Rebase v7.0 to support 6.6KV by akanksh5-intc · Pull Request #348 · intel-gpu/xekmd-backports

akanksh5-intc · 2026-04-21T06:49:37Z

No description provided.

knayana-gse · 2026-05-08T11:31:50Z

@@ -1,6 +1,5 @@
 /* SPDX-License-Identifier: GPL-2.0 */
 /*
- * workqueue.h --- work queue handling for Linux.


Removes the comment "workqueue.h --- work queue handling for Linux." — unrelated, unintended change.

Yes, reverted it back now.

Rebase the DRM_GPUVM conditional on the backport build flag so the xe-drm driver only selects GPUVM when the compat layer actually provides it. Also to resolve the failure need to fix the offset values. Reference: bdcdf968be314b6 drm, drm/xe: Fix xe userptr in the absence of CONFIG_DEVICE_PRIVATE Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

When the CONFIG_INTEL_EHL_PSE_IO was added, it caused the hunk failure due to mislignment of line offset. Reference: a0c83150eea5 platform/x86/intel: Introduce Intel Elkhart Lake PSE I/O Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

The memory allocation helper macros introduced were introduced from v7.0-rc1 kernel with the new allocation helpers being introduced, to simplify that default case. errors are listed below where compilation fails on older kernels ------------------------------------------------------------------- - ttm_sys_manager.c: implicit declaration of 'kzalloc_obj' - ttm_range_manager.c: implicit declaration of 'kzalloc_flex' - ttm_agp_backend.c: implicit declaration of ‘kmalloc_obj’ - vsec.c: implicit declaration of ‘kzalloc_objs’ - ttm_tt.c: implicit declaration of ‘kvzalloc_objs’ - sched_main.c: implicit declaration of ‘kmalloc_objs - xe_vm.c: implicit declaration of ‘kvmalloc_objs’ ------------------------------------------------------------------- Therefore added backport compatibility layer that provides fallback definitions for the following allocation helpers when missing: - kzalloc_obj() - Allocate and zero memory for single object - kzalloc_flex() - Allocate flexible array structure - kmalloc_obj(), kmalloc_objs() - kzalloc_objs(), kvzalloc_objs(), kvmalloc_objs() - default_gfp() helper macro The backport uses typeof() and size_mul() to maintain type safety and overflow protection, matching the upstream implementation. Reference: e19e1b480ac7 add default_gfp() helper macro and use it in the new *alloc_obj() helpers Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

The EMPTY_VMA_FLAGS macro changed from VM_NONE (0) to ((vma_flags_t){ }) when the vma_flags_t type was introduced. For older kernels lack both the vma_flags_t type and the EMPTY_VMA_FLAGS definition. error is listed below: ----------------------------------------------------------------- error: ‘EMPTY_VMA_FLAGS’ undeclared (first use in this function) ----------------------------------------------------------------- Implemented conditional definition that adapts to kernel capabilities: - For kernels without vma_flags_t: define EMPTY_VMA_FLAGS as 0 - For kernels with vma_flags_t: define as ((vma_flags_t){ }) - For v7.0-rc1kv onwards: use native EMPTY_VMA_FLAGS Reference: bae0ba7c7c0a02 mm: add basic VMA flag operation helper functions Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

For struct drm_device huge_mnt member introduced in v7.0kv to support transparent hugepage allocation in shmfs-backed gem objects. error: ------------------------------------------------------------------------- drm_gem.h:522: error: 'struct drm_device' has no member named 'huge_mnt' ------------------------------------------------------------------------- Added huge_mnt field to struct drm_device for transparent hugepage support. In kernels which are below v7.0 lack this infrastructure, causing direct member access to fail. - Implemented autoconf detection for huge_mnt member availability via struct compilation probe. - Return NULL for kernels without CONFIG_TRANSPARENT_HUGEPAGE support - Return NULL for kernels lacking huge_mnt member in drm_device Reference: 6e0b1b82017b9b drm/gem: Add huge tmpfs mountpoint helpers Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

In 7.0kv system_percpu_wq as a dedicated per-CPU workqueue. The gpu scheduler uses this for timeout handling, but relies on transitive includes the backport compatibility layer from activating. error: ------------------------------------------------------------------------------ ‘system_percpu_wq’ undeclared; did you mean ‘system_long_wq’? args->timeout_wq ------------------------------------------------------------------------------ Reference: 83646541d197 drm/sched: Replace use of system_wq with system_percpu_wq Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

