The backwards_tsc_observed global introduced in commit 16a96021 is never
reset to false. If a VM happens to be running while the host is suspended
(a common source of the TSC jumping backwards), master clock will never
be enabled again for any VM. In contrast, if no VM is running while the
host is suspended, master clock is unaffected. This is inconsistent and
unnecessarily strict. Let's track the backwards_tsc_observed variable
separately and let each VM start with a clean slate.

Real world impact: My Windows VMs get slower after my laptop undergoes a
suspend/resume cycle. The only way to get the perf back is unloading and
reloading the kvm module.
Signed-off-by: NLadi Prosek <lprosek@redhat.com>
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>

Adds the plumbing to disable A/D bits in the MMU based on a new role
bit, ad_disabled. When A/D is disabled, the MMU operates as though A/D
aren't available (i.e., using access tracking faults instead).

To avoid SP -> kvm_mmu_page.role.ad_disabled lookups all over the
place, A/D disablement is now stored in the SPTE. This state is stored
in the SPTE by tweaking the use of SPTE_SPECIAL_MASK for access
tracking. Rather than just setting SPTE_SPECIAL_MASK when an
access-tracking SPTE is non-present, we now always set
SPTE_SPECIAL_MASK for access-tracking SPTEs.
Signed-off-by: NPeter Feiner <pfeiner@google.com>
[Use role.ad_disabled even for direct (non-shadow) EPT page tables.  Add
 documentation and a few MMU_WARN_ONs. - Paolo]
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Marc Zyngier suggested that we define the arch specific VCPU request
base, rather than requiring each arch to remember to start from 8.
That suggestion, along with Radim Krcmar's recent VCPU request flag
addition, snowballed into defining something of an arch VCPU request
defining API.

No functional change.

(Looks like x86 is running out of arch VCPU request bits.  Maybe
 someday we'll need to extend to 64.)
Signed-off-by: NAndrew Jones <drjones@redhat.com>
Acked-by: NChristoffer Dall <cdall@linaro.org>
Signed-off-by: NChristoffer Dall <cdall@linaro.org>

In some fio benchmarks, halt_poll_ns=400000 caused CPU utilization to
increase heavily even in cases where the performance improvement was
small.  In particular, bandwidth divided by CPU usage was as much as
60% lower.

To some extent this is the expected effect of the patch, and the
additional CPU utilization is only visible when running the
benchmarks.  However, halving the threshold also halves the extra
CPU utilization (from +30-130% to +20-70%) and has no negative
effect on performance.
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>

When KVM updates accessed/dirty bits, this hook can be used
to invoke an arch specific function that implements/emulates
dirty logging such as PML.
Signed-off-by: NBandan Das <bsd@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

We needed the lock to avoid racing with creation of the irqchip on x86. As
kvm_set_irq_routing() calls srcu_synchronize_expedited(), this lock
might be held for a longer time.

Let's introduce an arch specific callback to check if we can actually
add irq routes. For x86, all we have to do is check if we have an
irqchip in the kernel. We don't need kvm->lock at that point as the
irqchip is marked as inititalized only when actually fully created.
Reported-by: NSteve Rutherford <srutherford@google.com>
Reviewed-by: NRadim Krčmář <rkrcmar@redhat.com>
Fixes: 1df6dded ("KVM: x86: race between KVM_SET_GSI_ROUTING and KVM_CREATE_IRQCHIP")
Signed-off-by: NDavid Hildenbrand <david@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

kvm_make_all_requests() provides a synchronization that waits until all
kicked VCPUs have acknowledged the kick.  This is important for
KVM_REQ_MMU_RELOAD as it prevents freeing while lockless paging is
underway.

This patch adds the synchronization property into all requests that are
currently being used with kvm_make_all_requests() in order to preserve
the current behavior and only introduce a new framework.  Removing it
from requests where it is not necessary is left for future patches.
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Some operations must ensure that the guest is not running with stale
data, but if the guest is halted, then the update can wait until another
event happens.  kvm_make_all_requests() currently doesn't wake up, so we
can mark all requests used with it.

