[PATCH 3.19-rc2 v13 4/5] ARM: Add support for on-demand backtrace of other CPUs

[Steven Rostedt]

On Fri, 9 Jan 2015 16:48:01 +0000
Russell King - ARM Linux <linux at arm.linux.org.uk> wrote:

> On Mon, Jan 05, 2015 at 10:19:25AM -0500, Steven Rostedt wrote:
> > On Mon,  5 Jan 2015 14:54:58 +0000
> > Daniel Thompson <daniel.thompson at linaro.org> wrote:
> > > +/* For reliability, we're prepared to waste bits here. */
> > > +static DECLARE_BITMAP(backtrace_mask, NR_CPUS) __read_mostly;
> > > +static  cpumask_t printtrace_mask;
> > > +
> > > +#define NMI_BUF_SIZE		4096
> > > +
> > > +struct nmi_seq_buf {
> > > +	unsigned char		buffer[NMI_BUF_SIZE];
> > > +	struct seq_buf		seq;
> > > +};
> 
> Am I missing something or does this limit us to 4096 characters of
> backtrace output per CPU?
> 
> > This is the same code as in x86. I wonder if we should move the
> > duplicate code into kernel/printk/ and have it compiled if the arch
> > requests it (CONFIG_ARCH_WANT_NMI_PRINTK or something). That way we
> > don't have 20 copies of the same nmi_vprintk() and later find that we
> > need to change it, and have to change it in 20 different archs.
> 
> Agreed, though I wonder about the buffer size.
> 

Have we had kernel back traces bigger than that? Since the stack size
is limited to page size, it would seem dangerous if backtraces filled
up a page size itself, as most function frames are bigger than the
typical 60 bytes of data per line.

We could change that hard coded 4096 to PAGE_SIZE, for those archs with
bigger pages.

Also, if the backtrace were to fill up that much. Most the pertinent
data from a back trace is at the beginning of the trace. Seldom do we
care about the top most callers (bottom of the output).

-- Steve