+/*
+ * stop tasklet and free all pending skb's
+ */
+static void tx_sched_stop(struct sge *sge)
+{
+ struct sched *s = sge->tx_sched;
+ int i;
+
+ tasklet_kill(&s->sched_tsk);
+
+ for (i = 0; i < MAX_NPORTS; i++)
+ __skb_queue_purge(&s->p[s->port].skbq);
+}
+
+/*
+ * t1_sched_update_parms() is called when the MTU or link speed changes. It
+ * re-computes scheduler parameters to scope with the change.
+ */
+unsigned int t1_sched_update_parms(struct sge *sge, unsigned int port,
+ unsigned int mtu, unsigned int speed)
+{
+ struct sched *s = sge->tx_sched;
+ struct sched_port *p = &s->p[port];
+ unsigned int max_avail_segs;
+
+ pr_debug("t1_sched_update_params mtu=%d speed=%d\n", mtu, speed);
+ if (speed)
+ p->speed = speed;
+ if (mtu)
+ p->mtu = mtu;
+
+ if (speed || mtu) {
+ unsigned long long drain = 1024ULL * p->speed * (p->mtu - 40);
+ do_div(drain, (p->mtu + 50) * 1000);
+ p->drain_bits_per_1024ns = (unsigned int) drain;
+
+ if (p->speed < 1000)
+ p->drain_bits_per_1024ns =
+ 90 * p->drain_bits_per_1024ns / 100;
+ }
+
+ if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204) {
+ p->drain_bits_per_1024ns -= 16;
+ s->max_avail = max(4096U, p->mtu + 16 + 14 + 4);
+ max_avail_segs = max(1U, 4096 / (p->mtu - 40));
+ } else {
+ s->max_avail = 16384;
+ max_avail_segs = max(1U, 9000 / (p->mtu - 40));
+ }
+
+ pr_debug("t1_sched_update_parms: mtu %u speed %u max_avail %u "
+ "max_avail_segs %u drain_bits_per_1024ns %u\n", p->mtu,
+ p->speed, s->max_avail, max_avail_segs,
+ p->drain_bits_per_1024ns);
+
+ return max_avail_segs * (p->mtu - 40);
+}
+
+/*
+ * t1_sched_max_avail_bytes() tells the scheduler the maximum amount of
+ * data that can be pushed per port.
+ */
+void t1_sched_set_max_avail_bytes(struct sge *sge, unsigned int val)
+{
+ struct sched *s = sge->tx_sched;
+ unsigned int i;
+
+ s->max_avail = val;
+ for (i = 0; i < MAX_NPORTS; i++)
+ t1_sched_update_parms(sge, i, 0, 0);
+}
+
+/*
+ * t1_sched_set_drain_bits_per_us() tells the scheduler at which rate a port
+ * is draining.
+ */
+void t1_sched_set_drain_bits_per_us(struct sge *sge, unsigned int port,
+ unsigned int val)
+{
+ struct sched *s = sge->tx_sched;
+ struct sched_port *p = &s->p[port];
+ p->drain_bits_per_1024ns = val * 1024 / 1000;
+ t1_sched_update_parms(sge, port, 0, 0);
+}
+
+
+/*
+ * get_clock() implements a ns clock (see ktime_get)
+ */
+static inline ktime_t get_clock(void)
+{
+ struct timespec ts;
+
+ ktime_get_ts(&ts);
+ return timespec_to_ktime(ts);
+}
+
+/*
+ * tx_sched_init() allocates resources and does basic initialization.
+ */
+static int tx_sched_init(struct sge *sge)
+{
+ struct sched *s;
+ int i;
+
+ s = kzalloc(sizeof (struct sched), GFP_KERNEL);
+ if (!s)
+ return -ENOMEM;
+
+ pr_debug("tx_sched_init\n");
+ tasklet_init(&s->sched_tsk, restart_sched, (unsigned long) sge);
+ sge->tx_sched = s;
+
+ for (i = 0; i < MAX_NPORTS; i++) {
+ skb_queue_head_init(&s->p[i].skbq);
+ t1_sched_update_parms(sge, i, 1500, 1000);
+ }
+
+ return 0;
+}
+
+/*
+ * sched_update_avail() computes the delta since the last time it was called
+ * and updates the per port quota (number of bits that can be sent to the any
+ * port).
+ */
+static inline int sched_update_avail(struct sge *sge)
+{
+ struct sched *s = sge->tx_sched;
+ ktime_t now = get_clock();
+ unsigned int i;
+ long long delta_time_ns;
+
+ delta_time_ns = ktime_to_ns(ktime_sub(now, s->last_updated));
+
+ pr_debug("sched_update_avail delta=%lld\n", delta_time_ns);
+ if (delta_time_ns < 15000)
+ return 0;
+
+ for (i = 0; i < MAX_NPORTS; i++) {
+ struct sched_port *p = &s->p[i];
+ unsigned int delta_avail;
+
+ delta_avail = (p->drain_bits_per_1024ns * delta_time_ns) >> 13;
+ p->avail = min(p->avail + delta_avail, s->max_avail);
+ }
+
+ s->last_updated = now;
+
+ return 1;
+}
+
+/*
+ * sched_skb() is called from two different places. In the tx path, any
+ * packet generating load on an output port will call sched_skb()
+ * (skb != NULL). In addition, sched_skb() is called from the irq/soft irq
+ * context (skb == NULL).
+ * The scheduler only returns a skb (which will then be sent) if the
+ * length of the skb is <= the current quota of the output port.
+ */
+static struct sk_buff *sched_skb(struct sge *sge, struct sk_buff *skb,
+ unsigned int credits)
+{
+ struct sched *s = sge->tx_sched;
+ struct sk_buff_head *skbq;
+ unsigned int i, len, update = 1;
+
+ pr_debug("sched_skb %p\n", skb);
+ if (!skb) {
+ if (!s->num)
+ return NULL;
+ } else {
+ skbq = &s->p[skb->dev->if_port].skbq;
+ __skb_queue_tail(skbq, skb);
+ s->num++;
+ skb = NULL;
+ }
+
+ if (credits < MAX_SKB_FRAGS + 1)
+ goto out;
+
+ again:
+ for (i = 0; i < MAX_NPORTS; i++) {
+ s->port = ++s->port & (MAX_NPORTS - 1);
+ skbq = &s->p[s->port].skbq;
+
+ skb = skb_peek(skbq);
+
+ if (!skb)
+ continue;
+
+ len = skb->len;
+ if (len <= s->p[s->port].avail) {
+ s->p[s->port].avail -= len;
+ s->num--;
+ __skb_unlink(skb, skbq);
+ goto out;
+ }
+ skb = NULL;
+ }
+
+ if (update-- && sched_update_avail(sge))
+ goto again;
+
+ out:
+ /* If there are more pending skbs, we use the hardware to schedule us
+ * again.
+ */
+ if (s->num && !skb) {
+ struct cmdQ *q = &sge->cmdQ[0];
+ clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
+ if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
+ set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
+ writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
+ }
+ }
+ pr_debug("sched_skb ret %p\n", skb);
+
+ return skb;
+}
+