Skip to contentSkip to conversation

MIPS: Alchemy: fix dbdma2

PATCH Applied Manuel Lauss 2 participants 2 messages May 11, 2023 – May 23, 2023 MIPS

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 MIPS: Alchemy: fix dbdma2

Various fixes for the Au1200/Au1550/Au1300 DBDMA2 code:

- skip cache invalidation if chip has working coherency circuitry.
- invalidate KSEG0-portion of the (physical) data address.
- force the dma channel doorbell write out to bus immediately with a sync.

--
This fixes Audio and SD/MMC on the DB1300 board.

---
 arch/mips/alchemy/common/dbdma.c | 27 +++++++++++++++------------
 1 file changed, 15 insertions(+), 12 deletions(-)
arch/mips/alchemy/common/dbdma.c +15 -12
index 5ab043000409..6a3c890f7bbf 100644
@@ -30,6 +30,7 @@
  *
  */
 
+#include <linux/dma-map-ops.h> /* for dma_default_coherent */
 #include <linux/init.h>
 #include <linux/kernel.h>
 #include <linux/slab.h>
@@ -623,17 +624,18 @@ u32 au1xxx_dbdma_put_source(u32 chanid, dma_addr_t buf, int nbytes, u32 flags)
 		dp->dscr_cmd0 &= ~DSCR_CMD0_IE;
 
 	/*
-	 * There is an errata on the Au1200/Au1550 parts that could result
-	 * in "stale" data being DMA'ed. It has to do with the snoop logic on
-	 * the cache eviction buffer.  DMA_NONCOHERENT is on by default for
-	 * these parts. If it is fixed in the future, these dma_cache_inv will
-	 * just be nothing more than empty macros. See io.h.
+	 * There is an erratum on certain Au1200/Au1550 revisions that could
+	 * result in "stale" data being DMA'ed. It has to do with the snoop
+	 * logic on the cache eviction buffer.  dma_default_coherent is set
+	 * to false on these parts.
 	 */
-	dma_cache_wback_inv((unsigned long)buf, nbytes);
+	if (!dma_default_coherent)
+		dma_cache_wback_inv(KSEG0ADDR(buf), nbytes);
 	dp->dscr_cmd0 |= DSCR_CMD0_V;	/* Let it rip */
 	wmb(); /* drain writebuffer */
 	dma_cache_wback_inv((unsigned long)dp, sizeof(*dp));
 	ctp->chan_ptr->ddma_dbell = 0;
+	wmb(); /* force doorbell write out to dma engine */
 
 	/* Get next descriptor pointer. */
 	ctp->put_ptr = phys_to_virt(DSCR_GET_NXTPTR(dp->dscr_nxtptr));
@@ -685,17 +687,18 @@ u32 au1xxx_dbdma_put_dest(u32 chanid, dma_addr_t buf, int nbytes, u32 flags)
 			  dp->dscr_source1, dp->dscr_dest0, dp->dscr_dest1);
 #endif
 	/*
-	 * There is an errata on the Au1200/Au1550 parts that could result in
-	 * "stale" data being DMA'ed. It has to do with the snoop logic on the
-	 * cache eviction buffer.  DMA_NONCOHERENT is on by default for these
-	 * parts. If it is fixed in the future, these dma_cache_inv will just
-	 * be nothing more than empty macros. See io.h.
+	 * There is an erratum on certain Au1200/Au1550 revisions that could
+	 * result in "stale" data being DMA'ed. It has to do with the snoop
+	 * logic on the cache eviction buffer.  dma_default_coherent is set
+	 * to false on these parts.
 	 */
-	dma_cache_inv((unsigned long)buf, nbytes);
+	if (!dma_default_coherent)
+		dma_cache_inv(KSEG0ADDR(buf), nbytes);
 	dp->dscr_cmd0 |= DSCR_CMD0_V;	/* Let it rip */
 	wmb(); /* drain writebuffer */
 	dma_cache_wback_inv((unsigned long)dp, sizeof(*dp));
 	ctp->chan_ptr->ddma_dbell = 0;
+	wmb(); /* force doorbell write out to dma engine */
 
 	/* Get next descriptor pointer. */
 	ctp->put_ptr = phys_to_virt(DSCR_GET_NXTPTR(dp->dscr_nxtptr));
Signature · 1 line
2.40.1
maintainer Applied
#1 lore raw
Quoted 7 lines · Manuel Lauss

Various fixes for the Au1200/Au1550/Au1300 DBDMA2 code:

- skip cache invalidation if chip has working coherency circuitry.
- invalidate KSEG0-portion of the (physical) data address.
- force the dma channel doorbell write out to bus immediately with a sync.

Quoted 7 lines

--
This fixes Audio and SD/MMC on the DB1300 board.

---
 arch/mips/alchemy/common/dbdma.c | 27 +++++++++++++++------------
 1 file changed, 15 insertions(+), 12 deletions(-)

applied to mips-fixes.

Thomas.

Signature · 2 lines
Crap can work. Given enough thrust pigs will fly, but it's not necessarily a
good idea.                                                [ RFC1925, 2.3 ]