On Fri, Jan 30, 2015 at 07:55:32PM +0200, Aaro Koskinen wrote:
> Hi,
>
> On Fri, Jan 30, 2015 at 02:59:57PM +0000, James Hogan wrote:
> > On 30/01/15 12:47, Maciej W. Rozycki wrote:
> > > On Fri, 30 Jan 2015, James Hogan wrote:
> > >
> > >>> Hmm, why can a call to `printk' cause a TLB miss, what's so special about
> > >>> this function? Does it use kernel mapped addresses for any purpose such
> > >>> as `vmalloc'?
> > >>
> > >> It would be the fact netconsole (or whatever other console is in use) is
> > >> built as a kernel module, memory for which is allocated from the vmalloc
> > >> area.
> > >
> > > Ah, I see, thanks for enlightening me. But in that case wouldn't it be
> > > possible to postpone console output from `printk' until it is safe to
> > > access the device? In a manner similar to how for example we handle calls
> > > to `printk' made from the hardirq context. That would make things less
> > > fragile.
> >
> > Hmm, kernel/printk/printk.c does have:
> >
> > static inline int can_use_console(unsigned int cpu)
> > {
> > return cpu_online(cpu) || have_callable_console();
> > }
> >
> > which should prevent it dumping printk buffer to console. CPU shouldn't
> > be marked online that early, which suggests that the console has the
> > CON_ANYTIME flag set, which it probably shouldn't if it depends on
> > module code. call_console_drivers() seems to ensure the CPU is online or
> > has CON_ANYTIME before calling the console write callback.
> >
> > A quick glance and I can't see any evidence of netconsole being able to
> > get CON_ANYTIME.
>
> It does not set the flag. But flags are kept in module's static data,
> so the original problem stays.
>
> A.
Ah yes, of course. This approach looks correct to me then.
Cheers
James
Attachment:
signature.asc
Description: Digital signature