On Thu, Jul 20, 2017 at 04:04:43PM +0100, Matt Redfearn wrote:
> Commit 1c3c5eab1715 ("sched/core: Enable might_sleep() and
> smp_processor_id() checks early") enables checks for might_sleep() and
> smp_processor_id() being used in preemptible code earlier in the boot
> than before. This results in a new BUG from
> pcibios_set_cache_line_size().
>
> BUG: using smp_processor_id() in preemptible [00000000] code:
> swapper/0/1
> caller is pcibios_set_cache_line_size+0x10/0x70
> CPU: 1 PID: 1 Comm: swapper/0 Not tainted 4.13.0-rc1-00007-g3ce3e4ba4275 #615
> Stack : 0000000000000000 ffffffff81189694 0000000000000000 ffffffff81822318
> 000000000000004e 0000000000000001 800000000e20bd08 20c49ba5e3540000
> 0000000000000000 0000000000000000 ffffffff818d0000 0000000000000000
> 0000000000000000 ffffffff81189328 ffffffff818ce692 0000000000000000
> 0000000000000000 ffffffff81189bc8 ffffffff818d0000 0000000000000000
> ffffffff81828907 ffffffff81769970 800000020ec78d80 ffffffff818c7b48
> 0000000000000001 0000000000000001 ffffffff818652b0 ffffffff81896268
> ffffffff818c0000 800000020ec7fb40 800000020ec7fc58 ffffffff81684cac
> 0000000000000000 ffffffff8118ab50 0000000000000030 ffffffff81769970
> 0000000000000001 ffffffff81122a58 0000000000000000 0000000000000000
> ...
> Call Trace:
> [<ffffffff81122a58>] show_stack+0x90/0xb0
> [<ffffffff81684cac>] dump_stack+0xac/0xf0
> [<ffffffff813f7050>] check_preemption_disabled+0x120/0x128
> [<ffffffff818855e8>] pcibios_set_cache_line_size+0x10/0x70
> [<ffffffff81100578>] do_one_initcall+0x48/0x140
> [<ffffffff81865dc4>] kernel_init_freeable+0x194/0x24c
> [<ffffffff8169c534>] kernel_init+0x14/0x118
> [<ffffffff8111ca84>] ret_from_kernel_thread+0x14/0x1c
>
> Fix this by using raw_current_cpu_data instead.
>
> Signed-off-by: Matt Redfearn <matt.redfearn@xxxxxxxxxx>
> ---
>
> In heteregenerous systems the more correct fix for this would be to
> iterate over CPUs checking each ones cache hierarchy. However, as no
> such systems currently exist that seems wasteful.
I fear this may return to bite us at some point.
How about switching to using cpu_*cache_line_size(), which uses
cpu_data[0]? That way we'll at least be able to grep for users when the
macros are later fixed or removed.
Admitedly there is no cpu_tcache_line_size(), but it could be trivially
added.
Cheers
James
>
> ---
> arch/mips/pci/pci.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/arch/mips/pci/pci.c b/arch/mips/pci/pci.c
> index bd67ac74fe2d..7ef8d97fa324 100644
> --- a/arch/mips/pci/pci.c
> +++ b/arch/mips/pci/pci.c
> @@ -28,7 +28,7 @@ EXPORT_SYMBOL(PCIBIOS_MIN_MEM);
>
> static int __init pcibios_set_cache_line_size(void)
> {
> - struct cpuinfo_mips *c = ¤t_cpu_data;
> + struct cpuinfo_mips *c = &raw_current_cpu_data;
> unsigned int lsize;
>
> /*
> --
> 2.7.4
>
>
Attachment:
signature.asc
Description: Digital signature
