cxgb4/cxgb4vf: For T5 use Packing and Padding Boundaries for SGE DMA transfers

T5 introduces the ability to have separate Packing and Padding Boundaries
for SGE DMA transfers from the chip to Host Memory. This change set takes
advantage of that to set up a smaller Padding Boundary to conserve PCI Link
and Memory Bandwidth with T5.
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

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) +
+
+	/* 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);

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));
 
+	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.
 	 *

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)

drivers/net/ethernet/chelsio/cxgb4vf/sge.c

@@ -2436,6 +2436,7 @@ 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
@@ -2460,8 +2461,34 @@ int t4vf_sge_init(struct adapter *adapter)
 	s->stat_len = ((sge_params->sge_control & EGRSTATUSPAGESIZE_MASK)
 			? 128 : 64);
 	s->pktshift = PKTSHIFT_GET(sge_params->sge_control);
-	s->fl_align = 1 << (INGPADBOUNDARY_GET(sge_params->sge_control) +
-			    SGE_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.  (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);
+	}
 
 	/*
 	 * Set up tasklet timers.

drivers/net/ethernet/chelsio/cxgb4vf/t4vf_common.h

@@ -134,6 +134,7 @@ 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 */

drivers/net/ethernet/chelsio/cxgb4vf/t4vf_hw.c

@@ -468,6 +468,29 @@ 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);