DMARC-Filter: OpenDMARC Filter v1.3.2 mx1.redhat.com 1BE4080472
Date: Tue, 28 Mar 2017 17:03:33 -0700
From: Jessica Yu <jeyu@redhat.com>
To: Miroslav Benes <mbenes@suse.cz>
Cc: zhouchengming <zhouchengming1@huawei.com>,
        live-patching@vger.kernel.org, linux-kernel@vger.kernel.org,
        jpoimboe@redhat.com, jikos@kernel.org, pmladek@suse.com,
        huawei.libin@huawei.com
Subject: Re: [PATCH] reduce the time of finding symbols for module
Message-ID: <20170329000333.zmafvp4uk34m3em6@jeyu>
References: <1490666535-19192-1-git-send-email-zhouchengming1@huawei.com>
 <alpine.LSU.2.20.1703281055550.30237@pobox.suse.cz>
 <58DA41EE.90607@huawei.com>
 <alpine.LSU.2.20.1703281308060.30237@pobox.suse.cz>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii; format=flowed
Content-Disposition: inline
In-Reply-To: <alpine.LSU.2.20.1703281308060.30237@pobox.suse.cz>
User-Agent: NeoMutt/20161126 (1.7.1)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2648
Lines: 65

+++ 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 :-)

Jessica