David Gibson [Thu, 14 Feb 2019 04:39:14 +0000 (15:39 +1100)]
virtio-balloon: Rework ballon_page() interface
This replaces the balloon_page() internal interface with
ballon_inflate_page(), with a slightly different interface. The new
interface will make future alterations simpler.
David Gibson [Thu, 14 Feb 2019 04:39:13 +0000 (15:39 +1100)]
virtio-balloon: Corrections to address verification
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
David Gibson [Thu, 14 Feb 2019 04:39:12 +0000 (15:39 +1100)]
virtio-balloon: Remove unnecessary MADV_WILLNEED on deflate
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.
Peter Xu [Tue, 12 Feb 2019 14:06:21 +0000 (15:06 +0100)]
i386/kvm: ignore masked irqs when update msi routes
When we are with intel-iommu device and with IR on, KVM will register
an IEC notifier to detect interrupt updates from the guest and we'll
kick off kvm_update_msi_routes_all() when it happens to make sure
kernel IRQ cache is matching the latest.
Though, kvm_update_msi_routes_all() is buggy in that it ignored the
mask bit of either MSI/MSIX messages and it tries to translate the
message even if the corresponding message was already masked by the
guest driver (hence the MSI/MSIX message will be invalid).
Without this patch, we can receive an error message when we reboot a
guest with both an assigned vfio-pci device and intel-iommu enabled:
qemu-system-x86_64: vtd_interrupt_remap_msi: MSI address low 32 bit invalid: 0x0
The error does not affect functionality of the guest since when we
failed to translate we'll just silently continue (which makes sense
since crashing the VM for this seems even worse), but still it's
better to fix it up.
While the git history remains bisectable, having a commit that changes
MSI/MSIX code but describes it as "fix vhost-user-blk compilation" is
rather confusing.
Revert the offending commit to properly apply both patches separately.
Wei Yang [Mon, 11 Feb 2019 06:46:29 +0000 (14:46 +0800)]
pc-dimm: use same mechanism for [get|set]_addr
[get|set]_addr are two counterpart to access PCDIMMDevice.addr.
Since we have already set up a property PC_DIMM_ADDR_PROP for this
field and use this mechanism in set_addr, it would be more proper to use
the same mechanism in get_addr.
This patch uses object_property_get_uint() to replace the direct memory
access to make [get|set]_addr with the same mechanism.
Laszlo Ersek [Mon, 4 Feb 2019 16:03:24 +0000 (17:03 +0100)]
tests/uefi-test-tools: add build scripts
Introduce the following build scripts under "tests/uefi-test-tools":
* "build.sh" builds a single module (a UEFI application) from
UefiTestToolsPkg, for a single QEMU emulation target.
"build.sh" relies on cross-compilers when the emulation target and the
build host architecture don't match. The cross-compiler prefix is
computed according to a fixed, Linux-specific pattern. No attempt is
made to copy or reimplement the GNU Make magic from "qemu/roms/Makefile"
for cross-compiler prefix determination. The reason is that the build
host OSes that are officially supported by edk2, and those that are
supported by QEMU, intersect only in Linux. (Note that the UNIXGCC
toolchain is being removed from edk2,
<https://bugzilla.tianocore.org/show_bug.cgi?id=1377>.)
* "Makefile" currently builds the "UefiTestToolsPkg/BiosTablesTest"
application, for arm, aarch64, i386, and x86_64, with the help of
"build.sh".
"Makefile" turns each resultant UEFI executable into a UEFI-bootable,
qcow2-compressed ISO image. The ISO images are output as
"tests/data/uefi-boot-images/bios-tables-test.<TARGET>.iso.qcow2".
Each ISO image should be passed to QEMU as follows:
Laszlo Ersek [Mon, 4 Feb 2019 16:03:23 +0000 (17:03 +0100)]
tests: introduce "uefi-test-tools" with the BiosTablesTest UEFI app
The "bios-tables-test" program in QEMU's test suite locates the RSD PTR
ACPI table in guest RAM, and (chasing pointers to other ACPI tables)
performs various sanity checks on the QEMU-generated and
firmware-installed tables.
Currently this set of test cases doesn't work with UEFI guests. The ACPI
spec defines distinct methods for OSPM to locate the RSD PTR on
traditional BIOS vs. UEFI platforms, and the UEFI method is more difficult
to implement from the hypervisor side with just raw guest memory access.
Add a UEFI application (to be booted in the UEFI guest) that populates a
small, MB-aligned structure in guest RAM. The structure begins with a
signature GUID. The hypervisor should loop over all MB-aligned pages in
guest RAM until one matches the signature GUID at offset 0, at which point
the hypervisor can fetch the RSDP address field(s) from the structure.
QEMU's test logic currently spins on a pre-determined guest address, until
that address assumes a magic value. The method described in this patch is
conceptually the same ("busy loop until match is found"), except there is
no hard-coded address. This plays a lot more nicely with UEFI guest
firmware (we'll be able to use the normal page allocation UEFI service).
Given the size of EFI_GUID (16 bytes -- 128 bits), mismatches should be
astronomically unlikely. In addition, given the typical guest RAM size for
such tests (128 MB), there are 128 locations to check in one iteration of
the "outer" loop, which shouldn't introduce an intolerable delay after the
guest stores the RSDP address(es), and then the GUID.
Note that in the patch, we define "gBiosTablesTestGuid" with all bits
inverted. This is a simple method to prevent the UEFI binary, which
incorporates "gBiosTablesTestGuid", from matching the actual GUID in guest
RAM.
The UEFI application is written against the edk2 framework, which was
introduced earlier as a git submodule. The next patch will provide build
scripts for maintainers.
The source code follows the edk2 coding style, and is licensed under the
2-clause BSDL (in case someone would like to include UefiTestToolsPkg
content in a different edk2 platform).
The "UefiTestToolsPkg.dsc" platform description file resolves the used
edk2 library classes to instances (= library implementations) such that
the UEFI binaries inherit no platform dependencies. They are expected to
run on any system that conforms to the UEFI-2.3.1 spec (which was released
in 2012). The arch-specific build options are carried over from edk2's
ArmVirtPkg and OvmfPkg platforms.
Laszlo Ersek [Mon, 4 Feb 2019 16:03:22 +0000 (17:03 +0100)]
roms: build the EfiRom utility from the roms/edk2 submodule
Building the EfiRom utility from "roms/edk2/BaseTools" should make
"roms/Makefile" more self-contained. Otherwise, we'd call the system-wide
EfiRom for building the combined iPXE option ROMs, but call the sibling
utilities from "roms/edk2/BaseTools" for building "roms/edk2" content.
Laszlo Ersek [Mon, 4 Feb 2019 16:03:21 +0000 (17:03 +0100)]
roms: add the edk2 project as a git submodule
The roms/edk2 submodule can help with three goals:
- build the OVMF and ArmVirtQemu virtual UEFI firmware platforms (to be
implemented later),
- build the EfiRom tool on the fly, which is used in roms/Makefile, for
building the "efirom" target,
- build UEFI test applications (to be run in guests), for qtest support.
The edk2 repository tracks some binary files that should not be removed by
QEMU's top-level "make clean"; exempt the full pathnames from the "find"
command.
Paolo Bonzini [Thu, 14 Feb 2019 17:35:55 +0000 (18:35 +0100)]
vhost-user-test: small changes to init_hugepagefs
After the conversion to qgraph, the equivalent of "main" will be in
a constructor and will run even if the tests are not being requested.
Therefore, it should not assert that init_hugepagefs succeeds and will
be called when creating the TestServer. This patch changes the prototype
of init_hugepagefs, this way the next patch looks nicer.
Paolo Bonzini [Thu, 14 Feb 2019 17:35:53 +0000 (18:35 +0100)]
vhost-net: revamp configure logic
Detect all invalid configurations (e.g. mingw32 with vhost-user,
non-Linux with vhost-kernel). As a collateral benefit, all vhost-kernel
backends can be now disabled if one wants to reduce the attack surface.
Paolo Bonzini [Thu, 14 Feb 2019 17:35:52 +0000 (18:35 +0100)]
vhost-net: compile it on all targets that have virtio-net.
This shows a preexisting bug: if a KVM target did not have virtio-net enabled,
it would fail with undefined symbols when vhost was enabled. This must now
be fixed, lest targets that have no virtio-net fail to compile.
Paolo Bonzini [Thu, 14 Feb 2019 17:35:51 +0000 (18:35 +0100)]
vhost-user: support cross-endian vnet headers
vhost-user already has a way to communicate the endianness of the guest
via the vring endianness messages. The vring endianness always matches
the vnet header endianness so there is no need to do anything else in
the backend.
Paolo Bonzini [Thu, 14 Feb 2019 17:35:50 +0000 (18:35 +0100)]
vhost: restrict Linux dependency to kernel vhost
vhost-user does not depend on Linux; it can run on any POSIX system. Restrict
vhost-kernel to Linux in hw/virtio/vhost-backend.c, everything else can be
compiled on all POSIX systems.
Paolo Bonzini [Thu, 14 Feb 2019 17:35:49 +0000 (18:35 +0100)]
vhost-net-user: add stubs for when no virtio-net device is present
hw/net/vhost_net.c needs functions that are declared in net/vhost-user.c: the
vhost-user code is always compiled into QEMU, only the constructor
net_init_vhost_user is unreachable. Also, net/vhost-user.c needs functions
declared in hw/virtio/vhost-stub.c even if no virtio device exists.
Break this dependency. First, add a minimal version of net/vhost-user.c,
with no functionality and no dependency on vhost code. Second, #ifdef out
the calls back to net/vhost-user.c from hw/net/vhost_net.c.
While at it, this patch fixes the CONFIG_VHOST_NET_USE*D* typo.
Paolo Bonzini [Thu, 14 Feb 2019 17:35:48 +0000 (18:35 +0100)]
vhost-net: move stubs to a separate file
There is no reason for CONFIG_VHOST_NET to be specific to a single target;
it is a host feature that can be add to all targets, as long as they support
the virtio-net device. Currently CONFIG_VHOST_NET depends on CONFIG_KVM,
but ioeventfd support is present in the core memory API and works with
other accelerators as well.
As a first step, move the vhost-net stubs to a separate file. Later, they
will become conditional on CONFIG_VIRTIO_NET, which is not available in .c
files.
* remotes/jnsnow/tags/bitmaps-pull-request:
blockdev: acquire aio_context for bitmap add/remove
block/dirty-bitmap: Documentation and Comment fixups
dirty-bitmap: Expose persistent flag to 'query-block'
* remotes/kraxel/tags/usb-20190220-pull-request:
usb: remove unnecessary NULL device check from usb_ep_get()
usb: add device checks before redirector calls to usb_ep_get()
usb: check device is not NULL before calling usb_ep_get()
uhci: check device is not NULL before calling usb_ep_get()
ohci: check device is not NULL before calling usb_ep_get()
ehci: check device is not NULL before calling usb_ep_get()
xhci: check device is not NULL before calling usb_ep_get()
xhci: add asserts to help with static code analysis
usb: rearrange usb_ep_get()
Liam Merwick [Wed, 6 Feb 2019 13:36:56 +0000 (13:36 +0000)]
usb: remove unnecessary NULL device check from usb_ep_get()
No caller of usb_ep_get() calls it with a NULL device (previous commits
have addressed the few remaining cases which didn't explicitly check).
Replace check for 'dev == NULL' with an assert instead.
Liam Merwick [Wed, 6 Feb 2019 13:36:54 +0000 (13:36 +0000)]
usb: check device is not NULL before calling usb_ep_get()
In musb_packet(), the call to usb_find_device() can return NULL
if it doesn't find a device matching 'addr' so explicitly check
the return value before passing it to usb_ep_get(). This then
allows the subsequent calculation of 'id' to be streamlined.
Liam Merwick [Wed, 6 Feb 2019 13:36:53 +0000 (13:36 +0000)]
uhci: check device is not NULL before calling usb_ep_get()
In uhci_handle_td(), the call to ehci_find_device() can return NULL
if it doesn't find a device matching 'addr' so explicitly check
the return value before passing it to usb_ep_get().
Liam Merwick [Wed, 6 Feb 2019 13:36:52 +0000 (13:36 +0000)]
ohci: check device is not NULL before calling usb_ep_get()
A call to ohci_find_device() can return NULL if it doesn't find a
device matching 'addr' so for the two callers, explicitly check
the return value before passing it to usb_ep_get().
Liam Merwick [Wed, 6 Feb 2019 13:36:51 +0000 (13:36 +0000)]
ehci: check device is not NULL before calling usb_ep_get()
In ehci_process_itd(), the call to ehci_find_device() can return NULL
if it doesn't find a device matching 'devaddr' so explicitly check
the return value before passing it to usb_ep_get().
Liam Merwick [Wed, 6 Feb 2019 13:36:49 +0000 (13:36 +0000)]
xhci: add asserts to help with static code analysis
Most callers of xhci_port_update() and xhci_wakeup() pass in a pointer
to an array entry and can never be NULL but add two defensive asserts
to protect against future changes (e.g. adding a new port speed, etc.)
adding a path through xhci_lookup_port() that could result in the
return of a NULL XHCIPort.
Liam Merwick [Wed, 6 Feb 2019 13:36:48 +0000 (13:36 +0000)]
usb: rearrange usb_ep_get()
There is no need to calculate the 'eps' variable in usb_ep_get()
if 'ep' is the control endpoint. Instead the calculation should
be done after validating the input before returning an entry
indexed by the endpoint 'ep'.
John Snow [Tue, 19 Feb 2019 22:49:43 +0000 (17:49 -0500)]
blockdev: acquire aio_context for bitmap add/remove
When bitmaps are persistent, they may incur a disk read or write when bitmaps
are added or removed. For configurations like virtio-dataplane, failing to
acquire this lock will abort QEMU when disk IO occurs.
We used to acquire aio_context as part of the bitmap lookup, so re-introduce
the lock for just the cases that have an IO penalty. Commit 2119882c removed
these locks, and I failed to notice this when we committed fd5ae4cc, so this
has been broken since persistent bitmaps were introduced.
Eric Blake [Tue, 19 Feb 2019 22:49:43 +0000 (17:49 -0500)]
dirty-bitmap: Expose persistent flag to 'query-block'
Since qemu currently doesn't flush persistent bitmaps to disk until
shutdown (which might be MUCH later), it's useful if 'query-block'
at least shows WHICH bitmaps will (eventually) make it to persistent
storage. Update affected iotests.
* remotes/dgibson/tags/ppc-for-4.0-20190219: (43 commits)
target/ppc: convert vmin* and vmax* to vector operations
target/ppc: convert vadd*s and vsub*s to vector operations
target/ppc: Split out VSCR_SAT to a vector field
target/ppc: Add set_vscr_sat
target/ppc: Use mtvscr/mfvscr for vmstate
target/ppc: Add helper_mfvscr
target/ppc: Remove vscr_nj and vscr_sat
target/ppc: Use helper_mtvscr for reset and gdb
target/ppc: Pass integer to helper_mtvscr
target/ppc: convert xxsel to vector operations
target/ppc: convert xxspltw to vector operations
target/ppc: convert xxspltib to vector operations
target/ppc: convert VSX logical operations to vector operations
target/ppc: convert vsplt[bhw] to use vector operations
target/ppc: convert vspltis[bhw] to use vector operations
target/ppc: convert vaddu[b,h,w,d] and vsubu[b,h,w,d] over to use vector operations
target/ppc: convert VMX logical instructions to use vector operations
xics: Drop the KVM ICS class
spapr/irq: Use the "simple" ICS class for KVM
xics: Handle KVM interrupt presentation from "simple" ICS code
...
Peter Maydell [Mon, 18 Feb 2019 14:23:13 +0000 (14:23 +0000)]
Merge remote-tracking branch 'remotes/armbru/tags/pull-qapi-2019-02-18' into staging
QAPI patches for 2019-02-18
# gpg: Signature made Mon 18 Feb 2019 13:44:30 GMT
# gpg: using RSA key 3870B400EB918653
# gpg: Good signature from "Markus Armbruster <[email protected]>" [full]
# gpg: aka "Markus Armbruster <[email protected]>" [full]
# Primary key fingerprint: 354B C8B3 D7EB 2A6B 6867 4E5F 3870 B400 EB91 8653
* remotes/armbru/tags/pull-qapi-2019-02-18:
qapi: move RTC_CHANGE to the target schema
qmp: Deprecate query-events in favor of query-qmp-schema
Revert "qapi-events: add 'if' condition to implicit event enum"
qapi: remove qmp_unregister_command()
qapi: make query-cpu-definitions depend on specific targets
qapi: make query-cpu-model-expansion depend on s390 or x86
qapi: make query-gic-capabilities depend on TARGET_ARM
target.json: add a note about query-cpu* not being s390x-specific
qapi: make s390 commands depend on TARGET_S390X
qapi: make rtc-reset-reinjection and SEV depend on TARGET_I386
qapi: New module target.json
build: Deal with all of QAPI's .o in qapi/Makefile.objs
build-sys: move qmp-introspect per target
qapi: Generate QAPIEvent stuff into separate files
qapi: Prepare for system modules other than 'builtin'
qapi: Clean up modular built-in code generation a bit
qapi: Fix up documentation for recent commit a95291007b2
qapi: Belatedly document modular code generation
A few targets don't emit RTC_CHANGE, we could restrict the event to
the tagets that do emit it.
