Remap ops have 3 parts: unmap, prev, and next. The commit step can fail
on any of these. Add a flag for each to these so the unwind is only done
the steps that have been committed.
v2: (Rodrigo) Use bit macros
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
On PVC unloading followed by reloading the module often results in a
completely dead machine (seems to be plaguing CI). Resetting the GuC
like we do at load seems to cure it at least when locally testing this.
v2:
- Move pc_fini into guc_fini. We want to do the GuC reset just after
calling pc_fini, otherwise we encounter communication failures. It
also seems like a good idea to do the reset before we start releasing
the various other GuC resources. In the case of pc_fini there is an
explicit stop, but for other stuff like logs, ads, ctb there is not.
References: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/542
References: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/597
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Some copy hardware engine instances are faster than others on PVC.
Use a virtual engine of these plus the reserved instance for the migrate
engine on PVC. The idea being if a fast instance is available it will be
used and the throughput of kernel copies, clears, and pagefault
servicing will be higher.
v2: Use OOB WA, use all copy engines if no WA is required
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Wa_16017236439 requires that we update BCS_SWCTRL
(via indirect context batch buffer) to set 64B
transfers when running on an even-numbered BCS
engine and 256B on an odd-numbered BCS engine.
v2: Move WA from engine_was[] to lrc_was[]
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Tejas Upadhyay <tejas.upadhyay@intel.com>
Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
The current only submission in the driver that doesn't use a vm is the
WA setup. We still pass a vm structure (the migration one), but we don't
actually use it at submission time and we instead have an hack to use
GGTT for this particular engine.
Instead of special-casing the WA engine, we can skip providing a VM and
use that as selector for whether to use GGTT or PPGTT. As part of this
change, we can drop the special engine flag for the WA engine and switch
the WA submission to use the standard job functions instead of dedicated
ones.
v2: rebased on s/engine/exec_queue
Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Link: https://lore.kernel.org/r/20230822173334.1664332-4-daniele.ceraolospurio@intel.com
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
If an engine is only destroyed on driver unload, we can skip its
clean-up steps with the GuC because the GuC is going to be tuned off as
well, so it doesn't matter if we're in sync with it or not. Currently,
we apply this optimization to all engines marked as kernel, but this
stops us to supporting kernel engines that don't stick around until
unload. To remove this limitation, add a separate flag to indicate if
the engine is expected to only be destryed on driver unload and use that
to trigger the optimzation.
While at it, add a small comment to explain what each engine flag
represents.
v2: s/XE_BUG_ON/XE_WARN_ON, s/ENGINE/EXEC_QUEUE
v3: rebased
Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Link: https://lore.kernel.org/r/20230822173334.1664332-3-daniele.ceraolospurio@intel.com
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
The only possible 64-bit register writes in the driver come from the
highly questionable MMIO ioctl. That ioctl's register write support
only operates for userspace running as root and cannot be used by any
real userspace; it exists solely to support the "xe_reg" debug tool in
IGT. Since the spec indicates that hardware does not officially support
64-bit register accesses, there's no reason to allow such 64-bit writes,
even for debugging.
Bspec: 60027
Reviewed-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: José Roberto de Souza <jose.souza@intel.com>
Link: https://lore.kernel.org/r/20230823003312.1356779-4-matthew.d.roper@intel.com
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Intel hardware officially only supports GTTMMADR register accesses of
32-bits or less (although 64-bit accesses to device memory and PTEs in
the GSM are fine). Even though we do usually seem to get back
reasonable values when performing readq() operations on registers in
BAR0, we shouldn't rely on this violation of the spec working
consistently. It's likely that even when we do get proper register
values back the hardware is internally satisfying the request via a
non-atomic sequence of two 32-bit reads, which can be problematic for
timestamps and counters if rollover of the lower bits is not considered.
