All of lore.kernel.org
 help / color / mirror / Atom feed
From: sashiko-bot@kernel.org
To: "FNU VISHWANATHA" <fnu.vishwanatha@intel.com>,
	intel-xe@lists.freedesktop.org
Subject: Re: [PATCH] drm/xe/pf: Restore PF config before resuming submission after GT reset
Date: Thu, 19 Nov 2026 00:02:25 +0000	[thread overview]
Message-ID: <sashiko-outbox-222221@kernel.org> (raw)
In-Reply-To: <20261118233946.275114-1-fnu.vishwanatha@intel.com>

Thank you for your contribution! Sashiko AI review found 2 potential
issues to consider:
- [High] drm/xe/pf: dma_fence deadlock and reclaim hazard in
  do_gt_restart()
- [Medium] drm/xe/pf: do_gt_restart() restores PF/VF config before GuC
  VGT policies
It also noticed 1 related pre-existing issue this patch does not
introduce:
- [High] drm/xe: loss of RING_IDLEDLY workarounds in adjust_idledly()
  across GT reset and resume [1]
--

commit 19aac698e33fa931358e6f2c835ac921dba22d3f
Author: FNU VISHWANATHA <fnu.vishwanatha@intel.com>

drm/xe/pf: Restore PF config before resuming submission after GT reset

This commit calls xe_gt_sriov_pf_config_restart() synchronously in
do_gt_restart() before xe_uc_start() resumes GuC submission after a GT
reset.

> diff --git a/drivers/gpu/drm/xe/xe_gt.c b/drivers/gpu/drm/xe/xe_gt.c
> index b751b5044f3ca..f0539440e7702 100644
> --- a/drivers/gpu/drm/xe/xe_gt.c
> +++ b/drivers/gpu/drm/xe/xe_gt.c
[ ... ]
> @@ -891,6 +892,18 @@ static int do_gt_restart(struct xe_gt *gt)
>  	/* Get CCS mode in sync between sw/hw */
>  	xe_gt_apply_ccs_mode(gt);
>  
> +	/*
> +	 * The GuC was reloaded and lost the PF configuration pushed to it,
> +	 * including the scheduling parameters. Restore it before
> +	 * xe_uc_start() resumes submission: exec queues re-registered with
> +	 * the GuC under its default configuration may be accepted but never
> +	 * dispatched, leaving their jobs unstarted and their fences unsignaled.
> +	 * The asynchronous PF restart below pushes the configuration again,
> +	 * which is harmless.
> +	 */
> +	if (IS_SRIOV_PF(gt_to_xe(gt)))
> +		xe_gt_sriov_pf_config_restart(gt);
> +
>  	err = xe_uc_start(&gt->uc);
>  	if (err)
>  		return err;

[Severity: High]
Can calling xe_gt_sriov_pf_config_restart() synchronously in do_gt_restart()
deadlock on xe->sriov.pf.master_lock and violate dma_fence reclaim rules?

Commit a4d1c5d0b99b ("drm/xe/pf: Move VFs reprovisioning to worker") moved
xe_gt_sriov_pf_config_restart() to an asynchronous worker because it takes
xe_gt_sriov_pf_master_mutex(gt), which is held across GFP_KERNEL allocations
and GPU migration dma_fence waits.  During a GT reset, xe_uc_stop() stops
GuC submission without signaling pending jobs on kernel or VM queues such as
tile->migrate->q; those jobs only complete after do_gt_restart() reaches
xe_uc_start().  Meanwhile, xe_gt_sriov_pf_stop_prepare() cancels only the PF
restart worker and leaves other master_lock holders running:

CPU0 (VF provisioning / FLR)      CPU1 (gt_reset_worker)
----------------------------      ----------------------
pf_provision_vf_lmem() /
xe_gt_sriov_pf_config_sanitize()
  mutex_lock(master_lock)
                                  gt_reset_worker()
                                    xe_uc_stop()
  xe_migrate_clear()
  dma_fence_wait(...)
                                    do_gt_restart()
                                      xe_gt_sriov_pf_config_restart()
                                        mutex_lock(master_lock)
                                      ...
                                      xe_uc_start()  <-- never reached

 *** DEADLOCK ***

In pf_provision_vf_lmem() (via TTM BO creation and drm_suballoc_new()) and
in pf_sanitize_lmem() when suballocating batch buffers for large VF LMEM
clears, the wait under master_lock is uninterruptible with
MAX_SCHEDULE_TIMEOUT, wedging gt_reset_worker() on gt->ordered_wq.

[Severity: Medium]
Does calling xe_gt_sriov_pf_config_restart() and xe_uc_start() here before
xe_gt_sriov_pf_policy_restart() runs in the async pf_restart() worker invert
the policy-before-config restore order?

When the GuC is reloaded in xe_uc_load_hw(), VGT policies reset to hardware
defaults (sched_if_idle disabled, default priority GUC_SCHED_PRIORITY_LOW,
and scheduler groups disabled).  However, encode_config_sched() consults the
software policy state when encoding PF and VF configs:

drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c:custom_sched_priority() {
	return xe_gt_sriov_pf_policy_get_sched_if_idle_locked(gt) ?
		priority != GUC_SCHED_PRIORITY_NORMAL :
		priority != GUC_SCHED_PRIORITY_LOW;
}

If sched_if_idle is enabled in software and a function uses
GUC_SCHED_PRIORITY_NORMAL, custom_sched_priority() returns false and omits
GUC_KLV_VF_CFG_SCHED_PRIORITY_KEY, assuming GUC_KLV_VGT_POLICY_SCHED_IF_IDLE_KEY
was already sent to the GuC.  Conversely, if a function has a custom
priority while sched_if_idle is enabled, the later async
xe_gt_sriov_pf_policy_restart() sends SCHED_IF_IDLE=1 and resets the GuC's
priority back to GUC_SCHED_PRIORITY_NORMAL while submission is active.
In addition, xe_uc_start() re-registers contexts before
pf_reprovision_sched_groups() runs, and as noted in
pf_provision_sched_groups(), the GuC silently ignores scheduler group
provisioning if any MLRC contexts are already registered.

[Severity: High]
This problem wasn't introduced by this patch, but in do_gt_restart() the
engine register restoration loop just above only reapplies hwe->reg_sr via
xe_reg_sr_apply_mmio() and does not call adjust_idledly() [1].

During initial probe, hw_engine_init() calls adjust_idledly() to apply
Wa_16023105232 and Wa_14025941587 directly via xe_mmio_write32() on
RING_IDLEDLY without recording the change in hwe->reg_sr.  Across a full GT
reset or system resume, RING_IDLEDLY reverts to hardware defaults and the
GuC ADS save/restore list is repopulated in xe_uc_load_hw(), so the bounded
IDLE_DELAY and cleared INHIBIT_SWITCH_UNTIL_PREEMPTED bit are lost, which
can cause preemption failures or GPU hangs.

[1] https://sashiko.dev/#/bug/linux-4061257d-241c-4f91-b214-9c272d968ab3

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261118233946.275114-1-fnu.vishwanatha@intel.com?part=1

  parent reply	other threads:[~2026-11-19  0:02 UTC|newest]

Thread overview: 9+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-18 23:39 [PATCH] drm/xe/pf: Restore PF config before resuming submission after GT reset FNU VISHWANATHA
2026-11-18 23:56 ` FNU VISHWANATHA
2026-11-19  1:23   ` Ceraolo Spurio, Daniele
2026-11-19 10:15     ` Piotr Piórkowski
2026-11-19 21:27       ` FNU VISHWANATHA
2026-11-19  0:02 ` sashiko-bot [this message]
2026-11-19  5:18 ` ✓ CI.KUnit: success for " Patchwork
2026-11-19  5:56 ` ✓ Xe.CI.BAT: " Patchwork
2026-11-20  7:18 ` ✓ Xe.CI.FULL: " 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=sashiko-outbox-222221@kernel.org \
    --to=sashiko-bot@kernel.org \
    --cc=fnu.vishwanatha@intel.com \
    --cc=intel-xe@lists.freedesktop.org \
    --cc=sashiko-reviews@lists.linux.dev \
    /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 an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.