aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/net/cxgb3/sge.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/net/cxgb3/sge.c')
-rw-r--r--drivers/net/cxgb3/sge.c170
1 files changed, 68 insertions, 102 deletions
diff --git a/drivers/net/cxgb3/sge.c b/drivers/net/cxgb3/sge.c
index 58a5f60521ed..069c1aca8a6b 100644
--- a/drivers/net/cxgb3/sge.c
+++ b/drivers/net/cxgb3/sge.c
@@ -591,9 +591,6 @@ void t3_free_qset(struct adapter *adapter, struct sge_qset *q)
591 q->rspq.desc, q->rspq.phys_addr); 591 q->rspq.desc, q->rspq.phys_addr);
592 } 592 }
593 593
594 if (q->netdev)
595 q->netdev->atalk_ptr = NULL;
596
597 memset(q, 0, sizeof(*q)); 594 memset(q, 0, sizeof(*q));
598} 595}
599 596
@@ -1074,7 +1071,7 @@ int t3_eth_xmit(struct sk_buff *skb, struct net_device *dev)
1074 unsigned int ndesc, pidx, credits, gen, compl; 1071 unsigned int ndesc, pidx, credits, gen, compl;
1075 const struct port_info *pi = netdev_priv(dev); 1072 const struct port_info *pi = netdev_priv(dev);
1076 struct adapter *adap = pi->adapter; 1073 struct adapter *adap = pi->adapter;
1077 struct sge_qset *qs = dev2qset(dev); 1074 struct sge_qset *qs = pi->qs;
1078 struct sge_txq *q = &qs->txq[TXQ_ETH]; 1075 struct sge_txq *q = &qs->txq[TXQ_ETH];
1079 1076
1080 /* 1077 /*
@@ -1326,13 +1323,12 @@ static void restart_ctrlq(unsigned long data)
1326 struct sk_buff *skb; 1323 struct sk_buff *skb;
1327 struct sge_qset *qs = (struct sge_qset *)data; 1324 struct sge_qset *qs = (struct sge_qset *)data;
1328 struct sge_txq *q = &qs->txq[TXQ_CTRL]; 1325 struct sge_txq *q = &qs->txq[TXQ_CTRL];
1329 const struct port_info *pi = netdev_priv(qs->netdev);
1330 struct adapter *adap = pi->adapter;
1331 1326
1332 spin_lock(&q->lock); 1327 spin_lock(&q->lock);
1333 again:reclaim_completed_tx_imm(q); 1328 again:reclaim_completed_tx_imm(q);
1334 1329
1335 while (q->in_use < q->size && (skb = __skb_dequeue(&q->sendq)) != NULL) { 1330 while (q->in_use < q->size &&
1331 (skb = __skb_dequeue(&q->sendq)) != NULL) {
1336 1332
1337 write_imm(&q->desc[q->pidx], skb, skb->len, q->gen); 1333 write_imm(&q->desc[q->pidx], skb, skb->len, q->gen);
1338 1334
@@ -1354,7 +1350,7 @@ static void restart_ctrlq(unsigned long data)
1354 } 1350 }
1355 1351
1356 spin_unlock(&q->lock); 1352 spin_unlock(&q->lock);
1357 t3_write_reg(adap, A_SG_KDOORBELL, 1353 t3_write_reg(qs->adap, A_SG_KDOORBELL,
1358 F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id)); 1354 F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1359} 1355}
1360 1356
@@ -1638,8 +1634,7 @@ static inline void offload_enqueue(struct sge_rspq *q, struct sk_buff *skb)
1638 else { 1634 else {
1639 struct sge_qset *qs = rspq_to_qset(q); 1635 struct sge_qset *qs = rspq_to_qset(q);
1640 1636
1641 if (__netif_rx_schedule_prep(qs->netdev)) 1637 napi_schedule(&qs->napi);
1642 __netif_rx_schedule(qs->netdev);
1643 q->rx_head = skb; 1638 q->rx_head = skb;
1644 } 1639 }
1645 q->rx_tail = skb; 1640 q->rx_tail = skb;
@@ -1675,34 +1670,30 @@ static inline void deliver_partial_bundle(struct t3cdev *tdev,
1675 * receive handler. Batches need to be of modest size as we do prefetches 1670 * receive handler. Batches need to be of modest size as we do prefetches
1676 * on the packets in each. 1671 * on the packets in each.
