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	.endif
519	.endif
520	.endif
521	.endif
522	.endif
523	/* Put function pointer in r0 */
524	moveli  r0, hw2_last(\c_routine)
525	shl16insli r0, r0, hw1(\c_routine)
526	{
527	 shl16insli r0, r0, hw0(\c_routine)
528	 j       \processing
529	}
530	.endif
531	ENDPROC(intvec_\vecname)
532
533#ifdef __COLLECT_LINKER_FEEDBACK__
534	.pushsection .text.intvec_feedback,"ax"
535	.org    (\vecnum << 5)
536	FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt, 1 << 8)
537	jrp     lr
538	.popsection
539#endif
540
541	.endm
542
543
544	/*
545	 * Save the rest of the registers that we didn't save in the actual
546	 * vector itself.  We can't use r0-r10 inclusive here.
547	 */
548	.macro  finish_interrupt_save, function
549
550	/* If it's a syscall, save a proper orig_r0, otherwise just zero. */
551	PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
552	{
553	 .ifc \function,handle_syscall
554	 st     r52, r0
555	 .else
556	 st     r52, zero
557	 .endif
558	 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
559	}
560	st      r52, tp
561	{
562	 mfspr  tp, CMPEXCH_VALUE
563	 PTREGS_PTR(r52, PTREGS_OFFSET_CMPEXCH)
564	}
565
566	/*
567	 * For ordinary syscalls, we save neither caller- nor callee-
568	 * save registers, since the syscall invoker doesn't expect the
569	 * caller-saves to be saved, and the called kernel functions will
570	 * take care of saving the callee-saves for us.
571	 *
572	 * For interrupts we save just the caller-save registers.  Saving
573	 * them is required (since the "caller" can't save them).  Again,
574	 * the called kernel functions will restore the callee-save
575	 * registers for us appropriately.
576	 *
577	 * On return, we normally restore nothing special for syscalls,
578	 * and just the caller-save registers for interrupts.
579	 *
580	 * However, there are some important caveats to all this:
581	 *
582	 * - We always save a few callee-save registers to give us
583	 *   some scratchpad registers to carry across function calls.
584	 *
585	 * - fork/vfork/etc require us to save all the callee-save
586	 *   registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
587	 *
588	 * - We always save r0..r5 and r10 for syscalls, since we need
589	 *   to reload them a bit later for the actual kernel call, and
590	 *   since we might need them for -ERESTARTNOINTR, etc.
591	 *
592	 * - Before invoking a signal handler, we save the unsaved
593	 *   callee-save registers so they are visible to the
594	 *   signal handler or any ptracer.
595	 *
596	 * - If the unsaved callee-save registers are modified, we set
597	 *   a bit in pt_regs so we know to reload them from pt_regs
598	 *   and not just rely on the kernel function unwinding.
599	 *   (Done for ptrace register writes and SA_SIGINFO handler.)
600	 */
601	{
602	 st     r52, tp
603	 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
604	}
605	wh64    r52    /* cache line 4 */
606	push_reg r33, r52
607	push_reg r32, r52
608	push_reg r31, r52
609	.ifc \function,handle_syscall
610	push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
611	push_reg TREG_SYSCALL_NR_NAME, r52, \
612	  PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
613	.else
614
615	push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
616	wh64    r52   /* cache line 3 */
617	push_reg r29, r52
618	push_reg r28, r52
619	push_reg r27, r52
620	push_reg r26, r52
621	push_reg r25, r52
622	push_reg r24, r52
623	push_reg r23, r52
624	push_reg r22, r52
625	wh64    r52   /* cache line 2 */
626	push_reg r21, r52
627	push_reg r20, r52
628	push_reg r19, r52
629	push_reg r18, r52
630	push_reg r17, r52
631	push_reg r16, r52
632	push_reg r15, r52
633	push_reg r14, r52
634	wh64    r52   /* cache line 1 */
635	push_reg r13, r52
636	push_reg r12, r52
637	push_reg r11, r52
638	push_reg r10, r52
639	push_reg r9, r52
640	push_reg r8, r52
641	push_reg r7, r52
642	push_reg r6, r52
643
644	.endif
645
646	push_reg r5, r52
647	st      r52, r4
648
649	/*
650	 * If we will be returning to the kernel, we will need to
651	 * reset the interrupt masks to the state they had before.
652	 * Set DISABLE_IRQ in flags iff we came from kernel pl with
653	 * irqs disabled.
654	 */
655	mfspr   r32, SPR_EX_CONTEXT_K_1
656	{
657	 IS_KERNEL_EX1(r22, r22)
658	 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
659	}
660	beqzt   r32, 1f       /* zero if from user space */
661	IRQS_DISABLED(r32)    /* zero if irqs enabled */
662#if PT_FLAGS_DISABLE_IRQ != 1
663# error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
664#endif
6651:
666	.ifnc \function,handle_syscall
667	/* Record the fact that we saved the caller-save registers above. */
668	ori     r32, r32, PT_FLAGS_CALLER_SAVES
669	.endif
670	st      r21, r32
671
672	/*
673	 * we've captured enough state to the stack (including in
674	 * particular our EX_CONTEXT state) that we can now release
675	 * the interrupt critical section and replace it with our
676	 * standard "interrupts disabled" mask value.  This allows
677	 * synchronous interrupts (and profile interrupts) to punch
678	 * through from this point onwards.
