• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1/*
2 * Copyright 2011 Tilera Corporation. All Rights Reserved.
3 *
4 *   This program is free software; you can redistribute it and/or
5 *   modify it under the terms of the GNU General Public License
6 *   as published by the Free Software Foundation, version 2.
7 *
8 *   This program is distributed in the hope that it will be useful, but
9 *   WITHOUT ANY WARRANTY; without even the implied warranty of
10 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
11 *   NON INFRINGEMENT.  See the GNU General Public License for
12 *   more details.
13 *
14 * Linux interrupt vectors.
15 */
16
17#include <linux/linkage.h>
18#include <linux/errno.h>
19#include <linux/unistd.h>
20#include <linux/init.h>
21#include <asm/ptrace.h>
22#include <asm/thread_info.h>
23#include <asm/irqflags.h>
24#include <asm/asm-offsets.h>
25#include <asm/types.h>
26#include <asm/traps.h>
27#include <asm/signal.h>
28#include <hv/hypervisor.h>
29#include <arch/abi.h>
30#include <arch/interrupts.h>
31#include <arch/spr_def.h>
32
33#define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
34
35#define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
36
37#if CONFIG_KERNEL_PL == 1 || CONFIG_KERNEL_PL == 2
38/*
39 * Set "result" non-zero if ex1 holds the PL of the kernel
40 * (with or without ICS being set).  Note this works only
41 * because we never find the PL at level 3.
42 */
43# define IS_KERNEL_EX1(result, ex1) andi result, ex1, CONFIG_KERNEL_PL
44#else
45# error Recode IS_KERNEL_EX1 for CONFIG_KERNEL_PL
46#endif
47
48	.macro  push_reg reg, ptr=sp, delta=-8
49	{
50	 st     \ptr, \reg
51	 addli  \ptr, \ptr, \delta
52	}
53	.endm
54
55	.macro  pop_reg reg, ptr=sp, delta=8
56	{
57	 ld     \reg, \ptr
58	 addli  \ptr, \ptr, \delta
59	}
60	.endm
61
62	.macro  pop_reg_zero reg, zreg, ptr=sp, delta=8
63	{
64	 move   \zreg, zero
65	 ld     \reg, \ptr
66	 addi   \ptr, \ptr, \delta
67	}
68	.endm
69
70	.macro  push_extra_callee_saves reg
71	PTREGS_PTR(\reg, PTREGS_OFFSET_REG(51))
72	push_reg r51, \reg
73	push_reg r50, \reg
74	push_reg r49, \reg
75	push_reg r48, \reg
76	push_reg r47, \reg
77	push_reg r46, \reg
78	push_reg r45, \reg
79	push_reg r44, \reg
80	push_reg r43, \reg
81	push_reg r42, \reg
82	push_reg r41, \reg
83	push_reg r40, \reg
84	push_reg r39, \reg
85	push_reg r38, \reg
86	push_reg r37, \reg
87	push_reg r36, \reg
88	push_reg r35, \reg
89	push_reg r34, \reg, PTREGS_OFFSET_BASE - PTREGS_OFFSET_REG(34)
90	.endm
91
92	.macro  panic str
93	.pushsection .rodata, "a"
941:
95	.asciz  "\str"
96	.popsection
97	{
98	 moveli r0, hw2_last(1b)
99	}
100	{
101	 shl16insli r0, r0, hw1(1b)
102	}
103	{
104	 shl16insli r0, r0, hw0(1b)
105	 jal    panic
106	}
107	.endm
108
109	/*
110	 * Unalign data exception fast handling: In order to handle
111	 * unaligned data access, a fast JIT version is generated and stored
112	 * in a specific area in user space. We first need to do a quick poke
113	 * to see if the JIT is available. We use certain bits in the fault
114	 * PC (3 to 9 is used for 16KB page size) as index to address the JIT
115	 * code area. The first 64bit word is the fault PC, and the 2nd one is
116	 * the fault bundle itself. If these 2 words both match, then we
117	 * directly "iret" to JIT code. If not, a slow path is invoked to
118	 * generate new JIT code. Note: the current JIT code WILL be
119	 * overwritten if it existed. So, ideally we can handle 128 unalign
120	 * fixups via JIT. For lookup efficiency and to effectively support
121	 * tight loops with multiple unaligned reference, a simple
122	 * direct-mapped cache is used.
123	 *
124	 * SPR_EX_CONTEXT_K_0 is modified to return to JIT code.
125	 * SPR_EX_CONTEXT_K_1 has ICS set.
126	 * SPR_EX_CONTEXT_0_0 is setup to user program's next PC.
127	 * SPR_EX_CONTEXT_0_1 = 0.
128	 */
129	.macro int_hand_unalign_fast  vecnum, vecname
130	.org  (\vecnum << 8)
131intvec_\vecname:
132	/* Put r3 in SPR_SYSTEM_SAVE_K_1.  */
133	mtspr   SPR_SYSTEM_SAVE_K_1, r3
134
135	mfspr   r3, SPR_EX_CONTEXT_K_1
136	/*
137	 * Examine if exception comes from user without ICS set.
138	 * If not, just go directly to the slow path.
139	 */
140	bnez    r3, hand_unalign_slow_nonuser
141
142	mfspr   r3, SPR_SYSTEM_SAVE_K_0
143
144	/* Get &thread_info->unalign_jit_tmp[0] in r3. */
145	bfexts  r3, r3, 0, CPU_SHIFT-1
146	mm      r3, zero, LOG2_THREAD_SIZE, 63
147	addli   r3, r3, THREAD_INFO_UNALIGN_JIT_TMP_OFFSET
148
149	/*
150	 * Save r0, r1, r2 into thread_info array r3 points to
151	 * from low to high memory in order.
152	 */
153	st_add  r3, r0, 8
154	st_add  r3, r1, 8
155	{
156	 st_add r3, r2, 8
157	 andi   r2, sp, 7
158	}
159
160	/* Save stored r3 value so we can revert it on a page fault. */
161	mfspr   r1, SPR_SYSTEM_SAVE_K_1
162	st      r3, r1
163
164	{
165	 /* Generate a SIGBUS if sp is not 8-byte aligned. */
166	 bnez   r2, hand_unalign_slow_badsp
167	}
168
169	/*
170	 * Get the thread_info in r0; load r1 with pc. Set the low bit of sp
171	 * as an indicator to the page fault code in case we fault.
172	 */
173	{
174	 ori    sp, sp, 1
175	 mfspr  r1, SPR_EX_CONTEXT_K_0
176	}
177
178	/* Add the jit_info offset in thread_info; extract r1 [3:9] into r2. */
179	{
180	 addli  r0, r3, THREAD_INFO_UNALIGN_JIT_BASE_OFFSET - \
181	  (THREAD_INFO_UNALIGN_JIT_TMP_OFFSET + (3 * 8))
182	 bfextu r2, r1, 3, (2 + PAGE_SHIFT - UNALIGN_JIT_SHIFT)
183	}
184
185	/* Load the jit_info; multiply r2 by 128. */
186	{
187	 ld     r0, r0
188	 shli   r2, r2, UNALIGN_JIT_SHIFT
189	}
190
191	/*
192	 * If r0 is NULL, the JIT page is not mapped, so go to slow path;
193	 * add offset r2 to r0 at the same time.
194	 */
195	{
196	 beqz   r0, hand_unalign_slow
197	 add    r2, r0, r2
198	}
199
200        /*
201	 * We are loading from userspace (both the JIT info PC and
202	 * instruction word, and the instruction word we executed)
203	 * and since either could fault while holding the interrupt
204	 * critical section, we must tag this region and check it in
205	 * do_page_fault() to handle it properly.
206	 */
207ENTRY(__start_unalign_asm_code)
208
209	/* Load first word of JIT in r0 and increment r2 by 8. */
210	ld_add  r0, r2, 8
211
212	/*
213	 * Compare the PC with the 1st word in JIT; load the fault bundle
214	 * into r1.
215	 */
216	{
217	 cmpeq  r0, r0, r1
218	 ld     r1, r1
219	}
220
221	/* Go to slow path if PC doesn't match. */
222	beqz    r0, hand_unalign_slow
223
224	/*
225	 * Load the 2nd word of JIT, which is supposed to be the fault
226	 * bundle for a cache hit. Increment r2; after this bundle r2 will
227	 * point to the potential start of the JIT code we want to run.
228	 */
229	ld_add  r0, r2, 8
230
231	/* No further accesses to userspace are done after this point. */
232ENTRY(__end_unalign_asm_code)
233
234	/* Compare the real bundle with what is saved in the JIT area. */
235	{
236	 cmpeq  r0, r1, r0
237	 mtspr  SPR_EX_CONTEXT_0_1, zero
238	}
239
240	/* Go to slow path if the fault bundle does not match. */
241	beqz    r0, hand_unalign_slow
242
243	/*
244	 * A cache hit is found.
245	 * r2 points to start of JIT code (3rd word).
246	 * r0 is the fault pc.
247	 * r1 is the fault bundle.
248	 * Reset the low bit of sp.
249	 */
250	{
251	 mfspr  r0, SPR_EX_CONTEXT_K_0
252	 andi   sp, sp, ~1
253	}
254
255	/* Write r2 into EX_CONTEXT_K_0 and increment PC. */
256	{
257	 mtspr  SPR_EX_CONTEXT_K_0, r2
258	 addi   r0, r0, 8
259	}
260
261	/*
262	 * Set ICS on kernel EX_CONTEXT_K_1 in order to "iret" to
263	 * user with ICS set. This way, if the JIT fixup causes another
264	 * unalign exception (which shouldn't be possible) the user
265	 * process will be terminated with SIGBUS. Also, our fixup will
266	 * run without interleaving with external interrupts.
267	 * Each fixup is at most 14 bundles, so it won't hold ICS for long.
268	 */
269	{
270	 movei  r1, PL_ICS_EX1(USER_PL, 1)
271	 mtspr  SPR_EX_CONTEXT_0_0, r0
272	}
273
274	{
275	 mtspr  SPR_EX_CONTEXT_K_1, r1
276	 addi   r3, r3, -(3 * 8)
277	}
278
279	/* Restore r0..r3. */
280	ld_add  r0, r3, 8
281	ld_add  r1, r3, 8
282	ld_add  r2, r3, 8
283	ld      r3, r3
284
285	iret
286	ENDPROC(intvec_\vecname)
287	.endm
288
289#ifdef __COLLECT_LINKER_FEEDBACK__
290	.pushsection .text.intvec_feedback,"ax"
291intvec_feedback:
292	.popsection
293#endif
294
295	/*
296	 * Default interrupt handler.
297	 *
298	 * vecnum is where we'll put this code.
299	 * c_routine is the C routine we'll call.
300	 *
301	 * The C routine is passed two arguments:
302	 * - A pointer to the pt_regs state.
303	 * - The interrupt vector number.
304	 *
305	 * The "processing" argument specifies the code for processing
306	 * the interrupt. Defaults to "handle_interrupt".
307	 */
308	.macro __int_hand vecnum, vecname, c_routine,processing=handle_interrupt
309intvec_\vecname:
310	/* Temporarily save a register so we have somewhere to work. */
311
312	mtspr   SPR_SYSTEM_SAVE_K_1, r0
313	mfspr   r0, SPR_EX_CONTEXT_K_1
314
315	/*
316	 * The unalign data fastpath code sets the low bit in sp to
317	 * force us to reset it here on fault.
318	 */
319	{
320	 blbs   sp, 2f
321	 IS_KERNEL_EX1(r0, r0)
322	}
323
324	.ifc    \vecnum, INT_DOUBLE_FAULT
325	/*
326	 * For double-faults from user-space, fall through to the normal
327	 * register save and stack setup path.  Otherwise, it's the
328	 * hypervisor giving us one last chance to dump diagnostics, and we
329	 * branch to the kernel_double_fault routine to do so.
330	 */
331	beqz    r0, 1f
332	j       _kernel_double_fault
3331:
334	.else
335	/*
336	 * If we're coming from user-space, then set sp to the top of
337	 * the kernel stack.  Otherwise, assume sp is already valid.
338	 */
339	{
340	 bnez   r0, 0f
341	 move   r0, sp
342	}
343	.endif
344
345	.ifc    \c_routine, do_page_fault
346	/*
347	 * The page_fault handler may be downcalled directly by the
348	 * hypervisor even when Linux is running and has ICS set.
349	 *
350	 * In this case the contents of EX_CONTEXT_K_1 reflect the
351	 * previous fault and can't be relied on to choose whether or
352	 * not to reinitialize the stack pointer.  So we add a test
353	 * to see whether SYSTEM_SAVE_K_2 has the high bit set,
354	 * and if so we don't reinitialize sp, since we must be coming
355	 * from Linux.  (In fact the precise case is !(val & ~1),
356	 * but any Linux PC has to have the high bit set.)
357	 *
358	 * Note that the hypervisor *always* sets SYSTEM_SAVE_K_2 for
359	 * any path that turns into a downcall to one of our TLB handlers.
360	 *
361	 * FIXME: if we end up never using this path, perhaps we should
362	 * prevent the hypervisor from generating downcalls in this case.
363	 * The advantage of getting a downcall is we can panic in Linux.
364	 */
365	mfspr   r0, SPR_SYSTEM_SAVE_K_2
366	{
367	 bltz   r0, 0f    /* high bit in S_S_1_2 is for a PC to use */
368	 move   r0, sp
369	}
370	.endif
371
3722:
373	/*
374	 * SYSTEM_SAVE_K_0 holds the cpu number in the high bits, and
375	 * the current stack top in the lower bits.  So we recover
376	 * our starting stack value by sign-extending the low bits, then
377	 * point sp at the top aligned address on the actual stack page.
378	 */
379	mfspr   r0, SPR_SYSTEM_SAVE_K_0
380	bfexts  r0, r0, 0, CPU_SHIFT-1
381
3820:
383	/*
384	 * Align the stack mod 64 so we can properly predict what
385	 * cache lines we need to write-hint to reduce memory fetch
386	 * latency as we enter the kernel.  The layout of memory is
387	 * as follows, with cache line 0 at the lowest VA, and cache
388	 * line 8 just below the r0 value this "andi" computes.
389	 * Note that we never write to cache line 8, and we skip
390	 * cache lines 1-3 for syscalls.
391	 *
392	 *    cache line 8: ptregs padding (two words)
393	 *    cache line 7: sp, lr, pc, ex1, faultnum, orig_r0, flags, cmpexch
394	 *    cache line 6: r46...r53 (tp)
395	 *    cache line 5: r38...r45
396	 *    cache line 4: r30...r37
397	 *    cache line 3: r22...r29
398	 *    cache line 2: r14...r21
399	 *    cache line 1: r6...r13
400	 *    cache line 0: 2 x frame, r0..r5
401	 */
402#if STACK_TOP_DELTA != 64
403#error STACK_TOP_DELTA must be 64 for assumptions here and in task_pt_regs()
404#endif
405	andi    r0, r0, -64
406
407	/*
408	 * Push the first four registers on the stack, so that we can set
409	 * them to vector-unique values before we jump to the common code.
410	 *
411	 * Registers are pushed on the stack as a struct pt_regs,
412	 * with the sp initially just above the struct, and when we're
413	 * done, sp points to the base of the struct, minus
414	 * C_ABI_SAVE_AREA_SIZE, so we can directly jal to C code.
415	 *
416	 * This routine saves just the first four registers, plus the
417	 * stack context so we can do proper backtracing right away,
418	 * and defers to handle_interrupt to save the rest.
419	 * The backtracer needs pc, ex1, lr, sp, r52, and faultnum,
420	 * and needs sp set to its final location at the bottom of
421	 * the stack frame.
422	 */
423	addli   r0, r0, PTREGS_OFFSET_LR - (PTREGS_SIZE + KSTK_PTREGS_GAP)
424	wh64    r0   /* cache line 7 */
425	{
426	 st     r0, lr
427	 addli  r0, r0, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
428	}
429	{
430	 st     r0, sp
431	 addli  sp, r0, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_SP
432	}
433	wh64    sp   /* cache line 6 */
434	{
435	 st     sp, r52
436	 addli  sp, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(52)
437	}
438	wh64    sp   /* cache line 0 */
439	{
440	 st     sp, r1
441	 addli  sp, sp, PTREGS_OFFSET_REG(2) - PTREGS_OFFSET_REG(1)
442	}
443	{
444	 st     sp, r2
445	 addli  sp, sp, PTREGS_OFFSET_REG(3) - PTREGS_OFFSET_REG(2)
446	}
447	{
448	 st     sp, r3
449	 addli  sp, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(3)
450	}
451	mfspr   r0, SPR_EX_CONTEXT_K_0
452	.ifc \processing,handle_syscall
453	/*
454	 * Bump the saved PC by one bundle so that when we return, we won't
455	 * execute the same swint instruction again.  We need to do this while
456	 * we're in the critical section.
457	 */
458	addi    r0, r0, 8
459	.endif
460	{
461	 st     sp, r0
462	 addli  sp, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
463	}
464	mfspr   r0, SPR_EX_CONTEXT_K_1
465	{
466	 st     sp, r0
467	 addi   sp, sp, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
468	/*
469	 * Use r0 for syscalls so it's a temporary; use r1 for interrupts
470	 * so that it gets passed through unchanged to the handler routine.
471	 * Note that the .if conditional confusingly spans bundles.
472	 */
473	 .ifc \processing,handle_syscall
474	 movei  r0, \vecnum
475	}
476	{
477	 st     sp, r0
478	 .else
479	 movei  r1, \vecnum
480	}
481	{
482	 st     sp, r1
483	 .endif
484	 addli  sp, sp, PTREGS_OFFSET_REG(0) - PTREGS_OFFSET_FAULTNUM
485	}
486	mfspr   r0, SPR_SYSTEM_SAVE_K_1    /* Original r0 */
487	{
488	 st     sp, r0
489	 addi   sp, sp, -PTREGS_OFFSET_REG(0) - 8
490	}
491	{
492	 st     sp, zero        /* write zero into "Next SP" frame pointer */
493	 addi   sp, sp, -8      /* leave SP pointing at bottom of frame */
494	}
495	.ifc \processing,handle_syscall
496	j       handle_syscall
497	.else
498	/* Capture per-interrupt SPR context to registers. */
499	.ifc \c_routine, do_page_fault
500	mfspr   r2, SPR_SYSTEM_SAVE_K_3   /* address of page fault */
501	mfspr   r3, SPR_SYSTEM_SAVE_K_2   /* info about page fault */
502	.else
503	.ifc \vecnum, INT_ILL_TRANS
504	mfspr   r2, ILL_VA_PC
505	.else
506	.ifc \vecnum, INT_DOUBLE_FAULT
507	mfspr   r2, SPR_SYSTEM_SAVE_K_2   /* double fault info from HV */
508	.else
509	.ifc \c_routine, do_trap
510	mfspr   r2, GPV_REASON
511	.else
512	.ifc \c_routine, handle_perf_interrupt
513	mfspr   r2, PERF_COUNT_STS
514	.else
515	.ifc \c_routine, handle_perf_interrupt
516	mfspr   r2, AUX_PERF_COUNT_STS
517	.endif
518	.ifc \c_routine, do_nmi
519	mfspr   r2, SPR_SYSTEM_SAVE_K_2   /* nmi type */
520	.else
521	.endif
522	.endif
523	.endif
524	.endif
525	.endif
526	.endif
527	/* Put function pointer in r0 */
528	moveli  r0, hw2_last(\c_routine)
529	shl16insli r0, r0, hw1(\c_routine)
530	{
531	 shl16insli r0, r0, hw0(\c_routine)
532	 j       \processing
533	}
534	.endif
535	ENDPROC(intvec_\vecname)
536
537#ifdef __COLLECT_LINKER_FEEDBACK__
538	.pushsection .text.intvec_feedback,"ax"
539	.org    (\vecnum << 5)
540	FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt, 1 << 8)
541	jrp     lr
542	.popsection
543#endif
544
545	.endm
546
547
548	/*
549	 * Save the rest of the registers that we didn't save in the actual
550	 * vector itself.  We can't use r0-r10 inclusive here.
551	 */
552	.macro  finish_interrupt_save, function
553
554	/* If it's a syscall, save a proper orig_r0, otherwise just zero. */
555	PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
556	{
557	 .ifc \function,handle_syscall
558	 st     r52, r0
559	 .else
560	 st     r52, zero
561	 .endif
562	 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
563	}
564	st      r52, tp
565	{
566	 mfspr  tp, CMPEXCH_VALUE
567	 PTREGS_PTR(r52, PTREGS_OFFSET_CMPEXCH)
568	}
569
570	/*
571	 * For ordinary syscalls, we save neither caller- nor callee-
572	 * save registers, since the syscall invoker doesn't expect the
573	 * caller-saves to be saved, and the called kernel functions will
574	 * take care of saving the callee-saves for us.
575	 *
576	 * For interrupts we save just the caller-save registers.  Saving
577	 * them is required (since the "caller" can't save them).  Again,
578	 * the called kernel functions will restore the callee-save
579	 * registers for us appropriately.
580	 *
581	 * On return, we normally restore nothing special for syscalls,
582	 * and just the caller-save registers for interrupts.
583	 *
584	 * However, there are some important caveats to all this:
585	 *
586	 * - We always save a few callee-save registers to give us
587	 *   some scratchpad registers to carry across function calls.
588	 *
589	 * - fork/vfork/etc require us to save all the callee-save
590	 *   registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
591	 *
592	 * - We always save r0..r5 and r10 for syscalls, since we need
593	 *   to reload them a bit later for the actual kernel call, and
594	 *   since we might need them for -ERESTARTNOINTR, etc.
595	 *
596	 * - Before invoking a signal handler, we save the unsaved
597	 *   callee-save registers so they are visible to the
598	 *   signal handler or any ptracer.
