]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - arch/tile/kernel/intvec_32.S
Merge remote-tracking branches 'regulator/topic/discharge', 'regulator/topic/fan53555...
[mirror_ubuntu-artful-kernel.git] / arch / tile / kernel / intvec_32.S
1 /*
2 * Copyright 2010 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/init.h>
20 #include <linux/unistd.h>
21 #include <asm/ptrace.h>
22 #include <asm/thread_info.h>
23 #include <asm/irqflags.h>
24 #include <asm/atomic_32.h>
25 #include <asm/asm-offsets.h>
26 #include <hv/hypervisor.h>
27 #include <arch/abi.h>
28 #include <arch/interrupts.h>
29 #include <arch/spr_def.h>
30
31 #define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
32
33 #define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
34
35 .macro push_reg reg, ptr=sp, delta=-4
36 {
37 sw \ptr, \reg
38 addli \ptr, \ptr, \delta
39 }
40 .endm
41
42 .macro pop_reg reg, ptr=sp, delta=4
43 {
44 lw \reg, \ptr
45 addli \ptr, \ptr, \delta
46 }
47 .endm
48
49 .macro pop_reg_zero reg, zreg, ptr=sp, delta=4
50 {
51 move \zreg, zero
52 lw \reg, \ptr
53 addi \ptr, \ptr, \delta
54 }
55 .endm
56
57 .macro push_extra_callee_saves reg
58 PTREGS_PTR(\reg, PTREGS_OFFSET_REG(51))
59 push_reg r51, \reg
60 push_reg r50, \reg
61 push_reg r49, \reg
62 push_reg r48, \reg
63 push_reg r47, \reg
64 push_reg r46, \reg
65 push_reg r45, \reg
66 push_reg r44, \reg
67 push_reg r43, \reg
68 push_reg r42, \reg
69 push_reg r41, \reg
70 push_reg r40, \reg
71 push_reg r39, \reg
72 push_reg r38, \reg
73 push_reg r37, \reg
74 push_reg r36, \reg
75 push_reg r35, \reg
76 push_reg r34, \reg, PTREGS_OFFSET_BASE - PTREGS_OFFSET_REG(34)
77 .endm
78
79 .macro panic str
80 .pushsection .rodata, "a"
81 1:
82 .asciz "\str"
83 .popsection
84 {
85 moveli r0, lo16(1b)
86 }
87 {
88 auli r0, r0, ha16(1b)
89 jal panic
90 }
91 .endm
92
93 #ifdef __COLLECT_LINKER_FEEDBACK__
94 .pushsection .text.intvec_feedback,"ax"
95 intvec_feedback:
96 .popsection
97 #endif
98
99 /*
100 * Default interrupt handler.
101 *
102 * vecnum is where we'll put this code.
103 * c_routine is the C routine we'll call.
104 *
105 * The C routine is passed two arguments:
106 * - A pointer to the pt_regs state.
107 * - The interrupt vector number.
108 *
109 * The "processing" argument specifies the code for processing
110 * the interrupt. Defaults to "handle_interrupt".
111 */
112 .macro int_hand vecnum, vecname, c_routine, processing=handle_interrupt
113 .org (\vecnum << 8)
114 intvec_\vecname:
115 .ifc \vecnum, INT_SWINT_1
116 blz TREG_SYSCALL_NR_NAME, sys_cmpxchg
117 .endif
118
119 /* Temporarily save a register so we have somewhere to work. */
120
121 mtspr SPR_SYSTEM_SAVE_K_1, r0
122 mfspr r0, SPR_EX_CONTEXT_K_1
123
124 /* The cmpxchg code clears sp to force us to reset it here on fault. */
125 {
126 bz sp, 2f
127 andi r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
128 }
129
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 bz 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 bnz r0, 0f
147 move r0, sp
148 }
149 .endif
150
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 mfspr r0, SPR_SYSTEM_SAVE_K_2
168 {
169 blz r0, 0f /* high bit in S_S_1_2 is for a PC to use */
170 move r0, sp
171 }
172 .endif
173
174 2:
175 /*
176 * SYSTEM_SAVE_K_0 holds the cpu number in the low bits, and
177 * the current stack top in the higher bits. So we recover
178 * our stack top by just masking off the low bits, then
179 * point sp at the top aligned address on the actual stack page.
180 */
181 mfspr r0, SPR_SYSTEM_SAVE_K_0
182 mm r0, r0, zero, LOG2_NR_CPU_IDS, 31
183
184 0:
185 /*
186 * Align the stack mod 64 so we can properly predict what
187 * cache lines we need to write-hint to reduce memory fetch
188 * latency as we enter the kernel. The layout of memory is
189 * as follows, with cache line 0 at the lowest VA, and cache
190 * line 4 just below the r0 value this "andi" computes.
191 * Note that we never write to cache line 4, and we skip
192 * cache line 1 for syscalls.
193 *
194 * cache line 4: ptregs padding (two words)
195 * cache line 3: r46...lr, pc, ex1, faultnum, orig_r0, flags, pad
196 * cache line 2: r30...r45
197 * cache line 1: r14...r29
198 * cache line 0: 2 x frame, r0..r13
199 */
200 #if STACK_TOP_DELTA != 64
201 #error STACK_TOP_DELTA must be 64 for assumptions here and in task_pt_regs()
202 #endif
203 andi r0, r0, -64
204
205 /*
206 * Push the first four registers on the stack, so that we can set
207 * them to vector-unique values before we jump to the common code.
208 *
209 * Registers are pushed on the stack as a struct pt_regs,
210 * with the sp initially just above the struct, and when we're
211 * done, sp points to the base of the struct, minus
212 * C_ABI_SAVE_AREA_SIZE, so we can directly jal to C code.
213 *
214 * This routine saves just the first four registers, plus the
215 * stack context so we can do proper backtracing right away,
216 * and defers to handle_interrupt to save the rest.
217 * The backtracer needs pc, ex1, lr, sp, r52, and faultnum.
218 */
219 addli r0, r0, PTREGS_OFFSET_LR - (PTREGS_SIZE + KSTK_PTREGS_GAP)
220 wh64 r0 /* cache line 3 */
221 {
222 sw r0, lr
223 addli r0, r0, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
224 }
225 {
226 sw r0, sp
227 addli sp, r0, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_SP
228 }
229 {
230 sw sp, r52
231 addli sp, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(52)
232 }
233 wh64 sp /* cache line 0 */
234 {
235 sw sp, r1
236 addli sp, sp, PTREGS_OFFSET_REG(2) - PTREGS_OFFSET_REG(1)
237 }
238 {
239 sw sp, r2
240 addli sp, sp, PTREGS_OFFSET_REG(3) - PTREGS_OFFSET_REG(2)
241 }
242 {
243 sw sp, r3
244 addli sp, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(3)
245 }
246 mfspr r0, SPR_EX_CONTEXT_K_0
247 .ifc \processing,handle_syscall
248 /*
249 * Bump the saved PC by one bundle so that when we return, we won't
250 * execute the same swint instruction again. We need to do this while
251 * we're in the critical section.
252 */
253 addi r0, r0, 8
254 .endif
255 {
256 sw sp, r0
257 addli sp, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
258 }
259 mfspr r0, SPR_EX_CONTEXT_K_1
260 {
261 sw sp, r0
262 addi sp, sp, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
263 /*
264 * Use r0 for syscalls so it's a temporary; use r1 for interrupts
265 * so that it gets passed through unchanged to the handler routine.
266 * Note that the .if conditional confusingly spans bundles.
267 */
268 .ifc \processing,handle_syscall
269 movei r0, \vecnum
270 }
271 {
272 sw sp, r0
273 .else
274 movei r1, \vecnum
275 }
276 {
277 sw sp, r1
278 .endif
279 addli sp, sp, PTREGS_OFFSET_REG(0) - PTREGS_OFFSET_FAULTNUM
280 }
281 mfspr r0, SPR_SYSTEM_SAVE_K_1 /* Original r0 */
282 {
283 sw sp, r0
284 addi sp, sp, -PTREGS_OFFSET_REG(0) - 4
285 }
286 {
287 sw sp, zero /* write zero into "Next SP" frame pointer */
288 addi sp, sp, -4 /* leave SP pointing at bottom of frame */
289 }
290 .ifc \processing,handle_syscall
291 j handle_syscall
292 .else
293 /*
294 * Capture per-interrupt SPR context to registers.
