Re: [Xen-devel] x86: PIE support and option to extend KASLR randomization

[Thomas Garnier <thgarnie@xxxxxxxxxx>]

On Fri, Aug 25, 2017 at 8:05 AM, Thomas Garnier <thgarnie@xxxxxxxxxx> wrote:
> On Fri, Aug 25, 2017 at 1:04 AM, Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>>
>> * Thomas Garnier <thgarnie@xxxxxxxxxx> wrote:
>>
>>> With the fix for function tracing, the hackbench results have an
>>> average of +0.8 to +1.4% (from +8% to +10% before). With a default
>>> configuration, the numbers are closer to 0.8%.
>>>
>>> On the .text size, with gcc 4.9 I see +0.8% on default configuration
>>> and +1.180% on the ubuntu configuration.
>>
>> A 1% text size increase is still significant. Could you look at the 
>> disassembly,
>> where does the size increase come from?
>
> I will take a look, in this current iteration I added the .got and
> .got.plt so removing them will remove a big (even if they are small,
> we don't use them to increase perf).
>
> What do you think about the perf numbers in general so far?

I looked at the size increase. I could identify two common cases:

1) PIE sometime needs two instructions to represent a single
instruction on mcmodel=kernel.

For example, this instruction plays on the sign extension (mcmodel=kernel):

mov    r9,QWORD PTR [r11*8-0x7e3da060] (8 bytes)

The address 0xffffffff81c25fa0 can be represented as -0x7e3da060 using
a 32S relocation.

with PIE:

lea    rbx,[rip+<off>] (7 bytes)
mov    r9,QWORD PTR [rbx+r11*8] (6 bytes)

2) GCC does not optimize switches in PIE in order to reduce relocations:

For example the switch in phy_modes [1]:

static inline const char *phy_modes(phy_interface_t interface)
{
    switch (interface) {
    case PHY_INTERFACE_MODE_NA:
        return "";
    case PHY_INTERFACE_MODE_INTERNAL:
        return "internal";
    case PHY_INTERFACE_MODE_MII:
        return "mii";

Without PIE (gcc 7.2.0), the whole table is optimize to be one instruction:

   0x000000000040045b <+27>:    mov    rdi,QWORD PTR [rax*8+0x400660]

With PIE (gcc 7.2.0):

   0x0000000000000641 <+33>:    movsxd rax,DWORD PTR [rdx+rax*4]
   0x0000000000000645 <+37>:    add    rax,rdx
   0x0000000000000648 <+40>:    jmp    rax
....
   0x000000000000065d <+61>:    lea    rdi,[rip+0x264]        # 0x8c8
   0x0000000000000664 <+68>:    jmp    0x651 <main+49>
   0x0000000000000666 <+70>:    lea    rdi,[rip+0x2bc]        # 0x929
   0x000000000000066d <+77>:    jmp    0x651 <main+49>
   0x000000000000066f <+79>:    lea    rdi,[rip+0x2a8]        # 0x91e
   0x0000000000000676 <+86>:    jmp    0x651 <main+49>
   0x0000000000000678 <+88>:    lea    rdi,[rip+0x294]        # 0x913
   0x000000000000067f <+95>:    jmp    0x651 <main+49>

That's a deliberate choice, clang is able to optimize it (clang-3.8):

   0x0000000000000963 <+19>:    lea    rcx,[rip+0x200406]        # 0x200d70
   0x000000000000096a <+26>:    mov    rdi,QWORD PTR [rcx+rax*8]

I checked gcc and the code deciding to fold the switch basically do
not do it for pic to reduce relocations [2].

The switches are the biggest increase on small functions but I don't
think they represent a large portion of the difference (number 1 is).

A side note, while testing gcc 7.2.0 on hackbench I have seen the PIE
kernel being faster by 1% across multiple runs (comparing 50 runs done
across 5 reboots twice). I don't think PIE is faster than a
mcmodel=kernel but recent versions of gcc makes them fairly similar.

[1] 
http://elixir.free-electrons.com/linux/v4.13-rc7/source/include/linux/phy.h#L113
[2] 
https://github.com/gcc-mirror/gcc/blob/7977b0509f07e42fbe0f06efcdead2b7e4a5135f/gcc/tree-switch-conversion.c#L828

>
>>
>> Thanks,
>>
>>         Ingo
>
>
>
> --
> Thomas



-- 
Thomas

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel