memory: hugepages: map numa node IDs to host and guest correctly
foreach_dimm() provides a guest numa node index, when used
in conjunction with the guest-to-host numa node topology
mapping one has to make sure that the correct host-side
indices are used.
This covers situations where the user defines a numaX with
hostnodes=Y with Y != X.
Both numa IDs 1 and 2 passed by foreach_dimm() have to be
mapped to host node 0.
Note that this also reverses the foreach_reverse_dimm() numa
node numbering as the current code, while walking sizes
backwards, walked the numa IDs inside each size forward,
which makes more sense. (Memory hot-unplug is still working
with this.)
this API call changes the config quite drastically, and as
such should not be possible while an operation that holds a
lock is ongoing (e.g., migration, backup, snapshot).
we have a few problems with hotplug at the moment:
qemu may add usb hubs when adding usb devices but fails to remove them
when removing the usb device (this is a qemu bug)
also when starting a guest with a usb device we add ehci and uchi
controllers, which we cannot hot unplug
with those devices, it is impossible to live migrate the guest
to another host, meaning even if you remove all usb devices,
the migrate fails
vm configuration
----------------
hugepages: (any|2|1024)
any: we'll try to allocate 1GB hugepage if possible, if not we use 2MB hugepage
2: we want to use 2MB hugepage
1024: we want to use 1GB hugepage. (memory need to be multiple of 1GB in this case)
optionnal host configuration for 1GB hugepages
----------------------------------------------
1GB hugepages can be allocated at boot if user want it.
hugepages need to be contiguous, so sometime it's not possible to reserve them on the fly
add them by default for qemu 2.6
(support is already present in qemu 2.5, but we don't want to break live migration for current running vm)
vpindex && runtime need host kernel 4.4
Theses 3 enlightements are needed by windows to use vmbus
http://searchwindowsserver.techtarget.com/definition/Microsoft-Virtual-Machine-Bus-VMBus
details :
- When Hyper-V "vpindex" is on, guest can use MSR HV_X64_MSR_VP_INDEX
to get virtual processor ID.
- Hyper-V "runtime" enlightement feature allows to use MSR
HV_X64_MSR_VP_RUNTIME to get the time the virtual processor consumes
running guest code, as well as the time the hypervisor spends running
code on behalf of that guest.
Restore previous behaviour and do not request a forward tunnel on
insecure migrations.
For the migrations of all kind this has no direct impact, they all
worked, but an port to much requested from an limited pool is still
not ideal. Also an open tunnel, if not needed.
This is a light regression introduced from commit 1c9d54b.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
Output all errors - if any - and add some log outputs on what we qmp
commands we do with which parameters, may be helpful when debugging
or analyzing a users problem.
Also check if the queried status is defined, as on a error this may
not be.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
We cannot guarantee when the SSH forward Tunnel really becomes
ready. The check with the mtunnel API call did not help for this
prolem as it only checked that the SSH connection itself works and
that the destination node has quorum but the forwarded tunnel itself
was not checked.
The Forward tunnel is a different channel in the SSH connection,
independent of the SSH `qm mtunnel` channel, so only if that works
it does not guarantees that our migration tunnel is up and ready.
When the node(s) where under load, or when we did parallel
migrations (migrateall), the migrate command was often started
before a tunnel was open and ready to receive data. This led to
a direct abortion of the migration and is the main cause in why
parallel migrations often leave two thirds or more VMs on the
source node.
The issue was tracked down to SSH after debugging the QEMU
process and enabling debug logging showed that the tunnel became
often to late available and ready, or not at all.
Fixing the TCP forward tunnel is quirky and not straight ahead, the
only way SSH gives as a possibility is to use -N (no command)
-f (background) and -o "ExitOnForwardFailure=yes", then it would
wait in the foreground until the tunnel is ready and only then
background itself. This is not quite the nicest way for our special
use case and our code base.
Waiting for the local port to become open and ready (through
/proc/net/tcp[6]] as a proof of concept is not enough, even if the
port is in the listening state and should theoretically accept
connections this still failed often as the tunnel was not yet fully
ready.
Further another problem would still be open if we tried to patch the
SSH Forward method we currently use - which we solve for free with
the approach of this patch - namely the problem that the method
to get an available port (next_migration_port) has a serious race
condition which could lead to multiple use of the same port on a
parallel migration (I observed this on my many test, seldom but if
it happens its really bad).
So lets now use UNIX sockets, which ssh supports since version 5.7.
The end points are UNIX socket bound to the VMID - thus no port so
no race and also no limitation of available ports (we reserved 50 for
migration).
The endpoints get created in /run/qemu-server/VMID.migrate and as
KVM/QEMU in current versions is able to use UNIX socket just as well
as TCP we have not to change much on the interaction with QEMU.
QEMU is started with the migrate_incoming url at the local
destination endpoint and creates the socket file, we then create
a listening socket on the source side and connect over SSH to the
destination.
Now the migration can be started by issuing the migrate qmp command
with an updated uri.
This breaks live migration from new to old, but *not* from old to
new, so there is a upgrade path.
If a live migration from new to old must be made (for whatever
reason), use the unsecure_migration setting (man datacenter.conf)
to allow this, although that should only be done in trusted network.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
With systemd-run qemu's --daemonize forks often happen
before systemd finishes setting up the scopes, which means
the limits we apply often don't work.
We now use enter_systemd_scope() to create the scope before
running qemu directly without systemd-run.
Note that vm_start() runs in a forked-worker or qm cli
command, so entering the scope in such a process should not
affect the rest of the pve daemon.
Dominik Csapak [Wed, 1 Jun 2016 09:12:42 +0000 (11:12 +0200)]
fix #1010: whitelist options for permissions
instead of defaulting to VM.Config.Options for
all options not checked seperately, we
now have lists for the different config permissions
and check them accordingly.
for everything not given, we require root access
this is important especially for usbN and hostpciN
since they can change the host configuration
otherwise, long kvm commands lead to systemd unit files with
very long lines, with confuses the systemd unit file parser.
apparently systemd has a length limit for unit file lines and
(line-)breaks the description string at that point. since
the rest of the description is probably not a valid key/value
pair, this leads to warnings. the default semantics of systemd-run
is to use the executed command as description unless a description
is specified explicitly.
note that this behaviour of systemd could allow an attacker
with access to the VM configuration to craft a kvm commandline
that starts or stops arbitrary systemd units.
fix #971: don't activate shared storage in offline migration
instead, just print a warning if the connection check fails.
as long as the storage is online on the target node, the VM
will start fine after migration.
previously, we did not check the file parameter of a disk,
allowing passthrough of a block device (by design)
with the change to the json parser for the disks, the format
became 'pve-volume-id' which is only valid for our volume ids
(and later we also allowed the value 'none')
this patch alternatively checks if the parameter is a path
or 'cdrom'
The volume_size_info() call was what actually failed, but
the error reported to the gui came from afterwards trying to
resize the disk to a garabge size.
projects / qemu-server.git / log