From 16d9dbf0c2bd167fdd942b83592d59696c7b73bd Mon Sep 17 00:00:00 2001 From: Jeremy Fitzhardinge Date: Fri, 24 Apr 2009 00:50:27 -0700 Subject: x86-64: move unlazy_fpu() into lazy cpu state part of context switch Make sure that unlazy_fpu()'s stts gets batched along with the other cpu state changes during context switch. (32-bit already does this.) This makes sure it gets batched when running paravirtualized. [ Impact: optimise paravirtual FPU context switch ] Signed-off-by: Jeremy Fitzhardinge Cc: Alok Kataria Cc: Rusty Russell --- arch/x86/kernel/process_64.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'arch/x86/kernel/process_64.c') diff --git a/arch/x86/kernel/process_64.c b/arch/x86/kernel/process_64.c index ebefb5407b9d..c9b8904736db 100644 --- a/arch/x86/kernel/process_64.c +++ b/arch/x86/kernel/process_64.c @@ -419,6 +419,9 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p) load_TLS(next, cpu); + /* Must be after DS reload */ + unlazy_fpu(prev_p); + /* * Leave lazy mode, flushing any hypercalls made here. * This must be done before restoring TLS segments so @@ -459,9 +462,6 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p) wrmsrl(MSR_KERNEL_GS_BASE, next->gs); prev->gsindex = gsindex; - /* Must be after DS reload */ - unlazy_fpu(prev_p); - /* * Switch the PDA and FPU contexts. */ -- cgit v1.2.3 From 17950c5b243f99cbabef173415ee988c52104d7e Mon Sep 17 00:00:00 2001 From: Jeremy Fitzhardinge Date: Fri, 24 Apr 2009 01:01:01 -0700 Subject: x86-64: move clts into batch cpu state updates when preloading fpu When a task is likely to be using the fpu, we preload its state during the context switch, rather than waiting for it to run an fpu instruction. Make sure the clts() happens while we're doing batched fpu state updates to optimise paravirtualized context switches. [ Impact: optimise paravirtual FPU context switch ] Signed-off-by: Jeremy Fitzhardinge Cc: Alok Kataria Cc: Rusty Russell --- arch/x86/kernel/process_64.c | 27 ++++++++++++++++++--------- 1 file changed, 18 insertions(+), 9 deletions(-) (limited to 'arch/x86/kernel/process_64.c') diff --git a/arch/x86/kernel/process_64.c b/arch/x86/kernel/process_64.c index c9b8904736db..a28279dbb07c 100644 --- a/arch/x86/kernel/process_64.c +++ b/arch/x86/kernel/process_64.c @@ -386,9 +386,17 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p) int cpu = smp_processor_id(); struct tss_struct *tss = &per_cpu(init_tss, cpu); unsigned fsindex, gsindex; + bool preload_fpu; + + /* + * If the task has used fpu the last 5 timeslices, just do a full + * restore of the math state immediately to avoid the trap; the + * chances of needing FPU soon are obviously high now + */ + preload_fpu = tsk_used_math(next_p) && next_p->fpu_counter > 5; /* we're going to use this soon, after a few expensive things */ - if (next_p->fpu_counter > 5) + if (preload_fpu) prefetch(next->xstate); /* @@ -422,6 +430,10 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p) /* Must be after DS reload */ unlazy_fpu(prev_p); + /* Make sure cpu is ready for new context */ + if (preload_fpu) + clts(); + /* * Leave lazy mode, flushing any hypercalls made here. * This must be done before restoring TLS segments so @@ -480,15 +492,12 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p) task_thread_info(prev_p)->flags & _TIF_WORK_CTXSW_PREV)) __switch_to_xtra(prev_p, next_p, tss); - /* If the task has used fpu the last 5 timeslices, just do a full - * restore of the math state immediately to avoid the trap; the - * chances of needing FPU soon are obviously high now - * - * tsk_used_math() checks prevent calling math_state_restore(), - * which can