public inbox for intel-xe@lists.freedesktop.org
 help / color / mirror / Atom feed
From: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
To: Matt Roper <matthew.d.roper@intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>,
	<intel-xe@lists.freedesktop.org>, <matthew.brost@intel.com>
Subject: Re: [PATCH v6 15/17] drm/xe/multi_queue: Support active group after primary is destroyed
Date: Tue, 30 Dec 2025 17:51:25 -0800	[thread overview]
Message-ID: <aVSBneeXYMPn7FJZ@nvishwa1-desk> (raw)
In-Reply-To: <20251219225333.GQ4164497@mdroper-desk1.amr.corp.intel.com>

On Fri, Dec 19, 2025 at 02:53:33PM -0800, Matt Roper wrote:
>On Fri, Dec 19, 2025 at 02:35:53PM -0800, Niranjana Vishwanathapura wrote:
>> On Fri, Dec 19, 2025 at 04:06:17PM -0500, Rodrigo Vivi wrote:
>> > On Wed, Dec 10, 2025 at 05:03:03PM -0800, Niranjana Vishwanathapura wrote:
>> > > Add support to keep the group active after the primary queue is
>> > > destroyed. Instead of killing the primary queue during exec_queue
>> > > destroy ioctl, kill it when all the secondary queues of the group
>> > > are killed.
>> > >
>> > > Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
>> > > Reviewed-by: Matthew Brost <matthew.brost@intel.com>
>> > > ---
>> > >  drivers/gpu/drm/xe/xe_device.c           |  7 ++-
>> > >  drivers/gpu/drm/xe/xe_exec_queue.c       | 55 +++++++++++++++++++++++-
>> > >  drivers/gpu/drm/xe/xe_exec_queue.h       |  2 +
>> > >  drivers/gpu/drm/xe/xe_exec_queue_types.h |  4 ++
>> > >  include/uapi/drm/xe_drm.h                |  4 ++
>> >
>> > Hi Niranjana,
>> >
>> > Where is the UMD ack for this? Who is using this uAPI?
>>
>> https://lists.freedesktop.org/archives/intel-xe/2025-November/105779.html
>
>I think Rodrigo's question was more about the specific
>DRM_XE_MULTI_GROUP_KEEP_ACTIVE flag added in this patch.  It doesn't
>appear that the compute-runtime PR linked in the cover letter makes use
>of that flag at all.  Is there a second PR that adds usage of that?
>

Ok, after speaking to the compute team it appears that it is not
a must have feature. Let me send out patch to revert this feature.
We can always add it back later if the requirement arises.

Thanks,
Niranjana

>
>Matt
>
>>
>> Niranjana
>>
>> >
>> > Thanks,
>> > Rodrigo.
>> >
>> > >  5 files changed, 69 insertions(+), 3 deletions(-)
>> > >
>> > > diff --git a/drivers/gpu/drm/xe/xe_device.c b/drivers/gpu/drm/xe/xe_device.c
>> > > index 7a498c8db7b1..24efb6a3e0ea 100644
>> > > --- a/drivers/gpu/drm/xe/xe_device.c
>> > > +++ b/drivers/gpu/drm/xe/xe_device.c
>> > > @@ -177,7 +177,12 @@ static void xe_file_close(struct drm_device *dev, struct drm_file *file)
>> > >  	xa_for_each(&xef->exec_queue.xa, idx, q) {
>> > >  		if (q->vm && q->hwe->hw_engine_group)
>> > >  			xe_hw_engine_group_del_exec_queue(q->hwe->hw_engine_group, q);
>> > > -		xe_exec_queue_kill(q);
>> > > +
>> > > +		if (xe_exec_queue_is_multi_queue_primary(q))
>> > > +			xe_exec_queue_group_kill_put(q->multi_queue.group);
>> > > +		else
>> > > +			xe_exec_queue_kill(q);
>> > > +
>> > >  		xe_exec_queue_put(q);
>> > >  	}
