Eric and Hugh have reported instant reboot due to my recent changes in decompression code.
The root cause is that I didn't realize that we need to adjust GOT to be able to run C code that early. The problem is only visible with old toolchain. Binutils >= 2.24 is able to eliminate GOT references by replacing them with RIP-relative address loads[1]. We need to adjust GOT two times: - before calling paging_prepare() with address the binary was loaded by bootloader - before relocating the kernel to the new place with relocation address [1] https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=commitdiff;h=80d873266dec Signed-off-by: Kirill A. Shutemov <[email protected]> Fixes: 194a9749c73d ("x86/boot/compressed/64: Handle 5-level paging boot if kernel is above 4G") Reported-by: Eric Dumazet <[email protected]> Reported-by: Hugh Dickins <[email protected]> --- arch/x86/boot/compressed/head_64.S | 66 ++++++++++++++++++++++++------ 1 file changed, 53 insertions(+), 13 deletions(-) diff --git a/arch/x86/boot/compressed/head_64.S b/arch/x86/boot/compressed/head_64.S index fca012baba19..6cbb2d64c91e 100644 --- a/arch/x86/boot/compressed/head_64.S +++ b/arch/x86/boot/compressed/head_64.S @@ -305,6 +305,21 @@ ENTRY(startup_64) /* Set up the stack */ leaq boot_stack_end(%rbx), %rsp + /* + * paging_prepare() and cleanup_trampoline() below can have GOT + * references. Adjust the table with address we are running at. + */ + + /* The GOP was not adjusted before */ + xorq %rax, %rax + + /* Calculate the address the binary is loaded at. */ + call 1f +1: popq %rdi + subq $1b, %rdi + + call adjust_gop + /* * At this point we are in long mode with 4-level paging enabled, * but we might want to enable 5-level paging or vice versa. @@ -381,6 +396,24 @@ trampoline_return: pushq $0 popfq + /* + * Previously we've adjusted the GOT with address the binary was + * loaded at. Now we need to re-adjust for relocation address. + */ + + /* + * Calculate the address the binary is loaded at. + * This address was used to adjust the table before and we need to + * undo the change. + */ + call 1f +1: popq %rax + subq $1b, %rax + + /* The new adjustment is relocation address */ + movq %rbx, %rdi + call adjust_gop + /* * Copy the compressed kernel to the end of our buffer * where decompression in place becomes safe. @@ -481,19 +514,6 @@ relocated: shrq $3, %rcx rep stosq -/* - * Adjust our own GOT - */ - leaq _got(%rip), %rdx - leaq _egot(%rip), %rcx -1: - cmpq %rcx, %rdx - jae 2f - addq %rbx, (%rdx) - addq $8, %rdx - jmp 1b -2: - /* * Do the extraction, and jump to the new kernel.. */ @@ -512,6 +532,26 @@ relocated: */ jmp *%rax +/* + * Adjust global offest table + * + * RAX is previous adjustment of the table to undo (0 if it's the first time we touch GOP). + * RDI is the new adjustment to apply. + */ +adjust_gop: + /* Walk through the GOT adding the address to the entries */ + leaq _got(%rip), %rdx + leaq _egot(%rip), %rcx +1: + cmpq %rcx, %rdx + jae 2f + subq %rax, (%rdx) /* Undo previous adjustment */ + addq %rdi, (%rdx) /* Apply the new adjustment */ + addq $8, %rdx + jmp 1b +2: + ret + .code32 /* * This is the 32-bit trampoline that will be copied over to low memory. -- 2.17.0
