From 8e8339a3a1069141985daaa2521ba304509ddecd Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Wed, 11 Dec 2013 11:09:53 +0100 Subject: [PATCH 1/4] sched: Initialize power_orig for overlapping groups Yinghai reported that he saw a /0 in sg_capacity on his EX parts. Make sure to always initialize power_orig now that we actually use it. Ideally build_sched_domains() -> init_sched_groups_power() would also initialize this; but for some yet unexplained reason some setups seem to miss updates there. Reported-by: Yinghai Lu Tested-by: Yinghai Lu Signed-off-by: Peter Zijlstra Link: http://lkml.kernel.org/n/tip-l8ng2m9uml6fhibln8wqpom7@git.kernel.org Signed-off-by: Ingo Molnar --- kernel/sched/core.c | 1 + 1 file changed, 1 insertion(+) diff --git a/kernel/sched/core.c b/kernel/sched/core.c index e85cda20ab2b..19af58f3a261 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -5112,6 +5112,7 @@ build_overlap_sched_groups(struct sched_domain *sd, int cpu) * die on a /0 trap. */ sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); + sg->sgp->power_orig = sg->sgp->power; /* * Make sure the first group of this domain contains the From ba1f14fbe70965ae0fb1655a5275a62723f65b77 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Thu, 28 Nov 2013 14:26:41 +0100 Subject: [PATCH 2/4] sched: Remove PREEMPT_NEED_RESCHED from generic code While hunting a preemption issue with Alexander, Ben noticed that the currently generic PREEMPT_NEED_RESCHED stuff is horribly broken for load-store architectures. We currently rely on the IPI to fold TIF_NEED_RESCHED into PREEMPT_NEED_RESCHED, but when this IPI lands while we already have a load for the preempt-count but before the store, the store will erase the PREEMPT_NEED_RESCHED change. The current preempt-count only works on load-store archs because interrupts are assumed to be completely balanced wrt their preempt_count fiddling; the previous preempt_count load will match the preempt_count state after the interrupt and therefore nothing gets lost. This patch removes the PREEMPT_NEED_RESCHED usage from generic code and pushes it into x86 arch code; the generic code goes back to relying on TIF_NEED_RESCHED. Boot tested on x86_64 and compile tested on ppc64. Reported-by: Benjamin Herrenschmidt Reported-and-Tested-by: Alexander Graf Signed-off-by: Peter Zijlstra Cc: Linus Torvalds Link: http://lkml.kernel.org/r/20131128132641.GP10022@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar --- arch/x86/include/asm/preempt.h | 11 +++++++++++ include/asm-generic/preempt.h | 35 +++++++++++----------------------- include/linux/sched.h | 2 -- 3 files changed, 22 insertions(+), 26 deletions(-) diff --git a/arch/x86/include/asm/preempt.h b/arch/x86/include/asm/preempt.h index 8729723636fd..c8b051933b1b 100644 --- a/arch/x86/include/asm/preempt.h +++ b/arch/x86/include/asm/preempt.h @@ -7,6 +7,12 @@ DECLARE_PER_CPU(int, __preempt_count); +/* + * We use the PREEMPT_NEED_RESCHED bit as an inverted NEED_RESCHED such + * that a decrement hitting 0 means we can and should reschedule. + */ +#define PREEMPT_ENABLED (0 + PREEMPT_NEED_RESCHED) + /* * We mask the PREEMPT_NEED_RESCHED bit so as not to confuse all current users * that think a non-zero value indicates we cannot preempt. @@ -74,6 +80,11 @@ static __always_inline void __preempt_count_sub(int val) __this_cpu_add_4(__preempt_count, -val); } +/* + * Because we keep PREEMPT_NEED_RESCHED set when we do _not_ need to reschedule + * a decrement which hits zero means we have no preempt_count and should + * reschedule. + */ static __always_inline bool __preempt_count_dec_and_test(void) { GEN_UNARY_RMWcc("decl", __preempt_count, __percpu_arg(0), "e"); diff --git a/include/asm-generic/preempt.h b/include/asm-generic/preempt.h index ddf2b420ac8f..1cd3f5d767a8 100644 --- a/include/asm-generic/preempt.h +++ b/include/asm-generic/preempt.h @@ -3,13 +3,11 @@ #include -/* - * We mask the PREEMPT_NEED_RESCHED bit so as not to confuse all current users - * that think a non-zero value indicates we cannot preempt. - */ +#define PREEMPT_ENABLED (0) + static __always_inline int preempt_count(void) { - return current_thread_info()->preempt_count & ~PREEMPT_NEED_RESCHED; + return current_thread_info()->preempt_count; } static __always_inline int *preempt_count_ptr(void) @@ -17,11 +15,6 @@ static __always_inline int *preempt_count_ptr(void) return ¤t_thread_info()->preempt_count; } -/* - * We now loose PREEMPT_NEED_RESCHED and cause an extra reschedule; however the - * alternative is loosing a reschedule. Better schedule too often -- also this - * should be a very rare operation. - */ static __always_inline void preempt_count_set(int pc) { *preempt_count_ptr() = pc; @@ -41,28 +34,17 @@ static __always_inline void preempt_count_set(int pc) task_thread_info(p)->preempt_count = PREEMPT_ENABLED; \ } while (0) -/* - * We fold the NEED_RESCHED bit into the preempt count such that - * preempt_enable() can decrement and test for needing to reschedule with a - * single instruction. - * - * We invert the actual bit, so that when the decrement hits 0 we know we both - * need to resched (the bit is cleared) and can resched (no preempt count). - */ - static __always_inline void set_preempt_need_resched(void) { - *preempt_count_ptr() &= ~PREEMPT_NEED_RESCHED; } static __always_inline void clear_preempt_need_resched(void) { - *preempt_count_ptr() |= PREEMPT_NEED_RESCHED; } static __always_inline bool test_preempt_need_resched(void) { - return !(*preempt_count_ptr() & PREEMPT_NEED_RESCHED); + return false; } /* @@ -81,7 +63,12 @@ static __always_inline void __preempt_count_sub(int val) static __always_inline bool __preempt_count_dec_and_test(void) { - return !--*preempt_count_ptr(); + /* + * Because of load-store architectures cannot do per-cpu atomic + * operations; we cannot use PREEMPT_NEED_RESCHED because it might get + * lost. + */ + return !--*preempt_count_ptr() && tif_need_resched(); } /* @@ -89,7 +76,7 @@ static __always_inline bool __preempt_count_dec_and_test(void) */ static __always_inline bool should_resched(void) { - return unlikely(!*preempt_count_ptr()); + return unlikely(!preempt_count() && tif_need_resched()); } #ifdef CONFIG_PREEMPT diff --git a/include/linux/sched.h b/include/linux/sched.h index 768b037dfacb..96d674ba3876 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -440,8 +440,6 @@ struct task_cputime { .sum_exec_runtime = 0, \ } -#define PREEMPT_ENABLED (PREEMPT_NEED_RESCHED) - #ifdef CONFIG_PREEMPT_COUNT #define PREEMPT_DISABLED (1 + PREEMPT_ENABLED) #else From be5e610c0fd6ef772cafb9e0bd4128134804aef3 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Mon, 18 Nov 2013 18:27:06 +0100 Subject: [PATCH 3/4] math64: Add mul_u64_u32_shr() Introduce mul_u64_u32_shr() as proposed by Andy a while back; it allows using 64x64->128 muls on 64bit archs and recent GCC which defines __SIZEOF_INT128__ and __int128. (This new method will be used by the scheduler.) Signed-off-by: Peter Zijlstra Cc: fweisbec@gmail.com Cc: Andy Lutomirski Cc: Linus Torvalds Link: http://lkml.kernel.org/n/tip-hxjoeuzmrcaumR0uZwjpe2pv@git.kernel.org Signed-off-by: Ingo Molnar --- arch/x86/Kconfig | 1 + include/linux/math64.h | 30 ++++++++++++++++++++++++++++++ init/Kconfig | 6 ++++++ 3 files changed, 37 insertions(+) diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig index e903c71f7e69..0952ecd60eca 100644 --- a/arch/x86/Kconfig +++ b/arch/x86/Kconfig @@ -26,6 +26,7 @@ config X86 select HAVE_AOUT if X86_32 select HAVE_UNSTABLE_SCHED_CLOCK select ARCH_SUPPORTS_NUMA_BALANCING + select ARCH_SUPPORTS_INT128 if X86_64 select ARCH_WANTS_PROT_NUMA_PROT_NONE select HAVE_IDE select HAVE_OPROFILE diff --git a/include/linux/math64.h b/include/linux/math64.h index 69ed5f5e9f6e..c45c089bfdac 100644 --- a/include/linux/math64.h +++ b/include/linux/math64.h @@ -133,4 +133,34 @@ __iter_div_u64_rem(u64 dividend, u32 divisor, u64 *remainder) return ret; } +#if defined(CONFIG_ARCH_SUPPORTS_INT128) && defined(__SIZEOF_INT128__) + +#ifndef mul_u64_u32_shr +static inline u64 mul_u64_u32_shr(u64 a, u32 mul, unsigned int shift) +{ + return (u64)(((unsigned __int128)a * mul) >> shift); +} +#endif /* mul_u64_u32_shr */ + +#else + +#ifndef mul_u64_u32_shr +static inline u64 mul_u64_u32_shr(u64 a, u32 mul, unsigned int shift) +{ + u32 ah, al; + u64 ret; + + al = a; + ah = a >> 32; + + ret = ((u64)al * mul) >> shift; + if (ah) + ret += ((u64)ah * mul) << (32 - shift); + + return ret; +} +#endif /* mul_u64_u32_shr */ + +#endif + #endif /* _LINUX_MATH64_H */ diff --git a/init/Kconfig b/init/Kconfig index 79383d3aa5dc..4e5d96ab2034 100644 --- a/init/Kconfig +++ b/init/Kconfig @@ -809,6 +809,12 @@ config GENERIC_SCHED_CLOCK config ARCH_SUPPORTS_NUMA_BALANCING bool +# +# For architectures that know their