cxgb4/cxgb4vf: Use new interfaces to calculate BAR2 SGE Queue Register addresses
Use BAR2 Going To Sleep (GTS) for T5 and later. Use new BAR2 User Doorbells for
T5 for both cxgb4 and cxgb4vf driver.
Based on original work by Casey Leedom <leedom@chelsio.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/drivers/net/ethernet/chelsio/cxgb4/cxgb4.h b/drivers/net/ethernet/chelsio/cxgb4/cxgb4.h
index 40905c6..c38a936 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/cxgb4.h
+++ b/drivers/net/ethernet/chelsio/cxgb4/cxgb4.h
@@ -438,7 +438,8 @@
struct rx_sw_desc *sdesc; /* address of SW Rx descriptor ring */
__be64 *desc; /* address of HW Rx descriptor ring */
dma_addr_t addr; /* bus address of HW ring start */
- u64 udb; /* BAR2 offset of User Doorbell area */
+ void __iomem *bar2_addr; /* address of BAR2 Queue registers */
+ unsigned int bar2_qid; /* Queue ID for BAR2 Queue registers */
};
/* A packet gather list */
@@ -468,7 +469,8 @@
u16 abs_id; /* absolute SGE id for the response q */
__be64 *desc; /* address of HW response ring */
dma_addr_t phys_addr; /* physical address of the ring */
- u64 udb; /* BAR2 offset of User Doorbell area */
+ void __iomem *bar2_addr; /* address of BAR2 Queue registers */
+ unsigned int bar2_qid; /* Queue ID for BAR2 Queue registers */
unsigned int iqe_len; /* entry size */
unsigned int size; /* capacity of response queue */
struct adapter *adap;
@@ -526,7 +528,8 @@
int db_disabled;
unsigned short db_pidx;
unsigned short db_pidx_inc;
- u64 udb; /* BAR2 offset of User Doorbell area */
+ void __iomem *bar2_addr; /* address of BAR2 Queue registers */
+ unsigned int bar2_qid; /* Queue ID for BAR2 Queue registers */
};
struct sge_eth_txq { /* state for an SGE Ethernet Tx queue */
diff --git a/drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c b/drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c
index e7342bc..4c26be9 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c
+++ b/drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c
@@ -3805,6 +3805,22 @@
}
EXPORT_SYMBOL(cxgb4_read_sge_timestamp);
+int cxgb4_bar2_sge_qregs(struct net_device *dev,
+ unsigned int qid,
+ enum cxgb4_bar2_qtype qtype,
+ u64 *pbar2_qoffset,
+ unsigned int *pbar2_qid)
+{
+ return t4_bar2_sge_qregs(netdev2adap(dev),
+ qid,
+ (qtype == CXGB4_BAR2_QTYPE_EGRESS
+ ? T4_BAR2_QTYPE_EGRESS
+ : T4_BAR2_QTYPE_INGRESS),
+ pbar2_qoffset,
+ pbar2_qid);
+}
+EXPORT_SYMBOL(cxgb4_bar2_sge_qregs);
+
static struct pci_driver cxgb4_driver;
static void check_neigh_update(struct neighbour *neigh)
@@ -3987,31 +4003,18 @@
u32 dropped_db = t4_read_reg(adap, 0x010ac);
u16 qid = (dropped_db >> 15) & 0x1ffff;
u16 pidx_inc = dropped_db & 0x1fff;
- unsigned int s_qpp;
- unsigned short udb_density;
- unsigned long qpshift;
- int page;
- u32 udb;
+ u64 bar2_qoffset;
+ unsigned int bar2_qid;
+ int ret;
- dev_warn(adap->pdev_dev,
- "Dropped DB 0x%x qid %d bar2 %d coalesce %d pidx %d\n",
- dropped_db, qid,
- (dropped_db >> 14) & 1,
- (dropped_db >> 13) & 1,
- pidx_inc);
-
- drain_db_fifo(adap, 1);
-
- 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);
- udb = qid << qpshift;
- udb &= PAGE_MASK;
- page = udb / PAGE_SIZE;
- udb += (qid - (page * udb_density)) * 128;
-
- writel(PIDX(pidx_inc), adap->bar2 + udb + 8);
+ ret = t4_bar2_sge_qregs(adap, qid, T4_BAR2_QTYPE_EGRESS,
+ &bar2_qoffset, &bar2_qid);
+ if (ret)
+ dev_err(adap->pdev_dev, "doorbell drop recovery: "
+ "qid=%d, pidx_inc=%d\n", qid, pidx_inc);
+ else
+ writel(PIDX_T5(pidx_inc) | QID(bar2_qid),
+ adap->bar2 + bar2_qoffset + SGE_UDB_KDOORBELL);
/* Re-enable BAR2 WC */
t4_set_reg_field(adap, 0x10b0, 1<<15, 1<<15);
@@ -4069,12 +4072,8 @@
lli.adapter_type = adap->params.chip;
lli.iscsi_iolen = MAXRXDATA_GET(t4_read_reg(adap, TP_PARA_REG2));
lli.cclk_ps = 1000000000 / adap->params.vpd.cclk;
- lli.udb_density = 1 << QUEUESPERPAGEPF0_GET(
- t4_read_reg(adap, SGE_EGRESS_QUEUES_PER_PAGE_PF) >>
- (adap->fn * 4));
- lli.ucq_density = 1 << QUEUESPERPAGEPF0_GET(
- t4_read_reg(adap, SGE_INGRESS_QUEUES_PER_PAGE_PF) >>
- (adap->fn * 4));
+ lli.udb_density = 1 << adap->params.sge.eq_qpp;
+ lli.ucq_density = 1 << adap->params.sge.iq_qpp;
lli.filt_mode = adap->params.tp.vlan_pri_map;
/* MODQ_REQ_MAP sets queues 0-3 to chan 0-3 */
for (i = 0; i < NCHAN; i++)
@@ -5926,6 +5925,7 @@
t4_load_mtus(adap, adap->params.mtus, adap->params.a_wnd,
adap->params.b_wnd);
}
+ t4_init_sge_params(adap);
t4_init_tp_params(adap);
adap->flags |= FW_OK;
return 0;
diff --git a/drivers/net/ethernet/chelsio/cxgb4/cxgb4_uld.h b/drivers/net/ethernet/chelsio/cxgb4/cxgb4_uld.h
index 4eba7cb..152b4c4 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/cxgb4_uld.h
+++ b/drivers/net/ethernet/chelsio/cxgb4/cxgb4_uld.h
@@ -305,4 +305,11 @@
int cxgb4_read_tpte(struct net_device *dev, u32 stag, __be32 *tpte);
u64 cxgb4_read_sge_timestamp(struct net_device *dev);
+enum cxgb4_bar2_qtype { CXGB4_BAR2_QTYPE_EGRESS, CXGB4_BAR2_QTYPE_INGRESS };
+int cxgb4_bar2_sge_qregs(struct net_device *dev,
+ unsigned int qid,
+ enum cxgb4_bar2_qtype qtype,
+ u64 *pbar2_qoffset,
+ unsigned int *pbar2_qid);
+
#endif /* !__CXGB4_OFLD_H */
diff --git a/drivers/net/ethernet/chelsio/cxgb4/sge.c b/drivers/net/ethernet/chelsio/cxgb4/sge.c
index 433560b..f12debd 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/sge.c
+++ b/drivers/net/ethernet/chelsio/cxgb4/sge.c
@@ -527,14 +527,16 @@
val |= DBPRIO(1);
wmb();
- /* If we're on T4, use the old doorbell mechanism; otherwise
- * use the new BAR2 mechanism.
+ /* If we don't have access to the new User Doorbell (T5+), use
+ * the old doorbell mechanism; otherwise use the new BAR2
+ * mechanism.
*/
- if (is_t4(adap->params.chip)) {
+ if (unlikely(q->bar2_addr == NULL)) {
t4_write_reg(adap, MYPF_REG(SGE_PF_KDOORBELL),
val | QID(q->cntxt_id));
} else {
- writel(val, adap->bar2 + q->udb + SGE_UDB_KDOORBELL);
+ writel(val | QID(q->bar2_qid),
+ q->bar2_addr + SGE_UDB_KDOORBELL);
/* This Write memory Barrier will force the write to
* the User Doorbell area to be flushed.
@@ -850,14 +852,13 @@
*end = 0;
}
-/* 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.
+/* This function copies 64 byte coalesced work request to
+ * memory mapped BAR2 space. For coalesced WR SGE fetches
+ * data from the FIFO instead of from Host.
*/
-static void cxgb_pio_copy(u64 __iomem *dst, struct tx_desc *desc)
+static void cxgb_pio_copy(u64 __iomem *dst, u64 *src)
{
- int count = sizeof(*desc) / sizeof(u64);
- u64 *src = (u64 *)desc;
+ int count = 8;
while (count) {
writeq(*src, dst);
@@ -879,7 +880,10 @@
{
wmb(); /* write descriptors before telling HW */
- if (is_t4(adap->params.chip)) {
+ /* If we don't have access to the new User Doorbell (T5+), use the old
+ * doorbell mechanism; otherwise use the new BAR2 mechanism.
+ */
+ if (unlikely(q->bar2_addr == NULL)) {
u32 val = PIDX(n);
unsigned long flags;
@@ -905,21 +909,22 @@
*/
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 we're only writing a single TX Descriptor and we can use
+ * Inferred QID registers, we can use the Write Combining
+ * Gather Buffer; otherwise we use the simple doorbell.
