| blob | 79c93e10ba27b1a11c939e216233ec5956b6b11a |
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 */
17 #include <linux/linkage.h>
18 #include <linux/errno.h>
19 #include <linux/unistd.h>
20 #include <asm/ptrace.h>
21 #include <asm/thread_info.h>
22 #include <asm/irqflags.h>
23 #include <asm/asm-offsets.h>
24 #include <asm/types.h>
25 #include <hv/hypervisor.h>
26 #include <arch/abi.h>
27 #include <arch/interrupts.h>
28 #include <arch/spr_def.h>
30 #ifdef CONFIG_PREEMPT
31 # error "No support for kernel preemption currently"
32 #endif
34 #define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
36 #define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
39 .macro push_reg reg, ptr=sp, delta=-8
40 {
41 st \ptr, \reg
42 addli \ptr, \ptr, \delta
43 }
44 .endm
46 .macro pop_reg reg, ptr=sp, delta=8
47 {
48 ld \reg, \ptr
49 addli \ptr, \ptr, \delta
50 }
51 .endm
53 .macro pop_reg_zero reg, zreg, ptr=sp, delta=8
54 {
55 move \zreg, zero
56 ld \reg, \ptr
57 addi \ptr, \ptr, \delta
58 }
59 .endm
61 .macro push_extra_callee_saves reg
62 PTREGS_PTR(\reg, PTREGS_OFFSET_REG(51))
63 push_reg r51, \reg
64 push_reg r50, \reg
65 push_reg r49, \reg
66 push_reg r48, \reg
67 push_reg r47, \reg
68 push_reg r46, \reg
69 push_reg r45, \reg
70 push_reg r44, \reg
71 push_reg r43, \reg
72 push_reg r42, \reg
73 push_reg r41, \reg
74 push_reg r40, \reg
75 push_reg r39, \reg
76 push_reg r38, \reg
77 push_reg r37, \reg
78 push_reg r36, \reg
79 push_reg r35, \reg
80 push_reg r34, \reg, PTREGS_OFFSET_BASE - PTREGS_OFFSET_REG(34)
81 .endm
83 .macro panic str
84 .pushsection .rodata, "a"
85 1:
86 .asciz "\str"
87 .popsection
88 {
89 moveli r0, hw2_last(1b)
90 }
91 {
92 shl16insli r0, r0, hw1(1b)
93 }
94 {
95 shl16insli r0, r0, hw0(1b)
96 jal panic
97 }
98 .endm
101 #ifdef __COLLECT_LINKER_FEEDBACK__
102 .pushsection .text.intvec_feedback,"ax"
103 intvec_feedback:
104 .popsection
105 #endif
107 /*
108 * Default interrupt handler.
109 *
110 * vecnum is where we'll put this code.
111 * c_routine is the C routine we'll call.
112 *
113 * The C routine is passed two arguments:
114 * - A pointer to the pt_regs state.
115 * - The interrupt vector number.
116 *
117 * The "processing" argument specifies the code for processing
118 * the interrupt. Defaults to "handle_interrupt".
119 */
120 .macro int_hand vecnum, vecname, c_routine, processing=handle_interrupt
121 .org (\vecnum << 8)
122 intvec_\vecname:
123 /* Temporarily save a register so we have somewhere to work. */
125 mtspr SPR_SYSTEM_SAVE_K_1, r0
126 mfspr r0, SPR_EX_CONTEXT_K_1
128 andi r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
130 .ifc \vecnum, INT_DOUBLE_FAULT
131 /*
132 * For double-faults from user-space, fall through to the normal
133 * register save and stack setup path. Otherwise, it's the
134 * hypervisor giving us one last chance to dump diagnostics, and we
135 * branch to the kernel_double_fault routine to do so.
136 */
137 beqz r0, 1f
138 j _kernel_double_fault
139 1:
140 .else
141 /*
142 * If we're coming from user-space, then set sp to the top of
143 * the kernel stack. Otherwise, assume sp is already valid.
144 */
145 {
146 bnez r0, 0f
147 move r0, sp
148 }
149 .endif
151 .ifc \c_routine, do_page_fault
152 /*
153 * The page_fault handler may be downcalled directly by the
154 * hypervisor even when Linux is running and has ICS set.
155 *
156 * In this case the contents of EX_CONTEXT_K_1 reflect the
157 * previous fault and can't be relied on to choose whether or
158 * not to reinitialize the stack pointer. So we add a test
159 * to see whether SYSTEM_SAVE_K_2 has the high bit set,
160 * and if so we don't reinitialize sp, since we must be coming
161 * from Linux. (In fact the precise case is !(val & ~1),
162 * but any Linux PC has to have the high bit set.)
163 *
164 * Note that the hypervisor *always* sets SYSTEM_SAVE_K_2 for
165 * any path that turns into a downcall to one of our TLB handlers.
166 *
167 * FIXME: if we end up never using this path, perhaps we should
168 * prevent the hypervisor from generating downcalls in this case.
169 * The advantage of getting a downcall is we can panic in Linux.
170 */
171 mfspr r0, SPR_SYSTEM_SAVE_K_2
172 {
173 bltz r0, 0f /* high bit in S_S_1_2 is for a PC to use */
174 move r0, sp
175 }
176 .endif
179 /*
180 * SYSTEM_SAVE_K_0 holds the cpu number in the low bits, and
181 * the current stack top in the higher bits. So we recover
182 * our stack top by just masking off the low bits, then
183 * point sp at the top aligned address on the actual stack page.
184 */
185 mfspr r0, SPR_SYSTEM_SAVE_K_0
186 mm r0, zero, LOG2_THREAD_SIZE, 63
188 0:
189 /*
190 * Align the stack mod 64 so we can properly predict what
191 * cache lines we need to write-hint to reduce memory fetch
192 * latency as we enter the kernel. The layout of memory is
193 * as follows, with cache line 0 at the lowest VA, and cache
194 * line 8 just below the r0 value this "andi" computes.
195 * Note that we never write to cache line 8, and we skip
196 * cache lines 1-3 for syscalls.
197 *
198 * cache line 8: ptregs padding (two words)
199 * cache line 7: sp, lr, pc, ex1, faultnum, orig_r0, flags, cmpexch
200 * cache line 6: r46...r53 (tp)
201 * cache line 5: r38...r45
202 * cache line 4: r30...r37
203 * cache line 3: r22...r29
204 * cache line 2: r14...r21
205 * cache line 1: r6...r13
206 * cache line 0: 2 x frame, r0..r5
207 */
208 andi r0, r0, -64
210 /*
211 * Push the first four registers on the stack, so that we can set
212 * them to vector-unique values before we jump to the common code.
213 *
214 * Registers are pushed on the stack as a struct pt_regs,
215 * with the sp initially just above the struct, and when we're
216 * done, sp points to the base of the struct, minus
217 * C_ABI_SAVE_AREA_SIZE, so we can directly jal to C code.
218 *
219 * This routine saves just the first four registers, plus the
220 * stack context so we can do proper backtracing right away,
221 * and defers to handle_interrupt to save the rest.
222 * The backtracer needs pc, ex1, lr, sp, r52, and faultnum.
223 */
224 addli r0, r0, PTREGS_OFFSET_LR - (PTREGS_SIZE + KSTK_PTREGS_GAP)
225 wh64 r0 /* cache line 7 */
226 {
227 st r0, lr
228 addli r0, r0, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
229 }
230 {
231 st r0, sp
232 addli sp, r0, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_SP
233 }
234 wh64 sp /* cache line 6 */
235 {
236 st sp, r52
237 addli sp, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(52)
238 }
239 wh64 sp /* cache line 0 */
240 {
241 st sp, r1
242 addli sp, sp, PTREGS_OFFSET_REG(2) - PTREGS_OFFSET_REG(1)
243 }
244 {
245 st sp, r2
246 addli sp, sp, PTREGS_OFFSET_REG(3) - PTREGS_OFFSET_REG(2)
247 }
248 {
249 st sp, r3
250 addli sp, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(3)
251 }
252 mfspr r0, SPR_EX_CONTEXT_K_0
253 .ifc \processing,handle_syscall
254 /*
255 * Bump the saved PC by one bundle so that when we return, we won't
256 * execute the same swint instruction again. We need to do this while
257 * we're in the critical section.
258 */
259 addi r0, r0, 8
260 .endif
261 {
262 st sp, r0
263 addli sp, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
264 }
265 mfspr r0, SPR_EX_CONTEXT_K_1
266 {
267 st sp, r0
268 addi sp, sp, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
269 /*
270 * Use r0 for syscalls so it's a temporary; use r1 for interrupts
271 * so that it gets passed through unchanged to the handler routine.
272 * Note that the .if conditional confusingly spans bundles.
273 */
274 .ifc \processing,handle_syscall
275 movei r0, \vecnum
276 }
277 {
278 st sp, r0
279 .else
280 movei r1, \vecnum
281 }
282 {
283 st sp, r1
284 .endif
285 addli sp, sp, PTREGS_OFFSET_REG(0) - PTREGS_OFFSET_FAULTNUM
286 }
287 mfspr r0, SPR_SYSTEM_SAVE_K_1 /* Original r0 */
288 {
289 st sp, r0
290 addi sp, sp, -PTREGS_OFFSET_REG(0) - 8
291 }
292 {
293 st sp, zero /* write zero into "Next SP" frame pointer */
294 addi sp, sp, -8 /* leave SP pointing at bottom of frame */
295 }
296 .ifc \processing,handle_syscall
297 j handle_syscall
298 .else
299 /* Capture per-interrupt SPR context to registers. */
300 .ifc \c_routine, do_page_fault
301 mfspr r2, SPR_SYSTEM_SAVE_K_3 /* address of page fault */
302 mfspr r3, SPR_SYSTEM_SAVE_K_2 /* info about page fault */
303 .else
304 .ifc \vecnum, INT_ILL_TRANS
305 mfspr r2, ILL_TRANS_REASON
306 .else
307 .ifc \vecnum, INT_DOUBLE_FAULT
308 mfspr r2, SPR_SYSTEM_SAVE_K_2 /* double fault info from HV */
309 .else
310 .ifc \c_routine, do_trap
311 mfspr r2, GPV_REASON
312 .else
313 .ifc \c_routine, op_handle_perf_interrupt
314 mfspr r2, PERF_COUNT_STS
315 #if CHIP_HAS_AUX_PERF_COUNTERS()
316 .else
317 .ifc \c_routine, op_handle_aux_perf_interrupt
318 mfspr r2, AUX_PERF_COUNT_STS
319 .endif
320 #endif
321 .endif
322 .endif
323 .endif
324 .endif
325 .endif
326 /* Put function pointer in r0 */
327 moveli r0, hw2_last(\c_routine)
328 shl16insli r0, r0, hw1(\c_routine)
329 {
330 shl16insli r0, r0, hw0(\c_routine)
331 j \processing
332 }
333 .endif
334 ENDPROC(intvec_\vecname)
336 #ifdef __COLLECT_LINKER_FEEDBACK__
337 .pushsection .text.intvec_feedback,"ax"
338 .org (\vecnum << 5)
339 FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt1, 1 << 8)
340 jrp lr
341 .popsection
342 #endif
344 .endm
347 /*
348 * Save the rest of the registers that we didn't save in the actual
349 * vector itself. We can't use r0-r10 inclusive here.
