A down_read on memory_lock is held when performing read/write accesses
to MMIO BAR space, including across the copy_to/from_user() callouts
which may fault. If the user buffer for these copies resides in an
mmap of device MMIO space, the mmap fault handler will acquire a
recursive read-lock on memory_lock. Avoid this by reducing the lock
granularity. Sequential accesses requiring multiple ioread/iowrite
cycles are expected to be rare, therefore typical accesses should not
see additional overhead.
VGA MMIO accesses are expected to be non-fatal regardless of the PCI
memory enable bit to allow legacy probing, this behavior remains with
a comment added. ioeventfds are now included in memory access testing,
with writes dropped while memory space is disabled.
Fixes: abafbc551f ("vfio-pci: Invalidate mmaps and block MMIO access on disabled memory")
Reported-by: Zhiyi Guo <zhguo@redhat.com>
Tested-by: Zhiyi Guo <zhguo@redhat.com>
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
The current generation of Intel® QuickAssist Technology devices
are not designed to run in an untrusted environment because of the
following issues reported in the document "Intel® QuickAssist Technology
(Intel® QAT) Software for Linux" (document number 336211-014):
QATE-39220 - GEN - Intel® QAT API submissions with bad addresses that
trigger DMA to invalid or unmapped addresses can cause a
platform hang
QATE-7495 - GEN - An incorrectly formatted request to Intel® QAT can
hang the entire Intel® QAT Endpoint
The document is downloadable from https://01.org/intel-quickassist-technology
at the following link:
https://01.org/sites/default/files/downloads/336211-014-qatforlinux-releasenotes-hwv1.7_0.pdf
This patch adds the following QAT devices to the denylist: DH895XCC,
C3XXX and C62X.
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Reviewed-by: Andy Shevchenko <andy.shevchenko@gmail.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Add denylist of devices that by default are not probed by vfio-pci.
Devices in this list may be susceptible to untrusted application, even
if the IOMMU is enabled. To be accessed via vfio-pci, the user has to
explicitly disable the denylist.
The denylist can be disabled via the module parameter disable_denylist.
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Reviewed-by: Andy Shevchenko <andy.shevchenko@gmail.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
No need to release and immediately re-acquire igate while clearing
out the eventfd ctxs.
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Intel document 333717-008, "Intel® Ethernet Controller X550
Specification Update", version 2.7, dated June 2020, includes errata
#22, added in version 2.1, May 2016, indicating X550 NICs suffer from
the same implementation deficiency as the 700-series NICs:
"The Interrupt Status bit in the Status register of the PCIe
configuration space is not implemented and is not set as described
in the PCIe specification."
Without the interrupt status bit, vfio-pci cannot determine when
these devices signal INTx. They are therefore added to the nointx
quirk.
Cc: Jesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
SR-IOV VFs do not implement the memory enable bit of the command
register, therefore this bit is not set in config space after
pci_enable_device(). This leads to an unintended difference
between PF and VF in hand-off state to the user. We can correct
this by setting the initial value of the memory enable bit in our
virtualized config space. There's really no need however to
ever fault a user on a VF though as this would only indicate an
error in the user's management of the enable bit, versus a PF
where the same access could trigger hardware faults.
Fixes: abafbc551f ("vfio-pci: Invalidate mmaps and block MMIO access on disabled memory")
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
The next use of the device will generate an underflow from the
stale reference.
Cc: Qian Cai <cai@lca.pw>
Fixes: 1518ac272e ("vfio/pci: fix memory leaks of eventfd ctx")
Reported-by: Daniel Wagner <dwagner@suse.de>
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Tested-by: Daniel Wagner <dwagner@suse.de>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Convert the last few remaining mmap_sem rwsem calls to use the new mmap
locking API. These were missed by coccinelle for some reason (I think
coccinelle does not support some of the preprocessor constructs in these
files ?)
[akpm@linux-foundation.org: convert linux-next leftovers]
[akpm@linux-foundation.org: more linux-next leftovers]
[akpm@linux-foundation.org: more linux-next leftovers]
Signed-off-by: Michel Lespinasse <walken@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com>
Reviewed-by: Laurent Dufour <ldufour@linux.ibm.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Davidlohr Bueso <dbueso@suse.de>
Cc: David Rientjes <rientjes@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Liam Howlett <Liam.Howlett@oracle.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ying Han <yinghan@google.com>
Link: http://lkml.kernel.org/r/20200520052908.204642-6-walken@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The PCI Code and ID Assignment Specification changed capability ID 0
from reserved to a NULL capability in the v1.1 revision. The NULL
capability is defined to include only the 16-bit capability header,
ie. only the ID and next pointer. Unfortunately vfio-pci creates a
map of config space, where ID 0 is used to reserve the standard type
0 header. Finding an actual capability with this ID therefore results
in a bogus range marked in that map and conflicts with subsequent
capabilities. As this seems to be a dummy capability anyway and we
already support dropping capabilities, let's hide this one rather than
delving into the potentially subtle dependencies within our map.
Seen on an NVIDIA Tesla T4.
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Accessing the disabled memory space of a PCI device would typically
result in a master abort response on conventional PCI, or an
unsupported request on PCI express. The user would generally see
these as a -1 response for the read return data and the write would be
silently discarded, possibly with an uncorrected, non-fatal AER error
triggered on the host. Some systems however take it upon themselves
to bring down the entire system when they see something that might
indicate a loss of data, such as this discarded write to a disabled
memory space.
To avoid this, we want to try to block the user from accessing memory
spaces while they're disabled. We start with a semaphore around the
memory enable bit, where writers modify the memory enable state and
must be serialized, while readers make use of the memory region and
can access in parallel. Writers include both direct manipulation via
the command register, as well as any reset path where the internal
mechanics of the reset may both explicitly and implicitly disable
memory access, and manipulation of the MSI-X configuration, where the
MSI-X vector table resides in MMIO space of the device. Readers
include the read and write file ops to access the vfio device fd
offsets as well as memory mapped access. In the latter case, we make
use of our new vma list support to zap, or invalidate, those memory
mappings in order to force them to be faulted back in on access.
Our semaphore usage will stall user access to MMIO spaces across
internal operations like reset, but the user might experience new
behavior when trying to access the MMIO space while disabled via the
PCI command register. Access via read or write while disabled will
return -EIO and access via memory maps will result in a SIGBUS. This
is expected to be compatible with known use cases and potentially
provides better error handling capabilities than present in the
hardware, while avoiding the more readily accessible and severe
platform error responses that might otherwise occur.
Fixes: CVE-2020-12888
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Rather than calling remap_pfn_range() when a region is mmap'd, setup
a vm_ops handler to support dynamic faulting of the range on access.
This allows us to manage a list of vmas actively mapping the area that
we can later use to invalidate those mappings. The open callback
invalidates the vma range so that all tracking is inserted in the
fault handler and removed in the close handler.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Older versions of skiboot only provide a single value in the device
tree property "ibm,mmio-atsd", even when multiple Address Translation
Shoot Down (ATSD) registers are present. This prevents NVLink2 devices
(other than the first) from being used with vfio-pci because vfio-pci
expects to be able to assign a dedicated ATSD register to each NVLink2
device.
However, ATSD registers can be shared among devices. This change
allows vfio-pci to fall back to sharing the register at index 0 if
necessary.
Fixes: 7f92891778 ("vfio_pci: Add NVIDIA GV100GL [Tesla V100 SXM2] subdriver")
Signed-off-by: Sam Bobroff <sbobroff@linux.ibm.com>
Reviewed-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
The cleanup is getting a tad long.
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
It currently results in messages like:
"vfio-pci 0000:03:00.0: vfio_pci: ..."
Which is quite a bit redundant.
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
With the VF Token interface we can now expect that a vfio userspace
driver must be in collaboration with the PF driver, an unwitting
userspace driver will not be able to get past the GET_DEVICE_FD step
in accessing the device. We can now move on to actually allowing
SR-IOV to be enabled by vfio-pci on the PF. Support for this is not
enabled by default in this commit, but it does provide a module option
for this to be enabled (enable_sriov=1). Enabling VFs is rather
straightforward, except we don't want to risk that a VF might get
autoprobed and bound to other drivers, so a bus notifier is used to
"capture" VFs to vfio-pci using the driver_override support. We
assume any later action to bind the device to other drivers is
condoned by the system admin and allow it with a log warning.
vfio-pci will disable SR-IOV on a PF before releasing the device,
allowing a VF driver to be assured other drivers cannot take over the
PF and that any other userspace driver must know the shared VF token.
This support also does not provide a mechanism for the PF userspace
driver itself to manipulate SR-IOV through the vfio API. With this
patch SR-IOV can only be enabled via the host sysfs interface and the
PF driver user cannot create or remove VFs.
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
The VFIO_DEVICE_FEATURE ioctl is meant to be a general purpose, device
agnostic ioctl for setting, retrieving, and probing device features.
This implementation provides a 16-bit field for specifying a feature
index, where the data porition of the ioctl is determined by the
semantics for the given feature. Additional flag bits indicate the
direction and nature of the operation; SET indicates user data is
provided into the device feature, GET indicates the device feature is
written out into user data. The PROBE flag augments determining
whether the given feature is supported, and if provided, whether the
given operation on the feature is supported.
The first user of this ioctl is for setting the vfio-pci VF token,
where the user provides a shared secret key (UUID) on a SR-IOV PF
device, which users must provide when opening associated VF devices.
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
If we enable SR-IOV on a vfio-pci owned PF, the resulting VFs are not
fully isolated from the PF. The PF can always cause a denial of service
to the VF, even if by simply resetting itself. The degree to which a PF
can access the data passed through a VF or interfere with its operation
is dependent on a given SR-IOV implementation. Therefore we want to
avoid a scenario where an existing vfio-pci based userspace driver might
assume the PF driver is trusted, for example assigning a PF to one VM
and VF to another with some expectation of isolation. IOMMU grouping
could be a solution to this, but imposes an unnecessarily strong
relationship between PF and VF drivers if they need to operate with the
same IOMMU context. Instead we introduce a "VF token", which is
essentially just a shared secret between PF and VF drivers, implemented
as a UUID.
The VF token can be set by a vfio-pci based PF driver and must be known
by the vfio-pci based VF driver in order to gain access to the device.
This allows the degree to which this VF token is considered secret to be
determined by the applications and environment. For example a VM might
generate a random UUID known only internally to the hypervisor while a
userspace networking appliance might use a shared, or even well know,
UUID among the application drivers.
To incorporate this VF token, the VFIO_GROUP_GET_DEVICE_FD interface is
extended to accept key=value pairs in addition to the device name. This
allows us to most easily deny user access to the device without risk
that existing userspace drivers assume region offsets, IRQs, and other
device features, leading to more elaborate error paths. The format of
these options are expected to take the form:
"$DEVICE_NAME $OPTION1=$VALUE1 $OPTION2=$VALUE2"
Where the device name is always provided first for compatibility and
additional options are specified in a space separated list. The
relation between and requirements for the additional options will be
vfio bus driver dependent, however unknown or unused option within this
schema should return error. This allow for future use of unknown
options as well as a positive indication to the user that an option is
used.
An example VF token option would take this form:
"0000:03:00.0 vf_token=2ab74924-c335-45f4-9b16-8569e5b08258"
When accessing a VF where the PF is making use of vfio-pci, the user
MUST provide the current vf_token. When accessing a PF, the user MUST
provide the current vf_token IF there are active VF users or MAY provide
a vf_token in order to set the current VF token when no VF users are
active. The former requirement assures VF users that an unassociated
driver cannot usurp the PF device. These semantics also imply that a
VF token MUST be set by a PF driver before VF drivers can access their
device, the default token is random and mechanisms to read the token are
not provided in order to protect the VF token of previous users. Use of
the vf_token option outside of these cases will return an error, as
discussed above.
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
This currently serves the same purpose as the default implementation
but will be expanded for additional functionality.
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Mmgrab was introduced in commit f1f1007644 ("mm: add new mmgrab()
helper") and most of the kernel was updated to use it. Update a
remaining file.
The semantic patch that makes this change is as follows:
(http://coccinelle.lip6.fr/)
<smpl>
@@ expression e; @@
- atomic_inc(&e->mm_count);
+ mmgrab(e);
</smpl>
Signed-off-by: Julia Lawall <Julia.Lawall@inria.fr>
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
The nvlink2 subdriver for IBM Witherspoon machines preregisters
GPU memory in the IOMMI API so KVM TCE code can map this memory
for DMA as well. This is done by mm_iommu_newdev() called from
vfio_pci_nvgpu_regops::mmap.
In an unlikely event of failure the data->mem remains NULL and
since mm_iommu_put() (which unregisters the region and unpins memory
if that was regular memory) does not expect mem=NULL, it should not be
called.
This adds a check to only call mm_iommu_put() for a valid data->mem.
Fixes: 7f92891778 ("vfio_pci: Add NVIDIA GV100GL [Tesla V100 SXM2] subdriver")
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
ioremap has provided non-cached semantics by default since the Linux 2.6
days, so remove the additional ioremap_nocache interface.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Since irq_bypass_register_producer() is called after request_irq(), we
should do tear-down in reverse order: irq_bypass_unregister_producer()
then free_irq().
Specifically free_irq() may release resources required by the
irqbypass del_producer() callback. Notably an example provided by
Marc Zyngier on arm64 with GICv4 that he indicates has the potential
to wedge the hardware:
free_irq(irq)
__free_irq(irq)
irq_domain_deactivate_irq(irq)
its_irq_domain_deactivate()
[unmap the VLPI from the ITS]
kvm_arch_irq_bypass_del_producer(cons, prod)
kvm_vgic_v4_unset_forwarding(kvm, irq, ...)
its_unmap_vlpi(irq)
[Unmap the VLPI from the ITS (again), remap the original LPI]
Signed-off-by: Jiang Yi <giangyi@amazon.com>
Cc: stable@vger.kernel.org # v4.4+
Fixes: 6d7425f109 ("vfio: Register/unregister irq_bypass_producer")
Link: https://lore.kernel.org/kvm/20191127164910.15888-1-giangyi@amazon.com
Reviewed-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
[aw: commit log]
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Code that iterates over all standard PCI BARs typically uses
PCI_STD_RESOURCE_END. However, that requires the unusual test
"i <= PCI_STD_RESOURCE_END" rather than something the typical
"i < PCI_STD_NUM_BARS".
Add a definition for PCI_STD_NUM_BARS and change loops to use the more
idiomatic C style to help avoid fencepost errors.
Link: https://lore.kernel.org/r/20190927234026.23342-1-efremov@linux.com
Link: https://lore.kernel.org/r/20190927234308.23935-1-efremov@linux.com
Link: https://lore.kernel.org/r/20190916204158.6889-3-efremov@linux.com
Signed-off-by: Denis Efremov <efremov@linux.com>
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Acked-by: Sebastian Ott <sebott@linux.ibm.com> # arch/s390/
Acked-by: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com> # video/fbdev/
Acked-by: Gustavo Pimentel <gustavo.pimentel@synopsys.com> # pci/controller/dwc/
Acked-by: Jack Wang <jinpu.wang@cloud.ionos.com> # scsi/pm8001/
Acked-by: Martin K. Petersen <martin.petersen@oracle.com> # scsi/pm8001/
Acked-by: Ulf Hansson <ulf.hansson@linaro.org> # memstick/
vfio_pci_enable() saves the device's initial configuration information
with the intent that it is restored in vfio_pci_disable(). However,
the commit referenced in Fixes: below replaced the call to
__pci_reset_function_locked(), which is not wrapped in a state save
and restore, with pci_try_reset_function(), which overwrites the
restored device state with the current state before applying it to the
device. Reinstate use of __pci_reset_function_locked() to return to
the desired behavior.
Fixes: 890ed578df ("vfio-pci: Use pci "try" reset interface")
Signed-off-by: hexin <hexin15@baidu.com>
Signed-off-by: Liu Qi <liuqi16@baidu.com>
Signed-off-by: Zhang Yu <zhangyu31@baidu.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Use a vma helper function to simply code.
Signed-off-by: Peng Hao <richard.peng@oppo.com>
Reviewed-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Based on 2 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license version 2 as
published by the free software foundation
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license version 2 as
published by the free software foundation #
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-only
has been chosen to replace the boilerplate/reference in 4122 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Enrico Weigelt <info@metux.net>
Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190604081206.933168790@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Add SPDX license identifiers to all Make/Kconfig files which:
- Have no license information of any form
These files fall under the project license, GPL v2 only. The resulting SPDX
license identifier is:
GPL-2.0-only
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Use dev_printk() when possible to make messages consistent with other
device-related messages.
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Acked-by: Eric Auger <eric.auger@redhat.com>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
When compiling with -Wformat, clang emits the following warnings:
drivers/vfio/pci/vfio_pci.c:1601:5: warning: format specifies type
'unsigned short' but the argument has type 'unsigned int' [-Wformat]
vendor, device, subvendor, subdevice,
^~~~~~
drivers/vfio/pci/vfio_pci.c:1601:13: warning: format specifies type
'unsigned short' but the argument has type 'unsigned int' [-Wformat]
vendor, device, subvendor, subdevice,
^~~~~~
drivers/vfio/pci/vfio_pci.c:1601:21: warning: format specifies type
'unsigned short' but the argument has type 'unsigned int' [-Wformat]
vendor, device, subvendor, subdevice,
^~~~~~~~~
drivers/vfio/pci/vfio_pci.c:1601:32: warning: format specifies type
'unsigned short' but the argument has type 'unsigned int' [-Wformat]
vendor, device, subvendor, subdevice,
^~~~~~~~~
drivers/vfio/pci/vfio_pci.c:1605:5: warning: format specifies type
'unsigned short' but the argument has type 'unsigned int' [-Wformat]
vendor, device, subvendor, subdevice,
^~~~~~
drivers/vfio/pci/vfio_pci.c:1605:13: warning: format specifies type
'unsigned short' but the argument has type 'unsigned int' [-Wformat]
vendor, device, subvendor, subdevice,
^~~~~~
drivers/vfio/pci/vfio_pci.c:1605:21: warning: format specifies type
'unsigned short' but the argument has type 'unsigned int' [-Wformat]
vendor, device, subvendor, subdevice,
^~~~~~~~~
drivers/vfio/pci/vfio_pci.c:1605:32: warning: format specifies type
'unsigned short' but the argument has type 'unsigned int' [-Wformat]
vendor, device, subvendor, subdevice,
^~~~~~~~~
The types of these arguments are unconditionally defined, so this patch
updates the format character to the correct ones for unsigned ints.
Link: https://github.com/ClangBuiltLinux/linux/issues/378
Signed-off-by: Louis Taylor <louis@kragniz.eu>
Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
pci_map_rom/pci_get_rom_size() performs memory access in the ROM.
In case the Memory Space accesses were disabled, readw() is likely
to trigger a synchronous external abort on some platforms.
In case memory accesses were disabled, re-enable them before the
call and disable them back again just after.
Fixes: 89e1f7d4c6 ("vfio: Add PCI device driver")
Signed-off-by: Eric Auger <eric.auger@redhat.com>
Suggested-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
PCI core handles save and restore of device state around reset, but
when using pci_set_power_state() we can unintentionally trigger a soft
reset of the device, where PCI core only restores the BAR state. If
we're using vfio-pci's idle D3 support to try to put devices into low
power when unused, this might trigger a reset when the device is woken
for use. Also power state management by the user, or within a guest,
can put the device into D3 power state with potentially limited
ability to restore the device if it should undergo a reset. The PCI
spec does not define the extent of a soft reset and many devices
reporting soft reset on D3->D0 transition do not undergo a PCI config
space reset. It's therefore assumed safe to unconditionally restore
the remainder of the state if the device indicates soft reset
support, even on a user initiated wakeup.
Implement a wrapper in vfio-pci to tag devices reporting PM reset
support, save their state on transitions into D3 and restore on
transitions back to D0.
Reported-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Using the {0} construct as a generic initializer is perfectly fine in C,
however due to a bug in old gcc there is a warning:
+ /kisskb/src/drivers/vfio/pci/vfio_pci_nvlink2.c: warning: (near
initialization for 'cap.header') [-Wmissing-braces]: => 181:9
Since for whatever reason we still want to compile the modern kernel
with such an old gcc without warnings, this changes the capabilities
initialization.
The gcc bugzilla: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53119
Fixes: 7f92891778 ("vfio_pci: Add NVIDIA GV100GL [Tesla V100 SXM2] subdriver")
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reported-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
The recently added nvlink2 VFIO driver introduced a license conflict in two
files. In both cases the SPDX license identifier is:
SPDX-License-Identifier: GPL-2.0+
but the files contain also the following license boiler plate text:
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 as
* published by the Free Software Foundation
The latter is GPL-2.9-only and not GPL-2.0=.
Looking deeper. The nvlink source file is derived from vfio_pci_igd.c which
is also licensed under GPL-2.0-only and it can be assumed that the file was
copied and modified. As the original file is licensed GPL-2.0-only it's not
possible to relicense derivative work to GPL-2.0-or-later.
Fix the SPDX identifier and remove the boiler plate as it is redundant.
Fixes: 7f92891778 ("vfio_pci: Add NVIDIA GV100GL [Tesla V100 SXM2] subdriver")
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Alexey Kardashevskiy <aik@ozlabs.ru>
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: kvm@vger.kernel.org
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
drivers/vfio/pci/vfio_pci_nvlink2.c cannot be compiled for in-tree
building.
CC drivers/vfio/pci/vfio_pci_nvlink2.o
In file included from drivers/vfio/pci/trace.h:102,
from drivers/vfio/pci/vfio_pci_nvlink2.c:29:
./include/trace/define_trace.h:89:42: fatal error: ./trace.h: No such file or directory
#include TRACE_INCLUDE(TRACE_INCLUDE_FILE)
^
compilation terminated.
make[1]: *** [scripts/Makefile.build;277: drivers/vfio/pci/vfio_pci_nvlink2.o] Error 1
To fix the build error, let's tell include/trace/define_trace.h the
location of drivers/vfio/pci/trace.h
Fixes: 7f92891778 ("vfio_pci: Add NVIDIA GV100GL [Tesla V100 SXM2] subdriver")
Reported-by: Laura Abbott <labbott@redhat.com>
Signed-off-by: Masahiro Yamada <yamada.masahiro@socionext.com>
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
- Replace global vfio-pci lock with per bus lock to allow concurrent
open and release (Alex Williamson)
- Declare mdev function as static (Paolo Cretaro)
- Convert char to u8 in mdev/mtty sample driver (Nathan Chancellor)
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.14 (GNU/Linux)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=HOtX
-----END PGP SIGNATURE-----
Merge tag 'vfio-v4.21-rc1' of git://github.com/awilliam/linux-vfio
Pull VFIO updates from Alex Williamson:
- Replace global vfio-pci lock with per bus lock to allow concurrent
open and release (Alex Williamson)
- Declare mdev function as static (Paolo Cretaro)
- Convert char to u8 in mdev/mtty sample driver (Nathan Chancellor)
* tag 'vfio-v4.21-rc1' of git://github.com/awilliam/linux-vfio:
vfio-mdev/samples: Use u8 instead of char for handle functions
vfio/mdev: add static modifier to add_mdev_supported_type
vfio/pci: Parallelize device open and release
POWER9 Witherspoon machines come with 4 or 6 V100 GPUs which are not
pluggable PCIe devices but still have PCIe links which are used
for config space and MMIO. In addition to that the GPUs have 6 NVLinks
which are connected to other GPUs and the POWER9 CPU. POWER9 chips
have a special unit on a die called an NPU which is an NVLink2 host bus
adapter with p2p connections to 2 to 3 GPUs, 3 or 2 NVLinks to each.
These systems also support ATS (address translation services) which is
a part of the NVLink2 protocol. Such GPUs also share on-board RAM
(16GB or 32GB) to the system via the same NVLink2 so a CPU has
cache-coherent access to a GPU RAM.
This exports GPU RAM to the userspace as a new VFIO device region. This
preregisters the new memory as device memory as it might be used for DMA.
This inserts pfns from the fault handler as the GPU memory is not onlined
until the vendor driver is loaded and trained the NVLinks so doing this
earlier causes low level errors which we fence in the firmware so
it does not hurt the host system but still better be avoided; for the same
reason this does not map GPU RAM into the host kernel (usual thing for
emulated access otherwise).
This exports an ATSD (Address Translation Shootdown) register of NPU which
allows TLB invalidations inside GPU for an operating system. The register
conveniently occupies a single 64k page. It is also presented to
the userspace as a new VFIO device region. One NPU has 8 ATSD registers,
each of them can be used for TLB invalidation in a GPU linked to this NPU.
This allocates one ATSD register per an NVLink bridge allowing passing
up to 6 registers. Due to the host firmware bug (just recently fixed),
only 1 ATSD register per NPU was actually advertised to the host system
so this passes that alone register via the first NVLink bridge device in
the group which is still enough as QEMU collects them all back and
presents to the guest via vPHB to mimic the emulated NPU PHB on the host.
In order to provide the userspace with the information about GPU-to-NVLink
connections, this exports an additional capability called "tgt"
(which is an abbreviated host system bus address). The "tgt" property
tells the GPU its own system address and allows the guest driver to
conglomerate the routing information so each GPU knows how to get directly
to the other GPUs.
For ATS to work, the nest MMU (an NVIDIA block in a P9 CPU) needs to
know LPID (a logical partition ID or a KVM guest hardware ID in other
words) and PID (a memory context ID of a userspace process, not to be
confused with a linux pid). This assigns a GPU to LPID in the NPU and
this is why this adds a listener for KVM on an IOMMU group. A PID comes
via NVLink from a GPU and NPU uses a PID wildcard to pass it through.
This requires coherent memory and ATSD to be available on the host as
the GPU vendor only supports configurations with both features enabled
and other configurations are known not to work. Because of this and
because of the ways the features are advertised to the host system
(which is a device tree with very platform specific properties),
this requires enabled POWERNV platform.
The V100 GPUs do not advertise any of these capabilities via the config
space and there are more than just one device ID so this relies on
the platform to tell whether these GPUs have special abilities such as
NVLinks.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
VFIO regions already support region capabilities with a limited set of
fields. However the subdriver might have to report to the userspace
additional bits.
This adds an add_capability() hook to vfio_pci_regops.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
So far we only allowed mapping of MMIO BARs to the userspace. However
there are GPUs with on-board coherent RAM accessible via side
channels which we also want to map to the userspace. The first client
for this is NVIDIA V100 GPU with NVLink2 direct links to a POWER9
NPU-enabled CPU; such GPUs have 16GB RAM which is coherently mapped
to the system address space, we are going to export these as an extra
PCI region.
We already support extra PCI regions and this adds support for mapping
them to the userspace.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
In commit 61d792562b ("vfio-pci: Use mutex around open, release, and
remove") a mutex was added to freeze the refcnt for a device so that
we can handle errors and perform bus resets on final close. However,
bus resets can be rather slow and a global mutex here is undesirable.
Evaluating the potential locking granularity, a per-device mutex
provides the best resolution but with multiple devices on a bus all
released concurrently, they'll race to acquire each other's mutex,
likely resulting in no reset at all if we use trylock. We therefore
lock at the granularity of the bus/slot reset as we're only attempting
a single reset for this group of devices anyway. This allows much
greater scaling as we're bounded in the number of devices protected by
a single reflck object.
Reported-by: Christian Ehrhardt <christian.ehrhardt@canonical.com>
Tested-by: Christian Ehrhardt <christian.ehrhardt@canonical.com>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Gustavo A. R. Silva