VirtioGpuDxe is now IOMMU-clean; it translates system memory addresses to
bus master device addresses. Negotiate VIRTIO_F_IOMMU_PLATFORM in parallel
with VIRTIO_F_VERSION_1. (Note: the VirtIo GPU device, and this driver,
are virtio-1.0 only (a.k.a. "modern-only").)
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
VirtioGpuDxe is a UEFI Bus driver (not a Device driver). This is because a
UEFI graphics driver is expected to produce its GraphicsOutput protocol
instance(s) on new child handle(s) of the video controller handle, one
child handle (plus GOP) per video output (or, one child handle plus GOP
per combination of multiple video outputs).
In VirtioGpuDxe, we support a single VirtIo GPU head (scanout), namely
head#0. This means that, with regard to a specific VirtIo GPU device, the
driver may be in one of three states, at any time:
[1] VirtioGpuDxe has not bound the device at all,
[2] VirtioGpuDxe has bound the device, but not produced the sole child
handle for head#0,
[3] VirtioGpuDxe has bound the device, and produced the sole child handle
for head#0, with a GOP instance on the child handle.
(Which state the driver is in wrt. a given VirtIo GPU device depends on
the VirtioGpuDriverBindingStart() / VirtioGpuDriverBindingStop()
invocations issued by the ConnectController() / DisconnectController()
boot services. In turn those come from BDS or e.g. the UEFI shell.)
The concept of "current video mode" is technically tied to the GOP (i.e.,
the child handle, state [3] only), not the VirtIo GPU controller handle.
This is why we manage the storage that backs the current video mode in our
EFI_GRAPHICS_OUTPUT_PROTOCOL.SetMode() member implementation.
GopSetMode() is first called *internally*, when we enter state [3] (that
is, when we produce the child handle + GOP on it):
VirtioGpuDriverBindingStart() [DriverBinding.c]
InitVgpuGop() [DriverBinding.c]
VgpuGop->Gop.SetMode() [Gop.c]
When this happens, we allocate the backing store *without* having a
preexistent backing store (due to no preexistent video mode and GOP).
Skipping VirtIo GPU details not relevant for this patch, we just note that
the backing store is exposed *permanently* to the VirtIo GPU device, with
the RESOURCE_ATTACH_BACKING command.
When external clients call the EFI_GRAPHICS_OUTPUT_PROTOCOL.Blt() member
function -- called GopBlt() in this driver --, in state [3], the function
operates on the backing store, and sends only small messages to the VirtIo
GPU device.
When external clients call GopSetMode() for switching between video modes
-- in state [3] --, then
- a new backing store is allocated and exposed to the device (attached to
a new host-side VirtIo GPU resource),
- head#0 is flipped to the new backing store,
- on success, the ReleaseGopResources() function both detaches the
previous backing store from the VirtIo GPU device, an releases it. The
new backing store address and size are saved in our GOP object. (In
other words, we "commit" to the new video mode.)
When the DisconnectController() boot service asks us to leave state [3] --
we can leave it directly only for state [2] --, then the
ReleaseGopResources() function is called on a different path:
VirtioGpuDriverBindingStop() [DriverBinding.c]
UninitVgpuGop() [DriverBinding.c]
ReleaseGopResources() [Gop.c]
In this case, the backing store being released is still in use (we're not
leaving it for a new mode -- head#0 has not been flipped "away" from it),
so in ReleaseGopResources() we disable head#0 first.
(The ReleaseGopResources() function is called the same way on the error
path in InitVgpuGop(), if the first -- internal -- VgpuGop->Gop.SetMode()
call succeeds, but the rest of InitVgpuGop() fails.)
Based on the above, for IOMMU-compatibility,
- in GopSetMode(), don't just allocate, but also map the backing store of
the nascent video mode to a device address, for bus master common buffer
operation,
- (the VirtioGpuAllocateZeroAndMapBackingStore() helper function
introduced in the last patch takes care of zeroing internally,)
- pass the device address to the VirtIo GPU device in the
RESOURCE_ATTACH_BACKING command,
- if GopSetMode() succeeds, save the mapping token,
- if GopSetMode() fails, don't just free but also unmap the still-born
backing store,
- in ReleaseGopResources(), don't just free but also unmap the backing
store -- which is the previous backing store if we're mode-switching,
and the current backing store if we're leaving state [3].
