The last patch in this series implements a new CONFIG_PRINTK_PERSIST option
that, when enabled, puts the printk buffer in a well-defined memory location
so that we can keep appending to it after a reboot. The upshot is that,
even after a kernel panic or non-panic hard lockup, on the next boot
userspace will be able to grab the kernel messages leading up to it. It
could then upload the messages to a server (for example) to keep crash
statistics.
The preceding patches in the series are mostly just things I fixed up while
working on that patch.
Some notes:
- I'm not totally sure of the locking or portability issues when calling
memblock or bootmem. This all happens really early, and I *think*
interrupts are still disabled at that time, so it's probably okay.
- Tested this version on x86 (kvm) and it works with soft reboot (ie. reboot
-f). Since some BIOSes wipe the memory during boot, you might not have
any luck. It should be great on many embedded systems, though, including
the MIPS system I've tested a variant of this patch on. (Our MIPS build
is based on a slightly older kernel so it's not 100% the same, but I think
this should behave identically.)
- The way we choose a well-defined memory location is slightly suspicious
(we just count down from the top of the address space) but I've tested it
pretty carefully, and it seems to be okay.
- In printk.c with CONFIG_PRINTK_PERSIST set, we're #defining words like
log_end. It might be cleaner to replace all instances of log_end with
LOG_END to make this more clear. This is also the reason the struct
logbits members start with _: because otherwise they conflict with the
macro. Suggestions welcome.