xref: /openbmc/linux/arch/ia64/kernel/entry.S (revision d9fd5a71)
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 * arch/ia64/kernel/entry.S
4 *
5 * Kernel entry points.
6 *
7 * Copyright (C) 1998-2003, 2005 Hewlett-Packard Co
8 *	David Mosberger-Tang <davidm@hpl.hp.com>
9 * Copyright (C) 1999, 2002-2003
10 *	Asit Mallick <Asit.K.Mallick@intel.com>
11 * 	Don Dugger <Don.Dugger@intel.com>
12 *	Suresh Siddha <suresh.b.siddha@intel.com>
13 *	Fenghua Yu <fenghua.yu@intel.com>
14 * Copyright (C) 1999 VA Linux Systems
15 * Copyright (C) 1999 Walt Drummond <drummond@valinux.com>
16 */
17/*
18 * ia64_switch_to now places correct virtual mapping in in TR2 for
19 * kernel stack. This allows us to handle interrupts without changing
20 * to physical mode.
21 *
22 * Jonathan Nicklin	<nicklin@missioncriticallinux.com>
23 * Patrick O'Rourke	<orourke@missioncriticallinux.com>
24 * 11/07/2000
25 */
26/*
27 * Copyright (c) 2008 Isaku Yamahata <yamahata at valinux co jp>
28 *                    VA Linux Systems Japan K.K.
29 *                    pv_ops.
30 */
31/*
32 * Global (preserved) predicate usage on syscall entry/exit path:
33 *
34 *	pKStk:		See entry.h.
35 *	pUStk:		See entry.h.
36 *	pSys:		See entry.h.
37 *	pNonSys:	!pSys
38 */
39
40
41#include <linux/pgtable.h>
42#include <asm/asmmacro.h>
43#include <asm/cache.h>
44#include <asm/errno.h>
45#include <asm/kregs.h>
46#include <asm/asm-offsets.h>
47#include <asm/percpu.h>
48#include <asm/processor.h>
49#include <asm/thread_info.h>
50#include <asm/unistd.h>
51#include <asm/ftrace.h>
52#include <asm/export.h>
53
54#include "minstate.h"
55
56	/*
57	 * execve() is special because in case of success, we need to
58	 * setup a null register window frame.
59	 */
60ENTRY(ia64_execve)
61	/*
62	 * Allocate 8 input registers since ptrace() may clobber them
63	 */
64	.prologue ASM_UNW_PRLG_RP|ASM_UNW_PRLG_PFS, ASM_UNW_PRLG_GRSAVE(8)
65	alloc loc1=ar.pfs,8,2,3,0
66	mov loc0=rp
67	.body
68	mov out0=in0			// filename
69	;;				// stop bit between alloc and call
70	mov out1=in1			// argv
71	mov out2=in2			// envp
72	br.call.sptk.many rp=sys_execve
73.ret0:
74	cmp4.ge p6,p7=r8,r0
75	mov ar.pfs=loc1			// restore ar.pfs
76	sxt4 r8=r8			// return 64-bit result
77	;;
78	stf.spill [sp]=f0
79	mov rp=loc0
80(p6)	mov ar.pfs=r0			// clear ar.pfs on success
81(p7)	br.ret.sptk.many rp
82
83	/*
84	 * In theory, we'd have to zap this state only to prevent leaking of
85	 * security sensitive state (e.g., if current->mm->dumpable is zero).  However,
86	 * this executes in less than 20 cycles even on Itanium, so it's not worth
87	 * optimizing for...).
88	 */
89	mov ar.unat=0; 		mov ar.lc=0
90	mov r4=0;		mov f2=f0;		mov b1=r0
91	mov r5=0;		mov f3=f0;		mov b2=r0
92	mov r6=0;		mov f4=f0;		mov b3=r0
93	mov r7=0;		mov f5=f0;		mov b4=r0
94	ldf.fill f12=[sp];	mov f13=f0;		mov b5=r0
95	ldf.fill f14=[sp];	ldf.fill f15=[sp];	mov f16=f0
96	ldf.fill f17=[sp];	ldf.fill f18=[sp];	mov f19=f0
97	ldf.fill f20=[sp];	ldf.fill f21=[sp];	mov f22=f0
98	ldf.fill f23=[sp];	ldf.fill f24=[sp];	mov f25=f0
99	ldf.fill f26=[sp];	ldf.fill f27=[sp];	mov f28=f0
100	ldf.fill f29=[sp];	ldf.fill f30=[sp];	mov f31=f0
101	br.ret.sptk.many rp
102END(ia64_execve)
103
104/*
105 * sys_clone2(u64 flags, u64 ustack_base, u64 ustack_size, u64 parent_tidptr, u64 child_tidptr,
106 *	      u64 tls)
107 */
108GLOBAL_ENTRY(sys_clone2)
109	/*
110	 * Allocate 8 input registers since ptrace() may clobber them
111	 */
112	.prologue ASM_UNW_PRLG_RP|ASM_UNW_PRLG_PFS, ASM_UNW_PRLG_GRSAVE(8)
113	alloc r16=ar.pfs,8,2,6,0
114	DO_SAVE_SWITCH_STACK
115	mov loc0=rp
116	mov loc1=r16                             // save ar.pfs across ia64_clone
117	.body
118	mov out0=in0
119	mov out1=in1
120	mov out2=in2
121	mov out3=in3
122	mov out4=in4
123	mov out5=in5
124	br.call.sptk.many rp=ia64_clone
125.ret1:	.restore sp
126	adds sp=IA64_SWITCH_STACK_SIZE,sp	// pop the switch stack
127	mov ar.pfs=loc1
128	mov rp=loc0
129	br.ret.sptk.many rp
130END(sys_clone2)
131
132/*
133 * sys_clone(u64 flags, u64 ustack_base, u64 parent_tidptr, u64 child_tidptr, u64 tls)
134 *	Deprecated.  Use sys_clone2() instead.
135 */
136GLOBAL_ENTRY(sys_clone)
137	/*
138	 * Allocate 8 input registers since ptrace() may clobber them
139	 */
140	.prologue ASM_UNW_PRLG_RP|ASM_UNW_PRLG_PFS, ASM_UNW_PRLG_GRSAVE(8)
141	alloc r16=ar.pfs,8,2,6,0
142	DO_SAVE_SWITCH_STACK
143	mov loc0=rp
144	mov loc1=r16                             // save ar.pfs across ia64_clone
145	.body
146	mov out0=in0
147	mov out1=in1
148	mov out2=16				// stacksize (compensates for 16-byte scratch area)
149	mov out3=in3
150	mov out4=in4
151	mov out5=in5
152	br.call.sptk.many rp=ia64_clone
153.ret2:	.restore sp
154	adds sp=IA64_SWITCH_STACK_SIZE,sp	// pop the switch stack
155	mov ar.pfs=loc1
156	mov rp=loc0
157	br.ret.sptk.many rp
158END(sys_clone)
159
160/*
161 * prev_task <- ia64_switch_to(struct task_struct *next)
162 *	With Ingo's new scheduler, interrupts are disabled when this routine gets
163 *	called.  The code starting at .map relies on this.  The rest of the code
164 *	doesn't care about the interrupt masking status.
165 */
166GLOBAL_ENTRY(ia64_switch_to)
167	.prologue
168	alloc r16=ar.pfs,1,0,0,0
169	DO_SAVE_SWITCH_STACK
170	.body
171
172	adds r22=IA64_TASK_THREAD_KSP_OFFSET,r13
173	movl r25=init_task
174	mov r27=IA64_KR(CURRENT_STACK)
175	adds r21=IA64_TASK_THREAD_KSP_OFFSET,in0
176	dep r20=0,in0,61,3		// physical address of "next"
177	;;
178	st8 [r22]=sp			// save kernel stack pointer of old task
179	shr.u r26=r20,IA64_GRANULE_SHIFT
180	cmp.eq p7,p6=r25,in0
181	;;
182	/*
183	 * If we've already mapped this task's page, we can skip doing it again.
