path: root/packages/linux/linux-mtx-2-2.4.27/44-dbdma-and-au1550_psc.diff

--- linux-old/include/asm-mips/au1xxx_dbdma.h	2006-05-09 18:24:17.000000000 +0200
+++ linux/include/asm-mips/au1xxx_dbdma.h	2006-05-09 14:05:48.000000000 +0200
@@ -34,6 +34,8 @@
 #ifndef _AU1000_DBDMA_H_
 #define _AU1000_DBDMA_H_
 
+#include <linux/config.h>
+
 #ifndef _LANGUAGE_ASSEMBLY
 
 /* The DMA base addresses.
@@ -43,7 +45,7 @@
 #define DDMA_GLOBAL_BASE	0xb4003000
 #define DDMA_CHANNEL_BASE	0xb4002000
 
-typedef struct dbdma_global {
+typedef volatile struct dbdma_global {
 	u32	ddma_config;
 	u32	ddma_intstat;
 	u32	ddma_throttle;
@@ -60,7 +62,7 @@
 
 /* The structure of a DMA Channel.
 */
-typedef struct au1xxx_dma_channel {
+typedef volatile struct au1xxx_dma_channel {
 	u32	ddma_cfg;	/* See below */
 	u32	ddma_desptr;	/* 32-byte aligned pointer to descriptor */
 	u32	ddma_statptr;	/* word aligned pointer to status word */
@@ -96,7 +98,7 @@
 /* "Standard" DDMA Descriptor.
  * Must be 32-byte aligned.
  */
-typedef struct au1xxx_ddma_desc {
+typedef volatile struct au1xxx_ddma_desc {
 	u32	dscr_cmd0;		/* See below */
 	u32	dscr_cmd1;		/* See below */
 	u32	dscr_source0;		/* source phys address */
@@ -105,6 +107,12 @@
 	u32	dscr_dest1;		/* See below */
 	u32	dscr_stat;		/* completion status */
 	u32	dscr_nxtptr;		/* Next descriptor pointer (mostly) */
+	/* First 32bytes are HW specific!!!
+	   Lets have some SW data following.. make sure its 32bytes
+	 */
+	u32	sw_status;
+	u32 	sw_context;
+	u32	sw_reserved[6];
 } au1x_ddma_desc_t;
 
 #define DSCR_CMD0_V		(1 << 31)	/* Descriptor valid */
@@ -123,6 +131,8 @@
 #define DSCR_CMD0_CV		(0x1 << 2)	/* Clear Valid when done */
 #define DSCR_CMD0_ST_MASK	(0x3 << 0)	/* Status instruction */
 
+#define SW_STATUS_INUSE		(1<<0)
+
 /* Command 0 device IDs.
 */
 #ifdef CONFIG_SOC_AU1550
@@ -169,8 +179,8 @@
 #define DSCR_CMD0_SDMS_RX0	9
 #define DSCR_CMD0_SDMS_TX1	10
 #define DSCR_CMD0_SDMS_RX1	11
-#define DSCR_CMD0_AES_TX	12
-#define DSCR_CMD0_AES_RX	13
+#define DSCR_CMD0_AES_TX	13
+#define DSCR_CMD0_AES_RX	12
 #define DSCR_CMD0_PSC0_TX	14
 #define DSCR_CMD0_PSC0_RX	15
 #define DSCR_CMD0_PSC1_TX	16
@@ -189,6 +199,10 @@
 #define DSCR_CMD0_THROTTLE	30
 #define DSCR_CMD0_ALWAYS	31
 #define DSCR_NDEV_IDS		32
+/* THis macro is used to find/create custom device types */
+#define DSCR_DEV2CUSTOM_ID(x,d)	(((((x)&0xFFFF)<<8)|0x32000000)|((d)&0xFF))
+#define DSCR_CUSTOM2DEV_ID(x)	((x)&0xFF)
+
 
 #define DSCR_CMD0_SID(x)	(((x) & 0x1f) << 25)
 #define DSCR_CMD0_DID(x)	(((x) & 0x1f) << 20)
@@ -277,6 +291,43 @@
 */
 #define NUM_DBDMA_CHANS	16
 
+/*
+ * Ddma API definitions
+ * FIXME: may not fit to this header file
+ */
+typedef struct dbdma_device_table {
+	u32		dev_id;
+	u32		dev_flags;
+	u32		dev_tsize;
+	u32		dev_devwidth;
+	u32		dev_physaddr;		/* If FIFO */
+	u32		dev_intlevel;
+	u32		dev_intpolarity;
+} dbdev_tab_t;
+
+
+typedef struct dbdma_chan_config {
+	spinlock_t      lock;
+
+	u32			chan_flags;
+	u32			chan_index;
+	dbdev_tab_t		*chan_src;
+	dbdev_tab_t		*chan_dest;
+	au1x_dma_chan_t		*chan_ptr;
+	au1x_ddma_desc_t	*chan_desc_base;
+	au1x_ddma_desc_t	*get_ptr, *put_ptr, *cur_ptr;
+	void			*chan_callparam;
+	void (*chan_callback)(int, void *, struct pt_regs *);
+} chan_tab_t;
+
+#define DEV_FLAGS_INUSE		(1 << 0)
+#define DEV_FLAGS_ANYUSE	(1 << 1)
+#define DEV_FLAGS_OUT		(1 << 2)
+#define DEV_FLAGS_IN		(1 << 3)
+#define DEV_FLAGS_BURSTABLE (1 << 4)
+#define DEV_FLAGS_SYNC		(1 << 5)
+/* end Ddma API definitions */
+
 /* External functions for drivers to use.
 */
 /* Use this to allocate a dbdma channel.  The device ids are one of the
@@ -299,8 +350,8 @@
 
 /* Put buffers on source/destination descriptors.
 */
-u32 au1xxx_dbdma_put_source(u32 chanid, void *buf, int nbytes);
-u32 au1xxx_dbdma_put_dest(u32 chanid, void *buf, int nbytes);
+u32 _au1xxx_dbdma_put_source(u32 chanid, void *buf, int nbytes, u32 flags);
+u32 _au1xxx_dbdma_put_dest(u32 chanid, void *buf, int nbytes, u32 flags);
 
 /* Get a buffer from the destination descriptor.
 */
@@ -314,5 +365,29 @@
 void au1xxx_dbdma_chan_free(u32 chanid);
 void au1xxx_dbdma_dump(u32 chanid);
 
+u32 au1xxx_dbdma_put_dscr(u32 chanid, au1x_ddma_desc_t *dscr );
+
+u32 au1xxx_ddma_add_device( dbdev_tab_t *dev );
+void * au1xxx_ddma_get_nextptr_virt(au1x_ddma_desc_t *dp);
+
+/*
+ 	Some compatibilty macros --
+		Needed to make changes to API without breaking existing drivers
+*/
+#define	au1xxx_dbdma_put_source(chanid,buf,nbytes)_au1xxx_dbdma_put_source(chanid, buf, nbytes, DDMA_FLAGS_IE)
+#define	au1xxx_dbdma_put_source_flags(chanid,buf,nbytes,flags) _au1xxx_dbdma_put_source(chanid, buf, nbytes, flags)
+#define	put_source_flags(chanid,buf,nbytes,flags) au1xxx_dbdma_put_source_flags(chanid,buf,nbytes,flags)
+
+
+#define au1xxx_dbdma_put_dest(chanid,buf,nbytes) _au1xxx_dbdma_put_dest(chanid, buf, nbytes, DDMA_FLAGS_IE)
+#define	au1xxx_dbdma_put_dest_flags(chanid,buf,nbytes,flags) _au1xxx_dbdma_put_dest(chanid, buf, nbytes, flags)
+#define	put_dest_flags(chanid,buf,nbytes,flags) au1xxx_dbdma_put_dest_flags(chanid,buf,nbytes,flags)
+
+/*
+ *	Flags for the put_source/put_dest functions.
+ */
+#define DDMA_FLAGS_IE	(1<<0)
+#define DDMA_FLAGS_NOIE (1<<1)
+
 #endif /* _LANGUAGE_ASSEMBLY */
 #endif /* _AU1000_DBDMA_H_ */
--- linux-old/arch/mips/au1000/common/dbdma.c	2006-05-09 18:23:18.000000000 +0200
+++ linux/arch/mips/au1000/common/dbdma.c	2006-05-09 14:04:44.000000000 +0200
@@ -30,6 +30,7 @@
  *
  */
 
+#include <linux/config.h>
 #include <linux/kernel.h>
 #include <linux/errno.h>
 #include <linux/sched.h>
@@ -37,13 +38,14 @@
 #include <linux/spinlock.h>
 #include <linux/string.h>
 #include <linux/delay.h>
+#include <linux/interrupt.h>
+#include <linux/module.h>
 #include <asm/au1000.h>
 #include <asm/au1xxx_dbdma.h>
 #include <asm/system.h>
 
-#include <linux/module.h>
 
-#if 1 // defined(CONFIG_SOC_AU1550) || defined(CONFIG_SOC_AU1200)
+#if defined(CONFIG_SOC_AU1550) || defined(CONFIG_SOC_AU1200)
 
 /*
  * The Descriptor Based DMA supports up to 16 channels.
@@ -62,37 +64,10 @@
 */
 #define ALIGN_ADDR(x, a)	((((u32)(x)) + (a-1)) & ~(a-1))
 
-static volatile dbdma_global_t *dbdma_gptr = (dbdma_global_t *)DDMA_GLOBAL_BASE;
-static int dbdma_initialized;
+static dbdma_global_t *dbdma_gptr = (dbdma_global_t *)DDMA_GLOBAL_BASE;
+static int dbdma_initialized=0;
 static void au1xxx_dbdma_init(void);
 
-typedef struct dbdma_device_table {
-	u32		dev_id;
-	u32		dev_flags;
-	u32		dev_tsize;
-	u32		dev_devwidth;
-	u32		dev_physaddr;		/* If FIFO */
-	u32		dev_intlevel;
-	u32		dev_intpolarity;
-} dbdev_tab_t;
-
-typedef struct dbdma_chan_config {
-	u32			chan_flags;
-	u32			chan_index;
-	dbdev_tab_t		*chan_src;
-	dbdev_tab_t		*chan_dest;
-	au1x_dma_chan_t		*chan_ptr;
-	au1x_ddma_desc_t	*chan_desc_base;
-	au1x_ddma_desc_t	*get_ptr, *put_ptr, *cur_ptr;
-	void			*chan_callparam;
-	void (*chan_callback)(int, void *, struct pt_regs *);
-} chan_tab_t;
-
-#define	DEV_FLAGS_INUSE		(1 << 0)
-#define	DEV_FLAGS_ANYUSE	(1 << 1)
-#define DEV_FLAGS_OUT		(1 << 2)
-#define DEV_FLAGS_IN		(1 << 3)
-
 static dbdev_tab_t dbdev_tab[] = {
 #ifdef CONFIG_SOC_AU1550
 	/* UARTS */
@@ -158,25 +133,25 @@
 	{ DSCR_CMD0_MAE_BOTH, DEV_FLAGS_ANYUSE, 0, 0, 0x00000000, 0, 0 },
 	{ DSCR_CMD0_LCD, DEV_FLAGS_ANYUSE, 0, 0, 0x00000000, 0, 0 },
 
