Re: [PATCH] init: Initialize jump labels before command line option parsing
From: Guenter Roeck
Date: Tue Apr 16 2019 - 20:39:17 EST
On Tue, Apr 16, 2019 at 5:04 PM Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
> On Tue, Apr 16, 2019 at 4:44 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> > On Tue, 16 Apr 2019 13:54:04 -0700 Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
> > > When a module option, or core kernel argument, toggles a static-key it
> > > requires jump labels to be initialized early. While x86, PowerPC, and
> > > ARM64 arrange for jump_label_init() to be called before parse_args(),
> > > ARM does not.
> > >
> > > Kernel command line: rdinit=/sbin/init page_alloc.shuffle=1 panic=-1 console=ttyAMA0,115200 page_alloc.shuffle=1
> > > ------------[ cut here ]------------
> > > WARNING: CPU: 0 PID: 0 at ./include/linux/jump_label.h:303
> > > page_alloc_shuffle+0x12c/0x1ac
> > > static_key_enable(): static key 'page_alloc_shuffle_key+0x0/0x4' used
> > > before call to jump_label_init()
> > > Modules linked in:
> > > CPU: 0 PID: 0 Comm: swapper Not tainted
> > > 5.1.0-rc4-next-20190410-00003-g3367c36ce744 #1
> > > Hardware name: ARM Integrator/CP (Device Tree)
> > > [<c0011c68>] (unwind_backtrace) from [<c000ec48>] (show_stack+0x10/0x18)
> > > [<c000ec48>] (show_stack) from [<c07e9710>] (dump_stack+0x18/0x24)
> > > [<c07e9710>] (dump_stack) from [<c001bb1c>] (__warn+0xe0/0x108)
> > > [<c001bb1c>] (__warn) from [<c001bb88>] (warn_slowpath_fmt+0x44/0x6c)
> > > [<c001bb88>] (warn_slowpath_fmt) from [<c0b0c4a8>]
> > > (page_alloc_shuffle+0x12c/0x1ac)
> > > [<c0b0c4a8>] (page_alloc_shuffle) from [<c0b0c550>] (shuffle_store+0x28/0x48)
> > > [<c0b0c550>] (shuffle_store) from [<c003e6a0>] (parse_args+0x1f4/0x350)
> > > [<c003e6a0>] (parse_args) from [<c0ac3c00>] (start_kernel+0x1c0/0x488)
> > >
> > > Move the fallback call to jump_label_init() to occur before
> > > parse_args(). The redundant calls to jump_label_init() in other archs
> > > are left intact in case they have static key toggling use cases that are
> > > even earlier than option parsing.
> > Has it been confirmed that this fixes
> > mm-shuffle-initial-free-memory-to-improve-memory-side-cache-utilization.patch
> > on beaglebone-black?
> This only fixes dynamically enabling the shuffling on 32-bit ARM.
> Guenter happened to run without the mm-only 'force-enable-always'
> patch and when he went to use the command line option to enable it he
> hit the jump-label warning.
For my part I have not seen the original failure; it seems that the
kernelci logs are no longer present. As such, I neither know how it
looks like nor how to (try to) reproduce it. I just thought it might
be worthwhile to run the patch through my boot tests to see if
anything pops up. From the feedback I got, though, it sounded like the
failure is/was very omap2 specific, so I would not be able to
reproduce it anyway.
> The original beaglebone-black failure was something different and
> avoided this case because the jump-label was never used.
> I am otherwise unable to recreate the failure on either the original
> failing -next, nor on v5.1-rc5 plus the latest state of the patches. I
> need from someone who is actually still seeing the failure so they can
> compare with the configuration that is working for me. For reference
> that's the Yocto beaglebone-black defconfig: