xref: /titanic_41/usr/src/uts/i86pc/ml/syscall_asm.s (revision 99ebb4ca412cb0a19d77a3899a87c055b9c30fa8)
1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
23 * Use is subject to license terms.
24 */
25
26/*	Copyright (c) 1990, 1991 UNIX System Laboratories, Inc.	*/
27/*	Copyright (c) 1984, 1986, 1987, 1988, 1989, 1990 AT&T	*/
28/*	  All Rights Reserved					*/
29
30/*	Copyright (c) 1987, 1988 Microsoft Corporation		*/
31/*	  All Rights Reserved					*/
32
33#pragma ident	"%Z%%M%	%I%	%E% SMI"
34
35#include <sys/asm_linkage.h>
36#include <sys/asm_misc.h>
37#include <sys/regset.h>
38#include <sys/psw.h>
39#include <sys/x86_archext.h>
40#include <sys/machbrand.h>
41
42#if defined(__lint)
43
44#include <sys/types.h>
45#include <sys/thread.h>
46#include <sys/systm.h>
47
48#else	/* __lint */
49
50#include <sys/segments.h>
51#include <sys/pcb.h>
52#include <sys/trap.h>
53#include <sys/ftrace.h>
54#include <sys/traptrace.h>
55#include <sys/clock.h>
56#include <sys/panic.h>
57#include "assym.h"
58
59#endif	/* __lint */
60
61/*
62 * We implement two flavours of system call entry points
63 *
64 * -	{int,lcall}/iret	(i386)
65 * -	sysenter/sysexit	(Pentium II and beyond)
66 *
67 * The basic pattern used in the handlers is to check to see if we can
68 * do fast (simple) version of the system call; if we can't we use various
69 * C routines that handle corner cases and debugging.
70 *
71 * To reduce the amount of assembler replication, yet keep the system call
72 * implementations vaguely comprehensible, the common code in the body
73 * of the handlers is broken up into a set of preprocessor definitions
74 * below.
75 */
76
77/*
78 * When we have SYSCALLTRACE defined, we sneak an extra
79 * predicate into a couple of tests.
80 */
81#if defined(SYSCALLTRACE)
82#define	ORL_SYSCALLTRACE(r32)	\
83	orl	syscalltrace, r32
84#else
85#define	ORL_SYSCALLTRACE(r32)
86#endif
87
88/*
89 * This check is false whenever we want to go fast i.e.
90 *
91 *	if (code >= NSYSCALL ||
92 *	    t->t_pre_sys || (t->t_proc_flag & TP_WATCHPT) != 0)
93 *		do full version
94 * #ifdef SYSCALLTRACE
95 *	if (syscalltrace)
96 *		do full version
97 * #endif
98 *
99 * Preconditions:
100 * -	t	curthread
101 * -	code	contains the syscall number
102 * Postconditions:
103 * -	%ecx and %edi are smashed
104 * -	condition code flag ZF is cleared if pre-sys is too complex
105 */
106#define	CHECK_PRESYS_NE(t, code)		\
107	movzbl	T_PRE_SYS(t), %edi;		\
108	movzwl	T_PROC_FLAG(t), %ecx;		\
109	andl	$TP_WATCHPT, %ecx;		\
110	orl	%ecx, %edi;			\
111	cmpl	$NSYSCALL, code;		\
112	setae	%cl;				\
113	movzbl	%cl, %ecx;			\
114	orl	%ecx, %edi;			\
115	ORL_SYSCALLTRACE(%edi)
116
117/*
118 * When the brand's callback is invoked, the stack will look like this:
119 *	   --------------------------------------
120 *         | 'scratch space'			|
121 *         | user's %ebx			|
122 *         | user's %gs selector		|
123 *         | kernel's %gs selector		|
124 *    |    | lwp brand data			|
125 *    |    | proc brand data			|
126 *    v    | user return address		|
127 *         | callback wrapper return addr 	|
128 *         --------------------------------------
129 *
130 * The lx brand (at least) uses each of these fields.
