From 38fef73c21d117cf992fb5ec6e30630e54e13f4f Mon Sep 17 00:00:00 2001 From: Max Filippov Date: Thu, 16 Jul 2015 10:37:31 +0300 Subject: xtensa: implement fake NMI In case perf IRQ is the highest of the medium-level IRQs, and is alone on its level, it may be treated as NMI: - LOCKLEVEL is defined to be one level less than EXCM level, - IRQ masking never lowers current IRQ level, - new fake exception cause code, EXCCAUSE_MAPPED_NMI is assigned to that IRQ; new second level exception handler, do_nmi, assigned to it handles it as NMI, - atomic operations in configurations without s32c1i still need to mask all interrupts. Cc: Peter Zijlstra Acked-by: Peter Zijlstra (Intel) Signed-off-by: Max Filippov --- arch/xtensa/include/asm/atomic.h | 10 ++-- arch/xtensa/include/asm/cmpxchg.h | 4 +- arch/xtensa/include/asm/irqflags.h | 22 ++++++++- arch/xtensa/include/asm/processor.h | 31 ++++++++++++- arch/xtensa/kernel/entry.S | 93 +++++++++++++++++++++++++++++++------ arch/xtensa/kernel/irq.c | 8 ++++ arch/xtensa/kernel/perf_event.c | 6 ++- arch/xtensa/kernel/traps.c | 26 +++++++++++ arch/xtensa/kernel/vectors.S | 10 +++- 9 files changed, 183 insertions(+), 27 deletions(-) (limited to 'arch') diff --git a/arch/xtensa/include/asm/atomic.h b/arch/xtensa/include/asm/atomic.h index 00b7d46b35b8..ebcd1f6fc8cb 100644 --- a/arch/xtensa/include/asm/atomic.h +++ b/arch/xtensa/include/asm/atomic.h @@ -29,7 +29,7 @@ * * Locking interrupts looks like this: * - * rsil a15, LOCKLEVEL + * rsil a15, TOPLEVEL * * wsr a15, PS * rsync @@ -106,7 +106,7 @@ static inline void atomic_##op(int i, atomic_t * v) \ unsigned int vval; \ \ __asm__ __volatile__( \ - " rsil a15, "__stringify(LOCKLEVEL)"\n"\ + " rsil a15, "__stringify(TOPLEVEL)"\n"\ " l32i %0, %2, 0\n" \ " " #op " %0, %0, %1\n" \ " s32i %0, %2, 0\n" \ @@ -124,7 +124,7 @@ static inline int atomic_##op##_return(int i, atomic_t * v) \ unsigned int vval; \ \ __asm__ __volatile__( \ - " rsil a15,"__stringify(LOCKLEVEL)"\n" \ + " rsil a15,"__stringify(TOPLEVEL)"\n" \ " l32i %0, %2, 0\n" \ " " #op " %0, %0, %1\n" \ " s32i %0, %2, 0\n" \ @@ -272,7 +272,7 @@ static inline void atomic_clear_mask(unsigned int mask, atomic_t *v) unsigned int vval; __asm__ __volatile__( - " rsil a15,"__stringify(LOCKLEVEL)"\n" + " rsil a15,"__stringify(TOPLEVEL)"\n" " l32i %0, %2, 0\n" " xor %1, %4, %3\n" " and %0, %0, %4\n" @@ -306,7 +306,7 @@ static inline void atomic_set_mask(unsigned int mask, atomic_t *v) unsigned int vval; __asm__ __volatile__( - " rsil a15,"__stringify(LOCKLEVEL)"\n" + " rsil a15,"__stringify(TOPLEVEL)"\n" " l32i %0, %2, 0\n" " or %0, %0, %1\n" " s32i %0, %2, 0\n" diff --git a/arch/xtensa/include/asm/cmpxchg.h b/arch/xtensa/include/asm/cmpxchg.h index 370b26f38414..201e9009efd8 100644 --- a/arch/xtensa/include/asm/cmpxchg.h +++ b/arch/xtensa/include/asm/cmpxchg.h @@ -34,7 +34,7 @@ __cmpxchg_u32(volatile int *p, int old, int new) return new; #else __asm__ __volatile__( - " rsil a15, "__stringify(LOCKLEVEL)"\n" + " rsil a15, "__stringify(TOPLEVEL)"\n" " l32i %0, %1, 0\n" " bne %0, %2, 1f\n" " s32i %3, %1, 0\n" @@ -123,7 +123,7 @@ static inline unsigned long xchg_u32(volatile int * m, unsigned long val) #else unsigned long tmp; __asm__ __volatile__( - " rsil a15, "__stringify(LOCKLEVEL)"\n" + " rsil a15, "__stringify(TOPLEVEL)"\n" " l32i %0, %1, 0\n" " s32i %2, %1, 0\n" " wsr a15, ps\n" diff --git a/arch/xtensa/include/asm/irqflags.h b/arch/xtensa/include/asm/irqflags.h index ea36674c6ec5..8e090c709046 100644 --- a/arch/xtensa/include/asm/irqflags.h +++ b/arch/xtensa/include/asm/irqflags.h @@ -6,6 +6,7 @@ * for more details. * * Copyright (C) 2001 - 2005 Tensilica Inc. + * Copyright (C) 2015 Cadence Design Systems Inc. */ #ifndef _XTENSA_IRQFLAGS_H @@ -23,8 +24,27 @@ static inline unsigned long arch_local_save_flags(void) static inline unsigned long arch_local_irq_save(void) { unsigned long flags; - asm volatile("rsil %0, "__stringify(LOCKLEVEL) +#if XTENSA_FAKE_NMI +#if defined(CONFIG_DEBUG_KERNEL) && (LOCKLEVEL | TOPLEVEL) >= XCHAL_DEBUGLEVEL + unsigned long tmp; + + asm volatile("rsr %0, ps\t\n" + "extui %1, %0, 0, 4\t\n" + "bgei %1, "__stringify(LOCKLEVEL)", 1f\t\n" + "rsil %0, "__stringify(LOCKLEVEL)"\n" + "1:" + : "=a" (flags), "=a" (tmp) :: "memory"); +#else + asm volatile("rsr %0, ps\t\n" + "or %0, %0, %1\t\n" + "xsr %0, ps\t\n" + "rsync" + : "=&a" (flags) : "a" (LOCKLEVEL) : "memory"); +#endif +#else + asm volatile("rsil %0, "__stringify(LOCKLEVEL) : "=a" (flags) :: "memory"); +#endif return flags; } diff --git a/arch/xtensa/include/asm/processor.h b/arch/xtensa/include/asm/processor.h index b61bdf0eea25..83e2e4bc01ba 100644 --- a/arch/xtensa/include/asm/processor.h +++ b/arch/xtensa/include/asm/processor.h @@ -1,11 +1,10 @@ /* - * include/asm-xtensa/processor.h - * * This file is subject to the terms and conditions of the GNU General Public * License. See the file "COPYING" in the main directory of this archive * for more details. * * Copyright (C) 2001 - 2008 Tensilica Inc. + * Copyright (C) 2015 Cadence Design Systems Inc. */ #ifndef _XTENSA_PROCESSOR_H @@ -44,6 +43,14 @@ #define STACK_TOP TASK_SIZE #define STACK_TOP_MAX STACK_TOP +/* + * General exception cause assigned to fake NMI. Fake NMI needs to be handled + * differently from other interrupts, but it uses common kernel entry/exit + * code. + */ + +#define EXCCAUSE_MAPPED_NMI 62 + /* * General exception cause assigned to debug exceptions. Debug exceptions go * to their own vector, rather than the general exception vectors (user, @@ -65,10 +72,30 @@ #define VALID_DOUBLE_EXCEPTION_ADDRESS 64 +#define XTENSA_INT_LEVEL(intno) _XTENSA_INT_LEVEL(intno) +#define _XTENSA_INT_LEVEL(intno) XCHAL_INT##intno##_LEVEL + +#define XTENSA_INTLEVEL_MASK(level) _XTENSA_INTLEVEL_MASK(level) +#define _XTENSA_INTLEVEL_MASK(level) (XCHAL_INTLEVEL##level##_MASK) + +#define IS_POW2(v) (((v) & ((v) - 1)) == 0) + +#define PROFILING_INTLEVEL XTENSA_INT_LEVEL(XCHAL_PROFILING_INTERRUPT) + /* LOCKLEVEL defines the interrupt level that masks all * general-purpose interrupts. */ +#if defined(CONFIG_XTENSA_VARIANT_HAVE_PERF_EVENTS) && \ + defined(XCHAL_PROFILING_INTERRUPT) && \ + PROFILING_INTLEVEL == XCHAL_EXCM_LEVEL && \ + XCHAL_EXCM_LEVEL > 1 && \ + IS_POW2(XTENSA_INTLEVEL_MASK(PROFILING_INTLEVEL)) +#define LOCKLEVEL (XCHAL_EXCM_LEVEL - 1) +#else #define LOCKLEVEL XCHAL_EXCM_LEVEL +#endif +#define TOPLEVEL XCHAL_EXCM_LEVEL +#define XTENSA_FAKE_NMI (LOCKLEVEL < TOPLEVEL) /* WSBITS and WBBITS are the width of the WINDOWSTART and WINDOWBASE * registers diff --git a/arch/xtensa/kernel/entry.S b/arch/xtensa/kernel/entry.S index 5703bc13a9ba..504130357597 100644 --- a/arch/xtensa/kernel/entry.S +++ b/arch/xtensa/kernel/entry.S @@ -1,6 +1,4 @@ /* - * arch/xtensa/kernel/entry.S - * * Low-level exception handling * * This file is subject to the terms and conditions of the GNU General Public @@ -8,6 +6,7 @@ * for more details. * * Copyright (C) 2004 - 2008 by Tensilica Inc. + * Copyright (C) 2015 Cadence Design Systems Inc. * * Chris Zankel * @@ -75,6 +74,27 @@ #endif .endm + + .macro irq_save flags tmp +#if XTENSA_FAKE_NMI +#if defined(CONFIG_DEBUG_KERNEL) && (LOCKLEVEL | TOPLEVEL) >= XCHAL_DEBUGLEVEL + rsr \flags, ps + extui \tmp, \flags, PS_INTLEVEL_SHIFT, PS_INTLEVEL_WIDTH + bgei \tmp, LOCKLEVEL, 99f + rsil \tmp, LOCKLEVEL +99: +#else + movi \tmp, LOCKLEVEL + rsr \flags, ps + or \flags, \flags, \tmp + xsr \flags, ps + rsync +#endif +#else + rsil \flags, LOCKLEVEL +#endif + .endm + /* ----------------- DEFAULT FIRST LEVEL EXCEPTION HANDLERS ----------------- */ /* @@ -352,11 +372,11 @@ common_exception: /* It is now save to restore the EXC_TABLE_FIXUP variable. */ - rsr a0, exccause + rsr a2, exccause movi a3, 0 - rsr a2, excsave1 - s32i a0, a1, PT_EXCCAUSE - s32i a3, a2, EXC_TABLE_FIXUP + rsr a0, excsave1 + s32i a2, a1, PT_EXCCAUSE + s32i a3, a0, EXC_TABLE_FIXUP /* All unrecoverable states are saved on stack, now, and a1 is valid. * Now we can allow exceptions again. In case we've got an interrupt @@ -367,19 +387,46 @@ common_exception: */ rsr a3, ps - addi a0, a0, -EXCCAUSE_LEVEL1_INTERRUPT - movi a2, LOCKLEVEL + s32i a3, a1, PT_PS # save ps + +#if XTENSA_FAKE_NMI + /* Correct PS needs to be saved in the PT_PS: + * - in case of exception or level-1 interrupt it's in the PS, + * and is already saved. + * - in case of medium level interrupt it's in the excsave2. + */ + movi a0, EXCCAUSE_MAPPED_NMI + extui a3, a3, PS_INTLEVEL_SHIFT, PS_INTLEVEL_WIDTH + beq a2, a0, .Lmedium_level_irq + bnei a2, EXCCAUSE_LEVEL1_INTERRUPT, .Lexception + beqz a3, .Llevel1_irq # level-1 IRQ sets ps.intlevel to 0 + +.Lmedium_level_irq: + rsr a0, excsave2 + s32i a0, a1, PT_PS # save medium-level interrupt ps + bgei a3, LOCKLEVEL, .Lexception + +.Llevel1_irq: + movi a3, LOCKLEVEL + +.Lexception: + movi a0, 1 << PS_WOE_BIT + or a3, a3, a0 +#else + addi a2, a2, -EXCCAUSE_LEVEL1_INTERRUPT + movi a0, LOCKLEVEL extui a3, a3, PS_INTLEVEL_SHIFT, PS_INTLEVEL_WIDTH # a3 = PS.INTLEVEL - moveqz a3, a2, a0 # a3 = LOCKLEVEL iff interrupt + moveqz a3, a0, a2 # a3 = LOCKLEVEL iff interrupt movi a2, 1 << PS_WOE_BIT or a3, a3, a2 rsr a2, exccause +#endif + /* restore return address (or 0 if return to userspace) */ rsr a0, depc - xsr a3, ps - - s32i a3, a1, PT_PS # save ps + wsr a3, ps + rsync # PS.WOE => rsync => overflow /* Save lbeg, lend */ @@ -417,8 +464,13 @@ common_exception: .global common_exception_return common_exception_return: +#if XTENSA_FAKE_NMI + l32i a2, a1, PT_EXCCAUSE + movi a3, EXCCAUSE_MAPPED_NMI + beq a2, a3, .LNMIexit +#endif 1: - rsil a2, LOCKLEVEL + irq_save a2, a3 #ifdef CONFIG_TRACE_IRQFLAGS movi a4, trace_hardirqs_off callx4 a4 @@ -481,6 +533,12 @@ common_exception_return: j 1b #endif +#if XTENSA_FAKE_NMI +.LNMIexit: + l32i a3, a1, PT_PS + _bbci.l a3, PS_UM_BIT, 4f +#endif + 5: #ifdef CONFIG_DEBUG_TLB_SANITY l32i a4, a1, PT_DEPC @@ -1564,6 +1622,13 @@ ENTRY(fast_second_level_miss) rfde 9: l32i a0, a1, TASK_ACTIVE_MM # unlikely case mm == 0 + bnez a0, 8b + + /* Even more unlikely case active_mm == 0. + * We can get here with NMI in the middle of context_switch that + * touches vmalloc area. + */ + movi a0, init_mm j 8b #if (DCACHE_WAY_SIZE > PAGE_SIZE) @@ -1867,7 +1932,7 @@ ENTRY(_switch_to) /* Disable ints while we manipulate the stack pointer. */ - rsil a14, LOCKLEVEL + irq_save a14, a3 rsync /* Switch CPENABLE */ diff --git a/arch/xtensa/kernel/irq.c b/arch/xtensa/kernel/irq.c index 32b60568ee45..91ba60563021 100644 --- a/arch/xtensa/kernel/irq.c +++ b/arch/xtensa/kernel/irq.c @@ -29,6 +29,7 @@ #include atomic_t irq_err_count; +DECLARE_PER_CPU(unsigned long, nmi_count); asmlinkage void do_IRQ(int hwirq, struct pt_regs *regs) { @@ -57,11 +58,18 @@ asmlinkage void do_IRQ(int hwirq, struct pt_regs *regs) int arch_show_interrupts(struct seq_file *p, int prec) { + unsigned cpu __maybe_unused; #ifdef CONFIG_SMP show_ipi_list(p, prec); #endif seq_printf(p, "%*s: ", prec, "ERR"); seq_printf(p, "%10u\n", atomic_read(&irq_err_count)); +#if XTENSA_FAKE_NMI + seq_printf(p, "%*s:", prec, "NMI"); + for_each_online_cpu(cpu) + seq_printf(p, " %10lu", per_cpu(nmi_count, cpu)); + seq_puts(p, " Non-maskable interrupts\n"); +#endif return 0; } diff --git a/arch/xtensa/kernel/perf_event.c b/arch/xtensa/kernel/perf_event.c index b44df3c8198f..54f01188c29c 100644 --- a/arch/xtensa/kernel/perf_event.c +++ b/arch/xtensa/kernel/perf_event.c @@ -359,7 +359,7 @@ void perf_event_print_debug(void) local_irq_restore(flags); } -static irqreturn_t xtensa_pmu_irq_handler(int irq, void *dev_id) +irqreturn_t