Finally, ExitBootServices() may be called when the driver is in either
state [1], [2] or [3], wrt. a given VirtIo GPU device. (Of course we are
only notified in states [2] and [3].) If we get the notification in state
[3], then the current video mode's backing store has to be unmapped, but
not released. (We must not change the UEFI memory map.)
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
Introduce the VirtioGpuAllocateZeroAndMapBackingStore() and
VirtioGpuUnmapAndFreeBackingStore() helper functions. These functions tie
together the allocation, zeroing and mapping, and unmapping and
deallocation, respectively, of memory that the virtio GPU will permanently
reference after receiving the RESOURCE_ATTACH_BACKING command.
With these functions we can keep the next patch simpler -- the GOP
implementation in "Gop.c" retains its error handling structure, and
remains oblivious to VIRTIO_DEVICE_PROTOCOL and VirtioLib.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
The RESOURCE_ATTACH_BACKING virtio GPU command assigns guest-side backing
pages to a host-side resource that was created earlier with the
RESOURCE_CREATE_2D command.
We compose the RESOURCE_ATTACH_BACKING command in the
VirtioGpuResourceAttachBacking() function. Currently this function takes
the parameter
IN VOID *FirstBackingPage
This is only appropriate as long as we pass a (guest-phys) system memory
address to the device. In preparation for a mapped bus master device
address, change the above parameter to
IN EFI_PHYSICAL_ADDRESS BackingStoreDeviceAddress
In order to keep the current call site functional, move the (VOID*) to
(UINTN) conversion out of the function, to the call site.
The "Request.Entry.Addr" field already has type UINT64.
This patch is similar to commit 4b725858de ("OvmfPkg/VirtioLib: change
the parameter of VirtioAppendDesc() to UINT64", 2017-08-23).
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
Every virtio GPU command used by VirtioGpuDxe is synchronous and formatted
as a two-descriptor chain: request, response. The internal workhorse
function that all the command-specific functions call for such messaging
is VirtioGpuSendCommand().
In VirtioGpuSendCommand(), map the request from system memory to bus
master device address for BusMasterRead operation, and map the response
from system memory to bus master device address for BusMasterWrite
operation.
Pass the bus master device addresses to VirtioAppendDesc(). (See also
commit 4b725858de, "OvmfPkg/VirtioLib: change the parameter of
VirtioAppendDesc() to UINT64", 2017-08-23.)
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
VirtioGpuDxe uses one virtio ring, for VIRTIO_GPU_CONTROL_QUEUE.
Map it for bus master common buffer operation with VirtioRingMap(), so
that it can be accessed equally by both guest and hypervisor even if an
IOMMU is used. (VirtioRingInit() already allocates the ring suitably for
this, see commit b0338c5329, "OvmfPkg/VirtioLib: alloc VRING buffer with
AllocateSharedPages()", 2017-08-23).
Pass the resultant translation offset ("RingBaseShift"), from system
memory address to bus master device address, to VIRTIO_SET_QUEUE_ADDRESS.
Unmap the ring in all contexts where the ring becomes unused (these
contexts are mutually exclusive):
- in VirtioGpuInit(): the ring has been mapped, but we cannot complete the
virtio initialization for another reason,
- in VirtioGpuUninit(): the virtio initialization has succeeded, but
VirtioGpuDriverBindingStart() fails for another reason, or
VirtioGpuDriverBindingStop() unbinds the device after use,
- in VirtioGpuExitBoot(): ExitBootServices() is called after
VirtioGpuDriverBindingStart() has successfully bound the device.
(Unmapping the ring does not change the UEFI memory map.)
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
Log all relevant IN parameters on entry. (There are only IN parameters.)
Beautify the format string.
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Brijesh Singh <brijesh.singh@amd.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
Log all relevant IN and IN OUT parameters on entry.
(Note that the HostAddress parameter is IN OUT rather than OUT due to
historical reasons. The "IN EFI_ALLOCATE_TYPE Type" parameter is now to be
ignored, but historically it could be set to AllocateMaxAddress for
example, and for that HostAddress had to be IN OUT.)
When exiting with success, log all relevant OUT parameters (i.e.,
HostAddress). Also log the new (internal) StashBuffer address, on which
IoMmuMap() and IoMmuUnmap() rely on, for BusMasterCommonBuffer operations
(in-place decryption and encryption, respectively).
