From: Jed D. <jl...@mo...> - 2013-07-20 04:47:05
|
On Mon, Jul 15, 2013 at 02:54:20PM +0100, Will Deacon wrote: > On Sat, Jul 13, 2013 at 04:18:20AM +0100, Jed Davis wrote: [...] > > Effects of this are probably limited to failure of EHABI unwinding when > > starting from a function that uses r7 to restore its stack pointer, but > > the possibility for further breakage (which would be invisible on > > non-Thumb kernels) is worrying. [...] > I'm struggling to understand exactly the problem that this patch is trying > to address. If it's just a code consistency issue, I don't think it's worth > it (I actually find it less confusing the way we currently have things) but > if there is a real bug, perhaps you could provide a testcase? There is a real bug here, but my commit message wasn't very clear. This was breaking PERF_COUNT_SW_CONTEXT_SWITCHES with CONFIG_THUMB2_KERNEL=y (with my other recently posted patch applied), because kernel/sched.c is built with -fno-omit-frame-pointer (which is wrong, but that's a problem for another patch) and so __schedule's EHABI entry uses 0x97 (mov sp, r7), and somewhere along the line the unwinder gets the r11 value instead. This would also apply to any function with a variable-length array, but __schedule happens to have the perf hook I was trying to use. I should add that this bug doesn't affect me directly at the moment, because we're not currently using CONFIG_THUMB2_KERNEL on Firefox OS, because our kernels are assorted older versions with hardware vendors' changes and there are some issues with it. But I felt it was worth tracking this down before trying to send changes upstream. The "right" thing to do here might be to just include all the registers, or at least {r4-pc}, in struct stackframe. The parts that aren't {fp, sp, lr, pc} could be ifdef'ed if we're concerned enough about the overhead in kernels using APCS frame pointer unwinding. I agree that a test case would be good -- I'm more than a little worried of regressions without one -- but I could use some advice on how best to do that. --Jed |