- 21 Dec, 2023 40 commits
-
-
Matt Roper authored
When recording the default LRC, the expectation is that the hardware's original state settings (both register and instruction) will be written out to the LRC upon first context switch. For many 3DSTATE_* state instructions that don't truly have "default" values, this translates to a simple instruction header (opcodes + dword length) being written to the LRC, followed by an appropriate number of blank dwords as a place holder. When userspace creates a context (which starts as a copy of the default LRC), they'll generally emit real 3DSTATE_* as part of their initialization to select the settings they desire. If they don't emit one of the 3DSTATE instructions, then the zeroed dwords that remain in their LRC image generally translate to various state remaining disabled. This will either be what userspace wants or will lead to very reproducible and easily-debugged problems (rendering glitches, engine hangs). It turns out that a subset of the 3DSTATE instructions, specifically those belonging to the SVG (State Variable - Global) unit, are not only emitting 0's for the instruction's "body" dwords, but also for the instruction header dword if no specific state has been explicitly set before context switch. This means that when the hardware switches to a context that hasn't explicitly provided an appropriate state setting, the hardware will just see a sequence of NOOPs in the spot reserved for that 3DSTATE instruction while executing the LRC, and the actual hardware state setting will unintentionally inherit the configuration used by the previously running context. Now when userspace makes a mistake and forgets to emit an important state instruction they no longer get consistent, easily-reproducible corruption/hangs, but rather erratic behavior where the presence/absence of a problem depends on what other workloads are running on the system and what order the contexts are scheduled on the engine. A specific example of this that came up recently related to mesh shading The OpenGL driver was not specifically emitting a 3DSTATE_MESH_CONTROL to disable mesh shading at context init, so on context switch, mesh shading would either be on or off depending on what the previous context had been doing. Vulkan apps _were_ enabling mesh shading, so running a Vulkan app and then context switching to an OpenGL app resulted in mesh shading still unexpectedly being enabled during OpenGL operation, and since other Mesh-related state was not properly initialized for that context a GPU hang was seen. Due to the specific ordering requirements (Vulkan app runs first, followed by OpenGL app), it took additional debug effort to track down the cause of the problem. There are various workarounds related to this behavior, with current implementations handled in the userspace drivers. E.g., Wa_14019789679 and Wa_22018402687. However it's been suggested that the kernel driver can help simplify things here by emitting zeroed SVG state with proper instruction headers as part of our default context creation (i.e., at the same point we apply LRC workarounds). This will help ensure that any future cases where a userspace driver does not emit an important state setting will result in consistent behavior. Bspec: 46261 Reviewed-by:
Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com> Link: https://lore.kernel.org/r/20231025151732.3461842-7-matthew.d.roper@intel.comSigned-off-by:
Matt Roper <matthew.d.roper@intel.com> Signed-off-by:
Rodrigo Vivi <rodrigo.vivi@intel.com>
-
Matt Roper authored
On some platforms we need to emit some non-register state while recording an engine class' default LRC. Add the infrastructure to support this; actual per-platform tables will be added in future patches. v2: - Checkpatch whitespace fix - Add extra assertion to ensure num_dw != 0. (Bala) Reviewed-by:
-
Balasubramani Vivekanandan authored
Make use of EVENT_CLASS to group similar trace events Signed-off-by:
Haridhar Kalvala <haridhar.kalvala@intel.com> Link: https://lore.kernel.org/intel-xe/20231019093140.1901665-3-balasubramani.vivekanandan@intel.com/Signed-off-by:
-
Balasubramani Vivekanandan authored
Event tracing enabled for CTB submissions. Additional minor refactor - Removed a unnecessary ct_to_xe() call. v2: Remove a unwanted comment (Hari) Add missing change to commit message Signed-off-by: -