>> > >  	xa_for_each(&xef->vm.xa, idx, vm)
>> > > diff --git a/drivers/gpu/drm/xe/xe_exec_queue.c b/drivers/gpu/drm/xe/xe_exec_queue.c
>> > > index d337b7bc2b80..3f4840d135a0 100644
>> > > --- a/drivers/gpu/drm/xe/xe_exec_queue.c
>> > > +++ b/drivers/gpu/drm/xe/xe_exec_queue.c
>> > > @@ -418,6 +418,26 @@ struct xe_exec_queue *xe_exec_queue_create_bind(struct xe_device *xe,
>> > >  }
>> > >  ALLOW_ERROR_INJECTION(xe_exec_queue_create_bind, ERRNO);
>> > >
>> > > +static void xe_exec_queue_group_kill(struct kref *ref)
>> > > +{
>> > > +	struct xe_exec_queue_group *group = container_of(ref, struct xe_exec_queue_group,
>> > > +							 kill_refcount);
>> > > +	xe_exec_queue_kill(group->primary);
>> > > +}
>> > > +
>> > > +static inline void xe_exec_queue_group_kill_get(struct xe_exec_queue_group *group)
>> > > +{
>> > > +	kref_get(&group->kill_refcount);
>> > > +}
>> > > +
>> > > +void xe_exec_queue_group_kill_put(struct xe_exec_queue_group *group)
>> > > +{
>> > > +	if (!group)
>> > > +		return;
>> > > +
>> > > +	kref_put(&group->kill_refcount, xe_exec_queue_group_kill);
>> > > +}
>> > > +
>> > >  void xe_exec_queue_destroy(struct kref *ref)
>> > >  {
>> > >  	struct xe_exec_queue *q = container_of(ref, struct xe_exec_queue, refcount);
>> > > @@ -650,6 +670,7 @@ static int xe_exec_queue_group_init(struct xe_device *xe, struct xe_exec_queue *
>> > >  	group->primary = q;
>> > >  	group->cgp_bo = bo;
>> > >  	INIT_LIST_HEAD(&group->list);
>> > > +	kref_init(&group->kill_refcount);
>> > >  	xa_init_flags(&group->xa, XA_FLAGS_ALLOC1);
>> > >  	mutex_init(&group->list_lock);
>> > >  	q->multi_queue.group = group;
>> > > @@ -725,6 +746,11 @@ static int xe_exec_queue_group_add(struct xe_device *xe, struct xe_exec_queue *q
>> > >
>> > >  	q->multi_queue.pos = pos;
>> > >
>> > > +	if (group->primary->multi_queue.keep_active) {
>> > > +		xe_exec_queue_group_kill_get(group);
>> > > +		q->multi_queue.keep_active = true;
>> > > +	}
>> > > +
>> > >  	return 0;
>> > >  }
>> > >
>> > > @@ -738,6 +764,11 @@ static void xe_exec_queue_group_delete(struct xe_device *xe, struct xe_exec_queu
>> > >  	lrc = xa_erase(&group->xa, q->multi_queue.pos);
>> > >  	xe_assert(xe, lrc);
>> > >  	xe_lrc_put(lrc);
>> > > +
>> > > +	if (q->multi_queue.keep_active) {
>> > > +		xe_exec_queue_group_kill_put(group);
>> > > +		q->multi_queue.keep_active = false;
>> > > +	}
>> > >  }
>> > >
>> > >  static int exec_queue_set_multi_group(struct xe_device *xe, struct xe_exec_queue *q,
>> > > @@ -759,12 +790,24 @@ static int exec_queue_set_multi_group(struct xe_device *xe, struct xe_exec_queue
>> > >  		return -EINVAL;
>> > >
>> > >  	if (value & DRM_XE_MULTI_GROUP_CREATE) {
>> > > -		if (XE_IOCTL_DBG(xe, value & ~DRM_XE_MULTI_GROUP_CREATE))
>> > > +		if (XE_IOCTL_DBG(xe, value & ~(DRM_XE_MULTI_GROUP_CREATE |
>> > > +					       DRM_XE_MULTI_GROUP_KEEP_ACTIVE)))
>> > > +			return -EINVAL;
>> > > +
>> > > +		/*
>> > > +		 * KEEP_ACTIVE is not supported in preempt fence mode as in that mode,
>> > > +		 * VM_DESTROY ioctl expects all exec queues of that VM are already killed.
>> > > +		 */
>> > > +		if (XE_IOCTL_DBG(xe, (value & DRM_XE_MULTI_GROUP_KEEP_ACTIVE) &&
>> > > +				 xe_vm_in_preempt_fence_mode(q->vm)))
>> > >  			return -EINVAL;
>> > >
>> > >  		q->multi_queue.valid = true;
>> > >  		q->multi_queue.is_primary = true;
>> > >  		q->multi_queue.pos = 0;
>> > > +		if (value & DRM_XE_MULTI_GROUP_KEEP_ACTIVE)
>> > > +			q->multi_queue.keep_active = true;
>> > > +
>> > >  		return 0;
>> > >  	}
>> > >
>> > > @@ -1312,6 +1355,11 @@ void xe_exec_queue_kill(struct xe_exec_queue *q)
>> > >
>> > >  	q->ops->kill(q);
>> > >  	xe_vm_remove_compute_exec_queue(q->vm, q);
>> > > +
>> > > +	if (!xe_exec_queue_is_multi_queue_primary(q) && q->multi_queue.keep_active) {
>> > > +		xe_exec_queue_group_kill_put(q->multi_queue.group);
>> > > +		q->multi_queue.keep_active = false;
>> > > +	}
>> > >  }
>> > >
>> > >  int xe_exec_queue_destroy_ioctl(struct drm_device *dev, void *data,
>> > > @@ -1338,7 +1386,10 @@ int xe_exec_queue_destroy_ioctl(struct drm_device *dev, void *data,
>> > >  	if (q->vm && q->hwe->hw_engine_group)
>> > >  		xe_hw_engine_group_del_exec_queue(q->hwe->hw_engine_group, q);
>> > >
>> > > -	xe_exec_queue_kill(q);
>> > > +	if (xe_exec_queue_is_multi_queue_primary(q))
>> > > +		xe_exec_queue_group_kill_put(q->multi_queue.group);
>> > > +	else
>> > > +		xe_exec_queue_kill(q);
>> > >
>> > >  	trace_xe_exec_queue_close(q);
>> > >  	xe_exec_queue_put(q);
>> > > diff --git a/drivers/gpu/drm/xe/xe_exec_queue.h b/drivers/gpu/drm/xe/xe_exec_queue.h
>> > > index ffcc1feb879e..10abed98fb6b 100644
>> > > --- a/drivers/gpu/drm/xe/xe_exec_queue.h
>> > > +++ b/drivers/gpu/drm/xe/xe_exec_queue.h
>> > > @@ -113,6 +113,8 @@ static inline struct xe_exec_queue *xe_exec_queue_multi_queue_primary(struct xe_
>> > >  	return xe_exec_queue_is_multi_queue(q) ? q->multi_queue.group->primary : q;
>> > >  }
>> > >
>> > > +void xe_exec_queue_group_kill_put(struct xe_exec_queue_group *group);
>> > > +
>> > >  bool xe_exec_queue_is_lr(struct xe_exec_queue *q);
>> > >
>> > >  bool xe_exec_queue_is_idle(struct xe_exec_queue *q);
>> > > diff --git a/drivers/gpu/drm/xe/xe_exec_queue_types.h b/drivers/gpu/drm/xe/xe_exec_queue_types.h
>> > > index 5fc516b0bb77..67ea5eebf70b 100644
>> > > --- a/drivers/gpu/drm/xe/xe_exec_queue_types.h
>> > > +++ b/drivers/gpu/drm/xe/xe_exec_queue_types.h
>> > > @@ -62,6 +62,8 @@ struct xe_exec_queue_group {
>> > >  	struct list_head list;
>> > >  	/** @list_lock: Secondary queue list lock */
>> > >  	struct mutex list_lock;
>> > > +	/** @kill_refcount: ref count to kill primary queue */
>> > > +	struct kref kill_refcount;
>> > >  	/** @sync_pending: CGP_SYNC_DONE g2h response pending */
>> > >  	bool sync_pending;
>> > >  	/** @banned: Group banned */
>> > > @@ -161,6 +163,8 @@ struct xe_exec_queue {
>> > >  		u8 valid:1;
>> > >  		/** @multi_queue.is_primary: Is primary queue (Q0) of the group */
>> > >  		u8 is_primary:1;
>> > > +		/** @multi_queue.keep_active: Keep the group active after primary is destroyed */
>> > > +		u8 keep_active:1;
>> > >  	} multi_queue;
>> > >
>> > >  	/** @sched_props: scheduling properties */
>> > > diff --git a/include/uapi/drm/xe_drm.h b/include/uapi/drm/xe_drm.h
>> > > index 705081bf0d81..bd6154e3b728 100644
>> > > --- a/include/uapi/drm/xe_drm.h
>> > > +++ b/include/uapi/drm/xe_drm.h
>> > > @@ -1280,6 +1280,9 @@ struct drm_xe_vm_bind {
>> > >   *    then a new multi-queue group is created with this queue as the primary queue
>> > >   *    (Q0). Otherwise, the queue gets added to the multi-queue group whose primary
>> > >   *    queue's exec_queue_id is specified in the lower 32 bits of the 'value' field.
>> > > + *    If the extension's 'value' field has %DRM_XE_MULTI_GROUP_KEEP_ACTIVE flag
>> > > + *    set, then the multi-queue group is kept active after the primary queue is
>> > > + *    destroyed.
>> > >   *    All the other non-relevant bits of extension's 'value' field while adding the
>> > >   *    primary or the secondary queues of the group must be set to 0.
>> > >   *  - %DRM_XE_EXEC_QUEUE_SET_PROPERTY_MULTI_QUEUE_PRIORITY - Set the queue
>> > > @@ -1328,6 +1331,7 @@ struct drm_xe_exec_queue_create {
>> > >  #define   DRM_XE_EXEC_QUEUE_SET_HANG_REPLAY_STATE		3
>> > >  #define   DRM_XE_EXEC_QUEUE_SET_PROPERTY_MULTI_GROUP		4
>> > >  #define     DRM_XE_MULTI_GROUP_CREATE				(1ull << 63)
>> > > +#define     DRM_XE_MULTI_GROUP_KEEP_ACTIVE			(1ull << 62)
>> > >  #define   DRM_XE_EXEC_QUEUE_SET_PROPERTY_MULTI_QUEUE_PRIORITY	5
>> > >  	/** @extensions: Pointer to the first extension struct, if any */
>> > >  	__u64 extensions;
>> > > --
>> > > 2.43.0
>> > >
>
>-- 
>Matt Roper
>Graphics Software Engineer
>Linux GPU Platform Enablement
>Intel Corporation

  reply	other threads:[~2025-12-31  1:51 UTC|newest]

Thread overview: 32+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2025-12-11  1:02 [PATCH v6 00/17] drm/xe: Multi Queue feature support Niranjana Vishwanathapura
2025-12-11  1:02 ` [PATCH v6 01/17] drm/xe/multi_queue: Add multi_queue_enable_mask to gt information Niranjana Vishwanathapura
2025-12-11  1:02 ` [PATCH v6 02/17] drm/xe/multi_queue: Add user interface for multi queue support Niranjana Vishwanathapura
2025-12-11  1:02 ` [PATCH v6 03/17] drm/xe/multi_queue: Add GuC " Niranjana Vishwanathapura
2025-12-11  1:02 ` [PATCH v6 04/17] drm/xe/multi_queue: Add multi queue priority property Niranjana Vishwanathapura
2025-12-11  1:02 ` [PATCH v6 05/17] drm/xe/multi_queue: Handle invalid exec queue property setting Niranjana Vishwanathapura
2025-12-11  1:02 ` [PATCH v6 06/17] drm/xe/multi_queue: Add exec_queue set_property ioctl support Niranjana Vishwanathapura
2026-01-19 16:57   ` Thomas Hellström
2026-01-20 21:06     ` Niranjana Vishwanathapura
2026-01-20 22:20       ` Matthew Brost
2026-01-20 22:30         ` Matthew Brost
2026-01-21 15:29         ` Thomas Hellström
2026-01-21 21:23           ` Niranjana Vishwanathapura
2025-12-11  1:02 ` [PATCH v6 07/17] drm/xe/multi_queue: Add support for multi queue dynamic priority change Niranjana Vishwanathapura
2025-12-11  1:02 ` [PATCH v6 08/17] drm/xe/multi_queue: Add multi queue information to guc_info dump Niranjana Vishwanathapura
2025-12-11  1:02 ` [PATCH v6 09/17] drm/xe/multi_queue: Handle tearing down of a multi queue Niranjana Vishwanathapura
2025-12-11  1:02 ` [PATCH v6 10/17] drm/xe/multi_queue: Set QUEUE_DRAIN_MODE for Multi Queue batches Niranjana Vishwanathapura
2025-12-11  1:02 ` [PATCH v6 11/17] drm/xe/multi_queue: Handle CGP context error Niranjana Vishwanathapura
2025-12-11  1:03 ` [PATCH v6 12/17] drm/xe/multi_queue: Reset GT upon CGP_SYNC failure Niranjana Vishwanathapura
2025-12-11  1:03 ` [PATCH v6 13/17] drm/xe/multi_queue: Teardown group upon job timeout Niranjana Vishwanathapura
2025-12-11  1:03 ` [PATCH v6 14/17] drm/xe/multi_queue: Tracepoint support Niranjana Vishwanathapura
2025-12-11  1:03 ` [PATCH v6 15/17] drm/xe/multi_queue: Support active group after primary is destroyed Niranjana Vishwanathapura
2025-12-19 21:06   ` Rodrigo Vivi
2025-12-19 22:35     ` Niranjana Vishwanathapura
2025-12-19 22:53       ` Matt Roper
2025-12-31  1:51         ` Niranjana Vishwanathapura [this message]
2025-12-11  1:03 ` [PATCH v6 16/17] drm/xe/doc: Add documentation for Multi Queue Group Niranjana Vishwanathapura
2025-12-11  1:03 ` [PATCH v6 17/17] drm/xe/doc: Add documentation for Multi Queue Group GuC interface Niranjana Vishwanathapura
2025-12-11  1:11 ` ✗ CI.checkpatch: warning for drm/xe: Multi Queue feature support (rev6) Patchwork
2025-12-11  1:12 ` ✓ CI.KUnit: success " Patchwork
2025-12-11  2:25 ` ✓ Xe.CI.BAT: " Patchwork
2025-12-11 10:04 ` ✗ 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=aVSBneeXYMPn7FJZ@nvishwa1-desk \
    --to=niranjana.vishwanathapura@intel.com \
    --cc=intel-xe@lists.freedesktop.org \
    --cc=matthew.brost@intel.com \
    --cc=matthew.d.roper@intel.com \
    --cc=rodrigo.vivi@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