firefly-linux-kernel-4.4.55.git
10 years agoKVM: x86: DR6/7.RTM cannot be written
Nadav Amit [Tue, 15 Jul 2014 14:37:46 +0000 (17:37 +0300)]
KVM: x86: DR6/7.RTM cannot be written

Haswell and newer Intel CPUs have support for RTM, and in that case DR6.RTM is
not fixed to 1 and DR7.RTM is not fixed to zero. That is not the case in the
current KVM implementation. This bug is apparent only if the MOV-DR instruction
is emulated or the host also debugs the guest.

This patch is a partial fix which enables DR6.RTM and DR7.RTM to be cleared and
set respectively. It also sets DR6.RTM upon every debug exception. Obviously,
it is not a complete fix, as debugging of RTM is still unsupported.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: nVMX: clean up nested_release_vmcs12 and code around it
Paolo Bonzini [Thu, 17 Jul 2014 09:55:46 +0000 (11:55 +0200)]
KVM: nVMX: clean up nested_release_vmcs12 and code around it

Make nested_release_vmcs12 idempotent.

Tested-by: Wanpeng Li <wanpeng.li@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: nVMX: fix lifetime issues for vmcs02
Paolo Bonzini [Thu, 17 Jul 2014 10:25:16 +0000 (12:25 +0200)]
KVM: nVMX: fix lifetime issues for vmcs02

free_nested needs the loaded_vmcs to be valid if it is a vmcs02, in
order to detach it from the shadow vmcs.  However, this is not
available anymore after commit 26a865f4aa8e (KVM: VMX: fix use after
free of vmx->loaded_vmcs, 2014-01-03).

Revert that patch, and fix its problem by forcing a vmcs01 as the
active VMCS before freeing all the nested VMX state.

Reported-by: Wanpeng Li <wanpeng.li@linux.intel.com>
Tested-by: Wanpeng Li <wanpeng.li@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Defining missing x86 vectors
Nadav Amit [Mon, 21 Jul 2014 11:37:24 +0000 (14:37 +0300)]
KVM: x86: Defining missing x86 vectors

Defining XE, XM and VE vector numbers.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: emulator injects #DB when RFLAGS.RF is set
Nadav Amit [Wed, 16 Jul 2014 22:19:31 +0000 (01:19 +0300)]
KVM: x86: emulator injects #DB when RFLAGS.RF is set

If the RFLAGS.RF is set, then no #DB should occur on instruction breakpoints.
However, the KVM emulator injects #DB regardless to RFLAGS.RF. This patch fixes
this behavior. KVM, however, still appears not to update RFLAGS.RF correctly,
regardless of this patch.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Cleanup of rflags.rf cleaning
Nadav Amit [Mon, 21 Jul 2014 11:37:30 +0000 (14:37 +0300)]
KVM: x86: Cleanup of rflags.rf cleaning

RFLAGS.RF was cleaned in several functions (e.g., syscall) in the x86 emulator.
Now that we clear it before the execution of an instruction in the emulator, we
can remove the specific cleanup of RFLAGS.RF.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Clear rflags.rf on emulated instructions
Nadav Amit [Mon, 21 Jul 2014 11:37:29 +0000 (14:37 +0300)]
KVM: x86: Clear rflags.rf on emulated instructions

When an instruction is emulated RFLAGS.RF should be cleared. KVM previously did
not do so. This patch clears RFLAGS.RF after interception is done.  If a fault
occurs during the instruction, RFLAGS.RF will be set by a previous patch.  This
patch does not handle the case of traps/interrupts during rep-strings. Traps
are only expected to occur on debug watchpoints, and those are anyhow not
handled by the emulator.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: popf emulation should not change RF
Nadav Amit [Mon, 21 Jul 2014 11:37:28 +0000 (14:37 +0300)]
KVM: x86: popf emulation should not change RF

RFLAGS.RF is always zero after popf. Therefore, popf should not updated RF, as
anyhow emulating popf, just as any other instruction should clear RFLAGS.RF.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Clearing rflags.rf upon skipped emulated instruction
Nadav Amit [Mon, 21 Jul 2014 11:37:26 +0000 (14:37 +0300)]
KVM: x86: Clearing rflags.rf upon skipped emulated instruction

When skipping an emulated instruction, rflags.rf should be cleared as it would
be on real x86 CPU.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoMerge tag 'kvm-s390-20140715' of git://git.kernel.org/pub/scm/linux/kernel/git/kvms39...
Paolo Bonzini [Mon, 21 Jul 2014 11:35:43 +0000 (13:35 +0200)]
Merge tag 'kvm-s390-20140715' of git://git./linux/kernel/git/kvms390/linux into kvm-next

This series enables the "KVM_(S|G)ET_MP_STATE" ioctls on s390 to make
the cpu state settable by user space.

This is necessary to avoid races in s390 SIGP/reset handling which
happen because some SIGPs are handled in QEMU, while others are
handled in the kernel. Together with the busy conditions as return
value of SIGP races happen especially in areas like starting and
stopping of CPUs. (For example, there is a program 'cpuplugd', that
runs on several s390 distros which does automatic onlining and
offlining on cpus.)

As soon as the MPSTATE interface is used, user space takes complete
control of the cpu states. Otherwise the kernel will use the old way.

Therefore, the new kernel continues to work fine with old QEMUs.

10 years agoKVM: nVMX: Fix virtual interrupt delivery injection
Wanpeng Li [Thu, 17 Jul 2014 11:03:00 +0000 (19:03 +0800)]
KVM: nVMX: Fix virtual interrupt delivery injection

This patch fix bug reported in https://bugzilla.kernel.org/show_bug.cgi?id=73331,
after the patch http://www.spinics.net/lists/kvm/msg105230.html applied, there is
some progress and the L2 can boot up, however, slowly. The original idea of this
fix vid injection patch is from "Zhang, Yang Z" <yang.z.zhang@intel.com>.

Interrupt which delivered by vid should be injected to L1 by L0 if current is in
L1, or should be injected to L2 by L0 through the old injection way if L1 doesn't
have set External-interrupt exiting bit. The current logic doen't consider these
cases. This patch fix it by vid intr to L1 if current is L1 or L2 through old
injection way if L1 doen't have External-interrupt exiting bit set.

Signed-off-by: Wanpeng Li <wanpeng.li@linux.intel.com>
Signed-off-by: "Zhang, Yang Z" <yang.z.zhang@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Emulator support for #UD on CPL>0
Nadav Amit [Wed, 18 Jun 2014 14:19:35 +0000 (17:19 +0300)]
KVM: x86: Emulator support for #UD on CPL>0

Certain instructions (e.g., mwait and monitor) cause a #UD exception when they
are executed in user mode. This is in contrast to the regular privileged
instructions which cause #GP. In order not to mess with SVM interception of
mwait and monitor which assumes privilege level assertions take place before
interception, a flag has been added.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Emulator flag for instruction that only support 16-bit addresses in real...
Nadav Amit [Wed, 18 Jun 2014 14:19:34 +0000 (17:19 +0300)]
KVM: x86: Emulator flag for instruction that only support 16-bit addresses in real mode

Certain instructions, such as monitor and xsave do not support big real mode
and cause a #GP exception if any of the accessed bytes effective address are
not within [0, 0xffff].  This patch introduces a flag to mark these
instructions, including the necassary checks.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: use kvm_read_guest_page for emulator accesses
Paolo Bonzini [Tue, 13 May 2014 12:02:13 +0000 (14:02 +0200)]
KVM: x86: use kvm_read_guest_page for emulator accesses

Emulator accesses are always done a page at a time, either by the emulator
itself (for fetches) or because we need to query the MMU for address
translations.  Speed up these accesses by using kvm_read_guest_page
and, in the case of fetches, by inlining kvm_read_guest_virt_helper and
dropping the loop around kvm_read_guest_page.

This final tweak saves 30-100 more clock cycles (4-10%), bringing the
count (as measured by kvm-unit-tests) down to 720-1100 clock cycles on
a Sandy Bridge Xeon host, compared to 2300-3200 before the whole series
and 925-1700 after the first two low-hanging fruit changes.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: ensure emulator fetches do not span multiple pages
Paolo Bonzini [Thu, 19 Jun 2014 09:37:06 +0000 (11:37 +0200)]
KVM: x86: ensure emulator fetches do not span multiple pages

When the CS base is not page-aligned, the linear address of the code could
get close to the page boundary (e.g. 0x...ffe) even if the EIP value is
not.  So we need to first linearize the address, and only then compute
the number of valid bytes that can be fetched.

This happens relatively often when executing real mode code.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: put pointers in the fetch_cache
Paolo Bonzini [Tue, 6 May 2014 14:33:01 +0000 (16:33 +0200)]
KVM: emulate: put pointers in the fetch_cache

This simplifies the code a bit, especially the overflow checks.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: avoid per-byte copying in instruction fetches
Paolo Bonzini [Tue, 6 May 2014 11:05:25 +0000 (13:05 +0200)]
KVM: emulate: avoid per-byte copying in instruction fetches

We do not need a memory copying loop anymore in insn_fetch; we
can use a byte-aligned pointer to access instruction fields directly
from the fetch_cache.  This eliminates 50-150 cycles (corresponding to
a 5-10% improvement in performance) from each instruction.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: avoid repeated calls to do_insn_fetch_bytes
Paolo Bonzini [Tue, 6 May 2014 11:05:25 +0000 (13:05 +0200)]
KVM: emulate: avoid repeated calls to do_insn_fetch_bytes

do_insn_fetch_bytes will only be called once in a given insn_fetch and
insn_fetch_arr, because in fact it will only be called at most twice
for any instruction and the first call is explicit in x86_decode_insn.
This observation lets us hoist the call out of the memory copying loop.
It does not buy performance, because most fetches are one byte long
anyway, but it prepares for the next patch.

The overflow check is tricky, but correct.  Because do_insn_fetch_bytes
has already been called once, we know that fc->end is at least 15.  So
it is okay to subtract the number of bytes we want to read.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: speed up do_insn_fetch
Paolo Bonzini [Tue, 6 May 2014 10:24:32 +0000 (12:24 +0200)]
KVM: emulate: speed up do_insn_fetch

Hoist the common case up from do_insn_fetch_byte to do_insn_fetch,
and prime the fetch_cache in x86_decode_insn.  This helps a bit the
compiler and the branch predictor, but above all it lays the
ground for further changes in the next few patches.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: do not initialize memopp
Bandan Das [Wed, 16 Apr 2014 16:46:14 +0000 (12:46 -0400)]
KVM: emulate: do not initialize memopp

rip_relative is only set if decode_modrm runs, and if you have ModRM
you will also have a memopp.  We can then access memopp unconditionally.
Note that rip_relative cannot be hoisted up to decode_modrm, or you
break "mov $0, xyz(%rip)".

Also, move typecast on "out of range value" of mem.ea to decode_modrm.

Together, all these optimizations save about 50 cycles on each emulated
instructions (4-6%).

Signed-off-by: Bandan Das <bsd@redhat.com>
[Fix immediate operands with rip-relative addressing. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: rework seg_override
Bandan Das [Wed, 16 Apr 2014 16:46:13 +0000 (12:46 -0400)]
KVM: emulate: rework seg_override

x86_decode_insn already sets a default for seg_override,
so remove it from the zeroed area. Also replace set/get functions
with direct access to the field.

Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: clean up initializations in init_decode_cache
Bandan Das [Wed, 16 Apr 2014 16:46:12 +0000 (12:46 -0400)]
KVM: emulate: clean up initializations in init_decode_cache

A lot of initializations are unnecessary as they get set to
appropriate values before actually being used. Optimize
placement of fields in x86_emulate_ctxt

Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: cleanup decode_modrm
Bandan Das [Wed, 16 Apr 2014 16:46:11 +0000 (12:46 -0400)]
KVM: emulate: cleanup decode_modrm

Remove the if conditional - that will help us avoid
an "else initialize to 0" Also, rearrange operators
for slightly better code.

Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: Remove ctxt->intercept and ctxt->check_perm checks
Bandan Das [Wed, 16 Apr 2014 16:46:10 +0000 (12:46 -0400)]
KVM: emulate: Remove ctxt->intercept and ctxt->check_perm checks

The same information can be gleaned from ctxt->d and avoids having
to zero/NULL initialize intercept and check_perm

Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: move init_decode_cache to emulate.c
Bandan Das [Wed, 16 Apr 2014 16:46:09 +0000 (12:46 -0400)]
KVM: emulate: move init_decode_cache to emulate.c

Core emulator functions all belong in emulator.c,
x86 should have no knowledge of emulator internals

Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: simplify writeback
Paolo Bonzini [Tue, 1 Apr 2014 11:23:24 +0000 (13:23 +0200)]
KVM: emulate: simplify writeback

The "if/return" checks are useless, because we return X86EMUL_CONTINUE
anyway if we do not return.

Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: speed up emulated moves
Paolo Bonzini [Thu, 27 Mar 2014 10:36:25 +0000 (11:36 +0100)]
KVM: emulate: speed up emulated moves

We can just blindly move all 16 bytes of ctxt->src's value to ctxt->dst.
write_register_operand will take care of writing only the lower bytes.

Avoiding a call to memcpy (the compiler optimizes it out) gains about
200 cycles on kvm-unit-tests for register-to-register moves, and makes
them about as fast as arithmetic instructions.

We could perhaps get a larger speedup by moving all instructions _except_
moves out of x86_emulate_insn, removing opcode_len, and replacing the
switch statement with an inlined em_mov.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: protect checks on ctxt->d by a common "if (unlikely())"
Paolo Bonzini [Thu, 27 Mar 2014 10:58:02 +0000 (11:58 +0100)]
KVM: emulate: protect checks on ctxt->d by a common "if (unlikely())"

There are several checks for "peculiar" aspects of instructions in both
x86_decode_insn and x86_emulate_insn.  Group them together, and guard
them with a single "if" that lets the processor quickly skip them all.
Make this more effective by adding two more flag bits that say whether the
.intercept and .check_perm fields are valid.  We will reuse these
flags later to avoid initializing fields of the emulate_ctxt struct.

This skims about 30 cycles for each emulated instructions, which is
approximately a 3% improvement.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: move around some checks
Paolo Bonzini [Thu, 27 Mar 2014 11:00:57 +0000 (12:00 +0100)]
KVM: emulate: move around some checks

The only purpose of this patch is to make the next patch simpler
to review.  No semantic change.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: avoid useless set of KVM_REQ_EVENT after emulation
Paolo Bonzini [Thu, 27 Mar 2014 10:29:28 +0000 (11:29 +0100)]
KVM: x86: avoid useless set of KVM_REQ_EVENT after emulation

Despite the provisions to emulate up to 130 consecutive instructions, in
practice KVM will emulate just one before exiting handle_invalid_guest_state,
because x86_emulate_instruction always sets KVM_REQ_EVENT.

However, we only need to do this if an interrupt could be injected,
which happens a) if an interrupt shadow bit (STI or MOV SS) has gone
away; b) if the interrupt flag has just been set (other instructions
than STI can set it without enabling an interrupt shadow).

This cuts another 700-900 cycles from the cost of emulating an
instruction (measured on a Sandy Bridge Xeon: 1650-2600 cycles
before the patch on kvm-unit-tests, 925-1700 afterwards).

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: return all bits from get_interrupt_shadow
Paolo Bonzini [Tue, 20 May 2014 12:29:47 +0000 (14:29 +0200)]
KVM: x86: return all bits from get_interrupt_shadow

For the next patch we will need to know the full state of the
interrupt shadow; we will then set KVM_REQ_EVENT when one bit
is cleared.

However, right now get_interrupt_shadow only returns the one
corresponding to the emulated instruction, or an unconditional
0 if the emulated instruction does not have an interrupt shadow.
This is confusing and does not allow us to check for cleared
bits as mentioned above.

Clean the callback up, and modify toggle_interruptibility to
match the comment above the call.  As a small result, the
call to set_interrupt_shadow will be skipped in the common
case where int_shadow == 0 && mask == 0.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: vmx: speed up emulation of invalid guest state
Paolo Bonzini [Thu, 27 Mar 2014 08:51:52 +0000 (09:51 +0100)]
KVM: vmx: speed up emulation of invalid guest state

About 25% of the time spent in emulation of invalid guest state
is wasted in checking whether emulation is required for the next
instruction.  However, this almost never changes except when a
segment register (or TR or LDTR) changes, or when there is a mode
transition (i.e. CR0 changes).

In fact, vmx_set_segment and vmx_set_cr0 already modify
vmx->emulation_required (except that the former for some reason
uses |= instead of just an assignment).  So there is no need to
call guest_state_valid in the emulation loop.

Emulation performance test results indicate 1650-2600 cycles
for common instructions, versus 2300-3200 before this patch on
a Sandy Bridge Xeon.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: svm: writes to MSR_K7_HWCR generates GPE in guest
Matthias Lange [Thu, 26 Jun 2014 11:50:15 +0000 (13:50 +0200)]
KVM: svm: writes to MSR_K7_HWCR generates GPE in guest

Since commit 575203 the MCE subsystem in the Linux kernel for AMD sets bit 18
in MSR_K7_HWCR. Running such a kernel as a guest in KVM on an AMD host results
in a GPE injected into the guest because kvm_set_msr_common returns 1. This
patch fixes this by masking bit 18 from the MSR value desired by the guest.

Signed-off-by: Matthias Lange <matthias.lange@kernkonzept.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Pending interrupt may be delivered after INIT
Nadav Amit [Mon, 30 Jun 2014 09:03:02 +0000 (12:03 +0300)]
KVM: x86: Pending interrupt may be delivered after INIT

We encountered a scenario in which after an INIT is delivered, a pending
interrupt is delivered, although it was sent before the INIT.  As the SDM
states in section 10.4.7.1, the ISR and the IRR should be cleared after INIT as
KVM does.  This also means that pending interrupts should be cleared.  This
patch clears upon reset (and INIT) the pending interrupts; and at the same
occassion clears the pending exceptions, since they may cause a similar issue.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: Synthesize G bit for all segments.
Jim Mattson [Tue, 8 Jul 2014 04:17:41 +0000 (09:47 +0530)]
KVM: Synthesize G bit for all segments.

We have noticed that qemu-kvm hangs early in the BIOS when runnning nested
under some versions of VMware ESXi.

The problem we believe is because KVM assumes that the platform preserves
the 'G' but for any segment register. The SVM specification itemizes the
segment attribute bits that are observed by the CPU, but the (G)ranularity bit
is not one of the bits itemized, for any segment. Though current AMD CPUs keep
track of the (G)ranularity bit for all segment registers other than CS, the
specification does not require it. VMware's virtual CPU may not track the
(G)ranularity bit for any segment register.

Since kvm already synthesizes the (G)ranularity bit for the CS segment. It
should do so for all segments. The patch below does that, and helps get rid of
the hangs. Patch applies on top of Linus' tree.

Signed-off-by: Jim Mattson <jmattson@vmware.com>
Signed-off-by: Alok N Kataria <akataria@vmware.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: s390: implement KVM_(S|G)ET_MP_STATE for user space state control
David Hildenbrand [Thu, 10 Apr 2014 15:35:00 +0000 (17:35 +0200)]
KVM: s390: implement KVM_(S|G)ET_MP_STATE for user space state control

This patch
- adds s390 specific MP states to linux headers and documents them
- implements the KVM_{SET,GET}_MP_STATE ioctls
- enables KVM_CAP_MP_STATE
- allows user space to control the VCPU state on s390.

If user space sets the VCPU state using the ioctl KVM_SET_MP_STATE, we can disable
manual changing of the VCPU state and trust user space to do the right thing.

Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com>
Reviewed-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Acked-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
10 years agoKVM: prepare for KVM_(S|G)ET_MP_STATE on other architectures
David Hildenbrand [Mon, 12 May 2014 14:05:13 +0000 (16:05 +0200)]
KVM: prepare for KVM_(S|G)ET_MP_STATE on other architectures

Highlight the aspects of the ioctls that are actually specific to x86
and ia64. As defined restrictions (irqchip) and mp states may not apply
to other architectures, these parts are flagged to belong to x86 and ia64.

In preparation for the use of KVM_(S|G)ET_MP_STATE by s390.
Fix a spelling error (KVM_SET_MP_STATE vs. KVM_SET_MPSTATE) on the way.

Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com>
Reviewed-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Acked-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
10 years agoKVM: s390: remove __cpu_is_stopped and expose is_vcpu_stopped
David Hildenbrand [Mon, 5 May 2014 14:26:19 +0000 (16:26 +0200)]
KVM: s390: remove __cpu_is_stopped and expose is_vcpu_stopped

The function "__cpu_is_stopped" is not used any more. Let's remove it and
expose the function "is_vcpu_stopped" instead, which is actually what we want.

This patch also converts an open coded check for CPUSTAT_STOPPED to
is_vcpu_stopped().

Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com>
Reviewed-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Acked-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
10 years agoKVM: s390: move finalization of SIGP STOP orders to kvm_s390_vcpu_stop
David Hildenbrand [Mon, 14 Apr 2014 10:40:03 +0000 (12:40 +0200)]
KVM: s390: move finalization of SIGP STOP orders to kvm_s390_vcpu_stop

Let's move the finalization of SIGP STOP and SIGP STOP AND STORE STATUS orders to
the point where the VCPU is actually stopped.

This change is needed to prepare for a user space driven VCPU state change. The
action_bits may only be cleared when setting the cpu state to STOPPED while
holding the local irq lock.

Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com>
Reviewed-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Reviewed-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
10 years agoKVM: s390: allow only one SIGP STOP (AND STORE STATUS) at a time
David Hildenbrand [Thu, 15 May 2014 12:25:25 +0000 (14:25 +0200)]
KVM: s390: allow only one SIGP STOP (AND STORE STATUS) at a time

A SIGP STOP (AND STORE STATUS) order is complete as soon as the VCPU has been
stopped. This patch makes sure that only one SIGP STOP (AND STORE STATUS) may
be pending at a time (as defined by the architecture). If the action_bits are
still set, a SIGP STOP has been issued but not completed yet. The VCPU is busy
for further SIGP STOP orders.

Also set the CPUSTAT_STOP_INT after the action_bits variable has been modified
(the same order that is used when injecting a KVM_S390_SIGP_STOP from
userspace).

Both changes are needed in preparation for a user space driven VCPU state change
(to avoid race conditions).

Signed-off-by: David Hildenbrand <dahi@linux.vnet.ibm.com>
Reviewed-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Reviewed-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
10 years agoKVM: MIPS: Document MIPS specifics of KVM API.
James Hogan [Fri, 4 Jul 2014 14:11:35 +0000 (15:11 +0100)]
KVM: MIPS: Document MIPS specifics of KVM API.

Document the MIPS specific parts of the KVM API, including:
 - The layout of the kvm_regs structure.
 - The interrupt number passed to KVM_INTERRUPT.
 - The registers supported by the KVM_{GET,SET}_ONE_REG interface, and
   the encoding of those register ids.
 - That KVM_INTERRUPT and KVM_GET_REG_LIST are supported on MIPS.

Signed-off-by: James Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: linux-doc@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: Reformat KVM_SET_ONE_REG register documentation
James Hogan [Fri, 4 Jul 2014 14:11:34 +0000 (15:11 +0100)]
KVM: Reformat KVM_SET_ONE_REG register documentation

Some of the MIPS registers that can be accessed with the
KVM_{GET,SET}_ONE_REG interface have fairly long names, so widen the
Register column of the table in the KVM_SET_ONE_REG documentation to
allow them to fit.

Tabs in the table are replaced with spaces at the same time for
consistency.

Signed-off-by: James Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: linux-doc@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: Document KVM_SET_SIGNAL_MASK as universal
James Hogan [Fri, 4 Jul 2014 14:11:33 +0000 (15:11 +0100)]
KVM: Document KVM_SET_SIGNAL_MASK as universal

KVM_SET_SIGNAL_MASK is implemented in generic code and isn't x86
specific, so document it as being applicable for all architectures.

Signed-off-by: James Hogan <james.hogan@imgtec.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Gleb Natapov <gleb@kernel.org>
Cc: kvm@vger.kernel.org
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: linux-doc@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Fix lapic.c debug prints
Nadav Amit [Sun, 29 Jun 2014 09:28:51 +0000 (12:28 +0300)]
KVM: x86: Fix lapic.c debug prints

In two cases lapic.c does not use the apic_debug macro correctly. This patch
fixes them.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: fix TSC matching
Tomasz Grabiec [Tue, 24 Jun 2014 07:42:43 +0000 (09:42 +0200)]
KVM: x86: fix TSC matching

I've observed kvmclock being marked as unstable on a modern
single-socket system with a stable TSC and qemu-1.6.2 or qemu-2.0.0.

The culprit was failure in TSC matching because of overflow of
kvm_arch::nr_vcpus_matched_tsc in case there were multiple TSC writes
in a single synchronization cycle.

Turns out that qemu does multiple TSC writes during init, below is the
evidence of that (qemu-2.0.0):

The first one:

 0xffffffffa08ff2b4 : vmx_write_tsc_offset+0xa4/0xb0 [kvm_intel]
 0xffffffffa04c9c05 : kvm_write_tsc+0x1a5/0x360 [kvm]
 0xffffffffa04cfd6b : kvm_arch_vcpu_postcreate+0x4b/0x80 [kvm]
 0xffffffffa04b8188 : kvm_vm_ioctl+0x418/0x750 [kvm]

The second one:

 0xffffffffa08ff2b4 : vmx_write_tsc_offset+0xa4/0xb0 [kvm_intel]
 0xffffffffa04c9c05 : kvm_write_tsc+0x1a5/0x360 [kvm]
 0xffffffffa090610d : vmx_set_msr+0x29d/0x350 [kvm_intel]
 0xffffffffa04be83b : do_set_msr+0x3b/0x60 [kvm]
 0xffffffffa04c10a8 : msr_io+0xc8/0x160 [kvm]
 0xffffffffa04caeb6 : kvm_arch_vcpu_ioctl+0xc86/0x1060 [kvm]
 0xffffffffa04b6797 : kvm_vcpu_ioctl+0xc7/0x5a0 [kvm]

 #0  kvm_vcpu_ioctl at /build/buildd/qemu-2.0.0+dfsg/kvm-all.c:1780
 #1  kvm_put_msrs at /build/buildd/qemu-2.0.0+dfsg/target-i386/kvm.c:1270
 #2  kvm_arch_put_registers at /build/buildd/qemu-2.0.0+dfsg/target-i386/kvm.c:1909
 #3  kvm_cpu_synchronize_post_init at /build/buildd/qemu-2.0.0+dfsg/kvm-all.c:1641
 #4  cpu_synchronize_post_init at /build/buildd/qemu-2.0.0+dfsg/include/sysemu/kvm.h:330
 #5  cpu_synchronize_all_post_init () at /build/buildd/qemu-2.0.0+dfsg/cpus.c:521
 #6  main at /build/buildd/qemu-2.0.0+dfsg/vl.c:4390

The third one:

 0xffffffffa08ff2b4 : vmx_write_tsc_offset+0xa4/0xb0 [kvm_intel]
 0xffffffffa04c9c05 : kvm_write_tsc+0x1a5/0x360 [kvm]
 0xffffffffa090610d : vmx_set_msr+0x29d/0x350 [kvm_intel]
 0xffffffffa04be83b : do_set_msr+0x3b/0x60 [kvm]
 0xffffffffa04c10a8 : msr_io+0xc8/0x160 [kvm]
 0xffffffffa04caeb6 : kvm_arch_vcpu_ioctl+0xc86/0x1060 [kvm]
 0xffffffffa04b6797 : kvm_vcpu_ioctl+0xc7/0x5a0 [kvm]

 #0  kvm_vcpu_ioctl at /build/buildd/qemu-2.0.0+dfsg/kvm-all.c:1780
 #1  kvm_put_msrs  at /build/buildd/qemu-2.0.0+dfsg/target-i386/kvm.c:1270
 #2  kvm_arch_put_registers  at /build/buildd/qemu-2.0.0+dfsg/target-i386/kvm.c:1909
 #3  kvm_cpu_synchronize_post_reset  at /build/buildd/qemu-2.0.0+dfsg/kvm-all.c:1635
 #4  cpu_synchronize_post_reset  at /build/buildd/qemu-2.0.0+dfsg/include/sysemu/kvm.h:323
 #5  cpu_synchronize_all_post_reset () at /build/buildd/qemu-2.0.0+dfsg/cpus.c:512
 #6  main  at /build/buildd/qemu-2.0.0+dfsg/vl.c:4482

The fix is to count each vCPU only once when matched, so that
nr_vcpus_matched_tsc holds the size of the matched set. This is
achieved by reusing generation counters. Every vCPU with
this_tsc_generation == cur_tsc_generation is in the matched set. The
match set is cleared by setting cur_tsc_generation to a value which no
other vCPU is set to (by incrementing it).

I needed to bump up the counter size form u8 to u64 to ensure it never
overflows. Otherwise in cases TSC is not written the same number of
times on each vCPU the counter could overflow and incorrectly indicate
some vCPUs as being in the matched set. This scenario seems unlikely
but I'm not sure if it can be disregarded.

Signed-off-by: Tomasz Grabiec <tgrabiec@cloudius-systems.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: nSVM: Set correct port for IOIO interception evaluation
Jan Kiszka [Mon, 30 Jun 2014 10:52:55 +0000 (12:52 +0200)]
KVM: nSVM: Set correct port for IOIO interception evaluation

Obtaining the port number from DX is bogus as a) there are immediate
port accesses and b) user space may have changed the register content
while processing the PIO access. Forward the correct value from the
instruction emulator instead.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: nSVM: Fix IOIO size reported on emulation
Jan Kiszka [Mon, 30 Jun 2014 09:07:05 +0000 (11:07 +0200)]
KVM: nSVM: Fix IOIO size reported on emulation

The access size of an in/ins is reported in dst_bytes, and that of
out/outs in src_bytes.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: nSVM: Fix IOIO bitmap evaluation
Jan Kiszka [Mon, 30 Jun 2014 08:54:17 +0000 (10:54 +0200)]
KVM: nSVM: Fix IOIO bitmap evaluation

First, kvm_read_guest returns 0 on success. And then we need to take the
access size into account when testing the bitmap: intercept if any of
bits corresponding to the access is set.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: nSVM: Do not report CLTS via SVM_EXIT_WRITE_CR0 to L1
Jan Kiszka [Sun, 29 Jun 2014 19:55:53 +0000 (21:55 +0200)]
KVM: nSVM: Do not report CLTS via SVM_EXIT_WRITE_CR0 to L1

CLTS only changes TS which is not monitored by selected CR0
interception. So skip any attempt to translate WRITE_CR0 to
CR0_SEL_WRITE for this instruction.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoarch: x86: kvm: x86.c: Cleaning up variable is set more than once
Rickard Strandqvist [Wed, 25 Jun 2014 12:25:58 +0000 (14:25 +0200)]
arch: x86: kvm: x86.c: Cleaning up variable is set more than once

A struct member variable is set to the same value more than once

This was found using a static code analysis program called cppcheck.

Signed-off-by: Rickard Strandqvist <rickard_strandqvist@spectrumdigital.se>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoMIPS: KVM: Remove dead code of TLB index error in kvm_mips_emul_tlbwr()
Deng-Cheng Zhu [Thu, 26 Jun 2014 19:11:40 +0000 (12:11 -0700)]
MIPS: KVM: Remove dead code of TLB index error in kvm_mips_emul_tlbwr()

It's impossible to fall into the error handling of the TLB index after
being masked by (KVM_MIPS_GUEST_TLB_SIZE - 1). Remove the dead code.

Reported-by: James Hogan <james.hogan@imgtec.com>
Signed-off-by: Deng-Cheng Zhu <dengcheng.zhu@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoMIPS: KVM: Skip memory cleaning in kvm_mips_commpage_init()
Deng-Cheng Zhu [Thu, 26 Jun 2014 19:11:39 +0000 (12:11 -0700)]
MIPS: KVM: Skip memory cleaning in kvm_mips_commpage_init()

The commpage is allocated using kzalloc(), so there's no need of cleaning
the memory of the kvm_mips_commpage struct and its internal mips_coproc.

Reviewed-by: James Hogan <james.hogan@imgtec.com>
Signed-off-by: Deng-Cheng Zhu <dengcheng.zhu@imgtec.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoMIPS: KVM: Rename files to remove the prefix "kvm_" and "kvm_mips_"
Deng-Cheng Zhu [Thu, 26 Jun 2014 19:11:38 +0000 (12:11 -0700)]
MIPS: KVM: Rename files to remove the prefix "kvm_" and "kvm_mips_"

Since all the files are in arch/mips/kvm/, there's no need of the prefixes
"kvm_" and "kvm_mips_".

Reviewed-by: James Hogan <james.hogan@imgtec.com>
Signed-off-by: Deng-Cheng Zhu <dengcheng.zhu@imgtec.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoMIPS: KVM: Remove unneeded volatile
Deng-Cheng Zhu [Thu, 26 Jun 2014 19:11:37 +0000 (12:11 -0700)]
MIPS: KVM: Remove unneeded volatile

The keyword volatile for idx in the TLB functions is unnecessary.

Reviewed-by: James Hogan <james.hogan@imgtec.com>
Signed-off-by: Deng-Cheng Zhu <dengcheng.zhu@imgtec.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoMIPS: KVM: Simplify functions by removing redundancy
Deng-Cheng Zhu [Thu, 26 Jun 2014 19:11:36 +0000 (12:11 -0700)]
MIPS: KVM: Simplify functions by removing redundancy

No logic changes inside.

Reviewed-by: James Hogan <james.hogan@imgtec.com>
Signed-off-by: Deng-Cheng Zhu <dengcheng.zhu@imgtec.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoMIPS: KVM: Use KVM internal logger
Deng-Cheng Zhu [Thu, 26 Jun 2014 19:11:35 +0000 (12:11 -0700)]
MIPS: KVM: Use KVM internal logger

Replace printks with kvm_[err|info|debug].

Signed-off-by: Deng-Cheng Zhu <dengcheng.zhu@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoMIPS: KVM: Reformat code and comments
Deng-Cheng Zhu [Thu, 26 Jun 2014 19:11:34 +0000 (12:11 -0700)]
MIPS: KVM: Reformat code and comments

No logic changes inside.

Signed-off-by: Deng-Cheng Zhu <dengcheng.zhu@imgtec.com>
Reviewed-by: James Hogan <james.hogan@imgtec.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoMerge tag 'kvms390-20140626' of git://git.kernel.org/pub/scm/linux/kernel/git/kvms390...
Paolo Bonzini [Mon, 30 Jun 2014 14:51:47 +0000 (16:51 +0200)]
Merge tag 'kvms390-20140626' of git://git./linux/kernel/git/kvms390/linux into HEAD

Fix sie.h header related problems introduced during the 3.16 development
cycle.

10 years agoMerge commit '33b458d276bb' into kvm-next
Paolo Bonzini [Mon, 30 Jun 2014 14:51:07 +0000 (16:51 +0200)]
Merge commit '33b458d276bb' into kvm-next

Fix bad x86 regression introduced during merge window.

10 years agoKVM: SVM: Fix CPL export via SS.DPL
Jan Kiszka [Sun, 29 Jun 2014 15:12:43 +0000 (17:12 +0200)]
KVM: SVM: Fix CPL export via SS.DPL

We import the CPL via SS.DPL since ae9fedc793. However, we fail to
export it this way so far. This caused spurious guest crashes, e.g. of
Linux when accessing the vmport from guest user space which triggered
register saving/restoring to/from host user space.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: s390: add sie.h uapi header file to Kbuild and remove header dependency
Heiko Carstens [Thu, 5 Jun 2014 11:22:49 +0000 (13:22 +0200)]
KVM: s390: add sie.h uapi header file to Kbuild and remove header dependency

sie.h was missing in arch/s390/include/uapi/asm/Kbuild and therefore missed
the "make headers_check" target.
If added it reveals that also arch/s390/include/asm/sigp.h would become uapi.
This is something we certainly do not want. So remove that dependency as well.

The header file was merged with ceae283bb2e0176c "KVM: s390: add sie exit
reasons tables", therefore we never had a kernel release with this commit and
can still change anything.

Acked-by: Alexander Yarygin <yarygin@linux.vnet.ibm.com>
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Cornelia Huck <cornelia.huck@de.ibm.com>
10 years agoKVM: vmx: vmx instructions handling does not consider cs.l
Nadav Amit [Wed, 18 Jun 2014 14:19:26 +0000 (17:19 +0300)]
KVM: vmx: vmx instructions handling does not consider cs.l

VMX instructions use 32-bit operands in 32-bit mode, and 64-bit operands in
64-bit mode.  The current implementation is broken since it does not use the
register operands correctly, and always uses 64-bit for reads and writes.
Moreover, write to memory in vmwrite only considers long-mode, so it ignores
cs.l. This patch fixes this behavior.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: vmx: handle_cr ignores 32/64-bit mode
Nadav Amit [Wed, 18 Jun 2014 14:19:25 +0000 (17:19 +0300)]
KVM: vmx: handle_cr ignores 32/64-bit mode

On 32-bit mode only bits [31:0] of the CR should be used for setting the CR
value.  Otherwise, the host may incorrectly assume the value is invalid if bits
[63:32] are not zero.  Moreover, the CR is currently being read twice when CR8
is used.  Last, nested mov-cr exiting is modified to handle the CR value
correctly as well.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Hypercall handling does not considers opsize correctly
Nadav Amit [Wed, 18 Jun 2014 14:19:24 +0000 (17:19 +0300)]
KVM: x86: Hypercall handling does not considers opsize correctly

Currently, the hypercall handling routine only considers LME as an indication
to whether the guest uses 32/64-bit mode. This is incosistent with hyperv
hypercalls handling and against the common sense of considering cs.l as well.
This patch uses is_64_bit_mode instead of is_long_mode for that matter. In
addition, the result is masked in respect to the guest execution mode. Last, it
changes kvm_hv_hypercall to use is_64_bit_mode as well to simplify the code.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: check DR6/7 high-bits are clear only on long-mode
Nadav Amit [Wed, 18 Jun 2014 14:19:23 +0000 (17:19 +0300)]
KVM: x86: check DR6/7 high-bits are clear only on long-mode

When the guest sets DR6 and DR7, KVM asserts the high 32-bits are clear, and
otherwise injects a #GP exception. This exception should only be injected only
if running in long-mode.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: nVMX: Fix returned value of MSR_IA32_VMX_VMCS_ENUM
Jan Kiszka [Mon, 16 Jun 2014 11:59:44 +0000 (13:59 +0200)]
KVM: nVMX: Fix returned value of MSR_IA32_VMX_VMCS_ENUM

Many real CPUs get this wrong as well, but ours is totally off: bits 9:1
define the highest index value.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: nVMX: Allow to disable VM_{ENTRY_LOAD,EXIT_SAVE}_DEBUG_CONTROLS
Jan Kiszka [Mon, 16 Jun 2014 11:59:43 +0000 (13:59 +0200)]
KVM: nVMX: Allow to disable VM_{ENTRY_LOAD,EXIT_SAVE}_DEBUG_CONTROLS

Allow L1 to "leak" its debug controls into L2, i.e. permit cleared
VM_{ENTRY_LOAD,EXIT_SAVE}_DEBUG_CONTROLS. This requires to manually
transfer the state of DR7 and IA32_DEBUGCTLMSR from L1 into L2 as both
run on different VMCS.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: nVMX: Fix returned value of MSR_IA32_VMX_PROCBASED_CTLS
Jan Kiszka [Mon, 16 Jun 2014 11:59:42 +0000 (13:59 +0200)]
KVM: nVMX: Fix returned value of MSR_IA32_VMX_PROCBASED_CTLS

SDM says bits 1, 4-6, 8, 13-16, and 26 have to be set.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: nVMX: Allow to disable CR3 access interception
Jan Kiszka [Mon, 16 Jun 2014 11:59:41 +0000 (13:59 +0200)]
KVM: nVMX: Allow to disable CR3 access interception

We already have this control enabled by exposing a broken
MSR_IA32_VMX_PROCBASED_CTLS value. This will properly advertise our
capability once the value is fixed by clearing the right bits in
MSR_IA32_VMX_TRUE_PROCBASED_CTLS. We also have to ensure to test the
right value on L2 entry.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: nVMX: Advertise support for MSR_IA32_VMX_TRUE_*_CTLS
Jan Kiszka [Mon, 16 Jun 2014 11:59:40 +0000 (13:59 +0200)]
KVM: nVMX: Advertise support for MSR_IA32_VMX_TRUE_*_CTLS

We already implemented them but failed to advertise them. Currently they
all return the identical values to the capability MSRs they are
augmenting. So there is no change in exposed features yet.

Drop related comments at this chance that are partially incorrect and
redundant anyway.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Fix constant value of VM_{EXIT_SAVE,ENTRY_LOAD}_DEBUG_CONTROLS
Jan Kiszka [Thu, 12 Jun 2014 17:40:32 +0000 (19:40 +0200)]
KVM: x86: Fix constant value of VM_{EXIT_SAVE,ENTRY_LOAD}_DEBUG_CONTROLS

The spec says those controls are at bit position 2 - makes 4 as value.

The impact of this mistake is effectively zero as we only use them to
ensure that these features are set at position 2 (or, previously, 1) in
MSR_IA32_VMX_{EXIT,ENTRY}_CTLS - which is and will be always true
according to the spec.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: NOP emulation clears (incorrectly) the high 32-bits of RAX
Nadav Amit [Sun, 15 Jun 2014 13:13:01 +0000 (16:13 +0300)]
KVM: x86: NOP emulation clears (incorrectly) the high 32-bits of RAX

On long-mode the current NOP (0x90) emulation still writes back to RAX.  As a
result, EAX is zero-extended and the high 32-bits of RAX are cleared.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: emulation of dword cmov on long-mode should clear [63:32]
Nadav Amit [Sun, 15 Jun 2014 13:13:00 +0000 (16:13 +0300)]
KVM: x86: emulation of dword cmov on long-mode should clear [63:32]

Even if the condition of cmov is not satisfied, bits[63:32] should be cleared.
This is clearly stated in Intel's CMOVcc documentation.  The solution is to
reassign the destination onto itself if the condition is unsatisfied.  For that
matter the original destination value needs to be read.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Inter-privilege level ret emulation is not implemeneted
Nadav Amit [Sun, 15 Jun 2014 13:12:59 +0000 (16:12 +0300)]
KVM: x86: Inter-privilege level ret emulation is not implemeneted

Return unhandlable error on inter-privilege level ret instruction.  This is
since the current emulation does not check the privilege level correctly when
loading the CS, and does not pop RSP/SS as needed.

Cc: stable@vger.kernel.org
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Wrong emulation on 'xadd X, X'
Nadav Amit [Sun, 15 Jun 2014 13:12:58 +0000 (16:12 +0300)]
KVM: x86: Wrong emulation on 'xadd X, X'

The emulator does not emulate the xadd instruction correctly if the two
operands are the same.  In this (unlikely) situation the result should be the
sum of X and X (2X) when it is currently X.  The solution is to first perform
writeback to the source, before writing to the destination.  The only
instruction which should be affected is xadd, as the other instructions that
perform writeback to the source use the extended accumlator (e.g., RAX:RDX).

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: bit-ops emulation ignores offset on 64-bit
Nadav Amit [Sun, 15 Jun 2014 13:12:57 +0000 (16:12 +0300)]
KVM: x86: bit-ops emulation ignores offset on 64-bit

The current emulation of bit operations ignores the offset from the destination
on 64-bit target memory operands. This patch fixes this behavior.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoarch/x86/kvm/vmx.c: use PAGE_ALIGNED instead of IS_ALIGNED(PAGE_SIZE
Fabian Frederick [Sat, 14 Jun 2014 21:44:29 +0000 (23:44 +0200)]
arch/x86/kvm/vmx.c: use PAGE_ALIGNED instead of IS_ALIGNED(PAGE_SIZE

use mm.h definition

Cc: Gleb Natapov <gleb@kernel.org>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: fix harmless typo in MMX decoding
Paolo Bonzini [Tue, 6 May 2014 12:03:29 +0000 (14:03 +0200)]
KVM: emulate: fix harmless typo in MMX decoding

It was using the wrong member of the union.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: simplify BitOp handling
Paolo Bonzini [Tue, 1 Apr 2014 12:54:19 +0000 (14:54 +0200)]
KVM: emulate: simplify BitOp handling

Memory is always the destination for BitOp instructions.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: emulate: POP SS triggers a MOV SS shadow too
Paolo Bonzini [Thu, 5 Jun 2014 15:29:34 +0000 (17:29 +0200)]
KVM: emulate: POP SS triggers a MOV SS shadow too

We did not do that when interruptibility was added to the emulator,
because at the time pop to segment was not implemented.  Now it is,
add it.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: smsw emulation is incorrect in 64-bit mode
Nadav Amit [Mon, 2 Jun 2014 15:34:11 +0000 (18:34 +0300)]
KVM: x86: smsw emulation is incorrect in 64-bit mode

In 64-bit mode, when the destination is a register, the assignment is done
according to the operand size. Otherwise (memory operand or no 64-bit mode), a
16-bit assignment is performed.

Currently, 16-bit assignment is always done to the destination.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Return error on cmpxchg16b emulation
Nadav Amit [Mon, 2 Jun 2014 15:34:10 +0000 (18:34 +0300)]
KVM: x86: Return error on cmpxchg16b emulation

cmpxchg16b is currently unimplemented in the emulator. The least we can do is
return error upon the emulation of this instruction.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: rdpmc emulation checks the counter incorrectly
Nadav Amit [Mon, 2 Jun 2014 15:34:09 +0000 (18:34 +0300)]
KVM: x86: rdpmc emulation checks the counter incorrectly

The rdpmc emulation checks that the counter (ECX) is not higher than 2, without
taking into considerations bits 30:31 role (e.g., bit 30 marks whether the
counter is fixed). The fix uses the pmu information for checking the validity
of the pmu counter.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: movnti minimum op size of 32-bit is not kept
Nadav Amit [Mon, 2 Jun 2014 15:34:08 +0000 (18:34 +0300)]
KVM: x86: movnti minimum op size of 32-bit is not kept

If the operand-size prefix (0x66) is used in 64-bit mode, the emulator would
assume the destination operand is 64-bit, when it should be 32-bit.

Reminder: movnti does not support 16-bit operands and its default operand size
is 32-bit.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: cmpxchg emulation should compare in reverse order
Nadav Amit [Mon, 2 Jun 2014 15:34:07 +0000 (18:34 +0300)]
KVM: x86: cmpxchg emulation should compare in reverse order

The current implementation of cmpxchg does not update the flags correctly,
since the accumulator should be compared with the destination and not the other
way around. The current implementation does not update the flags correctly.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: sgdt and sidt are not privilaged
Nadav Amit [Mon, 2 Jun 2014 15:34:06 +0000 (18:34 +0300)]
KVM: x86: sgdt and sidt are not privilaged

The SGDT and SIDT instructions are not privilaged, i.e. they can be executed
with CPL>0.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Loading segments on 64-bit mode may be wrong
Nadav Amit [Mon, 2 Jun 2014 15:34:05 +0000 (18:34 +0300)]
KVM: x86: Loading segments on 64-bit mode may be wrong

The current emulator implementation ignores the high 32 bits of the base in
long-mode.  During segment load from the LDT, the base of the LDT is calculated
incorrectly and may cause the wrong segment to be loaded.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Emulator ignores LDTR/TR extended base on LLDT/LTR
Nadav Amit [Mon, 2 Jun 2014 15:34:04 +0000 (18:34 +0300)]
KVM: x86: Emulator ignores LDTR/TR extended base on LLDT/LTR

The current implementation ignores the LDTR/TR base high 32-bits on long-mode.
As a result the loaded segment descriptor may be incorrect.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoKVM: x86: Mark VEX-prefix instructions emulation as unimplemented
Nadav Amit [Mon, 2 Jun 2014 15:34:03 +0000 (18:34 +0300)]
KVM: x86: Mark VEX-prefix instructions emulation as unimplemented

Currently the emulator does not recognize vex-prefix instructions.  However, it
may incorrectly decode lgdt/lidt instructions and try to execute them. This
patch returns unhandlable error on their emulation.

Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10 years agoLinux 3.16-rc1
Linus Torvalds [Mon, 16 Jun 2014 03:45:28 +0000 (17:45 -1000)]
Linux 3.16-rc1

10 years agoMerge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
Linus Torvalds [Mon, 16 Jun 2014 02:37:03 +0000 (16:37 -1000)]
Merge git://git./linux/kernel/git/davem/net

Pull networking fixes from David Miller:

 1) Fix checksumming regressions, from Tom Herbert.

 2) Undo unintentional permissions changes for SCTP rto_alpha and
    rto_beta sysfs knobs, from Denial Borkmann.

 3) VXLAN, like other IP tunnels, should advertize it's encapsulation
    size using dev->needed_headroom instead of dev->hard_header_len.
    From Cong Wang.

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net:
  net: sctp: fix permissions for rto_alpha and rto_beta knobs
  vxlan: Checksum fixes
  net: add skb_pop_rcv_encapsulation
  udp: call __skb_checksum_complete when doing full checksum
  net: Fix save software checksum complete
  net: Fix GSO constants to match NETIF flags
  udp: ipv4: do not waste time in __udp4_lib_mcast_demux_lookup
  vxlan: use dev->needed_headroom instead of dev->hard_header_len
  MAINTAINERS: update cxgb4 maintainer

10 years agoMerge tag 'clk-for-linus-3.16-part2' of git://git.linaro.org/people/mike.turquette...
Linus Torvalds [Mon, 16 Jun 2014 02:02:20 +0000 (16:02 -1000)]
Merge tag 'clk-for-linus-3.16-part2' of git://git.linaro.org/people/mike.turquette/linux

Pull more clock framework updates from Mike Turquette:
 "This contains the second half the of the clk changes for 3.16.

  They are simply fixes and code refactoring for the OMAP clock drivers.
  The sunxi clock driver changes include splitting out the one
  mega-driver into several smaller pieces and adding support for the A31
  SoC clocks"

* tag 'clk-for-linus-3.16-part2' of git://git.linaro.org/people/mike.turquette/linux: (25 commits)
  clk: sunxi: document PRCM clock compatible strings
  clk: sunxi: add PRCM (Power/Reset/Clock Management) clks support
  clk: sun6i: Protect SDRAM gating bit
  clk: sun6i: Protect CPU clock
  clk: sunxi: Rework clock protection code
  clk: sunxi: Move the GMAC clock to a file of its own
  clk: sunxi: Move the 24M oscillator to a file of its own
  clk: sunxi: Remove calls to clk_put
  clk: sunxi: document new A31 USB clock compatible
  clk: sunxi: Implement A31 USB clock
  ARM: dts: OMAP5/DRA7: use omap5-mpu-dpll-clock capable of dealing with higher frequencies
  CLK: TI: dpll: support OMAP5 MPU DPLL that need special handling for higher frequencies
  ARM: OMAP5+: dpll: support Duty Cycle Correction(DCC)
  CLK: TI: clk-54xx: Set the rate for dpll_abe_m2x2_ck
  CLK: TI: Driver for DRA7 ATL (Audio Tracking Logic)
  dt:/bindings: DRA7 ATL (Audio Tracking Logic) clock bindings
  ARM: dts: dra7xx-clocks: Correct name for atl clkin3 clock
  CLK: TI: gate: add composite interface clock to OMAP2 only build
  ARM: OMAP2: clock: add DT boot support for cpufreq_ck
  CLK: TI: OMAP2: add clock init support
  ...

10 years agoMerge git://git.infradead.org/users/willy/linux-nvme
Linus Torvalds [Mon, 16 Jun 2014 01:58:03 +0000 (15:58 -1000)]
Merge git://git.infradead.org/users/willy/linux-nvme

Pull NVMe update from Matthew Wilcox:
 "Mostly bugfixes again for the NVMe driver.  I'd like to call out the
  exported tracepoint in the block layer; I believe Keith has cleared
  this with Jens.

  We've had a few reports from people who're really pounding on NVMe
  devices at scale, hence the timeout changes (and new module
  parameters), hotplug cpu deadlock, tracepoints, and minor performance
  tweaks"

[ Jens hadn't seen that tracepoint thing, but is ok with it - it will
  end up going away when mq conversion happens ]

* git://git.infradead.org/users/willy/linux-nvme: (22 commits)
  NVMe: Fix START_STOP_UNIT Scsi->NVMe translation.
  NVMe: Use Log Page constants in SCSI emulation
  NVMe: Define Log Page constants
  NVMe: Fix hot cpu notification dead lock
  NVMe: Rename io_timeout to nvme_io_timeout
  NVMe: Use last bytes of f/w rev SCSI Inquiry
  NVMe: Adhere to request queue block accounting enable/disable
  NVMe: Fix nvme get/put queue semantics
  NVMe: Delete NVME_GET_FEAT_TEMP_THRESH
  NVMe: Make admin timeout a module parameter
  NVMe: Make iod bio timeout a parameter
  NVMe: Prevent possible NULL pointer dereference
  NVMe: Fix the buffer size passed in GetLogPage(CDW10.NUMD)
  NVMe: Update data structures for NVMe 1.2
  NVMe: Enable BUILD_BUG_ON checks
  NVMe: Update namespace and controller identify structures to the 1.1a spec
  NVMe: Flush with data support
  NVMe: Configure support for block flush
  NVMe: Add tracepoints
  NVMe: Protect against badly formatted CQEs
  ...

10 years agonet: sctp: fix permissions for rto_alpha and rto_beta knobs
Daniel Borkmann [Sat, 14 Jun 2014 22:59:14 +0000 (00:59 +0200)]
net: sctp: fix permissions for rto_alpha and rto_beta knobs

Commit 3fd091e73b81 ("[SCTP]: Remove multiple levels of msecs
to jiffies conversions.") has silently changed permissions for
rto_alpha and rto_beta knobs from 0644 to 0444. The purpose of
this was to discourage users from tweaking rto_alpha and
rto_beta knobs in production environments since they are key
to correctly compute rtt/srtt.

RFC4960 under section 6.3.1. RTO Calculation says regarding
rto_alpha and rto_beta under rule C3 and C4:

  [...]
  C3)  When a new RTT measurement R' is made, set

       RTTVAR <- (1 - RTO.Beta) * RTTVAR + RTO.Beta * |SRTT - R'|

       and

       SRTT <- (1 - RTO.Alpha) * SRTT + RTO.Alpha * R'

       Note: The value of SRTT used in the update to RTTVAR
       is its value before updating SRTT itself using the
       second assignment. After the computation, update
       RTO <- SRTT + 4 * RTTVAR.

  C4)  When data is in flight and when allowed by rule C5
       below, a new RTT measurement MUST be made each round
       trip. Furthermore, new RTT measurements SHOULD be
       made no more than once per round trip for a given
       destination transport address. There are two reasons
       for this recommendation: First, it appears that
       measuring more frequently often does not in practice
       yield any significant benefit [ALLMAN99]; second,
       if measurements are made more often, then the values
       of RTO.Alpha and RTO.Beta in rule C3 above should be
       adjusted so that SRTT and RTTVAR still adjust to
       changes at roughly the same rate (in terms of how many
       round trips it takes them to reflect new values) as
       they would if making only one measurement per
       round-trip and using RTO.Alpha and RTO.Beta as given
       in rule C3. However, the exact nature of these
       adjustments remains a research issue.
  [...]

While it is discouraged to adjust rto_alpha and rto_beta
and not further specified how to adjust them, the RFC also
doesn't explicitly forbid it, but rather gives a RECOMMENDED
default value (rto_alpha=3, rto_beta=2). We have a couple
of users relying on the old permissions before they got
changed. That said, if someone really has the urge to adjust
them, we could allow it with a warning in the log.

Fixes: 3fd091e73b81 ("[SCTP]: Remove multiple levels of msecs to jiffies conversions.")
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Cc: Vlad Yasevich <vyasevich@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoMerge branch 'csum_fixes'
David S. Miller [Sun, 15 Jun 2014 08:00:56 +0000 (01:00 -0700)]
Merge branch 'csum_fixes'

Tom Herbert says:

====================
Fixes related to some recent checksum modifications.

- Fix GSO constants to match NETIF flags
- Fix logic in saving checksum complete in __skb_checksum_complete
- Call __skb_checksum_complete from UDP if we are checksumming over
  whole packet in order to save checksum.
- Fixes to VXLAN to work correctly with checksum complete
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agovxlan: Checksum fixes
Tom Herbert [Sun, 15 Jun 2014 06:24:36 +0000 (23:24 -0700)]
vxlan: Checksum fixes

Call skb_pop_rcv_encapsulation and postpull_rcsum for the Ethernet
header to work properly with checksum complete.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: add skb_pop_rcv_encapsulation
Tom Herbert [Sun, 15 Jun 2014 06:24:28 +0000 (23:24 -0700)]
net: add skb_pop_rcv_encapsulation

This function is used by UDP encapsulation protocols in RX when
crossing encapsulation boundary. If ip_summed is set to
CHECKSUM_UNNECESSARY and encapsulation is not set, change to
CHECKSUM_NONE since the checksum has not been validated within the
encapsulation. Clears csum_valid by the same rationale.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agoudp: call __skb_checksum_complete when doing full checksum
Tom Herbert [Sun, 15 Jun 2014 06:24:20 +0000 (23:24 -0700)]
udp: call __skb_checksum_complete when doing full checksum

In __udp_lib_checksum_complete check if checksum is being done over all
the data (len is equal to skb->len) and if it is call
__skb_checksum_complete instead of __skb_checksum_complete_head. This
allows checksum to be saved in checksum complete.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: Fix save software checksum complete
Tom Herbert [Sun, 15 Jun 2014 06:24:03 +0000 (23:24 -0700)]
net: Fix save software checksum complete

Geert reported issues regarding checksum complete and UDP.
The logic introduced in commit 7e3cead5172927732f51fde
("net: Save software checksum complete") is not correct.

This patch:
1) Restores code in __skb_checksum_complete_header except for setting
   CHECKSUM_UNNECESSARY. This function may be calculating checksum on
   something less than skb->len.
2) Adds saving checksum to __skb_checksum_complete. The full packet
   checksum 0..skb->len is calculated without adding in pseudo header.
   This value is saved in skb->csum and then the pseudo header is added
   to that to derive the checksum for validation.
3) In both __skb_checksum_complete_header and __skb_checksum_complete,
   set skb->csum_valid to whether checksum of zero was computed. This
   allows skb_csum_unnecessary to return true without changing to
   CHECKSUM_UNNECESSARY which was done previously.
4) Copy new csum related bits in __copy_skb_header.

Reported-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
10 years agonet: Fix GSO constants to match NETIF flags
Tom Herbert [Sun, 15 Jun 2014 06:23:52 +0000 (23:23 -0700)]
net: Fix GSO constants to match NETIF flags

Joseph Gasparakis reported that VXLAN GSO offload stopped working with
i40e device after recent UDP changes. The problem is that the
SKB_GSO_* bits are out of sync with the corresponding NETIF flags. This
patch fixes that. Also, we add BUILD_BUG_ONs in net_gso_ok for several
GSO constants that were missing to avoid the problem in the future.

Reported-by: Joseph Gasparakis <joseph.gasparakis@intel.com>
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>