Shekhar Chauhan authored
Add L3SQCREG5 as part of HW recommended settings. The recommended value in Bspec is 00e0007f. For Xe2-LPG, bits 23:21 don't exist anymore, but it's confirmed with HW engineers that setting them doesn't do anything. They still exist on the media GT, Xe2-LPM, but they are already they are already set as per HW default value. So for Xe2 platform, the only bits that need to be set are 9:0 since HW's default is 0x1ff and the recommended value is 0x7f. Unlike most registers, which have the same relative offset on both the primary and media GT, this register has a different base offset on the media GT. On MTL the register only exists for the primary (graphics) GT, so there's no need to program it on the media gt. Also, it's part of the RCS engine's context, so it needs to be added as a LRC workaround. Bspec: 72161 Signed-off-by:
Shekhar Chauhan <shekhar.chauhan@intel.com> Reviewed-by:
Lucas De Marchi <lucas.demarchi@intel.com> Signed-off-by:
-
Dnyaneshwar Bhadane authored
Add the initial collection of gt/engine/lrc workarounds. While at it, add some newlines around the platform/IP comments to make them consistent across all workarounds. v2: - FF_MODE is an MCR register (Matt Roper) - Group 18032247524 with other Xe2 workarounds (Matt Roper) - Move WA changing PSS_CHICKEN to lrc_was[] as for Xe2 that register is part of the render context image (Matt Roper) - Apply WA 16020518922 only on render engine (Matt Roper) Signed-off-by:Dnyaneshwar Bhadane <dnyaneshwar.bhadane@intel.com> Signed-off-by:
-
Brian Welty authored
We can remove the unnecessary indirection thru xe->tiles[] to get the TTM VRAM manager. This code can be common for VRAM and STOLEN. Signed-off-by:
Brian Welty <brian.welty@intel.com> Reviewed-by:
Matthew Brost <matthew.brost@intel.com> Signed-off-by:
-
Matthew Auld authored
Bit 7 in the leaf node is normally programmed with pat[2], however with 2M/1G pages that same bit in the PDE/PDPE also toggles 2M/1G pages. For 2M/1G entries the pat[2] is rather moved to bit 12, which is now free given that the address must be aligned to 2M or 1G, leaving bit 7 for toggling 2M/1G pages. Bspec: 59510, 45038 Signed-off-by:
Matthew Auld <matthew.auld@intel.com> Cc: Lucas De Marchi <lucas.demarchi@intel.com> Cc: Matt Roper <matthew.d.roper@intel.com> Reviewed-by:
-
Daniele Ceraolo Spurio authored
The PVC GuC version that we're currently using (70.6.4) has a known issue that leads to dropping the disabling of contexts that have pending page faults. This is fixed in newer blobs, so we need to update to a more recent release. Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/696Signed-off-by:
Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Cc: John Harrison <John.C.Harrison@Intel.com> Reviewed-by:
John Harrison <John.C.Harrison@Intel.com> Signed-off-by:
-
Francois Dugast authored
This is used for the priority of an exec queue (not an engine) and should be named accordingly. Signed-off-by:
Francois Dugast <francois.dugast@intel.com> Signed-off-by:
-
Rodrigo Vivi authored
During the uapi review it was identified a possible confusion with the plural of acronym with a new acronym. So the recommendation is to go with gt_list instead. Suggested-by:
-
Rodrigo Vivi authored
We already have many bits reserved at the end already. Let's kill the unused ones. Signed-off-by:
-
Rodrigo Vivi authored
Let's have a single GT ID per GT within the PCI Device Card. Signed-off-by:
-
Rodrigo Vivi authored
Split drm_xe_query_gt out of the gt list one in order to better document it. No functional change at this point. Any actual change to the uapi should come in follow-up additions. v2: s/maks/mask Cc: Matt Roper <matthew.d.roper@intel.com> Signed-off-by:
-
Matthew Brost authored
A case existed where an out-sync of a later VM bind operation could signal before a previous one if the later operation results in a NOP (e.g. a unbind or prefetch to a VA range without any mappings). This breaks the ordering rules, fix this. This patch also lays the groundwork for users to pass in num_binds == 0 and out-syncs. Signed-off-by:
Thomas Hellström <thomas.hellstrom@linux.intel.com> Signed-off-by:
-
Matthew Brost authored
