diff mbox series

drm/xe/guc: Check sizing of guc_capture output

Message ID 20240116171237.118463-7-zhanjun.dong@intel.com (mailing list archive)
State New, archived
Headers show
Series drm/xe/guc: Check sizing of guc_capture output | expand

Commit Message

Dong, Zhanjun Jan. 16, 2024, 5:12 p.m. UTC
Add capture output size check function to provide a reasonable
minimum size for error capture region before allocating the shared
buffer.

Signed-off-by: Zhanjun Dong <zhanjun.dong@intel.com>
---
 drivers/gpu/drm/xe/xe_guc_capture.c | 76 +++++++++++++++++++++++++++++
 1 file changed, 76 insertions(+)
diff mbox series

Patch

diff --git a/drivers/gpu/drm/xe/xe_guc_capture.c b/drivers/gpu/drm/xe/xe_guc_capture.c
index dde3a269d114..f4153dc4ab86 100644
--- a/drivers/gpu/drm/xe/xe_guc_capture.c
+++ b/drivers/gpu/drm/xe/xe_guc_capture.c
@@ -559,6 +559,81 @@  xe_guc_capture_getnullheader(struct xe_guc *guc, void **outptr, size_t *size)
 	return 0;
 }
 
+static int
+guc_capture_output_min_size_est(struct xe_guc *guc)
+{
+	struct xe_gt *gt = guc_to_gt(guc);
+	struct xe_hw_engine *hwe;
+	enum xe_hw_engine_id id;
+
+	int worst_min_size = 0;
+	size_t tmp = 0;
+
+	if (!guc->capture)
+		return -ENODEV;
+
+	/*
+	 * If every single engine-instance suffered a failure in quick succession but
+	 * were all unrelated, then a burst of multiple error-capture events would dump
+	 * registers for every one engine instance, one at a time. In this case, GuC
+	 * would even dump the global-registers repeatedly.
+	 *
+	 * For each engine instance, there would be 1 x guc_state_capture_group_t output
+	 * followed by 3 x guc_state_capture_t lists. The latter is how the register
+	 * dumps are split across different register types (where the '3' are global vs class
+	 * vs instance).
+	 */
+	for_each_hw_engine(hwe, gt, id) {
+		worst_min_size += sizeof(struct guc_state_capture_group_header_t) +
+					 (3 * sizeof(struct guc_state_capture_header_t));
+
+		if (!guc_capture_getlistsize(guc, 0, GUC_CAPTURE_LIST_TYPE_GLOBAL, 0, &tmp, true))
+			worst_min_size += tmp;
+
+		if (!guc_capture_getlistsize(guc, 0, GUC_CAPTURE_LIST_TYPE_ENGINE_CLASS,
+					     hwe->class, &tmp, true)) {
+			worst_min_size += tmp;
+		}
+		if (!guc_capture_getlistsize(guc, 0, GUC_CAPTURE_LIST_TYPE_ENGINE_INSTANCE,
+					     hwe->class, &tmp, true)) {
+			worst_min_size += tmp;
+		}
+	}
+
+	return worst_min_size;
+}
+
+/*
+ * Add on a 3x multiplier to allow for multiple back-to-back captures occurring
+ * before the i915 can read the data out and process it
+ */
+#define GUC_CAPTURE_OVERBUFFER_MULTIPLIER 3
+
+static void check_guc_capture_size(struct xe_guc *guc)
+{
+	int min_size = guc_capture_output_min_size_est(guc);
+	int spare_size = min_size * GUC_CAPTURE_OVERBUFFER_MULTIPLIER;
+	u32 buffer_size = xe_guc_log_section_size_capture(&guc->log);
+
+	/*
+	 * NOTE: min_size is much smaller than the capture region allocation (DG2: <80K vs 1MB)
+	 * Additionally, its based on space needed to fit all engines getting reset at once
+	 * within the same G2H handler task slot. This is very unlikely. However, if GuC really
+	 * does run out of space for whatever reason, we will see an separate warning message
+	 * when processing the G2H event capture-notification, search for:
+	 * xe_guc_STATE_CAPTURE_EVENT_STATUS_NOSPACE.
+	 */
+	if (min_size < 0)
+		xe_gt_warn(guc_to_gt(guc), "Failed to calculate error state capture buffer minimum size: %d!\n",
+			   min_size);
+	else if (min_size > buffer_size)
+		xe_gt_warn(guc_to_gt(guc), "Error state capture buffer maybe small: %d < %d\n",
+			   buffer_size, min_size);
+	else if (spare_size > buffer_size)
+		xe_gt_dbg(guc_to_gt(guc), "Error state capture buffer lacks spare size: %d < %d (min = %d)\n",
+			  buffer_size, spare_size, min_size);
+}
+
 int xe_guc_capture_init(struct xe_guc *guc)
 {
 	guc->capture = kzalloc(sizeof(*guc->capture), GFP_KERNEL);
@@ -570,6 +645,7 @@  int xe_guc_capture_init(struct xe_guc *guc)
 	INIT_LIST_HEAD(&guc->capture->outlist);
 	INIT_LIST_HEAD(&guc->capture->cachelist);
 
+	check_guc_capture_size(guc);
 	return 0;
 }