679	 *
680	 * It's important that no code before this point touch memory
681	 * other than our own stack (to keep the invariant that this
682	 * is all that gets touched under ICS), and that no code after
683	 * this point reference any interrupt-specific SPR, in particular
684	 * the EX_CONTEXT_K_ values.
685	 */
686	.ifc \function,handle_nmi
687	IRQ_DISABLE_ALL(r20)
688	.else
689	IRQ_DISABLE(r20, r21)
690	.endif
691	mtspr   INTERRUPT_CRITICAL_SECTION, zero
692
693	/* Load tp with our per-cpu offset. */
694#ifdef CONFIG_SMP
695	{
696	 mfspr  r20, SPR_SYSTEM_SAVE_K_0
697	 moveli r21, hw2_last(__per_cpu_offset)
698	}
699	{
700	 shl16insli r21, r21, hw1(__per_cpu_offset)
701	 bfextu r20, r20, CPU_SHIFT, 63
702	}
703	shl16insli r21, r21, hw0(__per_cpu_offset)
704	shl3add r20, r20, r21
705	ld      tp, r20
706#else
707	move    tp, zero
708#endif
709
710#ifdef __COLLECT_LINKER_FEEDBACK__
711	/*
712	 * Notify the feedback routines that we were in the
713	 * appropriate fixed interrupt vector area.  Note that we
714	 * still have ICS set at this point, so we can't invoke any
715	 * atomic operations or we will panic.  The feedback
716	 * routines internally preserve r0..r10 and r30 up.
717	 */
718	.ifnc \function,handle_syscall
719	shli    r20, r1, 5
720	.else
721	moveli  r20, INT_SWINT_1 << 5
722	.endif
723	moveli  r21, hw2_last(intvec_feedback)
724	shl16insli r21, r21, hw1(intvec_feedback)
725	shl16insli r21, r21, hw0(intvec_feedback)
726	add     r20, r20, r21
727	jalr    r20
728
729	/* And now notify the feedback routines that we are here. */
730	FEEDBACK_ENTER(\function)
731#endif
732
733	/*
734	 * Prepare the first 256 stack bytes to be rapidly accessible
735	 * without having to fetch the background data.
736	 */
737	addi    r52, sp, -64
738	{
739	 wh64   r52
740	 addi   r52, r52, -64
741	}
742	{
743	 wh64   r52
744	 addi   r52, r52, -64
745	}
746	{
747	 wh64   r52
748	 addi   r52, r52, -64
749	}
750	wh64    r52
751
752#ifdef CONFIG_TRACE_IRQFLAGS
753	.ifnc \function,handle_nmi
754	/*
755	 * We finally have enough state set up to notify the irq
756	 * tracing code that irqs were disabled on entry to the handler.
757	 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
758	 * For syscalls, we already have the register state saved away
759	 * on the stack, so we don't bother to do any register saves here,
760	 * and later we pop the registers back off the kernel stack.
761	 * For interrupt handlers, save r0-r3 in callee-saved registers.
762	 */
763	.ifnc \function,handle_syscall
764	{ move r30, r0; move r31, r1 }
765	{ move r32, r2; move r33, r3 }
766	.endif
767	TRACE_IRQS_OFF
768	.ifnc \function,handle_syscall
769	{ move r0, r30; move r1, r31 }
770	{ move r2, r32; move r3, r33 }
771	.endif
772	.endif
773#endif
774
775	.endm
776
777	/*
778	 * Redispatch a downcall.
779	 */
780	.macro  dc_dispatch vecnum, vecname
781	.org    (\vecnum << 8)
782intvec_\vecname:
783	j       _hv_downcall_dispatch
784	ENDPROC(intvec_\vecname)
785	.endm
786
787	/*
788	 * Common code for most interrupts.  The C function we're eventually
789	 * going to is in r0, and the faultnum is in r1; the original
790	 * values for those registers are on the stack.
791	 */
792	.pushsection .text.handle_interrupt,"ax"
793handle_interrupt:
794	finish_interrupt_save handle_interrupt
795
796	/* Jump to the C routine; it should enable irqs as soon as possible. */
797	{
798	 jalr   r0
799	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
800	}
801	FEEDBACK_REENTER(handle_interrupt)
802	{
803	 movei  r30, 0   /* not an NMI */
804	 j      interrupt_return
805	}
806	STD_ENDPROC(handle_interrupt)
807
808/*
809 * This routine takes a boolean in r30 indicating if this is an NMI.
810 * If so, we also expect a boolean in r31 indicating whether to
811 * re-enable the oprofile interrupts.
812 *
813 * Note that .Lresume_userspace is jumped to directly in several
814 * places, and we need to make sure r30 is set correctly in those
815 * callers as well.
