[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH] stack overflow during pv-guest restore



I think this can be fixed with a much smaller change: Since the only caller
of cpu_initialize_context() is __cpu_up(), and since __cpu_up() invocation
is serialized, the variable in question could simply be static.

Jan

>>> Don Dutile <ddutile@xxxxxxxxxx> 31.01.08 21:05 >>>

When secondary cpus are initialized during an i386 pv-guest restore
(due to save/restore or live migration), and the guest has
a load that generates a fair number of interrupts (e.g., parallel kernel make),
a stack overflow can occur because cpu_initialize_context() has
a 2800 byte structure it declares on its stack.  linux-i386 has 4K stacks, by 
default.
Using 2800 bytes out of 4K by a single function in a call list isn't nice;
add the beginning of interrupt handling at just the right point, before the
switch to the interrupt stack is made... and... the scale tips...

Simple fix: malloc & free structure as needed.

Would fail save/restore testing of an i386-guest running a parallel, kernel 
make after
50->100 save/restores;  with the fix, haven't seen it fail after 650 
save/restores.

Note: this is a basic port of this function in Jeremy Fitzharinge's Xen 
implementation of pv-ops
       in upstream Linux, part of 15/32 patch, Xen SMP guest support.

Original patch done on rhel5; did a simple diff & merge to xen-unstable's 
version
of smpboot.c to generate the attached patch, so it cleanly applies; but
haven't built/run/tested the xen-unstable version.

Signed-off-by: Donald Dutile <ddutile@xxxxxxxxxx>




_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.