First 8 bits were arbitrarily reserved for request numbers.

Most uses of requests have the request type as a constant, so a compiler
will optimize the '&'.

An alternative would be to have an inline function that would return
whether the request needs a wake-up or not, but I like this one better
even though it might produce worse assembly.
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
Reviewed-by: NAndrew Jones <drjones@redhat.com>
Reviewed-by: NCornelia Huck <cornelia.huck@de.ibm.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Hardware support for faulting on the cpuid instruction is not required to
emulate it, because cpuid triggers a VM exit anyways. KVM handles the relevant
MSRs (MSR_PLATFORM_INFO and MSR_MISC_FEATURES_ENABLE) and upon a
cpuid-induced VM exit checks the cpuid faulting state and the CPL.
kvm_require_cpl is even kind enough to inject the GP fault for us.
Signed-off-by: NKyle Huey <khuey@kylehuey.com>
Reviewed-by: NDavid Matlack <dmatlack@google.com>
[Return "1" from kvm_emulate_cpuid, it's not void. - Paolo]
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Let's add a new mode and set it while we create the irqchip via
KVM_CREATE_IRQCHIP and KVM_CAP_SPLIT_IRQCHIP.

This mode will be used later to test if adding routes
(in kvm_set_routing_entry()) is already allowed.
Signed-off-by: NDavid Hildenbrand <david@redhat.com>
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>

Its value has never changed; we might as well make it part of the ABI instead
of using the return value of KVM_CHECK_EXTENSION(KVM_CAP_COALESCED_MMIO).

Because PPC does not always make MMIO available, the code has to be made
dependent on CONFIG_KVM_MMIO rather than KVM_COALESCED_MMIO_PAGE_OFFSET.
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>

Now use bit 6 of EPTP to optionally enable A/D bits for EPTP.  Another
thing to change is that, when EPT accessed and dirty bits are not in use,
VMX treats accesses to guest paging structures as data reads.  When they
are in use (bit 6 of EPTP is set), they are treated as writes and the
corresponding EPT dirty bit is set.  The MMU didn't know this detail,
so this patch adds it.

We also have to fix up the exit qualification.  It may be wrong because
KVM sets bit 6 but the guest might not.

L1 emulates EPT A/D bits using write permissions, so in principle it may
be possible for EPT A/D bits to be used by L1 even though not available
in hardware.  The problem is that guest page-table walks will be treated
as reads rather than writes, so they would not cause an EPT violation.
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>
[Fixed typo in walk_addr_generic() comment and changed bit clear +
 conditional-set pattern in handle_ept_violation() to conditional-clear]
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>

The FPU is always active now when running KVM.
Reviewed-by: NDavid Matlack <dmatlack@google.com>
Reviewed-by: NBandan Das <bsd@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Calls to apic_find_highest_irr are scanning IRR twice, once
in vmx_sync_pir_from_irr and once in apic_search_irr.  Change
sync_pir_from_irr to get the new maximum IRR from kvm_apic_update_irr;
now that it does the computation, it can also do the RVI write.

In order to avoid complications in svm.c, make the callback optional.
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

This statistic can be useful to estimate the cost of an IRQ injection
scenario, by comparing it with irq_injections.  For example the stat
shows that sti;hlt triggers more KVM_REQ_EVENT than sti;nop.
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

When a guest causes a NPF which requires emulation, KVM sometimes walks
the guest page tables to translate the GVA to a GPA. This is unnecessary
most of the time on AMD hardware since the hardware provides the GPA in
EXITINFO2.