816 */
817STD_ENTRY(interrupt_return)
818	/* If we're resuming to kernel space, don't check thread flags. */
819	{
820	 bnez   r30, .Lrestore_all  /* NMIs don't special-case user-space */
821	 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
822	}
823	ld      r29, r29
824	IS_KERNEL_EX1(r29, r29)
825	{
826	 beqzt  r29, .Lresume_userspace
827	 move   r29, sp
828	}
829
830#ifdef CONFIG_PREEMPT
831	/* Returning to kernel space. Check if we need preemption. */
832	EXTRACT_THREAD_INFO(r29)
833	addli   r28, r29, THREAD_INFO_FLAGS_OFFSET
834	{
835	 ld     r28, r28
836	 addli  r29, r29, THREAD_INFO_PREEMPT_COUNT_OFFSET
837	}
838	{
839	 andi   r28, r28, _TIF_NEED_RESCHED
840	 ld4s   r29, r29
841	}
842	beqzt   r28, 1f
843	bnez    r29, 1f
844	/* Disable interrupts explicitly for preemption. */
845	IRQ_DISABLE(r20,r21)
846	TRACE_IRQS_OFF
847	jal     preempt_schedule_irq
848	FEEDBACK_REENTER(interrupt_return)
8491:
850#endif
851
852	/* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
853	{
854	 moveli r27, hw2_last(_cpu_idle_nap)
855	 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
856	}
857	{
858	 ld     r28, r29
859	 shl16insli r27, r27, hw1(_cpu_idle_nap)
860	}
861	{
862	 shl16insli r27, r27, hw0(_cpu_idle_nap)
863	}
864	{
865	 cmpeq  r27, r27, r28
866	}
867	{
868	 blbc   r27, .Lrestore_all
869	 addi   r28, r28, 8
870	}
871	st      r29, r28
872	j       .Lrestore_all
873
874.Lresume_userspace:
875	FEEDBACK_REENTER(interrupt_return)
876
877	/*
878	 * Use r33 to hold whether we have already loaded the callee-saves
879	 * into ptregs.  We don't want to do it twice in this loop, since
880	 * then we'd clobber whatever changes are made by ptrace, etc.
881	 */
882	{
883	 movei  r33, 0
884	 move   r32, sp
885	}
886
887	/* Get base of stack in r32. */
888	EXTRACT_THREAD_INFO(r32)
889
890.Lretry_work_pending:
891	/*
892	 * Disable interrupts so as to make sure we don't
893	 * miss an interrupt that sets any of the thread flags (like
894	 * need_resched or sigpending) between sampling and the iret.
895	 * Routines like schedule() or do_signal() may re-enable
896	 * interrupts before returning.
897	 */
898	IRQ_DISABLE(r20, r21)
899	TRACE_IRQS_OFF  /* Note: clobbers registers r0-r29 */
900
901
902	/* Check to see if there is any work to do before returning to user. */
903	{
904	 addi   r29, r32, THREAD_INFO_FLAGS_OFFSET
905	 moveli r1, hw1_last(_TIF_ALLWORK_MASK)
906	}
907	{
908	 ld     r29, r29
909	 shl16insli r1, r1, hw0(_TIF_ALLWORK_MASK)
910	}
911	and     r1, r29, r1
912	beqzt   r1, .Lrestore_all
913
914	/*
915	 * Make sure we have all the registers saved for signal
916	 * handling or notify-resume.  Call out to C code to figure out
917	 * exactly what we need to do for each flag bit, then if
918	 * necessary, reload the flags and recheck.
919	 */
920	{
921	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
922	 bnez   r33, 1f
923	}
924	push_extra_callee_saves r0
925	movei   r33, 1
9261:	jal     do_work_pending
927	bnez    r0, .Lretry_work_pending
928
929	/*
930	 * In the NMI case we
931	 * omit the call to single_process_check_nohz, which normally checks
932	 * to see if we should start or stop the scheduler tick, because
933	 * we can't call arbitrary Linux code from an NMI context.
934	 * We always call the homecache TLB deferral code to re-trigger
935	 * the deferral mechanism.
936	 *
937	 * The other chunk of responsibility this code has is to reset the
938	 * interrupt masks appropriately to reset irqs and NMIs.  We have
939	 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
940	 * lockdep-type stuff, but we can't set ICS until afterwards, since
941	 * ICS can only be used in very tight chunks of code to avoid
942	 * tripping over various assertions that it is off.
943	 */
944.Lrestore_all:
945	PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
946	{
947	 ld      r0, r0
948	 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
949	}
950	{
951	 IS_KERNEL_EX1(r0, r0)
952	 ld     r32, r32
953	}
954	bnez    r0, 1f
955	j       2f
956#if PT_FLAGS_DISABLE_IRQ != 1
957# error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use blbct below
958#endif
9591:	blbct   r32, 2f
960	IRQ_DISABLE(r20,r21)
961	TRACE_IRQS_OFF
962	movei   r0, 1
963	mtspr   INTERRUPT_CRITICAL_SECTION, r0
964	beqzt   r30, .Lrestore_regs
965	j       3f
9662:	TRACE_IRQS_ON
967	IRQ_ENABLE_LOAD(r20, r21)
968	movei   r0, 1
969	mtspr   INTERRUPT_CRITICAL_SECTION, r0
970	IRQ_ENABLE_APPLY(r20, r21)
971	beqzt   r30, .Lrestore_regs
9723:
973
974#if INT_PERF_COUNT + 1 != INT_AUX_PERF_COUNT
975# error Bad interrupt assumption
976#endif
977	{
978	 movei  r0, 3   /* two adjacent bits for the PERF_COUNT mask */
979	 beqz   r31, .Lrestore_regs
980	}
981	shli    r0, r0, INT_PERF_COUNT
982	mtspr   SPR_INTERRUPT_MASK_RESET_K, r0
983
984	/*
985	 * We now commit to returning from this interrupt, since we will be
986	 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
987	 * frame.  No calls should be made to any other code after this point.
988	 * This code should only be entered with ICS set.
989	 * r32 must still be set to ptregs.flags.
990	 * We launch loads to each cache line separately first, so we can
991	 * get some parallelism out of the memory subsystem.
992	 * We start zeroing caller-saved registers throughout, since
993	 * that will save some cycles if this turns out to be a syscall.
994	 */
995.Lrestore_regs:
996
997	/*
998	 * Rotate so we have one high bit and one low bit to test.
999	 * - low bit says whether to restore all the callee-saved registers,
1000	 *   or just r30-r33, and r52 up.
1001	 * - high bit (i.e. sign bit) says whether to restore all the
1002	 *   caller-saved registers, or just r0.
1003	 */
1004#if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
1005# error Rotate trick does not work :-)
1006#endif
1007	{
1008	 rotli  r20, r32, 62
1009	 PTREGS_PTR(sp, PTREGS_OFFSET_REG(0))
1010	}
1011
1012	/*
1013	 * Load cache lines 0, 4, 6 and 7, in that order, then use
1014	 * the last loaded value, which makes it likely that the other
1015	 * cache lines have also loaded, at which point we should be
1016	 * able to safely read all the remaining words on those cache
1017	 * lines without waiting for the memory subsystem.
1018	 */
1019	pop_reg r0, sp, PTREGS_OFFSET_REG(30) - PTREGS_OFFSET_REG(0)
1020	pop_reg r30, sp, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_REG(30)
1021	pop_reg_zero r52, r3, sp, PTREGS_OFFSET_CMPEXCH - PTREGS_OFFSET_REG(52)
1022	pop_reg_zero r21, r27, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_CMPEXCH
1023	pop_reg_zero lr, r2, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_EX1
1024	{
1025	 mtspr  CMPEXCH_VALUE, r21
1026	 move   r4, zero
1027	}
1028	pop_reg r21, sp, PTREGS_OFFSET_REG(31) - PTREGS_OFFSET_PC
1029	{
1030	 mtspr  SPR_EX_CONTEXT_K_1, lr
1031	 IS_KERNEL_EX1(lr, lr)
1032	}
1033	{
1034	 mtspr  SPR_EX_CONTEXT_K_0, r21
1035	 move   r5, zero
1036	}
1037
1038	/* Restore callee-saveds that we actually use. */
1039	pop_reg_zero r31, r6
1040	pop_reg_zero r32, r7
1041	pop_reg_zero r33, r8, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(33)
1042
1043	/*
1044	 * If we modified other callee-saveds, restore them now.
1045	 * This is rare, but could be via ptrace or signal handler.
1046	 */
1047	{
1048	 move   r9, zero
1049	 blbs   r20, .Lrestore_callees
1050	}
1051.Lcontinue_restore_regs:
1052
1053	/* Check if we're returning from a syscall. */
1054	{
1055	 move   r10, zero
1056	 bltzt  r20, 1f  /* no, so go restore callee-save registers */
1057	}
1058
1059	/*
1060	 * Check if we're returning to userspace.
1061	 * Note that if we're not, we don't worry about zeroing everything.
1062	 */
1063	{
1064	 addli  sp, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(29)
1065	 bnez   lr, .Lkernel_return
1066	}
1067
1068	/*
1069	 * On return from syscall, we've restored r0 from pt_regs, but we
1070	 * clear the remainder of the caller-saved registers.  We could
1071	 * restore the syscall arguments, but there's not much point,
1072	 * and it ensures user programs aren't trying to use the
1073	 * caller-saves if we clear them, as well as avoiding leaking
1074	 * kernel pointers into userspace.
