we don't really have to do a whole lot if
448  * the thread is just mucking about in the kernel, switching on and
449  * off the cpu for whatever reason it feels like. And yet we still
450  * preserve fast syscalls, cause if we -don't- get descheduled,
451  * we never come here either.
452  */
453 
454 #define	VALID_LWP_DESC(udp) ((udp)->usd_type == SDT_MEMRWA && \
455 	    (udp)->usd_p == 1 && (udp)->usd_dpl == SEL_UPL)
456 
457 /*ARGSUSED*/
458 void
459 lwp_segregs_save(klwp_t *lwp)
460 {
461 #if defined(__amd64)
462 	pcb_t *pcb = &lwp->lwp_pcb;
463 	struct regs *rp;
464 
465 	ASSERT(VALID_LWP_DESC(&pcb->pcb_fsdesc));
466 	ASSERT(VALID_LWP_DESC(&pcb->pcb_gsdesc));
467 
468 	if (pcb->pcb_rupdate == 0) {
469 		rp = lwptoregs(lwp);
470 
471 		/*
472 		 * If there's no update already pending, capture the current
473 		 * %ds/%es/%fs/%gs values from lwp's regs in case the user
474 		 * changed them; %fsbase and %gsbase are privileged so the
475 		 * kernel versions of these registers in pcb_fsbase and
476 		 * pcb_gsbase are always up-to-date.
477 		 */
478 		pcb->pcb_ds = rp->r_ds;
479 		pcb->pcb_es = rp->r_es;
480 		pcb->pcb_fs = rp->r_fs;
481 		pcb->pcb_gs = rp->r_gs;
482 		pcb->pcb_rupdate = 1;
483 		lwp->lwp_thread->t_post_sys = 1;
484 	}
485 #endif	/* __amd64 */
486 
487 #if !defined(__xpv)	/* XXPV not sure if we can re-read gdt? */
488 	ASSERT(bcmp(&CPU->cpu_gdt[GDT_LWPFS], &lwp->lwp_pcb.pcb_fsdesc,
489 	    sizeof (lwp->lwp_pcb.pcb_fsdesc)) == 0);
490 	ASSERT(bcmp(&CPU->cpu_gdt[GDT_LWPGS], &lwp->lwp_pcb.pcb_gsdesc,
491 	    sizeof (lwp->lwp_pcb.pcb_gsdesc)) == 0);
492 #endif
493 }
494 
495 #if defined(__amd64)
496 
497 /*
498  * Update the segment registers with new values from the pcb.
499  *
500  * We have to do this carefully, and in the following order,
501  * in case any of the selectors points at a bogus descriptor.
502  * If they do, we'll catch trap with on_trap and return 1.
503  * returns 0 on success.
504  *
505  * This is particularly tricky for %gs.
506  * This routine must be executed under a cli.
507  */
508 int
509 update_sregs(struct regs *rp,  klwp_t *lwp)
510 {
511 	pcb_t *pcb = &lwp->lwp_pcb;
512 	ulong_t	kgsbase;
513 	on_trap_data_t	otd;
514 	int rc = 0;
515 
516 	if (!on_trap(&otd, OT_SEGMENT_ACCESS)) {
517 
518 #if defined(__xpv)
519 		/*
520 		 * On the hyervisor this is easy. The hypercall below will
521 		 * swapgs and load %gs with the user selector. If the user
522 		 * selector is bad the hypervisor will catch the fault and
523 		 * load %gs with the null selector instead. Either way the
524 		 * kernel's gsbase is not damaged.
525 		 */
526 		kgsbase = (ulong_t)CPU;
527 		if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL,
528 		    pcb->pcb_gs) != 0) {
529 				no_trap();
530 				return (1);
531 		}
532 
533 		rp->r_gs = pcb->pcb_gs;
534 		ASSERT((cpu_t *)kgsbase == CPU);
535 
536 #else	/* __xpv */
537 
538 		/*
539 		 * A little more complicated running native.
540 		 */
541 		kgsbase = (ulong_t)CPU;
542 		__set_gs(pcb->pcb_gs);
543 
544 		/*
545 		 * If __set_gs fails it's because the new %gs is a bad %gs,
546 		 * we'll be taking a trap but with the original %gs and %gsbase
547 		 * undamaged (i.e. pointing at curcpu).
548 		 *
549 		 * We've just mucked up the kernel's gsbase.  Oops.  In
550 		 * particular we can't take any traps at all.  Make the newly
551 		 * computed gsbase be the hidden gs via __swapgs, and fix
552 		 * the kernel's gsbase back again. Later, when we return to
553 		 * userland we'll swapgs again restoring gsbase just loaded
554 		 * above.
555 		 */
556 		__swapgs();
557 		rp->r_gs = pcb->pcb_gs;
558 
559 		/*
560 		 * restore kernel's gsbase
561 		 */
562 		wrmsr(MSR_AMD_GSBASE, kgsbase);
563 
564 #endif	/* __xpv */
565 
566 		/*
567 		 * Only override the descriptor base address if
568 		 * r_gs == LWPGS_SEL or if r_gs == NULL. A note on
569 		 * NULL descriptors -- 32-bit programs take faults
570 		 * if they deference NULL descriptors; however,
571 		 * when 64-bit programs load them into %fs or %gs,
572 		 * they DONT fault -- only the base address remains
573 		 * whatever it was from the last load.   Urk.
574 		 *
575 		 * XXX - note that lwp_setprivate now sets %fs/%gs to the
576 		 * null selector for 64 bit processes. Whereas before
577 		 * %fs/%gs were set to LWP(FS|GS)_SEL regardless of
578 		 * the process's data model. For now we check for both
579 		 * values so that the kernel can also support the older
580 		 * libc. This should be ripped out at some point in the
581 		 * future.
582 		 */
583 		if (pcb->pcb_gs == LWPGS_SEL || pcb->pcb_gs == 0) {
584 #if defined(__xpv)
585 			if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER,
586 			    pcb->pcb_gsbase)) {
587 				no_trap();
588 				return (1);
589 			}
590 #else
591 			wrmsr(MSR_AMD_KGSBASE, pcb->pcb_gsbase);
592 #endif
593 		}
594 
595 		__set_ds(pcb->pcb_ds);
596 		rp->r_ds = pcb->pcb_ds;
597 
598 		__set_es(pcb->pcb_es);
599 		rp->r_es = pcb->pcb_es;
600 
601 		__set_fs(pcb->pcb_fs);
602 		rp->r_fs = pcb->pcb_fs;
603 
604 		/*
605 		 * Same as for %gs
606 		 */
607 		if (pcb->pcb_fs == LWPFS_SEL || pcb->pcb_fs == 0) {
608 #if defined(__xpv)
609 			if (HYPERVISOR_set_segment_base(SEGBASE_FS,
610 			    pcb->pcb_fsbase)) {
611 				no_trap();
612 				return (1);
613 			}
614 #else
615 			wrmsr(MSR_AMD_FSBASE, pcb->pcb_fsbase);
616 #endif
617 		}
618 
619 	} else {
620 		cli();
621 		rc = 1;
622 	}
623 	no_trap();
624 	return (rc);
625 }
626 
627 /*
628  * Make sure any stale selectors are cleared from the segment registers
629  * by putting KDS_SEL (the kernel's default %ds gdt selector) into them.
630  * This is necessary because the kernel itself does not use %es, %fs, nor
631  * %ds. (%cs and %ss are necessary, and are set up by the kernel - along with
632  * %gs - to point to the current cpu struct.) If we enter kmdb while in the
633  * kernel and resume with a stale ldt or brandz selector sitting there in a
634  * segment register, kmdb will #gp fault if the stale selector points to,
635  * for example, an ldt in the context of another process.
636  *
637  * WARNING: Intel and AMD chips behave differently when storing
638  * the null selector into %fs and %gs while in long mode. On AMD
639  * chips fsbase and gsbase are not cleared. But on Intel chips, storing
640  * a null selector into %fs or %gs has the side effect of clearing
641  * fsbase or gsbase. For that reason we use KDS_SEL, which has
642  * consistent behavor between AMD and Intel.
643  *
644  * Caller responsible for preventing cpu migration.
645  */
646 void
647 reset_sregs(void)
648 {
649 	ulong_t kgsbase = (ulong_t)CPU;
650 
651 	ASSERT(curthread->t_preempt != 0 || getpil() >= DISP_LEVEL);
652 
653 	cli();
654 	__set_gs(KGS_SEL);
655 
656 	/*
657 	 * restore kernel gsbase
658 	 */
659 #if defined(__xpv)
660 	xen_set_segment_base(SEGBASE_GS_KERNEL, kgsbase);
661 #else
662 	wrmsr(MSR_AMD_GSBASE, kgsbase);
663 #endif
664 
665 	sti();
666 
667 	__set_ds(KDS_SEL);
668 	__set_es(0 | SEL_KPL);	/* selector RPL not ring 0 on hypervisor */
669 	__set_fs(KFS_SEL);
670 }
671 
672 #endif	/* __amd64 */
673 
674 #ifdef _SYSCALL32_IMPL
675 
676 /*
677  * Make it impossible for a process to change its data model.
