[PATCH 1/3] stack overflow safe kdump (2.6.18-rc1-i386) - safe_smp_processor_id

[kaos at ocs.com.au (Keith Owens)]

Fernando Luis =?ISO-8859-1?Q?V=E1zquez?= Cao (on Tue, 11 Jul 2006 13:21:01 +0900) wrote:
>That is a good idea, but I have on concern. In mach-default by default
>we use __send_IPI_shortcut (no_broadcast==0) instead of send_IPI_mask.
>Is it always safe to ignore the no_broadcast setting? In other words,
>can __send_IPI_shortcut be replaced by send_IPI_mask safely?

It is always safe to use send_IPI_mask.  It is not used by default
because of concerns that send_IPI_mask may be slower than using a
broadcast, although I do not know if anybody has measurements to back
up that concern.  OTOH I can guarantee that sending NMI as a broadcast
has problems, it breaks some Dell Xeon servers[1].  My fix was to never
broadcast NMI, from 2.6.18-rc1 NMI_VECTOR always uses a mask[2] and
crash was changed accordingly[3].

[1] http://marc.theaimsgroup.com/?t=114828920800003&r=1&w=2
[2] http://marc.theaimsgroup.com/?t=115103727400006&r=1&w=2
[3] http://marc.theaimsgroup.com/?t=115096703800003&r=1&w=2