On 02/11/2019 10:15 AM, Thomas Gleixner wrote:
On Mon, 11 Feb 2019, Alexandre Chartre wrote:
On 02/10/2019 10:23 PM, Thomas Gleixner wrote:
On Fri, 25 Jan 2019, Alexandre Chartre wrote:
Note that this issue has been observed and reproduced with a custom kernel
with some code mapped to different virtual addresses and using jump labels
As jump labels use text_poke_bp(), crashes were sometimes observed when
updating jump labels.
Rightfully so. text_poke_bp() pokes at the kernels text mapping which is
very well defined and unique. Why would you map the same text to different
virtual addresses and then use a randomly chosen one to poke at it?
As an example, we used to have per-CPU SYSCALL entry trampoline [1] where the
entry_SYSCALL_64_trampoline function was mapped to a different virtual address
for each CPU. So, a syscall would execute entry_SYSCALL_64_trampoline() from a
different virtual address depending on the CPU being used. With that code,
adding a jump label in entry_SYSCALL_64_trampoline() causes the described
issue.
Right, but we knew that and there was no reason to put a jump label into
that. AFAICT we don't have anything like this in the kernel.
In our particular case, we have introduced a jump label in JMP_NOSPEC (which
is used by entry_SYSCALL_64_trampoline()) to have the option to dynamically
enable/disable retpoline at runtime. So that's when we faced this issue.
That said, I'm not opposed to apply the patch as is, I just wanted to get a
better understanding about the background.
Sure. I am aware this is a corner case, and I was precisely looking for feedback
to check if it is worth fixing that issue. So I appreciate your reply, and I would
understand if the patch is rejected because that's a case that we are just not
expecting to happen.