[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [RFC 0/9] Introduce light weight PC platform pc-lite

From: Paolo Bonzini
Subject: Re: [Qemu-devel] [RFC 0/9] Introduce light weight PC platform pc-lite
Date: Thu, 23 Jun 2016 14:44:23 +0200
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Thunderbird/45.1.1

On 23/06/2016 10:32, Chao Peng wrote:
> The original usage model is to replace kvm-tool with QEMU for Clear
> Containers (https://clearlinux.org/features/clear-containers). It's not
> going to present the guest a real PC platform, but instead, a totally
> virtual platform.

It is not completely virtual; it has PCI for example.  Hyper-V is an
example of a completely virtual platform, even the LAPIC is customized
with paravirtual features.

qboot does basically four things: 1) relocate from ROM to 0xf0000; 2)
initialize PCI; 2) provide the ACPI and e820 tables; 3) boot.

If Linux can boot without initializing PCI bridges and without INTX, we
can remove that code from qboot.  The PCI scan is the most expensive
part, I think.  (2) and (3) are the same no matter if you run them in
QEMU or the guest.

That leaves out only relocation (PAM).

> Every little bit boot time saving is important because
> we are trying to achieve comparable result with that for Linux native
> container.
> With this usage model, I doubt introducing a firmware layer is a good
> idea:
>     On one side, even with optimized and compact qboot it still takes us
>     ~15ms.

Have you profiled it?  If it is code in QEMU that we can optimize (e.g.
memory.c), that would benefit all guests.

>     This is not a small value because current Linux kernel takes
>     only ~50ms (and we are still on the way to optimize it). And when
>     you look at the SeaBIOS or qboot, almost all the code are useless for
>     this usage model. They are doing things that is important for
>     traditional PC booting but cost 15ms doing useless things for us (It
>     is really not easy to save 15ms in other place, for example, in
>     Linux. Personally I tend to change the architecture for this new
>     usage model, e.g. eliminate firmware).
>     On the other side, even boot the new pc-lite platform with firmware,
>     it does not mean it can support non-Linux system like Windows. So
>     generally I don't see the benefit of introducing a firmware layer.

The main benefit is maintainability, by reducing the amount of code
specific to pc-lite.

> Besides, I'm also not quite sure if build around Q35 is the best
> solution:
>     The problem with Q35 is some features like SMM/SMRAM/PAM slow done
>     the booting even we actually never use them. While removing these
>     features can cause guest see different feature set for a same device
>     and it also prevents us to do further optimizations on that in guest.

Of these, qboot only uses PAM, and even that could be removed (PAM is
only necessary because of how qboot probes parallel flash).  SMRAM
should not slow down booting if you don't use them.  Do they?


reply via email to

[Prev in Thread] Current Thread [Next in Thread]