184	 */
185(p6)	cmp.eq p7,p6=r26,r27
186(p6)	br.cond.dpnt .map
187	;;
188.done:
189	ld8 sp=[r21]			// load kernel stack pointer of new task
190	MOV_TO_KR(CURRENT, in0, r8, r9)		// update "current" application register
191	mov r8=r13			// return pointer to previously running task
192	mov r13=in0			// set "current" pointer
193	;;
194	DO_LOAD_SWITCH_STACK
195
196#ifdef CONFIG_SMP
197	sync.i				// ensure "fc"s done by this CPU are visible on other CPUs
198#endif
199	br.ret.sptk.many rp		// boogie on out in new context
200
201.map:
202	RSM_PSR_IC(r25)			// interrupts (psr.i) are already disabled here
203	movl r25=PAGE_KERNEL
204	;;
205	srlz.d
206	or r23=r25,r20			// construct PA | page properties
207	mov r25=IA64_GRANULE_SHIFT<<2
208	;;
209	MOV_TO_ITIR(p0, r25, r8)
210	MOV_TO_IFA(in0, r8)		// VA of next task...
211	;;
212	mov r25=IA64_TR_CURRENT_STACK
213	MOV_TO_KR(CURRENT_STACK, r26, r8, r9)	// remember last page we mapped...
214	;;
215	itr.d dtr[r25]=r23		// wire in new mapping...
216	SSM_PSR_IC_AND_SRLZ_D(r8, r9)	// reenable the psr.ic bit
217	br.cond.sptk .done
218END(ia64_switch_to)
219
220/*
221 * Note that interrupts are enabled during save_switch_stack and load_switch_stack.  This
222 * means that we may get an interrupt with "sp" pointing to the new kernel stack while
223 * ar.bspstore is still pointing to the old kernel backing store area.  Since ar.rsc,
224 * ar.rnat, ar.bsp, and ar.bspstore are all preserved by interrupts, this is not a
225 * problem.  Also, we don't need to specify unwind information for preserved registers
226 * that are not modified in save_switch_stack as the right unwind information is already
227 * specified at the call-site of save_switch_stack.
228 */
229
230/*
231 * save_switch_stack:
232 *	- r16 holds ar.pfs
233 *	- b7 holds address to return to
234 *	- rp (b0) holds return address to save
235 */
236GLOBAL_ENTRY(save_switch_stack)
237	.prologue
238	.altrp b7
239	flushrs			// flush dirty regs to backing store (must be first in insn group)
240	.save @priunat,r17
241	mov r17=ar.unat		// preserve caller's
242	.body
243#ifdef CONFIG_ITANIUM
244	adds r2=16+128,sp
245	adds r3=16+64,sp
246	adds r14=SW(R4)+16,sp
247	;;
248	st8.spill [r14]=r4,16		// spill r4
249	lfetch.fault.excl.nt1 [r3],128
250	;;
251	lfetch.fault.excl.nt1 [r2],128
252	lfetch.fault.excl.nt1 [r3],128
253	;;
254	lfetch.fault.excl [r2]
255	lfetch.fault.excl [r3]
256	adds r15=SW(R5)+16,sp
257#else
258	add r2=16+3*128,sp
259	add r3=16,sp
260	add r14=SW(R4)+16,sp
261	;;
262	st8.spill [r14]=r4,SW(R6)-SW(R4)	// spill r4 and prefetch offset 0x1c0
263	lfetch.fault.excl.nt1 [r3],128	//		prefetch offset 0x010
264	;;
265	lfetch.fault.excl.nt1 [r3],128	//		prefetch offset 0x090
266	lfetch.fault.excl.nt1 [r2],128	//		prefetch offset 0x190
267	;;
268	lfetch.fault.excl.nt1 [r3]	//		prefetch offset 0x110
269	lfetch.fault.excl.nt1 [r2]	//		prefetch offset 0x210
270	adds r15=SW(R5)+16,sp
271#endif
272	;;
273	st8.spill [r15]=r5,SW(R7)-SW(R5)	// spill r5
274	mov.m ar.rsc=0			// put RSE in mode: enforced lazy, little endian, pl 0
275	add r2=SW(F2)+16,sp		// r2 = &sw->f2
276	;;
277	st8.spill [r14]=r6,SW(B0)-SW(R6)	// spill r6
278	mov.m r18=ar.fpsr		// preserve fpsr
279	add r3=SW(F3)+16,sp		// r3 = &sw->f3
280	;;
281	stf.spill [r2]=f2,32
282	mov.m r19=ar.rnat
283	mov r21=b0
284
285	stf.spill [r3]=f3,32
286	st8.spill [r15]=r7,SW(B2)-SW(R7)	// spill r7
287	mov r22=b1
288	;;
289	// since we're done with the spills, read and save ar.unat:
290	mov.m r29=ar.unat
291	mov.m r20=ar.bspstore
292	mov r23=b2
293	stf.spill [r2]=f4,32
294	stf.spill [r3]=f5,32
295	mov r24=b3
296	;;
297	st8 [r14]=r21,SW(B1)-SW(B0)		// save b0
298	st8 [r15]=r23,SW(B3)-SW(B2)		// save b2
299	mov r25=b4
300	mov r26=b5
301	;;
302	st8 [r14]=r22,SW(B4)-SW(B1)		// save b1
303	st8 [r15]=r24,SW(AR_PFS)-SW(B3)		// save b3
304	mov r21=ar.lc		// I-unit
305	stf.spill [r2]=f12,32
306	stf.spill [r3]=f13,32
307	;;
308	st8 [r14]=r25,SW(B5)-SW(B4)		// save b4
309	st8 [r15]=r16,SW(AR_LC)-SW(AR_PFS)	// save ar.pfs
310	stf.spill [r2]=f14,32
311	stf.spill [r3]=f15,32
312	;;
313	st8 [r14]=r26				// save b5
314	st8 [r15]=r21				// save ar.lc
315	stf.spill [r2]=f16,32
316	stf.spill [r3]=f17,32
317	;;
318	stf.spill [r2]=f18,32
319	stf.spill [r3]=f19,32
320	;;
321	stf.spill [r2]=f20,32
322	stf.spill [r3]=f21,32
323	;;
324	stf.spill [r2]=f22,32
325	stf.spill [r3]=f23,32
326	;;
327	stf.spill [r2]=f24,32
328	stf.spill [r3]=f25,32
329	;;
330	stf.spill [r2]=f26,32
331	stf.spill [r3]=f27,32
332	;;
333	stf.spill [r2]=f28,32
334	stf.spill [r3]=f29,32
335	;;
336	stf.spill [r2]=f30,SW(AR_UNAT)-SW(F30)
337	stf.spill [r3]=f31,SW(PR)-SW(F31)
338	add r14=SW(CALLER_UNAT)+16,sp
339	;;
340	st8 [r2]=r29,SW(AR_RNAT)-SW(AR_UNAT)	// save ar.unat
341	st8 [r14]=r17,SW(AR_FPSR)-SW(CALLER_UNAT) // save caller_unat
342	mov r21=pr
343	;;
344	st8 [r2]=r19,SW(AR_BSPSTORE)-SW(AR_RNAT) // save ar.rnat
345	st8 [r3]=r21				// save predicate registers
346	;;
347	st8 [r2]=r20				// save ar.bspstore
348	st8 [r14]=r18				// save fpsr
349	mov ar.rsc=3		// put RSE back into eager mode, pl 0
350	br.cond.sptk.many b7
351END(save_switch_stack)
352
353/*
354 * load_switch_stack:
355 *	- "invala" MUST be done at call site (normally in DO_LOAD_SWITCH_STACK)
356 *	- b7 holds address to return to
357 *	- must not touch r8-r11
358 */
359GLOBAL_ENTRY(load_switch_stack)
360	.prologue