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Brijesh Singh <brijesh.singh@amd.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
The only important external information for this function, and for the
human looking at the log, is the Mapping input parameter. Log it on entry.
Stop logging the contents of the MAP_INFO structure pointed-to by Mapping.
Thanks to the previous patch, we can now associate IoMmuUnmap() messages
with IoMmuMap() messages -- and thereby with MAP_INFO contents -- purely
via Mapping.
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Brijesh Singh <brijesh.singh@amd.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
Log all relevant IN and IN OUT parameters on entry.
When exiting with success, log all relevant OUT and IN OUT parameters.
Don't log OUT and IN OUT parameters that are never set or changed after
entering the function (i.e., *NumberOfBytes).
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Brijesh Singh <brijesh.singh@amd.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
Debug messages that start as natural (English) language phrases (after the
debug prefix) should uniformly begin with lower-case or upper-case. In
SetMemoryEncDec() we have a mixture now. Stick with lower-case.
(Upper-case is better for full sentences that also end with punctuation.)
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Brijesh Singh <brijesh.singh@amd.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
In SetMemoryEncDec(), we have four locations where we (a) log a message on
the DEBUG_WARN level that says "ERROR", (b) return the status code
RETURN_NO_MAPPING right after.
These messages clearly describe actual errors (bad PML4, PDPE, PDE, PTE).
Promote their debug levels to DEBUG_ERROR, and remove the word "ERROR"
from the messages.
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Brijesh Singh <brijesh.singh@amd.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
In the SetMemoryEncDec() function, the way we currently report
PhysicalAddress is not uniform:
- mostly we say "for %lx",
- in one spot we say "at %lx" (even though the 2MB page being split does
not live *at* PhysicalAddress, instead it maps PhysicalAddress),
- in another spot we don't log PhysicalAddress at all (when splitting a
1GB page).
Unify this, using the format string "for Physical=0x%Lx".
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Brijesh Singh <brijesh.singh@amd.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
None of the DEBUG macro invocations in SetMemoryEncDec() fit on a single
line. Break them to multiple lines, for (a) conforming to the coding style
spec, (b) easier modification in later patches.
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Brijesh Singh <brijesh.singh@amd.com>
Tested-by: Brijesh Singh <brijesh.singh@amd.com>
VirtioScsiDxe driver has been updated to use IOMMU-like member functions
from VIRTIO_DEVICE_PROTOCOL to translate the system physical address to
device address. We do not need to do anything special when
VIRTIO_F_IOMMU_PLATFORM bit is present hence treat it in parallel with
VIRTIO_F_VERSION_1.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Laszlo Ersek <lersek@redhat.com>
When device is behind the IOMMU, driver is require to pass the device
address of virtio request, response and any memory referenced by those
request/response to the bus master.
The patch uses IOMMU-like member functions from VIRTIO_DEVICE_PROTOCOL to
map request and response buffers system physical address to the device
address.
- If the buffer need to be accessed by both the processor and a bus
master then map with BusMasterCommonBuffer.
- If the buffer need to be accessed for a write operation by a bus master
then map with BusMasterWrite.
However, after a BusMasterWrite Unmap() failure, error reporting via
EFI_EXT_SCSI_PASS_THRU_SCSI_REQUEST_PACKET would be very complex,
therefore we map such buffers too with BusMasterCommonBuffer.
- If the buffer need to be accessed for a read operation by a bus master
then map with BusMasterRead.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
[lersek@redhat.com: restore lost sentence/paragraph in commit message]
[lersek@redhat.com: reindent/reflow "InDataBuffer" comment block]
[lersek@redhat.com: cast arg, not result, of EFI_SIZE_TO_PAGES() to UINTN]
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Laszlo Ersek <lersek@redhat.com>
When virtio request fails we return EFI_DEVICE_ERROR, as per the spec
EFI_EXT_SCSI_PASS_THRU_PROTOCOL.PassThru() member function is required
to implement elaborated error reporting.
The patch refactors out entire block of the code that creates the host
adapter error into a separate helper function (ReportHostAdapterError).
Suggested-by: Laszlo Ersek <lersek@redhat.com>
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
[lersek@redhat.com: fix style & typo in ReportHostAdapterError() comment]
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Laszlo Ersek <lersek@redhat.com>
When device is behind the IOMMU then driver need to pass the device
address when programing the bus master. The patch uses VirtioRingMap() to
map the VRING system physical address to device address.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Laszlo Ersek <lersek@redhat.com>
when "RequestIsWrite" is FALSE -- i.e., the CPU wants data from
the device, we map "Buffer" for VirtioOperationBusMasterWrite. In
this case, checking the return status of
Dev->VirtIo->UnmapSharedBuffer (Dev->VirtIo, BufferMapping);
is must. If the unmapping fails, then "Buffer" will not contain the
actual data from the device, and we must fail the request with
EFI_DEVICE_ERROR.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
[lersek@redhat.com: fix typos in subject]
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Laszlo Ersek <lersek@redhat.com>
There's a small window between
- AllocFwCfgDmaAccessBuffer() mapping the new FW_CFG_DMA_ACCESS object for
common buffer operation (i.e., decrypting it), and
- InternalQemuFwCfgDmaBytes() setting the fields of the object.
In this window, earlier garbage in the object is "leaked" to the
hypervisor. So zero the object before we decrypt it.
(This commit message references AMD SEV directly, because QemuFwCfgDxeLib
is not *generally* enabled for IOMMU operation just yet, unlike our goal
for the virtio infrastructure. Instead, QemuFwCfgDxeLib uses
MemEncryptSevLib explicitly to detect SEV, and then relies on IOMMU
protocol behavior that is specific to SEV. At this point, this is by
design.)
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Brijesh Singh <brijesh.singh@amd.com>
VirtioBlkDxe driver has been updated to use IOMMU-like member functions
from VIRTIO_DEVICE_PROTOCOL to translate the system physical address to
device address. We do not need to do anything special when
VIRTIO_F_IOMMU_PLATFORM bit is present hence treat it in parallel with
VIRTIO_F_VERSION_1.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
When device is behind the IOMMU, driver is require to pass the device
address of virtio request, response and any memory referenced by those
request/response to the bus master.
The patch uses IOMMU-like member functions from VIRTIO_DEVICE_PROTOCOL to
map request and response buffers system physical address to the device
address.
- If the buffer need to be accessed by both the processor and a bus
master then map with BusMasterCommonBuffer.
- If the buffer need to be accessed for a write operation by a bus master
then map with BusMasterWrite.
- If the buffer need to be accessed for a read operation by a bus master
then map with BusMasterRead.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
When device is behind the IOMMU then driver need to pass the device
address when programing the bus master. The patch uses VirtioRingMap() to
map the VRING system physical address to device address.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
This reverts commit ca56256d5e0b7e63325b049e90a6bd03f90e3598:
TianoCore BZ#671 <https://bugzilla.tianocore.org/show_bug.cgi?id=671> has
been fixed in commit 2f7f1e73c1 ("BaseTools: Add the missing -pie link
option in GCC tool chain", 2017-08-23), so we can return to the GCC5
toolchain with gcc-7.*.
Cc: Jordan Justen <jordan.l.justen@intel.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
VirtioRngDxe driver has been updated to use IOMMU-like member functions
from VIRTIO_DEVICE_PROTOCOL to translate the system physical address to
device address. We do not need to do anything special when
VIRTIO_F_IOMMU_PLATFORM bit is present hence treat it in parallel with
VIRTIO_F_VERSION_1.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
This feature indicates that the device is behind an IOMMU that translates
bus addresses from the device into physical addresses in memory. If this
feature bit is set to 0, then the device emits physical addresses which
are not translated further, even though an IOMMU may be present.
see [1] for more infromation
[1] https://lists.oasis-open.org/archives/virtio-dev/201610/msg00121.html
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
patch maps the host address to a device address for buffers (including
rings, device specifc request and response pointed by vring descriptor,
and any further memory reference by those request and response).
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
[lersek@redhat.com: return EFI_DEVICE_ERROR if mapping fails in GetRNG]
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
The patch change the "BufferPhysAddr" parameter of VirtioAppendDesc()
from type UINTN to UINT64.
UINTN is appropriate as long as we pass system memory references. After
the introduction of bus master device addresses, that's no longer the case
in general. Should we implement "real" IOMMU support at some point, UINTN
could break in 32-bit builds of OVMF.