350 */
351 .macro finish_interrupt_save, function
353 /* If it's a syscall, save a proper orig_r0, otherwise just zero. */
354 PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
355 {
356 .ifc \function,handle_syscall
357 st r52, r0
358 .else
359 st r52, zero
360 .endif
361 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
362 }
363 st r52, tp
364 {
365 mfspr tp, CMPEXCH_VALUE
366 PTREGS_PTR(r52, PTREGS_OFFSET_CMPEXCH)
367 }
369 /*
370 * For ordinary syscalls, we save neither caller- nor callee-
371 * save registers, since the syscall invoker doesn't expect the
372 * caller-saves to be saved, and the called kernel functions will
373 * take care of saving the callee-saves for us.
374 *
375 * For interrupts we save just the caller-save registers. Saving
376 * them is required (since the "caller" can't save them). Again,
377 * the called kernel functions will restore the callee-save
378 * registers for us appropriately.
379 *
380 * On return, we normally restore nothing special for syscalls,
381 * and just the caller-save registers for interrupts.
382 *
383 * However, there are some important caveats to all this:
384 *
385 * - We always save a few callee-save registers to give us
386 * some scratchpad registers to carry across function calls.
387 *
388 * - fork/vfork/etc require us to save all the callee-save
389 * registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
390 *
391 * - We always save r0..r5 and r10 for syscalls, since we need
392 * to reload them a bit later for the actual kernel call, and
393 * since we might need them for -ERESTARTNOINTR, etc.
394 *
395 * - Before invoking a signal handler, we save the unsaved
396 * callee-save registers so they are visible to the
397 * signal handler or any ptracer.
398 *
399 * - If the unsaved callee-save registers are modified, we set
400 * a bit in pt_regs so we know to reload them from pt_regs
401 * and not just rely on the kernel function unwinding.
402 * (Done for ptrace register writes and SA_SIGINFO handler.)
403 */
404 {
405 st r52, tp
406 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
407 }
408 wh64 r52 /* cache line 4 */
409 push_reg r33, r52
410 push_reg r32, r52
411 push_reg r31, r52
412 .ifc \function,handle_syscall
413 push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
414 push_reg TREG_SYSCALL_NR_NAME, r52, \
415 PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
416 .else
418 push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
419 wh64 r52 /* cache line 3 */
420 push_reg r29, r52
421 push_reg r28, r52
422 push_reg r27, r52
423 push_reg r26, r52
424 push_reg r25, r52
425 push_reg r24, r52
426 push_reg r23, r52
427 push_reg r22, r52
428 wh64 r52 /* cache line 2 */
429 push_reg r21, r52
430 push_reg r20, r52
431 push_reg r19, r52
432 push_reg r18, r52
433 push_reg r17, r52
434 push_reg r16, r52
435 push_reg r15, r52
436 push_reg r14, r52
437 wh64 r52 /* cache line 1 */
438 push_reg r13, r52
439 push_reg r12, r52
440 push_reg r11, r52
441 push_reg r10, r52
442 push_reg r9, r52
443 push_reg r8, r52
444 push_reg r7, r52
445 push_reg r6, r52
447 .endif
449 push_reg r5, r52
450 st r52, r4
452 /* Load tp with our per-cpu offset. */
453 #ifdef CONFIG_SMP
454 {
455 mfspr r20, SPR_SYSTEM_SAVE_K_0
456 moveli r21, hw2_last(__per_cpu_offset)
457 }
458 {
459 shl16insli r21, r21, hw1(__per_cpu_offset)
460 bfextu r20, r20, 0, LOG2_THREAD_SIZE-1
461 }
462 shl16insli r21, r21, hw0(__per_cpu_offset)
463 shl3add r20, r20, r21
464 ld tp, r20
465 #else
466 move tp, zero
467 #endif
469 /*
470 * If we will be returning to the kernel, we will need to
471 * reset the interrupt masks to the state they had before.
472 * Set DISABLE_IRQ in flags iff we came from PL1 with irqs disabled.
473 */
474 mfspr r32, SPR_EX_CONTEXT_K_1
475 {
476 andi r32, r32, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
477 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
478 }
479 beqzt r32, 1f /* zero if from user space */
480 IRQS_DISABLED(r32) /* zero if irqs enabled */
481 #if PT_FLAGS_DISABLE_IRQ != 1
482 # error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
483 #endif
484 1:
485 .ifnc \function,handle_syscall
486 /* Record the fact that we saved the caller-save registers above. */
487 ori r32, r32, PT_FLAGS_CALLER_SAVES
488 .endif
489 st r21, r32
491 #ifdef __COLLECT_LINKER_FEEDBACK__
492 /*
493 * Notify the feedback routines that we were in the
494 * appropriate fixed interrupt vector area. Note that we
495 * still have ICS set at this point, so we can't invoke any
496 * atomic operations or we will panic. The feedback
497 * routines internally preserve r0..r10 and r30 up.
498 */
499 .ifnc \function,handle_syscall
500 shli r20, r1, 5
501 .else
502 moveli r20, INT_SWINT_1 << 5
503 .endif
504 moveli r21, hw2_last(intvec_feedback)
505 shl16insli r21, r21, hw1(intvec_feedback)
506 shl16insli r21, r21, hw0(intvec_feedback)
507 add r20, r20, r21
508 jalr r20
510 /* And now notify the feedback routines that we are here. */
511 FEEDBACK_ENTER(\function)
512 #endif
514 /*
515 * we've captured enough state to the stack (including in
516 * particular our EX_CONTEXT state) that we can now release
517 * the interrupt critical section and replace it with our
518 * standard "interrupts disabled" mask value. This allows
519 * synchronous interrupts (and profile interrupts) to punch
520 * through from this point onwards.
521 */
522 .ifc \function,handle_nmi
523 IRQ_DISABLE_ALL(r20)
524 .else
525 IRQ_DISABLE(r20, r21)
526 .endif
527 mtspr INTERRUPT_CRITICAL_SECTION, zero
529 /*
530 * Prepare the first 256 stack bytes to be rapidly accessible
531 * without having to fetch the background data.
532 */
533 addi r52, sp, -64
534 {
535 wh64 r52
536 addi r52, r52, -64
537 }
538 {
539 wh64 r52
540 addi r52, r52, -64
541 }
542 {
543 wh64 r52
544 addi r52, r52, -64
545 }
546 wh64 r52
548 #ifdef CONFIG_TRACE_IRQFLAGS
549 .ifnc \function,handle_nmi
550 /*
551 * We finally have enough state set up to notify the irq
552 * tracing code that irqs were disabled on entry to the handler.
553 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
554 * For syscalls, we already have the register state saved away
555 * on the stack, so we don't bother to do any register saves here,
556 * and later we pop the registers back off the kernel stack.
557 * For interrupt handlers, save r0-r3 in callee-saved registers.
558 */
559 .ifnc \function,handle_syscall
560 { move r30, r0; move r31, r1 }
561 { move r32, r2; move r33, r3 }
562 .endif
563 TRACE_IRQS_OFF
564 .ifnc \function,handle_syscall
565 { move r0, r30; move r1, r31 }
566 { move r2, r32; move r3, r33 }
567 .endif
568 .endif
569 #endif
571 .endm
573 /*
574 * Redispatch a downcall.
575 */
576 .macro dc_dispatch vecnum, vecname
577 .org (\vecnum << 8)
578 intvec_\vecname:
579 j hv_downcall_dispatch
580 ENDPROC(intvec_\vecname)
581 .endm
583 /*
584 * Common code for most interrupts. The C function we're eventually
585 * going to is in r0, and the faultnum is in r1; the original
586 * values for those registers are on the stack.
587 */
588 .pushsection .text.handle_interrupt,"ax"
589 handle_interrupt:
590 finish_interrupt_save handle_interrupt
592 /* Jump to the C routine; it should enable irqs as soon as possible. */
593 {
594 jalr r0
595 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
596 }
597 FEEDBACK_REENTER(handle_interrupt)
598 {
599 movei r30, 0 /* not an NMI */
600 j interrupt_return
601 }
602 STD_ENDPROC(handle_interrupt)
604 /*
605 * This routine takes a boolean in r30 indicating if this is an NMI.
606 * If so, we also expect a boolean in r31 indicating whether to
607 * re-enable the oprofile interrupts.
608 */
609 STD_ENTRY(interrupt_return)
610 /* If we're resuming to kernel space, don't check thread flags. */
611 {
612 bnez r30, .Lrestore_all /* NMIs don't special-case user-space */
613 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
614 }
615 ld r29, r29
616 andi r29, r29, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
617 {
618 beqzt r29, .Lresume_userspace
619 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
620 }
622 /* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
623 moveli r27, hw2_last(_cpu_idle_nap)
624 {
625 ld r28, r29
626 shl16insli r27, r27, hw1(_cpu_idle_nap)
627 }
628 {
629 shl16insli r27, r27, hw0(_cpu_idle_nap)
630 }
631 {
632 cmpeq r27, r27, r28
633 }
634 {
635 blbc r27, .Lrestore_all
636 addi r28, r28, 8
637 }
638 st r29, r28
639 j .Lrestore_all
641 .Lresume_userspace:
642 FEEDBACK_REENTER(interrupt_return)
644 /*
645 * Disable interrupts so as to make sure we don't
646 * miss an interrupt that sets any of the thread flags (like
647 * need_resched or sigpending) between sampling and the iret.
648 * Routines like schedule() or do_signal() may re-enable
649 * interrupts before returning.
650 */
651 IRQ_DISABLE(r20, r21)
652 TRACE_IRQS_OFF /* Note: clobbers registers r0-r29 */
654 /* Get base of stack in r32; note r30/31 are used as arguments here. */
655 GET_THREAD_INFO(r32)
658 /* Check to see if there is any work to do before returning to user. */
659 {
660 addi r29, r32, THREAD_INFO_FLAGS_OFFSET
661 moveli r1, hw1_last(_TIF_ALLWORK_MASK)
662 }
663 {
664 ld r29, r29
665 shl16insli r1, r1, hw0(_TIF_ALLWORK_MASK)
666 }
667 and r1, r29, r1
668 beqzt r1, .Lrestore_all
670 /*
671 * Make sure we have all the registers saved for signal
672 * handling or single-step. Call out to C code to figure out
673 * exactly what we need to do for each flag bit, then if
674 * necessary, reload the flags and recheck.
675 */
676 push_extra_callee_saves r0
677 {
678 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
679 jal do_work_pending
680 }
681 bnez r0, .Lresume_userspace
683 /*
684 * In the NMI case we
685 * omit the call to single_process_check_nohz, which normally checks
686 * to see if we should start or stop the scheduler tick, because
687 * we can't call arbitrary Linux code from an NMI context.
688 * We always call the homecache TLB deferral code to re-trigger
689 * the deferral mechanism.
690 *
691 * The other chunk of responsibility this code has is to reset the
692 * interrupt masks appropriately to reset irqs and NMIs. We have
693 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
694 * lockdep-type stuff, but we can't set ICS until afterwards, since
695 * ICS can only be used in very tight chunks of code to avoid
696 * tripping over various assertions that it is off.
697 */
698 .Lrestore_all:
699 PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
700 {
701 ld r0, r0
702 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
703 }
704 {
705 andi r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK
706 ld r32, r32
707 }
708 bnez r0, 1f
709 j 2f
710 #if PT_FLAGS_DISABLE_IRQ != 1
711 # error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use blbct below
712 #endif
713 1: blbct r32, 2f
714 IRQ_DISABLE(r20,r21)
715 TRACE_IRQS_OFF
716 movei r0, 1
717 mtspr INTERRUPT_CRITICAL_SECTION, r0
718 beqzt r30, .Lrestore_regs
719 j 3f
720 2: TRACE_IRQS_ON
721 movei r0, 1
722 mtspr INTERRUPT_CRITICAL_SECTION, r0
723 IRQ_ENABLE(r20, r21)
724 beqzt r30, .Lrestore_regs
725 3:
728 /*
729 * We now commit to returning from this interrupt, since we will be
730 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
731 * frame. No calls should be made to any other code after this point.
732 * This code should only be entered with ICS set.
733 * r32 must still be set to ptregs.flags.
734 * We launch loads to each cache line separately first, so we can
735 * get some parallelism out of the memory subsystem.
736 * We start zeroing caller-saved registers throughout, since
737 * that will save some cycles if this turns out to be a syscall.
738 */
739 .Lrestore_regs:
740 FEEDBACK_REENTER(interrupt_return) /* called from elsewhere */
742 /*
743 * Rotate so we have one high bit and one low bit to test.
744 * - low bit says whether to restore all the callee-saved registers,
745 * or just r30-r33, and r52 up.
746 * - high bit (i.e. sign bit) says whether to restore all the
747 * caller-saved registers, or just r0.
748 */
749 #if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
750 # error Rotate trick does not work :-)
751 #endif
752 {
753 rotli r20, r32, 62
754 PTREGS_PTR(sp, PTREGS_OFFSET_REG(0))
755 }
757 /*
758 * Load cache lines 0, 4, 6 and 7, in that order, then use
759 * the last loaded value, which makes it likely that the other
760 * cache lines have also loaded, at which point we should be
761 * able to safely read all the remaining words on those cache
762 * lines without waiting for the memory subsystem.
763 */
764 pop_reg r0, sp, PTREGS_OFFSET_REG(30) - PTREGS_OFFSET_REG(0)
765 pop_reg r30, sp, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_REG(30)
766 pop_reg_zero r52, r3, sp, PTREGS_OFFSET_CMPEXCH - PTREGS_OFFSET_REG(52)
767 pop_reg_zero r21, r27, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_CMPEXCH
768 pop_reg_zero lr, r2, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_EX1
769 {
770 mtspr CMPEXCH_VALUE, r21
771 move r4, zero
772 }
773 pop_reg r21, sp, PTREGS_OFFSET_REG(31) - PTREGS_OFFSET_PC
774 {
775 mtspr SPR_EX_CONTEXT_K_1, lr
776 andi lr, lr, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
777 }
778 {
779 mtspr SPR_EX_CONTEXT_K_0, r21
780 move r5, zero
781 }
783 /* Restore callee-saveds that we actually use. */
784 pop_reg_zero r31, r6
785 pop_reg_zero r32, r7
786 pop_reg_zero r33, r8, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(33)
788 /*
789 * If we modified other callee-saveds, restore them now.
790 * This is rare, but could be via ptrace or signal handler.
791 */
792 {
793 move r9, zero
794 blbs r20, .Lrestore_callees
795 }
796 .Lcontinue_restore_regs:
798 /* Check if we're returning from a syscall. */
799 {
800 move r10, zero
801 bltzt r20, 1f /* no, so go restore callee-save registers */
802 }
804 /*
805 * Check if we're returning to userspace.
806 * Note that if we're not, we don't worry about zeroing everything.
807 */
808 {
809 addli sp, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(29)
810 bnez lr, .Lkernel_return
811 }
813 /*
814 * On return from syscall, we've restored r0 from pt_regs, but we
815 * clear the remainder of the caller-saved registers. We could
816 * restore the syscall arguments, but there's not much point,
817 * and it ensures user programs aren't trying to use the
818 * caller-saves if we clear them, as well as avoiding leaking
819 * kernel pointers into userspace.
820 */
821 pop_reg_zero lr, r11, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
822 pop_reg_zero tp, r12, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
823 {
824 ld sp, sp
825 move r13, zero
826 move r14, zero
827 }
828 { move r15, zero; move r16, zero }
829 { move r17, zero; move r18, zero }
830 { move r19, zero; move r20, zero }
831 { move r21, zero; move r22, zero }
832 { move r23, zero; move r24, zero }
833 { move r25, zero; move r26, zero }
835 /* Set r1 to errno if we are returning an error, otherwise zero. */
836 {
837 moveli r29, 4096
838 sub r1, zero, r0
839 }
840 {
841 move r28, zero
842 cmpltu r29, r1, r29
843 }
844 {
845 mnz r1, r29, r1
846 move r29, zero
847 }
848 iret
850 /*
851 * Not a syscall, so restore caller-saved registers.
852 * First kick off loads for cache lines 1-3, which we're touching
853 * for the first time here.
854 */
855 .align 64
856 1: pop_reg r29, sp, PTREGS_OFFSET_REG(21) - PTREGS_OFFSET_REG(29)
857 pop_reg r21, sp, PTREGS_OFFSET_REG(13) - PTREGS_OFFSET_REG(21)
858 pop_reg r13, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(13)
859 pop_reg r1
860 pop_reg r2
861 pop_reg r3
862 pop_reg r4
863 pop_reg r5
864 pop_reg r6
865 pop_reg r7
866 pop_reg r8
867 pop_reg r9
868 pop_reg r10
869 pop_reg r11
870 pop_reg r12, sp, 16
871 /* r13 already restored above */
872 pop_reg r14
873 pop_reg r15
874 pop_reg r16
875 pop_reg r17
876 pop_reg r18
877 pop_reg r19
878 pop_reg r20, sp, 16
879 /* r21 already restored above */
880 pop_reg r22
881 pop_reg r23
882 pop_reg r24
883 pop_reg r25
884 pop_reg r26
885 pop_reg r27
886 pop_reg r28, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(28)
887 /* r29 already restored above */
888 bnez lr, .Lkernel_return
889 pop_reg lr, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
890 pop_reg tp, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
891 ld sp, sp
892 iret
894 /*
895 * We can't restore tp when in kernel mode, since a thread might
896 * have migrated from another cpu and brought a stale tp value.
897 */
898 .Lkernel_return:
899 pop_reg lr, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
900 ld sp, sp
901 iret
903 /* Restore callee-saved registers from r34 to r51. */
904 .Lrestore_callees:
905 addli sp, sp, PTREGS_OFFSET_REG(34) - PTREGS_OFFSET_REG(29)
906 pop_reg r34
907 pop_reg r35
908 pop_reg r36
909 pop_reg r37
910 pop_reg r38
911 pop_reg r39
912 pop_reg r40
913 pop_reg r41
914 pop_reg r42
915 pop_reg r43
916 pop_reg r44
917 pop_reg r45
918 pop_reg r46
919 pop_reg r47
920 pop_reg r48
921 pop_reg r49
922 pop_reg r50
923 pop_reg r51, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(51)
924 j .Lcontinue_restore_regs
925 STD_ENDPROC(interrupt_return)
927 /*
928 * "NMI" interrupts mask ALL interrupts before calling the
929 * handler, and don't check thread flags, etc., on the way
930 * back out. In general, the only things we do here for NMIs
931 * are register save/restore and dataplane kernel-TLB management.
932 * We don't (for example) deal with start/stop of the sched tick.
933 */
934 .pushsection .text.handle_nmi,"ax"
935 handle_nmi:
936 finish_interrupt_save handle_nmi
937 {
938 jalr r0
939 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
940 }
941 FEEDBACK_REENTER(handle_nmi)
942 {
943 movei r30, 1
944 move r31, r0
945 }
946 j interrupt_return
947 STD_ENDPROC(handle_nmi)
949 /*
950 * Parallel code for syscalls to handle_interrupt.
951 */
952 .pushsection .text.handle_syscall,"ax"
953 handle_syscall:
954 finish_interrupt_save handle_syscall
956 /* Enable irqs. */
957 TRACE_IRQS_ON
958 IRQ_ENABLE(r20, r21)
960 /* Bump the counter for syscalls made on this tile. */
961 moveli r20, hw2_last(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
962 shl16insli r20, r20, hw1(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
963 shl16insli r20, r20, hw0(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
964 add r20, r20, tp
965 ld4s r21, r20
966 addi r21, r21, 1
967 st4 r20, r21
969 /* Trace syscalls, if requested. */
970 GET_THREAD_INFO(r31)
971 addi r31, r31, THREAD_INFO_FLAGS_OFFSET
972 ld r30, r31
973 andi r30, r30, _TIF_SYSCALL_TRACE
974 {
975 addi r30, r31, THREAD_INFO_STATUS_OFFSET - THREAD_INFO_FLAGS_OFFSET
976 beqzt r30, .Lrestore_syscall_regs
977 }
978 jal do_syscall_trace
979 FEEDBACK_REENTER(handle_syscall)
981 /*
982 * We always reload our registers from the stack at this
983 * point. They might be valid, if we didn't build with
984 * TRACE_IRQFLAGS, and this isn't a dataplane tile, and we're not
985 * doing syscall tracing, but there are enough cases now that it
986 * seems simplest just to do the reload unconditionally.
