Enable full vfio backport on 6.6 kernel by jagadeeshrb1 · Pull Request #384 · intel-gpu/xekmd-backports

jagadeeshrb1 · 2026-05-04T05:42:08Z

Updated below files to enable vfio driver.

copy-list -- Copy vfio driver to outfolder
defconfigs -- Enable required VFIO configs
Makefile.kernel -- Compile vfio driver folder
Kconfig.sources -- Source Kconfig

Signed-off-by: Rachapalli Bandi Jagadeesh jagadeesh.rachapalli.bandi@intel.com

smuqthya · 2026-05-07T12:00:54Z

+				[anon_inode_getfile_fmode() is not available])
+		])
+	])
+])


remove this junk line

smuqthya · 2026-05-07T12:04:14Z

+#endif	/* CONFIG_ARCH_SUPPORTS_PMD_PFNMAP */
+#endif
+
+#endif


end comment __BACKPORT_ASM_X86_PGTABLE_H

smuqthya · 2026-05-07T12:10:16Z

+				[iommu_paging_domain_alloc() is not available])
+		])
+	])
+])


smuqthya · 2026-05-07T12:11:26Z

@@ -0,0 +1,23 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2007-2008 Advanced Micro Devices, Inc.


Why AMD? are we copying?

Under GPL license, license notice be preserved in all copies and derivative works.

smuqthya · 2026-05-07T12:11:35Z

@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright (C) 2007-2008 Advanced Micro Devices, Inc.


Under GPL license, license notice be preserved in all copies and derivative works.

smuqthya · 2026-05-07T12:36:36Z

+{
+	return eventfd_signal(ctx, 1);
+}
+#define eventfd_signal bpm_eventfd_signal


use official macro.

smuqthya · 2026-05-07T12:47:21Z

 void follow_pfnmap_end(struct follow_pfnmap_args *args);
 #endif
+
+#ifdef BPM_VM_ALLOW_ANY_UNCACHED_NOT_PRESENT


Why do we need macro for this. can we use ifundef?

smuqthya · 2026-05-07T12:48:04Z

+static inline int bpm_irq_bypass_register_producer(struct irq_bypass_producer *producer,
+						struct eventfd_ctx *eventfd, int irq)
+{
+	return irq_bypass_register_producer(producer);


purpose of other arguments are ignored?

In 6.6, irq_bypass_register_producer() accept single arg.

smuqthya · 2026-05-07T12:48:58Z

+		],[
+			struct pci_dev *pdev = NULL;
+
+			if (pdev)


This condition will never work. it will always skip. allocate memory and check

smuqthya · 2026-05-07T12:53:24Z


-	BUILT_MODULE_NAME[10]="xe-vfio-pci"
-	BUILT_MODULE_LOCATION[10]="src/drivers/vfio/pci/xe"
+	BUILT_MODULE_NAME[10]="vfio"


do we need any control

Updated below files to enable vfio driver. 1. copy-list -- Copy vfio driver to outfolder 2. defconfigs -- Enable required VFIO configs 3. Makefile.kernel -- Compile vfio driver folder 4. Kconfig.sources -- Source Kconfig Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

struct fd accessor helpers fd_file() and fd_empty() were introduced in Linux 6.12. Kernels prior to 6.12 expose the file pointer directly as fd.file without these wrapper macros. Add a backport-include/linux/file.h shim that: 1. Pulls in the real kernel header via #include_next 2. Defines fd_file(f) and fd_empty(f) as inline macros mapping to the legacy .file member access pattern This unblocks compilation that use fd_file()/fd_empty() idioms introduced in Linux 6.12. Error: ------------------------------------------------------------------------------ error: implicit declaration of function ‘fd_empty’ error: implicit declaration of function ‘fd_file’ ------------------------------------------------------------------------------ Reference: 88a2f6468d01 struct fd: representation change Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

anon_inode_getfile_fmode() was introduced in Linux 6.10 as a convenience wrapper around anon_inode_getfile() that additionally sets caller-specified f_mode bits on the returned file. Kernels prior to 6.10 lack this helper. Add a compat implementation that: 1. Calls the existing anon_inode_getfile() to create the anon inode file 2. ORs the caller-provided f_mode into file->f_mode on success 3. Exports the symbol via EXPORT_SYMBOL_GPL for use by dependent modules This unblocks compilation of upstream drivers that rely on anon_inode_getfile_fmode() on kernels older than 6.10. Error: ------------------------------------------------------------------------------ error: implicit declaration of function ‘anon_inode_getfile_fmode’ ------------------------------------------------------------------------------ Reference: 55394d29c9e1 fs: Create anon_inode_getfile_fmode() Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

Add backport definitions and implementation for follow_pfnmap_args, follow_pfnmap_start(), and follow_pfnmap_end() when the kernel lacks the API. Errors: ------------------------------------------------------------------------------ error: variable ‘args’ has initializer but incomplete type error: ‘struct follow_pfnmap_args’ has no member named ‘vma’ error: ‘struct follow_pfnmap_args’ has no member named ‘address’ error: storage size of ‘args’ isn’t known error: implicit declaration of function ‘follow_pfnmap_start’ error: implicit declaration of function ‘follow_pfnmap_end’ ------------------------------------------------------------------------------ Reference: 6da8e9634bb7 mm: new follow_pfnmap API Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

iommu_paging_domain_alloc() was introduced in Linux 6.11 as an explicit API for allocating a paging-type IOMMU domain, replacing the more generic iommu_domain_alloc(). On older kernels, the function does not exist, causing the following build failure. Backport iommu_paging_domain_alloc() to compat/iommu.c for kernels below 6.11, implemented as a wrapper around iommu_domain_alloc(dev->bus), which is the functional equivalent on older kernels where all allocated domains are implicitly paging domains. Error: ------------------------------------------------------------------------------ error: implicit declaration of function ‘iommu_paging_domain_alloc’ ------------------------------------------------------------------------------ Reference: a27bf2743cb8 iommu: Add iommu_paging_domain_alloc() interface Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

Add compatibility wrappers for iommufd_device_attach(), iommufd_device_replace(), and iommufd_device_detach() so callers can use the PASID-aware interfaces on kernels that still provide the older signatures. Errors: ------------------------------------------------------------------------------ error: too many arguments to function ‘iommufd_device_attach’ error: too many arguments to function ‘iommufd_device_detach’ error: too many arguments to function ‘iommufd_device_replace’ ------------------------------------------------------------------------------ Reference: 2fb69c602d57 iommufd: Support pasid attach/replace Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

Add compatibility helpers for ida_find_first() and ida_exists() on kernels < 6.15 that do not provide these IDA interfaces. Errors: ------------------------------------------------------------------------------ error: implicit declaration of function ‘ida_find_first’ error: implicit declaration of function ‘ida_exists’ ------------------------------------------------------------------------------ Reference: 7fe6b987166b ida: Add ida_find_first_range() and ida_find_first()/ida_exists() variants Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

…NS args MODULE_IMPORT_NS() accepts a bare token, not a quoted string. Prior to 6.13, the macro internally used __stringify(ns) to convert the token to a string — passing a pre-quoted string like "IOMMUFD" caused __stringify to produce "\"IOMMUFD\"", which modpost could not match against exported namespace symbols. Remove the erroneous quotes so __stringify(IOMMUFD) correctly resolves to "IOMMUFD" on kernels < 6.13, and the refactored macro on 6.13+ likewise works correctly. Errors: ------------------------------------------------------------------------------ ERROR: modpost: module vfio uses symbol iommufd_vfio_compat_ioas_get_id from namespace IOMMUFD_VFIO, but does not import it. ERROR: modpost: module vfio uses symbol iommufd_access_unpin_pages from namespace IOMMUFD, but does not import it. ERROR: modpost: module vfio uses symbol iommufd_access_detach from namespace IOMMUFD, but does not import it. ERROR: modpost: module vfio uses symbol iommufd_device_attach from namespace IOMMUFD, but does not import it. ERROR: modpost: module vfio uses symbol iommufd_access_pin_pages from namespace IOMMUFD, but does not import it. ERROR: modpost: module vfio uses symbol iommufd_access_replace from namespace IOMMUFD, but does not import it. ------------------------------------------------------------------------------ Reference: cdd30ebb1b9f module: Convert symbol namespace to string literal Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

pte_offset_map() and pte_offset_map_lock() were converted from inline functions to out-of-line exported symbols (__pte_offset_map and __pte_offset_map_lock). The change also introduced the ability for these functions to return NULL when the PMD entry is concurrently modified. Add compat inline wrappers that delegate to the original inline pte_offset_map[_lock]() functions, preserving the new __-prefixed calling convention while building correctly on older kernels. Add a compat inline wrapper that delegates to pmd_ERROR() + pmd_clear() for kernels where pmd_clear_bad() is not yet an exported symbol, preserving build compatibility across kernel versions. Errors: ------------------------------------------------------------------------------ ERROR: modpost: "__pte_offset_map_lock" [src/compat/xe-compat.ko] undefined! ERROR: modpost: "__pte_offset_map" [src/compat/xe-compat.ko] undefined! ERROR: modpost: "pmd_clear_bad" [src/compat/xe-compat.ko] undefined! ------------------------------------------------------------------------------ Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

…nels iova_bitmap_alloc(), iova_bitmap_free(), and iova_bitmap_for_each() were introduced in Linux 6.7 as part of the VFIO dirty page tracking rework and are not available on older kernels. Add compat implementations in the backport compat layer to provide these symbols when building against kernels that predate the iova_bitmap API. Errors: ------------------------------------------------------------------------------ ERROR: modpost: "iova_bitmap_free" [src/drivers/vfio/vfio.ko] undefined! ERROR: modpost: "iova_bitmap_alloc" [src/drivers/vfio/vfio.ko] undefined! ERROR: modpost: "iova_bitmap_for_each" [src/drivers/vfio/vfio.ko] undefined! ------------------------------------------------------------------------------ Reference: 53f0b020218f vfio/iova_bitmap: Export more API symbols Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

struct vfio_pci_eventfd was removed or is not present in the 6.6 kernel, causing a compile failure when building the VFIO PCI backport: Add the struct definition to the compat layer to restore the missing type and allow vfio_pci_core.c to compile cleanly against 6.6 kernels. Error: ------------------------------------------------------------------------------ error: too few arguments to function ‘eventfd_signal’ ------------------------------------------------------------------------------ Reference: 3652117f8548 eventfd: simplify eventfd_signal() Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

VM_ALLOW_ANY_UNCACHED was introduced upstream in 6.9 to control uncached mapping permissions on ARM64. It is not present in below 6.9 kernels, causing a build failure. Backported VM_ALLOW_ANY_UNCACHED definition in mm.h to fix the build error. Error: ------------------------------------------------------------------------------ error: ‘VM_ALLOW_ANY_UNCACHED’ undeclared ------------------------------------------------------------------------------ Reference: 5c656fcdd6c6 mm: Introduce new flag to indicate wc safe Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

The upstream irq_bypass_register_producer() API was refactored to accept trigger and irq arguments directly, rather than requiring the caller to pre-populate fields on struct irq_bypass_producer before registering. The 6.6 kernel only exposes the 1-argument form, causing a build failure in the backported vfio_pci_intrs.c: Add a compat wrapper bpm_irq_bypass_register_producer() that accepts the new 2 & 3-argument signature, populates prod->token and prod->irq manually, then forwards to the single-argument kernel functio Error: ------------------------------------------------------------------------------ error: too many arguments to function ‘irq_bypass_register_producer’ ------------------------------------------------------------------------------ Reference: 2b521d86ee80 irqbypass: Take ownership of producer/consumer token tracking Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

…_PCI_MMAP) pci_dev->non_mappable_bars was introduced upstream in 6.15 to selectively block memory mapping of specific BARs on certain hardware. This field does not exist on pre-6.15 kernels, causing below build failure. Replace the runtime per-device check with a compile-time Kconfig guard using IS_ENABLED(CONFIG_VFIO_PCI_MMAP), which preserves the intended build-time gating behavior on kernels predating the introduction of this field. Error: ------------------------------------------------------------------------------ error: ‘struct pci_dev’ has no member named ‘non_mappable_bars’ ------------------------------------------------------------------------------ Reference: 888bd8322dfc s390/pci: Introduce pdev->non_mappable_bars and replace VFIO_PCI_MMAP Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

Adding below oot modules to the dkms deb package list. 1. vfio.ko 2. vfio_iommu_type1.ko 3. vfio-pci-core.ko 4. xe-vfio-pci.ko Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

The macro condition that disabled .match callback for kernel 6.6 caused a NULL function pointer dereference crash in vfio_pci_core_match+0x9a when QEMU issued a VFIO_GROUP_GET_DEVICE_FD ioctl during VF passthrough. Root cause: xe_vfio_pci_ops.match was conditionally excluded for 6.6, but vfio_pci_core_match() calls ops->match unconditionally without a NULL guard, resulting in a jump to address NULL. Call trace: ------------------------------------------------------------------------------- BUG: kernel NULL pointer dereference, address: 0000000000000000 Oops: 0010 [intel-gpu#1] PREEMPT SMP NOPTI CPU: 10 PID: 2587 Comm: qemu-system-x86 Kdump: loaded Tainted: G 6.6.126-lgci-xe+ RIP: 0010:0x0 Call Trace: <TASK> vfio_pci_core_match+0x9a/0x170 [vfio_pci_core] vfio_group_fops_unl_ioctl+0x195/0x680 [vfio] __x64_sys_ioctl+0xa0/0xf0 x64_sys_call+0x1b76/0x2210 do_syscall_64+0x56/0x90 ? irqentry_exit_to_user_mode+0x23/0x50 ? irqentry_exit+0x77/0xb0 ? exc_page_fault+0xad/0x230 entry_SYSCALL_64_after_hwframe+0x78/0xe2 ----------------------------------------------------------------------------- Reviewed-by: S A Muqthyar Ahmed <syed.abdul.muqthyar.ahmed@intel.com> Signed-off-by: Rachapalli Bandi Jagadeesh <jagadeesh.rachapalli.bandi@intel.com>

jagadeeshrb1 force-pushed the vfio_6_6 branch from e623ed4 to 31de2e8 Compare May 5, 2026 06:33

smuqthya requested changes May 7, 2026

View reviewed changes

jagadeeshrb1 force-pushed the vfio_6_6 branch 6 times, most recently from aaf24c1 to 0031832 Compare May 8, 2026 09:40

jagadeeshrb1 added 16 commits May 8, 2026 16:13

jagadeeshrb1 force-pushed the vfio_6_6 branch from 0031832 to 8b1c3e1 Compare May 8, 2026 10:44

smuqthya approved these changes May 8, 2026

View reviewed changes

Uh oh!

Conversation

jagadeeshrb1 commented May 4, 2026

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!

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!

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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants