In order to simplify MemoryManager::new() function. let's move the
memory configuration validation to its own function.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Move the PciSegment struct and the associated code to a new file. This
will allow some clearer separation between the core DeviceManager and
PCI handling.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Move the PCI related state from the DeviceManager struct to a PciSegment
struct inside the DeviceManager. This is in preparation for multiple
segment support. Currently this state is just the bus itself, the MMIO
and PIO config devices and hotplug related state.
The main change that this required is using the Arc<Mutex<PciBus>> in
the device addition logic in order to ensure that
the bus could be created earlier.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
When using PVH for booting (which we use for all firmwares and direct
kernel boot) the Linux kernel does not configure LA57 correctly. As such
we need to limit the address space to the maximum 4-level paging address
space.
If the user knows that their guest image can take advantage of the
5-level addressing and they need it for their workload then they can
increase the physical address space appropriately.
This PR removes the TDX specific handling as the new address space limit
is below the one that that code specified.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Whenever running TDX, we must pass the ACPI tables to the TDVF firmware
running in the guest. The proper way to do this is by adding the tables
to the TdHob as a TdVmmData type, so that TDVF will know how to access
these tables and expose them to the guest OS.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Instead of having the ACPI tables being created both in x86_64 and
aarch64 implementations of configure_system(), we can remove the
duplicated code by moving the ACPI tables creation in vm.rs inside the
boot() function.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
The argument `prefault` is provided in MemoryManager, but it can
only be used by SGX and restore.
With prefault (MAP_POPULATE) been set, subsequent page faults will
decrease during running, although it will make boot slower.
This commit adds `prefault` in MemoryConfig and MemoryZoneConfig.
To resolve conflict between memory and restore, argument
`prefault` has been changed from `bool` to `Option<bool>`, when
its value is None, config from memory will be used, otherwise
argument in Option will be used.
Signed-off-by: Yu Li <liyu.yukiteru@bytedance.com>
By using a single file for storing the memory ranges, we simplify the
way snapshot/restore works by avoiding multiples files, but the main and
more important point is that we have now a way to save only the ranges
that matter. In particular, the ranges related to virtio-mem regions are
not always fully hotplugged, meaning we don't want to save the entire
region. That's where the usage of memory ranges is interesting as it
lets us optimize the snapshot/restore process when one or multiple
virtio-mem regions are involved.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
The function memory_range_table() will be reused by the MemoryManager in
a following patch to describe all the ranges that we should snapshot.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Copy only the memory ranges that have been plugged through virtio-mem,
allowing for an interesting optimization regarding the time it takes to
migrate a large virtio-mem device. Even if the hotpluggable space is
very large (say 64GiB), if only 1GiB has been previously added to the
VM, only 1GiB will be sent to the destination VM, avoiding the transfer
of the remaining 63GiB which are unused.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
By creating the BlocksState object in the MemoryManager, we can directly
provide it to the virtio-mem device when being created. This will allow
the MemoryManager through each VirtioMemZone to have a handle onto the
blocks that are plugged at any point in time.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
This will be helpful to support the creation of a MemoryRangeTable from
virtio-mem, as it uses 2M pages.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Adding the snapshot/restore support along with migration as well,
allowing a VM with virtio-mem devices attached to be properly
migrated.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
The amount of memory plugged in the virtio-mem region should always be
kept up to date in the hotplugged_size field from VirtioMemZone.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
There's no need to duplicate the GuestMemory for snapshot purpose, as we
always have a handle onto the GuestMemory through the guest_memory
field.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Since we only support a single PCI bus right now advertise only a single
bus in the ACPI tables. This reduces the number of VM exits from probing
substantially.
Number of PCI config I/O port exits: 17871 -> 1551 (91% reduction) with
direct kernel boot.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Use a simpler method for extracting the affected slot on the eject
command. Also update the terminology to reflect that this a slot rather
than a bdf (which is what device id refers to elsewhere.)
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Refactor the serial buffer handling in order to write the serial
buffer's output to a PTY connected after the serial device stops being
written to by the guest.
This change moves the serial buffer initialization inside the serial
manager. That is done to allow the serial buffer to be made aware of
the PTY and epoll fds needed in order to modify the
EpollDispatch::File trigger. These are then used by the serial buffer
to trigger an epoll event when the PTY fd is writable and the buffer
has content in it. They are also used to remove the trigger when the
buffer is emptied in order to avoid unnecessary wake-ups.
Signed-off-by: William Douglas <william.douglas@intel.com>
Both read_exact_from() and write_all_to() functions from the GuestMemory
trait implementation in vm-memory are buggy. They should retry until
they wrote or read the amount of data that was expected, but instead
they simply return an error when this happens. This causes the migration
to fail when trying to send important amount of data through the
migration socket, due to large memory regions.
This should be eventually fixed in vm-memory, and here is the link to
follow up on the issue: https://github.com/rust-vmm/vm-memory/issues/174
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Implement the infrastructure that lets a virtio-mem device map the guest
memory into the device. This is necessary since with virtio-mem zones
memory can be added or removed and the vfio-user device must be
informed.
Fixes: #3025
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
By moving this from the VfioUserPciDevice to DeviceManager the client
can be reused for handling DMA mapping behind an IOMMU.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
For vfio-user the mapping handler is per device and needs to be removed
when the device in unplugged.
For VFIO the mapping handler is for the default VFIO container (used
when no vIOMMU is used - using a vIOMMU does not require mappings with
virtio-mem)
To represent these two use cases use an enum for the handlers that are
stored.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Looking up devices on the port I/O bus is time consuming during the
boot at there is an O(lg n) tree lookup and the overhead from taking a
lock on the bus contents.
Avoid this by adding a fast path uses the hardcoded port address and
size and directs PCI config requests directly to the device.
Command line:
target/release/cloud-hypervisor --kernel ~/src/linux/vmlinux --cmdline "root=/dev/vda1 console=ttyS0" --serial tty --console off --disk path=~/workloads/focal-server-cloudimg-amd64-custom-20210609-0.raw --api-socket /tmp/api
PIO exit: 17913
PCI fast path: 17871
Percentage on fast path: 99.8%
perf before:
marvin:~/src/cloud-hypervisor (main *)$ perf report -g | grep resolve
6.20% 6.20% vcpu0 cloud-hypervisor [.] vm_device:🚌:Bus::resolve
perf after:
marvin:~/src/cloud-hypervisor (2021-09-17-ioapic-fast-path *)$ perf report -g | grep resolve
0.08% 0.08% vcpu0 cloud-hypervisor [.] vm_device:🚌:Bus::resolve
The compromise required to implement this fast path is bringing the
creation of the PciConfigIo device into the DeviceManager::new() so that
it can be used in the VmmOps struct which is created before
DeviceManager::create_devices() is called.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
The MSI IOVA address on X86 and AArch64 is different.
This commit refactored the code to receive the MSI IOVA address and size
from device_manager, which provides the actual IOVA space data for both
architectures.
Signed-off-by: Michael Zhao <michael.zhao@arm.com>
Add a virtio-iommu node into FDT if iommu option is turned on. Now we
support only one virtio-iommu device.
Signed-off-by: Michael Zhao <michael.zhao@arm.com>
This change switches from handling serial input in the VMM thread to
its own thread controlled by the SerialManager.
The motivation for this change is to avoid the VMM thread being unable
to process events while serial input is happening and vice versa.
The change also makes future work flushing the serial buffer on PTY
connections easier.
Signed-off-by: William Douglas <william.douglas@intel.com>
This change adds a SerialManager with its own epoll handling that
should be created and run by the DeviceManager when creating an
appropriately configured console (serial tty or pty).
Both stdin and pty input are handled by the SerialManager. The stdin
and pty specific methods used by the VMM should be removed in a future
commit.
Signed-off-by: William Douglas <william.douglas@intel.com>
The clone method for PtyPair should have been an impl of the Clone
trait but the method ended up not being used. Future work will make
use of the trait however so correct the missing trait implementation.
Signed-off-by: William Douglas <william.douglas@intel.com>
For most use cases, there is no need to create multiple VFIO containers
as it causes unwanted behaviors. Especially when passing multiple
devices from the same IOMMU group, we need to use the same container so
that it can properly list the groups that have been already opened. The
correct logic was already there in vfio-ioctls, but it was incorrectly
used from our VMM implementation.
For the special case where we put a VFIO device behind a vIOMMU, we must
create one container per device, as we need to control the DMA mappings
per device, which is performed at the container level. Because we must
keep one container per device, the vIOMMU use case prevents multiple
devices attached to the same IOMMU group to be passed through the VM.
But this is a limitation that we are fine with, especially since the
vIOMMU doesn't let us group multiple devices in the same group from a
guest perspective.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
When a pty is resized (using the TIOCSWINSZ ioctl -- see ioctl_tty(2)),
the kernel will send a SIGWINCH signal to the pty's foreground process
group to notify it of the resize. This is the only way to be notified
by the kernel of a pty resize.
We can't just make the cloud-hypervisor process's process group the
foreground process group though, because a process can only set the
foreground process group of its controlling terminal, and
cloud-hypervisor's controlling terminal will often be the terminal the
user is running it in. To work around this, we fork a subprocess in a
new process group, and set its process group to be the foreground
process group of the pty. The subprocess additionally must be running
in a new session so that it can have a different controlling
terminal. This subprocess writes a byte to a pipe every time the pty
is resized, and the virtio-console device can listen for this in its
epoll loop.
Alternatives I considered were to have the subprocess just send
SIGWINCH to its parent, and to use an eventfd instead of a pipe.
I decided against the signal approach because re-purposing a signal
that has a very specific meaning (even if this use was only slightly
different to its normal meaning) felt unclean, and because it would
have required using pidfds to avoid race conditions if
cloud-hypervisor had terminated, which added complexity. I decided
against using an eventfd because using a pipe instead allows the child
to be notified (via poll(2)) when nothing is reading from the pipe any
more, meaning it can be reliably notified of parent death and
terminate itself immediately.
I used clone3(2) instead of fork(2) because without
CLONE_CLEAR_SIGHAND the subprocess would inherit signal-hook's signal
handlers, and there's no other straightforward way to restore all signal
handlers to their defaults in the child process. The only way to do
it would be to iterate through all possible signals, or maintain a
global list of monitored signals ourselves (vmm:vm::HANDLED_SIGNALS is
insufficient because it doesn't take into account e.g. the SIGSYS
signal handler that catches seccomp violations).
Signed-off-by: Alyssa Ross <hi@alyssa.is>
This prepares us to be able to handle console resizes in the console
device's epoll loop, which we'll have to do if the output is a pty,
since we won't get SIGWINCH from it.
Signed-off-by: Alyssa Ross <hi@alyssa.is>
Musl often uses mmap to allocate memory where Glibc would use brk.
This has caused seccomp violations for me on the API and signal
handling threads.
Signed-off-by: Alyssa Ross <hi@alyssa.is>
error: all if blocks contain the same code at the end
--> vmm/src/memory_manager.rs:884:9
|
884 | / Ok(mm)
885 | | }
| |_________^
Signed-off-by: Bo Chen <chen.bo@intel.com>
This concept ends up being broken with multiple types on input connected
e.g. console on TTY and serial on PTY. Already the code for checking for
injecting into the serial device checks that the serial is configured.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>