According to Intel Software Developer's Manual, Vol. 3B, Appendix H.4.2,
exit qualification should be of natural width. However, current code
uses u64 as the data type for this register, which occasionally
introduces invalid value to VMExit handling logics. This patch fixes
this bug.
I have tested Windows and Linux guest on i386 host, and they can boot
successfully with this patch.
Signed-off-by: Qing He <qing.he@intel.com> Signed-off-by: Avi Kivity <avi@qumranet.com>
Avi Kivity [Mon, 10 Sep 2007 08:28:17 +0000 (11:28 +0300)]
KVM: MMU: Don't do GFP_NOWAIT allocations
Before preempt notifiers, kvm needed to allocate memory with GFP_NOWAIT so
as not to have to enable preemption and take a heavyweight exit. On oom, we'd
fall back to a GFP_KERNEL allocation.
With preemption notifiers, we can do a GFP_KERNEL allocation, and perform
the heavyweight exit only if the kernel decides to put us to sleep.
Laurent Vivier [Thu, 30 Aug 2007 12:56:21 +0000 (14:56 +0200)]
KVM: Simplify memory allocation
The mutex->splinlock convertion alllows us to make some code simplifications.
As we can keep the lock longer, we don't have to release it and then
have to check if the environment has not been modified before re-taking it. We
can remove kvm->busy and kvm->memory_config_version.
Signed-off-by: Laurent Vivier <Laurent.Vivier@bull.net> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Sun, 9 Sep 2007 11:12:54 +0000 (14:12 +0300)]
KVM: Clean up unloved invlpg emulation
invlpg shouldn't fetch the "src" address, since it may not be valid,
however SVM's "solution" which neuters emulation of all group 7
instruction is horrible and breaks kvm-lite. The simplest fix is to
put a special check in for invlpg.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Qing He [Mon, 17 Sep 2007 06:47:13 +0000 (14:47 +0800)]
KVM: deliver PIC interrupt only to vcpu0
This patch changes the PIC interrupts delivery. Now it is only delivered
to vcpu0 when either condition is met (on vcpu0):
1. local APIC is hardware disabled
2. LVT0 is unmasked and configured to delivery mode ExtInt
It fixes the 2x faster wall clock on x86_64 and SMP i386 Linux guests
Signed-off-by: Eddie (Yaozu) Dong <eddie.dong@intel.com> Signed-off-by: Qing He <qing.he@intel.com> Signed-off-by: Avi Kivity <avi@qumranet.com>
Eddie Dong [Mon, 3 Sep 2007 13:56:58 +0000 (16:56 +0300)]
KVM: Keep track of missed timer irq injections
APIC timer IRQ is set every time when a certain period
expires at host time, but the guest may be descheduled
at that time and thus the irq be overwritten by later fire.
This patch keep track of firing irq numbers and decrease
only when the IRQ is injected to guest or buffered in
APIC.
Signed-off-by: Yaozu (Eddie) Dong <Eddie.Dong@intel.com> Signed-off-by: Qing He <qing.he@intel.com> Signed-off-by: Avi Kivity <avi@qumranet.com>
This patch enables TPR shadow of VMX on CR8 access. 64bit Windows using
CR8 access TPR frequently. The TPR shadow can improve the performance of
access TPR by not causing vmexit.
Signed-off-by: Sheng Yang <sheng.yang@intel.com> Signed-off-by: Yaozu (Eddie) Dong <eddie.dong@intel.com> Signed-off-by: Qing He <qing.he@intel.com> Signed-off-by: Avi Kivity <avi@qumranet.com>
Eddie Dong [Thu, 6 Sep 2007 09:22:56 +0000 (12:22 +0300)]
KVM: in-kernel LAPIC save and restore support
This patch adds a new vcpu-based IOCTL to save and restore the local
apic registers for a single vcpu. The kernel only copies the apic page as
a whole, extraction of registers is left to userspace side. On restore, the
APIC timer is restarted from the initial count, this introduces a little
delay, but works fine.
Signed-off-by: Yaozu (Eddie) Dong <eddie.dong@intel.com> Signed-off-by: Qing He <qing.he@intel.com> Signed-off-by: Avi Kivity <avi@qumranet.com>
He, Qing [Thu, 2 Aug 2007 11:03:07 +0000 (14:03 +0300)]
KVM: Bypass irq_pending get/set when using in kernel irqchip
vcpu->irq_pending is saved in get/set_sreg IOCTL, but when in-kernel
local APIC is used, doing this may occasionally overwrite vcpu->apic to
an invalid value, as in the vm restore path.
Eddie Dong [Wed, 12 Sep 2007 07:58:04 +0000 (10:58 +0300)]
KVM: Emulate local APIC in kernel
Because lightweight exits (exits which don't involve userspace) are many
times faster than heavyweight exits, it makes sense to emulate high usage
devices in the kernel. The local APIC is one such device, especially for
Windows and for SMP, so we add an APIC model to kvm.
It also allows in-kernel host-side drivers to inject interrupts without
going through userspace.
[compile fix on i386 from Jindrich Makovicka]
Signed-off-by: Yaozu (Eddie) Dong <Eddie.Dong@intel.com> Signed-off-by: Qing He <qing.he@intel.com> Signed-off-by: Avi Kivity <avi@qumranet.com>
Eddie Dong [Wed, 18 Jul 2007 08:34:57 +0000 (11:34 +0300)]
KVM: Define and use cr8 access functions
This patch is to wrap APIC base register and CR8 operation which can
provide a unique API for user level irqchip and kernel irqchip.
This is a preparation of merging lapic/ioapic patch.
Signed-off-by: Yaozu (Eddie) Dong <eddie.dong@intel.com> Signed-off-by: Avi Kivity <avi@qumranet.com>
Laurent Vivier [Thu, 23 Aug 2007 14:33:11 +0000 (16:33 +0200)]
KVM: VMX: Split segments reload in vmx_load_host_state()
vmx_load_host_state() bundles fs, gs, ldt, and tss reloading into
one in the hope that it is infrequent. With smp guests, fs reloading is
frequent due to fs being used by threads.
Unbundle the reloads so reduce expensive gs reloads.
Signed-off-by: Laurent Vivier <Laurent.Vivier@bull.net> Signed-off-by: Avi Kivity <avi@qumranet.com>
Avi Kivity [Wed, 15 Aug 2007 12:23:34 +0000 (15:23 +0300)]
KVM: Close minor race in signal handling
We need to check for signals inside the critical section, otherwise a
signal can be sent which we will not notice. Also move the check
before entry, so that if the signal happens before the first entry,
we exit immediately instead of waiting for something to happen to the
guest.
Laurent Vivier [Sun, 5 Aug 2007 07:36:40 +0000 (10:36 +0300)]
KVM: Cleanup string I/O instruction emulation
Both vmx and svm decode the I/O instructions, and both botch the job,
requiring the instruction prefixes to be fetched in order to completely
decode the instruction.
So, if we see a string I/O instruction, use the x86 emulator to decode it,
as it already has all the prefix decoding machinery.
This patch defines ins/outs opcodes in x86_emulate.c and calls
emulate_instruction() from io_interception() (svm.c) and from handle_io()
(vmx.c). It removes all vmx/svm prefix instruction decoders
(get_addr_size(), io_get_override(), io_address(), get_io_count())
Signed-off-by: Laurent Vivier <Laurent.Vivier@bull.net> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Wed, 1 Aug 2007 00:12:22 +0000 (10:12 +1000)]
KVM: Remove stat_set from debugfs
We shouldn't define stat_set on the debug attributes, since that will
cause silent failure on writing: without a set argument, userspace
will get -EACCESS.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Tue, 31 Jul 2007 10:42:42 +0000 (20:42 +1000)]
KVM: SVM: Make set_msr_interception more reliable
set_msr_interception() is used by svm to set up which MSRs should be
intercepted. It can only fail if someone has changed the code to try
to intercept an MSR without updating the array of ranges.
The return value is ignored anyway: it should just BUG() if it doesn't
work. (A build-time failure would be better, but that's tricky).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
All the physical CPUs on the board should support the same VMX feature
set. Add check_processor_compatibility to kvm_arch_ops for the consistency
check.
Signed-off-by: Sheng Yang <sheng.yang@intel.com> Signed-off-by: Avi Kivity <avi@qumranet.com>
kvm_vm_ioctl_get_dirty_log scans bitmap to see it it's all zero, but
doesn't use that information.
Avi says:
Looks like it was used to guard kvm_mmu_slot_remove_write_access();
optimizing the case where the guest just leaves the screen alone (which
it usually does, especially in benchmarks).
Rusty Russell [Mon, 30 Jul 2007 11:13:43 +0000 (21:13 +1000)]
KVM: Use alignment properties of vcpu to simplify FPU ops
Now we use a kmem cache for allocating vcpus, we can get the 16-byte
alignment required by fxsave & fxrstor instructions, and avoid
manually aligning the buffer.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Mon, 30 Jul 2007 11:12:19 +0000 (21:12 +1000)]
KVM: Use kmem cache for allocating vcpus
Avi wants the allocations of vcpus centralized again. The easiest way
is to add a "size" arg to kvm_init_arch, and expose the thus-prepared
cache to the modules.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Mon, 30 Jul 2007 10:08:05 +0000 (20:08 +1000)]
KVM: SVM: internal function name cleanup
Changes some svm.c internal function names:
1) io_adress -> io_address (de-germanify the spelling)
2) kvm_reput_irq -> reput_irq (it's not a generic kvm function)
3) kvm_do_inject_irq -> (it's not a generic kvm function)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Mon, 30 Jul 2007 10:07:08 +0000 (20:07 +1000)]
KVM: SVM: de-containization
container_of is wonderful, but not casting at all is better. This
patch changes svm.c's internal functions to pass "struct vcpu_svm"
instead of "struct kvm_vcpu" and using container_of.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Mon, 30 Jul 2007 06:31:43 +0000 (16:31 +1000)]
KVM: VMX: pass vcpu_vmx internally
container_of is wonderful, but not casting at all is better. This
patch changes vmx.c's internal functions to pass "struct vcpu_vmx"
instead of "struct kvm_vcpu" and using container_of.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Avi Kivity [Wed, 11 Jul 2007 15:17:21 +0000 (18:17 +0300)]
KVM: Use the scheduler preemption notifiers to make kvm preemptible
Current kvm disables preemption while the new virtualization registers are
in use. This of course is not very good for latency sensitive workloads (one
use of virtualization is to offload user interface and other latency
insensitive stuff to a container, so that it is easier to analyze the
remaining workload). This patch re-enables preemption for kvm; preemption
is now only disabled when switching the registers in and out, and during
the switch to guest mode and back.
