2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
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.
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
14 * Linux interrupt vectors.
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.h>
25 #include <asm/asm-offsets.h>
26 #include <hv/hypervisor.h>
28 #include <arch/interrupts.h>
29 #include <arch/spr_def.h>
32 # error "No support for kernel preemption currently"
35 #if INT_INTCTRL_K < 32 || INT_INTCTRL_K >= 48
36 # error INT_INTCTRL_K coded to set high interrupt mask
39 #define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
41 #define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
44 /* By making this an empty macro, we can use wh64 in the code. */
49 .macro push_reg reg, ptr=sp, delta=-4
52 addli \ptr, \ptr, \delta
56 .macro pop_reg reg, ptr=sp, delta=4
59 addli \ptr, \ptr, \delta
63 .macro pop_reg_zero reg, zreg, ptr=sp, delta=4
67 addi \ptr, \ptr, \delta
71 .macro push_extra_callee_saves reg
72 PTREGS_PTR(\reg, PTREGS_OFFSET_REG(51))
90 push_reg r34, \reg, PTREGS_OFFSET_BASE - PTREGS_OFFSET_REG(34)
94 .pushsection .rodata, "a"
102 auli r0, r0, ha16(1b)
107 #ifdef __COLLECT_LINKER_FEEDBACK__
108 .pushsection .text.intvec_feedback,"ax"
114 * Default interrupt handler.
116 * vecnum is where we'll put this code.
117 * c_routine is the C routine we'll call.
119 * The C routine is passed two arguments:
120 * - A pointer to the pt_regs state.
121 * - The interrupt vector number.
123 * The "processing" argument specifies the code for processing
124 * the interrupt. Defaults to "handle_interrupt".
126 .macro int_hand vecnum, vecname, c_routine, processing=handle_interrupt
129 .ifc \vecnum, INT_SWINT_1
130 blz TREG_SYSCALL_NR_NAME, sys_cmpxchg
133 /* Temporarily save a register so we have somewhere to work. */
135 mtspr SPR_SYSTEM_SAVE_K_1, r0
136 mfspr r0, SPR_EX_CONTEXT_K_1
138 /* The cmpxchg code clears sp to force us to reset it here on fault. */
141 andi r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
144 .ifc \vecnum, INT_DOUBLE_FAULT
146 * For double-faults from user-space, fall through to the normal
147 * register save and stack setup path. Otherwise, it's the
148 * hypervisor giving us one last chance to dump diagnostics, and we
149 * branch to the kernel_double_fault routine to do so.
152 j _kernel_double_fault
156 * If we're coming from user-space, then set sp to the top of
157 * the kernel stack. Otherwise, assume sp is already valid.
165 .ifc \c_routine, do_page_fault
167 * The page_fault handler may be downcalled directly by the
168 * hypervisor even when Linux is running and has ICS set.
170 * In this case the contents of EX_CONTEXT_K_1 reflect the
171 * previous fault and can't be relied on to choose whether or
172 * not to reinitialize the stack pointer. So we add a test
173 * to see whether SYSTEM_SAVE_K_2 has the high bit set,
174 * and if so we don't reinitialize sp, since we must be coming
175 * from Linux. (In fact the precise case is !(val & ~1),
176 * but any Linux PC has to have the high bit set.)
178 * Note that the hypervisor *always* sets SYSTEM_SAVE_K_2 for
179 * any path that turns into a downcall to one of our TLB handlers.
181 mfspr r0, SPR_SYSTEM_SAVE_K_2
183 blz r0, 0f /* high bit in S_S_1_2 is for a PC to use */
190 * SYSTEM_SAVE_K_0 holds the cpu number in the low bits, and
191 * the current stack top in the higher bits. So we recover
192 * our stack top by just masking off the low bits, then
193 * point sp at the top aligned address on the actual stack page.
195 mfspr r0, SPR_SYSTEM_SAVE_K_0
196 mm r0, r0, zero, LOG2_THREAD_SIZE, 31
200 * Align the stack mod 64 so we can properly predict what
201 * cache lines we need to write-hint to reduce memory fetch
202 * latency as we enter the kernel. The layout of memory is
203 * as follows, with cache line 0 at the lowest VA, and cache
204 * line 4 just below the r0 value this "andi" computes.
205 * Note that we never write to cache line 4, and we skip
206 * cache line 1 for syscalls.
208 * cache line 4: ptregs padding (two words)
209 * cache line 3: r46...lr, pc, ex1, faultnum, orig_r0, flags, pad
210 * cache line 2: r30...r45
211 * cache line 1: r14...r29
212 * cache line 0: 2 x frame, r0..r13
217 * Push the first four registers on the stack, so that we can set
218 * them to vector-unique values before we jump to the common code.
220 * Registers are pushed on the stack as a struct pt_regs,
221 * with the sp initially just above the struct, and when we're
222 * done, sp points to the base of the struct, minus
223 * C_ABI_SAVE_AREA_SIZE, so we can directly jal to C code.
225 * This routine saves just the first four registers, plus the
226 * stack context so we can do proper backtracing right away,
227 * and defers to handle_interrupt to save the rest.
228 * The backtracer needs pc, ex1, lr, sp, r52, and faultnum.
230 addli r0, r0, PTREGS_OFFSET_LR - (PTREGS_SIZE + KSTK_PTREGS_GAP)
231 wh64 r0 /* cache line 3 */
234 addli r0, r0, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
238 addli sp, r0, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_SP
242 addli sp, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(52)
244 wh64 sp /* cache line 0 */
247 addli sp, sp, PTREGS_OFFSET_REG(2) - PTREGS_OFFSET_REG(1)
251 addli sp, sp, PTREGS_OFFSET_REG(3) - PTREGS_OFFSET_REG(2)
255 addli sp, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(3)
257 mfspr r0, SPR_EX_CONTEXT_K_0
258 .ifc \processing,handle_syscall
260 * Bump the saved PC by one bundle so that when we return, we won't
261 * execute the same swint instruction again. We need to do this while
262 * we're in the critical section.
268 addli sp, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
270 mfspr r0, SPR_EX_CONTEXT_K_1
273 addi sp, sp, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
275 * Use r0 for syscalls so it's a temporary; use r1 for interrupts
276 * so that it gets passed through unchanged to the handler routine.
277 * Note that the .if conditional confusingly spans bundles.
279 .ifc \processing,handle_syscall
290 addli sp, sp, PTREGS_OFFSET_REG(0) - PTREGS_OFFSET_FAULTNUM
292 mfspr r0, SPR_SYSTEM_SAVE_K_1 /* Original r0 */
295 addi sp, sp, -PTREGS_OFFSET_REG(0) - 4
298 sw sp, zero /* write zero into "Next SP" frame pointer */
299 addi sp, sp, -4 /* leave SP pointing at bottom of frame */
301 .ifc \processing,handle_syscall
305 * Capture per-interrupt SPR context to registers.
306 * We overload the meaning of r3 on this path such that if its bit 31
307 * is set, we have to mask all interrupts including NMIs before
308 * clearing the interrupt critical section bit.
309 * See discussion below at "finish_interrupt_save".
311 .ifc \c_routine, do_page_fault
312 mfspr r2, SPR_SYSTEM_SAVE_K_3 /* address of page fault */
313 mfspr r3, SPR_SYSTEM_SAVE_K_2 /* info about page fault */
315 .ifc \vecnum, INT_DOUBLE_FAULT
317 mfspr r2, SPR_SYSTEM_SAVE_K_2 /* double fault info from HV */
321 .ifc \c_routine, do_trap
327 .ifc \c_routine, op_handle_perf_interrupt
329 mfspr r2, PERF_COUNT_STS
330 movei r3, -1 /* not used, but set for consistency */
333 #if CHIP_HAS_AUX_PERF_COUNTERS()
334 .ifc \c_routine, op_handle_aux_perf_interrupt
336 mfspr r2, AUX_PERF_COUNT_STS
337 movei r3, -1 /* not used, but set for consistency */
342 #if CHIP_HAS_AUX_PERF_COUNTERS()
349 /* Put function pointer in r0 */
350 moveli r0, lo16(\c_routine)
352 auli r0, r0, ha16(\c_routine)
356 ENDPROC(intvec_\vecname)
358 #ifdef __COLLECT_LINKER_FEEDBACK__
359 .pushsection .text.intvec_feedback,"ax"
361 FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt1, 1 << 8)
370 * Save the rest of the registers that we didn't save in the actual
371 * vector itself. We can't use r0-r10 inclusive here.
373 .macro finish_interrupt_save, function
375 /* If it's a syscall, save a proper orig_r0, otherwise just zero. */
376 PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
378 .ifc \function,handle_syscall
383 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
387 * For ordinary syscalls, we save neither caller- nor callee-
388 * save registers, since the syscall invoker doesn't expect the
389 * caller-saves to be saved, and the called kernel functions will
390 * take care of saving the callee-saves for us.
392 * For interrupts we save just the caller-save registers. Saving
393 * them is required (since the "caller" can't save them). Again,
394 * the called kernel functions will restore the callee-save
395 * registers for us appropriately.
397 * On return, we normally restore nothing special for syscalls,
398 * and just the caller-save registers for interrupts.
400 * However, there are some important caveats to all this:
402 * - We always save a few callee-save registers to give us
403 * some scratchpad registers to carry across function calls.
405 * - fork/vfork/etc require us to save all the callee-save
406 * registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
408 * - We always save r0..r5 and r10 for syscalls, since we need
409 * to reload them a bit later for the actual kernel call, and
410 * since we might need them for -ERESTARTNOINTR, etc.
412 * - Before invoking a signal handler, we save the unsaved
413 * callee-save registers so they are visible to the
414 * signal handler or any ptracer.
416 * - If the unsaved callee-save registers are modified, we set
417 * a bit in pt_regs so we know to reload them from pt_regs
418 * and not just rely on the kernel function unwinding.
419 * (Done for ptrace register writes and SA_SIGINFO handler.)
423 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
425 wh64 r52 /* cache line 2 */
429 .ifc \function,handle_syscall
430 push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
431 push_reg TREG_SYSCALL_NR_NAME, r52, \
432 PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
435 push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
436 wh64 r52 /* cache line 1 */
467 /* Load tp with our per-cpu offset. */
470 mfspr r20, SPR_SYSTEM_SAVE_K_0
471 moveli r21, lo16(__per_cpu_offset)
474 auli r21, r21, ha16(__per_cpu_offset)
475 mm r20, r20, zero, 0, LOG2_THREAD_SIZE-1
484 * If we will be returning to the kernel, we will need to
485 * reset the interrupt masks to the state they had before.
486 * Set DISABLE_IRQ in flags iff we came from PL1 with irqs disabled.
487 * We load flags in r32 here so we can jump to .Lrestore_regs
488 * directly after do_page_fault_ics() if necessary.
490 mfspr r32, SPR_EX_CONTEXT_K_1
492 andi r32, r32, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
493 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
495 bzt r32, 1f /* zero if from user space */
496 IRQS_DISABLED(r32) /* zero if irqs enabled */
497 #if PT_FLAGS_DISABLE_IRQ != 1
498 # error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
501 .ifnc \function,handle_syscall
502 /* Record the fact that we saved the caller-save registers above. */
503 ori r32, r32, PT_FLAGS_CALLER_SAVES
507 #ifdef __COLLECT_LINKER_FEEDBACK__
509 * Notify the feedback routines that we were in the
510 * appropriate fixed interrupt vector area. Note that we
511 * still have ICS set at this point, so we can't invoke any
512 * atomic operations or we will panic. The feedback
513 * routines internally preserve r0..r10 and r30 up.
515 .ifnc \function,handle_syscall
518 moveli r20, INT_SWINT_1 << 5
520 addli r20, r20, lo16(intvec_feedback)
521 auli r20, r20, ha16(intvec_feedback)
524 /* And now notify the feedback routines that we are here. */
525 FEEDBACK_ENTER(\function)
529 * we've captured enough state to the stack (including in
530 * particular our EX_CONTEXT state) that we can now release
531 * the interrupt critical section and replace it with our
532 * standard "interrupts disabled" mask value. This allows
533 * synchronous interrupts (and profile interrupts) to punch
534 * through from this point onwards.
536 * If bit 31 of r3 is set during a non-NMI interrupt, we know we
537 * are on the path where the hypervisor has punched through our
538 * ICS with a page fault, so we call out to do_page_fault_ics()
539 * to figure out what to do with it. If the fault was in
540 * an atomic op, we unlock the atomic lock, adjust the
541 * saved register state a little, and return "zero" in r4,
542 * falling through into the normal page-fault interrupt code.
543 * If the fault was in a kernel-space atomic operation, then
544 * do_page_fault_ics() resolves it itself, returns "one" in r4,
545 * and as a result goes directly to restoring registers and iret,
546 * without trying to adjust the interrupt masks at all.
547 * The do_page_fault_ics() API involves passing and returning
548 * a five-word struct (in registers) to avoid writing the
549 * save and restore code here.
551 .ifc \function,handle_nmi
554 .ifnc \function,handle_syscall
557 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
558 jal do_page_fault_ics
560 FEEDBACK_REENTER(\function)
565 IRQ_DISABLE(r20, r21)
567 mtspr INTERRUPT_CRITICAL_SECTION, zero
571 * Prepare the first 256 stack bytes to be rapidly accessible
572 * without having to fetch the background data. We don't really
573 * know how far to write-hint, but kernel stacks generally
574 * aren't that big, and write-hinting here does take some time.