1075	 */
1076	pop_reg_zero lr, r11, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1077	pop_reg_zero tp, r12, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1078	{
1079	 ld     sp, sp
1080	 move   r13, zero
1081	 move   r14, zero
1082	}
1083	{ move r15, zero; move r16, zero }
1084	{ move r17, zero; move r18, zero }
1085	{ move r19, zero; move r20, zero }
1086	{ move r21, zero; move r22, zero }
1087	{ move r23, zero; move r24, zero }
1088	{ move r25, zero; move r26, zero }
1089
1090	/* Set r1 to errno if we are returning an error, otherwise zero. */
1091	{
1092	 moveli r29, 4096
1093	 sub    r1, zero, r0
1094	}
1095	{
1096	 move   r28, zero
1097	 cmpltu r29, r1, r29
1098	}
1099	{
1100	 mnz    r1, r29, r1
1101	 move   r29, zero
1102	}
1103	iret
1104
1105	/*
1106	 * Not a syscall, so restore caller-saved registers.
1107	 * First kick off loads for cache lines 1-3, which we're touching
1108	 * for the first time here.
1109	 */
1110	.align 64
11111:	pop_reg r29, sp, PTREGS_OFFSET_REG(21) - PTREGS_OFFSET_REG(29)
1112	pop_reg r21, sp, PTREGS_OFFSET_REG(13) - PTREGS_OFFSET_REG(21)
1113	pop_reg r13, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(13)
1114	pop_reg r1
1115	pop_reg r2
1116	pop_reg r3
1117	pop_reg r4
1118	pop_reg r5
1119	pop_reg r6
1120	pop_reg r7
1121	pop_reg r8
1122	pop_reg r9
1123	pop_reg r10
1124	pop_reg r11
1125	pop_reg r12, sp, 16
1126	/* r13 already restored above */
1127	pop_reg r14
1128	pop_reg r15
1129	pop_reg r16
1130	pop_reg r17
1131	pop_reg r18
1132	pop_reg r19
1133	pop_reg r20, sp, 16
1134	/* r21 already restored above */
1135	pop_reg r22
1136	pop_reg r23
1137	pop_reg r24
1138	pop_reg r25
1139	pop_reg r26
1140	pop_reg r27
1141	pop_reg r28, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(28)
1142	/* r29 already restored above */
1143	bnez    lr, .Lkernel_return
1144	pop_reg lr, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1145	pop_reg tp, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1146	ld      sp, sp
1147	iret
1148
1149	/*
1150	 * We can't restore tp when in kernel mode, since a thread might
1151	 * have migrated from another cpu and brought a stale tp value.
1152	 */
1153.Lkernel_return:
1154	pop_reg lr, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
1155	ld      sp, sp
1156	iret
1157
1158	/* Restore callee-saved registers from r34 to r51. */
1159.Lrestore_callees:
1160	addli  sp, sp, PTREGS_OFFSET_REG(34) - PTREGS_OFFSET_REG(29)
1161	pop_reg r34
1162	pop_reg r35
1163	pop_reg r36
1164	pop_reg r37
1165	pop_reg r38
1166	pop_reg r39
1167	pop_reg r40
1168	pop_reg r41
1169	pop_reg r42
1170	pop_reg r43
1171	pop_reg r44
1172	pop_reg r45
1173	pop_reg r46
1174	pop_reg r47
1175	pop_reg r48
1176	pop_reg r49
1177	pop_reg r50
1178	pop_reg r51, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(51)
1179	j .Lcontinue_restore_regs
1180	STD_ENDPROC(interrupt_return)
1181
1182	/*
1183	 * "NMI" interrupts mask ALL interrupts before calling the
1184	 * handler, and don't check thread flags, etc., on the way
1185	 * back out.  In general, the only things we do here for NMIs
1186	 * are register save/restore and dataplane kernel-TLB management.
1187	 * We don't (for example) deal with start/stop of the sched tick.
1188	 */
1189	.pushsection .text.handle_nmi,"ax"
1190handle_nmi:
1191	finish_interrupt_save handle_nmi
1192	{
1193	 jalr   r0
1194	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1195	}
1196	FEEDBACK_REENTER(handle_nmi)
1197	{
1198	 movei  r30, 1
1199	 cmpeq  r31, r0, zero
1200	}
1201	j       interrupt_return
1202	STD_ENDPROC(handle_nmi)
1203
1204	/*
1205	 * Parallel code for syscalls to handle_interrupt.
1206	 */
1207	.pushsection .text.handle_syscall,"ax"
1208handle_syscall:
1209	finish_interrupt_save handle_syscall
1210
1211	/* Enable irqs. */
1212	TRACE_IRQS_ON
1213	IRQ_ENABLE(r20, r21)
1214
1215	/* Bump the counter for syscalls made on this tile. */
1216	moveli r20, hw2_last(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1217	shl16insli r20, r20, hw1(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1218	shl16insli r20, r20, hw0(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1219	add     r20, r20, tp
1220	ld4s    r21, r20
1221	{
1222	 addi   r21, r21, 1
1223	 move   r31, sp
1224	}
1225	{
1226	 st4    r20, r21
1227	 EXTRACT_THREAD_INFO(r31)
1228	}
1229
1230	/* Trace syscalls, if requested. */
1231	addi	r31, r31, THREAD_INFO_FLAGS_OFFSET
1232	{
1233	 ld     r30, r31
1234	 moveli r32, _TIF_SYSCALL_ENTRY_WORK
1235	}
1236	and     r30, r30, r32
1237	{
1238	 addi   r30, r31, THREAD_INFO_STATUS_OFFSET - THREAD_INFO_FLAGS_OFFSET
1239	 beqzt	r30, .Lrestore_syscall_regs
1240	}
1241	{
1242	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1243	 jal    do_syscall_trace_enter
1244	}
1245	FEEDBACK_REENTER(handle_syscall)
1246
1247	/*
1248	 * We always reload our registers from the stack at this
1249	 * point.  They might be valid, if we didn't build with
1250	 * TRACE_IRQFLAGS, and this isn't a dataplane tile, and we're not
1251	 * doing syscall tracing, but there are enough cases now that it
1252	 * seems simplest just to do the reload unconditionally.
