ioeventfd is a mechanism to register PIO/MMIO regions to trigger an eventfd
signal when written to by a guest. Host userspace can register any
arbitrary IO address with a corresponding eventfd and then pass the eventfd
to a specific end-point of interest for handling.
Normal IO requires a blocking round-trip since the operation may cause
side-effects in the emulated model or may return data to the caller.
Therefore, an IO in KVM traps from the guest to the host, causes a VMX/SVM
"heavy-weight" exit back to userspace, and is ultimately serviced by qemu's
device model synchronously before returning control back to the vcpu.
However, there is a subclass of IO which acts purely as a trigger for
other IO (such as to kick off an out-of-band DMA request, etc). For these
patterns, the synchronous call is particularly expensive since we really
only want to simply get our notification transmitted asychronously and
return as quickly as possible. All the sychronous infrastructure to ensure
proper data-dependencies are met in the normal IO case are just unecessary
overhead for signalling. This adds additional computational load on the
system, as well as latency to the signalling path.
Therefore, we provide a mechanism for registration of an in-kernel trigger
point that allows the VCPU to only require a very brief, lightweight
exit just long enough to signal an eventfd. This also means that any
clients compatible with the eventfd interface (which includes userspace
and kernelspace equally well) can now register to be notified. The end
result should be a more flexible and higher performance notification API
for the backend KVM hypervisor and perhipheral components.
To test this theory, we built a test-harness called "doorbell". This
module has a function called "doorbell_ring()" which simply increments a
counter for each time the doorbell is signaled. It supports signalling
from either an eventfd, or an ioctl().
We then wired up two paths to the doorbell: One via QEMU via a registered
io region and through the doorbell ioctl(). The other is direct via
ioeventfd.
You can download this test harness here:
ftp://ftp.novell.com/dev/ghaskins/doorbell.tar.bz2
The measured results are as follows:
qemu-mmio: 110000 iops, 9.09us rtt
ioeventfd-mmio: 200100 iops, 5.00us rtt
ioeventfd-pio: 367300 iops, 2.72us rtt
I didn't measure qemu-pio, because I have to figure out how to register a
PIO region with qemu's device model, and I got lazy. However, for now we
can extrapolate based on the data from the NULLIO runs of +2.56us for MMIO,
and -350ns for HC, we get:
qemu-pio: 153139 iops, 6.53us rtt
ioeventfd-hc: 412585 iops, 2.37us rtt
these are just for fun, for now, until I can gather more data.
Here is a graph for your convenience:
http://developer.novell.com/wiki/images/7/76/Iofd-chart.png
The conclusion to draw is that we save about 4us by skipping the userspace
hop.
--------------------
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Today kvm_io_bus_regsiter_dev() returns void and will internally BUG_ON
if it fails. We want to create dynamic MMIO/PIO entries driven from
userspace later in the series, so we need to enhance the code to be more
robust with the following changes:
1) Add a return value to the registration function
2) Fix up all the callsites to check the return code, handle any
failures, and percolate the error up to the caller.
3) Add an unregister function that collapses holes in the array
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
This changes bus accesses to use high-level kvm_io_bus_read/kvm_io_bus_write
functions. in_range now becomes unused so it is removed from device ops in
favor of read/write callbacks performing range checks internally.
This allows aliasing (mostly for in-kernel virtio), as well as better error
handling by making it possible to pass errors up to userspace.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Use slots_lock to protect device list on the bus. slots_lock is already
taken for read everywhere, so we only need to take it for write when
registering devices. This is in preparation to removing in_range and
kvm->lock around it.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Correct missing locking in a few places in x86's vm_ioctl handling path.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Since
commit 854b5338196b1175706e99d63be43a4f8d8ab607
Author: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>
KVM: s390: streamline memslot handling
s390 uses the values of the memslot instead of doing everything in the arch
ioctl handler of the KVM_SET_USER_MEMORY_REGION. Unfortunately we missed to
set the userspace_addr of our memslot due to our s390 ifdef in
__kvm_set_memory_region.
Old s390 userspace launchers did not notice, since they started the guest at
userspace address 0.
Because of CONFIG_DEFAULT_MMAP_MIN_ADDR we now put the guest at 1M userspace,
which does not work. This patch makes sure that new.userspace_addr is set
on s390.
This fix should go in quickly. Nevertheless, looking at the code we should
clean up that ifdef in the long term. Any kernel janitors?
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
This allows use of the powerful ftrace infrastructure.
See Documentation/trace/ for usage information.
[avi, stephen: various build fixes]
[sheng: fix control register breakage]
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Disable usage of 2M pages if VMX_EPT_2MB_PAGE_BIT (bit 16) is clear
in MSR_IA32_VMX_EPT_VPID_CAP and EPT is enabled.
[avi: s/largepages_disabled/largepages_enabled/ to avoid negative logic]
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Archs are free to use vcpu_id as they see fit. For x86 it is used as
vcpu's apic id. New ioctl is added to configure boot vcpu id that was
assumed to be 0 till now.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Protect irq injection/acking data structures with a separate irq_lock
mutex. This fixes the following deadlock:
CPU A CPU B
kvm_vm_ioctl_deassign_dev_irq()
mutex_lock(&kvm->lock); worker_thread()
-> kvm_deassign_irq() -> kvm_assigned_dev_interrupt_work_handler()
-> deassign_host_irq() mutex_lock(&kvm->lock);
-> cancel_work_sync() [blocked]
[gleb: fix ia64 path]
Reported-by: Alex Williamson <alex.williamson@hp.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Introduce irq_lock, and use to protect ioapic data structures.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
We modernize the io_device code so that we use container_of() instead of
dev->private, and move the vtable to a separate ops structure
(theoretically allows better caching for multiple instances of the same
ops structure)
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Acked-by: Chris Wright <chrisw@sous-sol.org>
Signed-off-by: Avi Kivity <avi@redhat.com>
Disable interrupt at interrupt handler and enable it when guest ack is for
the level triggered interrupt, to prevent reinjected interrupt. MSI/MSI-X don't
need it.
One possible problem is multiply same vector interrupt injected between irq
handler and scheduled work handler would be merged as one for MSI/MSI-X.
But AFAIK, the drivers handle it well.
The patch fixed the oplin card performance issue(MSI-X performance is half of
MSI/INTx).
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
KVM provides a complete virtual system environment for guests, including
support for injecting interrupts modeled after the real exception/interrupt
facilities present on the native platform (such as the IDT on x86).
Virtual interrupts can come from a variety of sources (emulated devices,
pass-through devices, etc) but all must be injected to the guest via
the KVM infrastructure. This patch adds a new mechanism to inject a specific
interrupt to a guest using a decoupled eventfd mechnanism: Any legal signal
on the irqfd (using eventfd semantics from either userspace or kernel) will
translate into an injected interrupt in the guest at the next available
interrupt window.
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
make_all_cpus_request contains a race condition which can
trigger false request completed status, as follows:
CPU0 CPU1
if (test_and_set_bit(req,&vcpu->requests))
.... if (test_and_set_bit(req,&vcpu->requests))
.. return
proceed to smp_call_function_many(wait=1)
Use a spinlock to serialize concurrent CPUs.
Cc: stable@kernel.org
Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
When slot is already allocated and being asked to be tracked we need
to break the large pages.
This code flush the mmu when someone ask a slot to start dirty bit
tracking.
Cc: stable@kernel.org
Signed-off-by: Izik Eidus <ieidus@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
If userspace specifies a memory slot that is larger than 8 petabytes, it
could overflow the largepages variable.
Cc: stable@kernel.org
Signed-off-by: Avi Kivity <avi@redhat.com>
If a slots guest physical address and host virtual address unequal (mod
large page size), then we would erronously try to back guest large pages
with host large pages. Detect this misalignment and diable large page
support for the trouble slot.
Cc: stable@kernel.org
Signed-off-by: Avi Kivity <avi@redhat.com>
kvm_handle_hva relies on mmu_lock protection to safely access
the memslot structures.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
kvm_assigned_dev_ack_irq is vulnerable to a race condition with the
interrupt handler function. It does:
if (dev->host_irq_disabled) {
enable_irq(dev->host_irq);
dev->host_irq_disabled = false;
}
If an interrupt triggers before the host->dev_irq_disabled assignment,
it will disable the interrupt and set dev->host_irq_disabled to true.
On return to kvm_assigned_dev_ack_irq, dev->host_irq_disabled is set to
false, and the next kvm_assigned_dev_ack_irq call will fail to reenable
it.
Other than that, having the interrupt handler and work handlers run in
parallel sounds like asking for trouble (could not spot any obvious
problem, but better not have to, its fragile).
