Re: [PATCH v3] sysfs: Unconditionally use vmalloc for buffer

[Kees Cook <keescook@chromium.org>]

On Thu, Apr 01, 2021 at 09:10:05AM +0200, Greg Kroah-Hartman wrote:
> On Wed, Mar 31, 2021 at 11:52:20PM -0700, Kees Cook wrote:
> > On Thu, Apr 01, 2021 at 07:16:56AM +0200, Greg Kroah-Hartman wrote:
> > > On Wed, Mar 31, 2021 at 07:21:45PM -0700, Kees Cook wrote:
> > > > The sysfs interface to seq_file continues to be rather fragile
> > > > (seq_get_buf() should not be used outside of seq_file), as seen with
> > > > some recent exploits[1]. Move the seq_file buffer to the vmap area
> > > > (while retaining the accounting flag), since it has guard pages that
> > > > will catch and stop linear overflows. This seems justified given that
> > > > sysfs's use of seq_file already uses kvmalloc(), is almost always using
> > > > a PAGE_SIZE or larger allocation, has normally short-lived allocations,
> > > > and is not normally on a performance critical path.
> > > > 
> > > > Once seq_get_buf() has been removed (and all sysfs callbacks using
> > > > seq_file directly), this change can also be removed.
> > > > 
> > > > [1] https://blog.grimm-co.com/2021/03/new-old-bugs-in-linux-kernel.html
> > > > 
> > > > Signed-off-by: Kees Cook <keescook@chromium.org>
> > > > ---
> > > > v3:
> > > > - Limit to only sysfs (instead of all of seq_file).
> > > > v2: https://lore.kernel.org/lkml/20210315174851.622228-1-keescook@chromium.org/
> > > > v1: https://lore.kernel.org/lkml/20210312205558.2947488-1-keescook@chromium.org/
> > > > ---
> > > >  fs/sysfs/file.c | 23 +++++++++++++++++++++++
> > > >  1 file changed, 23 insertions(+)
> > > > 
> > > > diff --git a/fs/sysfs/file.c b/fs/sysfs/file.c
> > > > index 9aefa7779b29..70e7a450e5d1 100644
> > > > --- a/fs/sysfs/file.c
> > > > +++ b/fs/sysfs/file.c
> > > > @@ -16,6 +16,7 @@
> > > >  #include <linux/mutex.h>
> > > >  #include <linux/seq_file.h>
> > > >  #include <linux/mm.h>
> > > > +#include <linux/vmalloc.h>
> > > >  
> > > >  #include "sysfs.h"
> > > >  
> > > > @@ -32,6 +33,25 @@ static const struct sysfs_ops *sysfs_file_ops(struct kernfs_node *kn)
> > > >  	return kobj->ktype ? kobj->ktype->sysfs_ops : NULL;
> > > >  }
> > > >  
> > > > +/*
> > > > + * To be proactively defensive against sysfs show() handlers that do not
> > > > + * correctly stay within their PAGE_SIZE buffer, use the vmap area to gain
> > > > + * the trailing guard page which will stop linear buffer overflows.
> > > > + */
> > > > +static void *sysfs_kf_seq_start(struct seq_file *sf, loff_t *ppos)
> > > > +{
> > > > +	struct kernfs_open_file *of = sf->private;
> > > > +	struct kernfs_node *kn = of->kn;
> > > > +
> > > > +	WARN_ON_ONCE(sf->buf);
> > > 
> > > How can buf ever not be NULL?  And if it is, we will leak memory in the
> > > next line so we shouldn't have _ONCE, we should always know, but not
> > > rebooting the machine would be nice.
> > 
> > It should never be possible. I did this because seq_file has some
> > unusual buf allocation patterns in the kernel, and I liked the cheap
> > leak check. I use _ONCE because spewing endlessly doesn't help most
> > cases. And if you want to trigger it again, you don't have to reboot:
> > https://www.kernel.org/doc/html/latest/admin-guide/clearing-warn-once.html
> 
> True, I was thinking of the panic-on-warn people, and the hesitation of
> adding new WARN_ON() to the kernel code.  If this really can happen,
> shouldn't we handle it properly?

It should never happen, but I hate silent bugs. Given the existing
pattern of "external preallocation", it seems like a fragile interface
worth asserting our expectations.

The panic_on_warn folks will get exactly what they wanted: immediate
feedback on "expected to be impossible" cases:
https://www.kernel.org/doc/html/latest/process/deprecated.html#bug-and-bug-on

> > > > +	sf->buf = __vmalloc(kn->attr.size, GFP_KERNEL_ACCOUNT);
> > > > +	if (!sf->buf)
> > > > +		return ERR_PTR(-ENOMEM);
> > > > +	sf->size = kn->attr.size;
> > > > +
> > > > +	return NULL + !*ppos;
> > > > +}
> > > 
> > > Will this also cause the vmalloc fragmentation/abuse that others have
> > > mentioned as userspace can trigger this?
> > 
> > If I understood the concern correctly, it was about it being a risk for
> > doing it for all seq_file uses. This version confines the changes to only
> > sysfs seq_file uses.
> 
> There are a few sysfs files that userspace can read from out there :)

Yes, but the vmap area is also used by default for process stacks, etc.
Malicious fragmentation is already possible. I understood the concern to
be about "regular" use. (And if I'm wrong, we can add a knob maybe?)

> > > And what code frees it?
> > 
> > The existing hooks to seq_release() handle this already. This kind of
> > "preallocation" of the seq_file buffer is done in a few places already
> > (hence my desire for the sanity checking WARN lest future seq_file
> > semantics change).
> 
> Ah, "magic", gotta love it...

Yeeeah. :P

-- 
Kees Cook