POWER8 has one virtual timebase (VTB) register per subcore, not one
per CPU thread.  The HV KVM code currently treats VTB as a per-thread
register, which can lead to spurious soft lockup messages from guests
which use the VTB as the time source for the soft lockup detector.
(CPUs before POWER8 did not have the VTB register.)

For HV KVM, this fixes the problem by making only the primary thread
in each virtual core save and restore the VTB value.  With this,
the VTB state becomes part of the kvmppc_vcore structure.  This
also means that "piggybacking" of multiple virtual cores onto one
subcore is not possible on POWER8, because then the virtual cores
would share a single VTB register.

PR KVM emulates a VTB register, which is per-vcpu because PR KVM
has no notion of CPU threads or SMT.  For PR KVM we move the VTB
state into the kvmppc_vcpu_book3s struct.

Cc: stable@vger.kernel.org # v3.14+
Reported-by: NThomas Huth <thuth@redhat.com>
Tested-by: NThomas Huth <thuth@redhat.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

Add VCPU stat counters to track affinity for passthrough
interrupts.

pthru_all: Counts all passthrough interrupts whose IRQ mappings are
           in the kvmppc_passthru_irq_map structure.
pthru_host: Counts all cached passthrough interrupts that were injected
	    from the host through kvm_set_irq (i.e. not handled in
	    real mode).
pthru_bad_aff: Counts how many cached passthrough interrupts have
               bad affinity (receiving CPU is not running VCPU that is
	       the target of the virtual interrupt in the guest).
Signed-off-by: NSuresh Warrier <warrier@linux.vnet.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

This patch introduces an IRQ mapping structure, the
kvmppc_passthru_irqmap structure that is to be used
to map the real hardware IRQ in the host with the virtual
hardware IRQ (gsi) that is injected into a guest by KVM for
passthrough adapters.

Currently, we assume a separate IRQ mapping structure for
each guest. Each kvmppc_passthru_irqmap has a mapping arrays,
containing all defined real<->virtual IRQs.

[paulus@ozlabs.org - removed irq_chip field from struct
 kvmppc_passthru_irqmap; changed parameter for
 kvmppc_get_passthru_irqmap from struct kvm_vcpu * to struct
 kvm *, removed small cached array.]
Signed-off-by: NSuresh Warrier <warrier@linux.vnet.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

vcpu stats are used to collect information about a vcpu which can be viewed
in the debugfs. For example halt_attempted_poll and halt_successful_poll
are used to keep track of the number of times the vcpu attempts to and
successfully polls. These stats are currently not used on powerpc.

Implement incrementation of the halt_attempted_poll and
halt_successful_poll vcpu stats for powerpc. Since these stats are summed
over all the vcpus for all running guests it doesn't matter which vcpu
they are attributed to, thus we choose the current runner vcpu of the
vcore.

Also add new vcpu stats: halt_poll_success_ns, halt_poll_fail_ns and
halt_wait_ns to be used to accumulate the total time spend polling
successfully, polling unsuccessfully and waiting respectively, and
halt_successful_wait to accumulate the number of times the vcpu waits.
Given that halt_poll_success_ns, halt_poll_fail_ns and halt_wait_ns are
expressed in nanoseconds it is necessary to represent these as 64-bit
quantities, otherwise they would overflow after only about 4 seconds.

Given that the total time spend either polling or waiting will be known and
the number of times that each was done, it will be possible to determine
the average poll and wait times. This will give the ability to tune the kvm
module parameters based on the calculated average wait and poll times.
Signed-off-by: NSuraj Jitindar Singh <sjitindarsingh@gmail.com>
Reviewed-by: NDavid Matlack <dmatlack@google.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

vms and vcpus have statistics associated with them which can be viewed
within the debugfs. Currently it is assumed within the vcpu_stat_get() and
vm_stat_get() functions that all of these statistics are represented as
u32s, however the next patch adds some u64 vcpu statistics.

Change all vcpu statistics to u64 and modify vcpu_stat_get() accordingly.
Since vcpu statistics are per vcpu, they will only be updated by a single
vcpu at a time so this shouldn't present a problem on 32-bit machines
which can't atomically increment 64-bit numbers. However vm statistics
could potentially be updated by multiple vcpus from that vm at a time.
To avoid the overhead of atomics make all vm statistics ulong such that
they are 64-bit on 64-bit systems where they can be atomically incremented
and are 32-bit on 32-bit systems which may not be able to atomically
increment 64-bit numbers. Modify vm_stat_get() to expect ulongs.
Signed-off-by: NSuraj Jitindar Singh <sjitindarsingh@gmail.com>
Reviewed-by: NDavid Matlack <dmatlack@google.com>
Acked-by: NChristian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

This patch introduces new halt polling functionality into the kvm_hv kernel
module. When a vcore is idle it will poll for some period of time before
scheduling itself out.

