Re: loading firmware while usermodehelper disabled.

[Ingo Molnar <mingo@elte.hu>]

* Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:

> > You mensioned earlier about not being able to tell the 
> > difference between a device that needs firmware and one that 
> > needs flash (e.g. they use exactly the same ids). It doesn't 
> > really matter - we just assume that it might need firmware 
> > and load it anyway. It uses more memory but is robust.
> 
> We only need to do that for the devices where order and not 
> blocking matters. There are a few (and some are video) where 
> the firmware sizes is megabytes, which on an embedded 
> controlling device is not acceptable. I don't believe any of 
> them are things where simply delaying the restoration will 
> cause problems however - its video, and DVB and the like not 
> wireless or serial.

Here's the size histogram/analysis of all *.fw, *.bin, *.dat, 
*.ucode, etc. files in /lib/firmware on a fully populated 
distro:

 476 firmware blobs total

The toplist:

 -rw-r--r--. 1 root root 2786404 Jul 24  2010 ./bcm70012fw.bin
 -rw-r--r--. 1 root root 1781048 Feb  9  2011 ./phanfw.bin
 -rw-r--r--. 1 root root  864276 Jul 24  2010 ./bcm70015fw.bin
 -rw-r--r--. 1 root root  844980 Feb  8  2011 ./asihpi/dsp6200.bin
 -rw-r--r--. 1 root root  636980 Feb  8  2011 ./asihpi/dsp6600.bin
 -rw-r--r--. 1 root root  627696 Feb  8  2011 ./asihpi/dsp6400.bin
 -rw-r--r--. 1 root root  563592 Aug  4 22:04 ./myri10ge_rss_ethp_z8e.dat
 -rw-r--r--. 1 root root  553192 Aug  4 22:04 ./myri10ge_rss_eth_z8e.dat
 -rw-r--r--. 1 root root  504916 Feb  8  2011 ./asihpi/dsp8900.bin
 -rw-r--r--. 1 root root  498128 Sep  7 20:14 ./ct2fw.bin

 - 2% of them, i.e. just a tiny fraction is over 512 KB.
 - 80% of the firmware blobs are below 100K.
 - 50% of them are below 16K.

So loading them into RAM is the obviously right solution.

Those few devices that absolutely want to load the firmware blob 
dynamically on some weird low-RAM system can do so *BEFORE* 
suspending.

There is nothing that prevents a low-RAM system from loading the 
firmware blob in an early suspend callback and making sure it's 
there at resume time - and then unloading it from RAM after 
resume.

I.e. large blobs can manage their RAM usage just fine - but the 
obscenity of the 1% should not control the design and sanity of 
the 99% case ...

> [...]
> 
> The world is heading this way more and more. It's moving from 
> the old PC model of 'user closes lid, clunk for 15 seconds, 
> enter suspend, user opens lid, churn churn, video, churn 
> clunk. resume' to suspend/resume being so fast it happens 
> between keystrokes.

Exactly!

Thanks,

	Ingo