All of lore.kernel.org
 help / color / mirror / Atom feed
From: Zhu Lingshan <lingshan.zhu@amd.com>
To: <Alexander.Deucher@amd.com>, <Christian.Koenig@amd.com>,
	<felix.kuehling@amd.com>
Cc: <Ray.Huang@amd.com>, <amd-gfx@lists.freedesktop.org>,
	Zhu Lingshan <lingshan.zhu@amd.com>
Subject: [PATCH 02/10] drm/amdgpu: keep the userq manager alive as long as its queues
Date: Fri, 28 Aug 2026 17:53:41 +0800	[thread overview]
Message-ID: <20260828095349.9797-3-lingshan.zhu@amd.com> (raw)
In-Reply-To: <20260828095349.9797-1-lingshan.zhu@amd.com>

The life cycle of a user queue is managed by its
kref. However when destroy a userq manager,
the kref_put of its queues in amdgpu_userq_mgr_fini
may not be the last put, therefore the queues
could be still alive after the userq manager
has been destroyed, resulting in
userq->userq_mgr use-after-free issues.

This commit fixes this problem by introduce a new
counter refs representing for the number of its queues,
and only free the userq_manager when refs == 0

Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 30 +++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h |  9 +++++++
 2 files changed, 39 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
index e0639f844a8e..f398986a61a5 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
@@ -27,6 +27,7 @@
 #include <linux/pm_runtime.h>
 #include <linux/overflow.h>
 #include <drm/drm_drv.h>
+#include <linux/wait_bit.h>
 
 #include "amdgpu.h"
 #include "amdgpu_reset.h"
@@ -533,6 +534,17 @@ amdgpu_userq_get_doorbell_index(struct amdgpu_userq_mgr *uq_mgr,
 	return r;
 }
 
+static void amdgpu_userq_mgr_inc_refs(struct amdgpu_userq_mgr *uq_mgr)
+{
+	atomic_inc(&uq_mgr->refs);
+}
+
+static void amdgpu_userq_mgr_dec_refs(struct amdgpu_userq_mgr *uq_mgr)
+{
+	if (atomic_dec_and_test(&uq_mgr->refs))
+		wake_up_var(&uq_mgr->refs);
+}
+
 static int
 amdgpu_userq_destroy(struct amdgpu_userq_mgr *uq_mgr, struct amdgpu_usermode_queue *queue)
 {
@@ -594,6 +606,8 @@ static void amdgpu_userq_kref_destroy(struct kref *kref)
 	r = amdgpu_userq_destroy(uq_mgr, queue);
 	if (r)
 		drm_file_err(uq_mgr->file, "Failed to destroy usermode queue %d\n", r);
+
+	amdgpu_userq_mgr_dec_refs(uq_mgr);
 }
 
 struct amdgpu_usermode_queue *amdgpu_userq_get(struct amdgpu_userq_mgr *uq_mgr, u32 qid)
@@ -707,6 +721,7 @@ amdgpu_userq_create(struct drm_file *filp, union drm_amdgpu_userq *args)
 	queue->xcp_id = (fpriv->xcp_id != AMDGPU_XCP_NO_PARTITION) ?
 				fpriv->xcp_id : 0;
 	queue->userq_mgr = uq_mgr;
+	amdgpu_userq_mgr_inc_refs(uq_mgr);
 	INIT_DELAYED_WORK(&queue->hang_detect_work,
 			  amdgpu_userq_hang_detect_work);
 
@@ -819,6 +834,7 @@ amdgpu_userq_create(struct drm_file *filp, union drm_amdgpu_userq *args)
 free_queue:
 	trace_amdgpu_userq_create_end(queue, r);
 	kfree(queue);
+	amdgpu_userq_mgr_dec_refs(uq_mgr);
 err_pm_runtime:
 	pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
 	return r;
@@ -1331,6 +1347,7 @@ int amdgpu_userq_mgr_init(struct amdgpu_userq_mgr *userq_mgr, struct drm_file *f
 {
 	mutex_init(&userq_mgr->userq_mutex);
 	xa_init_flags(&userq_mgr->userq_xa, XA_FLAGS_ALLOC);
+	atomic_set(&userq_mgr->refs, 0);
 	userq_mgr->adev = adev;
 	userq_mgr->file = file_priv;
 	userq_mgr->proc_ctx_allocated = false;
@@ -1380,6 +1397,19 @@ void amdgpu_userq_mgr_fini(struct amdgpu_userq_mgr *userq_mgr)
 		amdgpu_userq_put(queue);
 	}
 
