pci_bus_is_root() currently relies on a method in the PCIBusClass.
But it's always known if a PCI bus is a root bus when we create it, so
using a dynamic method is overkill.
This replaces it with an IS_ROOT bit in a new flags field, which is set on
root buses and otherwise clear. As a bonus this removes the special
is_root logic from pci_expander_bridge, since it already creates its bus
as a root bus.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Marcel Apfelbaum <marcel@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Message-Id: <20190424041959.4087-3-david@gibson.dropbear.id.au>
These functions have an explicit test for accesses above the device's
config size. But pci_host_config_{read,write}_common() which they're
about to call already have checks against the config space limit and
do the right thing. So, remove the redundant tests.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <20190424041959.4087-2-david@gibson.dropbear.id.au>
qemu_getrampagesize() works out the minimum host page size backing any of
guest RAM. This is required in a few places, such as for POWER8 PAPR KVM
guests, because limitations of the hardware virtualization mean the guest
can't use pagesizes larger than the host pages backing its memory.
However, it currently checks against *every* memory backend, whether or not
it is actually mapped into guest memory at the moment. This is incorrect.
This can cause a problem attempting to add memory to a POWER8 pseries KVM
guest which is configured to allow hugepages in the guest (e.g.
-machine cap-hpt-max-page-size=16m). If you attempt to add non-hugepage,
you can (correctly) create a memory backend, however it (correctly) will
throw an error when you attempt to map that memory into the guest by
'device_add'ing a pc-dimm.
What's not correct is that if you then reset the guest a startup check
against qemu_getrampagesize() will cause a fatal error because of the new
memory object, even though it's not mapped into the guest.
This patch corrects the problem by adjusting find_max_supported_pagesize()
(called from qemu_getrampagesize() via object_child_foreach) to exclude
non-mapped memory backends.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Igor Mammedov <imammedo@redhat.com>
Acked-by: David Hildenbrand <david@redhat.com>
27461d69a0 "ppc: add host-serial and host-model machine attributes
(CVE-2019-8934)" introduced 'host-serial' and 'host-model' machine
properties for spapr to explicitly control the values advertised to the
guest in device tree properties with the same names.
The previous behaviour on KVM was to unconditionally populate the device
tree with the real host serial number and model, which leaks possibly
sensitive information about the host to the guest.
To maintain compatibility for old machine types, we allowed those props
to be set to "passthrough" to take the value from the host as before. Or
they could be set to "none" to explicitly omit the device tree items.
Special casing specific values on what's otherwise a user supplied string
is very ugly. So, this patch simplifies things by implementing the
backwards compatibility in a different way: we have a machine class flag
set for the older machines, and we only load the host values into the
device tree if A) they're not set by the user and B) we have that flag set.
This does mean that the "passthrough" functionality is no longer available
with the current machine type. That's ok though: if a user or management
layer really wants the information passed through they can read it
themselves (OpenStack Nova already does something similar for x86).
It also means the user can't explicitly ask for the values to be omitted
on the old machine types. I think that's an acceptable trade-off: if you
care enough about not leaking the host information you can either move to
the new machine type, or use a dummy value for the properties.
For the new machine type, this also removes an odd inconsistency
between running on a POWER and non-POWER (or non-Linux) hosts: if the
host information couldn't be read from where we expect (in the host's
device tree as exposed by Linux), we'd fallback to omitting the guest
device tree items.
While we're there, improve some poorly worded comments, and the help text
for the properties.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Greg Kurz <groug@kaod.org>
176dccee "target/ppc/spapr: Clear partition table entry when allocating
hash table" reworked the H_REGISTER_PROCESS_TABLE hypercall, but
unfortunately due to a small error no longer correctly sets the LPCR[GTSE]
bit which allows the guest to directly execute (some types of) tlbie (TLB
flush) instructions without involving the hypervisor.
We got away with this, initially, because POWER9 did not have hypervisor
mode enabled in its msr_mask, which meant we didn't actually run hypervisor
privilege checks in TCG at all. However, da874d90 "target/ppc: add HV
support for POWER9" turned on HV support on POWER9 for the benefit of the
powernv machine type.
This exposed the earlier bug in H_REGISTER_PROCESS_TABLE, and causes guests
which rely on LPCR[GTSE] (i.e. basically all of them) to crash during early
boot when their first tlbie instruction causes an unexpected trap.
Fixes: 176dccee target/ppc/spapr: Clear partition table entry when allocating hash table
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Cleber Rosa <crosa@redhat.com>
Prior to f6deb6d9 "virtio-balloon: Remove unnecessary MADV_WILLNEED on
deflate", the balloon device issued an madvise() MADV_WILLNEED on
pages removed from the balloon. That would hint to the host kernel
that the pages were likely to be needed by the guest in the near
future.
It's unclear if this is actually valuable or not, and so f6deb6d9
removed this, essentially ignoring balloon deflate requests. However,
concerns have been raised that this might cause a performance
regression by causing extra latency for the guest in certain
configurations.
So, until we can get actual benchmark data to see if that's the case,
this restores the old behaviour, issuing a MADV_WILLNEED when a page is
removed from the balloon.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <20190306030601.21986-4-david@gibson.dropbear.id.au>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This fixes a balloon bug with a nasty consequence - potentially
corrupting guest memory - but which is extremely unlikely to be
triggered in practice.
The balloon always works in 4kiB units, but the host could have a
larger page size on certain platforms. Since ed48c59 "virtio-balloon:
Safely handle BALLOON_PAGE_SIZE < host page size" we've handled this
by accumulating requests to balloon 4kiB subpages until they formed a
full host page. Since f6deb6d "virtio-balloon: Remove unnecessary
MADV_WILLNEED on deflate" we essentially ignore deflate requests.
Suppose we have a host with 8kiB pages, and one host page has subpages
A & B. If we get this sequence of events -
inflate A
deflate A
inflate B
- the current logic will discard the whole host page. That's
incorrect because the guest has deflated subpage A, and could have
written important data to it.
This patch fixes the problem by adjusting our state information about
partially ballooned host pages when deflate requests are received.
Fixes: ed48c59 "virtio-balloon: Safely handle BALLOON_PAGE_SIZE < host page size"
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <20190306030601.21986-3-david@gibson.dropbear.id.au>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: David Hildenbrand <david@redhat.com>
ed48c59875 "virtio-balloon: Safely handle BALLOON_PAGE_SIZE < host
page size" introduced a new temporary data structure which tracks 4kiB
chunks which have been inserted into the balloon by the guest but
don't yet form a full host page which we can discard.
Unfortunately, I had a thinko and allocated that structure with
g_malloc0() but freed it with a plain free() rather than g_free().
This corrects the problem.
Fixes: ed48c59875
Reported-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <20190306030601.21986-2-david@gibson.dropbear.id.au>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
The new pseries-4.0 machine type defaults to enabling Spectre/Meltdown
mitigations. Unfortunately those mitigations aren't implemented for TCG
because we're not yet sure if they're necessary or how to implement them.
We don't fail fatally, but we do warn in this case, because it is quite
plausible that Spectre/Meltdown can be exploited through TCG (at least for
the guest to get access to the qemu address space).
This create noise in our testcases though. So, modify the affected tests
to explicitly disable the mitigations to suppress these warnings.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The qemu coding standard is to use CamelCase for type and structure names,
and the pseries code follows that... sort of. There are quite a lot of
places where we bend the rules in order to preserve the capitalization of
internal acronyms like "PHB", "TCE", "DIMM" and most commonly "sPAPR".
That was a bad idea - it frequently leads to names ending up with hard to
read clusters of capital letters, and means they don't catch the eye as
type identifiers, which is kind of the point of the CamelCase convention in
the first place.
In short, keeping type identifiers look like CamelCase is more important
than preserving standard capitalization of internal "words". So, this
patch renames a heap of spapr internal type names to a more standard
CamelCase.
In addition to case changes, we also make some other identifier renames:
VIOsPAPR* -> SpaprVio*
The reverse word ordering was only ever used to mitigate the capital
cluster, so revert to the natural ordering.
VIOsPAPRVTYDevice -> SpaprVioVty
VIOsPAPRVLANDevice -> SpaprVioVlan
Brevity, since the "Device" didn't add useful information
sPAPRDRConnector -> SpaprDrc
sPAPRDRConnectorClass -> SpaprDrcClass
Brevity, and makes it clearer this is the same thing as a "DRC"
mentioned in many other places in the code
This is 100% a mechanical search-and-replace patch. It will, however,
conflict with essentially any and all outstanding patches touching the
spapr code.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
SPAPR_MEMORY_BLOCK_SIZE is logically a difference in memory addresses, and
hence of type hwaddr which is 64-bit. Previously it wasn't marked as such
which means that it could be treated as 32-bit. That will work in some
circumstances but if multiplied by another 32-bit value it could lead to
a 32-bit overflow and an incorrect result.
One specific instance of this in spapr_lmb_dt_populate() was spotted by
Coverity (CID 1399145).
Reported-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The virtio-balloon always works in units of 4kiB (BALLOON_PAGE_SIZE), but
we can only actually discard memory in units of the host page size.
Now, we handle this very badly: we silently ignore balloon requests that
aren't host page aligned, and for requests that are host page aligned we
discard the entire host page. The latter can corrupt guest memory if its
page size is smaller than the host's.
The obvious choice would be to disable the balloon if the host page size is
not 4kiB. However, that would break the special case where host and guest
have the same page size, but that's larger than 4kiB. That case currently
works by accident[1] - and is used in practice on many production POWER
systems where 64kiB has long been the Linux default page size on both host
and guest.
To make the balloon safe, without breaking that useful special case, we
need to accumulate 4kiB balloon requests until we have a whole contiguous
host page to discard.
We could in principle do that across all guest memory, but it would require
a large bitmap to track. This patch represents a compromise: we track
ballooned subpages for a single contiguous host page at a time. This means
that if the guest discards all 4kiB chunks of a host page in succession,
we will discard it. This is the expected behaviour in the (host page) ==
(guest page) != 4kiB case we want to support.
If the guest scatters 4kiB requests across different host pages, we don't
discard anything, and issue a warning. Not ideal, but at least we don't
corrupt guest memory as the previous version could.
Warning reporting is kind of a compromise here. Determining whether we're
in a problematic state at realize() time is tricky, because we'd have to
look at the host pagesizes of all memory backends, but we can't really know
if some of those backends could be for special purpose memory that's not
subject to ballooning.
Reporting only when the guest tries to balloon a partial page also isn't
great because if the guest page size happens to line up it won't indicate
that we're in a non ideal situation. It could also cause alarming repeated
warnings whenever a migration is attempted.
So, what we do is warn the first time the guest attempts balloon a partial
host page, whether or not it will end up ballooning the rest of the page
immediately afterwards.
[1] Because when the guest attempts to balloon a page, it will submit
requests for each 4kiB subpage. Most will be ignored, but the one
which happens to be host page aligned will discard the whole lot.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Message-Id: <20190214043916.22128-6-david@gibson.dropbear.id.au>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Currently, virtio-balloon uses madvise() with MADV_DONTNEED to actually
discard RAM pages inserted into the balloon. This is basically a Linux
only interface (MADV_DONTNEED exists on some other platforms, but doesn't
always have the same semantics). It also doesn't work on hugepages and has
some other limitations.
It turns out that postcopy also needs to discard chunks of memory, and uses
a better interface for it: ram_block_discard_range(). It doesn't cover
every case, but it covers more than going direct to madvise() and this
gives us a single place to update for more possibilities in future.
There are some subtleties here to maintain the current balloon behaviour:
* For now, we just ignore requests to balloon in a hugepage backed region.
That matches current behaviour, because MADV_DONTNEED on a hugepage would
simply fail, and we ignore the error.
* If host page size is > BALLOON_PAGE_SIZE we can frequently call this on
non-host-page-aligned addresses. These would also fail in madvise(),
which we then ignored. ram_block_discard_range() error_report()s calls
on unaligned addresses, so we explicitly check that case to avoid
spamming the logs.
* We now call ram_block_discard_range() with the *host* page size, whereas
we previously called madvise() with BALLOON_PAGE_SIZE. Surprisingly,
this also matches existing behaviour. Although the kernel fails madvise
on unaligned addresses, it will round unaligned sizes *up* to the host
page size. Yes, this means that if BALLOON_PAGE_SIZE < guest page size
we can incorrectly discard more memory than the guest asked us to. I'm
planning to address that soon.
Errors other than the ones discussed above, will now be reported by
ram_block_discard_range(), rather than silently ignored, which means we
have a much better chance of seeing when something is going wrong.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Message-Id: <20190214043916.22128-5-david@gibson.dropbear.id.au>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
The virtio-balloon device's verification of the address given to it by the
guest has a number of faults:
* The addresses here are guest physical addresses, which should be
'hwaddr' rather than 'ram_addr_t' (the distinction is admittedly
pretty subtle and confusing)
* We don't check for section.mr being NULL, which is the main way that
memory_region_find() reports basic failures. We really need to check
that before looking at any other section fields, because
memory_region_find() doesn't initialize them on the failure path
* We're passing a length of '1' to memory_region_find(), but really the
guest is requesting that we put the entire page into the balloon,
so it makes more sense to call it with BALLOON_PAGE_SIZE
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Message-Id: <20190214043916.22128-3-david@gibson.dropbear.id.au>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
When the balloon is inflated, we discard memory place in it using madvise()
with MADV_DONTNEED. And when we deflate it we use MADV_WILLNEED, which
sounds like it makes sense but is actually unnecessary.
The misleadingly named MADV_DONTNEED just discards the memory in question,
it doesn't set any persistent state on it in-kernel; all that's necessary
to bring the memory back is to touch it. MADV_WILLNEED in contrast
specifically says that the memory will be used soon and faults it in.
This patch simplify's the balloon operation by dropping the madvise()
on deflate. This might have an impact on performance - it will move a
delay at deflate time until that memory is actually touched, which
might be more latency sensitive. However:
* Memory that's being given back to the guest by deflating the
balloon *might* be used soon, but it equally could just sit around
in the guest's pools until needed (or even be faulted out again if
the host is under memory pressure).
* Usually, the timescale over which you'll be adjusting the balloon
is long enough that a few extra faults after deflation aren't
going to make a difference.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Message-Id: <20190214043916.22128-2-david@gibson.dropbear.id.au>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
The (only) obvious use for these macros is constructing and parsing guest
visible register fields. But the way they're constructed, they're only
valid when used on a *host* long, whose size shouldn't be visible to the
guest at all.
They also have no current users, so just get rid of them.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
dtc v1.4.7 contains a bunch of improvements to make libfdt safer against
handling a corrupted or malicious tree, which is a good thing to have. It
also includes an explicit fdt checking function that we'll be wanting in
future.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
This function was introduced between v2.11 and v2.12 to replace obsolete
ways of specifying the NUMA nodes for DIMMs. It's used to find the correct
node for an LMB, by locating which DIMM object it lies within.
Unfortunately, one of the checks is inverted, so we check whether the
address is less than two different things, rather than actually checking
a range. This introduced a regression, meaning that after a reboot qemu
will advertise incorrect node information for memory to the guest.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Igor Mammedov <imammedo@redhat.com>
sam460ex_load_device_tree() handles nearly all possible errors by simply
exiting (within helper functions and macros). It handles two early error
cases by returning an error.
There's no particular point to this, so make it handle those directly as
well, removing the need for the caller to handle a failure. As a bonus it
gives us more specific error messages.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The qemu_fdt_*() helper functions already exit with a message instead of
returning errors, so we don't need to check for errors in the caller.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
In a couple of places sam460ex_load_device_tree() calls "raw" libfdt
functions which can fail, but doesn't check for error codes. At best,
if these fail the guest will be silently started in a non-standard state,
or it could fail entirely.
Fix this by using the _FDT() helper macro which aborts on a libfdt failure.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
At present the Sam460ex board is activated by the general CONFIG_PPC4XX
option. However that includes the board for both ppc-softmmu and
(deprecated) ppcemb-softmmu builds. As Sam460ex is developed, that would
require adding more things into ppcemb-softmmu, which we don't want to do.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Currently during KVM initialization on POWER, kvm_fixup_page_sizes()
rewrites a bunch of information in the cpu state to reflect the
capabilities of the host MMU and KVM. This overwrites the information
that's already there reflecting how the TCG implementation of the MMU will
operate.
This means that we can get guest-visibly different behaviour between KVM
and TCG (and between different KVM implementations). That's bad. It also
prevents migration between KVM and TCG.