*/
- if (n == 1) {
+ if (n == 1 && q->bar2_qid == 0) {
int index = (q->pidx
? (q->pidx - 1)
: (q->size - 1));
+ u64 *wr = (u64 *)&q->desc[index];
- cxgb_pio_copy(adap->bar2 + q->udb + SGE_UDB_WCDOORBELL,
- q->desc + index);
+ cxgb_pio_copy((u64 __iomem *)
+ (q->bar2_addr + SGE_UDB_WCDOORBELL),
+ wr);
} else {
- writel(val, adap->bar2 + q->udb + SGE_UDB_KDOORBELL);
+ writel(val | QID(q->bar2_qid),
+ q->bar2_addr + SGE_UDB_KDOORBELL);
}
/* This Write Memory Barrier will force the write to the User
@@ -1997,11 +2002,16 @@
params = QINTR_TIMER_IDX(7);
val = CIDXINC(work_done) | SEINTARM(params);
- if (is_t4(q->adap->params.chip)) {
+
+ /* If we don't have access to the new User GTS (T5+), use the old
+ * doorbell mechanism; otherwise use the new BAR2 mechanism.
+ */
+ if (unlikely(q->bar2_addr == NULL)) {
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);
+ writel(val | INGRESSQID(q->bar2_qid),
+ q->bar2_addr + SGE_UDB_GTS);
wmb();
}
return work_done;
@@ -2047,11 +2057,16 @@
}
val = CIDXINC(credits) | SEINTARM(q->intr_params);
- if (is_t4(adap->params.chip)) {
+
+ /* If we don't have access to the new User GTS (T5+), use the old
+ * doorbell mechanism; otherwise use the new BAR2 mechanism.
+ */
+ if (unlikely(q->bar2_addr == NULL)) {
t4_write_reg(adap, MYPF_REG(SGE_PF_GTS),
val | INGRESSQID(q->cntxt_id));
} else {
- writel(val, adap->bar2 + q->udb + SGE_UDB_GTS);
+ writel(val | INGRESSQID(q->bar2_qid),
+ q->bar2_addr + SGE_UDB_GTS);
wmb();
}
spin_unlock(&adap->sge.intrq_lock);
@@ -2235,48 +2250,32 @@
}
/**
- * 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)
+ * bar2_address - return the BAR2 address for an SGE Queue's Registers
+ * @adapter: the adapter
+ * @qid: the SGE Queue ID
+ * @qtype: the SGE Queue Type (Egress or Ingress)
+ * @pbar2_qid: BAR2 Queue ID or 0 for Queue ID inferred SGE Queues
*
- * 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.
+ * Returns the BAR2 address for the SGE Queue Registers associated with
+ * @qid. If BAR2 SGE Registers aren't available, returns NULL. Also
+ * returns the BAR2 Queue ID to be used with writes to the BAR2 SGE
+ * Queue Registers. If the BAR2 Queue ID is 0, then "Inferred Queue ID"
+ * Registers are supported (e.g. the Write Combining Doorbell Buffer).
*/
-static u64 udb_address(struct adapter *adap, unsigned int cntxt_id,
- unsigned int qpp)
+static void __iomem *bar2_address(struct adapter *adapter,
+ unsigned int qid,
+ enum t4_bar2_qtype qtype,
+ unsigned int *pbar2_qid)
{
- u64 udb;
- unsigned int s_qpp;
- unsigned short udb_density;
- unsigned long qpshift;
- int page;
+ u64 bar2_qoffset;
+ int ret;
- BUG_ON(is_t4(adap->params.chip));
+ ret = t4_bar2_sge_qregs(adapter, qid, qtype,
+ &bar2_qoffset, pbar2_qid);
+ if (ret)
+ return NULL;
- 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));
+ return adapter->bar2 + bar2_qoffset;
}
int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
@@ -2344,8 +2343,10 @@
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->bar2_addr = bar2_address(adap,
+ iq->cntxt_id,
+ T4_BAR2_QTYPE_INGRESS,
+ &iq->bar2_qid);
iq->size--; /* subtract status entry */
iq->netdev = dev;
iq->handler = hnd;
@@ -2362,11 +2363,13 @@
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!
+ /* Note, we must initialize the BAR2 Free List User Doorbell
+ * information before refilling the Free List!
*/
- if (!is_t4(adap->params.chip))
- fl->udb = udb_address_eq(adap, fl->cntxt_id);
+ fl->bar2_addr = bar2_address(adap,
+ fl->cntxt_id,
+ T4_BAR2_QTYPE_EGRESS,
+ &fl->bar2_qid);
refill_fl(adap, fl, fl_cap(fl), GFP_KERNEL);
}
return 0;
@@ -2392,9 +2395,10 @@
static void init_txq(struct adapter *adap, struct sge_txq *q, unsigned int id)
{
q->cntxt_id = id;
- if (!is_t4(adap->params.chip))
- q->udb = udb_address_eq(adap, q->cntxt_id);
-
+ q->bar2_addr = bar2_address(adap,
+ q->cntxt_id,
+ T4_BAR2_QTYPE_EGRESS,
+ &q->bar2_qid);
q->in_use = 0;
q->cidx = q->pidx = 0;
q->stops = q->restarts = 0;