On Sat, Jul 12, 2014 at 05:10:35PM +0800, Huacai Chen wrote: > Hi, Paul, > > You means my patch (http://patchwork.linux-mips.org/patch/7297/) is > the correct way? I believe you patch will fix the problem, but I think it would be better to remove the check for !preemptible() & the BUG_ON entirely. > Another question: Your patch > (http://patchwork.linux-mips.org/patch/7307/) remove > preempt_disable()/preempt_enable() in init_fpu(). It will cause > problems if there is another function call init_fpu() because it is > previously preempt-safe. Maybe introduce a new function (e.g. > __init_fpu()) is a better way? It may cause a problem if there were other callers, but there is only one caller of init_fpu (enable_restore_fp_context) and it needs to disable preemption for longer than the init_fpu function anyway. I see no value in keeping init_fpu as a wrapper that disables preemption when there would be nothing calling it. Thanks, Paul > Huacai > > On Sat, Jul 12, 2014 at 7:28 AM, Chen Jie <chenj@xxxxxxxxxx> wrote: > > 2014-07-11 23:56 GMT+08:00 Paul Burton <paul.burton@xxxxxxxxxx>: > >> On Fri, Jul 11, 2014 at 11:14:13AM +0800, chenj wrote: > >>> do_ade may be invoked with preempt enabled. do_cpu will be invoked with > >>> preempt enabled. When it's preempted(in do_ade/do_cpu), TIF_USEDFPU will be > >>> cleared, when it returns to do_ade/do_cpu, the fpu is actually disabled. > >>> > >>> e.g. > >>> In do_ade() > >>> emulate_load_store_insn(): > >>> BUG_ON(!is_fpu_owner()); <-- This assertion may be breaked. > >>> > >>> In do_cpu() > >>> enable_restore_fp_context(): > >>> was_fpu_owner = is_fpu_owner(); > >> > >> Preemption should indeed be disabled around the assignment & use of the > >> was_fpu_owner variable, but note that you can only hit the problem if > >> using MSA. One of the MSA fixes I just submitted also fixes this along > >> with another instance of the problem: > >> > >> http://patchwork.linux-mips.org/patch/7307/ > >> > >> I prefer my patch to this since it disables preemption for less time, > >> in addition to fixing the !used_math() case. > >> > >> In emulate_load_store_insn I believe the correct fix is simply to remove > >> that BUG_ON. The code is about to give up FPU ownership anyway, so it's > >> not like there is any requirement being violated if it was already lost. > > Yes, you're right. > > > > """ /* arch/mips/kernel/unaligned.c */ > > lose_fpu(1); /* Save FPU state for the emulator. */ > > res = fpu_emulator_cop1Handler(regs, ¤t->thread.fpu, 1, &fault_addr); > > own_fpu(1); /* Restore FPU state. */ > > """ > > > > Going deep into the code, I find lost_fpu(1) will save fpu context if > > owns fpu (otherwise, if preempted, the fpu context will be saved in > > process switch), then fpu_emulator_cop1Handler manipulates the saved > > fpu context, own_fpu(1) restores it. > > > > So, remove "BUG_ON(!is_fpu_owner())" is OK. > >
