This patch introduces a new module parameter for the KVM module; when it
is present, KVM attempts a bit of polling on every HLT before scheduling
itself out via kvm_vcpu_block.

This parameter helps a lot for latency-bound workloads---in particular
I tested it with O_DSYNC writes with a battery-backed disk in the host.
In this case, writes are fast (because the data doesn't have to go all
the way to the platters) but they cannot be merged by either the host or
the guest.  KVM's performance here is usually around 30% of bare metal,
or 50% if you use cache=directsync or cache=writethrough (these
parameters avoid that the guest sends pointless flush requests, and
at the same time they are not slow because of the battery-backed cache).
The bad performance happens because on every halt the host CPU decides
to halt itself too.  When the interrupt comes, the vCPU thread is then
migrated to a new physical CPU, and in general the latency is horrible
because the vCPU thread has to be scheduled back in.

With this patch performance reaches 60-65% of bare metal and, more
important, 99% of what you get if you use idle=poll in the guest.  This
means that the tunable gets rid of this particular bottleneck, and more
work can be done to improve performance in the kernel or QEMU.

Of course there is some price to pay; every time an otherwise idle vCPUs
is interrupted by an interrupt, it will poll unnecessarily and thus
impose a little load on the host.  The above results were obtained with
a mostly random value of the parameter (500000), and the load was around
1.5-2.5% CPU usage on one of the host's core for each idle guest vCPU.

The patch also adds a new stat, /sys/kernel/debug/kvm/halt_successful_poll,
that can be used to tune the parameter.  It counts how many HLT
instructions received an interrupt during the polling period; each
successful poll avoids that Linux schedules the VCPU thread out and back
in, and may also avoid a likely trip to C1 and back for the physical CPU.

While the VM is idle, a Linux 4 VCPU VM halts around 10 times per second.
Of these halts, almost all are failed polls.  During the benchmark,
instead, basically all halts end within the polling period, except a more
or less constant stream of 50 per second coming from vCPUs that are not
running the benchmark.  The wasted time is thus very low.  Things may
be slightly different for Windows VMs, which have a ~10 ms timer tick.

The effect is also visible on Marcelo's recently-introduced latency
test for the TSC deadline timer.  Though of course a non-RT kernel has
awful latency bounds, the latency of the timer is around 8000-10000 clock
cycles compared to 20000-120000 without setting halt_poll_ns.  For the TSC
deadline timer, thus, the effect is both a smaller average latency and
a smaller variance.
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

In the beggining was on_each_cpu(), which required an unused argument to
kvm_arch_ops.hardware_{en,dis}able, but this was soon forgotten.

Remove unnecessary arguments that stem from this.
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Using static inline is going to save few bytes and cycles.
For example on powerpc, the difference is 700 B after stripping.
(5 kB before)

This patch also deals with two overlooked empty functions:
kvm_arch_flush_shadow was not removed from arch/mips/kvm/mips.c
  2df72e9b KVM: split kvm_arch_flush_shadow
and kvm_arch_sched_in never made it into arch/ia64/kvm/kvm-ia64.c.
  e790d9ef KVM: add kvm_arch_sched_in
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Opaque KVM structs are useful for prototypes in asm/kvm_host.h, to avoid
"'struct foo' declared inside parameter list" warnings (and consequent
breakage due to conflicting types).

Move them from individual files to a generic place in linux/kvm_types.h.
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

No logic changes inside.
Reviewed-by: NJames Hogan <james.hogan@imgtec.com>
Signed-off-by: NDeng-Cheng Zhu <dengcheng.zhu@imgtec.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

No logic changes inside.
Signed-off-by: NDeng-Cheng Zhu <dengcheng.zhu@imgtec.com>
Reviewed-by: NJames Hogan <james.hogan@imgtec.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Fix whitespace in struct kvm_mips_callbacks function pointers.
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: linux-mips@linux-mips.org
Cc: Sanjay Lal <sanjayl@kymasys.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Expose the KVM guest CP0_Count frequency to userland via a new
KVM_REG_MIPS_COUNT_HZ register accessible with the KVM_{GET,SET}_ONE_REG
ioctls.