361	.altrp b7
362
363	.body
364	lfetch.fault.nt1 [sp]
365	adds r2=SW(AR_BSPSTORE)+16,sp
366	adds r3=SW(AR_UNAT)+16,sp
367	mov ar.rsc=0						// put RSE into enforced lazy mode
368	adds r14=SW(CALLER_UNAT)+16,sp
369	adds r15=SW(AR_FPSR)+16,sp
370	;;
371	ld8 r27=[r2],(SW(B0)-SW(AR_BSPSTORE))	// bspstore
372	ld8 r29=[r3],(SW(B1)-SW(AR_UNAT))	// unat
373	;;
374	ld8 r21=[r2],16		// restore b0
375	ld8 r22=[r3],16		// restore b1
376	;;
377	ld8 r23=[r2],16		// restore b2
378	ld8 r24=[r3],16		// restore b3
379	;;
380	ld8 r25=[r2],16		// restore b4
381	ld8 r26=[r3],16		// restore b5
382	;;
383	ld8 r16=[r2],(SW(PR)-SW(AR_PFS))	// ar.pfs
384	ld8 r17=[r3],(SW(AR_RNAT)-SW(AR_LC))	// ar.lc
385	;;
386	ld8 r28=[r2]		// restore pr
387	ld8 r30=[r3]		// restore rnat
388	;;
389	ld8 r18=[r14],16	// restore caller's unat
390	ld8 r19=[r15],24	// restore fpsr
391	;;
392	ldf.fill f2=[r14],32
393	ldf.fill f3=[r15],32
394	;;
395	ldf.fill f4=[r14],32
396	ldf.fill f5=[r15],32
397	;;
398	ldf.fill f12=[r14],32
399	ldf.fill f13=[r15],32
400	;;
401	ldf.fill f14=[r14],32
402	ldf.fill f15=[r15],32
403	;;
404	ldf.fill f16=[r14],32
405	ldf.fill f17=[r15],32
406	;;
407	ldf.fill f18=[r14],32
408	ldf.fill f19=[r15],32
409	mov b0=r21
410	;;
411	ldf.fill f20=[r14],32
412	ldf.fill f21=[r15],32
413	mov b1=r22
414	;;
415	ldf.fill f22=[r14],32
416	ldf.fill f23=[r15],32
417	mov b2=r23
418	;;
419	mov ar.bspstore=r27
420	mov ar.unat=r29		// establish unat holding the NaT bits for r4-r7
421	mov b3=r24
422	;;
423	ldf.fill f24=[r14],32
424	ldf.fill f25=[r15],32
425	mov b4=r25
426	;;
427	ldf.fill f26=[r14],32
428	ldf.fill f27=[r15],32
429	mov b5=r26
430	;;
431	ldf.fill f28=[r14],32
432	ldf.fill f29=[r15],32
433	mov ar.pfs=r16
434	;;
435	ldf.fill f30=[r14],32
436	ldf.fill f31=[r15],24
437	mov ar.lc=r17
438	;;
439	ld8.fill r4=[r14],16
440	ld8.fill r5=[r15],16
441	mov pr=r28,-1
442	;;
443	ld8.fill r6=[r14],16
444	ld8.fill r7=[r15],16
445
446	mov ar.unat=r18				// restore caller's unat
447	mov ar.rnat=r30				// must restore after bspstore but before rsc!
448	mov ar.fpsr=r19				// restore fpsr
449	mov ar.rsc=3				// put RSE back into eager mode, pl 0
450	br.cond.sptk.many b7
451END(load_switch_stack)
452
453	/*
454	 * Invoke a system call, but do some tracing before and after the call.
455	 * We MUST preserve the current register frame throughout this routine
456	 * because some system calls (such as ia64_execve) directly
457	 * manipulate ar.pfs.
458	 */
459GLOBAL_ENTRY(ia64_trace_syscall)
460	PT_REGS_UNWIND_INFO(0)
461	/*
462	 * We need to preserve the scratch registers f6-f11 in case the system
463	 * call is sigreturn.
464	 */
465	adds r16=PT(F6)+16,sp
466	adds r17=PT(F7)+16,sp
467	;;
468 	stf.spill [r16]=f6,32
469 	stf.spill [r17]=f7,32
470	;;
471 	stf.spill [r16]=f8,32
472 	stf.spill [r17]=f9,32
473	;;
474 	stf.spill [r16]=f10
475 	stf.spill [r17]=f11
476	br.call.sptk.many rp=syscall_trace_enter // give parent a chance to catch syscall args
477	cmp.lt p6,p0=r8,r0			// check tracehook
478	adds r2=PT(R8)+16,sp			// r2 = &pt_regs.r8
479	adds r3=PT(R10)+16,sp			// r3 = &pt_regs.r10
480	mov r10=0
481(p6)	br.cond.sptk strace_error		// syscall failed ->
482	adds r16=PT(F6)+16,sp
483	adds r17=PT(F7)+16,sp
484	;;
485	ldf.fill f6=[r16],32
486	ldf.fill f7=[r17],32
487	;;
488	ldf.fill f8=[r16],32
489	ldf.fill f9=[r17],32
490	;;
491	ldf.fill f10=[r16]
492	ldf.fill f11=[r17]
493	// the syscall number may have changed, so re-load it and re-calculate the
494	// syscall entry-point:
495	adds r15=PT(R15)+16,sp			// r15 = &pt_regs.r15 (syscall #)
496	;;
497	ld8 r15=[r15]
498	mov r3=NR_syscalls - 1
499	;;
500	adds r15=-1024,r15
501	movl r16=sys_call_table
502	;;
503	shladd r20=r15,3,r16			// r20 = sys_call_table + 8*(syscall-1024)
504	cmp.leu p6,p7=r15,r3
505	;;
506(p6)	ld8 r20=[r20]				// load address of syscall entry point
507(p7)	movl r20=sys_ni_syscall
508	;;
509	mov b6=r20
510	br.call.sptk.many rp=b6			// do the syscall
511.strace_check_retval:
512	cmp.lt p6,p0=r8,r0			// syscall failed?
513	adds r2=PT(R8)+16,sp			// r2 = &pt_regs.r8
514	adds r3=PT(R10)+16,sp			// r3 = &pt_regs.r10
515	mov r10=0
516(p6)	br.cond.sptk strace_error		// syscall failed ->
517	;;					// avoid RAW on r10
518.strace_save_retval:
519.mem.offset 0,0; st8.spill [r2]=r8		// store return value in slot for r8
520.mem.offset 8,0; st8.spill [r3]=r10		// clear error indication in slot for r10
521	br.call.sptk.many rp=syscall_trace_leave // give parent a chance to catch return value
522.ret3:
523(pUStk)	cmp.eq.unc p6,p0=r0,r0			// p6 <- pUStk
524(pUStk)	rsm psr.i				// disable interrupts
525	br.cond.sptk ia64_work_pending_syscall_end
526
527strace_error:
528	ld8 r3=[r2]				// load pt_regs.r8
529	sub r9=0,r8				// negate return value to get errno value
530	;;
531	cmp.ne p6,p0=r3,r0			// is pt_regs.r8!=0?
532	adds r3=16,r2				// r3=&pt_regs.r10
533	;;
534(p6)	mov r10=-1
535(p6)	mov r8=r9
536	br.cond.sptk .strace_save_retval
537END(ia64_trace_syscall)
538
539	/*
540	 * When traced and returning from sigreturn, we invoke syscall_trace but then
541	 * go straight to ia64_leave_kernel rather than ia64_leave_syscall.
542	 */
543GLOBAL_ENTRY(ia64_strace_leave_kernel)
544	PT_REGS_UNWIND_INFO(0)
545{	/*
546	 * Some versions of gas generate bad unwind info if the first instruction of a
547	 * procedure doesn't go into the first slot of a bundle.  This is a workaround.
548	 */
549	nop.m 0
550	nop.i 0
551	br.call.sptk.many rp=syscall_trace_leave // give parent a chance to catch return value
552}
553.ret4:	br.cond.sptk ia64_leave_kernel
554END(ia64_strace_leave_kernel)
555
556ENTRY(call_payload)
557	.prologue ASM_UNW_PRLG_RP|ASM_UNW_PRLG_PFS, ASM_UNW_PRLG_GRSAVE(0)
558	/* call the kernel_thread payload; fn is in r4, arg - in r5 */
559	alloc loc1=ar.pfs,0,3,1,0
560	mov loc0=rp
561	mov loc2=gp
562	mov out0=r5		// arg
563	ld8 r14 = [r4], 8	// fn.address
564	;;
565	mov b6 = r14
566	ld8 gp = [r4]		// fn.gp
567	;;
568	br.call.sptk.many rp=b6	// fn(arg)
569.ret12:	mov gp=loc2
570	mov rp=loc0
571	mov ar.pfs=loc1
572	/* ... and if it has returned, we are going to userland */
573	cmp.ne pKStk,pUStk=r0,r0
574	br.ret.sptk.many rp
575END(call_payload)
576
577GLOBAL_ENTRY(ia64_ret_from_clone)
578	PT_REGS_UNWIND_INFO(0)
579{	/*
580	 * Some versions of gas generate bad unwind info if the first instruction of a
581	 * procedure doesn't go into the first slot of a bundle.  This is a workaround.
582	 */
583	nop.m 0
584	nop.i 0
585	/*
586	 * We need to call schedule_tail() to complete the scheduling process.
587	 * Called by ia64_switch_to() after ia64_clone()->copy_thread().  r8 contains the
588	 * address of the previously executing task.
589	 */
590	br.call.sptk.many rp=ia64_invoke_schedule_tail
591}
592.ret8:
593(pKStk)	br.call.sptk.many rp=call_payload
594	adds r2=TI_FLAGS+IA64_TASK_SIZE,r13
595	;;
596	ld4 r2=[r2]
597	;;
598	mov r8=0
599	and r2=_TIF_SYSCALL_TRACEAUDIT,r2
600	;;
601	cmp.ne p6,p0=r2,r0
602(p6)	br.cond.spnt .strace_check_retval
603	;;					// added stop bits to prevent r8 dependency
604END(ia64_ret_from_clone)
605	// fall through
606GLOBAL_ENTRY(ia64_ret_from_syscall)
607	PT_REGS_UNWIND_INFO(0)
608	cmp.ge p6,p7=r8,r0			// syscall executed successfully?
609	adds r2=PT(R8)+16,sp			// r2 = &pt_regs.r8
610	mov r10=r0				// clear error indication in r10
611(p7)	br.cond.spnt handle_syscall_error	// handle potential syscall failure
612END(ia64_ret_from_syscall)
613	// fall through
614
615/*
616 * ia64_leave_syscall(): Same as ia64_leave_kernel, except that it doesn't
617 *	need to switch to bank 0 and doesn't restore the scratch registers.
618 *	To avoid leaking kernel bits, the scratch registers are set to
619 *	the following known-to-be-safe values:
620 *
621 *		  r1: restored (global pointer)
622 *		  r2: cleared
623 *		  r3: 1 (when returning to user-level)
624 *	      r8-r11: restored (syscall return value(s))
625 *		 r12: restored (user-level stack pointer)
626 *		 r13: restored (user-level thread pointer)
627 *		 r14: set to __kernel_syscall_via_epc
628 *		 r15: restored (syscall #)
629 *	     r16-r17: cleared
630 *		 r18: user-level b6
631 *		 r19: cleared
632 *		 r20: user-level ar.fpsr
633 *		 r21: user-level b0
634 *		 r22: cleared
635 *		 r23: user-level ar.bspstore
636 *		 r24: user-level ar.rnat
637 *		 r25: user-level ar.unat
638 *		 r26: user-level ar.pfs
639 *		 r27: user-level ar.rsc
640 *		 r28: user-level ip
641 *		 r29: user-level psr
642 *		 r30: user-level cfm
643 *		 r31: user-level pr
644 *	      f6-f11: cleared
645 *		  pr: restored (user-level pr)
646 *		  b0: restored (user-level rp)
647 *	          b6: restored
648 *		  b7: set to __kernel_syscall_via_epc
649 *	     ar.unat: restored (user-level ar.unat)
650 *	      ar.pfs: restored (user-level ar.pfs)
651 *	      ar.rsc: restored (user-level ar.rsc)
652 *	     ar.rnat: restored (user-level ar.rnat)
653 *	 ar.bspstore: restored (user-level ar.bspstore)
654 *	     ar.fpsr: restored (user-level ar.fpsr)
655 *	      ar.ccv: cleared
656 *	      ar.csd: cleared
657 *	      ar.ssd: cleared
658 */
659GLOBAL_ENTRY(ia64_leave_syscall)
660	PT_REGS_UNWIND_INFO(0)
661	/*
662	 * work.need_resched etc. mustn't get changed by this CPU before it returns to
663	 * user- or fsys-mode, hence we disable interrupts early on.
664	 *
665	 * p6 controls whether current_thread_info()->flags needs to be check for
666	 * extra work.  We always check for extra work when returning to user-level.
667	 * With CONFIG_PREEMPTION, we also check for extra work when the preempt_count
668	 * is 0.  After extra work processing has been completed, execution
669	 * resumes at ia64_work_processed_syscall with p6 set to 1 if the extra-work-check
670	 * needs to be redone.
671	 */
672#ifdef CONFIG_PREEMPTION
673	RSM_PSR_I(p0, r2, r18)			// disable interrupts
674	cmp.eq pLvSys,p0=r0,r0			// pLvSys=1: leave from syscall
675(pKStk) adds r20=TI_PRE_COUNT+IA64_TASK_SIZE,r13
676	;;
677	.pred.rel.mutex pUStk,pKStk
678(pKStk) ld4 r21=[r20]			// r21 <- preempt_count
679(pUStk)	mov r21=0			// r21 <- 0
680	;;
681	cmp.eq p6,p0=r21,r0		// p6 <- pUStk || (preempt_count == 0)
682#else /* !CONFIG_PREEMPTION */
683	RSM_PSR_I(pUStk, r2, r18)
684	cmp.eq pLvSys,p0=r0,r0		// pLvSys=1: leave from syscall
685(pUStk)	cmp.eq.unc p6,p0=r0,r0		// p6 <- pUStk
686#endif
687.global ia64_work_processed_syscall;
688ia64_work_processed_syscall:
689#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
690	adds r2=PT(LOADRS)+16,r12
691	MOV_FROM_ITC(pUStk, p9, r22, r19)	// fetch time at leave
692	adds r18=TI_FLAGS+IA64_TASK_SIZE,r13
693	;;
694(p6)	ld4 r31=[r18]				// load current_thread_info()->flags
695	ld8 r19=[r2],PT(B6)-PT(LOADRS)		// load ar.rsc value for "loadrs"
696	adds r3=PT(AR_BSPSTORE)+16,r12		// deferred
697	;;
698#else
699	adds r2=PT(LOADRS)+16,r12
700	adds r3=PT(AR_BSPSTORE)+16,r12
701	adds r18=TI_FLAGS+IA64_TASK_SIZE,r13
702	;;
703(p6)	ld4 r31=[r18]				// load current_thread_info()->flags
704	ld8 r19=[r2],PT(B6)-PT(LOADRS)		// load ar.rsc value for "loadrs"
705	nop.i 0
706	;;
707#endif
708	mov r16=ar.bsp				// M2  get existing backing store pointer
709	ld8 r18=[r2],PT(R9)-PT(B6)		// load b6
710(p6)	and r15=TIF_WORK_MASK,r31		// any work other than TIF_SYSCALL_TRACE?
711	;;
712	ld8 r23=[r3],PT(R11)-PT(AR_BSPSTORE)	// load ar.bspstore (may be garbage)
713(p6)	cmp4.ne.unc p6,p0=r15, r0		// any special work pending?
714(p6)	br.cond.spnt .work_pending_syscall
715	;;
716	// start restoring the state saved on the kernel stack (struct pt_regs):
717	ld8 r9=[r2],PT(CR_IPSR)-PT(R9)
718	ld8 r11=[r3],PT(CR_IIP)-PT(R11)
719(pNonSys) break 0		//      bug check: we shouldn't be here if pNonSys is TRUE!
720	;;
721	invala			// M0|1 invalidate ALAT
722	RSM_PSR_I_IC(r28, r29, r30)	// M2   turn off interrupts and interruption collection
723	cmp.eq p9,p0=r0,r0	// A    set p9 to indicate that we should restore cr.ifs
724
725	ld8 r29=[r2],16		// M0|1 load cr.ipsr
726	ld8 r28=[r3],16		// M0|1 load cr.iip
727#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
728(pUStk) add r14=TI_AC_LEAVE+IA64_TASK_SIZE,r13
729	;;
730	ld8 r30=[r2],16		// M0|1 load cr.ifs
731	ld8 r25=[r3],16		// M0|1 load ar.unat
732(pUStk) add r15=IA64_TASK_THREAD_ON_USTACK_OFFSET,r13
733	;;
734#else
735	mov r22=r0		// A    clear r22
736	;;
737	ld8 r30=[r2],16		// M0|1 load cr.ifs
738	ld8 r25=[r3],16		// M0|1 load ar.unat
739(pUStk) add r14=IA64_TASK_THREAD_ON_USTACK_OFFSET,r13
740	;;
741#endif
742	ld8 r26=[r2],PT(B0)-PT(AR_PFS)	// M0|1 load ar.pfs
743	MOV_FROM_PSR(pKStk, r22, r21)	// M2   read PSR now that interrupts are disabled
744	nop 0
745	;;
746	ld8 r21=[r2],PT(AR_RNAT)-PT(B0) // M0|1 load b0
747	ld8 r27=[r3],PT(PR)-PT(AR_RSC)	// M0|1 load ar.rsc
748	mov f6=f0			// F    clear f6
749	;;
750	ld8 r24=[r2],PT(AR_FPSR)-PT(AR_RNAT)	// M0|1 load ar.rnat (may be garbage)
751	ld8 r31=[r3],PT(R1)-PT(PR)		// M0|1 load predicates
752	mov f7=f0				// F    clear f7
753	;;
754	ld8 r20=[r2],PT(R12)-PT(AR_FPSR)	// M0|1 load ar.fpsr
755	ld8.fill r1=[r3],16			// M0|1 load r1
756(pUStk) mov r17=1				// A
757	;;
758#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
759(pUStk) st1 [r15]=r17				// M2|3
760#else
761(pUStk) st1 [r14]=r17				// M2|3
762#endif
763	ld8.fill r13=[r3],16			// M0|1
764	mov f8=f0				// F    clear f8
765	;;
766	ld8.fill r12=[r2]			// M0|1 restore r12 (sp)
767	ld8.fill r15=[r3]			// M0|1 restore r15
768	mov b6=r18				// I0   restore b6
769
770	LOAD_PHYS_STACK_REG_SIZE(r17)
771	mov f9=f0					// F    clear f9
772(pKStk) br.cond.dpnt.many skip_rbs_switch		// B
773
774	srlz.d				// M0   ensure interruption collection is off (for cover)
775	shr.u r18=r19,16		// I0|1 get byte size of existing "dirty" partition
776	COVER				// B    add current frame into dirty partition & set cr.ifs
777	;;
778#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
779	mov r19=ar.bsp			// M2   get new backing store pointer
780	st8 [r14]=r22			// M	save time at leave
781	mov f10=f0			// F    clear f10
782
783	mov r22=r0			// A	clear r22
784	movl r14=__kernel_syscall_via_epc // X
785	;;
786#else
787	mov r19=ar.bsp			// M2   get new backing store pointer
788	mov f10=f0			// F    clear f10
789
790	nop.m 0
791	movl r14=__kernel_syscall_via_epc // X
792	;;
793#endif
794	mov.m ar.csd=r0			// M2   clear ar.csd
795	mov.m ar.ccv=r0			// M2   clear ar.ccv
796	mov b7=r14			// I0   clear b7 (hint with __kernel_syscall_via_epc)
797
798	mov.m ar.ssd=r0			// M2   clear ar.ssd
799	mov f11=f0			// F    clear f11
800	br.cond.sptk.many rbs_switch	// B
801END(ia64_leave_syscall)
802
803GLOBAL_ENTRY(ia64_leave_kernel)
804	PT_REGS_UNWIND_INFO(0)
805	/*
806	 * work.need_resched etc. mustn't get changed by this CPU before it returns to
807	 * user- or fsys-mode, hence we disable interrupts early on.
808	 *
809	 * p6 controls whether current_thread_info()->flags needs to be check for
810	 * extra work.  We always check for extra work when returning to user-level.
811	 * With CONFIG_PREEMPTION, we also check for extra work when the preempt_count
812	 * is 0.  After extra work processing has been completed, execution
813	 * resumes at .work_processed_syscall with p6 set to 1 if the extra-work-check
814	 * needs to be redone.
815	 */
816#ifdef CONFIG_PREEMPTION
817	RSM_PSR_I(p0, r17, r31)			// disable interrupts
818	cmp.eq p0,pLvSys=r0,r0			// pLvSys=0: leave from kernel
819(pKStk)	adds r20=TI_PRE_COUNT+IA64_TASK_SIZE,r13
820	;;
821	.pred.rel.mutex pUStk,pKStk
822(pKStk)	ld4 r21=[r20]			// r21 <- preempt_count
823(pUStk)	mov r21=0			// r21 <- 0
824	;;
825	cmp.eq p6,p0=r21,r0		// p6 <- pUStk || (preempt_count == 0)
826#else
827	RSM_PSR_I(pUStk, r17, r31)
828	cmp.eq p0,pLvSys=r0,r0		// pLvSys=0: leave from kernel
829(pUStk)	cmp.eq.unc p6,p0=r0,r0		// p6 <- pUStk
830#endif
831.work_processed_kernel:
832	adds r17=TI_FLAGS+IA64_TASK_SIZE,r13
833	;;
834(p6)	ld4 r31=[r17]				// load current_thread_info()->flags
835	adds r21=PT(PR)+16,r12
836	;;
837
838	lfetch [r21],PT(CR_IPSR)-PT(PR)
839	adds r2=PT(B6)+16,r12
840	adds r3=PT(R16)+16,r12
841	;;
842	lfetch [r21]
843	ld8 r28=[r2],8		// load b6
844	adds r29=PT(R24)+16,r12
845
846	ld8.fill r16=[r3],PT(AR_CSD)-PT(R16)
847	adds r30=PT(AR_CCV)+16,r12
848(p6)	and r19=TIF_WORK_MASK,r31		// any work other than TIF_SYSCALL_TRACE?
849	;;
850	ld8.fill r24=[r29]
851	ld8 r15=[r30]		// load ar.ccv
852(p6)	cmp4.ne.unc p6,p0=r19, r0		// any special work pending?
853	;;
854	ld8 r29=[r2],16		// load b7
855	ld8 r30=[r3],16		// load ar.csd
856(p6)	br.cond.spnt .work_pending
857	;;
858	ld8 r31=[r2],16		// load ar.ssd
859	ld8.fill r8=[r3],16
860	;;
861	ld8.fill r9=[r2],16
862	ld8.fill r10=[r3],PT(R17)-PT(R10)
863	;;
864	ld8.fill r11=[r2],PT(R18)-PT(R11)
865	ld8.fill r17=[r3],16
866	;;
867	ld8.fill r18=[r2],16
868	ld8.fill r19=[r3],16
869	;;
870	ld8.fill r20=[r2],16
871	ld8.fill r21=[r3],16
872	mov ar.csd=r30
873	mov ar.ssd=r31
874	;;
875	RSM_PSR_I_IC(r23, r22, r25)	// initiate turning off of interrupt and interruption collection
876	invala			// invalidate ALAT
877	;;
878	ld8.fill r22=[r2],24
879	ld8.fill r23=[r3],24
880	mov b6=r28
881	;;
882	ld8.fill r25=[r2],16
883	ld8.fill r26=[r3],16
884	mov b7=r29
885	;;
886	ld8.fill r27=[r2],16
887	ld8.fill r28=[r3],16
888	;;
889	ld8.fill r29=[r2],16
890	ld8.fill r30=[r3],24
891	;;
892	ld8.fill r31=[r2],PT(F9)-PT(R31)
893	adds r3=PT(F10)-PT(F6),r3
894	;;
895	ldf.fill f9=[r2],PT(F6)-PT(F9)
896	ldf.fill f10=[r3],PT(F8)-PT(F10)
897	;;
898	ldf.fill f6=[r2],PT(F7)-PT(F6)
899	;;
900	ldf.fill f7=[r2],PT(F11)-PT(F7)
901	ldf.fill f8=[r3],32
902	;;
903	srlz.d	// ensure that inter. collection is off (VHPT is don't care, since text is pinned)
904	mov ar.ccv=r15
905	;;
906	ldf.fill f11=[r2]
907	BSW_0(r2, r3, r15)	// switch back to bank 0 (no stop bit required beforehand...)
908	;;
909(pUStk)	mov r18=IA64_KR(CURRENT)// M2 (12 cycle read latency)
910	adds r16=PT(CR_IPSR)+16,r12
911	adds r17=PT(CR_IIP)+16,r12
912
913#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
914	.pred.rel.mutex pUStk,pKStk
915	MOV_FROM_PSR(pKStk, r22, r29)	// M2 read PSR now that interrupts are disabled
916	MOV_FROM_ITC(pUStk, p9, r22, r29)	// M  fetch time at leave
917	nop.i 0
918	;;
919#else
920	MOV_FROM_PSR(pKStk, r22, r29)	// M2 read PSR now that interrupts are disabled
921	nop.i 0
922	nop.i 0
923	;;
924#endif
925	ld8 r29=[r16],16	// load cr.ipsr
926	ld8 r28=[r17],16	// load cr.iip
927	;;
928	ld8 r30=[r16],16	// load cr.ifs
929	ld8 r25=[r17],16	// load ar.unat
930	;;
931	ld8 r26=[r16],16	// load ar.pfs
932	ld8 r27=[r17],16	// load ar.rsc
933	cmp.eq p9,p0=r0,r0	// set p9 to indicate that we should restore cr.ifs
934	;;
935	ld8 r24=[r16],16	// load ar.rnat (may be garbage)
936	ld8 r23=[r17],16	// load ar.bspstore (may be garbage)
937	;;
938	ld8 r31=[r16],16	// load predicates
939	ld8 r21=[r17],16	// load b0
940	;;
941	ld8 r19=[r16],16	// load ar.rsc value for "loadrs"
942	ld8.fill r1=[r17],16	// load r1
943	;;
944	ld8.fill r12=[r16],16
945	ld8.fill r13=[r17],16
946#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
947(pUStk)	adds r3=TI_AC_LEAVE+IA64_TASK_SIZE,r18
948#else
949(pUStk)	adds r18=IA64_TASK_THREAD_ON_USTACK_OFFSET,r18
950#endif
951	;;
952	ld8 r20=[r16],16	// ar.fpsr
953	ld8.fill r15=[r17],16
954#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
955(pUStk)	adds r18=IA64_TASK_THREAD_ON_USTACK_OFFSET,r18	// deferred
956#endif
957	;;
958	ld8.fill r14=[r16],16
959	ld8.fill r2=[r17]
960(pUStk)	mov r17=1
961	;;
962#ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
963	//  mmi_ :  ld8 st1 shr;;         mmi_ : st8 st1 shr;;
964	//  mib  :  mov add br        ->  mib  : ld8 add br
965	//  bbb_ :  br  nop cover;;       mbb_ : mov br  cover;;
966	//
967	//  no one require bsp in r16 if (pKStk) branch is selected.
968(pUStk)	st8 [r3]=r22		// save time at leave
969(pUStk)	st1 [r18]=r17		// restore current->thread.on_ustack
970	shr.u r18=r19,16	// get byte size of existing "dirty" partition
971	;;
972	ld8.fill r3=[r16]	// deferred
973	LOAD_PHYS_STACK_REG_SIZE(r17)
974(pKStk)	br.cond.dpnt skip_rbs_switch
975	mov r16=ar.bsp		// get existing backing store pointer
976#else
977	ld8.fill r3=[r16]
978(pUStk)	st1 [r18]=r17		// restore current->thread.on_ustack
979	shr.u r18=r19,16	// get byte size of existing "dirty" partition
980	;;
981	mov r16=ar.bsp		// get existing backing store pointer
982	LOAD_PHYS_STACK_REG_SIZE(r17)
983(pKStk)	br.cond.dpnt skip_rbs_switch
984#endif
985
986	/*
987	 * Restore user backing store.
988	 *
989	 * NOTE: alloc, loadrs, and cover can't be predicated.
990	 */
991(pNonSys) br.cond.dpnt dont_preserve_current_frame
992	COVER				// add current frame into dirty partition and set cr.ifs
993	;;
994	mov r19=ar.bsp			// get new backing store pointer
995rbs_switch:
996	sub r16=r16,r18			// krbs = old bsp - size of dirty partition
997	cmp.ne p9,p0=r0,r0		// clear p9 to skip restore of cr.ifs
998	;;
999	sub r19=r19,r16			// calculate total byte size of dirty partition
1000	add r18=64,r18			// don't force in0-in7 into memory...
1001	;;
1002	shl r19=r19,16			// shift size of dirty partition into loadrs position
1003	;;
1004dont_preserve_current_frame:
1005	/*
1006	 * To prevent leaking bits between the kernel and user-space,
1007	 * we must clear the stacked registers in the "invalid" partition here.
1008	 * Not pretty, but at least it's fast (3.34 registers/cycle on Itanium,
1009	 * 5 registers/cycle on McKinley).