Async worker is gone. All jobs and memory allocations done in IOCTL to align with dma fencing rules. Async vs. sync now means when do bind operations complete relative to the IOCTL. Async completes when out-syncs signal while sync completes when the IOCTL returns. In-syncs and out-syncs are only allowed in async mode. If memory allocations fail in the job creation step the VM is killed. This is temporary, eventually a proper unwind will be done and VM will be usable. Signed-off-by:
-
Matthew Brost authored
This is not used nor does it align VM async document, kill this. Signed-off-by:
-
Rodrigo Vivi authored
This extension is currently not used and it is not aligned with the error handling on async VM_BIND. Let's remove it and along with that, since it was the only extension for the vm_create, remove VM extension entirely. v2: rebase on top of the removal of drm_xe_ext_exec_queue_set_property Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> Signed-off-by:
-
Ashutosh Dixit authored
There really is no difference between 'struct drm_xe_ext_vm_set_property' and 'struct drm_xe_ext_exec_queue_set_property', they are extensions which specify a <property, value> pair. Replace the two extensions with a single common 'struct drm_xe_ext_set_property' extension. The rationale is that rather than have each XE module (including future modules) invent their own property/value extensions, all XE modules use a common set_property extension when possible. Signed-off-by:
Ashutosh Dixit <ashutosh.dixit@intel.com> Signed-off-by:
-
Matthew Brost authored
Functionality of XE_EXEC_QUEUE_SET_PROPERTY_COMPUTE_MODE deprecated in a previous patch, drop from uAPI. The property is just simply inherented from the VM. v2: - Update commit message (Niranjana) Reviewed-by:
Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com> Signed-off-by:
-
Matthew Brost authored
Much better name and aligns with xe_vm_add_compute_exec_queue. As part of the rename, move the implementation from xe_exec_queue.c to xe_vm.c. Suggested-by:
-
Matthew Brost authored
We are going to remove XE_EXEC_QUEUE_SET_PROPERTY_COMPUTE_MODE from the uAPI, deprecate the implementation first by making XE_EXEC_QUEUE_SET_PROPERTY_COMPUTE_MODE a NOP. After removal of XE_EXEC_QUEUE_SET_PROPERTY_COMPUTE_MODE the proper is simply inherented from the VM. v2: - Update commit message with explaination of removal (Niranjana) Signed-off-by:
-
Matthew Brost authored
Last little piece to support parallel exec queue is compute mode. Signed-off-by:
-
Francois Dugast authored
This was previously used in UMD for timestamp correlation, which can now be done with DRM_XE_QUERY_CS_CYCLES. Link: https://lore.kernel.org/all/20230706042044.GR6953@mdroper-desk1.amr.corp.intel.com/ Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/636Signed-off-by:
José Roberto de Souza <jose.souza@intel.com> Reviewed-by:
-
Francois Dugast authored
This macro was necessary when bind operations were shifted but this is no longer the case, so removing to simplify code. Signed-off-by:
-
Francois Dugast authored
Use different members in the drm_xe_vm_bind_op for op and for flags as it is done in other structures. Type is left to u32 to leave enough room for future operations and flags. v2: Remove the XE_VM_BIND_* flags shift (Rodrigo Vivi) Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/303Signed-off-by:
-
Umesh Nerlige Ramappa authored
Perf measurements rely on CPU and engine timestamps to correlate events of interest across these time domains. Current mechanisms get these timestamps separately and the calculated delta between these timestamps lack enough accuracy. To improve the accuracy of these time measurements to within a few us, add a query that returns the engine and cpu timestamps captured as close to each other as possible. Mesa MR: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24591 v2: - Fix kernel-doc warnings (CI) - Document input params and group them together (Jose) - s/cs/engine/ (Jose) - Remove padding in the query (Ashutosh) Signed-off-by:
Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com> Reviewed-by:
-
Umesh Nerlige Ramappa authored
User engine class is of type u16. Set the same type for the array used to map xe engines to user engines. Signed-off-by:
-
Umesh Nerlige Ramappa authored
Queries are 0-indexed, so a query with value N is invalid if the ARRAY_SIZE is N. Modify the check to account for that. Fixes: dd08ebf6 ("drm/xe: Introduce a new DRM driver for Intel GPUs") Signed-off-by:
-
Matt Roper authored