599	 *
600	 * - If the unsaved callee-save registers are modified, we set
601	 *   a bit in pt_regs so we know to reload them from pt_regs
602	 *   and not just rely on the kernel function unwinding.
603	 *   (Done for ptrace register writes and SA_SIGINFO handler.)
604	 */
605	{
606	 st     r52, tp
607	 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
608	}
609	wh64    r52    /* cache line 4 */
610	push_reg r33, r52
611	push_reg r32, r52
612	push_reg r31, r52
613	.ifc \function,handle_syscall
614	push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
615	push_reg TREG_SYSCALL_NR_NAME, r52, \
616	  PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
617	.else
618
619	push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
620	wh64    r52   /* cache line 3 */
621	push_reg r29, r52
622	push_reg r28, r52
623	push_reg r27, r52
624	push_reg r26, r52
625	push_reg r25, r52
626	push_reg r24, r52
627	push_reg r23, r52
628	push_reg r22, r52
629	wh64    r52   /* cache line 2 */
630	push_reg r21, r52
631	push_reg r20, r52
632	push_reg r19, r52
633	push_reg r18, r52
634	push_reg r17, r52
635	push_reg r16, r52
636	push_reg r15, r52
637	push_reg r14, r52
638	wh64    r52   /* cache line 1 */
639	push_reg r13, r52
640	push_reg r12, r52
641	push_reg r11, r52
642	push_reg r10, r52
643	push_reg r9, r52
644	push_reg r8, r52
645	push_reg r7, r52
646	push_reg r6, r52
647
648	.endif
649
650	push_reg r5, r52
651	st      r52, r4
652
653	/*
654	 * If we will be returning to the kernel, we will need to
655	 * reset the interrupt masks to the state they had before.
656	 * Set DISABLE_IRQ in flags iff we came from kernel pl with
657	 * irqs disabled.
658	 */
659	mfspr   r32, SPR_EX_CONTEXT_K_1
660	{
661	 IS_KERNEL_EX1(r32, r32)
662	 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
663	}
664	beqzt   r32, 1f       /* zero if from user space */
665	IRQS_DISABLED(r32)    /* zero if irqs enabled */
666#if PT_FLAGS_DISABLE_IRQ != 1
667# error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
668#endif
6691:
670	.ifnc \function,handle_syscall
671	/* Record the fact that we saved the caller-save registers above. */
672	ori     r32, r32, PT_FLAGS_CALLER_SAVES
673	.endif
674	st      r21, r32
675
676	/*
677	 * we've captured enough state to the stack (including in
678	 * particular our EX_CONTEXT state) that we can now release
679	 * the interrupt critical section and replace it with our
680	 * standard "interrupts disabled" mask value.  This allows
681	 * synchronous interrupts (and profile interrupts) to punch
682	 * through from this point onwards.
683	 *
684	 * It's important that no code before this point touch memory
685	 * other than our own stack (to keep the invariant that this
686	 * is all that gets touched under ICS), and that no code after
687	 * this point reference any interrupt-specific SPR, in particular
688	 * the EX_CONTEXT_K_ values.
689	 */
690	.ifc \function,handle_nmi
691	IRQ_DISABLE_ALL(r20)
692	.else
693	IRQ_DISABLE(r20, r21)
694	.endif
695	mtspr   INTERRUPT_CRITICAL_SECTION, zero
696
697	/* Load tp with our per-cpu offset. */
698#ifdef CONFIG_SMP
699	{
700	 mfspr  r20, SPR_SYSTEM_SAVE_K_0
701	 moveli r21, hw2_last(__per_cpu_offset)
702	}
703	{
704	 shl16insli r21, r21, hw1(__per_cpu_offset)
705	 bfextu r20, r20, CPU_SHIFT, 63
706	}
707	shl16insli r21, r21, hw0(__per_cpu_offset)
708	shl3add r20, r20, r21
709	ld      tp, r20
710#else
711	move    tp, zero
712#endif
713
714#ifdef __COLLECT_LINKER_FEEDBACK__
715	/*
716	 * Notify the feedback routines that we were in the
717	 * appropriate fixed interrupt vector area.  Note that we
718	 * still have ICS set at this point, so we can't invoke any
719	 * atomic operations or we will panic.  The feedback
720	 * routines internally preserve r0..r10 and r30 up.
721	 */
722	.ifnc \function,handle_syscall
723	shli    r20, r1, 5
724	.else
725	moveli  r20, INT_SWINT_1 << 5
726	.endif
727	moveli  r21, hw2_last(intvec_feedback)
728	shl16insli r21, r21, hw1(intvec_feedback)
729	shl16insli r21, r21, hw0(intvec_feedback)
730	add     r20, r20, r21
731	jalr    r20
732
733	/* And now notify the feedback routines that we are here. */
734	FEEDBACK_ENTER(\function)
735#endif
736
737	/*
738	 * Prepare the first 256 stack bytes to be rapidly accessible
739	 * without having to fetch the background data.
740	 */
741	addi    r52, sp, -64
742	{
743	 wh64   r52
744	 addi   r52, r52, -64
745	}
746	{
747	 wh64   r52
748	 addi   r52, r52, -64
749	}
750	{
751	 wh64   r52
752	 addi   r52, r52, -64
753	}
754	wh64    r52
755
756#if defined(CONFIG_TRACE_IRQFLAGS) || defined(CONFIG_CONTEXT_TRACKING)
757	.ifnc \function,handle_nmi
758	/*
759	 * We finally have enough state set up to notify the irq
760	 * tracing code that irqs were disabled on entry to the handler.
761	 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
762	 * For syscalls, we already have the register state saved away
763	 * on the stack, so we don't bother to do any register saves here,
764	 * and later we pop the registers back off the kernel stack.
765	 * For interrupt handlers, save r0-r3 in callee-saved registers.
766	 */
767	.ifnc \function,handle_syscall
768	{ move r30, r0; move r31, r1 }
769	{ move r32, r2; move r33, r3 }
770	.endif
771	TRACE_IRQS_OFF
772#ifdef CONFIG_CONTEXT_TRACKING
773	jal     context_tracking_user_exit
774#endif
775	.ifnc \function,handle_syscall
776	{ move r0, r30; move r1, r31 }
777	{ move r2, r32; move r3, r33 }
778	.endif
779	.endif
780#endif
781
782	.endm
783
784	/*
785	 * Redispatch a downcall.
786	 */
787	.macro  dc_dispatch vecnum, vecname
788	.org    (\vecnum << 8)
789intvec_\vecname:
790	j       _hv_downcall_dispatch
791	ENDPROC(intvec_\vecname)
792	.endm
793
794	/*
795	 * Common code for most interrupts.  The C function we're eventually
796	 * going to is in r0, and the faultnum is in r1; the original
797	 * values for those registers are on the stack.
798	 */
799	.pushsection .text.handle_interrupt,"ax"
800handle_interrupt:
801	finish_interrupt_save handle_interrupt
802
803	/* Jump to the C routine; it should enable irqs as soon as possible. */
804	{
805	 jalr   r0
806	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
807	}
808	FEEDBACK_REENTER(handle_interrupt)
809	{
810	 movei  r30, 0   /* not an NMI */
811	 j      interrupt_return
812	}
813	STD_ENDPROC(handle_interrupt)
814
815/*
816 * This routine takes a boolean in r30 indicating if this is an NMI.
817 * If so, we also expect a boolean in r31 indicating whether to
818 * re-enable the oprofile interrupts.
819 *
820 * Note that .Lresume_userspace is jumped to directly in several
821 * places, and we need to make sure r30 is set correctly in those
822 * callers as well.
823 */
824STD_ENTRY(interrupt_return)
825	/* If we're resuming to kernel space, don't check thread flags. */
826	{
827	 bnez   r30, .Lrestore_all  /* NMIs don't special-case user-space */
828	 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
829	}
830	ld      r29, r29
831	IS_KERNEL_EX1(r29, r29)
832	{
833	 beqzt  r29, .Lresume_userspace
834	 move   r29, sp
835	}
836
837#ifdef CONFIG_PREEMPT
838	/* Returning to kernel space. Check if we need preemption. */
839	EXTRACT_THREAD_INFO(r29)
840	addli   r28, r29, THREAD_INFO_FLAGS_OFFSET
841	{
842	 ld     r28, r28
843	 addli  r29, r29, THREAD_INFO_PREEMPT_COUNT_OFFSET
844	}
845	{
846	 andi   r28, r28, _TIF_NEED_RESCHED
847	 ld4s   r29, r29
848	}
849	beqzt   r28, 1f
850	bnez    r29, 1f
851	/* Disable interrupts explicitly for preemption. */
852	IRQ_DISABLE(r20,r21)
853	TRACE_IRQS_OFF
854	jal     preempt_schedule_irq
855	FEEDBACK_REENTER(interrupt_return)
8561:
857#endif
858
859	/* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
860	{
861	 moveli r27, hw2_last(_cpu_idle_nap)
862	 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
863	}
864	{
865	 ld     r28, r29
866	 shl16insli r27, r27, hw1(_cpu_idle_nap)
867	}
868	{
869	 shl16insli r27, r27, hw0(_cpu_idle_nap)
870	}
871	{
872	 cmpeq  r27, r27, r28
873	}
874	{
875	 blbc   r27, .Lrestore_all
876	 addi   r28, r28, 8
877	}
878	st      r29, r28
879	j       .Lrestore_all
880
881.Lresume_userspace:
882	FEEDBACK_REENTER(interrupt_return)
883
884	/*
885	 * Disable interrupts so as to make sure we don't
886	 * miss an interrupt that sets any of the thread flags (like
887	 * need_resched or sigpending) between sampling and the iret.
888	 * Routines like schedule() or do_signal() may re-enable
889	 * interrupts before returning.
890	 */
891	IRQ_DISABLE(r20, r21)
892	TRACE_IRQS_OFF  /* Note: clobbers registers r0-r29 */
893
894	/*
895	 * See if there are any work items (including single-shot items)
896	 * to do.  If so, save the callee-save registers to pt_regs
897	 * and then dispatch to C code.
898	 */
899	move    r21, sp
900	EXTRACT_THREAD_INFO(r21)
901	{
902	 addi   r22, r21, THREAD_INFO_FLAGS_OFFSET
903	 moveli r20, hw1_last(_TIF_ALLWORK_MASK)
904	}
905	{
906	 ld     r22, r22
907	 shl16insli r20, r20, hw0(_TIF_ALLWORK_MASK)
908	}
909	and     r1, r22, r20
910	{
911	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
912	 beqzt  r1, .Lrestore_all
913	}
914	push_extra_callee_saves r0
915	jal     prepare_exit_to_usermode
916
917	/*
918	 * In the NMI case we
919	 * omit the call to single_process_check_nohz, which normally checks
920	 * to see if we should start or stop the scheduler tick, because
921	 * we can't call arbitrary Linux code from an NMI context.
922	 * We always call the homecache TLB deferral code to re-trigger
923	 * the deferral mechanism.
924	 *
925	 * The other chunk of responsibility this code has is to reset the
926	 * interrupt masks appropriately to reset irqs and NMIs.  We have
927	 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
928	 * lockdep-type stuff, but we can't set ICS until afterwards, since
929	 * ICS can only be used in very tight chunks of code to avoid
930	 * tripping over various assertions that it is off.
931	 */
932.Lrestore_all:
933	PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
934	{
935	 ld      r0, r0
936	 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
937	}
938	{
939	 IS_KERNEL_EX1(r0, r0)
940	 ld     r32, r32
941	}
942	bnez    r0, 1f
943	j       2f
944#if PT_FLAGS_DISABLE_IRQ != 1
945# error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use blbct below
946#endif
9471:	blbct   r32, 2f
948	IRQ_DISABLE(r20,r21)
949	TRACE_IRQS_OFF
950	movei   r0, 1
951	mtspr   INTERRUPT_CRITICAL_SECTION, r0
952	beqzt   r30, .Lrestore_regs
953	j       3f
9542:	TRACE_IRQS_ON
955	IRQ_ENABLE_LOAD(r20, r21)
956	movei   r0, 1
957	mtspr   INTERRUPT_CRITICAL_SECTION, r0
958	IRQ_ENABLE_APPLY(r20, r21)
959	beqzt   r30, .Lrestore_regs
9603:
961
962#if INT_PERF_COUNT + 1 != INT_AUX_PERF_COUNT
963# error Bad interrupt assumption
964#endif
965	{
966	 movei  r0, 3   /* two adjacent bits for the PERF_COUNT mask */
967	 beqz   r31, .Lrestore_regs
968	}
969	shli    r0, r0, INT_PERF_COUNT
970	mtspr   SPR_INTERRUPT_MASK_RESET_K, r0
971
972	/*
973	 * We now commit to returning from this interrupt, since we will be
974	 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
975	 * frame.  No calls should be made to any other code after this point.
976	 * This code should only be entered with ICS set.
977	 * r32 must still be set to ptregs.flags.
978	 * We launch loads to each cache line separately first, so we can
979	 * get some parallelism out of the memory subsystem.
980	 * We start zeroing caller-saved registers throughout, since
981	 * that will save some cycles if this turns out to be a syscall.
982	 */
983.Lrestore_regs:
984
985	/*
986	 * Rotate so we have one high bit and one low bit to test.
987	 * - low bit says whether to restore all the callee-saved registers,
988	 *   or just r30-r33, and r52 up.
989	 * - high bit (i.e. sign bit) says whether to restore all the
990	 *   caller-saved registers, or just r0.
991	 */
992#if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
993# error Rotate trick does not work :-)
994#endif
995	{
996	 rotli  r20, r32, 62
997	 PTREGS_PTR(sp, PTREGS_OFFSET_REG(0))
998	}
999
1000	/*
1001	 * Load cache lines 0, 4, 6 and 7, in that order, then use
1002	 * the last loaded value, which makes it likely that the other
1003	 * cache lines have also loaded, at which point we should be
1004	 * able to safely read all the remaining words on those cache
1005	 * lines without waiting for the memory subsystem.
1006	 */
1007	pop_reg r0, sp, PTREGS_OFFSET_REG(30) - PTREGS_OFFSET_REG(0)
1008	pop_reg r30, sp, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_REG(30)
1009	pop_reg_zero r52, r3, sp, PTREGS_OFFSET_CMPEXCH - PTREGS_OFFSET_REG(52)
1010	pop_reg_zero r21, r27, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_CMPEXCH
1011	pop_reg_zero lr, r2, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_EX1
1012	{
1013	 mtspr  CMPEXCH_VALUE, r21
1014	 move   r4, zero
1015	}
1016	pop_reg r21, sp, PTREGS_OFFSET_REG(31) - PTREGS_OFFSET_PC
1017	{
1018	 mtspr  SPR_EX_CONTEXT_K_1, lr
1019	 IS_KERNEL_EX1(lr, lr)
1020	}
1021	{
1022	 mtspr  SPR_EX_CONTEXT_K_0, r21
1023	 move   r5, zero
1024	}
1025
1026	/* Restore callee-saveds that we actually use. */
1027	pop_reg_zero r31, r6
1028	pop_reg_zero r32, r7
1029	pop_reg_zero r33, r8, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(33)
1030
1031	/*
1032	 * If we modified other callee-saveds, restore them now.
1033	 * This is rare, but could be via ptrace or signal handler.
1034	 */
1035	{
1036	 move   r9, zero
1037	 blbs   r20, .Lrestore_callees
1038	}
1039.Lcontinue_restore_regs:
1040
1041	/* Check if we're returning from a syscall. */
1042	{
1043	 move   r10, zero
1044	 bltzt  r20, 1f  /* no, so go restore callee-save registers */
1045	}
1046
1047	/*
1048	 * Check if we're returning to userspace.
1049	 * Note that if we're not, we don't worry about zeroing everything.
1050	 */
1051	{
1052	 addli  sp, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(29)
1053	 bnez   lr, .Lkernel_return
1054	}
1055
1056	/*
1057	 * On return from syscall, we've restored r0 from pt_regs, but we
1058	 * clear the remainder of the caller-saved registers.  We could
1059	 * restore the syscall arguments, but there's not much point,
1060	 * and it ensures user programs aren't trying to use the
1061	 * caller-saves if we clear them, as well as avoiding leaking
1062	 * kernel pointers into userspace.
1063	 */
1064	pop_reg_zero lr, r11, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1065	pop_reg_zero tp, r12, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1066	{
1067	 ld     sp, sp
1068	 move   r13, zero
1069	 move   r14, zero
1070	}
1071	{ move r15, zero; move r16, zero }
1072	{ move r17, zero; move r18, zero }
1073	{ move r19, zero; move r20, zero }
1074	{ move r21, zero; move r22, zero }
1075	{ move r23, zero; move r24, zero }
1076	{ move r25, zero; move r26, zero }
1077
1078	/* Set r1 to errno if we are returning an error, otherwise zero. */
1079	{
1080	 moveli r29, 4096
1081	 sub    r1, zero, r0
1082	}
1083	{
1084	 move   r28, zero
1085	 cmpltu r29, r1, r29
1086	}
1087	{
1088	 mnz    r1, r29, r1
1089	 move   r29, zero
1090	}
1091	iret
1092
1093	/*
1094	 * Not a syscall, so restore caller-saved registers.
1095	 * First kick off loads for cache lines 1-3, which we're touching
1096	 * for the first time here.
1097	 */
1098	.align 64
10991:	pop_reg r29, sp, PTREGS_OFFSET_REG(21) - PTREGS_OFFSET_REG(29)
1100	pop_reg r21, sp, PTREGS_OFFSET_REG(13) - PTREGS_OFFSET_REG(21)
1101	pop_reg r13, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(13)
1102	pop_reg r1
1103	pop_reg r2
1104	pop_reg r3
1105	pop_reg r4
1106	pop_reg r5
1107	pop_reg r6
1108	pop_reg r7
1109	pop_reg r8
1110	pop_reg r9
1111	pop_reg r10
1112	pop_reg r11
1113	pop_reg r12, sp, 16
1114	/* r13 already restored above */
1115	pop_reg r14
1116	pop_reg r15
1117	pop_reg r16
1118	pop_reg r17
1119	pop_reg r18
1120	pop_reg r19
1121	pop_reg r20, sp, 16
1122	/* r21 already restored above */
1123	pop_reg r22
1124	pop_reg r23
1125	pop_reg r24
1126	pop_reg r25
1127	pop_reg r26
1128	pop_reg r27
1129	pop_reg r28, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(28)
1130	/* r29 already restored above */
1131	bnez    lr, .Lkernel_return
1132	pop_reg lr, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1133	pop_reg tp, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1134	ld      sp, sp
1135	iret
1136
1137	/*
1138	 * We can't restore tp when in kernel mode, since a thread might
1139	 * have migrated from another cpu and brought a stale tp value.
1140	 */
1141.Lkernel_return:
1142	pop_reg lr, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
1143	ld      sp, sp
1144	iret
1145
1146	/* Restore callee-saved registers from r34 to r51. */
1147.Lrestore_callees:
1148	addli  sp, sp, PTREGS_OFFSET_REG(34) - PTREGS_OFFSET_REG(29)
1149	pop_reg r34
1150	pop_reg r35
1151	pop_reg r36
1152	pop_reg r37
1153	pop_reg r38
1154	pop_reg r39
1155	pop_reg r40
1156	pop_reg r41
1157	pop_reg r42
1158	pop_reg r43
1159	pop_reg r44
1160	pop_reg r45
1161	pop_reg r46
1162	pop_reg r47
1163	pop_reg r48
1164	pop_reg r49
1165	pop_reg r50
1166	pop_reg r51, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(51)
1167	j .Lcontinue_restore_regs
1168	STD_ENDPROC(interrupt_return)
1169
1170	/*
1171	 * "NMI" interrupts mask ALL interrupts before calling the
1172	 * handler, and don't check thread flags, etc., on the way
1173	 * back out.  In general, the only things we do here for NMIs
1174	 * are register save/restore and dataplane kernel-TLB management.
1175	 * We don't (for example) deal with start/stop of the sched tick.
1176	 */
1177	.pushsection .text.handle_nmi,"ax"
1178handle_nmi:
1179	finish_interrupt_save handle_nmi
1180	{
1181	 jalr   r0
1182	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1183	}
1184	FEEDBACK_REENTER(handle_nmi)
1185	{
1186	 movei  r30, 1
1187	 cmpeq  r31, r0, zero
1188	}
1189	j       interrupt_return
1190	STD_ENDPROC(handle_nmi)
1191
1192	/*
1193	 * Parallel code for syscalls to handle_interrupt.
1194	 */
1195	.pushsection .text.handle_syscall,"ax"
1196handle_syscall:
1197	finish_interrupt_save handle_syscall
1198
1199	/* Enable irqs. */
1200	TRACE_IRQS_ON
1201	IRQ_ENABLE(r20, r21)
1202
1203	/* Bump the counter for syscalls made on this tile. */
1204	moveli r20, hw2_last(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1205	shl16insli r20, r20, hw1(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1206	shl16insli r20, r20, hw0(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1207	add     r20, r20, tp
1208	ld4s    r21, r20
1209	{
1210	 addi   r21, r21, 1
1211	 move   r31, sp
1212	}
1213	{
1214	 st4    r20, r21
1215	 EXTRACT_THREAD_INFO(r31)
1216	}
1217
1218	/* Trace syscalls, if requested. */
1219	addi	r31, r31, THREAD_INFO_FLAGS_OFFSET
1220	{
1221	 ld     r30, r31
1222	 moveli r32, _TIF_SYSCALL_ENTRY_WORK
1223	}
1224	and     r30, r30, r32
1225	{
1226	 addi   r30, r31, THREAD_INFO_STATUS_OFFSET - THREAD_INFO_FLAGS_OFFSET
1227	 beqzt	r30, .Lrestore_syscall_regs
1228	}
1229	{
1230	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1231	 jal    do_syscall_trace_enter
1232	}
1233	FEEDBACK_REENTER(handle_syscall)
1234	bltz    r0, .Lsyscall_sigreturn_skip
1235
1236	/*
1237	 * We always reload our registers from the stack at this
1238	 * point.  They might be valid, if we didn't build with
1239	 * TRACE_IRQFLAGS, and this isn't a dataplane tile, and we're not
1240	 * doing syscall tracing, but there are enough cases now that it
1241	 * seems simplest just to do the reload unconditionally.
1242	 */
1243.Lrestore_syscall_regs:
1244	{
1245	 ld     r30, r30
1246	 PTREGS_PTR(r11, PTREGS_OFFSET_REG(0))
1247	}
1248	pop_reg r0,  r11
1249	pop_reg r1,  r11
1250	pop_reg r2,  r11
1251	pop_reg r3,  r11
1252	pop_reg r4,  r11
1253	pop_reg r5,  r11, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(5)
1254	{
1255	 ld     TREG_SYSCALL_NR_NAME, r11
1256	 moveli r21, __NR_syscalls
1257	}
1258
1259	/* Ensure that the syscall number is within the legal range. */
1260	{
1261	 moveli r20, hw2(sys_call_table)
1262#ifdef CONFIG_COMPAT
1263	 blbs   r30, .Lcompat_syscall
1264#endif
1265	}
1266	{
1267	 cmpltu r21, TREG_SYSCALL_NR_NAME, r21
1268	 shl16insli r20, r20, hw1(sys_call_table)
1269	}
1270	{
1271	 blbc   r21, .Linvalid_syscall
1272	 shl16insli r20, r20, hw0(sys_call_table)
1273	}
1274.Lload_syscall_pointer:
1275	shl3add r20, TREG_SYSCALL_NR_NAME, r20
1276	ld      r20, r20
1277
1278	/* Jump to syscall handler. */
1279	jalr    r20
1280.Lhandle_syscall_link: /* value of "lr" after "jalr r20" above */
1281
1282	/*
1283	 * Write our r0 onto the stack so it gets restored instead
1284	 * of whatever the user had there before.
1285	 * In compat mode, sign-extend r0 before storing it.
1286	 */
1287	{
1288	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1289	 blbct  r30, 1f
1290	}
1291	addxi   r0, r0, 0
12921:	st      r29, r0
1293
1294.Lsyscall_sigreturn_skip:
1295	FEEDBACK_REENTER(handle_syscall)
1296
1297	/* Do syscall trace again, if requested. */
1298	{
1299	 ld      r30, r31
1300	 moveli  r32, _TIF_SYSCALL_EXIT_WORK
1301	}
1302	and      r0, r30, r32
1303	{
1304	 andi    r0, r30, _TIF_SINGLESTEP
1305	 beqzt   r0, 1f
1306	}
1307	{
1308	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1309	 jal    do_syscall_trace_exit
1310	}
1311	FEEDBACK_REENTER(handle_syscall)
1312	andi    r0, r30, _TIF_SINGLESTEP
1313
13141:	beqzt	r0, 2f
1315
1316	/* Single stepping -- notify ptrace. */
1317	{
1318	 movei   r0, SIGTRAP
1319	 jal     ptrace_notify
1320	}
1321	FEEDBACK_REENTER(handle_syscall)
1322
13232:	{
1324	 movei  r30, 0               /* not an NMI */
1325	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1326	}
1327
1328#ifdef CONFIG_COMPAT
1329.Lcompat_syscall:
1330	/*
1331	 * Load the base of the compat syscall table in r20, and
1332	 * range-check the syscall number (duplicated from 64-bit path).
1333	 * Sign-extend all the user's passed arguments to make them consistent.
1334	 * Also save the original "r(n)" values away in "r(11+n)" in
1335	 * case the syscall table entry wants to validate them.
1336	 */
1337	moveli  r20, hw2(compat_sys_call_table)
1338	{
1339	 cmpltu r21, TREG_SYSCALL_NR_NAME, r21
1340	 shl16insli r20, r20, hw1(compat_sys_call_table)
1341	}
1342	{
1343	 blbc   r21, .Linvalid_syscall
1344	 shl16insli r20, r20, hw0(compat_sys_call_table)
1345	}
1346	{ move r11, r0; addxi r0, r0, 0 }
1347	{ move r12, r1; addxi r1, r1, 0 }
1348	{ move r13, r2; addxi r2, r2, 0 }
1349	{ move r14, r3; addxi r3, r3, 0 }
1350	{ move r15, r4; addxi r4, r4, 0 }
1351	{ move r16, r5; addxi r5, r5, 0 }
1352	j .Lload_syscall_pointer
1353#endif
1354
1355.Linvalid_syscall:
1356	/* Report an invalid syscall back to the user program */
1357	{
1358	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1359	 movei  r28, -ENOSYS
1360	}
1361	st      r29, r28
1362	{
1363	 movei  r30, 0               /* not an NMI */
1364	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1365	}
1366	STD_ENDPROC(handle_syscall)
1367
1368	/* Return the address for oprofile to suppress in backtraces. */
1369STD_ENTRY_SECTION(handle_syscall_link_address, .text.handle_syscall)
1370	lnk     r0
1371	{
1372	 addli  r0, r0, .Lhandle_syscall_link - .
1373	 jrp    lr
1374	}
1375	STD_ENDPROC(handle_syscall_link_address)
1376
1377STD_ENTRY(ret_from_fork)
1378	jal     sim_notify_fork
1379	jal     schedule_tail
1380	FEEDBACK_REENTER(ret_from_fork)
1381	{
1382	 movei  r30, 0               /* not an NMI */
1383	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1384	}
1385	STD_ENDPROC(ret_from_fork)
1386
1387STD_ENTRY(ret_from_kernel_thread)
1388	jal     sim_notify_fork
1389	jal     schedule_tail
1390	FEEDBACK_REENTER(ret_from_fork)
1391	{
1392	 move   r0, r31
1393	 jalr   r30
1394	}
1395	FEEDBACK_REENTER(ret_from_kernel_thread)
1396	{
1397	 movei  r30, 0               /* not an NMI */
1398	 j      interrupt_return
1399	}
1400	STD_ENDPROC(ret_from_kernel_thread)
1401
1402/* Various stub interrupt handlers and syscall handlers */
1403
1404STD_ENTRY_LOCAL(_kernel_double_fault)
1405	mfspr   r1, SPR_EX_CONTEXT_K_0
1406	move    r2, lr
1407	move    r3, sp
1408	move    r4, r52
1409	addi    sp, sp, -C_ABI_SAVE_AREA_SIZE
1410	j       kernel_double_fault
1411	STD_ENDPROC(_kernel_double_fault)
1412
1413STD_ENTRY_LOCAL(bad_intr)
1414	mfspr   r2, SPR_EX_CONTEXT_K_0
1415	panic   "Unhandled interrupt %#x: PC %#lx"
1416	STD_ENDPROC(bad_intr)
1417
1418/*
1419 * Special-case sigreturn to not write r0 to the stack on return.
1420 * This is technically more efficient, but it also avoids difficulties
1421 * in the 64-bit OS when handling 32-bit compat code, since we must not
1422 * sign-extend r0 for the sigreturn return-value case.
