[src/trunk]: src/sys/arch/alpha/pci According to section 8.1.2.2 of the Tsuna...

[thorpej <thorpej%NetBSD.org@localhost>]

details:   https://anonhg.NetBSD.org/src/rev/25569f895b17
branches:  trunk
changeset: 984708:25569f895b17
user:      thorpej <thorpej%NetBSD.org@localhost>
date:      Sun Jul 18 19:58:34 2021 +0000

description:
According to section 8.1.2.2 of the Tsunami/Typhoon hardware reference
manual (DS-0025A-TE), the SGMAP TLB is arranged as 168 locations of 4
consecutive quadwords.  It seems that on some revisions of the Pchip,
SGMAP translation is not perfectly reliable unless we align the DMA
segments to the TLBs natural boundaries (observed on the API CS20).

N.B. the Titan (as observed on a Compaq DS25) does not seem to have this
problem, but we'll play it safe and run this way on both variants.

PR port-alpha/40604.

diffstat:

 sys/arch/alpha/pci/tsp_dma.c |  23 +++++++++++++++++++++--
 1 files changed, 21 insertions(+), 2 deletions(-)

diffs (51 lines):

diff -r d26e4fd86b79 -r 25569f895b17 sys/arch/alpha/pci/tsp_dma.c
--- a/sys/arch/alpha/pci/tsp_dma.c      Sun Jul 18 11:45:31 2021 +0000
+++ b/sys/arch/alpha/pci/tsp_dma.c      Sun Jul 18 19:58:34 2021 +0000
@@ -1,4 +1,4 @@
-/* $NetBSD: tsp_dma.c,v 1.20 2021/07/18 05:09:47 thorpej Exp $ */
+/* $NetBSD: tsp_dma.c,v 1.21 2021/07/18 19:58:34 thorpej Exp $ */
 
 /*-
  * Copyright (c) 1997, 1998, 2021 The NetBSD Foundation, Inc.
@@ -61,7 +61,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: tsp_dma.c,v 1.20 2021/07/18 05:09:47 thorpej Exp $");
+__KERNEL_RCSID(0, "$NetBSD: tsp_dma.c,v 1.21 2021/07/18 19:58:34 thorpej Exp $");
 
 #include <sys/param.h>
 #include <sys/systm.h>
@@ -187,6 +187,20 @@
                t->_next_window = NULL;
                t->_boundary = 0;
                t->_sgmap = &pcp->pc_sgmap_hi;
+               /*
+                * According to section 8.1.2.2 of the Tsunami/Typhoon
+                * hardware reference manual (DS-0025A-TE), the SGMAP
+                * TLB is arranged as 168 locations of 4 consecutive
+                * quadwords.  It seems that on some revisions of the
+                * Pchip, SGMAP translation is not perfectly reliable
+                * unless we align the DMA segments to the TLBs natural
+                * boundaries (observed on the API CS20).
+                *
+                * N.B. the Titan (as observed on a Compaq DS25) does not
+                * seem to have this problem, but we'll play it safe and
+                * run this way on both variants.
+                */
+               t->_sgmap_minalign = PAGE_SIZE * 4;
                t->_pfthresh = TSP_SGMAP_PFTHRESH;
                t->_get_tag = tsp_dma_get_tag;
                t->_dmamap_create = alpha_sgmap_dmamap_create;
@@ -267,6 +281,11 @@
        t->_next_window = NULL;
        t->_boundary = 0;
        t->_sgmap = &pcp->pc_sgmap_lo;
+       /*
+        * This appears to be needed to make DMA on the ALI southbridge
+        * that's present in some systems happy.  ???
+        */
+       t->_sgmap_minalign = (64UL * 1024);
        t->_pfthresh = TSP_SGMAP_PFTHRESH;
        t->_get_tag = tsp_dma_get_tag;
        t->_dmamap_create = alpha_sgmap_dmamap_create;