Merge branch 'cxgb4-net'

Hariprasad Shenai says:

====================
cxgb4/cxgb4vf: Misc. fixes for cxgb4vf

For T5 use Packing and Padding Boundaries for SGE DMA transfers, move
fl_starve_thres to adpater structure, since they are different for each
adapter. The cxgb4vf driver's Free List Starvation Threshold needs to be larger
than the SGE's Egress Congestion Threshold or we'll end up in a mutual stall
where the driver waits for Ingress Packets to drive replacing Free List
Pointers and the SGE waits for Free List Pointers before pushing Ingress
Packets to the host.

The patches series is created against 'net' tree.
And includes patches on cxgb4 and cxgb4vf driver.

We have included all the maintainers of respective drivers. Kindly review the
change and let us know in case of any review comments.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

7 files changed, 209 insertions, 56 deletions

diff --git a/drivers/net/ethernet/chelsio/cxgb4/sge.c b/drivers/net/ethernet/chelsio/cxgb4/sge.c
index 5e1b314e11af..39f2b13e66c7 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/sge.c
+++ b/drivers/net/ethernet/chelsio/cxgb4/sge.c
@@ -2914,7 +2914,8 @@ static int t4_sge_init_hard(struct adapter *adap)
 int t4_sge_init(struct adapter *adap)
 {
 	struct sge *s = &adap->sge;
-	u32 sge_control, sge_conm_ctrl;
+	u32 sge_control, sge_control2, sge_conm_ctrl;
+	unsigned int ingpadboundary, ingpackboundary;
 	int ret, egress_threshold;
 
 	/*
@@ -2924,8 +2925,31 @@ int t4_sge_init(struct adapter *adap)
 	sge_control = t4_read_reg(adap, SGE_CONTROL);
 	s->pktshift = PKTSHIFT_GET(sge_control);
 	s->stat_len = (sge_control & EGRSTATUSPAGESIZE_MASK) ? 128 : 64;
-	s->fl_align = 1 << (INGPADBOUNDARY_GET(sge_control) +
-			    X_INGPADBOUNDARY_SHIFT);
+
+	/* T4 uses a single control field to specify both the PCIe Padding and
+	 * Packing Boundary.  T5 introduced the ability to specify these
+	 * separately.  The actual Ingress Packet Data alignment boundary
+	 * within Packed Buffer Mode is the maximum of these two
+	 * specifications.
+	 */
+	ingpadboundary = 1 << (INGPADBOUNDARY_GET(sge_control) +
+			       X_INGPADBOUNDARY_SHIFT);
+	if (is_t4(adap->params.chip)) {
+		s->fl_align = ingpadboundary;
+	} else {
+		/* T5 has a different interpretation of one of the PCIe Packing
+		 * Boundary values.
+		 */
+		sge_control2 = t4_read_reg(adap, SGE_CONTROL2_A);
+		ingpackboundary = INGPACKBOUNDARY_G(sge_control2);
+		if (ingpackboundary == INGPACKBOUNDARY_16B_X)
+			ingpackboundary = 16;
+		else
+			ingpackboundary = 1 << (ingpackboundary +
+						INGPACKBOUNDARY_SHIFT_X);
+
+		s->fl_align = max(ingpadboundary, ingpackboundary);
+	}
 
 	if (adap->flags & USING_SOFT_PARAMS)
 		ret = t4_sge_init_soft(adap);
diff --git a/drivers/net/ethernet/chelsio/cxgb4/t4_hw.c b/drivers/net/ethernet/chelsio/cxgb4/t4_hw.c
index a9d9d74e4f09..163a2a14948c 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/t4_hw.c
+++ b/drivers/net/ethernet/chelsio/cxgb4/t4_hw.c
@@ -3129,12 +3129,51 @@ int t4_fixup_host_params(struct adapter *adap, unsigned int page_size,
 		     HOSTPAGESIZEPF6(sge_hps) |
 		     HOSTPAGESIZEPF7(sge_hps));
 