xtensa_pmu_irq_handler(int irq, void *dev_id) { irqreturn_t rc = IRQ_NONE; struct xtensa_pmu_events *ev = this_cpu_ptr(&xtensa_pmu_events); @@ -436,10 +436,14 @@ static int __init xtensa_pmu_init(void) int irq = irq_create_mapping(NULL, XCHAL_PROFILING_INTERRUPT); perf_cpu_notifier(xtensa_pmu_notifier); +#if XTENSA_FAKE_NMI + enable_irq(irq); +#else ret = request_irq(irq, xtensa_pmu_irq_handler, IRQF_PERCPU, "pmu", NULL); if (ret < 0) return ret; +#endif ret = perf_pmu_register(&xtensa_pmu, "cpu", PERF_TYPE_RAW); if (ret) diff --git a/arch/xtensa/kernel/traps.c b/arch/xtensa/kernel/traps.c index a1b5bd237c71..42d441f7898b 100644 --- a/arch/xtensa/kernel/traps.c +++ b/arch/xtensa/kernel/traps.c @@ -62,6 +62,7 @@ extern void fast_coprocessor(void); extern void do_illegal_instruction (struct pt_regs*); extern void do_interrupt (struct pt_regs*); +extern void do_nmi(struct pt_regs *); extern void do_unaligned_user (struct pt_regs*); extern void do_multihit (struct pt_regs*, unsigned long); extern void do_page_fault (struct pt_regs*, unsigned long); @@ -146,6 +147,9 @@ COPROCESSOR(6), #if XTENSA_HAVE_COPROCESSOR(7) COPROCESSOR(7), #endif +#if XTENSA_FAKE_NMI +{ EXCCAUSE_MAPPED_NMI, 0, do_nmi }, +#endif { EXCCAUSE_MAPPED_DEBUG, 0, do_debug }, { -1, -1, 0 } @@ -199,6 +203,28 @@ void do_multihit(struct pt_regs *regs, unsigned long exccause) extern void do_IRQ(int, struct pt_regs *); +#if XTENSA_FAKE_NMI + +irqreturn_t xtensa_pmu_irq_handler(int irq, void *dev_id); + +DEFINE_PER_CPU(unsigned long, nmi_count); + +void do_nmi(struct pt_regs *regs) +{ + struct pt_regs *old_regs; + + if ((regs->ps & PS_INTLEVEL_MASK) < LOCKLEVEL) + trace_hardirqs_off(); + + old_regs = set_irq_regs(regs); + nmi_enter(); + ++*this_cpu_ptr(&nmi_count); + xtensa_pmu_irq_handler(0, NULL); + nmi_exit(); + set_irq_regs(old_regs); +} +#endif + void do_interrupt(struct pt_regs *regs) { static const unsigned int_level_mask[] = { diff --git a/arch/xtensa/kernel/vectors.S b/arch/xtensa/kernel/vectors.S index 1b397a902292..abcdb527f18a 100644 --- a/arch/xtensa/kernel/vectors.S +++ b/arch/xtensa/kernel/vectors.S @@ -627,7 +627,11 @@ ENTRY(_Level\level\()InterruptVector) wsr a0, excsave2 rsr a0, epc\level wsr a0, epc1 + .if \level <= LOCKLEVEL movi a0, EXCCAUSE_LEVEL1_INTERRUPT + .else + movi a0, EXCCAUSE_MAPPED_NMI + .endif wsr a0, exccause rsr a0, eps\level # branch to user or kernel vector @@ -682,11 +686,13 @@ ENDPROC(_WindowOverflow4) .align 4 _SimulateUserKernelVectorException: addi a0, a0, (1 << PS_EXCM_BIT) +#if !XTENSA_FAKE_NMI wsr a0, ps +#endif bbsi.l a0, PS_UM_BIT, 1f # branch if user mode - rsr a0, excsave2 # restore a0 + xsr a0, excsave2 # restore a0 j _KernelExceptionVector # simulate kernel vector exception -1: rsr a0, excsave2 # restore a0 +1: xsr a0, excsave2 # restore a0 j _UserExceptionVector # simulate user vector exception #endif -- cgit v1.2.3