Messages in this thread | | | Date | Wed, 12 Jan 2022 12:13:03 +0800 | From | Pingfan Liu <> | Subject | Re: [PATCH v2] tracing: Add test for user space strings when filtering on string pointers |
| |
Hi Steven,
Sorry that I am out of office, and not reply in time.
On Mon, Jan 10, 2022 at 12:24:36PM -0500, Steven Rostedt wrote: > On Mon, 10 Jan 2022 17:11:52 +0000 > David Laight <David.Laight@ACULAB.COM> wrote: > > > From: Steven Rostedt > > > Sent: 10 January 2022 16:56 > > > > > > From: Steven Rostedt <rostedt@goodmis.org> > > > > > > Pingfan reported that the following causes a fault: > > > > > > echo "filename ~ \"cpu\"" > events/syscalls/sys_enter_openat/filter > > > echo 1 > events/syscalls/sys_enter_at/enable > > > > > > The reason is that trace event filter treats the user space pointer > > > defined by "filename" as a normal pointer to compare against the "cpu" > > > string. If the string is not loaded into memory yet, it will trigger a > > > fault in kernel space:
For accurate commit log, the swapped-out user page is not the root cause of this bug is "supervisor read access in kernel mode". And it is trueth that swapped-out user page can trigger a bug here, but it should be a different a stack.
> > > > If a userspace pointer can end up the kernel structure then presumably > > a 'dodgy' user program can supply an arbitrary kernel address instead? > > This may give the user the ability to read arbitrary kernel addresses > > (including ones that are mapped to PCIe IO addresses). > > Doesn't sound good at all. > > Only root has access to the information read here. All tracing requires > root or those explicitly given access to the tracing data, which pretty > much allows all access to kernel internals (including all memory). So > nothing to worry about here ;-) >
I am not sure about the opposite way. Since kernel is not allowed to access userspace most of the time, then is it an leakage, which looks like: use tracepoint as trampoline to uaccess. read out user info from ustring_per_cpu
But any kernel code can call copy_from_user() function family freely, so it is not a problem caused by this patch, right? Or ustring_per_cpu should be zeroed out.
For V2, feel free to add "Tested-by"
Thanks,
Pingfan > > > > ... > > > + if (likely((unsigned long)str >= TASK_SIZE)) { > > > + /* For safety, do not trust the string pointer */ > > > + if (!strncpy_from_kernel_nofault(kstr, str, USTRING_BUF_SIZE)) > > > + return NULL; > > > + } else { > > > + /* user space address? */ > > > + ustr = (char __user *)str; > > > + if (!strncpy_from_user_nofault(kstr, ustr, USTRING_BUF_SIZE)) > > > + return NULL; > > > > Is that check against TASK_SIZE even correct for all architectures? > > copy_to/from_user() uses access_ok() - which is architecture dependant. > > The problem with access_ok() (which I tried first) is that it can't be used > from interrupt context, and this check can happen in interrupt context. > Either way, if we pick the wrong one for the arch, the only thing bad that > can happen is that it returns "fault" and the filter fails, just like if > the pointer was to bad memory. > > > > > I think you need to remember where the pointer came from. > > > > Why? > > -- Steve
| |