-	t4_set_reg_field(adap, SGE_CONTROL,
-			 INGPADBOUNDARY_MASK |
-			 EGRSTATUSPAGESIZE_MASK,
-			 INGPADBOUNDARY(fl_align_log - 5) |
-			 EGRSTATUSPAGESIZE(stat_len != 64));
-
+	if (is_t4(adap->params.chip)) {
+		t4_set_reg_field(adap, SGE_CONTROL,
+				 INGPADBOUNDARY_MASK |
+				 EGRSTATUSPAGESIZE_MASK,
+				 INGPADBOUNDARY(fl_align_log - 5) |
+				 EGRSTATUSPAGESIZE(stat_len != 64));
+	} else {
+		/* T5 introduced the separation of the Free List Padding and
+		 * Packing Boundaries.  Thus, we can select a smaller Padding
+		 * Boundary to avoid uselessly chewing up PCIe Link and Memory
+		 * Bandwidth, and use a Packing Boundary which is large enough
+		 * to avoid false sharing between CPUs, etc.
+		 *
+		 * For the PCI Link, the smaller the Padding Boundary the
+		 * better.  For the Memory Controller, a smaller Padding
+		 * Boundary is better until we cross under the Memory Line
+		 * Size (the minimum unit of transfer to/from Memory).  If we
+		 * have a Padding Boundary which is smaller than the Memory
+		 * Line Size, that'll involve a Read-Modify-Write cycle on the
+		 * Memory Controller which is never good.  For T5 the smallest
+		 * Padding Boundary which we can select is 32 bytes which is
+		 * larger than any known Memory Controller Line Size so we'll
+		 * use that.
+		 *
+		 * T5 has a different interpretation of the "0" value for the
+		 * Packing Boundary.  This corresponds to 16 bytes instead of
+		 * the expected 32 bytes.  We never have a Packing Boundary
+		 * less than 32 bytes so we can't use that special value but
+		 * on the other hand, if we wanted 32 bytes, the best we can
+		 * really do is 64 bytes.
+		*/
+		if (fl_align <= 32) {
+			fl_align = 64;
+			fl_align_log = 6;
+		}
+		t4_set_reg_field(adap, SGE_CONTROL,
+				 INGPADBOUNDARY_MASK |
+				 EGRSTATUSPAGESIZE_MASK,
+				 INGPADBOUNDARY(INGPCIEBOUNDARY_32B_X) |
+				 EGRSTATUSPAGESIZE(stat_len != 64));
+		t4_set_reg_field(adap, SGE_CONTROL2_A,
+				 INGPACKBOUNDARY_V(INGPACKBOUNDARY_M),
+				 INGPACKBOUNDARY_V(fl_align_log -
+						 INGPACKBOUNDARY_SHIFT_X));
+	}
 	/*
 	 * Adjust various SGE Free List Host Buffer Sizes.
 	 *
diff --git a/drivers/net/ethernet/chelsio/cxgb4/t4_regs.h b/drivers/net/ethernet/chelsio/cxgb4/t4_regs.h
index a1024db5dc13..8d2de1006b08 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/t4_regs.h
+++ b/drivers/net/ethernet/chelsio/cxgb4/t4_regs.h
@@ -95,6 +95,7 @@
 #define X_INGPADBOUNDARY_SHIFT 5
 
