Tegra194: ras: split up RAS error clear SMC call.
In order to make sure SMC call is within 25us, this patch reduces number of RAS
errors accessed to 8 at most for each SMC call and takes a input/output
parameter to specify in progress RAS error record index.
The measured SMC call latency is about 20us under Linux test kernel driver.
Change-Id: Ia1b57c8673e0193dc341a36af0b5c09fb48f965f
Signed-off-by: David Pu <dpu@nvidia.com>
diff --git a/plat/nvidia/tegra/soc/t194/plat_ras.c b/plat/nvidia/tegra/soc/t194/plat_ras.c
index 54c2924..0c4c6fa 100644
--- a/plat/nvidia/tegra/soc/t194/plat_ras.c
+++ b/plat/nvidia/tegra/soc/t194/plat_ras.c
@@ -9,6 +9,7 @@
#include <common/debug.h>
#include <lib/bakery_lock.h>
+#include <lib/cassert.h>
#include <lib/extensions/ras.h>
#include <lib/utils_def.h>
#include <services/sdei.h>
@@ -26,6 +27,17 @@
*/
#define ERR_FR_EN_BITS_MASK 0xFFFFFFFF00000000ULL
+/*
+ * Number of RAS errors will be cleared per 'tegra194_ras_corrected_err_clear'
+ * function call.
+ */
+#define RAS_ERRORS_PER_CALL 8
+
+/*
+ * the max possible RAS node index value.
+ */
+#define RAS_NODE_INDEX_MAX 0x1FFFFFFFU
+
/* bakery lock for platform RAS handler. */
static DEFINE_BAKERY_LOCK(ras_handler_lock);
#define ras_lock() bakery_lock_get(&ras_handler_lock)
@@ -151,12 +163,41 @@
/*
* Function to clear RAS ERR<n>STATUS for corrected RAS error.
- * This function ignores any new RAS error signaled during clearing; it is not
- * multi-core safe(no ras_lock is taken to reduce overhead).
+ *
+ * This function clears number of 'RAS_ERRORS_PER_CALL' RAS errors at most.
+ * 'cookie' - in/out cookie parameter to specify/store last visited RAS
+ * error record index. it is set to '0' to indicate no more RAS
+ * error record to clear.
*/
-void tegra194_ras_corrected_err_clear(void)
+void tegra194_ras_corrected_err_clear(uint64_t *cookie)
{
+ /*
+ * 'last_node' and 'last_idx' represent last visited RAS node index from
+ * previous function call. they are set to 0 when first smc call is made
+ * or all RAS error are visited by followed multipile smc calls.
+ */
+ union prev_record {
+ struct record {
+ uint32_t last_node;
+ uint32_t last_idx;
+ } rec;
+ uint64_t value;
+ } prev;
+
uint64_t clear_ce_status = 0ULL;
+ int32_t nerrs_per_call = RAS_ERRORS_PER_CALL;
+ uint32_t i;
+
+ if (cookie == NULL) {
+ return;
+ }
+
+ prev.value = *cookie;
+
+ if ((prev.rec.last_node >= RAS_NODE_INDEX_MAX) ||
+ (prev.rec.last_idx >= RAS_NODE_INDEX_MAX)) {
+ return;
+ }
ERR_STATUS_SET_FIELD(clear_ce_status, AV, 0x1UL);
ERR_STATUS_SET_FIELD(clear_ce_status, V, 0x1UL);
@@ -164,25 +205,56 @@
ERR_STATUS_SET_FIELD(clear_ce_status, MV, 0x1UL);
ERR_STATUS_SET_FIELD(clear_ce_status, CE, 0x3UL);
- for (uint32_t i = 0U; i < err_record_mappings.num_err_records; i++) {
+
+ for (i = prev.rec.last_node; i < err_record_mappings.num_err_records; i++) {
const struct err_record_info *info = &err_record_mappings.err_records[i];
uint32_t idx_start = info->sysreg.idx_start;
uint32_t num_idx = info->sysreg.num_idx;
+ uint32_t j;
+
- for (uint32_t j = 0U; j < num_idx; j++) {
+ j = (i == prev.rec.last_node && prev.value != 0UL) ?
+ (prev.rec.last_idx + 1U) : 0U;
+
+ for (; j < num_idx; j++) {
uint64_t status;
uint32_t err_idx = idx_start + j;
+ if (err_idx >= RAS_NODE_INDEX_MAX) {
+ return;
+ }
+
write_errselr_el1(err_idx);
status = read_erxstatus_el1();
if (ERR_STATUS_GET_FIELD(status, CE) != 0U) {
write_erxstatus_el1(clear_ce_status);
}
+
+ --nerrs_per_call;
+
+ /* only clear 'nerrs_per_call' errors each time. */
+ if (nerrs_per_call <= 0) {
+ prev.rec.last_idx = j;
+ prev.rec.last_node = i;
+ /* save last visited error record index
+ * into cookie.
+ */
+ *cookie = prev.value;
+
+ return;
+ }
}
}
+
+ /*
+ * finish if all ras error records are checked or provided index is out
+ * of range.
+ */
+ *cookie = 0ULL;
+ return;
}
/* Function to probe an error from error record group. */
@@ -330,18 +402,26 @@
static struct ras_aux_data per_core_ras_group[] = {
PER_CORE_RAS_GROUP_NODES
};
+CASSERT(ARRAY_SIZE(per_core_ras_group) < RAS_NODE_INDEX_MAX,
+ assert_max_per_core_ras_group_size);
static struct ras_aux_data per_cluster_ras_group[] = {
PER_CLUSTER_RAS_GROUP_NODES
};
+CASSERT(ARRAY_SIZE(per_cluster_ras_group) < RAS_NODE_INDEX_MAX,
+ assert_max_per_cluster_ras_group_size);
static struct ras_aux_data scf_l3_ras_group[] = {
SCF_L3_BANK_RAS_GROUP_NODES
};
+CASSERT(ARRAY_SIZE(scf_l3_ras_group) < RAS_NODE_INDEX_MAX,
+ assert_max_scf_l3_ras_group_size);
static struct ras_aux_data ccplex_ras_group[] = {
CCPLEX_RAS_GROUP_NODES
};
+CASSERT(ARRAY_SIZE(ccplex_ras_group) < RAS_NODE_INDEX_MAX,
+ assert_max_ccplex_ras_group_size);
/*
* We have same probe and handler for each error record group, use a macro to
@@ -395,6 +475,9 @@
ADD_ONE_ERR_GROUP(0x400, ccplex_ras_group),
};
+CASSERT(ARRAY_SIZE(carmel_ras_records) < RAS_NODE_INDEX_MAX,
+ assert_max_carmel_ras_records_size);
+
REGISTER_ERR_RECORD_INFO(carmel_ras_records);
/* dummy RAS interrupt */