Every caller has a readily available vfio_device pointer, use that instead
of passing in a generic struct device. Change vfio_dma_rw() to take in the
struct vfio_device and move the container users that would have been held
by vfio_group_get_external_user_from_dev() to vfio_dma_rw() directly, like
vfio_pin/unpin_pages().
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Link: https://lore.kernel.org/r/4-v4-8045e76bf00b+13d-vfio_mdev_no_group_jgg@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Instead of having an option to build the gvt code into the main i915
module, just move it into the kvmgt.ko module. This only requires
a new struct with three entries that the KVMGT modules needs to register
with the main i915 module, and a proper list of GVT-enabled devices
instead of global device pointer.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Zhi Wang <zhi.a.wang@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20220411141403.86980-7-hch@lst.de
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Reviewed-by: Zhi Wang <zhi.a.wang@intel.com>
To support the new mdev interfaces and the re-factor patches from
Christoph, which moves the GVT-g code into a dedicated module, the GVT-g
MMIO tracking table needs to be separated from GVT-g.
v9:
- Fix a problem might cause kernel panic.
- Remove the redaundant definitation of intel_get_device_type(). (Jani)
- Sort the list of header reference in intel_gvt_mmio.c (Jani)
- Include minimum header insted in intel_gvt_mmio.c (Jani)
v8:
- Use SPDX header in the intel_gvt_mmio_table.c
- Reference the gvt.h with path. (Jani)
- Add a missing fix on mmio emulation path during the debug.
- Fix a building problem on refreshed gvt-staging branch. (Christoph)
v7:
- Keep the marcos of device generation in GVT-g. (Christoph, Jani)
v6:
- Move the mmio_table.c into i915. (Christoph)
- Keep init_device_info and related structures in GVT-g. (Christoph)
- Refine the callbacks of the iterator. (Christoph)
- Move the flags of MMIO register defination to GVT-g. (Chrsitoph)
- Move the mmio block handling to GVT-g.
v5:
- Re-design the mmio table framework. (Christoph)
v4:
- Fix the errors of patch checking scripts.
v3:
- Fix the errors when CONFIG_DRM_I915_WERROR is turned on. (Jani)
v2:
- Implement a mmio table instead of generating it by marco in i915. (Jani)
Cc: Christoph Hellwig <hch@lst.de>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: Jani Nikula <jani.nikula@linux.intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Vivi Rodrigo <rodrigo.vivi@intel.com>
Cc: Zhenyu Wang <zhenyuw@linux.intel.com>
Cc: Zhi Wang <zhi.a.wang@intel.com>
Signed-off-by: Zhi Wang <zhi.a.wang@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Tested-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Acked-by: Jani Nikula <jani.nikula@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20220407071945.72148-2-zhi.a.wang@intel.com
As kvmgt module contains all handling for VFIO/mdev, leaving mdev attribute
groups in gvt module caused dependency issue. Although it was there for possible
other hypervisor usage, that turns out never to be true. So this moves all mdev
handling into kvmgt module completely to resolve dependency issue.
With this fix, no config workaround is required. So revert previous workaround
commits: adaeb718d4 ("vfio/gvt: fix DRM_I915_GVT dependency on VFIO_MDEV")
and 07e543f4f9 ("vfio/gvt: Make DRM_I915_GVT depend on VFIO_MDEV").
Reviewed-by: Colin Xu <colin.xu@intel.com>
Cc: Arnd Bergmann <arnd@kernel.org>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20210513083902.2822350-1-zhenyuw@linux.intel.com
Pull VFIO updates from Alex Williamson:
- Embed struct vfio_device into vfio driver structures (Jason
Gunthorpe)
- Make vfio_mdev type safe (Jason Gunthorpe)
- Remove vfio-pci NVLink2 extensions for POWER9 (Christoph Hellwig)
- Update vfio-pci IGD extensions for OpRegion 2.1+ (Fred Gao)
- Various spelling/blank line fixes (Zhen Lei, Zhou Wang, Bhaskar
Chowdhury)
- Simplify unpin_pages error handling (Shenming Lu)
- Fix i915 mdev Kconfig dependency (Arnd Bergmann)
- Remove unused structure member (Keqian Zhu)
* tag 'vfio-v5.13-rc1' of git://github.com/awilliam/linux-vfio: (43 commits)
vfio/gvt: fix DRM_I915_GVT dependency on VFIO_MDEV
vfio/iommu_type1: Remove unused pinned_page_dirty_scope in vfio_iommu
vfio/mdev: Correct the function signatures for the mdev_type_attributes
vfio/mdev: Remove kobj from mdev_parent_ops->create()
vfio/gvt: Use mdev_get_type_group_id()
vfio/gvt: Make DRM_I915_GVT depend on VFIO_MDEV
vfio/mbochs: Use mdev_get_type_group_id()
vfio/mdpy: Use mdev_get_type_group_id()
vfio/mtty: Use mdev_get_type_group_id()
vfio/mdev: Add mdev/mtype_get_type_group_id()
vfio/mdev: Remove duplicate storage of parent in mdev_device
vfio/mdev: Add missing error handling to dev_set_name()
vfio/mdev: Reorganize mdev_device_create()
vfio/mdev: Add missing reference counting to mdev_type
vfio/mdev: Expose mdev_get/put_parent to mdev_private.h
vfio/mdev: Use struct mdev_type in struct mdev_device
vfio/mdev: Simplify driver registration
vfio/mdev: Add missing typesafety around mdev_device
vfio/mdev: Do not allow a mdev_type to have a NULL parent pointer
vfio/mdev: Fix missing static's on MDEV_TYPE_ATTR's
...
Current vblank emulator uses single hrtimer at 16ms period for all vGPUs,
which introduces three major issues:
- 16ms matches the refresh rate at 62.5Hz (instead of 60Hz) which
doesn't follow standard timing. This leads to some frame drop or glitch
issue during video playback. SW expects a vsync interval of 16.667ms or
higher precision for an accurate 60Hz refresh rate. However current
vblank emulator only works at 16ms.
- Doesn't respect the fact that with current virtual EDID timing set,
not all resolutions are running at 60Hz. For example, current virtual
EDID also supports refresh rate at 56Hz, 59.97Hz, 60Hz, 75Hz, etc.
- Current vblank emulator use single hrtimer for all vGPUs. Regardsless
the possibility that different guests could run in different
resolutions, all vsync interrupts are injected at 16ms interval with
same hrtimer.
Based on previous patch which decode guest expected refresh rate from
vreg, the vblank emulator refactor patch makes following changes:
- Change the vblank emulator hrtimer from gvt global to per-vGPU.
By doing this, each vGPU display can operates at different refresh
rates. Currently only one dislay is supported for each vGPU so per-vGPU
hrtimer is enough. If multiple displays are supported per-vGPU in
future, we can expand to per-PIPE further.
- Change the fixed hrtimer period from 16ms to dynamic based on vreg.
GVT is expected to emulate the HW as close as possible. So reflacting
the accurate vsync interrupt interval is more correct than fixed 16ms.
- Change the vblank timer period and start the timer on PIPECONF change.
The initial period is updated to 16666667 based on 60Hz refresh rate.
According to PRM, PIPECONF controls the timing generator of the
connected display on this pipe, so it's safe to stop hrtimer on
PIPECONF disabling, and re-start hrtimer at new period on enabling.
Other changes including:
- Move vblank_timer_fn from irq.c into display.c.
- Clean per-vGPU vblank timer at clean_display instead of clean_irq.
To run quick test, launch a web browser and goto URL: www.displayhz.com
The actual refresh rate from guest can now always match guest settings.
V2:
Rebase to 5.11.
Remove unused intel_gvt_clean_irq().
Simplify enable logic in update_vblank_emulation(). (zhenyu)
Loop all vGPU by idr when check all vblank timer. (zhenyu)
Signed-off-by: Colin Xu <colin.xu@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20210226044630.284269-1-colin.xu@intel.com
Reviewed-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Guest OS builds up its timing mode list based on the virtual EDID as
simulated by GVT. However since there are several timings supported in
the virtual EDID, and each timing can also support several modes
(resolution and refresh rate), current emulated vblank period (16ms)
may not always be correct and could lead to miss-sync behavior in guest.
Guest driver will setup new resolution and program vregs accordingly and
it should always follows GEN PRM. Based on the simulated display regs by
GVT, it's safe to decode the actual refresh rate using by guest from
vreg only.
Current implementation only enables PIPE_A and PIPE_A is always tied to
TRANSCODER_A in HW. GVT may simulate DP monitor on PORT_B or PORT_D
based on the caller. So we can find out which DPLL is used by PORT_x
which connected to TRANSCODER_A and calculate the DP bit rate from the
DPLL frequency. Then DP stream clock (pixel clock) can be calculated
from DP link M/N and DP bit rate. Finally, get the refresh rate from
pixel clock, H total and V total.
The per-vGPU accurate refresh rate is not used yet but only stored,
until per-vGPU vblank timer is enabled. Then each vGPU can have
different and accurate refresh rate per-guest driver configuration.
Refer to PRM for GEN display and VESA timing standard for more details.
V2:
Rebase to 5.11.
Correctly calculate DP link rate for BDW and BXT.
Use GVT_DEFAULT_REFRESH_RATE instead of hardcoded to 60 as init refresh.
Typo fix. (zhenyu)
Signed-off-by: Colin Xu <colin.xu@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20210226044559.283622-1-colin.xu@intel.com
Reviewed-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Logical Context is actually a big batch buffer consisting of multiple
LRI commands + saved registers. It comprises Ring Context (the first
0x50 dwords) and Engine Context. The registers defined in Engine Context
are command accessible, and safe to execute in VM Context.
However, not all of them are currently tracked in existing register
whitelist. Here we kick hardware to generate a dummy Engine Context and
then scan the dummy Engine context to update whitelist dynamically. Based
on updated list, later patches will audit future VM Engine Contexts to
disallow undesired LRIs (if out of what hardware generates).
Cc: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Wang Zhi <zhi.a.wang@intel.com>
Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20201223034430.16934-1-yan.y.zhao@intel.com
Reviewed-by: Zhenyu Wang <zhenyuw@linux.intel.com>
This patch save/restore necessary GVT info during i915 suspend/resume so
that GVT enabled QEMU VM can continue running.
Only GGTT and fence regs are saved/restored now. GVT will save GGTT
entries on each host_entry update, restore the saved dirty entries
and re-init fence regs in resume routine.
V2:
- Change kzalloc/kfree to vzalloc/vfree since the space allocated
from kmalloc may not enough for all saved GGTT entries.
- Keep gvt suspend/resume wrapper in intel_gvt.h/intel_gvt.c and
move the actual implementation to gvt.h/gvt.c. (zhenyu)
- Check gvt config on and active with intel_gvt_active(). (zhenyu)
V3: (zhenyu)
- Incorrect copy length. Should be num entries * entry size.
- Use memcpy_toio()/memcpy_fromio() instead of memcpy for iomem.
- Add F_PM_SAVE flags to indicate which MMIOs to save/restore for PM.
V4:
Rebase.
V5:
Fail intel_gvt_save_ggtt as -ENOMEM if fail to alloc memory to save
ggtt. Free allocated ggtt_entries on failure.
V6:
Save host entry to per-vGPU gtt.ggtt_mm on each host_entry update.
V7:
Restore GGTT entry based on present bit.
Split fence restore and mmio restore in different functions.
Reviewed-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Signed-off-by: Hang Yuan <hang.yuan@linux.intel.com>
Signed-off-by: Colin Xu <colin.xu@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20201027045308.158955-1-colin.xu@intel.com
When system enters S3 state, device enters D3 state while RAM remains
powered. From vGPU/GVT perspective, ppgtt_mm is residual in guest memory
during vGPU in D3 state, so that when guest state transits from S3->S0,
ppgtt_mm can be re-used and no need rebuild.
Previous implementation invalidate and destroy ppgtt_mm at DMLR,
regardless the power state transition is S0->S3->S0 (guest suspend or
resume) or OFF->S0 (normal boot/reboot), invalidate and destroy ppgtt_mm
is unnecessary in the former transition case.
The patch saves the vGPU D3/D0 transition state when guest writes the
PCI_PM_CTRL in vGPU's configure space, then in later DMLR, GVT can decide
whether or not invalidate and destroy ppgtt_mm is required. The
d3_entered flags is reset after DMLR.
To test this feature, make sure S3 is enabled in QEMU parameters:
i440fx: PIIX4_PM.disable_s3=0
q35: ICH9-LPC.disable_s3=0
Also need enable sleep option in guest OS if it's disabled.
v2:
- Revise commit message to more accurate description. (Kevin)
- Split patch by logic. (Zhenyu)
Reviewed-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Signed-off-by: Hang Yuan <hang.yuan@linux.intel.com>
Signed-off-by: Colin Xu <colin.xu@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200709071002.247960-2-colin.xu@intel.com
Software is not expected to populate engine context except when using
restore inhibit bit or golden state to initialize it for the first time.
Therefore, if a newly submitted guest context is the same as the last
shadowed one, no need to populate its engine context from guest again.
Currently using lrca + ring_context_gpa to identify whether two guest
contexts are the same.
The reason of why context id is not included as an identifier is that
i915 recently changed the code and context id is only unique for a
context when OA is enabled. And when OA is on, context id is generated
based on lrca. Therefore, in that case, if two contexts are of the same
lrca, they have identical context ids as well.
(This patch also works with old guest kernel like 4.20.)
for guest context, if its ggtt entry is modified after last context
shadowing, it is also deemed as not the same context as last shadowed one.
v7:
-removed local variable "valid". use the one in s->last_ctx diretly
v6:
-change type of lrca of last ctx to be u32. as currently it's all
protected by vgpu lock (Kevin Tian)
-reset valid of last ctx to false once it needs to be repopulated before
population completes successfully (Kevin Tian)
v5:
-merge all 3 patches into one patch (Zhenyu Wang)
v4:
- split the series into 3 patches.
- don't turn on optimization until last patch in this series (Kevin Tian)
- define lrca to be atomic in this patch rather than update its type in
the second patch (Kevin Tian)
v3: updated commit message to describe engine context and context id
clearly (Kevin Tian)
v2: rebased to 5.6.0-rc4+Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
Reviewed-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Cc: Kevin Tian <kevin.tian@intel.com>
Suggested-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200417091334.32628-1-yan.y.zhao@intel.com
Factor out tlb and mocs register offset table to fix the issues reported
by klocwork, #512 and #550. Mostly, the reason why the klocwork reports
these problems is because there can be possbilities for platforms, which
have more rings than the ring offset table, to take the dirty data from
the stack as the register offset. It results to a random HW register
offset writting in this scenairo when doing context switch between vGPUs.
After the factoring, the ring offset table of TLB and MOCS should be per
platform.
v2:
- Enable TLB register switch for GEN8. (Zhenyu)
Reviewed-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Signed-off-by: Zhi Wang <zhi.a.wang@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Our eventual goal is to rid request construction of struct_mutex, with
the short term step of lifting the struct_mutex requirements into the
higher levels (i.e. the caller must ensure that the context is already
pinned into the GTT). In this patch, we pin GVT's shadow context upon
allocation and so keep them pinned into the GGTT for as long as the
virtual machine is alive, and so we can use the simpler request
construction path safe in the knowledge that the hard work is already
done.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Zhenyu Wang <zhenyuw@linux.intel.com>
Acked-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190426163336.15906-1-chris@chris-wilson.co.uk
