[PATCH 0/4] CRASH_ZEROIZE: Wipe secrets before kdump

[Jan Sebastian Götte]

On 8/2/26 07:08, Dave Young wrote:
> On 8/2/26 12:31 AM, Jan Sebastian Götte wrote:
>> On 8/1/26 16:03, Baoquan He wrote:
>>> Note that we usually dont' want to run a lot of work after panic and
>>> before jumping into kdump kernel.
>>
>> I understand. For this reason, I think it's best to keep this default-off. As-is, the notifier list call is timed and on the (slow) ARM64 target I'm using, it takes about 3-5 ms to run. I took the "try lock, skip if locked" approach to keep the risk of this code crashing during panic minimal. In my application, the kdump payload is code that then does a full wipe, taking a couple hundred milliseconds.
> 
> Not only about the time used, the panicked kernel is not reliable, any more extra logic can make it even not reliable, any pre-kdump extra logic is not a good idea unless it is a must to ensure kdump working.
> 
> Cleaning up secret data can be done with makedumpfile + eppic scripts (see the manual of makedumpfile), or it is even possible to do so in kdump kernel with Tao Liu's improvments for makedumpfile previously  (I don't know the status, probably dropped for the time being, but it is possible, cced him).

Thank you for the pointer!

There's two scenarios worth considering. First, in the standard scenario 
where you enable this option, then drop into a standard kdump kernel, I 
don't think it makes a big difference *when* you do this cleanup since 
someone is going to have to dereference these pointers. IMHO a good 
reason to do it in the old kernel is that there, the code knows about 
the layout of all the data structures. To retroactively do this in the 
kdump kernel is much more complicated, since there you have to 
reconstruct the structure layouts from symbols or hardcoded struct 
layouts, and you have to keep this symbol/layout information perfectly 
in sync with the running kernel.

The second scenario is what I'm working on here: I'm not using a normal 
kdump kernel, but instead a custom payload that wipes all RAM from start 
to end. This payload will wipe all these keys too, but my critical 
concern is speed: On the embedded SoCs I'm targeting, the full memory 
wipe takes too long (hundreds of ms) for an HSM application, so I want 
to do a targeted wipe of just the keys first. The old kernel I think is 
the natural place to do this. Adding to that, in my scenario the most 
likely trigger of a panic is not something like memory corruption, but a 
trigger of the system's tamper alarms, which would leave the old kernel 
relatively stable during panic.

I can imagine several possible mitigations for the stability concerns 
beyond the default off config option:

* Since the wipe handlers are all really simple, it would be possible to 
manually guard every memory access there to ensure they can't fail and 
that they don't write to sensitive areas like the dump kernel or the 
remaining panic'ing stack. Doing that would only rely on information 
(more or less intact stack pointer, kdump kernel area boundaries) that 
would be necessary for kdump to succeed anyway.

* And/Or I could extend the patchset to include a mechanism similar to 
that in Bradley Morgan's patchset that catches segfaults during wipe, 
and then skips the handler causing the fault.

* A last option would be to have the alive kernel prepare some kind of 
"wipe this first" structure in its memory during normal operation that 
the dump kernel then can pick up to do the actual dirty work. I disfavor 
that since it adds double bookkeeping to a lot of places, some of which 
could be performance critical.

I've also picked up that I should remove the timing logic since that 
could cause instability.

Thanks,
Jan Sebastian