* Jay Lan <jlan at sgi.com> [2007-08-21 15:18]: > Vivek Goyal wrote: > > On Thu, Aug 16, 2007 at 06:26:35PM +0900, Takenori Nagano wrote: > >> Vivek Goyal wrote: > >> > So for the time being I think we can put RAS tools on die notifier list > >>> and if it runs into issues we can always think of creating a separate list. > >>> > >>> Few things come to mind. > >>> > >>> - Why there is a separate panic_notifier_list? Can't it be merged with > >>> die_chain? die_val already got one of the event type as PANIC. If there > >>> are no specific reasons then we should merge the two lists. Registering > >>> RAS tools on a single list is easier. > >> I think it is difficult, because die_chain is defined by each architecture. > >> > > > > I think die_chain is arch independent definition (kernel/die_notifier.c)? > > But anyway, to begin with it can be done only for panic_notifier. > > > >>> - Modify Kdump to register on die_chain list. > >>> - Modify Kdb to register on die_chain list. > >>> - Export all the registered members of die_chain through sysfs along with > >>> their priorities. Priorities should be modifiable. Most likely one > >>> shall have to introduce additional field in struct notifier_block. This > >>> field will be a string as an identifier of the user registerd. e.g > >>> "Kdump", "Kdb" etc. > >>> > >>> Now user will be able to view all the die_chain users through sysfs and > >>> be able to modify the order in which these should run by modifying their > >>> priority. Hence all the RAS tools can co-exist. > >> This is my image of your proposal. > >> > >> - Print current order > >> > >> # cat /sys/class/misc/debug/panic_notifier_list > >> priority name > >> 1 IPMI > >> 2 watchdog > >> 3 Kdb > >> 4 Kdump > >> > > > > I think Bernhard's suggestion looks better here. I noticed that > > /sys/kernel/debug is already present. So how about following. > > > > /sys/kernel/debug/kdump/priority > > /sys/kernel/debug/kdb/priority > > /sys/kernel/debug/IPMI/priority > > Why separate priority files is better than a central file? > At least i think you get a grand picture of priority being > defined for all parties with a central file? Well, it's more intuitive to set the priority in that case. You don't have to know a special syntax. However, it may be a good idea to implement a second read-only file that lists the sorted priorities in that order the kernel executues the handlers the handlers. > What do we decide priority if more than one component has > the same priority value? You can check this and return EINVAL in that case. Thanks, Bernhard