678  * We do this by toggling the present bits for the 32 and
679  * 64-bit user code descriptors. That way if a user lwp attempts
680  * to change its data model (by using the wrong code descriptor in
681  * %cs) it will fault immediately. This also allows us to simplify
682  * assertions and checks in the kernel.
683  */
684 
685 static void
686 gdt_ucode_model(model_t model)
687 {
688 	kpreempt_disable();
689 	if (model == DATAMODEL_NATIVE) {
690 		gdt_update_usegd(GDT_UCODE, &ucs_on);
691 		gdt_update_usegd(GDT_U32CODE, &ucs32_off);
692 	} else {
693 		gdt_update_usegd(GDT_U32CODE, &ucs32_on);
694 		gdt_update_usegd(GDT_UCODE, &ucs_off);
695 	}
696 	kpreempt_enable();
697 }
698 
699 #endif	/* _SYSCALL32_IMPL */
700 
701 /*
702  * Restore lwp private fs and gs segment descriptors
703  * on current cpu's GDT.
704  */
705 static void
706 lwp_segregs_restore(klwp_t *lwp)
707 {
708 	pcb_t *pcb = &lwp->lwp_pcb;
709 
710 	ASSERT(VALID_LWP_DESC(&pcb->pcb_fsdesc));
711 	ASSERT(VALID_LWP_DESC(&pcb->pcb_gsdesc));
712 
713 #ifdef	_SYSCALL32_IMPL
714 	gdt_ucode_model(DATAMODEL_NATIVE);
715 #endif
716 
717 	gdt_update_usegd(GDT_LWPFS, &pcb->pcb_fsdesc);
718 	gdt_update_usegd(GDT_LWPGS, &pcb->pcb_gsdesc);
719 
720 }
721 
722 #ifdef _SYSCALL32_IMPL
723 
724 static void
725 lwp_segregs_restore32(klwp_t *lwp)
726 {
727 	/*LINTED*/
728 	cpu_t *cpu = CPU;
729 	pcb_t *pcb = &lwp->lwp_pcb;
730 
731 	ASSERT(VALID_LWP_DESC(&lwp->lwp_pcb.pcb_fsdesc));
732 	ASSERT(VALID_LWP_DESC(&lwp->lwp_pcb.pcb_gsdesc));
733 
734 	gdt_ucode_model(DATAMODEL_ILP32);
735 	gdt_update_usegd(GDT_LWPFS, &pcb->pcb_fsdesc);
736 	gdt_update_usegd(GDT_LWPGS, &pcb->pcb_gsdesc);
737 }
738 
739 #endif	/* _SYSCALL32_IMPL */
740 
741 /*
742  * If this is a process in a branded zone, then we want it to use the brand
743  * syscall entry points instead of the standard Solaris entry points.  This
744  * routine must be called when a new lwp is created within a branded zone
745  * or when an existing lwp moves into a branded zone via a zone_enter()
746  * operation.
747  */
748 void
749 lwp_attach_brand_hdlrs(klwp_t *lwp)
750 {
751 	kthread_t *t = lwptot(lwp);
752 
753 	ASSERT(PROC_IS_BRANDED(lwptoproc(lwp)));
754 
755 	ASSERT(removectx(t, NULL, brand_interpositioning_disable,
756 	    brand_interpositioning_enable, NULL, NULL,
757 	    brand_interpositioning_disable, NULL) == 0);
758 	installctx(t, NULL, brand_interpositioning_disable,
759 	    brand_interpositioning_enable, NULL, NULL,
760 	    brand_interpositioning_disable, NULL);
761 
762 	if (t == curthread) {
763 		kpreempt_disable();
764 		brand_interpositioning_enable();
765 		kpreempt_enable();
766 	}
767 }
768 
769 /*
770  * If this is a process in a branded zone, then we want it to disable the
771  * brand syscall entry points.  This routine must be called when the last
772  * lwp in a process is exiting in proc_exit().
773  */
774 void
775 lwp_detach_brand_hdlrs(klwp_t *lwp)
776 {
777 	kthread_t *t = lwptot(lwp);
778 
779 	ASSERT(PROC_IS_BRANDED(lwptoproc(lwp)));
780 	if (t == curthread)
781 		kpreempt_disable();
782 
783 	/* Remove the original context handlers */
784 	VERIFY(removectx(t, NULL, brand_interpositioning_disable,
785 	    brand_interpositioning_enable, NULL, NULL,
786 	    brand_interpositioning_disable, NULL) != 0);
787 
788 	if (t == curthread) {
789 		/* Cleanup our MSR and IDT entries. */
790 		brand_interpositioning_disable();
791 		kpreempt_enable();
792 	}
793 }
794 
795 /*
796  * Add any lwp-associated context handlers to the lwp at the beginning
797  * of the lwp's useful life.
798  *
799  * All paths which create lwp's invoke lwp_create(); lwp_create()
800  * invokes lwp_stk_init() which initializes the stack, sets up
801  * lwp_regs, and invokes this routine.
802  *
803  * All paths which destroy lwp's invoke lwp_exit() to rip the lwp
804  * apart and put it on 'lwp_deathrow'; if the lwp is destroyed it
805  * ends up in thread_free() which invokes freectx(t, 0) before
806  * invoking lwp_stk_fini().  When the lwp is recycled from death
807  * row, lwp_stk_fini() is invoked, then thread_free(), and thus
808  * freectx(t, 0) as before.
809  *
810  * In the case of exec, the surviving lwp is thoroughly scrubbed
811  * clean; exec invokes freectx(t, 1) to destroy associated contexts.
812  * On the way back to the new image, it invokes setregs() which
813  * in turn invokes this routine.
814  */
815 void
816 lwp_installctx(klwp_t *lwp)
817 {
818 	kthread_t *t = lwptot(lwp);
819 	int thisthread = t == curthread;
820 #ifdef _SYSCALL32_IMPL
821 	void (*restop)(klwp_t *) = lwp_getdatamodel(lwp) == DATAMODEL_NATIVE ?
822 	    lwp_segregs_restore : lwp_segregs_restore32;
823 #else
824 	void (*restop)(klwp_t *) = lwp_segregs_restore;
825 #endif
826 
827 	/*
828 	 * Install the basic lwp context handlers on each lwp.
829 	 *
830 	 * On the amd64 kernel, the context handlers are responsible for
831 	 * virtualizing %ds, %es, %fs, and %gs to the lwp.  The register
832 	 * values are only ever changed via sys_rtt when the
833 	 * pcb->pcb_rupdate == 1.  Only sys_rtt gets to clear the bit.
834 	 *
835 	 * On the i386 kernel, the context handlers are responsible for
836 	 * virtualizing %gs/%fs to the lwp by updating the per-cpu GDTs
837 	 */
838 	ASSERT(removectx(t, lwp, lwp_segregs_save, restop,
839 	    NULL, NULL, NULL, NULL) == 0);
840 	if (thisthread)
841 		kpreempt_disable();
842 	installctx(t, lwp, lwp_segregs_save, restop,
843 	    NULL, NULL, NULL, NULL);
844 	if (thisthread) {
845 		/*
846 		 * Since we're the right thread, set the values in the GDT
847 		 */
848 		restop(lwp);
849 		kpreempt_enable();
850 	}
851 
852 	/*
853 	 * If we have sysenter/sysexit instructions enabled, we need
854 	 * to ensure that the hardware mechanism is kept up-to-date with the
855 	 * lwp's kernel stack pointer across context switches.
856 	 *
857 	 * sep_save zeros the sysenter stack pointer msr; sep_restore sets
858 	 * it to the lwp's kernel stack pointer (kstktop).
859 	 */
860 	if (is_x86_feature(x86_featureset, X86FSET_SEP)) {
861 #if defined(__amd64)
862 		caddr_t kstktop = (caddr_t)lwp->lwp_regs;
863 #elif defined(__i386)
864 		caddr_t kstktop = ((caddr_t)lwp->lwp_regs - MINFRAME) +
865 		    SA(sizeof (struct regs) + MINFRAME);
866 #endif
867 		ASSERT(removectx(t, kstktop,
868 		    sep_save, sep_restore, NULL, NULL, NULL, NULL) == 0);
869 
870 		if (thisthread)
871 			kpreempt_disable();
872 		installctx(t, kstktop,
873 		    sep_save, sep_restore, NULL, NULL, NULL, NULL);
874 		if (thisthread) {
875 			/*
876 			 * We're the right thread, so set the stack pointer
877 			 * for the first sysenter instruction to use
878 			 */
879 			sep_restore(kstktop);
880 			kpreempt_enable();
881 		}
882 	}
883 
884 	if (PROC_IS_BRANDED(ttoproc(t)))
885 		lwp_attach_brand_hdlrs(lwp);
886 }
887 
888 /*
889  * Clear registers on exec(2).
890  */
891 void
892 setregs(uarg_t *args)
893 {
894 	struct regs *rp;
895 	kthread_t *t = curthread;
896 	klwp_t *lwp = ttolwp(t);
897 	pcb_t *pcb = &lwp->lwp_pcb;
898 	greg_t sp;
899 
900 	/*
901 	 * Initialize user registers
902 	 */
903 	(void) save_syscall_args();	/* copy args from registers first */
904 	rp = lwptoregs(lwp);
905 	sp = rp->r_sp;
906 	bzero(rp, sizeof (*rp));
907 
908 	rp->r_ss = UDS_SEL;
909 	rp->r_sp = sp;
910 	rp->r_pc = args->entry;
911 	rp->r_ps = PSL_USER;
912 
913 #if defined(__amd64)
914 
915 	pcb->pcb_fs = pcb->pcb_gs = 0;
916 	pcb->pcb_fsbase = pcb->pcb_gsbase = 0;
917 
918 	if (ttoproc(t)->p_model == DATAMODEL_NATIVE) {
919 
920 		rp->r_cs = UCS_SEL;
921 
922 		/*
923 		 * Only allow 64-bit user code descriptor to be present.
924 		 */
925 		gdt_ucode_model(DATAMODEL_NATIVE);
926 
927 		/*
928 		 * Arrange that the virtualized %fs and %gs GDT descriptors
929 		 * have a well-defined initial state (present, ring 3
930 		 * and of type data).
931 		 */
932 		pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_udesc;
933 
934 		/*
935 		 * thrptr is either NULL or a value used by DTrace.
936 		 * 64-bit processes use %fs as their "thread" register.
937 		 */
938 		if (args->thrptr)
939 			(void) lwp_setprivate(lwp, _LWP_FSBASE, args->thrptr);
940 
941 	} else {
942 
943 		rp->r_cs = U32CS_SEL;
944 		rp->r_ds = rp->r_es = UDS_SEL;
945 
946 		/*
947 		 * only allow 32-bit user code selector to be present.
948 		 */
949 		gdt_ucode_model(DATAMODEL_ILP32);
950 
951 		pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_u32desc;
952 
953 		/*
954 		 * thrptr is either NULL or a value used by DTrace.
955 		 * 32-bit processes use %gs as their "thread" register.
956 		 */
957 		if (args->thrptr)
958 			(void) lwp_setprivate(lwp, _LWP_GSBASE, args->thrptr);
959 
960 	}
961 
962 	pcb->pcb_ds = rp->r_ds;
963 	pcb->pcb_es = rp->r_es;
964 	pcb->pcb_rupdate = 1;
965 
966 #elif defined(__i386)
967 
968 	rp->r_cs = UCS_SEL;
969 	rp->r_ds = rp->r_es = UDS_SEL;
970 
971 	/*
972 	 * Arrange that the virtualized %fs and %gs GDT descriptors
973 	 * have a well-defined initial state (present, ring 3
974 	 * and of type data).
975 	 */
976 	pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_udesc;
977 
978 	/*
979 	 * For %gs we need to reset LWP_GSBASE in pcb and the
980 	 * per-cpu GDT descriptor. thrptr is either NULL
981 	 * or a value used by DTrace.
982 	 */
983 	if (args->thrptr)
984 		(void) lwp_setprivate(lwp, _LWP_GSBASE, args->thrptr);
985 #endif
986 
987 	lwp->lwp_eosys = JUSTRETURN;
988 	t->t_post_sys = 1;
989 
990 	/*
991 	 * Here we initialize minimal fpu state.
992 	 * The rest is done at the first floating
993 	 * point instruction that a process executes.
994 	 */
995 	pcb->pcb_fpu.fpu_flags = 0;
996 
997 	/*
998 	 * Add the lwp context handlers that virtualize segment registers,
999 	 * and/or system call stacks etc.
1000 	 */
1001 	lwp_installctx(lwp);
1002 }
1003 
1004 user_desc_t *
1005 cpu_get_gdt(void)
1006 {
1007 	return (CPU->cpu_gdt);
1008 }
1009 
1010 
1011 #if !defined(lwp_getdatamodel)
1012 
1013 /*
1014  * Return the datamodel of the given lwp.
1015  */
1016 /*ARGSUSED*/
1017 model_t
1018 lwp_getdatamodel(klwp_t *lwp)
1019 {
1020 	return (lwp->lwp_procp->p_model);
1021 }
1022 
1023 #endif	/* !lwp_getdatamodel */
1024 
1025 #if !defined(get_udatamodel)
1026 
1027 model_t
1028 get_udatamodel(void)
1029 {
1030 	return (curproc->p_model);
1031 }
1032 
1033 #endif	/* !get_udatamodel */
1034