+	/*
+	 * The above amdgpu_userq_put() may not be the last put
+	 * of the kref of a user queue, therefore there could
+	 * be some queues still alive even when the userq manager
+	 * has been destroyed. This wait_evet() blocks
+	 * amdgpu_userq_mgr_fini(), so keep userq_mgr alive
+	 * while any queues holding it.
+	 *
+	 * This prevents queue->userq_mgr use-after-free issues.
+	 */
+	wait_var_event(&userq_mgr->refs,
+		       !atomic_read_acquire(&userq_mgr->refs));
+
 	xa_destroy(&userq_mgr->userq_xa);
 
 	/*
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h
index 8fc73862f64e..a13d8d4dd5c7 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h
@@ -126,6 +126,15 @@ struct amdgpu_userq_mgr {
 	 */
 	struct xarray			userq_xa;
 	struct mutex			userq_mutex;
+
+	/**
+	 * @refs:
+	 *
+	 * Each queue increases this counter when join this manager,
+	 * and decreases it when leave this manager.
+	 */
+	atomic_t			refs;
+
 	struct amdgpu_device		*adev;
 	struct delayed_work		resume_work;
 	struct drm_file			*file;
-- 
2.53.0


  parent reply	other threads:[~2026-08-28  9:53 UTC|newest]

Thread overview: 18+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-08-28  9:53 [PATCH 00/10] drm/amdgpu: secure userq lifecycle by its kref Zhu Lingshan
2026-08-28  9:53 ` [PATCH 01/10] drm/amdgpu: introduce amdgpu_lookup_queue_by_doorbell Zhu Lingshan
2026-08-28 13:08   ` Christian König
2026-08-28 15:59     ` Zhu, Lingshan
2026-08-28  9:53 ` Zhu Lingshan [this message]
2026-08-28 13:09   ` [PATCH 02/10] drm/amdgpu: keep the userq manager alive as long as its queues Christian König
2026-08-28 15:59     ` Zhu, Lingshan
2026-08-28 16:26       ` Christian König
2026-08-28  9:53 ` [PATCH 03/10] drm/amdgpu/gfx11: hold userq refs in private fault worker Zhu Lingshan
2026-08-28 13:11   ` Christian König
2026-08-28 15:59     ` Zhu, Lingshan
2026-08-28  9:53 ` [PATCH 04/10] drm/amdgpu/gfx12: " Zhu Lingshan
2026-08-28  9:53 ` [PATCH 05/10] drm/amdgpu: implement asynchronous userq destruction routine Zhu Lingshan
2026-08-28  9:53 ` [PATCH 06/10] drm/amdgpu: hold userq kref in MES reset Zhu Lingshan
2026-08-28  9:53 ` [PATCH 07/10] drm/amdgpu: hold userq kref during isolation scheduling Zhu Lingshan
2026-08-28  9:53 ` [PATCH 08/10] drm/amdgpu: hold userq kref during suspend and resume Zhu Lingshan
2026-08-28  9:53 ` [PATCH 09/10] drm/amdgpu: free userq by kref_put when fails to create Zhu Lingshan
2026-08-28  9:53 ` [PATCH 10/10] drm/amdgpu: take queue kref in userq_create to avoid UAF Zhu Lingshan

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=20260828095349.9797-3-lingshan.zhu@amd.com \
    --to=lingshan.zhu@amd.com \
    --cc=Alexander.Deucher@amd.com \
    --cc=Christian.Koenig@amd.com \
    --cc=Ray.Huang@amd.com \
    --cc=amd-gfx@lists.freedesktop.org \
    --cc=felix.kuehling@amd.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 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.