When the frequency is altered the bias is adjusted such that the guest
CP0_Count doesn't jump discontinuously or lose any timer interrupts.
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: linux-mips@linux-mips.org
Cc: David Daney <david.daney@cavium.com>
Cc: Sanjay Lal <sanjayl@kymasys.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Expose two new virtual registers to userland via the
KVM_{GET,SET}_ONE_REG ioctls.

KVM_REG_MIPS_COUNT_CTL is for timer configuration fields and just
contains a master disable count bit. This can be used by userland to
freeze the timer in order to read a consistent state from the timer
count value and timer interrupt pending bit. This cannot be done with
the CP0_Cause.DC bit because the timer interrupt pending bit (TI) is
also in CP0_Cause so it would be impossible to stop the timer without
also risking a race with an hrtimer interrupt and having to explicitly
check whether an interrupt should have occurred.

When the timer is re-enabled it resumes without losing time, i.e. the
CP0_Count value jumps to what it would have been had the timer not been
disabled, which would also be impossible to do from userland with
CP0_Cause.DC. The timer interrupt also cannot be lost, i.e. if a timer
interrupt would have occurred had the timer not been disabled it is
queued when the timer is re-enabled.

This works by storing the nanosecond monotonic time when the master
disable is set, and using it for various operations instead of the
current monotonic time (e.g. when recalculating the bias when the
CP0_Count is set), until the master disable is cleared again, i.e. the
timer state is read/written as it would have been at that time. This
state is exposed to userland via the read-only KVM_REG_MIPS_COUNT_RESUME
virtual register so that userland can determine the exact time the
master disable took effect.

This should allow userland to atomically save the state of the timer,
and later restore it.
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: linux-mips@linux-mips.org
Cc: David Daney <david.daney@cavium.com>
Cc: Sanjay Lal <sanjayl@kymasys.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Previously the emulation of the CPU timer was just enough to get a Linux
guest running but some shortcuts were taken:
 - The guest timer interrupt was hard coded to always happen every 10 ms
   rather than being timed to when CP0_Count would match CP0_Compare.
 - The guest's CP0_Count register was based on the host's CP0_Count
   register. This isn't very portable and fails on cores without a
   CP_Count register implemented such as Ingenic XBurst. It also meant
   that the guest's CP0_Cause.DC bit to disable the CP0_Count register
   took no effect.
 - The guest's CP0_Count register was emulated by just dividing the
   host's CP0_Count register by 4. This resulted in continuity problems
   when used as a clock source, since when the host CP0_Count overflows
   from 0x7fffffff to 0x80000000, the guest CP0_Count transitions
   discontinuously from 0x1fffffff to 0xe0000000.

Therefore rewrite & fix emulation of the guest timer based on the
monotonic kernel time (i.e. ktime_get()). Internally a 32-bit count_bias
value is added to the frequency scaled nanosecond monotonic time to get
the guest's CP0_Count. The frequency of the timer is initialised to
100MHz and cannot yet be changed, but a later patch will allow the
frequency to be configured via the KVM_{GET,SET}_ONE_REG ioctl
interface.

The timer can now be stopped via the CP0_Cause.DC bit (by the guest or
via the KVM_SET_ONE_REG ioctl interface), at which point the current
CP0_Count is stored and can be read directly. When it is restarted the
bias is recalculated such that the CP0_Count value is continuous.

Due to the nature of hrtimer interrupts any read of the guest's
CP0_Count register while it is running triggers a check for whether the
hrtimer has expired, so that the guest/userland cannot observe the
CP0_Count passing CP0_Compare without queuing a timer interrupt. This is
also taken advantage of when stopping the timer to ensure that a pending
timer interrupt is queued.