Contains fixes from Shaohua Li <shaohua.li@intel.com>.
Jeff Dike [Mon, 16 Jul 2007 19:24:47 +0000 (15:24 -0400)]
KVM: add hypercall nr to kvm_run
Add the hypercall number to kvm_run and initialize it. This changes the ABI,
but as this particular ABI was unusable before this no users are affected.
Signed-off-by: Jeff Dike <jdike@linux.intel.com> Signed-off-by: Avi Kivity <avi@qumranet.com>
KVM: VMX: Improve the method of writing vmcs control
Put cpu feature detecting part in hardware_setup, and stored the vmcs
condition in global variable for further check.
[glommer: fix for some i386-only machines not supporting CR8 load/store
exiting]
Signed-off-by: Sheng Yang <sheng.yang@intel.com> Signed-off-by: Glauber de Oliveira Costa <gcosta@redhat.com> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Fri, 27 Jul 2007 07:16:56 +0000 (17:16 +1000)]
KVM: Dynamically allocate vcpus
This patch converts the vcpus array in "struct kvm" to a pointer
array, and changes the "vcpu_create" and "vcpu_setup" hooks into one
"vcpu_create" call which does the allocation and initialization of the
vcpu (calling back into the kvm_vcpu_init core helper).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Wed, 25 Jul 2007 03:29:51 +0000 (13:29 +1000)]
KVM: load_pdptrs() cleanups
load_pdptrs can be handed an invalid cr3, and it should not oops.
This can happen because we injected #gp in set_cr3() after we set
vcpu->cr3 to the invalid value, or from kvm_vcpu_ioctl_set_sregs(), or
memory configuration changes after the guest did set_cr3().
We should also copy the pdpte array once, before checking and
assigning, otherwise an SMP guest can potentially alter the values
between the check and the set.
Finally one nitpick: ret = 1 should be done as late as possible: this
allows GCC to check for unset "ret" should the function change in
future.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Wed, 18 Jul 2007 03:05:58 +0000 (13:05 +1000)]
KVM: Trivial: Use standard BITMAP macros, open-code userspace-exposed header
Creating one's own BITMAP macro seems suboptimal: if we use manual
arithmetic in the one place exposed to userspace, we can use standard
macros elsewhere.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Tue, 17 Jul 2007 13:34:16 +0000 (23:34 +1000)]
KVM: Use standard CR4 flags, tighten checking
On this machine (Intel), writing to the CR4 bits 0x00000800 and
0x00001000 cause a GPF. The Intel manual is a little unclear, but
AFIACT they're reserved, too.
Also fix spelling of CR4_RESEVED_BITS.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Rusty Russell [Tue, 17 Jul 2007 13:16:11 +0000 (23:16 +1000)]
KVM: Trivial: Make decode_register() static
I have shied away from touching x86_emulate.c (it could definitely use
some love, but it is forked from the Xen code, and it would be more
productive to cross-merge fixes).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Avi Kivity <avi@qumranet.com>
Avi Kivity [Thu, 27 Sep 2007 08:07:04 +0000 (10:07 +0200)]
i386: Expose IOAPIC register definitions even if CONFIG_X86_IO_APIC is not set
KVM reuses the IOAPIC register definitions, and needs them even if the
host is not compiled with IOAPIC support. Move the #ifdef below so that only
the IOAPIC variables and functions are protected, and the register definitions
are available to all.
Nick Piggin [Sat, 13 Oct 2007 01:07:38 +0000 (03:07 +0200)]
x86: optimise barriers
According to latest memory ordering specification documents from Intel
and AMD, both manufacturers are committed to in-order loads from
cacheable memory for the x86 architecture. Hence, smp_rmb() may be a
simple barrier.
Also according to those documents, and according to existing practice in
Linux (eg. spin_unlock doesn't enforce ordering), stores to cacheable
memory are visible in program order too. Special string stores are safe
-- their constituent stores may be out of order, but they must complete
in order WRT surrounding stores. Nontemporal stores to WB memory can go
out of order, and so they should be fenced explicitly to make them
appear in-order WRT other stores. Hence, smp_wmb() may be a simple
barrier.
In userspace microbenchmarks on a core2 system, fence instructions range
anywhere from around 15 cycles to 50, which may not be totally
insignificant in performance critical paths (code size will go down
too).
However the primary motivation for this is to have the canonical barrier
implementation for x86 architecture.
smp_rmb on buggy pentium pros remains a locked op, which is apparently
required.
Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
projects / linux-2.6 / log