linux-kernel - [PATCH v2 3/3] drm/lima: mask irqs in timeout path before hard reset

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20240405152951.1531555-4-nunes.erico@gmail.com>
Date: Fri,  5 Apr 2024 17:29:51 +0200
From: Erico Nunes <nunes.erico@...il.com>
To: Qiang Yu <yuq825@...il.com>,
	anarsoul@...il.com,
	dri-devel@...ts.freedesktop.org,
	lima@...ts.freedesktop.org
Cc: Maarten Lankhorst <maarten.lankhorst@...ux.intel.com>,
	Maxime Ripard <mripard@...nel.org>,
	Thomas Zimmermann <tzimmermann@...e.de>,
	David Airlie <airlied@...il.com>,
	Daniel Vetter <daniel@...ll.ch>,
	christian.koenig@....com,
	megi@....cz,
	linux-kernel@...r.kernel.org,
	Erico Nunes <nunes.erico@...il.com>
Subject: [PATCH v2 3/3] drm/lima: mask irqs in timeout path before hard reset

There is a race condition in which a rendering job might take just long
enough to trigger the drm sched job timeout handler but also still
complete before the hard reset is done by the timeout handler.
This runs into race conditions not expected by the timeout handler.
In some very specific cases it currently may result in a refcount
imbalance on lima_pm_idle, with a stack dump such as:

[10136.669170] WARNING: CPU: 0 PID: 0 at drivers/gpu/drm/lima/lima_devfreq.c:205 lima_devfreq_record_idle+0xa0/0xb0
..
[10136.669459] pc : lima_devfreq_record_idle+0xa0/0xb0
..
[10136.669628] Call trace:
[10136.669634]  lima_devfreq_record_idle+0xa0/0xb0
[10136.669646]  lima_sched_pipe_task_done+0x5c/0xb0
[10136.669656]  lima_gp_irq_handler+0xa8/0x120
[10136.669666]  __handle_irq_event_percpu+0x48/0x160
[10136.669679]  handle_irq_event+0x4c/0xc0

We can prevent that race condition entirely by masking the irqs at the
beginning of the timeout handler, at which point we give up on waiting
for that job entirely.
The irqs will be enabled again at the next hard reset which is already
done as a recovery by the timeout handler.

Signed-off-by: Erico Nunes <nunes.erico@...il.com>
Reviewed-by: Qiang Yu <yuq825@...il.com>
---
 drivers/gpu/drm/lima/lima_sched.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/drivers/gpu/drm/lima/lima_sched.c b/drivers/gpu/drm/lima/lima_sched.c
index 66841503a618..bbf3f8feab94 100644
--- a/drivers/gpu/drm/lima/lima_sched.c
+++ b/drivers/gpu/drm/lima/lima_sched.c
@@ -430,6 +430,13 @@ static enum drm_gpu_sched_stat lima_sched_timedout_job(struct drm_sched_job *job
 		return DRM_GPU_SCHED_STAT_NOMINAL;
 	}
 
+	/*
+	 * The task might still finish while this timeout handler runs.
+	 * To prevent a race condition on its completion, mask all irqs
+	 * on the running core until the next hard reset completes.
+	 */
+	pipe->task_mask_irq(pipe);
+
 	if (!pipe->error)
 		DRM_ERROR("%s job timeout\n", lima_ip_name(ip));
 
-- 
2.44.0