 #define SGE_CONTROL 0x1008
+#define SGE_CONTROL2_A		0x1124
 #define  DCASYSTYPE             0x00080000U
 #define  RXPKTCPLMODE_MASK      0x00040000U
 #define  RXPKTCPLMODE_SHIFT     18
@@ -106,6 +107,7 @@
 #define  PKTSHIFT_SHIFT         10
 #define  PKTSHIFT(x)            ((x) << PKTSHIFT_SHIFT)
 #define  PKTSHIFT_GET(x)	(((x) & PKTSHIFT_MASK) >> PKTSHIFT_SHIFT)
+#define  INGPCIEBOUNDARY_32B_X	0
 #define  INGPCIEBOUNDARY_MASK   0x00000380U
 #define  INGPCIEBOUNDARY_SHIFT  7
 #define  INGPCIEBOUNDARY(x)     ((x) << INGPCIEBOUNDARY_SHIFT)
@@ -114,6 +116,14 @@
 #define  INGPADBOUNDARY(x)      ((x) << INGPADBOUNDARY_SHIFT)
 #define  INGPADBOUNDARY_GET(x)	(((x) & INGPADBOUNDARY_MASK) \
 				 >> INGPADBOUNDARY_SHIFT)
+#define  INGPACKBOUNDARY_16B_X	0
+#define  INGPACKBOUNDARY_SHIFT_X 5
+
+#define  INGPACKBOUNDARY_S	16
+#define  INGPACKBOUNDARY_M	0x7U
+#define  INGPACKBOUNDARY_V(x)	((x) << INGPACKBOUNDARY_S)
+#define  INGPACKBOUNDARY_G(x)	(((x) >> INGPACKBOUNDARY_S) \
+				 & INGPACKBOUNDARY_M)
 #define  EGRPCIEBOUNDARY_MASK   0x0000000eU
 #define  EGRPCIEBOUNDARY_SHIFT  1
 #define  EGRPCIEBOUNDARY(x)     ((x) << EGRPCIEBOUNDARY_SHIFT)
diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/adapter.h b/drivers/net/ethernet/chelsio/cxgb4vf/adapter.h
index 68eaa9c88c7d..3d06e77d7121 100644
--- a/drivers/net/ethernet/chelsio/cxgb4vf/adapter.h
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/adapter.h
@@ -299,6 +299,14 @@ struct sge {
 	u16 timer_val[SGE_NTIMERS];	/* interrupt holdoff timer array */
 	u8 counter_val[SGE_NCOUNTERS];	/* interrupt RX threshold array */
 
+	/* Decoded Adapter Parameters.
+	 */
+	u32 fl_pg_order;		/* large page allocation size */
+	u32 stat_len;			/* length of status page at ring end */
+	u32 pktshift;			/* padding between CPL & packet data */
+	u32 fl_align;			/* response queue message alignment */
+	u32 fl_starve_thres;		/* Free List starvation threshold */
+
 	/*
 	 * Reverse maps from Absolute Queue IDs to associated queue pointers.
 	 * The absolute Queue IDs are in a compact range which start at a
diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/sge.c b/drivers/net/ethernet/chelsio/cxgb4vf/sge.c
index 85036e6b42c4..fdd078d7d82c 100644
--- a/drivers/net/ethernet/chelsio/cxgb4vf/sge.c
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/sge.c
@@ -51,14 +51,6 @@
 #include "../cxgb4/t4_msg.h"
 
 /*
- * Decoded Adapter Parameters.
- */
-static u32 FL_PG_ORDER;		/* large page allocation size */
-static u32 STAT_LEN;		/* length of status page at ring end */
-static u32 PKTSHIFT;		/* padding between CPL and packet data */
-static u32 FL_ALIGN;		/* response queue message alignment */
-
-/*
  * Constants ...
  */
 enum {
@@ -102,12 +94,6 @@ enum {
 	MAX_TIMER_TX_RECLAIM = 100,
 
 	/*
-	 * An FL with <= FL_STARVE_THRES buffers is starving and a periodic
-	 * timer will attempt to refill it.
-	 */
-	FL_STARVE_THRES = 4,
-
-	/*
 	 * Suspend an Ethernet TX queue with fewer available descriptors than
 	 * this.  We always want to have room for a maximum sized packet:
 	 * inline immediate data + MAX_SKB_FRAGS. This is the same as
@@ -264,15 +250,19 @@ static inline unsigned int fl_cap(const struct sge_fl *fl)
 
 /**
  *	fl_starving - return whether a Free List is starving.
+ *	@adapter: pointer to the adapter
  *	@fl: the Free List
  *
  *	Tests specified Free List to see whether the number of buffers
  *	available to the hardware has falled below our "starvation"
  *	threshold.
  */
-static inline bool fl_starving(const struct sge_fl *fl)
+static inline bool fl_starving(const struct adapter *adapter,
+			       const struct sge_fl *fl)
 {
-	return fl->avail - fl->pend_cred <= FL_STARVE_THRES;
+	const struct sge *s = &adapter->sge;
+
+	return fl->avail - fl->pend_cred <= s->fl_starve_thres;
 }
 
 /**
@@ -457,13 +447,16 @@ static inline void reclaim_completed_tx(struct adapter *adapter,
 
 /**
  *	get_buf_size - return the size of an RX Free List buffer.
+ *	@adapter: pointer to the associated adapter
  *	@sdesc: pointer to the software buffer descriptor
  */
-static inline int get_buf_size(const struct rx_sw_desc *sdesc)
+static inline int get_buf_size(const struct adapter *adapter,
+			       const struct rx_sw_desc *sdesc)
 {
-	return FL_PG_ORDER > 0 && (sdesc->dma_addr & RX_LARGE_BUF)
-		? (PAGE_SIZE << FL_PG_ORDER)
-		: PAGE_SIZE;
+	const struct sge *s = &adapter->sge;
+
+	return (s->fl_pg_order > 0 && (sdesc->dma_addr & RX_LARGE_BUF)
+		? (PAGE_SIZE << s->fl_pg_order) : PAGE_SIZE);
 }
 
 /**
@@ -483,7 +476,8 @@ static void free_rx_bufs(struct adapter *adapter, struct sge_fl *fl, int n)
 
 		if (is_buf_mapped(sdesc))
 			dma_unmap_page(adapter->pdev_dev, get_buf_addr(sdesc),
-				       get_buf_size(sdesc), PCI_DMA_FROMDEVICE);
+				       get_buf_size(adapter, sdesc),
+				       PCI_DMA_FROMDEVICE);
 		put_page(sdesc->page);
 		sdesc->page = NULL;
 		if (++fl->cidx == fl->size)
@@ -511,7 +505,8 @@ static void unmap_rx_buf(struct adapter *adapter, struct sge_fl *fl)
 
 	if (is_buf_mapped(sdesc))
 		dma_unmap_page(adapter->pdev_dev, get_buf_addr(sdesc),
-			       get_buf_size(sdesc), PCI_DMA_FROMDEVICE);
+			       get_buf_size(adapter, sdesc),
+			       PCI_DMA_FROMDEVICE);
 	sdesc->page = NULL;
 	if (++fl->cidx == fl->size)
 		fl->cidx = 0;
@@ -589,6 +584,7 @@ static inline void poison_buf(struct page *page, size_t sz)
 static unsigned int refill_fl(struct adapter *adapter, struct sge_fl *fl,
 			      int n, gfp_t gfp)
 {
+	struct sge *s = &adapter->sge;
 	struct page *page;
 	dma_addr_t dma_addr;
 	unsigned int cred = fl->avail;
@@ -608,12 +604,12 @@ static unsigned int refill_fl(struct adapter *adapter, struct sge_fl *fl,
 	 * If we don't support large pages, drop directly into the small page
 	 * allocation code.
 	 */
-	if (FL_PG_ORDER == 0)
+	if (s->fl_pg_order == 0)
 		goto alloc_small_pages;
 
 	while (n) {
 		page = alloc_pages(gfp | __GFP_COMP | __GFP_NOWARN,
-				   FL_PG_ORDER);
+				   s->fl_pg_order);
 		if (unlikely(!page)) {
 			/*
 			 * We've failed inour attempt to allocate a "large
@@ -623,10 +619,10 @@ static unsigned int refill_fl(struct adapter *adapter, struct sge_fl *fl,
 			fl->large_alloc_failed++;
 			break;
 		}
-		poison_buf(page, PAGE_SIZE << FL_PG_ORDER);
+		poison_buf(page, PAGE_SIZE << s->fl_pg_order);
 
 		dma_addr = dma_map_page(adapter->pdev_dev, page, 0,
-					PAGE_SIZE << FL_PG_ORDER,
+					PAGE_SIZE << s->fl_pg_order,
 					PCI_DMA_FROMDEVICE);
 		if (unlikely(dma_mapping_error(adapter->pdev_dev, dma_addr))) {
 			/*
@@ -637,7 +633,7 @@ static unsigned int refill_fl(struct adapter *adapter, struct sge_fl *fl,
 			 * because DMA mapping resources are typically
 			 * critical resources once they become scarse.
 			 */
-			__free_pages(page, FL_PG_ORDER);
+			__free_pages(page, s->fl_pg_order);
 			goto out;
 		}
 		dma_addr |= RX_LARGE_BUF;
@@ -693,7 +689,7 @@ out:
 	fl->pend_cred += cred;
 	ring_fl_db(adapter, fl);
 
-	if (unlikely(fl_starving(fl))) {
+	if (unlikely(fl_starving(adapter, fl))) {
 		smp_wmb();
 		set_bit(fl->cntxt_id, adapter->sge.starving_fl);
 	}
@@ -1468,6 +1464,8 @@ static void t4vf_pktgl_free(const struct pkt_gl *gl)
 static void do_gro(struct sge_eth_rxq *rxq, const struct pkt_gl *gl,
 		   const struct cpl_rx_pkt *pkt)
 {
+	struct adapter *adapter = rxq->rspq.adapter;
+	struct sge *s = &adapter->sge;
 	int ret;
 	struct sk_buff *skb;
 
@@ -1478,8 +1476,8 @@ static void do_gro(struct sge_eth_rxq *rxq, const struct pkt_gl *gl,
 		return;
 	}
 
-	copy_frags(skb, gl, PKTSHIFT);
-	skb->len = gl->tot_len - PKTSHIFT;
+	copy_frags(skb, gl, s->pktshift);
+	skb->len = gl->tot_len - s->pktshift;
 	skb->data_len = skb->len;
 	skb->truesize += skb->data_len;
 	skb->ip_summed = CHECKSUM_UNNECESSARY;
@@ -1516,6 +1514,8 @@ int t4vf_ethrx_handler(struct sge_rspq *rspq, const __be64 *rsp,
 	bool csum_ok = pkt->csum_calc && !pkt->err_vec &&
 		       (rspq->netdev->features & NETIF_F_RXCSUM);
 	struct sge_eth_rxq *rxq = container_of(rspq, struct sge_eth_rxq, rspq);
+	struct adapter *adapter = rspq->adapter;
+	struct sge *s = &adapter->sge;
 
 	/*
 	 * If this is a good TCP packet and we have Generic Receive Offload
@@ -1537,7 +1537,7 @@ int t4vf_ethrx_handler(struct sge_rspq *rspq, const __be64 *rsp,
 		rxq->stats.rx_drops++;
 		return 0;
 	}
-	__skb_pull(skb, PKTSHIFT);
+	__skb_pull(skb, s->pktshift);
 	skb->protocol = eth_type_trans(skb, rspq->netdev);
 	skb_record_rx_queue(skb, rspq->idx);
 	rxq->stats.pkts++;
@@ -1648,6 +1648,8 @@ static inline void rspq_next(struct sge_rspq *rspq)
 static int process_responses(struct sge_rspq *rspq, int budget)
 {
 	struct sge_eth_rxq *rxq = container_of(rspq, struct sge_eth_rxq, rspq);
+	struct adapter *adapter = rspq->adapter;
+	struct sge *s = &adapter->sge;
 	int budget_left = budget;
 
