Re: [v9,09/10] drm/xe/pci: Introduce PCIe FLR

["Laguna, Lukasz" <lukasz.laguna@intel.com>]

On 7/1/2026 10:29, Raag Jadav wrote:
> With bare minimum pieces in place, we can finally introduce PCIe Function
> Level Reset (FLR) support which re-initializes hardware state without the
> need for reloading the driver from userspace. All VRAM contents are lost
> along with hardware state and driver takes care of recreating the required
> kernel bos as part of re-initialization, but user still needs to recreate
> user bos and reload context after PCIe FLR.
>
> Signed-off-by: Raag Jadav <raag.jadav@intel.com>
> Tested-by: Lukasz Laguna <lukasz.laguna@intel.com>
> Acked-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
> ---
> v2: Spell out Function Level Reset (Jani)
> v5: Prevent PM ref leak for wedged device (Matthew Brost)
> v6: Add PCIe FLR documentation (Daniele)
> v7: Refine PCIe FLR documentation (Daniele)
>      Introduce xe_pci_reset_skip() helper (Lukasz)
> v9: Add 'Xe' prefix to document title (Rodrigo)
> ---
>   drivers/gpu/drm/xe/Makefile          |   1 +
>   drivers/gpu/drm/xe/xe_device_types.h |   3 +
>   drivers/gpu/drm/xe/xe_pci.c          |   1 +
>   drivers/gpu/drm/xe/xe_pci.h          |   2 +
>   drivers/gpu/drm/xe/xe_pci_error.c    | 129 +++++++++++++++++++++++++++
>   5 files changed, 136 insertions(+)
>   create mode 100644 drivers/gpu/drm/xe/xe_pci_error.c
>
> diff --git a/drivers/gpu/drm/xe/Makefile b/drivers/gpu/drm/xe/Makefile
> index 8e7b146880f4..3c001b2a4aec 100644
> --- a/drivers/gpu/drm/xe/Makefile
> +++ b/drivers/gpu/drm/xe/Makefile
> @@ -101,6 +101,7 @@ xe-y += xe_bb.o \
>   	xe_page_reclaim.o \
>   	xe_pat.o \
>   	xe_pci.o \
> +	xe_pci_error.o \
>   	xe_pci_rebar.o \
>   	xe_pcode.o \
>   	xe_pm.o \
> diff --git a/drivers/gpu/drm/xe/xe_device_types.h b/drivers/gpu/drm/xe/xe_device_types.h
> index 80fa60821a47..3ccfa8af482c 100644
> --- a/drivers/gpu/drm/xe/xe_device_types.h
> +++ b/drivers/gpu/drm/xe/xe_device_types.h
> @@ -480,6 +480,9 @@ struct xe_device {
>   	/** @pxp: Encapsulate Protected Xe Path support */
>   	struct xe_pxp *pxp;
>   
> +	/** @flr_prepared: Prepared for function-reset */
> +	bool flr_prepared;
> +
>   	/** @needs_flr_on_fini: requests function-reset on fini */
>   	bool needs_flr_on_fini;
>   
> diff --git a/drivers/gpu/drm/xe/xe_pci.c b/drivers/gpu/drm/xe/xe_pci.c
> index 03362480e3e0..d11188790eca 100644
> --- a/drivers/gpu/drm/xe/xe_pci.c
> +++ b/drivers/gpu/drm/xe/xe_pci.c
> @@ -1353,6 +1353,7 @@ static struct pci_driver xe_pci_driver = {
>   #ifdef CONFIG_PM_SLEEP
>   	.driver.pm = &xe_pm_ops,
>   #endif
> +	.err_handler = &xe_pci_error_handlers,
>   };
>   
>   /**
> diff --git a/drivers/gpu/drm/xe/xe_pci.h b/drivers/gpu/drm/xe/xe_pci.h
> index 11bcc5fe2c5b..24e51a71a959 100644
> --- a/drivers/gpu/drm/xe/xe_pci.h
> +++ b/drivers/gpu/drm/xe/xe_pci.h
> @@ -8,6 +8,8 @@
>   
>   struct pci_dev;
>   
> +extern const struct pci_error_handlers xe_pci_error_handlers;
> +
>   int xe_register_pci_driver(void);
>   void xe_unregister_pci_driver(void);
>   struct xe_device *xe_pci_to_pf_device(struct pci_dev *pdev);
> diff --git a/drivers/gpu/drm/xe/xe_pci_error.c b/drivers/gpu/drm/xe/xe_pci_error.c
> new file mode 100644
> index 000000000000..544005af4d3f
> --- /dev/null
> +++ b/drivers/gpu/drm/xe/xe_pci_error.c
> @@ -0,0 +1,129 @@
> +// SPDX-License-Identifier: MIT
> +/*
> + * Copyright © 2026 Intel Corporation
> + */
> +
> +#include "xe_device.h"
> +#include "xe_printk.h"
> +#include "xe_sriov_pf_helpers.h"
> +
> +/**
> + * DOC: Xe PCI Error Handling
> + *
> + * Xe driver registers PCI callbacks which are called by PCI core in case of
> + * bus errors or resets.
> + *
> + * Currently only Function Level Reset (FLR) callbacks are supported. PCIe FLR

This is outdated and needs to be updated.

> + * wipes the VRAM and resets the state of all the hardware units. Therefore, the
> + * contents of all exec queues and BOs in VRAM are lost and the hardware needs a
> + * full re-initialization. The way Xe driver handles it, is pretty much similar
> + * to system suspend/resume flow with a few notable exceptions.
> + *
> + * Prepare phase:
> + *
> + * - Temporarily wedge the device to prevent userspace access
> + * - Kill exec queues which signals all fences and frees in-flight jobs
> + * - Stop the scheduler and all submissions to GuC
> + * - The fact that FLR is needed is because hardware could be in corrupted state
> + *   and access unreliable, so skip memory eviction due to untrustworthy VRAM
> + *   contents
> + * - Remove all memory mappings since VRAM contents will be lost
> + *
> + * Re-initialization phase:
> + *
> + * - Recreate kernel BOs due to skipped memory eviction in prepare phase
> + * - Restore kernel queues which were killed in prepare phase
> + * - Reload all uC firmwares
> + * - Bring up all hardware units
> + * - Unwedge the device to allow userspace access
> + *
> + * Since VRAM contents are lost, the user is expected to recreate user memory
> + * and reload context.
> + *
> + * TODO: Add PCIe error handling callbacks using similar flow.

This is outdated and needs to be updated.

> + *
> + * Current implementation is only limited to re-initializing GT. This needs to
> + * be extended for a lot of components listed below.
> + *
> + * - Proper re-initialization of GSC and PXP for integrated platforms
> + * - SR-IOV cases which need PF and VF synchronization
> + * - Re-initialization of all child devices registered by Xe
> + * - MM corner cases
> + * - Display
> + */
> +
> +static inline bool xe_pci_reset_skip(struct xe_device *xe)
> +{
> +	return !IS_DGFX(xe) || IS_SRIOV_VF(xe) || xe_sriov_pf_num_vfs(xe) || xe->info.probe_display;
> +}
> +
> +static void xe_pci_reset_prepare(struct pci_dev *pdev)
> +{
> +	struct xe_device *xe = pdev_to_xe_device(pdev);
> +
> +	if (xe_pci_reset_skip(xe)) {
> +		xe_err(xe, "PCIe FLR not supported\n");
> +		return;
> +	}
> +
> +	if (xe_device_wedged(xe)) {
> +		xe_err(xe, "PCIe FLR aborted, device in unexpected state\n");
> +		return;
> +	}
> +
> +	/* Wedge the device to prevent userspace access but don't send the event yet */
> +	xe_device_wedged_get(xe);
> +
> +	/*
> +	 * The hardware could be in corrupted state and access unreliable, but we try to
> +	 * update data structures and cleanup any pending work to avoid side effects during
> +	 * PCIe FLR. This will be similar to system suspend flow but without eviction.
> +	 */
> +	if (xe_device_suspend(xe, true)) {
> +		xe_err(xe, "Failed to prepare for PCIe FLR\n");
> +		xe_device_wedged_put(xe);
> +		return;
> +	}
> +
> +	xe->flr_prepared = true;
> +	xe_info(xe, "Prepared for PCIe FLR\n");
> +}
> +
> +static void xe_pci_reset_done(struct pci_dev *pdev)
> +{
> +	struct xe_device *xe = pdev_to_xe_device(pdev);
> +
> +	if (xe_pci_reset_skip(xe))
> +		goto out;
> +
> +	if (!xe_device_wedged(xe) || !xe->flr_prepared)
> +		goto out;
> +
> +	/* Unprepare early in case we fail */
> +	xe->flr_prepared = false;
> +
> +	/*
> +	 * We already have the data structures intact, so try to re-initialize the device.
> +	 * This will be similar to system resume flow, except we'll also need to recreate
> +	 * kernel bos and restore kernel queues.
> +	 */
> +	if (xe_device_resume(xe, true)) {
> +		xe_err(xe, "Re-initialization failed\n");
> +		xe_device_wedged_put(xe);

I'm not sure if we should bother, but it introduces a small window where 
the device is unwedged while still broken.

Lukasz

> +		goto out;
> +	}
> +
> +	/* Unwedge to allow userspace access */
> +	xe_device_wedged_put(xe);
> +	xe_info(xe, "Re-initialization success\n");
> +
> +	return;
> +out:
> +	/* Most likely the device is unusable and there's nothing we can do about it */
> +	xe_device_declare_wedged(xe);
> +}
> +
> +const struct pci_error_handlers xe_pci_error_handlers = {
> +	.reset_prepare = xe_pci_reset_prepare,
> +	.reset_done = xe_pci_reset_done,
> +};