What it is: vhost net is a character device that can be used to reduce
the number of system calls involved in virtio networking.
Existing virtio net code is used in the guest without modification.
There's similarity with vringfd, with some differences and reduced scope
- uses eventfd for signalling
- structures can be moved around in memory at any time (good for
migration, bug work-arounds in userspace)
- write logging is supported (good for migration)
- support memory table and not just an offset (needed for kvm)
common virtio related code has been put in a separate file vhost.c and
can be made into a separate module if/when more backends appear. I used
Rusty's lguest.c as the source for developing this part : this supplied
me with witty comments I wouldn't be able to write myself.
What it is not: vhost net is not a bus, and not a generic new system
call. No assumptions are made on how guest performs hypercalls.
Userspace hypervisors are supported as well as kvm.
How it works: Basically, we connect virtio frontend (configured by
userspace) to a backend. The backend could be a network device, or a tap
device. Backend is also configured by userspace, including vlan/mac
etc.
Status: This works for me, and I haven't see any crashes.
Compared to userspace, people reported improved latency (as I save up to
4 system calls per packet), as well as better bandwidth and CPU
utilization.
Features that I plan to look at in the future:
- mergeable buffers
- zero copy
- scalability tuning: figure out the best threading model to use
Note on RCU usage (this is also documented in vhost.h, near
private_pointer which is the value protected by this variant of RCU):
what is happening is that the rcu_dereference() is being used in a
workqueue item. The role of rcu_read_lock() is taken on by the start of
execution of the workqueue item, of rcu_read_unlock() by the end of
execution of the workqueue item, and of synchronize_rcu() by
flush_workqueue()/flush_work(). In the future we might need to apply
some gcc attribute or sparse annotation to the function passed to
INIT_WORK(). Paul's ack below is for this RCU usage.
(Includes fixes by Alan Cox <alan@linux.intel.com>,
David L Stevens <dlstevens@us.ibm.com>,
Chris Wright <chrisw@redhat.com>)
Acked-by: Rusty Russell <rusty@rustcorp.com.au>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
User space may not want to overwrite asynchronously changing VCPU event
states on write-back. So allow to skip nmi.pending and sipi_vector by
setting corresponding bits in the flags field of kvm_vcpu_events.
[avi: advertise the bits in KVM_GET_VCPU_EVENTS]
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
The vcpus are initialized with irr_pending set to false, but
loading the LAPIC registers with pending IRR fails to reset
the irr_pending variable.
Cc: stable@kernel.org
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
The invlpg prefault optimization breaks Windows 2008 R2 occasionally.
The visible effect is that the invlpg handler instantiates a pte which
is, microseconds later, written with a different gfn by another vcpu.
The OS could have other mechanisms to prevent a present translation from
being used, which the hypervisor is unaware of.
While the documentation states that the cpu is at liberty to prefetch tlb
entries, it looks like this is not heeded, so remove tlb prefetch from
invlpg.
Cc: stable@kernel.org
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu: (34 commits)
m68k: rename global variable vmalloc_end to m68k_vmalloc_end
percpu: add missing per_cpu_ptr_to_phys() definition for UP
percpu: Fix kdump failure if booted with percpu_alloc=page
percpu: make misc percpu symbols unique
percpu: make percpu symbols in ia64 unique
percpu: make percpu symbols in powerpc unique
percpu: make percpu symbols in x86 unique
percpu: make percpu symbols in xen unique
percpu: make percpu symbols in cpufreq unique
percpu: make percpu symbols in oprofile unique
percpu: make percpu symbols in tracer unique
percpu: make percpu symbols under kernel/ and mm/ unique
percpu: remove some sparse warnings
percpu: make alloc_percpu() handle array types
vmalloc: fix use of non-existent percpu variable in put_cpu_var()
this_cpu: Use this_cpu_xx in trace_functions_graph.c
this_cpu: Use this_cpu_xx for ftrace
this_cpu: Use this_cpu_xx in nmi handling
this_cpu: Use this_cpu operations in RCU
this_cpu: Use this_cpu ops for VM statistics
...