131 * If the brand code returns, we assume that we are meant to execute the
132 * normal system call path.
133 */
134#define	BRAND_CALLBACK(callback_id)					    \
135	subl	$4, %esp		/* save some scratch space	*/ ;\
136	pushl	%ebx			/* save %ebx to use for scratch	*/ ;\
137	pushl	%gs			/* save the user %gs		*/ ;\
138	movl	$KGS_SEL, %ebx						   ;\
139	pushl	%ebx			/* push kernel's %gs		*/ ;\
140	movw	%bx, %gs		/* switch to the kernel's %gs	*/ ;\
141	movl	%gs:CPU_THREAD, %ebx	/* load the thread pointer	*/ ;\
142	movl	T_LWP(%ebx), %ebx	/* load the lwp pointer		*/ ;\
143	pushl	LWP_BRAND(%ebx)		/* push the lwp's brand data	*/ ;\
144	movl	LWP_PROCP(%ebx), %ebx	/* load the proc pointer	*/ ;\
145	pushl	P_BRAND_DATA(%ebx)	/* push the proc's brand data	*/ ;\
146	movl	P_BRAND(%ebx), %ebx	/* load the brand pointer	*/ ;\
147	movl	B_MACHOPS(%ebx), %ebx	/* load the machops pointer	*/ ;\
148	movl	_CONST(_MUL(callback_id, CPTRSIZE))(%ebx), %ebx		   ;\
149	cmpl	$0, %ebx						   ;\
150	je	1f							   ;\
151	movl	%ebx, 20(%esp)		/* save callback to scratch	*/ ;\
152	movl	12(%esp), %ebx		/* grab the the user %gs	*/ ;\
153	movw	%bx, %gs		/* restore the user %gs		*/ ;\
154	movl	16(%esp), %ebx		/* restore %ebx			*/ ;\
155	pushl	24(%esp)		/* push the return address	*/ ;\
156	call	*24(%esp)		/* call callback		*/ ;\
157	addl	$4, %esp		/* get rid of ret addr		*/ ;\
1581:	movl	12(%esp), %ebx		/* grab the the user %gs	*/ ;\
159	movw	%bx, %gs		/* restore the user %gs		*/ ;\
160	movl	16(%esp), %ebx		/* restore user's %ebx		*/ ;\
161	addl	$24, %esp		/* restore stack ptr		*/
162
163#define	MSTATE_TRANSITION(from, to)		\
164	pushl	$to;				\
165	pushl	$from;				\
166	call	syscall_mstate;			\
167	addl	$0x8, %esp
168
169/*
170 * aka CPU_STATS_ADDQ(CPU, sys.syscall, 1)
171 * This must be called with interrupts or preemption disabled.
172 */
173#define	CPU_STATS_SYS_SYSCALL_INC			\
174	addl	$1, %gs:CPU_STATS_SYS_SYSCALL;		\
175	adcl	$0, %gs:CPU_STATS_SYS_SYSCALL+4;
176
177#if !defined(__lint)
178
179/*
180 * ASSERT(lwptoregs(lwp) == rp);
181 *
182 * this may seem obvious, but very odd things happen if this
183 * assertion is false
184 *
185 * Preconditions:
186 *	-none-
187 * Postconditions (if assertion is true):
188 *	%esi and %edi are smashed
189 */
190#if defined(DEBUG)
191
192__lwptoregs_msg:
193	.string	"%M%:%d lwptoregs(%p) [%p] != rp [%p]"
194
195#define	ASSERT_LWPTOREGS(t, rp)				\
196	movl	T_LWP(t), %esi;				\
197	movl	LWP_REGS(%esi), %edi;			\
198	cmpl	rp, %edi;				\
199	je	7f;					\
200	pushl	rp;					\
201	pushl	%edi;					\
202	pushl	%esi;					\
203	pushl	$__LINE__;				\
204	pushl	$__lwptoregs_msg;			\
205	call	panic;					\
2067:
207#else
208#define	ASSERT_LWPTOREGS(t, rp)
209#endif
210
211#endif	/* __lint */
212
213/*
214 * This is an assembler version of this fragment:
215 *
216 * lwp->lwp_state = LWP_SYS;
217 * lwp->lwp_ru.sysc++;
218 * lwp->lwp_eosys = NORMALRETURN;
219 * lwp->lwp_ap = argp;
220 *
221 * Preconditions:
222 *	-none-
223 * Postconditions:
224 *	-none-
225 */
226#define	SET_LWP(lwp, argp)				\
227	movb	$LWP_SYS, LWP_STATE(lwp);		\
228	addl	$1, LWP_RU_SYSC(lwp);			\
229	adcl	$0, LWP_RU_SYSC+4(lwp);			\
230	movb	$NORMALRETURN, LWP_EOSYS(lwp);		\
231	movl	argp, LWP_AP(lwp)
232
233/*
234 * Set up the thread, lwp, find the handler, and copy
235 * in the arguments from userland to the kernel stack.