1677 */ 1672 */
1678static int ofld_poll(struct net_device *dev, int *budget) 1673static int ofld_poll(struct napi_struct *napi, int budget)
1679{ 1674{
1680 const struct port_info *pi = netdev_priv(dev); 1675 struct sge_qset *qs = container_of(napi, struct sge_qset, napi);
1681 struct adapter *adapter = pi->adapter;
1682 struct sge_qset *qs = dev2qset(dev);
1683 struct sge_rspq *q = &qs->rspq; 1676 struct sge_rspq *q = &qs->rspq;
1684 int work_done, limit = min(*budget, dev->quota), avail = limit; 1677 struct adapter *adapter = qs->adap;
1678 int work_done = 0;
1685 1679
1686 while (avail) { 1680 while (work_done < budget) {
1687 struct sk_buff *head, *tail, *skbs[RX_BUNDLE_SIZE]; 1681 struct sk_buff *head, *tail, *skbs[RX_BUNDLE_SIZE];
1688 int ngathered; 1682 int ngathered;
1689 1683
1690 spin_lock_irq(&q->lock); 1684 spin_lock_irq(&q->lock);
1691 head = q->rx_head; 1685 head = q->rx_head;
1692 if (!head) { 1686 if (!head) {
1693 work_done = limit - avail; 1687 napi_complete(napi);
1694 *budget -= work_done;
1695 dev->quota -= work_done;
1696 __netif_rx_complete(dev);
1697 spin_unlock_irq(&q->lock); 1688 spin_unlock_irq(&q->lock);
1698 return 0; 1689 return work_done;
1699 } 1690 }
1700 1691
1701 tail = q->rx_tail; 1692 tail = q->rx_tail;
1702 q->rx_head = q->rx_tail = NULL; 1693 q->rx_head = q->rx_tail = NULL;
1703 spin_unlock_irq(&q->lock); 1694 spin_unlock_irq(&q->lock);
1704 1695
1705 for (ngathered = 0; avail && head; avail--) { 1696 for (ngathered = 0; work_done < budget && head; work_done++) {
1706 prefetch(head->data); 1697 prefetch(head->data);
1707 skbs[ngathered] = head; 1698 skbs[ngathered] = head;
1708 head = head->next; 1699 head = head->next;
@@ -1724,10 +1715,8 @@ static int ofld_poll(struct net_device *dev, int *budget)
1724 } 1715 }
1725 deliver_partial_bundle(&adapter->tdev, q, skbs, ngathered); 1716 deliver_partial_bundle(&adapter->tdev, q, skbs, ngathered);
1726 } 1717 }
1727 work_done = limit - avail; 1718
1728 *budget -= work_done; 1719 return work_done;
1729 dev->quota -= work_done;
1730 return 1;
1731} 1720}
1732 1721
1733/** 1722/**
@@ -2071,50 +2060,47 @@ static inline int is_pure_response(const struct rsp_desc *r)
2071 2060
2072/** 2061/**
2073 * napi_rx_handler - the NAPI handler for Rx processing 2062 * napi_rx_handler - the NAPI handler for Rx processing
2074 * @dev: the net device 2063 * @napi: the napi instance
2075 * @budget: how many packets we can process in this round 2064 * @budget: how many packets we can process in this round
2076 * 2065 *
2077 * Handler for new data events when using NAPI. 2066 * Handler for new data events when using NAPI.
2078 */ 2067 */
2079static int napi_rx_handler(struct net_device *dev, int *budget) 2068static int napi_rx_handler(struct napi_struct *napi, int budget)
2080{ 2069{
2081 const struct port_info *pi = netdev_priv(dev); 2070 struct sge_qset *qs = container_of(napi, struct sge_qset, napi);
2082 struct adapter *adap = pi->adapter; 2071 struct adapter *adap = qs->adap;
2083 struct sge_qset *qs = dev2qset(dev); 2072 int work_done = process_responses(adap, qs, budget);
2084 int effective_budget = min(*budget, dev->quota);
2085
2086 int work_done = process_responses(adap, qs, effective_budget);
2087 *budget -= work_done;
2088 dev->quota -= work_done;
2089 2073
2090 if (work_done >= effective_budget) 2074 if (likely(work_done < budget)) {
2091 return 1; 2075 napi_complete(napi);
2092
2093 netif_rx_complete(dev);
2094 2076
2095 /* 2077 /*
2096 * Because we don't atomically flush the following write it is 2078 * Because we don't atomically flush the following
2097 * possible that in very rare cases it can reach the device in a way 2079 * write it is possible that in very rare cases it can
2098 * that races with a new response being written plus an error interrupt 2080 * reach the device in a way that races with a new
2099 * causing the NAPI interrupt handler below to return unhandled status 2081 * response being written plus an error interrupt
2100 * to the OS. To protect against this would require flushing the write 2082 * causing the NAPI interrupt handler below to return
2101 * and doing both the write and the flush with interrupts off. Way too 2083 * unhandled status to the OS. To protect against
2102 * expensive and unjustifiable given the rarity of the race. 2084 * this would require flushing the write and doing
2103 * 2085 * both the write and the flush with interrupts off.
