On Tue, 3 Jun 2014, Peter Zijlstra wrote: > On Mon, Jun 02, 2014 at 12:56:40PM -0700, James Bottomley wrote: > > Architecturally, there is a way we could emulate the atomic exchange > > instructions. We could have a special section of memory that always > > triggers a page trap. In the Q state dtlb trap handlers we could > > recognise the "atomic" section of memory and wrap the attempted > > modification in a semaphore. This would add a bit of overhead, but not > > a huge amount if we do it in the trap handlers like the TMPALIAS > > flushes. This involves a lot of work for us because we have to decode > > the instructions in software, recognise the operations and manually > > apply the hashed semaphores around them. If we did it like this, all > > we'd need by way of mainline support is that variables treated as > > atomically exchangeable should be in a separate section (because it's a > > page fault handler effectively, we need them all separated from "normal" > > code). This would probably require some type of variable marker and if > > we ever saw a xchg or cmpxchg on a variable without the marker, we could > > break the build. > > Cute, but I don't think that's entirely feasible given how these things > can be embedded in other structures (some dynamically allocated etc..). We could deliberately misalign all the atomic variables - then, we would take the alignment trap (that is already written) and take the atomic spinlock in it. I've got another idea - we could stop the other CPUs while xchg or cmpxchg is being executed. But there is a problem if the other CPU has interrupts disabled. Could we mask interrupts on PA-RISC in such a way that they are all disabled except one IPI that stops the CPU temporarily? Maybe do not mask interrupts with PSW I-bit and mask them with EIEM instead (leaving the one interrupt for cmpxchg IPI enabled)? Mikulas -- To unsubscribe from this list: send the line "unsubscribe linux-parisc" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
