| Message ID | 20160916192318.14030-2-chris@chris-wilson.co.uk (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On pe, 2016-09-16 at 20:23 +0100, Chris Wilson wrote: > void intel_lr_context_resume(struct drm_i915_private *dev_priv) > { > > - struct i915_gem_context *ctx = dev_priv->kernel_context; > > struct intel_engine_cs *engine; > > + struct i915_gem_context *ctx; > + > > + /* Because we emit WA_TAIL_DWORDS there may be a disparity > > + * between our bookkeeping in ce->ring->head and ce->ring->tail and > + * that stored in context. As we only write new comamnds from "new commands" > + * ce->ring->tail onwards, everything before that is junk. If the GPU > > + * starts reading from its RING_HEAD from the context, it may try to > > + * execute that junk and die. > > + * > > + * So to avoid that we reset the context images upon resume. For > > + * simplicity, we just zero everything out. > > + */ > > + list_for_each_entry(ctx, &dev_priv->context_list, link) { > > + for_each_engine(engine, dev_priv) { > > + struct intel_context *ce = &ctx->engine[engine->id]; > > + u32 *reg_state; > > > - for_each_engine(engine, dev_priv) { > > - struct intel_context *ce = &ctx->engine[engine->id]; > > - void *vaddr; > > - uint32_t *reg_state; > - > > - if (!ce->state) > > - continue; > - > > - vaddr = i915_gem_object_pin_map(ce->state->obj, I915_MAP_WB); > > - if (WARN_ON(IS_ERR(vaddr))) > > - continue; > > + if (!ce->state) > > + continue; > > > - reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE; > > + reg_state = i915_gem_object_pin_map(ce->state->obj, > > + I915_MAP_WB); > > + if (WARN_ON(IS_ERR(reg_state))) > > + continue; > > > - reg_state[CTX_RING_HEAD+1] = 0; > > - reg_state[CTX_RING_TAIL+1] = 0; > + reg_state += LRC_STATE_PN * PAGE_SIZE / sizeof(u32); / sizeof(*reg_state) I presume. Also add a newline here. With those; Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Regards, Joonas
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c index bc1786478f8e..ec2ad603be86 100644 --- a/drivers/gpu/drm/i915/intel_lrc.c +++ b/drivers/gpu/drm/i915/intel_lrc.c @@ -2100,30 +2100,41 @@ error_deref_obj: void intel_lr_context_resume(struct drm_i915_private *dev_priv) { - struct i915_gem_context *ctx = dev_priv->kernel_context; struct intel_engine_cs *engine; + struct i915_gem_context *ctx; + + /* Because we emit WA_TAIL_DWORDS there may be a disparity + * between our bookkeeping in ce->ring->head and ce->ring->tail and + * that stored in context. As we only write new comamnds from + * ce->ring->tail onwards, everything before that is junk. If the GPU + * starts reading from its RING_HEAD from the context, it may try to + * execute that junk and die. + * + * So to avoid that we reset the context images upon resume. For + * simplicity, we just zero everything out. + */ + list_for_each_entry(ctx, &dev_priv->context_list, link) { + for_each_engine(engine, dev_priv) { + struct intel_context *ce = &ctx->engine[engine->id]; + u32 *reg_state; - for_each_engine(engine, dev_priv) { - struct intel_context *ce = &ctx->engine[engine->id]; - void *vaddr; - uint32_t *reg_state; - - if (!ce->state) - continue; - - vaddr = i915_gem_object_pin_map(ce->state->obj, I915_MAP_WB); - if (WARN_ON(IS_ERR(vaddr))) - continue; + if (!ce->state) + continue; - reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE; + reg_state = i915_gem_object_pin_map(ce->state->obj, + I915_MAP_WB); + if (WARN_ON(IS_ERR(reg_state))) + continue; - reg_state[CTX_RING_HEAD+1] = 0; - reg_state[CTX_RING_TAIL+1] = 0; + reg_state += LRC_STATE_PN * PAGE_SIZE / sizeof(u32); + reg_state[CTX_RING_HEAD+1] = 0; + reg_state[CTX_RING_TAIL+1] = 0; - ce->state->obj->dirty = true; - i915_gem_object_unpin_map(ce->state->obj); + ce->state->obj->dirty = true; + i915_gem_object_unpin_map(ce->state->obj); - ce->ring->head = 0; - ce->ring->tail = 0; + ce->ring->head = ce->ring->tail = 0; + ce->ring->last_retired_head = -1; + } } }
There is a disparity in the context image saved to disk and our own bookkeeping - that is we presume the RING_HEAD and RING_TAIL match our stored ce->ring->tail value. However, as we emit WA_TAIL_DWORDS into the ring but may not tell the GPU about them, the GPU may be lagging behind our bookkeeping. Upon hibernation we do not save stolen pages, presuming that their contents are volatile. This means that although we start writing into the ring at tail, the GPU starts executing from its HEAD and there may be some garbage in between and so the GPU promptly hangs upon resume. Testcase: igt/gem_exec_suspend/basic-S4 Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96526 Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> --- drivers/gpu/drm/i915/intel_lrc.c | 49 ++++++++++++++++++++++++---------------- 1 file changed, 30 insertions(+), 19 deletions(-)