Unloading the nouveau module while the GPU is asleep (e.g. on dual GPU laptops) leads to an infinite loop in nvkm_timer_wait_eq() because the timer read out is 0xffffffffffffffff so the condition of the while loop becomes -1 - (-1) < nsec and stays like that unless the GPU is woken up.
Use the kernel timer as fallback in this unlikely event. Synchronize the kernel timer and GPU timer in nv04_timer_init() / gk20a_timer_init() which should get called once on driver initialization and on every resume.
Even with this fix applied, unloading the module takes a whopping 167 seconds. This could be reduced by changing the NV_WAIT_DEFAULT timeout from the current (maybe excessive?) 2 seconds to 200 ms.
A WARN_ON is spewed out at nouveau_bo.c:398 after 81 seconds and a null pointer dereference occurs in nouveau_cli_destroy(), so there's more to fix here.
This patch might also be needed to properly handle a GPU connected via Thunderbolt which is suddenly unplugged.
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=88861 Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=61115 Tested-by: Paul Hordiienko pvt.gord@gmail.com [MBP 6,2 2010 intel ILK + nvidia GT216 pre-retina] Tested-by: William Brown william@blackhats.net.au [MBP 8,2 2011 intel SNB + amd turks pre-retina] Tested-by: Lukas Wunner lukas@wunner.de [MBP 9,1 2012 intel IVB + nvidia GK107 pre-retina] Tested-by: Bruno Bierbaumer bruno@bierbaumer.net [MBP 11,3 2013 intel HSW + nvidia GK107 retina -- work in progress]
Signed-off-by: Lukas Wunner lukas@wunner.de --- drivers/gpu/drm/nouveau/nvkm/subdev/timer/gk20a.c | 4 ++++ drivers/gpu/drm/nouveau/nvkm/subdev/timer/nv04.c | 9 +++++++++ drivers/gpu/drm/nouveau/nvkm/subdev/timer/nv04.h | 1 + 3 files changed, 14 insertions(+)
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/gk20a.c b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/gk20a.c index 80e3806..28d27ff 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/gk20a.c +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/gk20a.c @@ -41,6 +41,10 @@ gk20a_timer_init(struct nvkm_object *object) /* restore the time before suspend */ nv_wr32(priv, NV04_PTIMER_TIME_1, hi); nv_wr32(priv, NV04_PTIMER_TIME_0, lo); + + /* save kernel time as fallback */ + priv->suspend_ktime = ktime_to_ns(ktime_get()) - priv->suspend_time; + return 0; }
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/nv04.c b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/nv04.c index 6b7facb..228749d 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/nv04.c +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/nv04.c @@ -36,6 +36,11 @@ nv04_timer_read(struct nvkm_timer *ptimer) lo = nv_rd32(priv, NV04_PTIMER_TIME_0); } while (hi != nv_rd32(priv, NV04_PTIMER_TIME_1));
+ if (unlikely(hi == -1 && lo == -1)) { + nv_spam(priv, "read failed, falling back to kernel timer\n"); + return ktime_to_ns(ktime_get()) - priv->suspend_ktime; + } + return ((u64)hi << 32 | lo); }
@@ -216,6 +221,10 @@ nv04_timer_init(struct nvkm_object *object) nv_wr32(priv, NV04_PTIMER_INTR_EN_0, 0x00000000); nv_wr32(priv, NV04_PTIMER_TIME_1, hi); nv_wr32(priv, NV04_PTIMER_TIME_0, lo); + + /* save kernel time as fallback */ + priv->suspend_ktime = ktime_to_ns(ktime_get()) - priv->suspend_time; + return 0; }
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/nv04.h b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/nv04.h index 89996a9..1b83a0f 100644 --- a/drivers/gpu/drm/nouveau/nvkm/subdev/timer/nv04.h +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/timer/nv04.h @@ -15,6 +15,7 @@ struct nv04_timer_priv { struct list_head alarms; spinlock_t lock; u64 suspend_time; + u64 suspend_ktime; };
int nv04_timer_ctor(struct nvkm_object *, struct nvkm_object *,