On Fri, Apr 19, 2024 at 01:35:44PM +0200, Jiri Slaby wrote:
On 18. 04. 24, 22:17, Ard Biesheuvel wrote:
From: Ard Biesheuvel ardb@kernel.org
On x86, the ordinary, position dependent 'small' and 'kernel' code models only support placement of the executable in 32-bit addressable memory, due to the use of 32-bit signed immediates to generate references to global variables. For the kernel, this implies that all global variables must reside in the top 2 GiB of the kernel virtual address space, where the implicit address bits 63:32 are equal to sign bit 31.
This means the kernel code model is not suitable for other bare metal executables such as the kexec purgatory, which can be placed arbitrarily in the physical address space, where its address may no longer be representable as a sign extended 32-bit quantity. For this reason, commit
e16c2983fba0 ("x86/purgatory: Change compiler flags from -mcmodel=kernel to -mcmodel=large to fix kexec relocation errors")
switched to the 'large' code model, which uses 64-bit immediates for all symbol references, including function calls, in order to avoid relying on any assumptions regarding proximity of symbols in the final executable.
The large code model is rarely used, clunky and the least likely to operate in a similar fashion when comparing GCC and Clang, so it is best avoided. This is especially true now that Clang 18 has started to emit executable code in two separate sections (.text and .ltext), which triggers an issue in the kexec loading code at runtime.
Instead, use the position independent small code model, which makes no assumptions about placement but only about proximity, where all referenced symbols must be within -/+ 2 GiB, i.e., in range for a RIP-relative reference. Use hidden visibility to suppress the use of a GOT, which carries absolute addresses that are not covered by static ELF relocations, and is therefore incompatible with the kexec loader's relocation logic.
FWIW:
Interesting. I thought gcc doesn't have problems here yet and was talking to Matz on Thu about it and it seems he's forgotten about his statement too that "you should simply stop using -mcmodel=large. Noone should use it." :-)
Ok then, lemme queue Ard's fix.
I guess we want it in stable too but as it is pretty much the case with toolchain fixes I'd prefer if we queued it the normal path for 6.10 and then backport it after it sees some testing and nothing breaks. And not rush it to Linus now.
Thx.