Yi Sun [Tue, 5 Mar 2019 02:34:55 +0000 (10:34 +0800)]
intel_iommu: add scalable-mode option to make scalable mode work
This patch adds an option to provide flexibility for user to expose
Scalable Mode to guest. User could expose Scalable Mode to guest by
the config as below:
Liu, Yi L [Tue, 5 Mar 2019 02:34:53 +0000 (10:34 +0800)]
intel_iommu: scalable mode emulation
Intel(R) VT-d 3.0 spec introduces scalable mode address translation to
replace extended context mode. This patch extends current emulator to
support Scalable Mode which includes root table, context table and new
pasid table format change. Now intel_iommu emulates both legacy mode
and scalable mode (with legacy-equivalent capability set).
The key points are below:
1. Extend root table operations to support both legacy mode and scalable
mode.
2. Extend context table operations to support both legacy mode and
scalable mode.
3. Add pasid tabled operations to support scalable mode.
Since commit 2566378d6d13bf4d28c7770bdbda5f7682594bbe, libvhost-user
no longer panics on disconnect (rc == 0), and instead silently ignores
an invalid VHOST_USER_NONE message.
Without extra work from the API user, this will simply busy-loop on
HUP events. The obvious thing to do is to exit(0) instead, while
additional or different work can be done by overriding
iface->process_msg().
vhost-user: define conventions for vhost-user backends
As discussed during "[PATCH v4 00/29] vhost-user for input & GPU"
review, let's define a common set of backend conventions to help with
management layer implementation, and interoperability.
contrib/libvhost-user/libvhost-user.c:953:20: error: implicit conversion from enumeration type 'enum VhostUserSlaveRequest' to different enumeration type 'VhostUserRequest' (aka 'enum VhostUserRequest') [-Werror,-Wenum-conversion]
.request = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
David Gibson [Wed, 6 Mar 2019 03:06:01 +0000 (14:06 +1100)]
virtio-balloon: Restore MADV_WILLNEED hint on balloon deflate
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.
David Gibson [Wed, 6 Mar 2019 03:06:00 +0000 (14:06 +1100)]
virtio-balloon: Fix possible guest memory corruption with inflates & deflates
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.
ed48c59875b6 "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.
Wei Wang [Tue, 12 Mar 2019 09:34:40 +0000 (17:34 +0800)]
virtio-balloon: fix a use-after-free case
The elem could theorically contain both outbuf and inbufs. We move the
free operation to the end of this function to avoid using elem->in_sg
while elem has been freed.
* remotes/kraxel/tags/audio-20190312-pull-request:
audio: -audiodev command line option: cleanup
wavaudio: port to -audiodev config
spiceaudio: port to -audiodev config
sdlaudio: port to -audiodev config
paaudio: port to -audiodev config
ossaudio: port to -audiodev config
noaudio: port to -audiodev config
dsoundaudio: port to -audiodev config
coreaudio: port to -audiodev config
alsaaudio: port to -audiodev config
audio: -audiodev command line option basic implementation
audio: -audiodev command line option: documentation
audio: use qapi AudioFormat instead of audfmt_e
qapi: qapi for audio backends
Signed-off-by: Peter Maydell <[email protected]>
# Conflicts:
# qemu-deprecated.texi
# gpg: Signature made Tue 12 Mar 2019 00:57:41 GMT
# gpg: using RSA key 2807936F984DC5A6
# gpg: Good signature from "Eduardo Habkost <[email protected]>" [full]
# Primary key fingerprint: 5A32 2FD5 ABC4 D3DB ACCF D1AA 2807 936F 984D C5A6
* remotes/ehabkost/tags/machine-next-pull-request:
memfd: improve error messages
memfd: set up correct errno if not supported
memfd: always check for MFD_CLOEXEC
hostmem-memfd: disable for systems without sealing support
machine: Move nvdimms state into struct MachineState
nvdimm: Rename AcpiNVDIMMState into NVDIMMState
hostmem-file: reject invalid pmem file sizes
* remotes/vivier2/tags/trivial-branch-pull-request:
hw/nvram/fw_cfg: Use the ldst API
hw/arm/virt: Remove null-check in virt_build_smbios()
hw/i386: Remove unused include
hw/nvram/fw_cfg: Remove the unnecessary boot_splash_filedata_size
thunk: improve readability of allocation loop
Peter Maydell [Tue, 12 Mar 2019 10:15:00 +0000 (10:15 +0000)]
Merge remote-tracking branch 'remotes/dgibson/tags/ppc-for-4.0-20190312' into staging
ppc patch queue for 2019-03-10
This pull requests supersedes ppc-for-4.0-20190310. Changes are:
* Fixed a bunch of minor style problems
* Suppressed warnings about Spectre/Meltdown mitigations with TCG
* Added one more patch, a preliminary fix towards the not-quite-ready
support for NVLink VFIO passthrough.
