> > Hi, > > I was just in a discussion where we realized KVM_HINTS_REALTIME is a > > little underdocumented. Here is attempt to address that. Please correct > > me if there are inaccuracies or reply if you have additional questions: > > > > KVM_HINTS_REALTIME (akaalso QEMU kvm-hint-dedicated) is defined as follows > > in Documentation/virt/kvm/cpuid.rst: > > > > guest checks this feature bit to determine that vCPUs are never > > preempted for an unlimited time allowing optimizations > > > > Users or management tools set this flag themselves (it is not set > > automatically). This raises the question of what effects this flag has > > and when it should be set. > > > > When should I set KVM_HINTS_REALTIME? > > ------------------------------------- > > When vCPUs are pinned to dedicated pCPUs. Even better if the isolcpus= > > kernel parameter is used on the host so there are no disturbances. > > > > Is the flag guest-wide or per-vCPU? > > ----------------------------------- > > This flag is guest-wide so all vCPUs should be dedicated, not just some > > of them. > > > > Which Linux guest features are affected? > > ---------------------------------------- > > PV spinlocks, PV TLB flush, and PV sched yield are disabled by > > KVM_HINTS_REALTIME. This is because no other vCPUs or host tasks will be > > running on the pCPUs, so there is no benefit in involving the host. > > Do we need to mention "halt_poll_ns" at host side also will also be disabled? with KVM_FEATURE_POLL_CONTROL Sorry, pressed enter quickly in previous email. > > > > > The cpuidle-haltpoll driver is enabled by KVM_HINTS_REALTIME. This > > driver performs busy waiting inside the guest before halting the CPU in > > order to avoid the vCPU's wakeup latency. This driver also has a boolean > > "force" module parameter if you wish to enable it without setting > > KVM_HINTS_REALTIME. > > > > When KVM_HINTS_REALTIME is set, the KVM_CAP_X86_DISABLE_EXITS capability > > can also be used to disable MWAIT/HLT/PAUSE/CSTATE exits. This improves > > the latency of these operations. The user or management tools need to > > disable these exits themselves, e.g. with QEMU's -overcommit cpu-pm=on. > > This looks good. Thank you. > > Thanks, > Pankaj