Add drm_dumb_buffers.h and drm_pagemap_util.h for xe-oot compatability error: ---------------------------------------------------------------------- xe_vm_types.h: error: drm/drm_pagemap_util.h: No such file or directory xe_bo.c: error: drm/drm_dumb_buffers.h: No such file or directory ---------------------------------------------------------------------- for drm_dumb_buffers.h Reference: fb24aaf5415c drm/dumb-buffers: Provide helper to set pitch and size for drm_pagemap_util.h Reference: 77f14f2f2d73f drm/pagemap: Add a drm_pagemap cache and shrinker Above mentioned files are introduced from v7.0-rc1 kernel. Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

Implement compatibility layer for get_region_info_caps callback in struct vfio_device_ops introduced in v7.0kv to support vfio-pci operations on older kernels. error: ------------------------------------------------------------------------------- error: ‘const struct vfio_device_ops’ has no member named ‘get_region_info_caps’ ------------------------------------------------------------------------------- In 7.0kv get_region_info_caps introduced to call struct vfio_device_ops inorder to enable drivers to provide custom region capability information. But in older kernels this mechanism lack this callback member in the vfio_device_ops structure. So, implemented autoconf detection and conditional compilation to detect get_region_info_caps member availability via struct compilation probe. Reference: 775f726a742a vfio: Add get_region_info_caps op Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

ACQUIRE and ACQUIRE_ERR cleanup guard macros in v7.0 as part of the cleanup.h infrastructure for automatic resource management. The xe driver uses these for PM runtime locking in ioctl handlers. error: ----------------------------------------------------------------------------- error: implicit declaration of function ACQUIRE(xe_pm_runtime_ioctl, pm)(xe); error: implicit declaration of function ACQUIRE_ERR(xe_pm_runtime_ioctl, &pm); ----------------------------------------------------------------------------- Implemented autoconf detection and conditional fallback pattern to detect ACQUIRE macro availability via cleanup.h compilation probe. Reference: 857d18f23ab1 cleanup: Introduce ACQUIRE() and ACQUIRE_ERR() for conditional locks Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

Implement compatibility layer for DEFINE_GUARD_COND macro signature change from 3 to 4 arguments introduced in v7.0kv to support xe-power management guard helpers. error: ------------------------------------------------------------------------------- error: macro "DEFINE_GUARD_COND" passed 4 arguments, but takes just 3 DEFINE_GUARD_COND(xe_pm_runtime, _ioctl, xe_pm_runtime_get_ioctl(_T), _RET >= 0) ------------------------------------------------------------------------------- In 7.0kv the implementation extended DEFINE_GUARD_COND from 3 to 4 arguments by adding a condition parameter (_RET >= 0) to determine guard success. The xe PM runtime ioctl guard uses the 4-argument form to check return values. But in older kernel it support the 3-argument form without condition checking. Implemented autoconf detection and conditional macro usage: - Use 3-argument form (without condition check) on older kernels - Use 4-argument form (with _RET >= 0 condition) on v7.0+ kernels Reference: e4ab322fbaaaf cleanup: Add conditional guard support Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

Other subsystem(s) may want to reuse the for_each_if() macro. Move it to util_macros.h to make it globally available. error: -------------------------------------------------------------------------------- error: implicit declaration of function for_each_if(!(__entity) || (__job)->entity == (__entity)) -------------------------------------------------------------------------------- The for_each_if macro was introduced to linux/util_macros.h for conditional iteration filtering. DRM GPU scheduler uses this in drm_sched_for_each_pending_job() iterator to filter jobs by entity. Implemented autoconf detection to detect for_each_if availability via util_macros.h compilation probe and define as conditional empty-block pattern when unavailable. Reference: b2108fc82a0a drm: Move for_each_if() to util_macros.h for wider use Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

Implemented compatibility layer for init_llist_node function introduced in v6.7 kernel onwards to support DRM GPUVM lock-less list initialization. error: ------------------------------------------------------------------------------- implicit declaration of function ‘init_llist_head’ ------------------------------------------------------------------------------- The function init_llist_node() in v6.7 for proper initialization of lock-less list nodes. The DRM GPUVM uses this to initialize deferred BO list entries. So in v6.6 kernels lack this helper function. Therefore, implemented autoconf to detect init_llist_node availability through llist.h and provided inline implementation using WRITE_ONCE when unavailable. Reference: d6b3358a2813 llist: add interface to check if a node is on a list Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