When all of the runnable vcpus on a vcore have ceded (and thus the vcore is
idle) we schedule ourselves out to allow something else to run. In the
event that we need to wake up very quickly (for example an interrupt
arrives), we are required to wait until we get scheduled again.

Implement halt polling so that when a vcore is idle, and before scheduling
ourselves, we poll for vcpus in the runnable_threads list which have
pending exceptions or which leave the ceded state. If we poll successfully
then we can get back into the guest very quickly without ever scheduling
ourselves, otherwise we schedule ourselves out as before.

There exists generic halt_polling code in virt/kvm_main.c, however on
powerpc the polling conditions are different to the generic case. It would
be nice if we could just implement an arch specific kvm_check_block()
function, but there is still other arch specific things which need to be
done for kvm_hv (for example manipulating vcore states) which means that a
separate implementation is the best option.

Testing of this patch with a TCP round robin test between two guests with
virtio network interfaces has found a decrease in round trip time of ~15us
on average. A performance gain is only seen when going out of and
back into the guest often and quickly, otherwise there is no net benefit
from the polling. The polling interval is adjusted such that when we are
often scheduled out for long periods of time it is reduced, and when we
often poll successfully it is increased. The rate at which the polling
interval increases or decreases, and the maximum polling interval, can
be set through module parameters.

Based on the implementation in the generic kvm module by Wanpeng Li and
Paolo Bonzini, and on direction from Paul Mackerras.
Signed-off-by: NSuraj Jitindar Singh <sjitindarsingh@gmail.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

The struct kvmppc_vcore is a structure used to store various information
about a virtual core for a kvm guest. The runnable_threads element of the
struct provides a list of all of the currently runnable vcpus on the core
(those in the KVMPPC_VCPU_RUNNABLE state). The previous implementation of
this list was a linked_list. The next patch requires that the list be able
to be iterated over without holding the vcore lock.

Reimplement the runnable_threads list in the kvmppc_vcore struct as an
array. Implement function to iterate over valid entries in the array and
update access sites accordingly.
Signed-off-by: NSuraj Jitindar Singh <sjitindarsingh@gmail.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

The next commit will introduce a member to the kvmppc_vcore struct which
references MAX_SMT_THREADS which is defined in kvm_book3s_asm.h, however
this file isn't included in kvm_host.h directly. Thus compiling for
certain platforms such as pmac32_defconfig and ppc64e_defconfig with KVM
fails due to MAX_SMT_THREADS not being defined.

Move the struct kvmppc_vcore definition to kvm_book3s.h which explicitly
includes kvm_book3s_asm.h.
Signed-off-by: NSuraj Jitindar Singh <sjitindarsingh@gmail.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

It doesn't make sense to create irqfds for a VM that doesn't have
in-kernel interrupt controller emulation.  There is an existing
interface for architecture code to tell the irqfd code whether or
not any interrupt controller has been initialized, called
kvm_arch_intc_initialized(), so let's implement that for powerpc.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

Some wakeups should not be considered a sucessful poll. For example on
s390 I/O interrupts are usually floating, which means that _ALL_ CPUs
would be considered runnable - letting all vCPUs poll all the time for
transactional like workload, even if one vCPU would be enough.
This can result in huge CPU usage for large guests.
This patch lets architectures provide a way to qualify wakeups if they
should be considered a good/bad wakeups in regard to polls.

For s390 the implementation will fence of halt polling for anything but
known good, single vCPU events. The s390 implementation for floating
interrupts does a wakeup for one vCPU, but the interrupt will be delivered
by whatever CPU checks first for a pending interrupt. We prefer the
woken up CPU by marking the poll of this CPU as "good" poll.
This code will also mark several other wakeup reasons like IPI or
expired timers as "good". This will of course also mark some events as
not sucessful. As  KVM on z runs always as a 2nd level hypervisor,
we prefer to not poll, unless we are really sure, though.

This patch successfully limits the CPU usage for cases like uperf 1byte
transactional ping pong workload or wakeup heavy workload like OLTP
while still providing a proper speedup.

This also introduced a new vcpu stat "halt_poll_no_tuning" that marks
wakeups that are considered not good for polling.
Signed-off-by: NChristian Borntraeger <borntraeger@de.ibm.com>
Acked-by: Radim Krčmář <rkrcmar@redhat.com> (for an earlier version)
Cc: David Matlack <dmatlack@google.com>
Cc: Wanpeng Li <kernellwp@gmail.com>
[Rename config symbol. - Paolo]
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

The KVM_MAX_VCPUS define provides the maximum number of vCPUs per guest, and
also the upper limit for vCPU ids. This is okay for all archs except PowerPC
which can have higher ids, depending on the cpu/core/thread topology. In the
worst case (single threaded guest, host with 8 threads per core), it limits
the maximum number of vCPUS to KVM_MAX_VCPUS / 8.