The only exception cases involve string operations involving rep or
operations that use two memory locations. With rep, the GPA will only be
the value of the initial NPF and with dual memory locations we won't know
which memory address was translated into EXITINFO2.
Signed-off-by: NTom Lendacky <thomas.lendacky@amd.com>
Reviewed-by: NBorislav Petkov <bp@suse.de>
Signed-off-by: NBrijesh Singh <brijesh.singh@amd.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

This change implements lockless access tracking for Intel CPUs without EPT
A bits. This is achieved by marking the PTEs as not-present (but not
completely clearing them) when clear_flush_young() is called after marking
the pages as accessed. When an EPT Violation is generated as a result of
the VM accessing those pages, the PTEs are restored to their original values.
Signed-off-by: NJunaid Shahid <junaids@google.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

MMIO SPTEs currently set both bits 62 and 63 to distinguish them as special
PTEs. However, bit 63 is used as the SVE bit in Intel EPT PTEs. The SVE bit
is ignored for misconfigured PTEs but not necessarily for not-Present PTEs.
Since MMIO SPTEs use an EPT misconfiguration, so using bit 63 for them is
acceptable. However, the upcoming fast access tracking feature adds another
type of special tracking PTE, which uses not-Present PTEs and hence should
not set bit 63.

In order to use common bits to distinguish both type of special PTEs, we
now use only bit 62 as the special bit.
Signed-off-by: NJunaid Shahid <junaids@google.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

When using two-dimensional paging, the mmu_page_hash (which provides
lookups for existing kvm_mmu_page structs), becomes imbalanced; with
too many collisions in buckets 0 and 512. This has been seen to cause
mmu_lock to be held for multiple milliseconds in kvm_mmu_get_page on
VMs with a large amount of RAM mapped with 4K pages.

The current hash function uses the lower 10 bits of gfn to index into
mmu_page_hash. When doing shadow paging, gfn is the address of the
guest page table being shadow. These tables are 4K-aligned, which
makes the low bits of gfn a good hash. However, with two-dimensional
paging, no guest page tables are being shadowed, so gfn is the base
address that is mapped by the table. Thus page tables (level=1) have
a 2MB aligned gfn, page directories (level=2) have a 1GB aligned gfn,
etc. This means hashes will only differ in their 10th bit.

hash_64() provides a better hash. For example, on a VM with ~200G
(99458 direct=1 kvm_mmu_page structs):

hash            max_mmu_page_hash_collisions
--------------------------------------------
low 10 bits     49847
hash_64         105
perfect         97

While we're changing the hash, increase the table size by 4x to better
support large VMs (further reduces number of collisions in 200G VM to
29).

Note that hash_64() does not provide a good distribution prior to commit
ef703f49 ("Eliminate bad hash multipliers from hash_32() and
hash_64()").
Signed-off-by: NDavid Matlack <dmatlack@google.com>
Change-Id: I5aa6b13c834722813c6cca46b8b1ed6f53368ade
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Report the maximum number of mmu_page_hash collisions as a per-VM stat.
This will make it easy to identify problems with the mmu_page_hash in
the future.
Signed-off-by: NDavid Matlack <dmatlack@google.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

irqchip_in_kernel() tried to save a bit by reusing pic_irqchip(), but it
just complicated the code.
Add a separate state for the irqchip mode.
Reviewed-by: NDavid Hildenbrand <david@redhat.com>
[Used Paolo's version of condition in irqchip_in_kernel().]
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>

There is no point in having an extra type for extra confusion. u64 is
unambiguous.

Conversion was done with the following coccinelle script:

@rem@
@@
-typedef u64 cycle_t;

@fix@
typedef cycle_t;
@@
-cycle_t
+u64
Signed-off-by: NThomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: John Stultz <john.stultz@linaro.org>

Introduce a new mutex to avoid an AB-BA deadlock between kvm->lock and
vcpu->mutex.  Protect accesses in kvm_hv_setup_tsc_page too, as suggested
by Roman.
Reported-by: NDmitry Vyukov <dvyukov@google.com>
Reviewed-by: NRoman Kagan <rkagan@virtuozzo.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Loading CR3 as part of emulating vmentry is different from regular CR3 loads,
as implemented in kvm_set_cr3, in several ways.

* different rules are followed to check CR3 and it is desirable for the caller
to distinguish between the possible failures
* PDPTRs are not loaded if PAE paging and nested EPT are both enabled
* many MMU operations are not necessary

This patch introduces nested_vmx_load_cr3 suitable for CR3 loads as part of
nested vmentry and vmexit, and makes use of it on the nested vmentry path.
Signed-off-by: NLadi Prosek <lprosek@redhat.com>
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>

kvm_skip_emulated_instruction calls both
kvm_x86_ops->skip_emulated_instruction and kvm_vcpu_check_singlestep,
skipping the emulated instruction and generating a trap if necessary.

