pasemi_mac: jumbo frame support
First cut at jumbo frame support. To support large MTU, one or several
separate channels must be allocated to calculate the TCP/UDP checksum
separately, since the mac lacks enough buffers to hold a whole packet
while it's being calculated.
Furthermore, it seems that a single function channel is not quite
enough to feed one of the 10Gig links, so allocate two channels for
XAUI interfaces.
Signed-off-by: Olof Johansson <olof@lixom.net>
Acked-by: Jeff Garzik <jgarzik@pobox.com>
diff --git a/drivers/net/pasemi_mac.c b/drivers/net/pasemi_mac.c
index 6ea822a..54904ad 100644
--- a/drivers/net/pasemi_mac.c
+++ b/drivers/net/pasemi_mac.c
@@ -59,11 +59,12 @@
/* Must be a power of two */
#define RX_RING_SIZE 2048
#define TX_RING_SIZE 4096
+#define CS_RING_SIZE (TX_RING_SIZE*2)
#define LRO_MAX_AGGR 64
#define PE_MIN_MTU 64
-#define PE_MAX_MTU 1500
+#define PE_MAX_MTU 9000
#define PE_DEF_MTU ETH_DATA_LEN
#define DEFAULT_MSG_ENABLE \
@@ -81,6 +82,7 @@
#define RX_DESC(rx, num) ((rx)->chan.ring_virt[(num) & (RX_RING_SIZE-1)])
#define RX_DESC_INFO(rx, num) ((rx)->ring_info[(num) & (RX_RING_SIZE-1)])
#define RX_BUFF(rx, num) ((rx)->buffers[(num) & (RX_RING_SIZE-1)])
+#define CS_DESC(cs, num) ((cs)->chan.ring_virt[(num) & (CS_RING_SIZE-1)])
#define RING_USED(ring) (((ring)->next_to_fill - (ring)->next_to_clean) \
& ((ring)->size - 1))
@@ -322,6 +324,103 @@
return (nfrags + 3) & ~1;
}
+static struct pasemi_mac_csring *pasemi_mac_setup_csring(struct pasemi_mac *mac)
+{
+ struct pasemi_mac_csring *ring;
+ u32 val;
+ unsigned int cfg;
+ int chno;
+
+ ring = pasemi_dma_alloc_chan(TXCHAN, sizeof(struct pasemi_mac_csring),
+ offsetof(struct pasemi_mac_csring, chan));
+
+ if (!ring) {
+ dev_err(&mac->pdev->dev, "Can't allocate checksum channel\n");
+ goto out_chan;
+ }
+
+ chno = ring->chan.chno;
+
+ ring->size = CS_RING_SIZE;
+ ring->next_to_fill = 0;
+
+ /* Allocate descriptors */
+ if (pasemi_dma_alloc_ring(&ring->chan, CS_RING_SIZE))
+ goto out_ring_desc;
+
+ write_dma_reg(PAS_DMA_TXCHAN_BASEL(chno),
+ PAS_DMA_TXCHAN_BASEL_BRBL(ring->chan.ring_dma));
+ val = PAS_DMA_TXCHAN_BASEU_BRBH(ring->chan.ring_dma >> 32);
+ val |= PAS_DMA_TXCHAN_BASEU_SIZ(CS_RING_SIZE >> 3);
+
+ write_dma_reg(PAS_DMA_TXCHAN_BASEU(chno), val);
+
+ ring->events[0] = pasemi_dma_alloc_flag();
+ ring->events[1] = pasemi_dma_alloc_flag();
+ if (ring->events[0] < 0 || ring->events[1] < 0)
+ goto out_flags;
+
+ pasemi_dma_clear_flag(ring->events[0]);
+ pasemi_dma_clear_flag(ring->events[1]);
+
+ ring->fun = pasemi_dma_alloc_fun();
+ if (ring->fun < 0)
+ goto out_fun;
+
+ cfg = PAS_DMA_TXCHAN_CFG_TY_FUNC | PAS_DMA_TXCHAN_CFG_UP |
+ PAS_DMA_TXCHAN_CFG_TATTR(ring->fun) |
+ PAS_DMA_TXCHAN_CFG_LPSQ | PAS_DMA_TXCHAN_CFG_LPDQ;
+
+ if (translation_enabled())
+ cfg |= PAS_DMA_TXCHAN_CFG_TRD | PAS_DMA_TXCHAN_CFG_TRR;
+
+ write_dma_reg(PAS_DMA_TXCHAN_CFG(chno), cfg);
+
+ /* enable channel */
+ pasemi_dma_start_chan(&ring->chan, PAS_DMA_TXCHAN_TCMDSTA_SZ |
+ PAS_DMA_TXCHAN_TCMDSTA_DB |
+ PAS_DMA_TXCHAN_TCMDSTA_DE |
+ PAS_DMA_TXCHAN_TCMDSTA_DA);
+
+ return ring;
+
+out_fun:
+out_flags:
+ if (ring->events[0] >= 0)
+ pasemi_dma_free_flag(ring->events[0]);
+ if (ring->events[1] >= 0)
+ pasemi_dma_free_flag(ring->events[1]);
+ pasemi_dma_free_ring(&ring->chan);
+out_ring_desc:
+ pasemi_dma_free_chan(&ring->chan);
+out_chan:
+
+ return NULL;
+}
+
+static void pasemi_mac_setup_csrings(struct pasemi_mac *mac)
+{
+ int i;
+ mac->cs[0] = pasemi_mac_setup_csring(mac);
+ if (mac->type == MAC_TYPE_XAUI)
+ mac->cs[1] = pasemi_mac_setup_csring(mac);
+ else
+ mac->cs[1] = 0;
+
+ for (i = 0; i < MAX_CS; i++)
+ if (mac->cs[i])
+ mac->num_cs++;
+}
+
+static void pasemi_mac_free_csring(struct pasemi_mac_csring *csring)
+{
+ pasemi_dma_stop_chan(&csring->chan);
+ pasemi_dma_free_flag(csring->events[0]);
+ pasemi_dma_free_flag(csring->events[1]);
+ pasemi_dma_free_ring(&csring->chan);
+ pasemi_dma_free_chan(&csring->chan);
+}
+
static int pasemi_mac_setup_rx_resources(const struct net_device *dev)
{
struct pasemi_mac_rxring *ring;
@@ -445,7 +544,7 @@
cfg = PAS_DMA_TXCHAN_CFG_TY_IFACE |
PAS_DMA_TXCHAN_CFG_TATTR(mac->dma_if) |
PAS_DMA_TXCHAN_CFG_UP |
- PAS_DMA_TXCHAN_CFG_WT(2);
+ PAS_DMA_TXCHAN_CFG_WT(4);
if (translation_enabled())
cfg |= PAS_DMA_TXCHAN_CFG_TRD | PAS_DMA_TXCHAN_CFG_TRR;
@@ -810,13 +909,21 @@
u64 mactx = TX_DESC(txring, i);
struct sk_buff *skb;
- skb = TX_DESC_INFO(txring, i+1).skb;
- nr_frags = TX_DESC_INFO(txring, i).dma;
-
if ((mactx & XCT_MACTX_E) ||
(*chan->status & PAS_STATUS_ERROR))
pasemi_mac_tx_error(mac, mactx);
+ /* Skip over control descriptors */
+ if (!(mactx & XCT_MACTX_LLEN_M)) {
+ TX_DESC(txring, i) = 0;
+ TX_DESC(txring, i+1) = 0;
+ buf_count = 2;
+ continue;
+ }
+
+ skb = TX_DESC_INFO(txring, i+1).skb;
+ nr_frags = TX_DESC_INFO(txring, i).dma;
+
if (unlikely(mactx & XCT_MACTX_O))