RCS and CCS engines include several non-register gfxpipe commands in their LRC images. Include these in the dump output so that we can see exactly what's inside the context snapshot. v2: - Include raw instruction header in output - Add 3DSTATE_AMFS_TEXTURE_POINTERS and 3DSTATE_MONOFILTER_SIZE. The first was supposed to be removed in Xe_HPG, and the second by gen12, but both still show up in the RCS LRC. v3: - Sanity check that we don't have numdw > remaining_dw. (Lucas) Cc: Lucas De Marchi <lucas.demarchi@intel.com> Reviewed-by:
-
Matt Roper authored
For non-RCS engines, nearly all of the LRC state is composed of MI instructions (specifically MI_LOAD_REGISTER_IMM). Providing a dump interface allows us to verify that the context image layout matches what's documented in the bspec, and also allows us to check whether LRC workarounds are being properly captured by the default state we record at startup. For now, the non-MI instructions found in the RCS and CCS engines will dump as "unknown;" parsing of those will be added in a follow-up patch. v2: - Add raw instruction header as well as decoded meaning. (Lucas) - Check that num_dw isn't greater than remaining_dw for instructions that have a "# dwords" field. (Lucas) - Clarify comment about skipping over ppHWSP. (Lucas) Bspec: 64993 Cc: Lucas De Marchi <lucas.demarchi@intel.com> Reviewed-by:
-
Matt Roper authored
Extracting the common MI_* instructions that can be used with any engine to their own header will make it easier as we add additional engine instructions in upcoming patches. Also, since the majority of GPU instructions (both MI and non-MI) have a "length" field in bits 7:0 of the instruction header, a common define is added for that. Instruction-specific length fields are still defined for special case instructions that have larger/smaller length fields. v2: - Use "instr" instead of "inst" as the short form of "instruction" everywhere. (Lucas) - Include xe_reg_defs.h instead of the i915 compat header. (Lucas) Reviewed-by:
-
Matt Roper authored
MI_STORE_DATA_IMM can store either dword values or qword values, and can store more than one value if the instruction's length field is large enough. Create explicit defines to specify the number of dwords/qwords to be stored, which will set the instruction length correctly and, if necessary, turn on the 'store qword' bit. While we're here, also replace an open-coded version of MI_STORE_DATA_IMM with the common macros. Bspec: 60246 Reviewed-by:
-
Matt Roper authored
Keeping the number of registers to be loaded as a separate macro from the instruction opcode will simplify some upcoming LRC parsing code. Reviewed-by:
-
Matt Roper authored
Despite its name, MI_FLUSH_DW instruction can write an immediate value of either dword size or qword size, depending on the 'length' field of the instruction. Since "length" excludes the first two dwords of the instruction, a value of 2 in the length field implies a dword write and a value of 3 implies a qword write. Even in cases where the flush instruction's post-sync operation is set to "no write" we're still expected to size the overall instruction as if we were doing a dword or qword write (i.e., a length of 1 shouldn't be used on modern platforms). Rather than baking a size of "1" into the #define and then adding another unexplained "+ 1" at all the spots where the definition gets used, lets just create MI_FLUSH_IMM_DW and MI_FLUSH_IMM_QW definitions that should be OR'd into the instruction header to make it more explicit what behavior we're requesting. Bspec: 60229 Reviewed-by:
-
Vitaly Lubart authored
Create mei-gscfi auxiliary device and configure interrupts to be consumed by mei-gsc device driver. Reviewed-by:
Vitaly Lubart <vitaly.lubart@intel.com> Signed-off-by:
Alexander Usyskin <alexander.usyskin@intel.com> Signed-off-by:
-
Vitaly Lubart authored
Mark support of MEI-GSC interaction per device. Add has_heci_gscfi indication to xe_device and xe_pci structures. Mark DG1 and DG2 devices as supported. Reviewed-by:
-
Vitaly Lubart authored
Add HECI2 register offsets for DG1 and DG2 to regs/xe_regs.h Reviewed-by:
-
Shekhar Chauhan authored
The bspec was recently updated to remove PCI ID 0x5698; this ID is actually reserved for future use and should not be treated as DG2-G11. BSpec: 44477 Signed-off-by:
-
Shekhar Chauhan authored
Add recently added PCI IDs for DG2 BSpec: 44477 Signed-off-by:
-