Suggested-by: Laszlo Ersek <lersek@redhat.com>
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
[lersek@redhat.com: clarify commit message]
[lersek@redhat.com: balance parens in VirtioAppendDesc() comment blocks]
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
The VRING buffer is a communication area between guest and hypervisor.
Allocate it using VIRTIO_DEVICE_PROTOCOL.AllocateSharedPages() so that
it can be mapped later with VirtioRingMap() for bi-directional access.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
[lersek@redhat.com: correct typo in VirtioRingInit() comment blocks]
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
Add a function to map the ring buffer with BusMasterCommonBuffer so that
ring can be accessed by both guest and hypervisor.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
[lersek@redhat.com: fix typo in commit message]
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
virtio drivers use VIRTIO_DEVICE_PROTOCOL.MapSharedBuffer() to map the
ring buffer host address to a device address. If an IOMMU is present then
RingBaseShift contains the offset from the host address.
Suggested-by: Laszlo Ersek <lersek@redhat.com>
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
For the case when an IOMMU is used for translating system physical
addresses to DMA bus master addresses, the transport-independent
virtio device drivers will be required to map their VRING areas to
bus addresses with VIRTIO_DEVICE_PROTOCOL.MapSharedBuffer() calls.
- MMIO and legacy virtio transport do not support IOMMU to translate the
addresses hence RingBaseShift will always be set to zero.
- modern virtio transport supports IOMMU to translate the address, in
next patch we will update the Virtio10Dxe to use RingBaseShift offset.
Suggested-by: Laszlo Ersek <lersek@redhat.com>
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
[lersek@redhat.com: remove commit msg paragraph with VirtioLib reference]
[lersek@redhat.com: fix typo in VIRTIO_SET_QUEUE_ADDRESS comment block]
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
The function can be used for mapping the system physical address to virtio
device address using VIRTIO_DEVICE_PROTOCOL.MapSharedBuffer (). The
function helps with centralizing error handling, and it allows the caller
to pass in constant or other evaluated expressions for NumberOfBytes.
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
[lersek@redhat.com: s/This/VirtIo/ in the new function's comment blocks]
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
The patch extends VIRTIO_DEVICE_PROTOCOL to provide the following new
member functions:
- AllocateSharedPages : allocate a memory region suitable for sharing
between guest and hypervisor (e.g ring buffer).
- FreeSharedPages: free the memory allocated using AllocateSharedPages ().
- MapSharedBuffer: map a host address to device address suitable to share
with device for bus master operations.
- UnmapSharedBuffer: unmap the device address obtained through the
MapSharedBuffer().
We're free to extend the protocol structure without changing the protocol
GUID, or bumping any protocol version fields (of which we currently have
none), because VIRTIO_DEVICE_PROTOCOL is internal to edk2 by design --
see the disclaimers in "VirtioDevice.h".
The patch implements Laszlo's recommendation [1].
[1] http://mid.mail-archive.com/841bec5f-6f6e-8b1f-25ba-0fd37a915b72@redhat.com
Suggested-by: Laszlo Ersek <lersek@redhat.com>
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Laszlo Ersek <lersek@redhat.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Brijesh Singh <brijesh.singh@amd.com>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
Regression-tested-by: Laszlo Ersek <lersek@redhat.com>
Now that we have dropped QemuVideoDxe from all QEMU targeted builds
under ArmVirtPkg, we can revert the ARM specific changes to it.
This partially reverts commits 84a75f70e9 (SVN 16890) and
05a5379458.
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-by: Laszlo Ersek <lersek@redhat.com>
When UefiCpuPkg/MpInitLib is built for X64 with gcc-7, using the DEBUG
build target and the GCC5 toolchain settings, a C-language assignment is
miscompiled such that the initial AP startup hangs in CpuMpPei (X64) or
CpuDxe (Ia32X64). See <https://bugzilla.tianocore.org/show_bug.cgi?id=671>
for a detailed analysis of the symptoms, and for mailing list links.
This issue has been reported several times (one example is
<https://bugzilla.tianocore.org/show_bug.cgi?id=657>). Until we (or the
upstream gcc developers) figure out how to dissuade gcc-7 from the
miscompilation, pick the GCC49 toolchain in "build.sh" for gcc-7.*.
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
In commit 8057622527 ("OvmfPkg/AcpiPlatformDxe: save fw_cfg boot script
with QemuFwCfgS3Lib", 2017-02-23), we replaced the explicit S3 boot script
manipulation in TransferS3ContextToBootScript() with a call to
QemuFwCfgS3CallWhenBootScriptReady(). (Passing AppendFwCfgBootScript() as
callback.)
QemuFwCfgS3CallWhenBootScriptReady() checks for fw_cfg DMA up-front, and
bails with RETURN_NOT_FOUND if fw_cfg DMA is missing.
(This is justified as the goal of QemuFwCfgS3Lib is to "enable[] driver
modules [...] to produce fw_cfg DMA operations that are to be replayed at
S3 resume time".)
In turn, if QemuFwCfgS3CallWhenBootScriptReady() fails, then
OvmfPkg/AcpiPlatformDxe rolls back any earlier linker/loader script
processing, and falls back to the built-in ACPI tables.
(This is also justified because failure to save WRITE_POINTER commands for
replaying at S3 resume implies failure to process the linker/loader script
comprehensively.)
Calling QemuFwCfgS3CallWhenBootScriptReady() from
TransferS3ContextToBootScript() *unconditionally* is wrong however. For
the case when the linker/loader script contains no WRITE_POINTER commands,
the call perpetuated an earlier side effect, and introduced another one:
(1) On machine types that provide fw_cfg DMA (i.e., 2.5+),
QemuFwCfgS3CallWhenBootScriptReady() would succeed, and allocate
workspace for the boot script opcodes in reserved memory. However, no
opcodes would actually be produced in the AppendFwCfgBootScript()
callback, due to lack of any WRITE_POINTER commands.
This waste of reserved memory had been introduced in earlier commit
df73df138d ("OvmfPkg/AcpiPlatformDxe: replay
QEMU_LOADER_WRITE_POINTER commands at S3", 2017-02-09).
(2) On machine types that lack fw_cfg DMA (i.e., 2.4 and earlier),
TransferS3ContextToBootScript() would now fail the linker/loader
script for no reason.
(Note that QEMU itself prevents adding devices that depend on
WRITE_POINTER if the machine type lacks fw_cfg DMA:
$ qemu-system-x86_64 -M pc-q35-2.4 -device vmgenid
qemu-system-x86_64: -device vmgenid: vmgenid requires DMA write
support in fw_cfg, which this machine type does not provide)
Short-circuit an empty S3_CONTEXT in TransferS3ContextToBootScript() by
dropping S3_CONTEXT on the floor. This is compatible with the current
contract of the function as it constitutes a transfer of ownership.
Regression (2) was found and reported by Dhiru Kholia as an OSX guest boot
failure on the "pc-q35-2.4" machine type:
http://mid.mail-archive.com/CANO7a6x6EaWNZ8y=MvLU=w_LjRLXserO3NmsgHvaYE0aUCCWzg@mail.gmail.com
Dhiru bisected the issue to commit 8057622527.
Cc: Dhiru Kholia <dhiru.kholia@gmail.com>
Cc: Jordan Justen <jordan.l.justen@intel.com>
Fixes: df73df138d
Fixes: 8057622527
Reported-by: Dhiru Kholia <dhiru.kholia@gmail.com>
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Tested-by: Dhiru Kholia <dhiru.kholia@gmail.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
In OVMF we currently get the upper (>=4GB) memory size with the
GetSystemMemorySizeAbove4gb() function.
The GetSystemMemorySizeAbove4gb() function is used in two places:
(1) It is the starting point of the calculations in GetFirstNonAddress().
GetFirstNonAddress() in turn
- determines the placement of the 64-bit PCI MMIO aperture,
- provides input for the GCD memory space map's sizing (see
AddressWidthInitialization(), and the CPU HOB in
MiscInitialization()),
- influences the permanent PEI RAM cap (the DXE core's page tables,
built in permanent PEI RAM, grow as the RAM to map grows).
(2) In QemuInitializeRam(), GetSystemMemorySizeAbove4gb() determines the
single memory descriptor HOB that we produce for the upper memory.
Respectively, there are two problems with GetSystemMemorySizeAbove4gb():
(1) It reads a 24-bit count of 64KB RAM chunks from the CMOS, and
therefore cannot return a larger value than one terabyte.