Fix up trivial (famous last words) global per-cpu naming conflicts in
arch/x86/kvm/svm.c
mm/slab.c
* 'kvm-updates/2.6.33' of git://git.kernel.org/pub/scm/virt/kvm/kvm: (84 commits)
KVM: VMX: Fix comparison of guest efer with stale host value
KVM: s390: Fix prefix register checking in arch/s390/kvm/sigp.c
KVM: Drop user return notifier when disabling virtualization on a cpu
KVM: VMX: Disable unrestricted guest when EPT disabled
KVM: x86 emulator: limit instructions to 15 bytes
KVM: s390: Make psw available on all exits, not just a subset
KVM: x86: Add KVM_GET/SET_VCPU_EVENTS
KVM: VMX: Report unexpected simultaneous exceptions as internal errors
KVM: Allow internal errors reported to userspace to carry extra data
KVM: Reorder IOCTLs in main kvm.h
KVM: x86: Polish exception injection via KVM_SET_GUEST_DEBUG
KVM: only clear irq_source_id if irqchip is present
KVM: x86: disallow KVM_{SET,GET}_LAPIC without allocated in-kernel lapic
KVM: x86: disallow multiple KVM_CREATE_IRQCHIP
KVM: VMX: Remove vmx->msr_offset_efer
KVM: MMU: update invlpg handler comment
KVM: VMX: move CR3/PDPTR update to vmx_set_cr3
KVM: remove duplicated task_switch check
KVM: powerpc: Fix BUILD_BUG_ON condition
KVM: VMX: Use shared msr infrastructure
...
Trivial conflicts due to new Kconfig options in arch/Kconfig and kernel/Makefile
update_transition_efer() masks out some efer bits when deciding whether
to switch the msr during guest entry; for example, NX is emulated using the
mmu so we don't need to disable it, and LMA/LME are handled by the hardware.
However, with shared msrs, the comparison is made against a stale value;
at the time of the guest switch we may be running with another guest's efer.
Fix by deferring the mask/compare to the actual point of guest entry.
Noted by Marcelo.
Signed-off-by: Avi Kivity <avi@redhat.com>
This way, we don't leave a dangling notifier on cpu hotunplug or module
unload. In particular, module unload leaves the notifier pointing into
freed memory.
Signed-off-by: Avi Kivity <avi@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Otherwise would cause VMEntry failure when using ept=0 on unrestricted guest
supported processors.
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
While we are never normally passed an instruction that exceeds 15 bytes,
smp games can cause us to attempt to interpret one, which will cause
large latencies in non-preempt hosts.
Cc: stable@kernel.org
Signed-off-by: Avi Kivity <avi@redhat.com>
This new IOCTL exports all yet user-invisible states related to
exceptions, interrupts, and NMIs. Together with appropriate user space
changes, this fixes sporadic problems of vmsave/restore, live migration
and system reset.
[avi: future-proof abi by adding a flags field]
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
These happen when we trap an exception when another exception is being
delivered; we only expect these with MCEs and page faults. If something
unexpected happens, things probably went south and we're better off reporting
an internal error and freezing.
Signed-off-by: Avi Kivity <avi@redhat.com>
Usually userspace will freeze the guest so we can inspect it, but some
internal state is not available. Add extra data to internal error
reporting so we can expose it to the debugger. Extra data is specific
to the suberror.
Signed-off-by: Avi Kivity <avi@redhat.com>
Decouple KVM_GUESTDBG_INJECT_DB and KVM_GUESTDBG_INJECT_BP from
KVM_GUESTDBG_ENABLE, their are actually orthogonal. At this chance,
avoid triggering the WARN_ON in kvm_queue_exception if there is already
an exception pending and reject such invalid requests.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Otherwise kvm will leak memory on multiple KVM_CREATE_IRQCHIP.
Also serialize multiple accesses with kvm->lock.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Large page translations are always synchronized (either in level 3
or level 2), so its not necessary to properly deal with them
in the invlpg handler.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
GUEST_CR3 is updated via kvm_set_cr3 whenever CR3 is modified from
outside guest context. Similarly pdptrs are updated via load_pdptrs.
Let kvm_set_cr3 perform the update, removing it from the vcpu_run
fast path.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Acked-by: Acked-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Instead of reloading syscall MSRs on every preemption, use the new shared
msr infrastructure to reload them at the last possible minute (just before
exit to userspace).
Improves vcpu/idle/vcpu switches by about 2000 cycles (when EFER needs to be
reloaded as well).
[jan: fix slot index missing indirection]
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
The various syscall-related MSRs are fairly expensive to switch. Currently
we switch them on every vcpu preemption, which is far too often:
- if we're switching to a kernel thread (idle task, threaded interrupt,
kernel-mode virtio server (vhost-net), for example) and back, then
there's no need to switch those MSRs since kernel threasd won't
be exiting to userspace.
