[tip:x86/pti] x86/entry/64: Get rid of the ALLOC_PT_GPREGS_ON_STACK and SAVE_AND_CLEAR_REGS macros

From: tip-bot for Dominik Brodowski
Date: Mon Feb 12 2018 - 05:20:14 EST


Commit-ID: f09d160992d129b6d8fada7543761d40f5b33596
Gitweb: https://git.kernel.org/tip/f09d160992d129b6d8fada7543761d40f5b33596
Author: Dominik Brodowski <linux@xxxxxxxxxxxxxxxxxxxx>
AuthorDate: Sun, 11 Feb 2018 11:49:47 +0100
Committer: Ingo Molnar <mingo@xxxxxxxxxx>
CommitDate: Mon, 12 Feb 2018 08:06:36 +0100

x86/entry/64: Get rid of the ALLOC_PT_GPREGS_ON_STACK and SAVE_AND_CLEAR_REGS macros

Previously, error_entry() and paranoid_entry() saved the GP registers
onto stack space previously allocated by its callers. Combine these two
steps in the callers, and use the generic PUSH_AND_CLEAR_REGS macro
for that.

This adds a significant amount ot text size. However, Ingo Molnar points
out that:

"these numbers also _very_ significantly over-represent the
extra footprint. The assumptions that resulted in
us compressing the IRQ entry code have changed very
significantly with the new x86 IRQ allocation code we
introduced in the last year:

- IRQ vectors are usually populated in tightly clustered
groups.

With our new vector allocator code the typical per CPU
allocation percentage on x86 systems is ~3 device vectors
and ~10 fixed vectors out of ~220 vectors - i.e. a very
low ~6% utilization (!). [...]

The days where we allocated a lot of vectors on every
CPU and the compression of the IRQ entry code text
mattered are over.

- Another issue is that only a small minority of vectors
is frequent enough to actually matter to cache utilization
in practice: 3-4 key IPIs and 1-2 device IRQs at most - and
those vectors tend to be tightly clustered as well into about
two groups, and are probably already on 2-3 cache lines in
practice.

For the common case of 'cache cold' IRQs it's the depth of
the call chain and the fragmentation of the resulting I$
that should be the main performance limit - not the overall
size of it.

- The CPU side cost of IRQ delivery is still very expensive
even in the best, most cached case, as in 'over a thousand
cycles'. So much stuff is done that maybe contemporary x86
IRQ entry microcode already prefetches the IDT entry and its
expected call target address."[*]

[*] http://lkml.kernel.org/r/20180208094710.qnjixhm6hybebdv7@xxxxxxxxx

The "testb $3, CS(%rsp)" instruction in the idtentry macro does not need
modification. Previously, %rsp was manually decreased by 15*8; with
this patch, %rsp is decreased by 15 pushq instructions.

[jpoimboe@xxxxxxxxxx: unwind hint improvements]

Suggested-by: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
Signed-off-by: Dominik Brodowski <linux@xxxxxxxxxxxxxxxxxxxx>
Cc: Andy Lutomirski <luto@xxxxxxxxxx>
Cc: Borislav Petkov <bp@xxxxxxxxx>
Cc: Brian Gerst <brgerst@xxxxxxxxx>
Cc: Denys Vlasenko <dvlasenk@xxxxxxxxxx>
Cc: H. Peter Anvin <hpa@xxxxxxxxx>
Cc: Josh Poimboeuf <jpoimboe@xxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Cc: dan.j.williams@xxxxxxxxx
Link: http://lkml.kernel.org/r/20180211104949.12992-7-linux@xxxxxxxxxxxxxxxxxxxx
Signed-off-by: Ingo Molnar <mingo@xxxxxxxxxx>
---
arch/x86/entry/calling.h | 42 +-----------------------------------------
arch/x86/entry/entry_64.S | 20 +++++++++-----------
2 files changed, 10 insertions(+), 52 deletions(-)

diff --git a/arch/x86/entry/calling.h b/arch/x86/entry/calling.h
index d6a97e2..5967501 100644
--- a/arch/x86/entry/calling.h
+++ b/arch/x86/entry/calling.h
@@ -97,46 +97,6 @@ For 32-bit we have the following conventions - kernel is built with

#define SIZEOF_PTREGS 21*8

- .macro ALLOC_PT_GPREGS_ON_STACK
- addq $-(15*8), %rsp
- .endm
-
- .macro SAVE_AND_CLEAR_REGS offset=0
- /*
- * Save registers and sanitize registers of values that a
- * speculation attack might otherwise want to exploit. The
- * lower registers are likely clobbered well before they
- * could be put to use in a speculative execution gadget.
- * Interleave XOR with PUSH for better uop scheduling:
- */
- movq %rdi, 14*8+\offset(%rsp)
- movq %rsi, 13*8+\offset(%rsp)
- movq %rdx, 12*8+\offset(%rsp)
- movq %rcx, 11*8+\offset(%rsp)
- movq %rax, 10*8+\offset(%rsp)
- movq %r8, 9*8+\offset(%rsp)
- xorq %r8, %r8 /* nospec r8 */
- movq %r9, 8*8+\offset(%rsp)
- xorq %r9, %r9 /* nospec r9 */
- movq %r10, 7*8+\offset(%rsp)
- xorq %r10, %r10 /* nospec r10 */
- movq %r11, 6*8+\offset(%rsp)
- xorq %r11, %r11 /* nospec r11 */
- movq %rbx, 5*8+\offset(%rsp)
- xorl %ebx, %ebx /* nospec rbx */
- movq %rbp, 4*8+\offset(%rsp)
- xorl %ebp, %ebp /* nospec rbp */
- movq %r12, 3*8+\offset(%rsp)
- xorq %r12, %r12 /* nospec r12 */
- movq %r13, 2*8+\offset(%rsp)
- xorq %r13, %r13 /* nospec r13 */
- movq %r14, 1*8+\offset(%rsp)
- xorq %r14, %r14 /* nospec r14 */
- movq %r15, 0*8+\offset(%rsp)
- xorq %r15, %r15 /* nospec r15 */
- UNWIND_HINT_REGS offset=\offset
- .endm
-
.macro PUSH_AND_CLEAR_REGS rdx=%rdx rax=%rax
/*
* Push registers and sanitize registers of values that a
@@ -211,7 +171,7 @@ For 32-bit we have the following conventions - kernel is built with
* is just setting the LSB, which makes it an invalid stack address and is also
* a signal to the unwinder that it's a pt_regs pointer in disguise.
*
- * NOTE: This macro must be used *after* SAVE_AND_CLEAR_REGS because it corrupts
+ * NOTE: This macro must be used *after* PUSH_AND_CLEAR_REGS because it corrupts
* the original rbp.
*/
.macro ENCODE_FRAME_POINTER ptregs_offset=0
diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S
index b06a4b5..cfbf433 100644
--- a/arch/x86/entry/entry_64.S
+++ b/arch/x86/entry/entry_64.S
@@ -871,7 +871,9 @@ ENTRY(\sym)
pushq $-1 /* ORIG_RAX: no syscall to restart */
.endif

- ALLOC_PT_GPREGS_ON_STACK
+ /* Save all registers in pt_regs */
+ PUSH_AND_CLEAR_REGS
+ ENCODE_FRAME_POINTER

.if \paranoid < 2
testb $3, CS(%rsp) /* If coming from userspace, switch stacks */
@@ -1121,15 +1123,12 @@ idtentry machine_check do_mce has_error_code=0 paranoid=1
#endif

/*
- * Save all registers in pt_regs, and switch gs if needed.
+ * Switch gs if needed.
* Use slow, but surefire "are we in kernel?" check.
* Return: ebx=0: need swapgs on exit, ebx=1: otherwise
*/
ENTRY(paranoid_entry)
- UNWIND_HINT_FUNC
cld
- SAVE_AND_CLEAR_REGS 8
- ENCODE_FRAME_POINTER 8
movl $1, %ebx
movl $MSR_GS_BASE, %ecx
rdmsr
@@ -1142,7 +1141,7 @@ ENTRY(paranoid_entry)
SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14

ret
-END(paranoid_entry)
+ENDPROC(paranoid_entry)

/*
* "Paranoid" exit path from exception stack. This is invoked
@@ -1173,14 +1172,12 @@ ENTRY(paranoid_exit)
END(paranoid_exit)

/*
- * Save all registers in pt_regs, and switch gs if needed.
+ * Switch gs if needed.
* Return: EBX=0: came from user mode; EBX=1: otherwise
*/
ENTRY(error_entry)
- UNWIND_HINT_FUNC
+ UNWIND_HINT_REGS offset=8
cld
- SAVE_AND_CLEAR_REGS 8
- ENCODE_FRAME_POINTER 8
testb $3, CS+8(%rsp)
jz .Lerror_kernelspace

@@ -1571,7 +1568,8 @@ end_repeat_nmi:
* frame to point back to repeat_nmi.
*/
pushq $-1 /* ORIG_RAX: no syscall to restart */
- ALLOC_PT_GPREGS_ON_STACK
+ PUSH_AND_CLEAR_REGS
+ ENCODE_FRAME_POINTER

/*
* Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit