Skip to contentSkip to conversation

Withdraw: [PATCH] x86-64: fix another kernel data leak to 32-bit processes

DISCUSSION Jan Beulich 2 participants 2 messages Oct 20, 2009

lore
Share & tools
lore
#0 lore raw
Quoted 10 lines

Unfortunately I didn't realize that the other instances of branches to int_ret_from_sys_call also need fixing when preparing the previous similar patch. The issue fixed here was in fact introduced by an earlier patch of mine (295286a89107c353b9677bc604361c537fd6a1c0, i.e. in 2.6.28, but through stable now also present in 2.6.27), making kernel stack contents potentially visible through R8...R11 when an this or earlier syscall got interrupted prior to the handler being able to decrement the stack pointer (such that the space normally used by those registers within pt_regs would get overwritten by the interrupt handler stub).

That analysis wasn't right after all - there's a CLEAR_RREGS in each of the modified paths already, so the change is unnecessary (and adding redundant code). Please don't apply it.

Quoted 3 lines

While touching the code, I also swapped the branch pairs so that the static branch prediction logic would consider the syscall-number-in- range case the taken path.

If this would seem a worthwhile change, I can re-send it as a separate, lower priority patch...

Jan

#1 lore raw
Quoted 15 lines · Jan Beulich
Quoted 10 lines

Unfortunately I didn't realize that the other instances of branches to int_ret_from_sys_call also need fixing when preparing the previous similar patch. The issue fixed here was in fact introduced by an earlier patch of mine (295286a89107c353b9677bc604361c537fd6a1c0, i.e. in 2.6.28, but through stable now also present in 2.6.27), making kernel stack contents potentially visible through R8...R11 when an this or earlier syscall got interrupted prior to the handler being able to decrement the stack pointer (such that the space normally used by those registers within pt_regs would get overwritten by the interrupt handler stub).

That analysis wasn't right after all - there's a CLEAR_RREGS in each of the modified paths already, so the change is unnecessary (and adding redundant code). Please don't apply it.

Quoted 6 lines
Quoted 3 lines

While touching the code, I also swapped the branch pairs so that the static branch prediction logic would consider the syscall-number-in- range case the taken path.

If this would seem a worthwhile change, I can re-send it as a separate, lower priority patch...

Please do - thanks!

Ingo