In v7.0-rc1 shmem_writeout() was renamed from shmem_writepage() to shmem_writeout() as a simplified helper for writing shmem folios to swap. The TTM backup module uses this for efficient memory reclaim during buffer object swapping. error: ------------------------------------------------------------------------------- error: implicit declaration of function shmem_writeout(to_folio, NULL, NULL); ERROR: modpost: "shmem_writeout" [~/src/drivers/gpu/drm/ttm/ttm.ko] undefined! ------------------------------------------------------------------------------- Reference: 7b73c12c6ebf0 shmem: Add shmem_writeout() Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

The poll_timeout_us macro was introduced to linux/iopoll.h for simplified polling operations with microsecond precision. XE driver uses this extensively for GuC firmware load status checks and power management state transitions. Legacy kernels lack this utility macro. error: -------------------------------------------------------------------------------- implicit declaration of function poll_timeout_us(load_result = guc_load_done(gt, &status, &tries), load_result, -------------------------------------------------------------------------------- Provided backport implementation in iopoll.h wrapper: - Define poll_timeout_us macro with full upstream semantics - Support sleep between polls with usleep_range() - Support optional sleep-before-first-poll behavior - Return 0 on success, -ETIMEDOUT on timeout - Include barrier() for proper memory ordering Reference: 9df8043a546d iopoll: Generalize read_poll_timeout() into poll_timeout_us() Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

…bility In v7.0-rc1 kernel onwards dma_fence_check_and_signal_locked() was introduced as an atomic operation that checks if a fence is already signaled and signals it if not, all while holding the fence lock. The XE driver uses this for hardware fence cleanup paths. error: ------------------------------------------------------------------------ implicit declaration of function ‘dma_fence_check_and_signal_locked’; XE_WARN_ON(dma_fence_check_and_signal_locked(&fence->dma)); ------------------------------------------------------------------------ Implementation: - Detect dma_fence_check_and_signal_locked availability via compilation probe - Provide implementation combining test_signaled_flag and signal_locked - Export symbol for module usage - Declare function prototype in backport header Reference: c891b99d25dd dma-buf/dma-fence: Add dma_fence_check_and_signal() Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

knayana-gse

conditional approve.
Please take care of latest comments.
Please check compilation of supported OSVs/kernels and add BAT results here as well.

knayana-gse · 2026-05-22T06:46:32Z

 u32 pci_iov_vf_bar_get_sizes(struct pci_dev *dev, int resno, int num_vfs);
 #endif

+#ifdef BPM_PCI_RESIZE_RESOURCE_ARG4_NOT_PRESENT


can you check with these change pci Resizable BAR is working fine or not.
steps are mentioned in jira: VLK-83694

Status before installing dkms

command: sudo lspci -vv -s 03:00.0
Capabilities: [420 v1] Physical Resizable BAR
BAR 2: current size: 16GB, supported: 256MB 512MB 1GB 2GB 4GB 8GB 16GB
Capabilities: [220 v1] Virtual Resizable BAR
BAR 2: current size: 2GB, supported: 256MB 512MB 1GB 2GB 4GB 8GB 16GB
Capabilities: [320 v1] Single Root I/O Virtualization (SR-IOV)
IOVCap: Migration- 10BitTagReq+ Interrupt Message Number: 000
IOVCtl: Enable- Migration- Interrupt- MSE- ARIHierarchy+ 10BitTagReq-
IOVSta: Migration-
Initial VFs: 12, Total VFs: 12, Number of VFs: 0, Function Dependency Link: 00
VF offset: 1, stride: 1, Device ID: e20b
Supported Page Size: 00000553, System Page Size: 00000001
Region 0: Memory at 0000004601000000 (64-bit, prefetchable)
Region 2: Memory at 0000004c00000000 (64-bit, prefetchable)
VF Migration: offset: 00000000, BIR: 0

Status after installing dkms

command: sudo lspci -vv -s 03:00.0
Capabilities: [420 v1] Physical Resizable BAR
BAR 2: current size: 16GB, supported: 256MB 512MB 1GB 2GB 4GB 8GB 16GB
Capabilities: [220 v1] Virtual Resizable BAR
BAR 2: current size: 16GB, supported: 256MB 512MB 1GB 2GB 4GB 8GB 16GB

knayana-gse · 2026-05-22T06:46:48Z

@@ -0,0 +1,73 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*


remove these empty lines.

knayana-gse · 2026-05-22T06:48:43Z

 static void zconfprint(const char *err, ...);
 static void zconf_error(const char *err, ...);
 static bool zconf_endtoken(const struct kconf_id *id, int starttoken, int endtoken);
+static const struct kconf_id *kconf_id_lookup(const char *str, unsigned int len);


typo
Moveed --> moved

Below are the compilation issues which are faced while compiling against 6.6 headers error: ------------------------------------------------------------------------------- error: too many arguments to function pci_resize_resource(pdev, resno, bar_size, 0); error: implicit declaration of function pci_rebar_size_supported, pci_rebar_size_to_bytes, pci_rebar_get_max_size ------------------------------------------------------------------------------- Extended pci_resize_resource() from 3 to 4 arguments in v6.19, adding an exclude_bars parameter to control which BARs should be excluded from resizing. The xe driver uses the 4-argument form for PCBAR expansion. But in older kernels only support the 3-argument form without the exclude_bars parameter. Similarly, there are three another compilation issues those are - pci_rebar_size_supported() - validate if size is supported for BAR - pci_rebar_size_to_bytes() - convert size encoding to byte values - pci_rebar_get_max_size() - query maximum supported BAR size Implementation: - pci_resize_resource: detect pci_resize_resource 4-argument support via compilation probe - pci_rebar_size_supported: check if size is in supported bitmap - pci_rebar_size_to_bytes: convert encoding to bytes using bit shift - pci_rebar_get_max_size: find highest set bit in size bitmap Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

Implement compatibility layer for __assign_str tracing macro signature change from 1 to 2 arguments that occurred between kernel versions to support xe driver tracepoints across all target kernels. error: ------------------------------------------------------------------------------- error: macro "__assign_str" requires 2 arguments, but only 1 given __assign_str undeclared (first use in this function) ------------------------------------------------------------------------------- The kernel tracing subsystem changed __assign_str macro signature between versions. But in older kernels use single-argument form:__assign_str(field); From v7.0 kernels require two-argument form with explicit source value: __assign_str(field, source_value); Reference: c759e609030c tracing: Remove __assign_str_len() Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

Implement compatibility layer for configfs_group_operations.is_visible member introduced in upstream 6.12 kernels to allow dynamic visibility control of configfs attributes. This uses this for controlling visibility of device and SR-IOV configuration groups based on device capabilities. error: ----------------------------------------------------------------------------- error: ‘struct configfs_group_operations’ has no member named ‘is_visible’ .is_visible = xe_config_device_is_visible, ----------------------------------------------------------------------------- So implemeted and conditional probe if is_visible member availability via struct then include member assignment only when supported but member isn't assigned on old kernels which holds original implementation. Reference: 0e6a35b93745bb fs/configfs: Add a callback to determine attribute visibility Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

Add a compatibility macro for drm_warn_once() which is not available in older kernel versions e.g in 6.12kv and below. The fallback implementation uses dev_warn_once() to provide equivalent functionality. This resolves compilation errors in drm_dumb_buffers.c and other files that use drm_warn_once() when building against kernels that don't provide this macro natively. error: ------------------------------------------------------------------------------- error: implicit declaration of function ‘drm_warn_once’; drm_warn_once(dev, modpost: "drm_mode_size_dumb" [~/src/drivers/gpu/drm/xe/xe.ko] undefined! ------------------------------------------------------------------------------- Reference: fb6c7ab8718e drm/print: introduce new struct drm_device based logging macros Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

Add detection and fallback implementation for drm_pagemap_put() which was introduced in kernel v7.0. For older kernels, provide a static inline implementation that uses the drm_pagemap operations release callback via kref_put(). The compatibility is handled entirely in the header file with a static inline function that properly decrements the reference count and invokes the destroy operation when the reference reaches zero. This function is used extensively in xe_vm_madvise.c and xe_svm.c for managing DRM pagemap lifecycle. error: ------------------------------------------------------------------------------- modpost: "drm_pagemap_put" [~/src/drivers/gpu/drm/xe/xe.ko] undefined! ------------------------------------------------------------------------------- Reference: a599b98607de drm/pagemap, drm/xe: Add refcounting to struct drm_pagemap Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

…n for lock For drm_gpuvm_bo_obtain_locked() and drm_gpuvm_madvise_ops_create() which are used by the xe driver for GPU virtual memory management but not available in kernels below v7.0 kernels. error: ------------------------------------------------------------------------------- modpost: "drm_gpuvm_bo_obtain_locked" [~/src/drivers/gpu/drm/xe/xe.ko] modpost: "drm_gpuvm_madvise_ops_create" [~/src/drivers/gpu/drm/xe/xe.ko] ------------------------------------------------------------------------------- drm_gpuvm_bo_obtain_locked(): This function obtains or creates a drm_gpuvm_bo instance representing the combination of a drm_gpuvm and drm_gem_object. The backport implementation uses drm_gpuvm_bo_find() to search for existing instances, and drm_gpuvm_bo_create() to allocate new ones if needed. drm_gpuvm_madvise_ops_create(): This function creates operations for memory advisory hints. For kernels without madvise support, the implementation falls back to drm_gpuvm_sm_map_ops_create() which provides compatible mapping operations. Both functions are properly exported with EXPORT_SYMBOL_GPL for use by the xe.ko module. Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

Moved the #include "kconf_id.c" from the mid-grammar %{ %} block to the bottom of the file alongside the other implementation includes. The mid-grammar placement caused kconf_id.c to be included before the bison-generated token constants were defined, leading to build failures with certain bison/flex versions. Added a forward declaration for kconf_id_lookup() in the top-level declarations block so the parser rules can reference it before the definition is included at the end of the file. Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

akanksh5-intc · 2026-05-22T16:19:10Z

conditional approve. Please take care of latest comments. Please check compilation of supported OSVs/kernels and add BAT results here as well.

BAT results: https://gtax-ril-fm.intel.com/#/jobs/4737065

akanksh5-intc force-pushed the oot-backport/main branch 4 times, most recently from 90e2aa3 to a26e551 Compare April 29, 2026 06:19

akanksh5-intc force-pushed the oot-backport/main branch 2 times, most recently from 6e05b38 to 0a77c00 Compare April 30, 2026 06:50

knayana-gse requested changes May 4, 2026

View reviewed changes

akanksh5-intc force-pushed the oot-backport/main branch 3 times, most recently from d00c7f5 to 61711a9 Compare May 6, 2026 17:22

knayana-gse requested changes May 8, 2026

View reviewed changes

akanksh5-intc added 2 commits May 11, 2026 09:11

Fix hunk offsets in Makefile for v7.0-rc1

0d9bce4

When the CONFIG_INTEL_EHL_PSE_IO was added, it caused the hunk failure due to mislignment of line offset. Reference: a0c83150eea5 platform/x86/intel: Introduce Intel Elkhart Lake PSE I/O Signed-off-by: Akanksha Hotta <akanksha.hotta@intel.com>

akanksh5-intc force-pushed the oot-backport/main branch 2 times, most recently from fbc358a to 487d1df Compare May 11, 2026 04:28

akanksh5-intc added 3 commits May 11, 2026 09:59

akanksh5-intc force-pushed the oot-backport/main branch from 487d1df to e187203 Compare May 11, 2026 06:01

akanksh5-intc added 3 commits May 11, 2026 11:38

akanksh5-intc force-pushed the oot-backport/main branch from e187203 to a8742e8 Compare May 11, 2026 08:38

akanksh5-intc added 7 commits May 14, 2026 14:49

akanksh5-intc force-pushed the oot-backport/main branch 2 times, most recently from 20b32be to 2240d5c Compare May 18, 2026 11:17

akanksh5-intc changed the title ~~Resolve hunk failure for v7.0-rc6~~ Rebase v7.0 to support 6.6KV May 21, 2026

knayana-gse approved these changes May 22, 2026

View reviewed changes

akanksh5-intc added 7 commits May 22, 2026 14:55

smuqthya force-pushed the oot-backport/main branch from 2240d5c to bf46199 Compare May 22, 2026 10:02

Uh oh!

Conversation

akanksh5-intc commented Apr 21, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

knayana-gse left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

akanksh5-intc commented May 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants