Am Do., 6. Aug. 2020 um 02:30 Uhr schrieb Russell King - ARM Linux admin <li...@armlinux.org.uk>: > > On Thu, Aug 06, 2020 at 01:05:55AM +0200, Norbert Lange wrote: > > diff --git a/arch/arm/boot/compressed/head.S > > b/arch/arm/boot/compressed/head.S > > index 434a16982e34..1af01bfe6638 100644 > > --- a/arch/arm/boot/compressed/head.S > > +++ b/arch/arm/boot/compressed/head.S > > @@ -614,7 +614,11 @@ not_relocated: mov r0, #0 > > */ > > mov r0, r4 > > mov r1, sp @ malloc space above stack > > +#if defined(CONFIG_KERNEL_ZSTD) > > + add r2, sp, #0x30000 @ Context needs ~160K > > That's going to mess up kexec: > > /* > * The zImage length does not include its stack (4k) or its > * malloc space (64k). Include this. > */ > len += 0x11000; > > I guess we need to add this to the information provided to kexec.
Ouch, I guess it's rather impossible to load a new ZSTD kernel with a old kexec version in that case. Some ideas to fix that would be: - Increase the padding to "192K + 4K is enough for everyone" in kexec - Add the required/recommended stack+heap size to zImage so kexec can use it. - change boot logic to place heap before/after the *output* range. with the various ways of booting arm this will likely cause other issues in some combinations. - try to reduce zstd context size safetly below 64K. I believe the entropy tables are computed once and never modified, so those could be part of the zImage (at some loss of compression rate of course). Would need some host tool to inject those. Norbert