> On a system which is configured for kexec crash dumps, some memory is reserved at boot time for a second Linux kernel (the “kdump kernel”). On startup, the system uses kexec_load(2) to load a kernel image into this reserved memory region. If a panic occurs, all CPUs are halted and control is transferred to the kdump kernel.
Is there really no better way to do this than to have an entire second kernel ready to take over? Like a more specialized piece of code that only handles kernel coredumps?
> Like a more specialized piece of code that only handles kernel coredumps?
Can't the kdump kernel be exactly that?
You can build a second kernel separate from the normal one, that only has the necessary options enabled to handle the coredump stuff. This way it can be relatively small and you can reuse the required driver code.
That is in fact how crash dump works on illumos systems. Once you enter a kernel panic, interrupts are disabled, all other CPUs are brought to rest, etc. Then the panic code takes the pages to write to the dump and passes them to a special routine in the driver for the dump device: usually either a slice on a disk or a zvol). That driver is responsible for getting the device into a state where polled I/O can be performed, and to write the pages out. At the end the system reboots, and on the next boot a program called savecore inspects the dump device to see if there is a new dump, and if there is it transfers it into a new file.
FreeBSD does the same thing (probably some shared 80s BSD heritage?). One challenge is actually providing that special routine for the dump device in a way that is safe at crash time. You don't know how bad a state your crashed kernel is in, and some buses may not have a polled IO mode (USB?).
The Linux kexec process kind of elegantly restarts into a known state, with the known drawback of hogging some memory. It's definitely a different set of tradeoffs and both seem vaguely reasonable.
> Is there really no better way to do this than to have an entire second kernel
> ready to take over? Like a more specialized piece of code that only handles
> kernel coredumps?
So the requirements of this software are: It must be able to boot a system, write to the specific sections of memory, have some method of initializing hardware, some ability to write the other memory to disk or network connections. It sounds like you're talking about an OS.
A crash dumper does not have to be implemented with a reboot; this is just something Linux chooses to do. I have actually worked on kernel code that performs crash dumps to disk and over the network, if you find that kind of expertise relevant.
Is there really no better way to do this than to have an entire second kernel ready to take over? Like a more specialized piece of code that only handles kernel coredumps?