This is a final pull request before the 4.0 soft freeze. Changes
include:
* A Great Renaming to use camel case properly in spapr code
* Optimization of some vector instructions
* Support for POWER9 cpus in the powernv machine
* Fixes a regression from the last pull request in handling VSX
instructions with mixed operands from the FPR and VMX parts of the
register array
* Optimization hack to avoid scanning all the (empty) entries on a
new IOMMU window
* Add FSL I2C controller model for E500
* Support for KVM acceleration of the H_PAGE_INIT hypercall on spapr
* Update u-boot image for E500
* Enable Specre/Meltdown mitigations by default on the new machine type
* Enable large decrementer support for POWER9
* remotes/dgibson/tags/ppc-for-4.0-20190312: (62 commits)
vfio: Make vfio_get_region_info_cap public
Suppress test warnings about missing Spectre/Meltdown mitigations with TCG
spapr: Use CamelCase properly
target/ppc: Optimize x[sv]xsigdp using deposit_i64()
target/ppc: Optimize xviexpdp() using deposit_i64()
target/ppc: add HV support for POWER9
ppc/pnv: add a "ibm,opal/power-mgt" device tree node on POWER9
ppc/pnv: add more dummy XSCOM addresses
ppc/pnv: activate XSCOM tests for POWER9
ppc/pnv: POWER9 XSCOM quad support
ppc/pnv: extend XSCOM core support for POWER9
ppc/pnv: add a OCC model for POWER9
ppc/pnv: add a OCC model class
ppc/pnv: add SerIRQ routing registers
ppc/pnv: add a LPC Controller model for POWER9
ppc/pnv: add a 'dt_isa_nodename' to the chip
ppc/pnv: add a LPC Controller class model
ppc/pnv: lpc: fix OPB address ranges
ppc/pnv: add a PSI bridge model for POWER9
ppc/pnv: add a PSI bridge class model
...
David Gibson [Tue, 12 Mar 2019 05:07:14 +0000 (16:07 +1100)]
Suppress test warnings about missing Spectre/Meltdown mitigations with TCG
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.
David Gibson [Wed, 6 Mar 2019 04:35:37 +0000 (15:35 +1100)]
spapr: Use CamelCase properly
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.
The POWER9 processor does not support per-core frequency control. The
cores are arranged in groups of four, along with their respective L2
and L3 caches, into a structure known as a Quad. The frequency must be
managed at the Quad level.
Provide a basic Quad model to fake the settings done by the firmware
on the Non-Cacheable Unit (NCU). Each core pair (EX) needs a special
BAR setting for the TIMA area of XIVE because it resides on the same
address on all chips.
Provide a new class attribute to define XSCOM operations per CPU
family and add a couple of XSCOM addresses controlling the power
management states of the core on POWER9.
The LPC Controller on POWER9 is very similar to the one found on
POWER8 but accesses are now done via on MMIOs, without the XSCOM and
ECCB logic. The device tree is populated differently so we add a
specific POWER9 routine for the purpose.
The ISA bus has a different DT nodename on POWER9. Compute the name
when the PnvChip is realized, that is before it is used by the machine
to populate the device tree with the ISA devices.
The PowerNV LPC Controller exposes different sets of registers for
each of the functional units it encompasses, among which the OPB
(On-Chip Peripheral Bus) Master and Arbitrer and the LPC HOST
Controller.
The mapping addresses of each register range are correct but the sizes
are too large. Fix the sizes and define the OPB Arbitrer range to fill
the gap between the OPB Master registers and the LPC HOST Controller
registers.
The PSI bridge on POWER9 is very similar to POWER8. The BAR is still
set through XSCOM but the controls are now entirely done with MMIOs.
More interrupts are defined and the interrupt controller interface has
changed to XIVE. The POWER9 model is a first example of the usage of
the notify() handler of the XiveNotifier interface, linking the PSI
XiveSource to its owning device model.
To ease the introduction of the PSI bridge model for POWER9, abstract
the POWER chip differences in a PnvPsi class model and introduce a
specific Pnv8Psi type for POWER8. POWER8 interface to the interrupt
controller is still XICS whereas POWER9 uses the new XIVE model.
target/ppc: introduce vsr64_offset() to simplify get_cpu_vsr{l,h}() and set_cpu_vsr{l,h}()
Now that all VSX registers are stored in host endian order, there is no need
to go via different accessors depending upon the register number. Instead we
introduce vsr64_offset() and use it directly from within get_cpu_vsr{l,h}() and
set_cpu_vsr{l,h}().
This also allows us to rewrite avr64_offset() and fpr_offset() in terms of the
new vsr64_offset() function to more clearly express the relationship between the
VSX, FPR and VMX registers, and also remove vsrl_offset() which is no longer
required.
target/ppc: switch fpr/vsrl registers so all VSX registers are in host endian order
When VSX support was initially added, the fpr registers were added at
offset 0 of the VSR register and the vsrl registers were added at offset
1. This is in contrast to the VMX registers (the last 32 VSX registers) which
are stored in host-endian order.
Switch the fpr/vsrl registers so that the lower 32 VSX registers are now also
stored in host endian order to match the VMX registers. This ensures that TCG
vector operations involving mixed VMX and VSX registers will function
correctly.
target/ppc: improve avr64_offset() and use it to simplify get_avr64()/set_avr64()
By using the VsrD macro in avr64_offset() the same offset calculation can be
used regardless of the host endian. This allows get_avr64() and set_avr64() to
be simplified accordingly.
All TCG vector operations require pointers to the base address of the vector
rather than separate access to the top and bottom 64-bits. Convert the VMX TCG
instructions to use a new avr_full_offset() function instead of avr64_offset()
which can then itself be written as a simple wrapper onto vsr_full_offset().
This same function can also reused in cpu_avr_ptr() to avoid having more than
one copy of the offset calculation logic.
spapr_iommu: Do not replay mappings from just created DMA window
On sPAPR vfio_listener_region_add() is called in 2 situations:
1. a new listener is registered from vfio_connect_container();
2. a new IOMMU Memory Region is added from rtas_ibm_create_pe_dma_window().
In both cases vfio_listener_region_add() calls
memory_region_iommu_replay() to notify newly registered IOMMU notifiers
about existing mappings which is totally desirable for case 1.
However for case 2 it is nothing but noop as the window has just been
created and has no valid mappings so replaying those does not do anything.
It is barely noticeable with usual guests but if the window happens to be
really big, such no-op replay might take minutes and trigger RCU stall
warnings in the guest.
For example, a upcoming GPU RAM memory region mapped at 64TiB (right
after SPAPR_PCI_LIMIT) causes a 64bit DMA window to be at least 128TiB
which is (128<<40)/0x10000=2.147.483.648 TCEs to replay.
This mitigates the problem by adding an "skipping_replay" flag to
sPAPRTCETable and defining sPAPR own IOMMU MR replay() hook which does
exactly the same thing as the generic one except it returns early if
@skipping_replay==true.
Another way of fixing this would be delaying replay till the very first
H_PUT_TCE but this does not work if in-kernel H_PUT_TCE handler is
enabled (a likely case).
When "ibm,create-pe-dma-window" is complete, the guest will map only
required regions of the huge DMA window.
The NSR register of the HV ring has a different, although similar, bit
layout. TM_QW3_NSR_HE_PHYS bit should now be raised when the
Hypervisor interrupt line is signaled. Other bits TM_QW3_NSR_HE_POOL
and TM_QW3_NSR_HE_LSI are not modeled. LSI are for special interrupts
reserved for HW bringup and the POOL bit is used when signaling a
group of VPs. This is not currently implemented in Linux but it is in
pHyp.
The most important special commands on the HV TIMA page are added to
let the core manage interrupts : acking and changing the CPU priority.
ppc/pnv: add a XIVE interrupt controller model for POWER9
This is a simple model of the POWER9 XIVE interrupt controller for the
PowerNV machine which only addresses the needs of the skiboot
firmware. The PowerNV model reuses the common XIVE framework developed
for sPAPR as the fundamentals aspects are quite the same. The
difference are outlined below.
The controller initial BAR configuration is performed using the XSCOM
bus from there, MMIO are used for further configuration.
The MMIO regions exposed are :
- Interrupt controller registers
- ESB pages for IPIs and ENDs
- Presenter MMIO (Not used)
- Thread Interrupt Management Area MMIO, direct and indirect
The virtualization controller MMIO region containing the IPI ESB pages
and END ESB pages is sub-divided into "sets" which map portions of the
VC region to the different ESB pages. These are modeled with custom
address spaces and the XiveSource and XiveENDSource objects are sized
to the maximum allowed by HW. The memory regions are resized at
run-time using the configuration of EDT set translation table provided
by the firmware.
The XIVE virtualization structure tables (EAT, ENDT, NVTT) are now in
the machine RAM and not in the hypervisor anymore. The firmware
(skiboot) configures these tables using Virtual Structure Descriptor
defining the characteristics of each table : SBE, EAS, END and
NVT. These are later used to access the virtual interrupt entries. The
internal cache of these tables in the interrupt controller is updated
and invalidated using a set of registers.
Still to address to complete the model but not fully required is the
support for block grouping. Escalation support will be necessary for
KVM guests.
The PowerNV machine can perform indirect loads and stores on the TIMA
on behalf of another CPU. Give the controller the possibility to call
the TIMA memory accessors with a XiveTCTX of its choice.
ppc/xive: hardwire the Physical CAM line of the thread context
By default on P9, the HW CAM line (23bits) is hardwired to :
0x000||0b1||4Bit chip number||7Bit Thread number.
When the block group mode is enabled at the controller level (PowerNV),
the CAM line is changed for CAM compares to :
4Bit chip number||0x001||7Bit Thread number
This will require changes in xive_presenter_tctx_match() possibly.
This is a lowlevel functionality of the HW controller and it is not
strictly needed. Leave it for later.
PPC: E500: Add FSL I2C controller and integrate RTC with it
Original commit message:
This patch adds an emulation model for i2c controller found on most of the FSL SoCs.
It also integrates the RTC (ds1338) that sits on the i2c Bus with e500 machine model.
Patch was originally written by Amit Singh Tomar <[email protected]>
see http://patchwork.ozlabs.org/patch/431475/
I only fixed it enough for application on top of current qemu master 20b084c4b1401b7f8fbc385649d48c67b6f43d44, and hopefully fixed checkpatch errors
Tested by booting Linux kernel 4.20.12. Now e500 machine doesn't need
network time protocol daemon because it will have working RTC
(before all timestamps on files were from 2016)
David Gibson [Wed, 6 Mar 2019 03:15:26 +0000 (14:15 +1100)]
spapr: Force SPAPR_MEMORY_BLOCK_SIZE to be a hwaddr (64-bit)
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).
target/ppc/spapr: Clear partition table entry when allocating hash table
If we allocate a hash page table then we know that the guest won't be
using process tables, so set the partition table entry maintained for
the guest to zero. If this isn't done, then the guest radix bit will
remain set in the entry. This means that when the guest calls
H_REGISTER_PROCESS_TABLE there will be a mismatch between then flags
and the value in spapr->patb_entry, and the call will fail. The guest
will then panic:
Failed to register process table (rc=-4)
kernel BUG at arch/powerpc/platforms/pseries/lpar.c:959
The result being that it isn't possible to boot a hash guest on a P9
system.
Also fix a bug in the flags parsing in h_register_process_table() which
was introduced by the same patch, and simplify the handling to make it
less likely that errors will be introduced in the future. The effect
would have been setting the host radix bit LPCR_HR for a hash guest
using process tables, which currently isn't supported and so couldn't
have been triggered.
Alexander Graf [Mon, 4 Mar 2019 10:39:30 +0000 (11:39 +0100)]
PPC: E500: Update u-boot to v2019.01
Quite a while has passed since we last updated U-Boot for e500. This patch
bumps it to the last released version 2019.01 to make sure users don't feel
like they're using out of date software.
Fabiano Rosas [Thu, 28 Feb 2019 22:57:58 +0000 (19:57 -0300)]
target/ppc: Refactor kvm_handle_debug
There are four scenarios being handled in this function:
- single stepping
- hardware breakpoints
- software breakpoints
- fallback (no debug supported)
A future patch will add code to handle specific single step and
software breakpoints cases so let's split each scenario into its own
function now to avoid hurting readability.
target/ppc/spapr: Enable mitigations by default for pseries-4.0 machine type
There are currently 3 mitigations the availability of which is controlled
by the spapr-caps mechanism, cap-cfpc, cap-sbbc, and cap-ibs. Enable these
mitigations by default for the pseries-4.0 machine type.
By now machine firmware should have been upgraded to allow these
settings.
target/ppc/tcg: make spapr_caps apply cap-[cfpc/sbbc/ibs] non-fatal for tcg
The spapr_caps cap-cfpc, cap-sbbc and cap-ibs are used to control the
availability of certain mitigations to the guest. These haven't been
implemented under TCG, it is unlikely they ever will be, and it is unclear
as to whether they even need to be.
As such, make failure to apply these capabilities under TCG non-fatal.
Instead we print a warning message to the user but still allow the guest
to continue.
Introduce a new spapr_cap SPAPR_CAP_CCF_ASSIST to be used to indicate
the requirement for a hw-assisted version of the count cache flush
workaround.
The count cache flush workaround is a software workaround which can be
used to flush the count cache on context switch. Some revisions of
hardware may have a hardware accelerated flush, in which case the
software flush can be shortened. This cap is used to set the
availability of such hardware acceleration for the count cache flush
routine.
The availability of such hardware acceleration is indicated by the
H_CPU_CHAR_BCCTR_FLUSH_ASSIST flag being set in the characteristics
returned from the KVM_PPC_GET_CPU_CHAR ioctl.
target/ppc/spapr: Add workaround option to SPAPR_CAP_IBS
The spapr_cap SPAPR_CAP_IBS is used to indicate the level of capability
for mitigations for indirect branch speculation. Currently the available
values are broken (default), fixed-ibs (fixed by serialising indirect
branches) and fixed-ccd (fixed by diabling the count cache).
Introduce a new value for this capability denoted workaround, meaning that
software can work around the issue by flushing the count cache on
context switch. This option is available if the hypervisor sets the
H_CPU_BEHAV_FLUSH_COUNT_CACHE flag in the cpu behaviours returned from
the KVM_PPC_GET_CPU_CHAR ioctl.
target/ppc/spapr: Enable the large decrementer for pseries-4.0
Enable the large decrementer by default for the pseries-4.0 machine type.
It is disabled again by default_caps_with_cpu() for pre-POWER9 cpus
since they don't support the large decrementer.
target/ppc: Implement large decrementer support for KVM
Implement support to allow KVM guests to take advantage of the large
decrementer introduced on POWER9 cpus.
To determine if the host can support the requested large decrementer
size, we check it matches that specified in the ibm,dec-bits device-tree
property. We also need to enable it in KVM by setting the LPCR_LD bit in
the LPCR. Note that to do this we need to try and set the bit, then read
it back to check the host allowed us to set it, if so we can use it but
if we were unable to set it the host cannot support it and we must not
use the large decrementer.
target/ppc: Implement large decrementer support for TCG
Prior to POWER9 the decrementer was a 32-bit register which decremented
with each tick of the timebase. From POWER9 onwards the decrementer can
be set to operate in a mode called large decrementer where it acts as a
n-bit decrementing register which is visible as a 64-bit register, that
is the value of the decrementer is sign extended to 64 bits (where n is
implementation dependant).
The mode in which the decrementer operates is controlled by the LPCR_LD
bit in the logical paritition control register (LPCR).
>From POWER9 onwards the HDEC (hypervisor decrementer) was enlarged to
h-bits, also sign extended to 64 bits (where h is implementation
dependant). Note this isn't configurable and is always enabled.
On POWER9 the large decrementer and hdec are both 56 bits, as
represented by the lrg_decr_bits cpu class property. Since they are the
same size we only add one property for now, which could be extended in
the case they ever differ in the future.
We also add the lrg_decr_bits property for POWER5+/7/8 since it is used
to determine the size of the hdec, which is only generated on the
POWER5+ processor and later. On these processors it is 32 bits.
Greg Kurz [Fri, 1 Mar 2019 19:32:31 +0000 (20:32 +0100)]
Revert "spapr: support memory unplug for qtest"
Commit b8165118f52c broke CPU hotplug tests for old machine types:
$ QTEST_QEMU_BINARY=ppc64-softmmu/qemu-system-ppc64 ./tests/cpu-plug-test -m=slow
/ppc64/cpu-plug/pseries-3.1/device-add/2x3x1&maxcpus=6: OK
/ppc64/cpu-plug/pseries-2.12-sxxm/device-add/2x3x1&maxcpus=6: OK
/ppc64/cpu-plug/pseries-3.0/device-add/2x3x1&maxcpus=6: OK
/ppc64/cpu-plug/pseries-2.10/device-add/2x3x1&maxcpus=6: OK
/ppc64/cpu-plug/pseries-2.11/device-add/2x3x1&maxcpus=6: OK
/ppc64/cpu-plug/pseries-2.12/device-add/2x3x1&maxcpus=6: OK
/ppc64/cpu-plug/pseries-2.9/device-add/2x3x1&maxcpus=6: OK
/ppc64/cpu-plug/pseries-2.7/device-add/2x3x1&maxcpus=6: **
ERROR:/home/thuth/devel/qemu/hw/ppc/spapr_events.c:313:rtas_event_log_to_source: assertion failed: (source->enabled)
Broken pipe
/home/thuth/devel/qemu/tests/libqtest.c:143: kill_qemu() detected QEMU death from signal 6 (Aborted) (core dumped)
Aborted (core dumped)
The approach of faking the availability of OV5_HP_EVT causes the
code to assume the hotplug event source is enabled, which is wrong
for older machines.
Greg Kurz [Fri, 1 Mar 2019 19:32:37 +0000 (20:32 +0100)]
spapr: Simulate CAS for qtest
The RTAS event hotplug code for machine types 2.8 and newer depends on
the CAS negotiated ov5 in order to work properly. However, there's no
CAS when running under qtest. There has been a tentative to trick the
code by faking the OV5_HP_EVT bit, but it turned out to break other
assumptions in the code and the change got reverted.
Go for a more general approach and simulate a CAS when running under
qtest. For simplicity, this pseudo CAS simple simulates the case where
the guest supports the same features as the machine. It is done at
reset time, just before we reset the DRCs, which could potentially
exercise the unplug code.
This allows to test unplug on spapr with both older and newer machine
types.
The "systempagesize" name suggests that it is the host system page size
while it is the smallest page size of memory backing the guest RAM so
let's rename it to stop confusion. This should cause no behavioral change.
The current code assumes that we can address more bits on a PCI bus
for DMA than we really can but there is no way knowing the actual limit.
This makes a better guess for the number of levels and if the kernel
fails to allocate that, this increases the level numbers till succeeded
or reached the 64bit limit.
This adds levels to the trace point.
This may cause the kernel to warn about failed allocation:
[65122.837458] Failed to allocate a TCE memory, level shift=28
which might happen if MAX_ORDER is not large enough as it can vary:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/powerpc/Kconfig?h=v5.0-rc2#n727
The PC machines put firmware in ROM by default. To get it put into
flash memory (required by OVMF), you have to use -drive
if=pflash,unit=0,... and optionally -drive if=pflash,unit=1,...
Why two -drive? This permits setting up one part of the flash memory
read-only, and the other part read/write. It also makes upgrading
firmware on the host easier. Below the hood, it creates two separate
flash devices, because we were too lazy to improve our flash device
models to support sector protection.
The problem at hand is to do the same with -blockdev somehow, as one
more step towards deprecating -drive.
Mapping -drive if=none,... to -blockdev is a solved problem. With
if=T other than if=none, -drive additionally configures a block device
frontend. For non-onboard devices, that part maps to -device. Also a
solved problem. For onboard devices such as PC flash memory, we have
an unsolved problem.
This is actually an instance of a wider problem: our general device
configuration interface doesn't cover onboard devices. Instead, we have
a zoo of ad hoc interfaces that are much more limited. One of them is
-drive, which we'd rather deprecate, but can't until we have suitable
replacements for all its uses.
Sadly, I can't attack the wider problem today. So back to the narrow
problem.
My first idea was to reduce it to its solved buddy by using pluggable
instead of onboard devices for the flash memory. Workable, but it
requires some extra smarts in firmware descriptors and libvirt. Paolo
had an idea that is simpler for libvirt: keep the devices onboard, and
add machine properties for their block backends.
The implementation is less than straightforward, I'm afraid.
First, block backend properties are *qdev* properties. Machines can't
have those, as they're not devices. I could duplicate these qdev
properties as QOM properties, but I hate that.
More seriously, the properties do not belong to the machine, they
belong to the onboard flash devices. Adding them to the machine would
then require bad magic to somehow transfer them to the flash devices.
Fortunately, QOM provides the means to handle exactly this case: add
alias properties to the machine that forward to the onboard devices'
properties.
Properties need to be created in .instance_init() methods. For PC
machines, that's pc_machine_initfn(). To make alias properties work,
we need to create the onboard flash devices there, too. Requires
several bug fixes, in the previous commits. We also have to realize
the devices. More on that below.
If the user sets pflash0, firmware resides in flash memory.
pc_system_firmware_init() maps and realizes the flash devices.
Else, firmware resides in ROM. The onboard flash devices aren't used
then. pc_system_firmware_init() destroys them unrealized, along with
the alias properties.
The existing code to pick up drives defined with -drive if=pflash is
replaced by code to desugar into the machine properties.
pc_sysfw: Pass PCMachineState to pc_system_firmware_init()
pc_system_firmware_init() parameter @isapc_ram_fw is PCMachineState
member pci_enabled negated. The next commit will need more of
PCMachineState. To prepare for that, pass a PCMachineState *, and
drop the now redundant parameter @isapc_ram_fw.
vl: Create block backends before setting machine properties
qemu-system-FOO's main() acts on command line arguments in its own
idiosyncratic order. There's not much method to its madness.
Whenever we find a case where one kind of command line argument needs
to refer to something created for another kind later, we rejigger the
order.
Block devices get created long after machine properties get processed.
Therefore, block device machine properties can be created, but not
set. No such properties exist. But the next commit will create some.
Time to rejigger again: create block devices earlier.
The first call of sysbus_get_default() creates the main system bus and
stores it in QOM as "/machine/unattached/sysbus". This must not
happen before main() creates "/machine", or else container_get() would
"helpfully" create it as "container" object, and the real creation of
"/machine" would later abort with "attempt to add duplicate property
'machine' to object (type 'container')". Has been that way ever since
we wired up busses in QOM (commit f968fc6892d, v1.2.0).
I believe the bug is latent. I got it to bite by trying to
qdev_create() a sysbus device from a machine's .instance_init()
method.
The fix is obvious: store the main system bus in QOM right after
creating "/machine".
vl: Fix latent bug with -global and onboard devices
main() registers the user's -global only after we create the machine
object, i.e. too late for devices created in the machine's
.instance_init().
Fortunately, we know the bug is only latent: the commit before
previous fixed a bug that would've crashed any attempt to create a
device in an .instance_init().
qdev: Fix latent bug with compat_props and onboard devices
Compatibility properties started life as a qdev property thing: we
supported them only for qdev properties, and implemented them with the
machinery backing command line option -global.
Recent commit fa0cb34d221 put them to use (tacitly) with memory
backend objects (subtypes of TYPE_MEMORY_BACKEND). To make that
possible, we first moved the work of applying them from the -global
machinery into TYPE_DEVICE's .instance_post_init() method
device_post_init(), in commits ea9ce8934c5 and b66bbee39f6, then made
it available to TYPE_MEMORY_BACKEND's .instance_post_init() method
host_memory_backend_post_init() as object_apply_compat_props(), in
commit 1c3994f6d2a.
Note the code smell: we now have function name starting with object_
in hw/core/qdev.c. It has to be there rather than in qom/, because it
calls qdev_get_machine() to find the current accelerator's and
machine's compat_props.
Turns out calling qdev_get_machine() there is problematic. If we
qdev_create() from a machine's .instance_init() method, we call
device_post_init() and thus qdev_get_machine() before main() can
create "/machine" in QOM. qdev_get_machine() tries to get it with
container_get(), which "helpfully" creates it as "container" object,
and returns that. object_apply_compat_props() tries to paper over the
problem by doing nothing when the value of qdev_get_machine() isn't a
TYPE_MACHINE. But the damage is done already: when main() later
attempts to create the real "/machine", it fails with "attempt to add
duplicate property 'machine' to object (type 'container')", and
aborts.
Since no machine .instance_init() calls qdev_create() so far, the bug
is latent. But since I want to do that, I get to fix the bug first.
Observe that object_apply_compat_props() doesn't actually need the
MachineState, only its the compat_props member of its MachineClass and
AccelClass. This permits a simple fix: register MachineClass and
AccelClass compat_props with the object_apply_compat_props() machinery
right after these classes get selected.
This is actually similar to how things worked before commits ea9ce8934c5 and b66bbee39f6, except we now register much earlier. The
old code registered them only after the machine's .instance_init()
ran, which would've broken compatibility properties for any devices
created there.
Our pflash devices are simplistically modelled has having
"num-blocks" sectors of equal size "sector-length". Real hardware
commonly has sectors of different sizes. How our "sector-length"
property is related to the physical device's multiple sector sizes
is unclear.
Helper functions pflash_cfi01_register() and pflash_cfi02_register()
create a pflash device, set properties including "sector-length" and
"num-blocks", and realize. They take parameters @size, @sector_len
and @nb_blocs.
QOMification left parameter @size unused. Obviously, @size should
match @sector_len and @nb_blocs, i.e. size == sector_len * nb_blocs.
All callers satisfy this.
Remove @nb_blocs and compute it from @size and @sector_len.
projects / qemu.git / log