-	{ DSCR_CMD0_SDMS_TX0, DEV_FLAGS_OUT, 0, 0, 0x00000000, 0, 0 },
-	{ DSCR_CMD0_SDMS_RX0, DEV_FLAGS_IN, 0, 0, 0x00000000, 0, 0 },
-	{ DSCR_CMD0_SDMS_TX1, DEV_FLAGS_OUT, 0, 0, 0x00000000, 0, 0 },
-	{ DSCR_CMD0_SDMS_RX1, DEV_FLAGS_IN, 0, 0, 0x00000000, 0, 0 },
+	{ DSCR_CMD0_SDMS_TX0, DEV_FLAGS_OUT, 4, 8, 0x10600000, 0, 0 },
+	{ DSCR_CMD0_SDMS_RX0, DEV_FLAGS_IN, 4, 8, 0x10600004, 0, 0 },
+	{ DSCR_CMD0_SDMS_TX1, DEV_FLAGS_OUT, 4, 8, 0x10680000, 0, 0 },
+	{ DSCR_CMD0_SDMS_RX1, DEV_FLAGS_IN, 4, 8, 0x10680004, 0, 0 },
 
-	{ DSCR_CMD0_AES_TX, DEV_FLAGS_OUT, 0, 0, 0x00000000, 0, 0 },
-	{ DSCR_CMD0_AES_RX, DEV_FLAGS_IN, 0, 0, 0x00000000, 0, 0 },
+	{ DSCR_CMD0_AES_RX, DEV_FLAGS_IN , 4, 32, 0x10300008, 0, 0 },
+	{ DSCR_CMD0_AES_TX, DEV_FLAGS_OUT, 4, 32, 0x10300004, 0, 0 },
 
-	{ DSCR_CMD0_PSC0_TX, DEV_FLAGS_OUT, 0, 0, 0x11a0001c, 0, 0 },
-	{ DSCR_CMD0_PSC0_RX, DEV_FLAGS_IN, 0, 0, 0x11a0001c, 0, 0 },
+	{ DSCR_CMD0_PSC0_TX, DEV_FLAGS_OUT, 0, 16, 0x11a0001c, 0, 0 },
+	{ DSCR_CMD0_PSC0_RX, DEV_FLAGS_IN, 0, 16, 0x11a0001c, 0, 0 },
 	{ DSCR_CMD0_PSC0_SYNC, DEV_FLAGS_ANYUSE, 0, 0, 0x00000000, 0, 0 },
 
-	{ DSCR_CMD0_PSC1_TX, DEV_FLAGS_OUT, 0, 0, 0x11b0001c, 0, 0 },
-	{ DSCR_CMD0_PSC1_RX, DEV_FLAGS_IN, 0, 0, 0x11b0001c, 0, 0 },
+	{ DSCR_CMD0_PSC1_TX, DEV_FLAGS_OUT, 0, 16, 0x11b0001c, 0, 0 },
+	{ DSCR_CMD0_PSC1_RX, DEV_FLAGS_IN, 0, 16, 0x11b0001c, 0, 0 },
 	{ DSCR_CMD0_PSC1_SYNC, DEV_FLAGS_ANYUSE, 0, 0, 0x00000000, 0, 0 },
 
-	{ DSCR_CMD0_CIM_RXA, DEV_FLAGS_IN, 0, 0, 0x00000000, 0, 0 },
-	{ DSCR_CMD0_CIM_RXB, DEV_FLAGS_IN, 0, 0, 0x00000000, 0, 0 },
-	{ DSCR_CMD0_CIM_RXC, DEV_FLAGS_IN, 0, 0, 0x00000000, 0, 0 },
+	{ DSCR_CMD0_CIM_RXA, DEV_FLAGS_IN, 0, 32, 0x14004020, 0, 0 },
+	{ DSCR_CMD0_CIM_RXB, DEV_FLAGS_IN, 0, 32, 0x14004040, 0, 0 },
+	{ DSCR_CMD0_CIM_RXC, DEV_FLAGS_IN, 0, 32, 0x14004060, 0, 0 },
 	{ DSCR_CMD0_CIM_SYNC, DEV_FLAGS_ANYUSE, 0, 0, 0x00000000, 0, 0 },
 
 	{ DSCR_CMD0_NAND_FLASH, DEV_FLAGS_IN, 0, 0, 0x00000000, 0, 0 },
@@ -185,6 +160,24 @@
 
 	{ DSCR_CMD0_THROTTLE, DEV_FLAGS_ANYUSE, 0, 0, 0x00000000, 0, 0 },
 	{ DSCR_CMD0_ALWAYS, DEV_FLAGS_ANYUSE, 0, 0, 0x00000000, 0, 0 },
+
+	/* Provide 16 user definable device types */
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
+	{ 0, 0, 0, 0, 0, 0, 0 },
 };
 