987 */
988 .Lrestore_syscall_regs:
989 {
990 ld r30, r30
991 PTREGS_PTR(r11, PTREGS_OFFSET_REG(0))
992 }
993 pop_reg r0, r11
994 pop_reg r1, r11
995 pop_reg r2, r11
996 pop_reg r3, r11
997 pop_reg r4, r11
998 pop_reg r5, r11, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(5)
999 {
1000 ld TREG_SYSCALL_NR_NAME, r11
1001 moveli r21, __NR_syscalls
1002 }
1004 /* Ensure that the syscall number is within the legal range. */
1005 {
1006 moveli r20, hw2(sys_call_table)
1007 blbs r30, .Lcompat_syscall
1008 }
1009 {
1010 cmpltu r21, TREG_SYSCALL_NR_NAME, r21
1011 shl16insli r20, r20, hw1(sys_call_table)
1012 }
1013 {
1014 blbc r21, .Linvalid_syscall
1015 shl16insli r20, r20, hw0(sys_call_table)
1016 }
1017 .Lload_syscall_pointer:
1018 shl3add r20, TREG_SYSCALL_NR_NAME, r20
1019 ld r20, r20
1021 /* Jump to syscall handler. */
1022 jalr r20
1023 .Lhandle_syscall_link: /* value of "lr" after "jalr r20" above */
1025 /*
1026 * Write our r0 onto the stack so it gets restored instead
1027 * of whatever the user had there before.
1028 * In compat mode, sign-extend r0 before storing it.
1029 */
1030 {
1031 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1032 blbct r30, 1f
1033 }
1034 addxi r0, r0, 0
1035 1: st r29, r0
1037 .Lsyscall_sigreturn_skip:
1038 FEEDBACK_REENTER(handle_syscall)
1040 /* Do syscall trace again, if requested. */
1041 ld r30, r31
1042 andi r30, r30, _TIF_SYSCALL_TRACE
1043 beqzt r30, 1f
1044 jal do_syscall_trace
1045 FEEDBACK_REENTER(handle_syscall)
1046 1: j .Lresume_userspace /* jump into middle of interrupt_return */
1048 .Lcompat_syscall:
1049 /*
1050 * Load the base of the compat syscall table in r20, and
1051 * range-check the syscall number (duplicated from 64-bit path).
1052 * Sign-extend all the user's passed arguments to make them consistent.
1053 * Also save the original "r(n)" values away in "r(11+n)" in
1054 * case the syscall table entry wants to validate them.
1055 */
1056 moveli r20, hw2(compat_sys_call_table)
1057 {
1058 cmpltu r21, TREG_SYSCALL_NR_NAME, r21
1059 shl16insli r20, r20, hw1(compat_sys_call_table)
1060 }
1061 {
1062 blbc r21, .Linvalid_syscall
1063 shl16insli r20, r20, hw0(compat_sys_call_table)
1064 }
1065 { move r11, r0; addxi r0, r0, 0 }
1066 { move r12, r1; addxi r1, r1, 0 }
1067 { move r13, r2; addxi r2, r2, 0 }
1068 { move r14, r3; addxi r3, r3, 0 }
1069 { move r15, r4; addxi r4, r4, 0 }
1070 { move r16, r5; addxi r5, r5, 0 }
1071 j .Lload_syscall_pointer
1073 .Linvalid_syscall:
1074 /* Report an invalid syscall back to the user program */
1075 {
1076 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1077 movei r28, -ENOSYS
1078 }
1079 st r29, r28
1080 j .Lresume_userspace /* jump into middle of interrupt_return */
1081 STD_ENDPROC(handle_syscall)
1083 /* Return the address for oprofile to suppress in backtraces. */
1084 STD_ENTRY_SECTION(handle_syscall_link_address, .text.handle_syscall)
1085 lnk r0
1086 {
1087 addli r0, r0, .Lhandle_syscall_link - .
1088 jrp lr
1089 }
1090 STD_ENDPROC(handle_syscall_link_address)
1092 STD_ENTRY(ret_from_fork)
1093 jal sim_notify_fork
1094 jal schedule_tail
1095 FEEDBACK_REENTER(ret_from_fork)
1096 j .Lresume_userspace
1097 STD_ENDPROC(ret_from_fork)
1099 /* Various stub interrupt handlers and syscall handlers */
1101 STD_ENTRY_LOCAL(_kernel_double_fault)
1102 mfspr r1, SPR_EX_CONTEXT_K_0
1103 move r2, lr
1104 move r3, sp
1105 move r4, r52
1106 addi sp, sp, -C_ABI_SAVE_AREA_SIZE
1107 j kernel_double_fault
1108 STD_ENDPROC(_kernel_double_fault)
1110 STD_ENTRY_LOCAL(bad_intr)
1111 mfspr r2, SPR_EX_CONTEXT_K_0
1112 panic "Unhandled interrupt %#x: PC %#lx"
1113 STD_ENDPROC(bad_intr)
1115 /* Put address of pt_regs in reg and jump. */
1116 #define PTREGS_SYSCALL(x, reg) \
1117 STD_ENTRY(_##x); \
1118 { \
1119 PTREGS_PTR(reg, PTREGS_OFFSET_BASE); \
1120 j x \
1121 }; \
1122 STD_ENDPROC(_##x)
1124 /*
1125 * Special-case sigreturn to not write r0 to the stack on return.
1126 * This is technically more efficient, but it also avoids difficulties
1127 * in the 64-bit OS when handling 32-bit compat code, since we must not
1128 * sign-extend r0 for the sigreturn return-value case.
