Re: [PATCH] kho: check fdt_phys in is_kho_boot()

[Pasha Tatashin]

I am planning to take this patch, and change the commit message as 
follows to address Mike's concern:

> During arm64 setup_arch(), efi_init() -> reserve_regions() calls
> is_kho_boot() before arm64_memblock_init() initializes memstart_addr.
> At this point memstart_addr still holds its initial sentinel value (-1).
> 
> When booting via kexec with KHO enabled, kho_populate() sets
> kho_in.fdt_phys from the /chosen FDT node. Calling is_kho_boot() ->
> kho_get_fdt() invokes phys_to_virt(kho_in.fdt_phys), which evaluates
> PHYS_OFFSET and triggers VM_BUG_ON(memstart_addr & 1) on arm64 kernels
> built with CONFIG_DEBUG_VM=y:

Repalce the above two parapgraphs with:

is_kho_boot() is called early in boot before the linear memory map is 
initialized, so calling phys_to_virt() via kho_get_fdt() there is 
invalid.

On arm64 (and similarly in upcoming riscv64 support), this triggers 
BUG_ONs in debug builds, such as:

> 
>   kernel BUG at ./arch/arm64/include/asm/memory.h:377!
>   Internal error: Oops - BUG: 00000000f2000800 [#1] SMP
>   Call trace:
>    is_kho_boot+0x2c/0x38 (P)
>    efi_init+0x94/0xf8
>    setup_arch+0xf0/0x240
>    start_kernel+0x84/0x490
>    __primary_switched+0x88/0x98
> 
> Check kho_in.fdt_phys directly in is_kho_boot() instead of calling
> kho_get_fdt() so that early boot callers do not invoke phys_to_virt()
> before the linear map offset is initialized.

Replace the above with:

Use kho_in.fdt_phys in is_kho_boot() directly instead of translating it 
to a virtual address.

-- 
Pasha