lkml.org 
[lkml]   [2018]   [Feb]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v3 8/7] TESTING_ONLY x86/entry: reduce static footprint of idtentry

* Dominik Brodowski <linux@dominikbrodowski.net> wrote:

> Play a little trick in the generic PUSH_AND_CLEAR_REGS macro
> to insert the GP registers "above" the original return address.
> This allows us to (re-)insert the macro in error_entry() and
> paranoid_entry() and to remove it from the idtentry macro. This
> reduces the static footprint significantly.
>
> NOTE: It still needs to be evaluated whether the reduced static
> footprint at the cost of code readability is really needed here.
>
> Co-developed-by: Linus Torvalds <torvalds@linux-foundation.org>
> Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
> ---
> arch/x86/entry/calling.h | 11 ++++++++++-
> arch/x86/entry/entry_64.S | 21 ++++++++++-----------
> 2 files changed, 20 insertions(+), 12 deletions(-)
>
> diff --git a/arch/x86/entry/calling.h b/arch/x86/entry/calling.h
> index 6985440c68fa..2d4523ba04a8 100644
> --- a/arch/x86/entry/calling.h
> +++ b/arch/x86/entry/calling.h
> @@ -97,7 +97,7 @@ For 32-bit we have the following conventions - kernel is built with
>
> #define SIZEOF_PTREGS 21*8
>
> -.macro PUSH_AND_CLEAR_REGS rdx=%rdx rax=%rax
> +.macro PUSH_AND_CLEAR_REGS rdx=%rdx rax=%rax save_ret=0
> /*
> * Push registers and sanitize registers of values that a
> * speculation attack might otherwise want to exploit. The
> @@ -105,8 +105,14 @@ For 32-bit we have the following conventions - kernel is built with
> * could be put to use in a speculative execution gadget.
> * Interleave XOR with PUSH for better uop scheduling:
> */
> + .if \save_ret
> + pushq %rsi /* pt_regs->si */
> + movq 8(%rsp), %rsi /* temporarily store ret address in %rsi */
> + movq %rdi, 8(%rsp) /* pt_regs->di (overwriting original ret) */
> + .else
> pushq %rdi /* pt_regs->di */
> pushq %rsi /* pt_regs->si */
> + .endif
> pushq \rdx /* pt_regs->dx */
> pushq %rcx /* pt_regs->cx */
> pushq \rax /* pt_regs->ax */
> @@ -131,6 +137,9 @@ For 32-bit we have the following conventions - kernel is built with
> pushq %r15 /* pt_regs->r15 */
> xorq %r15, %r15 /* nospec r15*/
> UNWIND_HINT_REGS
> + .if \save_ret
> + pushq %rsi /* return address on top of stack */
> + .endif
> .endm
>
> .macro POP_REGS pop_rdi=1 skip_r11rcx=0
> diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S
> index b868433d7e68..a2e41177e390 100644
> --- a/arch/x86/entry/entry_64.S
> +++ b/arch/x86/entry/entry_64.S
> @@ -871,12 +871,8 @@ ENTRY(\sym)
> pushq $-1 /* ORIG_RAX: no syscall to restart */
> .endif
>
> - /* 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 */
> + testb $3, CS-ORIG_RAX(%rsp) /* If coming from userspace, switch stacks */
> jnz .Lfrom_usermode_switch_stack_\@
> .endif
>
> @@ -1123,12 +1119,15 @@ idtentry machine_check do_mce has_error_code=0 paranoid=1
> #endif
>
> /*
> - * Switch gs if needed.
> + * Save all registers in pt_regs, and 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
> + PUSH_AND_CLEAR_REGS save_ret=1
> + ENCODE_FRAME_POINTER 8
> movl $1, %ebx
> movl $MSR_GS_BASE, %ecx
> rdmsr
> @@ -1141,7 +1140,7 @@ ENTRY(paranoid_entry)
> SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14
>
> ret
> -ENDPROC(paranoid_entry)
> +END(paranoid_entry)
>
> /*
> * "Paranoid" exit path from exception stack. This is invoked
> @@ -1172,12 +1171,14 @@ ENTRY(paranoid_exit)
> END(paranoid_exit)
>
> /*
> - * Switch gs if needed.
> + * Save all registers in pt_regs, and switch gs if needed.
> * Return: EBX=0: came from user mode; EBX=1: otherwise
> */
> ENTRY(error_entry)
> - UNWIND_HINT_REGS offset=8
> + UNWIND_HINT_FUNC
> cld
> + PUSH_AND_CLEAR_REGS save_ret=1
> + ENCODE_FRAME_POINTER 8
> testb $3, CS+8(%rsp)
> jz .Lerror_kernelspace
>
> @@ -1568,8 +1569,6 @@ end_repeat_nmi:
> * frame to point back to repeat_nmi.
> */
> pushq $-1 /* ORIG_RAX: no syscall to restart */
> - PUSH_AND_CLEAR_REGS
> - ENCODE_FRAME_POINTER
>
> /*
> * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit

Ok, so this does not look _that_ complicated, and the .text savings are
significant:

text data bss dec hex filename
24182 0 0 24182 5e76 entry_64.o.before
20878 0 0 20878 518e entry_64.o.after

But, let's try to estimate the dynamic I$ impact through the static entry points.

I'm using the following list of 'static' IDT entry points:

divide_error
overflow
bounds
invalid_op
device_not_available
double_fault
coprocessor_segment_overrun
invalid_TSS
segment_not_present
spurious_interrupt_bug
coprocessor_error
alignment_check
simd_coprocessor_error
debug
int3
stack_segment
general_protection
page_fault
machine_check
nmi
apic_timer_interrupt
x86_platform_ipi
threshold_interrupt
deferred_error_interrupt
thermal_interrupt
call_function_single_interrupt
call_function_interrupt
reschedule_interrupt
error_interrupt
spurious_interrupt
irq_work_interrupt

Which are present in the vmlinux with well-known addresses - this makes I$
estimates easier.

But let's first filter these entries - a big chunk of them are legacy vectors,
never used on a modern x86 CPU built in the last decade or so. Another chunk are
infrequent entries like <machine_check>.

The entry points that are relevant to an I$ analysis are just 11 entry points:

divide_error
# overflow
# bounds
# invalid_op
# device_not_available
# double_fault
# coprocessor_segment_overrun
# invalid_TSS
# segment_not_present
# spurious_interrupt_bug
# coprocessor_error
# alignment_check
# simd_coprocessor_error
debug
int3
# stack_segment
general_protection
page_fault
# machine_check
nmi
apic_timer_interrupt
# x86_platform_ipi
# threshold_interrupt
# deferred_error_interrupt
# thermal_interrupt
call_function_single_interrupt
call_function_interrupt
reschedule_interrupt
# error_interrupt
# spurious_interrupt
irq_work_interrupt


( Also note that even this list is probably generous: for example
<apic_timer_interrupt> will only arrive once per timer IRQ, by which time it's
probably already out of the L1 cache - and <debug> only matters to a small
minority of users. )

Here's the x86-64 defconfig addresses of these entry points, on a 'vanilla' kernel
that does not have any entry point bloating (and debloating) patches applied, sha1
b7a60525e83a:

address symbol # size # offs
---------------- ------------------------------------ -------------
ffffffff81a00e70: <divide_error> # 40 # 0
ffffffff81a01250: <debug> # 60 # 3e0
ffffffff81a012b0: <int3> # 60 # 60
ffffffff81a01370: <general_protection> # 60 # c0
ffffffff81a013d0: <page_fault> # 60 # 60
ffffffff81a01680: <nmi> # cc # 2b0
ffffffff81a01a40: <apic_timer_interrupt> # c0 # 3c0
ffffffff81a01e00: <call_function_single_interrupt> # c0 # 3c0
ffffffff81a01ec0: <call_function_interrupt> # c0 # c0
ffffffff81a01f80: <reschedule_interrupt> # c0 # c0
ffffffff81a021c0: <irq_work_interrupt> # c0 # 240

The 'size' column is the size of the entry function, the 'offs' column shows the
symbol's offset from the previous entry.

As can be seen in the table:

- even in the 'compressed' layout the entries are already far enough from each
other to make the dynamic I$ footprint 27 cachelines, assuming 64 byte
cachelines:

27 == 1 + 2 + 2 + 2 + 2 + 3 + 3 + 3 + 3 + 3

After applying the simplification patches (sha: 5a10e729bc0a) the static footprint
of the entry code becomes more bloated:

text data bss dec hex filename
19693 0 0 19693 4ced arch/x86/entry/entry_64.o.before
24212 0 0 24212 5e94 arch/x86/entry/entry_64.o.after (+4519 bytes)

But the dynamic I$ footprint does not change nearly as much:

address symbol # size # offs
---------------- ------------------------------------ -------------
ffffffff81a00d00: <divide_error> # 70 # 0
ffffffff81a01320: <debug> # 90 # 620
ffffffff81a013b0: <int3> # 90 # 90
ffffffff81a014c0: <general_protection> # 80 # 110
ffffffff81a01540: <page_fault> # 80 # 80
ffffffff81a01780: <nmi> # cc # 240
ffffffff81a01b70: <apic_timer_interrupt> # 80 # 3f0
ffffffff81a01df0: <call_function_single_interrupt> # 80 # 280
ffffffff81a01e70: <call_function_interrupt> # 80 # 80
ffffffff81a01ef0: <reschedule_interrupt> # 80 # 80
ffffffff81a02070: <irq_work_interrupt> # 80 # 180

The footprint is in fact 25 cache lines:

25 == 2 + 3 + 3 + 2 + 2 + 3 + 2 + 2 + 2 + 2 + 2

The reason is that while exception entry size increased, the IRQ entry size
actually shrunk, because the PUSH based sequences are more compact than the MOVs
they used before:

# BEFORE:

ffffffff81a01f80 <reschedule_interrupt>:
[...]
ffffffff81a01f98: 48 83 c4 88 add $0xffffffffffffff88,%rsp
ffffffff81a01f9c: 4c 89 5c 24 30 mov %r11,0x30(%rsp)
ffffffff81a01fa1: 4c 89 54 24 38 mov %r10,0x38(%rsp)
ffffffff81a01fa6: 4c 89 4c 24 40 mov %r9,0x40(%rsp)
ffffffff81a01fab: 4c 89 44 24 48 mov %r8,0x48(%rsp)
ffffffff81a01fb0: 48 89 44 24 50 mov %rax,0x50(%rsp)
ffffffff81a01fb5: 48 89 4c 24 58 mov %rcx,0x58(%rsp)
ffffffff81a01fba: 48 89 54 24 60 mov %rdx,0x60(%rsp)
ffffffff81a01fbf: 48 89 74 24 68 mov %rsi,0x68(%rsp)
ffffffff81a01fc4: 48 89 7c 24 70 mov %rdi,0x70(%rsp)
ffffffff81a01fc9: 4c 89 3c 24 mov %r15,(%rsp)
ffffffff81a01fcd: 4c 89 74 24 08 mov %r14,0x8(%rsp)
ffffffff81a01fd2: 4c 89 6c 24 10 mov %r13,0x10(%rsp)
ffffffff81a01fd7: 4c 89 64 24 18 mov %r12,0x18(%rsp)
ffffffff81a01fdc: 48 89 6c 24 20 mov %rbp,0x20(%rsp)
ffffffff81a01fe1: 48 89 5c 24 28 mov %rbx,0x28(%rsp)
ffffffff81a01fe6: 31 ed xor %ebp,%ebp
ffffffff81a01fe8: 31 db xor %ebx,%ebx
ffffffff81a01fea: 4d 31 c0 xor %r8,%r8
ffffffff81a01fed: 4d 31 c9 xor %r9,%r9
ffffffff81a01ff0: 4d 31 d2 xor %r10,%r10
ffffffff81a01ff3: 4d 31 db xor %r11,%r11
ffffffff81a01ff6: 4d 31 e4 xor %r12,%r12
ffffffff81a01ff9: 4d 31 ed xor %r13,%r13
ffffffff81a01ffc: 4d 31 f6 xor %r14,%r14
ffffffff81a01fff: 4d 31 ff xor %r15,%r15

# AFTER:

ffffffff81a01ef0 <reschedule_interrupt>:
[...]
ffffffff81a01f08: 57 push %rdi
ffffffff81a01f09: 56 push %rsi
ffffffff81a01f0a: 52 push %rdx
ffffffff81a01f0b: 51 push %rcx
ffffffff81a01f0c: 50 push %rax
ffffffff81a01f0d: 41 50 push %r8
ffffffff81a01f0f: 4d 31 c0 xor %r8,%r8
ffffffff81a01f12: 41 51 push %r9
ffffffff81a01f14: 4d 31 c9 xor %r9,%r9
ffffffff81a01f17: 41 52 push %r10
ffffffff81a01f19: 4d 31 d2 xor %r10,%r10
ffffffff81a01f1c: 41 53 push %r11
ffffffff81a01f1e: 4d 31 db xor %r11,%r11
ffffffff81a01f21: 53 push %rbx
ffffffff81a01f22: 31 db xor %ebx,%ebx
ffffffff81a01f24: 55 push %rbp
ffffffff81a01f25: 31 ed xor %ebp,%ebp
ffffffff81a01f27: 41 54 push %r12
ffffffff81a01f29: 4d 31 e4 xor %r12,%r12
ffffffff81a01f2c: 41 55 push %r13
ffffffff81a01f2e: 4d 31 ed xor %r13,%r13
ffffffff81a01f31: 41 56 push %r14
ffffffff81a01f33: 4d 31 f6 xor %r14,%r14
ffffffff81a01f36: 41 57 push %r15
ffffffff81a01f38: 4d 31 ff xor %r15,%r15

Finally, applying the 8/7 debloating patch gives the following result:

address symbol # size # offs
---------------- ------------------------------------ -------------
ffffffff81a00d00: <divide_error> # 40 # 0
ffffffff81a010b0: <debug> # 50 # 3b0
ffffffff81a01100: <int3> # 50 # 50
ffffffff81a011a0: <general_protection> # 50 # a0
ffffffff81a011f0: <page_fault> # 50 # 50
ffffffff81a01450: <nmi> # cc # 260
ffffffff81a01810: <apic_timer_interrupt> # 80 # 3c0
ffffffff81a01a90: <call_function_single_interrupt> # 80 # 280
ffffffff81a01b10: <call_function_interrupt> # 80 # 80
ffffffff81a01b90: <reschedule_interrupt> # 80 # 80
ffffffff81a01d10: <irq_work_interrupt> # 80 # 180

That's 22 cachelines:

22 == 1 + 2 + 2 + 2 + 2 + 3 + 2 + 2 + 2 + 2 + 2
25 == 2 + 3 + 3 + 2 + 2 + 3 + 2 + 2 + 2 + 2 + 2

So it's a 3 cachelines improvements, coming from the smaller size of these three
entry points:

ffffffff81a00d00: <divide_error> # 40 # 0
ffffffff81a010b0: <debug> # 50 # 3b0
ffffffff81a01100: <int3> # 50 # 50

... which are all debugging/instrumentation related and are thus performance
critical only for a small minority of our users.

So I'm torn whether to apply this last patch, but the interesting, unintuitive
result is that the existing MOV->PUSH simplification patches already appear to
have improved the dynamic I$ footprint from 27 cachelines to 22 cachelines!

Thanks,

Ingo

\
 
 \ /
  Last update: 2018-02-12 10:38    [W:0.191 / U:1.516 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site