 #define DBDEV_TAB_SIZE (sizeof(dbdev_tab) / sizeof(dbdev_tab_t))
@@ -204,6 +197,36 @@
 	return NULL;
 }
 
+void * au1xxx_ddma_get_nextptr_virt(au1x_ddma_desc_t *dp)
+{
+        return phys_to_virt(DSCR_GET_NXTPTR(dp->dscr_nxtptr));
+}
+EXPORT_SYMBOL(au1xxx_ddma_get_nextptr_virt);
+
+u32
+au1xxx_ddma_add_device(dbdev_tab_t *dev)
+{
+	u32 ret = 0;
+	dbdev_tab_t *p=NULL;
+	static u16 new_id=0x1000;
+
+	p = find_dbdev_id(0);
+	if ( NULL != p )
+	{
+		memcpy(p, dev, sizeof(dbdev_tab_t));
+		p->dev_id = DSCR_DEV2CUSTOM_ID(new_id,dev->dev_id);
+		ret = p->dev_id;
+		new_id++;
+#if 0
+		printk("add_device: id:%x flags:%x padd:%x\n",
+				p->dev_id, p->dev_flags, p->dev_physaddr );
+#endif
+	}
+
+	return ret;
+}
+EXPORT_SYMBOL(au1xxx_ddma_add_device);
+
 /* Allocate a channel and return a non-zero descriptor if successful.
 */
 u32
@@ -216,7 +239,7 @@
 	int		i;
 	dbdev_tab_t	*stp, *dtp;
 	chan_tab_t	*ctp;
-	volatile au1x_dma_chan_t *cp;
+	au1x_dma_chan_t *cp;
 
 	/* We do the intialization on the first channel allocation.
 	 * We have to wait because of the interrupt handler initialization
@@ -226,9 +249,6 @@
 		au1xxx_dbdma_init();
 	dbdma_initialized = 1;
 
-	if ((srcid > DSCR_NDEV_IDS) || (destid > DSCR_NDEV_IDS))
-		return 0;
-	
 	if ((stp = find_dbdev_id(srcid)) == NULL) return 0;
 	if ((dtp = find_dbdev_id(destid)) == NULL) return 0;
 
@@ -240,7 +260,7 @@
 	spin_lock_irqsave(&au1xxx_dbdma_spin_lock, flags);
 	if (!(stp->dev_flags & DEV_FLAGS_INUSE) ||
 	     (stp->dev_flags & DEV_FLAGS_ANYUSE)) {
-	     	/* Got source */
+		/* Got source */
 		stp->dev_flags |= DEV_FLAGS_INUSE;
 		if (!(dtp->dev_flags & DEV_FLAGS_INUSE) ||
 		     (dtp->dev_flags & DEV_FLAGS_ANYUSE)) {
@@ -270,9 +290,8 @@
 				/* If kmalloc fails, it is caught below same
 				 * as a channel not available.
 				 */
-				ctp = (chan_tab_t *)kmalloc(sizeof(chan_tab_t), GFP_KERNEL);
+				ctp = kmalloc(sizeof(chan_tab_t), GFP_KERNEL);
 				chan_tab_ptr[i] = ctp;
-				ctp->chan_index = chan = i;
 				break;
 			}
 		}
@@ -280,10 +299,11 @@
 
 		if (ctp != NULL) {
 			memset(ctp, 0, sizeof(chan_tab_t));
+			ctp->chan_index = chan = i;
 			dcp = DDMA_CHANNEL_BASE;
 			dcp += (0x0100 * chan);
 			ctp->chan_ptr = (au1x_dma_chan_t *)dcp;
-			cp = (volatile au1x_dma_chan_t *)dcp;
+			cp = (au1x_dma_chan_t *)dcp;
 			ctp->chan_src = stp;
 			ctp->chan_dest = dtp;
 			ctp->chan_callback = callback;
@@ -300,6 +320,9 @@
 				i |= DDMA_CFG_DED;
 			if (dtp->dev_intpolarity)
 				i |= DDMA_CFG_DP;
+			if ((stp->dev_flags & DEV_FLAGS_SYNC) ||
+				(dtp->dev_flags & DEV_FLAGS_SYNC))
+					i |= DDMA_CFG_SYNC;
 			cp->ddma_cfg = i;
 			au_sync();
 