1423 */
1424#define PTREGS_SYSCALL_SIGRETURN(x, reg)                \
1425	STD_ENTRY(_##x);                                \
1426	addli   lr, lr, .Lsyscall_sigreturn_skip - .Lhandle_syscall_link; \
1427	{                                               \
1428	 PTREGS_PTR(reg, PTREGS_OFFSET_BASE);           \
1429	 j      x                                       \
1430	};                                              \
1431	STD_ENDPROC(_##x)
1432
1433PTREGS_SYSCALL_SIGRETURN(sys_rt_sigreturn, r0)
1434#ifdef CONFIG_COMPAT
1435PTREGS_SYSCALL_SIGRETURN(compat_sys_rt_sigreturn, r0)
1436#endif
1437
1438/* Save additional callee-saves to pt_regs and jump to standard function. */
1439STD_ENTRY(_sys_clone)
1440	push_extra_callee_saves r4
1441	j       sys_clone
1442	STD_ENDPROC(_sys_clone)
1443
1444	/*
1445	 * Recover r3, r2, r1 and r0 here saved by unalign fast vector.
1446	 * The vector area limit is 32 bundles, so we handle the reload here.
1447	 * r0, r1, r2 are in thread_info from low to high memory in order.
1448	 * r3 points to location the original r3 was saved.
1449	 * We put this code in the __HEAD section so it can be reached
1450	 * via a conditional branch from the fast path.
1451	 */
1452	__HEAD
1453hand_unalign_slow:
1454	andi    sp, sp, ~1
1455hand_unalign_slow_badsp:
1456	addi    r3, r3, -(3 * 8)
1457	ld_add  r0, r3, 8
1458	ld_add  r1, r3, 8
1459	ld      r2, r3
1460hand_unalign_slow_nonuser:
1461	mfspr   r3, SPR_SYSTEM_SAVE_K_1
1462	__int_hand     INT_UNALIGN_DATA, UNALIGN_DATA_SLOW, int_unalign
1463
1464/* The unaligned data support needs to read all the registers. */
1465int_unalign:
1466	push_extra_callee_saves r0
1467	j       do_unaligned
1468ENDPROC(hand_unalign_slow)
1469
1470/* Fill the return address stack with nonzero entries. */
1471STD_ENTRY(fill_ra_stack)
1472	{
1473	 move	r0, lr
1474	 jal	1f
1475	}
14761:	jal	2f
14772:	jal	3f
14783:	jal	4f
14794:	jrp	r0
1480	STD_ENDPROC(fill_ra_stack)
1481
1482	.macro int_hand  vecnum, vecname, c_routine, processing=handle_interrupt
1483	.org   (\vecnum << 8)
1484		__int_hand   \vecnum, \vecname, \c_routine, \processing
1485	.endm
1486
1487/* Include .intrpt array of interrupt vectors */
1488	.section ".intrpt", "ax"
1489	.global intrpt_start
1490intrpt_start:
1491
1492#ifndef CONFIG_USE_PMC
1493#define handle_perf_interrupt bad_intr
1494#endif
1495
1496#ifndef CONFIG_HARDWALL
1497#define do_hardwall_trap bad_intr
1498#endif
1499
1500	int_hand     INT_MEM_ERROR, MEM_ERROR, do_trap
1501	int_hand     INT_SINGLE_STEP_3, SINGLE_STEP_3, bad_intr
1502#if CONFIG_KERNEL_PL == 2
1503	int_hand     INT_SINGLE_STEP_2, SINGLE_STEP_2, gx_singlestep_handle
1504	int_hand     INT_SINGLE_STEP_1, SINGLE_STEP_1, bad_intr
1505#else
1506	int_hand     INT_SINGLE_STEP_2, SINGLE_STEP_2, bad_intr
1507	int_hand     INT_SINGLE_STEP_1, SINGLE_STEP_1, gx_singlestep_handle
1508#endif
1509	int_hand     INT_SINGLE_STEP_0, SINGLE_STEP_0, bad_intr
1510	int_hand     INT_IDN_COMPLETE, IDN_COMPLETE, bad_intr
1511	int_hand     INT_UDN_COMPLETE, UDN_COMPLETE, bad_intr
1512	int_hand     INT_ITLB_MISS, ITLB_MISS, do_page_fault
1513	int_hand     INT_ILL, ILL, do_trap
1514	int_hand     INT_GPV, GPV, do_trap
1515	int_hand     INT_IDN_ACCESS, IDN_ACCESS, do_trap
1516	int_hand     INT_UDN_ACCESS, UDN_ACCESS, do_trap
1517	int_hand     INT_SWINT_3, SWINT_3, do_trap
1518	int_hand     INT_SWINT_2, SWINT_2, do_trap
1519	int_hand     INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall
1520	int_hand     INT_SWINT_0, SWINT_0, do_trap
1521	int_hand     INT_ILL_TRANS, ILL_TRANS, do_trap
1522	int_hand_unalign_fast INT_UNALIGN_DATA, UNALIGN_DATA
1523	int_hand     INT_DTLB_MISS, DTLB_MISS, do_page_fault
1524	int_hand     INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault
1525	int_hand     INT_IDN_FIREWALL, IDN_FIREWALL, do_hardwall_trap
1526	int_hand     INT_UDN_FIREWALL, UDN_FIREWALL, do_hardwall_trap
1527	int_hand     INT_TILE_TIMER, TILE_TIMER, do_timer_interrupt
1528	int_hand     INT_IDN_TIMER, IDN_TIMER, bad_intr
1529	int_hand     INT_UDN_TIMER, UDN_TIMER, bad_intr
1530	int_hand     INT_IDN_AVAIL, IDN_AVAIL, bad_intr
1531	int_hand     INT_UDN_AVAIL, UDN_AVAIL, bad_intr
1532	int_hand     INT_IPI_3, IPI_3, bad_intr
1533#if CONFIG_KERNEL_PL == 2
1534	int_hand     INT_IPI_2, IPI_2, tile_dev_intr
1535	int_hand     INT_IPI_1, IPI_1, bad_intr
1536#else
1537	int_hand     INT_IPI_2, IPI_2, bad_intr
1538	int_hand     INT_IPI_1, IPI_1, tile_dev_intr
1539#endif
1540	int_hand     INT_IPI_0, IPI_0, bad_intr
1541	int_hand     INT_PERF_COUNT, PERF_COUNT, \
1542		     handle_perf_interrupt, handle_nmi
1543	int_hand     INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \
1544		     handle_perf_interrupt, handle_nmi
1545	int_hand     INT_INTCTRL_3, INTCTRL_3, bad_intr
1546#if CONFIG_KERNEL_PL == 2
1547	dc_dispatch  INT_INTCTRL_2, INTCTRL_2
1548	int_hand     INT_INTCTRL_1, INTCTRL_1, bad_intr
1549#else
1550	int_hand     INT_INTCTRL_2, INTCTRL_2, bad_intr
1551	dc_dispatch  INT_INTCTRL_1, INTCTRL_1
1552#endif
1553	int_hand     INT_INTCTRL_0, INTCTRL_0, bad_intr
1554	int_hand     INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
1555		     hv_message_intr
1556	int_hand     INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, bad_intr
1557	int_hand     INT_I_ASID, I_ASID, bad_intr
1558	int_hand     INT_D_ASID, D_ASID, bad_intr
1559	int_hand     INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
1560
1561	/* Synthetic interrupt delivered only by the simulator */
1562	int_hand     INT_BREAKPOINT, BREAKPOINT, do_breakpoint
1563	/* Synthetic interrupt delivered by hv */
1564	int_hand     INT_NMI_DWNCL, NMI_DWNCL, do_nmi, handle_nmi
1565