Note: There is a lot more of events & commands that we could restrict
to capable targets, with the cost of some additional complexity, but
the benefit of added correctness and better introspection.
qmp: Deprecate query-events in favor of query-qmp-schema
query-events doesn't reflect compile-time configuration. Instead of
fixing that, deprecate the command in favor of query-qmp-schema.
Libvirt prefers query-qmp-schema as of commit 22d7222ec0 "qemu: caps:
Don't call 'query-events' when we probe events from QMP schema".
It'll be in the next release.
The commit applied the events' conditions to the members of enum
QAPIEvent. Awkward, because it renders QAPIEvent unusable in
target-independent code as soon as we make an event target-dependent.
Reverting this has the following effects:
* ui/vnc.c can remain target independent.
* monitor_qapi_event_conf[] doesn't have to muck around with #ifdef.
* query-events again doesn't reflect conditionals. I'm going to
deprecate it in favor of query-qmp-schema.
Another option would be to split target-dependent parts off enum
QAPIEvent into a target-dependent enum. Doesn't seem worthwhile right
now.
We can't add appropriate target-specific conditionals to misc.json,
because that would make all of misc.json unusable in
target-independent code. To keep misc.json target-independent, we
need to split off target-dependent target.json.
This commit doesn't actually split off anything, it merely creates the
empty module. The next few patches will move stuff from misc.json
there.
build: Deal with all of QAPI's .o in qapi/Makefile.objs
Adding QAPI's .o to util-obj-y, common-obj-y and obj-y is spread over
three places: Makefile.objs takes care of target-independent generated
code, Makefile.target of target-dependent generated code, and
qapi/Makefile.objs of (target-independent) hand-written code.
The following patches are going to introduce per-target #ifdef in the
schemas.
The introspection data is statically generated once, and must thus be
built per-target to reflect target-specific configuration.
Drop "do_test_visitor_in_qmp_introspect(&qmp_schema_qlit)" since the
schema is no longer in a common object. It is covered by the per-target
query-qmp-schema test instead.
qapi: Generate QAPIEvent stuff into separate files
Having to include qapi-events.h just for QAPIEvent is suboptimal, but
quite tolerable now. It'll become problematic when we have events
conditional on the target, because then qapi-events.h won't be usable
from target-independent code anymore. Avoid that by generating it
into separate files.
qapi: Prepare for system modules other than 'builtin'
The next commit wants to generate qapi-emit-events.{c.h}. To enable
that, extend QAPISchemaModularCVisitor to support additional "system
modules", i.e. modules that don't correspond to a (user-defined) QAPI
schema module.
qapi: Clean up modular built-in code generation a bit
We neglect to call .visit_module() for the special module we use for
built-ins. Harmless, but clean it up anyway. The
tests/qapi-schema/*.out now show the built-in module as 'module None'.
Subclasses of QAPISchemaModularCVisitor need to ._add_module() this
special module to enable code generation for built-ins. When this
hasn't been done, QAPISchemaModularCVisitor.visit_module() does
nothing for the special module. That looks like built-ins could
accidentally be generated into the wrong module when a subclass
neglects to call ._add_module(). Can't happen, because built-ins are
all visited before any other module. But that's non-obvious. Switch
off code generation explicitly.
Rename QAPISchemaModularCVisitor._begin_module() to
._begin_user_module().
New QAPISchemaModularCVisitor._is_builtin_module(), for clarity.
We generate code for built-ins and sub-modules into separate files
since commit cdb6610ae42 and 252dc3105fc (v2.12.0). Both commits
neglected to update documentation. Do that now.
Peter Maydell [Mon, 18 Feb 2019 11:32:00 +0000 (11:32 +0000)]
Merge remote-tracking branch 'remotes/cohuck/tags/s390x-20190218' into staging
s390x updates:
- tcg: implement STCK and friends for CONFIG_USER_ONLY
- add zpci to qemu cpu model, as pci is now always built
- add mepoch to default z14 cpu model
- add cpu model for z14 GA2
- various improvements
* remotes/cohuck/tags/s390x-20190218:
s390x: upgrade status of KVM cores to "supported"
s390x/kvm: add tracepoint to ioeventfd interface
s390x/cpumodel: add z14 GA2 model
s390x/cpumodel: default enable mepoch for z14 and later
s390x/cpumodel: mepochptff: warn when no mepoch and re-align group init
s390x: add zPCI feature to "qemu" CPU model
target/s390x: Implement STCK et al for CONFIG_USER_ONLY
target/s390x: Split out s390-tod.h
s390x: always provide pci support
s390x: Fix the confusing contributions-after-2012 license statements
Collin Walling [Tue, 12 Feb 2019 01:16:56 +0000 (20:16 -0500)]
s390x/cpumodel: default enable mepoch for z14 and later
Latest systems and host kernels support mepoch, which is a
feature that was meant to be supported for z14 GA1 from the
get-go. Let's copy it to the z14 GA1 default CPU model.
Machines s390-ccw-virtio-3.1 and older will retain the old CPU
models and will not provide this bit nor the extended PTFF
functions in the default model.
Collin Walling [Tue, 12 Feb 2019 01:16:55 +0000 (20:16 -0500)]
s390x/cpumodel: mepochptff: warn when no mepoch and re-align group init
The extended PTFF features (qsie, qtoue, stoe, stoue) are dependent
on the multiple-epoch facility (mepoch). Let's print a warning if these
features are enabled without mepoch.
While we're at it, let's move the FEAT_GROUP_INIT for mepochptff down
the s390_feature_groups list so it can be properly indexed with its
generated S390FeatGroup enum.
target/s390x: Implement STCK et al for CONFIG_USER_ONLY
This is a non-privileged instruction that was only implemented
for system mode. However, the stck instruction is used by glibc,
so this was causing SIGILL for programs run under debian stretch.
Cornelia Huck [Mon, 11 Feb 2019 11:32:55 +0000 (12:32 +0100)]
s390x: always provide pci support
We tried to make pci support optional on s390x in the past;
unfortunately, we still require the s390 phb to be created
unconditionally due to backwards compatibility issues.
Instead of sinking more effort into this (including compat
handling for older machines etc.) for non-obvious gains, let's
just make CONFIG_PCI something that is always set on s390x.
Note that you can still fence off pci for the _guest_ if you
provide a cpu model without the zpci feature.
Thomas Huth [Wed, 6 Feb 2019 12:41:33 +0000 (13:41 +0100)]
s390x: Fix the confusing contributions-after-2012 license statements
The license information in these files is rather confusing. The text
declares LGPL first, but then says that contributions after 2012 are
licensed under the GPL instead. How should the average user who just
downloaded the release tarball know which part is now GPL and which
is LGPL?
Looking at the text of the LGPL (see COPYING.LIB in the top directory),
the license clearly states how this should be done instead:
"3. You may opt to apply the terms of the ordinary GNU General Public
License instead of this License to a given copy of the Library. To do
this, you must alter all the notices that refer to this License, so
that they refer to the ordinary GNU General Public License, version 2,
instead of to this License."
Thus let's clean up the confusing statements and use the proper GPL
text only.
Greg Kurz [Fri, 15 Feb 2019 11:40:30 +0000 (12:40 +0100)]
xics: Handle KVM interrupt presentation from "simple" ICS code
We want to use the "simple" ICS type in both KVM and non-KVM setups.
Teach the "simple" ICS how to present interrupts to KVM and adapt
sPAPR accordingly.
Greg Kurz [Fri, 15 Feb 2019 11:40:24 +0000 (12:40 +0100)]
xics: Handle KVM ICS reset from the "simple" ICS code
The KVM ICS reset handler simply writes the ICS state to KVM. This
doesn't need the overkill parent_reset logic we have today. Also
we want to use the same ICS type for the KVM and non-KVM case with
pseries.
Call icp_set_kvm_state() from the "simple" ICS reset function.
Greg Kurz [Fri, 15 Feb 2019 11:40:18 +0000 (12:40 +0100)]
xics: Explicitely call KVM ICS methods from the common code
The pre_save(), post_load() and synchronize_state() methods of the
ICSStateClass type are really KVM only things. Make that obvious
by dropping the indirections and directly calling the KVM functions
instead.
Greg Kurz [Fri, 15 Feb 2019 11:40:00 +0000 (12:40 +0100)]
xics: Handle KVM ICP realize from the common code
The realization of KVM ICP currently follows the parent_realize logic,
which is a bit overkill here. Also we want to get rid of the KVM ICP
class. Explicitely call icp_kvm_realize() from the base ICP realize
function.
Note that ICPStateClass::parent_realize is retained because powernv
needs it.
Greg Kurz [Fri, 15 Feb 2019 11:39:54 +0000 (12:39 +0100)]
xics: Handle KVM ICP reset from the common code
The KVM ICP reset handler simply writes the ICP state to KVM. This
doesn't need the overkill parent_reset logic we have today. Call
icp_set_kvm_state() from the base ICP reset function instead.
Since there are no other users for ICPStateClass::parent_reset, and
it isn't currently expected to change, drop it as well.
Greg Kurz [Fri, 15 Feb 2019 11:39:48 +0000 (12:39 +0100)]
xics: Explicitely call KVM ICP methods from the common code
The pre_save(), post_load() and synchronize_state() methods of the
ICPStateClass type are really KVM only things. Make that obvious
by dropping the indirections and directly calling the KVM functions
instead.
Cédric Le Goater [Wed, 13 Feb 2019 21:07:56 +0000 (22:07 +0100)]
spapr/irq: remove the XICS offset adjustment
Now that we have changed the XICS and the XIVE interrupt backend to
have different size for their IRQ number space, we do not need to
align their source numbers anymore. Remove the offset adjustment and
wire the dual 'qirq' handler to the 'qirq' handler of the current
interrupt mode in use.
Cédric Le Goater [Wed, 13 Feb 2019 21:07:55 +0000 (22:07 +0100)]
spapr/irq: add an 'nr_irq' parameter to initialize the backend.
When using the 'dual' interrupt mode, the source numbers of both sPAPR
IRQ backends are aligned to share a common IRQ number space and to use
a similar mapping of the machine qemu_irq array which is indexed by
the source number.
The XICS IRQ number range initially being [ 0x1000 - 0x2000 ], this
requires to change the XICS ICSState offset to 0 and to provision for
an extra 4K of source numbers and qemu_irqs which will never be used
by the machine when running under the XICS interrupt mode. This is not
an optimal solution.
Change the init() method to allocate an IRQ number space of the
expected size for the XICS sPAPR IRQ backend. It breaks the interrupt
signaling when under the 'dual' mode because source numbers have
unexpected values but next patch will fix that.
This causes the allocated buffer 'int_buf' to be smaller than expected
and we eventually overwrite some of glibc's control structures (see
"chunk" in https://sourceware.org/glibc/wiki/MallocInternals)
The following error is seen while trying to free int_buf:
Michael Roth [Tue, 12 Feb 2019 18:24:59 +0000 (19:24 +0100)]
qdev: pass an Object * to qbus_set_hotplug_handler()
Certain devices types, like memory/CPU, are now being handled using a
hotplug interface provided by a top-level MachineClass. Hotpluggable
host bridges are another such device where it makes sense to use a
machine-level hotplug handler. However, unlike those devices,
host-bridges have a parent bus (the main system bus), and devices with
a parent bus use a different mechanism for registering their hotplug
handlers: qbus_set_hotplug_handler(). This interface currently expects
a handler to be a subclass of DeviceClass, but this is not the case
for MachineClass, which derives directly from ObjectClass.
Internally, the interface only requires an ObjectClass, so expose that
in qbus_set_hotplug_handler().
Greg Kurz [Tue, 12 Feb 2019 18:24:06 +0000 (19:24 +0100)]
xive: Only set source type for LSIs
MSI is the default and LSI specific code is guarded by the
xive_source_irq_is_lsi() helper. The xive_source_irq_set()
helper is a nop for MSIs.
Simplify the code by turning xive_source_irq_set() into
xive_source_irq_set_lsi() and only call it for LSIs. The
call to xive_source_irq_set(false) in spapr_xive_irq_free()
is also a nop. Just drop it.
Roman Kapl [Tue, 12 Feb 2019 12:12:55 +0000 (13:12 +0100)]
ppc: fix crash during branch stepping
The PPC BRANCH exception could bubble up, but this is an QEMU internal exception
and QEMU then crased. Instead it should trigger TRACE exception, according to
PPC 2.07 book. It could happen only when using branch stepping, which is not
commonly used.
Change gen_prep_dbgex do do trigger TRACE. The excp, argument is now removed,
since the type of exception can be inferred from the singlestep_enabled flags.
removed the guards around gen_exception, since they are unnecessary.
projects / qemu.git / log