xref: /titanic_41/usr/src/uts/i86pc/ml/syscall_asm_amd64.s (revision 3441f6a1af86b9b2f883f3323bf02c9dd0f7a94d)
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#pragma ident	"%Z%%M%	%I%	%E% SMI"
27
28#include <sys/asm_linkage.h>
29#include <sys/asm_misc.h>
30#include <sys/regset.h>
31#include <sys/privregs.h>
32#include <sys/psw.h>
33#include <sys/machbrand.h>
34
35#if defined(__lint)
36
37#include <sys/types.h>
38#include <sys/thread.h>
39#include <sys/systm.h>
40
41#else	/* __lint */
42
43#include <sys/segments.h>
44#include <sys/pcb.h>
45#include <sys/trap.h>
46#include <sys/ftrace.h>
47#include <sys/traptrace.h>
48#include <sys/clock.h>
49#include <sys/model.h>
50#include <sys/panic.h>
51#include "assym.h"
52
53#endif	/* __lint */
54
55/*
56 * We implement five flavours of system call entry points
57 *
58 * -	syscall/sysretq		(amd64 generic)
59 * -	syscall/sysretl		(i386 plus SYSC bit)
60 * -	sysenter/sysexit	(i386 plus SEP bit)
61 * -	int/iret		(i386 generic)
62 * -	lcall/iret		(i386 generic)
63 *
64 * The current libc included in Solaris uses int/iret as the base unoptimized
65 * kernel entry method. Older libc implementations and legacy binaries may use
66 * the lcall call gate, so it must continue to be supported.
67 *
68 * System calls that use an lcall call gate are processed in trap() via a
69 * segment-not-present trap, i.e. lcalls are extremely slow(!).
70 *
71 * The basic pattern used in the 32-bit SYSC handler at this point in time is
72 * to have the bare minimum of assembler, and get to the C handlers as
73 * quickly as possible.
74 *
75 * The 64-bit handler is much closer to the sparcv9 handler; that's
76 * because of passing arguments in registers.  The 32-bit world still
77 * passes arguments on the stack -- that makes that handler substantially
78 * more complex.
79 *
80 * The two handlers share a few code fragments which are broken
81 * out into preprocessor macros below.
82 *
83 * XX64	come back and speed all this up later.  The 32-bit stuff looks
84 * especially easy to speed up the argument copying part ..
85 *
86 *
87 * Notes about segment register usage (c.f. the 32-bit kernel)
88 *
89 * In the 32-bit kernel, segment registers are dutifully saved and
90 * restored on all mode transitions because the kernel uses them directly.
91 * When the processor is running in 64-bit mode, segment registers are
92 * largely ignored.
93 *
94 * %cs and %ss
95 *	controlled by the hardware mechanisms that make mode transitions
96 *
97 * The remaining segment registers have to either be pointing at a valid
98 * descriptor i.e. with the 'present' bit set, or they can NULL descriptors
99 *
100 * %ds and %es
101 *	always ignored
102 *
103 * %fs and %gs
104 *	fsbase and gsbase are used to control the place they really point at.
105 *	The kernel only depends on %gs, and controls its own gsbase via swapgs
106 *
107 * Note that loading segment registers is still costly because the GDT
108 * lookup still happens (this is because the hardware can't know that we're
109 * not setting up these segment registers for a 32-bit program).  Thus we
110 * avoid doing this in the syscall path, and defer them to lwp context switch
111 * handlers, so the register values remain virtualized to the lwp.
112 */
113
114#if defined(SYSCALLTRACE)
115#define	ORL_SYSCALLTRACE(r32)		\
116	orl	syscalltrace(%rip), r32
117#else
118#define	ORL_SYSCALLTRACE(r32)
119#endif
120
121/*
122 * In the 32-bit kernel, we do absolutely nothing before getting into the
123 * brand callback checks.  In 64-bit land, we do swapgs and then come here.
124 * We assume that the %rsp- and %r15-stashing fields in the CPU structure
125 * are still unused.
126 *
127 * When the callback is invoked, we will be on the user's %gs and
128 * the stack will look like this:
129 *
130 * stack:  --------------------------------------
131 *         | callback pointer			|
132 *    |    | user stack pointer			|
133 *    |    | lwp brand data			|
134 *    |    | proc brand data			|
135 *    v    | userland return address		|
136 *         | callback wrapper return addr	|
137 *         --------------------------------------
138 *
139 */
140#define	BRAND_CALLBACK(callback_id)					    \
141	movq	%rsp, %gs:CPU_RTMP_RSP	/* save the stack pointer	*/ ;\
142	movq	%r15, %gs:CPU_RTMP_R15	/* save %r15			*/ ;\
143	movq	%gs:CPU_THREAD, %r15	/* load the thread pointer	*/ ;\
144	movq	T_STACK(%r15), %rsp	/* switch to the kernel stack	*/ ;\
145	subq	$16, %rsp		/* save space for two pointers	*/ ;\
146	pushq	%r14			/* save %r14			*/ ;\
147	movq	%gs:CPU_RTMP_RSP, %r14					   ;\
148	movq	%r14, 8(%rsp)		/* stash the user stack pointer	*/ ;\
149	popq	%r14			/* restore %r14			*/ ;\
150	movq	T_LWP(%r15), %r15	/* load the lwp pointer		*/ ;\
151	pushq	LWP_BRAND(%r15)		/* push the lwp's brand data	*/ ;\
152	movq	LWP_PROCP(%r15), %r15	/* load the proc pointer	*/ ;\
153	pushq	P_BRAND_DATA(%r15)	/* push the proc's brand data	*/ ;\
154	movq	P_BRAND(%r15), %r15	/* load the brand pointer	*/ ;\
155	movq	B_MACHOPS(%r15), %r15	/* load the machops pointer	*/ ;\
156	movq	_CONST(_MUL(callback_id, CPTRSIZE))(%r15), %r15		   ;\
157	cmpq	$0, %r15						   ;\
158	je	1f							   ;\
159	movq	%r15, 24(%rsp)		/* save the callback pointer	*/ ;\
160	movq	%gs:CPU_RTMP_RSP, %r15	/* grab the user stack pointer	*/ ;\
161	pushq	(%r15)			/* push the return address	*/ ;\
162	movq	%gs:CPU_RTMP_R15, %r15	/* restore %r15			*/ ;\
163	swapgs								   ;\
164	call	*32(%rsp)		/* call callback		*/ ;\
165	swapgs								   ;\
1661:	movq	%gs:CPU_RTMP_R15, %r15	/* restore %r15			*/ ;\
167	movq	%gs:CPU_RTMP_RSP, %rsp	/* restore the stack pointer	*/
168
169#define	MSTATE_TRANSITION(from, to)		\
170	movl	$from, %edi;			\
171	movl	$to, %esi;			\
172	call	syscall_mstate
173
174/*
175 * Check to see if a simple (direct) return is possible i.e.
176 *
177 *	if (t->t_post_sys_ast | syscalltrace |
178 *	    lwp->lwp_pcb.pcb_rupdate == 1)
179 *		do full version	;
180 *
181 * Preconditions:
182 * -	t is curthread
183 * Postconditions:
184 * -	condition code NE is set if post-sys is too complex
185 * -	rtmp is zeroed if it isn't (we rely on this!)
186 * -	ltmp is smashed
187 */
188#define	CHECK_POSTSYS_NE(t, ltmp, rtmp)			\
189	movq	T_LWP(t), ltmp;				\
190	movzbl	PCB_RUPDATE(ltmp), rtmp;		\
191	ORL_SYSCALLTRACE(rtmp);				\
192	orl	T_POST_SYS_AST(t), rtmp;		\
193	cmpl	$0, rtmp
194
195/*
196 * Fix up the lwp, thread, and eflags for a successful return
197 *
198 * Preconditions:
199 * -	zwreg contains zero
200 */
201#define	SIMPLE_SYSCALL_POSTSYS(t, lwp, zwreg)		\
202	movb	$LWP_USER, LWP_STATE(lwp);		\
203	movw	zwreg, T_SYSNUM(t);			\
204	andb	$_CONST(0xffff - PS_C), REGOFF_RFL(%rsp)
205
206/*
207 * ASSERT(lwptoregs(lwp) == rp);
208 *
209 * This may seem obvious, but very odd things happen if this
210 * assertion is false
211 *
212 * Preconditions:
213 *	(%rsp is ready for normal call sequence)
214 * Postconditions (if assertion is true):
215 *	%r11 is smashed
216 *
217 * ASSERT(rp->r_cs == descnum)
218 *
219 * The code selector is written into the regs structure when the
220 * lwp stack is created.  We use this ASSERT to validate that
221 * the regs structure really matches how we came in.
222 *
223 * Preconditions:
224 *	(%rsp is ready for normal call sequence)
225 * Postconditions (if assertion is true):
226 *	-none-
227 *
228 * ASSERT(lwp->lwp_pcb.pcb_rupdate == 0);
229 *
230 * If this is false, it meant that we returned to userland without
231 * updating the segment registers as we were supposed to.
232 *
233 * Note that we must ensure no interrupts or other traps intervene
234 * between entering privileged mode and performing the assertion,
235 * otherwise we may perform a context switch on the thread, which
236 * will end up setting pcb_rupdate to 1 again.
237 */
238#if defined(DEBUG)
239
240#if !defined(__lint)
241
242__lwptoregs_msg:
243	.string	"syscall_asm_amd64.s:%d lwptoregs(%p) [%p] != rp [%p]"
244
245__codesel_msg:
246	.string	"syscall_asm_amd64.s:%d rp->r_cs [%ld] != %ld"
247
248__no_rupdate_msg:
249	.string	"syscall_asm_amd64.s:%d lwp %p, pcb_rupdate != 0"
250
251#endif	/* !__lint */
252
253#define	ASSERT_LWPTOREGS(lwp, rp)			\
254	movq	LWP_REGS(lwp), %r11;			\
255	cmpq	rp, %r11;				\
256	je	7f;					\
257	leaq	__lwptoregs_msg(%rip), %rdi;		\
258	movl	$__LINE__, %esi;			\
259	movq	lwp, %rdx;				\
260	movq	%r11, %rcx;				\
261	movq	rp, %r8;				\
262	xorl	%eax, %eax;				\
263	call	panic;					\
2647:
265
266#define	ASSERT_NO_RUPDATE_PENDING(lwp)			\
267	testb	$0x1, PCB_RUPDATE(lwp);			\
268	je	8f;					\
269	movq	lwp, %rdx;				\
270	leaq	__no_rupdate_msg(%rip), %rdi;		\
271	movl	$__LINE__, %esi;			\
272	xorl	%eax, %eax;				\
273	call	panic;					\
2748:
275
276#else
277#define	ASSERT_LWPTOREGS(lwp, rp)
278#define	ASSERT_NO_RUPDATE_PENDING(lwp)
279#endif
280
281/*
282 * Do the traptrace thing and restore any registers we used
283 * in situ.  Assumes that %rsp is pointing at the base of
284 * the struct regs, obviously ..
285 */
286#ifdef TRAPTRACE
287#define	SYSCALL_TRAPTRACE(ttype)				\
288	TRACE_PTR(%rdi, %rbx, %ebx, %rcx, ttype);		\
289	TRACE_REGS(%rdi, %rsp, %rbx, %rcx);			\
290	TRACE_STAMP(%rdi);	/* rdtsc clobbers %eax, %edx */	\
291	movq	REGOFF_RAX(%rsp), %rax;				\
292	movq	REGOFF_RBX(%rsp), %rbx;				\
293	movq	REGOFF_RCX(%rsp), %rcx;				\
294	movq	REGOFF_RDX(%rsp), %rdx;				\
295	movl	%eax, TTR_SYSNUM(%rdi);				\
296	movq	REGOFF_RDI(%rsp), %rdi
297
298#define	SYSCALL_TRAPTRACE32(ttype)				\
299	SYSCALL_TRAPTRACE(ttype);				\
300	/* paranoia: clean the top 32-bits of the registers */	\
301	orl	%eax, %eax;					\
302	orl	%ebx, %ebx;					\
303	orl	%ecx, %ecx;					\
304	orl	%edx, %edx;					\
305	orl	%edi, %edi
306#else	/* TRAPTRACE */
307#define	SYSCALL_TRAPTRACE(ttype)
308#define	SYSCALL_TRAPTRACE32(ttype)
309#endif	/* TRAPTRACE */
310
311/*
312 * The 64-bit libc syscall wrapper does this:
313 *
314 * fn(<args>)
315 * {
316 *	movq	%rcx, %r10	-- because syscall smashes %rcx
317 *	movl	$CODE, %eax
318 *	syscall
319 *	<error processing>
320 * }
321 *
322 * Thus when we come into the kernel:
323 *
324 *	%rdi, %rsi, %rdx, %r10, %r8, %r9 contain first six args
325 *	%rax is the syscall number
326 *	%r12-%r15 contain caller state
327 *
328 * The syscall instruction arranges that:
329 *
330 *	%rcx contains the return %rip
331 *	%r11d contains bottom 32-bits of %rflags
332 *	%rflags is masked (as determined by the SFMASK msr)
333 *	%cs is set to UCS_SEL (as determined by the STAR msr)
334 *	%ss is set to UDS_SEL (as determined by the STAR msr)
335 *	%rip is set to sys_syscall (as determined by the LSTAR msr)
336 *
337 * Or in other words, we have no registers available at all.
338 * Only swapgs can save us!
339 */
340
341#if defined(__lint)
342
343/*ARGSUSED*/
344void
345sys_syscall()
346{}
347
348void
349_allsyscalls()
350{}
351
352size_t _allsyscalls_size;
353
354#else	/* __lint */
355
356	ENTRY_NP2(brand_sys_syscall,_allsyscalls)
357	SWAPGS
358	BRAND_CALLBACK(BRAND_CB_SYSCALL)
359	SWAPGS
360
361	ALTENTRY(sys_syscall)
362	SWAPGS
363	movq	%rsp, %gs:CPU_RTMP_RSP
364	movq	%r15, %gs:CPU_RTMP_R15
365	movq	%gs:CPU_THREAD, %r15
366	movq	T_STACK(%r15), %rsp
367
368	movl	$UCS_SEL, REGOFF_CS(%rsp)
369	movq	%rcx, REGOFF_RIP(%rsp)		/* syscall: %rip -> %rcx */
370	movq	%r11, REGOFF_RFL(%rsp)		/* syscall: %rfl -> %r11d */
371	movl	$UDS_SEL, REGOFF_SS(%rsp)
372
373	movl	%eax, %eax			/* wrapper: sysc# -> %eax */
374	movq	%rdi, REGOFF_RDI(%rsp)
375	movq	%rsi, REGOFF_RSI(%rsp)
376	movq	%rdx, REGOFF_RDX(%rsp)
377	movq	%r10, REGOFF_RCX(%rsp)		/* wrapper: %rcx -> %r10 */
378	movq	%r10, %rcx			/* arg[3] for direct calls */
379
380	movq	%r8, REGOFF_R8(%rsp)
381	movq	%r9, REGOFF_R9(%rsp)
382	movq	%rax, REGOFF_RAX(%rsp)
383	movq	%rbx, REGOFF_RBX(%rsp)
384
385	movq	%rbp, REGOFF_RBP(%rsp)
386	movq	%r10, REGOFF_R10(%rsp)
387	movq	%gs:CPU_RTMP_RSP, %r11
388	movq	%r11, REGOFF_RSP(%rsp)
389	movq	%r12, REGOFF_R12(%rsp)
390
391	movq	%r13, REGOFF_R13(%rsp)
392	movq	%r14, REGOFF_R14(%rsp)
393	movq	%gs:CPU_RTMP_R15, %r10
394	movq	%r10, REGOFF_R15(%rsp)
395	movq	$0, REGOFF_SAVFP(%rsp)
396	movq	$0, REGOFF_SAVPC(%rsp)
397
398	/*
399	 * Copy these registers here in case we end up stopped with
400	 * someone (like, say, /proc) messing with our register state.
401	 * We don't -restore- them unless we have to in update_sregs.
402	 *
403	 * Since userland -can't- change fsbase or gsbase directly,
404	 * and capturing them involves two serializing instructions,
405	 * we don't bother to capture them here.
406	 */
407	xorl	%ebx, %ebx
408	movw	%ds, %bx
409	movq	%rbx, REGOFF_DS(%rsp)
410	movw	%es, %bx
411	movq	%rbx, REGOFF_ES(%rsp)
412	movw	%fs, %bx
413	movq	%rbx, REGOFF_FS(%rsp)
414	movw	%gs, %bx
415	movq	%rbx, REGOFF_GS(%rsp)
416
417	/*
418	 * Machine state saved in the regs structure on the stack
419	 * First six args in %rdi, %rsi, %rdx, %rcx, %r8, %r9
420	 * %eax is the syscall number
421	 * %rsp is the thread's stack, %r15 is curthread
422	 * REG_RSP(%rsp) is the user's stack
423	 */
424
425	SYSCALL_TRAPTRACE($TT_SYSC64)
426
427	movq	%rsp, %rbp
428
429	movq	T_LWP(%r15), %r14
430	ASSERT_NO_RUPDATE_PENDING(%r14)
431	ENABLE_INTR_FLAGS
432
433	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
434	movl	REGOFF_RAX(%rsp), %eax	/* (%rax damaged by mstate call) */
435
436	ASSERT_LWPTOREGS(%r14, %rsp)
437
438	movb	$LWP_SYS, LWP_STATE(%r14)
439	incq	LWP_RU_SYSC(%r14)
440	movb	$NORMALRETURN, LWP_EOSYS(%r14)
441
442	incq	%gs:CPU_STATS_SYS_SYSCALL
443
444	movw	%ax, T_SYSNUM(%r15)
445	movzbl	T_PRE_SYS(%r15), %ebx
446	ORL_SYSCALLTRACE(%ebx)
447	testl	%ebx, %ebx
448	jne	_syscall_pre
449
450_syscall_invoke:
451	movq	REGOFF_RDI(%rbp), %rdi
452	movq	REGOFF_RSI(%rbp), %rsi
453	movq	REGOFF_RDX(%rbp), %rdx
454	movq	REGOFF_RCX(%rbp), %rcx
455	movq	REGOFF_R8(%rbp), %r8
456	movq	REGOFF_R9(%rbp), %r9
457
458	cmpl	$NSYSCALL, %eax
459	jae	_syscall_ill
460	shll	$SYSENT_SIZE_SHIFT, %eax
461	leaq	sysent(%rax), %rbx
462
463	call	*SY_CALLC(%rbx)
464
465	movq	%rax, %r12
466	movq	%rdx, %r13
467
468	/*
469	 * If the handler returns two ints, then we need to split the
470	 * 64-bit return value into two 32-bit values.
471	 */
472	testw	$SE_32RVAL2, SY_FLAGS(%rbx)
473	je	5f
474	movq	%r12, %r13
475	shrq	$32, %r13	/* upper 32-bits into %edx */
476	movl	%r12d, %r12d	/* lower 32-bits into %eax */
4775:
478	/*
479	 * Optimistically assume that there's no post-syscall
480	 * work to do.  (This is to avoid having to call syscall_mstate()
481	 * with interrupts disabled)
482	 */
483	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
484
485	/*
486	 * We must protect ourselves from being descheduled here;
487	 * If we were, and we ended up on another cpu, or another
488	 * lwp got in ahead of us, it could change the segment
489	 * registers without us noticing before we return to userland.
490	 */
491	CLI(%r14)
492	CHECK_POSTSYS_NE(%r15, %r14, %ebx)
493	jne	_syscall_post
494	SIMPLE_SYSCALL_POSTSYS(%r15, %r14, %bx)
495
496	movq	%r12, REGOFF_RAX(%rsp)
497	movq	%r13, REGOFF_RDX(%rsp)
498
499	/*
500	 * To get back to userland, we need the return %rip in %rcx and
501	 * the return %rfl in %r11d.  The sysretq instruction also arranges
502	 * to fix up %cs and %ss; everything else is our responsibility.
503	 */
504	movq	REGOFF_RDI(%rsp), %rdi
505	movq	REGOFF_RSI(%rsp), %rsi
506	movq	REGOFF_RDX(%rsp), %rdx
507	/* %rcx used to restore %rip value */
508
509	movq	REGOFF_R8(%rsp), %r8
510	movq	REGOFF_R9(%rsp), %r9
511	movq	REGOFF_RAX(%rsp), %rax
512	movq	REGOFF_RBX(%rsp), %rbx
513
514	movq	REGOFF_RBP(%rsp), %rbp
515	movq	REGOFF_R10(%rsp), %r10
516	/* %r11 used to restore %rfl value */
517	movq	REGOFF_R12(%rsp), %r12
518
519	movq	REGOFF_R13(%rsp), %r13
520	movq	REGOFF_R14(%rsp), %r14
521	movq	REGOFF_R15(%rsp), %r15
522
523	movq	REGOFF_RIP(%rsp), %rcx
524	movl	REGOFF_RFL(%rsp), %r11d
525	movq	REGOFF_RSP(%rsp), %rsp
526	SWAPGS
527	sysretq
528
529_syscall_pre:
530	call	pre_syscall
531	movl	%eax, %r12d
532	testl	%eax, %eax
533	jne	_syscall_post_call
534	/*
535	 * Didn't abort, so reload the syscall args and invoke the handler.
536	 */
537	movzwl	T_SYSNUM(%r15), %eax
538	jmp	_syscall_invoke
539
540_syscall_ill:
541	call	nosys
542	movq	%rax, %r12
543	movq	%rdx, %r13
544	jmp	_syscall_post_call
545
546_syscall_post:
547	STI
548	/*
549	 * Sigh, our optimism wasn't justified, put it back to LMS_SYSTEM
550	 * so that we can account for the extra work it takes us to finish.
551	 */
552	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
553_syscall_post_call:
554	movq	%r12, %rdi
555	movq	%r13, %rsi
556	call	post_syscall
557	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
558	jmp	_sys_rtt
559	SET_SIZE(sys_syscall)
560	SET_SIZE(brand_sys_syscall)
561
562#endif	/* __lint */
563
564#if defined(__lint)
565
566/*ARGSUSED*/
567void
568sys_syscall32()
569{}
570
571#else	/* __lint */
572
573	ENTRY_NP(brand_sys_syscall32)
574	SWAPGS
575	BRAND_CALLBACK(BRAND_CB_SYSCALL32)
576	SWAPGS
577
578	ALTENTRY(sys_syscall32)
579	SWAPGS
580	movl	%esp, %r10d
581	movq	%gs:CPU_THREAD, %r15
582	movq	T_STACK(%r15), %rsp
583	movl	%eax, %eax
584
585	movl	$U32CS_SEL, REGOFF_CS(%rsp)
586	movl	%ecx, REGOFF_RIP(%rsp)		/* syscall: %rip -> %rcx */
587	movq	%r11, REGOFF_RFL(%rsp)		/* syscall: %rfl -> %r11d */
588	movq	%r10, REGOFF_RSP(%rsp)
589	movl	$UDS_SEL, REGOFF_SS(%rsp)
590
591_syscall32_save:
592	movl	%edi, REGOFF_RDI(%rsp)
593	movl	%esi, REGOFF_RSI(%rsp)
594	movl	%ebp, REGOFF_RBP(%rsp)
595	movl	%ebx, REGOFF_RBX(%rsp)
596	movl	%edx, REGOFF_RDX(%rsp)
597	movl	%ecx, REGOFF_RCX(%rsp)
598	movl	%eax, REGOFF_RAX(%rsp)		/* wrapper: sysc# -> %eax */
599	movq	$0, REGOFF_SAVFP(%rsp)
600	movq	$0, REGOFF_SAVPC(%rsp)
601
602	/*
603	 * Copy these registers here in case we end up stopped with
604	 * someone (like, say, /proc) messing with our register state.
605	 * We don't -restore- them unless we have to in update_sregs.
606	 *
607	 * Since userland -can't- change fsbase or gsbase directly,
608	 * we don't bother to capture them here.
609	 */
610	xorl	%ebx, %ebx
611	movw	%ds, %bx
612	movq	%rbx, REGOFF_DS(%rsp)
613	movw	%es, %bx
614	movq	%rbx, REGOFF_ES(%rsp)
615	movw	%fs, %bx
616	movq	%rbx, REGOFF_FS(%rsp)
617	movw	%gs, %bx
618	movq	%rbx, REGOFF_GS(%rsp)
619
620	/*
621	 * Application state saved in the regs structure on the stack
622	 * %eax is the syscall number
623	 * %rsp is the thread's stack, %r15 is curthread
624	 * REG_RSP(%rsp) is the user's stack
625	 */
626
627	SYSCALL_TRAPTRACE32($TT_SYSC)
628
629	movq	%rsp, %rbp
630
631	movq	T_LWP(%r15), %r14
632	ASSERT_NO_RUPDATE_PENDING(%r14)
633
634	ENABLE_INTR_FLAGS
635
636	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
637	movl	REGOFF_RAX(%rsp), %eax	/* (%rax damaged by mstate call) */
638
639	ASSERT_LWPTOREGS(%r14, %rsp)
640
641	incq	 %gs:CPU_STATS_SYS_SYSCALL
642
643	/*
644	 * Make some space for MAXSYSARGS (currently 8) 32-bit args placed
645	 * into 64-bit (long) arg slots, maintaining 16 byte alignment.  Or
646	 * more succinctly:
647	 *
648	 *	SA(MAXSYSARGS * sizeof (long)) == 64
649	 */
650#define	SYS_DROP	64			/* drop for args */
651	subq	$SYS_DROP, %rsp
652	movb	$LWP_SYS, LWP_STATE(%r14)
653	movq	%r15, %rdi
654	movq	%rsp, %rsi
655	call	syscall_entry
656
657	/*
658	 * Fetch the arguments copied onto the kernel stack and put
659	 * them in the right registers to invoke a C-style syscall handler.
660	 * %rax contains the handler address.
661	 *
662	 * Ideas for making all this go faster of course include simply
663	 * forcibly fetching 6 arguments from the user stack under lofault
664	 * protection, reverting to copyin_args only when watchpoints
665	 * are in effect.
666	 *
667	 * (If we do this, make sure that exec and libthread leave
668	 * enough space at the top of the stack to ensure that we'll
669	 * never do a fetch from an invalid page.)
670	 *
671	 * Lots of ideas here, but they won't really help with bringup B-)
672	 * Correctness can't wait, performance can wait a little longer ..
673	 */
674
675	movq	%rax, %rbx
676	movl	0(%rsp), %edi
677	movl	8(%rsp), %esi
678	movl	0x10(%rsp), %edx
679	movl	0x18(%rsp), %ecx
680	movl	0x20(%rsp), %r8d
681	movl	0x28(%rsp), %r9d
682
683	call	*SY_CALLC(%rbx)
684
685	movq	%rbp, %rsp	/* pop the args */
686
687	/*
688	 * amd64 syscall handlers -always- return a 64-bit value in %rax.
689	 * On the 32-bit kernel, they always return that value in %eax:%edx
690	 * as required by the 32-bit ABI.
691	 *
692	 * Simulate the same behaviour by unconditionally splitting the
693	 * return value in the same way.
694	 */
695	movq	%rax, %r13
696	shrq	$32, %r13	/* upper 32-bits into %edx */
697	movl	%eax, %r12d	/* lower 32-bits into %eax */
698
699	/*
700	 * Optimistically assume that there's no post-syscall
701	 * work to do.  (This is to avoid having to call syscall_mstate()
702	 * with interrupts disabled)
703	 */
704	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
705
706	/*
707	 * We must protect ourselves from being descheduled here;
708	 * If we were, and we ended up on another cpu, or another
709	 * lwp got in ahead of us, it could change the segment
710	 * registers without us noticing before we return to userland.
711	 */
712	CLI(%r14)
713	CHECK_POSTSYS_NE(%r15, %r14, %ebx)
714	jne	_full_syscall_postsys32
715	SIMPLE_SYSCALL_POSTSYS(%r15, %r14, %bx)
716
717	/*
718	 * To get back to userland, we need to put the return %rip in %rcx and
719	 * the return %rfl in %r11d.  The sysret instruction also arranges
720	 * to fix up %cs and %ss; everything else is our responsibility.
721	 */
722
723	movl	%r12d, %eax			/* %eax: rval1 */
724	movl	REGOFF_RBX(%rsp), %ebx
725	/* %ecx used for return pointer */
726	movl	%r13d, %edx			/* %edx: rval2 */
727	movl	REGOFF_RBP(%rsp), %ebp
728	movl	REGOFF_RSI(%rsp), %esi
729	movl	REGOFF_RDI(%rsp), %edi
730
731	movl	REGOFF_RFL(%rsp), %r11d		/* %r11 -> eflags */
732	movl	REGOFF_RIP(%rsp), %ecx		/* %ecx -> %eip */
733	movl	REGOFF_RSP(%rsp), %esp
734
735	swapgs
736	sysretl
737
738_full_syscall_postsys32:
739	STI
740	/*
741	 * Sigh, our optimism wasn't justified, put it back to LMS_SYSTEM
742	 * so that we can account for the extra work it takes us to finish.
743	 */
744	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
745	movq	%r15, %rdi
746	movq	%r12, %rsi			/* rval1 - %eax */
747	movq	%r13, %rdx			/* rval2 - %edx */
748	call	syscall_exit
749	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
750	jmp	_sys_rtt
751	SET_SIZE(sys_syscall32)
752	SET_SIZE(brand_sys_syscall32)
753
754#endif	/* __lint */
755
756/*
757 * System call handler via the sysenter instruction
758 * Used only for 32-bit system calls on the 64-bit kernel.
759 *
760 * The caller in userland has arranged that:
761 *
762 * -	%eax contains the syscall number
763 * -	%ecx contains the user %esp
764 * -	%edx contains the return %eip
765 * -	the user stack contains the args to the syscall
766 *
767 * Hardware and (privileged) initialization code have arranged that by
768 * the time the sysenter instructions completes:
769 *
770 * - %rip is pointing to sys_sysenter (below).
771 * - %cs and %ss are set to kernel text and stack (data) selectors.
772 * - %rsp is pointing at the lwp's stack
773 * - interrupts have been disabled.
774 *
775 * Note that we are unable to return both "rvals" to userland with
776 * this call, as %edx is used by the sysexit instruction.
777 *
778 * One final complication in this routine is its interaction with
779 * single-stepping in a debugger.  For most of the system call mechanisms,
780 * the CPU automatically clears the single-step flag before we enter the
781 * kernel.  The sysenter mechanism does not clear the flag, so a user
782 * single-stepping through a libc routine may suddenly find him/herself
783 * single-stepping through the kernel.  To detect this, kmdb compares the
784 * trap %pc to the [brand_]sys_enter addresses on each single-step trap.
785 * If it finds that we have single-stepped to a sysenter entry point, it
786 * explicitly clears the flag and executes the sys_sysenter routine.
787 *
788 * One final complication in this final complication is the fact that we
789 * have two different entry points for sysenter: brand_sys_sysenter and
790 * sys_sysenter.  If we enter at brand_sys_sysenter and start single-stepping
791 * through the kernel with kmdb, we will eventually hit the instruction at
792 * sys_sysenter.  kmdb cannot distinguish between that valid single-step
793 * and the undesirable one mentioned above.  To avoid this situation, we
794 * simply add a jump over the instruction at sys_sysenter to make it
795 * impossible to single-step to it.
796 */
797#if defined(__lint)
798
799void
800sys_sysenter()
801{}
802
803#else	/* __lint */
804
805	ENTRY_NP(brand_sys_sysenter)
806	SWAPGS
807
808	ALTENTRY(_brand_sys_sysenter_post_swapgs)
809	BRAND_CALLBACK(BRAND_CB_SYSENTER)
810	/*
811	 * Jump over sys_sysenter to allow single-stepping as described
812	 * above.
813	 */
814	jmp	_sys_sysenter_post_swapgs
815
816	ALTENTRY(sys_sysenter)
817	SWAPGS
818
819	ALTENTRY(_sys_sysenter_post_swapgs)
820	movq	%gs:CPU_THREAD, %r15
821
822	movl	$U32CS_SEL, REGOFF_CS(%rsp)
823	movl	%ecx, REGOFF_RSP(%rsp)		/* wrapper: %esp -> %ecx */
824	movl	%edx, REGOFF_RIP(%rsp)		/* wrapper: %eip -> %edx */
825	pushfq
826	popq	%r10
827	movl	$UDS_SEL, REGOFF_SS(%rsp)
828
829	/*
830	 * Set the interrupt flag before storing the flags to the
831	 * flags image on the stack so we can return to user with
832	 * interrupts enabled if we return via sys_rtt_syscall32
833	 */
834	orq	$PS_IE, %r10
835	movq	%r10, REGOFF_RFL(%rsp)
836
837	movl	%edi, REGOFF_RDI(%rsp)
838	movl	%esi, REGOFF_RSI(%rsp)
839	movl	%ebp, REGOFF_RBP(%rsp)
840	movl	%ebx, REGOFF_RBX(%rsp)
841	movl	%edx, REGOFF_RDX(%rsp)
842	movl	%ecx, REGOFF_RCX(%rsp)
843	movl	%eax, REGOFF_RAX(%rsp)		/* wrapper: sysc# -> %eax */
844	movq	$0, REGOFF_SAVFP(%rsp)
845	movq	$0, REGOFF_SAVPC(%rsp)
846
847	/*
848	 * Copy these registers here in case we end up stopped with
849	 * someone (like, say, /proc) messing with our register state.
850	 * We don't -restore- them unless we have to in update_sregs.
851	 *
852	 * Since userland -can't- change fsbase or gsbase directly,
853	 * we don't bother to capture them here.
854	 */
855	xorl	%ebx, %ebx
856	movw	%ds, %bx
857	movq	%rbx, REGOFF_DS(%rsp)
858	movw	%es, %bx
859	movq	%rbx, REGOFF_ES(%rsp)
860	movw	%fs, %bx
861	movq	%rbx, REGOFF_FS(%rsp)
862	movw	%gs, %bx
863	movq	%rbx, REGOFF_GS(%rsp)
864
865	/*
866	 * Application state saved in the regs structure on the stack
867	 * %eax is the syscall number
868	 * %rsp is the thread's stack, %r15 is curthread
869	 * REG_RSP(%rsp) is the user's stack
870	 */
871
872	SYSCALL_TRAPTRACE($TT_SYSENTER)
873
874	movq	%rsp, %rbp
875
876	movq	T_LWP(%r15), %r14
877	ASSERT_NO_RUPDATE_PENDING(%r14)
878
879	ENABLE_INTR_FLAGS
880
881	/*
882	 * Catch 64-bit process trying to issue sysenter instruction
883	 * on Nocona based systems.
884	 */
885	movq	LWP_PROCP(%r14), %rax
886	cmpq	$DATAMODEL_ILP32, P_MODEL(%rax)
887	je	7f
888
889	/*
890	 * For a non-32-bit process, simulate a #ud, since that's what
891	 * native hardware does.  The traptrace entry (above) will
892	 * let you know what really happened.
893	 */
894	movq	$T_ILLINST, REGOFF_TRAPNO(%rsp)
895	movq	REGOFF_CS(%rsp), %rdi
896	movq	%rdi, REGOFF_ERR(%rsp)
897	movq	%rsp, %rdi
898	movq	REGOFF_RIP(%rsp), %rsi
899	movl	%gs:CPU_ID, %edx
900	call	trap
901	jmp	_sys_rtt
9027:
903
904	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
905	movl	REGOFF_RAX(%rsp), %eax	/* (%rax damaged by mstate calls) */
906
907	ASSERT_LWPTOREGS(%r14, %rsp)
908
909	incq	%gs:CPU_STATS_SYS_SYSCALL
910
911	/*
912	 * Make some space for MAXSYSARGS (currently 8) 32-bit args
913	 * placed into 64-bit (long) arg slots, plus one 64-bit
914	 * (long) arg count, maintaining 16 byte alignment.
915	 */
916	subq	$SYS_DROP, %rsp
917	movb	$LWP_SYS, LWP_STATE(%r14)
918	movq	%r15, %rdi
919	movq	%rsp, %rsi
920	call	syscall_entry
921
922	/*
923	 * Fetch the arguments copied onto the kernel stack and put
924	 * them in the right registers to invoke a C-style syscall handler.
925	 * %rax contains the handler address.
926	 */
927	movq	%rax, %rbx
928	movl	0(%rsp), %edi
929	movl	8(%rsp), %esi
930	movl	0x10(%rsp), %edx
931	movl	0x18(%rsp), %ecx
932	movl	0x20(%rsp), %r8d
933	movl	0x28(%rsp), %r9d
934
935	call	*SY_CALLC(%rbx)
936
937	movq	%rbp, %rsp	/* pop the args */
938
939	/*
940	 * amd64 syscall handlers -always- return a 64-bit value in %rax.
941	 * On the 32-bit kernel, the always return that value in %eax:%edx
942	 * as required by the 32-bit ABI.
943	 *
944	 * Simulate the same behaviour by unconditionally splitting the
945	 * return value in the same way.
946	 */
947	movq	%rax, %r13
948	shrq	$32, %r13	/* upper 32-bits into %edx */
949	movl	%eax, %r12d	/* lower 32-bits into %eax */
950
951	/*
952	 * Optimistically assume that there's no post-syscall
953	 * work to do.  (This is to avoid having to call syscall_mstate()
954	 * with interrupts disabled)
955	 */
956	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
957
958	/*
959	 * We must protect ourselves from being descheduled here;
960	 * If we were, and we ended up on another cpu, or another
961	 * lwp got int ahead of us, it could change the segment
962	 * registers without us noticing before we return to userland.
963	 */
964	cli
965	CHECK_POSTSYS_NE(%r15, %r14, %ebx)
966	jne	_full_syscall_postsys32
967	SIMPLE_SYSCALL_POSTSYS(%r15, %r14, %bx)
968
969	/*
970	 * To get back to userland, load up the 32-bit registers and
971	 * sysexit back where we came from.
972	 */
973
974	/*
975	 * Interrupts will be turned on by the 'sti' executed just before
976	 * sysexit.  The following ensures that restoring the user's rflags
977	 * doesn't enable interrupts too soon.
978	 */
979	andq	$_BITNOT(PS_IE), REGOFF_RFL(%rsp)
980
981	/*
982	 * (There's no point in loading up %edx because the sysexit
983	 * mechanism smashes it.)
984	 */
985	movl	%r12d, %eax
986	movl	REGOFF_RBX(%rsp), %ebx
987	movl	REGOFF_RBP(%rsp), %ebp
988	movl	REGOFF_RSI(%rsp), %esi
989	movl	REGOFF_RDI(%rsp), %edi
990
991	movl	REGOFF_RIP(%rsp), %edx	/* sysexit: %edx -> %eip */
992	pushq	REGOFF_RFL(%rsp)
993	popfq
994	movl	REGOFF_RSP(%rsp), %ecx	/* sysexit: %ecx -> %esp */
995	swapgs
996	sti
997	sysexit
998	SET_SIZE(sys_sysenter)
999	SET_SIZE(_sys_sysenter_post_swapgs)
1000	SET_SIZE(brand_sys_sysenter)
1001
1002#endif	/* __lint */
1003
1004#if defined(__lint)
1005/*
1006 * System call via an int80.  This entry point is only used by the Linux
1007 * application environment.  Unlike the other entry points, there is no
1008 * default action to take if no callback is registered for this process.
1009 */
1010void
1011sys_int80()
1012{}
1013
1014#else	/* __lint */
1015
1016	ENTRY_NP(brand_sys_int80)
1017	swapgs
1018	BRAND_CALLBACK(BRAND_CB_INT80)
1019	swapgs
1020
1021	ENTRY_NP(sys_int80)
1022	/*
1023	 * We hit an int80, but this process isn't of a brand with an int80
1024	 * handler.  Bad process!  Make it look as if the INT failed.
1025	 * Modify %eip to point before the INT, push the expected error
1026	 * code and fake a GP fault.
1027	 *
1028	 */
1029	subq	$2, (%rsp)	/* int insn 2-bytes */
1030	pushq	$_CONST(_MUL(T_INT80, GATE_DESC_SIZE) + 2)
1031	jmp	gptrap			/ GP fault
1032	SET_SIZE(sys_int80)
1033	SET_SIZE(brand_sys_int80)
1034#endif	/* __lint */
1035
1036
1037/*
1038 * This is the destination of the "int $T_SYSCALLINT" interrupt gate, used by
1039 * the generic i386 libc to do system calls. We do a small amount of setup
1040 * before jumping into the existing sys_syscall32 path.
1041 */
1042#if defined(__lint)
1043
1044/*ARGSUSED*/
1045void
1046sys_syscall_int()
1047{}
1048
1049#else	/* __lint */
1050
1051	ENTRY_NP(brand_sys_syscall_int)
1052	SWAPGS
1053	BRAND_CALLBACK(BRAND_CB_INT91)
1054	swapgs
1055
1056	ALTENTRY(sys_syscall_int)
1057	swapgs
1058	movq	%gs:CPU_THREAD, %r15
1059	movq	T_STACK(%r15), %rsp
1060	movl	%eax, %eax
1061	/*
1062	 * Set t_post_sys on this thread to force ourselves out via the slow
1063	 * path. It might be possible at some later date to optimize this out
1064	 * and use a faster return mechanism.
1065	 */
1066	movb	$1, T_POST_SYS(%r15)
1067	CLEAN_CS
1068	jmp	_syscall32_save
1069	SET_SIZE(sys_syscall_int)
1070	SET_SIZE(brand_sys_syscall_int)
1071
1072#endif	/* __lint */
1073
1074/*
1075 * Legacy 32-bit applications and old libc implementations do lcalls;
1076 * we should never get here because the LDT entry containing the syscall
1077 * segment descriptor has the "segment present" bit cleared, which means
1078 * we end up processing those system calls in trap() via a not-present trap.
1079 *
1080 * We do it this way because a call gate unhelpfully does -nothing- to the
1081 * interrupt flag bit, so an interrupt can run us just after the lcall
1082 * completes, but just before the swapgs takes effect.   Thus the INTR_PUSH and
1083 * INTR_POP paths would have to be slightly more complex to dance around
1084 * this problem, and end up depending explicitly on the first
1085 * instruction of this handler being either swapgs or cli.
1086 */
1087
1088#if defined(__lint)
1089
1090/*ARGSUSED*/
1091void
1092sys_lcall32()
1093{}
1094
1095#else	/* __lint */
1096
1097	ENTRY_NP(sys_lcall32)
1098	SWAPGS
1099	pushq	$0
1100	pushq	%rbp
1101	movq	%rsp, %rbp
1102	leaq	__lcall_panic_str(%rip), %rdi
1103	xorl	%eax, %eax
1104	call	panic
1105	SET_SIZE(sys_lcall32)
1106
1107__lcall_panic_str:
1108	.string	"sys_lcall32: shouldn't be here!"
1109
1110/*
1111 * Declare a uintptr_t which covers the entire pc range of syscall
1112 * handlers for the stack walkers that need this.
1113 */
1114	.align	CPTRSIZE
1115	.globl	_allsyscalls_size
1116	.type	_allsyscalls_size, @object
1117_allsyscalls_size:
1118	.NWORD	. - _allsyscalls
1119	SET_SIZE(_allsyscalls_size)
1120
1121#endif	/* __lint */
1122
1123/*
1124 * These are the thread context handlers for lwps using sysenter/sysexit.
1125 */
1126
1127#if defined(__lint)
1128
1129/*ARGSUSED*/
1130void
1131sep_save(void *ksp)
1132{}
1133
1134/*ARGSUSED*/
1135void
1136sep_restore(void *ksp)
1137{}
1138
1139#else	/* __lint */
1140
1141	/*
1142	 * setting this value to zero as we switch away causes the
1143	 * stack-pointer-on-sysenter to be NULL, ensuring that we
1144	 * don't silently corrupt another (preempted) thread stack
1145	 * when running an lwp that (somehow) didn't get sep_restore'd
1146	 */
1147	ENTRY_NP(sep_save)
1148	xorl	%edx, %edx
1149	xorl	%eax, %eax
1150	movl	$MSR_INTC_SEP_ESP, %ecx
1151	wrmsr
1152	ret
1153	SET_SIZE(sep_save)
1154
1155	/*
1156	 * Update the kernel stack pointer as we resume onto this cpu.
1157	 */
1158	ENTRY_NP(sep_restore)
1159	movq	%rdi, %rdx
1160	shrq	$32, %rdx
1161	movl	%edi, %eax
1162	movl	$MSR_INTC_SEP_ESP, %ecx
1163	wrmsr
1164	ret
1165	SET_SIZE(sep_restore)
1166
1167#endif	/* __lint */
1168