On Mon, Nov 27, 2023 at 12:16:43PM +0100, Peter Zijlstra wrote:
On Fri, Nov 24, 2023 at 09:51:53PM -0500, Guo Ren wrote:
On Fri, Nov 24, 2023 at 11:15:19AM +0100, Peter Zijlstra wrote:
On Fri, Nov 24, 2023 at 08:21:37AM +0100, Christoph Muellner wrote:
From: Christoph Müllner christoph.muellner@vrull.eu
The upcoming RISC-V Ssdtso specification introduces a bit in the senvcfg CSR to switch the memory consistency model at run-time from RVWMO to TSO (and back). The active consistency model can therefore be switched on a per-hart base and managed by the kernel on a per-process/thread base.
You guys, computers are hartless, nobody told ya?
This patch implements basic Ssdtso support and adds a prctl API on top so that user-space processes can switch to a stronger memory consistency model (than the kernel was written for) at run-time.
I am not sure if other architectures support switching the memory consistency model at run-time, but designing the prctl API in an arch-independent way allows reusing it in the future.
IIRC some Sparc chips could do this, but I don't think anybody ever exposed this to userspace (or used it much).
IA64 had planned to do this, except they messed it up and did it the wrong way around (strong first and then relax it later), which lead to the discovery that all existing software broke (d'uh).
I think ARM64 approached this problem by adding the load-acquire/store-release instructions and for TSO based code, translate into those (eg. x86 -> arm64 transpilers).
Keeping global TSO order is easier and faster than mixing acquire/release and regular load/store. That means when ssdtso is enabled, the transpiler's load-acquire/store-release becomes regular load/store. Some micro-arch hardwares could speed up the performance.
Why is it faster? Because the release+acquire thing becomes RcSC instead of RcTSO? Surely that can be fixed with a weaker store-release variant ot something?
The "ld.acq + st.rel" could only be close to the ideal RCtso because maintaining "ld.acq + st.rel + ld + st" is more complex in LSU than "ld + st" by global TSO. So, that is why we want a global TSO flag to simplify the micro-arch implementation, especially for some small processors in the big-little system.
The problem I have with all of this is that you need to context switch this state and that you need to deal with exceptions, which must be written for the weak model but then end up running in the tso model -- possibly slower than desired.
The s-mode TSO is useless for the riscv Linux kernel and this patch only uses u-mode TSO. So, the exception handler and the whole kernel always run in WMO.
Two years ago, we worried about stuff like io_uring, which means io_uring userspace is in TSO, but the kernel side is in WMO. But it still seems like no problem because every side has a different implementation, but they all ensure their order. So, there should be no problem between TSO & WMO io_uring communication. The only things we need to prevent are: 1. Do not let the WMO code run in TSO mode, which is inefficient. (you mentioned) 2. Do not let the TSO code run in WMO mode, which is incorrect.
If OTOH you only have a single model, everything becomes so much simpler. You just need to be able to express exactly what you want.
The ssdtso is no harm to the current WMO; it's just a tradeoff for micro-arch implementation. You still could use "ld + st" are "ld.acq + st.rl", but they are the same in the global tso state.