This patch separates the vCPU numbering from the total number of vCPUs, with
the introduction of KVM_MAX_VCPU_ID, as the maximal valid value for vCPU ids
plus one.

The corresponding KVM_CAP_MAX_VCPU_ID allows userspace to validate vCPU ids
before passing them to KVM_CREATE_VCPU.

This patch only implements KVM_MAX_VCPU_ID with a specific value for PowerPC.
Other archs continue to return KVM_MAX_VCPUS instead.
Suggested-by: NRadim Krcmar <rkrcmar@redhat.com>
Signed-off-by: NGreg Kurz <gkurz@linux.vnet.ibm.com>
Reviewed-by: NCornelia Huck <cornelia.huck@de.ibm.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

This enables userspace view of TCE tables to start from non-zero offset
on a bus. This will be used for huge DMA windows.

This only changes the internal structure, the user interface needs to
change in order to use an offset.
Signed-off-by: NAlexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: NDavid Gibson <david@gibson.dropbear.id.au>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

At the moment the kvmppc_spapr_tce_table struct can only describe
4GB windows and handle fixed size (4K) pages. Dynamic DMA windows
support more so these limits need to be extended.

This replaces window_size (in bytes, 4GB max) with page_shift (32bit)
and size (64bit, in pages).

This should cause no behavioural change as this is changing
the internal structures only - the user interface still only
allows one to create a 32-bit table with 4KiB pages at this stage.
Signed-off-by: NAlexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: NDavid Gibson <david@gibson.dropbear.id.au>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

The problem:

On -rt, an emulated LAPIC timer instances has the following path:

1) hard interrupt
2) ksoftirqd is scheduled
3) ksoftirqd wakes up vcpu thread
4) vcpu thread is scheduled

This extra context switch introduces unnecessary latency in the
LAPIC path for a KVM guest.

The solution:

Allow waking up vcpu thread from hardirq context,
thus avoiding the need for ksoftirqd to be scheduled.

Normal waitqueues make use of spinlocks, which on -RT
are sleepable locks. Therefore, waking up a waitqueue
waiter involves locking a sleeping lock, which
is not allowed from hard interrupt context.

cyclictest command line:

This patch reduces the average latency in my tests from 14us to 11us.

Daniel writes:
Paolo asked for numbers from kvm-unit-tests/tscdeadline_latency
benchmark on mainline. The test was run 1000 times on
tip/sched/core 4.4.0-rc8-01134-g0905f04e:

  ./x86-run x86/tscdeadline_latency.flat -cpu host

with idle=poll.

The test seems not to deliver really stable numbers though most of
them are smaller. Paolo write:

"Anything above ~10000 cycles means that the host went to C1 or
lower---the number means more or less nothing in that case.

The mean shows an improvement indeed."

Before:

               min             max         mean           std
count  1000.000000     1000.000000  1000.000000   1000.000000
mean   5162.596000  2019270.084000  5824.491541  20681.645558
std      75.431231   622607.723969    89.575700   6492.272062
min    4466.000000    23928.000000  5537.926500    585.864966
25%    5163.000000  16132529.750000  5790.132275  16683.745433
50%    5175.000000  2281919.000000  5834.654000  23151.990026
75%    5190.000000  2382865.750000  5861.412950  24148.206168
max    5228.000000  4175158.000000  6254.827300  46481.048691

After
               min            max         mean           std
count  1000.000000     1000.00000  1000.000000   1000.000000
mean   5143.511000  2076886.10300  5813.312474  21207.357565
std      77.668322   610413.09583    86.541500   6331.915127
min    4427.000000    25103.00000  5529.756600    559.187707
25%    5148.000000  1691272.75000  5784.889825  17473.518244
50%    5160.000000  2308328.50000  5832.025000  23464.837068
75%    5172.000000  2393037.75000  5853.177675  24223.969976
max    5222.000000  3922458.00000  6186.720500  42520.379830

[Patch was originaly based on the swait implementation found in the -rt
 tree. Daniel ported it to mainline's version and gathered the
 benchmark numbers for tscdeadline_latency test.]
Signed-off-by: NDaniel Wagner <daniel.wagner@bmw-carit.de>
Acked-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: linux-rt-users@vger.kernel.org
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Paul Gortmaker <paul.gortmaker@windriver.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/1455871601-27484-4-git-send-email-wagi@monom.orgSigned-off-by: NThomas Gleixner <tglx@linutronix.de>

At the moment only spapr_tce_tables updates are protected against races
but not lookups. This fixes missing protection by using RCU for the list.
As lookups also happen in real mode, this uses
list_for_each_entry_lockless() (which is expected not to access any
vmalloc'd memory).