This replaces the implementation of:
 - Guest read of CP0_Count
 - Guest write of CP0_Count
 - Guest write of CP0_Compare
 - Guest write of CP0_Cause
 - Guest read of HWR 2 (CC) with RDHWR
 - Host read of CP0_Count via KVM_GET_ONE_REG ioctl interface
 - Host write of CP0_Count via KVM_SET_ONE_REG ioctl interface
 - Host write of CP0_Compare via KVM_SET_ONE_REG ioctl interface
 - Host write of CP0_Cause via KVM_SET_ONE_REG ioctl interface
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: linux-mips@linux-mips.org
Cc: Sanjay Lal <sanjayl@kymasys.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

The hrtimer callback for guest timer timeouts sets the guest's
CP0_Cause.TI bit to indicate to the guest that a timer interrupt is
pending, however there is no mutual exclusion implemented to prevent
this occurring while the guest's CP0_Cause register is being
read-modify-written elsewhere.

When this occurs the setting of the CP0_Cause.TI bit is undone and the
guest misses the timer interrupt and doesn't reprogram the CP0_Compare
register for the next timeout. Currently another timer interrupt will be
triggered again in another 10ms anyway due to the way timers are
emulated, but after the MIPS timer emulation is fixed this would result
in Linux guest time standing still and the guest scheduler not being
invoked until the guest CP0_Count has looped around again, which at
100MHz takes just under 43 seconds.

Currently this is the only asynchronous modification of guest registers,
therefore it is fixed by adjusting the implementations of the
kvm_set_c0_guest_cause(), kvm_clear_c0_guest_cause(), and
kvm_change_c0_guest_cause() macros which are used for modifying the
guest CP0_Cause register to use ll/sc to ensure atomic modification.
This should work in both UP and SMP cases without requiring interrupts
to be disabled.
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: linux-mips@linux-mips.org
Cc: Sanjay Lal <sanjayl@kymasys.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Implement KVM_{GET,SET}_ONE_REG ioctl based access to the guest CP0
UserLocal register. This is so that userland can save and restore its
value.
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: linux-mips@linux-mips.org
Cc: David Daney <david.daney@cavium.com>
Cc: Sanjay Lal <sanjayl@kymasys.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Implement KVM_{GET,SET}_ONE_REG ioctl based access to the guest CP0
Count and Compare registers. These registers are special in that writing
to them has side effects (adjusting the time until the next timer
interrupt) and reading of Count depends on the time. Therefore add a
couple of callbacks so that different implementations (trap & emulate or
VZ) can implement them differently depending on what the hardware
provides.

The trap & emulate versions mostly duplicate what happens when a T&E
guest reads or writes these registers, so it inherits the same
limitations which can be fixed in later patches.
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: linux-mips@linux-mips.org
Cc: David Daney <david.daney@cavium.com>
Cc: Sanjay Lal <sanjayl@kymasys.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Move the KVM_{GET,SET}_ONE_REG MIPS register id definitions out of
kvm_mips.c to kvm_host.h so that they can be shared between multiple
source files. This allows register access to be indirected depending on
the underlying implementation (trap & emulate or VZ).
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: linux-mips@linux-mips.org
Cc: David Daney <david.daney@cavium.com>
Cc: Sanjay Lal <sanjayl@kymasys.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

MIPS KVM uses mips32_SyncICache to synchronise the icache with the
dcache after dynamically modifying guest instructions or writing guest
exception vector. However this uses rdhwr to get the SYNCI step, which
causes a reserved instruction exception on Ingenic XBurst cores.

It would seem to make more sense to use local_flush_icache_range()
instead which does the same thing but is more portable.
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: linux-mips@linux-mips.org
Cc: Sanjay Lal <sanjayl@kymasys.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

The ability to read hardware registers from userland with the RDHWR
instruction should depend upon the corresponding bit of the HWREna
register being set, otherwise a reserved instruction exception should be
generated.

However KVM's current emulation ignores the guest's HWREna and always
emulates RDHWR instructions even if the guest OS has disallowed them.

Therefore rework the RDHWR emulation code to check for privilege or the
corresponding bit in the guest HWREna bit. Also remove the #if 0 case
for the UserLocal register. I presume it was there for debug purposes
but it seems unnecessary now that the guest can control whether it
causes a guest exception.
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Sanjay Lal <sanjayl@kymasys.com>
Cc: linux-mips@linux-mips.org
Cc: kvm@vger.kernel.org
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

The whitespace in asm/kvm_host.h is quite inconsistent in places. Clean
up the whole file to use tabs more consistently.

When you use the --ignore-space-change argument to git diff this patch
only changes line wrapping in TLB_IS_GLOBAL and TLB_IS_VALID macros.
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Sanjay Lal <sanjayl@kymasys.com>
Cc: linux-mips@linux-mips.org
Cc: kvm@vger.kernel.org
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

The kvm_mips_init_shadow_tlb() function is called from
kvm_arch_vcpu_init() and initialises entries 0 to
current_cpu_data.tlbsize-1 of the virtual cpu's shadow_tlb[64] array.

However newer cores with FTLBs can have a tlbsize > 64, for example the
ProAptiv I'm testing on has a total tlbsize of 576. This causes
kvm_mips_init_shadow_tlb() to overflow the shadow_tlb[64] array and
overwrite the comparecount_timer among other things, causing a lock up
when starting a KVM guest.

Aside from kvm_mips_init_shadow_tlb() which only initialises it, the
shadow_tlb[64] array is only actually used by the following functions:
 - kvm_shadow_tlb_put() & kvm_shadow_tlb_load()
     These are never called. The only call sites are #if 0'd out.
 - kvm_mips_dump_shadow_tlbs()
     This is never called.

It was originally added for trap & emulate, but turned out to be
unnecessary so it was disabled.

So instead of fixing the shadow_tlb initialisation code, lets just
remove the shadow_tlb[64] array and the above functions entirely. The
only functional change here is the removal of broken shadow_tlb
initialisation. The rest just deletes dead code.
Signed-off-by: NJames Hogan <james.hogan@imgtec.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: linux-mips@linux-mips.org
Cc: Gleb Natapov <gleb@redhat.com>
Cc: kvm@vger.kernel.org
Cc: Sanjay Lal <sanjayl@kymasys.com>
Acked-by: NPaolo Bonzini <pbonzini@redhat.com>
Signed-off-by: NJohn Crispin <blogic@openwrt.org>
Patchwork: http://patchwork.linux-mips.org/patch/6384/

Now when the main kvm code relying on these defines has been moved to
the x86 specific part of the world, we can get rid of these.
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>
Signed-off-by: NGleb Natapov <gleb@redhat.com>

Because not all 256 CP0 registers are ever implemented, we need a
different method of manipulating them.  Use the
KVM_SET_ONE_REG/KVM_GET_ONE_REG mechanism.

Now unused code and definitions are removed.
Signed-off-by: NDavid Daney <david.daney@cavium.com>
Acked-by: NSanjay Lal <sanjayl@kymasys.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

This reverts commit d532f3d2.

The original commit has several problems:

1) Doesn't work with 64-bit kernels.

2) Calls TLBMISS_HANDLER_SETUP() before the code is generated.

3) Calls TLBMISS_HANDLER_SETUP() twice in per_cpu_trap_init() when
   only one call is needed.

[ralf@linux-mips.org: Also revert the bits of the ASID patch which were
hidden in the KVM merge.]
Signed-off-by: NDavid Daney <david.daney@cavium.com>
Cc: linux-mips@linux-mips.org
Cc: linux-kernel@vger.kernel.org
Cc: "Steven J. Hill" <Steven.Hill@imgtec.com>
Cc: David Daney <david.daney@cavium.com>
Patchwork: https://patchwork.linux-mips.org/patch/5242/Signed-off-by: NRalf Baechle <ralf@linux-mips.org>

Signed-off-by: NSanjay Lal <sanjayl@kymasys.com>
Cc: kvm@vger.kernel.org
Cc: linux-mips@linux-mips.org
Signed-off-by: NRalf Baechle <ralf@linux-mips.org>