Replacing skip_emulated_instruction calls with
kvm_skip_emulated_instruction is straightforward, except for:

- ICEBP, which is already inside a trap, so avoid triggering another trap.
- Instructions that can trigger exits to userspace, such as the IO insns,
  MOVs to CR8, and HALT. If kvm_skip_emulated_instruction does trigger a
  KVM_GUESTDBG_SINGLESTEP exit, and the handling code for
  IN/OUT/MOV CR8/HALT also triggers an exit to userspace, the latter will
  take precedence. The singlestep will be triggered again on the next
  instruction, which is the current behavior.
- Task switch instructions which would require additional handling (e.g.
  the task switch bit) and are instead left alone.
- Cases where VMLAUNCH/VMRESUME do not proceed to the next instruction,
  which do not trigger singlestep traps as mentioned previously.
Signed-off-by: NKyle Huey <khuey@kylehuey.com>
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>

Once skipping the emulated instruction can potentially trigger an exit to
userspace (via KVM_GUESTDBG_SINGLESTEP) kvm_emulate_cpuid will need to
propagate a return value.
Signed-off-by: NKyle Huey <khuey@kylehuey.com>
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>

Update the I/O interception support to add the kvm_fast_pio_in function
to speed up the in instruction similar to the out instruction.
Signed-off-by: NTom Lendacky <thomas.lendacky@amd.com>
Reviewed-by: NBorislav Petkov <bp@suse.de>
Signed-off-by: NBrijesh Singh <brijesh.singh@amd.com>
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>

AMD hardware adds two additional bits to aid in nested page fault handling.

Bit 32 - NPF occurred while translating the guest's final physical address
Bit 33 - NPF occurred while translating the guest page tables

The guest page tables fault indicator can be used as an aid for nested
virtualization. Using V0 for the host, V1 for the first level guest and
V2 for the second level guest, when both V1 and V2 are using nested paging
there are currently a number of unnecessary instruction emulations. When
V2 is launched shadow paging is used in V1 for the nested tables of V2. As
a result, KVM marks these pages as RO in the host nested page tables. When
V2 exits and we resume V1, these pages are still marked RO.

Every nested walk for a guest page table is treated as a user-level write
access and this causes a lot of NPFs because the V1 page tables are marked
RO in the V0 nested tables. While executing V1, when these NPFs occur KVM
sees a write to a read-only page, emulates the V1 instruction and unprotects
the page (marking it RW). This patch looks for cases where we get a NPF due
to a guest page table walk where the page was marked RO. It immediately
unprotects the page and resumes the guest, leading to far fewer instruction
emulations when nested virtualization is used.
Signed-off-by: NTom Lendacky <thomas.lendacky@amd.com>
Reviewed-by: NBorislav Petkov <bp@suse.de>
Signed-off-by: NBrijesh Singh <brijesh.singh@amd.com>
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>

Since commit a545ab6a ("kvm: x86: add tsc_offset field to struct
kvm_vcpu_arch", 2016-09-07) the offset between host and L1 TSC is
cached and need not be fished out of the VMCS or VMCB.  This means
that we can implement adjust_tsc_offset_guest and read_l1_tsc
entirely in generic code.  The simplification is particularly
significant for VMX code, where vmx->nested.vmcs01_tsc_offset
was duplicating what is now in vcpu->arch.tsc_offset.  Therefore
the vmcs01_tsc_offset can be dropped completely.

More importantly, this fixes KVM_GET_CLOCK/KVM_SET_CLOCK
which, after commit 108b249c ("KVM: x86: introduce get_kvmclock_ns",
2016-09-01) called read_l1_tsc while the VMCS was not loaded.
It thus returned bogus values on Intel CPUs.

Fixes: 108b249cReported-by: NRoman Kagan <rkagan@virtuozzo.com>
Reviewed-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Lately tsc page was implemented but filled with empty
values. This patch setup tsc page scale and offset based
on vcpu tsc, tsc_khz and  HV_X64_MSR_TIME_REF_COUNT value.

The valid tsc page drops HV_X64_MSR_TIME_REF_COUNT msr
reads count to zero which potentially improves performance.
Signed-off-by: NAndrey Smetanin <asmetanin@virtuozzo.com>
Reviewed-by: NPeter Hornyack <peterhornyack@google.com>
Reviewed-by: NRadim Krčmář <rkrcmar@redhat.com>
CC: Paolo Bonzini <pbonzini@redhat.com>
CC: Roman Kagan <rkagan@virtuozzo.com>
CC: Denis V. Lunev <den@openvz.org>
[Computation of TSC page parameters rewritten to use the Linux timekeeper
 parameters. - Paolo]
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

The TSC offset can now be read directly from struct kvm_arch_vcpu.
Signed-off-by: NLuiz Capitulino <lcapitulino@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

A future commit will want to easily read a vCPU's TSC offset,
so we store it in struct kvm_arch_vcpu_arch for easy access.
Signed-off-by: NLuiz Capitulino <lcapitulino@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

This patch introduces avic_ga_log_notifier, which will be called
by IOMMU driver whenever it handles the Guest vAPIC (GA) log entry.
Reviewed-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NSuravee Suthikulpanit <suravee.suthikulpanit@amd.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Introduces per-VM AVIC ID and helper functions to manage the IDs.
Currently, the ID will be used to implement 32-bit AVIC IOMMU GA tag.

The ID is 24-bit one-based indexing value, and is managed via helper
functions to get the next ID, or to free an ID once a VM is destroyed.
There should be no ID conflict for any active VMs.
Reviewed-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NSuravee Suthikulpanit <suravee.suthikulpanit@amd.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

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>

kzalloc was replaced with kvm_kvzalloc to allow non-contiguous areas and
rcu had to be modified to cope with it.

The practical limit for KVM_MAX_VCPU_ID right now is INT_MAX, but lower
value was chosen in case there were bugs.  1023 is sufficient maximum
APIC ID for 288 VCPUs.
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

288 is in high demand because of Knights Landing CPU.
We cannot set the limit to 640k, because that would be wasting space.
Reviewed-by: NPaolo Bonzini <pbonzini@redhat.com>
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Add KVM_X2APIC_API_DISABLE_BROADCAST_QUIRK as a feature flag to
KVM_CAP_X2APIC_API.

The quirk made KVM interpret 0xff as a broadcast even in x2APIC mode.
The enableable capability is needed in order to support standard x2APIC and
remain backward compatible.
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
[Expand kvm_apic_mda comment. - Paolo]
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

KVM_CAP_X2APIC_API is a capability for features related to x2APIC
enablement.  KVM_X2APIC_API_32BIT_FORMAT feature can be enabled to
extend APIC ID in get/set ioctl and MSI addresses to 32 bits.
Both are needed to support x2APIC.

The feature has to be enableable and disabled by default, because
get/set ioctl shifted and truncated APIC ID to 8 bits by using a
non-standard protocol inspired by xAPIC and the change is not
backward-compatible.

Changes to MSI addresses follow the format used by interrupt remapping
unit.  The upper address word, that used to be 0, contains upper 24 bits
of the LAPIC address in its upper 24 bits.  Lower 8 bits are reserved as
0.  Using the upper address word is not backward-compatible either as we
didn't check that userspace zeroed the word.  Reserved bits are still
not explicitly checked, but non-zero data will affect LAPIC addresses,
which will cause a bug.
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

x2APIC supports up to 2^32-1 LAPICs, but most guest in coming years will
probably has fewer VCPUs.  Dynamic size saves memory at the cost of
turning one constant into a variable.

apic_map mutex had to be moved before allocation to avoid races with cpu
hotplug.
Signed-off-by: NRadim Krčmář <rkrcmar@redhat.com>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>