(2) It cannot express discontiguous high RAM.
Starting with version 1.7.0, QEMU has provided the fw_cfg file called
"etc/e820". Refer to the following QEMU commits:
- 0624c7f916b4 ("e820: pass high memory too.", 2013-10-10),
- 7d67110f2d9a ("pc: add etc/e820 fw_cfg file", 2013-10-18)
- 7db16f2480db ("pc: register e820 entries for ram", 2013-10-10)
Ever since these commits in v1.7.0 -- with the last QEMU release being
v2.9.0, and v2.10.0 under development --, the only two RAM entries added
to this E820 map correspond to the below-4GB RAM range, and the above-4GB
RAM range. And, the above-4GB range exactly matches the CMOS registers in
question; see the use of "pcms->above_4g_mem_size":
pc_q35_init() | pc_init1()
pc_memory_init()
e820_add_entry(0x100000000ULL, pcms->above_4g_mem_size, E820_RAM);
pc_cmos_init()
val = pcms->above_4g_mem_size / 65536;
rtc_set_memory(s, 0x5b, val);
rtc_set_memory(s, 0x5c, val >> 8);
rtc_set_memory(s, 0x5d, val >> 16);
Therefore, remedy the above OVMF limitations as follows:
(1) Start off GetFirstNonAddress() by scanning the E820 map for the
highest exclusive >=4GB RAM address. Fall back to the CMOS if the E820
map is unavailable. Base all further calculations (such as 64-bit PCI
MMIO aperture placement, GCD sizing etc) on this value.
At the moment, the only difference this change makes is that we can
have more than 1TB above 4GB -- given that the sole "high RAM" entry
in the E820 map matches the CMOS exactly, modulo the most significant
bits (see above).
However, Igor plans to add discontiguous (cold-plugged) high RAM to
the fw_cfg E820 RAM map later on, and then this scanning will adapt
automatically.
(2) In QemuInitializeRam(), describe the high RAM regions from the E820
map one by one with memory HOBs. Fall back to the CMOS only if the
E820 map is missing.
Again, right now this change only makes a difference if there is at
least 1TB high RAM. Later on it will adapt to discontiguous high RAM
(regardless of its size) automatically.
-*-
Implementation details: introduce the ScanOrAdd64BitE820Ram() function,
which reads the E820 entries from fw_cfg, and finds the highest exclusive
>=4GB RAM address, or produces memory resource descriptor HOBs for RAM
entries that start at or above 4GB. The RAM map is not read in a single
go, because its size can vary, and in PlatformPei we should stay away from
dynamic memory allocation, for the following reasons:
- "Pool" allocations are limited to ~64KB, are served from HOBs, and
cannot be released ever.
- "Page" allocations are seriously limited before PlatformPei installs the
permanent PEI RAM. Furthermore, page allocations can only be released in
DXE, with dedicated code (so the address would have to be passed on with
a HOB or PCD).
- Raw memory allocation HOBs would require the same freeing in DXE.
Therefore we process each E820 entry as soon as it is read from fw_cfg.
-*-
Considering the impact of high RAM on the DXE core:
A few years ago, installing high RAM as *tested* would cause the DXE core
to inhabit such ranges rather than carving out its home from the permanent
PEI RAM. Fortunately, this was fixed in the following edk2 commit:
3a05b13106, "MdeModulePkg DxeCore: Take the range in resource HOB for
PHIT as higher priority", 2015-09-18
which I regression-tested at the time:
http://mid.mail-archive.com/55FC27B0.4070807@redhat.com
Later on, OVMF was changed to install its high RAM as tested (effectively
"arming" the earlier DXE core change for OVMF), in the following edk2
commit:
035ce3b37c, "OvmfPkg/PlatformPei: Add memory above 4GB as tested",
2016-04-21
which I also regression-tested at the time:
http://mid.mail-archive.com/571E8B90.1020102@redhat.com
Therefore adding more "tested memory" HOBs is safe.
Cc: Jordan Justen <jordan.l.justen@intel.com>
Ref: https://bugzilla.redhat.com/show_bug.cgi?id=1468526
Contributed-under: TianoCore Contribution Agreement 1.1
Signed-off-by: Laszlo Ersek <lersek@redhat.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Laszlo Ersek
Brijesh Singh
Ard Biesheuvel