sge.c

linux-2.6.15.6

}

/*
 * Main interrupt handler, optimized assuming that we took a 'DATA'
 * interrupt.
 *
 * 1. Clear the interrupt
 * 2. Loop while we find valid descriptors and process them; accumulate
 *      information that can be processed after the loop
 * 3. Tell the SGE at which index we stopped processing descriptors
 * 4. Bookkeeping; free TX buffers, ring doorbell if there are any
 *      outstanding TX buffers waiting, replenish RX buffers, potentially
 *      reenable upper layers if they were turned off due to lack of TX
 *      resources which are available again.
 * 5. If we took an interrupt, but no valid respQ descriptors was found we
 *      let the slow_intr_handler run and do error handling.
 */
static irqreturn_t t1_interrupt(int irq, void *cookie, struct pt_regs *regs)
{
	int work_done;
	struct respQ_e *e;
	struct adapter *adapter = cookie;
	struct respQ *Q = &adapter->sge->respQ;

	spin_lock(&adapter->async_lock);
	e = &Q->entries[Q->cidx];
	prefetch(e);

	writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);

	if (likely(e->GenerationBit == Q->genbit))
		work_done = process_responses(adapter, -1);
	else
		work_done = t1_slow_intr_handler(adapter);

	/*
	 * The unconditional clearing of the PL_CAUSE above may have raced
	 * with DMA completion and the corresponding generation of a response
	 * to cause us to miss the resulting data interrupt.  The next write
	 * is also unconditional to recover the missed interrupt and render
	 * this race harmless.
	 */
	writel(Q->cidx, adapter->regs + A_SG_SLEEPING);

	if (!work_done)
		adapter->sge->stats.unhandled_irqs++;
	spin_unlock(&adapter->async_lock);
	return IRQ_RETVAL(work_done != 0);
}

intr_handler_t t1_select_intr_handler(adapter_t *adapter)
{
	return adapter->params.sge.polling ? t1_interrupt_napi : t1_interrupt;
}

/*
 * Enqueues the sk_buff onto the cmdQ[qid] and has hardware fetch it.
 *
 * The code figures out how many entries the sk_buff will require in the
 * cmdQ and updates the cmdQ data structure with the state once the enqueue
 * has complete. Then, it doesn't access the global structure anymore, but
 * uses the corresponding fields on the stack. In conjuction with a spinlock
 * around that code, we can make the function reentrant without holding the
 * lock when we actually enqueue (which might be expensive, especially on
 * architectures with IO MMUs).
 *
 * This runs with softirqs disabled.
 */
unsigned int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter,
		       unsigned int qid, struct net_device *dev)
{
	struct sge *sge = adapter->sge;
	struct cmdQ *q = &sge->cmdQ[qid];
	unsigned int credits, pidx, genbit, count;

	spin_lock(&q->lock);
	reclaim_completed_tx(sge, q);

	pidx = q->pidx;
	credits = q->size - q->in_use;
	count = 1 + skb_shinfo(skb)->nr_frags;

	{	/* Ethernet packet */
	 	if (unlikely(credits < count)) {
			netif_stop_queue(dev);
			set_bit(dev->if_port, &sge->stopped_tx_queues);
			sge->stats.cmdQ_full[3]++;
			spin_unlock(&q->lock);
			CH_ERR("%s: Tx ring full while queue awake!\n",
			       adapter->name);
			return 1;
		}
		if (unlikely(credits - count < q->stop_thres)) {
			sge->stats.cmdQ_full[3]++;
			netif_stop_queue(dev);
			set_bit(dev->if_port, &sge->stopped_tx_queues);
		}
	}
	q->in_use += count;
	genbit = q->genbit;
	q->pidx += count;
	if (q->pidx >= q->size) {
		q->pidx -= q->size;
		q->genbit ^= 1;
	}
	spin_unlock(&q->lock);

	write_tx_descs(adapter, skb, pidx, genbit, q);

	/*
	 * We always ring the doorbell for cmdQ1.  For cmdQ0, we only ring
	 * the doorbell if the Q is asleep. There is a natural race, where
	 * the hardware is going to sleep just after we checked, however,
	 * then the interrupt handler will detect the outstanding TX packet
	 * and ring the doorbell for us.
	 */
	if (qid)
		doorbell_pio(adapter, F_CMDQ1_ENABLE);
	else {
		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, adapter->regs + A_SG_DOORBELL);
		}
	}
	return 0;
}

#define MK_ETH_TYPE_MSS(type, mss) (((mss) & 0x3FFF) | ((type) << 14))

/*
 *	eth_hdr_len - return the length of an Ethernet header
 *	@data: pointer to the start of the Ethernet header
 *
 *	Returns the length of an Ethernet header, including optional VLAN tag.
 */
static inline int eth_hdr_len(const void *data)
{
	const struct ethhdr *e = data;

	return e->h_proto == htons(ETH_P_8021Q) ? VLAN_ETH_HLEN : ETH_HLEN;
}

/*
 * Adds the CPL header to the sk_buff and passes it to t1_sge_tx.
 */
int t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
{
	struct adapter *adapter = dev->priv;
	struct sge_port_stats *st = &adapter->sge->port_stats[dev->if_port];
	struct sge *sge = adapter->sge;
	struct cpl_tx_pkt *cpl;

#ifdef NETIF_F_TSO
	if (skb_shinfo(skb)->tso_size) {
		int eth_type;
		struct cpl_tx_pkt_lso *hdr;

		st->tso++;

		eth_type = skb->nh.raw - skb->data == ETH_HLEN ?
			CPL_ETH_II : CPL_ETH_II_VLAN;

		hdr = (struct cpl_tx_pkt_lso *)skb_push(skb, sizeof(*hdr));
		hdr->opcode = CPL_TX_PKT_LSO;
		hdr->ip_csum_dis = hdr->l4_csum_dis = 0;
		hdr->ip_hdr_words = skb->nh.iph->ihl;
		hdr->tcp_hdr_words = skb->h.th->doff;
		hdr->eth_type_mss = htons(MK_ETH_TYPE_MSS(eth_type,
						skb_shinfo(skb)->tso_size));
		hdr->len = htonl(skb->len - sizeof(*hdr));
		cpl = (struct cpl_tx_pkt *)hdr;
		sge->stats.tx_lso_pkts++;
	} else
#endif
	{
		/*
	 	 * Packets shorter than ETH_HLEN can break the MAC, drop them
		 * early.  Also, we may get oversized packets because some
		 * parts of the kernel don't handle our unusual hard_header_len
		 * right, drop those too.
		 */
		if (unlikely(skb->len < ETH_HLEN ||
			     skb->len > dev->mtu + eth_hdr_len(skb->data))) {
			dev_kfree_skb_any(skb);
			return NET_XMIT_SUCCESS;
		}

		/*
		 * We are using a non-standard hard_header_len and some kernel
		 * components, such as pktgen, do not handle it right.
		 * Complain when this happens but try to fix things up.
		 */
		if (unlikely(skb_headroom(skb) <
			     dev->hard_header_len - ETH_HLEN)) {
			struct sk_buff *orig_skb = skb;

			if (net_ratelimit())
				printk(KERN_ERR "%s: inadequate headroom in "
				       "Tx packet\n", dev->name);
			skb = skb_realloc_headroom(skb, sizeof(*cpl));
			dev_kfree_skb_any(orig_skb);
			if (!skb)
				return -ENOMEM;
		}

		if (!(adapter->flags & UDP_CSUM_CAPABLE) &&
		    skb->ip_summed == CHECKSUM_HW &&
		    skb->nh.iph->protocol == IPPROTO_UDP)
			if (unlikely(skb_checksum_help(skb, 0))) {
				dev_kfree_skb_any(skb);
				return -ENOMEM;
			}

		/* Hmmm, assuming to catch the gratious arp... and we'll use
		 * it to flush out stuck espi packets...
		  */
		if (unlikely(!adapter->sge->espibug_skb)) {
			if (skb->protocol == htons(ETH_P_ARP) &&
			    skb->nh.arph->ar_op == htons(ARPOP_REQUEST)) {
				adapter->sge->espibug_skb = skb;
				/* We want to re-use this skb later. We
				 * simply bump the reference count and it
				 * will not be freed...
				 */
				skb = skb_get(skb);
			}
		}

		cpl = (struct cpl_tx_pkt *)__skb_push(skb, sizeof(*cpl));
		cpl->opcode = CPL_TX_PKT;
		cpl->ip_csum_dis = 1;    /* SW calculates IP csum */
		cpl->l4_csum_dis = skb->ip_summed == CHECKSUM_HW ? 0 : 1;
		/* the length field isn't used so don't bother setting it */

		st->tx_cso += (skb->ip_summed == CHECKSUM_HW);
		sge->stats.tx_do_cksum += (skb->ip_summed == CHECKSUM_HW);
		sge->stats.tx_reg_pkts++;
	}
	cpl->iff = dev->if_port;

#if defined(CONFIG_VLAN_8021Q) || defined(CONFIG_VLAN_8021Q_MODULE)
	if (adapter->vlan_grp && vlan_tx_tag_present(skb)) {
		cpl->vlan_valid = 1;
		cpl->vlan = htons(vlan_tx_tag_get(skb));
		st->vlan_insert++;
	} else
#endif
		cpl->vlan_valid = 0;

	dev->trans_start = jiffies;
	return t1_sge_tx(skb, adapter, 0, dev);
}

/*
 * Callback for the Tx buffer reclaim timer.  Runs with softirqs disabled.
 */
static void sge_tx_reclaim_cb(unsigned long data)
{
	int i;
	struct sge *sge = (struct sge *)data;

	for (i = 0; i < SGE_CMDQ_N; ++i) {
		struct cmdQ *q = &sge->cmdQ[i];

		if (!spin_trylock(&q->lock))
			continue;

		reclaim_completed_tx(sge, q);
		if (i == 0 && q->in_use)   /* flush pending credits */
			writel(F_CMDQ0_ENABLE,
				sge->adapter->regs + A_SG_DOORBELL);

		spin_unlock(&q->lock);
	}
	mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
}

/*
 * Propagate changes of the SGE coalescing parameters to the HW.
 */
int t1_sge_set_coalesce_params(struct sge *sge, struct sge_params *p)
{
	sge->netdev->poll = t1_poll;
	sge->fixed_intrtimer = p->rx_coalesce_usecs *
		core_ticks_per_usec(sge->adapter);
	writel(sge->fixed_intrtimer, sge->adapter->regs + A_SG_INTRTIMER);
	return 0;
}

/*
 * Allocates both RX and TX resources and configures the SGE. However,
 * the hardware is not enabled yet.
 */
int t1_sge_configure(struct sge *sge, struct sge_params *p)
{
	if (alloc_rx_resources(sge, p))
		return -ENOMEM;
	if (alloc_tx_resources(sge, p)) {
		free_rx_resources(sge);
		return -ENOMEM;
	}
	configure_sge(sge, p);

	/*
	 * Now that we have sized the free lists calculate the payload
	 * capacity of the large buffers.  Other parts of the driver use
	 * this to set the max offload coalescing size so that RX packets
	 * do not overflow our large buffers.
	 */
	p->large_buf_capacity = jumbo_payload_capacity(sge);
	return 0;
}

/*
 * Disables the DMA engine.
 */
void t1_sge_stop(struct sge *sge)
{
	writel(0, sge->adapter->regs + A_SG_CONTROL);
	(void) readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
	if (is_T2(sge->adapter))
		del_timer_sync(&sge->espibug_timer);
	del_timer_sync(&sge->tx_reclaim_timer);
}

/*
 * Enables the DMA engine.
 */
void t1_sge_start(struct sge *sge)
{
	refill_free_list(sge, &sge->freelQ[0]);
	refill_free_list(sge, &sge->freelQ[1]);

	writel(sge->sge_control, sge->adapter->regs + A_SG_CONTROL);
	doorbell_pio(sge->adapter, F_FL0_ENABLE | F_FL1_ENABLE);
	(void) readl(sge->adapter->regs + A_SG_CONTROL); /* flush */

	mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);

	if (is_T2(sge->adapter)) 
		mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
}

/*
 * Callback for the T2 ESPI 'stuck packet feature' workaorund
 */
static void espibug_workaround(void *data)
{
	struct adapter *adapter = (struct adapter *)data;
	struct sge *sge = adapter->sge;

	if (netif_running(adapter->port[0].dev)) {
		struct sk_buff *skb = sge->espibug_skb;

		u32 seop = t1_espi_get_mon(adapter, 0x930, 0);

		if ((seop & 0xfff0fff) == 0xfff && skb) {
			if (!skb->cb[0]) {
				u8 ch_mac_addr[ETH_ALEN] =
				    {0x0, 0x7, 0x43, 0x0, 0x0, 0x0};
				memcpy(skb->data + sizeof(struct cpl_tx_pkt),
				    ch_mac_addr, ETH_ALEN);
				memcpy(skb->data + skb->len - 10, ch_mac_addr,
				    ETH_ALEN);
				skb->cb[0] = 0xff;
			}

			/* bump the reference count to avoid freeing of the
			 * skb once the DMA has completed.
			 */
			skb = skb_get(skb);
			t1_sge_tx(skb, adapter, 0, adapter->port[0].dev);
		}
	}
	mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
}

/*
 * Creates a t1_sge structure and returns suggested resource parameters.
 */
struct sge * __devinit t1_sge_create(struct adapter *adapter,
				     struct sge_params *p)
{
	struct sge *sge = kmalloc(sizeof(*sge), GFP_KERNEL);

	if (!sge)
		return NULL;
	memset(sge, 0, sizeof(*sge));

	sge->adapter = adapter;
	sge->netdev = adapter->port[0].dev;
	sge->rx_pkt_pad = t1_is_T1B(adapter) ? 0 : 2;
	sge->jumbo_fl = t1_is_T1B(adapter) ? 1 : 0;

	init_timer(&sge->tx_reclaim_timer);
	sge->tx_reclaim_timer.data = (unsigned long)sge;
	sge->tx_reclaim_timer.function = sge_tx_reclaim_cb;

	if (is_T2(sge->adapter)) {
		init_timer(&sge->espibug_timer);
		sge->espibug_timer.function = (void *)&espibug_workaround;
		sge->espibug_timer.data = (unsigned long)sge->adapter;
		sge->espibug_timeout = 1;
	}
	 

	p->cmdQ_size[0] = SGE_CMDQ0_E_N;
	p->cmdQ_size[1] = SGE_CMDQ1_E_N;
	p->freelQ_size[!sge->jumbo_fl] = SGE_FREEL_SIZE;
	p->freelQ_size[sge->jumbo_fl] = SGE_JUMBO_FREEL_SIZE;
	p->rx_coalesce_usecs =  50;
	p->coalesce_enable = 0;
	p->sample_interval_usecs = 0;
	p->polling = 0;

	return sge;
}