On Mon, Jul 10, 2017 at 01:00:29PM -0700, Palmer Dabbelt wrote: > On Thu, 06 Jul 2017 08:45:13 PDT (-0700), will.deacon@xxxxxxx wrote: > > On Thu, Jul 06, 2017 at 08:34:27AM -0700, Christoph Hellwig wrote: > >> On Thu, Jul 06, 2017 at 09:55:03AM +0100, Will Deacon wrote: > >> > Agreed on the indirection; it feels like this is something that should be in > >> > the vDSO, which could use the cmpxchg instruction if it's available, or > >> > otherwise just uses plain loads and stores. > > These are already in the vDSO, and use the corresponding atomic instructions on > systems with the A extension. The vDSO routines call the system calls in non-A > systems. As far as I can tell that's necessary to preserve atomicity, which we > currently do by disabling scheduling. If there's a way to do this without > entering the kernel then I'd be happy to support it, but I'm not sure how we > could maintain atomicity using only regular loads and stores. Take a look at the ARM code I mentioned. You can do away with the syscall if you notice that you preempt a thread inside the critical section of the vDSO, and, in that case you resume execution at a known "restart" address. > >> Even that seems like a lot of indirection for something that is in > >> the critical fast path for synchronization. I really can't understand > >> how a new ISA / ABI could even come up with an idea as stupid as making > >> essential synchronization primitives optional. > > > > No disagreement there! > > The default set of multilibs on Linux are: > > * rv32imac: 32-bit; Multiply, Atomic, and Compressed extensions > * rv32imafdc: like above, but with single+double float > * rv64imac: 64-bit, Multiply, Atomic and Compressed > * rv64imafdc: like above, but with single+double float > > all of which support the A extension. We certainly don't plan on building any > systems that support Linux without the A extension at SiFive, so I'm fine > removing the system call -- this was originally added by a user, so there was > at least enough interest for someone to add the system call. > > We've found people are retrofitting other cores to run RISC-V, and I could > certainly imagine an older design that lacks a beefy enough memory system to > support our atomics (which are LR/SC based) being a design that might arise. > There's a lot of systems where people don't seem to care that much about the > performance and just want something to work -- if they're on such a tiny system > they can't implement the A extension then they're probably not going to be > doing a lot of atomics anyway, so maybe it doesn't matter if atomics are slow. > As the cost for supporting these A-less systems seems fairly small, it seemed > like the right thing to do -- one of the points of making RISC-V have many > optional extensions was to let people pick the ones they view as important. > Since I don't know the performance constraints of their systems or the cost of > implementing the A extension in their design, I'm not really qualified to tell > them a cmpxchg syscall is a bad idea. The problem is that by supporting these hypothetical designs that can't do atomics, you hurt sensible designs that *can* do the atomics because you force them to take an additional indirection that could otherwise be avoided. Will
