diff options
Diffstat (limited to 'drivers/net/ethernet/chelsio/cxgb4/sge.c')
-rw-r--r-- | drivers/net/ethernet/chelsio/cxgb4/sge.c | 218 |
1 files changed, 170 insertions, 48 deletions
diff --git a/drivers/net/ethernet/chelsio/cxgb4/sge.c b/drivers/net/ethernet/chelsio/cxgb4/sge.c index d22d728d4e5c..fab4c84a1da4 100644 --- a/drivers/net/ethernet/chelsio/cxgb4/sge.c +++ b/drivers/net/ethernet/chelsio/cxgb4/sge.c @@ -203,6 +203,9 @@ enum { RX_LARGE_MTU_BUF = 0x3, /* large MTU buffer */ }; +static int timer_pkt_quota[] = {1, 1, 2, 3, 4, 5}; +#define MIN_NAPI_WORK 1 + static inline dma_addr_t get_buf_addr(const struct rx_sw_desc *d) { return d->dma_addr & ~(dma_addr_t)RX_BUF_FLAGS; @@ -521,9 +524,23 @@ static inline void ring_fl_db(struct adapter *adap, struct sge_fl *q) val = PIDX(q->pend_cred / 8); if (!is_t4(adap->params.chip)) val |= DBTYPE(1); + val |= DBPRIO(1); wmb(); - t4_write_reg(adap, MYPF_REG(SGE_PF_KDOORBELL), DBPRIO(1) | - QID(q->cntxt_id) | val); + + /* If we're on T4, use the old doorbell mechanism; otherwise + * use the new BAR2 mechanism. + */ + if (is_t4(adap->params.chip)) { + t4_write_reg(adap, MYPF_REG(SGE_PF_KDOORBELL), + val | QID(q->cntxt_id)); + } else { + writel(val, adap->bar2 + q->udb + SGE_UDB_KDOORBELL); + + /* This Write memory Barrier will force the write to + * the User Doorbell area to be flushed. + */ + wmb(); + } q->pend_cred &= 7; } } @@ -833,13 +850,14 @@ static void write_sgl(const struct sk_buff *skb, struct sge_txq *q, *end = 0; } -/* This function copies 64 byte coalesced work request to - * memory mapped BAR2 space(user space writes). - * For coalesced WR SGE, fetches data from the FIFO instead of from Host. +/* This function copies a tx_desc struct to memory mapped BAR2 space(user space + * writes). For coalesced WR SGE, fetches data from the FIFO instead of from + * Host. */ -static void cxgb_pio_copy(u64 __iomem *dst, u64 *src) +static void cxgb_pio_copy(u64 __iomem *dst, struct tx_desc *desc) { - int count = 8; + int count = sizeof(*desc) / sizeof(u64); + u64 *src = (u64 *)desc; while (count) { writeq(*src, dst); @@ -859,30 +877,63 @@ static void cxgb_pio_copy(u64 __iomem *dst, u64 *src) */ static inline void ring_tx_db(struct adapter *adap, struct sge_txq *q, int n) { - unsigned int *wr, index; - unsigned long flags; - wmb(); /* write descriptors before telling HW */ - spin_lock_irqsave(&q->db_lock, flags); - if (!q->db_disabled) { - if (is_t4(adap->params.chip)) { + + if (is_t4(adap->params.chip)) { + u32 val = PIDX(n); + unsigned long flags; + + /* For T4 we need to participate in the Doorbell Recovery + * mechanism. + */ + spin_lock_irqsave(&q->db_lock, flags); + if (!q->db_disabled) t4_write_reg(adap, MYPF_REG(SGE_PF_KDOORBELL), - QID(q->cntxt_id) | PIDX(n)); + QID(q->cntxt_id) | val); + else + q->db_pidx_inc += n; + q->db_pidx = q->pidx; + spin_unlock_irqrestore(&q->db_lock, flags); + } else { + u32 val = PIDX_T5(n); + + /* T4 and later chips share the same PIDX field offset within + * the doorbell, but T5 and later shrank the field in order to + * gain a bit for Doorbell Priority. The field was absurdly + * large in the first place (14 bits) so we just use the T5 + * and later limits and warn if a Queue ID is too large. + */ + WARN_ON(val & DBPRIO(1)); + + /* For T5 and later we use the Write-Combine mapped BAR2 User + * Doorbell mechanism. If we're only writing a single TX + * Descriptor and TX Write Combining hasn't been disabled, we + * can use the Write Combining Gather Buffer; otherwise we use + * the simple doorbell. + */ + if (n == 1) { + int index = (q->pidx + ? (q->pidx - 1) + : (q->size - 1)); + + cxgb_pio_copy(adap->bar2 + q->udb + SGE_UDB_WCDOORBELL, + q->desc + index); } else { - if (n == 1) { - index = q->pidx ? (q->pidx - 1) : (q->size - 1); - wr = (unsigned int *)&q->desc[index]; - cxgb_pio_copy((u64 __iomem *) - (adap->bar2 + q->udb + 64), - (u64 *)wr); - } else - writel(n, adap->bar2 + q->udb + 8); - wmb(); + writel(val, adap->bar2 + q->udb + SGE_UDB_KDOORBELL); } - } else - q->db_pidx_inc += n; - q->db_pidx = q->pidx; - spin_unlock_irqrestore(&q->db_lock, flags); + + /* This Write Memory Barrier will force the write to the User + * Doorbell area to be flushed. This is needed to prevent + * writes on different CPUs for the same queue from hitting + * the adapter out of order. This is required when some Work + * Requests take the Write Combine Gather Buffer path (user + * doorbell area offset [SGE_UDB_WCDOORBELL..+63]) and some + * take the traditional path where we simply increment the + * PIDX (User Doorbell area SGE_UDB_KDOORBELL) and have the + * hardware DMA read the actual Work Request. + */ + wmb(); + } } /** @@ -1916,16 +1967,40 @@ static int napi_rx_handler(struct napi_struct *napi, int budget) unsigned int params; struct sge_rspq *q = container_of(napi, struct sge_rspq, napi); int work_done = process_responses(q, budget); + u32 val; if (likely(work_done < budget)) { + int timer_index; + napi_complete(napi); - params = q->next_intr_params; - q->next_intr_params = q->intr_params; + timer_index = QINTR_TIMER_IDX_GET(q->next_intr_params); + + if (q->adaptive_rx) { + if (work_done > max(timer_pkt_quota[timer_index], + MIN_NAPI_WORK)) + timer_index = (timer_index + 1); + else + timer_index = timer_index - 1; + + timer_index = clamp(timer_index, 0, SGE_TIMERREGS - 1); + q->next_intr_params = QINTR_TIMER_IDX(timer_index) | + V_QINTR_CNT_EN; + params = q->next_intr_params; + } else { + params = q->next_intr_params; + q->next_intr_params = q->intr_params; + } } else params = QINTR_TIMER_IDX(7); - t4_write_reg(q->adap, MYPF_REG(SGE_PF_GTS), CIDXINC(work_done) | - INGRESSQID((u32)q->cntxt_id) | SEINTARM(params)); + val = CIDXINC(work_done) | SEINTARM(params); + if (is_t4(q->adap->params.chip)) { + t4_write_reg(q->adap, MYPF_REG(SGE_PF_GTS), + val | INGRESSQID((u32)q->cntxt_id)); + } else { + writel(val, q->adap->bar2 + q->udb + SGE_UDB_GTS); + wmb(); + } return work_done; } @@ -1949,6 +2024,7 @@ static unsigned int process_intrq(struct adapter *adap) unsigned int credits; const struct rsp_ctrl *rc; struct sge_rspq *q = &adap->sge.intrq; + u32 val; spin_lock(&adap->sge.intrq_lock); for (credits = 0; ; credits++) { @@ -1967,8 +2043,14 @@ static unsigned int process_intrq(struct adapter *adap) rspq_next(q); } - t4_write_reg(adap, MYPF_REG(SGE_PF_GTS), CIDXINC(credits) | - INGRESSQID(q->cntxt_id) | SEINTARM(q->intr_params)); + val = CIDXINC(credits) | SEINTARM(q->intr_params); + if (is_t4(adap->params.chip)) { + t4_write_reg(adap, MYPF_REG(SGE_PF_GTS), + val | INGRESSQID(q->cntxt_id)); + } else { + writel(val, adap->bar2 + q->udb + SGE_UDB_GTS); + wmb(); + } spin_unlock(&adap->sge.intrq_lock); return credits; } @@ -2149,6 +2231,51 @@ static void sge_tx_timer_cb(unsigned long data) mod_timer(&s->tx_timer, jiffies + (budget ? TX_QCHECK_PERIOD : 2)); } +/** + * udb_address - return the BAR2 User Doorbell address for a Queue + * @adap: the adapter + * @cntxt_id: the Queue Context ID + * @qpp: Queues Per Page (for all PFs) + * + * Returns the BAR2 address of the user Doorbell associated with the + * indicated Queue Context ID. Note that this is only applicable + * for T5 and later. + */ +static u64 udb_address(struct adapter *adap, unsigned int cntxt_id, + unsigned int qpp) +{ + u64 udb; + unsigned int s_qpp; + unsigned short udb_density; + unsigned long qpshift; + int page; + + BUG_ON(is_t4(adap->params.chip)); + + s_qpp = (QUEUESPERPAGEPF0 + + (QUEUESPERPAGEPF1 - QUEUESPERPAGEPF0) * adap->fn); + udb_density = 1 << ((qpp >> s_qpp) & QUEUESPERPAGEPF0_MASK); + qpshift = PAGE_SHIFT - ilog2(udb_density); + udb = (u64)cntxt_id << qpshift; + udb &= PAGE_MASK; + page = udb / PAGE_SIZE; + udb += (cntxt_id - (page * udb_density)) * SGE_UDB_SIZE; + + return udb; +} + +static u64 udb_address_eq(struct adapter *adap, unsigned int cntxt_id) +{ + return udb_address(adap, cntxt_id, + t4_read_reg(adap, SGE_EGRESS_QUEUES_PER_PAGE_PF)); +} + +static u64 udb_address_iq(struct adapter *adap, unsigned int cntxt_id) +{ + return udb_address(adap, cntxt_id, + t4_read_reg(adap, SGE_INGRESS_QUEUES_PER_PAGE_PF)); +} + int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq, struct net_device *dev, int intr_idx, struct sge_fl *fl, rspq_handler_t hnd) @@ -2214,6 +2341,8 @@ int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq, iq->next_intr_params = iq->intr_params; iq->cntxt_id = ntohs(c.iqid); iq->abs_id = ntohs(c.physiqid); + if (!is_t4(adap->params.chip)) + iq->udb = udb_address_iq(adap, iq->cntxt_id); iq->size--; /* subtract status entry */ iq->netdev = dev; iq->handler = hnd; @@ -2229,6 +2358,12 @@ int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq, fl->pidx = fl->cidx = 0; fl->alloc_failed = fl->large_alloc_failed = fl->starving = 0; adap->sge.egr_map[fl->cntxt_id - adap->sge.egr_start] = fl; + + /* Note, we must initialize the Free List User Doorbell + * address before refilling the Free List! + */ + if (!is_t4(adap->params.chip)) + fl->udb = udb_address_eq(adap, fl->cntxt_id); refill_fl(adap, fl, fl_cap(fl), GFP_KERNEL); } return 0; @@ -2254,21 +2389,8 @@ err: static void init_txq(struct adapter *adap, struct sge_txq *q, unsigned int id) { q->cntxt_id = id; - if (!is_t4(adap->params.chip)) { - unsigned int s_qpp; - unsigned short udb_density; - unsigned long qpshift; - int page; - - s_qpp = QUEUESPERPAGEPF1 * adap->fn; - udb_density = 1 << QUEUESPERPAGEPF0_GET((t4_read_reg(adap, - SGE_EGRESS_QUEUES_PER_PAGE_PF) >> s_qpp)); - qpshift = PAGE_SHIFT - ilog2(udb_density); - q->udb = q->cntxt_id << qpshift; - q->udb &= PAGE_MASK; - page = q->udb / PAGE_SIZE; - q->udb += (q->cntxt_id - (page * udb_density)) * 128; - } + if (!is_t4(adap->params.chip)) + q->udb = udb_address_eq(adap, q->cntxt_id); q->in_use = 0; q->cidx = q->pidx = 0; |