2104 * The race cannot happen at all with MSI-X. 2086 * Way too expensive and unjustifiable given the
2105 */ 2087 * rarity of the race.
2106 t3_write_reg(adap, A_SG_GTS, V_RSPQ(qs->rspq.cntxt_id) | 2088 *
2107 V_NEWTIMER(qs->rspq.next_holdoff) | 2089 * The race cannot happen at all with MSI-X.
2108 V_NEWINDEX(qs->rspq.cidx)); 2090 */
2109 return 0; 2091 t3_write_reg(adap, A_SG_GTS, V_RSPQ(qs->rspq.cntxt_id) |
2092 V_NEWTIMER(qs->rspq.next_holdoff) |
2093 V_NEWINDEX(qs->rspq.cidx));
2094 }
2095 return work_done;
2110} 2096}
2111 2097
2112/* 2098/*
2113 * Returns true if the device is already scheduled for polling. 2099 * Returns true if the device is already scheduled for polling.
2114 */ 2100 */
2115static inline int napi_is_scheduled(struct net_device *dev) 2101static inline int napi_is_scheduled(struct napi_struct *napi)
2116{ 2102{
2117 return test_bit(__LINK_STATE_RX_SCHED, &dev->state); 2103 return test_bit(NAPI_STATE_SCHED, &napi->state);
2118} 2104}
2119 2105
2120/** 2106/**
@@ -2197,8 +2183,7 @@ static inline int handle_responses(struct adapter *adap, struct sge_rspq *q)
2197 V_NEWTIMER(q->holdoff_tmr) | V_NEWINDEX(q->cidx)); 2183 V_NEWTIMER(q->holdoff_tmr) | V_NEWINDEX(q->cidx));
2198 return 0; 2184 return 0;
2199 } 2185 }
2200 if (likely(__netif_rx_schedule_prep(qs->netdev))) 2186 napi_schedule(&qs->napi);
2201 __netif_rx_schedule(qs->netdev);
2202 return 1; 2187 return 1;
2203} 2188}
2204 2189
@@ -2209,8 +2194,7 @@ static inline int handle_responses(struct adapter *adap, struct sge_rspq *q)
2209irqreturn_t t3_sge_intr_msix(int irq, void *cookie) 2194irqreturn_t t3_sge_intr_msix(int irq, void *cookie)
2210{ 2195{
2211 struct sge_qset *qs = cookie; 2196 struct sge_qset *qs = cookie;
2212 const struct port_info *pi = netdev_priv(qs->netdev); 2197 struct adapter *adap = qs->adap;
2213 struct adapter *adap = pi->adapter;
2214 struct sge_rspq *q = &qs->rspq; 2198 struct sge_rspq *q = &qs->rspq;
2215 2199
2216 spin_lock(&q->lock); 2200 spin_lock(&q->lock);
@@ -2229,13 +2213,11 @@ irqreturn_t t3_sge_intr_msix(int irq, void *cookie)
2229irqreturn_t t3_sge_intr_msix_napi(int irq, void *cookie) 2213irqreturn_t t3_sge_intr_msix_napi(int irq, void *cookie)
2230{ 2214{
2231 struct sge_qset *qs = cookie; 2215 struct sge_qset *qs = cookie;
2232 const struct port_info *pi = netdev_priv(qs->netdev);
2233 struct adapter *adap = pi->adapter;
2234 struct sge_rspq *q = &qs->rspq; 2216 struct sge_rspq *q = &qs->rspq;
2235 2217
2236 spin_lock(&q->lock); 2218 spin_lock(&q->lock);
2237 2219
2238 if (handle_responses(adap, q) < 0) 2220 if (handle_responses(qs->adap, q) < 0)
2239 q->unhandled_irqs++; 2221 q->unhandled_irqs++;
2240 spin_unlock(&q->lock); 2222 spin_unlock(&q->lock);
2241 return IRQ_HANDLED; 2223 return IRQ_HANDLED;
@@ -2278,11 +2260,13 @@ static irqreturn_t t3_intr_msi(int irq, void *cookie)
2278 return IRQ_HANDLED; 2260 return IRQ_HANDLED;
2279} 2261}
2280 2262
2281static int rspq_check_napi(struct net_device *dev, struct sge_rspq *q) 2263static int rspq_check_napi(struct sge_qset *qs)
2282{ 2264{
2283 if (!napi_is_scheduled(dev) && is_new_response(&q->desc[q->cidx], q)) { 2265 struct sge_rspq *q = &qs->rspq;
2284 if (likely(__netif_rx_schedule_prep(dev))) 2266
2285 __netif_rx_schedule(dev); 2267 if (!napi_is_scheduled(&qs->napi) &&
2268 is_new_response(&q->desc[q->cidx], q)) {
2269 napi_schedule(&qs->napi);
2286 return 1; 2270 return 1;
2287 } 2271 }
2288 return 0; 2272 return 0;
@@ -2303,10 +2287,9 @@ irqreturn_t t3_intr_msi_napi(int irq, void *cookie)
2303 2287
2304 spin_lock(&q->lock); 2288 spin_lock(&q->lock);
2305 2289
2306 new_packets = rspq_check_napi(adap->sge.qs[0].netdev, q); 2290 new_packets = rspq_check_napi(&adap->sge.qs[0]);
2307 if (adap->params.nports == 2) 2291 if (adap->params.nports == 2)
2308 new_packets += rspq_check_napi(adap->sge.qs[1].netdev, 2292 new_packets += rspq_check_napi(&adap->sge.qs[1]);
2309 &adap->sge.qs[1].rspq);
2310 if (!new_packets && t3_slow_intr_handler(adap) == 0) 2293 if (!new_packets && t3_slow_intr_handler(adap) == 0)
2311 q->unhandled_irqs++; 2294 q->unhandled_irqs++;
2312 2295
@@ -2409,9 +2392,9 @@ static irqreturn_t t3b_intr(int irq, void *cookie)
2409static irqreturn_t t3b_intr_napi(int irq, void *cookie) 2392static irqreturn_t t3b_intr_napi(int irq, void *cookie)
2410{ 2393{
2411 u32 map; 2394 u32 map;
2412 struct net_device *dev;
2413 struct adapter *adap = cookie; 2395 struct adapter *adap = cookie;
2414 struct sge_rspq *q0 = &adap->sge.qs[0].rspq; 2396 struct sge_qset *qs0 = &adap->sge.qs[0];
2397 struct sge_rspq *q0 = &qs0->rspq;
2415 2398
2416 t3_write_reg(adap, A_PL_CLI, 0); 2399 t3_write_reg(adap, A_PL_CLI, 0);
2417 map = t3_read_reg(adap, A_SG_DATA_INTR); 2400 map = t3_read_reg(adap, A_SG_DATA_INTR);
@@ -2424,18 +2407,11 @@ static irqreturn_t t3b_intr_napi(int irq, void *cookie)
2424 if (unlikely(map & F_ERRINTR)) 2407 if (unlikely(map & F_ERRINTR))
2425 t3_slow_intr_handler(adap); 2408 t3_slow_intr_handler(adap);
2426 2409
2427 if (likely(map & 1)) { 2410 if (likely(map & 1))
2428 dev = adap->sge.qs[0].netdev; 2411 napi_schedule(&qs0->napi);
2429
2430 if (likely(__netif_rx_schedule_prep(dev)))
2431 __netif_rx_schedule(dev);
2432 }
2433 if (map & 2) {
2434 dev = adap->sge.qs[1].netdev;
2435 2412
2436 if (likely(__netif_rx_schedule_prep(dev))) 2413 if (map & 2)
2437 __netif_rx_schedule(dev); 2414 napi_schedule(&adap->sge.qs[1].napi);
2438 }
2439 2415
2440 spin_unlock(&q0->lock); 2416 spin_unlock(&q0->lock);
2441 return IRQ_HANDLED; 2417 return IRQ_HANDLED;
@@ -2514,8 +2490,7 @@ static void sge_timer_cb(unsigned long data)
2514{ 2490{
2515 spinlock_t *lock; 2491 spinlock_t *lock;
2516 struct sge_qset *qs = (struct sge_qset *)data; 2492 struct sge_qset *qs = (struct sge_qset *)data;
2517 const struct port_info *pi = netdev_priv(qs->netdev); 2493 struct adapter *adap = qs->adap;
2518 struct adapter *adap = pi->adapter;
2519 2494
2520 if (spin_trylock(&qs->txq[TXQ_ETH].lock)) { 2495 if (spin_trylock(&qs->txq[TXQ_ETH].lock)) {
2521 reclaim_completed_tx(adap, &qs->txq[TXQ_ETH]); 2496 reclaim_completed_tx(adap, &qs->txq[TXQ_ETH]);
@@ -2526,9 +2501,9 @@ static void sge_timer_cb(unsigned long data)
2526 spin_unlock(&qs->txq[TXQ_OFLD].lock); 2501 spin_unlock(&qs->txq[TXQ_OFLD].lock);
2527 } 2502 }
2528 lock = (adap->flags & USING_MSIX) ? &qs->rspq.lock : 2503 lock = (adap->flags & USING_MSIX) ? &qs->rspq.lock :
2529 &adap->sge.qs[0].rspq.lock; 2504 &adap->sge.qs[0].rspq.lock;
2530 if (spin_trylock_irq(lock)) { 2505 if (spin_trylock_irq(lock)) {
2531 if (!napi_is_scheduled(qs->netdev)) { 2506 if (!napi_is_scheduled(&qs->napi)) {
2532 u32 status = t3_read_reg(adap, A_SG_RSPQ_FL_STATUS); 2507 u32 status = t3_read_reg(adap, A_SG_RSPQ_FL_STATUS);
2533 2508
2534 if (qs->fl[0].credits < qs->fl[0].size) 2509 if (qs->fl[0].credits < qs->fl[0].size)
@@ -2562,12 +2537,9 @@ static void sge_timer_cb(unsigned long data)
2562 */ 2537 */
2563void t3_update_qset_coalesce(struct sge_qset *qs, const struct qset_params *p) 2538void t3_update_qset_coalesce(struct sge_qset *qs, const struct qset_params *p)
2564{ 2539{
2565 if (!qs->netdev)
2566 return;
2567
2568 qs->rspq.holdoff_tmr = max(p->coalesce_usecs * 10, 1U);/* can't be 0 */ 2540 qs->rspq.holdoff_tmr = max(p->coalesce_usecs * 10, 1U);/* can't be 0 */
2569 qs->rspq.polling = p->polling; 2541 qs->rspq.polling = p->polling;
2570 qs->netdev->poll = p->polling ? napi_rx_handler : ofld_poll; 2542 qs->napi.poll = p->polling ? napi_rx_handler : ofld_poll;
2571} 2543}
2572 2544
2573/** 2545/**
@@ -2587,7 +2559,7 @@ void t3_update_qset_coalesce(struct sge_qset *qs, const struct qset_params *p)
2587 */ 2559 */
2588int t3_sge_alloc_qset(struct adapter *adapter, unsigned int id, int nports, 2560int t3_sge_alloc_qset(struct adapter *adapter, unsigned int id, int nports,
2589 int irq_vec_idx, const struct qset_params *p, 2561 int irq_vec_idx, const struct qset_params *p,
2590 int ntxq, struct net_device *netdev) 2562 int ntxq, struct net_device *dev)
2591{ 2563{
2592 int i, ret = -ENOMEM; 2564 int i, ret = -ENOMEM;
2593 struct sge_qset *q = &adapter->sge.qs[id]; 2565 struct sge_qset *q = &adapter->sge.qs[id];
@@ -2708,16 +2680,10 @@ int t3_sge_alloc_qset(struct adapter *adapter, unsigned int id, int nports,
2708 } 2680 }
2709 2681
2710 spin_unlock(&adapter->sge.reg_lock); 2682 spin_unlock(&adapter->sge.reg_lock);
2711 q->netdev = netdev;
2712 t3_update_qset_coalesce(q, p);
2713 2683
2714 /* 2684 q->adap = adapter;
2715 * We use atalk_ptr as a backpointer to a qset. In case a device is 2685 q->netdev = dev;
2716 * associated with multiple queue sets only the first one sets 2686 t3_update_qset_coalesce(q, p);
2717 * atalk_ptr.
2718 */
2719 if (netdev->atalk_ptr == NULL)
2720 netdev->atalk_ptr = q;
2721 2687
2722 refill_fl(adapter, &q->fl[0], q->fl[0].size, GFP_KERNEL); 2688 refill_fl(adapter, &q->fl[0], q->fl[0].size, GFP_KERNEL);
2723 refill_fl(adapter, &q->fl[1], q->fl[1].size, GFP_KERNEL); 2689 refill_fl(adapter, &q->fl[1], q->fl[1].size, GFP_KERNEL);