On 15/01/20 15:59, Alexander Graf wrote: > On 15.01.20 15:43, milanpa@xxxxxxxxxx wrote: >>>> Let's expose new interface to userspace for garhering these >>>> statistics with one ioctl. >>>> >>>> Userspace application can read counter description once using >>>> KVM_GET_SUPPORTED_DEBUGFS_STAT and periodically invoke the >>>> KVM_GET_DEBUGFS_VALUES to get value update. >>> >>> This is an interface that requires a lot of logic and buffers from >>> user space to retrieve individual, explicit counters. What if I just >>> wanted to monitor the number of exits on every user space exit? >> >> In case we want to cover such latency sensitive use cases solution b), >> with mmap'ed structs you suggested, would be a way to go, IMO. >> >>> Also, we're suddenly making the debugfs names a full ABI, because >>> through this interface we only identify the individual stats through >>> their names. That means we can not remove stats or change their >>> names, because people may rely on them, no? Thining about this again, >>> maybe they already are an ABI because people rely on them in debugfs >>> though? In theory not, in practice I have treated them as a kind of "soft" ABI: if the meaning changes you should rename them, and removing them is fine, but you shouldn't for example change the unit of measure (which is not hard since they are all counters :) but perhaps you could have nanoseconds vs TSC cycles in the future for some counters). >>> I see two alternatives to this approach here: >>> >>> a) ONE_REG >>> >>> We can just add a new DEBUG arch in ONE_REG and expose ONE_REG per VM >>> as well (if we really have to?). That gives us explicit identifiers >>> for each stat with an explicit path to introduce new ones with very >>> unique identifiers. ONE_REG would force us to define constants for each counter, and would make it hard to retire them. I don't like this. >>> b) part of the mmap'ed vcpu struct Same here. Even if we say the semantics of the struct would be exposed to userspace via KVM_GET_SUPPORTED_DEBUGFS_STAT, someone might end up getting this wrong and expecting a particular layout. Milan's proposed API has the big advantage of being hard to get wrong for userspace. And pushing the aggregation to userspace is not a huge chore, but it's still a chore. So unless someone has a usecase for latency-sensitive monitoring I'd prefer to keep it simple (usually these kind of stats can even make sense if you gather them over relatively large period of time, because then you'll probably use something else like tracepoints to actually pinpoint what's going on). >>> 2) vcpu counters >>> >>> Most of the counters count on vcpu granularity, but debugfs only >>> gives us a full VM view. Whatever we do to improve the situation, we >>> should definitely try to build something that allows us to get the >>> counters per vcpu (as well). >>> >>> The main purpose of these counters is monitoring. It can be quite >>> important to know that only a single vCPU is going wild, compared to >>> all of them for example. >> >> I agree, exposing per vcpu counters can be useful. I guess it didn't >> make much sense exposing them through debugfs so aggregation was done >> in kernel. However if we chose to go with approach 1-b) mmap counters >> struct in userspace, we could do this. > > The reason I dislike the debugfs/statfs approach is that it generates a > completely separate permission and access paths to the stats. That's > great for full system monitoring, but really bad when you have multiple > individual tenants on a single host. I agree, anything in sysfs is complementary to vmfd/vcpufd access. Paolo