1129 */
1130 #define PTREGS_SYSCALL_SIGRETURN(x, reg) \
1131 STD_ENTRY(_##x); \
1132 addli lr, lr, .Lsyscall_sigreturn_skip - .Lhandle_syscall_link; \
1133 { \
1134 PTREGS_PTR(reg, PTREGS_OFFSET_BASE); \
1135 j x \
1136 }; \
1137 STD_ENDPROC(_##x)
1139 PTREGS_SYSCALL(sys_execve, r3)
1140 PTREGS_SYSCALL(sys_sigaltstack, r2)
1141 PTREGS_SYSCALL_SIGRETURN(sys_rt_sigreturn, r0)
1142 #ifdef CONFIG_COMPAT
1143 PTREGS_SYSCALL(compat_sys_execve, r3)
1144 PTREGS_SYSCALL(compat_sys_sigaltstack, r2)
1145 PTREGS_SYSCALL_SIGRETURN(compat_sys_rt_sigreturn, r0)
1146 #endif
1148 /* Save additional callee-saves to pt_regs, put address in r4 and jump. */
1149 STD_ENTRY(_sys_clone)
1150 push_extra_callee_saves r4
1151 j sys_clone
1152 STD_ENDPROC(_sys_clone)
1154 /* The single-step support may need to read all the registers. */
1155 int_unalign:
1156 push_extra_callee_saves r0
1157 j do_trap
1159 /* Include .intrpt1 array of interrupt vectors */
1160 .section ".intrpt1", "ax"
1162 #define op_handle_perf_interrupt bad_intr
1163 #define op_handle_aux_perf_interrupt bad_intr
1165 #ifndef CONFIG_HARDWALL
1166 #define do_hardwall_trap bad_intr
1167 #endif
1169 int_hand INT_MEM_ERROR, MEM_ERROR, bad_intr
1170 int_hand INT_SINGLE_STEP_3, SINGLE_STEP_3, bad_intr
1171 #if CONFIG_KERNEL_PL == 2
1172 int_hand INT_SINGLE_STEP_2, SINGLE_STEP_2, gx_singlestep_handle
1173 int_hand INT_SINGLE_STEP_1, SINGLE_STEP_1, bad_intr
1174 #else
1175 int_hand INT_SINGLE_STEP_2, SINGLE_STEP_2, bad_intr
1176 int_hand INT_SINGLE_STEP_1, SINGLE_STEP_1, gx_singlestep_handle
1177 #endif
1178 int_hand INT_SINGLE_STEP_0, SINGLE_STEP_0, bad_intr
1179 int_hand INT_IDN_COMPLETE, IDN_COMPLETE, bad_intr
1180 int_hand INT_UDN_COMPLETE, UDN_COMPLETE, bad_intr
1181 int_hand INT_ITLB_MISS, ITLB_MISS, do_page_fault
1182 int_hand INT_ILL, ILL, do_trap
1183 int_hand INT_GPV, GPV, do_trap
1184 int_hand INT_IDN_ACCESS, IDN_ACCESS, do_trap
1185 int_hand INT_UDN_ACCESS, UDN_ACCESS, do_trap
1186 int_hand INT_SWINT_3, SWINT_3, do_trap
1187 int_hand INT_SWINT_2, SWINT_2, do_trap
1188 int_hand INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall
1189 int_hand INT_SWINT_0, SWINT_0, do_trap
1190 int_hand INT_ILL_TRANS, ILL_TRANS, do_trap
1191 int_hand INT_UNALIGN_DATA, UNALIGN_DATA, int_unalign
1192 int_hand INT_DTLB_MISS, DTLB_MISS, do_page_fault
1193 int_hand INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault
1194 int_hand INT_IDN_FIREWALL, IDN_FIREWALL, bad_intr
1195 int_hand INT_UDN_FIREWALL, UDN_FIREWALL, do_hardwall_trap
1196 int_hand INT_TILE_TIMER, TILE_TIMER, do_timer_interrupt
1197 int_hand INT_IDN_TIMER, IDN_TIMER, bad_intr
1198 int_hand INT_UDN_TIMER, UDN_TIMER, bad_intr
1199 int_hand INT_IDN_AVAIL, IDN_AVAIL, bad_intr
1200 int_hand INT_UDN_AVAIL, UDN_AVAIL, bad_intr
1201 int_hand INT_IPI_3, IPI_3, bad_intr
1202 #if CONFIG_KERNEL_PL == 2
1203 int_hand INT_IPI_2, IPI_2, tile_dev_intr
1204 int_hand INT_IPI_1, IPI_1, bad_intr
1205 #else
1206 int_hand INT_IPI_2, IPI_2, bad_intr
1207 int_hand INT_IPI_1, IPI_1, tile_dev_intr
1208 #endif
1209 int_hand INT_IPI_0, IPI_0, bad_intr
1210 int_hand INT_PERF_COUNT, PERF_COUNT, \
1211 op_handle_perf_interrupt, handle_nmi
1212 int_hand INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \
1213 op_handle_perf_interrupt, handle_nmi
1214 int_hand INT_INTCTRL_3, INTCTRL_3, bad_intr
1215 #if CONFIG_KERNEL_PL == 2
1216 dc_dispatch INT_INTCTRL_2, INTCTRL_2
1217 int_hand INT_INTCTRL_1, INTCTRL_1, bad_intr
1218 #else
1219 int_hand INT_INTCTRL_2, INTCTRL_2, bad_intr
1220 dc_dispatch INT_INTCTRL_1, INTCTRL_1
1221 #endif
1222 int_hand INT_INTCTRL_0, INTCTRL_0, bad_intr
1223 int_hand INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
1224 hv_message_intr
1225 int_hand INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, bad_intr
1226 int_hand INT_I_ASID, I_ASID, bad_intr
1227 int_hand INT_D_ASID, D_ASID, bad_intr
1228 int_hand INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
1230 /* Synthetic interrupt delivered only by the simulator */
1231 int_hand INT_BREAKPOINT, BREAKPOINT, do_breakpoint