sleep in the case of !tsk_used_math() + /* + * Preload the FPU context, now that we've determined that the + * task is likely to be using it. */ - if (tsk_used_math(next_p) && next_p->fpu_counter > 5) - math_state_restore(); + if (preload_fpu) + __math_state_restore(); return prev_p; } -- cgit v1.2.3 From bdf977b37418cdf8a2252504779a7e12a09b7575 Mon Sep 17 00:00:00 2001 From: Tejun Heo Date: Mon, 3 Aug 2009 14:12:19 +0900 Subject: x86, percpu: Collect hot percpu variables into one cacheline On x86_64, percpu variables current_task and kernel_stack are used for get_current() and current_thread_info() respectively and thus are often used close to each other. Move definition of current_task to kernel/cpu/common.c right above kernel_stack definition and align it to cacheline so that they always fall into the same cacheline. Two percpu variables defined there together - irq_stack_ptr and irq_count - are also pretty hot and will benefit from sharing the cacheline. For consistency, current_task definition for x86_32 is also moved to kernel/cpu/common.c. Putting current_task and kernel_stack into the same cacheline was suggested by Linus Torvalds. Signed-off-by: Tejun Heo Cc: Linus Torvalds Cc: Ingo Molnar Signed-off-by: H. Peter Anvin --- arch/x86/kernel/cpu/common.c | 15 +++++++++++++-- arch/x86/kernel/process_32.c | 3 --- arch/x86/kernel/process_64.c | 3 --- 3 files changed, 13 insertions(+), 8 deletions(-) (limited to 'arch/x86/kernel/process_64.c') diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c index 12493c5485e5..1bd88ed978bd 100644 --- a/arch/x86/kernel/cpu/common.c +++ b/arch/x86/kernel/cpu/common.c @@ -987,13 +987,21 @@ struct desc_ptr idt_descr = { 256 * 16 - 1, (unsigned long) idt_table }; DEFINE_PER_CPU_FIRST(union irq_stack_union, irq_stack_union) __aligned(PAGE_SIZE); -DEFINE_PER_CPU(char *, irq_stack_ptr) = - init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE - 64; +/* + * The following four percpu variables are hot. Align current_task to + * cacheline size such that all four fall in the same cacheline. + */ +DEFINE_PER_CPU(struct task_struct *, current_task) ____cacheline_aligned = + &init_task; +EXPORT_PER_CPU_SYMBOL(current_task); DEFINE_PER_CPU(unsigned long, kernel_stack) = (unsigned long)&init_thread_union - KERNEL_STACK_OFFSET + THREAD_SIZE; EXPORT_PER_CPU_SYMBOL(kernel_stack); +DEFINE_PER_CPU(char *, irq_stack_ptr) = + init_per_cpu_var(irq_stack_union.irq_stack) + IRQ_STACK_SIZE - 64; + DEFINE_PER_CPU(unsigned int, irq_count) = -1; /* @@ -1041,6 +1049,9 @@ DEFINE_PER_CPU(struct orig_ist, orig_ist); #else /* CONFIG_X86_64 */ +DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task; +EXPORT_PER_CPU_SYMBOL(current_task); + #ifdef CONFIG_CC_STACKPROTECTOR DEFINE_PER_CPU(unsigned long, stack_canary); #endif diff --git a/arch/x86/kernel/process_32.c b/arch/x86/kernel/process_32.c index 59f4524984af..daa4107be3b4 100644 --- a/arch/x86/kernel/process_32.c +++ b/arch/x86/kernel/process_32.c @@ -61,9 +61,6 @@ asmlinkage void ret_from_fork(void) __asm__("ret_from_fork"); -DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task; -EXPORT_PER_CPU_SYMBOL(current_task); - /* * Return saved PC of a blocked thread. */ diff --git a/arch/x86/kernel/process_64.c b/arch/x86/kernel/process_64.c index ebefb5407b9d..c4c675d5ba1a 100644 --- a/arch/x86/kernel/process_64.c +++ b/arch/x86/kernel/process_64.c @@ -55,9 +55,6 @@ asmlinkage extern void ret_from_fork(void); -DEFINE_PER_CPU(struct task_struct *, current_task) = &init_task; -EXPORT_PER_CPU_SYMBOL(current_task); - DEFINE_PER_CPU(unsigned long, old_rsp); static DEFINE_PER_CPU(unsigned char, is_idle); -- cgit v1.2.3