The NVIDIA V100 GPU has an onboard RAM that is mapped into the
host memory and accessible as normal RAM via an NVLink2 bridge. When
passed through in a guest, QEMU puts the NVIDIA RAM window in a
non-contiguous area, above the PCI MMIO area that starts at 32TiB.
This means that the NVIDIA RAM window starts at 64TiB and go all the
way to 128TiB.
This means that the guest might request a 64-bit window, for each PCI
Host Bridge, that goes all the way to 128TiB. However, the NVIDIA RAM
window isn't counted as regular RAM, thus this window is considered
only for the allocation of the Translation and Control Entry (TCE).
For more information about how NVLink2 support works in QEMU,
refer to the accepted implementation [1].
This memory layout differs from the existing VFIO case, requiring its
own formula. This patch changes the PPC64 code of
@qemuDomainGetMemLockLimitBytes to:
- detect if we have a NVLink2 bridge being passed through to the
guest. This is done by using the @ppc64VFIODeviceIsNV2Bridge function
added in the previous patch. The existence of the NVLink2 bridge in
the guest means that we are dealing with the NVLink2 memory layout;
- if an IBM NVLink2 bridge exists, passthroughLimit is calculated in a
different way to account for the extra memory the TCE table can alloc.
The 64TiB..128TiB window is more than enough to fit all possible
GPUs, thus the memLimit is the same regardless of passing through 1 or
multiple V100 GPUs.
Further reading explaining the background
[1] https://lists.gnu.org/archive/html/qemu-devel/2019-03/msg03700.html
[2] https://www.redhat.com/archives/libvir-list/2019-March/msg00660.html
[3] https://www.redhat.com/archives/libvir-list/2019-April/msg00527.html
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Reviewed-by: Erik Skultety <eskultet@redhat.com>
The NVLink2 support in QEMU implements the detection of NVLink2
capable devices by verifying the attributes of the VFIO mem region
QEMU allocates for the NVIDIA GPUs. To properly allocate an
adequate amount of memLock, Libvirt needs this information before
a QEMU instance is even created, thus querying QEMU is not
possible and opening a VFIO window is too much.
An alternative is presented in this patch. Making the following
assumptions:
- if we want GPU RAM to be available in the guest, an NVLink2 bridge
must be passed through;
- an unknown PCI device can be classified as a NVLink2 bridge
if its device tree node has 'ibm,gpu', 'ibm,nvlink',
'ibm,nvlink-speed' and 'memory-region'.
This patch introduces a helper called @ppc64VFIODeviceIsNV2Bridge
that checks the device tree node of a given PCI device and
check if it meets the criteria to be a NVLink2 bridge. This
new function will be used in a follow-up patch that, using the
first assumption, will set up the rlimits of the guest
accordingly.
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
This does not cause a problem in usual scenarios thanks to us allowing
CAP_DAC_OVERRIDE for the qemu process, however in some scenarios this might be
an issue because the directory is created with mkdtemp(3) which explicitly
creates that with 0700 permissions and qemu running as non-root cannot access
that.
The scenarios include:
- Builds without CAPNG
- Running libvirtd in certain container configurations [1]
- and possibly others.
[1] https://github.com/kubevirt/kubevirt/pull/2181#issuecomment-481840304
Signed-off-by: Martin Kletzander <mkletzan@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
The new virHostCPUGetMSR internal API will try to read the MSR from
/dev/cpu/0/msr and if it is not possible (the device does not exist or
libvirt is running unprivileged), it will fallback to asking KVM for the
MSR using KVM_GET_MSRS ioctl.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
This patch adds an inline python code for reading MSR features. Since
reading MSRs is a privileged operation, we have to read them from
/dev/cpu/*/msr if it is readable (i.e., the script runs as root) or
fallback to using KVM ioctl which can be done by any user that can start
virtual machines.
The python code is inlined rather than provided in a separate script
because whenever there's an issue with proper detection of CPU features,
we ask the reporter to run cpu-gather.sh script to give us all data we
need to know about the host CPU. Asking them to run several scripts
would likely result in one of them being ignored or forgotten.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
The parseMapFeature for parsing features from CPU map XML can be easily
generalized to support more feature types.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Let's make sure the current CPUID specific code is only applied to CPUID
features.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
This will let us simplify the code since the dictionary keys will match
attribute names in various XMLs.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
leaf["eax"] & eax > 0 check works correctly only if there's at most 1
bit set in eax. Luckily that's been always the case, but fixing this
could save us from future surprises.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
The function will have to deal with both CPUID and MSR features.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
We don't really need to parse CPU data from QEMU older than 2.9 (i.e.,
before query-cpu-model-expansion) at this point. But even if there's a
need to do so, we can always use an older version of this script to do
the conversion.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
They are static and we will need to call them a little bit closer to the
beginning of the file.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
The structure can only be used for CPUID data now. Adding a type
indicator and moving the data into a union will let us store alternative
data types.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
The function now works on virCPUx86DataItem and it's called
virCPUx86DataItemMatch.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
The function is renamed as virCPUx86DataItemMatchMasked to reflect the
change in parameter types.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
The function now works on virCPUx86DataItem and it's renamed as
virCPUx86DataItemAndBits.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
The parameters changed from virCPUx86CPUID to virCPUx86DataItem and the
function is now called virCPUx86DataItemClearBits.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
The function is renamed as virCPUx86DataItemSetBits and it works on
virCPUx86DataItem now.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
virCPUx86DataSorter already compares two virCPUx86DataItem structs.
Let's add a tiny wrapper around it called virCPUx86DataCmp and use it
instead of open coded comparisons.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
It is called virCPUx86DataSorter since the function will work on any CPU
data type.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
The function is now called virCPUx86DataNext to reflect its purpose: it
is an iterator over CPU data (both CPUID and MSR in the near future).
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Although vendor string is always reported by CPUID, the container struct
is used for consistency and thus "cpuid" name is not a good fit anymore.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
The following patches introduce CPU features read from MSR in addition
to those queried via CPUID instruction. Let's introduce a container
struct which will be able to describe either feature type.
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Introduced in QEMU 3.1.0 by commit
c7a88b52f62b30c04158eeb07f73e3f72221b6a8
Signed-off-by: Jiri Denemark <jdenemar@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Vim has trouble figuring out the filetype automatically because
the name doesn't follow existing conventions; annotations like
the ones we already have in Makefile.ci help it out.
Signed-off-by: Andrea Bolognani <abologna@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Unfortunately the data reported by pkg-config is not completely
accurate, so until the issue has been fixed in readline we need
to work around it in libvirt.
The good news is that we only need the fix to land in FreeBSD
ports and macOS homebrew before we can drop the kludge, so
we're talking months rather than years.
Signed-off-by: Andrea Bolognani <abologna@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
With the 7.0 release, readline has finally started shipping
pkg-config support in the form of a readline.pc file.
Unfortunately, most downstreams have yet to catch up with this
change: among Linux distributions in particular, Fedora Rawhide
seems to be the only one installing it at the moment.
Non-Linux operating systems have been faring much better in
this regard: both FreeBSD (through ports) and macOS (through
homebrew) include pkg-config support in their readline package.
This is great news for us, since those are the platforms where
pkg-config is more useful on account of them installing headers
and libraries outside of the respective default search paths.
Our implementation checks whether readline is registered as a
pkg-config package, and if so obtains CFLAGS and LIBS using the
tool; if not, we just keep using the existing logic.
This commit is best viewed with 'git show -w'.
Signed-off-by: Andrea Bolognani <abologna@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
The first implementation of this logic was introduced with
commit 2ec759fc58 all the way back in 2007; looking at the
build logs from our CI environment, however, it's apparent
that none of the platforms we currently target are actually
using it, so we can assume whatever issue it was working
around has been fixed at some point in the last 12 years.
Signed-off-by: Andrea Bolognani <abologna@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
The current code is a bit awkward, and we're going to need
to share it later anyway. We can drop the call to AC_SUBST()
while we're at it, since LIBVIRT_CHECK_LIB() already marks
READLINE_CFLAGS for substitution.
The new code goes to some extra length to avoid setting
-D_FUNCTION_DEF twice: this is mostly for cosmetic reasons,
and it's necessary because LIBVIRT_CHECK_READLINE() is called
twice: once on its own, and then once more as part of
LIBVIRT_CHECK_BASH_COMPLETION().
Signed-off-by: Andrea Bolognani <abologna@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
The check was added in 74416b1d48 without offering any
explanation outside of the commit message. Introduce a comment
to make digging through the git history unnecessary.
Signed-off-by: Andrea Bolognani <abologna@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
This code is needed to use readline older than 4.1, but all
our target platforms ship with at least 6.0 these days so we
can safely get rid of it.
Signed-off-by: Andrea Bolognani <abologna@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Firstly, virCommandRun() does report an error on failure (which
in most cases is more accurate than what we overwrite it with).
Secondly, usually errno is not set (or gets overwritten in the
cleanup code) which makes virReportSystemError() report useless
error messages. Drop all virReportSystemError() calls in cases
like this (I've found three occurrences).
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
The latter is deprecated and will be removed soon. The advised
replacement is '-overcommit mem-lock=on|off'.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Added in QEMU commit of v3.0.0-rc0~48^2~9 (then fixed by
v3.1.0-rc0~119^2~37) QEMU is replacing '-realtime mlock' with
'-overcommit mem-lock'. Add a capability to tell if we're dealing
new new enough qemu to use the replacement.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
The '-realtime mlock' cmd line argument was introduced in QEMU
commit v1.5.0-rc0~190 which matches minimal QEMU version we
require. Therefore, the capability will always be present.
Apparently, nearly none of our xml2argv test cases had the
capability hence slightly bigger change under qemuxml2argvdata/.
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Ján Tomko <jtomko@redhat.com>
Test the memory locking command line with different QEMU versions
to prepare for changing it for latest QEMU.
Signed-off-by: Ján Tomko <jtomko@redhat.com>