The pseries machine type now has filtering of the pagesizes it allows the
guest to use which means it can present a consistent model of the MMU
across all accelerators.
So, we can now replace kvm_fixup_page_sizes() with kvm_check_mmu() which
merely verifies that the expected cpu model can be faithfully handled by
KVM, rather than updating the cpu model to match KVM.
We call kvm_check_mmu() from the spapr cpu reset code. This is a hack:
conceptually it makes more sense where fixup_page_sizes() was - in the KVM
cpu init path. However, doing that would require moving the platform's
pagesize filtering much earlier, which would require a lot of work making
further adjustments. There wouldn't be a lot of concrete point to doing
that, since the only KVM implementation which has the awkward MMU
restrictions is KVM HV, which can only work with an spapr guest anyway.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
KVM HV has some limitations (deriving from the hardware) that mean not all
host-cpu supported pagesizes may be usable in the guest. At present this
means that KVM guests and TCG guests may see different available page sizes
even if they notionally have the same vcpu model. This is confusing and
also prevents migration between TCG and KVM.
This patch makes the environment consistent by always allowing the same set
of pagesizes. Since we can't remove the KVM limitations, we do this by
always applying the same limitations it has, even to TCG guests.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
The paravirtualized PAPR platform sometimes needs to restrict the guest to
using only some of the page sizes actually supported by the host's MMU.
At the moment this is handled in KVM specific code, but for consistency we
want to apply the same limitations to all accelerators.
This makes a start on this by providing a helper function in the cpu code
to allow platform code to remove some of the cpu's page size definitions
via a caller supplied callback.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
The way we used to handle KVM allowable guest pagesizes for PAPR guests
required some convoluted checking of memory attached to the guest.
The allowable pagesizes advertised to the guest cpus depended on the memory
which was attached at boot, but then we needed to ensure that any memory
later hotplugged didn't change which pagesizes were allowed.
Now that we have an explicit machine option to control the allowable
maximum pagesize we can simplify this. We just check all memory backends
against that declared pagesize. We check base and cold-plugged memory at
reset time, and hotplugged memory at pre_plug() time.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
The way the POWER Hash Page Table (HPT) MMU is virtualized by KVM HV means
that every page that the guest puts in the pagetables must be truly
physically contiguous, not just GPA-contiguous. In effect this means that
an HPT guest can't use any pagesizes greater than the host page size used
to back its memory.
At present we handle this by changing what we advertise to the guest based
on the backing pagesizes. This is pretty bad, because it means the guest
sees a different environment depending on what should be host configuration
details.
As a start on fixing this, we add a new capability parameter to the
pseries machine type which gives the maximum allowed pagesizes for an
HPT guest. For now we just create and validate the parameter without
making it do anything.
For backwards compatibility, on older machine types we set it to the max
available page size for the host. For the 3.0 machine type, we fix it to
16, the intention being to only allow HPT pagesizes up to 64kiB by default
in future.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
KVM HV has a restriction that for HPT mode guests, guest pages must be hpa
contiguous as well as gpa contiguous. We have to account for that in
various places. We determine whether we're subject to this restriction
from the SMMU information exposed by KVM.
Planned cleanups to the way we handle this will require knowing whether
this restriction is in play in wider parts of the code. So, expose a
helper function which returns it.
This does mean some redundant calls to kvm_get_smmu_info(), but they'll go
away again with future cleanups.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
spapr capabilities have an apply hook to actually activate (or deactivate)
the feature in the system at reset time. However, a number of capabilities
affect the setup of cpus, and need to be applied to each of them -
including hotplugged cpus for extra complication. To make this simpler,
add an optional cpu_apply hook that is called from spapr_cpu_reset().
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Previously, the effective values of the various spapr capability flags
were only determined at machine reset time. That was a lazy way of making
sure it was after cpu initialization so it could use the cpu object to
inform the defaults.
But we've now improved the compat checking code so that we don't need to
instantiate the cpus to use it. That lets us move the resolution of the
capability defaults much earlier.
This is going to be necessary for some future capabilities.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
ppc_check_compat() is used in a number of places to check if a cpu object
supports a certain compatiblity mode, subject to various constraints.
It takes a PowerPCCPU *, however it really only depends on the cpu's class.
We have upcoming cases where it would be useful to make compatibility
checks before we fully instantiate the cpu objects.
ppc_type_check_compat() will now make an equivalent check, but based on a
CPU's QOM typename instead of an instantiated CPU object.
We make use of the new interface in several places in spapr, where we're
essentially making a global check, rather than one specific to a particular
cpu. This avoids some ugly uses of first_cpu to grab a "representative"
instance.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
CPUPPCState currently contains a number of fields containing the state of
the VPA. The VPA is a PAPR specific concept covering several guest/host
shared memory areas used to communicate some information with the
hypervisor.
As a PAPR concept this is really machine specific information, although it
is per-cpu, so it doesn't really belong in the core CPU state structure.
There's also other information that's per-cpu, but platform/machine
specific. So create a (void *)machine_data in PowerPCCPU which can be
used by the machine to locate per-cpu data. Intialization, lifetime and
cleanup of machine_data is entirely up to the machine type.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Greg Kurz <groug@kaod.org>
Currently we don't have any unrealize path for pnv cpu cores. We get away
with this because we don't yet support cpu hotplug for pnv.
However, we're going to want it eventually, and in the meantime, it makes
it non-obvious why there are a bunch of allocations on the realize() path
that don't have matching frees.
So, implement the missing unrealize path.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
pnv_cpu_init() is only called from the the pnv cpu core realize path, and
really only can be called from there. So fold it into its caller, which
we also rename for brevity.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Currently, we allocate space for all the cpu objects within a single core
in one big block. This was copied from an older version of the spapr code
and requires some ugly pointer manipulation to extract the individual
objects.
This design was due to a misunderstanding of qemu lifetime conventions and
has already been changed in spapr (in 94ad93bd "spapr_cpu_core: instantiate
CPUs separately".
Make an equivalent change in pnv_core to get rid of the nasty pointer
arithmetic.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
In pnv_core_realize() we call two functions with an Error * parameter in
succession, which will go badly if they both cause errors. In fact, a
failure in either of them indicates a qemu internal error, so we can just
use &error_abort in both cases.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
spapr_cpu_init() and spapr_cpu_destroy() are only called from the spapr
cpu core realize/unrealize paths, and really can only be called from there.
Those are all short functions, so fold the pairs together for simplicity.
While we're there rename some functions and change some parameter types
for brevity and clarity.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
By default, the IOMMU model built into the spapr virtual PCI host bridge
supports 4kiB and 64kiB IOMMU page sizes. However this can be overridden
which may be desirable to allow larger IOMMU page sizes when running a
guest with hugepage backing and passthrough devices. For that reason a
warning was printed when the device wasn't configured to allow the pagesize
with which guest RAM is backed.
Experience has proven, however, that this message is more confusing than
useful. Worse it sometimes makes little sense when the host-available page
sizes don't match those available on the guest, which can happen with
a POWER8 guest running on a POWER9 KVM host.
Long term we do want better handling to allow large IOMMU page sizes to be
used, but for now this parameter and warning don't really accomplish it.
So, remove the message, pending a better solution.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
To prevent spurious wakeups on cpus that are supposed to be disabled, we
need to clear the LPCR bits which control certain wakeup events.
spapr_cpu_reset() has separate cases here for boot and non-boot (initially
inactive) cpus. rtas_start_cpu() then turns the LPCR bits on when the
non-boot cpus are activated.
But explicit checks against first_cpu are not how we usually do things:
instead spapr_cpu_reset() generally sets things up for non-boot (inactive)
cpus, then spapr_machine_reset() and/or rtas_start_cpu() override as
necessary.
So, do that instead. Because the LPCR activation is identical for boot
cpus and non-boot cpus just activated with rtas_start_cpu() we can put the
code common in spapr_cpu_set_entry_state().
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Tested-by: Cédric Le Goater <clg@kaod.org>
cpu_ppc_set_papr() does several things:
1) it sets up the virtual hypervisor interface
2) it prevents the cpu from ever entering hypervisor mode
3) it tells KVM that we're emulating a cpu in PAPR mode
and 4) it configures the LPCR and AMOR (hypervisor privileged registers)
so that TCG will behave correctly for PAPR guests, without
attempting to emulate the cpu in hypervisor mode
(1) & (2) make sense for any virtual hypervisor (if another one ever
exists).
(3) belongs more properly in the machine type specific to a PAPR guest, so
move it to spapr_cpu_init(). While we're at it, remove an ugly test on
kvm_enabled() by making kvmppc_set_papr() a safe no-op on non-KVM.
(4) also belongs more properly in the machine type specific code. (4) is
done by mangling the default values of the SPRs, so that they will be set
correctly at reset time. Manipulating usually-static parameters of the cpu
model like this is kind of ugly, especially since the values used really
have more to do with the platform than the cpu.
The spapr code already has places for PAPR specific initializations of
register state in spapr_cpu_reset(), so move this handling there.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Tested-by: Cédric Le Goater <clg@kaod.org>
In cpu_ppc_set_papr() the UPRT and GTSE bits of the LPCR default value are
initialized based on on ppc64_radix_guest(). Which seems reasonable,
except that ppc64_radix_guest() is based on spapr->patb_entry which is
only set up in spapr_machine_reset, called _after_ cpu_ppc_set_papr() for
boot cpus. Well, and the fact that modifying the SPR default value for an
instance rather than a class is kind of yucky.
The initialization here is really only necessary or valid for
hotplugged cpus; the base cpu initialization already sets a value
that's good enough for the boot cpus until the guest uses an hcall to
configure it's preferred MMU mode.
So, move this initialization to the rtas_start_cpu() path, at which point
ppc64_radix_guest() will have a sensible value, to make sure secondary cpus
come up in an MMU mode matching the existing cpus.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Tested-by: Cédric Le Goater <clg@kaod.org>
There are several places in spapr_hcall.c where we need to update the LPCR
value on all CPUs. We do this with the set_spr() helper. That's not
really correct because this directly sets the SPR value, without going
through the ppc_store_lpcr() helper which may need to update state based
on the LPCR change.
In fact, set_spr() is only ever used for the LPCR, so replace it with an
explicit LPCR updated which uses the right low-level helper. While we're
there, move the CPU_FOREACH() which was in every one of the callers into
the new helper: set_all_lpcrs().
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Tested-by: Cédric Le Goater <clg@kaod.org>
Under PAPR, only the boot CPU is active when the system starts. Other cpus
must be explicitly activated using an RTAS call. The entry state for the
boot and secondary cpus isn't identical, but it has some things in common.
We're going to add a bit more common setup later, too, so to simplify
make a helper which sets up the common entry state for both boot and
secondary cpu threads.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
Tested-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
rtas_start_cpu() calls spapr_cpu_update_tb_offset() and
spapr_cpu_set_endianness() to initialize certain things in the new cpu's
state. This is the only caller of those helpers, and they're each only
a few lines long, so we might as well just fold them into the caller.
In addition, those helpers initialize state on the new cpu to match that of
the first cpu. That will generally work, but might be at least logically
incorrect if the first cpu has been set offline by the guest. So, instead
base the state on that of the cpu invoking the RTAS call, which is
obviously active already.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Tested-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
This makes several minor cleanups to these functions:
* Follow usual convention of an early exit on error, rather than having
most of the body in an if
* Clearer naming of cpu and cpu_. Now callcpu is the cpu from which the
RTAS call is invoked, newcpu is the cpu which we're starting
* Use cpu_synchronize_state() instead of kvm_cpu_synchronize_state()
directly
* Remove pointless comment describing what cpu_synchronize_state() does
* Use ppc_store_lpcr() instead of directly writing the register field
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Tested-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
There are some fields in the cpu state which need to be updated when the
LPCR register is changed, which is done by ppc_hash64_update_rmls() and
ppc_hash64_update_vrma(). Code which alters env->spr[SPR_LPCR] needs to
call them afterwards to make sure the state is up to date.
That's easy to get wrong. The normal way of dealing with sitautions like
that is to use a helper which both updates the basic register value and the
derived state.
So, do that.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Tested-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Current POWER cpus allow for a VRMA, a special mapping which describes a
guest's view of memory when in real mode (MMU off, from the guest's point
of view). Older cpus didn't have that which meant that to support a guest
a special host-contiguous region of memory was needed to give the guest its
Real Mode Area (RMA).
KVM used to provide special calls to allocate a contiguous RMA for those
cases. This was useful in the early days of KVM on Power to allow it to be
tested on PowerPC 970 chips as used in Macintosh G5 machines. Now, those
machines are so old as to be almost irrelevant.
The normal qemu deprecation process would require this to be marked
deprecated then removed in 2 releases. However, this can only be used
with corresponding support in the host kernel - which was dropped
years ago (in c17b98cf "KVM: PPC: Book3S HV: Remove code for PPC970
processors" of 2014-12-03 to be precise). Therefore it should be ok
to drop this immediately.
Just to be clear this only affects *KVM HV* guests with PowerPC 970,
and those already require an ancient host kernel. TCG and KVM PR
guests with PowerPC 970 should still work.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Acked-by: Thomas Huth <thuth@redhat.com>
If the -mem-path option is set, we attempt to map the guest's RAM from a
file in the given path; it's usually used to back guest RAM with hugepages.
If we're unable to (e.g. not enough free hugepages) then we fall back to
allocating normal anonymous pages. This behaviour can be surprising, but a
comment in allocate_system_memory_nonnuma() suggests it's legacy behaviour
we can't change.
What really isn't ok, though, is that in this case we leave mem_path set.
That means functions which attempt to determine the pagesize of main RAM
can erroneously think it is hugepage based on the requested path, even
though it's not.
This is particular bad for the pseries machine type. KVM HV limitations
mean the guest can't use pagesizes larger than the host page size used to
back RAM. That means that such a fallback, rather than merely giving
poorer performance than expected will cause the guest to freeze up early in
boot as it attempts to use large page mappings that can't work.
This patch addresses the problem by clearing the mem_path variable when we
fall back to anonymous pages, meaning that subsequent attempts to
determine the RAM page size will get an accurate result.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Although the order doesn't really matter at the moment, it's possible
other initializastions could depend on the compatiblity mode, so make sure
we set it first in spapr_cpu_reset().
While we're at it drop the test against first_cpu. Setting the compat mode
to the value it already has is redundant, but harmless, so we might as well
make a small simplification to the code.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
cpu_ppc_set_papr() removes the EP and HV bits from the MSR mask. While
removing the HV bit makes sense (a cpu in PAPR mode should never be
emulated in hypervisor mode), the EP bit is just bizarre. Although it's
true that a papr mode guest shouldn't be able to change the exception
prefix, the MSR[EP] bit doesn't even exist on the cpus supported for PAPR
mode, so it's pointless to do anything with it here.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Thomas Huth <thuth@redhat.com>
The env->slb_nr field gives the size of the SLB (Segment Lookaside Buffer).
This is another static-after-initialization parameter of the specific
version of the 64-bit hash MMU in the CPU. So, this patch folds the field
into PPCHash64Options with the other hash MMU options.
This is a bit more complicated that the things previously put in there,
because slb_nr was foolishly included in the migration stream. So we need
some of the usual dance to handle backwards compatible migration.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
These macros were introduced to deal with the fact that the mmu_model
field has bit flags mixed in with what's otherwise an enum of various mmu
types.
We've now eliminated all those flags except for one, and that one -
POWERPC_MMU_64 - is already included/compared in the MMU_VER macros. So,
we can get rid of those macros and just directly compare mmu_model values
in the places it was used.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
The only place we test this flag is in conjunction with
ppc64_use_proc_tbl(). That checks for the LPCR_UPRT bit, which we already
ensure can't be set except on a machine with a v3 MMU (i.e. POWER9).
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
The ci_large_pages boolean in CPUPPCState is only relevant to 64-bit hash
MMU machines, indicating whether it's possible to map large (> 4kiB) pages
as cache-inhibitied (i.e. for IO, rather than memory). Fold it as another
flag into the PPCHash64Options structure.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Currently env->mmu_model is a bit of an unholy mess of an enum of distinct
MMU types, with various flag bits as well. This makes which bits of the
field should be compared pretty confusing.
Make a start on cleaning that up by moving two of the flags bits -
POWERPC_MMU_1TSEG and POWERPC_MMU_AMR - which are specific to the 64-bit
hash MMU into a new flags field in PPCHash64Options structure.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Currently some cpus set the hash64_opts field in the class structure, with
specific details of their variant of the 64-bit hash mmu. For the
remaining cpus with that mmu, ppc_hash64_realize() fills in defaults.
But there are only a couple of cpus that use those fallbacks, so just have
them to set the has64_opts field instead, simplifying the logic.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
env->sps contains page size encoding information as an embedded structure.
Since this information is specific to 64-bit hash MMUs, split it out into
a separately allocated structure, to reduce the basic env size for other
cpus. Along the way we make a few other cleanups:
* Rename to PPCHash64Options which is more in line with qemu name
conventions, and reflects that we're going to merge some more hash64
mmu specific details in there in future. Also rename its
substructures to match qemu conventions.
* Move structure definitions to the mmu-hash64.[ch] files.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Initialization of the env->sps structure at the end of instance_init is
specific to the 64-bit hash MMU, so move the code into a helper function
in mmu-hash64.c.
We also create a corresponding function to be called at finalize time -
it's empty for now, but we'll need it shortly.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
CPU definitions for cpus with the 64-bit hash MMU can include a table of
available pagesizes. If this isn't supplied ppc_cpu_instance_init() will
fill it in a fallback table based on the POWERPC_MMU_64K bit in mmu_model.
However, it turns out all the cpus which support 64K pages already include
an explicit table of page sizes, so there's no point to the fallback table
including 64k pages.
That removes the only place which tests POWERPC_MMU_64K, so we can remove
it. Which in turn allows some logic to be removed from
kvm_fixup_page_sizes().
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
In most cases we prefer to pass a PowerPCCPU rather than the (embedded)
CPUPPCState.
For ppc_hash64_update_{rmls,vrma}() change to take "cpu" instead of "env".
For ppc_hash64_set_{dsi,isi}() remove the redundant "env" parameter.
In theory this makes more work for the functions, but since "cs", "cpu"
and "env" are related by at most constant offsets, the compiler should be
able to optimize out the difference at effectively zero cost.
helper_*() functions are left alone - since they're more closely tied to
the TCG generated code, passing "env" is still the standard there.
While we're there, fix an incorrect indentation.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
As a rule we prefer to pass PowerPCCPU instead of CPUPPCState, and this
change will make some things simpler later on.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
The #if isn't necessary, because there's a suitable one inside
ppc_cpu_is_valid(). We've already filtered for suitable cpu models in the
functions that search and register them. So by the time we get to realize
having an invalid one indicates a code error, not a user error, so an
assert() is more appropriate than error_setg().
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Thomas Huth <thuth@redhat.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Because of the various hooks called some variant on "init" - and the rather
greater number that used to exist, I'm always wondering when a function
called simply "*_init" or "*_initfn" will be called.
To make it easier on myself, and maybe others, rename the instance_init
hooks for ppc cpus to *_instance_init(). While we're at it rename the
realize time hooks to *_realize() (from *_realizefn()) which seems to be
the more common current convention.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
There are a couple places (one generic, one target specific) where we need
to get the host page size associated with a particular memory backend. I
have some upcoming code which will add another place which wants this. So,
for convenience, add a helper function to calculate this.
host_memory_backend_pagesize() returns the host pagesize for a given
HostMemoryBackend object.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Acked-by: Paolo Bonzini <pbonzini@redhat.com>
qemu_mempath_getpagesize() gets the effective (host side) page size for
a block of memory backed by an mmap()ed file on the host. It requires
the mem_path parameter to be non-NULL.
This ends up meaning all the callers need a different case for handling
anonymous memory (for memory-backend-ram or default memory with -mem-path
is not specified).
We can make all those callers a little simpler by having
qemu_mempath_getpagesize() accept NULL, and treat that as the anonymous
memory case.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Acked-by: Paolo Bonzini <pbonzini@redhat.com>
21b786f "PowerPC: Add TS bits into msr_mask" added the transaction states
to msr_mask for recent POWER CPUs to allow correct migration of machines
that are in certain interim transactional memory states.
This was correct, but unfortunately breaks backwards of pseries-2.7 and
earlier machine types which (stupidly) transferred the msr_mask in the
migration stream and failed if it wasn't equal on each end.
This works around the problem by masking out the new MSR bits in the
compatibility code to send the msr_mask on old machine types.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Greg Kurz <groug@kaod.org>
Tested-by: Lukáš Doktor <ldoktor@redhat.com>
ppc_tr_init_disas_context() correctly sets lazy_tlb_flush to true on
certain CPU models. However, it leaves it uninitialized, instead of
setting it to false on all others.
It wasn't caught before now because we didn't have examples in the tests
that exercised this path. However it can now be caught using clang's
undefined behaviour sanitizer and the sam460ex board.
Suggested-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Thomas Huth <thuth@redhat.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
fa98fbfc "PC: KVM: Support machine option to set VSMT mode" introduced the
"vsmt" parameter for the pseries machine type, which controls the spacing
of the vcpu ids of thread 0 for each virtual core. This was done to bring
some consistency and stability to how that was done, while still allowing
backwards compatibility for migration and otherwise.
The default value we used for vsmt was set to the max of the host's
advertised default number of threads and the number of vthreads per vcore
in the guest. This was done to continue running without extra parameters
on older KVM versions which don't allow the VSMT value to be changed.
Unfortunately, even that smaller than before leakage of host configuration
into guest visible configuration still breaks things. Specifically a guest
with 4 (or less) vthread/vcore will get a different vsmt value when
running on a POWER8 (vsmt==8) and POWER9 (vsmt==4) host. That means the
vcpu ids don't line up so you can't migrate between them, though you should
be able to.
