drm/i915: Handle log buffer flush interrupt event from GuC
authorSagar Arun Kamble <sagar.a.kamble@intel.com>
Wed, 12 Oct 2016 16:24:32 +0000 (21:54 +0530)
committerTvrtko Ursulin <tvrtko.ursulin@intel.com>
Tue, 25 Oct 2016 08:34:06 +0000 (09:34 +0100)
GuC ukernel sends an interrupt to Host to flush the log buffer
and expects Host to correspondingly update the read pointer
information in the state structure, once it has consumed the
log buffer contents by copying them to a file or buffer.
Even if Host couldn't copy the contents, it can still update the
read pointer so that logging state is not disturbed on GuC side.

v2:
- Use a dedicated workqueue for handling flush interrupt. (Tvrtko)
- Reduce the overall log buffer copying time by skipping the copy of
  crash buffer area for regular cases and copying only the state
  structure data in first page.

v3:
 - Create a vmalloc mapping of log buffer. (Chris)
 - Cover the flush acknowledgment under rpm get & put.(Chris)
 - Revert the change of skipping the copy of crash dump area, as
   not really needed, will be covered by subsequent patch.

v4:
 - Destroy the wq under the same condition in which it was created,
   pass dev_piv pointer instead of dev to newly added GuC function,
   add more comments & rename variable for clarity. (Tvrtko)

v5:
- Allocate & destroy the dedicated wq, for handling flush interrupt,
  from the setup/teardown routines of GuC logging. (Chris)
- Validate the log buffer size value retrieved from state structure
  and do some minor cleanup. (Tvrtko)
- Fix error/warnings reported by checkpatch. (Tvrtko)
- Rebase.

v6:
 - Remove the interrupts_enabled check from guc_capture_logs_work, need
   to process that last work item also, queued just before disabling the
   interrupt as log buffer flush interrupt handling is a bit different
   case where GuC is actually expecting an ACK from host, which should be
   provided to keep the logging going.
   Sync against the work will be done by caller disabling the interrupt.
 - Don't sample the log buffer size value from state structure, directly
   use the expected value to move the pointer & do the copy and that cannot
   go wrong (out of bounds) as Driver only allocated the log buffer and the
   relay buffers. Driver should refrain from interpreting the log packet,
   as much possible and let Userspace parser detect the anomaly. (Chris)

v7:
- Use switch statement instead of 'if else' for retrieving the GuC log
  buffer size. (Tvrtko)
- Refactored the log buffer copying function and shortended the name of
  couple of variables for better readability. (Tvrtko)

v8:
- Make the dedicated wq as a high priority one to further reduce the
  turnaround time of handing log buffer flush event from GuC.

Signed-off-by: Sagar Arun Kamble <sagar.a.kamble@intel.com>
Signed-off-by: Akash Goel <akash.goel@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
drivers/gpu/drm/i915/i915_guc_submission.c
drivers/gpu/drm/i915/i915_irq.c
drivers/gpu/drm/i915/intel_guc.h

index 6b77af2533272a21101e67d7e9c2d64b038c25c2..1f27de2c0e1b66dd8881f8cf797f8b2559b1fd5c 100644 (file)
@@ -170,6 +170,15 @@ static int host2guc_sample_forcewake(struct intel_guc *guc,
        return host2guc_action(guc, data, ARRAY_SIZE(data));
 }
 
+static int host2guc_logbuffer_flush_complete(struct intel_guc *guc)
+{
+       u32 data[1];
+
+       data[0] = HOST2GUC_ACTION_LOG_BUFFER_FILE_FLUSH_COMPLETE;
+
+       return host2guc_action(guc, data, 1);
+}
+
 /*
  * Initialise, update, or clear doorbell data shared with the GuC
  *
@@ -847,6 +856,163 @@ err:
        return NULL;
 }
 
+static void guc_move_to_next_buf(struct intel_guc *guc)
+{
+}
+
+static void *guc_get_write_buffer(struct intel_guc *guc)
+{
+       return NULL;
+}
+
+static unsigned int guc_get_log_buffer_size(enum guc_log_buffer_type type)
+{
+       switch (type) {
+       case GUC_ISR_LOG_BUFFER:
+               return (GUC_LOG_ISR_PAGES + 1) * PAGE_SIZE;
+       case GUC_DPC_LOG_BUFFER:
+               return (GUC_LOG_DPC_PAGES + 1) * PAGE_SIZE;
+       case GUC_CRASH_DUMP_LOG_BUFFER:
+               return (GUC_LOG_CRASH_PAGES + 1) * PAGE_SIZE;
+       default:
+               MISSING_CASE(type);
+       }
+
+       return 0;
+}
+
+static void guc_read_update_log_buffer(struct intel_guc *guc)
+{
+       struct guc_log_buffer_state *log_buf_state, *log_buf_snapshot_state;
+       struct guc_log_buffer_state log_buf_state_local;
+       unsigned int buffer_size, write_offset;
+       enum guc_log_buffer_type type;
+       void *src_data, *dst_data;
+
+       if (WARN_ON(!guc->log.buf_addr))
+               return;
+
+       /* Get the pointer to shared GuC log buffer */
+       log_buf_state = src_data = guc->log.buf_addr;
+
+       /* Get the pointer to local buffer to store the logs */
+       log_buf_snapshot_state = dst_data = guc_get_write_buffer(guc);
+
+       /* Actual logs are present from the 2nd page */
+       src_data += PAGE_SIZE;
+       dst_data += PAGE_SIZE;
+
+       for (type = GUC_ISR_LOG_BUFFER; type < GUC_MAX_LOG_BUFFER; type++) {
+               /* Make a copy of the state structure, inside GuC log buffer
+                * (which is uncached mapped), on the stack to avoid reading
+                * from it multiple times.
+                */
+               memcpy(&log_buf_state_local, log_buf_state,
+                      sizeof(struct guc_log_buffer_state));
+               buffer_size = guc_get_log_buffer_size(type);
+               write_offset = log_buf_state_local.sampled_write_ptr;
+
+               /* Update the state of shared log buffer */
+               log_buf_state->read_ptr = write_offset;
+               log_buf_state->flush_to_file = 0;
+               log_buf_state++;
+
+               if (unlikely(!log_buf_snapshot_state))
+                       continue;
+
+               /* First copy the state structure in snapshot buffer */
+               memcpy(log_buf_snapshot_state, &log_buf_state_local,
+                      sizeof(struct guc_log_buffer_state));
+
+               /* The write pointer could have been updated by GuC firmware,
+                * after sending the flush interrupt to Host, for consistency
+                * set write pointer value to same value of sampled_write_ptr
+                * in the snapshot buffer.
+                */
+               log_buf_snapshot_state->write_ptr = write_offset;
+               log_buf_snapshot_state++;
+
+               /* Now copy the actual logs. */
+               memcpy(dst_data, src_data, buffer_size);
+
+               src_data += buffer_size;
+               dst_data += buffer_size;
+
+               /* FIXME: invalidate/flush for log buffer needed */
+       }
+
+       if (log_buf_snapshot_state)
+               guc_move_to_next_buf(guc);
+}
+
+static void guc_capture_logs_work(struct work_struct *work)
+{
+       struct drm_i915_private *dev_priv =
+               container_of(work, struct drm_i915_private, guc.log.flush_work);
+
+       i915_guc_capture_logs(dev_priv);
+}
+
+static void guc_log_cleanup(struct intel_guc *guc)
+{
+       struct drm_i915_private *dev_priv = guc_to_i915(guc);
+
+       lockdep_assert_held(&dev_priv->drm.struct_mutex);
+
+       /* First disable the flush interrupt */
+       gen9_disable_guc_interrupts(dev_priv);
+
+       if (guc->log.flush_wq)
+               destroy_workqueue(guc->log.flush_wq);
+
+       guc->log.flush_wq = NULL;
+
+       if (guc->log.buf_addr)
+               i915_gem_object_unpin_map(guc->log.vma->obj);
+
+       guc->log.buf_addr = NULL;
+}
+
+static int guc_log_create_extras(struct intel_guc *guc)
+{
+       struct drm_i915_private *dev_priv = guc_to_i915(guc);
+       void *vaddr;
+       int ret;
+
+       lockdep_assert_held(&dev_priv->drm.struct_mutex);
+
+       /* Nothing to do */
+       if (i915.guc_log_level < 0)
+               return 0;
+
+       if (!guc->log.buf_addr) {
+               /* Create a vmalloc mapping of log buffer pages */
+               vaddr = i915_gem_object_pin_map(guc->log.vma->obj, I915_MAP_WB);
+               if (IS_ERR(vaddr)) {
+                       ret = PTR_ERR(vaddr);
+                       DRM_ERROR("Couldn't map log buffer pages %d\n", ret);
+                       return ret;
+               }
+
+               guc->log.buf_addr = vaddr;
+       }
+
+       if (!guc->log.flush_wq) {
+               INIT_WORK(&guc->log.flush_work, guc_capture_logs_work);
+
+               /* Need a dedicated wq to process log buffer flush interrupts
+                * from GuC without much delay so as to avoid any loss of logs.
+                */
+               guc->log.flush_wq = alloc_ordered_workqueue("i915-guc_log", WQ_HIGHPRI);
+               if (guc->log.flush_wq == NULL) {
+                       DRM_ERROR("Couldn't allocate the wq for GuC logging\n");
+                       return -ENOMEM;
+               }
+       }
+
+       return 0;
+}
+
 static void guc_log_create(struct intel_guc *guc)
 {
        struct i915_vma *vma;
@@ -872,6 +1038,13 @@ static void guc_log_create(struct intel_guc *guc)
                }
 
                guc->log.vma = vma;
+
+               if (guc_log_create_extras(guc)) {
+                       guc_log_cleanup(guc);
+                       i915_vma_unpin_and_release(&guc->log.vma);
+                       i915.guc_log_level = -1;
+                       return;
+               }
        }
 
        /* each allocated unit is a page */
@@ -1065,6 +1238,7 @@ void i915_guc_submission_fini(struct drm_i915_private *dev_priv)
        struct intel_guc *guc = &dev_priv->guc;
 
        i915_vma_unpin_and_release(&guc->ads_vma);
+       guc_log_cleanup(guc);
        i915_vma_unpin_and_release(&guc->log.vma);
 
        if (guc->ctx_pool_vma)
@@ -1126,3 +1300,15 @@ int intel_guc_resume(struct drm_device *dev)
 
        return host2guc_action(guc, data, ARRAY_SIZE(data));
 }
+
+void i915_guc_capture_logs(struct drm_i915_private *dev_priv)
+{
+       guc_read_update_log_buffer(&dev_priv->guc);
+
+       /* Generally device is expected to be active only at this
+        * time, so get/put should be really quick.
+        */
+       intel_runtime_pm_get(dev_priv);
+       host2guc_logbuffer_flush_complete(&dev_priv->guc);
+       intel_runtime_pm_put(dev_priv);
+}
index b37959fdfb0e9048460835876ba1a9a5d88c8394..253c22109c3fc99702779b48ac7ad0e1e58f849d 100644 (file)
@@ -1669,7 +1669,33 @@ static void gen6_rps_irq_handler(struct drm_i915_private *dev_priv, u32 pm_iir)
 static void gen9_guc_irq_handler(struct drm_i915_private *dev_priv, u32 gt_iir)
 {
        if (gt_iir & GEN9_GUC_TO_HOST_INT_EVENT) {
-               /* TODO: Handle events for which GuC interrupted host */
+               /* Sample the log buffer flush related bits & clear them out now
+                * itself from the message identity register to minimize the
+                * probability of losing a flush interrupt, when there are back
+                * to back flush interrupts.
+                * There can be a new flush interrupt, for different log buffer
+                * type (like for ISR), whilst Host is handling one (for DPC).
+                * Since same bit is used in message register for ISR & DPC, it
+                * could happen that GuC sets the bit for 2nd interrupt but Host
+                * clears out the bit on handling the 1st interrupt.
+                */
+               u32 msg, flush;
+
+               msg = I915_READ(SOFT_SCRATCH(15));
+               flush = msg & (GUC2HOST_MSG_CRASH_DUMP_POSTED |
+                              GUC2HOST_MSG_FLUSH_LOG_BUFFER);
+               if (flush) {
+                       /* Clear the message bits that are handled */
+                       I915_WRITE(SOFT_SCRATCH(15), msg & ~flush);
+
+                       /* Handle flush interrupt in bottom half */
+                       queue_work(dev_priv->guc.log.flush_wq,
+                                  &dev_priv->guc.log.flush_work);
+               } else {
+                       /* Not clearing of unhandled event bits won't result in
+                        * re-triggering of the interrupt.
+                        */
+               }
        }
 }
 
index 81bd345c7598dd476e4fc437b0d42e8d4ce0cedb..fb72313c8649884ddcc95f4cd886516448816c2d 100644 (file)
@@ -126,6 +126,9 @@ struct intel_guc_fw {
 struct intel_guc_log {
        uint32_t flags;
        struct i915_vma *vma;
+       void *buf_addr;
+       struct workqueue_struct *flush_wq;
+       struct work_struct flush_work;
 };
 
 struct intel_guc {
@@ -170,5 +173,6 @@ int i915_guc_wq_reserve(struct drm_i915_gem_request *rq);
 void i915_guc_wq_unreserve(struct drm_i915_gem_request *request);
 void i915_guc_submission_disable(struct drm_i915_private *dev_priv);
 void i915_guc_submission_fini(struct drm_i915_private *dev_priv);
+void i915_guc_capture_logs(struct drm_i915_private *dev_priv);
 
 #endif