Re: [RFC] rust: kernel: Add KUnit tests for ARCH_WARN_ASM bug table emission
flat view
From: Mark Rutland <mark.rutland@arm.com>
Date: 2026-09-24 12:47:27
Also in:
linux-riscv, linux-s390, lkml, llvm, rust-for-linux
On Thu, Sep 24, 2026 at 01:15:57PM +0200, Peter Zijlstra wrote:
On Wed, Sep 23, 2026 at 11:58:13AM +0530, Mukesh Kumar Chaurasiya wrote:quoted
On Tue, Sep 22, 2026 at 08:47:35AM +0200, Peter Zijlstra wrote:quoted
On Tue, Sep 22, 2026 at 11:20:00AM +0530, Mukesh Kumar Chaurasiya (IBM) wrote:quoted
Verify that the __bug_table entry emitted by ARCH_WARN_ASM has a correct bug_addr displacement — i.e. the arch's trap label reference resolves to the trap instruction — by calling find_bug() with the exact virtual address of the trap, mirroring what the real trap handler does. To support all architectures, each arch that implements ARCH_WARN_ASM now defines ARCH_WARN_ASM_TRAP_LABEL, a string constant naming the local label at which the trap instruction is placed: x86 "1" (ud2 at label 1:) powerpc "1" (twi at label 1:) riscv "1" (ebreak at label 1:) arm64 "14471" (brk placed at 14471: by __BUG_ENTRY_END) s390 "0" (mc at label 0:) The label is used consistently: in ARCH_WARN_ASM itself, in the bug_addr back-reference inside __BUG_ENTRY / _EMIT_BUG_ENTRY,
quoted hunk ↗ jump to hunk
quoted
quoted
Not really a fan of that. And I can't really tell what you're doing with it either. The kunit is in Rust and thus unreadable :-( I would rather you fix up is_valid_bugaddr(), some architectures seem to have an always true stub because of the callchains always being from the break instruction.Hey Peter, thanks for the suggestion. I will fix up the is_valid_bugaddr for powerpc. I am not very good with other archs asm but i can do it for ppc32 and ppc64.Untested thingies for arm64 and s390x.diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c index 914282016069..3d8a969eab5f 100644 --- a/arch/arm64/kernel/traps.c +++ b/arch/arm64/kernel/traps.c@@ -987,14 +987,13 @@ void do_serror(struct pt_regs *regs, unsigned long esr) #ifdef CONFIG_GENERIC_BUG int is_valid_bugaddr(unsigned long addr) { - /* - * bug_brk_handler() only called for BRK #BUG_BRK_IMM. - * So the answer is trivial -- any spurious instances with no - * bug table entry will be rejected by report_bug() and passed - * back to the debug-monitors code and handled as a fatal - * unexpected debug exception. - */ - return 1; + u32 insn; + + if (aarch64_insn_read((u32 *)addr, &insn)) + return 0; + + /* Match ASM_BUG_FLAGS() / BUG() instructions. */ + return insn == (0xd4200000 | (BUG_BRK_IMM << 5)); } #endif
What problem is this trying to solve? Mukesh's original patch seems to be trying to check whether we compiled things correctly (such that the bug_table entry points at the expected insturction), and I don't think that warrants a runtime check, but (as per my reply to him) I don't understand why that's a thing to check in the first place. On arm64 we only call is_valid_bugaddr() under do_el1_brk64() -> call_el1_break_hook() -> bug_brk_handler(). We only call bug_brk_handler() when HW has taken a BRK exception with the BUG_BRK_IMM immediate, and that can only be triggered by executing the right instruction (BRK #BUG_BRK_IMM), which we ONLY use for BUG(). and WARN(). So I don't see why any change is necessary here. If the HW has somehow taken a BRK exception from any other instruction (which would presumably be a HW bug), there won't be a bug_table entry, and so __report_bug() will fail to find an entry, and will return BUG_TRAP_TYPE_NONE, same as for the !is_valid_bugaddr() case. Mark.