 	while (likely(budget_left)) {
@@ -1697,7 +1699,7 @@ static int process_responses(struct sge_rspq *rspq, int budget)
 				BUG_ON(frag >= MAX_SKB_FRAGS);
 				BUG_ON(rxq->fl.avail == 0);
 				sdesc = &rxq->fl.sdesc[rxq->fl.cidx];
-				bufsz = get_buf_size(sdesc);
+				bufsz = get_buf_size(adapter, sdesc);
 				fp->page = sdesc->page;
 				fp->offset = rspq->offset;
 				fp->size = min(bufsz, len);
@@ -1726,7 +1728,7 @@ static int process_responses(struct sge_rspq *rspq, int budget)
 			 */
 			ret = rspq->handler(rspq, rspq->cur_desc, &gl);
 			if (likely(ret == 0))
-				rspq->offset += ALIGN(fp->size, FL_ALIGN);
+				rspq->offset += ALIGN(fp->size, s->fl_align);
 			else
 				restore_rx_bufs(&gl, &rxq->fl, frag);
 		} else if (likely(rsp_type == RSP_TYPE_CPL)) {
@@ -1963,7 +1965,7 @@ static void sge_rx_timer_cb(unsigned long data)
 			 * schedule napi but the FL is no longer starving.
 			 * No biggie.
 			 */
-			if (fl_starving(fl)) {
+			if (fl_starving(adapter, fl)) {
 				struct sge_eth_rxq *rxq;
 
 				rxq = container_of(fl, struct sge_eth_rxq, fl);
@@ -2047,6 +2049,7 @@ int t4vf_sge_alloc_rxq(struct adapter *adapter, struct sge_rspq *rspq,
 		       int intr_dest,
 		       struct sge_fl *fl, rspq_handler_t hnd)
 {
+	struct sge *s = &adapter->sge;
 	struct port_info *pi = netdev_priv(dev);
 	struct fw_iq_cmd cmd, rpl;
 	int ret, iqandst, flsz = 0;
@@ -2117,7 +2120,7 @@ int t4vf_sge_alloc_rxq(struct adapter *adapter, struct sge_rspq *rspq,
 		fl->size = roundup(fl->size, FL_PER_EQ_UNIT);
 		fl->desc = alloc_ring(adapter->pdev_dev, fl->size,
 				      sizeof(__be64), sizeof(struct rx_sw_desc),
-				      &fl->addr, &fl->sdesc, STAT_LEN);
+				      &fl->addr, &fl->sdesc, s->stat_len);
 		if (!fl->desc) {
 			ret = -ENOMEM;
 			goto err;
@@ -2129,7 +2132,7 @@ int t4vf_sge_alloc_rxq(struct adapter *adapter, struct sge_rspq *rspq,
 		 * free list ring) in Egress Queue Units.
 		 */
 		flsz = (fl->size / FL_PER_EQ_UNIT +
-			STAT_LEN / EQ_UNIT);
+			s->stat_len / EQ_UNIT);
 
 		/*
 		 * Fill in all the relevant firmware Ingress Queue Command
@@ -2217,6 +2220,7 @@ int t4vf_sge_alloc_eth_txq(struct adapter *adapter, struct sge_eth_txq *txq,
 			   struct net_device *dev, struct netdev_queue *devq,
 			   unsigned int iqid)
 {
+	struct sge *s = &adapter->sge;
 	int ret, nentries;
 	struct fw_eq_eth_cmd cmd, rpl;
 	struct port_info *pi = netdev_priv(dev);
@@ -2225,7 +2229,7 @@ int t4vf_sge_alloc_eth_txq(struct adapter *adapter, struct sge_eth_txq *txq,
 	 * Calculate the size of the hardware TX Queue (including the Status
 	 * Page on the end of the TX Queue) in units of TX Descriptors.
 	 */
-	nentries = txq->q.size + STAT_LEN / sizeof(struct tx_desc);
+	nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
 
 	/*
 	 * Allocate the hardware ring for the TX ring (with space for its
@@ -2234,7 +2238,7 @@ int t4vf_sge_alloc_eth_txq(struct adapter *adapter, struct sge_eth_txq *txq,
 	txq->q.desc = alloc_ring(adapter->pdev_dev, txq->q.size,
 				 sizeof(struct tx_desc),
 				 sizeof(struct tx_sw_desc),
-				 &txq->q.phys_addr, &txq->q.sdesc, STAT_LEN);
+				 &txq->q.phys_addr, &txq->q.sdesc, s->stat_len);
 	if (!txq->q.desc)
 		return -ENOMEM;
 
@@ -2307,8 +2311,10 @@ int t4vf_sge_alloc_eth_txq(struct adapter *adapter, struct sge_eth_txq *txq,
  */
 static void free_txq(struct adapter *adapter, struct sge_txq *tq)
 {
+	struct sge *s = &adapter->sge;
+
 	dma_free_coherent(adapter->pdev_dev,
-			  tq->size * sizeof(*tq->desc) + STAT_LEN,
+			  tq->size * sizeof(*tq->desc) + s->stat_len,
 			  tq->desc, tq->phys_addr);
 	tq->cntxt_id = 0;
 	tq->sdesc = NULL;
@@ -2322,6 +2328,7 @@ static void free_txq(struct adapter *adapter, struct sge_txq *tq)
 static void free_rspq_fl(struct adapter *adapter, struct sge_rspq *rspq,
 			 struct sge_fl *fl)
 {
+	struct sge *s = &adapter->sge;
 	unsigned int flid = fl ? fl->cntxt_id : 0xffff;
 
 	t4vf_iq_free(adapter, FW_IQ_TYPE_FL_INT_CAP,
@@ -2337,7 +2344,7 @@ static void free_rspq_fl(struct adapter *adapter, struct sge_rspq *rspq,
 	if (fl) {
 		free_rx_bufs(adapter, fl, fl->avail);
 		dma_free_coherent(adapter->pdev_dev,
-				  fl->size * sizeof(*fl->desc) + STAT_LEN,
+				  fl->size * sizeof(*fl->desc) + s->stat_len,
 				  fl->desc, fl->addr);
 		kfree(fl->sdesc);
 		fl->sdesc = NULL;
@@ -2423,6 +2430,7 @@ int t4vf_sge_init(struct adapter *adapter)
 	u32 fl0 = sge_params->sge_fl_buffer_size[0];
 	u32 fl1 = sge_params->sge_fl_buffer_size[1];
 	struct sge *s = &adapter->sge;
+	unsigned int ingpadboundary, ingpackboundary;
 
 	/*
 	 * Start by vetting the basic SGE parameters which have been set up by
@@ -2443,12 +2451,48 @@ int t4vf_sge_init(struct adapter *adapter)
 	 * Now translate the adapter parameters into our internal forms.
 	 */
 	if (fl1)
-		FL_PG_ORDER = ilog2(fl1) - PAGE_SHIFT;
-	STAT_LEN = ((sge_params->sge_control & EGRSTATUSPAGESIZE_MASK)
-		    ? 128 : 64);
-	PKTSHIFT = PKTSHIFT_GET(sge_params->sge_control);
-	FL_ALIGN = 1 << (INGPADBOUNDARY_GET(sge_params->sge_control) +
-			 SGE_INGPADBOUNDARY_SHIFT);
+		s->fl_pg_order = ilog2(fl1) - PAGE_SHIFT;
+	s->stat_len = ((sge_params->sge_control & EGRSTATUSPAGESIZE_MASK)
+			? 128 : 64);
+	s->pktshift = PKTSHIFT_GET(sge_params->sge_control);
+
+	/* T4 uses a single control field to specify both the PCIe Padding and
+	 * Packing Boundary.  T5 introduced the ability to specify these
+	 * separately.  The actual Ingress Packet Data alignment boundary
+	 * within Packed Buffer Mode is the maximum of these two
+	 * specifications.  (Note that it makes no real practical sense to
+	 * have the Pading Boudary be larger than the Packing Boundary but you
+	 * could set the chip up that way and, in fact, legacy T4 code would
+	 * end doing this because it would initialize the Padding Boundary and
+	 * leave the Packing Boundary initialized to 0 (16 bytes).)
+	 */
+	ingpadboundary = 1 << (INGPADBOUNDARY_GET(sge_params->sge_control) +
+			       X_INGPADBOUNDARY_SHIFT);
+	if (is_t4(adapter->params.chip)) {
+		s->fl_align = ingpadboundary;
+	} else {
+		/* T5 has a different interpretation of one of the PCIe Packing
+		 * Boundary values.
+		 */
+		ingpackboundary = INGPACKBOUNDARY_G(sge_params->sge_control2);
+		if (ingpackboundary == INGPACKBOUNDARY_16B_X)
+			ingpackboundary = 16;
+		else
+			ingpackboundary = 1 << (ingpackboundary +
+						INGPACKBOUNDARY_SHIFT_X);
+
+		s->fl_align = max(ingpadboundary, ingpackboundary);
+	}
+
+	/* A FL with <= fl_starve_thres buffers is starving and a periodic
+	 * timer will attempt to refill it.  This needs to be larger than the
+	 * SGE's Egress Congestion Threshold.  If it isn't, then we can get
+	 * stuck waiting for new packets while the SGE is waiting for us to
+	 * give it more Free List entries.  (Note that the SGE's Egress
+	 * Congestion Threshold is in units of 2 Free List pointers.)
+	 */
+	s->fl_starve_thres
+		= EGRTHRESHOLD_GET(sge_params->sge_congestion_control)*2 + 1;
 
 	/*
 	 * Set up tasklet timers.
diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_common.h b/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_common.h
index 95df61dcb4ce..4b6a6d14d86d 100644
--- a/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_common.h
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_common.h
@@ -134,11 +134,13 @@ struct dev_params {
  */
 struct sge_params {
 	u32 sge_control;		/* padding, boundaries, lengths, etc. */
+	u32 sge_control2;		/* T5: more of the same */
 	u32 sge_host_page_size;		/* RDMA page sizes */
 	u32 sge_queues_per_page;	/* RDMA queues/page */
 	u32 sge_user_mode_limits;	/* limits for BAR2 user mode accesses */
 	u32 sge_fl_buffer_size[16];	/* free list buffer sizes */
 	u32 sge_ingress_rx_threshold;	/* RX counter interrupt threshold[4] */
+	u32 sge_congestion_control;     /* congestion thresholds, etc. */
 	u32 sge_timer_value_0_and_1;	/* interrupt coalescing timer values */
 	u32 sge_timer_value_2_and_3;
 	u32 sge_timer_value_4_and_5;
diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_hw.c b/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_hw.c
index e984fdc48ba2..1e896b923234 100644
--- a/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_hw.c
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/t4vf_hw.c
@@ -468,12 +468,38 @@ int t4vf_get_sge_params(struct adapter *adapter)
 	sge_params->sge_timer_value_2_and_3 = vals[5];
 	sge_params->sge_timer_value_4_and_5 = vals[6];
 
+	/* T4 uses a single control field to specify both the PCIe Padding and
+	 * Packing Boundary.  T5 introduced the ability to specify these
+	 * separately with the Padding Boundary in SGE_CONTROL and and Packing
+	 * Boundary in SGE_CONTROL2.  So for T5 and later we need to grab
+	 * SGE_CONTROL in order to determine how ingress packet data will be
+	 * laid out in Packed Buffer Mode.  Unfortunately, older versions of
+	 * the firmware won't let us retrieve SGE_CONTROL2 so if we get a
+	 * failure grabbing it we throw an error since we can't figure out the
+	 * right value.
+	 */
+	if (!is_t4(adapter->params.chip)) {
+		params[0] = (FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
+			     FW_PARAMS_PARAM_XYZ(SGE_CONTROL2_A));
+		v = t4vf_query_params(adapter, 1, params, vals);
+		if (v != FW_SUCCESS) {
+			dev_err(adapter->pdev_dev,
+				"Unable to get SGE Control2; "
+				"probably old firmware.\n");
+			return v;
+		}
+		sge_params->sge_control2 = vals[0];
+	}
+
 	params[0] = (FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
 		     FW_PARAMS_PARAM_XYZ(SGE_INGRESS_RX_THRESHOLD));
-	v = t4vf_query_params(adapter, 1, params, vals);
+	params[1] = (FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) |
+		     FW_PARAMS_PARAM_XYZ(SGE_CONM_CTRL));
+	v = t4vf_query_params(adapter, 2, params, vals);
 	if (v)
 		return v;
 	sge_params->sge_ingress_rx_threshold = vals[0];
+	sge_params->sge_congestion_control = vals[1];
 
 	return 0;
 }