CC: sheng.yang@intel.com
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Intel TXT(Trusted Execution Technology) required VMX off for all cpu to work
when system shutdown.
CC: Joseph Cihula <joseph.cihula@intel.com>
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
kvm_vcpu_block() unhalts vpu on an interrupt/timer without checking
if interrupt window is actually opened.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Currently timer events are processed before entering guest mode. Move it
to main vcpu event loop since timer events should be processed even while
vcpu is halted. Timer may cause interrupt/nmi to be injected and only then
vcpu will be unhalted.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
free_mmu_pages() should only undo what alloc_mmu_pages() does.
Free mmu pages from the generic VM destruction function, kvm_destroy_vm().
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
After discussion with Marcelo, we decided to rework device assignment framework
together. The old problems are kernel logic is unnecessary complex. So Marcelo
suggest to split it into a more elegant way:
1. Split host IRQ assign and guest IRQ assign. And userspace determine the
combination. Also discard msi2intx parameter, userspace can specific
KVM_DEV_IRQ_HOST_MSI | KVM_DEV_IRQ_GUEST_INTX in assigned_irq->flags to
enable MSI to INTx convertion.
2. Split assign IRQ and deassign IRQ. Import two new ioctls:
KVM_ASSIGN_DEV_IRQ and KVM_DEASSIGN_DEV_IRQ.
This patch also fixed the reversed _IOR vs _IOW in definition(by deprecated the
old interface).
[avi: replace homemade bitcount() by hweight_long()]
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
This patch finally enable MSI-X.
What we need for MSI-X:
1. Intercept one page in MMIO region of device. So that we can get guest desired
MSI-X table and set up the real one. Now this have been done by guest, and
transfer to kernel using ioctl KVM_SET_MSIX_NR and KVM_SET_MSIX_ENTRY.
2. Information for incoming interrupt. Now one device can have more than one
interrupt, and they are all handled by one workqueue structure. So we need to
identify them. The previous patch enable gsi_msg_pending_bitmap get this done.
3. Mapping from host IRQ to guest gsi as well as guest gsi to real MSI/MSI-X
message address/data. We used same entry number for the host and guest here, so
that it's easy to find the correlated guest gsi.
What we lack for now:
1. The PCI spec said nothing can existed with MSI-X table in the same page of
MMIO region, except pending bits. The patch ignore pending bits as the first
step (so they are always 0 - no pending).
2. The PCI spec allowed to change MSI-X table dynamically. That means, the OS
can enable MSI-X, then mask one MSI-X entry, modify it, and unmask it. The patch
didn't support this, and Linux also don't work in this way.
3. The patch didn't implement MSI-X mask all and mask single entry. I would
implement the former in driver/pci/msi.c later. And for single entry, userspace
should have reposibility to handle it.
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
We have to handle more than one interrupt with one handler for MSI-X. Avi
suggested to use a flag to indicate the pending. So here is it.
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Introduce KVM_SET_MSIX_NR and KVM_SET_MSIX_ENTRY two ioctls.
This two ioctls are used by userspace to specific guest device MSI-X entry
number and correlate MSI-X entry with GSI during the initialization stage.
MSI-X should be well initialzed before enabling.
Don't support change MSI-X entry number for now.
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Under CONFIG_MAXSMP, cpus_hardware_enabled is allocated from the heap and
not statically initialized. This causes a crash on reboot when kvm thinks
vmx is enabled on random nonexistent cpus and accesses nonexistent percpu
lists.
Fix by explicitly clearing the variable.
Cc: stable@kernel.org
Reported-and-tested-by: Yinghai Lu <yinghai@kernel.org>
Signed-off-by: Avi Kivity <avi@redhat.com>
When checking for overlapping slots on registration of a new one, kvm
currently also considers zero-length (ie. deleted) slots and rejects
requests incorrectly. This finally denies user space from joining slots.
Fix the check by skipping deleted slots and advertise this via a
KVM_CAP_JOIN_MEMORY_REGIONS_WORKS.
Cc: stable@kernel.org
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
The large page initialization code concludes there are two large pages spanned
by a slot covering 1 (small) page starting at gfn 1. This is incorrect, and
also results in incorrect write_count initialization in some cases (base = 1,
npages = 513 for example).
