On 13.06.25 09:31, Xin Li wrote:

On 6/11/2025 5:58 AM, Juergen Gross wrote:

quoted

quoted

Here is a patch I cooked.  I added an ALTERNATIVE() hack because the new 
instructions can't be more than 6 bytes long.  But with the patch you
just sent, it shouldn't be needed.

I have meanwhile dropped the patch copying the original indirect call.

Reason is that I'm seeing a potential risk with current alternative
patching when using ALTERNATIVE_[23](): depending on the tested features
it might happen that an instruction sequence not suitable for the current
runtime environment is patched in as an intermediate step. In case there
is an interrupt happening just then AND the handling of the interrupt is
using the patch site, this could result in crashes or undefined behavior.

Oh, I had assumed that Linux disables interrupts during the patching
process. Just out of curiosity, why are interrupts allowed in this case?

Interrupts are disabled within text_poke_early() while patching a single
instance.

I guess keeping interrupts disabled during the complete apply_alternatives()
handling would potentially result in a too long period without handling any
interrupts.


Juergen