On 29 March 2016 at 09:31, Sergey Fedorov <serge.f...@gmail.com> wrote: > On 29/03/16 01:12, Richard Henderson wrote: >> On 03/24/2016 08:11 AM, Paolo Bonzini wrote: >>> There is also a case where a TB jumps to itself; it then appears twice >>> in the list with different values in the low bits, such as this: >>> >>> tb->jmp_list_first = tb | 0; >>> .--------------------' | >>> | .-------' >>> tb->jmp_list_next[0] = tb | 2; >> >> Of course, it begs the question of why TB would be in its own list, >> even if it does jump to itself. We only need the points-to list in >> order to invalidate a TB and unlink it. But if TB is being >> invalidated, we don't need to reset the jump within TB itself. > > If we're going to move tb_phys_invalidate() outside of tb_lock, we > probably need to reset all jumps to the TB, even if it jumps to itself, > so that it eventually finish its execution.
This is likely also the historical reason for the current code -- originally we handled requesting a CPU exit by unlinking the TB, so you needed to be able to detach jumps-to-self (these days we do it by checking a flag at the start of each TB). thanks -- PMM
