Re: [PATCH] fs/proc: allow larger /proc/<pid>/cmdline output

[Jarod Wilson]

On Thu, Apr 09, 2015 at 09:12:00PM -0700, Andrew Morton wrote:
> On Thu,  9 Apr 2015 23:59:02 -0400 Jarod Wilson <jarod@xxxxxxxxxx> wrote:
> 
> > There are people who run java. Sometimes, when it misbehaves, they try to
> > figure out what's going on by dumping /proc/<pid>/cmdline, but the length
> > of that output is currently capped by PAGE_SIZE (so x86_64's 4k, in most
> > cases), and sometimes, java command lines are longer than 4k characters.
> > 
> > This change allows the user to request a larger max length, up to 4x
> > PAGE_SIZE, but the default out-of-the-box setting should keep things the
> > same as they ever were. The 4x maximum is somewhat arbitrary, but seemed
> > like it should be more than enough, because really, if you have more than
> > 16k characters on your command line, you're probably doing it wrong...
> > 
> > I've tested this lightly with non-java shell commands with really long
> > parameters, and things are perfectly stable after several hundred
> > iterations of exercising things on a system booted with both
> > proc_pid_maxlen=8192 and 16384. I wouldn't call my testing exhaustive,
> > and I may not have considered something that will blow up horribly here,
> > so comments and clues welcomed.
> > 
> > Using single_open_size() looked less messy than giving proc_pid_cmdline()
> > its own .start op that would allow multiple buffers.
> > 
> > Note: I've only added this extended sizing for /proc/<pid>/cmdline output,
> > rather than for all /proc/<pid>/foo elements, thinking that nothing else
> > should ever really be that long, but anything that is can simply switch
> > from using the ONE() macro to the ONE_SIZE() macro.
> 
> Why have an upper limit at all?

Just trying to be conservative and keep people from doing anything too
insane, but I didn't really have any particularly good reason beyond that
for capping it. I'll remove the upper bound next iteration.

> > --- a/fs/proc/base.c
> > +++ b/fs/proc/base.c
> > @@ -134,6 +134,30 @@ struct pid_entry {
> >  	NOD(NAME, (S_IFREG|(MODE)), 			\
> >  		NULL, &proc_single_file_operations,	\
> >  		{ .proc_show = show } )
> > +#define ONE_SIZE(NAME, MODE, show)				\
> > +	NOD(NAME, (S_IFREG|(MODE)), 				\
> > +		NULL, &proc_single_file_size_operations,	\
> > +		{ .proc_show = show } )
> > +
> > +/*
> > + * Its hideous, but some java gunk winds up with a cmdline that is longer
> > + * than PAGE_SIZE, and some people want to be able to see all of it for
> > + * debugging purposes. Allocate at least PAGE_SIZE, and allow the user to
> > + * ask for up to PAGE_SIZE << 2 (4x) to help with that situation.
> > + */
> > +static unsigned long proc_pid_maxlen = PAGE_SIZE;
> > +static int set_proc_pid_maxlen(char *str)
> > +{
> > +	if (!str)
> > +		return 0;
> > +
> > +	proc_pid_maxlen = simple_strtoul(str, &str, 0);
> > +	proc_pid_maxlen = max(PAGE_SIZE, proc_pid_maxlen);
> > +	proc_pid_maxlen = min(PAGE_SIZE << 2, proc_pid_maxlen);
> > +
> > +	return 1;
> > +}
> > +__setup("proc_pid_maxlen=", set_proc_pid_maxlen);
> 
> This permits 4k-16k on x86 and 64k-256k on powerpc.  This makes the
> kernel interface inconsistent across architectures, which is not good -
> some applications will work OK on one arch but will fail when moved to
> a different arch.
> 
> s/PAGE_SIZE/4096/g would fix that.

I was going for minimal disturbance to the status quo, which is that
everything is structured around PAGE_SIZE. proc_pid_cmdline() has a
BUG_ON(m->size < PAGE_SIZE) in it and the default initial allocation for
every seq_file using single_open() is PAGE_SIZE, making it a bit more
invasive to change it. Perhaps add a default size #define to seq_file.h
and use that for seq_file everywhere instead of PAGE_SIZE to get
consistency across all arches?

-- 
Jarod Wilson
jarod@xxxxxxxxxx

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/