Re: [PATCH 2/4] x86: get rid of KERNEL_STACK_OFFSET

[Denys Vlasenko <dvlasenk@redhat.com>]

On 02/25/2015 01:48 PM, Denys Vlasenko wrote:
> On Wed, Feb 25, 2015 at 9:53 AM, Ingo Molnar <mingo@kernel.org> wrote:
>> But the fix should be to not touch RSP in SAVE_ARGS, to
>> keep percpu::kernel_stack as an optimized entry point -
>> with KERNEL_STACK_OFFSET pointing to.
>>
>> So NAK - this should be fixed for real.
> 
> IOW, the proposal is to set KERNEL_STACK_OFFSET
> to SIZEOF_PTREGS. I can do that.
> 
> However.
> 
> There is an ortogonal idea we were discussing: to save
> registers and construct iret frame using PUSH insns, not MOVs.
> IIRC Andy and Linus liked it. I am ambivalent: the code will be smaller,
> but might get slower (at least on some CPUs).
> If we go that way, we will require KERNEL_STACK_OFFSET = 0
> (IOW: the current patch).
> 
> The decision on how exactly we should fix KERNEL_STACK_OFFSET
> (set it to SIZEOF_PTREGS or to zero) depends on whether
> we switch to using PUSHes, or not. What do you think?

A data point. I implemented push-based creation of pt_regs
and benchmarked it. The patch is on top of all my latest
patches sent to ML.

On SandyBridge CPU, it does not get slower: seems to be 1 cycle
faster per syscall.

We lose a number of large insns there:

    text           data     bss     dec     hex filename
-   9863              0       0    9863    2687 entry_64.o
+   9671              0       0    9671    25c7 entry_64.o


diff --git a/arch/x86/kernel/entry_64.S b/arch/x86/kernel/entry_64.S
index f505cb6..d97bd92 100644
--- a/arch/x86/kernel/entry_64.S
+++ b/arch/x86/kernel/entry_64.S
@@ -128,8 +128,6 @@ ENDPROC(native_usergs_sysret64)
  * manipulation.
  */
 	.macro FIXUP_TOP_OF_STACK tmp offset=0
-	movq $__USER_DS,SS+\offset(%rsp)
-	movq $__USER_CS,CS+\offset(%rsp)
 	movq RIP+\offset(%rsp),\tmp  /* get rip */
 	movq \tmp,RCX+\offset(%rsp)  /* copy it to rcx as sysret would do */
 	movq EFLAGS+\offset(%rsp),\tmp /* ditto for rflags->r11 */
@@ -245,14 +243,22 @@ GLOBAL(system_call_after_swapgs)
 	 * and short:
 	 */
 	ENABLE_INTERRUPTS(CLBR_NONE)
-	ALLOC_PT_GPREGS_ON_STACK 6*8 /* 6*8: space for orig_ax and iret frame */
-	movq	%rcx,RIP(%rsp)
-	movq	%r11,EFLAGS(%rsp)
-	movq    PER_CPU_VAR(old_rsp),%rcx
-	movq    %rcx,RSP(%rsp)
-	movq_cfi rax,ORIG_RAX
-	SAVE_C_REGS_EXCEPT_RAX_RCX_R11
-	movq	$-ENOSYS,RAX(%rsp)
+	/* Construct struct pt_regs on stack */
+	pushq	$__USER_DS		/* pt_regs->ss */
+	pushq	PER_CPU_VAR(old_rsp)	/* pt_regs->sp */
+	pushq	%r11			/* pt_regs->flags */
+	pushq	$__USER_CS		/* pt_regs->cs */
+	pushq	%rcx			/* pt_regs->ip */
+	pushq	%rax			/* pt_regs->orig_ax */
+	pushq	%rdi			/* pt_regs->di */
+	pushq	%rsi			/* pt_regs->si */
+	pushq	%rdx			/* pt_regs->dx */
+	pushq	%rcx			/* pt_regs->cx */
+	pushq	$-ENOSYS		/* pt_regs->ax */
+	pushq	%r8			/* pt_regs->r8 */
+	pushq	%r9			/* pt_regs->r9 */
+	pushq	%r10			/* pt_regs->r10 */
+	sub	$(7*8),%rsp /* pt_regs->r11,bp,bx,r12-15 */
 	CFI_REL_OFFSET rip,RIP
 	testl $_TIF_WORK_SYSCALL_ENTRY,TI_flags+THREAD_INFO(%rsp,SIZEOF_PTREGS)
 	jnz tracesys