592 #ifdef CONFIG_TRACE_IRQFLAGS
593 .ifnc \function,handle_nmi
595 * We finally have enough state set up to notify the irq
596 * tracing code that irqs were disabled on entry to the handler.
597 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
598 * For syscalls, we already have the register state saved away
599 * on the stack, so we don't bother to do any register saves here,
600 * and later we pop the registers back off the kernel stack.
601 * For interrupt handlers, save r0-r3 in callee-saved registers.
603 .ifnc \function,handle_syscall
604 { move r30, r0; move r31, r1 }
605 { move r32, r2; move r33, r3 }
608 .ifnc \function,handle_syscall
609 { move r0, r30; move r1, r31 }
610 { move r2, r32; move r3, r33 }
617 .macro check_single_stepping, kind, not_single_stepping
619 * Check for single stepping in user-level priv
620 * kind can be "normal", "ill", or "syscall"
621 * At end, if fall-thru
622 * r29: thread_info->step_state
625 * r26: thread_info->step_state->buffer
628 /* Check for single stepping */
631 /* Get pointer to field holding step state */
632 addi r29, r29, THREAD_INFO_STEP_STATE_OFFSET
634 /* Get pointer to EX1 in register state */
635 PTREGS_PTR(r27, PTREGS_OFFSET_EX1)
638 /* Get pointer to field holding PC */
639 PTREGS_PTR(r28, PTREGS_OFFSET_PC)
641 /* Load the pointer to the step state */
647 /* Points to flags */
648 addi r23, r29, SINGLESTEP_STATE_FLAGS_OFFSET
650 /* No single stepping if there is no step state structure */
651 bzt r29, \not_single_stepping
654 /* mask off ICS and any other high bits */
655 andi r27, r27, SPR_EX_CONTEXT_1_1__PL_MASK
657 /* Load pointer to single step instruction buffer */
660 /* Check priv state */
661 bnz r27, \not_single_stepping
666 /* Branch if single-step mode not enabled */
667 bbnst r22, \not_single_stepping
669 /* Clear enabled flag */
670 andi r22, r22, ~SINGLESTEP_STATE_MASK_IS_ENABLED
677 /* Point to the entry containing the original PC */
678 addi r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
681 /* Disable single stepping flag */
685 /* Get the original pc */
688 /* See if the PC is at the start of the single step buffer */
692 * NOTE: it is really expected that the PC be in the single step buffer
695 bzt r25, \not_single_stepping
697 /* Restore the original PC */
705 /* Point to the entry containing the next PC */
706 addi r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
709 /* Increment the stopped PC by the bundle size */
712 /* Disable single stepping flag */
716 /* Get the next pc */
720 * See if the PC is one bundle past the start of the
727 * NOTE: it is really expected that the PC be in the
728 * single step buffer at this point
730 bzt r25, \not_single_stepping
732 /* Set to the next PC */
736 /* Point to 3rd bundle in buffer */
743 /* Disable single stepping flag */
746 /* See if the PC is in the single step buffer */
753 * NOTE: it is really expected that the PC be in the
754 * single step buffer at this point
756 bzt r24, \not_single_stepping
758 bzt r25, \not_single_stepping
764 * Redispatch a downcall.
766 .macro dc_dispatch vecnum, vecname
769 j hv_downcall_dispatch
770 ENDPROC(intvec_\vecname)
774 * Common code for most interrupts. The C function we're eventually
775 * going to is in r0, and the faultnum is in r1; the original
776 * values for those registers are on the stack.
778 .pushsection .text.handle_interrupt,"ax"
780 finish_interrupt_save handle_interrupt
783 * Check for if we are single stepping in user level. If so, then
784 * we need to restore the PC.
787 check_single_stepping normal, .Ldispatch_interrupt
788 .Ldispatch_interrupt:
790 /* Jump to the C routine; it should enable irqs as soon as possible. */
793 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
795 FEEDBACK_REENTER(handle_interrupt)
797 movei r30, 0 /* not an NMI */
800 STD_ENDPROC(handle_interrupt)
803 * This routine takes a boolean in r30 indicating if this is an NMI.
804 * If so, we also expect a boolean in r31 indicating whether to
805 * re-enable the oprofile interrupts.
807 STD_ENTRY(interrupt_return)
808 /* If we're resuming to kernel space, don't check thread flags. */
810 bnz r30, .Lrestore_all /* NMIs don't special-case user-space */
811 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
814 andi r29, r29, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
816 bzt r29, .Lresume_userspace
817 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
820 /* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
823 moveli r27, lo16(_cpu_idle_nap)
826 auli r27, r27, ha16(_cpu_idle_nap)
832 bbns r27, .Lrestore_all
839 FEEDBACK_REENTER(interrupt_return)
842 * Disable interrupts so as to make sure we don't
843 * miss an interrupt that sets any of the thread flags (like
844 * need_resched or sigpending) between sampling and the iret.
845 * Routines like schedule() or do_signal() may re-enable
846 * interrupts before returning.
848 IRQ_DISABLE(r20, r21)
849 TRACE_IRQS_OFF /* Note: clobbers registers r0-r29 */
851 /* Get base of stack in r32; note r30/31 are used as arguments here. */
855 /* Check to see if there is any work to do before returning to user. */
857 addi r29, r32, THREAD_INFO_FLAGS_OFFSET
858 moveli r28, lo16(_TIF_ALLWORK_MASK)
862 auli r28, r28, ha16(_TIF_ALLWORK_MASK)
865 bnz r28, .Lwork_pending
869 * omit the call to single_process_check_nohz, which normally checks
870 * to see if we should start or stop the scheduler tick, because
871 * we can't call arbitrary Linux code from an NMI context.
872 * We always call the homecache TLB deferral code to re-trigger
873 * the deferral mechanism.
875 * The other chunk of responsibility this code has is to reset the
876 * interrupt masks appropriately to reset irqs and NMIs. We have
877 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
878 * lockdep-type stuff, but we can't set ICS until afterwards, since
879 * ICS can only be used in very tight chunks of code to avoid
880 * tripping over various assertions that it is off.
882 * (There is what looks like a window of vulnerability here since
883 * we might take a profile interrupt between the two SPR writes
884 * that set the mask, but since we write the low SPR word first,
885 * and our interrupt entry code checks the low SPR word, any
886 * profile interrupt will actually disable interrupts in both SPRs
887 * before returning, which is OK.)