This converts release_spapr_tce_table() to a RCU scheduled handler.
Signed-off-by: NAlexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: NDavid Gibson <david@gibson.dropbear.id.au>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Since the numbers now overlap, it makes sense to enumerate
them in asm/kvm_host.h rather than linux/kvm_host.h.  Functions
that refer to architecture-specific requests are also moved
to arch/.
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Only using 32 memslots for KVM on powerpc is way too low, you can
nowadays hit this limit quite fast by adding a couple of PCI devices
and/or pluggable memory DIMMs to the guest.

x86 already increased the KVM_USER_MEM_SLOTS to 509, to satisfy 256
pluggable DIMM slots, 3 private slots and 253 slots for other things
like PCI devices (i.e. resulting in 256 + 3 + 253 = 512 slots in
total). We should do something similar for powerpc, and since we do
not use private slots here, we can set the value to 512 directly.

While we're at it, also remove the KVM_MEM_SLOTS_NUM definition
from the powerpc-specific header since this gets defined in the
generic kvm_host.h header anyway.
Signed-off-by: NThomas Huth <thuth@redhat.com>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Some times it is useful for architecture implementations of KVM to know
when the VCPU thread is about to block or when it comes back from
blocking (arm/arm64 needs to know this to properly implement timers, for
example).

Therefore provide a generic architecture callback function in line with
what we do elsewhere for KVM generic-arch interactions.
Reviewed-by: NMarc Zyngier <marc.zyngier@arm.com>
Signed-off-by: NChristoffer Dall <christoffer.dall@linaro.org>

This reverts commit 9678cdaa ("Use the POWER8 Micro Partition
Prefetch Engine in KVM HV on POWER8") because the original commit had
multiple, partly self-cancelling bugs, that could cause occasional
memory corruption.

In fact the logmpp instruction was incorrectly using register r0 as the
source of the buffer address and operation code, and depending on what
was in r0, it would either do nothing or corrupt the 64k page pointed to
by r0.

The logmpp instruction encoding and the operation code definitions could
be corrected, but then there is the problem that there is no clearly
defined way to know when the hardware has finished writing to the
buffer.

The original commit attempted to work around this by aborting the
write-out before starting the prefetch, but this is ineffective in the
case where the virtual core is now executing on a different physical
core from the one where the write-out was initiated.

These problems plus advice from the hardware designers not to use the
function (since the measured performance improvement from using the
feature was actually mostly negative), mean that reverting the code is
the best option.

Fixes: 9678cdaa ("Use the POWER8 Micro Partition Prefetch Engine in KVM HV on POWER8")
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

We observed some performance degradation on s390x with dynamic
halt polling. Until we can provide a proper fix, let's enable
halt_poll_ns as default only for supported architectures.

Architectures are now free to set their own halt_poll_ns
default value.
Signed-off-by: NDavid Hildenbrand <dahi@linux.vnet.ibm.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

This new statistic can help diagnosing VCPUs that, for any reason,
trigger bad behavior of halt_poll_ns autotuning.

For example, say halt_poll_ns = 480000, and wakeups are spaced exactly
like 479us, 481us, 479us, 481us. Then KVM always fails polling and wastes
10+20+40+80+160+320+480 = 1110 microseconds out of every
479+481+479+481+479+481+479 = 3359 microseconds. The VCPU then
is consuming about 30% more CPU than it would use without
polling.  This would show as an abnormally high number of
attempted polling compared to the successful polls.

Acked-by: Christian Borntraeger <borntraeger@de.ibm.com<
Reviewed-by: NDavid Matlack <dmatlack@google.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

The size of the Problem State Priority Boost Register is only
32 bits, but the kvm_vcpu_arch->pspb variable is declared as
"ulong", ie. 64-bit. However, the assembler code accesses this
variable with 32-bit accesses, and the KVM_REG_PPC_PSPB macro
is defined with SIZE_U32, too, so that the current code is
broken on big endian hosts: kvmppc_get_one_reg_hv() will only
return zero for this register since it is using the wrong half
of the pspb variable. Let's fix this problem by adjusting the
size of the pspb field in the kvm_vcpu_arch structure.
Signed-off-by: NThomas Huth <thuth@redhat.com>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

This fixes a bug in the tracking of pages that get modified by the
guest.  If the guest creates a large-page HPTE, writes to memory
somewhere within the large page, and then removes the HPTE, we only
record the modified state for the first normal page within the large
page, when in fact the guest might have modified some other normal
page within the large page.

