xref: /linux/drivers/gpu/drm/xe/xe_guc_submit.c (revision 570f7e331f5febb30f1384817463c7e42b65ca7d)

// SPDX-License-Identifier: MIT
/*
 * Copyright © 2022 Intel Corporation
 */

#include "xe_guc_submit.h"

#include <linux/bitfield.h>
#include <linux/bitmap.h>
#include <linux/circ_buf.h>
#include <linux/dma-fence-array.h>

#include <drm/drm_drv.h>
#include <drm/drm_managed.h>

#include "abi/guc_actions_abi.h"
#include "abi/guc_actions_slpc_abi.h"
#include "abi/guc_klvs_abi.h"
#include "xe_assert.h"
#include "xe_bo.h"
#include "xe_devcoredump.h"
#include "xe_device.h"
#include "xe_exec_queue.h"
#include "xe_force_wake.h"
#include "xe_gpu_scheduler.h"
#include "xe_gt.h"
#include "xe_gt_clock.h"
#include "xe_gt_printk.h"
#include "xe_guc.h"
#include "xe_guc_capture.h"
#include "xe_guc_ct.h"
#include "xe_guc_exec_queue_types.h"
#include "xe_guc_id_mgr.h"
#include "xe_guc_klv_helpers.h"
#include "xe_guc_submit_types.h"
#include "xe_hw_engine.h"
#include "xe_lrc.h"
#include "xe_macros.h"
#include "xe_map.h"
#include "xe_mocs.h"
#include "xe_module.h"
#include "xe_pm.h"
#include "xe_ring_ops_types.h"
#include "xe_sched_job.h"
#include "xe_sleep.h"
#include "xe_trace.h"
#include "xe_uc_fw.h"
#include "xe_vm.h"

#define XE_GUC_EXEC_QUEUE_CGP_CONTEXT_ERROR_LEN		6

static int guc_submit_reset_prepare(struct xe_guc *guc);

static struct xe_guc *
exec_queue_to_guc(struct xe_exec_queue *q)
{
	return &q->gt->uc.guc;
}

/*
 * Helpers for engine state, using an atomic as some of the bits can transition
 * as the same time (e.g. a suspend can be happning at the same time as schedule
 * engine done being processed).
 */
#define EXEC_QUEUE_STATE_REGISTERED		(1 << 0)
#define EXEC_QUEUE_STATE_ENABLED		(1 << 1)
#define EXEC_QUEUE_STATE_PENDING_ENABLE		(1 << 2)
#define EXEC_QUEUE_STATE_PENDING_DISABLE	(1 << 3)
#define EXEC_QUEUE_STATE_DESTROYED		(1 << 4)
#define EXEC_QUEUE_STATE_SUSPENDED		(1 << 5)
#define EXEC_QUEUE_STATE_RESET			(1 << 6)
#define EXEC_QUEUE_STATE_KILLED			(1 << 7)
#define EXEC_QUEUE_STATE_WEDGED			(1 << 8)
#define EXEC_QUEUE_STATE_BANNED			(1 << 9)
#define EXEC_QUEUE_STATE_PENDING_RESUME		(1 << 10)

static bool exec_queue_registered(struct xe_exec_queue *q)
{
	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_REGISTERED;
}

static void set_exec_queue_registered(struct xe_exec_queue *q)
{
	atomic_or(EXEC_QUEUE_STATE_REGISTERED, &q->guc->state);
}

static void clear_exec_queue_registered(struct xe_exec_queue *q)
{
	atomic_and(~EXEC_QUEUE_STATE_REGISTERED, &q->guc->state);
}

static bool exec_queue_enabled(struct xe_exec_queue *q)
{
	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_ENABLED;
}

static void set_exec_queue_enabled(struct xe_exec_queue *q)
{
	atomic_or(EXEC_QUEUE_STATE_ENABLED, &q->guc->state);
}

static void clear_exec_queue_enabled(struct xe_exec_queue *q)
{
	atomic_and(~EXEC_QUEUE_STATE_ENABLED, &q->guc->state);
}

static bool exec_queue_pending_enable(struct xe_exec_queue *q)
{
	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_PENDING_ENABLE;
}

static void set_exec_queue_pending_enable(struct xe_exec_queue *q)
{
	atomic_or(EXEC_QUEUE_STATE_PENDING_ENABLE, &q->guc->state);
}

static void clear_exec_queue_pending_enable(struct xe_exec_queue *q)
{
	atomic_and(~EXEC_QUEUE_STATE_PENDING_ENABLE, &q->guc->state);
}

static bool exec_queue_pending_disable(struct xe_exec_queue *q)
{
	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_PENDING_DISABLE;
}

static void set_exec_queue_pending_disable(struct xe_exec_queue *q)
{
	atomic_or(EXEC_QUEUE_STATE_PENDING_DISABLE, &q->guc->state);
}

static void clear_exec_queue_pending_disable(struct xe_exec_queue *q)
{
	atomic_and(~EXEC_QUEUE_STATE_PENDING_DISABLE, &q->guc->state);
}

static bool exec_queue_destroyed(struct xe_exec_queue *q)
{
	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_DESTROYED;
}

static void set_exec_queue_destroyed(struct xe_exec_queue *q)
{
	atomic_or(EXEC_QUEUE_STATE_DESTROYED, &q->guc->state);
}

static void clear_exec_queue_destroyed(struct xe_exec_queue *q)
{
	atomic_and(~EXEC_QUEUE_STATE_DESTROYED, &q->guc->state);
}

static bool exec_queue_banned(struct xe_exec_queue *q)
{
	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_BANNED;
}

static void set_exec_queue_banned(struct xe_exec_queue *q)
{
	atomic_or(EXEC_QUEUE_STATE_BANNED, &q->guc->state);
}

static void clear_exec_queue_banned(struct xe_exec_queue *q)
{
	atomic_andnot(EXEC_QUEUE_STATE_BANNED, &q->guc->state);
}

static bool exec_queue_suspended(struct xe_exec_queue *q)
{
	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_SUSPENDED;
}

static void set_exec_queue_suspended(struct xe_exec_queue *q)
{
	atomic_or(EXEC_QUEUE_STATE_SUSPENDED, &q->guc->state);
}

static void clear_exec_queue_suspended(struct xe_exec_queue *q)
{
	atomic_and(~EXEC_QUEUE_STATE_SUSPENDED, &q->guc->state);
}

static bool exec_queue_reset(struct xe_exec_queue *q)
{
	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_RESET;
}

static void set_exec_queue_reset(struct xe_exec_queue *q)
{
	atomic_or(EXEC_QUEUE_STATE_RESET, &q->guc->state);
}

static bool exec_queue_killed(struct xe_exec_queue *q)
{
	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_KILLED;
}

static void set_exec_queue_killed(struct xe_exec_queue *q)
{
	atomic_or(EXEC_QUEUE_STATE_KILLED, &q->guc->state);
}

static bool exec_queue_wedged(struct xe_exec_queue *q)
{
	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_WEDGED;
}

static void set_exec_queue_wedged(struct xe_exec_queue *q)
{
	atomic_or(EXEC_QUEUE_STATE_WEDGED, &q->guc->state);
}

static bool exec_queue_pending_resume(struct xe_exec_queue *q)
{
	return atomic_read(&q->guc->state) & EXEC_QUEUE_STATE_PENDING_RESUME;
}

static void set_exec_queue_pending_resume(struct xe_exec_queue *q)
{
	atomic_or(EXEC_QUEUE_STATE_PENDING_RESUME, &q->guc->state);
}

static void clear_exec_queue_pending_resume(struct xe_exec_queue *q)
{
	atomic_and(~EXEC_QUEUE_STATE_PENDING_RESUME, &q->guc->state);
}

static bool exec_queue_killed_or_banned_or_wedged(struct xe_exec_queue *q)
{
	return (atomic_read(&q->guc->state) &
		(EXEC_QUEUE_STATE_WEDGED | EXEC_QUEUE_STATE_KILLED |
		 EXEC_QUEUE_STATE_BANNED));
}

static void guc_submit_sw_fini(struct drm_device *drm, void *arg)
{
	struct xe_guc *guc = arg;
	struct xe_gt *gt = guc_to_gt(guc);

	xe_gt_assert(gt, xa_empty(&guc->submission_state.exec_queue_lookup));

	xa_destroy(&guc->submission_state.exec_queue_lookup);
}

static void guc_submit_fini(void *arg)
{
	struct xe_guc *guc = arg;
	struct xe_exec_queue *q;
	unsigned long index;

	/* Drop any wedged queue refs */
	mutex_lock(&guc->submission_state.lock);
	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q) {
		if (exec_queue_wedged(q)) {
			mutex_unlock(&guc->submission_state.lock);
			xe_exec_queue_put(q);
			mutex_lock(&guc->submission_state.lock);
		}
	}
	mutex_unlock(&guc->submission_state.lock);

	/* Forcefully kill any remaining exec queues */
	xe_guc_ct_stop(&guc->ct);
	guc_submit_reset_prepare(guc);
	xe_guc_softreset(guc);
	xe_guc_submit_stop(guc);
	xe_uc_fw_sanitize(&guc->fw);
	xe_guc_submit_pause_abort(guc);
}

static const struct xe_exec_queue_ops guc_exec_queue_ops;

static void primelockdep(struct xe_guc *guc)
{
	if (!IS_ENABLED(CONFIG_LOCKDEP))
		return;

	fs_reclaim_acquire(GFP_KERNEL);

	mutex_lock(&guc->submission_state.lock);
	mutex_unlock(&guc->submission_state.lock);

	fs_reclaim_release(GFP_KERNEL);
}

/**
 * xe_guc_submit_init() - Initialize GuC submission.
 * @guc: the &xe_guc to initialize
 * @num_ids: number of GuC context IDs to use
 *
 * The bare-metal or PF driver can pass ~0 as &num_ids to indicate that all
 * GuC context IDs supported by the GuC firmware should be used for submission.
 *
 * Only VF drivers will have to provide explicit number of GuC context IDs
 * that they can use for submission.
 *
 * Return: 0 on success or a negative error code on failure.
 */
int xe_guc_submit_init(struct xe_guc *guc, unsigned int num_ids)
{
	struct xe_device *xe = guc_to_xe(guc);
	struct xe_gt *gt = guc_to_gt(guc);
	int err;

	err = drmm_mutex_init(&xe->drm, &guc->submission_state.lock);
	if (err)
		return err;

	err = xe_guc_id_mgr_init(&guc->submission_state.idm, num_ids);
	if (err)
		return err;

	gt->exec_queue_ops = &guc_exec_queue_ops;

	xa_init(&guc->submission_state.exec_queue_lookup);

	primelockdep(guc);

	guc->submission_state.initialized = true;

	err = drmm_add_action_or_reset(&xe->drm, guc_submit_sw_fini, guc);
	if (err)
		return err;

	return devm_add_action_or_reset(xe->drm.dev, guc_submit_fini, guc);
}

/*
 * Given that we want to guarantee enough RCS throughput to avoid missing
 * frames, we set the yield policy to 20% of each 80ms interval.
 */
#define RC_YIELD_DURATION	80	/* in ms */
#define RC_YIELD_RATIO		20	/* in percent */
static u32 *emit_render_compute_yield_klv(u32 *emit)
{
	*emit++ = PREP_GUC_KLV_TAG(SCHEDULING_POLICIES_RENDER_COMPUTE_YIELD);
	*emit++ = RC_YIELD_DURATION;
	*emit++ = RC_YIELD_RATIO;

	return emit;
}

#define SCHEDULING_POLICY_MAX_DWORDS 16
static int guc_init_global_schedule_policy(struct xe_guc *guc)
{
	u32 data[SCHEDULING_POLICY_MAX_DWORDS];
	u32 *emit = data;
	u32 count = 0;
	int ret;

	if (GUC_SUBMIT_VER(guc) < MAKE_GUC_VER(1, 1, 0))
		return 0;

	*emit++ = XE_GUC_ACTION_UPDATE_SCHEDULING_POLICIES_KLV;

	if (CCS_INSTANCES(guc_to_gt(guc)))
		emit = emit_render_compute_yield_klv(emit);

	count = emit - data;
	if (count > 1) {
		xe_assert(guc_to_xe(guc), count <= SCHEDULING_POLICY_MAX_DWORDS);

		ret = xe_guc_ct_send_block(&guc->ct, data, count);
		if (ret < 0) {
			xe_gt_err(guc_to_gt(guc),
				  "failed to enable GuC scheduling policies: %pe\n",
				  ERR_PTR(ret));
			return ret;
		}
	}

	return 0;
}

int xe_guc_submit_enable(struct xe_guc *guc)
{
	int ret;

	ret = guc_init_global_schedule_policy(guc);
	if (ret)
		return ret;

	guc->submission_state.enabled = true;

	return 0;
}

void xe_guc_submit_disable(struct xe_guc *guc)
{
	guc->submission_state.enabled = false;
}

static void __release_guc_id(struct xe_guc *guc, struct xe_exec_queue *q,
			     int count)
{
	int i;

	mutex_lock(&guc->submission_state.lock);

	for (i = 0; i < count; ++i)
		xa_erase(&guc->submission_state.exec_queue_lookup,
			 q->guc->id + i);

	xe_guc_id_mgr_release_locked(&guc->submission_state.idm,
				     q->guc->id, q->width);

	mutex_unlock(&guc->submission_state.lock);
}

static int alloc_guc_id(struct xe_guc *guc, struct xe_exec_queue *q)
{
	int ret, i;

	mutex_lock(&guc->submission_state.lock);
	ret = xe_guc_id_mgr_reserve_locked(&guc->submission_state.idm,
					   q->width);
	mutex_unlock(&guc->submission_state.lock);
	if (ret < 0)
		return ret;

	q->guc->id = ret;

	/* Reserve empty slots. */
	for (i = 0; i < q->width; ++i) {
		ret = xa_insert(&guc->submission_state.exec_queue_lookup,
				 q->guc->id + i, NULL, GFP_KERNEL);
		if (ret)
			goto err_release;
	}

	return 0;

err_release:
	__release_guc_id(guc, q, i);

	return ret;
}

static void publish_guc_id(struct xe_guc *guc, struct xe_exec_queue *q)
{
	int i;

	lockdep_assert_held(&guc->submission_state.lock);

	for (i = 0; i < q->width; ++i) {
		void *old;

		old = xa_store(&guc->submission_state.exec_queue_lookup,
			       q->guc->id + i, q, GFP_NOWAIT);
		XE_WARN_ON(old || xa_is_err(old));
	}
}

static void release_guc_id(struct xe_guc *guc, struct xe_exec_queue *q)
{
	__release_guc_id(guc, q, q->width);
}

struct exec_queue_policy {
	u32 count;
	struct guc_update_exec_queue_policy h2g;
};

static u32 __guc_exec_queue_policy_action_size(struct exec_queue_policy *policy)
{
	size_t bytes = sizeof(policy->h2g.header) +
		       (sizeof(policy->h2g.klv[0]) * policy->count);

	return bytes / sizeof(u32);
}

static void __guc_exec_queue_policy_start_klv(struct exec_queue_policy *policy,
					      u16 guc_id)
{
	policy->h2g.header.action =
		XE_GUC_ACTION_HOST2GUC_UPDATE_CONTEXT_POLICIES;
	policy->h2g.header.guc_id = guc_id;
	policy->count = 0;
}

#define MAKE_EXEC_QUEUE_POLICY_ADD(func, id) \
static void __guc_exec_queue_policy_add_##func(struct exec_queue_policy *policy, \
					   u32 data) \
{ \
	XE_WARN_ON(policy->count >= GUC_CONTEXT_POLICIES_KLV_NUM_IDS); \
\
	policy->h2g.klv[policy->count].kl = \
		FIELD_PREP(GUC_KLV_0_KEY, \
			   GUC_CONTEXT_POLICIES_KLV_ID_##id) | \
		FIELD_PREP(GUC_KLV_0_LEN, 1); \
	policy->h2g.klv[policy->count].value = data; \
	policy->count++; \
}

MAKE_EXEC_QUEUE_POLICY_ADD(execution_quantum, EXECUTION_QUANTUM)
MAKE_EXEC_QUEUE_POLICY_ADD(preemption_timeout, PREEMPTION_TIMEOUT)
MAKE_EXEC_QUEUE_POLICY_ADD(priority, SCHEDULING_PRIORITY)
MAKE_EXEC_QUEUE_POLICY_ADD(slpc_exec_queue_freq_req, SLPM_GT_FREQUENCY)
#undef MAKE_EXEC_QUEUE_POLICY_ADD

static const int xe_exec_queue_prio_to_guc[] = {
	[XE_EXEC_QUEUE_PRIORITY_LOW] = GUC_CLIENT_PRIORITY_NORMAL,
	[XE_EXEC_QUEUE_PRIORITY_NORMAL] = GUC_CLIENT_PRIORITY_KMD_NORMAL,
	[XE_EXEC_QUEUE_PRIORITY_HIGH] = GUC_CLIENT_PRIORITY_HIGH,
	[XE_EXEC_QUEUE_PRIORITY_KERNEL] = GUC_CLIENT_PRIORITY_KMD_HIGH,
};

static void init_policies(struct xe_guc *guc, struct xe_exec_queue *q)
{
	struct exec_queue_policy policy;
	enum xe_exec_queue_priority prio = q->sched_props.priority;
	u32 timeslice_us = q->sched_props.timeslice_us;
	u32 slpc_exec_queue_freq_req = 0;
	u32 preempt_timeout_us = q->sched_props.preempt_timeout_us;

	xe_gt_assert(guc_to_gt(guc), exec_queue_registered(q) &&
		     !xe_exec_queue_is_multi_queue_secondary(q));

	if (q->flags & EXEC_QUEUE_FLAG_LOW_LATENCY)
		slpc_exec_queue_freq_req |= SLPC_CTX_FREQ_REQ_IS_COMPUTE;