1010	 */
1011#	define pRecurse	p6
1012#	define pReturn	p7
1013#ifdef CONFIG_ITANIUM
1014#	define Nregs	10
1015#else
1016#	define Nregs	14
1017#endif
1018	alloc loc0=ar.pfs,2,Nregs-2,2,0
1019	shr.u loc1=r18,9		// RNaTslots <= floor(dirtySize / (64*8))
1020	sub r17=r17,r18			// r17 = (physStackedSize + 8) - dirtySize
1021	;;
1022	mov ar.rsc=r19			// load ar.rsc to be used for "loadrs"
1023	shladd in0=loc1,3,r17
1024	mov in1=0
1025	;;
1026	TEXT_ALIGN(32)
1027rse_clear_invalid:
1028#ifdef CONFIG_ITANIUM
1029	// cycle 0
1030 { .mii
1031	alloc loc0=ar.pfs,2,Nregs-2,2,0
1032	cmp.lt pRecurse,p0=Nregs*8,in0	// if more than Nregs regs left to clear, (re)curse
1033	add out0=-Nregs*8,in0
1034}{ .mfb
1035	add out1=1,in1			// increment recursion count
1036	nop.f 0
1037	nop.b 0				// can't do br.call here because of alloc (WAW on CFM)
1038	;;
1039}{ .mfi	// cycle 1
1040	mov loc1=0
1041	nop.f 0
1042	mov loc2=0
1043}{ .mib
1044	mov loc3=0
1045	mov loc4=0
1046(pRecurse) br.call.sptk.many b0=rse_clear_invalid
1047
1048}{ .mfi	// cycle 2
1049	mov loc5=0
1050	nop.f 0
1051	cmp.ne pReturn,p0=r0,in1	// if recursion count != 0, we need to do a br.ret
1052}{ .mib
1053	mov loc6=0
1054	mov loc7=0
1055(pReturn) br.ret.sptk.many b0
1056}
1057#else /* !CONFIG_ITANIUM */
1058	alloc loc0=ar.pfs,2,Nregs-2,2,0
1059	cmp.lt pRecurse,p0=Nregs*8,in0	// if more than Nregs regs left to clear, (re)curse
1060	add out0=-Nregs*8,in0
1061	add out1=1,in1			// increment recursion count
1062	mov loc1=0
1063	mov loc2=0
1064	;;
1065	mov loc3=0
1066	mov loc4=0
1067	mov loc5=0
1068	mov loc6=0
1069	mov loc7=0
1070(pRecurse) br.call.dptk.few b0=rse_clear_invalid
1071	;;
1072	mov loc8=0
1073	mov loc9=0
1074	cmp.ne pReturn,p0=r0,in1	// if recursion count != 0, we need to do a br.ret
1075	mov loc10=0
1076	mov loc11=0
1077(pReturn) br.ret.dptk.many b0
1078#endif /* !CONFIG_ITANIUM */
1079#	undef pRecurse
1080#	undef pReturn
1081	;;
1082	alloc r17=ar.pfs,0,0,0,0	// drop current register frame
1083	;;
1084	loadrs
1085	;;
1086skip_rbs_switch:
1087	mov ar.unat=r25		// M2
1088(pKStk)	extr.u r22=r22,21,1	// I0 extract current value of psr.pp from r22
1089(pLvSys)mov r19=r0		// A  clear r19 for leave_syscall, no-op otherwise
1090	;;
1091(pUStk)	mov ar.bspstore=r23	// M2
1092(pKStk)	dep r29=r22,r29,21,1	// I0 update ipsr.pp with psr.pp
1093(pLvSys)mov r16=r0		// A  clear r16 for leave_syscall, no-op otherwise
1094	;;
1095	MOV_TO_IPSR(p0, r29, r25)	// M2
1096	mov ar.pfs=r26		// I0
1097(pLvSys)mov r17=r0		// A  clear r17 for leave_syscall, no-op otherwise
1098
1099	MOV_TO_IFS(p9, r30, r25)// M2
1100	mov b0=r21		// I0
1101(pLvSys)mov r18=r0		// A  clear r18 for leave_syscall, no-op otherwise
1102
1103	mov ar.fpsr=r20		// M2
1104	MOV_TO_IIP(r28, r25)	// M2
1105	nop 0
1106	;;
1107(pUStk)	mov ar.rnat=r24		// M2 must happen with RSE in lazy mode
1108	nop 0
1109(pLvSys)mov r2=r0
1110
1111	mov ar.rsc=r27		// M2
1112	mov pr=r31,-1		// I0
1113	RFI			// B
1114
1115	/*
1116	 * On entry:
1117	 *	r20 = &current->thread_info->pre_count (if CONFIG_PREEMPTION)
1118	 *	r31 = current->thread_info->flags
1119	 * On exit:
1120	 *	p6 = TRUE if work-pending-check needs to be redone
1121	 *
1122	 * Interrupts are disabled on entry, reenabled depend on work, and
1123	 * disabled on exit.
1124	 */
1125.work_pending_syscall:
1126	add r2=-8,r2
1127	add r3=-8,r3
1128	;;
1129	st8 [r2]=r8
1130	st8 [r3]=r10
1131.work_pending:
1132	tbit.z p6,p0=r31,TIF_NEED_RESCHED	// is resched not needed?
1133(p6)	br.cond.sptk.few .notify
1134	br.call.spnt.many rp=preempt_schedule_irq
1135.ret9:	cmp.eq p6,p0=r0,r0	// p6 <- 1 (re-check)
1136(pLvSys)br.cond.sptk.few  ia64_work_pending_syscall_end
1137	br.cond.sptk.many .work_processed_kernel
1138
1139.notify:
1140(pUStk)	br.call.spnt.many rp=notify_resume_user
1141.ret10:	cmp.ne p6,p0=r0,r0	// p6 <- 0 (don't re-check)
1142(pLvSys)br.cond.sptk.few  ia64_work_pending_syscall_end
1143	br.cond.sptk.many .work_processed_kernel
1144
1145.global ia64_work_pending_syscall_end;
1146ia64_work_pending_syscall_end:
1147	adds r2=PT(R8)+16,r12
1148	adds r3=PT(R10)+16,r12
1149	;;
1150	ld8 r8=[r2]
1151	ld8 r10=[r3]
1152	br.cond.sptk.many ia64_work_processed_syscall
1153END(ia64_leave_kernel)
1154
1155ENTRY(handle_syscall_error)
1156	/*
1157	 * Some system calls (e.g., ptrace, mmap) can return arbitrary values which could
1158	 * lead us to mistake a negative return value as a failed syscall.  Those syscall
1159	 * must deposit a non-zero value in pt_regs.r8 to indicate an error.  If
1160	 * pt_regs.r8 is zero, we assume that the call completed successfully.
1161	 */
1162	PT_REGS_UNWIND_INFO(0)
1163	ld8 r3=[r2]		// load pt_regs.r8
1164	;;
1165	cmp.eq p6,p7=r3,r0	// is pt_regs.r8==0?
1166	;;
1167(p7)	mov r10=-1
1168(p7)	sub r8=0,r8		// negate return value to get errno
1169	br.cond.sptk ia64_leave_syscall
1170END(handle_syscall_error)
1171
1172	/*
1173	 * Invoke schedule_tail(task) while preserving in0-in7, which may be needed
1174	 * in case a system call gets restarted.
1175	 */
1176GLOBAL_ENTRY(ia64_invoke_schedule_tail)
1177	.prologue ASM_UNW_PRLG_RP|ASM_UNW_PRLG_PFS, ASM_UNW_PRLG_GRSAVE(8)
1178	alloc loc1=ar.pfs,8,2,1,0
1179	mov loc0=rp
1180	mov out0=r8				// Address of previous task
1181	;;
1182	br.call.sptk.many rp=schedule_tail
1183.ret11:	mov ar.pfs=loc1
1184	mov rp=loc0
1185	br.ret.sptk.many rp
1186END(ia64_invoke_schedule_tail)
1187
1188	/*
1189	 * Setup stack and call do_notify_resume_user(), keeping interrupts
1190	 * disabled.
1191	 *
1192	 * Note that pSys and pNonSys need to be set up by the caller.
1193	 * We declare 8 input registers so the system call args get preserved,
1194	 * in case we need to restart a system call.
1195	 */
1196GLOBAL_ENTRY(notify_resume_user)
1197	.prologue ASM_UNW_PRLG_RP|ASM_UNW_PRLG_PFS, ASM_UNW_PRLG_GRSAVE(8)
1198	alloc loc1=ar.pfs,8,2,3,0 // preserve all eight input regs in case of syscall restart!
1199	mov r9=ar.unat
1200	mov loc0=rp				// save return address
1201	mov out0=0				// there is no "oldset"
1202	adds out1=8,sp				// out1=&sigscratch->ar_pfs
1203(pSys)	mov out2=1				// out2==1 => we're in a syscall
1204	;;
1205(pNonSys) mov out2=0				// out2==0 => not a syscall
1206	.fframe 16
1207	.spillsp ar.unat, 16
1208	st8 [sp]=r9,-16				// allocate space for ar.unat and save it
1209	st8 [out1]=loc1,-8			// save ar.pfs, out1=&sigscratch
1210	.body
1211	br.call.sptk.many rp=do_notify_resume_user
1212.ret15:	.restore sp
1213	adds sp=16,sp				// pop scratch stack space
1214	;;
1215	ld8 r9=[sp]				// load new unat from sigscratch->scratch_unat
1216	mov rp=loc0
1217	;;
1218	mov ar.unat=r9
1219	mov ar.pfs=loc1
1220	br.ret.sptk.many rp
1221END(notify_resume_user)
1222
1223ENTRY(sys_rt_sigreturn)
1224	PT_REGS_UNWIND_INFO(0)
1225	/*
1226	 * Allocate 8 input registers since ptrace() may clobber them
1227	 */
1228	alloc r2=ar.pfs,8,0,1,0
1229	.prologue
1230	PT_REGS_SAVES(16)
1231	adds sp=-16,sp
1232	.body
1233	cmp.eq pNonSys,pSys=r0,r0		// sigreturn isn't a normal syscall...