1253	 */
1254.Lrestore_syscall_regs:
1255	{
1256	 ld     r30, r30
1257	 PTREGS_PTR(r11, PTREGS_OFFSET_REG(0))
1258	}
1259	pop_reg r0,  r11
1260	pop_reg r1,  r11
1261	pop_reg r2,  r11
1262	pop_reg r3,  r11
1263	pop_reg r4,  r11
1264	pop_reg r5,  r11, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(5)
1265	{
1266	 ld     TREG_SYSCALL_NR_NAME, r11
1267	 moveli r21, __NR_syscalls
1268	}
1269
1270	/* Ensure that the syscall number is within the legal range. */
1271	{
1272	 moveli r20, hw2(sys_call_table)
1273#ifdef CONFIG_COMPAT
1274	 blbs   r30, .Lcompat_syscall
1275#endif
1276	}
1277	{
1278	 cmpltu r21, TREG_SYSCALL_NR_NAME, r21
1279	 shl16insli r20, r20, hw1(sys_call_table)
1280	}
1281	{
1282	 blbc   r21, .Linvalid_syscall
1283	 shl16insli r20, r20, hw0(sys_call_table)
1284	}
1285.Lload_syscall_pointer:
1286	shl3add r20, TREG_SYSCALL_NR_NAME, r20
1287	ld      r20, r20
1288
1289	/* Jump to syscall handler. */
1290	jalr    r20
1291.Lhandle_syscall_link: /* value of "lr" after "jalr r20" above */
1292
1293	/*
1294	 * Write our r0 onto the stack so it gets restored instead
1295	 * of whatever the user had there before.
1296	 * In compat mode, sign-extend r0 before storing it.
1297	 */
1298	{
1299	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1300	 blbct  r30, 1f
1301	}
1302	addxi   r0, r0, 0
13031:	st      r29, r0
1304
1305.Lsyscall_sigreturn_skip:
1306	FEEDBACK_REENTER(handle_syscall)
1307
1308	/* Do syscall trace again, if requested. */
1309	{
1310	 ld      r30, r31
1311	 moveli  r32, _TIF_SYSCALL_EXIT_WORK
1312	}
1313	and      r0, r30, r32
1314	{
1315	 andi    r0, r30, _TIF_SINGLESTEP
1316	 beqzt   r0, 1f
1317	}
1318	{
1319	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1320	 jal    do_syscall_trace_exit
1321	}
1322	FEEDBACK_REENTER(handle_syscall)
1323	andi    r0, r30, _TIF_SINGLESTEP
1324
13251:	beqzt	r0, 2f
1326
1327	/* Single stepping -- notify ptrace. */
1328	{
1329	 movei   r0, SIGTRAP
1330	 jal     ptrace_notify
1331	}
1332	FEEDBACK_REENTER(handle_syscall)
1333
13342:	{
1335	 movei  r30, 0               /* not an NMI */
1336	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1337	}
1338
1339#ifdef CONFIG_COMPAT
1340.Lcompat_syscall:
1341	/*
1342	 * Load the base of the compat syscall table in r20, and
1343	 * range-check the syscall number (duplicated from 64-bit path).
1344	 * Sign-extend all the user's passed arguments to make them consistent.
1345	 * Also save the original "r(n)" values away in "r(11+n)" in
1346	 * case the syscall table entry wants to validate them.
1347	 */
1348	moveli  r20, hw2(compat_sys_call_table)
1349	{
1350	 cmpltu r21, TREG_SYSCALL_NR_NAME, r21
1351	 shl16insli r20, r20, hw1(compat_sys_call_table)
1352	}
1353	{
1354	 blbc   r21, .Linvalid_syscall
1355	 shl16insli r20, r20, hw0(compat_sys_call_table)
1356	}
1357	{ move r11, r0; addxi r0, r0, 0 }
1358	{ move r12, r1; addxi r1, r1, 0 }
1359	{ move r13, r2; addxi r2, r2, 0 }
1360	{ move r14, r3; addxi r3, r3, 0 }
1361	{ move r15, r4; addxi r4, r4, 0 }
1362	{ move r16, r5; addxi r5, r5, 0 }
1363	j .Lload_syscall_pointer
1364#endif
1365
1366.Linvalid_syscall:
1367	/* Report an invalid syscall back to the user program */
1368	{
1369	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1370	 movei  r28, -ENOSYS
1371	}
1372	st      r29, r28
1373	{
1374	 movei  r30, 0               /* not an NMI */
1375	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1376	}
1377	STD_ENDPROC(handle_syscall)
1378
1379	/* Return the address for oprofile to suppress in backtraces. */
1380STD_ENTRY_SECTION(handle_syscall_link_address, .text.handle_syscall)
1381	lnk     r0
1382	{
1383	 addli  r0, r0, .Lhandle_syscall_link - .
1384	 jrp    lr
1385	}
1386	STD_ENDPROC(handle_syscall_link_address)
1387
1388STD_ENTRY(ret_from_fork)
1389	jal     sim_notify_fork
1390	jal     schedule_tail
1391	FEEDBACK_REENTER(ret_from_fork)
1392	{
1393	 movei  r30, 0               /* not an NMI */
1394	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1395	}
1396	STD_ENDPROC(ret_from_fork)
1397
1398STD_ENTRY(ret_from_kernel_thread)
1399	jal     sim_notify_fork
1400	jal     schedule_tail
1401	FEEDBACK_REENTER(ret_from_fork)
1402	{
1403	 move   r0, r31
1404	 jalr   r30
1405	}
1406	FEEDBACK_REENTER(ret_from_kernel_thread)
1407	{
1408	 movei  r30, 0               /* not an NMI */
1409	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1410	}
1411	STD_ENDPROC(ret_from_kernel_thread)
1412
1413/* Various stub interrupt handlers and syscall handlers */
1414
1415STD_ENTRY_LOCAL(_kernel_double_fault)
1416	mfspr   r1, SPR_EX_CONTEXT_K_0
1417	move    r2, lr
1418	move    r3, sp
1419	move    r4, r52
1420	addi    sp, sp, -C_ABI_SAVE_AREA_SIZE
1421	j       kernel_double_fault
1422	STD_ENDPROC(_kernel_double_fault)
1423
1424STD_ENTRY_LOCAL(bad_intr)
1425	mfspr   r2, SPR_EX_CONTEXT_K_0
1426	panic   "Unhandled interrupt %#x: PC %#lx"
1427	STD_ENDPROC(bad_intr)
1428
1429/*
1430 * Special-case sigreturn to not write r0 to the stack on return.
1431 * This is technically more efficient, but it also avoids difficulties
1432 * in the 64-bit OS when handling 32-bit compat code, since we must not
1433 * sign-extend r0 for the sigreturn return-value case.
1434 */
1435#define PTREGS_SYSCALL_SIGRETURN(x, reg)                \
1436	STD_ENTRY(_##x);                                \
1437	addli   lr, lr, .Lsyscall_sigreturn_skip - .Lhandle_syscall_link; \
1438	{                                               \
1439	 PTREGS_PTR(reg, PTREGS_OFFSET_BASE);           \
1440	 j      x                                       \
1441	};                                              \
1442	STD_ENDPROC(_##x)
1443
1444PTREGS_SYSCALL_SIGRETURN(sys_rt_sigreturn, r0)
1445#ifdef CONFIG_COMPAT
1446PTREGS_SYSCALL_SIGRETURN(compat_sys_rt_sigreturn, r0)
1447#endif
1448
1449/* Save additional callee-saves to pt_regs and jump to standard function. */
1450STD_ENTRY(_sys_clone)
1451	push_extra_callee_saves r4
1452	j       sys_clone
1453	STD_ENDPROC(_sys_clone)
1454
1455	/*
1456	 * Recover r3, r2, r1 and r0 here saved by unalign fast vector.
1457	 * The vector area limit is 32 bundles, so we handle the reload here.
1458	 * r0, r1, r2 are in thread_info from low to high memory in order.
1459	 * r3 points to location the original r3 was saved.
1460	 * We put this code in the __HEAD section so it can be reached
1461	 * via a conditional branch from the fast path.
1462	 */
1463	__HEAD
1464hand_unalign_slow:
1465	andi    sp, sp, ~1
1466hand_unalign_slow_badsp:
1467	addi    r3, r3, -(3 * 8)
1468	ld_add  r0, r3, 8
1469	ld_add  r1, r3, 8
1470	ld      r2, r3
1471hand_unalign_slow_nonuser:
1472	mfspr   r3, SPR_SYSTEM_SAVE_K_1
1473	__int_hand     INT_UNALIGN_DATA, UNALIGN_DATA_SLOW, int_unalign
1474
1475/* The unaligned data support needs to read all the registers. */
1476int_unalign:
1477	push_extra_callee_saves r0
1478	j       do_unaligned
1479ENDPROC(hand_unalign_slow)
1480
1481/* Fill the return address stack with nonzero entries. */
1482STD_ENTRY(fill_ra_stack)
1483	{
1484	 move	r0, lr
1485	 jal	1f
1486	}
14871:	jal	2f
14882:	jal	3f
14893:	jal	4f
14904:	jrp	r0
1491	STD_ENDPROC(fill_ra_stack)
1492
1493	.macro int_hand  vecnum, vecname, c_routine, processing=handle_interrupt
1494	.org   (\vecnum << 8)
1495		__int_hand   \vecnum, \vecname, \c_routine, \processing
1496	.endm
1497
1498/* Include .intrpt array of interrupt vectors */
1499	.section ".intrpt", "ax"
1500	.global intrpt_start
1501intrpt_start:
1502
1503#ifndef CONFIG_USE_PMC
1504#define handle_perf_interrupt bad_intr
1505#endif
1506
1507#ifndef CONFIG_HARDWALL
1508#define do_hardwall_trap bad_intr
1509#endif
1510
1511	int_hand     INT_MEM_ERROR, MEM_ERROR, do_trap
1512	int_hand     INT_SINGLE_STEP_3, SINGLE_STEP_3, bad_intr
1513#if CONFIG_KERNEL_PL == 2
1514	int_hand     INT_SINGLE_STEP_2, SINGLE_STEP_2, gx_singlestep_handle
1515	int_hand     INT_SINGLE_STEP_1, SINGLE_STEP_1, bad_intr
1516#else
1517	int_hand     INT_SINGLE_STEP_2, SINGLE_STEP_2, bad_intr
1518	int_hand     INT_SINGLE_STEP_1, SINGLE_STEP_1, gx_singlestep_handle
1519#endif
1520	int_hand     INT_SINGLE_STEP_0, SINGLE_STEP_0, bad_intr
1521	int_hand     INT_IDN_COMPLETE, IDN_COMPLETE, bad_intr
1522	int_hand     INT_UDN_COMPLETE, UDN_COMPLETE, bad_intr
1523	int_hand     INT_ITLB_MISS, ITLB_MISS, do_page_fault
1524	int_hand     INT_ILL, ILL, do_trap
1525	int_hand     INT_GPV, GPV, do_trap
1526	int_hand     INT_IDN_ACCESS, IDN_ACCESS, do_trap
1527	int_hand     INT_UDN_ACCESS, UDN_ACCESS, do_trap
1528	int_hand     INT_SWINT_3, SWINT_3, do_trap
1529	int_hand     INT_SWINT_2, SWINT_2, do_trap
1530	int_hand     INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall
1531	int_hand     INT_SWINT_0, SWINT_0, do_trap
1532	int_hand     INT_ILL_TRANS, ILL_TRANS, do_trap
1533	int_hand_unalign_fast INT_UNALIGN_DATA, UNALIGN_DATA
1534	int_hand     INT_DTLB_MISS, DTLB_MISS, do_page_fault
1535	int_hand     INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault
1536	int_hand     INT_IDN_FIREWALL, IDN_FIREWALL, do_hardwall_trap
1537	int_hand     INT_UDN_FIREWALL, UDN_FIREWALL, do_hardwall_trap
1538	int_hand     INT_TILE_TIMER, TILE_TIMER, do_timer_interrupt
1539	int_hand     INT_IDN_TIMER, IDN_TIMER, bad_intr
1540	int_hand     INT_UDN_TIMER, UDN_TIMER, bad_intr
1541	int_hand     INT_IDN_AVAIL, IDN_AVAIL, bad_intr
1542	int_hand     INT_UDN_AVAIL, UDN_AVAIL, bad_intr
1543	int_hand     INT_IPI_3, IPI_3, bad_intr
1544#if CONFIG_KERNEL_PL == 2
1545	int_hand     INT_IPI_2, IPI_2, tile_dev_intr
1546	int_hand     INT_IPI_1, IPI_1, bad_intr
1547#else
1548	int_hand     INT_IPI_2, IPI_2, bad_intr
1549	int_hand     INT_IPI_1, IPI_1, tile_dev_intr
1550#endif
1551	int_hand     INT_IPI_0, IPI_0, bad_intr
1552	int_hand     INT_PERF_COUNT, PERF_COUNT, \
1553		     handle_perf_interrupt, handle_nmi
1554	int_hand     INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \
1555		     handle_perf_interrupt, handle_nmi
1556	int_hand     INT_INTCTRL_3, INTCTRL_3, bad_intr
1557#if CONFIG_KERNEL_PL == 2
1558	dc_dispatch  INT_INTCTRL_2, INTCTRL_2
1559	int_hand     INT_INTCTRL_1, INTCTRL_1, bad_intr
1560#else
1561	int_hand     INT_INTCTRL_2, INTCTRL_2, bad_intr
1562	dc_dispatch  INT_INTCTRL_1, INTCTRL_1
1563#endif
1564	int_hand     INT_INTCTRL_0, INTCTRL_0, bad_intr
1565	int_hand     INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
1566		     hv_message_intr
1567	int_hand     INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, bad_intr
1568	int_hand     INT_I_ASID, I_ASID, bad_intr
1569	int_hand     INT_D_ASID, D_ASID, bad_intr
1570	int_hand     INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
1571
1572	/* Synthetic interrupt delivered only by the simulator */
1573	int_hand     INT_BREAKPOINT, BREAKPOINT, do_breakpoint
1574