public inbox for intel-xe@lists.freedesktop.org
 help / color / mirror / Atom feed
From: Matthew Brost <matthew.brost@intel.com>
To: Tomasz Lis <tomasz.lis@intel.com>
Cc: intel-xe@lists.freedesktop.org,
	"Michał Winiarski" <michal.winiarski@intel.com>,
	"Michał Wajdeczko" <michal.wajdeczko@intel.com>,
	"Piotr Piórkowski" <piotr.piorkowski@intel.com>,
	"Lucas De Marchi" <lucas.demarchi@intel.com>
Subject: Re: [PATCH v1 3/4] drm/xe/vf: Wait for default LRCs fixups before using
Date: Fri, 6 Feb 2026 10:11:59 -0800	[thread overview]
Message-ID: <aYYu7wFpES1UXoWm@lstrano-desk.jf.intel.com> (raw)
In-Reply-To: <20260206145334.674679-4-tomasz.lis@intel.com>

On Fri, Feb 06, 2026 at 03:53:33PM +0100, Tomasz Lis wrote:
> When a context is being created during save/restore, the LRC creation
> needs to wait for GGTT address space to be shifted. But it also needs
> to have fixed default LRCs. This is mandatory to avoid the situation
> where LRC will be created based on data from before the fixups, but
> reference within exec queue will be set too late for fixups.
> 
> This fixes an issue where contexts created during save/restore have
> a large chance of having one unfixed LRC, due to the xe_lrc_create()
> being synced for equal start to race with default LRC fixups.
> 

This indeed looks like a potential problem.

This whole flow is a bit weak (my fault — I meant to fix this but lost
track of it). Maybe we should make this a bit more robust and ordered by
the GuC backend.

Roughly:

- In guc_exec_queue_unpause_prepare, if at least one job has been
  submitted, we fix the LRCs on the queue. This would happen prior to
  emit_job. I believe the 'at least one job has been submited check' is
  exec_queue_registered.

- In guc_exec_queue_run_job, upon running the first job (hook in
  register_exec_queue) and if VF migration is enabled, we fix the LRC.

What do you think? I believe this would work, provide stronger ordering
guarantees, and let us drop the READ_ONCE/WRITE_ONCE and wait‑queue
semantics, which are questionable at best. We could probably revert the
flipping of q->ops->init and the LRC creation order back to the original
flow as well.

Matt

> Signed-off-by: Tomasz Lis <tomasz.lis@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_exec_queue.c        |  2 +-
>  drivers/gpu/drm/xe/xe_gt_sriov_vf.c       | 24 +++++++++++------------
>  drivers/gpu/drm/xe/xe_gt_sriov_vf.h       |  2 +-
>  drivers/gpu/drm/xe/xe_gt_sriov_vf_types.h |  4 ++--
>  4 files changed, 15 insertions(+), 17 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_exec_queue.c b/drivers/gpu/drm/xe/xe_exec_queue.c
> index e9396ad3390a..6eb561086e1c 100644
> --- a/drivers/gpu/drm/xe/xe_exec_queue.c
> +++ b/drivers/gpu/drm/xe/xe_exec_queue.c
> @@ -309,7 +309,7 @@ static int __xe_exec_queue_init(struct xe_exec_queue *q, u32 exec_queue_flags)
>  	for (i = 0; i < q->width; ++i) {
>  		struct xe_lrc *lrc;
>  
> -		xe_gt_sriov_vf_wait_valid_ggtt(q->gt);
> +		xe_gt_sriov_vf_wait_valid_default_lrc(q->gt);
>  		lrc = xe_lrc_create(q->hwe, q->vm, q->replay_state,
>  				    xe_lrc_ring_size(), q->msix_vec, flags);
>  		if (IS_ERR(lrc)) {
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_vf.c b/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
> index 30e8c2cf5f09..1edccee84c76 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_vf.c
> @@ -529,12 +529,6 @@ static int vf_get_ggtt_info(struct xe_gt *gt)
>  		xe_tile_sriov_vf_fixup_ggtt_nodes_locked(gt_to_tile(gt), shift);
>  	}
>  
> -	if (xe_sriov_vf_migration_supported(gt_to_xe(gt))) {
> -		WRITE_ONCE(gt->sriov.vf.migration.ggtt_need_fixes, false);
> -		smp_wmb();	/* Ensure above write visible before wake */
> -		wake_up_all(&gt->sriov.vf.migration.wq);
> -	}
> -
>  	return 0;
>  }
>  
> @@ -837,6 +831,10 @@ static void xe_gt_sriov_vf_default_lrcs_hwsp_rebase(struct xe_gt *gt)
>  
>  	for_each_hw_engine(hwe, gt, id)
>  		xe_default_lrc_update_memirq_regs_with_address(hwe);
> +
> +	WRITE_ONCE(gt->sriov.vf.migration.default_lrcs_need_fixes, false);
> +	smp_wmb();	/* Ensure above write visible before wake */
> +	wake_up_all(&gt->sriov.vf.migration.wq);
>  }
>  
>  static void vf_start_migration_recovery(struct xe_gt *gt)
> @@ -851,7 +849,7 @@ static void vf_start_migration_recovery(struct xe_gt *gt)
>  	    !gt->sriov.vf.migration.recovery_teardown) {
>  		gt->sriov.vf.migration.recovery_queued = true;
>  		WRITE_ONCE(gt->sriov.vf.migration.recovery_inprogress, true);
> -		WRITE_ONCE(gt->sriov.vf.migration.ggtt_need_fixes, true);
> +		WRITE_ONCE(gt->sriov.vf.migration.default_lrcs_need_fixes, true);
>  		smp_wmb();	/* Ensure above writes visible before wake */
>  
>  		xe_guc_ct_wake_waiters(&gt->uc.guc.ct);
> @@ -1296,7 +1294,7 @@ static void vf_post_migration_abort(struct xe_gt *gt)
>  {
>  	spin_lock_irq(&gt->sriov.vf.migration.lock);
>  	WRITE_ONCE(gt->sriov.vf.migration.recovery_inprogress, false);
> -	WRITE_ONCE(gt->sriov.vf.migration.ggtt_need_fixes, false);
> +	WRITE_ONCE(gt->sriov.vf.migration.default_lrcs_need_fixes, false);
>  	spin_unlock_irq(&gt->sriov.vf.migration.lock);
>  
>  	wake_up_all(&gt->sriov.vf.migration.wq);
> @@ -1492,7 +1490,7 @@ bool xe_gt_sriov_vf_recovery_pending(struct xe_gt *gt)
>  	return READ_ONCE(gt->sriov.vf.migration.recovery_inprogress);
>  }
>  
> -static bool vf_valid_ggtt(struct xe_gt *gt)
> +static bool vf_valid_default_lrc(struct xe_gt *gt)
>  {
>  	struct xe_memirq *memirq = &gt_to_tile(gt)->memirq;
>  	bool irq_pending = xe_device_uses_memirq(gt_to_xe(gt)) &&
> @@ -1500,17 +1498,17 @@ static bool vf_valid_ggtt(struct xe_gt *gt)
>  
>  	xe_gt_assert(gt, IS_SRIOV_VF(gt_to_xe(gt)));
>  
> -	if (irq_pending || READ_ONCE(gt->sriov.vf.migration.ggtt_need_fixes))
> +	if (irq_pending || READ_ONCE(gt->sriov.vf.migration.default_lrcs_need_fixes))
>  		return false;
>  
>  	return true;
>  }
>  
>  /**
> - * xe_gt_sriov_vf_wait_valid_ggtt() - VF wait for valid GGTT addresses
> + * xe_gt_sriov_vf_wait_valid_default_lrc() - wait for valid GGTT refs in default LRCs
>   * @gt: the &xe_gt
>   */
> -void xe_gt_sriov_vf_wait_valid_ggtt(struct xe_gt *gt)
> +void xe_gt_sriov_vf_wait_valid_default_lrc(struct xe_gt *gt)
>  {
>  	int ret;
>  
> @@ -1519,7 +1517,7 @@ void xe_gt_sriov_vf_wait_valid_ggtt(struct xe_gt *gt)
>  		return;
>  
>  	ret = wait_event_interruptible_timeout(gt->sriov.vf.migration.wq,
> -					       vf_valid_ggtt(gt),
> +					       vf_valid_default_lrc(gt),
>  					       HZ * 5);
>  	xe_gt_WARN_ON(gt, !ret);
>  }
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_vf.h b/drivers/gpu/drm/xe/xe_gt_sriov_vf.h
> index 7d97189c2d3d..70232dc38f9a 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_vf.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_vf.h
> @@ -39,6 +39,6 @@ void xe_gt_sriov_vf_print_config(struct xe_gt *gt, struct drm_printer *p);
>  void xe_gt_sriov_vf_print_runtime(struct xe_gt *gt, struct drm_printer *p);
>  void xe_gt_sriov_vf_print_version(struct xe_gt *gt, struct drm_printer *p);
>  
> -void xe_gt_sriov_vf_wait_valid_ggtt(struct xe_gt *gt);
> +void xe_gt_sriov_vf_wait_valid_default_lrc(struct xe_gt *gt);
>  
>  #endif
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_vf_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_vf_types.h
> index 4ef881b9b662..8be181bf3cf3 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_vf_types.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_vf_types.h
> @@ -73,8 +73,8 @@ struct xe_gt_sriov_vf_migration {
>  	bool recovery_queued;
>  	/** @recovery_inprogress: VF post migration recovery in progress */
>  	bool recovery_inprogress;
> -	/** @ggtt_need_fixes: VF GGTT needs fixes */
> -	bool ggtt_need_fixes;
> +	/** @default_lrcs_need_fixes: GGTT refs within default LRCs need fixes */
> +	bool default_lrcs_need_fixes;
>  };
>  
>  /**
> -- 
> 2.25.1
> 

  reply	other threads:[~2026-02-06 18:12 UTC|newest]

Thread overview: 14+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-02-06 14:53 [PATCH v1 0/4] drm/xe/vf: Fix exec queue creation during post-migration recovery Tomasz Lis
2026-02-06 14:53 ` [PATCH v1 1/4] drm/xe/queue: Call fini on exec queue creation fail Tomasz Lis
2026-02-06 17:38   ` Matthew Brost
2026-02-06 14:53 ` [PATCH v1 2/4] drm/xe/vf: Avoid LRC being freed while applying fixups Tomasz Lis
2026-02-06 17:46   ` Matthew Brost
2026-02-10 20:16     ` Lis, Tomasz
2026-02-06 14:53 ` [PATCH v1 3/4] drm/xe/vf: Wait for default LRCs fixups before using Tomasz Lis
2026-02-06 18:11   ` Matthew Brost [this message]
2026-02-10 20:11     ` Lis, Tomasz
2026-02-18 23:15       ` Lis, Tomasz
2026-02-06 14:53 ` [PATCH v1 4/4] drm/xe/vf: Redo LRC creation while in VF fixups Tomasz Lis
2026-02-06 14:56 ` ✓ CI.KUnit: success for drm/xe/vf: Fix exec queue creation during post-migration recovery Patchwork
2026-02-06 15:29 ` ✓ Xe.CI.BAT: " Patchwork
2026-02-07 15:42 ` ✗ Xe.CI.FULL: failure " Patchwork

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=aYYu7wFpES1UXoWm@lstrano-desk.jf.intel.com \
    --to=matthew.brost@intel.com \
    --cc=intel-xe@lists.freedesktop.org \
    --cc=lucas.demarchi@intel.com \
    --cc=michal.wajdeczko@intel.com \
    --cc=michal.winiarski@intel.com \
    --cc=piotr.piorkowski@intel.com \
    --cc=tomasz.lis@intel.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox