On Sat, 25 May 2013 16:31:58 +0800 Zhang Yanfei <zhangyanfei.yes at gmail.com> wrote: [snip] > For s390, if we put swap info into the elf header, This will > change /sbin/kexec. But at this point, copy_oldmem_page is still > doing the swap when we try to read the pages among [0 - OLDMEM_SIZE] > and [OLDMEM_BASE - OLDMEM_BASE + OLDMEM_SIZE]. So removing the swap > in copy_oldmem_page should be done at the same time. New kexec with > old kernels would fail and old kexec with new kernels would fail too. > > So could you please explain more about the ""backwards compatible". > And please correct me if I am wrong. Hello all, I think Zhang is right and in theory we are not backwards compatible because with the ELF header fix old kexec tools would no longer work with new kernels. But: For s390 we normally do *not* create the ELF header in the old kernel with kexec. Instead the new kernel does all the memory and CPU detection and creates the ELF header in the new memory. See also our current discussion with Vivek: https://lkml.org/lkml/2013/5/24/164 The main reason why we did this for s390 was that we can have many CPU hotplug events because of a s390 specific daemon called "cpuplugd" (s390-tools). We wanted to avoid the kdump reloading with kexec triggered by CPU and memory hotplug events. For s390 distributions the kexec udev rules are disabled. Besides of the newmem mechanism, for completeness, we also implemented the oldmem ELF header mechansim in kexec. But this is disabled by default. See: "#ifdef WITH_ELF_HEADER" in kexec/arch/s390/crashdump-s390.c Currently no distribution uses the oldmem mechanism. Therefore, if necessary, IMHO we can switch to the ELF header memory swap mechanism for s390 in the kernel. Of course we would then also have to adjust the (disabled) kexec code. Best Regards, Michael
