Commit 96bfe9ff authored by Dave Airlie's avatar Dave Airlie
Browse files

Merge tag 'drm-xe-fixes-2026-03-05' of https://gitlab.freedesktop.org/drm/xe/kernel into drm-fixes 



Cross-subsystem Changes:
 - mm: Fix a hmm_range_fault() livelock / starvation problem (Thomas)

Core Changes:
 - Revert "drm/pagemap: Disable device-to-device migration" (Thomas)

Driver Changes:
 - Do not preempt fence signaling CS instructions (Brost)
 - Some leak and finalization fixes (Shuicheng, Tomasz, Varun, Zhanjun)
 - Workaround fix (Roper)

Signed-off-by: default avatarDave Airlie <airlied@redhat.com>

From: Rodrigo Vivi <rodrigo.vivi@intel.com>
Link: https://patch.msgid.link/aamGvvGRBRtX8-6u@intel.com
parents 43198996 0cfe9c48

drivers/gpu/drm/drm_pagemap.c

+2 −12
Original line number Diff line number Diff line
@@ -480,18 +480,8 @@ int drm_pagemap_migrate_to_devmem(struct drm_pagemap_devmem *devmem_allocation, 
		.start		= start,

		.end		= end,

		.pgmap_owner	= pagemap->owner,

		/*

		 * FIXME: MIGRATE_VMA_SELECT_DEVICE_PRIVATE intermittently

		 * causes 'xe_exec_system_allocator --r *race*no*' to trigger aa

		 * engine reset and a hard hang due to getting stuck on a folio

		 * lock. This should work and needs to be root-caused. The only

		 * downside of not selecting MIGRATE_VMA_SELECT_DEVICE_PRIVATE

		 * is that device-to-device migrations won’t work; instead,

		 * memory will bounce through system memory. This path should be

		 * rare and only occur when the madvise attributes of memory are

		 * changed or atomics are being used.

		 */

		.flags		= MIGRATE_VMA_SELECT_SYSTEM | MIGRATE_VMA_SELECT_DEVICE_COHERENT,

		.flags		= MIGRATE_VMA_SELECT_SYSTEM | MIGRATE_VMA_SELECT_DEVICE_COHERENT |

		MIGRATE_VMA_SELECT_DEVICE_PRIVATE,

	};

	unsigned long i, npages = npages_in_range(start, end);

	unsigned long own_pages = 0, migrated_pages = 0;

drivers/gpu/drm/xe/xe_configfs.c

+1 −0
Original line number Diff line number Diff line
@@ -830,6 +830,7 @@ static void xe_config_device_release(struct config_item *item) 


	mutex_destroy(&dev->lock);



	kfree(dev->config.ctx_restore_mid_bb[0].cs);

	kfree(dev->config.ctx_restore_post_bb[0].cs);

	kfree(dev);

}

drivers/gpu/drm/xe/xe_exec_queue.c

+11 −12
Original line number Diff line number Diff line
@@ -266,6 +266,16 @@ static struct xe_exec_queue *__xe_exec_queue_alloc(struct xe_device *xe, 
	return q;

}



static void __xe_exec_queue_fini(struct xe_exec_queue *q)

{

	int i;



	q->ops->fini(q);



	for (i = 0; i < q->width; ++i)

		xe_lrc_put(q->lrc[i]);

}



static int __xe_exec_queue_init(struct xe_exec_queue *q, u32 exec_queue_flags)

{

	int i, err;
@@ -320,21 +330,10 @@ static int __xe_exec_queue_init(struct xe_exec_queue *q, u32 exec_queue_flags) 
	return 0;



err_lrc:

	for (i = i - 1; i >= 0; --i)

		xe_lrc_put(q->lrc[i]);

	__xe_exec_queue_fini(q);

	return err;

}



static void __xe_exec_queue_fini(struct xe_exec_queue *q)

{

	int i;



	q->ops->fini(q);



	for (i = 0; i < q->width; ++i)

		xe_lrc_put(q->lrc[i]);

}



struct xe_exec_queue *xe_exec_queue_create(struct xe_device *xe, struct xe_vm *vm,

					   u32 logical_mask, u16 width,

					   struct xe_hw_engine *hwe, u32 flags,

drivers/gpu/drm/xe/xe_gsc_proxy.c

+35 −8
Original line number Diff line number Diff line
@@ -435,15 +435,11 @@ static int proxy_channel_alloc(struct xe_gsc *gsc) 
	return 0;

}



static void xe_gsc_proxy_remove(void *arg)

static void xe_gsc_proxy_stop(struct xe_gsc *gsc)

{

	struct xe_gsc *gsc = arg;

	struct xe_gt *gt = gsc_to_gt(gsc);

	struct xe_device *xe = gt_to_xe(gt);



	if (!gsc->proxy.component_added)

		return;



	/* disable HECI2 IRQs */

	scoped_guard(xe_pm_runtime, xe) {

		CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GSC);
@@ -455,6 +451,30 @@ static void xe_gsc_proxy_remove(void *arg) 
	}



	xe_gsc_wait_for_worker_completion(gsc);

	gsc->proxy.started = false;

}



static void xe_gsc_proxy_remove(void *arg)

{

	struct xe_gsc *gsc = arg;

	struct xe_gt *gt = gsc_to_gt(gsc);

	struct xe_device *xe = gt_to_xe(gt);



	if (!gsc->proxy.component_added)

		return;



	/*

	 * GSC proxy start is an async process that can be ongoing during

	 * Xe module load/unload. Using devm managed action to register

	 * xe_gsc_proxy_stop could cause issues if Xe module unload has

	 * already started when the action is registered, potentially leading

	 * to the cleanup being called at the wrong time. Therefore, instead

	 * of registering a separate devm action to undo what is done in

	 * proxy start, we call it from here, but only if the start has

	 * completed successfully (tracked with the 'started' flag).

	 */

	if (gsc->proxy.started)

		xe_gsc_proxy_stop(gsc);



	component_del(xe->drm.dev, &xe_gsc_proxy_component_ops);

	gsc->proxy.component_added = false;
@@ -510,6 +530,7 @@ int xe_gsc_proxy_init(struct xe_gsc *gsc) 
 */

int xe_gsc_proxy_start(struct xe_gsc *gsc)

{

	struct xe_gt *gt = gsc_to_gt(gsc);

	int err;



	/* enable the proxy interrupt in the GSC shim layer */
@@ -521,12 +542,18 @@ int xe_gsc_proxy_start(struct xe_gsc *gsc) 
	 */

	err = xe_gsc_proxy_request_handler(gsc);

	if (err)

		return err;

		goto err_irq_disable;



	if (!xe_gsc_proxy_init_done(gsc)) {

		xe_gt_err(gsc_to_gt(gsc), "GSC FW reports proxy init not completed\n");

		return -EIO;

		xe_gt_err(gt, "GSC FW reports proxy init not completed\n");

		err = -EIO;

		goto err_irq_disable;

	}



	gsc->proxy.started = true;

	return 0;



err_irq_disable:

	gsc_proxy_irq_toggle(gsc, false);

	return err;

}

drivers/gpu/drm/xe/xe_gsc_types.h

+2 −0
Original line number Diff line number Diff line
@@ -58,6 +58,8 @@ struct xe_gsc { 
		struct mutex mutex;

		/** @proxy.component_added: whether the component has been added */

		bool component_added;

		/** @proxy.started: whether the proxy has been started */

		bool started;

		/** @proxy.bo: object to store message to and from the GSC */

		struct xe_bo *bo;

		/** @proxy.to_gsc: map of the memory used to send messages to the GSC */