To fix this we use some unused bits in the rmap entry to record the
order (log base 2) of the size of the page that was modified, when
removing an HPTE.  Then in kvm_test_clear_dirty_npages() we use that
order to return the correct number of modified pages.

The same thing could in principle happen when removing a HPTE at the
host's request, i.e. when paging out a page, except that we never
page out large pages, and the guest can only create large-page HPTEs
if the guest RAM is backed by large pages.  However, we also fix
this case for the sake of future-proofing.

The reference bit is also subject to the same loss of information.  We
don't make the same fix here for the reference bit because there isn't
an interface for userspace to find out which pages the guest has
referenced, whereas there is one for userspace to find out which pages
the guest has modified.  Because of this loss of information, the
kvm_age_hva_hv() and kvm_test_age_hva_hv() functions might incorrectly
say that a page has not been referenced when it has, but that doesn't
matter greatly because we never page or swap out large pages.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

This builds on the ability to run more than one vcore on a physical
core by using the micro-threading (split-core) modes of the POWER8
chip.  Previously, only vcores from the same VM could be run together,
and (on POWER8) only if they had just one thread per core.  With the
ability to split the core on guest entry and unsplit it on guest exit,
we can run up to 8 vcpu threads from up to 4 different VMs, and we can
run multiple vcores with 2 or 4 vcpus per vcore.

Dynamic micro-threading is only available if the static configuration
of the cores is whole-core mode (unsplit), and only on POWER8.

To manage this, we introduce a new kvm_split_mode struct which is
shared across all of the subcores in the core, with a pointer in the
paca on each thread.  In addition we extend the core_info struct to
have information on each subcore.  When deciding whether to add a
vcore to the set already on the core, we now have two possibilities:
(a) piggyback the vcore onto an existing subcore, or (b) start a new
subcore.

Currently, when any vcpu needs to exit the guest and switch to host
virtual mode, we interrupt all the threads in all subcores and switch
the core back to whole-core mode.  It may be possible in future to
allow some of the subcores to keep executing in the guest while
subcore 0 switches to the host, but that is not implemented in this
patch.

This adds a module parameter called dynamic_mt_modes which controls
which micro-threading (split-core) modes the code will consider, as a
bitmap.  In other words, if it is 0, no micro-threading mode is
considered; if it is 2, only 2-way micro-threading is considered; if
it is 4, only 4-way, and if it is 6, both 2-way and 4-way
micro-threading mode will be considered.  The default is 6.

With this, we now have secondary threads which are the primary thread
for their subcore and therefore need to do the MMU switch.  These
threads will need to be started even if they have no vcpu to run, so
we use the vcore pointer in the PACA rather than the vcpu pointer to
trigger them.

It is now possible for thread 0 to find that an exit has been
requested before it gets to switch the subcore state to the guest.  In
that case we haven't added the guest's timebase offset to the
timebase, so we need to be careful not to subtract the offset in the
guest exit path.  In fact we just skip the whole path that switches
back to host context, since we haven't switched to the guest context.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

When running a virtual core of a guest that is configured with fewer
threads per core than the physical cores have, the extra physical
threads are currently unused.  This makes it possible to use them to
run one or more other virtual cores from the same guest when certain
conditions are met.  This applies on POWER7, and on POWER8 to guests
with one thread per virtual core.  (It doesn't apply to POWER8 guests
with multiple threads per vcore because they require a 1-1 virtual to
physical thread mapping in order to be able to use msgsndp and the
TIR.)

The idea is that we maintain a list of preempted vcores for each
physical cpu (i.e. each core, since the host runs single-threaded).
Then, when a vcore is about to run, it checks to see if there are
any vcores on the list for its physical cpu that could be
piggybacked onto this vcore's execution.  If so, those additional
vcores are put into state VCORE_PIGGYBACK and their runnable VCPU
threads are started as well as the original vcore, which is called
the master vcore.

After the vcores have exited the guest, the extra ones are put back
onto the preempted list if any of their VCPUs are still runnable and
not idle.

