drm/i915: increase GPU wedging timeout
Subsystems (2)
PATCH 1/2 drm/i915: increase GPU wedging timeout
Currently our wedge timeout is 5 seconds. Hangcheck needs atleast three runs to declare a hang with 1500ms timer tick period.
To make sure that gpu can be wedged in the first place, define wedge timeout as multiple of hangcheck timer periods to ensure that it is always greater than hang detection time.
This commit increases wedging period from 5 seconds to 8 seconds.
v2: better name for macro (Chris Wilson)
--- drivers/gpu/drm/i915/i915_drv.c | 3 ++- drivers/gpu/drm/i915/i915_drv.h | 3 +++ 2 files changed, 5 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/i915_drv.c +2 -1
@@ -815,7 +815,8 @@ int i915_reset(struct drm_device *dev) simulated = dev_priv->gpu_error.stop_rings !=0 ; -if (!simulated &&get_seconds () - dev_priv->gpu_error.last_reset <5 ) { +if (!simulated &&get_seconds () - dev_priv->gpu_error.last_reset < +DRM_I915_WEDGE_PERIOD ) {DRM_ERROR ("GPU hanging too fast, declaring wedged!\n" ); ret = -ENODEV ; }else {
drivers/gpu/drm/i915/i915_drv.h +3 -0
@@ -976,6 +976,9 @@ struct i915_gpu_error {/* For hangcheck timer */ #define DRM_I915_HANGCHECK_PERIOD 1500/* in ms */ #define DRM_I915_HANGCHECK_JIFFIES msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD) +/* Hangcheck needs >2 periods to declare a hang */ +#define DRM_I915_WEDGE_PERIOD DIV_ROUND_UP(5*DRM_I915_HANGCHECK_PERIOD, 1000) +struct timer_list hangcheck_timer;/* For reset and error_state handling. */
Signature · 1 line
1.7.9.5
PATCH 2/2 drm/i915: ban badly behaving contexts
Now when we have mechanism in place to track which context was guilty of hanging the gpu, it is possible to punish for bad behaviour.
If context has recently submitted a faulty batchbuffers guilty of gpu hang and submits another batch which hangs gpu in a 16 seconds time window, ban it permanently.
Ban means that batchbuffers from this fd/context will not be accepted anymore for execution.
v2: Store guilty ban status bool in gpu_error instead of pointers
that might become danling before hang is declared.
v3: Use return value for banned status instead of stashing state
into gpu_error (Chris Wilson)
v4: - rebase on top of fixed hang stats api
- add define for ban period
- rename commit and improve commit msg
--- drivers/gpu/drm/i915/i915_drv.c | 6 +++-- drivers/gpu/drm/i915/i915_drv.h | 10 +++++++- drivers/gpu/drm/i915/i915_gem.c | 35 +++++++++++++++++++++------- drivers/gpu/drm/i915/i915_gem_execbuffer.c | 12 ++++++++++ 4 files changed, 51 insertions(+), 12 deletions(-)
drivers/gpu/drm/i915/i915_drv.c +4 -2
@@ -804,6 +804,7 @@ int i915_reset(struct drm_device *dev) {drm_i915_private_t *dev_priv = dev->dev_private;bool simulated; +bool ctx_banned;int ret;if (!i915_try_reset)
@@ -811,11 +812,12 @@ int i915_reset(struct drm_device *dev)mutex_lock (&dev->struct_mutex); -i915_gem_reset (dev); + ctx_banned =i915_gem_reset (dev); simulated = dev_priv->gpu_error.stop_rings !=0 ; -if (!simulated &&get_seconds () - dev_priv->gpu_error.last_reset < +if (!(simulated || ctx_banned) && +get_seconds () - dev_priv->gpu_error.last_reset <DRM_I915_WEDGE_PERIOD ) {DRM_ERROR ("GPU hanging too fast, declaring wedged!\n" ); ret = -ENODEV ;
drivers/gpu/drm/i915/i915_drv.h +9 -1
@@ -575,6 +575,12 @@ struct i915_ctx_hang_stats {/* This context had batch active when hang was declared */ unsigned batch_active; + +/* Time when this context was last blamed for a GPU reset */ +unsigned long batch_active_reset_ts; + +/* This context is banned to submit more work */ +bool banned; };/* This must match up with the value previously used for execbuf2.rsvd1. */
@@ -978,6 +984,8 @@ struct i915_gpu_error {#define DRM_I915_HANGCHECK_JIFFIES msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD)/* Hangcheck needs >2 periods to declare a hang */ #define DRM_I915_WEDGE_PERIOD DIV_ROUND_UP(5*DRM_I915_HANGCHECK_PERIOD, 1000) +/* Hang gpu twice in this window and your context gets banned */ +#define DRM_I915_CONTEXT_BAN_PERIOD (2 * DRM_I915_WEDGE_PERIOD)struct timer_list hangcheck_timer;
@@ -1922,7 +1930,7 @@ static inline bool i915_terminally_wedged(struct i915_gpu_error *error)return atomic_read (&error->reset_counter) ==I915_WEDGED ; } -void i915_gem_reset (struct drm_device *dev); +bool i915_gem_reset (struct drm_device *dev);bool i915_gem_clflush_object (struct drm_i915_gem_object *obj,bool force);int __must_check i915_gem_object_finish_gpu (struct drm_i915_gem_object *obj);int __must_check i915_gem_init (struct drm_device *dev);
drivers/gpu/drm/i915/i915_gem.c +26 -9
@@ -2201,16 +2201,16 @@ static bool i915_request_guilty(struct drm_i915_gem_request *request,return false; } -static void i915_set_reset_status (struct intel_ring_buffer *ring, +static bool i915_set_reset_status (struct intel_ring_buffer *ring,struct drm_i915_gem_request *request,u32 acthd) {struct i915_ctx_hang_stats *hs =NULL ; -bool inside, guilty; +bool inside, guilty, banned;unsigned long offset =0 ;/* Innocent until proven guilty */ - guilty = false; + guilty = banned = false;if (request->batch_obj) offset = i915_gem_obj_offset(request->batch_obj,
@@ -2237,11 +2237,21 @@ static void i915_set_reset_status(struct intel_ring_buffer *ring, hs = &request->file_priv->hang_stats;if (hs) { -if (guilty) +if (guilty) { +if (!hs->banned && +get_seconds () - hs->batch_active_reset_ts < +DRM_I915_CONTEXT_BAN_PERIOD ) { + hs->banned = banned = true; +DRM_ERROR ("context hanging too fast, declaring banned\n" ); + } hs->batch_active++; -else + hs->batch_active_reset_ts =get_seconds (); + }else { hs->batch_pending++; + } } + +return banned; }static void i915_gem_free_request (struct drm_i915_gem_request *request)
@@ -2255,11 +2265,12 @@ static void i915_gem_free_request(struct drm_i915_gem_request *request)kfree (request); } -static void i915_gem_reset_ring_lists (struct drm_i915_private *dev_priv, +static bool i915_gem_reset_ring_lists (struct drm_i915_private *dev_priv,struct intel_ring_buffer *ring) {u32 completed_seqno;u32 acthd; +bool ctx_banned = false; acthd =intel_ring_get_active_head (ring); completed_seqno = ring->get_seqno (ring, false);
@@ -2272,7 +2283,8 @@ static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv, list);if (request->seqno > completed_seqno) -i915_set_reset_status (ring, request, acthd); + ctx_banned |=i915_set_reset_status (ring, + request, acthd);i915_gem_free_request (request); }
@@ -2286,6 +2298,8 @@ static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,i915_gem_object_move_to_inactive (obj); } + +return ctx_banned; }void i915_gem_restore_fences (struct drm_device *dev)
@@ -2309,16 +2323,19 @@ void i915_gem_restore_fences(struct drm_device *dev) } } -void i915_gem_reset (struct drm_device *dev) +bool i915_gem_reset (struct drm_device *dev) {struct drm_i915_private *dev_priv = dev->dev_private;struct intel_ring_buffer *ring;int i; +bool ctx_banned = false;for_each_ring (ring, dev_priv, i) -i915_gem_reset_ring_lists (dev_priv, ring); + ctx_banned |=i915_gem_reset_ring_lists (dev_priv, ring);i915_gem_restore_fences (dev); + +return ctx_banned; }/**
drivers/gpu/drm/i915/i915_gem_execbuffer.c +12 -0
@@ -886,6 +886,7 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,struct drm_i915_gem_object *batch_obj;struct drm_clip_rect *cliprects =NULL ;struct intel_ring_buffer *ring; +struct i915_ctx_hang_stats *hs;u32 ctx_id =i915_execbuffer2_get_context_id (*args);u32 exec_start, exec_len;u32 mask, flags;
@@ -1077,6 +1078,17 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,if (ret) goto err; + hs =i915_gem_context_get_hang_stats (dev, file, ctx_id); +if (IS_ERR (hs)) { + ret =PTR_ERR (hs); + goto err; + } + +if (hs->banned) { + ret = -EIO ; + goto err; + } + ret =i915_switch_context (ring, file, ctx_id);if (ret) goto err;
Signature · 1 line
1.7.9.5
Quoted 11 lines · Mika Kuoppala
Currently our wedge timeout is 5 seconds. Hangcheck needs atleast three runs to declare a hang with 1500ms timer tick period.
To make sure that gpu can be wedged in the first place, define wedge timeout as multiple of hangcheck timer periods to ensure that it is always greater than hang detection time.
This commit increases wedging period from 5 seconds to 8 seconds.
v2: better name for macro (Chris Wilson)
Forget both of these. I will post an improved series soon. -Mika