On Tue, 2003-03-25 at 19:28, Benjamin Herrenschmidt wrote:

You "fixed" it by using GFP_ATOMIC but didn't test the result of
kmalloc. That is very bad. GFP_ATOMIC can fail (return NULL), thus
you will crash the kernel under high memory pressure.

I think the proper fix is, as you asked me, using a workqueue,
that way, you can both use GFP_KERNEL allocations, and avoid
the spinlock you added to fbmem.c, thus letting the fb_sync()
ops on fbdev's be able to block.

Well, actually, creating a workqueue would be overhead since
it involves one kernel thread per CPU. After more thinking &
discussion, I beleive you shall rather use keventd existing
workqueue (schedule_work() will do that)

Ben.



-------------------------------------------------------
This SF.net email is sponsored by:
The Definitive IT and Networking Event. Be There!
NetWorld+Interop Las Vegas 2003 -- Register today!
http://ads.sourceforge.net/cgi-bin/redirect.pl?keyn0001en