/* Not yet transmitted */
break;
@@ -1058,6 +1165,12 @@
if (!mac->tx)
goto out_tx_ring;
+ if (dev->mtu > 1500) {
+ pasemi_mac_setup_csrings(mac);
+ if (!mac->num_cs)
+ goto out_tx_ring;
+ }
+
/* 0x3ff with 33MHz clock is about 31us */
write_iob_reg(PAS_IOB_DMA_COM_TIMEOUTCFG,
PAS_IOB_DMA_COM_TIMEOUTCFG_TCNT(0x3ff));
@@ -1241,7 +1354,7 @@
{
struct pasemi_mac *mac = netdev_priv(dev);
unsigned int sta;
- int rxch, txch;
+ int rxch, txch, i;
rxch = rx_ring(mac)->chan.chno;
txch = tx_ring(mac)->chan.chno;
@@ -1286,6 +1399,9 @@
free_irq(mac->tx->chan.irq, mac->tx);
free_irq(mac->rx->chan.irq, mac->rx);
+ for (i = 0; i < mac->num_cs; i++)
+ pasemi_mac_free_csring(mac->cs[i]);
+
/* Free resources */
pasemi_mac_free_rx_resources(mac);
pasemi_mac_free_tx_resources(mac);
@@ -1293,36 +1409,114 @@
return 0;
}
+static void pasemi_mac_queue_csdesc(const struct sk_buff *skb,
+ const dma_addr_t *map,
+ const unsigned int *map_size,
+ struct pasemi_mac_txring *txring,
+ struct pasemi_mac_csring *csring)
+{
+ u64 fund;
+ dma_addr_t cs_dest;
+ const int nh_off = skb_network_offset(skb);
+ const int nh_len = skb_network_header_len(skb);
+ const int nfrags = skb_shinfo(skb)->nr_frags;
+ int cs_size, i, fill, hdr, cpyhdr, evt;
+ dma_addr_t csdma;
+
+ fund = XCT_FUN_ST | XCT_FUN_RR_8BRES |
+ XCT_FUN_O | XCT_FUN_FUN(csring->fun) |
+ XCT_FUN_CRM_SIG | XCT_FUN_LLEN(skb->len - nh_off) |
+ XCT_FUN_SHL(nh_len >> 2) | XCT_FUN_SE;
+
+ switch (ip_hdr(skb)->protocol) {
+ case IPPROTO_TCP:
+ fund |= XCT_FUN_SIG_TCP4;
+ /* TCP checksum is 16 bytes into the header */
+ cs_dest = map[0] + skb_transport_offset(skb) + 16;
+ break;
+ case IPPROTO_UDP:
+ fund |= XCT_FUN_SIG_UDP4;
+ /* UDP checksum is 6 bytes into the header */
+ cs_dest = map[0] + skb_transport_offset(skb) + 6;
+ break;
+ default:
+ BUG();
+ }
+
+ /* Do the checksum offloaded */
+ fill = csring->next_to_fill;
+ hdr = fill;
+
+ CS_DESC(csring, fill++) = fund;
+ /* Room for 8BRES. Checksum result is really 2 bytes into it */
+ csdma = csring->chan.ring_dma + (fill & (CS_RING_SIZE-1)) * 8 + 2;
+ CS_DESC(csring, fill++) = 0;
+
+ CS_DESC(csring, fill) = XCT_PTR_LEN(map_size[0]-nh_off) | XCT_PTR_ADDR(map[0]+nh_off);
+ for (i = 1; i <= nfrags; i++)
+ CS_DESC(csring, fill+i) = XCT_PTR_LEN(map_size[i]) | XCT_PTR_ADDR(map[i]);
+
+ fill += i;
+ if (fill & 1)
+ fill++;
+
+ /* Copy the result into the TCP packet */
+ cpyhdr = fill;
+ CS_DESC(csring, fill++) = XCT_FUN_O | XCT_FUN_FUN(csring->fun) |
+ XCT_FUN_LLEN(2) | XCT_FUN_SE;
+ CS_DESC(csring, fill++) = XCT_PTR_LEN(2) | XCT_PTR_ADDR(cs_dest) | XCT_PTR_T;
+ CS_DESC(csring, fill++) = XCT_PTR_LEN(2) | XCT_PTR_ADDR(csdma);
+ fill++;
+
+ evt = !csring->last_event;
+ csring->last_event = evt;
+
+ /* Event handshaking with MAC TX */
+ CS_DESC(csring, fill++) = CTRL_CMD_T | CTRL_CMD_META_EVT | CTRL_CMD_O |
+ CTRL_CMD_ETYPE_SET | CTRL_CMD_REG(csring->events[evt]);
+ CS_DESC(csring, fill++) = 0;
+ CS_DESC(csring, fill++) = CTRL_CMD_T | CTRL_CMD_META_EVT | CTRL_CMD_O |
+ CTRL_CMD_ETYPE_WCLR | CTRL_CMD_REG(csring->events[!evt]);
+ CS_DESC(csring, fill++) = 0;
+ csring->next_to_fill = fill & (CS_RING_SIZE-1);
+
+ cs_size = fill - hdr;
+ write_dma_reg(PAS_DMA_TXCHAN_INCR(csring->chan.chno), (cs_size) >> 1);
+
+ /* TX-side event handshaking */
+ fill = txring->next_to_fill;
+ TX_DESC(txring, fill++) = CTRL_CMD_T | CTRL_CMD_META_EVT | CTRL_CMD_O |
+ CTRL_CMD_ETYPE_WSET | CTRL_CMD_REG(csring->events[evt]);
+ TX_DESC(txring, fill++) = 0;
+ TX_DESC(txring, fill++) = CTRL_CMD_T | CTRL_CMD_META_EVT | CTRL_CMD_O |
+ CTRL_CMD_ETYPE_CLR | CTRL_CMD_REG(csring->events[!evt]);
+ TX_DESC(txring, fill++) = 0;
+ txring->next_to_fill = fill;
+
+ write_dma_reg(PAS_DMA_TXCHAN_INCR(txring->chan.chno), 2);
+
+ return;
+}
+
static int pasemi_mac_start_tx(struct sk_buff *skb, struct net_device *dev)
{
- struct pasemi_mac *mac = netdev_priv(dev);
- struct pasemi_mac_txring *txring;
- u64 dflags, mactx;
+ struct pasemi_mac * const mac = netdev_priv(dev);
+ struct pasemi_mac_txring * const txring = tx_ring(mac);
+ struct pasemi_mac_csring *csring;
+ u64 dflags = 0;
+ u64 mactx;
dma_addr_t map[MAX_SKB_FRAGS+1];
unsigned int map_size[MAX_SKB_FRAGS+1];
unsigned long flags;
int i, nfrags;
int fill;
+ const int nh_off = skb_network_offset(skb);
+ const int nh_len = skb_network_header_len(skb);
+
+ prefetch(&txring->ring_info);
dflags = XCT_MACTX_O | XCT_MACTX_ST | XCT_MACTX_CRC_PAD;
- if (skb->ip_summed == CHECKSUM_PARTIAL) {
- const unsigned char *nh = skb_network_header(skb);
-
- switch (ip_hdr(skb)->protocol) {
- case IPPROTO_TCP:
- dflags |= XCT_MACTX_CSUM_TCP;
- dflags |= XCT_MACTX_IPH(skb_network_header_len(skb) >> 2);
- dflags |= XCT_MACTX_IPO(nh - skb->data);
- break;
- case IPPROTO_UDP:
- dflags |= XCT_MACTX_CSUM_UDP;
- dflags |= XCT_MACTX_IPH(skb_network_header_len(skb) >> 2);
- dflags |= XCT_MACTX_IPO(nh - skb->data);
- break;
- }
- }
-
nfrags = skb_shinfo(skb)->nr_frags;
map[0] = pci_map_single(mac->dma_pdev, skb->data, skb_headlen(skb),
@@ -1344,24 +1538,46 @@
}
}
+ if (skb->ip_summed == CHECKSUM_PARTIAL && skb->len <= 1540) {
+ switch (ip_hdr(skb)->protocol) {
+ case IPPROTO_TCP:
+ dflags |= XCT_MACTX_CSUM_TCP;
+ dflags |= XCT_MACTX_IPH(nh_len >> 2);
+ dflags |= XCT_MACTX_IPO(nh_off);
+ break;
+ case IPPROTO_UDP:
+ dflags |= XCT_MACTX_CSUM_UDP;
+ dflags |= XCT_MACTX_IPH(nh_len >> 2);
+ dflags |= XCT_MACTX_IPO(nh_off);
+ break;
+ default:
+ WARN_ON(1);
+ }
+ }
+
mactx = dflags | XCT_MACTX_LLEN(skb->len);
- txring = tx_ring(mac);
-
spin_lock_irqsave(&txring->lock, flags);
- fill = txring->next_to_fill;
-
/* Avoid stepping on the same cache line that the DMA controller
* is currently about to send, so leave at least 8 words available.
* Total free space needed is mactx + fragments + 8
*/
- if (RING_AVAIL(txring) < nfrags + 10) {
+ if (RING_AVAIL(txring) < nfrags + 14) {
/* no room -- stop the queue and wait for tx intr */
netif_stop_queue(dev);
goto out_err;
}
+ /* Queue up checksum + event descriptors, if needed */
+ if (mac->num_cs && skb->ip_summed == CHECKSUM_PARTIAL && skb->len > 1540) {
+ csring = mac->cs[mac->last_cs];
+ mac->last_cs = (mac->last_cs + 1) % mac->num_cs;
+
+ pasemi_mac_queue_csdesc(skb, map, map_size, txring, csring);
+ }
+
+ fill = txring->next_to_fill;
TX_DESC(txring, fill) = mactx;
TX_DESC_INFO(txring, fill).dma = nfrags;
fill++;
@@ -1439,8 +1655,9 @@
{
struct pasemi_mac *mac = netdev_priv(dev);
unsigned int reg;
- unsigned int rcmdsta;
+ unsigned int rcmdsta = 0;
int running;
+ int ret = 0;
if (new_mtu < PE_MIN_MTU || new_mtu > PE_MAX_MTU)
return -EINVAL;
@@ -1462,6 +1679,16 @@
pasemi_mac_pause_rxint(mac);
pasemi_mac_clean_rx(rx_ring(mac), RX_RING_SIZE);
pasemi_mac_free_rx_buffers(mac);
+
+ }
+
+ /* Setup checksum channels if large MTU and none already allocated */
+ if (new_mtu > 1500 && !mac->num_cs) {
+ pasemi_mac_setup_csrings(mac);
+ if (!mac->num_cs) {
+ ret = -ENOMEM;
+ goto out;
+ }
}
/* Change maxf, i.e. what size frames are accepted.
@@ -1476,6 +1703,7 @@
/* MTU + ETH_HLEN + VLAN_HLEN + 2 64B cachelines */
mac->bufsz = new_mtu + ETH_HLEN + ETH_FCS_LEN + LOCAL_SKB_ALIGN + 128;
+out:
if (running) {
write_dma_reg(PAS_DMA_RXINT_RCMDSTA(mac->dma_if),
rcmdsta | PAS_DMA_RXINT_RCMDSTA_EN);
@@ -1488,7 +1716,7 @@
pasemi_mac_intf_enable(mac);
}
- return 0;
+ return ret;
}
static int __devinit