Long term we really want to make vsmt == smp_threads for sufficiently
new machine types. However, that means that qemu will then require a
sufficiently recent KVM (one which supports changing VSMT) - that's still
not widely enough deployed to be really comfortable to do.
In the meantime we need some default that will work as often as
possible. This patch changes that default to 8 in all circumstances.
This does change guest visible behaviour (including for existing
machine versions) for many cases - just not the most common/important
case.
Following is case by case justification for why this is still the least
worst option. Note that any of the old behaviours can still be duplicated
after this patch, it's just that it requires manual intervention by
setting the vsmt property on the command line.
KVM HV on POWER8 host:
This is the overwhelmingly common case in production setups, and is
unchanged by design. POWER8 hosts will advertise a default VSMT mode
of 8, and > 8 vthreads/vcore isn't permitted
KVM HV on POWER7 host:
Will break, but POWER7s allowing KVM were never released to the public.
KVM HV on POWER9 host:
Not yet released to the public, breaking this now will reduce other
breakage later.
KVM HV on PowerPC 970:
Will theoretically break it, but it was barely supported to begin with
and already required various user visible hacks to work. Also so old
that I just don't care.
TCG:
This is the nastiest one; it means migration of TCG guests (without
manual vsmt setting) will break. Since TCG is rarely used in production
I think this is worth it for the other benefits. It does also remove
one more barrier to TCG<->KVM migration which could be interesting for
debugging applications.
KVM PR:
As with TCG, this will break migration of existing configurations,
without adding extra manual vsmt options. As with TCG, it is rare in
production so I think the benefits outweigh breakages.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Laurent Vivier <lvivier@redhat.com>
Reviewed-by: Jose Ricardo Ziviani <joserz@linux.vnet.ibm.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
At present if we require a vsmt mode that's not equal to the kernel's
default, and the kernel doesn't let us change it (e.g. because it's an old
kernel without support) then we always fail.
But in fact we can cope with the kernel having a different vsmt as long as
a) it's >= the actual number of vthreads/vcore (so that guest threads
that are supposed to be on the same core act like it)
b) it's a submultiple of the requested vsmt mode (so that guest threads
spaced by the vsmt value will act like they're on different cores)
Allowing this case gives us a bit more freedom to adjust the vsmt behaviour
without breaking existing cases.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Laurent Vivier <lvivier@redhat.com>
Tested-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
We recently had some discussions that were sidetracked for a while, because
nearly everyone misapprehended the purpose of the 'max_threads' field in
the compatiblity modes table. It's all about guest expectations, not host
expectations or support (that's handled elsewhere).
In an attempt to avoid a repeat of that confusion, rename the field to
'max_vthreads' and add an explanatory comment.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Laurent Vivier <lvivier@redhat.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Jose Ricardo Ziviani <joserz@linux.vnet.ibm.com>
The options field here is intended to list the available values for the
capability. It's not used yet, because the existing capabilities are
boolean.
We're going to add capabilities that aren't, but in that case the info on
the possible values can be folded into the .description field.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Decimal Floating Point has been available on POWER7 and later (server)
cpus. However, it can be disabled on the hypervisor, meaning that it's
not available to guests.
We currently handle this by conditionally advertising DFP support in the
device tree depending on whether the guest CPU model supports it - which
can also depend on what's allowed in the host for -cpu host. That can lead
to confusion on migration, since host properties are silently affecting
guest visible properties.
This patch handles it by treating it as an optional capability for the
pseries machine type.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
We currently have some conditionals in the spapr device tree code to decide
whether or not to advertise the availability of the VMX (aka Altivec) and
VSX vector extensions to the guest, based on whether the guest cpu has
those features.
This can lead to confusion and subtle failures on migration, since it makes
a guest visible change based only on host capabilities. We now have a
better mechanism for this, in spapr capabilities flags, which explicitly
depend on user options rather than host capabilities.
Rework the advertisement of VSX and VMX based on a new VSX capability. We
no longer bother with a conditional for VMX support, because every CPU
that's ever been supported by the pseries machine type supports VMX.
NOTE: Some userspace distributions (e.g. RHEL7.4) already rely on
availability of VSX in libc, so using cap-vsx=off may lead to a fatal
SIGILL in init.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
When constructing the "host" cpu class we modify whether the VMX and VSX
vector extensions and DFP (Decimal Floating Point) are available
based on whether KVM can support those instructions. This can depend on
policy in the host kernel as well as on the actual host cpu capabilities.
However, the way we probe for this is not very nice: we explicitly check
the host's device tree. That works in practice, but it's not really
correct, since the device tree is a property of the host kernel's platform
which we don't really know about. We get away with it because the only
modern POWER platforms happen to encode VMX, VSX and DFP availability in
the device tree in the same way.
Arguably we should have an explicit KVM capability for this, but we haven't
needed one so far. Barring specific KVM policies which don't yet exist,
each of these instruction classes will be available in the guest if and
only if they're available in the qemu userspace process. We can determine
that from the ELF AUX vector we're supplied with.
Once reworked like this, there are no more callers for kvmppc_get_vmx() and
kvmppc_get_dfp() so remove them.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Now that the "pseries" machine type implements optional capabilities (well,
one so far) there's the possibility of having different capabilities
available at either end of a migration. Although arguably a user error,
it would be nice to catch this situation and fail as gracefully as we can.
This adds code to migrate the capabilities flags. These aren't pulled
directly into the destination's configuration since what the user has
specified on the destination command line should take precedence. However,
they are checked against the destination capabilities.
If the source was using a capability which is absent on the destination,
we fail the migration, since that could easily cause a guest crash or other
bad behaviour. If the source lacked a capability which is present on the
destination we warn, but allow the migration to proceed.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
David Gibson