236 *
237 * Preconditions:
238 * -	%eax contains the syscall number
239 * Postconditions:
240 * -	%eax contains a pointer to the sysent structure
241 * -	%ecx is zeroed
242 * -	%esi, %edi are smashed
243 * -	%esp is SYS_DROPped ready for the syscall
244 */
245#define	SIMPLE_SYSCALL_PRESYS(t, faultlabel)		\
246	movl	T_LWP(t), %esi;				\
247	movw	%ax, T_SYSNUM(t);			\
248	subl	$SYS_DROP, %esp;			\
249	shll	$SYSENT_SIZE_SHIFT, %eax;			\
250	SET_LWP(%esi, %esp);				\
251	leal	sysent(%eax), %eax;			\
252	movzbl	SY_NARG(%eax), %ecx;			\
253	testl	%ecx, %ecx;				\
254	jz	4f;					\
255	movl	%esp, %edi;				\
256	movl	SYS_DROP + REGOFF_UESP(%esp), %esi;	\
257	movl	$faultlabel, T_LOFAULT(t);		\
258	addl	$4, %esi;				\
259	rep;						\
260	  smovl;					\
261	movl	%ecx, T_LOFAULT(t);			\
2624:
263
264/*
265 * Check to see if a simple return is possible i.e.
266 *
267 *	if ((t->t_post_sys_ast | syscalltrace) != 0)
268 *		do full version;
269 *
270 * Preconditions:
271 * -	t is curthread
272 * Postconditions:
273 * -	condition code NE is set if post-sys is too complex
274 * -	rtmp is zeroed if it isn't (we rely on this!)
275 */
276#define	CHECK_POSTSYS_NE(t, rtmp)			\
277	xorl	rtmp, rtmp;				\
278	ORL_SYSCALLTRACE(rtmp);				\
279	orl	T_POST_SYS_AST(t), rtmp;		\
280	cmpl	$0, rtmp
281
282/*
283 * Fix up the lwp, thread, and eflags for a successful return
284 *
285 * Preconditions:
286 * -	zwreg contains zero
287 * Postconditions:
288 * -	%esp has been unSYS_DROPped
289 * -	%esi is smashed (points to lwp)
290 */
291#define	SIMPLE_SYSCALL_POSTSYS(t, zwreg)		\
292	movl	T_LWP(t), %esi;				\
293	addl	$SYS_DROP, %esp;			\
294	movw	zwreg, T_SYSNUM(t);			\
295	movb	$LWP_USER, LWP_STATE(%esi);		\
296	andb	$_CONST(0xffff - PS_C), REGOFF_EFL(%esp)
297
298/*
299 * System call handler.  This is the destination of both the call
300 * gate (lcall 0x27) _and_ the interrupt gate (int 0x91). For our purposes,
301 * there are two significant differences between an interrupt gate and a call
302 * gate:
303 *
304 * 1) An interrupt gate runs the handler with interrupts disabled, whereas a
305 * call gate runs the handler with whatever EFLAGS settings were in effect at
306 * the time of the call.
307 *
308 * 2) An interrupt gate pushes the contents of the EFLAGS register at the time
309 * of the interrupt onto the stack, whereas a call gate does not.
310 *
311 * Because we use the following code sequence to handle system calls made from
312 * _both_ a call gate _and_ an interrupt gate, these two differences must be
313 * respected. In regards to number 1) above, the handler must ensure that a sane
314 * EFLAGS snapshot is stored on the stack so that when the kernel returns back
315 * to the user via iret (which returns to user with the EFLAGS value saved on
316 * the stack), interrupts are re-enabled.
317 *
318 * In regards to number 2) above, the handler must always put a current snapshot
319 * of EFLAGS onto the stack in the appropriate place. If we came in via an
320 * interrupt gate, we will be clobbering the EFLAGS value that was pushed by
321 * the interrupt gate. This is OK, as the only bit that was changed by the
322 * hardware was the IE (interrupt enable) bit, which for an interrupt gate is
323 * now off. If we were to do nothing, the stack would contain an EFLAGS with
324 * IE off, resulting in us eventually returning back to the user with interrupts
325 * disabled. The solution is to turn on the IE bit in the EFLAGS value saved on
326 * the stack.
327 *
328 * Another subtlety which deserves mention is the difference between the two
329 * descriptors. The call gate descriptor is set to instruct the hardware to copy
330 * one parameter from the user stack to the kernel stack, whereas the interrupt
331 * gate descriptor doesn't use the parameter passing mechanism at all. The
332 * kernel doesn't actually use the parameter that is copied by the hardware; the
333 * only reason it does this is so that there is a space on the stack large
334 * enough to hold an EFLAGS register value, which happens to be in the correct
335 * place for use by iret when we go back to userland. How convenient.
336 *
337 * Stack frame description in syscall() and callees.
338 *
339 * |------------|
340 * | regs	| +(8*4)+4	registers
341 * |------------|
342 * | 8 args	| <- %esp	MAXSYSARGS (currently 8) arguments
343 * |------------|
344 *
345 */
346#define	SYS_DROP	_CONST(_MUL(MAXSYSARGS, 4))
347
348#if defined(__lint)
349
350/*ARGSUSED*/
351void
352sys_call()
353{}
354
355void
356_allsyscalls()
357{}
358
359size_t _allsyscalls_size;
360
361#else	/* __lint */
362
363	ENTRY_NP2(brand_sys_call, _allsyscalls)
364	BRAND_CALLBACK(BRAND_CB_SYSCALL)
365
366	ALTENTRY(sys_call)
367	/ on entry	eax = system call number
368
369	/ set up the stack to look as in reg.h
370	subl    $8, %esp        / pad the stack with ERRCODE and TRAPNO
371
372	SYSCALL_PUSH
373
374#ifdef TRAPTRACE
375	TRACE_PTR(%edi, %ebx, %ebx, %ecx, $TT_SYSCALL) / Uses labels "8" and "9"
376	TRACE_REGS(%edi, %esp, %ebx, %ecx)	/ Uses label "9"
377	pushl	%eax
378	TRACE_STAMP(%edi)		/ Clobbers %eax, %edx, uses "9"
379	popl	%eax
380	movl	%eax, TTR_SYSNUM(%edi)
381#endif
382
383_watch_do_syscall:
384	movl	%esp, %ebp
385
386	pushl	%eax				/ preserve across mstate call
387	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
388	popl	%eax
389
390	movl	%gs:CPU_THREAD, %ebx
391
392	/ Interrupts are enabled here, so we must make sure this thread doesn't
393	/ migrate off the CPU while it updates the CPU stats.
394	addb	$1, T_PREEMPT(%ebx)
395	CPU_STATS_SYS_SYSCALL_INC
396	subb	$1, T_PREEMPT(%ebx)
397
398	/ Set EFLAGS to standard kernel settings.