295 * We overload the meaning of r3 on this path such that if its bit 31
296 * is set, we have to mask all interrupts including NMIs before
297 * clearing the interrupt critical section bit.
298 * See discussion below at "finish_interrupt_save".
299 */
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_DOUBLE_FAULT
305 {
306 mfspr r2, SPR_SYSTEM_SAVE_K_2 /* double fault info from HV */
307 movei r3, 0
308 }
309 .else
310 .ifc \c_routine, do_trap
311 {
312 mfspr r2, GPV_REASON
313 movei r3, 0
314 }
315 .else
316 .ifc \c_routine, handle_perf_interrupt
317 {
318 mfspr r2, PERF_COUNT_STS
319 movei r3, -1 /* not used, but set for consistency */
320 }
321 .else
322 .ifc \c_routine, handle_perf_interrupt
323 {
324 mfspr r2, AUX_PERF_COUNT_STS
325 movei r3, -1 /* not used, but set for consistency */
326 }
327 .else
328 movei r3, 0
329 .endif
330 .endif
331 .endif
332 .endif
333 .endif
334 /* Put function pointer in r0 */
335 moveli r0, lo16(\c_routine)
336 {
337 auli r0, r0, ha16(\c_routine)
338 j \processing
339 }
340 .endif
341 ENDPROC(intvec_\vecname)
342
343 #ifdef __COLLECT_LINKER_FEEDBACK__
344 .pushsection .text.intvec_feedback,"ax"
345 .org (\vecnum << 5)
346 FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt, 1 << 8)
347 jrp lr
348 .popsection
349 #endif
350
351 .endm
352
353
354 /*
355 * Save the rest of the registers that we didn't save in the actual
356 * vector itself. We can't use r0-r10 inclusive here.
357 */
358 .macro finish_interrupt_save, function
359
360 /* If it's a syscall, save a proper orig_r0, otherwise just zero. */
361 PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
362 {
363 .ifc \function,handle_syscall
364 sw r52, r0
365 .else
366 sw r52, zero
367 .endif
368 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
369 }
370
371 /*
372 * For ordinary syscalls, we save neither caller- nor callee-
373 * save registers, since the syscall invoker doesn't expect the
374 * caller-saves to be saved, and the called kernel functions will
375 * take care of saving the callee-saves for us.
376 *
377 * For interrupts we save just the caller-save registers. Saving
378 * them is required (since the "caller" can't save them). Again,
379 * the called kernel functions will restore the callee-save
380 * registers for us appropriately.
381 *
382 * On return, we normally restore nothing special for syscalls,
383 * and just the caller-save registers for interrupts.
384 *
385 * However, there are some important caveats to all this:
386 *
387 * - We always save a few callee-save registers to give us
388 * some scratchpad registers to carry across function calls.
389 *
390 * - fork/vfork/etc require us to save all the callee-save
391 * registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
392 *
393 * - We always save r0..r5 and r10 for syscalls, since we need
394 * to reload them a bit later for the actual kernel call, and
395 * since we might need them for -ERESTARTNOINTR, etc.
396 *
397 * - Before invoking a signal handler, we save the unsaved
398 * callee-save registers so they are visible to the
399 * signal handler or any ptracer.
400 *
401 * - If the unsaved callee-save registers are modified, we set
402 * a bit in pt_regs so we know to reload them from pt_regs
403 * and not just rely on the kernel function unwinding.
404 * (Done for ptrace register writes and SA_SIGINFO handler.)
405 */
406 {
407 sw r52, tp
408 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
409 }
410 wh64 r52 /* cache line 2 */
411 push_reg r33, r52
412 push_reg r32, r52
413 push_reg r31, r52
414 .ifc \function,handle_syscall
415 push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
416 push_reg TREG_SYSCALL_NR_NAME, r52, \
417 PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
418 .else
419
420 push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
421 wh64 r52 /* cache line 1 */
422 push_reg r29, r52
423 push_reg r28, r52
424 push_reg r27, r52
425 push_reg r26, r52
426 push_reg r25, r52
427 push_reg r24, r52
428 push_reg r23, r52
429 push_reg r22, r52
430 push_reg r21, r52
431 push_reg r20, r52
432 push_reg r19, r52
433 push_reg r18, r52
434 push_reg r17, r52
435 push_reg r16, r52
436 push_reg r15, r52
437 push_reg r14, r52
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
446
447 .endif
448
449 push_reg r5, r52
450 sw r52, r4
451
452 /* Load tp with our per-cpu offset. */
453 #ifdef CONFIG_SMP
454 {
455 mfspr r20, SPR_SYSTEM_SAVE_K_0
456 moveli r21, lo16(__per_cpu_offset)
457 }
458 {
459 auli r21, r21, ha16(__per_cpu_offset)
460 mm r20, r20, zero, 0, LOG2_NR_CPU_IDS-1
461 }
462 s2a r20, r20, r21
463 lw tp, r20
464 #else
465 move tp, zero
466 #endif
467
468 /*
469 * If we will be returning to the kernel, we will need to
470 * reset the interrupt masks to the state they had before.
471 * Set DISABLE_IRQ in flags iff we came from PL1 with irqs disabled.
472 * We load flags in r32 here so we can jump to .Lrestore_regs
473 * directly after do_page_fault_ics() if necessary.
474 */
475 mfspr r32, SPR_EX_CONTEXT_K_1
476 {
477 andi r32, r32, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
478 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
479 }
480 bzt r32, 1f /* zero if from user space */
481 IRQS_DISABLED(r32) /* zero if irqs enabled */
482 #if PT_FLAGS_DISABLE_IRQ != 1
483 # error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
484 #endif
485 1:
486 .ifnc \function,handle_syscall
487 /* Record the fact that we saved the caller-save registers above. */
488 ori r32, r32, PT_FLAGS_CALLER_SAVES
489 .endif
490 sw r21, r32
491
492 #ifdef __COLLECT_LINKER_FEEDBACK__
493 /*
494 * Notify the feedback routines that we were in the
495 * appropriate fixed interrupt vector area. Note that we
496 * still have ICS set at this point, so we can't invoke any
497 * atomic operations or we will panic. The feedback
498 * routines internally preserve r0..r10 and r30 up.
499 */
500 .ifnc \function,handle_syscall
501 shli r20, r1, 5
502 .else
503 moveli r20, INT_SWINT_1 << 5
504 .endif
505 addli r20, r20, lo16(intvec_feedback)
506 auli r20, r20, ha16(intvec_feedback)
507 jalr r20
508
509 /* And now notify the feedback routines that we are here. */
510 FEEDBACK_ENTER(\function)
511 #endif
512
513 /*
514 * we've captured enough state to the stack (including in
515 * particular our EX_CONTEXT state) that we can now release
516 * the interrupt critical section and replace it with our
517 * standard "interrupts disabled" mask value. This allows
518 * synchronous interrupts (and profile interrupts) to punch
519 * through from this point onwards.
520 *
521 * If bit 31 of r3 is set during a non-NMI interrupt, we know we
522 * are on the path where the hypervisor has punched through our
523 * ICS with a page fault, so we call out to do_page_fault_ics()
524 * to figure out what to do with it. If the fault was in
525 * an atomic op, we unlock the atomic lock, adjust the
526 * saved register state a little, and return "zero" in r4,
527 * falling through into the normal page-fault interrupt code.
528 * If the fault was in a kernel-space atomic operation, then
529 * do_page_fault_ics() resolves it itself, returns "one" in r4,
530 * and as a result goes directly to restoring registers and iret,
531 * without trying to adjust the interrupt masks at all.
532 * The do_page_fault_ics() API involves passing and returning
533 * a five-word struct (in registers) to avoid writing the
534 * save and restore code here.
535 */
536 .ifc \function,handle_nmi
537 IRQ_DISABLE_ALL(r20)
538 .else
539 .ifnc \function,handle_syscall
540 bgezt r3, 1f
541 {
542 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
543 jal do_page_fault_ics
544 }
545 FEEDBACK_REENTER(\function)
546 bzt r4, 1f
547 j .Lrestore_regs
548 1:
549 .endif
550 IRQ_DISABLE(r20, r21)
551 .endif
552 mtspr INTERRUPT_CRITICAL_SECTION, zero
553
554 /*
555 * Prepare the first 256 stack bytes to be rapidly accessible
556 * without having to fetch the background data. We don't really
557 * know how far to write-hint, but kernel stacks generally
558 * aren't that big, and write-hinting here does take some time.
559 */
560 addi r52, sp, -64
561 {
562 wh64 r52
563 addi r52, r52, -64
564 }
565 {
566 wh64 r52
567 addi r52, r52, -64
568 }
569 {
570 wh64 r52
571 addi r52, r52, -64
572 }
573 wh64 r52
574
575 #if defined(CONFIG_TRACE_IRQFLAGS) || defined(CONFIG_CONTEXT_TRACKING)
576 .ifnc \function,handle_nmi
577 /*
578 * We finally have enough state set up to notify the irq
579 * tracing code that irqs were disabled on entry to the handler.
580 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
581 * For syscalls, we already have the register state saved away
582 * on the stack, so we don't bother to do any register saves here,
583 * and later we pop the registers back off the kernel stack.
584 * For interrupt handlers, save r0-r3 in callee-saved registers.
585 */
586 .ifnc \function,handle_syscall
587 { move r30, r0; move r31, r1 }
588 { move r32, r2; move r33, r3 }
589 .endif
590 TRACE_IRQS_OFF
591 #ifdef CONFIG_CONTEXT_TRACKING
592 jal context_tracking_user_exit
593 #endif
594 .ifnc \function,handle_syscall
595 { move r0, r30; move r1, r31 }
596 { move r2, r32; move r3, r33 }
597 .endif
598 .endif
599 #endif
600
601 .endm
602
603 .macro check_single_stepping, kind, not_single_stepping
604 /*
605 * Check for single stepping in user-level priv
606 * kind can be "normal", "ill", or "syscall"
607 * At end, if fall-thru
608 * r29: thread_info->step_state
609 * r28: &pt_regs->pc
610 * r27: pt_regs->pc
611 * r26: thread_info->step_state->buffer
612 */
613
614 /* Check for single stepping */
615 GET_THREAD_INFO(r29)
616 {
617 /* Get pointer to field holding step state */
618 addi r29, r29, THREAD_INFO_STEP_STATE_OFFSET
619
620 /* Get pointer to EX1 in register state */
621 PTREGS_PTR(r27, PTREGS_OFFSET_EX1)
622 }
623 {
624 /* Get pointer to field holding PC */
625 PTREGS_PTR(r28, PTREGS_OFFSET_PC)
626
627 /* Load the pointer to the step state */
628 lw r29, r29
629 }
630 /* Load EX1 */
631 lw r27, r27
632 {
633 /* Points to flags */
634 addi r23, r29, SINGLESTEP_STATE_FLAGS_OFFSET
635
636 /* No single stepping if there is no step state structure */
637 bzt r29, \not_single_stepping
638 }
639 {
640 /* mask off ICS and any other high bits */
641 andi r27, r27, SPR_EX_CONTEXT_1_1__PL_MASK
642
643 /* Load pointer to single step instruction buffer */
644 lw r26, r29
645 }
646 /* Check priv state */
647 bnz r27, \not_single_stepping
648
649 /* Get flags */
650 lw r22, r23
651 {
652 /* Branch if single-step mode not enabled */
653 bbnst r22, \not_single_stepping
654
655 /* Clear enabled flag */
656 andi r22, r22, ~SINGLESTEP_STATE_MASK_IS_ENABLED
657 }
658 .ifc \kind,normal
659 {
660 /* Load PC */
661 lw r27, r28
662
663 /* Point to the entry containing the original PC */
664 addi r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
665 }
666 {
667 /* Disable single stepping flag */
668 sw r23, r22
669 }
670 {
671 /* Get the original pc */
672 lw r24, r24
673
674 /* See if the PC is at the start of the single step buffer */
675 seq r25, r26, r27
676 }
677 /*
678 * NOTE: it is really expected that the PC be in the single step buffer
679 * at this point
680 */
681 bzt r25, \not_single_stepping
682
683 /* Restore the original PC */
684 sw r28, r24
685 .else
686 .ifc \kind,syscall
687 {
688 /* Load PC */
689 lw r27, r28
690
691 /* Point to the entry containing the next PC */
692 addi r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
693 }
694 {
695 /* Increment the stopped PC by the bundle size */
696 addi r26, r26, 8
697
698 /* Disable single stepping flag */
699 sw r23, r22
700 }
701 {
702 /* Get the next pc */
703 lw r24, r24
704
705 /*
706 * See if the PC is one bundle past the start of the
707 * single step buffer
708 */
709 seq r25, r26, r27
710 }
711 {
712 /*
713 * NOTE: it is really expected that the PC be in the
714 * single step buffer at this point
715 */
716 bzt r25, \not_single_stepping
717 }
718 /* Set to the next PC */
719 sw r28, r24
720 .else
721 {
722 /* Point to 3rd bundle in buffer */
723 addi r25, r26, 16
724
725 /* Load PC */
726 lw r27, r28
727 }
728 {
729 /* Disable single stepping flag */
730 sw r23, r22
731
732 /* See if the PC is in the single step buffer */
733 slte_u r24, r26, r27
734 }
735 {
736 slte_u r25, r27, r25
737
738 /*
739 * NOTE: it is really expected that the PC be in the
740 * single step buffer at this point
741 */
742 bzt r24, \not_single_stepping
743 }
744 bzt r25, \not_single_stepping
745 .endif
746 .endif
747 .endm
748
749 /*
750 * Redispatch a downcall.
751 */
752 .macro dc_dispatch vecnum, vecname
753 .org (\vecnum << 8)
754 intvec_\vecname:
755 j _hv_downcall_dispatch
756 ENDPROC(intvec_\vecname)
757 .endm
758
759 /*
760 * Common code for most interrupts. The C function we're eventually
761 * going to is in r0, and the faultnum is in r1; the original
762 * values for those registers are on the stack.
763 */
764 .pushsection .text.handle_interrupt,"ax"
765 handle_interrupt:
766 finish_interrupt_save handle_interrupt
767
768 /*
769 * Check for if we are single stepping in user level. If so, then
770 * we need to restore the PC.
771 */
772
773 check_single_stepping normal, .Ldispatch_interrupt
774 .Ldispatch_interrupt:
775
776 /* Jump to the C routine; it should enable irqs as soon as possible. */
777 {
778 jalr r0
779 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
780 }
781 FEEDBACK_REENTER(handle_interrupt)
782 {
783 movei r30, 0 /* not an NMI */
784 j interrupt_return
785 }
786 STD_ENDPROC(handle_interrupt)
787
788 /*
789 * This routine takes a boolean in r30 indicating if this is an NMI.
790 * If so, we also expect a boolean in r31 indicating whether to
791 * re-enable the oprofile interrupts.
792 *
793 * Note that .Lresume_userspace is jumped to directly in several
794 * places, and we need to make sure r30 is set correctly in those
795 * callers as well.
796 */
797 STD_ENTRY(interrupt_return)
798 /* If we're resuming to kernel space, don't check thread flags. */
799 {
800 bnz r30, .Lrestore_all /* NMIs don't special-case user-space */
801 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
802 }
803 lw r29, r29
804 andi r29, r29, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
805 bzt r29, .Lresume_userspace
806
807 #ifdef CONFIG_PREEMPT
808 /* Returning to kernel space. Check if we need preemption. */
809 GET_THREAD_INFO(r29)
810 addli r28, r29, THREAD_INFO_FLAGS_OFFSET
811 {
812 lw r28, r28
813 addli r29, r29, THREAD_INFO_PREEMPT_COUNT_OFFSET
814 }
815 {
816 andi r28, r28, _TIF_NEED_RESCHED
817 lw r29, r29
818 }
819 bzt r28, 1f
820 bnz r29, 1f
821 /* Disable interrupts explicitly for preemption. */
822 IRQ_DISABLE(r20,r21)
823 TRACE_IRQS_OFF
824 jal preempt_schedule_irq
825 FEEDBACK_REENTER(interrupt_return)
826 1:
827 #endif
828
829 /* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
830 {
831 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
832 moveli r27, lo16(_cpu_idle_nap)
833 }
834 {
835 lw r28, r29
836 auli r27, r27, ha16(_cpu_idle_nap)
837 }
838 {
839 seq r27, r27, r28
840 }
841 {
842 bbns r27, .Lrestore_all
843 addi r28, r28, 8
844 }
845 sw r29, r28
846 j .Lrestore_all
847
848 .Lresume_userspace:
849 FEEDBACK_REENTER(interrupt_return)
850
851 /*
852 * Disable interrupts so as to make sure we don't
853 * miss an interrupt that sets any of the thread flags (like
854 * need_resched or sigpending) between sampling and the iret.
855 * Routines like schedule() or do_signal() may re-enable
856 * interrupts before returning.
857 */
858 IRQ_DISABLE(r20, r21)
859 TRACE_IRQS_OFF /* Note: clobbers registers r0-r29 */
860
861 /*
862 * See if there are any work items (including single-shot items)
863 * to do. If so, save the callee-save registers to pt_regs
864 * and then dispatch to C code.
865 */
866 GET_THREAD_INFO(r21)
867 {
868 addi r22, r21, THREAD_INFO_FLAGS_OFFSET
869 moveli r20, lo16(_TIF_ALLWORK_MASK)
870 }
871 {
872 lw r22, r22
873 auli r20, r20, ha16(_TIF_ALLWORK_MASK)
874 }
875 and r1, r22, r20
876 {
877 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
878 bzt r1, .Lrestore_all
879 }
880 push_extra_callee_saves r0
881 jal prepare_exit_to_usermode
882
883 /*
884 * In the NMI case we
885 * omit the call to single_process_check_nohz, which normally checks
886 * to see if we should start or stop the scheduler tick, because
887 * we can't call arbitrary Linux code from an NMI context.
888 * We always call the homecache TLB deferral code to re-trigger
889 * the deferral mechanism.
890 *
891 * The other chunk of responsibility this code has is to reset the
892 * interrupt masks appropriately to reset irqs and NMIs. We have
893 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
894 * lockdep-type stuff, but we can't set ICS until afterwards, since
895 * ICS can only be used in very tight chunks of code to avoid
896 * tripping over various assertions that it is off.
897 *
898 * (There is what looks like a window of vulnerability here since
899 * we might take a profile interrupt between the two SPR writes
900 * that set the mask, but since we write the low SPR word first,
901 * and our interrupt entry code checks the low SPR word, any
902 * profile interrupt will actually disable interrupts in both SPRs
903 * before returning, which is OK.)
904 */
905 .Lrestore_all:
906 PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
907 {
908 lw r0, r0
909 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
910 }
911 {
912 andi r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK
913 lw r32, r32
914 }
915 bnz r0, 1f
916 j 2f
917 #if PT_FLAGS_DISABLE_IRQ != 1
918 # error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use bbnst below
919 #endif
920 1: bbnst r32, 2f
921 IRQ_DISABLE(r20,r21)
922 TRACE_IRQS_OFF
923 movei r0, 1
924 mtspr INTERRUPT_CRITICAL_SECTION, r0
925 bzt r30, .Lrestore_regs
926 j 3f
927 2: TRACE_IRQS_ON
928 movei r0, 1
929 mtspr INTERRUPT_CRITICAL_SECTION, r0
930 IRQ_ENABLE(r20, r21)
931 bzt r30, .Lrestore_regs
932 3:
933
934 /* We are relying on INT_PERF_COUNT at 33, and AUX_PERF_COUNT at 48 */
935 {
936 moveli r0, lo16(1 << (INT_PERF_COUNT - 32))
937 bz r31, .Lrestore_regs
938 }
939 auli r0, r0, ha16(1 << (INT_AUX_PERF_COUNT - 32))
940 mtspr SPR_INTERRUPT_MASK_RESET_K_1, r0
941
942 /*
943 * We now commit to returning from this interrupt, since we will be
944 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
945 * frame. No calls should be made to any other code after this point.
946 * This code should only be entered with ICS set.
947 * r32 must still be set to ptregs.flags.
948 * We launch loads to each cache line separately first, so we can
949 * get some parallelism out of the memory subsystem.
950 * We start zeroing caller-saved registers throughout, since
951 * that will save some cycles if this turns out to be a syscall.
952 */
953 .Lrestore_regs:
954 FEEDBACK_REENTER(interrupt_return) /* called from elsewhere */
955
956 /*
957 * Rotate so we have one high bit and one low bit to test.
958 * - low bit says whether to restore all the callee-saved registers,
959 * or just r30-r33, and r52 up.
960 * - high bit (i.e. sign bit) says whether to restore all the
961 * caller-saved registers, or just r0.
962 */
963 #if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
964 # error Rotate trick does not work :-)
965 #endif
966 {
967 rli r20, r32, 30
968 PTREGS_PTR(sp, PTREGS_OFFSET_REG(0))
969 }
970
971 /*
972 * Load cache lines 0, 2, and 3 in that order, then use
973 * the last loaded value, which makes it likely that the other
974 * cache lines have also loaded, at which point we should be
975 * able to safely read all the remaining words on those cache
976 * lines without waiting for the memory subsystem.
977 */
978 pop_reg_zero r0, r28, sp, PTREGS_OFFSET_REG(30) - PTREGS_OFFSET_REG(0)
979 pop_reg_zero r30, r2, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(30)
980 pop_reg_zero r21, r3, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
981 pop_reg_zero lr, r4, sp, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_EX1
982 {
983 mtspr SPR_EX_CONTEXT_K_0, r21
984 move r5, zero
985 }
986 {
987 mtspr SPR_EX_CONTEXT_K_1, lr
988 andi lr, lr, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
989 }
990
991 /* Restore callee-saveds that we actually use. */
992 pop_reg_zero r52, r6, sp, PTREGS_OFFSET_REG(31) - PTREGS_OFFSET_REG(52)
993 pop_reg_zero r31, r7
994 pop_reg_zero r32, r8
995 pop_reg_zero r33, r9, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(33)
996
997 /*
998 * If we modified other callee-saveds, restore them now.
999 * This is rare, but could be via ptrace or signal handler.
1000 */
1001 {
1002 move r10, zero
1003 bbs r20, .Lrestore_callees
1004 }
1005 .Lcontinue_restore_regs:
1006
1007 /* Check if we're returning from a syscall. */
1008 {
1009 move r11, zero
1010 blzt r20, 1f /* no, so go restore callee-save registers */
1011 }
1012
1013 /*
1014 * Check if we're returning to userspace.
1015 * Note that if we're not, we don't worry about zeroing everything.
1016 */
1017 {
1018 addli sp, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(29)
1019 bnz lr, .Lkernel_return
1020 }
1021
1022 /*
1023 * On return from syscall, we've restored r0 from pt_regs, but we
1024 * clear the remainder of the caller-saved registers. We could
1025 * restore the syscall arguments, but there's not much point,
1026 * and it ensures user programs aren't trying to use the
1027 * caller-saves if we clear them, as well as avoiding leaking
1028 * kernel pointers into userspace.
1029 */
1030 pop_reg_zero lr, r12, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1031 pop_reg_zero tp, r13, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1032 {
1033 lw sp, sp
1034 move r14, zero
1035 move r15, zero
1036 }
1037 { move r16, zero; move r17, zero }
1038 { move r18, zero; move r19, zero }
1039 { move r20, zero; move r21, zero }
1040 { move r22, zero; move r23, zero }
1041 { move r24, zero; move r25, zero }
1042 { move r26, zero; move r27, zero }
1043
1044 /* Set r1 to errno if we are returning an error, otherwise zero. */
1045 {
1046 moveli r29, 4096
1047 sub r1, zero, r0
1048 }
1049 slt_u r29, r1, r29
1050 {
1051 mnz r1, r29, r1
1052 move r29, zero
1053 }
1054 iret
1055
1056 /*
1057 * Not a syscall, so restore caller-saved registers.
1058 * First kick off a load for cache line 1, which we're touching
1059 * for the first time here.
1060 */
1061 .align 64
1062 1: pop_reg r29, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(29)
1063 pop_reg r1
1064 pop_reg r2
1065 pop_reg r3
1066 pop_reg r4
1067 pop_reg r5
1068 pop_reg r6
1069 pop_reg r7
1070 pop_reg r8
1071 pop_reg r9
1072 pop_reg r10
1073 pop_reg r11
1074 pop_reg r12
1075 pop_reg r13
1076 pop_reg r14
1077 pop_reg r15
1078 pop_reg r16
1079 pop_reg r17
1080 pop_reg r18
1081 pop_reg r19
1082 pop_reg r20
1083 pop_reg r21
1084 pop_reg r22
1085 pop_reg r23
1086 pop_reg r24
1087 pop_reg r25
1088 pop_reg r26
1089 pop_reg r27
1090 pop_reg r28, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(28)
1091 /* r29 already restored above */
1092 bnz lr, .Lkernel_return
1093 pop_reg lr, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1094 pop_reg tp, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1095 lw sp, sp
1096 iret
1097
1098 /*
1099 * We can't restore tp when in kernel mode, since a thread might
1100 * have migrated from another cpu and brought a stale tp value.
1101 */
1102 .Lkernel_return:
1103 pop_reg lr, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
1104 lw sp, sp
1105 iret
1106
1107 /* Restore callee-saved registers from r34 to r51. */
1108 .Lrestore_callees:
1109 addli sp, sp, PTREGS_OFFSET_REG(34) - PTREGS_OFFSET_REG(29)
1110 pop_reg r34
1111 pop_reg r35
1112 pop_reg r36
1113 pop_reg r37
1114 pop_reg r38
1115 pop_reg r39
1116 pop_reg r40
1117 pop_reg r41
1118 pop_reg r42
1119 pop_reg r43
1120 pop_reg r44
1121 pop_reg r45
1122 pop_reg r46
1123 pop_reg r47
1124 pop_reg r48
1125 pop_reg r49
1126 pop_reg r50
1127 pop_reg r51, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(51)
1128 j .Lcontinue_restore_regs
1129 STD_ENDPROC(interrupt_return)
1130
1131 /*
1132 * Some interrupts don't check for single stepping
1133 */
1134 .pushsection .text.handle_interrupt_no_single_step,"ax"
1135 handle_interrupt_no_single_step:
1136 finish_interrupt_save handle_interrupt_no_single_step
1137 {
1138 jalr r0
1139 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1140 }
1141 FEEDBACK_REENTER(handle_interrupt_no_single_step)
1142 {
1143 movei r30, 0 /* not an NMI */
1144 j interrupt_return
1145 }
1146 STD_ENDPROC(handle_interrupt_no_single_step)
1147
1148 /*
1149 * "NMI" interrupts mask ALL interrupts before calling the
1150 * handler, and don't check thread flags, etc., on the way
1151 * back out. In general, the only things we do here for NMIs
1152 * are the register save/restore, fixing the PC if we were
1153 * doing single step, and the dataplane kernel-TLB management.
1154 * We don't (for example) deal with start/stop of the sched tick.
1155 */
1156 .pushsection .text.handle_nmi,"ax"
1157 handle_nmi:
1158 finish_interrupt_save handle_nmi
1159 check_single_stepping normal, .Ldispatch_nmi
1160 .Ldispatch_nmi:
1161 {
1162 jalr r0
1163 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1164 }
1165 FEEDBACK_REENTER(handle_nmi)
1166 {
1167 movei r30, 1
1168 seq r31, r0, zero
1169 }
1170 j interrupt_return
1171 STD_ENDPROC(handle_nmi)
1172
1173 /*
1174 * Parallel code for syscalls to handle_interrupt.
1175 */
1176 .pushsection .text.handle_syscall,"ax"
1177 handle_syscall:
1178 finish_interrupt_save handle_syscall
1179
1180 /*
1181 * Check for if we are single stepping in user level. If so, then
1182 * we need to restore the PC.
1183 */
1184 check_single_stepping syscall, .Ldispatch_syscall
1185 .Ldispatch_syscall:
1186
1187 /* Enable irqs. */
1188 TRACE_IRQS_ON
1189 IRQ_ENABLE(r20, r21)
1190
1191 /* Bump the counter for syscalls made on this tile. */
1192 moveli r20, lo16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1193 auli r20, r20, ha16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1194 add r20, r20, tp
1195 lw r21, r20
1196 addi r21, r21, 1
1197 {
1198 sw r20, r21
1199 GET_THREAD_INFO(r31)
1200 }
1201
1202 /* Trace syscalls, if requested. */
1203 addi r31, r31, THREAD_INFO_FLAGS_OFFSET
1204 lw r30, r31
1205 andi r30, r30, _TIF_SYSCALL_TRACE
1206 bzt r30, .Lrestore_syscall_regs
1207 {
1208 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1209 jal do_syscall_trace_enter
1210 }
1211 FEEDBACK_REENTER(handle_syscall)
1212 blz r0, .Lsyscall_sigreturn_skip
1213
1214 /*
1215 * We always reload our registers from the stack at this
1216 * point. They might be valid, if we didn't build with
1217 * TRACE_IRQFLAGS, and this isn't a dataplane tile, and we're not
1218 * doing syscall tracing, but there are enough cases now that it
1219 * seems simplest just to do the reload unconditionally.
1220 */
1221 .Lrestore_syscall_regs:
1222 PTREGS_PTR(r11, PTREGS_OFFSET_REG(0))
1223 pop_reg r0, r11
1224 pop_reg r1, r11
1225 pop_reg r2, r11
1226 pop_reg r3, r11
1227 pop_reg r4, r11
1228 pop_reg r5, r11, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(5)
1229 pop_reg TREG_SYSCALL_NR_NAME, r11
1230
1231 /* Ensure that the syscall number is within the legal range. */
1232 moveli r21, __NR_syscalls
1233 {
1234 slt_u r21, TREG_SYSCALL_NR_NAME, r21
1235 moveli r20, lo16(sys_call_table)
1236 }
1237 {
1238 bbns r21, .Linvalid_syscall
1239 auli r20, r20, ha16(sys_call_table)
1240 }
1241 s2a r20, TREG_SYSCALL_NR_NAME, r20
1242 lw r20, r20
1243
1244 /* Jump to syscall handler. */
1245 jalr r20
1246 .Lhandle_syscall_link: /* value of "lr" after "jalr r20" above */
1247
1248 /*
1249 * Write our r0 onto the stack so it gets restored instead
1250 * of whatever the user had there before.
1251 */
1252 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1253 sw r29, r0
1254
1255 .Lsyscall_sigreturn_skip:
1256 FEEDBACK_REENTER(handle_syscall)
1257
1258 /* Do syscall trace again, if requested. */
1259 lw r30, r31
1260 andi r30, r30, _TIF_SYSCALL_TRACE
1261 bzt r30, 1f
1262 {
1263 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1264 jal do_syscall_trace_exit
1265 }
1266 FEEDBACK_REENTER(handle_syscall)
1267 1: {
1268 movei r30, 0 /* not an NMI */
1269 j .Lresume_userspace /* jump into middle of interrupt_return */
1270 }
1271
1272 .Linvalid_syscall:
1273 /* Report an invalid syscall back to the user program */
1274 {
1275 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1276 movei r28, -ENOSYS
1277 }
1278 sw r29, r28
1279 {
1280 movei r30, 0 /* not an NMI */
1281 j .Lresume_userspace /* jump into middle of interrupt_return */
1282 }
1283 STD_ENDPROC(handle_syscall)
1284
1285 /* Return the address for oprofile to suppress in backtraces. */
1286 STD_ENTRY_SECTION(handle_syscall_link_address, .text.handle_syscall)
1287 lnk r0
1288 {
1289 addli r0, r0, .Lhandle_syscall_link - .
1290 jrp lr
1291 }
1292 STD_ENDPROC(handle_syscall_link_address)
1293
1294 STD_ENTRY(ret_from_fork)
1295 jal sim_notify_fork
1296 jal schedule_tail
1297 FEEDBACK_REENTER(ret_from_fork)
1298 {
1299 movei r30, 0 /* not an NMI */
1300 j .Lresume_userspace /* jump into middle of interrupt_return */
1301 }
1302 STD_ENDPROC(ret_from_fork)
1303
1304 STD_ENTRY(ret_from_kernel_thread)
1305 jal sim_notify_fork
1306 jal schedule_tail
1307 FEEDBACK_REENTER(ret_from_fork)
1308 {
1309 move r0, r31
1310 jalr r30
1311 }
1312 FEEDBACK_REENTER(ret_from_kernel_thread)
1313 {
1314 movei r30, 0 /* not an NMI */
1315 j interrupt_return
1316 }
1317 STD_ENDPROC(ret_from_kernel_thread)
1318
1319 /*
1320 * Code for ill interrupt.
1321 */
1322 .pushsection .text.handle_ill,"ax"
1323 handle_ill:
1324 finish_interrupt_save handle_ill
1325
1326 /*
1327 * Check for if we are single stepping in user level. If so, then
1328 * we need to restore the PC.
1329 */
1330 check_single_stepping ill, .Ldispatch_normal_ill
1331
1332 {
1333 /* See if the PC is the 1st bundle in the buffer */
1334 seq r25, r27, r26
1335
1336 /* Point to the 2nd bundle in the buffer */
1337 addi r26, r26, 8
1338 }
1339 {
1340 /* Point to the original pc */
1341 addi r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
1342
1343 /* Branch if the PC is the 1st bundle in the buffer */
1344 bnz r25, 3f
1345 }
1346 {
1347 /* See if the PC is the 2nd bundle of the buffer */
1348 seq r25, r27, r26
1349
1350 /* Set PC to next instruction */
1351 addi r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
1352 }
1353 {
1354 /* Point to flags */
1355 addi r25, r29, SINGLESTEP_STATE_FLAGS_OFFSET
1356
1357 /* Branch if PC is in the second bundle */
1358 bz r25, 2f
1359 }
1360 /* Load flags */
1361 lw r25, r25
1362 {
1363 /*
1364 * Get the offset for the register to restore
1365 * Note: the lower bound is 2, so we have implicit scaling by 4.
1366 * No multiplication of the register number by the size of a register
1367 * is needed.
1368 */
1369 mm r27, r25, zero, SINGLESTEP_STATE_TARGET_LB, \
1370 SINGLESTEP_STATE_TARGET_UB
1371
1372 /* Mask Rewrite_LR */
1373 andi r25, r25, SINGLESTEP_STATE_MASK_UPDATE
1374 }
1375 {
1376 addi r29, r29, SINGLESTEP_STATE_UPDATE_VALUE_OFFSET
1377
1378 /* Don't rewrite temp register */
1379 bz r25, 3f
1380 }
1381 {
1382 /* Get the temp value */
1383 lw r29, r29
1384
1385 /* Point to where the register is stored */
1386 add r27, r27, sp
1387 }
1388
1389 /* Add in the C ABI save area size to the register offset */
1390 addi r27, r27, C_ABI_SAVE_AREA_SIZE
1391
1392 /* Restore the user's register with the temp value */
1393 sw r27, r29
1394 j 3f
1395
1396 2:
1397 /* Must be in the third bundle */
1398 addi r24, r29, SINGLESTEP_STATE_BRANCH_NEXT_PC_OFFSET
1399
1400 3:
1401 /* set PC and continue */
1402 lw r26, r24
1403 {
1404 sw r28, r26
1405 GET_THREAD_INFO(r0)
1406 }
1407
1408 /*
1409 * Clear TIF_SINGLESTEP to prevent recursion if we execute an ill.
1410 * The normal non-arch flow redundantly clears TIF_SINGLESTEP, but we
1411 * need to clear it here and can't really impose on all other arches.
1412 * So what's another write between friends?
1413 */
1414
1415 addi r1, r0, THREAD_INFO_FLAGS_OFFSET
1416 {
1417 lw r2, r1
1418 addi r0, r0, THREAD_INFO_TASK_OFFSET /* currently a no-op */
1419 }
1420 andi r2, r2, ~_TIF_SINGLESTEP
1421 sw r1, r2
1422
1423 /* Issue a sigtrap */
1424 {
1425 lw r0, r0 /* indirect thru thread_info to get task_info*/
1426 addi r1, sp, C_ABI_SAVE_AREA_SIZE /* put ptregs pointer into r1 */
1427 }
1428
1429 jal send_sigtrap /* issue a SIGTRAP */
1430 FEEDBACK_REENTER(handle_ill)
1431 {
1432 movei r30, 0 /* not an NMI */
1433 j .Lresume_userspace /* jump into middle of interrupt_return */
1434 }
1435
1436 .Ldispatch_normal_ill:
1437 {
1438 jalr r0
1439 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1440 }
1441 FEEDBACK_REENTER(handle_ill)
1442 {
1443 movei r30, 0 /* not an NMI */
1444 j interrupt_return
1445 }
1446 STD_ENDPROC(handle_ill)
1447
1448 /* Various stub interrupt handlers and syscall handlers */
1449
1450 STD_ENTRY_LOCAL(_kernel_double_fault)
1451 mfspr r1, SPR_EX_CONTEXT_K_0
1452 move r2, lr
1453 move r3, sp
1454 move r4, r52
1455 addi sp, sp, -C_ABI_SAVE_AREA_SIZE
1456 j kernel_double_fault
1457 STD_ENDPROC(_kernel_double_fault)
1458
1459 STD_ENTRY_LOCAL(bad_intr)
1460 mfspr r2, SPR_EX_CONTEXT_K_0
1461 panic "Unhandled interrupt %#x: PC %#lx"
1462 STD_ENDPROC(bad_intr)
1463
1464 /*
1465 * Special-case sigreturn to not write r0 to the stack on return.
1466 * This is technically more efficient, but it also avoids difficulties
1467 * in the 64-bit OS when handling 32-bit compat code, since we must not
1468 * sign-extend r0 for the sigreturn return-value case.
1469 */
1470 #define PTREGS_SYSCALL_SIGRETURN(x, reg) \
1471 STD_ENTRY(_##x); \
1472 addli lr, lr, .Lsyscall_sigreturn_skip - .Lhandle_syscall_link; \
1473 { \
1474 PTREGS_PTR(reg, PTREGS_OFFSET_BASE); \
1475 j x \
1476 }; \
1477 STD_ENDPROC(_##x)
1478
1479 PTREGS_SYSCALL_SIGRETURN(sys_rt_sigreturn, r0)
1480
1481 /* Save additional callee-saves to pt_regs and jump to standard function. */
1482 STD_ENTRY(_sys_clone)
1483 push_extra_callee_saves r4
1484 j sys_clone
1485 STD_ENDPROC(_sys_clone)
1486
1487 /*
1488 * This entrypoint is taken for the cmpxchg and atomic_update fast
1489 * swints. We may wish to generalize it to other fast swints at some
1490 * point, but for now there are just two very similar ones, which
1491 * makes it faster.
1492 *
1493 * The fast swint code is designed to have a small footprint. It does
1494 * not save or restore any GPRs, counting on the caller-save registers
1495 * to be available to it on entry. It does not modify any callee-save
1496 * registers (including "lr"). It does not check what PL it is being
1497 * called at, so you'd better not call it other than at PL0.
1498 * The <atomic.h> wrapper assumes it only clobbers r20-r29, so if
1499 * it ever is necessary to use more registers, be aware.
1500 *
1501 * It does not use the stack, but since it might be re-interrupted by
1502 * a page fault which would assume the stack was valid, it does
1503 * save/restore the stack pointer and zero it out to make sure it gets reset.
1504 * Since we always keep interrupts disabled, the hypervisor won't
1505 * clobber our EX_CONTEXT_K_x registers, so we don't save/restore them
1506 * (other than to advance the PC on return).
1507 *
1508 * We have to manually validate the user vs kernel address range
1509 * (since at PL1 we can read/write both), and for performance reasons
1510 * we don't allow cmpxchg on the fc000000 memory region, since we only
1511 * validate that the user address is below PAGE_OFFSET.
1512 *
1513 * We place it in the __HEAD section to ensure it is relatively
1514 * near to the intvec_SWINT_1 code (reachable by a conditional branch).
1515 *
1516 * Our use of ATOMIC_LOCK_REG here must match do_page_fault_ics().
1517 *
1518 * As we do in lib/atomic_asm_32.S, we bypass a store if the value we
1519 * would store is the same as the value we just loaded.
