Right thanks. I will pull/try it shortly and give you feedback.
I will only be able to try it on my Intel systems – no AMDs at the moment.
From: Keir Fraser
Sent: Tuesday, September 16, 2008 8:45 AM
To: Ross Philipson; xen-devel@xxxxxxxxxxxxxxxxxxx
Cc: Christoph Egger
Subject: Re: [Xen-devel] Crash in update microcode changes - change set
The better fix was not to do the update in
interrupt context at all, but instead 'continue' the hypercall on each online
processor in turn. This is what I implemented in c/s 18487. It'd be nice to
know that this works for you (and also for Christoph).
In earlier changesets I've also cleaned up the microcode quite a lot and
reformatted for Xen coding style. My thinking is that microcode-update logic is
not all the complex, and the original code not really all that great, so I'm
not that bothered about keeping closely in sync with the Linux original
version. In this case I'd rather have it in a format I'm happy to hack on
On 16/9/08 13:37, "Ross Philipson" <Ross.Philipson@xxxxxxxxxx>
Actually perhaps it is easy to fix. The copying of
chunks out of the larger microcode buffer seems to just be a convenience.
Perhaps just returning offset pointers into the original buffer from the
get_next_ucode_from_buffer() function would get rid of the vmalloc() call. Just
a thought as I looked at it further.
On Behalf Of Ross Philipson
Sent: Monday, September 15, 2008 5:35 PM
Subject: [Xen-devel] Crash in update microcode changes - change set
The changes for
CPU microcode loading that were done recently (change set 18475 in unstable
staging) seem to be causing a crash. I am using an Intel system and I get an
assertion in Xen during the DOM0 boot. This is what I believe is going on.
In xen/arch/x86/microcode.c the routine do_microcode_update() is dispatching
the update work to each CPU with on_each_cpu() which in turn uses an IPI to
dispatch the callback vector on each CPU. The microcode update routine passed
in is called in the IPI context on the target CPU (including irq_enter() before
calling the ucode function). Within the ucode function the calls eventually get
down to the Intel specific calls in microcode_intel.c. Specifically:
Within the last call, vmalloc() is called and eventually _xmalloc() asserts on
ASSERT(!in_irq()). I checked and the earlier code, though it dispatched work to
different CPUs with IPIs, did not try to dynamically allocate memory. I am not
sure how to fix this easily without redoing how the whole new microcode
framework works. Also I didn’t look closely at AMD but it may have the same
issue. I can take a crack at fixing it but maybe someone will see a simple
Senior Software Engineer
Citrix Systems, Inc
14 Crosby Drive
Bedford, MA 01730
Xen-devel mailing list