Commit 5f07b3c5 authored by Sathya Perla's avatar Sathya Perla Committed by David S. Miller

be2net: support TX batching using skb->xmit_more flag

This patch uses skb->xmit_more flag to batch TX requests.
TX is flushed either when xmit_more is false or there is
no more space in the TXQ.

Skyhawk-R and BEx chips require an even number of wrbs to be posted.
So, when a batch of TX requests is accumulated, the last header wrb
may need to be fixed with an extra dummy wrb.

This patch refactors be_xmit() routine as a sequence of be_xmit_enqueue()
and be_xmit_flush() calls. The Tx completion code is also
updated to be able to unmap/free a batch of skbs rather than a single
skb.
Signed-off-by: default avatarSathya Perla <sathya.perla@emulex.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent 889ee2c7
......@@ -243,7 +243,6 @@ struct be_tx_stats {
u64 tx_bytes;
u64 tx_pkts;
u64 tx_reqs;
u64 tx_wrbs;
u64 tx_compl;
ulong tx_jiffies;
u32 tx_stops;
......@@ -266,6 +265,9 @@ struct be_tx_obj {
/* Remember the skbs that were transmitted */
struct sk_buff *sent_skb_list[TX_Q_LEN];
struct be_tx_stats stats;
u16 pend_wrb_cnt; /* Number of WRBs yet to be given to HW */
u16 last_req_wrb_cnt; /* wrb cnt of the last req in the Q */
u16 last_req_hdr; /* index of the last req's hdr-wrb */
} ____cacheline_aligned_in_smp;
/* Struct to remember the pages posted for rx frags */
......
......@@ -193,8 +193,6 @@ static const struct be_ethtool_stat et_tx_stats[] = {
{DRVSTAT_TX_INFO(tx_pkts)},
/* Number of skbs queued for trasmission by the driver */
{DRVSTAT_TX_INFO(tx_reqs)},
/* Number of TX work request blocks DMAed to HW */
{DRVSTAT_TX_INFO(tx_wrbs)},
/* Number of times the TX queue was stopped due to lack
* of spaces in the TXQ.
*/
......
......@@ -311,6 +311,11 @@ struct amap_eth_hdr_wrb {
u8 vlan_tag[16];
} __packed;
#define TX_HDR_WRB_COMPL 1 /* word 2 */
#define TX_HDR_WRB_EVT (1 << 1) /* word 2 */
#define TX_HDR_WRB_NUM_SHIFT 13 /* word 2: bits 13:17 */
#define TX_HDR_WRB_NUM_MASK 0x1F /* word 2: bits 13:17 */
struct be_eth_hdr_wrb {
u32 dw[4];
};
......
......@@ -662,41 +662,22 @@ void be_link_status_update(struct be_adapter *adapter, u8 link_status)
netif_carrier_off(netdev);
}
static void be_tx_stats_update(struct be_tx_obj *txo,
u32 wrb_cnt, u32 copied, u32 gso_segs,
bool stopped)
static void be_tx_stats_update(struct be_tx_obj *txo, struct sk_buff *skb)
{
struct be_tx_stats *stats = tx_stats(txo);
u64_stats_update_begin(&stats->sync);
stats->tx_reqs++;
stats->tx_wrbs += wrb_cnt;
stats->tx_bytes += copied;
stats->tx_pkts += (gso_segs ? gso_segs : 1);
if (stopped)
stats->tx_stops++;
stats->tx_bytes += skb->len;
stats->tx_pkts += (skb_shinfo(skb)->gso_segs ? : 1);
u64_stats_update_end(&stats->sync);
}
/* Determine number of WRB entries needed to xmit data in an skb */
static u32 wrb_cnt_for_skb(struct be_adapter *adapter, struct sk_buff *skb,
bool *dummy)
/* Returns number of WRBs needed for the skb */
static u32 skb_wrb_cnt(struct sk_buff *skb)
{
int cnt = (skb->len > skb->data_len);
cnt += skb_shinfo(skb)->nr_frags;
/* to account for hdr wrb */
cnt++;
if (lancer_chip(adapter) || !(cnt & 1)) {
*dummy = false;
} else {
/* add a dummy to make it an even num */
cnt++;
*dummy = true;
}
BUG_ON(cnt > BE_MAX_TX_FRAG_COUNT);
return cnt;
/* +1 for the header wrb */
return 1 + (skb_headlen(skb) ? 1 : 0) + skb_shinfo(skb)->nr_frags;
}
static inline void wrb_fill(struct be_eth_wrb *wrb, u64 addr, int len)
......@@ -770,11 +751,14 @@ static void wrb_fill_hdr(struct be_adapter *adapter, struct be_eth_hdr_wrb *hdr,
SET_TX_WRB_HDR_BITS(vlan_tag, hdr, vlan_tag);
}
/* To skip HW VLAN tagging: evt = 1, compl = 0 */
SET_TX_WRB_HDR_BITS(complete, hdr, !skip_hw_vlan);
SET_TX_WRB_HDR_BITS(event, hdr, 1);
SET_TX_WRB_HDR_BITS(num_wrb, hdr, wrb_cnt);
SET_TX_WRB_HDR_BITS(len, hdr, len);
/* Hack to skip HW VLAN tagging needs evt = 1, compl = 0
* When this hack is not needed, the evt bit is set while ringing DB
*/
if (skip_hw_vlan)
SET_TX_WRB_HDR_BITS(event, hdr, 1);
}
static void unmap_tx_frag(struct device *dev, struct be_eth_wrb *wrb,
......@@ -794,22 +778,24 @@ static void unmap_tx_frag(struct device *dev, struct be_eth_wrb *wrb,
}
}
static int make_tx_wrbs(struct be_adapter *adapter, struct be_queue_info *txq,
struct sk_buff *skb, u32 wrb_cnt, bool dummy_wrb,
bool skip_hw_vlan)
/* Returns the number of WRBs used up by the skb */
static u32 be_xmit_enqueue(struct be_adapter *adapter, struct be_tx_obj *txo,
struct sk_buff *skb, bool skip_hw_vlan)
{
dma_addr_t busaddr;
int i, copied = 0;
u32 i, copied = 0, wrb_cnt = skb_wrb_cnt(skb);
struct device *dev = &adapter->pdev->dev;
struct sk_buff *first_skb = skb;
struct be_eth_wrb *wrb;
struct be_queue_info *txq = &txo->q;
struct be_eth_hdr_wrb *hdr;
bool map_single = false;
u16 map_head;
struct be_eth_wrb *wrb;
dma_addr_t busaddr;
u16 head = txq->head;
hdr = queue_head_node(txq);
wrb_fill_hdr(adapter, hdr, skb, wrb_cnt, skb->len, skip_hw_vlan);
be_dws_cpu_to_le(hdr, sizeof(*hdr));
queue_head_inc(txq);
map_head = txq->head;
if (skb->len > skb->data_len) {
int len = skb_headlen(skb);
......@@ -839,19 +825,23 @@ static int make_tx_wrbs(struct be_adapter *adapter, struct be_queue_info *txq,
copied += skb_frag_size(frag);
}
if (dummy_wrb) {
wrb = queue_head_node(txq);
wrb_fill(wrb, 0, 0);
be_dws_cpu_to_le(wrb, sizeof(*wrb));
queue_head_inc(txq);
}
BUG_ON(txo->sent_skb_list[head]);
txo->sent_skb_list[head] = skb;
txo->last_req_hdr = head;
atomic_add(wrb_cnt, &txq->used);
txo->last_req_wrb_cnt = wrb_cnt;
txo->pend_wrb_cnt += wrb_cnt;
wrb_fill_hdr(adapter, hdr, first_skb, wrb_cnt, copied, skip_hw_vlan);
be_dws_cpu_to_le(hdr, sizeof(*hdr));
be_tx_stats_update(txo, skb);
return wrb_cnt;
return copied;
dma_err:
txq->head = map_head;
/* Bring the queue back to the state it was in before this
* routine was invoked.
*/
txq->head = head;
/* skip the first wrb (hdr); it's not mapped */
queue_head_inc(txq);
while (copied) {
wrb = queue_head_node(txq);
unmap_tx_frag(dev, wrb, map_single);
......@@ -860,6 +850,7 @@ static int make_tx_wrbs(struct be_adapter *adapter, struct be_queue_info *txq,
adapter->drv_stats.dma_map_errors++;
queue_head_inc(txq);
}
txq->head = head;
return 0;
}
......@@ -1030,52 +1021,64 @@ static struct sk_buff *be_xmit_workarounds(struct be_adapter *adapter,
return skb;
}
static void be_xmit_flush(struct be_adapter *adapter, struct be_tx_obj *txo)
{
struct be_queue_info *txq = &txo->q;
struct be_eth_hdr_wrb *hdr = queue_index_node(txq, txo->last_req_hdr);
/* Mark the last request eventable if it hasn't been marked already */
if (!(hdr->dw[2] & cpu_to_le32(TX_HDR_WRB_EVT)))
hdr->dw[2] |= cpu_to_le32(TX_HDR_WRB_EVT | TX_HDR_WRB_COMPL);
/* compose a dummy wrb if there are odd set of wrbs to notify */
if (!lancer_chip(adapter) && (txo->pend_wrb_cnt & 1)) {
wrb_fill(queue_head_node(txq), 0, 0);
queue_head_inc(txq);
atomic_inc(&txq->used);
txo->pend_wrb_cnt++;
hdr->dw[2] &= ~cpu_to_le32(TX_HDR_WRB_NUM_MASK <<
TX_HDR_WRB_NUM_SHIFT);
hdr->dw[2] |= cpu_to_le32((txo->last_req_wrb_cnt + 1) <<
TX_HDR_WRB_NUM_SHIFT);
}
be_txq_notify(adapter, txo, txo->pend_wrb_cnt);
txo->pend_wrb_cnt = 0;
}
static netdev_tx_t be_xmit(struct sk_buff *skb, struct net_device *netdev)
{
bool skip_hw_vlan = false, flush = !skb->xmit_more;
struct be_adapter *adapter = netdev_priv(netdev);
struct be_tx_obj *txo = &adapter->tx_obj[skb_get_queue_mapping(skb)];
u16 q_idx = skb_get_queue_mapping(skb);
struct be_tx_obj *txo = &adapter->tx_obj[q_idx];
struct be_queue_info *txq = &txo->q;
bool dummy_wrb, stopped = false;
u32 wrb_cnt = 0, copied = 0;
bool skip_hw_vlan = false;
u32 start = txq->head;
u16 wrb_cnt;
skb = be_xmit_workarounds(adapter, skb, &skip_hw_vlan);
if (!skb) {
tx_stats(txo)->tx_drv_drops++;
return NETDEV_TX_OK;
}
wrb_cnt = wrb_cnt_for_skb(adapter, skb, &dummy_wrb);
if (unlikely(!skb))
goto drop;
copied = make_tx_wrbs(adapter, txq, skb, wrb_cnt, dummy_wrb,
skip_hw_vlan);
if (copied) {
int gso_segs = skb_shinfo(skb)->gso_segs;
wrb_cnt = be_xmit_enqueue(adapter, txo, skb, skip_hw_vlan);
if (unlikely(!wrb_cnt)) {
dev_kfree_skb_any(skb);
goto drop;
}
/* record the sent skb in the sent_skb table */
BUG_ON(txo->sent_skb_list[start]);
txo->sent_skb_list[start] = skb;
if ((atomic_read(&txq->used) + BE_MAX_TX_FRAG_COUNT) >= txq->len) {
netif_stop_subqueue(netdev, q_idx);
tx_stats(txo)->tx_stops++;
}
/* Ensure txq has space for the next skb; Else stop the queue
* *BEFORE* ringing the tx doorbell, so that we serialze the
* tx compls of the current transmit which'll wake up the queue
*/
atomic_add(wrb_cnt, &txq->used);
if ((BE_MAX_TX_FRAG_COUNT + atomic_read(&txq->used)) >=
txq->len) {
netif_stop_subqueue(netdev, skb_get_queue_mapping(skb));
stopped = true;
}
if (flush || __netif_subqueue_stopped(netdev, q_idx))
be_xmit_flush(adapter, txo);
be_txq_notify(adapter, txo, wrb_cnt);
return NETDEV_TX_OK;
drop:
tx_stats(txo)->tx_drv_drops++;
/* Flush the already enqueued tx requests */
if (flush && txo->pend_wrb_cnt)
be_xmit_flush(adapter, txo);
be_tx_stats_update(txo, wrb_cnt, copied, gso_segs, stopped);
} else {
txq->head = start;
tx_stats(txo)->tx_drv_drops++;
dev_kfree_skb_any(skb);
