xref: /illumos-gate/usr/src/uts/i86pc/ml/syscall_asm_amd64.S (revision 41e0a469c3dbc14deb2b200f6ca6f6e00b5865d0)
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) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2019 Joyent, Inc.
24 * Copyright (c) 2016 by Delphix. All rights reserved.
25 */
26
27#include <sys/asm_linkage.h>
28#include <sys/asm_misc.h>
29#include <sys/regset.h>
30#include <sys/privregs.h>
31#include <sys/psw.h>
32#include <sys/machbrand.h>
33
34#include <sys/segments.h>
35#include <sys/pcb.h>
36#include <sys/trap.h>
37#include <sys/ftrace.h>
38#include <sys/traptrace.h>
39#include <sys/clock.h>
40#include <sys/model.h>
41#include <sys/panic.h>
42
43#if defined(__xpv)
44#include <sys/hypervisor.h>
45#endif
46
47#include "assym.h"
48
49/*
50 * We implement five flavours of system call entry points
51 *
52 * -	syscall/sysretq		(amd64 generic)
53 * -	syscall/sysretl		(i386 plus SYSC bit)
54 * -	sysenter/sysexit	(i386 plus SEP bit)
55 * -	int/iret		(i386 generic)
56 * -	lcall/iret		(i386 generic)
57 *
58 * The current libc included in Solaris uses int/iret as the base unoptimized
59 * kernel entry method. Older libc implementations and legacy binaries may use
60 * the lcall call gate, so it must continue to be supported.
61 *
62 * System calls that use an lcall call gate are processed in trap() via a
63 * segment-not-present trap, i.e. lcalls are extremely slow(!).
64 *
65 * The basic pattern used in the 32-bit SYSC handler at this point in time is
66 * to have the bare minimum of assembler, and get to the C handlers as
67 * quickly as possible.
68 *
69 * The 64-bit handler is much closer to the sparcv9 handler; that's
70 * because of passing arguments in registers.  The 32-bit world still
71 * passes arguments on the stack -- that makes that handler substantially
72 * more complex.
73 *
74 * The two handlers share a few code fragments which are broken
75 * out into preprocessor macros below.
76 *
77 * XX64	come back and speed all this up later.  The 32-bit stuff looks
78 * especially easy to speed up the argument copying part ..
79 *
80 *
81 * Notes about segment register usage (c.f. the 32-bit kernel)
82 *
83 * In the 32-bit kernel, segment registers are dutifully saved and
84 * restored on all mode transitions because the kernel uses them directly.
85 * When the processor is running in 64-bit mode, segment registers are
86 * largely ignored.
87 *
88 * %cs and %ss
89 *	controlled by the hardware mechanisms that make mode transitions
90 *
91 * The remaining segment registers have to either be pointing at a valid
92 * descriptor i.e. with the 'present' bit set, or they can NULL descriptors
93 *
94 * %ds and %es
95 *	always ignored
96 *
97 * %fs and %gs
98 *	fsbase and gsbase are used to control the place they really point at.
99 *	The kernel only depends on %gs, and controls its own gsbase via swapgs
100 *
101 * Note that loading segment registers is still costly because the GDT
102 * lookup still happens (this is because the hardware can't know that we're
103 * not setting up these segment registers for a 32-bit program).  Thus we
104 * avoid doing this in the syscall path, and defer them to lwp context switch
105 * handlers, so the register values remain virtualized to the lwp.
106 */
107
108#if defined(SYSCALLTRACE)
109#define	ORL_SYSCALLTRACE(r32)		\
110	orl	syscalltrace(%rip), r32
111#else
112#define	ORL_SYSCALLTRACE(r32)
113#endif
114
115/*
116 * In the 32-bit kernel, we do absolutely nothing before getting into the
117 * brand callback checks.  In 64-bit land, we do swapgs and then come here.
118 * We assume that the %rsp- and %r15-stashing fields in the CPU structure
119 * are still unused.
120 *
121 * Check if a brand_mach_ops callback is defined for the specified callback_id
122 * type.  If so invoke it with the kernel's %gs value loaded and the following
123 * data on the stack:
124 *
125 * stack:  --------------------------------------
126 *      32 | callback pointer			|
127 *    | 24 | user (or interrupt) stack pointer	|
128 *    | 16 | lwp pointer			|
129 *    v  8 | userland return address		|
130 *       0 | callback wrapper return addr	|
131 *         --------------------------------------
132 *
133 * Since we're pushing the userland return address onto the kernel stack
134 * we need to get that address without accessing the user's stack (since we
135 * can't trust that data).  There are different ways to get the userland
136 * return address depending on how the syscall trap was made:
137 *
138 * a) For sys_syscall and sys_syscall32 the return address is in %rcx.
139 * b) For sys_sysenter the return address is in %rdx.
140 * c) For sys_int80 and sys_syscall_int (int91), upon entry into the macro,
141 *    the stack pointer points at the state saved when we took the interrupt:
142 *	 ------------------------
143 *    |  | user's %ss		|
144 *    |  | user's %esp		|
145 *    |  | EFLAGS register	|
146 *    v  | user's %cs		|
147 *       | user's %eip		|
148 *	 ------------------------
149 *
150 * The 2nd parameter to the BRAND_CALLBACK macro is either the
151 * BRAND_URET_FROM_REG or BRAND_URET_FROM_INTR_STACK macro.  These macros are
152 * used to generate the proper code to get the userland return address for
153 * each syscall entry point.
154 *
155 * The interface to the brand callbacks on the 64-bit kernel assumes %r15
156 * is available as a scratch register within the callback.  If the callback
157 * returns within the kernel then this macro will restore %r15.  If the
158 * callback is going to return directly to userland then it should restore
159 * %r15 before returning to userland.
160 */
161#define BRAND_URET_FROM_REG(rip_reg)					\
162	pushq	rip_reg			/* push the return address	*/
163
164/*
165 * The interrupt stack pointer we saved on entry to the BRAND_CALLBACK macro
166 * is currently pointing at the user return address (%eip).
167 */
168#define BRAND_URET_FROM_INTR_STACK()					\
169	movq	%gs:CPU_RTMP_RSP, %r15	/* grab the intr. stack pointer	*/ ;\
170	pushq	(%r15)			/* push the return address	*/
171
172#define	BRAND_CALLBACK(callback_id, push_userland_ret)			    \
173	movq	%rsp, %gs:CPU_RTMP_RSP	/* save the stack pointer	*/ ;\
174	movq	%r15, %gs:CPU_RTMP_R15	/* save %r15			*/ ;\
175	movq	%gs:CPU_THREAD, %r15	/* load the thread pointer	*/ ;\
176	movq	T_STACK(%r15), %rsp	/* switch to the kernel stack	*/ ;\
177	subq	$16, %rsp		/* save space for 2 pointers	*/ ;\
178	pushq	%r14			/* save %r14			*/ ;\
179	movq	%gs:CPU_RTMP_RSP, %r14					   ;\
180	movq	%r14, 8(%rsp)		/* stash the user stack pointer	*/ ;\
181	popq	%r14			/* restore %r14			*/ ;\
182	movq	T_LWP(%r15), %r15	/* load the lwp pointer		*/ ;\
183	pushq	%r15			/* push the lwp pointer		*/ ;\
184	movq	LWP_PROCP(%r15), %r15	/* load the proc pointer	*/ ;\
185	movq	P_BRAND(%r15), %r15	/* load the brand pointer	*/ ;\
186	movq	B_MACHOPS(%r15), %r15	/* load the machops pointer	*/ ;\
187	movq	_CONST(_MUL(callback_id, CPTRSIZE))(%r15), %r15		   ;\
188	cmpq	$0, %r15						   ;\
189	je	1f							   ;\
190	movq	%r15, 16(%rsp)		/* save the callback pointer	*/ ;\
191	push_userland_ret		/* push the return address	*/ ;\
192	movq	24(%rsp), %r15		/* load callback pointer	*/ ;\
193	INDIRECT_CALL_REG(r15)		/* call callback		*/ ;\
1941:	movq	%gs:CPU_RTMP_R15, %r15	/* restore %r15			*/ ;\
195	movq	%gs:CPU_RTMP_RSP, %rsp	/* restore the stack pointer	*/
196
197#define	MSTATE_TRANSITION(from, to)		\
198	movl	$from, %edi;			\
199	movl	$to, %esi;			\
200	call	syscall_mstate
201
202/*
203 * Check to see if a simple (direct) return is possible i.e.
204 *
205 *	if (t->t_post_sys_ast | syscalltrace |
206 *	    lwp->lwp_pcb.pcb_rupdate == 1)
207 *		do full version	;
208 *
209 * Preconditions:
210 * -	t is curthread
211 * Postconditions:
212 * -	condition code NE is set if post-sys is too complex
213 * -	rtmp is zeroed if it isn't (we rely on this!)
214 * -	ltmp is smashed
215 */
216#define	CHECK_POSTSYS_NE(t, ltmp, rtmp)			\
217	movq	T_LWP(t), ltmp;				\
218	movzbl	PCB_RUPDATE(ltmp), rtmp;		\
219	ORL_SYSCALLTRACE(rtmp);				\
220	orl	T_POST_SYS_AST(t), rtmp;		\
221	cmpl	$0, rtmp
222
223/*
224 * Fix up the lwp, thread, and eflags for a successful return
225 *
226 * Preconditions:
227 * -	zwreg contains zero
228 */
229#define	SIMPLE_SYSCALL_POSTSYS(t, lwp, zwreg)		\
230	movb	$LWP_USER, LWP_STATE(lwp);		\
231	movw	zwreg, T_SYSNUM(t);			\
232	andb	$_CONST(0xffff - PS_C), REGOFF_RFL(%rsp)
233
234/*
235 * ASSERT(lwptoregs(lwp) == rp);
236 *
237 * This may seem obvious, but very odd things happen if this
238 * assertion is false
239 *
240 * Preconditions:
241 *	(%rsp is ready for normal call sequence)
242 * Postconditions (if assertion is true):
243 *	%r11 is smashed
244 *
245 * ASSERT(rp->r_cs == descnum)
246 *
247 * The code selector is written into the regs structure when the
248 * lwp stack is created.  We use this ASSERT to validate that
249 * the regs structure really matches how we came in.
250 *
251 * Preconditions:
252 *	(%rsp is ready for normal call sequence)
253 * Postconditions (if assertion is true):
254 *	-none-
255 *
256 * ASSERT(lwp->lwp_pcb.pcb_rupdate == 0);
257 *
258 * If this is false, it meant that we returned to userland without
259 * updating the segment registers as we were supposed to.
260 *
261 * Note that we must ensure no interrupts or other traps intervene
262 * between entering privileged mode and performing the assertion,
263 * otherwise we may perform a context switch on the thread, which
264 * will end up setting pcb_rupdate to 1 again.
265 *
266 * ASSERT(%cr0 & CR0_TS == 0);
267 * Preconditions:
268 *	(%rsp is ready for normal call sequence)
269 * Postconditions (if assertion is true):
270 *      (specified register is clobbered)
271 *
272 * Check to make sure that we are returning to user land and that CR0.TS
273 * is not set. This is required as part of the eager FPU (see
274 * uts/intel/os/fpu.c for more information).
275 */
276
277#if defined(DEBUG)
278
279__lwptoregs_msg:
280	.string	"syscall_asm_amd64.s:%d lwptoregs(%p) [%p] != rp [%p]"
281
282__codesel_msg:
283	.string	"syscall_asm_amd64.s:%d rp->r_cs [%ld] != %ld"
284
285__no_rupdate_msg:
286	.string	"syscall_asm_amd64.s:%d lwp %p, pcb_rupdate != 0"
287
288__bad_ts_msg:
289	.string "syscall_asm_amd64.s:%d CR0.TS set on user return"
290
291#define	ASSERT_LWPTOREGS(lwp, rp)			\
292	movq	LWP_REGS(lwp), %r11;			\
293	cmpq	rp, %r11;				\
294	je	7f;					\
295	leaq	__lwptoregs_msg(%rip), %rdi;		\
296	movl	$__LINE__, %esi;			\
297	movq	lwp, %rdx;				\
298	movq	%r11, %rcx;				\
299	movq	rp, %r8;				\
300	xorl	%eax, %eax;				\
301	call	panic;					\
3027:
303
304#define	ASSERT_NO_RUPDATE_PENDING(lwp)			\
305	testb	$0x1, PCB_RUPDATE(lwp);			\
306	je	8f;					\
307	movq	lwp, %rdx;				\
308	leaq	__no_rupdate_msg(%rip), %rdi;		\
309	movl	$__LINE__, %esi;			\
310	xorl	%eax, %eax;				\
311	call	panic;					\
3128:
313
314#define	ASSERT_CR0TS_ZERO(reg)				\
315	movq	%cr0, reg;				\
316	testq	$CR0_TS, reg;				\
317	jz	9f;					\
318	leaq	__bad_ts_msg(%rip), %rdi;		\
319	movl	$__LINE__, %esi;			\
320	xorl	%eax, %eax;				\
321	call	panic;					\
3229:
323
324#else
325#define	ASSERT_LWPTOREGS(lwp, rp)
326#define	ASSERT_NO_RUPDATE_PENDING(lwp)
327#define	ASSERT_CR0TS_ZERO(reg)
328#endif
329
330/*
331 * Do the traptrace thing and restore any registers we used
332 * in situ.  Assumes that %rsp is pointing at the base of
333 * the struct regs, obviously ..
334 */
335#ifdef TRAPTRACE
336#define	SYSCALL_TRAPTRACE(ttype)				\
337	TRACE_PTR(%rdi, %rbx, %ebx, %rcx, ttype);		\
338	TRACE_REGS(%rdi, %rsp, %rbx, %rcx);			\
339	TRACE_STAMP(%rdi);	/* rdtsc clobbers %eax, %edx */	\
340	movq	REGOFF_RAX(%rsp), %rax;				\
341	movq	REGOFF_RBX(%rsp), %rbx;				\
342	movq	REGOFF_RCX(%rsp), %rcx;				\
343	movq	REGOFF_RDX(%rsp), %rdx;				\
344	movl	%eax, TTR_SYSNUM(%rdi);				\
345	movq	REGOFF_RDI(%rsp), %rdi
346
347#define	SYSCALL_TRAPTRACE32(ttype)				\
348	SYSCALL_TRAPTRACE(ttype);				\
349	/* paranoia: clean the top 32-bits of the registers */	\
350	orl	%eax, %eax;					\
351	orl	%ebx, %ebx;					\
352	orl	%ecx, %ecx;					\
353	orl	%edx, %edx;					\
354	orl	%edi, %edi
355#else	/* TRAPTRACE */
356#define	SYSCALL_TRAPTRACE(ttype)
357#define	SYSCALL_TRAPTRACE32(ttype)
358#endif	/* TRAPTRACE */
359
360/*
361 * The 64-bit libc syscall wrapper does this:
362 *
363 * fn(<args>)
364 * {
365 *	movq	%rcx, %r10	-- because syscall smashes %rcx
366 *	movl	$CODE, %eax
367 *	syscall
368 *	<error processing>
369 * }
370 *
371 * Thus when we come into the kernel:
372 *
373 *	%rdi, %rsi, %rdx, %r10, %r8, %r9 contain first six args
374 *	%rax is the syscall number
375 *	%r12-%r15 contain caller state
376 *
377 * The syscall instruction arranges that:
378 *
379 *	%rcx contains the return %rip
380 *	%r11d contains bottom 32-bits of %rflags
381 *	%rflags is masked (as determined by the SFMASK msr)
382 *	%cs is set to UCS_SEL (as determined by the STAR msr)
383 *	%ss is set to UDS_SEL (as determined by the STAR msr)
384 *	%rip is set to sys_syscall (as determined by the LSTAR msr)
385 *
386 * Or in other words, we have no registers available at all.
387 * Only swapgs can save us!
388 *
389 * Under the hypervisor, the swapgs has happened already.  However, the
390 * state of the world is very different from that we're familiar with.
391 *
392 * In particular, we have a stack structure like that for interrupt
393 * gates, except that the %cs and %ss registers are modified for reasons
394 * that are not entirely clear.  Critically, the %rcx/%r11 values do
395 * *not* reflect the usage of those registers under a 'real' syscall[1];
396 * the stack, therefore, looks like this:
397 *
398 *	0x0(rsp)	potentially junk %rcx
399 *	0x8(rsp)	potentially junk %r11
400 *	0x10(rsp)	user %rip
401 *	0x18(rsp)	modified %cs
402 *	0x20(rsp)	user %rflags
403 *	0x28(rsp)	user %rsp
404 *	0x30(rsp)	modified %ss
405 *
406 *
407 * and before continuing on, we must load the %rip into %rcx and the
408 * %rflags into %r11.
409 *
410 * [1] They used to, and we relied on it, but this was broken in 3.1.1.
411 * Sigh.
412 */
413#if defined(__xpv)
414#define	XPV_SYSCALL_PROD						\
415	movq	0x10(%rsp), %rcx;					\
416	movq	0x20(%rsp), %r11;					\
417	movq	0x28(%rsp), %rsp
418#else
419#define	XPV_SYSCALL_PROD /* nothing */
420#endif
421
422	ENTRY_NP2(brand_sys_syscall,_allsyscalls)
423	SWAPGS				/* kernel gsbase */
424	XPV_SYSCALL_PROD
425	BRAND_CALLBACK(BRAND_CB_SYSCALL, BRAND_URET_FROM_REG(%rcx))
426	jmp	noprod_sys_syscall
427
428	ALTENTRY(sys_syscall)
429	SWAPGS				/* kernel gsbase */
430	XPV_SYSCALL_PROD
431
432noprod_sys_syscall:
433	movq	%r15, %gs:CPU_RTMP_R15
434	movq	%rsp, %gs:CPU_RTMP_RSP
435
436	movq	%gs:CPU_THREAD, %r15
437	movq	T_STACK(%r15), %rsp	/* switch from user to kernel stack */
438
439	ASSERT_UPCALL_MASK_IS_SET
440
441	movl	$UCS_SEL, REGOFF_CS(%rsp)
442	movq	%rcx, REGOFF_RIP(%rsp)		/* syscall: %rip -> %rcx */
443	movq	%r11, REGOFF_RFL(%rsp)		/* syscall: %rfl -> %r11d */
444	movl	$UDS_SEL, REGOFF_SS(%rsp)
445
446	movl	%eax, %eax			/* wrapper: sysc# -> %eax */
447	movq	%rdi, REGOFF_RDI(%rsp)
448	movq	%rsi, REGOFF_RSI(%rsp)
449	movq	%rdx, REGOFF_RDX(%rsp)
450	movq	%r10, REGOFF_RCX(%rsp)		/* wrapper: %rcx -> %r10 */
451	movq	%r10, %rcx			/* arg[3] for direct calls */
452
453	movq	%r8, REGOFF_R8(%rsp)
454	movq	%r9, REGOFF_R9(%rsp)
455	movq	%rax, REGOFF_RAX(%rsp)
456	movq	%rbx, REGOFF_RBX(%rsp)
457
458	movq	%rbp, REGOFF_RBP(%rsp)
459	movq	%r10, REGOFF_R10(%rsp)
460	movq	%gs:CPU_RTMP_RSP, %r11
461	movq	%r11, REGOFF_RSP(%rsp)
462	movq	%r12, REGOFF_R12(%rsp)
463
464	movq	%r13, REGOFF_R13(%rsp)
465	movq	%r14, REGOFF_R14(%rsp)
466	movq	%gs:CPU_RTMP_R15, %r10
467	movq	%r10, REGOFF_R15(%rsp)
468	movq	$0, REGOFF_SAVFP(%rsp)
469	movq	$0, REGOFF_SAVPC(%rsp)
470
471	/*
472	 * Copy these registers here in case we end up stopped with
473	 * someone (like, say, /proc) messing with our register state.
474	 * We don't -restore- them unless we have to in update_sregs.
475	 *
476	 * Since userland -can't- change fsbase or gsbase directly,
477	 * and capturing them involves two serializing instructions,
478	 * we don't bother to capture them here.
479	 */
480	xorl	%ebx, %ebx
481	movw	%ds, %bx
482	movq	%rbx, REGOFF_DS(%rsp)
483	movw	%es, %bx
484	movq	%rbx, REGOFF_ES(%rsp)
485	movw	%fs, %bx
486	movq	%rbx, REGOFF_FS(%rsp)
487	movw	%gs, %bx
488	movq	%rbx, REGOFF_GS(%rsp)
489
490	/*
491	 * If we're trying to use TRAPTRACE though, I take that back: we're
492	 * probably debugging some problem in the SWAPGS logic and want to know
493	 * what the incoming gsbase was.
494	 *
495	 * Since we already did SWAPGS, record the KGSBASE.
496	 */
497#if defined(DEBUG) && defined(TRAPTRACE) && !defined(__xpv)
498	movl	$MSR_AMD_KGSBASE, %ecx
499	rdmsr
500	movl	%eax, REGOFF_GSBASE(%rsp)
501	movl	%edx, REGOFF_GSBASE+4(%rsp)
502#endif
503
504	/*
505	 * Machine state saved in the regs structure on the stack
506	 * First six args in %rdi, %rsi, %rdx, %rcx, %r8, %r9
507	 * %eax is the syscall number
508	 * %rsp is the thread's stack, %r15 is curthread
509	 * REG_RSP(%rsp) is the user's stack
510	 */
511
512	SYSCALL_TRAPTRACE($TT_SYSC64)
513
514	movq	%rsp, %rbp
515
516	movq	T_LWP(%r15), %r14
517	ASSERT_NO_RUPDATE_PENDING(%r14)
518	ENABLE_INTR_FLAGS
519
520	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
521	movl	REGOFF_RAX(%rsp), %eax	/* (%rax damaged by mstate call) */
522
523	ASSERT_LWPTOREGS(%r14, %rsp)
524
525	movb	$LWP_SYS, LWP_STATE(%r14)
526	incq	LWP_RU_SYSC(%r14)
527	movb	$NORMALRETURN, LWP_EOSYS(%r14)
528
529	incq	%gs:CPU_STATS_SYS_SYSCALL
530
531	movw	%ax, T_SYSNUM(%r15)
532	movzbl	T_PRE_SYS(%r15), %ebx
533	ORL_SYSCALLTRACE(%ebx)
534	testl	%ebx, %ebx
535	jne	_syscall_pre
536
537_syscall_invoke:
538	movq	REGOFF_RDI(%rbp), %rdi
539	movq	REGOFF_RSI(%rbp), %rsi
540	movq	REGOFF_RDX(%rbp), %rdx
541	movq	REGOFF_RCX(%rbp), %rcx
542	movq	REGOFF_R8(%rbp), %r8
543	movq	REGOFF_R9(%rbp), %r9
544
545	cmpl	$NSYSCALL, %eax
546	jae	_syscall_ill
547	shll	$SYSENT_SIZE_SHIFT, %eax
548	leaq	sysent(%rax), %rbx
549
550	movq	SY_CALLC(%rbx), %rax
551	INDIRECT_CALL_REG(rax)
552
553	movq	%rax, %r12
554	movq	%rdx, %r13
555
556	/*
557	 * If the handler returns two ints, then we need to split the
558	 * 64-bit return value into two 32-bit values.
559	 */
560	testw	$SE_32RVAL2, SY_FLAGS(%rbx)
561	je	5f
562	movq	%r12, %r13
563	shrq	$32, %r13	/* upper 32-bits into %edx */
564	movl	%r12d, %r12d	/* lower 32-bits into %eax */
5655:
566	/*
567	 * Optimistically assume that there's no post-syscall
568	 * work to do.  (This is to avoid having to call syscall_mstate()
569	 * with interrupts disabled)
570	 */
571	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
572
573	/*
574	 * We must protect ourselves from being descheduled here;
575	 * If we were, and we ended up on another cpu, or another
576	 * lwp got in ahead of us, it could change the segment
577	 * registers without us noticing before we return to userland.
578	 */
579	CLI(%r14)
580	CHECK_POSTSYS_NE(%r15, %r14, %ebx)
581	jne	_syscall_post
582
583	/*
584	 * We need to protect ourselves against non-canonical return values
585	 * because Intel doesn't check for them on sysret (AMD does).  Canonical
586	 * addresses on current amd64 processors only use 48-bits for VAs; an
587	 * address is canonical if all upper bits (47-63) are identical. If we
588	 * find a non-canonical %rip, we opt to go through the full
589	 * _syscall_post path which takes us into an iretq which is not
590	 * susceptible to the same problems sysret is.
591	 *
592	 * We're checking for a canonical address by first doing an arithmetic
593	 * shift. This will fill in the remaining bits with the value of bit 63.
594	 * If the address were canonical, the register would now have either all
595	 * zeroes or all ones in it. Therefore we add one (inducing overflow)
596	 * and compare against 1. A canonical address will either be zero or one
597	 * at this point, hence the use of ja.
598	 *
599	 * At this point, r12 and r13 have the return value so we can't use
600	 * those registers.
601	 */
602	movq	REGOFF_RIP(%rsp), %rcx
603	sarq	$47, %rcx
604	incq	%rcx
605	cmpq	$1, %rcx
606	ja	_syscall_post
607
608
609	SIMPLE_SYSCALL_POSTSYS(%r15, %r14, %bx)
610
611	movq	%r12, REGOFF_RAX(%rsp)
612	movq	%r13, REGOFF_RDX(%rsp)
613
614	/*
615	 * Clobber %r11 as we check CR0.TS.
616	 */
617	ASSERT_CR0TS_ZERO(%r11)
618
619	/*
620	 * Unlike other cases, because we need to restore the user stack pointer
621	 * before exiting the kernel we must clear the microarch state before
622	 * getting here. This should be safe because it means that the only
623	 * values on the bus after this are based on the user's registers and
624	 * potentially the addresses where we stored them. Given the constraints
625	 * of sysret, that's how it has to be.
626	 */
627	call	x86_md_clear
628
629	/*
630	 * To get back to userland, we need the return %rip in %rcx and
631	 * the return %rfl in %r11d.  The sysretq instruction also arranges
632	 * to fix up %cs and %ss; everything else is our responsibility.
633	 */
634	movq	REGOFF_RDI(%rsp), %rdi
635	movq	REGOFF_RSI(%rsp), %rsi
636	movq	REGOFF_RDX(%rsp), %rdx
637	/* %rcx used to restore %rip value */
638
639	movq	REGOFF_R8(%rsp), %r8
640	movq	REGOFF_R9(%rsp), %r9
641	movq	REGOFF_RAX(%rsp), %rax
642	movq	REGOFF_RBX(%rsp), %rbx
643
644	movq	REGOFF_RBP(%rsp), %rbp
645	movq	REGOFF_R10(%rsp), %r10
646	/* %r11 used to restore %rfl value */
647	movq	REGOFF_R12(%rsp), %r12
648
649	movq	REGOFF_R13(%rsp), %r13
650	movq	REGOFF_R14(%rsp), %r14
651	movq	REGOFF_R15(%rsp), %r15
652
653	movq	REGOFF_RIP(%rsp), %rcx
654	movl	REGOFF_RFL(%rsp), %r11d
655
656#if defined(__xpv)
657	addq	$REGOFF_RIP, %rsp
658#else
659	movq	REGOFF_RSP(%rsp), %rsp
660#endif
661
662        /*
663         * There can be no instructions between the ALTENTRY below and
664	 * SYSRET or we could end up breaking brand support. See label usage
665         * in sn1_brand_syscall_callback for an example.
666         */
667	ASSERT_UPCALL_MASK_IS_SET
668#if defined(__xpv)
669	SYSRETQ
670        ALTENTRY(nopop_sys_syscall_swapgs_sysretq)
671
672	/*
673	 * We can only get here after executing a brand syscall
674	 * interposition callback handler and simply need to
675	 * "sysretq" back to userland. On the hypervisor this
676	 * involves the iret hypercall which requires us to construct
677	 * just enough of the stack needed for the hypercall.
678	 * (rip, cs, rflags, rsp, ss).
679	 */
680	movq    %rsp, %gs:CPU_RTMP_RSP		/* save user's rsp */
681	movq	%gs:CPU_THREAD, %r11
682	movq	T_STACK(%r11), %rsp
683
684	movq	%rcx, REGOFF_RIP(%rsp)
685	movl	$UCS_SEL, REGOFF_CS(%rsp)
686	movq	%gs:CPU_RTMP_RSP, %r11
687	movq	%r11, REGOFF_RSP(%rsp)
688	pushfq
689	popq	%r11				/* hypercall enables ints */
690	movq	%r11, REGOFF_RFL(%rsp)
691	movl	$UDS_SEL, REGOFF_SS(%rsp)
692	addq	$REGOFF_RIP, %rsp
693	/*
694	 * XXPV: see comment in SYSRETQ definition for future optimization
695	 *       we could take.
696	 */
697	ASSERT_UPCALL_MASK_IS_SET
698	SYSRETQ
699#else
700        ALTENTRY(nopop_sys_syscall_swapgs_sysretq)
701	jmp	tr_sysretq
702#endif
703        /*NOTREACHED*/
704        SET_SIZE(nopop_sys_syscall_swapgs_sysretq)
705
706_syscall_pre:
707	call	pre_syscall
708	movl	%eax, %r12d
709	testl	%eax, %eax
710	jne	_syscall_post_call
711	/*
712	 * Didn't abort, so reload the syscall args and invoke the handler.
713	 */
714	movzwl	T_SYSNUM(%r15), %eax
715	jmp	_syscall_invoke
716
717_syscall_ill:
718	call	nosys
719	movq	%rax, %r12
720	movq	%rdx, %r13
721	jmp	_syscall_post_call
722
723_syscall_post:
724	STI
725	/*
726	 * Sigh, our optimism wasn't justified, put it back to LMS_SYSTEM
727	 * so that we can account for the extra work it takes us to finish.
728	 */
729	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
730_syscall_post_call:
731	movq	%r12, %rdi
732	movq	%r13, %rsi
733	call	post_syscall
734	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
735	jmp	_sys_rtt
736	SET_SIZE(sys_syscall)
737	SET_SIZE(brand_sys_syscall)
738
739	ENTRY_NP(brand_sys_syscall32)
740	SWAPGS				/* kernel gsbase */
741	XPV_TRAP_POP
742	BRAND_CALLBACK(BRAND_CB_SYSCALL32, BRAND_URET_FROM_REG(%rcx))
743	jmp	nopop_sys_syscall32
744
745	ALTENTRY(sys_syscall32)
746	SWAPGS				/* kernel gsbase */
747	XPV_TRAP_POP
748
749nopop_sys_syscall32:
750	movl	%esp, %r10d
751	movq	%gs:CPU_THREAD, %r15
752	movq	T_STACK(%r15), %rsp
753	movl	%eax, %eax
754
755	movl	$U32CS_SEL, REGOFF_CS(%rsp)
756	movl	%ecx, REGOFF_RIP(%rsp)		/* syscall: %rip -> %rcx */
757	movq	%r11, REGOFF_RFL(%rsp)		/* syscall: %rfl -> %r11d */
758	movq	%r10, REGOFF_RSP(%rsp)
759	movl	$UDS_SEL, REGOFF_SS(%rsp)
760
761_syscall32_save:
762	movl	%edi, REGOFF_RDI(%rsp)
763	movl	%esi, REGOFF_RSI(%rsp)
764	movl	%ebp, REGOFF_RBP(%rsp)
765	movl	%ebx, REGOFF_RBX(%rsp)
766	movl	%edx, REGOFF_RDX(%rsp)
767	movl	%ecx, REGOFF_RCX(%rsp)
768	movl	%eax, REGOFF_RAX(%rsp)		/* wrapper: sysc# -> %eax */
769	movq	$0, REGOFF_SAVFP(%rsp)
770	movq	$0, REGOFF_SAVPC(%rsp)
771
772	/*
773	 * Copy these registers here in case we end up stopped with
774	 * someone (like, say, /proc) messing with our register state.
775	 * We don't -restore- them unless we have to in update_sregs.
776	 *
777	 * Since userland -can't- change fsbase or gsbase directly,
778	 * we don't bother to capture them here.
779	 */
780	xorl	%ebx, %ebx
781	movw	%ds, %bx
782	movq	%rbx, REGOFF_DS(%rsp)
783	movw	%es, %bx
784	movq	%rbx, REGOFF_ES(%rsp)
785	movw	%fs, %bx
786	movq	%rbx, REGOFF_FS(%rsp)
787	movw	%gs, %bx
788	movq	%rbx, REGOFF_GS(%rsp)
789
790	/*
791	 * If we're trying to use TRAPTRACE though, I take that back: we're
792	 * probably debugging some problem in the SWAPGS logic and want to know
793	 * what the incoming gsbase was.
794	 *
795	 * Since we already did SWAPGS, record the KGSBASE.
796	 */
797#if defined(DEBUG) && defined(TRAPTRACE) && !defined(__xpv)
798	movl	$MSR_AMD_KGSBASE, %ecx
799	rdmsr
800	movl	%eax, REGOFF_GSBASE(%rsp)
801	movl	%edx, REGOFF_GSBASE+4(%rsp)
802#endif
803
804	/*
805	 * Application state saved in the regs structure on the stack
806	 * %eax is the syscall number
807	 * %rsp is the thread's stack, %r15 is curthread
808	 * REG_RSP(%rsp) is the user's stack
809	 */
810
811	SYSCALL_TRAPTRACE32($TT_SYSC)
812
813	movq	%rsp, %rbp
814
815	movq	T_LWP(%r15), %r14
816	ASSERT_NO_RUPDATE_PENDING(%r14)
817
818	ENABLE_INTR_FLAGS
819
820	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
821	movl	REGOFF_RAX(%rsp), %eax	/* (%rax damaged by mstate call) */
822
823	ASSERT_LWPTOREGS(%r14, %rsp)
824
825	incq	 %gs:CPU_STATS_SYS_SYSCALL
826
827	/*
828	 * Make some space for MAXSYSARGS (currently 8) 32-bit args placed
829	 * into 64-bit (long) arg slots, maintaining 16 byte alignment.  Or
830	 * more succinctly:
831	 *
832	 *	SA(MAXSYSARGS * sizeof (long)) == 64
833	 */
834#define	SYS_DROP	64			/* drop for args */
835	subq	$SYS_DROP, %rsp
836	movb	$LWP_SYS, LWP_STATE(%r14)
837	movq	%r15, %rdi
838	movq	%rsp, %rsi
839	call	syscall_entry
840
841	/*
842	 * Fetch the arguments copied onto the kernel stack and put
843	 * them in the right registers to invoke a C-style syscall handler.
844	 * %rax contains the handler address.
845	 *
846	 * Ideas for making all this go faster of course include simply
847	 * forcibly fetching 6 arguments from the user stack under lofault
848	 * protection, reverting to copyin_args only when watchpoints
849	 * are in effect.
850	 *
851	 * (If we do this, make sure that exec and libthread leave
852	 * enough space at the top of the stack to ensure that we'll
853	 * never do a fetch from an invalid page.)
854	 *
855	 * Lots of ideas here, but they won't really help with bringup B-)
856	 * Correctness can't wait, performance can wait a little longer ..
857	 */
858
859	movq	%rax, %rbx
860	movl	0(%rsp), %edi
861	movl	8(%rsp), %esi
862	movl	0x10(%rsp), %edx
863	movl	0x18(%rsp), %ecx
864	movl	0x20(%rsp), %r8d
865	movl	0x28(%rsp), %r9d
866
867	movq	SY_CALLC(%rbx), %rax
868	INDIRECT_CALL_REG(rax)
869
870	movq	%rbp, %rsp	/* pop the args */
871
872	/*
873	 * amd64 syscall handlers -always- return a 64-bit value in %rax.
874	 * On the 32-bit kernel, they always return that value in %eax:%edx
875	 * as required by the 32-bit ABI.
876	 *
877	 * Simulate the same behaviour by unconditionally splitting the
878	 * return value in the same way.
879	 */
880	movq	%rax, %r13
881	shrq	$32, %r13	/* upper 32-bits into %edx */
882	movl	%eax, %r12d	/* lower 32-bits into %eax */
883
884	/*
885	 * Optimistically assume that there's no post-syscall
886	 * work to do.  (This is to avoid having to call syscall_mstate()
887	 * with interrupts disabled)
888	 */
889	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
890
891	/*
892	 * We must protect ourselves from being descheduled here;
893	 * If we were, and we ended up on another cpu, or another
894	 * lwp got in ahead of us, it could change the segment
895	 * registers without us noticing before we return to userland.
896	 */
897	CLI(%r14)
898	CHECK_POSTSYS_NE(%r15, %r14, %ebx)
899	jne	_full_syscall_postsys32
900	SIMPLE_SYSCALL_POSTSYS(%r15, %r14, %bx)
901
902	/*
903	 * Clobber %r11 as we check CR0.TS.
904	 */
905	ASSERT_CR0TS_ZERO(%r11)
906
907	/*
908	 * Unlike other cases, because we need to restore the user stack pointer
909	 * before exiting the kernel we must clear the microarch state before
910	 * getting here. This should be safe because it means that the only
911	 * values on the bus after this are based on the user's registers and
912	 * potentially the addresses where we stored them. Given the constraints
913	 * of sysret, that's how it has to be.
914	 */
915	call	x86_md_clear
916
917	/*
918	 * To get back to userland, we need to put the return %rip in %rcx and
919	 * the return %rfl in %r11d.  The sysret instruction also arranges
920	 * to fix up %cs and %ss; everything else is our responsibility.
921	 */
922
923	movl	%r12d, %eax			/* %eax: rval1 */
924	movl	REGOFF_RBX(%rsp), %ebx
925	/* %ecx used for return pointer */
926	movl	%r13d, %edx			/* %edx: rval2 */
927	movl	REGOFF_RBP(%rsp), %ebp
928	movl	REGOFF_RSI(%rsp), %esi
929	movl	REGOFF_RDI(%rsp), %edi
930
931	movl	REGOFF_RFL(%rsp), %r11d		/* %r11 -> eflags */
932	movl	REGOFF_RIP(%rsp), %ecx		/* %ecx -> %eip */
933	movl	REGOFF_RSP(%rsp), %esp
934
935	ASSERT_UPCALL_MASK_IS_SET
936        ALTENTRY(nopop_sys_syscall32_swapgs_sysretl)
937	jmp	tr_sysretl
938        SET_SIZE(nopop_sys_syscall32_swapgs_sysretl)
939	/*NOTREACHED*/
940
941_full_syscall_postsys32:
942	STI
943	/*
944	 * Sigh, our optimism wasn't justified, put it back to LMS_SYSTEM
945	 * so that we can account for the extra work it takes us to finish.
946	 */
947	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
948	movq	%r15, %rdi
949	movq	%r12, %rsi			/* rval1 - %eax */
950	movq	%r13, %rdx			/* rval2 - %edx */
951	call	syscall_exit
952	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
953	jmp	_sys_rtt
954	SET_SIZE(sys_syscall32)
955	SET_SIZE(brand_sys_syscall32)
956
957/*
958 * System call handler via the sysenter instruction
959 * Used only for 32-bit system calls on the 64-bit kernel.
960 *
961 * The caller in userland has arranged that:
962 *
963 * -	%eax contains the syscall number
964 * -	%ecx contains the user %esp
965 * -	%edx contains the return %eip
966 * -	the user stack contains the args to the syscall
967 *
968 * Hardware and (privileged) initialization code have arranged that by
969 * the time the sysenter instructions completes:
970 *
971 * - %rip is pointing to sys_sysenter (below).
972 * - %cs and %ss are set to kernel text and stack (data) selectors.
973 * - %rsp is pointing at the lwp's stack
974 * - interrupts have been disabled.
975 *
976 * Note that we are unable to return both "rvals" to userland with
977 * this call, as %edx is used by the sysexit instruction.
978 *
979 * One final complication in this routine is its interaction with
980 * single-stepping in a debugger.  For most of the system call mechanisms, the
981 * CPU automatically clears the single-step flag before we enter the kernel.
982 * The sysenter mechanism does not clear the flag, so a user single-stepping
983 * through a libc routine may suddenly find themself single-stepping through the
984 * kernel.  To detect this, kmdb and trap() both compare the trap %pc to the
985 * [brand_]sys_enter addresses on each single-step trap.  If it finds that we
986 * have single-stepped to a sysenter entry point, it explicitly clears the flag
987 * and executes the sys_sysenter routine.
988 *
989 * One final complication in this final complication is the fact that we have
990 * two different entry points for sysenter: brand_sys_sysenter and sys_sysenter.
991 * If we enter at brand_sys_sysenter and start single-stepping through the
992 * kernel with kmdb, we will eventually hit the instruction at sys_sysenter.
993 * kmdb cannot distinguish between that valid single-step and the undesirable
994 * one mentioned above.  To avoid this situation, we simply add a jump over the
995 * instruction at sys_sysenter to make it impossible to single-step to it.
996 */
997
998	ENTRY_NP(brand_sys_sysenter)
999	SWAPGS				/* kernel gsbase */
1000	ALTENTRY(_brand_sys_sysenter_post_swapgs)
1001
1002	BRAND_CALLBACK(BRAND_CB_SYSENTER, BRAND_URET_FROM_REG(%rdx))
1003	/*
1004	 * Jump over sys_sysenter to allow single-stepping as described
1005	 * above.
1006	 */
1007	jmp	_sys_sysenter_post_swapgs
1008
1009	ALTENTRY(sys_sysenter)
1010	SWAPGS				/* kernel gsbase */
1011	ALTENTRY(_sys_sysenter_post_swapgs)
1012
1013	movq	%gs:CPU_THREAD, %r15
1014
1015	movl	$U32CS_SEL, REGOFF_CS(%rsp)
1016	movl	%ecx, REGOFF_RSP(%rsp)		/* wrapper: %esp -> %ecx */
1017	movl	%edx, REGOFF_RIP(%rsp)		/* wrapper: %eip -> %edx */
1018	/*
1019	 * NOTE: none of the instructions that run before we get here should
1020	 * clobber bits in (R)FLAGS! This includes the kpti trampoline.
1021	 */
1022	pushfq
1023	popq	%r10
1024	movl	$UDS_SEL, REGOFF_SS(%rsp)
1025
1026	/*
1027	 * Set the interrupt flag before storing the flags to the
1028	 * flags image on the stack so we can return to user with
1029	 * interrupts enabled if we return via sys_rtt_syscall32
1030	 */
1031	orq	$PS_IE, %r10
1032	movq	%r10, REGOFF_RFL(%rsp)
1033
1034	movl	%edi, REGOFF_RDI(%rsp)
1035	movl	%esi, REGOFF_RSI(%rsp)
1036	movl	%ebp, REGOFF_RBP(%rsp)
1037	movl	%ebx, REGOFF_RBX(%rsp)
1038	movl	%edx, REGOFF_RDX(%rsp)
1039	movl	%ecx, REGOFF_RCX(%rsp)
1040	movl	%eax, REGOFF_RAX(%rsp)		/* wrapper: sysc# -> %eax */
1041	movq	$0, REGOFF_SAVFP(%rsp)
1042	movq	$0, REGOFF_SAVPC(%rsp)
1043
1044	/*
1045	 * Copy these registers here in case we end up stopped with
1046	 * someone (like, say, /proc) messing with our register state.
1047	 * We don't -restore- them unless we have to in update_sregs.
1048	 *
1049	 * Since userland -can't- change fsbase or gsbase directly,
1050	 * we don't bother to capture them here.
1051	 */
1052	xorl	%ebx, %ebx
1053	movw	%ds, %bx
1054	movq	%rbx, REGOFF_DS(%rsp)
1055	movw	%es, %bx
1056	movq	%rbx, REGOFF_ES(%rsp)
1057	movw	%fs, %bx
1058	movq	%rbx, REGOFF_FS(%rsp)
1059	movw	%gs, %bx
1060	movq	%rbx, REGOFF_GS(%rsp)
1061
1062	/*
1063	 * If we're trying to use TRAPTRACE though, I take that back: we're
1064	 * probably debugging some problem in the SWAPGS logic and want to know
1065	 * what the incoming gsbase was.
1066	 *
1067	 * Since we already did SWAPGS, record the KGSBASE.
1068	 */
1069#if defined(DEBUG) && defined(TRAPTRACE) && !defined(__xpv)
1070	movl	$MSR_AMD_KGSBASE, %ecx
1071	rdmsr
1072	movl	%eax, REGOFF_GSBASE(%rsp)
1073	movl	%edx, REGOFF_GSBASE+4(%rsp)
1074#endif
1075
1076	/*
1077	 * Application state saved in the regs structure on the stack
1078	 * %eax is the syscall number
1079	 * %rsp is the thread's stack, %r15 is curthread
1080	 * REG_RSP(%rsp) is the user's stack
1081	 */
1082
1083	SYSCALL_TRAPTRACE($TT_SYSENTER)
1084
1085	movq	%rsp, %rbp
1086
1087	movq	T_LWP(%r15), %r14
1088	ASSERT_NO_RUPDATE_PENDING(%r14)
1089
1090	ENABLE_INTR_FLAGS
1091
1092	/*
1093	 * Catch 64-bit process trying to issue sysenter instruction
1094	 * on Nocona based systems.
1095	 */
1096	movq	LWP_PROCP(%r14), %rax
1097	cmpq	$DATAMODEL_ILP32, P_MODEL(%rax)
1098	je	7f
1099
1100	/*
1101	 * For a non-32-bit process, simulate a #ud, since that's what
1102	 * native hardware does.  The traptrace entry (above) will
1103	 * let you know what really happened.
1104	 */
1105	movq	$T_ILLINST, REGOFF_TRAPNO(%rsp)
1106	movq	REGOFF_CS(%rsp), %rdi
1107	movq	%rdi, REGOFF_ERR(%rsp)
1108	movq	%rsp, %rdi
1109	movq	REGOFF_RIP(%rsp), %rsi
1110	movl	%gs:CPU_ID, %edx
1111	call	trap
1112	jmp	_sys_rtt
11137:
1114
1115	MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
1116	movl	REGOFF_RAX(%rsp), %eax	/* (%rax damaged by mstate calls) */
1117
1118	ASSERT_LWPTOREGS(%r14, %rsp)
1119
1120	incq	%gs:CPU_STATS_SYS_SYSCALL
1121
1122	/*
1123	 * Make some space for MAXSYSARGS (currently 8) 32-bit args
1124	 * placed into 64-bit (long) arg slots, plus one 64-bit
1125	 * (long) arg count, maintaining 16 byte alignment.
1126	 */
1127	subq	$SYS_DROP, %rsp
1128	movb	$LWP_SYS, LWP_STATE(%r14)
1129	movq	%r15, %rdi
1130	movq	%rsp, %rsi
1131	call	syscall_entry
1132
1133	/*
1134	 * Fetch the arguments copied onto the kernel stack and put
1135	 * them in the right registers to invoke a C-style syscall handler.
1136	 * %rax contains the handler address.
1137	 */
1138	movq	%rax, %rbx
1139	movl	0(%rsp), %edi
1140	movl	8(%rsp), %esi
1141	movl	0x10(%rsp), %edx
1142	movl	0x18(%rsp), %ecx
1143	movl	0x20(%rsp), %r8d
1144	movl	0x28(%rsp), %r9d
1145
1146	movq	SY_CALLC(%rbx), %rax
1147	INDIRECT_CALL_REG(rax)
1148
1149	movq	%rbp, %rsp	/* pop the args */
1150
1151	/*
1152	 * amd64 syscall handlers -always- return a 64-bit value in %rax.
1153	 * On the 32-bit kernel, the always return that value in %eax:%edx
1154	 * as required by the 32-bit ABI.
1155	 *
1156	 * Simulate the same behaviour by unconditionally splitting the
1157	 * return value in the same way.
1158	 */
1159	movq	%rax, %r13
1160	shrq	$32, %r13	/* upper 32-bits into %edx */
1161	movl	%eax, %r12d	/* lower 32-bits into %eax */
1162
1163	/*
1164	 * Optimistically assume that there's no post-syscall
1165	 * work to do.  (This is to avoid having to call syscall_mstate()
1166	 * with interrupts disabled)
1167	 */
1168	MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
1169
1170	/*
1171	 * We must protect ourselves from being descheduled here;
1172	 * If we were, and we ended up on another cpu, or another
1173	 * lwp got int ahead of us, it could change the segment
1174	 * registers without us noticing before we return to userland.
1175	 *
1176	 * This cli is undone in the tr_sysexit trampoline code.
1177	 */
1178	cli
1179	CHECK_POSTSYS_NE(%r15, %r14, %ebx)
1180	jne	_full_syscall_postsys32
1181	SIMPLE_SYSCALL_POSTSYS(%r15, %r14, %bx)
1182
1183	/*
1184	 * To get back to userland, load up the 32-bit registers and
1185	 * sysexit back where we came from.
1186	 */
1187
1188	/*
1189	 * Interrupts will be turned on by the 'sti' executed just before
1190	 * sysexit.  The following ensures that restoring the user's rflags
1191	 * doesn't enable interrupts too soon.
1192	 */
1193	andq	$_BITNOT(PS_IE), REGOFF_RFL(%rsp)
1194
1195	/*
1196	 * Clobber %r11 as we check CR0.TS.
1197	 */
1198	ASSERT_CR0TS_ZERO(%r11)
1199
1200	/*
1201	 * (There's no point in loading up %edx because the sysexit
1202	 * mechanism smashes it.)
1203	 */
1204	movl	%r12d, %eax
1205	movl	REGOFF_RBX(%rsp), %ebx
1206	movl	REGOFF_RBP(%rsp), %ebp
1207	movl	REGOFF_RSI(%rsp), %esi
1208	movl	REGOFF_RDI(%rsp), %edi
1209
1210	movl	REGOFF_RIP(%rsp), %edx	/* sysexit: %edx -> %eip */
1211	pushq	REGOFF_RFL(%rsp)
1212	popfq
1213	movl	REGOFF_RSP(%rsp), %ecx	/* sysexit: %ecx -> %esp */
1214        ALTENTRY(sys_sysenter_swapgs_sysexit)
1215	call	x86_md_clear
1216	jmp	tr_sysexit
1217	SET_SIZE(sys_sysenter_swapgs_sysexit)
1218	SET_SIZE(sys_sysenter)
1219	SET_SIZE(_sys_sysenter_post_swapgs)
1220	SET_SIZE(brand_sys_sysenter)
1221
1222/*
1223 * This is the destination of the "int $T_SYSCALLINT" interrupt gate, used by
1224 * the generic i386 libc to do system calls. We do a small amount of setup
1225 * before jumping into the existing sys_syscall32 path.
1226 */
1227
1228	ENTRY_NP(brand_sys_syscall_int)
1229	SWAPGS				/* kernel gsbase */
1230	XPV_TRAP_POP
1231	call	smap_enable
1232	BRAND_CALLBACK(BRAND_CB_INT91, BRAND_URET_FROM_INTR_STACK())
1233	jmp	nopop_syscall_int
1234
1235	ALTENTRY(sys_syscall_int)
1236	SWAPGS				/* kernel gsbase */
1237	XPV_TRAP_POP
1238	call	smap_enable
1239
1240nopop_syscall_int:
1241	movq	%gs:CPU_THREAD, %r15
1242	movq	T_STACK(%r15), %rsp
1243	movl	%eax, %eax
1244	/*
1245	 * Set t_post_sys on this thread to force ourselves out via the slow
1246	 * path. It might be possible at some later date to optimize this out
1247	 * and use a faster return mechanism.
1248	 */
1249	movb	$1, T_POST_SYS(%r15)
1250	CLEAN_CS
1251	jmp	_syscall32_save
1252	/*
1253	 * There should be no instructions between this label and SWAPGS/IRET
1254	 * or we could end up breaking branded zone support. See the usage of
1255	 * this label in lx_brand_int80_callback and sn1_brand_int91_callback
1256	 * for examples.
1257	 *
1258	 * We want to swapgs to maintain the invariant that all entries into
1259	 * tr_iret_user are done on the user gsbase.
1260	 */
1261	ALTENTRY(sys_sysint_swapgs_iret)
1262	call	x86_md_clear
1263	SWAPGS
1264	jmp	tr_iret_user
1265	/*NOTREACHED*/
1266	SET_SIZE(sys_sysint_swapgs_iret)
1267	SET_SIZE(sys_syscall_int)
1268	SET_SIZE(brand_sys_syscall_int)
1269
1270/*
1271 * Legacy 32-bit applications and old libc implementations do lcalls;
1272 * we should never get here because the LDT entry containing the syscall
1273 * segment descriptor has the "segment present" bit cleared, which means
1274 * we end up processing those system calls in trap() via a not-present trap.
1275 *
1276 * We do it this way because a call gate unhelpfully does -nothing- to the
1277 * interrupt flag bit, so an interrupt can run us just after the lcall
1278 * completes, but just before the swapgs takes effect.   Thus the INTR_PUSH and
1279 * INTR_POP paths would have to be slightly more complex to dance around
1280 * this problem, and end up depending explicitly on the first
1281 * instruction of this handler being either swapgs or cli.
1282 */
1283
1284	ENTRY_NP(sys_lcall32)
1285	SWAPGS				/* kernel gsbase */
1286	pushq	$0
1287	pushq	%rbp
1288	movq	%rsp, %rbp
1289	leaq	__lcall_panic_str(%rip), %rdi
1290	xorl	%eax, %eax
1291	call	panic
1292	SET_SIZE(sys_lcall32)
1293
1294__lcall_panic_str:
1295	.string	"sys_lcall32: shouldn't be here!"
1296
1297/*
1298 * Declare a uintptr_t which covers the entire pc range of syscall
1299 * handlers for the stack walkers that need this.
1300 */
1301	.align	CPTRSIZE
1302	.globl	_allsyscalls_size
1303	.type	_allsyscalls_size, @object
1304_allsyscalls_size:
1305	.NWORD	. - _allsyscalls
1306	SET_SIZE(_allsyscalls_size)
1307
1308/*
1309 * These are the thread context handlers for lwps using sysenter/sysexit.
1310 */
1311
1312	/*
1313	 * setting this value to zero as we switch away causes the
1314	 * stack-pointer-on-sysenter to be NULL, ensuring that we
1315	 * don't silently corrupt another (preempted) thread stack
1316	 * when running an lwp that (somehow) didn't get sep_restore'd
1317	 */
1318	ENTRY_NP(sep_save)
1319	xorl	%edx, %edx
1320	xorl	%eax, %eax
1321	movl	$MSR_INTC_SEP_ESP, %ecx
1322	wrmsr
1323	ret
1324	SET_SIZE(sep_save)
1325
1326	/*
1327	 * Update the kernel stack pointer as we resume onto this cpu.
1328	 */
1329	ENTRY_NP(sep_restore)
1330	movq	%rdi, %rdx
1331	shrq	$32, %rdx
1332	movl	%edi, %eax
1333	movl	$MSR_INTC_SEP_ESP, %ecx
1334	wrmsr
1335	ret
1336	SET_SIZE(sep_restore)
1337