On Tue, Nov 21, 2023 at 02:41:08PM +0800, Baoquan He wrote: > > > > > > Still failed. And I found the normal reboot does't work either. I will > > > do more testing tomorrow, e.g use the tdx-tools's own rhel9 kernel > > > config and rebuild, and update host kernel too. > > I did more tests, resuls are summarized as below: > > 1) kexec reboot works, but always fallback to 1 cpu even though multiple > cpus are specified; That's expected. Until you have new BIOS. See below. > 2) kdump kernel need more crashkernel memory to boot up, > crashkernel=512M works well in our case. I guess it is due to SWIOTLB memory which requres at least 64M. > But it failed in vmcore > saving process, either makedumpfile or cp can't access the 1st > kernel's old memory; Will look into it. > 3) Normal reboot always failed; It is expected. TD vCPUs are not resettable. So we need to destroy TD and construct a new one to emulate "reboot". I guess we can try to add some glue on QEMU side to make reboot more seamless. > My colleague Tao helped to double check this, he got the same testing result. > Plesae > 1) what can we do to enable the multiple cpu support for kexec reboot? You would need a patched BIOS image. I've hacked one[1] for my testing. But it only works if kernel runs in 4-level paging mode (specify no5lvl in kernel command line). BIOS folks work on proper patch, but it is not ready yet. [1] https://gist.github.com/kiryl/e1dc1719e0c990b3ceee5d8de8dbf332 > 2) anything missing to allow makedumpfile/cp access 1st kernel's memory? It worked before for us, but I have not checked for a while. I expected you've dropped my "if (crash) return;" debug patch I asked you to tested before, right? If not, failure is expected. > 3) not sure if this is particular case on the system we tested on. -- Kiryl Shutsemau / Kirill A. Shutemov _______________________________________________ kexec mailing list kexec@xxxxxxxxxxxxxxxxxxx http://lists.infradead.org/mailman/listinfo/kexec