}
return NETDEV_TX_OK;
}
......@@ -1959,32 +1962,34 @@ static struct be_eth_tx_compl *be_tx_compl_get(struct be_queue_info *tx_cq)
static u16 be_tx_compl_process(struct be_adapter *adapter,
struct be_tx_obj *txo, u16 last_index)
{
struct sk_buff **sent_skbs = txo->sent_skb_list;
struct be_queue_info *txq = &txo->q;
u16 frag_index, num_wrbs = 0;
struct sk_buff *skb = NULL;
bool unmap_skb_hdr = false;
struct be_eth_wrb *wrb;
struct sk_buff **sent_skbs = txo->sent_skb_list;
struct sk_buff *sent_skb;
u16 cur_index, num_wrbs = 1; /* account for hdr wrb */
bool unmap_skb_hdr = true;
sent_skb = sent_skbs[txq->tail];
BUG_ON(!sent_skb);
sent_skbs[txq->tail] = NULL;
/* skip header wrb */
queue_tail_inc(txq);
do {
cur_index = txq->tail;
if (sent_skbs[txq->tail]) {
/* Free skb from prev req */
if (skb)
dev_consume_skb_any(skb);
skb = sent_skbs[txq->tail];
sent_skbs[txq->tail] = NULL;
queue_tail_inc(txq); /* skip hdr wrb */
num_wrbs++;
unmap_skb_hdr = true;
}
wrb = queue_tail_node(txq);
frag_index = txq->tail;
unmap_tx_frag(&adapter->pdev->dev, wrb,
(unmap_skb_hdr && skb_headlen(sent_skb)));
(unmap_skb_hdr && skb_headlen(skb)));
unmap_skb_hdr = false;
num_wrbs++;
queue_tail_inc(txq);
} while (cur_index != last_index);
num_wrbs++;
} while (frag_index != last_index);
dev_consume_skb_any(skb);
dev_consume_skb_any(sent_skb);
return num_wrbs;
}
......@@ -2068,12 +2073,11 @@ static void be_rx_cq_clean(struct be_rx_obj *rxo)
static void be_tx_compl_clean(struct be_adapter *adapter)
{
u16 end_idx, notified_idx, cmpl = 0, timeo = 0, num_wrbs = 0;
struct device *dev = &adapter->pdev->dev;
struct be_tx_obj *txo;
struct be_queue_info *txq;
struct be_eth_tx_compl *txcp;
u16 end_idx, cmpl = 0, timeo = 0, num_wrbs = 0;
struct sk_buff *sent_skb;
bool dummy_wrb;
int i, pending_txqs;
/* Stop polling for compls when HW has been silent for 10ms */
......@@ -2095,7 +2099,7 @@ static void be_tx_compl_clean(struct be_adapter *adapter)
atomic_sub(num_wrbs, &txq->used);
timeo = 0;
}
if (atomic_read(&txq->used) == 0)
if (atomic_read(&txq->used) == txo->pend_wrb_cnt)
pending_txqs--;
}
......@@ -2105,21 +2109,29 @@ static void be_tx_compl_clean(struct be_adapter *adapter)
mdelay(1);
} while (true);
/* Free enqueued TX that was never notified to HW */
for_all_tx_queues(adapter, txo, i) {
txq = &txo->q;
if (atomic_read(&txq->used))
dev_err(&adapter->pdev->dev, "%d pending tx-compls\n",
atomic_read(&txq->used));
/* free posted tx for which compls will never arrive */
while (atomic_read(&txq->used)) {
sent_skb = txo->sent_skb_list[txq->tail];
if (atomic_read(&txq->used)) {
dev_info(dev, "txq%d: cleaning %d pending tx-wrbs\n",
i, atomic_read(&txq->used));
notified_idx = txq->tail;
end_idx = txq->tail;
num_wrbs = wrb_cnt_for_skb(adapter, sent_skb,
&dummy_wrb);
index_adv(&end_idx, num_wrbs - 1, txq->len);
index_adv(&end_idx, atomic_read(&txq->used) - 1,
txq->len);
/* Use the tx-compl process logic to handle requests
* that were not sent to the HW.
*/
num_wrbs = be_tx_compl_process(adapter, txo, end_idx);
atomic_sub(num_wrbs, &txq->used);
BUG_ON(atomic_read(&txq->used));
txo->pend_wrb_cnt = 0;
/* Since hw was never notified of these requests,
* reset TXQ indices
*/
txq->head = notified_idx;
txq->tail = notified_idx;
}
}
}
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment