Re: [PATCH] kprobe: safely access memory specified by userspace

[Changbin Du]

On Thu, Feb 14, 2019 at 08:04:18AM -0500, Steven Rostedt wrote:
> On Thu, 14 Feb 2019 08:05:24 +0800
> Changbin Du <changbin.du@xxxxxxxxx> wrote:
> 
> > On Wed, Feb 13, 2019 at 10:41:43AM -0500, Steven Rostedt wrote:
> > > On Wed, 13 Feb 2019 22:36:40 +0800
> > > Changbin Du <changbin.du@xxxxxxxxx> wrote:
> > >   
> > > > Hi Steven,
> > > > I think this is a critical issue. Could you give priority to this fix?
> > > > 
> > > > On Fri, Jan 25, 2019 at 11:10:50PM +0800, Changbin Du wrote:  
> > > > > The userspace can ask kprobe to intercept strings at any memory address,
> > > > > including invalid kernel address. In this case, fetch_store_strlen()
> > > > > would crash since it uses general usercopy function.
> > > > > 
> > > > > For example, we can crash the kernel by doing something as below:
> > > > > 
> > > > > $ sudo kprobe 'p:do_sys_open +0(+0(%si)):string'  
> > > 
> > > Note, I'm not able to reproduce this.
> > > 
> > > I just get:
> > > 
> > >         sendmail-1085  [001] ....   277.344573: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1550  [003] ....   279.879011: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1550  [003] ....   279.879056: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1550  [003] ....   279.879079: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1550  [003] ....   279.879132: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1550  [003] ....   279.879683: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1550  [003] ....   279.881521: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1550  [003] ....   279.881541: open: (do_sys_open+0x0/0x210) arg1=""
> > >            <...>-1597  [005] ....   280.907662: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1597  [005] ....   280.907694: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1597  [005] ....   280.907772: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1597  [005] ....   280.907825: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1597  [005] ....   280.907891: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > >            <...>-1597  [005] ....   280.907947: open: (do_sys_open+0x0/0x210) arg1=(fault)
> > > 
> > >   
> > > > > 
> > > > > [  103.620391] BUG: GPF in non-whitelisted uaccess (non-canonical address?)
> > > > > [  103.622104] general protection fault: 0000 [#1] SMP PTI
> > > > > [  103.623424] CPU: 10 PID: 1046 Comm: cat Not tainted 5.0.0-rc3-00130-gd73aba1-dirty #96
> > > > > [  103.625321] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.12.0-2-g628b2e6-dirty-20190104_103505-linux 04/01/2014
> > > > > [  103.628284] RIP: 0010:process_fetch_insn+0x1ab/0x4b0  
> > > 
> > > What line number is the RIP on?
> > >  
> > I still can reproduce this bug on mainline (1f947a7a011fcceb14cb912f5481a53b18f1879a).
> > But it seems your linux has already fix this issue.
> 
> 
> No I didn't have the fix. I was running an older kernel actually. One
> before commit 9da3f2b74054406f87dff7101a569217ffceb29b was added.
> There's nothing actually wrong with that code, since kprobes is allowed
> to poke at anything. But that commit considers the kernel using copy
> from user to poke kernel address space is a security bug.
> 
Glade to know that. And I wonder wether all such cases have been
disclosed. I noticed the uprobe code also uses some usercopy functions.

> So yeah, I agree your patch should be added with a stable tag, with a
> Fixes: with that commit, since that commit is what causes it to bug.
> 
> I'll apply it and start testing it.
> 
> -- Steve

-- 
Cheers,
Changbin Du