The current EQ delay calculation for AIC is based only on RX packet rate.
This fails to be effective when there's only TX and no RX.
This patch inclues:
- Calculating EQ-delay based on both RX and TX pps.
- Modifying EQ-delay of all EQs via one cmd, instead of issuing a separate
cmd for each EQ.
- A new structure to store interrupt coalescing parameters, in a separate
cache-line from the EQ-obj.
Signed-off-by: Sathya Perla <sathya.perla@emulex.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
#define BE_MIN_MTU 256
#define BE_NUM_VLANS_SUPPORTED 64
-#define BE_MAX_EQD 96u
+#define BE_MAX_EQD 128u
#define BE_MAX_TX_FRAG_COUNT 30
#define EVNT_Q_LEN 1024
struct be_adapter *adapter;
} ____cacheline_aligned_in_smp;
+struct be_aic_obj { /* Adaptive interrupt coalescing (AIC) info */
+ bool enable;
+ u32 min_eqd; /* in usecs */
+ u32 max_eqd; /* in usecs */
+ u32 prev_eqd; /* in usecs */
+ u32 et_eqd; /* configured val when aic is off */
+ ulong jiffies;
+ u64 rx_pkts_prev; /* Used to calculate RX pps */
+ u64 tx_reqs_prev; /* Used to calculate TX pps */
+};
+
struct be_mcc_obj {
struct be_queue_info q;
struct be_queue_info cq;
struct be_rx_stats {
u64 rx_bytes;
u64 rx_pkts;
- u64 rx_pkts_prev;
- ulong rx_jiffies;
u32 rx_drops_no_skbs; /* skb allocation errors */
u32 rx_drops_no_frags; /* HW has no fetched frags */
u32 rx_post_fail; /* page post alloc failures */
u32 rx_compl;
u32 rx_mcast_pkts;
u32 rx_compl_err; /* completions with err set */
- u32 rx_pps; /* pkts per second */
struct u64_stats_sync sync;
};
u32 big_page_size; /* Compounded page size shared by rx wrbs */
struct be_drv_stats drv_stats;
+ struct be_aic_obj aic_obj[MAX_EVT_QS];
u16 vlans_added;
u8 vlan_tag[VLAN_N_VID];
u8 vlan_prio_bmap; /* Available Priority BitMap */
/* set the EQ delay interval of an EQ to specified value
* Uses async mcc
*/
-int be_cmd_modify_eqd(struct be_adapter *adapter, u32 eq_id, u32 eqd)
+int be_cmd_modify_eqd(struct be_adapter *adapter, struct be_set_eqd *set_eqd,
+ int num)
{
struct be_mcc_wrb *wrb;
struct be_cmd_req_modify_eq_delay *req;
- int status = 0;
+ int status = 0, i;
spin_lock_bh(&adapter->mcc_lock);
be_wrb_cmd_hdr_prepare(&req->hdr, CMD_SUBSYSTEM_COMMON,
OPCODE_COMMON_MODIFY_EQ_DELAY, sizeof(*req), wrb, NULL);
- req->num_eq = cpu_to_le32(1);
- req->delay[0].eq_id = cpu_to_le32(eq_id);
- req->delay[0].phase = 0;
- req->delay[0].delay_multiplier = cpu_to_le32(eqd);
+ req->num_eq = cpu_to_le32(num);
+ for (i = 0; i < num; i++) {
+ req->set_eqd[i].eq_id = cpu_to_le32(set_eqd[i].eq_id);
+ req->set_eqd[i].phase = 0;
+ req->set_eqd[i].delay_multiplier =
+ cpu_to_le32(set_eqd[i].delay_multiplier);
+ }
be_mcc_notify(adapter);
-
err:
spin_unlock_bh(&adapter->mcc_lock);
return status;
} __packed;
/******************** Modify EQ Delay *******************/
+struct be_set_eqd {
+ u32 eq_id;
+ u32 phase;
+ u32 delay_multiplier;
+};
+
struct be_cmd_req_modify_eq_delay {
struct be_cmd_req_hdr hdr;
u32 num_eq;
- struct {
- u32 eq_id;
- u32 phase;
- u32 delay_multiplier;
- } delay[8];
+ struct be_set_eqd set_eqd[MAX_EVT_QS];
} __packed;
struct be_cmd_resp_modify_eq_delay {
struct be_dma_mem *nonemb_cmd);
int be_cmd_get_fw_ver(struct be_adapter *adapter, char *fw_ver,
char *fw_on_flash);
-
-int be_cmd_modify_eqd(struct be_adapter *adapter, u32 eq_id, u32 eqd);
+int be_cmd_modify_eqd(struct be_adapter *adapter, struct be_set_eqd *, int num);
int be_cmd_vlan_config(struct be_adapter *adapter, u32 if_id, u16 *vtag_array,
u32 num, bool untagged, bool promiscuous);
int be_cmd_rx_filter(struct be_adapter *adapter, u32 flags, u32 status);
struct ethtool_coalesce *et)
{
struct be_adapter *adapter = netdev_priv(netdev);
- struct be_eq_obj *eqo = &adapter->eq_obj[0];
+ struct be_aic_obj *aic = &adapter->aic_obj[0];
- et->rx_coalesce_usecs = eqo->cur_eqd;
- et->rx_coalesce_usecs_high = eqo->max_eqd;
- et->rx_coalesce_usecs_low = eqo->min_eqd;
+ et->rx_coalesce_usecs = aic->prev_eqd;
+ et->rx_coalesce_usecs_high = aic->max_eqd;
+ et->rx_coalesce_usecs_low = aic->min_eqd;
- et->tx_coalesce_usecs = eqo->cur_eqd;
- et->tx_coalesce_usecs_high = eqo->max_eqd;
- et->tx_coalesce_usecs_low = eqo->min_eqd;
+ et->tx_coalesce_usecs = aic->prev_eqd;
+ et->tx_coalesce_usecs_high = aic->max_eqd;
+ et->tx_coalesce_usecs_low = aic->min_eqd;
- et->use_adaptive_rx_coalesce = eqo->enable_aic;
- et->use_adaptive_tx_coalesce = eqo->enable_aic;
+ et->use_adaptive_rx_coalesce = aic->enable;
+ et->use_adaptive_tx_coalesce = aic->enable;
return 0;
}
struct ethtool_coalesce *et)
{
struct be_adapter *adapter = netdev_priv(netdev);
+ struct be_aic_obj *aic = &adapter->aic_obj[0];
struct be_eq_obj *eqo;
int i;
for_all_evt_queues(adapter, eqo, i) {
- eqo->enable_aic = et->use_adaptive_rx_coalesce;
- eqo->max_eqd = min(et->rx_coalesce_usecs_high, BE_MAX_EQD);
- eqo->min_eqd = min(et->rx_coalesce_usecs_low, eqo->max_eqd);
- eqo->eqd = et->rx_coalesce_usecs;
+ aic->enable = et->use_adaptive_rx_coalesce;
+ aic->max_eqd = min(et->rx_coalesce_usecs_high, BE_MAX_EQD);
+ aic->min_eqd = min(et->rx_coalesce_usecs_low, aic->max_eqd);
+ aic->et_eqd = min(et->rx_coalesce_usecs, aic->max_eqd);
+ aic->et_eqd = max(aic->et_eqd, aic->min_eqd);
+ aic++;
}
return 0;
return status;
}
-static void be_eqd_update(struct be_adapter *adapter, struct be_eq_obj *eqo)
+static void be_aic_update(struct be_aic_obj *aic, u64 rx_pkts, u64 tx_pkts,
+ ulong now)
{
- struct be_rx_stats *stats = rx_stats(&adapter->rx_obj[eqo->idx]);
- ulong now = jiffies;
- ulong delta = now - stats->rx_jiffies;
- u64 pkts;
- unsigned int start, eqd;
+ aic->rx_pkts_prev = rx_pkts;
+ aic->tx_reqs_prev = tx_pkts;
+ aic->jiffies = now;
+}
- if (!eqo->enable_aic) {
- eqd = eqo->eqd;
- goto modify_eqd;
- }
+static void be_eqd_update(struct be_adapter *adapter)
+{
+ struct be_set_eqd set_eqd[MAX_EVT_QS];
+ int eqd, i, num = 0, start;
+ struct be_aic_obj *aic;
+ struct be_eq_obj *eqo;
+ struct be_rx_obj *rxo;
+ struct be_tx_obj *txo;
+ u64 rx_pkts, tx_pkts;
+ ulong now;
+ u32 pps, delta;
- if (eqo->idx >= adapter->num_rx_qs)
- return;
