Re: Warning from free_init_pages with large initrd
From: Sitsofe Wheeler
Date: Tue Sep 27 2016 - 17:21:52 EST
On 27 September 2016 at 19:36, H. Peter Anvin <hpa@xxxxxxxxx> wrote:
> On September 27, 2016 3:20:06 AM PDT, Sitsofe Wheeler <sitsofe@xxxxxxxxx> wrote:
>>(See http://www.gossamer-threads.com/lists/linux/kernel/2534175 for
>>the history of this thread )
>>On 26 September 2016 at 20:00, Randy Dunlap <rdunlap@xxxxxxxxxxxxx>
>>> but the warning in free_init_pages() is about alignment, not size...
>>> Maybe the concatenation is bad?
>>What would l have to pull apart to be able to tell?
>>Having said that I've just noticed that newer versions of the script
>>concatenate an ISO to the initrd (rather than another cpio) -
>>. Unfortunately after rebuilding the initrd with the
>>livecd-iso-to-pxeboot version from Fedora 24 the warning persists.
>>CC'ing a few of the tool's authors on this email.
> This really doesn't seem like a good approach, and certainly may cause memory exhaustion at boot time. Still would be good to track down the problem.
> Overall the idea of downloading the whole shebang in the initial TFTP transaction ID a really bad idea. TFTP is slow and unreliable, and the PXE ROMs are often complete crap in terms of quality. It would be way better to bring up a small initramfs and then fetch the rest using http or a similar protocol from the very high quality Linux TCP stack. As far as I know, Fedora has supported this mode for *ages*.
I wouldn't mind tracking down the problem too but I admit I don't know
where to start on this alignment issue.
As for the process, it can be excruciating slow and can cause memory
exhaustion but the initrd and vmlinuz are actually downloaded over
HTTP courtesy of iPXE (so grabbing 1GByte happens in under a minute
over HTTP) and these days machines have many gigabytes of memory so it
actually works quite well once booted...
I was going to say that Fedora doesn't allow PXE booting of its live
images (I know it always supported something for server installs) but
I've just come across this
https://bugzilla.redhat.com/show_bug.cgi?id=1317780 so perhaps there's
a another route to take in addition to this next time round...
Sitsofe | http://sucs.org/~sits/