	__guc_exec_queue_policy_start_klv(&policy, q->guc->id);
	__guc_exec_queue_policy_add_priority(&policy, xe_exec_queue_prio_to_guc[prio]);
	__guc_exec_queue_policy_add_execution_quantum(&policy, timeslice_us);
	__guc_exec_queue_policy_add_preemption_timeout(&policy, preempt_timeout_us);
	__guc_exec_queue_policy_add_slpc_exec_queue_freq_req(&policy,
							     slpc_exec_queue_freq_req);

	xe_guc_ct_send(&guc->ct, (u32 *)&policy.h2g,
		       __guc_exec_queue_policy_action_size(&policy), 0, 0);
}

static void set_min_preemption_timeout(struct xe_guc *guc, struct xe_exec_queue *q)
{
	struct exec_queue_policy policy;

	xe_assert(guc_to_xe(guc), !xe_exec_queue_is_multi_queue_secondary(q));

	__guc_exec_queue_policy_start_klv(&policy, q->guc->id);
	__guc_exec_queue_policy_add_preemption_timeout(&policy, 1);

	xe_guc_ct_send(&guc->ct, (u32 *)&policy.h2g,
		       __guc_exec_queue_policy_action_size(&policy), 0, 0);
}

static bool vf_recovery(struct xe_guc *guc)
{
	return xe_gt_recovery_pending(guc_to_gt(guc));
}

static void xe_guc_exec_queue_trigger_cleanup(struct xe_exec_queue *q)
{
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_device *xe = guc_to_xe(guc);

	/** to wakeup xe_wait_user_fence ioctl if exec queue is reset */
	wake_up_all(&xe->ufence_wq);

	xe_sched_tdr_queue_imm(&q->guc->sched);
}

static void xe_guc_exec_queue_group_stop(struct xe_exec_queue *q)
{
	struct xe_exec_queue *primary = xe_exec_queue_multi_queue_primary(q);
	struct xe_exec_queue_group *group = q->multi_queue.group;
	struct xe_exec_queue *eq, *next;
	LIST_HEAD(tmp);

	xe_gt_assert(guc_to_gt(exec_queue_to_guc(q)),
		     xe_exec_queue_is_multi_queue(q));

	mutex_lock(&group->list_lock);

	/*
	 * Stop all future queues being from executing while group is stopped.
	 */
	group->stopped = true;

	list_for_each_entry_safe(eq, next, &group->list, multi_queue.link)
		/*
		 * Refcount prevents an attempted removal from &group->list,
		 * temporary list allows safe iteration after dropping
		 * &group->list_lock.
		 */
		if (xe_exec_queue_get_unless_zero(eq))
			list_move_tail(&eq->multi_queue.link, &tmp);

	mutex_unlock(&group->list_lock);

	/* We cannot stop under list lock without getting inversions */
	xe_sched_submission_stop(&primary->guc->sched);
	list_for_each_entry(eq, &tmp, multi_queue.link)
		xe_sched_submission_stop(&eq->guc->sched);

	mutex_lock(&group->list_lock);
	list_for_each_entry_safe(eq, next, &tmp, multi_queue.link) {
		/*
		 * Corner where we got banned while stopping and not on
		 * &group->list
		 */
		if (READ_ONCE(group->banned))
			xe_guc_exec_queue_trigger_cleanup(eq);

		list_move_tail(&eq->multi_queue.link, &group->list);
		xe_exec_queue_put(eq);
	}
	mutex_unlock(&group->list_lock);
}

static void xe_guc_exec_queue_group_start(struct xe_exec_queue *q)
{
	struct xe_exec_queue *primary = xe_exec_queue_multi_queue_primary(q);
	struct xe_exec_queue_group *group = q->multi_queue.group;
	struct xe_exec_queue *eq;

	xe_gt_assert(guc_to_gt(exec_queue_to_guc(q)),
		     xe_exec_queue_is_multi_queue(q));

	xe_sched_submission_start(&primary->guc->sched);

	mutex_lock(&group->list_lock);
	group->stopped = false;
	list_for_each_entry(eq, &group->list, multi_queue.link)
		xe_sched_submission_start(&eq->guc->sched);
	mutex_unlock(&group->list_lock);
}

static void xe_guc_exec_queue_group_trigger_cleanup(struct xe_exec_queue *q)
{
	struct xe_exec_queue *primary = xe_exec_queue_multi_queue_primary(q);
	struct xe_exec_queue_group *group = q->multi_queue.group;
	struct xe_exec_queue *eq;

	xe_gt_assert(guc_to_gt(exec_queue_to_guc(q)),
		     xe_exec_queue_is_multi_queue(q));

	/* Group banned, skip timeout check in TDR */
	WRITE_ONCE(group->banned, true);
	xe_guc_exec_queue_trigger_cleanup(primary);

	mutex_lock(&group->list_lock);
	list_for_each_entry(eq, &group->list, multi_queue.link)
		xe_guc_exec_queue_trigger_cleanup(eq);
	mutex_unlock(&group->list_lock);
}

static void xe_guc_exec_queue_reset_trigger_cleanup(struct xe_exec_queue *q)
{
	if (xe_exec_queue_is_multi_queue(q)) {
		struct xe_exec_queue *primary = xe_exec_queue_multi_queue_primary(q);
		struct xe_exec_queue_group *group = q->multi_queue.group;
		struct xe_exec_queue *eq;

		/* Group banned, skip timeout check in TDR */
		WRITE_ONCE(group->banned, true);

		set_exec_queue_reset(primary);
		if (!exec_queue_banned(primary))
			xe_guc_exec_queue_trigger_cleanup(primary);

		mutex_lock(&group->list_lock);
		list_for_each_entry(eq, &group->list, multi_queue.link) {
			set_exec_queue_reset(eq);
			if (!exec_queue_banned(eq))
				xe_guc_exec_queue_trigger_cleanup(eq);
		}
		mutex_unlock(&group->list_lock);
	} else {
		set_exec_queue_reset(q);
		if (!exec_queue_banned(q))
			xe_guc_exec_queue_trigger_cleanup(q);
	}
}

static void set_exec_queue_group_banned(struct xe_exec_queue *q)
{
	struct xe_exec_queue *primary = xe_exec_queue_multi_queue_primary(q);
	struct xe_exec_queue_group *group = q->multi_queue.group;
	struct xe_exec_queue *eq;

	/* Ban all queues of the multi-queue group */
	xe_gt_assert(guc_to_gt(exec_queue_to_guc(q)),
		     xe_exec_queue_is_multi_queue(q));
	set_exec_queue_banned(primary);

	mutex_lock(&group->list_lock);
	list_for_each_entry(eq, &group->list, multi_queue.link)
		set_exec_queue_banned(eq);
	mutex_unlock(&group->list_lock);
}

/* Helper for context registration H2G */
struct guc_ctxt_registration_info {
	u32 flags;
	u32 context_idx;
	u32 engine_class;
	u32 engine_submit_mask;
	u32 wq_desc_lo;
	u32 wq_desc_hi;
	u32 wq_base_lo;
	u32 wq_base_hi;
	u32 wq_size;
	u32 cgp_lo;
	u32 cgp_hi;
	u32 hwlrca_lo;
	u32 hwlrca_hi;
};

#define parallel_read(xe_, map_, field_) \
	xe_map_rd_field(xe_, &map_, 0, struct guc_submit_parallel_scratch, \
			field_)
#define parallel_write(xe_, map_, field_, val_) \
	xe_map_wr_field(xe_, &map_, 0, struct guc_submit_parallel_scratch, \
			field_, val_)

/**
 * DOC: Multi Queue Group GuC interface
 *
 * The multi queue group coordination between KMD and GuC is through a software
 * construct called Context Group Page (CGP). The CGP is a KMD managed 4KB page
 * allocated in the global GTT.
 *
 * CGP format:
 *
 * +-----------+---------------------------+---------------------------------------------+
 * | DWORD     | Name                      | Description                                 |
 * +-----------+---------------------------+---------------------------------------------+
 * | 0         | Version                   | Bits [15:8]=Major ver, [7:0]=Minor ver      |
 * +-----------+---------------------------+---------------------------------------------+
 * | 1..15     | RESERVED                  | MBZ                                         |
 * +-----------+---------------------------+---------------------------------------------+
 * | 16        | KMD_QUEUE_UPDATE_MASK_DW0 | KMD queue mask for queues 31..0             |
 * +-----------+---------------------------+---------------------------------------------+
 * | 17        | KMD_QUEUE_UPDATE_MASK_DW1 | KMD queue mask for queues 63..32            |
 * +-----------+---------------------------+---------------------------------------------+
 * | 18..31    | RESERVED                  | MBZ                                         |
 * +-----------+---------------------------+---------------------------------------------+
 * | 32        | Q0CD_DW0                  | Queue 0 context LRC descriptor lower DWORD  |
 * +-----------+---------------------------+---------------------------------------------+
 * | 33        | Q0ContextIndex            | Context ID for Queue 0                      |
 * +-----------+---------------------------+---------------------------------------------+
 * | 34        | Q1CD_DW0                  | Queue 1 context LRC descriptor lower DWORD  |
 * +-----------+---------------------------+---------------------------------------------+
 * | 35        | Q1ContextIndex            | Context ID for Queue 1                      |
 * +-----------+---------------------------+---------------------------------------------+
 * | ...       |...                        | ...                                         |
 * +-----------+---------------------------+---------------------------------------------+
 * | 158       | Q63CD_DW0                 | Queue 63 context LRC descriptor lower DWORD |
 * +-----------+---------------------------+---------------------------------------------+
 * | 159       | Q63ContextIndex           | Context ID for Queue 63                     |
 * +-----------+---------------------------+---------------------------------------------+
 * | 160..1024 | RESERVED                  | MBZ                                         |
 * +-----------+---------------------------+---------------------------------------------+
 *
 * While registering Q0 with GuC, CGP is updated with Q0 entry and GuC is notified
 * through XE_GUC_ACTION_REGISTER_CONTEXT_MULTI_QUEUE H2G message which specifies
 * the CGP address. When the secondary queues are added to the group, the CGP is
 * updated with entry for that queue and GuC is notified through the H2G interface
 * XE_GUC_ACTION_MULTI_QUEUE_CONTEXT_CGP_SYNC. GuC responds to these H2G messages
 * with a XE_GUC_ACTION_NOTIFY_MULTIQ_CONTEXT_CGP_SYNC_DONE G2H message. GuC also
 * sends a XE_GUC_ACTION_NOTIFY_MULTI_QUEUE_CGP_CONTEXT_ERROR notification for any
 * error in the CGP. Only one of these CGP update messages can be outstanding
 * (waiting for GuC response) at any time. The bits in KMD_QUEUE_UPDATE_MASK_DW*
 * fields indicate which queue entry is being updated in the CGP.
 *
 * The primary queue (Q0) represents the multi queue group context in GuC and
 * submission on any queue of the group must be through Q0 GuC interface only.
 *
 * As it is not required to register secondary queues with GuC, the secondary queue
 * context ids in the CGP are populated with Q0 context id.
 */

#define CGP_VERSION_MAJOR_SHIFT	8

static void xe_guc_exec_queue_group_cgp_update(struct xe_device *xe,
					       struct xe_exec_queue *q)
{
	struct xe_exec_queue_group *group = q->multi_queue.group;
	u32 guc_id = group->primary->guc->id;

	/* Currently implementing CGP version 1.0 */
	xe_map_wr(xe, &group->cgp_bo->vmap, 0, u32,
		  1 << CGP_VERSION_MAJOR_SHIFT);

	xe_map_wr(xe, &group->cgp_bo->vmap,
		  (32 + q->multi_queue.pos * 2) * sizeof(u32),
		  u32, lower_32_bits(xe_lrc_descriptor(q->lrc[0])));

	xe_map_wr(xe, &group->cgp_bo->vmap,
		  (33 + q->multi_queue.pos * 2) * sizeof(u32),
		  u32, guc_id);

	if (q->multi_queue.pos / 32) {
		xe_map_wr(xe, &group->cgp_bo->vmap, 17 * sizeof(u32),
			  u32, BIT(q->multi_queue.pos % 32));
		xe_map_wr(xe, &group->cgp_bo->vmap, 16 * sizeof(u32), u32, 0);
	} else {
		xe_map_wr(xe, &group->cgp_bo->vmap, 16 * sizeof(u32),
			  u32, BIT(q->multi_queue.pos));
		xe_map_wr(xe, &group->cgp_bo->vmap, 17 * sizeof(u32), u32, 0);
	}
}

static void xe_guc_exec_queue_group_cgp_sync(struct xe_guc *guc,
					     struct xe_exec_queue *q,
					     const u32 *action, u32 len)
{
	struct xe_exec_queue_group *group = q->multi_queue.group;
	struct xe_device *xe = guc_to_xe(guc);
	enum xe_multi_queue_priority priority;
	long ret;

	/*
	 * As all queues of a multi queue group use single drm scheduler
	 * submit workqueue, CGP synchronization with GuC are serialized.
	 * Hence, no locking is required here.
	 * Wait for any pending CGP_SYNC_DONE response before updating the
	 * CGP page and sending CGP_SYNC message.
	 *
	 * FIXME: Support VF migration
	 */
	ret = wait_event_timeout(guc->ct.wq,
				 !READ_ONCE(group->sync_pending) ||
				 xe_guc_read_stopped(guc), HZ);
	if (!ret || xe_guc_read_stopped(guc)) {
		/* CGP_SYNC failed. Reset gt, cleanup the group */
		xe_gt_warn(guc_to_gt(guc), "Wait for CGP_SYNC_DONE response failed!\n");
		set_exec_queue_group_banned(q);
		xe_gt_reset_async(q->gt);
		xe_guc_exec_queue_group_trigger_cleanup(q);
		return;
	}

	scoped_guard(spinlock, &q->multi_queue.lock)
		priority = q->multi_queue.priority;

	xe_lrc_set_multi_queue_priority(q->lrc[0], priority);
	xe_guc_exec_queue_group_cgp_update(xe, q);

	WRITE_ONCE(group->sync_pending, true);
	xe_guc_ct_send(&guc->ct, action, len, G2H_LEN_DW_MULTI_QUEUE_CONTEXT, 1);
}

static void guc_exec_queue_send_cgp_sync(struct xe_exec_queue *q)
{
#define MAX_MULTI_QUEUE_CGP_SYNC_SIZE	(2)
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_exec_queue_group *group = q->multi_queue.group;
	u32 action[MAX_MULTI_QUEUE_CGP_SYNC_SIZE];
	int len = 0;

	action[len++] = XE_GUC_ACTION_MULTI_QUEUE_CONTEXT_CGP_SYNC;
	action[len++] = group->primary->guc->id;

	xe_gt_assert(guc_to_gt(guc), len <= MAX_MULTI_QUEUE_CGP_SYNC_SIZE);
#undef MAX_MULTI_QUEUE_CGP_SYNC_SIZE

	xe_guc_exec_queue_group_cgp_sync(guc, q, action, len);
}

static void __register_exec_queue_group(struct xe_exec_queue *q,
					struct guc_ctxt_registration_info *info)
{
	struct xe_guc *guc = exec_queue_to_guc(q);
#define MAX_MULTI_QUEUE_REG_SIZE	(8)
	u32 action[MAX_MULTI_QUEUE_REG_SIZE];
	int len = 0;

	action[len++] = XE_GUC_ACTION_REGISTER_CONTEXT_MULTI_QUEUE;
	action[len++] = info->flags;
	action[len++] = info->context_idx;
	action[len++] = info->engine_class;
	action[len++] = info->engine_submit_mask;
	action[len++] = 0; /* Reserved */
	action[len++] = info->cgp_lo;
	action[len++] = info->cgp_hi;

	xe_gt_assert(guc_to_gt(guc), len <= MAX_MULTI_QUEUE_REG_SIZE);
#undef MAX_MULTI_QUEUE_REG_SIZE

	/*
	 * The above XE_GUC_ACTION_REGISTER_CONTEXT_MULTI_QUEUE do expect a
	 * XE_GUC_ACTION_NOTIFY_MULTI_QUEUE_CONTEXT_CGP_SYNC_DONE response
	 * from guc.
	 */
	xe_guc_exec_queue_group_cgp_sync(guc, q, action, len);
}

static void __register_mlrc_exec_queue(struct xe_guc *guc,
				       struct xe_exec_queue *q,
				       struct guc_ctxt_registration_info *info)
{
#define MAX_MLRC_REG_SIZE      (13 + XE_HW_ENGINE_MAX_INSTANCE * 2)
	u32 action[MAX_MLRC_REG_SIZE];
	int len = 0;
	int i;

	xe_gt_assert(guc_to_gt(guc), xe_exec_queue_is_parallel(q));

	action[len++] = XE_GUC_ACTION_REGISTER_CONTEXT_MULTI_LRC;
	action[len++] = info->flags;
	action[len++] = info->context_idx;
	action[len++] = info->engine_class;
	action[len++] = info->engine_submit_mask;
	action[len++] = info->wq_desc_lo;
	action[len++] = info->wq_desc_hi;
	action[len++] = info->wq_base_lo;
	action[len++] = info->wq_base_hi;
	action[len++] = info->wq_size;
	action[len++] = q->width;
	action[len++] = info->hwlrca_lo;
	action[len++] = info->hwlrca_hi;

	for (i = 1; i < q->width; ++i) {
		struct xe_lrc *lrc = q->lrc[i];

		action[len++] = lower_32_bits(xe_lrc_descriptor(lrc));
		action[len++] = upper_32_bits(xe_lrc_descriptor(lrc));
	}

	/* explicitly checks some fields that we might fixup later */
	xe_gt_assert(guc_to_gt(guc), info->wq_desc_lo ==
		     action[XE_GUC_REGISTER_CONTEXT_MULTI_LRC_DATA_5_WQ_DESC_ADDR_LOWER]);
	xe_gt_assert(guc_to_gt(guc), info->wq_base_lo ==
		     action[XE_GUC_REGISTER_CONTEXT_MULTI_LRC_DATA_7_WQ_BUF_BASE_LOWER]);
	xe_gt_assert(guc_to_gt(guc), q->width ==
		     action[XE_GUC_REGISTER_CONTEXT_MULTI_LRC_DATA_10_NUM_CTXS]);
	xe_gt_assert(guc_to_gt(guc), info->hwlrca_lo ==
		     action[XE_GUC_REGISTER_CONTEXT_MULTI_LRC_DATA_11_HW_LRC_ADDR]);
	xe_gt_assert(guc_to_gt(guc), len <= MAX_MLRC_REG_SIZE);
#undef MAX_MLRC_REG_SIZE

	xe_guc_ct_send(&guc->ct, action, len, 0, 0);
}

static void __register_exec_queue(struct xe_guc *guc,
				  struct guc_ctxt_registration_info *info)
{
	u32 action[] = {
		XE_GUC_ACTION_REGISTER_CONTEXT,
		info->flags,
		info->context_idx,
		info->engine_class,
		info->engine_submit_mask,
		info->wq_desc_lo,
		info->wq_desc_hi,
		info->wq_base_lo,
		info->wq_base_hi,
		info->wq_size,
		info->hwlrca_lo,
		info->hwlrca_hi,
	};

	/* explicitly checks some fields that we might fixup later */
	xe_gt_assert(guc_to_gt(guc), info->wq_desc_lo ==
		     action[XE_GUC_REGISTER_CONTEXT_DATA_5_WQ_DESC_ADDR_LOWER]);
	xe_gt_assert(guc_to_gt(guc), info->wq_base_lo ==
		     action[XE_GUC_REGISTER_CONTEXT_DATA_7_WQ_BUF_BASE_LOWER]);
	xe_gt_assert(guc_to_gt(guc), info->hwlrca_lo ==
		     action[XE_GUC_REGISTER_CONTEXT_DATA_10_HW_LRC_ADDR]);

	xe_guc_ct_send(&guc->ct, action, ARRAY_SIZE(action), 0, 0);
}

static u32 xe_hwe_guc_logical_to_submit_mask(struct xe_hw_engine *hwe, u32 logical_mask)
{
	struct xe_gt *gt = hwe->gt;

	if (xe_gt_is_usm_hwe(gt, hwe)) {
		int shift = gt->usm.paging_hwe0->logical_instance;
		u32 paging_logical_mask = gt->usm.paging_logical_mask;

		xe_gt_assert(gt, (logical_mask & paging_logical_mask) == logical_mask);

		/*
		 * Remap to GUC_PAGING_CLASS logical instance mask, if
		 * applicable.
		 */
		if (xe_guc_has_paging_engine(&hwe->gt->uc.guc))
			return logical_mask >> shift;
	}

	return logical_mask;
}

static void register_exec_queue(struct xe_exec_queue *q, int ctx_type)
{
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_device *xe = guc_to_xe(guc);
	struct xe_lrc *lrc = q->lrc[0];
	struct guc_ctxt_registration_info info;

	xe_gt_assert(guc_to_gt(guc), !exec_queue_registered(q));
	xe_gt_assert(guc_to_gt(guc), ctx_type < GUC_CONTEXT_COUNT);

	memset(&info, 0, sizeof(info));
	info.context_idx = q->guc->id;
	info.engine_class = xe_hwe_to_guc_class(q->hwe);
	info.engine_submit_mask =
		xe_hwe_guc_logical_to_submit_mask(q->hwe, q->logical_mask);
	info.hwlrca_lo = lower_32_bits(xe_lrc_descriptor(lrc));
	info.hwlrca_hi = upper_32_bits(xe_lrc_descriptor(lrc));
	info.flags = CONTEXT_REGISTRATION_FLAG_KMD |
		FIELD_PREP(CONTEXT_REGISTRATION_FLAG_TYPE, ctx_type);

	if (xe_exec_queue_is_multi_queue(q)) {
		struct xe_exec_queue_group *group = q->multi_queue.group;

		info.cgp_lo = xe_bo_ggtt_addr(group->cgp_bo);
		info.cgp_hi = 0;
	}

	if (xe_exec_queue_is_parallel(q)) {
		u64 ggtt_addr = xe_lrc_parallel_ggtt_addr(lrc);
		struct iosys_map map = xe_lrc_parallel_map(lrc);

		info.wq_desc_lo = lower_32_bits(ggtt_addr +
			offsetof(struct guc_submit_parallel_scratch, wq_desc));
		info.wq_desc_hi = upper_32_bits(ggtt_addr +
			offsetof(struct guc_submit_parallel_scratch, wq_desc));
		info.wq_base_lo = lower_32_bits(ggtt_addr +
			offsetof(struct guc_submit_parallel_scratch, wq[0]));
		info.wq_base_hi = upper_32_bits(ggtt_addr +
			offsetof(struct guc_submit_parallel_scratch, wq[0]));
		info.wq_size = WQ_SIZE;

		q->guc->wqi_head = 0;
		q->guc->wqi_tail = 0;
		xe_map_memset(xe, &map, 0, 0, PARALLEL_SCRATCH_SIZE - WQ_SIZE);
		parallel_write(xe, map, wq_desc.wq_status, WQ_STATUS_ACTIVE);
	}

	set_exec_queue_registered(q);
	trace_xe_exec_queue_register(q);
	if (xe_exec_queue_is_multi_queue_primary(q))
		__register_exec_queue_group(q, &info);
	else if (xe_exec_queue_is_parallel(q))
		__register_mlrc_exec_queue(guc, q, &info);
	else if (!xe_exec_queue_is_multi_queue_secondary(q))
		__register_exec_queue(guc, &info);

	if (!xe_exec_queue_is_multi_queue_secondary(q))
		init_policies(guc, q);

	if (xe_exec_queue_is_multi_queue_secondary(q))
		guc_exec_queue_send_cgp_sync(q);
}

static u32 wq_space_until_wrap(struct xe_exec_queue *q)
{
	return (WQ_SIZE - q->guc->wqi_tail);
}

static int wq_wait_for_space(struct xe_exec_queue *q, u32 wqi_size)
{
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_device *xe = guc_to_xe(guc);
	struct iosys_map map = xe_lrc_parallel_map(q->lrc[0]);
	unsigned int sleep_period_ms = 1, sleep_total_ms = 0;

#define AVAILABLE_SPACE \
	CIRC_SPACE(q->guc->wqi_tail, q->guc->wqi_head, WQ_SIZE)
	if (wqi_size > AVAILABLE_SPACE && !vf_recovery(guc)) {
try_again:
		q->guc->wqi_head = parallel_read(xe, map, wq_desc.head);
		if (wqi_size > AVAILABLE_SPACE && !vf_recovery(guc)) {
			if (sleep_total_ms > 2000) {
				xe_gt_reset_async(q->gt);
				return -ENODEV;
			}

			sleep_total_ms += xe_sleep_exponential_ms(&sleep_period_ms, 64);
			goto try_again;
		}
	}
#undef AVAILABLE_SPACE

	return 0;
}

static int wq_noop_append(struct xe_exec_queue *q)
{
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_device *xe = guc_to_xe(guc);
	struct iosys_map map = xe_lrc_parallel_map(q->lrc[0]);
	u32 len_dw = wq_space_until_wrap(q) / sizeof(u32) - 1;

	if (wq_wait_for_space(q, wq_space_until_wrap(q)))
		return -ENODEV;

	xe_gt_assert(guc_to_gt(guc), FIELD_FIT(WQ_LEN_MASK, len_dw));

	parallel_write(xe, map, wq[q->guc->wqi_tail / sizeof(u32)],
		       FIELD_PREP(WQ_TYPE_MASK, WQ_TYPE_NOOP) |
		       FIELD_PREP(WQ_LEN_MASK, len_dw));
	q->guc->wqi_tail = 0;

	return 0;
}

static void wq_item_append(struct xe_exec_queue *q)
{
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_device *xe = guc_to_xe(guc);
	struct iosys_map map = xe_lrc_parallel_map(q->lrc[0]);
#define WQ_HEADER_SIZE	4	/* Includes 1 LRC address too */
	u32 wqi[XE_HW_ENGINE_MAX_INSTANCE + (WQ_HEADER_SIZE - 1)];
	u32 wqi_size = (q->width + (WQ_HEADER_SIZE - 1)) * sizeof(u32);
	u32 len_dw = (wqi_size / sizeof(u32)) - 1;
	int i = 0, j;

	if (wqi_size > wq_space_until_wrap(q)) {
		if (wq_noop_append(q))
			return;
	}
	if (wq_wait_for_space(q, wqi_size))
		return;

	wqi[i++] = FIELD_PREP(WQ_TYPE_MASK, WQ_TYPE_MULTI_LRC) |
		FIELD_PREP(WQ_LEN_MASK, len_dw);
	wqi[i++] = xe_lrc_descriptor(q->lrc[0]);
	wqi[i++] = FIELD_PREP(WQ_GUC_ID_MASK, q->guc->id) |
		FIELD_PREP(WQ_RING_TAIL_MASK, q->lrc[0]->ring.tail / sizeof(u64));
	wqi[i++] = 0;
	for (j = 1; j < q->width; ++j) {
		struct xe_lrc *lrc = q->lrc[j];

		wqi[i++] = lrc->ring.tail / sizeof(u64);
	}

	xe_gt_assert(guc_to_gt(guc), i == wqi_size / sizeof(u32));

	iosys_map_incr(&map, offsetof(struct guc_submit_parallel_scratch,
				      wq[q->guc->wqi_tail / sizeof(u32)]));
	xe_map_memcpy_to(xe, &map, 0, wqi, wqi_size);
	q->guc->wqi_tail += wqi_size;
	xe_gt_assert(guc_to_gt(guc), q->guc->wqi_tail <= WQ_SIZE);

	xe_device_wmb(xe);

	map = xe_lrc_parallel_map(q->lrc[0]);
	parallel_write(xe, map, wq_desc.tail, q->guc->wqi_tail);
}

#define RESUME_PENDING	~0x0ull
static void submit_exec_queue(struct xe_exec_queue *q, struct xe_sched_job *job)
{
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_lrc *lrc = q->lrc[0];
	u32 action[3];
	u32 g2h_len = 0;
	u32 num_g2h = 0;
	int len = 0;
	bool extra_submit = false;

	xe_gt_assert(guc_to_gt(guc), exec_queue_registered(q));

	if (!job->restore_replay || job->last_replay) {
		if (xe_exec_queue_is_parallel(q))
			wq_item_append(q);
		else
			xe_lrc_set_ring_tail(lrc, lrc->ring.tail);
		job->last_replay = false;
	}

	if (exec_queue_suspended(q) && !xe_exec_queue_is_parallel(q))
		return;

	/*
	 * All queues in a multi-queue group will use the primary queue
	 * of the group to interface with GuC. If primay is suspended,
	 * just return. Jobs will get scheduled once primary is resumed.
	 */
	q = xe_exec_queue_multi_queue_primary(q);
	if (exec_queue_suspended(q))
		return;

	if (!exec_queue_enabled(q)) {
		action[len++] = XE_GUC_ACTION_SCHED_CONTEXT_MODE_SET;
		action[len++] = q->guc->id;
		action[len++] = GUC_CONTEXT_ENABLE;
		g2h_len = G2H_LEN_DW_SCHED_CONTEXT_MODE_SET;
		num_g2h = 1;
		if (xe_exec_queue_is_parallel(q))
			extra_submit = true;

		q->guc->resume_time = RESUME_PENDING;
		set_exec_queue_pending_enable(q);
		set_exec_queue_enabled(q);
		trace_xe_exec_queue_scheduling_enable(q);
	} else {
		action[len++] = XE_GUC_ACTION_SCHED_CONTEXT;
		action[len++] = q->guc->id;
		trace_xe_exec_queue_submit(q);
	}

	xe_guc_ct_send(&guc->ct, action, len, g2h_len, num_g2h);

	if (extra_submit) {
		len = 0;
		action[len++] = XE_GUC_ACTION_SCHED_CONTEXT;
		action[len++] = q->guc->id;
		trace_xe_exec_queue_submit(q);

		xe_guc_ct_send(&guc->ct, action, len, 0, 0);
	}
}

static struct dma_fence *
guc_exec_queue_run_job(struct drm_sched_job *drm_job)
{
	struct xe_sched_job *job = to_xe_sched_job(drm_job);
	struct xe_exec_queue *q = job->q;
	struct xe_guc *guc = exec_queue_to_guc(q);
	bool killed_or_banned_or_wedged =
		exec_queue_killed_or_banned_or_wedged(q);

	xe_gt_assert(guc_to_gt(guc), !(exec_queue_destroyed(q) || exec_queue_pending_disable(q)) ||
		     exec_queue_banned(q) || exec_queue_suspended(q));

	trace_xe_sched_job_run(job);

	if (!killed_or_banned_or_wedged && !xe_sched_job_is_error(job)) {
		if (xe_exec_queue_is_multi_queue_secondary(q)) {
			struct xe_exec_queue *primary = xe_exec_queue_multi_queue_primary(q);

			if (exec_queue_killed_or_banned_or_wedged(primary))
				goto run_job_out;

			if (!exec_queue_registered(primary))
				register_exec_queue(primary, GUC_CONTEXT_NORMAL);
		}

		if (!exec_queue_registered(q))
			register_exec_queue(q, GUC_CONTEXT_NORMAL);
		if (!job->restore_replay)
			q->ring_ops->emit_job(job);
		submit_exec_queue(q, job);
		job->restore_replay = false;
	}

run_job_out:

	return job->fence;
}

static void guc_exec_queue_free_job(struct drm_sched_job *drm_job)
{
	struct xe_sched_job *job = to_xe_sched_job(drm_job);

	trace_xe_sched_job_free(job);
	xe_sched_job_put(job);
}

int xe_guc_read_stopped(struct xe_guc *guc)
{
	return atomic_read(&guc->submission_state.stopped);
}

static void handle_multi_queue_secondary_sched_done(struct xe_guc *guc,
						    struct xe_exec_queue *q,
						    u32 runnable_state);
static void handle_deregister_done(struct xe_guc *guc, struct xe_exec_queue *q);

#define MAKE_SCHED_CONTEXT_ACTION(q, enable_disable)			\
	u32 action[] = {						\
		XE_GUC_ACTION_SCHED_CONTEXT_MODE_SET,			\
		q->guc->id,						\
		GUC_CONTEXT_##enable_disable,				\
	}

static void disable_scheduling_deregister(struct xe_guc *guc,
					  struct xe_exec_queue *q)
{
	MAKE_SCHED_CONTEXT_ACTION(q, DISABLE);
	int ret;

	if (!xe_exec_queue_is_multi_queue_secondary(q))
		set_min_preemption_timeout(guc, q);

	smp_rmb();
	ret = wait_event_timeout(guc->ct.wq,
				 (!exec_queue_pending_enable(q) &&
				  !exec_queue_pending_disable(q)) ||
					 xe_guc_read_stopped(guc) ||
					 vf_recovery(guc),
				 HZ * 5);
	if (!ret && !vf_recovery(guc)) {
		struct xe_gpu_scheduler *sched = &q->guc->sched;

		xe_gt_warn(q->gt, "Pending enable/disable failed to respond\n");
		xe_sched_submission_start(sched);
		xe_gt_reset_async(q->gt);
		xe_sched_tdr_queue_imm(sched);
		return;
	}

	clear_exec_queue_enabled(q);
	set_exec_queue_pending_disable(q);
	set_exec_queue_destroyed(q);
	trace_xe_exec_queue_scheduling_disable(q);

	/*
	 * Reserve space for both G2H here as the 2nd G2H is sent from a G2H
	 * handler and we are not allowed to reserved G2H space in handlers.
	 */
	if (xe_exec_queue_is_multi_queue_secondary(q))
		handle_multi_queue_secondary_sched_done(guc, q, 0);
	else
		xe_guc_ct_send(&guc->ct, action, ARRAY_SIZE(action),
			       G2H_LEN_DW_SCHED_CONTEXT_MODE_SET +
			       G2H_LEN_DW_DEREGISTER_CONTEXT, 2);
}

/**
 * xe_guc_submit_wedge() - Wedge GuC submission
 * @guc: the GuC object
 *
 * Save exec queue's registered with GuC state by taking a ref to each queue.
 * Register a DRMM handler to drop refs upon driver unload.
 */
void xe_guc_submit_wedge(struct xe_guc *guc)
{
	struct xe_device *xe = guc_to_xe(guc);
	struct xe_exec_queue *q;
	unsigned long index;

	xe_gt_assert(guc_to_gt(guc), guc_to_xe(guc)->wedged.mode);

	/*
	 * If device is being wedged even before submission_state is
	 * initialized, there's nothing to do here.
	 */
	if (!guc->submission_state.initialized)
		return;

	if (xe->wedged.mode == XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET) {
		mutex_lock(&guc->submission_state.lock);
		xa_for_each(&guc->submission_state.exec_queue_lookup, index, q)
			if (xe_exec_queue_get_unless_zero(q))
				set_exec_queue_wedged(q);
		mutex_unlock(&guc->submission_state.lock);
	} else {
		/* Forcefully kill any remaining exec queues, signal fences */
		guc_submit_reset_prepare(guc);
		xe_guc_submit_stop(guc);
		xe_guc_softreset(guc);
		xe_uc_fw_sanitize(&guc->fw);
		xe_guc_submit_pause_abort(guc);
	}
}

static bool guc_submit_hint_wedged(struct xe_guc *guc)
{
	struct xe_device *xe = guc_to_xe(guc);

	if (xe->wedged.mode != XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET)
		return false;

	if (xe_device_wedged(xe))
		return true;

	xe_device_declare_wedged(xe);

	return true;
}

#define ADJUST_FIVE_PERCENT(__t)	mul_u64_u32_div(__t, 105, 100)

static bool check_timeout(struct xe_exec_queue *q, struct xe_sched_job *job)
{
	struct xe_gt *gt = guc_to_gt(exec_queue_to_guc(q));
	u32 ctx_timestamp, ctx_job_timestamp;
	u32 timeout_ms = q->sched_props.job_timeout_ms;
	u32 diff;
	u64 running_time_ms;

	if (!xe_sched_job_started(job)) {
		xe_gt_warn(gt, "Check job timeout: seqno=%u, lrc_seqno=%u, guc_id=%d, not started",
			   xe_sched_job_seqno(job), xe_sched_job_lrc_seqno(job),
			   q->guc->id);

		/* GuC never scheduled this job - let the caller trigger a GT reset. */
		return true;
	}

	ctx_timestamp = lower_32_bits(xe_lrc_timestamp(q->lrc[0]));
	if (ctx_timestamp == job->sample_timestamp) {
		if (IS_SRIOV_VF(gt_to_xe(gt)))
			xe_gt_notice(gt, "Check job timeout: seqno=%u, lrc_seqno=%u, guc_id=%d, timestamp stuck",
				     xe_sched_job_seqno(job),
				     xe_sched_job_lrc_seqno(job), q->guc->id);
		else
			xe_gt_warn(gt, "Check job timeout: seqno=%u, lrc_seqno=%u, guc_id=%d, timestamp stuck",
				   xe_sched_job_seqno(job),
				   xe_sched_job_lrc_seqno(job), q->guc->id);

		return xe_sched_invalidate_job(job, 0);
	}

	job->sample_timestamp = ctx_timestamp;
	ctx_job_timestamp = xe_lrc_ctx_job_timestamp(q->lrc[0]);

	/*
	 * Counter wraps at ~223s at the usual 19.2MHz, be paranoid catch
	 * possible overflows with a high timeout.
	 */
	xe_gt_assert(gt, timeout_ms < 100 * MSEC_PER_SEC);

	diff = ctx_timestamp - ctx_job_timestamp;

	/*
	 * Ensure timeout is within 5% to account for an GuC scheduling latency
	 */
	running_time_ms =
		ADJUST_FIVE_PERCENT(xe_gt_clock_interval_to_ms(gt, diff));

	xe_gt_dbg(gt,
		  "Check job timeout: seqno=%u, lrc_seqno=%u, guc_id=%d, running_time_ms=%llu, timeout_ms=%u, diff=0x%08x",
		  xe_sched_job_seqno(job), xe_sched_job_lrc_seqno(job),
		  q->guc->id, running_time_ms, timeout_ms, diff);

	return running_time_ms >= timeout_ms;
}

static void enable_scheduling(struct xe_exec_queue *q)
{
	MAKE_SCHED_CONTEXT_ACTION(q, ENABLE);
	struct xe_guc *guc = exec_queue_to_guc(q);
	int ret;

	xe_gt_assert(guc_to_gt(guc), !exec_queue_destroyed(q));
	xe_gt_assert(guc_to_gt(guc), exec_queue_registered(q));
	xe_gt_assert(guc_to_gt(guc), !exec_queue_pending_disable(q));
	xe_gt_assert(guc_to_gt(guc), !exec_queue_pending_enable(q));

	set_exec_queue_pending_enable(q);
	set_exec_queue_enabled(q);
	trace_xe_exec_queue_scheduling_enable(q);

	if (xe_exec_queue_is_multi_queue_secondary(q))
		handle_multi_queue_secondary_sched_done(guc, q, 1);
	else
		xe_guc_ct_send(&guc->ct, action, ARRAY_SIZE(action),
			       G2H_LEN_DW_SCHED_CONTEXT_MODE_SET, 1);

	ret = wait_event_timeout(guc->ct.wq,
				 !exec_queue_pending_enable(q) ||
				 xe_guc_read_stopped(guc) ||
				 vf_recovery(guc), HZ * 5);
	if ((!ret && !vf_recovery(guc)) || xe_guc_read_stopped(guc)) {
		xe_gt_warn(guc_to_gt(guc), "Schedule enable failed to respond");
		set_exec_queue_banned(q);
		xe_gt_reset_async(q->gt);
		xe_sched_tdr_queue_imm(&q->guc->sched);
	}
}

static void disable_scheduling(struct xe_exec_queue *q, bool immediate)
{
	MAKE_SCHED_CONTEXT_ACTION(q, DISABLE);
	struct xe_guc *guc = exec_queue_to_guc(q);

	xe_gt_assert(guc_to_gt(guc), !exec_queue_destroyed(q));
	xe_gt_assert(guc_to_gt(guc), exec_queue_registered(q));
	xe_gt_assert(guc_to_gt(guc), !exec_queue_pending_disable(q));

	if (immediate && !xe_exec_queue_is_multi_queue_secondary(q))
		set_min_preemption_timeout(guc, q);
	clear_exec_queue_enabled(q);
	set_exec_queue_pending_disable(q);
	trace_xe_exec_queue_scheduling_disable(q);

	if (xe_exec_queue_is_multi_queue_secondary(q))
		handle_multi_queue_secondary_sched_done(guc, q, 0);
	else
		xe_guc_ct_send(&guc->ct, action, ARRAY_SIZE(action),
			       G2H_LEN_DW_SCHED_CONTEXT_MODE_SET, 1);
}

/*
 * Recover via GT reset for a kernel queue, or for a GuC scheduling failure (job
 * never started) on a queue that was not already killed or banned. An already
 * banned queue must stay banned, so its unstarted jobs do not clear the ban or
 * trigger a reset.
 */
static bool timeout_needs_gt_reset(struct xe_exec_queue *q, struct xe_sched_job *job,
				   bool skip_timeout_check)
{
	if (q->flags & EXEC_QUEUE_FLAG_KERNEL)
		return true;

	return !skip_timeout_check && !xe_sched_job_started(job);
}

static enum drm_gpu_sched_stat
guc_exec_queue_timedout_job(struct drm_sched_job *drm_job)
{
	struct xe_sched_job *job = to_xe_sched_job(drm_job);
	struct drm_sched_job *tmp_job;
	struct xe_exec_queue *q = job->q, *primary;
	struct xe_gpu_scheduler *sched = &q->guc->sched;
	struct xe_guc *guc = exec_queue_to_guc(q);
	const char *process_name = "no process";
	struct xe_device *xe = guc_to_xe(guc);
	int err = -ETIME;
	pid_t pid = -1;
	bool wedged = false, wedge_device = false, skip_timeout_check;

	xe_gt_assert(guc_to_gt(guc), !exec_queue_destroyed(q));

	primary = xe_exec_queue_multi_queue_primary(q);

	/*
	 * TDR has fired before free job worker. Common if exec queue
	 * immediately closed after last fence signaled. Add back to pending
	 * list so job can be freed and kick scheduler ensuring free job is not
	 * lost.
	 */
	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &job->fence->flags) ||
	    vf_recovery(guc))
		return DRM_GPU_SCHED_STAT_NO_HANG;

	/* Kill the run_job entry point */
	if (xe_exec_queue_is_multi_queue(q))
		xe_guc_exec_queue_group_stop(q);
	else
		xe_sched_submission_stop(sched);

	/* Must check all state after stopping scheduler */
	skip_timeout_check = exec_queue_reset(q) ||
		exec_queue_killed_or_banned_or_wedged(q);

	/* Skip timeout check if multi-queue group is banned */
	if (xe_exec_queue_is_multi_queue(q) &&
	    READ_ONCE(q->multi_queue.group->banned))
		skip_timeout_check = true;

	/* LR jobs can only get here if queue has been killed or hit an error */
	if (xe_exec_queue_is_lr(q))
		xe_gt_assert(guc_to_gt(guc), skip_timeout_check);

	/*
	 * If devcoredump not captured and GuC capture for the job is not ready
	 * do manual capture first and decide later if we need to use it
	 */
	if (!xe_device_is_in_reset(xe) && !exec_queue_killed(q) && !xe->devcoredump.captured &&
	    !xe_guc_capture_get_matching_and_lock(q)) {
		/* take force wake before engine register manual capture */
		CLASS(xe_force_wake, fw_ref)(gt_to_fw(q->gt), XE_FORCEWAKE_ALL);
		if (!xe_force_wake_ref_has_domain(fw_ref.domains, XE_FORCEWAKE_ALL))
			xe_gt_info(q->gt, "failed to get forcewake for coredump capture\n");

		xe_engine_snapshot_capture_for_queue(q);
	}

	/*
	 * Check if job is actually timed out, if so restart job execution and TDR
	 */
	if (!skip_timeout_check && !check_timeout(q, job))
		goto rearm;

	if (!exec_queue_killed(q))
		wedged = guc_submit_hint_wedged(exec_queue_to_guc(q));

	set_exec_queue_banned(q);

	/* Kick job / queue off hardware */
	if (!xe_device_is_in_reset(xe) && !wedged &&
	    (exec_queue_enabled(primary) || exec_queue_pending_disable(primary))) {
		int ret;

		if (exec_queue_reset(primary))
			err = -EIO;

		if (xe_uc_fw_is_running(&guc->fw)) {
			/*
			 * Wait for any pending G2H to flush out before
			 * modifying state
			 */
			ret = wait_event_timeout(guc->ct.wq,
						 (!exec_queue_pending_enable(primary) &&
						  !exec_queue_pending_disable(primary)) ||
						 xe_guc_read_stopped(guc) ||
						 vf_recovery(guc), HZ * 5);
			if (vf_recovery(guc))
				goto handle_vf_resume;
			if (!ret || xe_guc_read_stopped(guc))
				goto trigger_reset;

			disable_scheduling(primary, skip_timeout_check);
		}

		/*
		 * Must wait for scheduling to be disabled before signalling
		 * any fences, if GT broken the GT reset code should signal us.
		 *
		 * FIXME: Tests can generate a ton of 0x6000 (IOMMU CAT fault
		 * error) messages which can cause the schedule disable to get
		 * lost. If this occurs, trigger a GT reset to recover.
		 */
		smp_rmb();
		ret = wait_event_timeout(guc->ct.wq,
					 !xe_uc_fw_is_running(&guc->fw) ||
					 !exec_queue_pending_disable(primary) ||
					 xe_guc_read_stopped(guc) ||
					 vf_recovery(guc), HZ * 5);
		if (vf_recovery(guc))
			goto handle_vf_resume;
		if (!ret || xe_guc_read_stopped(guc)) {
trigger_reset:
			if (!ret)
				xe_gt_warn(guc_to_gt(guc),
					   "Schedule disable failed to respond, guc_id=%d",
					   primary->guc->id);
			xe_devcoredump(primary, job,
				       "Schedule disable failed to respond, guc_id=%d, ret=%d, guc_read=%d",
				       primary->guc->id, ret, xe_guc_read_stopped(guc));
			xe_gt_reset_async(primary->gt);
			xe_sched_tdr_queue_imm(sched);
			goto rearm;
		}
	}

	if (q->vm && q->vm->xef) {
		process_name = q->vm->xef->process_name;
		pid = q->vm->xef->pid;
	}

	if (!exec_queue_killed(q))
		xe_gt_notice(guc_to_gt(guc),
			     "Timedout job: seqno=%u, lrc_seqno=%u, guc_id=%d, flags=0x%lx in %s [%d]",
			     xe_sched_job_seqno(job), xe_sched_job_lrc_seqno(job),
			     q->guc->id, q->flags, process_name, pid);

	trace_xe_sched_job_timedout(job);

	/* Do not access device if in reset */
	if (!xe_device_is_in_reset(xe) && !exec_queue_killed(q))
		xe_devcoredump(q, job,
			       "Timedout job - seqno=%u, lrc_seqno=%u, guc_id=%d, flags=0x%lx",
			       xe_sched_job_seqno(job), xe_sched_job_lrc_seqno(job),
			       q->guc->id, q->flags);

	if (!wedged) {
		if (timeout_needs_gt_reset(q, job, skip_timeout_check)) {
			if (!xe_sched_invalidate_job(job, 2)) {
				clear_exec_queue_banned(q);
				xe_gt_reset_async(q->gt);
				goto rearm;
			}
			if (q->flags & EXEC_QUEUE_FLAG_KERNEL) {
				xe_gt_WARN(q->gt, true, "Kernel-submitted job timed out\n");
				wedge_device = true;
			}
		} else if (q->flags & EXEC_QUEUE_FLAG_VM && !exec_queue_killed(q)) {
			xe_gt_WARN(q->gt, true, "VM job timed out on non-killed execqueue\n");
		}
	}

	/* Mark all outstanding jobs as bad, thus completing them */
	xe_sched_job_set_error(job, err);
	drm_sched_for_each_pending_job(tmp_job, &sched->base, NULL)
		xe_sched_job_set_error(to_xe_sched_job(tmp_job), -ECANCELED);

	if (xe_exec_queue_is_multi_queue(q)) {
		xe_guc_exec_queue_group_start(q);
		xe_guc_exec_queue_group_trigger_cleanup(q);
	} else {
		xe_sched_submission_start(sched);
		xe_guc_exec_queue_trigger_cleanup(q);
	}

	if (wedge_device)
		xe_device_declare_wedged(gt_to_xe(q->gt));

	/*
	 * We want the job added back to the pending list so it gets freed; this
	 * is what DRM_GPU_SCHED_STAT_NO_HANG does.
	 */
	return DRM_GPU_SCHED_STAT_NO_HANG;

rearm:
	/*
	 * XXX: Ideally want to adjust timeout based on current execution time
	 * but there is not currently an easy way to do in DRM scheduler. With
	 * some thought, do this in a follow up.
	 */
	if (xe_exec_queue_is_multi_queue(q))
		xe_guc_exec_queue_group_start(q);
	else
		xe_sched_submission_start(sched);
handle_vf_resume:
	return DRM_GPU_SCHED_STAT_NO_HANG;
}

static void guc_exec_queue_multi_queue_drop_suspend(struct xe_exec_queue *q);
static int guc_exec_queue_suspend_wait_blocking(struct xe_exec_queue *q);

static void guc_exec_queue_fini(struct xe_exec_queue *q)
{
	struct xe_guc_exec_queue *ge = q->guc;
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct drm_device *drm = &guc_to_xe(guc)->drm;

	/*
	 * A secondary can leave the group while still preempt suspended (e.g.
	 * xe_vm_remove_compute_exec_queue() forces its preempt fence to signal,
	 * which suspends it). It holds one forwarded suspend reference on the
	 * primary, so drop it and resume the primary if it was the last member
	 * that had it suspended. Primaries forward to nobody, so they don't need
	 * this.
	 *
	 * First make sure the primary's forwarded suspend has completed. If the
	 * secondary was killed/reset before its preempt fence worker ran, that
	 * worker skips suspend_wait() (see preempt_fence_work_func()), leaving
	 * the primary's suspend possibly in flight. drop_suspend() runs under a
	 * spinlock and cannot wait, so drain it here with the uninterruptible
	 * blocking wait; otherwise resuming the primary in drop_suspend() could
	 * trip the !suspend_pending assert.
	 */
	if (xe_exec_queue_is_multi_queue_secondary(q)) {
		if (READ_ONCE(q->guc->suspend_count))
			guc_exec_queue_suspend_wait_blocking(q);
		guc_exec_queue_multi_queue_drop_suspend(q);
	}

	if (xe_exec_queue_is_multi_queue_secondary(q)) {
		struct xe_exec_queue_group *group = q->multi_queue.group;

		mutex_lock(&group->list_lock);
		list_del(&q->multi_queue.link);
		mutex_unlock(&group->list_lock);
	}

	release_guc_id(guc, q);
	xe_sched_entity_fini(&ge->entity);
	xe_sched_fini(&ge->sched);

	/*
	 * RCU free due sched being exported via DRM scheduler fences
	 * (timeline name).
	 */
	kfree_rcu(ge, rcu);

	drm_dev_put(drm);
}

static void guc_exec_queue_do_destroy(struct xe_exec_queue *q)
{
	struct xe_guc_exec_queue *ge = q->guc;
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_device *xe = guc_to_xe(guc);
	struct drm_device *drm = &xe->drm;

	/*
	 * guc_exec_queue_fini() drops the queue's drm_device ref.
	 * Keep the device alive until the PM-runtime guard unwinds.
	 */
	drm_dev_get(drm);

	scoped_guard(xe_pm_runtime, xe) {
		trace_xe_exec_queue_destroy(q);

		/* Confirm no work left behind accessing device structures */
		cancel_delayed_work_sync(&ge->sched.base.work_tdr);

		xe_exec_queue_fini(q);
	}

	drm_dev_put(drm);
}

static void __guc_exec_queue_destroy_async(struct work_struct *w)
{
	struct xe_guc_exec_queue *ge =
		container_of(w, struct xe_guc_exec_queue, destroy_async);

	guc_exec_queue_do_destroy(ge->q);
}

static void guc_exec_queue_destroy_async(struct xe_exec_queue *q)
{
	INIT_WORK(&q->guc->destroy_async, __guc_exec_queue_destroy_async);

	/* We must block on kernel engines so slabs are empty on driver unload */
	if (q->flags & EXEC_QUEUE_FLAG_PERMANENT || exec_queue_wedged(q))
		guc_exec_queue_do_destroy(q);
	else
		xe_destroy_wq_queue(&q->guc->destroy_async);
}

static void __guc_exec_queue_destroy(struct xe_guc *guc, struct xe_exec_queue *q)
{
	/*
	 * Might be done from within the GPU scheduler, need to do async as we
	 * fini the scheduler when the engine is fini'd, the scheduler can't
	 * complete fini within itself (circular dependency). Async resolves
	 * this we and don't really care when everything is fini'd, just that it
	 * is.
	 */
	guc_exec_queue_destroy_async(q);
}

static void __guc_exec_queue_process_msg_cleanup(struct xe_sched_msg *msg)
{
	struct xe_exec_queue *q = msg->private_data;
	struct xe_guc *guc = exec_queue_to_guc(q);

	xe_gt_assert(guc_to_gt(guc), !(q->flags & EXEC_QUEUE_FLAG_PERMANENT));
	trace_xe_exec_queue_cleanup_entity(q);

	/*
	 * Expected state transitions for cleanup:
	 * - If the exec queue is registered and GuC firmware is running, we must first
	 *   disable scheduling and deregister the queue to ensure proper teardown and
	 *   resource release in the GuC, then destroy the exec queue on driver side.
	 * - If the GuC is already stopped (e.g., during driver unload or GPU reset),
	 *   we cannot expect a response for the deregister request. In this case,
	 *   it is safe to directly destroy the exec queue on driver side, as the GuC
	 *   will not process further requests and all resources must be cleaned up locally.
	 */
	if (exec_queue_registered(q) && xe_uc_fw_is_running(&guc->fw))
		disable_scheduling_deregister(guc, q);
	else
		__guc_exec_queue_destroy(guc, q);
}

static bool guc_exec_queue_allowed_to_change_state(struct xe_exec_queue *q)
{
	return !exec_queue_killed_or_banned_or_wedged(q) && exec_queue_registered(q);
}

static void __guc_exec_queue_process_msg_set_sched_props(struct xe_sched_msg *msg)
{
	struct xe_exec_queue *q = msg->private_data;
	struct xe_guc *guc = exec_queue_to_guc(q);

	if (guc_exec_queue_allowed_to_change_state(q))
		init_policies(guc, q);
	kfree(msg);
}

static void __suspend_fence_signal(struct xe_exec_queue *q)
{
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_device *xe = guc_to_xe(guc);

	if (!q->guc->suspend_pending)
		return;

	WRITE_ONCE(q->guc->suspend_pending, false);

	/*
	 * We use a GuC shared wait queue for VFs because the VF resfix start
	 * interrupt must be able to wake all instances of suspend_wait. This
	 * prevents the VF migration worker from being starved during
	 * scheduling.
	 */
	if (IS_SRIOV_VF(xe))
		wake_up_all(&guc->ct.wq);
	else
		wake_up(&q->guc->suspend_wait);
}

static void suspend_fence_signal(struct xe_exec_queue *q)
{
	struct xe_guc *guc = exec_queue_to_guc(q);

	xe_gt_assert(guc_to_gt(guc), exec_queue_suspended(q) || exec_queue_killed(q) ||
		     xe_guc_read_stopped(guc));
	xe_gt_assert(guc_to_gt(guc), q->guc->suspend_pending);

	__suspend_fence_signal(q);
}

static void __guc_exec_queue_process_msg_suspend(struct xe_sched_msg *msg)
{
	struct xe_exec_queue *q = msg->private_data;
	struct xe_guc *guc = exec_queue_to_guc(q);

	if (guc_exec_queue_allowed_to_change_state(q) && !exec_queue_suspended(q) &&
	    exec_queue_enabled(q)) {
		wait_event(guc->ct.wq, vf_recovery(guc) ||
			   ((q->guc->resume_time != RESUME_PENDING ||
			   xe_guc_read_stopped(guc)) && !exec_queue_pending_disable(q)));

		if (!xe_guc_read_stopped(guc)) {
			s64 since_resume_ms =
				ktime_ms_delta(ktime_get(),
					       q->guc->resume_time);
			s64 wait_ms = q->vm->preempt.min_run_period_ms -
				since_resume_ms;

			if (wait_ms > 0 && q->guc->resume_time)
				xe_sleep_relaxed_ms(wait_ms);

			set_exec_queue_suspended(q);
			disable_scheduling(q, false);
		}
	} else if (q->guc->suspend_pending) {
		set_exec_queue_suspended(q);
		suspend_fence_signal(q);
	}
}

static void __guc_exec_queue_process_msg_resume(struct xe_sched_msg *msg)
{
	struct xe_exec_queue *q = msg->private_data;

	if (guc_exec_queue_allowed_to_change_state(q)) {
		clear_exec_queue_suspended(q);
		if (!exec_queue_enabled(q)) {
			q->guc->resume_time = RESUME_PENDING;
			set_exec_queue_pending_resume(q);
			enable_scheduling(q);
		}
	} else {
		clear_exec_queue_suspended(q);
	}
}

static void __guc_exec_queue_process_msg_set_multi_queue_priority(struct xe_sched_msg *msg)
{
	struct xe_exec_queue *q = msg->private_data;

	if (guc_exec_queue_allowed_to_change_state(q))
		guc_exec_queue_send_cgp_sync(q);

	kfree(msg);
}

#define CLEANUP				1	/* Non-zero values to catch uninitialized msg */
#define SET_SCHED_PROPS			2
#define SUSPEND				3
#define RESUME				4
#define SET_MULTI_QUEUE_PRIORITY	5
#define OPCODE_MASK	0xf
#define MSG_LOCKED	BIT(8)
#define MSG_HEAD	BIT(9)

static void guc_exec_queue_process_msg(struct xe_sched_msg *msg)
{
	struct xe_device *xe = guc_to_xe(exec_queue_to_guc(msg->private_data));

	trace_xe_sched_msg_recv(msg);

	switch (msg->opcode) {
	case CLEANUP:
		__guc_exec_queue_process_msg_cleanup(msg);
		break;
	case SET_SCHED_PROPS:
		__guc_exec_queue_process_msg_set_sched_props(msg);
		break;
	case SUSPEND:
		__guc_exec_queue_process_msg_suspend(msg);
		break;
	case RESUME:
		__guc_exec_queue_process_msg_resume(msg);
		break;
	case SET_MULTI_QUEUE_PRIORITY:
		__guc_exec_queue_process_msg_set_multi_queue_priority(msg);
		break;
	default:
		XE_WARN_ON("Unknown message type");
	}

	xe_pm_runtime_put(xe);
}

static const struct drm_sched_backend_ops drm_sched_ops = {
	.run_job = guc_exec_queue_run_job,
	.free_job = guc_exec_queue_free_job,
	.timedout_job = guc_exec_queue_timedout_job,
};

static const struct xe_sched_backend_ops xe_sched_ops = {
	.process_msg = guc_exec_queue_process_msg,
};

static int guc_exec_queue_init(struct xe_exec_queue *q)
{
	struct xe_gpu_scheduler *sched;
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct drm_device *drm = &guc_to_xe(guc)->drm;
	struct workqueue_struct *submit_wq = NULL;
	struct xe_guc_exec_queue *ge;
	long timeout;
	int err, i;

	xe_gt_assert(guc_to_gt(guc), xe_device_uc_enabled(guc_to_xe(guc)));

	ge = kzalloc_obj(*ge);
	if (!ge)
		return -ENOMEM;

	drm_dev_get(drm);

	q->guc = ge;
	ge->q = q;
	init_rcu_head(&ge->rcu);
	init_waitqueue_head(&ge->suspend_wait);

	for (i = 0; i < MAX_STATIC_MSG_TYPE; ++i)
		INIT_LIST_HEAD(&ge->static_msgs[i].link);

	timeout = (q->vm && xe_vm_in_lr_mode(q->vm)) ? MAX_SCHEDULE_TIMEOUT :
		  msecs_to_jiffies(q->sched_props.job_timeout_ms);

	err = alloc_guc_id(guc, q);
	if (err)
		goto err_free;

	xe_exec_queue_assign_name(q, q->guc->id);

	strscpy(ge->name, q->name, sizeof(ge->name));

	/*
	 * Use primary queue's submit_wq for all secondary queues of a
	 * multi queue group. This serialization avoids any locking around
	 * CGP synchronization with GuC.
	 */
	if (xe_exec_queue_is_multi_queue_secondary(q)) {
		struct xe_exec_queue *primary = xe_exec_queue_multi_queue_primary(q);

		submit_wq = primary->guc->sched.base.submit_wq;
	}

	err = xe_sched_init(&ge->sched, &drm_sched_ops, &xe_sched_ops,
			    submit_wq, xe_lrc_ring_size() / MAX_JOB_SIZE_BYTES, 64,
			    timeout, guc_to_gt(guc)->ordered_wq, NULL,
			    ge->name, gt_to_xe(q->gt)->drm.dev);
	if (err)
		goto err_release_id;

	sched = &ge->sched;
	err = xe_sched_entity_init(&ge->entity, sched);
	if (err)
		goto err_sched;

	q->entity = &ge->entity;

	mutex_lock(&guc->submission_state.lock);
	if (xe_guc_read_stopped(guc) || vf_recovery(guc))
		xe_sched_stop(sched);
	publish_guc_id(guc, q);
	mutex_unlock(&guc->submission_state.lock);

	/*
	 * Maintain secondary queues of the multi queue group in a list
	 * for handling dependencies across the queues in the group.
	 */
	if (xe_exec_queue_is_multi_queue_secondary(q)) {
		struct xe_exec_queue_group *group = q->multi_queue.group;

		INIT_LIST_HEAD(&q->multi_queue.link);
		mutex_lock(&group->list_lock);
		if (group->stopped)
			WRITE_ONCE(q->guc->sched.base.pause_submit, true);
		list_add_tail(&q->multi_queue.link, &group->list);
		mutex_unlock(&group->list_lock);
	}

	if (xe_exec_queue_is_multi_queue(q))
		trace_xe_exec_queue_create_multi_queue(q);
	else
		trace_xe_exec_queue_create(q);

	return 0;

err_sched:
	xe_sched_fini(&ge->sched);
err_release_id:
	release_guc_id(guc, q);
err_free:
	kfree(ge);
	drm_dev_put(drm);

	return err;
}

static void guc_exec_queue_kill(struct xe_exec_queue *q)
{
	trace_xe_exec_queue_kill(q);
	set_exec_queue_killed(q);
	__suspend_fence_signal(q);
	xe_guc_exec_queue_trigger_cleanup(q);
}

static void guc_exec_queue_add_msg(struct xe_exec_queue *q, struct xe_sched_msg *msg,
				   u32 opcode)
{
	xe_pm_runtime_get_noresume(guc_to_xe(exec_queue_to_guc(q)));

	INIT_LIST_HEAD(&msg->link);
	msg->opcode = opcode & OPCODE_MASK;
	msg->private_data = q;

	trace_xe_sched_msg_add(msg);
	if (opcode & MSG_HEAD)
		xe_sched_add_msg_head(&q->guc->sched, msg);
	else if (opcode & MSG_LOCKED)
		xe_sched_add_msg_locked(&q->guc->sched, msg);
	else
		xe_sched_add_msg(&q->guc->sched, msg);
}

static void guc_exec_queue_try_add_msg_head(struct xe_exec_queue *q,
					    struct xe_sched_msg *msg,
					    u32 opcode)
{
	if (!list_empty(&msg->link))
		return;

	guc_exec_queue_add_msg(q, msg, opcode | MSG_LOCKED | MSG_HEAD);
}

static bool guc_exec_queue_try_add_msg(struct xe_exec_queue *q,
				       struct xe_sched_msg *msg,
				       u32 opcode)
{
	if (!list_empty(&msg->link))
		return false;

	guc_exec_queue_add_msg(q, msg, opcode | MSG_LOCKED);

	return true;
}

#define STATIC_MSG_CLEANUP	0
#define STATIC_MSG_SUSPEND	1
#define STATIC_MSG_RESUME	2
static void guc_exec_queue_destroy(struct xe_exec_queue *q)
{
	struct xe_sched_msg *msg = q->guc->static_msgs + STATIC_MSG_CLEANUP;

	if (!(q->flags & EXEC_QUEUE_FLAG_PERMANENT) && !exec_queue_wedged(q))
		guc_exec_queue_add_msg(q, msg, CLEANUP);
	else
		__guc_exec_queue_destroy(exec_queue_to_guc(q), q);
}

static int guc_exec_queue_set_priority(struct xe_exec_queue *q,
				       enum xe_exec_queue_priority priority)
{
	struct xe_sched_msg *msg;

	if (q->sched_props.priority == priority ||
	    exec_queue_killed_or_banned_or_wedged(q))
		return 0;

	msg = kmalloc_obj(*msg);
	if (!msg)
		return -ENOMEM;

	q->sched_props.priority = priority;
	guc_exec_queue_add_msg(q, msg, SET_SCHED_PROPS);

	return 0;
}

static int guc_exec_queue_set_timeslice(struct xe_exec_queue *q, u32 timeslice_us)
{
	struct xe_sched_msg *msg;

	if (q->sched_props.timeslice_us == timeslice_us ||
	    exec_queue_killed_or_banned_or_wedged(q))
		return 0;

	msg = kmalloc_obj(*msg);
	if (!msg)
		return -ENOMEM;

	q->sched_props.timeslice_us = timeslice_us;
	guc_exec_queue_add_msg(q, msg, SET_SCHED_PROPS);

	return 0;
}

static int guc_exec_queue_set_preempt_timeout(struct xe_exec_queue *q,
					      u32 preempt_timeout_us)
{
	struct xe_sched_msg *msg;

	if (q->sched_props.preempt_timeout_us == preempt_timeout_us ||
	    exec_queue_killed_or_banned_or_wedged(q))
		return 0;

	msg = kmalloc_obj(*msg);
	if (!msg)
		return -ENOMEM;

	q->sched_props.preempt_timeout_us = preempt_timeout_us;
	guc_exec_queue_add_msg(q, msg, SET_SCHED_PROPS);

	return 0;
}

static int guc_exec_queue_set_multi_queue_priority(struct xe_exec_queue *q,
						   enum xe_multi_queue_priority priority)
{
	struct xe_sched_msg *msg;

	xe_gt_assert(guc_to_gt(exec_queue_to_guc(q)), xe_exec_queue_is_multi_queue(q));

	if (exec_queue_killed_or_banned_or_wedged(q))
		return 0;

	msg = kmalloc_obj(*msg);
	if (!msg)
		return -ENOMEM;

	scoped_guard(spinlock, &q->multi_queue.lock) {
		if (q->multi_queue.priority == priority) {
			kfree(msg);
			return 0;
		}

		q->multi_queue.priority = priority;
	}

	guc_exec_queue_add_msg(q, msg, SET_MULTI_QUEUE_PRIORITY);

	return 0;
}

/*
 * Core suspend: take a suspend reference on @q and, on the first reference,
 * disable its GuC context so the GPU is actually preempted. Caller must have
 * ensured @q is not killed/banned/wedged. Returns true if this was the first
 * suspend reference (the 0->1 transition).
 */
static bool __guc_exec_queue_suspend(struct xe_exec_queue *q)
{
	struct xe_guc_exec_queue *ge = q->guc;
	struct xe_gpu_scheduler *sched = &ge->sched;
	struct xe_sched_msg *msg = ge->static_msgs + STATIC_MSG_SUSPEND;
	bool first;

	xe_sched_msg_lock(sched);
	first = (++ge->suspend_count == 1);
	if (first) {
		bool added = guc_exec_queue_try_add_msg(q, msg, SUSPEND);

		/* slot must be free at 0->1 */
		xe_gt_assert(guc_to_gt(exec_queue_to_guc(q)), added);
		ge->suspend_pending = true;
	}
	xe_sched_msg_unlock(sched);

	return first;
}

/*
 * Core resume: drop a suspend reference on @q and, on the last reference,
 * re-enable its GuC context. Returns true if this dropped the last suspend
 * reference (the 1->0 transition).
 */
static bool __guc_exec_queue_resume(struct xe_exec_queue *q)
{
	struct xe_guc_exec_queue *ge = q->guc;
	struct xe_gpu_scheduler *sched = &ge->sched;
	struct xe_sched_msg *msg = ge->static_msgs + STATIC_MSG_RESUME;
	struct xe_guc *guc = exec_queue_to_guc(q);
	bool last;

	xe_sched_msg_lock(sched);
	xe_gt_assert(guc_to_gt(guc), !ge->suspend_pending);
	xe_gt_assert(guc_to_gt(guc), ge->suspend_count > 0);
	last = (--ge->suspend_count == 0);
	if (last) {
		bool added = guc_exec_queue_try_add_msg(q, msg, RESUME);

		/* slot must be free at 1->0 */
		xe_gt_assert(guc_to_gt(guc), added);
	}
	xe_sched_msg_unlock(sched);

	return last;
}

static int guc_exec_queue_suspend(struct xe_exec_queue *q)
{
	if (exec_queue_killed_or_banned_or_wedged(q))
		return -EINVAL;

	/*
	 * Non-multi-queue queues and multi-queue primaries suspend themselves
	 * directly: their own msg_lock makes the suspend_count 0->1 transition
	 * and the suspend_pending update atomic, so no group level serialization
	 * is needed.
	 */
	if (!xe_exec_queue_is_multi_queue_secondary(q)) {
		__guc_exec_queue_suspend(q);
		return 0;
	}

	/*
	 * A secondary's suspend is meaningless once the primary - which owns the
	 * group's GuC context - is gone, so fail it too. This keeps the
	 * secondary's effective state consistent with guc_exec_queue_reset_status(),
	 * which already reports the primary's killed/banned/wedged state for
	 * secondaries. A primary killed *after* this check is still handled at
	 * message-processing time, where the SUSPEND is a no-op for a killed
	 * context; this only covers an already-dead primary.
	 */
	if (exec_queue_killed_or_banned_or_wedged(xe_exec_queue_multi_queue_primary(q)))
		return -EINVAL;

	/*
	 * A secondary doesn't interface with GuC: suspend it like any other
	 * queue (its own suspend_count drives its internally handled scheduler
	 * state) and, only on its own 0->1 transition, forward the suspend to the
	 * primary so the GPU is actually preempted. Hold @suspend_lock so that
	 * observing the secondary's transition and forwarding it to the primary
	 * happen atomically; this keeps the primary's refcount paired with member
	 * transitions even if the same secondary is suspended and resumed
	 * concurrently across rebind cycles.
	 */
	scoped_guard(spinlock, &q->multi_queue.group->suspend_lock) {
		if (__guc_exec_queue_suspend(q))
			__guc_exec_queue_suspend(xe_exec_queue_multi_queue_primary(q));
	}

	return 0;
}

static void guc_exec_queue_suspend_timeout_ban(struct xe_exec_queue *q)
{
	struct xe_guc *guc = exec_queue_to_guc(q);

	xe_gt_warn(guc_to_gt(guc),
		   "Suspend fence, guc_id=%d, failed to respond, banning queue",
		   q->guc->id);
	/*
	 * The GuC failed to respond to the suspend within the timeout. This is
	 * not recoverable for this context, so ban it and tear it down via
	 * cleanup rather than leave it suspended forever. __suspend_fence_signal
	 * clears suspend_pending and wakes any waiter.
	 *
	 * @q is the primary here; it owns the group's GuC context, so a failure
	 * to suspend it wedges the whole group. Ban and tear down the entire
	 * group in the multi-queue case.
	 */
	if (xe_exec_queue_is_multi_queue(q)) {
		set_exec_queue_group_banned(q);
		__suspend_fence_signal(q);
		xe_guc_exec_queue_group_trigger_cleanup(q);
	} else {
		set_exec_queue_banned(q);
		__suspend_fence_signal(q);
		xe_guc_exec_queue_trigger_cleanup(q);
	}
}

/*
 * Wait for @q's own suspend to complete: suspend_pending cleared, or the queue
 * killed / GuC stopped. With @blocking, wait uninterruptibly and do not handle
 * VF recovery (for callers that must complete on behalf of a possibly
 * cross-process queue); otherwise wait interruptibly.
 *
 * Returns 0 on completion or -ETIME on timeout. Interruptible waits may also
 * return -EAGAIN (VF recovery in progress, retry) or -ERESTARTSYS (aborted by a
 * signal; suspend_pending may still be set, so callers must not resume()
 * without re-confirming the suspend).
 */
static int guc_exec_queue_wait_suspend_done(struct xe_exec_queue *q, bool blocking)
{
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_device *xe = guc_to_xe(guc);
	int ret;

	/*
	 * Likely don't need to check exec_queue_killed() as we clear
	 * suspend_pending upon kill but to be paranoid but races in which
	 * suspend_pending is set after kill also check kill here.
	 */
#define WAIT_COND \
	(!READ_ONCE(q->guc->suspend_pending) ||	exec_queue_killed(q) || \
	 xe_guc_read_stopped(guc))

retry:
	if (blocking) {
		if (IS_SRIOV_VF(xe))
			ret = wait_event_timeout(guc->ct.wq, WAIT_COND, HZ * 5);
		else
			ret = wait_event_timeout(q->guc->suspend_wait, WAIT_COND,
						 HZ * 5);
	} else if (IS_SRIOV_VF(xe)) {
		ret = wait_event_interruptible_timeout(guc->ct.wq, WAIT_COND ||
						       vf_recovery(guc), HZ * 5);
	} else {
		ret = wait_event_interruptible_timeout(q->guc->suspend_wait,
						       WAIT_COND, HZ * 5);
	}

	if (!blocking && vf_recovery(guc) && !xe_device_wedged(xe))
		return -EAGAIN;

	if (!ret)
		return -ETIME;
	else if (!blocking && IS_SRIOV_VF(xe) && !WAIT_COND)
		/* Corner case on RESFIX DONE where vf_recovery() changes */
		goto retry;

#undef WAIT_COND

	return ret < 0 ? ret : 0;
}

static int guc_exec_queue_suspend_wait_common(struct xe_exec_queue *q, bool blocking)
{
	int ret;

	/*
	 * A secondary's suspend rides the sched-message worker (short-circuited,
	 * no GuC round-trip) and so is not synchronous with
	 * guc_exec_queue_suspend(): its own suspend_pending may still be set
	 * here. Waiting on the primary alone is not sufficient - if the primary
	 * was already suspended, the forward is a refcount-only transition that
	 * queues no new primary SUSPEND and leaves the primary's suspend_pending
	 * clear, so the primary wait would return immediately while the
	 * secondary's suspend is still in flight, and a later resume() would trip
	 * the secondary's !suspend_pending assert. So first wait for the
	 * secondary's own suspend to complete, then wait on the primary.
	 *
	 * A timeout on either bans the queue (being multi-queue, that tears down
	 * the whole group). A secondary suspend has no real GuC round-trip, so
	 * its timeout is a software scheduler stall rather than a GuC fault, but
	 * banning is still the safe recovery: otherwise the queue is left with
	 * suspend_pending set and a subsequent resume() trips the !suspend_pending
	 * assert.
	 */
	if (xe_exec_queue_is_multi_queue_secondary(q)) {
		ret = guc_exec_queue_wait_suspend_done(q, blocking);
		if (ret == -ETIME)
			guc_exec_queue_suspend_timeout_ban(q);
		if (ret)
			return ret;
	}

	q = xe_exec_queue_multi_queue_primary(q);
	ret = guc_exec_queue_wait_suspend_done(q, blocking);
	if (ret == -ETIME)
		guc_exec_queue_suspend_timeout_ban(q);

	return ret;
}

static int guc_exec_queue_suspend_wait(struct xe_exec_queue *q)
{
	return guc_exec_queue_suspend_wait_common(q, false);
}

/*
 * Uninterruptible variant of guc_exec_queue_suspend_wait() for callers that
 * must complete the wait on behalf of a queue possibly owned by a different
 * process (e.g. cleanup/undo paths). An interruptible wait could return
 * -ERESTARTSYS if the calling task is signalled, leaving that queue suspended
 * forever (cross-process DoS). VF recovery is deliberately not handled (no
 * -EAGAIN) since a blocking caller cannot retry.
 */
static int guc_exec_queue_suspend_wait_blocking(struct xe_exec_queue *q)
{
	return guc_exec_queue_suspend_wait_common(q, true);
}

static void guc_exec_queue_resume(struct xe_exec_queue *q)
{
	/*
	 * Non-multi-queue queues and multi-queue primaries resume themselves
	 * directly; their own msg_lock is sufficient.
	 */
	if (!xe_exec_queue_is_multi_queue_secondary(q)) {
		__guc_exec_queue_resume(q);
		return;
	}

	/*
	 * Mirror of guc_exec_queue_suspend(): resume the secondary like any
	 * other queue and, only on its own 1->0 transition, forward the resume
	 * to the primary so the primary's GuC context is re-enabled once the
	 * last member that suspended it resumes. @suspend_lock keeps the
	 * secondary transition and the primary forward atomic.
	 */
	scoped_guard(spinlock, &q->multi_queue.group->suspend_lock) {
		if (__guc_exec_queue_resume(q))
			__guc_exec_queue_resume(xe_exec_queue_multi_queue_primary(q));
	}
}

/*
 * Drop a leaving secondary's forwarded suspend reference on the primary and
 * resume the primary if this was the last member that had it suspended.
 * See guc_exec_queue_fini().
 */
static void guc_exec_queue_multi_queue_drop_suspend(struct xe_exec_queue *q)
{
	scoped_guard(spinlock, &q->multi_queue.group->suspend_lock) {
		struct xe_exec_queue *primary = xe_exec_queue_multi_queue_primary(q);

		/*
		 * A suspended secondary holds exactly one suspend reference on the
		 * primary (forwarded on its 0->1 transition). If it leaves while
		 * still suspended, release that reference so the primary is not
		 * kept disabled forever.
		 */
		if (!READ_ONCE(q->guc->suspend_count))
			break;

		if (exec_queue_killed_or_banned_or_wedged(primary))
			break;

		/*
		 * No suspend_wait() here (and we can't - suspend_lock is a
		 * spinlock). guc_exec_queue_fini() has already drained the
		 * primary's forwarded suspend with the blocking wait, so its
		 * suspend has completed (suspend_pending cleared) by the time we
		 * resume it here. __guc_exec_queue_resume() asserts this.
		 */
		__guc_exec_queue_resume(primary);
	}
}

static bool guc_exec_queue_reset_status(struct xe_exec_queue *q)
{
	if (xe_exec_queue_is_multi_queue_secondary(q) &&
	    guc_exec_queue_reset_status(xe_exec_queue_multi_queue_primary(q)))
		return true;

	return exec_queue_reset(q) || exec_queue_killed_or_banned_or_wedged(q);
}

/*
 * All of these functions are an abstraction layer which other parts of Xe can
 * use to trap into the GuC backend. All of these functions, aside from init,
 * really shouldn't do much other than trap into the DRM scheduler which
 * synchronizes these operations.
 */
static const struct xe_exec_queue_ops guc_exec_queue_ops = {
	.init = guc_exec_queue_init,
	.kill = guc_exec_queue_kill,
	.fini = guc_exec_queue_fini,
	.destroy = guc_exec_queue_destroy,
	.set_priority = guc_exec_queue_set_priority,
	.set_timeslice = guc_exec_queue_set_timeslice,
	.set_preempt_timeout = guc_exec_queue_set_preempt_timeout,
	.set_multi_queue_priority = guc_exec_queue_set_multi_queue_priority,
	.suspend = guc_exec_queue_suspend,
	.suspend_wait = guc_exec_queue_suspend_wait,
	.suspend_wait_blocking = guc_exec_queue_suspend_wait_blocking,
	.resume = guc_exec_queue_resume,
	.reset_status = guc_exec_queue_reset_status,
};

static void guc_exec_queue_stop(struct xe_guc *guc, struct xe_exec_queue *q)
{
	struct xe_gpu_scheduler *sched = &q->guc->sched;
	bool do_destroy = false;

	/* Stop scheduling + flush any DRM scheduler operations */
	xe_sched_submission_stop(sched);

	/* Clean up lost G2H + reset engine state */
	if (exec_queue_registered(q)) {
		if (exec_queue_destroyed(q))
			do_destroy = true;
	}
	if (q->guc->suspend_pending) {
		set_exec_queue_suspended(q);
		suspend_fence_signal(q);
	}
	atomic_and(EXEC_QUEUE_STATE_WEDGED | EXEC_QUEUE_STATE_BANNED |
		   EXEC_QUEUE_STATE_KILLED | EXEC_QUEUE_STATE_DESTROYED |
		   EXEC_QUEUE_STATE_SUSPENDED,
		   &q->guc->state);
	q->guc->resume_time = 0;
	trace_xe_exec_queue_stop(q);

	/*
	 * Ban any engine (aside from kernel and engines used for VM ops) with a
	 * started but not complete job or if a job has gone through a GT reset
	 * more than twice.
	 */
	if (!(q->flags & (EXEC_QUEUE_FLAG_KERNEL | EXEC_QUEUE_FLAG_VM))) {
		struct xe_sched_job *job = xe_sched_first_pending_job(sched);
		bool ban = false;

		if (job) {
			if ((xe_sched_job_started(job) &&
			    !xe_sched_job_completed(job)) ||
			    xe_sched_invalidate_job(job, 2)) {
				trace_xe_sched_job_ban(job);
				ban = true;
			}
		}

		if (ban) {
			set_exec_queue_banned(q);
			xe_guc_exec_queue_trigger_cleanup(q);
		}
	}

	if (do_destroy)
		__guc_exec_queue_destroy(guc, q);
}

static int guc_submit_reset_prepare(struct xe_guc *guc)
{
	int ret;

	/*
	 * Using an atomic here rather than submission_state.lock as this
	 * function can be called while holding the CT lock (engine reset
	 * failure). submission_state.lock needs the CT lock to resubmit jobs.
	 * Atomic is not ideal, but it works to prevent against concurrent reset
	 * and releasing any TDRs waiting on guc->submission_state.stopped.
	 */
	ret = atomic_fetch_or(1, &guc->submission_state.stopped);
	smp_wmb();
	wake_up_all(&guc->ct.wq);

	return ret;
}

int xe_guc_submit_reset_prepare(struct xe_guc *guc)
{
	if (xe_gt_WARN_ON(guc_to_gt(guc), vf_recovery(guc)))
		return 0;

	if (!guc->submission_state.initialized)
		return 0;

	return guc_submit_reset_prepare(guc);
}

void xe_guc_submit_reset_wait(struct xe_guc *guc)
{
	wait_event(guc->ct.wq, xe_device_wedged(guc_to_xe(guc)) ||
		   !xe_guc_read_stopped(guc));
}

void xe_guc_submit_stop(struct xe_guc *guc)
{
	struct xe_exec_queue *q;
	unsigned long index;

	xe_gt_assert(guc_to_gt(guc), xe_guc_read_stopped(guc) == 1);

	mutex_lock(&guc->submission_state.lock);

	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q) {
		/* Prevent redundant attempts to stop parallel queues */
		if (q->guc->id != index)
			continue;

		guc_exec_queue_stop(guc, q);
	}

	mutex_unlock(&guc->submission_state.lock);

	/*
	 * No one can enter the backend at this point, aside from new engine
	 * creation which is protected by guc->submission_state.lock.
	 */

}

static void guc_exec_queue_revert_pending_state_change(struct xe_guc *guc,
						       struct xe_exec_queue *q)
{
	bool pending_enable, pending_disable, pending_resume;

	pending_enable = exec_queue_pending_enable(q);
	pending_resume = exec_queue_pending_resume(q);

	if (pending_enable && pending_resume) {
		q->guc->needs_resume = true;
		xe_gt_dbg(guc_to_gt(guc), "Replay RESUME - guc_id=%d",
			  q->guc->id);
	}

	if (pending_enable && !pending_resume) {
		clear_exec_queue_registered(q);
		xe_gt_dbg(guc_to_gt(guc), "Replay REGISTER - guc_id=%d",
			  q->guc->id);
	}

	if (pending_enable) {
		clear_exec_queue_enabled(q);
		clear_exec_queue_pending_resume(q);
		clear_exec_queue_pending_enable(q);
		xe_gt_dbg(guc_to_gt(guc), "Replay ENABLE - guc_id=%d",
			  q->guc->id);
	}

	if (exec_queue_destroyed(q) && exec_queue_registered(q)) {
		clear_exec_queue_destroyed(q);
		q->guc->needs_cleanup = true;
		xe_gt_dbg(guc_to_gt(guc), "Replay CLEANUP - guc_id=%d",
			  q->guc->id);
	}

	pending_disable = exec_queue_pending_disable(q);

	if (pending_disable && exec_queue_suspended(q)) {
		clear_exec_queue_suspended(q);
		q->guc->needs_suspend = true;
		xe_gt_dbg(guc_to_gt(guc), "Replay SUSPEND - guc_id=%d",
			  q->guc->id);
	}

	if (pending_disable) {
		if (!pending_enable)
			set_exec_queue_enabled(q);
		clear_exec_queue_pending_disable(q);
		xe_gt_dbg(guc_to_gt(guc), "Replay DISABLE - guc_id=%d",
			  q->guc->id);
	}

	q->guc->resume_time = 0;
}

static void lrc_parallel_clear(struct xe_lrc *lrc)
{
	struct xe_device *xe = gt_to_xe(lrc->gt);
	struct iosys_map map = xe_lrc_parallel_map(lrc);
	int i;

	for (i = 0; i < WQ_SIZE / sizeof(u32); ++i)
		parallel_write(xe, map, wq[i],
			       FIELD_PREP(WQ_TYPE_MASK, WQ_TYPE_NOOP) |
			       FIELD_PREP(WQ_LEN_MASK, 0));
}

/*
 * This function is quite complex but only real way to ensure no state is lost
 * during VF resume flows. The function scans the queue state, make adjustments
 * as needed, and queues jobs / messages which replayed upon unpause.
 */
static void guc_exec_queue_pause(struct xe_guc *guc, struct xe_exec_queue *q)
{
	struct xe_gpu_scheduler *sched = &q->guc->sched;
	struct xe_sched_job *job;
	int i;

	lockdep_assert_held(&guc->submission_state.lock);

	/* Stop scheduling + flush any DRM scheduler operations */
	xe_sched_submission_stop(sched);
	cancel_delayed_work_sync(&sched->base.work_tdr);

	guc_exec_queue_revert_pending_state_change(guc, q);

	if (xe_exec_queue_is_parallel(q)) {
		/* Pairs with WRITE_ONCE in __xe_exec_queue_init  */
		struct xe_lrc *lrc = READ_ONCE(q->lrc[0]);

		/*
		 * NOP existing WQ commands that may contain stale GGTT
		 * addresses. These will be replayed upon unpause. The hardware
		 * seems to get confused if the WQ head/tail pointers are
		 * adjusted.
		 */
		if (lrc)
			lrc_parallel_clear(lrc);
	}

	job = xe_sched_first_pending_job(sched);
	if (job) {
		job->restore_replay = true;

		/*
		 * Adjust software tail so jobs submitted overwrite previous
		 * position in ring buffer with new GGTT addresses.
		 */
		for (i = 0; i < q->width; ++i)
			q->lrc[i]->ring.tail = job->ptrs[i].head;
	}
}

/**
 * xe_guc_submit_pause - Stop further runs of submission tasks on given GuC.
 * @guc: the &xe_guc struct instance whose scheduler is to be disabled
 */
void xe_guc_submit_pause(struct xe_guc *guc)
{
	struct xe_exec_queue *q;
	unsigned long index;

	mutex_lock(&guc->submission_state.lock);
	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q)
		xe_sched_submission_stop(&q->guc->sched);
	mutex_unlock(&guc->submission_state.lock);
}

/**
 * xe_guc_submit_pause_vf - Stop further runs of submission tasks for VF.
 * @guc: the &xe_guc struct instance whose scheduler is to be disabled
 */
void xe_guc_submit_pause_vf(struct xe_guc *guc)
{
	struct xe_exec_queue *q;
	unsigned long index;

	xe_gt_assert(guc_to_gt(guc), IS_SRIOV_VF(guc_to_xe(guc)));
	xe_gt_assert(guc_to_gt(guc), vf_recovery(guc));

	mutex_lock(&guc->submission_state.lock);
	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q) {
		/* Prevent redundant attempts to stop parallel queues */
		if (q->guc->id != index)
			continue;

		guc_exec_queue_pause(guc, q);
	}
	mutex_unlock(&guc->submission_state.lock);
}

static void guc_exec_queue_start(struct xe_exec_queue *q)
{
	struct xe_gpu_scheduler *sched = &q->guc->sched;

	if (!exec_queue_killed_or_banned_or_wedged(q)) {
		struct xe_sched_job *job = xe_sched_first_pending_job(sched);
		int i;

		trace_xe_exec_queue_resubmit(q);
		if (job) {
			for (i = 0; i < q->width; ++i) {
				/*
				 * The GuC context is unregistered at this point
				 * time, adjusting software ring tail ensures
				 * jobs are rewritten in original placement,
				 * adjusting LRC tail ensures the newly loaded
				 * GuC / contexts only view the LRC tail
				 * increasing as jobs are written out.
				 */
				q->lrc[i]->ring.tail = job->ptrs[i].head;
				xe_lrc_set_ring_tail(q->lrc[i],
						     xe_lrc_ring_head(q->lrc[i]));
			}
		}
		xe_sched_resubmit_jobs(sched);
	}

	xe_sched_submission_start(sched);
	xe_sched_submission_resume_tdr(sched);
}

int xe_guc_submit_start(struct xe_guc *guc)
{
	struct xe_exec_queue *q;
	unsigned long index;

	xe_gt_assert(guc_to_gt(guc), xe_guc_read_stopped(guc) == 1);

	mutex_lock(&guc->submission_state.lock);
	atomic_dec(&guc->submission_state.stopped);
	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q) {
		/* Prevent redundant attempts to start parallel queues */
		if (q->guc->id != index)
			continue;

		guc_exec_queue_start(q);
	}
	mutex_unlock(&guc->submission_state.lock);

	wake_up_all(&guc->ct.wq);

	return 0;
}

static void guc_exec_queue_unpause_prepare(struct xe_guc *guc,
					   struct xe_exec_queue *q)
{
	struct xe_gpu_scheduler *sched = &q->guc->sched;
	struct xe_sched_job *job = NULL;
	struct drm_sched_job *s_job;
	bool restore_replay = false;

	drm_sched_for_each_pending_job(s_job, &sched->base, NULL) {
		job = to_xe_sched_job(s_job);
		restore_replay |= job->restore_replay;
		if (restore_replay) {
			xe_gt_dbg(guc_to_gt(guc), "Replay JOB - guc_id=%d, seqno=%d",
				  q->guc->id, xe_sched_job_seqno(job));

			q->ring_ops->emit_job(job);
			job->restore_replay = true;
		}
	}

	if (job)
		job->last_replay = true;
}

/**
 * xe_guc_submit_unpause_prepare_vf - Prepare unpause submission tasks for VF.
 * @guc: the &xe_guc struct instance whose scheduler is to be prepared for unpause
 */
void xe_guc_submit_unpause_prepare_vf(struct xe_guc *guc)
{
	struct xe_exec_queue *q;
	unsigned long index;

	xe_gt_assert(guc_to_gt(guc), IS_SRIOV_VF(guc_to_xe(guc)));
	xe_gt_assert(guc_to_gt(guc), vf_recovery(guc));

	mutex_lock(&guc->submission_state.lock);
	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q) {
		/* Prevent redundant attempts to stop parallel queues */
		if (q->guc->id != index)
			continue;

		guc_exec_queue_unpause_prepare(guc, q);
	}
	mutex_unlock(&guc->submission_state.lock);
}

static void guc_exec_queue_replay_pending_state_change(struct xe_exec_queue *q)
{
	struct xe_gpu_scheduler *sched = &q->guc->sched;
	struct xe_sched_msg *msg;

	if (q->guc->needs_cleanup) {
		msg = q->guc->static_msgs + STATIC_MSG_CLEANUP;

		guc_exec_queue_add_msg(q, msg, CLEANUP);
		q->guc->needs_cleanup = false;
	}

	if (q->guc->needs_suspend) {
		msg = q->guc->static_msgs + STATIC_MSG_SUSPEND;

		xe_sched_msg_lock(sched);
		guc_exec_queue_try_add_msg_head(q, msg, SUSPEND);
		xe_sched_msg_unlock(sched);

		q->guc->needs_suspend = false;
	}

	/*
	 * The resume must be in the message queue before the suspend as it is
	 * not possible for a resume to be issued if a suspend pending is, but
	 * the inverse is possible.
	 */
	if (q->guc->needs_resume) {
		msg = q->guc->static_msgs + STATIC_MSG_RESUME;

		xe_sched_msg_lock(sched);
		guc_exec_queue_try_add_msg_head(q, msg, RESUME);
		xe_sched_msg_unlock(sched);

		q->guc->needs_resume = false;
	}
}

static void guc_exec_queue_unpause(struct xe_guc *guc, struct xe_exec_queue *q)
{
	struct xe_gpu_scheduler *sched = &q->guc->sched;
	bool needs_tdr = exec_queue_killed_or_banned_or_wedged(q);

	lockdep_assert_held(&guc->submission_state.lock);

	xe_sched_resubmit_jobs(sched);
	guc_exec_queue_replay_pending_state_change(q);
	xe_sched_submission_start(sched);
	if (needs_tdr)
		xe_guc_exec_queue_trigger_cleanup(q);
	xe_sched_submission_resume_tdr(sched);
}

/**
 * xe_guc_submit_unpause - Allow further runs of submission tasks on given GuC.
 * @guc: the &xe_guc struct instance whose scheduler is to be enabled
 */
void xe_guc_submit_unpause(struct xe_guc *guc)
{
	struct xe_exec_queue *q;
	unsigned long index;

	mutex_lock(&guc->submission_state.lock);
	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q)
		xe_sched_submission_start(&q->guc->sched);
	mutex_unlock(&guc->submission_state.lock);
}

/**
 * xe_guc_submit_unpause_vf - Allow further runs of submission tasks for VF.
 * @guc: the &xe_guc struct instance whose scheduler is to be enabled
 */
void xe_guc_submit_unpause_vf(struct xe_guc *guc)
{
	struct xe_exec_queue *q;
	unsigned long index;

	xe_gt_assert(guc_to_gt(guc), IS_SRIOV_VF(guc_to_xe(guc)));

	mutex_lock(&guc->submission_state.lock);
	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q) {
		/*
		 * Prevent redundant attempts to stop parallel queues, or queues
		 * created after resfix done.
		 */
		if (q->guc->id != index ||
		    !drm_sched_is_stopped(&q->guc->sched.base))
			continue;

		guc_exec_queue_unpause(guc, q);
	}
	mutex_unlock(&guc->submission_state.lock);
}

/**
 * xe_guc_submit_pause_abort - Abort all paused submission task on given GuC.
 * @guc: the &xe_guc struct instance whose scheduler is to be aborted
 */
void xe_guc_submit_pause_abort(struct xe_guc *guc)
{
	struct xe_exec_queue *q;
	unsigned long index;

	mutex_lock(&guc->submission_state.lock);
	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q) {
		struct xe_gpu_scheduler *sched = &q->guc->sched;

		/* Prevent redundant attempts to stop parallel queues */
		if (q->guc->id != index)
			continue;

		xe_sched_submission_start(sched);
		guc_exec_queue_kill(q);
	}
	mutex_unlock(&guc->submission_state.lock);
}

static struct xe_exec_queue *
g2h_exec_queue_lookup(struct xe_guc *guc, u32 guc_id)
{
	struct xe_gt *gt = guc_to_gt(guc);
	struct xe_exec_queue *q;

	if (unlikely(guc_id >= GUC_ID_MAX)) {
		xe_gt_err(gt, "Invalid guc_id %u\n", guc_id);
		return NULL;
	}

	q = xa_load(&guc->submission_state.exec_queue_lookup, guc_id);
	if (unlikely(!q)) {
		xe_gt_err(gt, "No exec queue found for guc_id %u\n", guc_id);
		return NULL;
	}

	xe_gt_assert(guc_to_gt(guc), guc_id >= q->guc->id);
	xe_gt_assert(guc_to_gt(guc), guc_id < (q->guc->id + q->width));

	return q;
}

static void deregister_exec_queue(struct xe_guc *guc, struct xe_exec_queue *q)
{
	u32 action[] = {
		XE_GUC_ACTION_DEREGISTER_CONTEXT,
		q->guc->id,
	};

	xe_gt_assert(guc_to_gt(guc), exec_queue_destroyed(q));
	xe_gt_assert(guc_to_gt(guc), exec_queue_registered(q));
	xe_gt_assert(guc_to_gt(guc), !exec_queue_pending_disable(q));
	xe_gt_assert(guc_to_gt(guc), !exec_queue_pending_enable(q));

	trace_xe_exec_queue_deregister(q);

	if (xe_exec_queue_is_multi_queue_secondary(q))
		handle_deregister_done(guc, q);
	else
		xe_guc_ct_send_g2h_handler(&guc->ct, action,
					   ARRAY_SIZE(action));
}

static void handle_sched_done(struct xe_guc *guc, struct xe_exec_queue *q,
			      u32 runnable_state)
{
	trace_xe_exec_queue_scheduling_done(q);

	if (runnable_state == 1) {
		xe_gt_assert(guc_to_gt(guc), exec_queue_pending_enable(q));

		q->guc->resume_time = ktime_get();
		clear_exec_queue_pending_resume(q);
		clear_exec_queue_pending_enable(q);
		smp_wmb();
		wake_up_all(&guc->ct.wq);
	} else {
		xe_gt_assert(guc_to_gt(guc), runnable_state == 0);
		xe_gt_assert(guc_to_gt(guc), exec_queue_pending_disable(q));

		if (q->guc->suspend_pending) {
			clear_exec_queue_pending_disable(q);
			suspend_fence_signal(q);
		} else {
			if (exec_queue_banned(q)) {
				smp_wmb();
				wake_up_all(&guc->ct.wq);
			}
			if (exec_queue_destroyed(q)) {
				/*
				 * Make sure to clear the pending_disable only
				 * after sampling the destroyed state. We want
				 * to ensure we don't trigger the unregister too
				 * early with something intending to only
				 * disable scheduling. The caller doing the
				 * destroy must wait for an ongoing
				 * pending_disable before marking as destroyed.
				 */
				clear_exec_queue_pending_disable(q);
				deregister_exec_queue(guc, q);
			} else {
				clear_exec_queue_pending_disable(q);
			}
		}
	}
}

static void handle_multi_queue_secondary_sched_done(struct xe_guc *guc,
						    struct xe_exec_queue *q,
						    u32 runnable_state)
{
	/* Take CT lock here as handle_sched_done() do send a h2g message */
	mutex_lock(&guc->ct.lock);
	handle_sched_done(guc, q, runnable_state);
	mutex_unlock(&guc->ct.lock);
}

int xe_guc_sched_done_handler(struct xe_guc *guc, u32 *msg, u32 len)
{
	struct xe_exec_queue *q;
	u32 guc_id, runnable_state;

	if (unlikely(len < 2))
		return -EPROTO;

	guc_id = msg[0];
	runnable_state = msg[1];

	q = g2h_exec_queue_lookup(guc, guc_id);
	if (unlikely(!q))
		return -EPROTO;

	if (unlikely(!exec_queue_pending_enable(q) &&
		     !exec_queue_pending_disable(q))) {
		xe_gt_err(guc_to_gt(guc),
			  "SCHED_DONE: Unexpected engine state 0x%04x, guc_id=%d, runnable_state=%u",
			  atomic_read(&q->guc->state), q->guc->id,
			  runnable_state);
		return -EPROTO;
	}

	handle_sched_done(guc, q, runnable_state);

	return 0;
}

static void handle_deregister_done(struct xe_guc *guc, struct xe_exec_queue *q)
{
	trace_xe_exec_queue_deregister_done(q);

	clear_exec_queue_registered(q);
	__guc_exec_queue_destroy(guc, q);
}

int xe_guc_deregister_done_handler(struct xe_guc *guc, u32 *msg, u32 len)
{
	struct xe_exec_queue *q;
	u32 guc_id;

	if (unlikely(len < 1))
		return -EPROTO;

	guc_id = msg[0];

	q = g2h_exec_queue_lookup(guc, guc_id);
	if (unlikely(!q))
		return -EPROTO;

	if (!exec_queue_destroyed(q) || exec_queue_pending_disable(q) ||
	    exec_queue_pending_enable(q) || exec_queue_enabled(q)) {
		xe_gt_err(guc_to_gt(guc),
			  "DEREGISTER_DONE: Unexpected engine state 0x%04x, guc_id=%d",
			  atomic_read(&q->guc->state), q->guc->id);
		return -EPROTO;
	}

	handle_deregister_done(guc, q);

	return 0;
}

int xe_guc_exec_queue_reset_handler(struct xe_guc *guc, u32 *msg, u32 len)
{
	struct xe_gt *gt = guc_to_gt(guc);
	struct xe_exec_queue *q;
	u32 guc_id;

	if (unlikely(len < 1))
		return -EPROTO;

	guc_id = msg[0];

	q = g2h_exec_queue_lookup(guc, guc_id);
	if (unlikely(!q))
		return -EPROTO;

	if (!exec_queue_killed(q))
		xe_gt_info(gt, "Engine reset: engine_class=%s, logical_mask: 0x%x, guc_id=%d, state=0x%0x",
			   xe_hw_engine_class_to_str(q->class), q->logical_mask, guc_id,
			   atomic_read(&q->guc->state));

	trace_xe_exec_queue_reset(q);

	/*
	 * A banned engine is a NOP at this point (came from
	 * guc_exec_queue_timedout_job). Otherwise, kick drm scheduler to cancel
	 * jobs by setting timeout of the job to the minimum value kicking
	 * guc_exec_queue_timedout_job.
	 */
	xe_guc_exec_queue_reset_trigger_cleanup(q);

	return 0;
}

/*
 * xe_guc_error_capture_handler - Handler of GuC captured message
 * @guc: The GuC object
 * @msg: Point to the message
 * @len: The message length
 *
 * When GuC captured data is ready, GuC will send message
 * XE_GUC_ACTION_STATE_CAPTURE_NOTIFICATION to host, this function will be
 * called 1st to check status before process the data comes with the message.
 *
 * Returns: error code. 0 if success
 */
int xe_guc_error_capture_handler(struct xe_guc *guc, u32 *msg, u32 len)
{
	u32 status;

	if (unlikely(len != XE_GUC_ACTION_STATE_CAPTURE_NOTIFICATION_DATA_LEN))
		return -EPROTO;

	status = msg[0] & XE_GUC_STATE_CAPTURE_EVENT_STATUS_MASK;
	if (status == XE_GUC_STATE_CAPTURE_EVENT_STATUS_NOSPACE)
		xe_gt_warn(guc_to_gt(guc), "G2H-Error capture no space");

	xe_guc_capture_process(guc);

	return 0;
}

int xe_guc_exec_queue_memory_cat_error_handler(struct xe_guc *guc, u32 *msg,
					       u32 len)
{
	struct xe_gt *gt = guc_to_gt(guc);
	struct xe_exec_queue *q;
	u32 guc_id;
	u32 type = XE_GUC_CAT_ERR_TYPE_INVALID;

	if (unlikely(!len || len > 2))
		return -EPROTO;

	guc_id = msg[0];

	if (len == 2)
		type = msg[1];

	if (guc_id == GUC_ID_UNKNOWN) {
		/*
		 * GuC uses GUC_ID_UNKNOWN if it can not map the CAT fault to any PF/VF
		 * context. In such case only PF will be notified about that fault.
		 */
		xe_gt_err_ratelimited(gt, "Memory CAT error reported by GuC!\n");
		return 0;
	}

	q = g2h_exec_queue_lookup(guc, guc_id);
	if (unlikely(!q))
		return -EPROTO;

	/*
	 * The type is HW-defined and changes based on platform, so we don't
	 * decode it in the kernel and only check if it is valid.
	 * See bspec 54047 and 72187 for details.
	 */
	if (type != XE_GUC_CAT_ERR_TYPE_INVALID)
		xe_gt_info(gt,
			   "Engine memory CAT error [%u]: class=%s, logical_mask: 0x%x, guc_id=%d",
			   type, xe_hw_engine_class_to_str(q->class), q->logical_mask, guc_id);
	else
		xe_gt_info(gt,
			   "Engine memory CAT error: class=%s, logical_mask: 0x%x, guc_id=%d",
			   xe_hw_engine_class_to_str(q->class), q->logical_mask, guc_id);

	trace_xe_exec_queue_memory_cat_error(q);

	/* Treat the same as engine reset */
	xe_guc_exec_queue_reset_trigger_cleanup(q);

	return 0;
}

int xe_guc_uncorrectable_error_handler(struct xe_guc *guc, u32 *msg, u32 len)
{
	struct xe_gt *gt = guc_to_gt(guc);
	struct xe_exec_queue *q;
	u32 guc_id;

	if (unlikely(!len || len > 1))
		return -EPROTO;

	guc_id = msg[0];

	if (guc_id == GUC_ID_UNKNOWN) {
		xe_gt_err(gt, "GuC: Uncorrectable local error with unknown GuC id\n");
		return 0;
	}

	q = g2h_exec_queue_lookup(guc, guc_id);
	if (unlikely(!q))
		return -EPROTO;

	xe_gt_err(gt,
		  "GuC: Uncorrectable local error! guc_id=%d class=%s, logical_mask=0x%x",
		  guc_id, xe_hw_engine_class_to_str(q->class), q->logical_mask);

	trace_xe_guc_uncorrectable_error(q);

	/* Treat the same as engine reset */
	xe_guc_exec_queue_reset_trigger_cleanup(q);

	return 0;
}

int xe_guc_exec_queue_reset_failure_handler(struct xe_guc *guc, u32 *msg, u32 len)
{
	struct xe_gt *gt = guc_to_gt(guc);
	u8 guc_class, instance;
	u32 reason;

	if (unlikely(len != 3))
		return -EPROTO;

	guc_class = msg[0];
	instance = msg[1];
	reason = msg[2];

	/* Unexpected failure of a hardware feature, log an actual error */
	xe_gt_err(gt, "GuC engine reset request failed on %d:%d because 0x%08X",
		  guc_class, instance, reason);

	xe_gt_reset_async(gt);

	return 0;
}

int xe_guc_exec_queue_cgp_context_error_handler(struct xe_guc *guc, u32 *msg,
						u32 len)
{
	struct xe_gt *gt = guc_to_gt(guc);
	struct xe_device *xe = guc_to_xe(guc);
	struct xe_exec_queue *q;
	u32 guc_id = msg[2];

	if (unlikely(len != XE_GUC_EXEC_QUEUE_CGP_CONTEXT_ERROR_LEN)) {
		drm_err(&xe->drm, "Invalid length %u", len);
		return -EPROTO;
	}

	q = g2h_exec_queue_lookup(guc, guc_id);
	if (unlikely(!q))
		return -EPROTO;

	xe_gt_dbg(gt,
		  "CGP context error: [%s] err=0x%x, q0_id=0x%x LRCA=0x%x guc_id=0x%x",
		  msg[0] & 1 ? "uc" : "kmd", msg[1], msg[2], msg[3], msg[4]);

	trace_xe_exec_queue_cgp_context_error(q);

	/* Treat the same as engine reset */
	xe_guc_exec_queue_reset_trigger_cleanup(q);

	return 0;
}

/**
 * xe_guc_exec_queue_cgp_sync_done_handler - CGP synchronization done handler
 * @guc: guc
 * @msg: message indicating CGP sync done
 * @len: length of message
 *
 * Set multi queue group's sync_pending flag to false and wakeup anyone waiting
 * for CGP synchronization to complete.
 *
 * Return: 0 on success, -EPROTO for malformed messages.
 */
int xe_guc_exec_queue_cgp_sync_done_handler(struct xe_guc *guc, u32 *msg, u32 len)
{
	struct xe_device *xe = guc_to_xe(guc);
	struct xe_exec_queue *q;
	u32 guc_id = msg[0];

	if (unlikely(len < 1)) {
		drm_err(&xe->drm, "Invalid CGP_SYNC_DONE length %u", len);
		return -EPROTO;
	}

	q = g2h_exec_queue_lookup(guc, guc_id);
	if (unlikely(!q))
		return -EPROTO;

	if (!xe_exec_queue_is_multi_queue_primary(q)) {
		drm_err(&xe->drm, "Unexpected CGP_SYNC_DONE response");
		return -EPROTO;
	}

	/* Wakeup the serialized cgp update wait */
	WRITE_ONCE(q->multi_queue.group->sync_pending, false);
	xe_guc_ct_wake_waiters(&guc->ct);

	return 0;
}

static void
guc_exec_queue_wq_snapshot_capture(struct xe_exec_queue *q,
				   struct xe_guc_submit_exec_queue_snapshot *snapshot)
{
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_device *xe = guc_to_xe(guc);
	struct iosys_map map = xe_lrc_parallel_map(q->lrc[0]);
	int i;

	snapshot->guc.wqi_head = q->guc->wqi_head;
	snapshot->guc.wqi_tail = q->guc->wqi_tail;
	snapshot->parallel.wq_desc.head = parallel_read(xe, map, wq_desc.head);
	snapshot->parallel.wq_desc.tail = parallel_read(xe, map, wq_desc.tail);
	snapshot->parallel.wq_desc.status = parallel_read(xe, map,
							  wq_desc.wq_status);

	if (snapshot->parallel.wq_desc.head !=
	    snapshot->parallel.wq_desc.tail) {
		for (i = snapshot->parallel.wq_desc.head;
		     i != snapshot->parallel.wq_desc.tail;
		     i = (i + sizeof(u32)) % WQ_SIZE)
			snapshot->parallel.wq[i / sizeof(u32)] =
				parallel_read(xe, map, wq[i / sizeof(u32)]);
	}
}

static void
guc_exec_queue_wq_snapshot_print(struct xe_guc_submit_exec_queue_snapshot *snapshot,
				 struct drm_printer *p)
{
	int i;

	drm_printf(p, "\tWQ head: %u (internal), %d (memory)\n",
		   snapshot->guc.wqi_head, snapshot->parallel.wq_desc.head);
	drm_printf(p, "\tWQ tail: %u (internal), %d (memory)\n",
		   snapshot->guc.wqi_tail, snapshot->parallel.wq_desc.tail);
	drm_printf(p, "\tWQ status: %u\n", snapshot->parallel.wq_desc.status);

	if (snapshot->parallel.wq_desc.head !=
	    snapshot->parallel.wq_desc.tail) {
		for (i = snapshot->parallel.wq_desc.head;
		     i != snapshot->parallel.wq_desc.tail;
		     i = (i + sizeof(u32)) % WQ_SIZE)
			drm_printf(p, "\tWQ[%zu]: 0x%08x\n", i / sizeof(u32),
				   snapshot->parallel.wq[i / sizeof(u32)]);
	}
}

/**
 * xe_guc_exec_queue_snapshot_capture - Take a quick snapshot of the GuC Engine.
 * @q: faulty exec queue
 *
 * This can be printed out in a later stage like during dev_coredump
 * analysis.
 *
 * Returns: a GuC Submit Engine snapshot object that must be freed by the
 * caller, using `xe_guc_exec_queue_snapshot_free`.
 */
struct xe_guc_submit_exec_queue_snapshot *
xe_guc_exec_queue_snapshot_capture(struct xe_exec_queue *q)
{
	struct xe_gpu_scheduler *sched = &q->guc->sched;
	struct xe_guc_submit_exec_queue_snapshot *snapshot;
	int i;

	snapshot = kzalloc_obj(*snapshot, GFP_ATOMIC);

	if (!snapshot)
		return NULL;

	snapshot->guc.id = q->guc->id;
	memcpy(&snapshot->name, &q->name, sizeof(snapshot->name));
	snapshot->class = q->class;
	snapshot->logical_mask = q->logical_mask;
	snapshot->width = q->width;
	snapshot->refcount = kref_read(&q->refcount);
	snapshot->sched_timeout = sched->base.timeout;
	snapshot->sched_props.timeslice_us = q->sched_props.timeslice_us;
	snapshot->sched_props.preempt_timeout_us =
		q->sched_props.preempt_timeout_us;

	snapshot->lrc = kmalloc_objs(struct xe_lrc_snapshot *, q->width,
				     GFP_ATOMIC);

	if (snapshot->lrc) {
		for (i = 0; i < q->width; ++i) {
			struct xe_lrc *lrc = q->lrc[i];

			snapshot->lrc[i] = xe_lrc_snapshot_capture(lrc);
		}
	}

	snapshot->schedule_state = atomic_read(&q->guc->state);
	snapshot->exec_queue_flags = q->flags;

	snapshot->parallel_execution = xe_exec_queue_is_parallel(q);
	if (snapshot->parallel_execution)
		guc_exec_queue_wq_snapshot_capture(q, snapshot);

	if (xe_exec_queue_is_multi_queue(q)) {
		snapshot->multi_queue.valid = true;
		snapshot->multi_queue.primary = xe_exec_queue_multi_queue_primary(q)->guc->id;
		snapshot->multi_queue.pos = q->multi_queue.pos;
	}

	return snapshot;
}

/**
 * xe_guc_exec_queue_snapshot_capture_delayed - Take delayed part of snapshot of the GuC Engine.
 * @snapshot: Previously captured snapshot of job.
 *
 * This captures some data that requires taking some locks, so it cannot be done in signaling path.
 */
void
xe_guc_exec_queue_snapshot_capture_delayed(struct xe_guc_submit_exec_queue_snapshot *snapshot)
{
	int i;

	if (!snapshot || !snapshot->lrc)
		return;

	for (i = 0; i < snapshot->width; ++i)
		xe_lrc_snapshot_capture_delayed(snapshot->lrc[i]);
}

/**
 * xe_guc_exec_queue_snapshot_print - Print out a given GuC Engine snapshot.
 * @snapshot: GuC Submit Engine snapshot object.
 * @p: drm_printer where it will be printed out.
 *
 * This function prints out a given GuC Submit Engine snapshot object.
 */
void
xe_guc_exec_queue_snapshot_print(struct xe_guc_submit_exec_queue_snapshot *snapshot,
				 struct drm_printer *p)
{
	int i;

	if (!snapshot)
		return;

	drm_printf(p, "GuC ID: %d\n", snapshot->guc.id);
	drm_printf(p, "\tName: %s\n", snapshot->name);
	drm_printf(p, "\tClass: %d\n", snapshot->class);
	drm_printf(p, "\tLogical mask: 0x%x\n", snapshot->logical_mask);
	drm_printf(p, "\tWidth: %d\n", snapshot->width);
	drm_printf(p, "\tRef: %d\n", snapshot->refcount);
	drm_printf(p, "\tTimeout: %ld (ms)\n", snapshot->sched_timeout);
	drm_printf(p, "\tTimeslice: %u (us)\n",
		   snapshot->sched_props.timeslice_us);
	drm_printf(p, "\tPreempt timeout: %u (us)\n",
		   snapshot->sched_props.preempt_timeout_us);

	for (i = 0; snapshot->lrc && i < snapshot->width; ++i)
		xe_lrc_snapshot_print(snapshot->lrc[i], p);

	drm_printf(p, "\tSchedule State: 0x%x\n", snapshot->schedule_state);
	drm_printf(p, "\tFlags: 0x%lx\n", snapshot->exec_queue_flags);

	if (snapshot->parallel_execution)
		guc_exec_queue_wq_snapshot_print(snapshot, p);

	if (snapshot->multi_queue.valid) {
		drm_printf(p, "\tMulti queue primary GuC ID: %d\n", snapshot->multi_queue.primary);
		drm_printf(p, "\tMulti queue position: %d\n", snapshot->multi_queue.pos);
	}
}

/**
 * xe_guc_exec_queue_snapshot_free - Free all allocated objects for a given
 * snapshot.
 * @snapshot: GuC Submit Engine snapshot object.
 *
 * This function free all the memory that needed to be allocated at capture
 * time.
 */
void xe_guc_exec_queue_snapshot_free(struct xe_guc_submit_exec_queue_snapshot *snapshot)
{
	int i;

	if (!snapshot)
		return;

	if (snapshot->lrc) {
		for (i = 0; i < snapshot->width; i++)
			xe_lrc_snapshot_free(snapshot->lrc[i]);
		kfree(snapshot->lrc);
	}
	kfree(snapshot);
}

static void guc_exec_queue_print(struct xe_exec_queue *q, struct drm_printer *p)
{
	struct xe_guc_submit_exec_queue_snapshot *snapshot;

	snapshot = xe_guc_exec_queue_snapshot_capture(q);
	xe_guc_exec_queue_snapshot_print(snapshot, p);
	xe_guc_exec_queue_snapshot_free(snapshot);
}

/**
 * xe_guc_register_vf_exec_queue - Register exec queue for a given context type.
 * @q: Execution queue
 * @ctx_type: Type of the context
 *
 * This function registers the execution queue with the guc. Special context
 * types like GUC_CONTEXT_COMPRESSION_SAVE and GUC_CONTEXT_COMPRESSION_RESTORE
 * are only applicable for IGPU and in the VF.
 * Submits the execution queue to GUC after registering it.
 *
 * Returns - None.
 */
void xe_guc_register_vf_exec_queue(struct xe_exec_queue *q, int ctx_type)
{
	struct xe_guc *guc = exec_queue_to_guc(q);
	struct xe_device *xe = guc_to_xe(guc);
	struct xe_gt *gt = guc_to_gt(guc);

	xe_gt_assert(gt, IS_SRIOV_VF(xe));
	xe_gt_assert(gt, !IS_DGFX(xe));
	xe_gt_assert(gt, ctx_type == GUC_CONTEXT_COMPRESSION_SAVE ||
		     ctx_type == GUC_CONTEXT_COMPRESSION_RESTORE);
	xe_gt_assert(gt, GUC_SUBMIT_VER(guc) >= MAKE_GUC_VER(1, 23, 0));

	register_exec_queue(q, ctx_type);
	enable_scheduling(q);
}

/**
 * xe_guc_submit_print - GuC Submit Print.
 * @guc: GuC.
 * @p: drm_printer where it will be printed out.
 *
 * This function capture and prints snapshots of **all** GuC Engines.
 */
void xe_guc_submit_print(struct xe_guc *guc, struct drm_printer *p)
{
	struct xe_exec_queue *q;
	unsigned long index;

	if (!xe_device_uc_enabled(guc_to_xe(guc)))
		return;

	mutex_lock(&guc->submission_state.lock);
	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q)
		guc_exec_queue_print(q, p);
	mutex_unlock(&guc->submission_state.lock);
}

/**
 * xe_guc_has_registered_mlrc_queues - check whether there are any MLRC queues
 * registered with the GuC
 * @guc: GuC.
 *
 * Return: true if any MLRC queue is registered with the GuC, false otherwise.
 */
bool xe_guc_has_registered_mlrc_queues(struct xe_guc *guc)
{
	struct xe_exec_queue *q;
	unsigned long index;

	guard(mutex)(&guc->submission_state.lock);

	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q)
		if (q->width > 1)
			return true;

	return false;
}

/**
 * xe_guc_contexts_hwsp_rebase - Re-compute GGTT references within all
 * exec queues registered to given GuC.
 * @guc: the &xe_guc struct instance
 * @scratch: scratch buffer to be used as temporary storage
 *
 * Returns: zero on success, negative error code on failure.
 */
int xe_guc_contexts_hwsp_rebase(struct xe_guc *guc, void *scratch)
{
	struct xe_exec_queue *q;
	unsigned long index;
	int err = 0;

	mutex_lock(&guc->submission_state.lock);
	xa_for_each(&guc->submission_state.exec_queue_lookup, index, q) {
		/* Prevent redundant attempts to stop parallel queues */
		if (q->guc->id != index)
			continue;

		err = xe_exec_queue_contexts_hwsp_rebase(q, scratch);
		if (err)
			break;
	}
	mutex_unlock(&guc->submission_state.lock);

	return err;
}