890 PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
893 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
896 andi r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK
901 #if PT_FLAGS_DISABLE_IRQ != 1
902 # error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use bbnst below
908 mtspr INTERRUPT_CRITICAL_SECTION, r0
909 bzt r30, .Lrestore_regs
913 mtspr INTERRUPT_CRITICAL_SECTION, r0
915 bzt r30, .Lrestore_regs
920 * We now commit to returning from this interrupt, since we will be
921 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
922 * frame. No calls should be made to any other code after this point.
923 * This code should only be entered with ICS set.
924 * r32 must still be set to ptregs.flags.
925 * We launch loads to each cache line separately first, so we can
926 * get some parallelism out of the memory subsystem.
927 * We start zeroing caller-saved registers throughout, since
928 * that will save some cycles if this turns out to be a syscall.
931 FEEDBACK_REENTER(interrupt_return) /* called from elsewhere */
934 * Rotate so we have one high bit and one low bit to test.
935 * - low bit says whether to restore all the callee-saved registers,
936 * or just r30-r33, and r52 up.
937 * - high bit (i.e. sign bit) says whether to restore all the
938 * caller-saved registers, or just r0.
940 #if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
941 # error Rotate trick does not work :-)
945 PTREGS_PTR(sp, PTREGS_OFFSET_REG(0))
949 * Load cache lines 0, 2, and 3 in that order, then use
950 * the last loaded value, which makes it likely that the other
951 * cache lines have also loaded, at which point we should be
952 * able to safely read all the remaining words on those cache
953 * lines without waiting for the memory subsystem.
955 pop_reg_zero r0, r28, sp, PTREGS_OFFSET_REG(30) - PTREGS_OFFSET_REG(0)
956 pop_reg_zero r30, r2, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(30)
957 pop_reg_zero r21, r3, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
958 pop_reg_zero lr, r4, sp, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_EX1
960 mtspr SPR_EX_CONTEXT_K_0, r21
964 mtspr SPR_EX_CONTEXT_K_1, lr
965 andi lr, lr, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
968 /* Restore callee-saveds that we actually use. */
969 pop_reg_zero r52, r6, sp, PTREGS_OFFSET_REG(31) - PTREGS_OFFSET_REG(52)
972 pop_reg_zero r33, r9, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(33)
975 * If we modified other callee-saveds, restore them now.
976 * This is rare, but could be via ptrace or signal handler.
980 bbs r20, .Lrestore_callees
982 .Lcontinue_restore_regs:
984 /* Check if we're returning from a syscall. */
987 blzt r20, 1f /* no, so go restore callee-save registers */
991 * Check if we're returning to userspace.
992 * Note that if we're not, we don't worry about zeroing everything.
995 addli sp, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(29)
996 bnz lr, .Lkernel_return
1000 * On return from syscall, we've restored r0 from pt_regs, but we
1001 * clear the remainder of the caller-saved registers. We could
1002 * restore the syscall arguments, but there's not much point,
1003 * and it ensures user programs aren't trying to use the
1004 * caller-saves if we clear them, as well as avoiding leaking
1005 * kernel pointers into userspace.
1007 pop_reg_zero lr, r12, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1008 pop_reg_zero tp, r13, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1014 { move r16, zero; move r17, zero }
1015 { move r18, zero; move r19, zero }
1016 { move r20, zero; move r21, zero }
1017 { move r22, zero; move r23, zero }
1018 { move r24, zero; move r25, zero }
1019 { move r26, zero; move r27, zero }
1021 /* Set r1 to errno if we are returning an error, otherwise zero. */
1034 * Not a syscall, so restore caller-saved registers.
1035 * First kick off a load for cache line 1, which we're touching
1036 * for the first time here.
1039 1: pop_reg r29, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(29)
1067 pop_reg r28, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(28)
1068 /* r29 already restored above */
1069 bnz lr, .Lkernel_return
1070 pop_reg lr, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1071 pop_reg tp, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1076 * We can't restore tp when in kernel mode, since a thread might
1077 * have migrated from another cpu and brought a stale tp value.
1080 pop_reg lr, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
1084 /* Restore callee-saved registers from r34 to r51. */
1086 addli sp, sp, PTREGS_OFFSET_REG(34) - PTREGS_OFFSET_REG(29)
1104 pop_reg r51, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(51)
1105 j .Lcontinue_restore_regs
1108 /* Mask the reschedule flag */
1109 andi r28, r29, _TIF_NEED_RESCHED
1113 * If the NEED_RESCHED flag is called, we call schedule(), which
1114 * may drop this context right here and go do something else.
1115 * On return, jump back to .Lresume_userspace and recheck.
1119 /* Mask the async-tlb flag */
1120 andi r28, r29, _TIF_ASYNC_TLB
1124 FEEDBACK_REENTER(interrupt_return)
1126 /* Reload the flags and check again */
1127 j .Lresume_userspace
1131 bz r28, .Lneed_sigpending
1133 /* Mask the sigpending flag */
1134 andi r28, r29, _TIF_SIGPENDING
1137 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1138 jal do_async_page_fault
1139 FEEDBACK_REENTER(interrupt_return)
1142 * Go restart the "resume userspace" process. We may have
1143 * fired a signal, and we need to disable interrupts again.
1145 j .Lresume_userspace
1149 * At this point we are either doing signal handling or single-step,
1150 * so either way make sure we have all the registers saved.
1152 push_extra_callee_saves r0
1155 /* If no signal pending, skip to singlestep check */
1156 bz r28, .Lneed_singlestep
1158 /* Mask the singlestep flag */
1159 andi r28, r29, _TIF_SINGLESTEP
1163 FEEDBACK_REENTER(interrupt_return)
1165 /* Reload the flags and check again */
1166 j .Lresume_userspace
1170 /* Get a pointer to the EX1 field */
1171 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
1173 /* If we get here, our bit must be set. */
1174 bz r28, .Lwork_confusion
1176 /* If we are in priv mode, don't single step */
1178 andi r28, r28, SPR_EX_CONTEXT_1_1__PL_MASK /* mask off ICS */
1179 bnz r28, .Lrestore_all
1181 /* Allow interrupts within the single step code */
1182 TRACE_IRQS_ON /* Note: clobbers registers r0-r29 */
1183 IRQ_ENABLE(r20, r21)
1185 /* try to single-step the current instruction */
1186 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1187 jal single_step_once
1188 FEEDBACK_REENTER(interrupt_return)
1190 /* Re-disable interrupts. TRACE_IRQS_OFF in .Lrestore_all. */
1191 IRQ_DISABLE(r20,r21)
1197 panic "thread_info allwork flags unhandled on userspace resume: %#x"
1199 STD_ENDPROC(interrupt_return)
1202 * This interrupt variant clears the INT_INTCTRL_K interrupt mask bit
1203 * before returning, so we can properly get more downcalls.
1205 .pushsection .text.handle_interrupt_downcall,"ax"
1206 handle_interrupt_downcall:
1207 finish_interrupt_save handle_interrupt_downcall
1208 check_single_stepping normal, .Ldispatch_downcall
1209 .Ldispatch_downcall:
1211 /* Clear INTCTRL_K from the set of interrupts we ever enable. */
1212 GET_INTERRUPTS_ENABLED_MASK_PTR(r30)
1215 movei r31, INT_MASK(INT_INTCTRL_K)
1226 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1228 FEEDBACK_REENTER(handle_interrupt_downcall)
1230 /* Allow INTCTRL_K to be enabled next time we enable interrupts. */
1236 movei r30, 0 /* not an NMI */
1239 STD_ENDPROC(handle_interrupt_downcall)
1242 * Some interrupts don't check for single stepping
1244 .pushsection .text.handle_interrupt_no_single_step,"ax"
1245 handle_interrupt_no_single_step:
1246 finish_interrupt_save handle_interrupt_no_single_step
1249 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1251 FEEDBACK_REENTER(handle_interrupt_no_single_step)
1253 movei r30, 0 /* not an NMI */
1256 STD_ENDPROC(handle_interrupt_no_single_step)
1259 * "NMI" interrupts mask ALL interrupts before calling the
1260 * handler, and don't check thread flags, etc., on the way
1261 * back out. In general, the only things we do here for NMIs
1262 * are the register save/restore, fixing the PC if we were
1263 * doing single step, and the dataplane kernel-TLB management.
1264 * We don't (for example) deal with start/stop of the sched tick.
1266 .pushsection .text.handle_nmi,"ax"
1268 finish_interrupt_save handle_nmi
1269 check_single_stepping normal, .Ldispatch_nmi
1273 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1275 FEEDBACK_REENTER(handle_nmi)
1277 STD_ENDPROC(handle_nmi)
1280 * Parallel code for syscalls to handle_interrupt.
1282 .pushsection .text.handle_syscall,"ax"
1284 finish_interrupt_save handle_syscall
1287 * Check for if we are single stepping in user level. If so, then
1288 * we need to restore the PC.
1290 check_single_stepping syscall, .Ldispatch_syscall
1295 IRQ_ENABLE(r20, r21)
1297 /* Bump the counter for syscalls made on this tile. */
1298 moveli r20, lo16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1299 auli r20, r20, ha16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1305 /* Trace syscalls, if requested. */
1306 GET_THREAD_INFO(r31)
1307 addi r31, r31, THREAD_INFO_FLAGS_OFFSET
1309 andi r30, r30, _TIF_SYSCALL_TRACE
1310 bzt r30, .Lrestore_syscall_regs
1311 jal do_syscall_trace
1312 FEEDBACK_REENTER(handle_syscall)
1315 * We always reload our registers from the stack at this
1316 * point. They might be valid, if we didn't build with
1317 * TRACE_IRQFLAGS, and this isn't a dataplane tile, and we're not
1318 * doing syscall tracing, but there are enough cases now that it
1319 * seems simplest just to do the reload unconditionally.
1321 .Lrestore_syscall_regs:
1322 PTREGS_PTR(r11, PTREGS_OFFSET_REG(0))
1328 pop_reg r5, r11, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(5)
1329 pop_reg TREG_SYSCALL_NR_NAME, r11
1331 /* Ensure that the syscall number is within the legal range. */
1332 moveli r21, __NR_syscalls
1334 slt_u r21, TREG_SYSCALL_NR_NAME, r21
1335 moveli r20, lo16(sys_call_table)
1338 bbns r21, .Linvalid_syscall
1339 auli r20, r20, ha16(sys_call_table)
1341 s2a r20, TREG_SYSCALL_NR_NAME, r20
1344 /* Jump to syscall handler. */
1346 .Lhandle_syscall_link: /* value of "lr" after "jalr r20" above */
1349 * Write our r0 onto the stack so it gets restored instead
1350 * of whatever the user had there before.
1352 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1355 .Lsyscall_sigreturn_skip:
1356 FEEDBACK_REENTER(handle_syscall)
1358 /* Do syscall trace again, if requested. */
1360 andi r30, r30, _TIF_SYSCALL_TRACE
1362 jal do_syscall_trace
1363 FEEDBACK_REENTER(handle_syscall)
1364 1: j .Lresume_userspace /* jump into middle of interrupt_return */
1367 /* Report an invalid syscall back to the user program */
1369 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1373 j .Lresume_userspace /* jump into middle of interrupt_return */
1374 STD_ENDPROC(handle_syscall)
1376 /* Return the address for oprofile to suppress in backtraces. */
1377 STD_ENTRY_SECTION(handle_syscall_link_address, .text.handle_syscall)
1380 addli r0, r0, .Lhandle_syscall_link - .
1383 STD_ENDPROC(handle_syscall_link_address)
1385 STD_ENTRY(ret_from_fork)
1388 FEEDBACK_REENTER(ret_from_fork)
1389 j .Lresume_userspace /* jump into middle of interrupt_return */
1390 STD_ENDPROC(ret_from_fork)
1393 * Code for ill interrupt.
1395 .pushsection .text.handle_ill,"ax"
1397 finish_interrupt_save handle_ill
1400 * Check for if we are single stepping in user level. If so, then
1401 * we need to restore the PC.
1403 check_single_stepping ill, .Ldispatch_normal_ill
1406 /* See if the PC is the 1st bundle in the buffer */
1409 /* Point to the 2nd bundle in the buffer */
1413 /* Point to the original pc */
1414 addi r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
1416 /* Branch if the PC is the 1st bundle in the buffer */
1420 /* See if the PC is the 2nd bundle of the buffer */
1423 /* Set PC to next instruction */
1424 addi r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
1427 /* Point to flags */
1428 addi r25, r29, SINGLESTEP_STATE_FLAGS_OFFSET
1430 /* Branch if PC is in the second bundle */
1437 * Get the offset for the register to restore
1438 * Note: the lower bound is 2, so we have implicit scaling by 4.
1439 * No multiplication of the register number by the size of a register
1442 mm r27, r25, zero, SINGLESTEP_STATE_TARGET_LB, \
1443 SINGLESTEP_STATE_TARGET_UB
1445 /* Mask Rewrite_LR */
1446 andi r25, r25, SINGLESTEP_STATE_MASK_UPDATE
1449 addi r29, r29, SINGLESTEP_STATE_UPDATE_VALUE_OFFSET
1451 /* Don't rewrite temp register */
1455 /* Get the temp value */
1458 /* Point to where the register is stored */
1462 /* Add in the C ABI save area size to the register offset */
1463 addi r27, r27, C_ABI_SAVE_AREA_SIZE
1465 /* Restore the user's register with the temp value */
1470 /* Must be in the third bundle */
1471 addi r24, r29, SINGLESTEP_STATE_BRANCH_NEXT_PC_OFFSET
1474 /* set PC and continue */
1479 * Clear TIF_SINGLESTEP to prevent recursion if we execute an ill.
1480 * The normal non-arch flow redundantly clears TIF_SINGLESTEP, but we
1481 * need to clear it here and can't really impose on all other arches.
1482 * So what's another write between friends?
1486 addi r1, r0, THREAD_INFO_FLAGS_OFFSET
1489 addi r0, r0, THREAD_INFO_TASK_OFFSET /* currently a no-op */
1491 andi r2, r2, ~_TIF_SINGLESTEP
1494 /* Issue a sigtrap */
1496 lw r0, r0 /* indirect thru thread_info to get task_info*/
1497 addi r1, sp, C_ABI_SAVE_AREA_SIZE /* put ptregs pointer into r1 */
1498 move r2, zero /* load error code into r2 */
1501 jal send_sigtrap /* issue a SIGTRAP */
1502 FEEDBACK_REENTER(handle_ill)
1503 j .Lresume_userspace /* jump into middle of interrupt_return */
1505 .Ldispatch_normal_ill:
1508 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1510 FEEDBACK_REENTER(handle_ill)
1512 movei r30, 0 /* not an NMI */
1515 STD_ENDPROC(handle_ill)
1517 /* Various stub interrupt handlers and syscall handlers */
1519 STD_ENTRY_LOCAL(_kernel_double_fault)
1520 mfspr r1, SPR_EX_CONTEXT_K_0
1524 addi sp, sp, -C_ABI_SAVE_AREA_SIZE
1525 j kernel_double_fault
1526 STD_ENDPROC(_kernel_double_fault)
1528 STD_ENTRY_LOCAL(bad_intr)
1529 mfspr r2, SPR_EX_CONTEXT_K_0
1530 panic "Unhandled interrupt %#x: PC %#lx"
1531 STD_ENDPROC(bad_intr)
1533 /* Put address of pt_regs in reg and jump. */
1534 #define PTREGS_SYSCALL(x, reg) \
1537 PTREGS_PTR(reg, PTREGS_OFFSET_BASE); \
1543 * Special-case sigreturn to not write r0 to the stack on return.
1544 * This is technically more efficient, but it also avoids difficulties
1545 * in the 64-bit OS when handling 32-bit compat code, since we must not
1546 * sign-extend r0 for the sigreturn return-value case.
1548 #define PTREGS_SYSCALL_SIGRETURN(x, reg) \
1550 addli lr, lr, .Lsyscall_sigreturn_skip - .Lhandle_syscall_link; \
1552 PTREGS_PTR(reg, PTREGS_OFFSET_BASE); \
1557 PTREGS_SYSCALL(sys_execve, r3)
1558 PTREGS_SYSCALL(sys_sigaltstack, r2)
1559 PTREGS_SYSCALL_SIGRETURN(sys_rt_sigreturn, r0)
1560 PTREGS_SYSCALL(sys_cmpxchg_badaddr, r1)
1562 /* Save additional callee-saves to pt_regs, put address in r4 and jump. */
1563 STD_ENTRY(_sys_clone)
1564 push_extra_callee_saves r4
1566 STD_ENDPROC(_sys_clone)
1569 * This entrypoint is taken for the cmpxchg and atomic_update fast
1570 * swints. We may wish to generalize it to other fast swints at some
1571 * point, but for now there are just two very similar ones, which
1574 * The fast swint code is designed to have a small footprint. It does
1575 * not save or restore any GPRs, counting on the caller-save registers
1576 * to be available to it on entry. It does not modify any callee-save
1577 * registers (including "lr"). It does not check what PL it is being
1578 * called at, so you'd better not call it other than at PL0.
1579 * The <atomic.h> wrapper assumes it only clobbers r20-r29, so if
1580 * it ever is necessary to use more registers, be aware.
1582 * It does not use the stack, but since it might be re-interrupted by
1583 * a page fault which would assume the stack was valid, it does
1584 * save/restore the stack pointer and zero it out to make sure it gets reset.
1585 * Since we always keep interrupts disabled, the hypervisor won't
1586 * clobber our EX_CONTEXT_K_x registers, so we don't save/restore them
1587 * (other than to advance the PC on return).
1589 * We have to manually validate the user vs kernel address range
1590 * (since at PL1 we can read/write both), and for performance reasons
1591 * we don't allow cmpxchg on the fc000000 memory region, since we only
1592 * validate that the user address is below PAGE_OFFSET.
1594 * We place it in the __HEAD section to ensure it is relatively
1595 * near to the intvec_SWINT_1 code (reachable by a conditional branch).
1597 * Must match register usage in do_page_fault().
1601 /* Align much later jump on the start of a cache line. */
1602 #if !ATOMIC_LOCKS_FOUND_VIA_TABLE()
1608 * Save "sp" and set it zero for any possible page fault.
1610 * HACK: We want to both zero sp and check r0's alignment,
1611 * so we do both at once. If "sp" becomes nonzero we
1612 * know r0 is unaligned and branch to the error handler that
1613 * restores sp, so this is OK.
1615 * ICS is disabled right now so having a garbage but nonzero
1616 * sp is OK, since we won't execute any faulting instructions
1617 * when it is nonzero.
1625 * Get the lock address in ATOMIC_LOCK_REG, and also validate that the
1626 * address is less than PAGE_OFFSET, since that won't trap at PL1.
1627 * We only use bits less than PAGE_SHIFT to avoid having to worry
1628 * about aliasing among multiple mappings of the same physical page,
1629 * and we ignore the low 3 bits so we have one lock that covers
1630 * both a cmpxchg64() and a cmpxchg() on either its low or high word.
1631 * NOTE: this code must match __atomic_hashed_lock() in lib/atomic.c.
1634 #if ATOMIC_LOCKS_FOUND_VIA_TABLE()
1636 /* Check for unaligned input. */
1637 bnz sp, .Lcmpxchg_badaddr
1638 mm r25, r0, zero, 3, PAGE_SHIFT-1
1641 crc32_32 r25, zero, r25
1642 moveli r21, lo16(atomic_lock_ptr)
1645 auli r21, r21, ha16(atomic_lock_ptr)
1646 auli r23, zero, hi16(PAGE_OFFSET) /* hugepage-aligned */
1649 shri r20, r25, 32 - ATOMIC_HASH_L1_SHIFT
1653 * Ensure that the TLB is loaded before we take out the lock.
1654 * On TILEPro, this will start fetching the value all the way
1655 * into our L1 as well (and if it gets modified before we
1656 * grab the lock, it will be invalidated from our cache
1657 * before we reload it). On tile64, we'll start fetching it
1658 * into our L1 if we're the home, and if we're not, we'll
1659 * still at least start fetching it into the home's L2.
1665 bbns r23, .Lcmpxchg_badaddr
1669 seqi r23, TREG_SYSCALL_NR_NAME, __NR_FAST_cmpxchg64
1670 andi r25, r25, ATOMIC_HASH_L2_SIZE - 1
1673 /* Branch away at this point if we're doing a 64-bit cmpxchg. */
1674 bbs r23, .Lcmpxchg64
1675 andi r23, r0, 7 /* Precompute alignment for cmpxchg64. */
1680 * We very carefully align the code that actually runs with
1681 * the lock held (nine bundles) so that we know it is all in
1682 * the icache when we start. This instruction (the jump) is
1683 * at the start of the first cache line, address zero mod 64;
1684 * we jump to somewhere in the second cache line to issue the
1685 * tns, then jump back to finish up.
1687 s2a ATOMIC_LOCK_REG_NAME, r25, r21
1691 #else /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */
1693 /* Check for unaligned input. */
1694 bnz sp, .Lcmpxchg_badaddr
1695 auli r23, zero, hi16(PAGE_OFFSET) /* hugepage-aligned */
1699 * Slide bits into position for 'mm'. We want to ignore
1700 * the low 3 bits of r0, and consider only the next
1701 * ATOMIC_HASH_SHIFT bits.
1702 * Because of C pointer arithmetic, we want to compute this:
1704 * ((char*)atomic_locks +
1705 * (((r0 >> 3) & (1 << (ATOMIC_HASH_SIZE - 1))) << 2))
1707 * Instead of two shifts we just ">> 1", and use 'mm'
1708 * to ignore the low and high bits we don't want.
1715 * Ensure that the TLB is loaded before we take out the lock.
1716 * On tilepro, this will start fetching the value all the way
1717 * into our L1 as well (and if it gets modified before we
1718 * grab the lock, it will be invalidated from our cache
1719 * before we reload it). On tile64, we'll start fetching it
1720 * into our L1 if we're the home, and if we're not, we'll
1721 * still at least start fetching it into the home's L2.
1726 /* atomic_locks is page aligned so this suffices to get its addr. */
1727 auli r21, zero, hi16(atomic_locks)
1729 bbns r23, .Lcmpxchg_badaddr
1733 * Insert the hash bits into the page-aligned pointer.
1734 * ATOMIC_HASH_SHIFT is so big that we don't actually hash
1735 * the unmasked address bits, as that may cause unnecessary
1738 mm ATOMIC_LOCK_REG_NAME, r25, r21, 2, (ATOMIC_HASH_SHIFT + 2) - 1
1740 seqi r23, TREG_SYSCALL_NR_NAME, __NR_FAST_cmpxchg64
1743 /* Branch away at this point if we're doing a 64-bit cmpxchg. */
1744 bbs r23, .Lcmpxchg64
1745 andi r23, r0, 7 /* Precompute alignment for cmpxchg64. */
1749 * We very carefully align the code that actually runs with
1750 * the lock held (nine bundles) so that we know it is all in
1751 * the icache when we start. This instruction (the jump) is
1752 * at the start of the first cache line, address zero mod 64;
1753 * we jump to somewhere in the second cache line to issue the
1754 * tns, then jump back to finish up.
1759 #endif /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */
1761 ENTRY(__sys_cmpxchg_grab_lock)
1764 * Perform the actual cmpxchg or atomic_update.
1765 * Note that __futex_mark_unlocked() in uClibc relies on
1766 * atomic_update() to always perform an "mf", so don't make
1767 * it optional or conditional without modifying that code.
1772 seqi r23, TREG_SYSCALL_NR_NAME, __NR_FAST_atomic_update
1776 seq r22, r21, r1 /* See if cmpxchg matches. */
1777 and r25, r21, r1 /* If atomic_update, compute (*mem & mask) */
1780 or r22, r22, r23 /* Skip compare branch for atomic_update. */
1781 add r25, r25, r2 /* Compute (*mem & mask) + addend. */
1784 mvnz r24, r23, r25 /* Use atomic_update value if appropriate. */
1785 bbns r22, .Lcmpxchg32_mismatch
1789 /* Do slow mtspr here so the following "mf" waits less. */
1792 mtspr SPR_EX_CONTEXT_K_0, r28
1796 /* The following instruction is the start of the second cache line. */
1799 sw ATOMIC_LOCK_REG_NAME, zero
1803 /* Duplicated code here in the case where we don't overlap "mf" */
1804 .Lcmpxchg32_mismatch:
1807 sw ATOMIC_LOCK_REG_NAME, zero
1811 mtspr SPR_EX_CONTEXT_K_0, r28
1816 * The locking code is the same for 32-bit cmpxchg/atomic_update,
1817 * and for 64-bit cmpxchg. We provide it as a macro and put
1818 * it into both versions. We can't share the code literally
1819 * since it depends on having the right branch-back address.
1820 * Note that the first few instructions should share the cache
1821 * line with the second half of the actual locked code.
1823 .macro cmpxchg_lock, bitwidth
1825 /* Lock; if we succeed, jump back up to the read-modify-write. */
1827 tns r21, ATOMIC_LOCK_REG_NAME
1830 * Non-SMP preserves all the lock infrastructure, to keep the
1831 * code simpler for the interesting (SMP) case. However, we do
1832 * one small optimization here and in atomic_asm.S, which is
1833 * to fake out acquiring the actual lock in the atomic_lock table.
1838 /* Issue the slow SPR here while the tns result is in flight. */
1839 mfspr r28, SPR_EX_CONTEXT_K_0
1842 addi r28, r28, 8 /* return to the instruction after the swint1 */
1843 bzt r21, .Ldo_cmpxchg\bitwidth
1846 * The preceding instruction is the last thing that must be
1847 * on the second cache line.
1852 * We failed to acquire the tns lock on our first try. Now use
1853 * bounded exponential backoff to retry, like __atomic_spinlock().
1856 moveli r23, 2048 /* maximum backoff time in cycles */
1857 moveli r25, 32 /* starting backoff time in cycles */
1859 1: mfspr r26, CYCLE_LOW /* get start point for this backoff */
1860 2: mfspr r22, CYCLE_LOW /* test to see if we've backed off enough */
1865 shli r25, r25, 1 /* double the backoff; retry the tns */
1866 tns r21, ATOMIC_LOCK_REG_NAME
1868 slt r26, r23, r25 /* is the proposed backoff too big? */
1871 bzt r21, .Ldo_cmpxchg\bitwidth
1874 #endif /* CONFIG_SMP */
1881 * This code is invoked from sys_cmpxchg after most of the
1882 * preconditions have been checked. We still need to check
1883 * that r0 is 8-byte aligned, since if it's not we won't
1884 * actually be atomic. However, ATOMIC_LOCK_REG has the atomic
1885 * lock pointer and r27/r28 have the saved SP/PC.
1886 * r23 is holding "r0 & 7" so we can test for alignment.
1887 * The compare value is in r2/r3; the new value is in r4/r5.
1888 * On return, we must put the old value in r0/r1.
1893 #if ATOMIC_LOCKS_FOUND_VIA_TABLE()
1894 s2a ATOMIC_LOCK_REG_NAME, r25, r21
1896 bzt r23, .Lcmpxchg64_tns
1910 bz r26, .Lcmpxchg64_mismatch
1914 bz r26, .Lcmpxchg64_mismatch
1920 * The 32-bit path provides optimized "match" and "mismatch"
1921 * iret paths, but we don't have enough bundles in this cache line
1922 * to do that, so we just make even the "mismatch" path do an "mf".
1924 .Lcmpxchg64_mismatch:
1927 mtspr SPR_EX_CONTEXT_K_0, r28
1932 sw ATOMIC_LOCK_REG_NAME, zero
1941 * Reset sp and revector to sys_cmpxchg_badaddr(), which will
1942 * just raise the appropriate signal and exit. Doing it this
1943 * way means we don't have to duplicate the code in intvec.S's
1944 * int_hand macro that locates the top of the stack.
1948 moveli TREG_SYSCALL_NR_NAME, __NR_cmpxchg_badaddr
1952 ENDPROC(sys_cmpxchg)
1953 ENTRY(__sys_cmpxchg_end)
1956 /* The single-step support may need to read all the registers. */
1958 push_extra_callee_saves r0
1961 /* Include .intrpt1 array of interrupt vectors */
1962 .section ".intrpt1", "ax"
1964 #define op_handle_perf_interrupt bad_intr
1965 #define op_handle_aux_perf_interrupt bad_intr
1967 #ifndef CONFIG_HARDWALL
1968 #define do_hardwall_trap bad_intr
1971 int_hand INT_ITLB_MISS, ITLB_MISS, \
1972 do_page_fault, handle_interrupt_no_single_step
1973 int_hand INT_MEM_ERROR, MEM_ERROR, bad_intr
1974 int_hand INT_ILL, ILL, do_trap, handle_ill
1975 int_hand INT_GPV, GPV, do_trap
1976 int_hand INT_SN_ACCESS, SN_ACCESS, do_trap
1977 int_hand INT_IDN_ACCESS, IDN_ACCESS, do_trap
1978 int_hand INT_UDN_ACCESS, UDN_ACCESS, do_trap
1979 int_hand INT_IDN_REFILL, IDN_REFILL, bad_intr
1980 int_hand INT_UDN_REFILL, UDN_REFILL, bad_intr
1981 int_hand INT_IDN_COMPLETE, IDN_COMPLETE, bad_intr
1982 int_hand INT_UDN_COMPLETE, UDN_COMPLETE, bad_intr
1983 int_hand INT_SWINT_3, SWINT_3, do_trap
1984 int_hand INT_SWINT_2, SWINT_2, do_trap
1985 int_hand INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall
1986 int_hand INT_SWINT_0, SWINT_0, do_trap
1987 int_hand INT_UNALIGN_DATA, UNALIGN_DATA, int_unalign
1988 int_hand INT_DTLB_MISS, DTLB_MISS, do_page_fault
1989 int_hand INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault
1990 int_hand INT_DMATLB_MISS, DMATLB_MISS, do_page_fault
1991 int_hand INT_DMATLB_ACCESS, DMATLB_ACCESS, do_page_fault
1992 int_hand INT_SNITLB_MISS, SNITLB_MISS, do_page_fault
1993 int_hand INT_SN_NOTIFY, SN_NOTIFY, bad_intr
1994 int_hand INT_SN_FIREWALL, SN_FIREWALL, do_hardwall_trap
1995 int_hand INT_IDN_FIREWALL, IDN_FIREWALL, bad_intr
1996 int_hand INT_UDN_FIREWALL, UDN_FIREWALL, do_hardwall_trap
1997 int_hand INT_TILE_TIMER, TILE_TIMER, do_timer_interrupt
1998 int_hand INT_IDN_TIMER, IDN_TIMER, bad_intr
1999 int_hand INT_UDN_TIMER, UDN_TIMER, bad_intr
2000 int_hand INT_DMA_NOTIFY, DMA_NOTIFY, bad_intr
2001 int_hand INT_IDN_CA, IDN_CA, bad_intr
2002 int_hand INT_UDN_CA, UDN_CA, bad_intr
2003 int_hand INT_IDN_AVAIL, IDN_AVAIL, bad_intr
2004 int_hand INT_UDN_AVAIL, UDN_AVAIL, bad_intr
2005 int_hand INT_PERF_COUNT, PERF_COUNT, \
2006 op_handle_perf_interrupt, handle_nmi
2007 int_hand INT_INTCTRL_3, INTCTRL_3, bad_intr
2008 #if CONFIG_KERNEL_PL == 2
2009 dc_dispatch INT_INTCTRL_2, INTCTRL_2
2010 int_hand INT_INTCTRL_1, INTCTRL_1, bad_intr
2012 int_hand INT_INTCTRL_2, INTCTRL_2, bad_intr
2013 dc_dispatch INT_INTCTRL_1, INTCTRL_1
2015 int_hand INT_INTCTRL_0, INTCTRL_0, bad_intr
2016 int_hand INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
2017 hv_message_intr, handle_interrupt_downcall
2018 int_hand INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, \
2019 tile_dev_intr, handle_interrupt_downcall
2020 int_hand INT_I_ASID, I_ASID, bad_intr
2021 int_hand INT_D_ASID, D_ASID, bad_intr
2022 int_hand INT_DMATLB_MISS_DWNCL, DMATLB_MISS_DWNCL, \
2023 do_page_fault, handle_interrupt_downcall
2024 int_hand INT_SNITLB_MISS_DWNCL, SNITLB_MISS_DWNCL, \
2025 do_page_fault, handle_interrupt_downcall
2026 int_hand INT_DMATLB_ACCESS_DWNCL, DMATLB_ACCESS_DWNCL, \
2027 do_page_fault, handle_interrupt_downcall
2028 int_hand INT_SN_CPL, SN_CPL, bad_intr
2029 int_hand INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
2030 #if CHIP_HAS_AUX_PERF_COUNTERS()
2031 int_hand INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \
2032 op_handle_aux_perf_interrupt, handle_nmi
2035 /* Synthetic interrupt delivered only by the simulator */
2036 int_hand INT_BREAKPOINT, BREAKPOINT, do_breakpoint