Re: [for-next][PATCH 04/26] bootconfig: Add Extra Boot Config support

[Masami Hiramatsu]

On Tue, 18 Feb 2020 14:27:24 +0100
Borislav Petkov <bp@xxxxxxxxx> wrote:

> On Tue, Feb 11, 2020 at 11:02:07AM +0900, Masami Hiramatsu wrote:
> > Sorry, you might misunderstand that the bootconfig replaces legacy command
> > line. No, the legacy command line is still there and at least the part of
> > suppremental command line parts of bootconfig are merged into the command
> > line and shown in dmesg. So even if the user use bootconfig, their kernel
> > command line was just updated, they can grub it and simply paste on other
> > machine (if it accepts longer command line)
> 
> Ok, I did just play with the thing. I used this boot config for my guest:
> 
> ---
> kernel {
> 	root=/dev/sda1
> 	resume=/dev/sda2
> 	debug
> 	ignore_loglevel
> 	log_buf_len=16M
> 	earlyprintk=ttyS0,115200
> 	console=ttyS0,115200

Ah, bootconfig syntax interprets the "," as a delimiter of array elements.
So those must be quoted by "" (or '') as below.

	earlyprintk="ttyS0,115200"
	console="ttyS0,115200"

> 	console=tty0

Oops, I forgot to support double parameters.

> 	no_console_suspend
> }
> 
> init {
> 	net.ifnames=0
> }
> ---
> 
> and added it to an initrd which I created especially for this. Which
> brings me to the next issue I see: AFAIU, bootconfig requires people
> to have initrd. And the majority of deployments do have initrd enabled
> but there are use cases where people don't need it. So if bootconfig is
> going to be default Y, it better support that case.

Yes, I think we can embed bootconfig into the kernel image while
building it as a text data. I think appending the bootconfig to kernel
image will be hard because some kernel image will be compressed and self-
extracted on memory. It might break original data.

> For example and for a similar reason I support builtin microcode in the
> microcode loader - when you don't want to enable initrd.

Thanks for the good information! I'll check it.

> Then, I did add the above bootconfig to the initrd:
> 
> ./tools/bootconfig/bootconfig -a bootconfig /tmp/initrd.img
> Apply bootconfig to /tmp/initrd.img
>         Number of nodes: 21
>         Size: 194 bytes
>         Checksum: 16268
> 
> and booted my guest.
> 
> The thing I noticed is that the guest doesn't show any output in its
> window anymore but only shows the login window when it is done booting.
> 
> And that is probably because of the console= thing. /proc/bootconfig has
> 
> $ cat /proc/bootconfig
> kernel.root = "/dev/sda1"
> kernel.resume = "/dev/sda2"
> kernel.debug = ""
> kernel.ignore_loglevel = ""
> kernel.log_buf_len = "16M"
> kernel.earlyprintk = "ttyS0", "115200"
> kernel.console = "ttyS0", "115200", "tty0"
> kernel.no_console_suspend = ""
> init.net.ifnames = "0"
> 
> and bootconfig parsing has merged the console= parameters. And the same
> thing happens if I boot the guest with "console=ttyS0,115200,tty0" on
> the normal cmdline. But
> 
> console=ttyS0,115200 console=tty0
> 
> is not the same as
> 
> console=ttyS0,115200,tty0

Yes, that is my mistake. Since the legacy command line doesn't support
array, it should be expanded. If you write it correctly, I will ensure
it to
console=ttyS0,115200 console=tty0

> 
> Former has been this way since forever if you wanna have serial output
> to ttyS* *and* tty output at the same time.
> 
> Also, I wouldn't want to rule out any other changes in behavior in
> other command line options due to this change in parsing and merging of
> options.

For the above example, I can fix it. I also would not like to change the
behaviors. Could you share your concerning items if you have more?

> 
> So again, I'm still not convinced bootconfig should be used everywhere
> and by default.
> 
> In addition to the initrd requirement, it needs a special tool to glue
> it to the end of the initrd. Yeah yeah, it is part of the kernel and the
> glueing can be made part of the initrd creation tool but still, it is
> one more and new step.
> 
> Now, if bootconfig can *supplement* the normal command line so that you
> can add more complex configuration directives with it, in *addition* to
> the normal command line, then sure, by all means.
> 
> And it seems that it does that: I did
> 
> ---
> kernel {
>         initcall_debug
> }
> 
> init {
> }
> ---
> 
> and added it to the cmdline:
> 
> [    0.047358] Kernel command line: initcall_debug root=/dev/sda1 resume=/dev/sda2 debug ignore_loglevel log_buf_len=16M earlyprintk=ttyS0,115200 console=ttyS0,115200 console=tty0 no_console_suspend net.ifnames=0  bootconfig
> 
> and you can see "initcall_debug" there and it did enable the initcall
> debugging output so yeah, that seems to work.
> 
> Btw, if you remove the boot config with the tool (-d) and still boot
> with "bootconfig" it says:
> 
> [    0.043958] bootconfig size -23483140 greater than max size 32767
> 
> so you need to check presence of bootconfig blob and limits in the
> parsing code too or so.

Agreed, that's confusing message.

> 
> In any case, this is only my opinion, of course, and I might very well
> be missing something.

Thank you very much for your opinion and information!
That's very helpful to me.

Thank you,

-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>