Replace xe_mmio_read64() with xe_mmio_read64_2x32() that implements
64-bit register reads as two 32-bit reads and attempts to ensure that
the upper dword has stabilized to avoid problematic rollovers for
counter and timestamp registers.
v2:
- Move function from xe_mmio.h to xe_mmio.c. (Lucas)
- Convert comment to kerneldoc and note that it shouldn't be used on
registers where reads may trigger side effects. (Lucas)
Bspec: 60027
Reviewed-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: José Roberto de Souza <jose.souza@intel.com>
Link: https://lore.kernel.org/r/20230823003312.1356779-3-matthew.d.roper@intel.com
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
As with PVC, Xe2 platforms require that the index of an uncached MOCS
entry be programmed into the GUC_SHIM_CONTROL register. This will
likely be needed on future platforms as well.
Xe2 also extends the size of the MOCS index register field from two bits
to four bits. Since these extra bits were unused on PVC, it should be
safe to just increase the size of the mask.
Bspec: 60592
Cc: Haridhar Kalvala <haridhar.kalvala@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Starting with Xe2, a 5-level page table is always used, regardless of
the actual virtual address range supported by the platform. The two
values need to be tracked separately in the device descriptor since Xe2
platforms only have a 48 bit virtual address range.
Bspec: 59505, 65637, 70817
Cc: Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Define a common set of Xe2 graphics feature flags and definitions that
will be used for all platforms in this family.
Several of the feature flags are inherited unchanged from Xe_HP and/or
Xe_HPC platforms:
- dma_mask_size remains 46 (Bspec 70817)
- supports_usm=1 (Bspec 59651)
- has_flatccs=1 (Bspec 58797)
- has_asid=1 (Bspec 59654, 59265, 60288)
- has_range_tlb_invalidate=1 (Bspec 71126)
However some of them still need proper implementation in the driver to
be used, so they are disabled.
Notable Xe2-specific changes:
- All Xe2 platforms use a five-level page table, regardless of the
virtual address space for the platform. (Bspec 59505)
The graphics engine mask represents the Xe2 architecture engines (Bspec
60149), but individual platforms may have a reduced set of usable
engines, as reflected by their fusing.
Cc: Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Starting with Xe2, all platforms (including igpu platforms) use FlatCCS
compression rather than AuxCCS. Similar to PVC, any future platforms
that don't support FlatCCS should not attempt to fall back to AuxCCS
programming.
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Engine register state layout has changed a bit on Xe2. We'll also
explicitly define a BCS layout to ensure BLIT_SWCTL and BLIT_CCTL are
included.
Bspec: 65182, 60184, 55793
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Matt Atwood <matthew.s.atwood@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Xe2 media has a few types of MCR registers, but all except for "GPMXMT"
can safely steer to instance 0,0. GPMXMT follows the same rules that
MTL's OADDRM ranges did, so it can re-use the same enum value.
Bspec: 71186
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Matt Atwood <matthew.s.atwood@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Xe2 uses the same steering control register and steering semaphore
register as MTL. As with recent platforms, group/instance 0,0 is
sufficient to target a non-terminated instance for most classes of MCR
registers; the only types of ranges that need to consider platform
fusing to find a non-terminated instance are SLICE/DSS ranges and a new
SQIDI_PSMI type of range.
Note that the range of valid bits in XE2_NODE_ENABLE_MASK may be reduced
for some Xe2 SKUs. However the lowest bits are always valid and only
the lowest instance is obtained via __ffs(), so there's no need to
complicate the masking with extra platform/subplatform checks.
Also note that Wa_14017387313 suggests skipping MCR lock acquisition
around GAM and GAMWKR registers to prevent MCR register accesses in an
interrupt handler from deadlocking when the steering semaphore is
already held outside the interrupt context. At this time Xe never
issues MCR accesses from within an interrupt handler so the workaround
is not currently needed.
v2:
- [0x008700-0x0087FF] range to extend up to 0x887F (Matt Attwood)
- [0x00EF00-0x00F4FF] -> [0x00F000, 0xFFFF] to follow latest
bspec version (Bala)
Bspec: 71185
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com>
Reviewed-by: Matt Atwood <matthew.s.atwood@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
The render and compute context are significantly smaller on Xe2 than on
previous platforms.
Registers:
- Render: 3008 dwords -> 12032 bytes -> round to 3 pages
- Compute: 1424 dwords -> 5696 bytes -> round to 2 pages
We also allocate one additional page for the HWSP, so the total
allocation sizes for render and compute are 4 and 3 pages respectively.
Bspec: 65182, 56578, 55793
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Matt Atwood <matthew.s.atwood@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
The choice of Y-major tiling format (either the legacy "TileY" or the
newer "Tile4") is based on graphics IP version (12.50 and beyond have
Tile4, earlier platforms have TileY). The tracking in xe was originally
added to allow re-using display from i915. However as of i915 commit
4ebf43d048 ("drm/i915: Eliminate has_4tile feature flag"), the display
code determines TileY vs Tile4 itself, so this can be removed from xe.
Reviewed-by: Lucas De Marchi <lucas.demarchi@intel.com>
Link: https://lore.kernel.org/r/20230817230407.909816-10-matthew.d.roper@intel.com
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Print both device physical address range and CPU io range
of vram. Also print vram's actual size, usable size excluding
stolen memory, and CPU io accessible size.
V1:
- Add back small BAR device info (Matt)
Signed-off-by: Oak Zeng <oak.zeng@intel.com>
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Make a xe_mem_region structure which will be used in the
coming patches. The new structure is used in both xe device
level (xe->mem.vram) and xe_tile level (tile->vram).
Make the definition of xe_mem_region.dpa_base to be the DPA
base of this memory region and change codes according to
this new definition.
v1:
- rename xe_mem_region.base to dpa_base per conversation with Mike
Ruhl
Signed-off-by: Oak Zeng <oak.zeng@intel.com>
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Usually we call __guc_exec_queue_fini_async via a worker as the
exec_queue fini can be done from within the GPU scheduler which creates
a circular dependency without a worker. Kernel exec_queues are fini'd at
driver unload (not from within the GPU scheduler) so it is safe to
directly call __guc_exec_queue_fini_async.
Suggested-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
If we are closing a vm, mark each vma as XE_VMA_DESTROYED and skip
touching the rebind_list if this is seen on the eviction path. That way
we can safely drop the vm dma-resv lock on the close path without
needing to worry about racing with the eviction path trying to add stuff
to the rebind_list which can corrupt our contended list, since the
destroy and rebind links are the same list entry underneath.
References: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/514
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Rather check if the engine is still registered before proceeding with
deregister steps. Also the engine being marked as disabled doesn't mean
the engine has been disabled or deregistered from GuC pov, and here we
are signalling fences so we need to be sure GuC is not still using this
context.
v2:
- Drop the read_stopped() for this path. Since we are signalling
fences on error here, best play it safe and wait for the GT reset to
mark the engine as disabled, rather than it just being queued.
v3 (Matt Brost):
- Keep the read_stopped() on the wait event, since there is no need to
wait for an already scheduled GT reset. If it is set we can then just
bail without signalling anything.
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
The workaround database tells us to set this bit, even though the bspec
indicates the bit doesn't exist on these platforms. Since this is a
write-only register, we also can't read back its value to verify whether
it's actually working or not. For now we'll trust that the workaround
database knows what it's talking about; if not, the hardware will just
ignore the attempt to write to a non-existent bit and it shouldn't cause
any problems.
Reviewed-by: Matt Atwood <matthew.s.atwood@intel.com>
Link: https://lore.kernel.org/r/20230727220920.2291913-2-matthew.d.roper@intel.com
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
To workaround a HW bug on DG2, driver is required to map the whole
ppgtt virtual address space before GPU workload submission. Thus
set the XE_VM_FLAG_SCRATCH_PAGE flag during vm create so the whole
address space is mapped to point to scratch page.
v1:
- Move the workaround implementation from xe_vm_create to
xe_vm_create_ioctl - Brian
- Reorder error checking in xe_vm_create_ioctl - Jose
- Implement WA only for DG2-G10 and DG2-G12
Signed-off-by: Oak Zeng <oak.zeng@intel.com>
Reviewed-by: Brian Welty <brian.welty@intel.com>
Reviewed-by: José Roberto de Souza <jose.souza@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