399	ENABLE_INTR_FLAGS
400
401	ASSERT_LWPTOREGS(%ebx, %esp)
402
403	CHECK_PRESYS_NE(%ebx, %eax)
404	jne	_full_syscall_presys
405	SIMPLE_SYSCALL_PRESYS(%ebx, _syscall_fault)
406
407_syslcall_call:
408	call	*SY_CALLC(%eax)
409
410_syslcall_done:
411	CHECK_POSTSYS_NE(%ebx, %ecx)
412	jne	_full_syscall_postsys
413	SIMPLE_SYSCALL_POSTSYS(%ebx, %cx)
414	movl	%eax, REGOFF_EAX(%esp)
415	movl	%edx, REGOFF_EDX(%esp)
416
417	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
418
419	/
420	/ get back via iret
421	/
422	cli
423	jmp	set_user_regs
424
425_full_syscall_presys:
426	movl	T_LWP(%ebx), %esi
427	subl	$SYS_DROP, %esp
428	movb	$LWP_SYS, LWP_STATE(%esi)
429	pushl	%esp
430	pushl	%ebx
431	call	syscall_entry
432	addl	$8, %esp
433	jmp	_syslcall_call
434
435_full_syscall_postsys:
436	addl	$SYS_DROP, %esp
437	pushl	%edx
438	pushl	%eax
439	pushl	%ebx
440	call	syscall_exit
441	addl	$12, %esp
442	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
443	jmp	sys_rtt_syscall
444
445_syscall_fault:
446	push	$0xe			/ EFAULT
447	call	set_errno
448	addl	$4, %esp
449	xorl	%eax, %eax		/ fake syscall_err()
450	xorl	%edx, %edx
451	jmp	_syslcall_done
452	SET_SIZE(sys_call)
453	SET_SIZE(brand_sys_call)
454
455#endif	/* __lint */
456
457/*
458 * System call handler via the sysenter instruction
459 *
460 * Here's how syscall entry usually works (see sys_call for details).
461 *
462 * There, the caller (lcall or int) in userland has arranged that:
463 *
464 * -	%eax contains the syscall number
465 * -	the user stack contains the args to the syscall
466 *
467 * Normally the lcall instruction into the call gate causes the processor
468 * to push %ss, %esp, <top-of-stack>, %cs, %eip onto the kernel stack.
469 * The sys_call handler then leaves space for r_trapno and r_err, and
470 * pusha's {%eax, %ecx, %edx, %ebx, %esp, %ebp, %esi, %edi}, followed
471 * by %ds, %es, %fs and %gs to capture a 'struct regs' on the stack.
472 * Then the kernel sets %ds, %es and %gs to kernel selectors, and finally
473 * extracts %efl and puts it into r_efl (which happens to live at the offset
474 * that <top-of-stack> was copied into). Note that the value in r_efl has
475 * the IF (interrupt enable) flag turned on. (The int instruction into the
476 * interrupt gate does essentially the same thing, only instead of
477 * <top-of-stack> we get eflags - see comment above.)
478 *
479 * In the sysenter case, things are a lot more primitive.
480 *
481 * The caller in userland has arranged that:
482 *
483 * -	%eax contains the syscall number
484 * -	%ecx contains the user %esp
485 * -	%edx contains the return %eip
486 * -	the user stack contains the args to the syscall
487 *
488 * e.g.
489 *	<args on the stack>
490 *	mov	$SYS_callnum, %eax
491 *	mov	$1f, %edx	/ return %eip
492 *	mov	%esp, %ecx	/ return %esp
493 *	sysenter
494 * 1:
495 *
496 * Hardware and (privileged) initialization code have arranged that by
497 * the time the sysenter instructions completes:
498 *
499 * - %eip is pointing to sys_sysenter (below).
500 * - %cs and %ss are set to kernel text and stack (data) selectors.
501 * - %esp is pointing at the lwp's stack
502 * - Interrupts have been disabled.
503 *
504 * The task for the sysenter handler is:
505 *
506 * -	recreate the same regs structure on the stack and the same
507 *	kernel state as if we'd come in on an lcall
508 * -	do the normal work of a syscall
509 * -	execute the system call epilogue, use sysexit to return to userland.
510 *
511 * Note that we are unable to return both "rvals" to userland with this
512 * call, as %edx is used by the sysexit instruction.
513 *
514 * One final complication in this routine is its interaction with
515 * single-stepping in a debugger.  For most of the system call mechanisms,
516 * the CPU automatically clears the single-step flag before we enter the
517 * kernel.  The sysenter mechanism does not clear the flag, so a user
518 * single-stepping through a libc routine may suddenly find him/herself
519 * single-stepping through the kernel.  To detect this, kmdb compares the
520 * trap %pc to the [brand_]sys_enter addresses on each single-step trap.
521 * If it finds that we have single-stepped to a sysenter entry point, it
522 * explicitly clears the flag and executes the sys_sysenter routine.
523 *
524 * One final complication in this final complication is the fact that we
525 * have two different entry points for sysenter: brand_sys_sysenter and
526 * sys_sysenter.  If we enter at brand_sys_sysenter and start single-stepping
527 * through the kernel with kmdb, we will eventually hit the instruction at
528 * sys_sysenter.  kmdb cannot distinguish between that valid single-step
529 * and the undesirable one mentioned above.  To avoid this situation, we
530 * simply add a jump over the instruction at sys_sysenter to make it
531 * impossible to single-step to it.
532 */
533#if defined(__lint)
534
535void
536sys_sysenter()
537{}
538
539#else	/* __lint */
540
541	ENTRY_NP(brand_sys_sysenter)
542	pushl	%edx
543	BRAND_CALLBACK(BRAND_CB_SYSENTER)
544	popl	%edx
545	/*
546	 * Jump over sys_sysenter to allow single-stepping as described
547	 * above.
548	 */
549	ja	1f
550
551	ALTENTRY(sys_sysenter)
552	nop
5531:
554	/
555	/ do what the call gate would've done to the stack ..
556	/
557	pushl	$UDS_SEL	/ (really %ss, but it's the same ..)
558	pushl	%ecx		/ userland makes this a copy of %esp
559	pushfl
560	orl	$PS_IE, (%esp)	/ turn interrupts on when we return to user
561	pushl	$UCS_SEL
562	pushl	%edx		/ userland makes this a copy of %eip
563	/
564	/ done.  finish building the stack frame
565	/
566	subl	$8, %esp	/ leave space for ERR and TRAPNO
567
568	SYSENTER_PUSH
569
570#ifdef TRAPTRACE
571	TRACE_PTR(%edi, %ebx, %ebx, %ecx, $TT_SYSENTER)	/ uses labels 8 and 9
572	TRACE_REGS(%edi, %esp, %ebx, %ecx)		/ uses label 9
573	pushl	%eax
574	TRACE_STAMP(%edi)		/ clobbers %eax, %edx, uses label 9
575	popl	%eax
576	movl	%eax, TTR_SYSNUM(%edi)
577#endif
578	movl	%esp, %ebp
579
580	CPU_STATS_SYS_SYSCALL_INC
581
582	ENABLE_INTR_FLAGS
583
584	pushl	%eax				/ preserve across mstate call
585	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
586	popl	%eax
587
588	movl	%gs:CPU_THREAD, %ebx
589
590	ASSERT_LWPTOREGS(%ebx, %esp)
591
592	CHECK_PRESYS_NE(%ebx, %eax)
593	jne	_full_syscall_presys
594	SIMPLE_SYSCALL_PRESYS(%ebx, _syscall_fault)
595
596_sysenter_call:
597	call	*SY_CALLC(%eax)
598
599_sysenter_done:
600	CHECK_POSTSYS_NE(%ebx, %ecx)
601	jne	_full_syscall_postsys
602	SIMPLE_SYSCALL_POSTSYS(%ebx, %cx)
603	/
604	/ sysexit uses %edx to restore %eip, so we can't use it
605	/ to return a value, sigh.
606	/
607	movl	%eax, REGOFF_EAX(%esp)
608	/ movl	%edx, REGOFF_EDX(%esp)
609
610	/ Interrupts will be turned on by the 'sti' executed just before
611	/ sysexit. The following ensures that restoring the user's EFLAGS
612	/ doesn't enable interrupts too soon.
613	andl	$_BITNOT(PS_IE), REGOFF_EFL(%esp)
614
615	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
616
617	cli
618
619	SYSCALL_POP
620
621	popl	%edx			/ sysexit: %edx -> %eip
622	addl	$4, %esp		/ get CS off the stack
623	popfl				/ EFL
624	popl	%ecx			/ sysexit: %ecx -> %esp
625	sti
626	sysexit
627	SET_SIZE(sys_sysenter)
628	SET_SIZE(brand_sys_sysenter)
629
630#endif	/* __lint */
631
632#if defined(__lint)
633/*
634 * System call via an int80.  This entry point is only used by the Linux
635 * application environment.  Unlike the sysenter path, there is no default
636 * action to take if no callback is registered for this process.
637 */
638void
639sys_int80()
640{}
641
642#else	/* __lint */
643
644	ENTRY_NP(brand_sys_int80)
645	BRAND_CALLBACK(BRAND_CB_INT80)
646
647	ALTENTRY(sys_int80)
648	/*
649	 * We hit an int80, but this process isn't of a brand with an int80
650	 * handler.  Bad process!  Make it look as if the INT failed.
651	 * Modify %eip to point before the INT, push the expected error
652	 * code and fake a GP fault.
653	 *
654	 */
655	subl	$2, (%esp)	/* int insn 2-bytes */
656	pushl	$_CONST(_MUL(T_INT80, GATE_DESC_SIZE) + 2)
657	jmp	gptrap			/ GP fault
658	SET_SIZE(sys_int80)
659	SET_SIZE(brand_sys_int80)
660
661/*
662 * Declare a uintptr_t which covers the entire pc range of syscall
663 * handlers for the stack walkers that need this.
664 */
665	.align	CPTRSIZE
666	.globl	_allsyscalls_size
667	.type	_allsyscalls_size, @object
668_allsyscalls_size:
669	.NWORD	. - _allsyscalls
670	SET_SIZE(_allsyscalls_size)
671
672#endif	/* __lint */
673
674/*
675 * These are the thread context handlers for lwps using sysenter/sysexit.
676 */
677
678#if defined(__lint)
679
680/*ARGSUSED*/
681void
682sep_save(void *ksp)
683{}
684
685/*ARGSUSED*/
686void
687sep_restore(void *ksp)
688{}
689
690#else	/* __lint */
691
692	/*
693	 * setting this value to zero as we switch away causes the
694	 * stack-pointer-on-sysenter to be NULL, ensuring that we
695	 * don't silently corrupt another (preempted) thread stack
696	 * when running an lwp that (somehow) didn't get sep_restore'd
697	 */
698	ENTRY_NP(sep_save)
699	xorl	%edx, %edx
700	xorl	%eax, %eax
701	movl	$MSR_INTC_SEP_ESP, %ecx
702	wrmsr
703	ret
704	SET_SIZE(sep_save)
705
706	/*
707	 * Update the kernel stack pointer as we resume onto this cpu.
708	 */
709	ENTRY_NP(sep_restore)
710	movl	4(%esp), %eax			/* per-lwp kernel sp */
711	xorl	%edx, %edx
712	movl	$MSR_INTC_SEP_ESP, %ecx
713	wrmsr
714	ret
715	SET_SIZE(sep_restore)
716
717#endif	/* __lint */
718
719/*
720 * Call syscall().  Called from trap() on watchpoint at lcall 0,7
721 */
722
723#if defined(__lint)
724
725void
726watch_syscall(void)
727{}
728
729#else	/* __lint */
730
731	ENTRY_NP(watch_syscall)
732	movl	%gs:CPU_THREAD, %ebx
733	movl	T_STACK(%ebx), %esp		/ switch to the thread stack
734	movl	REGOFF_EAX(%esp), %eax		/ recover original syscall#
735	jmp	_watch_do_syscall
736	SET_SIZE(watch_syscall)
737
738#endif	/* __lint */
739