Re: [PATCH] reduce the time of finding symbols for module

[Li Bin]

Hi,

on 2017/3/29 8:03, Jessica Yu wrote:
> +++ Miroslav Benes [28/03/17 13:16 +0200]:
>> On Tue, 28 Mar 2017, zhouchengming wrote:
>>
>>> On 2017/3/28 17:00, Miroslav Benes wrote:
>>> >
>>> > Hi,
>>> >
>>> > On Tue, 28 Mar 2017, Zhou Chengming wrote:
>>> >
>>> > > It's reported that the time of insmoding a klp.ko for one of our
>>> > > out-tree modules is too long.
>>> > >
>>> > > ~ time sudo insmod klp.ko
>>> > > real    0m23.799s
>>> > > user    0m0.036s
>>> > > sys    0m21.256s
>>> >
>>> > Is this stable through several (>=10) runs? 23 seconds are really
>>> > suspicious. Yes, there is a linear search through all the kallsyms in
>>> > kallsyms_on_each_symbol(), but there are something like 70k symbols on my
>>> > machine (that is, way less than 1M). 23 seconds are somewhat unexpected.
>>> >
>>>
>>> Yes, it's stable through several runs.
>>>
>>> I think the big reason is that our out-tree module used a lot of static local
>>> variables. We can see '.rela.kpatch.dynrelas' contains many entries, so it
>>> will
>>> waste a lot of time if we use kallsyms_on_each_symbol() to find these symbols
>>> of module.
>>
>> Ok, it means that you have a lot of relocation records which reference
>> your out-of-tree module. Then for each such entry klp_resolve_symbol()
>> is called and then klp_find_object_symbol() to actually resolve it. So if
>> you have 20k entries, you walk through vmlinux kallsyms table 20k times.
>> It is unneeded and that is why your fix works.
>>
>> But if there were 20k modules loaded, the problem would still be there.
>>
>> I think it would be really nice to fix kallsyms :). Replace ordinary array
>> and the linear search with a hash table.
>>
>>> Relocation section '.rela.kpatch.funcs' at offset 0x382e0 contains 3 entries:
>>>   Offset          Info           Type           Sym. Value    Sym. Name +
>>> Addend
>>> 000000000000  003300000101 R_AARCH64_ABS64   0000000000000000 value_show + 0
>>> 000000000020  000b00000101 R_AARCH64_ABS64   0000000000000000 .kpatch.strings
>>> + 8
>>> 000000000028  000b00000101 R_AARCH64_ABS64   0000000000000000 .kpatch.strings
>>> + 0
>>
>> Hm, we do not have aarch64 support in upstream (yet). There is even no
>> dynamic ftrace with regs yet (if I am not mistaken).
> 
> I'm curious, how was this tested? Since there is no dynamic ftrace
> with regs and no livepatch stubs (klp_arch_set_pc, etc) implemented
> yet for aarch64. Also, livepatch has switched from klp_relocs/dynrelas
> to .klp.rela. sections since 4.7, so I'm curious how your patch module
> has a .kpatch.dynrelas section working with livepatch.
> 
> Unrelated to this patch, if there is a working aarch64 livepatch port (and
> kpatch build tool, it seems) floating out there, it would be
> wonderful to push that upstream :-)

Yeah, from 2014, we started to work on livepatch support on aarch64, and
in May 2015, we pushed the solution to the livepatch community[1] and gcc
community (mfentry feature on aarch64)[2]. And then, there were an another
gcc solution from linaro [3], which proposes to implement a new option
-fprolog-pad=N that generate a pad of N nops at the beginning of each
function, and AFAIK, Torsten Duwe from SUSE is still discussing this method
with gcc community.

At this stage, we are validating the livepatch support on aarch64 based on
aarch64 mfentry feature. When the community has a clear plan, we are happy
to make adaptation and contribute our related work to the community, including
the kpatch-build support :-)

[1] livepatch: add support on arm64
https://lkml.org/lkml/2015/5/28/54
[2] [AArch64] support -mfentry feature for arm64
https://gcc.gnu.org/ml/gcc-patches/2016-03/msg00756.html
[3] Kernel livepatching support in GCC
https://gcc.gnu.org/ml/gcc/2015-05/msg00267.html
[4] arm64: ftrace with regs for livepatch support
http://lists.infradead.org/pipermail/linux-arm-kernel/2016-January/401352.html

Thanks,
Li Bin

> 
> Jessica
> 
> .
>