Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
[cascardo/linux.git] / drivers / net / ethernet / chelsio / cxgb4 / t4_hw.c
index 2bb4efa..4d32df5 100644 (file)
@@ -3130,12 +3130,51 @@ 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.
         *