1234	;;
1235	/*
1236	 * leave_kernel() restores f6-f11 from pt_regs, but since the streamlined
1237	 * syscall-entry path does not save them we save them here instead.  Note: we
1238	 * don't need to save any other registers that are not saved by the stream-lined
1239	 * syscall path, because restore_sigcontext() restores them.
1240	 */
1241	adds r16=PT(F6)+32,sp
1242	adds r17=PT(F7)+32,sp
1243	;;
1244 	stf.spill [r16]=f6,32
1245 	stf.spill [r17]=f7,32
1246	;;
1247 	stf.spill [r16]=f8,32
1248 	stf.spill [r17]=f9,32
1249	;;
1250 	stf.spill [r16]=f10
1251 	stf.spill [r17]=f11
1252	adds out0=16,sp				// out0 = &sigscratch
1253	br.call.sptk.many rp=ia64_rt_sigreturn
1254.ret19:	.restore sp,0
1255	adds sp=16,sp
1256	;;
1257	ld8 r9=[sp]				// load new ar.unat
1258	mov.sptk b7=r8,ia64_leave_kernel
1259	;;
1260	mov ar.unat=r9
1261	br.many b7
1262END(sys_rt_sigreturn)
1263
1264GLOBAL_ENTRY(ia64_prepare_handle_unaligned)
1265	.prologue
1266	/*
1267	 * r16 = fake ar.pfs, we simply need to make sure privilege is still 0
1268	 */
1269	mov r16=r0
1270	DO_SAVE_SWITCH_STACK
1271	br.call.sptk.many rp=ia64_handle_unaligned	// stack frame setup in ivt
1272.ret21:	.body
1273	DO_LOAD_SWITCH_STACK
1274	br.cond.sptk.many rp				// goes to ia64_leave_kernel
1275END(ia64_prepare_handle_unaligned)
1276
1277	//
1278	// unw_init_running(void (*callback)(info, arg), void *arg)
1279	//
1280#	define EXTRA_FRAME_SIZE	((UNW_FRAME_INFO_SIZE+15)&~15)
1281
1282GLOBAL_ENTRY(unw_init_running)
1283	.prologue ASM_UNW_PRLG_RP|ASM_UNW_PRLG_PFS, ASM_UNW_PRLG_GRSAVE(2)
1284	alloc loc1=ar.pfs,2,3,3,0
1285	;;
1286	ld8 loc2=[in0],8
1287	mov loc0=rp
1288	mov r16=loc1
1289	DO_SAVE_SWITCH_STACK
1290	.body
1291
1292	.prologue ASM_UNW_PRLG_RP|ASM_UNW_PRLG_PFS, ASM_UNW_PRLG_GRSAVE(2)
1293	.fframe IA64_SWITCH_STACK_SIZE+EXTRA_FRAME_SIZE
1294	SWITCH_STACK_SAVES(EXTRA_FRAME_SIZE)
1295	adds sp=-EXTRA_FRAME_SIZE,sp
1296	.body
1297	;;
1298	adds out0=16,sp				// &info
1299	mov out1=r13				// current
1300	adds out2=16+EXTRA_FRAME_SIZE,sp	// &switch_stack
1301	br.call.sptk.many rp=unw_init_frame_info
13021:	adds out0=16,sp				// &info
1303	mov b6=loc2
1304	mov loc2=gp				// save gp across indirect function call
1305	;;
1306	ld8 gp=[in0]
1307	mov out1=in1				// arg
1308	br.call.sptk.many rp=b6			// invoke the callback function
13091:	mov gp=loc2				// restore gp
1310
1311	// For now, we don't allow changing registers from within
1312	// unw_init_running; if we ever want to allow that, we'd
1313	// have to do a load_switch_stack here:
1314	.restore sp
1315	adds sp=IA64_SWITCH_STACK_SIZE+EXTRA_FRAME_SIZE,sp
1316
1317	mov ar.pfs=loc1
1318	mov rp=loc0
1319	br.ret.sptk.many rp
1320END(unw_init_running)
1321EXPORT_SYMBOL(unw_init_running)
1322
1323#ifdef CONFIG_FUNCTION_TRACER
1324#ifdef CONFIG_DYNAMIC_FTRACE
1325GLOBAL_ENTRY(_mcount)
1326	br ftrace_stub
1327END(_mcount)
1328EXPORT_SYMBOL(_mcount)
1329
1330.here:
1331	br.ret.sptk.many b0
1332
1333GLOBAL_ENTRY(ftrace_caller)
1334	alloc out0 = ar.pfs, 8, 0, 4, 0
1335	mov out3 = r0
1336	;;
1337	mov out2 = b0
1338	add r3 = 0x20, r3
1339	mov out1 = r1;
1340	br.call.sptk.many b0 = ftrace_patch_gp
1341	//this might be called from module, so we must patch gp
1342ftrace_patch_gp:
1343	movl gp=__gp
1344	mov b0 = r3
1345	;;
1346.global ftrace_call;
1347ftrace_call:
1348{
1349	.mlx
1350	nop.m 0x0
1351	movl r3 = .here;;
1352}
1353	alloc loc0 = ar.pfs, 4, 4, 2, 0
1354	;;
1355	mov loc1 = b0
1356	mov out0 = b0
1357	mov loc2 = r8
1358	mov loc3 = r15
1359	;;
1360	adds out0 = -MCOUNT_INSN_SIZE, out0
1361	mov out1 = in2
1362	mov b6 = r3
1363
1364	br.call.sptk.many b0 = b6
1365	;;
1366	mov ar.pfs = loc0
1367	mov b0 = loc1
1368	mov r8 = loc2
1369	mov r15 = loc3
1370	br ftrace_stub
1371	;;
1372END(ftrace_caller)
1373
1374#else
1375GLOBAL_ENTRY(_mcount)
1376	movl r2 = ftrace_stub
1377	movl r3 = ftrace_trace_function;;
1378	ld8 r3 = [r3];;
1379	ld8 r3 = [r3];;
1380	cmp.eq p7,p0 = r2, r3
1381(p7)	br.sptk.many ftrace_stub
1382	;;
1383
1384	alloc loc0 = ar.pfs, 4, 4, 2, 0
1385	;;
1386	mov loc1 = b0
1387	mov out0 = b0
1388	mov loc2 = r8
1389	mov loc3 = r15
1390	;;
1391	adds out0 = -MCOUNT_INSN_SIZE, out0
1392	mov out1 = in2
1393	mov b6 = r3
1394
1395	br.call.sptk.many b0 = b6
1396	;;
1397	mov ar.pfs = loc0
1398	mov b0 = loc1
1399	mov r8 = loc2
1400	mov r15 = loc3
1401	br ftrace_stub
1402	;;
1403END(_mcount)
1404#endif
1405
1406GLOBAL_ENTRY(ftrace_stub)
1407	mov r3 = b0
1408	movl r2 = _mcount_ret_helper
1409	;;
1410	mov b6 = r2
1411	mov b7 = r3
1412	br.ret.sptk.many b6
1413
1414_mcount_ret_helper:
1415	mov b0 = r42
1416	mov r1 = r41
1417	mov ar.pfs = r40
1418	br b7
1419END(ftrace_stub)
1420
1421#endif /* CONFIG_FUNCTION_TRACER */
1422
1423#define __SYSCALL(nr, entry, nargs) data8 entry
1424	.rodata
1425	.align 8
1426	.globl sys_call_table
1427sys_call_table:
1428#include <asm/syscall_table.h>
1429#undef __SYSCALL
1430