On Fri, Sep 22, 2017 at 08:23:02AM +0200, Jan Kiszka wrote: > On 2017-09-22 03:19, Marcelo Tosatti wrote: > > On Thu, Sep 21, 2017 at 07:45:32PM +0200, Jan Kiszka wrote: > >> On 2017-09-21 13:38, Marcelo Tosatti wrote: > >>> When executing guest vcpu-0 with FIFO:1 priority, which is necessary to > >>> deal with the following situation: > >>> > >>> VCPU-0 (housekeeping VCPU) VCPU-1 (realtime VCPU) > >>> > >>> raw_spin_lock(A) > >>> interrupted, schedule task T-1 raw_spin_lock(A) (spin) > >>> > >>> raw_spin_unlock(A) > >>> > >>> Certain operations must interrupt guest vcpu-0 (see trace below). > >>> > >>> To fix this issue, only change guest vcpu-0 to FIFO priority > >>> on spinlock critical sections (see patch). > >>> > >>> Hang trace > >>> ========== > >>> > >>> Without FIFO priority: > >>> > >>> qemu-kvm-6705 [002] ....1.. 767785.648964: kvm_exit: reason IO_INSTRUCTION rip 0xe8fe info 1f00039 0 > >>> qemu-kvm-6705 [002] ....1.. 767785.648965: kvm_exit: reason IO_INSTRUCTION rip 0xe911 info 3f60008 0 > >>> qemu-kvm-6705 [002] ....1.. 767785.648968: kvm_exit: reason IO_INSTRUCTION rip 0x8984 info 608000b 0 > >>> qemu-kvm-6705 [002] ....1.. 767785.648971: kvm_exit: reason IO_INSTRUCTION rip 0xb313 info 1f70008 0 > >>> qemu-kvm-6705 [002] ....1.. 767785.648974: kvm_exit: reason IO_INSTRUCTION rip 0xb514 info 3f60000 0 > >>> qemu-kvm-6705 [002] ....1.. 767785.648977: kvm_exit: reason PENDING_INTERRUPT rip 0x8052 info 0 0 > >>> qemu-kvm-6705 [002] ....1.. 767785.648980: kvm_exit: reason IO_INSTRUCTION rip 0xeee6 info 200040 0 > >>> qemu-kvm-6705 [002] ....1.. 767785.648999: kvm_exit: reason EPT_MISCONFIG rip 0x2120 info 0 0 > >>> > >>> With FIFO priority: > >>> > >>> qemu-kvm-7636 [002] ....1.. 768218.205065: kvm_exit: reason IO_INSTRUCTION rip 0xb313 info 1f70008 0 > >>> qemu-kvm-7636 [002] ....1.. 768218.205068: kvm_exit: reason IO_INSTRUCTION rip 0x8984 info 608000b 0 > >>> qemu-kvm-7636 [002] ....1.. 768218.205071: kvm_exit: reason IO_INSTRUCTION rip 0xb313 info 1f70008 0 > >>> qemu-kvm-7636 [002] ....1.. 768218.205074: kvm_exit: reason IO_INSTRUCTION rip 0x8984 info 608000b 0 > >>> qemu-kvm-7636 [002] ....1.. 768218.205077: kvm_exit: reason IO_INSTRUCTION rip 0xb313 info 1f70008 0 > >>> .. > >>> > >>> Performance numbers (kernel compilation with make -j2) > >>> ====================================================== > >>> > >>> With hypercall: 4:40. (make -j2) > >>> Without hypercall: 3:38. (make -j2) > >>> > >>> Note for NFV workloads spinlock performance is not relevant > >>> since DPDK should not enter the kernel (and housekeeping vcpu > >>> performance is far from a key factor). > >>> > >>> Signed-off-by: Marcelo Tosatti <mtosatti@xxxxxxxxxx> > >>> > >> > >> That sounds familiar, though not yet the same: :) > >> > >> http://git.kiszka.org/?p=linux-kvm.git;a=shortlog;h=refs/heads/queues/paravirt-sched > >> (paper: http://lwn.net/images/conf/rtlws11/papers/proc/p18.pdf) > >> > >> I suppose your goal is not to enable the host to follow the guest > >> scheduler priority completely but only to have prio-ceiling for such > >> short critical sections. Maybe still useful to think ahead about future > >> extensions when actually introducing such an interface. > > > > Hi Jan! > > > > Hum... I'll take a look at your interface/paper and get back to you. > > > >> But shouldn't there be some limits on the maximum prio the guest can select? > > > > The SCHED_FIFO prio is fixed, selectable when QEMU starts. Do you > > envision any other use case than a fixed priority value selectable > > at QEMU initialization? > > Oh, indeed, this is a pure prio-ceiling variant with host-defined > ceiling value. > > But it's very inefficient to use a hypercall for entering and leaving > each and every sections. I would strongly recommend using a lazy scheme > where the guest writes the desired state into a shared memory page, and > the host only evaluates that prior to taking a scheduling decision, or > at least only on real vmexits. We're using such scheme successfully to > accelerate the fast path of prio-ceiling for pthread mutexes in the > Xenomai real-time extension. Yes, a faster scheme was envisioned, but not developed.