1520 */
1521 __HEAD
1522 .align 64
1523 /* Align much later jump on the start of a cache line. */
1524 nop
1525 #if PAGE_SIZE >= 0x10000
1526 nop
1527 #endif
1528 ENTRY(sys_cmpxchg)
1529
1530 /*
1531 * Save "sp" and set it zero for any possible page fault.
1532 *
1533 * HACK: We want to both zero sp and check r0's alignment,
1534 * so we do both at once. If "sp" becomes nonzero we
1535 * know r0 is unaligned and branch to the error handler that
1536 * restores sp, so this is OK.
1537 *
1538 * ICS is disabled right now so having a garbage but nonzero
1539 * sp is OK, since we won't execute any faulting instructions
1540 * when it is nonzero.
1541 */
1542 {
1543 move r27, sp
1544 andi sp, r0, 3
1545 }
1546
1547 /*
1548 * Get the lock address in ATOMIC_LOCK_REG, and also validate that the
1549 * address is less than PAGE_OFFSET, since that won't trap at PL1.
1550 * We only use bits less than PAGE_SHIFT to avoid having to worry
1551 * about aliasing among multiple mappings of the same physical page,
1552 * and we ignore the low 3 bits so we have one lock that covers
1553 * both a cmpxchg64() and a cmpxchg() on either its low or high word.
1554 * NOTE: this must match __atomic_hashed_lock() in lib/atomic_32.c.
1555 */
1556
1557 #if (PAGE_OFFSET & 0xffff) != 0
1558 # error Code here assumes PAGE_OFFSET can be loaded with just hi16()
1559 #endif
1560
1561 {
1562 /* Check for unaligned input. */
1563 bnz sp, .Lcmpxchg_badaddr
1564 auli r23, zero, hi16(PAGE_OFFSET) /* hugepage-aligned */
1565 }
1566 {
1567 /*
1568 * Slide bits into position for 'mm'. We want to ignore
1569 * the low 3 bits of r0, and consider only the next
1570 * ATOMIC_HASH_SHIFT bits.
1571 * Because of C pointer arithmetic, we want to compute this:
1572 *
1573 * ((char*)atomic_locks +
1574 * (((r0 >> 3) & ((1 << ATOMIC_HASH_SHIFT) - 1)) << 2))
1575 *
1576 * Instead of two shifts we just ">> 1", and use 'mm'
1577 * to ignore the low and high bits we don't want.
1578 */
1579 shri r25, r0, 1
1580
1581 slt_u r23, r0, r23
1582
1583 /*
1584 * Ensure that the TLB is loaded before we take out the lock.
1585 * This will start fetching the value all the way into our L1
1586 * as well (and if it gets modified before we grab the lock,
1587 * it will be invalidated from our cache before we reload it).
1588 */
1589 lw r26, r0
1590 }
1591 {
1592 auli r21, zero, ha16(atomic_locks)
1593
1594 bbns r23, .Lcmpxchg_badaddr
1595 }
1596 #if PAGE_SIZE < 0x10000
1597 /* atomic_locks is page-aligned so for big pages we don't need this. */
1598 addli r21, r21, lo16(atomic_locks)
1599 #endif
1600 {
1601 /*
1602 * Insert the hash bits into the page-aligned pointer.
1603 * ATOMIC_HASH_SHIFT is so big that we don't actually hash
1604 * the unmasked address bits, as that may cause unnecessary
1605 * collisions.
1606 */
1607 mm ATOMIC_LOCK_REG_NAME, r25, r21, 2, (ATOMIC_HASH_SHIFT + 2) - 1
1608
1609 seqi r23, TREG_SYSCALL_NR_NAME, __NR_FAST_cmpxchg64
1610 }
1611 {
1612 /* Branch away at this point if we're doing a 64-bit cmpxchg. */
1613 bbs r23, .Lcmpxchg64
1614 andi r23, r0, 7 /* Precompute alignment for cmpxchg64. */
1615 }
1616 {
1617 /*
1618 * We very carefully align the code that actually runs with
1619 * the lock held (twelve bundles) so that we know it is all in
1620 * the icache when we start. This instruction (the jump) is
1621 * at the start of the first cache line, address zero mod 64;
1622 * we jump to the very end of the second cache line to get that
1623 * line loaded in the icache, then fall through to issue the tns
1624 * in the third cache line, at which point it's all cached.
1625 * Note that is for performance, not correctness.
1626 */
1627 j .Lcmpxchg32_tns
1628 }
1629
1630 /* Symbol for do_page_fault_ics() to use to compare against the PC. */
1631 .global __sys_cmpxchg_grab_lock
1632 __sys_cmpxchg_grab_lock:
1633
1634 /*
1635 * Perform the actual cmpxchg or atomic_update.
1636 */
1637 .Ldo_cmpxchg32:
1638 {
1639 lw r21, r0
1640 seqi r23, TREG_SYSCALL_NR_NAME, __NR_FAST_atomic_update
1641 move r24, r2
1642 }
1643 {
1644 seq r22, r21, r1 /* See if cmpxchg matches. */
1645 and r25, r21, r1 /* If atomic_update, compute (*mem & mask) */
1646 }
1647 {
1648 or r22, r22, r23 /* Skip compare branch for atomic_update. */
1649 add r25, r25, r2 /* Compute (*mem & mask) + addend. */
1650 }
1651 {
1652 mvnz r24, r23, r25 /* Use atomic_update value if appropriate. */
1653 bbns r22, .Lcmpxchg32_nostore
1654 }
1655 seq r22, r24, r21 /* Are we storing the value we loaded? */
1656 bbs r22, .Lcmpxchg32_nostore
1657 sw r0, r24
1658
1659 /* The following instruction is the start of the second cache line. */
1660 /* Do slow mtspr here so the following "mf" waits less. */
1661 {
1662 move sp, r27
1663 mtspr SPR_EX_CONTEXT_K_0, r28
1664 }
1665 mf
1666
1667 {
1668 move r0, r21
1669 sw ATOMIC_LOCK_REG_NAME, zero
1670 }
1671 iret
1672
1673 /* Duplicated code here in the case where we don't overlap "mf" */
1674 .Lcmpxchg32_nostore:
1675 {
1676 move r0, r21
1677 sw ATOMIC_LOCK_REG_NAME, zero
1678 }
1679 {
1680 move sp, r27
1681 mtspr SPR_EX_CONTEXT_K_0, r28
1682 }
1683 iret
1684
1685 /*
1686 * The locking code is the same for 32-bit cmpxchg/atomic_update,
1687 * and for 64-bit cmpxchg. We provide it as a macro and put
1688 * it into both versions. We can't share the code literally
1689 * since it depends on having the right branch-back address.
1690 */
1691 .macro cmpxchg_lock, bitwidth
1692
1693 /* Lock; if we succeed, jump back up to the read-modify-write. */
1694 #ifdef CONFIG_SMP
1695 tns r21, ATOMIC_LOCK_REG_NAME
1696 #else
1697 /*
1698 * Non-SMP preserves all the lock infrastructure, to keep the
1699 * code simpler for the interesting (SMP) case. However, we do
1700 * one small optimization here and in atomic_asm.S, which is
1701 * to fake out acquiring the actual lock in the atomic_lock table.
1702 */
1703 movei r21, 0
1704 #endif
1705
1706 /* Issue the slow SPR here while the tns result is in flight. */
1707 mfspr r28, SPR_EX_CONTEXT_K_0
1708
1709 {
1710 addi r28, r28, 8 /* return to the instruction after the swint1 */
1711 bzt r21, .Ldo_cmpxchg\bitwidth
1712 }
1713 /*
1714 * The preceding instruction is the last thing that must be
1715 * hot in the icache before we do the "tns" above.
1716 */
1717
1718 #ifdef CONFIG_SMP
1719 /*
1720 * We failed to acquire the tns lock on our first try. Now use
1721 * bounded exponential backoff to retry, like __atomic_spinlock().
1722 */
1723 {
1724 moveli r23, 2048 /* maximum backoff time in cycles */
1725 moveli r25, 32 /* starting backoff time in cycles */
1726 }
1727 1: mfspr r26, CYCLE_LOW /* get start point for this backoff */
1728 2: mfspr r22, CYCLE_LOW /* test to see if we've backed off enough */
1729 sub r22, r22, r26
1730 slt r22, r22, r25
1731 bbst r22, 2b
1732 {
1733 shli r25, r25, 1 /* double the backoff; retry the tns */
1734 tns r21, ATOMIC_LOCK_REG_NAME
1735 }
1736 slt r26, r23, r25 /* is the proposed backoff too big? */
1737 {
1738 mvnz r25, r26, r23
1739 bzt r21, .Ldo_cmpxchg\bitwidth
1740 }
1741 j 1b
1742 #endif /* CONFIG_SMP */
1743 .endm
1744
1745 .Lcmpxchg32_tns:
1746 /*
1747 * This is the last instruction on the second cache line.
1748 * The nop here loads the second line, then we fall through
1749 * to the tns to load the third line before we take the lock.
1750 */
1751 nop
1752 cmpxchg_lock 32
1753
1754 /*
1755 * This code is invoked from sys_cmpxchg after most of the
1756 * preconditions have been checked. We still need to check
1757 * that r0 is 8-byte aligned, since if it's not we won't
1758 * actually be atomic. However, ATOMIC_LOCK_REG has the atomic
1759 * lock pointer and r27/r28 have the saved SP/PC.
1760 * r23 is holding "r0 & 7" so we can test for alignment.
1761 * The compare value is in r2/r3; the new value is in r4/r5.
1762 * On return, we must put the old value in r0/r1.
1763 */
1764 .align 64
1765 .Lcmpxchg64:
1766 {
1767 bzt r23, .Lcmpxchg64_tns
1768 }
1769 j .Lcmpxchg_badaddr
1770
1771 .Ldo_cmpxchg64:
1772 {
1773 lw r21, r0
1774 addi r25, r0, 4
1775 }
1776 {
1777 lw r1, r25
1778 }
1779 seq r26, r21, r2
1780 {
1781 bz r26, .Lcmpxchg64_mismatch
1782 seq r26, r1, r3
1783 }
1784 {
1785 bz r26, .Lcmpxchg64_mismatch
1786 }
1787 sw r0, r4
1788 sw r25, r5
1789
1790 /*
1791 * The 32-bit path provides optimized "match" and "mismatch"
1792 * iret paths, but we don't have enough bundles in this cache line
1793 * to do that, so we just make even the "mismatch" path do an "mf".
1794 */
1795 .Lcmpxchg64_mismatch:
1796 {
1797 move sp, r27
1798 mtspr SPR_EX_CONTEXT_K_0, r28
1799 }
1800 mf
1801 {
1802 move r0, r21
1803 sw ATOMIC_LOCK_REG_NAME, zero
1804 }
1805 iret
1806
1807 .Lcmpxchg64_tns:
1808 cmpxchg_lock 64
1809
1810
1811 /*
1812 * Reset sp and revector to sys_cmpxchg_badaddr(), which will
1813 * just raise the appropriate signal and exit. Doing it this
1814 * way means we don't have to duplicate the code in intvec.S's
1815 * int_hand macro that locates the top of the stack.
1816 */
1817 .Lcmpxchg_badaddr:
1818 {
1819 moveli TREG_SYSCALL_NR_NAME, __NR_cmpxchg_badaddr
1820 move sp, r27
1821 }
1822 j intvec_SWINT_1
1823 ENDPROC(sys_cmpxchg)
1824 ENTRY(__sys_cmpxchg_end)
1825
1826
1827 /* The single-step support may need to read all the registers. */
1828 int_unalign:
1829 push_extra_callee_saves r0
1830 j do_trap
1831
1832 /* Include .intrpt array of interrupt vectors */
1833 .section ".intrpt", "ax"
1834
1835 #ifndef CONFIG_USE_PMC
1836 #define handle_perf_interrupt bad_intr
1837 #endif
1838
1839 #ifndef CONFIG_HARDWALL
1840 #define do_hardwall_trap bad_intr
1841 #endif
1842
1843 int_hand INT_ITLB_MISS, ITLB_MISS, \
1844 do_page_fault, handle_interrupt_no_single_step
1845 int_hand INT_MEM_ERROR, MEM_ERROR, bad_intr
1846 int_hand INT_ILL, ILL, do_trap, handle_ill
1847 int_hand INT_GPV, GPV, do_trap
1848 int_hand INT_SN_ACCESS, SN_ACCESS, do_trap
1849 int_hand INT_IDN_ACCESS, IDN_ACCESS, do_trap
1850 int_hand INT_UDN_ACCESS, UDN_ACCESS, do_trap
1851 int_hand INT_IDN_REFILL, IDN_REFILL, bad_intr
1852 int_hand INT_UDN_REFILL, UDN_REFILL, bad_intr
1853 int_hand INT_IDN_COMPLETE, IDN_COMPLETE, bad_intr
1854 int_hand INT_UDN_COMPLETE, UDN_COMPLETE, bad_intr
1855 int_hand INT_SWINT_3, SWINT_3, do_trap
1856 int_hand INT_SWINT_2, SWINT_2, do_trap
1857 int_hand INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall
1858 int_hand INT_SWINT_0, SWINT_0, do_trap
1859 int_hand INT_UNALIGN_DATA, UNALIGN_DATA, int_unalign
1860 int_hand INT_DTLB_MISS, DTLB_MISS, do_page_fault
1861 int_hand INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault
1862 int_hand INT_DMATLB_MISS, DMATLB_MISS, do_page_fault
1863 int_hand INT_DMATLB_ACCESS, DMATLB_ACCESS, do_page_fault
1864 int_hand INT_SNITLB_MISS, SNITLB_MISS, do_page_fault
1865 int_hand INT_SN_NOTIFY, SN_NOTIFY, bad_intr
1866 int_hand INT_SN_FIREWALL, SN_FIREWALL, do_hardwall_trap
1867 int_hand INT_IDN_FIREWALL, IDN_FIREWALL, bad_intr
1868 int_hand INT_UDN_FIREWALL, UDN_FIREWALL, do_hardwall_trap
1869 int_hand INT_TILE_TIMER, TILE_TIMER, do_timer_interrupt
1870 int_hand INT_IDN_TIMER, IDN_TIMER, bad_intr
1871 int_hand INT_UDN_TIMER, UDN_TIMER, bad_intr
1872 int_hand INT_DMA_NOTIFY, DMA_NOTIFY, bad_intr
1873 int_hand INT_IDN_CA, IDN_CA, bad_intr
1874 int_hand INT_UDN_CA, UDN_CA, bad_intr
1875 int_hand INT_IDN_AVAIL, IDN_AVAIL, bad_intr
1876 int_hand INT_UDN_AVAIL, UDN_AVAIL, bad_intr
1877 int_hand INT_PERF_COUNT, PERF_COUNT, \
1878 handle_perf_interrupt, handle_nmi
1879 int_hand INT_INTCTRL_3, INTCTRL_3, bad_intr
1880 #if CONFIG_KERNEL_PL == 2
1881 dc_dispatch INT_INTCTRL_2, INTCTRL_2
1882 int_hand INT_INTCTRL_1, INTCTRL_1, bad_intr
1883 #else
1884 int_hand INT_INTCTRL_2, INTCTRL_2, bad_intr
1885 dc_dispatch INT_INTCTRL_1, INTCTRL_1
1886 #endif
1887 int_hand INT_INTCTRL_0, INTCTRL_0, bad_intr
1888 int_hand INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
1889 hv_message_intr
1890 int_hand INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, \
1891 tile_dev_intr
1892 int_hand INT_I_ASID, I_ASID, bad_intr
1893 int_hand INT_D_ASID, D_ASID, bad_intr
1894 int_hand INT_DMATLB_MISS_DWNCL, DMATLB_MISS_DWNCL, \
1895 do_page_fault
1896 int_hand INT_SNITLB_MISS_DWNCL, SNITLB_MISS_DWNCL, \
1897 do_page_fault
1898 int_hand INT_DMATLB_ACCESS_DWNCL, DMATLB_ACCESS_DWNCL, \
1899 do_page_fault
1900 int_hand INT_SN_CPL, SN_CPL, bad_intr
1901 int_hand INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
1902 int_hand INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \
1903 handle_perf_interrupt, handle_nmi
1904
1905 /* Synthetic interrupt delivered only by the simulator */
1906 int_hand INT_BREAKPOINT, BREAKPOINT, do_breakpoint