On 2019/11/7 23:46, Josh Poimboeuf wrote:
On Thu, Nov 07, 2019 at 10:32:01PM +0800, shile.zhang@xxxxxxxxxxxxxxxxx wrote:
From: Shile Zhang <shile.zhang@xxxxxxxxxxxxxxxxx>
Hi,
I found the unwind_init taken long time (more than 90ms) in kernel
booting, mainly spent on sorting the two ORC unwind tables, orc_unwind
and orc_unwind_ip.
I also noticed that this issued has reported and discussed last year:
https://lkml.org/lkml/2018/10/8/342
But seems no final solution until now, I tried to sort the ORC tables at
build time, followed the helpful hints from Josh and Ingo in that thread.
And mainly referred the implementation of 'sortextable' tool:
https://lore.kernel.org/linux-mips/1334872799-14589-1-git-send-email-ddaney.cavm@xxxxxxxxx/
What I did:
- Add a Kconfig to control build-time sorting or runtime sorting;
- Referred 'sortextable', create a similar helper tool 'sortorctable',
help to sort the ORC unwind tables at vmlinux link process.
One potential improvement is to sort the module ORC tables in future.
Thanks!
Thanks a lot for working on this!
I'd say the new config option isn't needed. The runtime ORC sorting
logic is unconditionally bad and the code should just be removed. I saw
recently that it's one of the main offenders for boot time latency.
Hi, Josh, Thanks very much for your quickly response!
I'll refactor the code as your advice.
Yes, the run-time sorting cost is bigger for currently Cloud products,
such as serverless products, which needs boot up ASAP.
I also agree with Peter that we should try to reduce the link-time
penalty as much as possible. But it's a necessary evil to a certain
extent.
agree!
Thanks again and looking forwards more advice from you!