- if we're switching to another guest running an identical OS, most likely
those MSRs will have the same value, so there's little point in reloading
them.
- if we're running the same OS on the guest and host, the MSRs will have
identical values and reloading is unnecessary.
This patch uses the new user return notifiers to implement last-minute
switching, and checks the msr values to avoid unnecessary reloading.
Signed-off-by: Avi Kivity <avi@redhat.com>
Currently MSR_KERNEL_GS_BASE is saved and restored as part of the
guest/host msr reloading. Since we wish to lazy-restore all the other
msrs, save and reload MSR_KERNEL_GS_BASE explicitly instead of using
the common code.
Signed-off-by: Avi Kivity <avi@redhat.com>
The svm_set_cr0() call will initialize save->cr0 properly even when npt is
enabled, clearing the NW and CD bits as expected, so we don't need to
initialize it manually for npt_enabled anymore.
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
svm_vcpu_reset() was not properly resetting the contents of the guest-visible
cr0 register, causing the following issue:
https://bugzilla.redhat.com/show_bug.cgi?id=525699
Without resetting cr0 properly, the vcpu was running the SIPI bootstrap routine
with paging enabled, making the vcpu get a pagefault exception while trying to
run it.
Instead of setting vmcb->save.cr0 directly, the new code just resets
kvm->arch.cr0 and calls kvm_set_cr0(). The bits that were set/cleared on
vmcb->save.cr0 (PG, WP, !CD, !NW) will be set properly by svm_set_cr0().
kvm_set_cr0() is used instead of calling svm_set_cr0() directly to make sure
kvm_mmu_reset_context() is called to reset the mmu to nonpaging mode.
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
This should have no effect, it is just to make the code clearer.
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
When we migrate a kvm guest that uses pvclock between two hosts, we may
suffer a large skew. This is because there can be significant differences
between the monotonic clock of the hosts involved. When a new host with
a much larger monotonic time starts running the guest, the view of time
will be significantly impacted.
Situation is much worse when we do the opposite, and migrate to a host with
a smaller monotonic clock.
This proposed ioctl will allow userspace to inform us what is the monotonic
clock value in the source host, so we can keep the time skew short, and
more importantly, never goes backwards. Userspace may also need to trigger
the current data, since from the first migration onwards, it won't be
reflected by a simple call to clock_gettime() anymore.
[marcelo: future-proof abi with a flags field]
[jan: fix KVM_GET_CLOCK by clearing flags field instead of checking it]
Signed-off-by: Glauber Costa <glommer@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Commit 705c5323 opened the doors of hell by unconditionally injecting
single-step flags as long as guest_debug signaled this. This doesn't
work when the guest branches into some interrupt or exception handler
and triggers a vmexit with flag reloading.
Fix it by saving cs:rip when user space requests single-stepping and
restricting the trace flag injection to this guest code position.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Support for Xen PV-on-HVM guests can be implemented almost entirely in
userspace, except for handling one annoying MSR that maps a Xen
hypercall blob into guest address space.
A generic mechanism to delegate MSR writes to userspace seems overkill
and risks encouraging similar MSR abuse in the future. Thus this patch
adds special support for the Xen HVM MSR.
I implemented a new ioctl, KVM_XEN_HVM_CONFIG, that lets userspace tell
KVM which MSR the guest will write to, as well as the starting address
and size of the hypercall blobs (one each for 32-bit and 64-bit) that
userspace has loaded from files. When the guest writes to the MSR, KVM
copies one page of the blob from userspace to the guest.
I've tested this patch with a hacked-up version of Gerd's userspace
code, booting a number of guests (CentOS 5.3 i386 and x86_64, and
FreeBSD 8.0-RC1 amd64) and exercising PV network and block devices.
[jan: fix i386 build warning]
[avi: future proof abi with a flags field]
Signed-off-by: Ed Swierk <eswierk@aristanetworks.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
This (broken) check dates back to the days when this code was shared
across architectures. x86 has IOMEM, so drop it.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
If cpufreq can't determine the CPU khz, or cpufreq is not compiled in,
we should fallback to the measured TSC khz.
Signed-off-by: Zachary Amsden <zamsden@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
New AMD processors (Family 0x10 models 8+) support the Pause
Filter Feature. This feature creates a new field in the VMCB
called Pause Filter Count. If Pause Filter Count is greater
than 0 and intercepting PAUSEs is enabled, the processor will
increment an internal counter when a PAUSE instruction occurs
instead of intercepting. When the internal counter reaches the
Pause Filter Count value, a PAUSE intercept will occur.
This feature can be used to detect contended spinlocks,
especially when the lock holding VCPU is not scheduled.
Rescheduling another VCPU prevents the VCPU seeking the
lock from wasting its quantum by spinning idly.
Experimental results show that most spinlocks are held
for less than 1000 PAUSE cycles or more than a few
thousand. Default the Pause Filter Counter to 3000 to
detect the contended spinlocks.
Processor support for this feature is indicated by a CPUID
bit.
On a 24 core system running 4 guests each with 16 VCPUs,
this patch improved overall performance of each guest's
32 job kernbench by approximately 3-5% when combined
with a scheduler algorithm thati caused the VCPU to
sleep for a brief period. Further performance improvement
may be possible with a more sophisticated yield algorithm.
Signed-off-by: Mark Langsdorf <mark.langsdorf@amd.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
New NHM processors will support Pause-Loop Exiting by adding 2 VM-execution
control fields:
PLE_Gap - upper bound on the amount of time between two successive
executions of PAUSE in a loop.
PLE_Window - upper bound on the amount of time a guest is allowed to execute in
a PAUSE loop
If the time, between this execution of PAUSE and previous one, exceeds the
PLE_Gap, processor consider this PAUSE belongs to a new loop.
Otherwise, processor determins the the total execution time of this loop(since
1st PAUSE in this loop), and triggers a VM exit if total time exceeds the
PLE_Window.
* Refer SDM volume 3b section 21.6.13 & 22.1.3.
Pause-Loop Exiting can be used to detect Lock-Holder Preemption, where one VP
is sched-out after hold a spinlock, then other VPs for same lock are sched-in
to waste the CPU time.
Our tests indicate that most spinlocks are held for less than 212 cycles.
Performance tests show that with 2X LP over-commitment we can get +2% perf
improvement for kernel build(Even more perf gain with more LPs).
Signed-off-by: Zhai Edwin <edwin.zhai@intel.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
With all important informations now delivered through
tracepoints we can savely remove the nsvm_printk debugging
code for nested svm.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This patch adds a tracepoint for the event that the guest
executed the SKINIT instruction. This information is
important because SKINIT is an SVM extenstion not yet
implemented by nested SVM and we may need this information
for debugging hypervisors that do not yet run on nested SVM.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This patch adds a tracepoint for the event that the guest
executed the INVLPGA instruction.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This patch adds a special tracepoint for the event that a
nested #vmexit is injected because kvm wants to inject an
interrupt into the guest.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This patch adds a tracepoint for a nested #vmexit that gets
re-injected to the guest.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This patch adds a tracepoint for every #vmexit we get from a
nested guest.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This patch adds a dedicated kvm tracepoint for a nested
vmrun.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
The nested SVM code emulates a #vmexit caused by a request
to open the irq window right in the request function. This
is a bug because the request function runs with preemption
and interrupts disabled but the #vmexit emulation might
sleep. This can cause a schedule()-while-atomic bug and is
fixed with this patch.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
If event_inj is valid on a #vmexit the host CPU would write
the contents to exit_int_info, so the hypervisor knows that
the event wasn't injected.
We don't do this in nested SVM by now which is a bug and
fixed by this patch.
Signed-off-by: Alexander Graf <agraf@suse.de>
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
For a while now, we are issuing a rdmsr instruction to find out which
msrs in our save list are really supported by the underlying machine.
However, it fails to account for kvm-specific msrs, such as the pvclock
ones.
This patch moves then to the beginning of the list, and skip testing them.
Cc: stable@kernel.org
Signed-off-by: Glauber Costa <glommer@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Push TF and RF injection and filtering on guest single-stepping into the
vender get/set_rflags callbacks. This makes the whole mechanism more
robust wrt user space IOCTL order and instruction emulations.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Disable paravirt MMU capability reporting, so that new (or rebooted)
guests switch to native operation.
Paravirt MMU is a burden to maintain and does not bring significant
advantages compared to shadow anymore.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Much of so far vendor-specific code for setting up guest debug can
actually be handled by the generic code. This also fixes a minor deficit
in the SVM part /wrt processing KVM_GUESTDBG_ENABLE.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
