Hi, This is the v7 of series introducing xe-vfio-pci driver variant supporting migration. Thanks for all the review feedback.
It looks like all the r-b's are in place and we need a final ack from Alex. We also got a confirmation that DRM will be able to take the series once everything is ready. The biggest changes in this revision are: generating VFIO helpers using a macro, addition of XE_VFIO_PCI build dependency on PCI_IOV, and a bug fix for the issue where we would skip calling xe_sriov_vfio_wait_flr_done() if the VF doesn't support migration. Also, note that GuC v70.54.0 is now available in linux-firmware: https://gitlab.com/kernel-firmware/linux-firmware/-/commit/23568a4b942079631c9378011aa0a9b29de19370 Full changelog, including URLs to previous revisions can be found below. Cover letter from the previous revision: Xe is a DRM driver supporting Intel GPUs and for SR-IOV capable devices, it enables the creation of SR-IOV VFs. This series adds xe-vfio-pci driver variant that interacts with Xe driver to control VF device state and read/write migration data, allowing it to extend regular vfio-pci functionality with VFIO migration capability. The driver doesn't expose PRE_COPY support, as currently supported hardware lacks the capability to track dirty pages. While Xe driver already had the capability to manage VF device state, management of migration data was something that needed to be implemented and constitutes the majority of the series. The migration data is processed asynchronously by the Xe driver, and is organized into multiple migration data packet types representing the hardware interfaces of the device (GGTT / MMIO / GuC FW / VRAM). Since the VRAM can potentially be larger than available system memory, it is copied in multiple chunks. The metadata needed for migration compatibility decisions is added as part of descriptor packet (currently limited to PCI device ID / revision). Xe driver abstracts away the internals of packet processing and takes care of tracking the position within individual packets. The API exported to VFIO is similar to API exported by VFIO to userspace, a simple .read()/.write(). Note that some of the VF resources are not virtualized (e.g. GGTT - the GFX device global virtual address space). This means that the VF driver needs to be aware that migration has occurred in order to properly relocate (patching or reemiting data that contains references to GGTT addresses) before resuming operation. The code to handle that is already present in upstream Linux and in production VF drivers for other OSes. Links to previous revisions for reference. v1: https://lore.kernel.org/lkml/[email protected]/ v2: https://lore.kernel.org/lkml/[email protected]/ v3: https://lore.kernel.org/lkml/[email protected]/ v4: https://lore.kernel.org/lkml/[email protected]/ v5: https://lore.kernel.org/lkml/[email protected]/ v6: https://lore.kernel.org/lkml/[email protected]/ v6 -> v7: * Update commit msg with details on debug flag behavior (Michał) * Generate VFIO migration helpers with a macro (Alex) * Depend on PCI_IOV (Michał) * Reorder migration file struct members (Alex) * Fix xe_sriov_vfio_wait_flr_done() behavior (Alex) * Set state in .open_device() / .close_device() (Alex) * Remove redundant "disable_file" in release_file (Alex) * Use GFP_KERNEL_ACCOUNT alloc flag (Alex) * Use pci_iov_vf_id() directly in vdev->vfid assignment (Alex) v5 -> v6: * Exclude the patches already merged through drm-tip * Add logging when migration is enabled in debug mode (Michał) * Rename the xe_pf_get_pf helper (Michał) * Don't use "vendor specific" (yet again) (Michał) * Kerneldoc tweaks (Michał) * Use guard(xe_pm_runtime_noresume) instead of assert (Michał) * Check for num_vfs rather than total_vfs (Michał) v4 -> v5: * Require GuC version >= 70.54.0 * Fix VFIO migration migf disable * Fix null-ptr-deref on save_read error * Don't use "vendor specific" (again) (Kevin) * Introduce xe_sriov_packet_types.h (Michał) * Kernel-doc fixes (Michał) * Use tile_id / gt_id instead of tile / gt in packet header (Michał) * Don't use struct_group() in packet (Michał) * And other, more minor changes v3 -> v4: * Add error handling on data_read / data_write path * Don't match on PCI class, use PCI_DRIVER_OVERRIDE_DEVICE_VFIO helper instead (Lucas De Marchi) * Use proper node VMA size inside GGTT save / restore helper (Michał) * Improve data tracking set_bit / clear_bit wrapper names (Michał) * Improve packet dump helper (Michał) * Use drmm for migration mutex init (Michał) * Rename the pf_device access helper (Michał) * Use non-interruptible sleep in VRAM copy (Matt) * Rename xe_sriov_migration_data to xe_sriov_packet along with relevant functions (Michał) * Rename per-vf device-level data to xe_sriov_migration_state (Michał) * Use struct name that matches component name instead of anonymous struct (Michał) * Don't add XE_GT_SRIOV_STATE_MAX to state enum, use a helper macro instead (Michał) * Kernel-doc fixes (Michał) v2 -> v3: * Bind xe-vfio-pci to specific devices instead of using vendor and class (Christoph Hellwig / Jason Gunthorpe) * Don't refer to the driver as "vendor specific" (Christoph) * Use pci_iov_get_pf_drvdata and change the interface to take xe_device (Jason) * Update the RUNNING_P2P comment (Jason / Kevin Tian) * Add state_mutex to protect device state transitions (Kevin) * Implement .error_detected (Kevin) * Drop redundant comments (Kevin) * Explain 1-based indexing and wait_flr_done (Kevin) * Add a missing get_file() (Kevin) * Drop redundant state transitions when p2p is supported (Kevin) * Update run/stop naming to match other drivers (Kevin) * Fix error state handling (Kevin) * Fix SAVE state diagram rendering (Michał Wajdeczko) * Control state machine flipping PROCESS / WAIT logic (Michał Wajdeczko) * Drop GUC / GGTT / MMIO / VRAM from SAVE control state machine * Use devm instead of drmm for migration-related allocations (Michał) * Use GGTT node for size calculations (Michał) * Use mutex guards consistently (Michał) * Fix build break on 32-bit (lkp) * Kernel-doc updates (Michał) * And other, more minor changes v1 -> v2: * Do not require debug flag to support migration on PTL/BMG * Fix PCI class match on VFIO side * Reorganized PF Control state machine (Michał Wajdeczko) * Kerneldoc tidying (Michał Wajdeczko) * Return NULL instead of -ENODATA for produce/consume (Michał Wajdeczko) * guc_buf s/sync/sync_read (Matt Brost) * Squash patch 03 (Matt Brost) * Assert on PM ref instead of taking it (Matt Brost) * Remove CCS completely (Matt Brost) * Return ptr on guc_buf_sync_read (Michał Wajdeczko) * Define default guc_buf size (Michał Wajdeczko) * Drop CONFIG_PCI_IOV=n stubs where not needed (Michał Wajdeczko) * And other, more minor changes Michał Winiarski (4): drm/xe/pf: Enable SR-IOV VF migration drm/xe/pci: Introduce a helper to allow VF access to PF xe_device drm/xe/pf: Export helpers for VFIO vfio/xe: Add device specific vfio_pci driver variant for Intel graphics MAINTAINERS | 7 + drivers/gpu/drm/xe/Makefile | 4 + drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 9 + drivers/gpu/drm/xe/xe_pci.c | 17 + drivers/gpu/drm/xe/xe_pci.h | 3 + drivers/gpu/drm/xe/xe_sriov_pf_migration.c | 35 +- drivers/gpu/drm/xe/xe_sriov_pf_migration.h | 1 + .../gpu/drm/xe/xe_sriov_pf_migration_types.h | 4 +- drivers/gpu/drm/xe/xe_sriov_vfio.c | 80 +++ drivers/vfio/pci/Kconfig | 2 + drivers/vfio/pci/Makefile | 2 + drivers/vfio/pci/xe/Kconfig | 12 + drivers/vfio/pci/xe/Makefile | 3 + drivers/vfio/pci/xe/main.c | 573 ++++++++++++++++++ include/drm/intel/xe_sriov_vfio.h | 143 +++++ 15 files changed, 888 insertions(+), 7 deletions(-) create mode 100644 drivers/gpu/drm/xe/xe_sriov_vfio.c create mode 100644 drivers/vfio/pci/xe/Kconfig create mode 100644 drivers/vfio/pci/xe/Makefile create mode 100644 drivers/vfio/pci/xe/main.c create mode 100644 include/drm/intel/xe_sriov_vfio.h -- 2.51.2