GCC __int128 support is sound +# +config ARCH_SUPPORTS_INT128 + bool + # For architectures that (ab)use NUMA to represent different memory regions # all cpu-local but of different latencies, such as SuperH. # From 9dbdb155532395ba000c5d5d187658b0e17e529f Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Mon, 18 Nov 2013 18:27:06 +0100 Subject: [PATCH 4/4] sched/fair: Rework sched_fair time accounting Christian suffers from a bad BIOS that wrecks his i5's TSC sync. This results in him occasionally seeing time going backwards - which crashes the scheduler ... Most of our time accounting can actually handle that except the most common one; the tick time update of sched_fair. There is a further problem with that code; previously we assumed that because we get a tick every TICK_NSEC our time delta could never exceed 32bits and math was simpler. However, ever since Frederic managed to get NO_HZ_FULL merged; this is no longer the case since now a task can run for a long time indeed without getting a tick. It only takes about ~4.2 seconds to overflow our u32 in nanoseconds. This means we not only need to better deal with time going backwards; but also means we need to be able to deal with large deltas. This patch reworks the entire code and uses mul_u64_u32_shr() as proposed by Andy a long while ago. We express our virtual time scale factor in a u32 multiplier and shift right and the 32bit mul_u64_u32_shr() implementation reduces to a single 32x32->64 multiply if the time delta is still short (common case). For 64bit a 64x64->128 multiply can be used if ARCH_SUPPORTS_INT128. Reported-and-Tested-by: Christian Engelmayer Signed-off-by: Peter Zijlstra Cc: fweisbec@gmail.com Cc: Paul Turner Cc: Stanislaw Gruszka Cc: Andy Lutomirski Cc: Linus Torvalds Cc: Andrew Morton Link: http://lkml.kernel.org/r/20131118172706.GI3866@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar --- include/linux/sched.h | 3 +- kernel/sched/fair.c | 146 +++++++++++++++++++----------------------- 2 files changed, 67 insertions(+), 82 deletions(-) diff --git a/include/linux/sched.h b/include/linux/sched.h index 96d674ba3876..53f97eb8dbc7 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -930,7 +930,8 @@ struct pipe_inode_info; struct uts_namespace; struct load_weight { - unsigned long weight, inv_weight; + unsigned long weight; + u32 inv_weight; }; struct sched_avg { diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index fd773ade1a31..9030da7bcb15 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -178,59 +178,61 @@ void sched_init_granularity(void) update_sysctl(); } -#if BITS_PER_LONG == 32 -# define WMULT_CONST (~0UL) -#else -# define WMULT_CONST (1UL << 32) -#endif - +#define WMULT_CONST (~0U) #define WMULT_SHIFT 32 -/* - * Shift right and round: - */ -#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) - -/* - * delta *= weight / lw - */ -static unsigned long -calc_delta_mine(unsigned long delta_exec, unsigned long weight, - struct load_weight *lw) +static void __update_inv_weight(struct load_weight *lw) { - u64 tmp; + unsigned long w; - /* - * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched - * entities since MIN_SHARES = 2. Treat weight as 1 if less than - * 2^SCHED_LOAD_RESOLUTION. - */ - if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION))) - tmp = (u64)delta_exec * scale_load_down(weight); + if (likely(lw->inv_weight)) + return; + + w = scale_load_down(lw->weight); + + if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) + lw->inv_weight = 1; + else if (unlikely(!w)) + lw->inv_weight = WMULT_CONST; else - tmp = (u64)delta_exec; + lw->inv_weight = WMULT_CONST / w; +} - if (!lw->inv_weight) { - unsigned long w = scale_load_down(lw->weight); +/* + * delta_exec * weight / lw.weight + * OR + * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT + * + * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case + * we're guaranteed shift stays positive because inv_weight is guaranteed to + * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. + * + * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus + * weight/lw.weight <= 1, and therefore our shift will also be positive. + */ +static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) +{ + u64 fact = scale_load_down(weight); + int shift = WMULT_SHIFT; - if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) - lw->inv_weight = 1; - else if (unlikely(!w)) - lw->inv_weight = WMULT_CONST; - else - lw->inv_weight = WMULT_CONST / w; + __update_inv_weight(lw); + + if (unlikely(fact >> 32)) { + while (fact >> 32) { + fact >>= 1; + shift--; + } } - /* - * Check whether we'd overflow the 64-bit multiplication: - */ - if (unlikely(tmp > WMULT_CONST)) - tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, - WMULT_SHIFT/2); - else - tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); + /* hint to use a 32x32->64 mul */ + fact = (u64)(u32)fact * lw->inv_weight; - return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); + while (fact >> 32) { + fact >>= 1; + shift--; + } + + return mul_u64_u32_shr(delta_exec, fact, shift); } @@ -443,7 +445,7 @@ find_matching_se(struct sched_entity **se, struct sched_entity **pse) #endif /* CONFIG_FAIR_GROUP_SCHED */ static __always_inline -void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec); +void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); /************************************************************** * Scheduling class tree data structure manipulation methods: @@ -612,11 +614,10 @@ int sched_proc_update_handler(struct ctl_table *table, int write, /* * delta /= w */ -static inline unsigned long -calc_delta_fair(unsigned long delta, struct sched_entity *se) +static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) { if (unlikely(se->load.weight != NICE_0_LOAD)) - delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); + delta = __calc_delta(delta, NICE_0_LOAD, &se->load); return delta; } @@ -665,7 +666,7 @@ static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) update_load_add(&lw, se->load.weight); load = &lw; } - slice = calc_delta_mine(slice, se->load.weight, load); + slice = __calc_delta(slice, se->load.weight, load); } return slice; } @@ -703,47 +704,32 @@ void init_task_runnable_average(struct task_struct *p) #endif /* - * Update the current task's runtime statistics. Skip current tasks that - * are not in our scheduling class. + * Update the current task's runtime statistics. */ -static inline void -__update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, - unsigned long delta_exec) -{ - unsigned long delta_exec_weighted; - - schedstat_set(curr->statistics.exec_max, - max((u64)delta_exec, curr->statistics.exec_max)); - - curr->sum_exec_runtime += delta_exec; - schedstat_add(cfs_rq, exec_clock, delta_exec); - delta_exec_weighted = calc_delta_fair(delta_exec, curr); - - curr->vruntime += delta_exec_weighted; - update_min_vruntime(cfs_rq); -} - static void update_curr(struct cfs_rq *cfs_rq) { struct sched_entity *curr = cfs_rq->curr; u64 now = rq_clock_task(rq_of(cfs_rq)); - unsigned long delta_exec; + u64 delta_exec; if (unlikely(!curr)) return; - /* - * Get the amount of time the current task was running - * since the last time we changed load (this cannot - * overflow on 32 bits): - */ - delta_exec = (unsigned long)(now - curr->exec_start); - if (!delta_exec) + delta_exec = now - curr->exec_start; + if (unlikely((s64)delta_exec <= 0)) return; - __update_curr(cfs_rq, curr, delta_exec); curr->exec_start = now; + schedstat_set(curr->statistics.exec_max, + max(delta_exec, curr->statistics.exec_max)); + + curr->sum_exec_runtime += delta_exec; + schedstat_add(cfs_rq, exec_clock, delta_exec); + + curr->vruntime += calc_delta_fair(delta_exec, curr); + update_min_vruntime(cfs_rq); + if (entity_is_task(curr)) { struct task_struct *curtask = task_of(curr); @@ -3015,8 +3001,7 @@ static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) } } -static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, - unsigned long delta_exec) +static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) { /* dock delta_exec before expiring quota (as it could span periods) */ cfs_rq->runtime_remaining -= delta_exec; @@ -3034,7 +3019,7 @@ static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, } static __always_inline -void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) +void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) { if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) return; @@ -3574,8 +3559,7 @@ static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) return rq_clock_task(rq_of(cfs_rq)); } -static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, - unsigned long delta_exec) {} +static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}