Cc: stable@kernel.org
Signed-off-by: Avi Kivity <avi@redhat.com>
only need to set assigned_dev_id for deassignment, use
match->flags to judge and deassign it.
Acked-by: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Weidong Han <weidong.han@intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
The function kvm_is_mmio_pfn is called before put_page is called on a
page by KVM. This is a problem when when this function is called on some
struct page which is part of a compund page. It does not test the
reserved flag of the compound page but of the struct page within the
compount page. This is a problem when KVM works with hugepages allocated
at boot time. These pages have the reserved bit set in all tail pages.
Only the flag in the compount head is cleared. KVM would not put such a
page which results in a memory leak.
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Acked-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Merge MSI userspace interface with IRQ routing table. Notice the API have been
changed, and using IRQ routing table would be the only interface kvm-userspace
supported.
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Currently KVM has a static routing from GSI numbers to interrupts (namely,
0-15 are mapped 1:1 to both PIC and IOAPIC, and 16:23 are mapped 1:1 to
the IOAPIC). This is insufficient for several reasons:
- HPET requires non 1:1 mapping for the timer interrupt
- MSIs need a new method to assign interrupt numbers and dispatch them
- ACPI APIC mode needs to be able to reassign the PCI LINK interrupts to the
ioapics
This patch implements an interrupt routing table (as a linked list, but this
can be easily changed) and a userspace interface to replace the table. The
routing table is initialized according to the current hardwired mapping.
Signed-off-by: Avi Kivity <avi@redhat.com>
Allow clients to request notifications when the guest masks or unmasks a
particular irq line. This complements irq ack notifications, as the guest
will not ack an irq line that is masked.
Currently implemented for the ioapic only.
Signed-off-by: Avi Kivity <avi@redhat.com>
MSI is always enabled by default for msi2intx=1. But if msi2intx=0, we
have to disable MSI if guest require to do so.
The patch also discard unnecessary msi2intx judgment if guest want to update
MSI state.
Notice KVM_DEV_IRQ_ASSIGN_MSI_ACTION is a mask which should cover all MSI
related operations, though we only got one for now.
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
This rips out the support for KVM_DEBUG_GUEST and introduces a new IOCTL
instead: KVM_SET_GUEST_DEBUG. The IOCTL payload consists of a generic
part, controlling the "main switch" and the single-step feature. The
arch specific part adds an x86 interface for intercepting both types of
debug exceptions separately and re-injecting them when the host was not
interested. Moveover, the foundation for guest debugging via debug
registers is layed.
To signal breakpoint events properly back to userland, an arch-specific
data block is now returned along KVM_EXIT_DEBUG. For x86, the arch block
contains the PC, the debug exception, and relevant debug registers to
tell debug events properly apart.
The availability of this new interface is signaled by
KVM_CAP_SET_GUEST_DEBUG. Empty stubs for not yet supported archs are
provided.
Note that both SVM and VTX are supported, but only the latter was tested
yet. Based on the experience with all those VTX corner case, I would be
fairly surprised if SVM will work out of the box.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
kvm->slots_lock is outer to kvm->lock, so take slots_lock
in kvm_vm_ioctl_assign_device() before taking kvm->lock,
rather than taking it in kvm_iommu_map_memslots().
Cc: stable@kernel.org
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Acked-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
In the past, kvm_get_kvm() and kvm_put_kvm() was called in assigned device irq
handler and interrupt_work, in order to prevent cancel_work_sync() in
kvm_free_assigned_irq got a illegal state when waiting for interrupt_work done.
But it's tricky and still got two problems:
1. A bug ignored two conditions that cancel_work_sync() would return true result
in a additional kvm_put_kvm().
2. If interrupt type is MSI, we would got a window between cancel_work_sync()
and free_irq(), which interrupt would be injected again...
This patch discard the reference count used for irq handler and interrupt_work,
and ensure the legal state by moving the free function at the very beginning of
kvm_destroy_vm(). And the patch fix the second bug by disable irq before
cancel_work_sync(), which may result in nested disable of irq but OK for we are
going to free it.
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
kvm_arch_sync_events is introduced to quiet down all other events may happen
contemporary with VM destroy process, like IRQ handler and work struct for
assigned device.
For kvm_arch_sync_events is called at the very beginning of kvm_destroy_vm(), so
the state of KVM here is legal and can provide a environment to quiet down other
events.
Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Gregory Haskins