This means that vcpu->arch.ptid is no longer necessarily the same as
the physical thread that the vcpu runs on.  In order to make it easier
for code that wants to send an IPI to know which CPU to target, we
now store that in a new field in struct vcpu_arch, called thread_cpu.
Reviewed-by: NDavid Gibson <david@gibson.dropbear.id.au>
Tested-by: NLaurent Vivier <lvivier@redhat.com>
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

Prepare for the case of multiple address spaces.
Reviewed-by: NRadim Krcmar <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Currently, the entry_exit_count field in the kvmppc_vcore struct
contains two 8-bit counts, one of the threads that have started entering
the guest, and one of the threads that have started exiting the guest.
This changes it to an entry_exit_map field which contains two bitmaps
of 8 bits each.  The advantage of doing this is that it gives us a
bitmap of which threads need to be signalled when exiting the guest.
That means that we no longer need to use the trick of setting the
HDEC to 0 to pull the other threads out of the guest, which led in
some cases to a spurious HDEC interrupt on the next guest entry.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

We can tell when a secondary thread has finished running a guest by
the fact that it clears its kvm_hstate.kvm_vcpu pointer, so there
is no real need for the nap_count field in the kvmppc_vcore struct.
This changes kvmppc_wait_for_nap to poll the kvm_hstate.kvm_vcpu
pointers of the secondary threads rather than polling vc->nap_count.
Besides reducing the size of the kvmppc_vcore struct by 8 bytes,
this also means that we can tell which secondary threads have got
stuck and thus print a more informative error message.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

Rather than calling cond_resched() in kvmppc_run_core() before doing
the post-processing for the vcpus that we have just run (that is,
calling kvmppc_handle_exit_hv(), kvmppc_set_timer(), etc.), we now do
that post-processing before calling cond_resched(), and that post-
processing is moved out into its own function, post_guest_process().

The reschedule point is now in kvmppc_run_vcpu() and we define a new
vcore state, VCORE_PREEMPT, to indicate that that the vcore's runner
task is runnable but not running.  (Doing the reschedule with the
vcore in VCORE_INACTIVE state would be bad because there are potentially
other vcpus waiting for the runner in kvmppc_wait_for_exec() which
then wouldn't get woken up.)

Also, we make use of the handy cond_resched_lock() function, which
unlocks and relocks vc->lock for us around the reschedule.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

* Remove unused kvmppc_vcore::n_busy field.
* Remove setting of RMOR, since it was only used on PPC970 and the
  PPC970 KVM support has been removed.
* Don't use r1 or r2 in setting the runlatch since they are
  conventionally reserved for other things; use r0 instead.
* Streamline the code a little and remove the ext_interrupt_to_host
  label.
* Add some comments about register usage.
* hcall_try_real_mode doesn't need to be global, and can't be
  called from C code anyway.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

Previously, if kvmppc_run_core() was running a VCPU that needed a VPA
update (i.e. one of its 3 virtual processor areas needed to be pinned
in memory so the host real mode code can update it on guest entry and
exit), we would drop the vcore lock and do the update there and then.
Future changes will make it inconvenient to drop the lock, so instead
we now remove it from the list of runnable VCPUs and wake up its
VCPU task.  This will have the effect that the VCPU task will exit
kvmppc_run_vcpu(), go around the do loop in kvmppc_vcpu_run_hv(), and
re-enter kvmppc_run_vcpu(), whereupon it will do the necessary call
to kvmppc_update_vpas() and then rejoin the vcore.

The one complication is that the runner VCPU (whose VCPU task is the
current task) might be one of the ones that gets removed from the
runnable list.  In that case we just return from kvmppc_run_core()
and let the code in kvmppc_run_vcpu() wake up another VCPU task to be
the runner if necessary.

This all means that the VCORE_STARTING state is no longer used, so we
remove it.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

This reads the timebase at various points in the real-mode guest
entry/exit code and uses that to accumulate total, minimum and
maximum time spent in those parts of the code.  Currently these
times are accumulated per vcpu in 5 parts of the code:

* rm_entry - time taken from the start of kvmppc_hv_entry() until
  just before entering the guest.
* rm_intr - time from when we take a hypervisor interrupt in the
  guest until we either re-enter the guest or decide to exit to the
  host.  This includes time spent handling hcalls in real mode.
* rm_exit - time from when we decide to exit the guest until the
  return from kvmppc_hv_entry().
* guest - time spend in the guest
* cede - time spent napping in real mode due to an H_CEDE hcall
  while other threads in the same vcore are active.

These times are exposed in debugfs in a directory per vcpu that
contains a file called "timings".  This file contains one line for
each of the 5 timings above, with the name followed by a colon and
4 numbers, which are the count (number of times the code has been
executed), the total time, the minimum time, and the maximum time,
all in nanoseconds.

The overhead of the extra code amounts to about 30ns for an hcall that
is handled in real mode (e.g. H_SET_DABR), which is about 25%.  Since
production environments may not wish to incur this overhead, the new
code is conditional on a new config symbol,
CONFIG_KVM_BOOK3S_HV_EXIT_TIMING.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

This creates a debugfs directory for each HV guest (assuming debugfs
is enabled in the kernel config), and within that directory, a file
by which the contents of the guest's HPT (hashed page table) can be
read.  The directory is named vmnnnn, where nnnn is the PID of the
process that created the guest.  The file is named "htab".  This is
intended to help in debugging problems in the host's management
of guest memory.

