Re: [PATCH net-next v3 1/2] bpf: Add bpf_copy_to_user BPF helper to be called in tracers (kprobes)

From: Daniel Borkmann
Date: Tue Jul 19 2016 - 19:20:04 EST


On 07/19/2016 06:34 PM, Alexei Starovoitov wrote:
On Tue, Jul 19, 2016 at 01:17:53PM +0200, Daniel Borkmann wrote:
+ return -EINVAL;
+
+ /* Is this a user address, or a kernel address? */
+ if (!access_ok(VERIFY_WRITE, to, size))
+ return -EINVAL;
+
+ return probe_kernel_write(to, from, size);

I'm still worried that this can lead to all kind of hard to find
bugs or races for user processes, if you make this writable to entire
user address space (which is the only thing that access_ok() checks
for). What if the BPF program has bugs and writes to wrong addresses
for example, introducing bugs in some other, non-intended processes
elsewhere? Can this be limited to syscalls only? And if so, to the
passed memory only?

my understanding that above code will write only to memory of current process,
so impact is contained and in that sense buggy kprobe program is no different
from buggy seccomp prorgram.

Compared to seccomp, you might not notice that a race has happened,
in seccomp case you might have killed your process, which is visible.
But ok, in ptrace() case it might be similar issue perhaps ...

The asm-generic version does __access_ok(..) { return 1; } for nommu
case, I haven't checked closely enough whether there's actually an arch
that uses this, but for example arm nommu with only one addr space would
certainly result in access_ok() as 1, and then you could also go into
probe_kernel_write(), no?

Don't know that code well enough, but I believe the check would only
ensure in normal use-cases that user process doesn't fiddle with kernel
address space, but not necessarily guarantee that this really only
belongs to the process address space.

x86 code comments this with "note that, depending on architecture,
this function probably just checks that the pointer is in the user
space range - after calling this function, memory access functions may
still return -EFAULT".

Also, what happens in case of kernel thread?

As it stands, it does ...

if (unlikely(in_interrupt()))
return -EINVAL;
if (unlikely(!task || !task->pid))
return -EINVAL;

So up to here, irq/sirq, NULL current and that current is not the 'idle'
process is being checked (still fail to see the point for the !task->pid,
I believe the intend here is different).

/* Is this a user address, or a kernel address? */
if (!access_ok(VERIFY_WRITE, to, size))
return -EINVAL;

Now here. What if it's a kernel thread? You'll have KERNEL_DS segment,
task->pid was non-zero as well for the kthread, so access_ok() will
pass and you can still execute probe_kernel_write() ...

Limiting this to syscalls will make it too limited.
I'm in favor of this change, because it allows us to experiment
with restartable sequences and lock-free algorithms that need ultrafast
access to cpuid without burdening the kernel with stable abi.

Have you played around with ptrace() to check whether you could
achieve similar functionality (was thinking about things like [1],
PTRACE_PEEK{TEXT,DATA} / PTRACE_POKE{TEXT,DATA}). If not, why can't
this be limited to a similar functionality for only the current task.
ptrace() utilizes helpers like access_process_vm(), maybe this can
similarly be adapted here, too (under the circumstances that sleeping
is not allowed)?

If we hack access_process_vm I think at the end it will look like
probe_kernel_write. Walking mm requires semaphore, so we would only
be able to do it in task_work and there we can do normal copy_to_user
just as well, but it will complicate the programs quite a bit, since
writes will be asynchronous and batched.
Looks like with access_ok+probe_write we can achieve the same
with a lot less code.

I believe it may not quite be the same as it currently stands. No
fundamental objection, just that this needs to be made "safe" to the
limits you state above yourself. ;)

As far as races between user and bpf program, yeah, if process
is multithreaded, the other threads may access the same mem that
bpf is writing to, but that's no different from reading.
For tracing we walk complex datastructures and pointers. They
can be changed by user space on the fly and bpf will see garbage.
Like we have uprobe+bpf that walks clang c++ internal datastructures
to figure out how long it takes clang to process .h headers.
There is a lot of fragility in the bpf script, yet it's pretty
useful to quickly analyze compile times.
I see bpf_copy_to_user to be hugely valuable too, not as a stable
interface, but rather as a tool to quickly debug and experiment.

Right, so maybe there should be a warn once to the dmesg log that this
is just experimental?

Thanks,
Daniel