On 10.07.2013, at 20:42, Scott Wood wrote: > On 07/10/2013 05:15:09 AM, Alexander Graf wrote: >> On 10.07.2013, at 02:06, Scott Wood wrote: >> > On 07/09/2013 04:44:24 PM, Alexander Graf wrote: >> >> On 09.07.2013, at 20:46, Scott Wood wrote: >> >> > I suspect that tlbsx is faster, or at worst similar. And unlike >> >> > comparing tlbsx to lwepx (not counting a fix for the threading >> >> > problem), we don't already have code to search the guest TLB, so >> >> > testing would be more work. >> >> We have code to walk the guest TLB for TLB misses. This really is just >> >> the TLB miss search without host TLB injection. >> >> So let's say we're using the shadow TLB. The guest always has its say 64 >> >> TLB entries that it can count on - we never evict anything by accident, >> >> because we store all of the 64 entries in our guest TLB cache. When the >> >> guest faults at an address, the first thing we do is we check the cache >> >> whether we have that page already mapped. >> >> However, with this method we now have 2 enumeration methods for guest TLB >> >> searches. We have the tlbsx one which searches the host TLB and we have >> >> our guest TLB cache. The guest TLB cache might still contain an entry for >> >> an address that we already invalidated on the host. Would that impose a >> >> problem? >> >> I guess not because we're swizzling the exit code around to instead be an >> >> instruction miss which means we restore the TLB entry into our host's TLB >> >> so that when we resume, we land here and the tlbsx hits. But it feels >> >> backwards. >> > >> > Any better way? Searching the guest TLB won't work for the LRAT case, so >> > we'd need to have this logic around anyway. We shouldn't add a second >> > codepath unless it's a clear performance gain -- and again, I suspect it >> > would be the opposite, especially if the entry is not in TLB0 or in one of >> > the first few entries searched in TLB1. The tlbsx miss case is not what >> > we should optimize for. >> Hrm. >> So let's redesign this thing theoretically. We would have an exit that >> requires an instruction fetch. We would override kvmppc_get_last_inst() to >> always do kvmppc_ld_inst(). That one can fail because it can't find the TLB >> entry in the host TLB. When it fails, we have to abort the emulation and >> resume the guest at the same IP. >> Now the guest gets the TLB miss, we populate, go back into the guest. The >> guest hits the emulation failure again. We go back to kvmppc_ld_inst() which >> succeeds this time and we can emulate the instruction. > > That's pretty much what this patch does, except that it goes immediately to > the TLB miss code rather than having the extra round-trip back to the guest. > Is there any benefit from adding that extra round-trip? Rewriting the exit > type instead doesn't seem that bad...
It's pretty bad. I want to have code that is easy to follow - and I don't care whether the very rare case of a TLB entry getting evicted by a random other thread right when we execute the exit path is slower by a few percent if we get cleaner code for that. > >> I think this works. Just make sure that the gateway to the instruction fetch >> is kvmppc_get_last_inst() and make that failable. Then the difference >> between looking for the TLB entry in the host's TLB or in the guest's TLB >> cache is hopefully negligible. > > I don't follow here. What does this have to do with looking in the guest TLB? I want to hide the fact that we're cheating as much as possible, that's it. Alex -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html