The contents of the file consist of a series of lines like this:

  3f48 4000d032bf003505 0000000bd7ff1196 00000003b5c71196

The first field is the index of the entry in the HPT, the second and
third are the HPT entry, so the third entry contains the real page
number that is mapped by the entry if the entry's valid bit is set.
The fourth field is the guest's view of the second doubleword of the
entry, so it contains the guest physical address.  (The format of the
second through fourth fields are described in the Power ISA and also
in arch/powerpc/include/asm/mmu-hash64.h.)
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

We don't support real-mode areas now that 970 support is removed.
Remove the remaining details of rma from the code.  Also rename
rma_setup_done to hpte_setup_done to better reflect the changes.
Signed-off-by: NAneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

Sometimes the KVM code on powerpc needs to emulate load or store
instructions from the guest, which can include both normal and byte
reversed forms.

We currently (AFAICT) handle this correctly, but some variable names are
very misleading.  In particular we use "is_bigendian" in several places to
actually mean "is the IO the same endian as the host", but we now support
little-endian powerpc hosts.  This also ties into the misleadingly named
ld_le*() and st_le*() functions, which in fact always byteswap, even on
an LE host.

This patch cleans this up by renaming to more accurate "host_swabbed", and
uses the generic swab*() functions instead of the powerpc specific and
misleadingly named ld_le*() and st_le*() functions.
Signed-off-by: NDavid Gibson <david@gibson.dropbear.id.au>
Reviewed-by: NAlexander Graf <agraf@suse.de>
Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>

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>

Currently the H_CONFER hcall is implemented in kernel virtual mode,
meaning that whenever a guest thread does an H_CONFER, all the threads
in that virtual core have to exit the guest.  This is bad for
performance because it interrupts the other threads even if they
are doing useful work.

The H_CONFER hcall is called by a guest VCPU when it is spinning on a
spinlock and it detects that the spinlock is held by a guest VCPU that
is currently not running on a physical CPU.  The idea is to give this
VCPU's time slice to the holder VCPU so that it can make progress
towards releasing the lock.

To avoid having the other threads exit the guest unnecessarily,
we add a real-mode implementation of H_CONFER that checks whether
the other threads are doing anything.  If all the other threads
are idle (i.e. in H_CEDE) or trying to confer (i.e. in H_CONFER),
it returns H_TOO_HARD which causes a guest exit and allows the
H_CONFER to be handled in virtual mode.

Otherwise it spins for a short time (up to 10 microseconds) to give
other threads the chance to observe that this thread is trying to
confer.  The spin loop also terminates when any thread exits the guest
or when all other threads are idle or trying to confer.  If the
timeout is reached, the H_CONFER returns H_SUCCESS.  In this case the
guest VCPU will recheck the spinlock word and most likely call
H_CONFER again.

This also improves the implementation of the H_CONFER virtual mode
handler.  If the VCPU is part of a virtual core (vcore) which is
runnable, there will be a 'runner' VCPU which has taken responsibility
for running the vcore.  In this case we yield to the runner VCPU
rather than the target VCPU.

We also introduce a check on the target VCPU's yield count: if it
differs from the yield count passed to H_CONFER, the target VCPU
has run since H_CONFER was called and may have already released
the lock.  This check is required by PAPR.
Signed-off-by: NSam Bobroff <sam.bobroff@au1.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

There are two ways in which a guest instruction can be obtained from
the guest in the guest exit code in book3s_hv_rmhandlers.S.  If the
exit was caused by a Hypervisor Emulation interrupt (i.e. an illegal
instruction), the offending instruction is in the HEIR register
(Hypervisor Emulation Instruction Register).  If the exit was caused
by a load or store to an emulated MMIO device, we load the instruction
from the guest by turning data relocation on and loading the instruction
with an lwz instruction.

Unfortunately, in the case where the guest has opposite endianness to
the host, these two methods give results of different endianness, but
both get put into vcpu->arch.last_inst.  The HEIR value has been loaded
using guest endianness, whereas the lwz will load the instruction using
host endianness.  The rest of the code that uses vcpu->arch.last_inst
assumes it was loaded using host endianness.

To fix this, we define a new vcpu field to store the HEIR value.  Then,
in kvmppc_handle_exit_hv(), we transfer the value from this new field to
vcpu->arch.last_inst, doing a byte-swap if the guest and host endianness
differ.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

This removes the code that was added to enable HV KVM to work
on PPC970 processors.  The PPC970 is an old CPU that doesn't
support virtualizing guest memory.  Removing PPC970 support also
lets us remove the code for allocating and managing contiguous
real-mode areas, the code for the !kvm->arch.using_mmu_notifiers
case, the code for pinning pages of guest memory when first
accessed and keeping track of which pages have been pinned, and
the code for handling H_ENTER hypercalls in virtual mode.

Book3S HV KVM is now supported only on POWER7 and POWER8 processors.
The KVM_CAP_PPC_RMA capability now always returns 0.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>

Currently the calculations of stolen time for PPC Book3S HV guests
uses fields in both the vcpu struct and the kvmppc_vcore struct.  The
fields in the kvmppc_vcore struct are protected by the
vcpu->arch.tbacct_lock of the vcpu that has taken responsibility for
running the virtual core.  This works correctly but confuses lockdep,
because it sees that the code takes the tbacct_lock for a vcpu in
kvmppc_remove_runnable() and then takes another vcpu's tbacct_lock in
vcore_stolen_time(), and it thinks there is a possibility of deadlock,
causing it to print reports like this:

=============================================
[ INFO: possible recursive locking detected ]
3.18.0-rc7-kvm-00016-g8db4bc6 #89 Not tainted
---------------------------------------------
qemu-system-ppc/6188 is trying to acquire lock:
 (&(&vcpu->arch.tbacct_lock)->rlock){......}, at: [<d00000000ecb1fe8>] .vcore_stolen_time+0x48/0xd0 [kvm_hv]

but task is already holding lock:
 (&(&vcpu->arch.tbacct_lock)->rlock){......}, at: [<d00000000ecb25a0>] .kvmppc_remove_runnable.part.3+0x30/0xd0 [kvm_hv]

other info that might help us debug this:
 Possible unsafe locking scenario:

       CPU0
       ----
  lock(&(&vcpu->arch.tbacct_lock)->rlock);
  lock(&(&vcpu->arch.tbacct_lock)->rlock);

 *** DEADLOCK ***

 May be due to missing lock nesting notation

3 locks held by qemu-system-ppc/6188:
 #0:  (&vcpu->mutex){+.+.+.}, at: [<d00000000eb93f98>] .vcpu_load+0x28/0xe0 [kvm]
 #1:  (&(&vcore->lock)->rlock){+.+...}, at: [<d00000000ecb41b0>] .kvmppc_vcpu_run_hv+0x530/0x1530 [kvm_hv]
 #2:  (&(&vcpu->arch.tbacct_lock)->rlock){......}, at: [<d00000000ecb25a0>] .kvmppc_remove_runnable.part.3+0x30/0xd0 [kvm_hv]

stack backtrace:
CPU: 40 PID: 6188 Comm: qemu-system-ppc Not tainted 3.18.0-rc7-kvm-00016-g8db4bc6 #89
Call Trace:
[c000000b2754f3f0] [c000000000b31b6c] .dump_stack+0x88/0xb4 (unreliable)
[c000000b2754f470] [c0000000000faeb8] .__lock_acquire+0x1878/0x2190
[c000000b2754f600] [c0000000000fbf0c] .lock_acquire+0xcc/0x1a0
[c000000b2754f6d0] [c000000000b2954c] ._raw_spin_lock_irq+0x4c/0x70
[c000000b2754f760] [d00000000ecb1fe8] .vcore_stolen_time+0x48/0xd0 [kvm_hv]
[c000000b2754f7f0] [d00000000ecb25b4] .kvmppc_remove_runnable.part.3+0x44/0xd0 [kvm_hv]
[c000000b2754f880] [d00000000ecb43ec] .kvmppc_vcpu_run_hv+0x76c/0x1530 [kvm_hv]
[c000000b2754f9f0] [d00000000eb9f46c] .kvmppc_vcpu_run+0x2c/0x40 [kvm]
[c000000b2754fa60] [d00000000eb9c9a4] .kvm_arch_vcpu_ioctl_run+0x54/0x160 [kvm]
[c000000b2754faf0] [d00000000eb94538] .kvm_vcpu_ioctl+0x498/0x760 [kvm]
[c000000b2754fcb0] [c000000000267eb4] .do_vfs_ioctl+0x444/0x770
[c000000b2754fd90] [c0000000002682a4] .SyS_ioctl+0xc4/0xe0
[c000000b2754fe30] [c0000000000092e4] syscall_exit+0x0/0x98

In order to make the locking easier to analyse, we change the code to
use a spinlock in the kvmppc_vcore struct to protect the stolen_tb and
preempt_tb fields.  This lock needs to be an irq-safe lock since it is
used in the kvmppc_core_vcpu_load_hv() and kvmppc_core_vcpu_put_hv()
functions, which are called with the scheduler rq lock held, which is
an irq-safe lock.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NAlexander Graf <agraf@suse.de>