On Sep 17, 2015, at 7:33 PM, James Morse wrote:
Hi James and Will,
> Hi Will,
>
> On 16/09/15 12:25, Will Deacon wrote:
>> On Sun, Sep 13, 2015 at 03:42:17PM +0100, Jungseok Lee wrote:
>>> diff --git a/arch/arm64/include/asm/thread_info.h
>>> b/arch/arm64/include/asm/thread_info.h
>>> index dcd06d1..44839c0 100644
>>> --- a/arch/arm64/include/asm/thread_info.h
>>> +++ b/arch/arm64/include/asm/thread_info.h
>>> @@ -73,8 +73,11 @@ static inline struct thread_info
>>> *current_thread_info(void) __attribute_const__;
>>>
>>> static inline struct thread_info *current_thread_info(void)
>>> {
>>> - return (struct thread_info *)
>>> - (current_stack_pointer & ~(THREAD_SIZE - 1));
>>> + unsigned long sp_el0;
>>> +
>>> + asm volatile("mrs %0, sp_el0" : "=r" (sp_el0));
>>> +
>>> + return (struct thread_info *)(sp_el0 & ~(THREAD_SIZE - 1));
>>
>> This looks like it will generate worse code than our current implementation,
>> thanks to the asm volatile. Maybe just add something like a global
>> current_stack_pointer_el0?
>
> Like current_stack_pointer does?:
>> register unsigned long current_stack_pointer_el0 asm ("sp_el0");
>
> Unfortunately the compiler won't accept this, as it doesn't like the
> register name, it also won't accept instructions in this asm string.
>
> Dropping the 'volatile' has the desired affect[0]. This would only cause a
> problem over a call to cpu_switch_to(), which writes to sp_el0, but also
> save/restores the callee-saved registers, so they will always be consistent.
>
>
> James
>
>
>
>
> [0] A fictitious example printk:
>> printk("%p%p%u%p", get_fs(), current_thread_info(),
>> smp_processor_id(), current);
>
> With this patch compiles to:
> 5f8: d5384101 mrs x1, sp_el0
> 5fc: d5384100 mrs x0, sp_el0
> 600: d5384103 mrs x3, sp_el0
> 604: d5384104 mrs x4, sp_el0
> 608: 9272c484 and x4, x4, #0xffffffffffffc000
> 60c: 9272c463 and x3, x3, #0xffffffffffffc000
> 610: 9272c421 and x1, x1, #0xffffffffffffc000
> 614: aa0403e2 mov x2, x4
> 618: 90000000 adrp x0, 0 <do_bad>
> 61c: f9400884 ldr x4, [x4,#16]
> 620: 91000000 add x0, x0, #0x0
> 624: b9401c63 ldr w3, [x3,#28]
> 628: f9400421 ldr x1, [x1,#8]
> 62c: 94000000 bl 0 <printk>
>
> Removing the volatile:
> 5e4: d5384102 mrs x2, sp_el0
> 5e8: f9400844 ldr x4, [x2,#16]
> 5ec: 91000000 add x0, x0, #0x0
> 5f0: b9401c43 ldr w3, [x2,#28]
> 5f4: f9400441 ldr x1, [x2,#8]
> 5f8: 94000000 bl 0 <printk>
>
>
As Will pointed out, if "worse" means "bigger text size", the change generates
worse codes than current implementation. A data based on System.map is as
follows.
GCC version: aarch64-linux-gnu-gcc (Linaro GCC 2014.11) 4.9.3 20141031
(prerelease)
[1] 4.3-rc1
ffffffc000080000 T _text
ffffffc0007f1524 R _etext
[2] 4.3-rc1 + this patch
ffffffc000080000 T _text
ffffffc0007f8504 R _etext
[3] 4.3-rc1 + this patch + the following hunk
ffffffc000080000 T _text
ffffffc0007ef514 R _etext
diff --git a/arch/arm64/include/asm/thread_info.h
b/arch/arm64/include/asm/thread_info.h
index 44839c0..4ab08a1 100644
--- a/arch/arm64/include/asm/thread_info.h
+++ b/arch/arm64/include/asm/thread_info.h
@@ -77,7 +77,7 @@ static inline struct thread_info *current_thread_info(void)
asm volatile("mrs %0, sp_el0" : "=r" (sp_el0));
- return (struct thread_info *)(sp_el0 & ~(THREAD_SIZE - 1));
+ return (struct thread_info *)sp_el0;
}
#define thread_saved_pc(tsk) \
diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S
index c156540..314ac81 100644
--- a/arch/arm64/kernel/entry.S
+++ b/arch/arm64/kernel/entry.S
@@ -88,7 +88,8 @@
.if \el == 0
mrs x21, sp_el0
- get_thread_info \el, tsk // Ensure MDSCR_EL1.SS is clear,
+ mov tsk, sp
+ and tsk, tsk, #~(THREAD_SIZE - 1) // Ensure MDSCR_EL1.SS is clear,
ldr x19, [tsk, #TI_FLAGS] // since we can unmask debug
disable_step_tsk x19, x20 // exceptions when scheduling.
.else
@@ -105,8 +106,7 @@
.if \el == 0
mvn x21, xzr
str x21, [sp, #S_SYSCALLNO]
- mov x25, sp
- msr sp_el0, x25
+ msr sp_el0, tsk
.endif
/*
@@ -165,13 +165,8 @@ alternative_endif
eret // return to kernel
.endm
- .macro get_thread_info, el, rd
- .if \el == 0
- mov \rd, sp
- .else
+ .macro get_thread_info, rd
mrs \rd, sp_el0
- .endif
- and \rd, \rd, #~(THREAD_SIZE - 1) // bottom of thread stack
.endm
.macro get_irq_stack
@@ -400,7 +395,7 @@ el1_irq:
irq_handler
#ifdef CONFIG_PREEMPT
- get_thread_info 1, tsk
+ get_thread_info tsk
ldr w24, [tsk, #TI_PREEMPT] // get preempt count
cbnz w24, 1f // preempt count != 0
ldr x0, [tsk, #TI_FLAGS] // get flags
@@ -636,6 +631,7 @@ ENTRY(cpu_switch_to)
ldp x29, x9, [x8], #16
ldr lr, [x8]
mov sp, x9
+ and x9, x9, #~(THREAD_SIZE - 1)
msr sp_el0, x9
ret
ENDPROC(cpu_switch_to)
@@ -695,7 +691,7 @@ ENTRY(ret_from_fork)
cbz x19, 1f // not a kernel thread
mov x0, x20
blr x19
-1: get_thread_info 1, tsk
+1: get_thread_info tsk
b ret_to_user
ENDPROC(ret_from_fork)
diff --git a/arch/arm64/kernel/head.S b/arch/arm64/kernel/head.S
index cb13290..213df0b 100644
--- a/arch/arm64/kernel/head.S
+++ b/arch/arm64/kernel/head.S
@@ -442,6 +442,7 @@ __mmap_switched:
2:
adr_l sp, initial_sp, x4
mov x4, sp
+ and x4, x4, #~(THREAD_SIZE - 1)
msr sp_el0, x4
str_l x21, __fdt_pointer, x5 // Save FDT pointer
str_l x24, memstart_addr, x6 // Save PHYS_OFFSET
@@ -615,6 +616,7 @@ ENDPROC(secondary_startup)
ENTRY(__secondary_switched)
ldr x0, [x21] // get secondary_data.stack
mov sp, x0
+ and x0, x0, #~(THREAD_SIZE - 1)
msr sp_el0, x0
mov x29, #0
b secondary_start_kernel
If struct thread_info address is directly stored into sp_el0, we can avoid
masking operation in many places. It helps to decrease a kernel text size.
This idea comes from James's comment in v1 patch.
Best Regards
Jungseok Lee--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
