Hi Suzuki, Thank you for having a look!
On 9/22/21 11:11, Suzuki K Poulose wrote: > On 25/08/2021 17:17, Alexandru Elisei wrote: >> This is v4 of the SPE series posted at [1]. v2 can be found at [2], and the >> original series at [3]. >> >> Statistical Profiling Extension (SPE) is an optional feature added in >> ARMv8.2. It allows sampling at regular intervals of the operations executed >> by the PE and storing a record of each operation in a memory buffer. A high >> level overview of the extension is presented in an article on arm.com [4]. >> >> This is another complete rewrite of the series, and nothing is set in >> stone. If you think of a better way to do things, please suggest it. >> >> >> Features added >> ============== >> >> The rewrite enabled me to add support for several features not >> present in the previous iteration: >> >> - Support for heterogeneous systems, where only some of the CPUs support SPE. >> This is accomplished via the KVM_ARM_VCPU_SUPPORTED_CPUS VCPU ioctl. >> >> - Support for VM migration with the >> KVM_ARM_VCPU_SPE_CTRL(KVM_ARM_VCPU_SPE_STOP) >> VCPU ioctl. >> >> - The requirement for userspace to mlock() the guest memory has been removed, >> and now userspace can make changes to memory contents after the memory is >> mapped at stage 2. >> >> - Better debugging of guest memory pinning by printing a warning when we >> get an unexpected read or write fault. This helped me catch several bugs >> during development, it has already proven very useful. Many thanks to >> James who suggested when reviewing v3. >> >> >> Missing features >> ================ >> >> I've tried to keep the series as small as possible to make it easier to >> review, >> while implementing the core functionality needed for the SPE emulation. As >> such, >> I've chosen to not implement several features: >> >> - Host profiling a guest which has the SPE feature bit set (see open >> questions). >> >> - No errata workarounds have been implemented yet, and there are quite a few >> of >> them for Neoverse N1 and Neoverse V1. >> >> - Disabling CONFIG_NUMA_BALANCING is a hack to get KVM SPE to work and I am >> investigating other ways to get around automatic numa balancing, like >> requiring userspace to disable it via set_mempolicy(). I am also going to >> look at how VFIO gets around it. Suggestions welcome. >> >> - There's plenty of room for optimization. Off the top of my head, using >> block mappings at stage 2, batch pinning of pages (similar to what VFIO >> does), optimize the way KVM keeps track of pinned pages (using a linked >> list triples the memory usage), context-switch the SPE registers on >> vcpu_load/vcpu_put on VHE if the host is not profiling, locking >> optimizations, etc, etc. >> >> - ...and others. I'm sure I'm missing at least a few things which are >> important for someone. >> >> >> Known issues >> ============ >> >> This is an RFC, so keep in mind that almost definitely there will be scary >> bugs. For example, below is a list of known issues which don't affect the >> correctness of the emulation, and which I'm planning to fix in a future >> iteration: >> >> - With CONFIG_PROVE_LOCKING=y, lockdep complains about lock contention when >> the VCPU executes the dcache clean pending ops. >> >> - With CONFIG_PROVE_LOCKING=y, KVM will hit a BUG at >> kvm_lock_all_vcpus()->mutex_trylock(&vcpu->mutex) with more than 48 >> VCPUs. >> >> This BUG statement can also be triggered with mainline. To reproduce it, >> compile kvmtool from this branch [5] and follow the instruction in the >> kvmtool commit message. >> >> One workaround could be to stop trying to lock all VCPUs when locking a >> memslot and document the fact that it is required that no VCPUs are run >> before the ioctl completes, otherwise bad things might happen to the VM. >> >> >> Open questions >> ============== >> >> 1. Implementing support for host profiling a guest with the SPE feature >> means setting the profiling buffer owning regime to EL2. While that is in >> effect, PMBIDR_EL1.P will equal 1. This has two consequences: if the guest >> probes SPE during this time, the driver will fail; and the guest will be >> able to determine when it is profiled. I see two options here: > > This doesn't mean the EL2 is owning the SPE. It only tells you that a > higher level EL is owning the SPE. It could as well be EL3. (e.g, > MDCR_EL3.NSPB > == 0 or 1). So I think this is architecturally correct, > as long as we trap the guest access to other SPE registers and inject > and UNDEF. You are right, I was wrong about the part about the guest being able to determine when it is profiled, I forgot that PMBIDR_EL1.P can also be 1 if EL3 is owning the buffer. But I don't understand why you are suggesting that KVM injects UNDEF in this case. I was thinking that when the host is profiling the guest, KVM can store guest writes to the SPE registers to the shadow copy of the registers. On a world switch to guest, KVM will not restore the registers onto the hardware as to not interfere with the profiling operation performed by the host. When profiling ends, KVM can then let the guest use SPE again by restoring the guest register values onto the hardware at every world switch and setting the buffer owning regime to EL1 again. I put this as an open question because if the guest sees PMBIDR_EL1.P set when the guest SPE driver probes, the driver will likely decide that SPE is not available. But the VM has been created with SPE enabled because the host userspace wants the guest to have access to SPE, and the host userspace might not realize that the act of profiling a guest can make SPE unusable by that guest. I am inclined now to let the host's userspace do whatever it wishes and allow it to profile a guest if it so desires, and mention this possible side effect in the KVM documentation, as it might be surprising for someone who isn't familiar with the inner workings of KVM's SPE emulation. Thanks, Alex > > > Thanks > Suzuki _______________________________________________ kvmarm mailing list kvmarm@lists.cs.columbia.edu https://lists.cs.columbia.edu/mailman/listinfo/kvmarm