On Mon, 2024-11-04 at 14:52 +0000, Andrew Cooper wrote: > > Unfortunately, that's all we have right now in the official > > documentation. > > > > I've put up some notes in > > https://amitshah.net/2024/11/eraps-reduces-software-tax-for-hardware-bugs/ > > I appreciate the attempt to get a few details out, but this is very > confused on bunch of details. > > Most importantly, you've described Intel RSB underflows, but named it > AMD BTC. > > "Retbleed" is two totally different things. I begged the > discoverers > to give it two names, and I also begged the x86 maintainers to not > alias > them in Linux's view of the world, but alas. > > AMD's BTC comes from a bad branch type prediction, and a late resteer > from the ret uop executing. It has nothing to do with RAS/RSB > underflow conditions. BTC indeed is only branch-type confusion. The point I wanted to make there is that to entirely get rid of X86_FEATURE_RSB_CTXW, I had to confirm that AMD CPUs do not speculate return addresses from the BTB or BHB since BTC was fixed. (Or, in other words, to clarify the previous comments there that said that AMD predicts from the BTB/BHB in every case). So - the only point in saying BTC_NO is relevant here is me confirming that AMD is not going to speculate return addresses from outside of the RSB. And that comment can now reflect reality. Amit