+ for_all_evt_queues(adapter, eqo, i) {
+ aic = &adapter->aic_obj[eqo->idx];
+ if (!aic->enable) {
+ if (aic->jiffies)
+ aic->jiffies = 0;
+ eqd = aic->et_eqd;
+ goto modify_eqd;
+ }
- stats = rx_stats(&adapter->rx_obj[eqo->idx]);
+ rxo = &adapter->rx_obj[eqo->idx];
+ do {
+ start = u64_stats_fetch_begin_bh(&rxo->stats.sync);
+ rx_pkts = rxo->stats.rx_pkts;
+ } while (u64_stats_fetch_retry_bh(&rxo->stats.sync, start));
- /* Wrapped around */
- if (time_before(now, stats->rx_jiffies)) {
- stats->rx_jiffies = now;
- return;
- }
+ txo = &adapter->tx_obj[eqo->idx];
+ do {
+ start = u64_stats_fetch_begin_bh(&txo->stats.sync);
+ tx_pkts = txo->stats.tx_reqs;
+ } while (u64_stats_fetch_retry_bh(&txo->stats.sync, start));
- /* Update once a second */
- if (delta < HZ)
- return;
- do {
- start = u64_stats_fetch_begin_bh(&stats->sync);
- pkts = stats->rx_pkts;
- } while (u64_stats_fetch_retry_bh(&stats->sync, start));
-
- stats->rx_pps = (unsigned long)(pkts - stats->rx_pkts_prev) / (delta / HZ);
- stats->rx_pkts_prev = pkts;
- stats->rx_jiffies = now;
- eqd = (stats->rx_pps / 110000) << 3;
- eqd = min(eqd, eqo->max_eqd);
- eqd = max(eqd, eqo->min_eqd);
- if (eqd < 10)
- eqd = 0;
+ /* Skip, if wrapped around or first calculation */
+ now = jiffies;
+ if (!aic->jiffies || time_before(now, aic->jiffies) ||
+ rx_pkts < aic->rx_pkts_prev ||
+ tx_pkts < aic->tx_reqs_prev) {
+ be_aic_update(aic, rx_pkts, tx_pkts, now);
+ continue;
+ }
+
+ delta = jiffies_to_msecs(now - aic->jiffies);
+ pps = (((u32)(rx_pkts - aic->rx_pkts_prev) * 1000) / delta) +
+ (((u32)(tx_pkts - aic->tx_reqs_prev) * 1000) / delta);
+ eqd = (pps / 15000) << 2;
+ if (eqd < 8)
+ eqd = 0;
+ eqd = min_t(u32, eqd, aic->max_eqd);
+ eqd = max_t(u32, eqd, aic->min_eqd);
+
+ be_aic_update(aic, rx_pkts, tx_pkts, now);
modify_eqd:
- if (eqd != eqo->cur_eqd) {
- be_cmd_modify_eqd(adapter, eqo->q.id, eqd);
- eqo->cur_eqd = eqd;
+ if (eqd != aic->prev_eqd) {
+ set_eqd[num].delay_multiplier = (eqd * 65)/100;
+ set_eqd[num].eq_id = eqo->q.id;
+ aic->prev_eqd = eqd;
+ num++;
+ }
}
+
+ if (num)
+ be_cmd_modify_eqd(adapter, set_eqd, num);
}
static void be_rx_stats_update(struct be_rx_obj *rxo,
{
struct be_queue_info *eq;
struct be_eq_obj *eqo;
+ struct be_aic_obj *aic;
int i, rc;
adapter->num_evt_qs = min_t(u16, num_irqs(adapter),
for_all_evt_queues(adapter, eqo, i) {
netif_napi_add(adapter->netdev, &eqo->napi, be_poll,
BE_NAPI_WEIGHT);
+ aic = &adapter->aic_obj[i];
eqo->adapter = adapter;
eqo->tx_budget = BE_TX_BUDGET;
eqo->idx = i;
- eqo->max_eqd = BE_MAX_EQD;
- eqo->enable_aic = true;
+ aic->max_eqd = BE_MAX_EQD;
+ aic->enable = true;
eq = &eqo->q;
rc = be_queue_alloc(adapter, eq, EVNT_Q_LEN,
struct be_adapter *adapter =
container_of(work, struct be_adapter, work.work);
struct be_rx_obj *rxo;
- struct be_eq_obj *eqo;
int i;
/* when interrupts are not yet enabled, just reap any pending
}
}
- for_all_evt_queues(adapter, eqo, i)
- be_eqd_update(adapter, eqo);
+ be_eqd_update(adapter);
reschedule:
adapter->work_counter++;