@@ -310,14 +333,14 @@
 			rv = (u32)(&chan_tab_ptr[chan]);
 		}
 		else {
-			/* Release devices.
-			*/
+			/* Release devices */
 			stp->dev_flags &= ~DEV_FLAGS_INUSE;
 			dtp->dev_flags &= ~DEV_FLAGS_INUSE;
 		}
 	}
 	return rv;
 }
+EXPORT_SYMBOL(au1xxx_dbdma_chan_alloc);
 
 /* Set the device width if source or destination is a FIFO.
  * Should be 8, 16, or 32 bits.
@@ -345,6 +368,7 @@
 
 	return rv;
 }
+EXPORT_SYMBOL(au1xxx_dbdma_set_devwidth);
 
 /* Allocate a descriptor ring, initializing as much as possible.
 */
@@ -371,10 +395,11 @@
 	 * and if we try that first we are likely to not waste larger
 	 * slabs of memory.
 	 */
-	desc_base = (u32)kmalloc(entries * sizeof(au1x_ddma_desc_t), GFP_KERNEL);
+	desc_base = (u32)kmalloc(entries * sizeof(au1x_ddma_desc_t),
+			GFP_KERNEL|GFP_DMA);
 	if (desc_base == 0)
 		return 0;
-	
+
 	if (desc_base & 0x1f) {
 		/* Lost....do it again, allocate extra, and round
 		 * the address base.
@@ -382,7 +407,7 @@
 		kfree((const void *)desc_base);
 		i = entries * sizeof(au1x_ddma_desc_t);
 		i += (sizeof(au1x_ddma_desc_t) - 1);
-		if ((desc_base = (u32)kmalloc(i, GFP_KERNEL)) == 0)
+		if ((desc_base = (u32)kmalloc(i, GFP_KERNEL|GFP_DMA)) == 0)
 			return 0;
 
 		desc_base = ALIGN_ADDR(desc_base, sizeof(au1x_ddma_desc_t));
@@ -404,7 +429,13 @@
 	cmd0 |= DSCR_CMD0_SID(srcid);
 	cmd0 |= DSCR_CMD0_DID(destid);
 	cmd0 |= DSCR_CMD0_IE | DSCR_CMD0_CV;
-	cmd0 |= DSCR_CMD0_ST(DSCR_CMD0_ST_CURRENT);
+	cmd0 |= DSCR_CMD0_ST(DSCR_CMD0_ST_NOCHANGE);
+
+        /* is it mem to mem transfer? */
+        if(((DSCR_CUSTOM2DEV_ID(srcid) == DSCR_CMD0_THROTTLE) || (DSCR_CUSTOM2DEV_ID(srcid) == DSCR_CMD0_ALWAYS)) &&
+           ((DSCR_CUSTOM2DEV_ID(destid) == DSCR_CMD0_THROTTLE) || (DSCR_CUSTOM2DEV_ID(destid) == DSCR_CMD0_ALWAYS))) {
+               cmd0 |= DSCR_CMD0_MEM;
+        }
 
 	switch (stp->dev_devwidth) {
 	case 8:
@@ -462,9 +493,14 @@
 	/* If source input is fifo, set static address.
 	*/
 	if (stp->dev_flags & DEV_FLAGS_IN) {
-		src0 = stp->dev_physaddr;
+		if ( stp->dev_flags & DEV_FLAGS_BURSTABLE )
+			src1 |= DSCR_SRC1_SAM(DSCR_xAM_BURST);
+		else
 		src1 |= DSCR_SRC1_SAM(DSCR_xAM_STATIC);
+
 	}
+	if (stp->dev_physaddr)
+		src0 = stp->dev_physaddr;
 
 	/* Set up dest1.  For now, assume no stride and increment.
 	 * A channel attribute update can change this later.
@@ -488,10 +524,18 @@
 	/* If destination output is fifo, set static address.
 	*/
 	if (dtp->dev_flags & DEV_FLAGS_OUT) {
-		dest0 = dtp->dev_physaddr;
+		if ( dtp->dev_flags & DEV_FLAGS_BURSTABLE )
+	                dest1 |= DSCR_DEST1_DAM(DSCR_xAM_BURST);
+				else
 		dest1 |= DSCR_DEST1_DAM(DSCR_xAM_STATIC);
 	}
-	
+	if (dtp->dev_physaddr)
+		dest0 = dtp->dev_physaddr;
+
+#if 0
+		printk("did:%x sid:%x cmd0:%x cmd1:%x source0:%x source1:%x dest0:%x dest1:%x\n",
+			dtp->dev_id, stp->dev_id, cmd0, cmd1, src0, src1, dest0, dest1 );
+#endif
 	for (i=0; i<entries; i++) {
 		dp->dscr_cmd0 = cmd0;
 		dp->dscr_cmd1 = cmd1;
@@ -500,10 +544,12 @@
 		dp->dscr_dest0 = dest0;
 		dp->dscr_dest1 = dest1;
 		dp->dscr_stat = 0;
+		dp->sw_context = 0;
+		dp->sw_status = 0;
 		dp->dscr_nxtptr = DSCR_NXTPTR(virt_to_phys(dp + 1));
 		dp++;
 	}
-	
+
 	/* Make last descrptor point to the first.
 	*/
 	dp--;
@@ -512,13 +558,14 @@
 
 	return (u32)(ctp->chan_desc_base);
 }
+EXPORT_SYMBOL(au1xxx_dbdma_ring_alloc);
 
 /* Put a source buffer into the DMA ring.
  * This updates the source pointer and byte count.  Normally used
  * for memory to fifo transfers.
  */
 u32
-au1xxx_dbdma_put_source(u32 chanid, void *buf, int nbytes)
+_au1xxx_dbdma_put_source(u32 chanid, void *buf, int nbytes, u32 flags)
 {
 	chan_tab_t		*ctp;
 	au1x_ddma_desc_t	*dp;
@@ -540,14 +587,30 @@
 	if (dp->dscr_cmd0 & DSCR_CMD0_V) {
 		return 0;
 	}
-	
+
 	/* Load up buffer address and byte count.
 	*/
 	dp->dscr_source0 = virt_to_phys(buf);
 	dp->dscr_cmd1 = nbytes;
-	dp->dscr_cmd0 |= DSCR_CMD0_V;	/* Let it rip */
-	ctp->chan_ptr->ddma_dbell = 0xffffffff;	/* Make it go */
-	
+	/* Check flags  */
+	if (flags & DDMA_FLAGS_IE)
+		dp->dscr_cmd0 |= DSCR_CMD0_IE;
+	if (flags & DDMA_FLAGS_NOIE)
+		dp->dscr_cmd0 &= ~DSCR_CMD0_IE;
+
+	/*
+	 * There is an errata on the Au1200/Au1550 parts that could result
+	 * in "stale" data being DMA'd. It has to do with the snoop logic on
+	 * the dache eviction buffer.  NONCOHERENT_IO is on by default for
+	 * these parts. If it is fixedin the future, these dma_cache_inv will
+	 * just be nothing more than empty macros. See io.h.
+	 * */
+	dma_cache_wback_inv((unsigned long)buf, nbytes);
+        dp->dscr_cmd0 |= DSCR_CMD0_V;        /* Let it rip */
+	au_sync();
+	dma_cache_wback_inv((unsigned long)dp, sizeof(dp));
+        ctp->chan_ptr->ddma_dbell = 0;
+
 	/* Get next descriptor pointer.
 	*/
 	ctp->put_ptr = phys_to_virt(DSCR_GET_NXTPTR(dp->dscr_nxtptr));
@@ -556,13 +619,14 @@
 	*/
 	return nbytes;
 }
+EXPORT_SYMBOL(_au1xxx_dbdma_put_source);
 
 /* Put a destination buffer into the DMA ring.
  * This updates the destination pointer and byte count.  Normally used
  * to place an empty buffer into the ring for fifo to memory transfers.
  */
 u32
-au1xxx_dbdma_put_dest(u32 chanid, void *buf, int nbytes)
+_au1xxx_dbdma_put_dest(u32 chanid, void *buf, int nbytes, u32 flags)
 {
 	chan_tab_t		*ctp;
 	au1x_ddma_desc_t	*dp;
@@ -581,15 +645,38 @@
 	/* If the descriptor is valid, we are way ahead of the DMA
 	 * engine, so just return an error condition.
 	 */
-	if (dp->dscr_cmd0 & DSCR_CMD0_V)
+	if (dp->dscr_cmd0 & DSCR_CMD0_V) {
 		return 0;
-	
-	/* Load up buffer address and byte count.
-	*/
+	}
+
+	/* Load up buffer address and byte count */
+
+	/* Check flags  */
+	if (flags & DDMA_FLAGS_IE)
+		dp->dscr_cmd0 |= DSCR_CMD0_IE;
+	if (flags & DDMA_FLAGS_NOIE)
+		dp->dscr_cmd0 &= ~DSCR_CMD0_IE;
+
 	dp->dscr_dest0 = virt_to_phys(buf);
 	dp->dscr_cmd1 = nbytes;
+#if 0
+	printk("cmd0:%x cmd1:%x source0:%x source1:%x dest0:%x dest1:%x\n",
+			dp->dscr_cmd0, dp->dscr_cmd1, dp->dscr_source0,
+			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'd. It has to do with the snoop logic on the
+	 * dache eviction buffer. NONCOHERENT_IO is on by default for these
+	 * parts. If it is fixedin the future, these dma_cache_inv will just
+	 * be nothing more than empty macros. See io.h.
+	 * */
+	dma_cache_inv((unsigned long)buf,nbytes);
 	dp->dscr_cmd0 |= DSCR_CMD0_V;	/* Let it rip */
-	
+	au_sync();
+	dma_cache_wback_inv((unsigned long)dp, sizeof(dp));
+        ctp->chan_ptr->ddma_dbell = 0;
+
 	/* Get next descriptor pointer.
 	*/
 	ctp->put_ptr = phys_to_virt(DSCR_GET_NXTPTR(dp->dscr_nxtptr));
@@ -598,6 +685,7 @@
 	*/
 	return nbytes;
 }
+EXPORT_SYMBOL(_au1xxx_dbdma_put_dest);
 
 /* Get a destination buffer into the DMA ring.
  * Normally used to get a full buffer from the ring during fifo
@@ -625,29 +713,31 @@
 	/* If the descriptor is valid, we are way ahead of the DMA
 	 * engine, so just return an error condition.
 	 */
-	if (dp->dscr_cmd0 & DSCR_CMD0_V)
+	if (dp->dscr_cmd0 & DSCR_CMD0_V) {
 		return 0;
-	
+	}
+
 	/* Return buffer address and byte count.
 	*/
 	*buf = (void *)(phys_to_virt(dp->dscr_dest0));
 	*nbytes = dp->dscr_cmd1;
 	rv = dp->dscr_stat;
-	
+
 	/* Get next descriptor pointer.
 	*/
 	ctp->get_ptr = phys_to_virt(DSCR_GET_NXTPTR(dp->dscr_nxtptr));
 
 	/* return something not zero.
 	*/
-	return rv;
+	return *nbytes;
 }
+EXPORT_SYMBOL(au1xxx_dbdma_get_dest);
 
 void
 au1xxx_dbdma_stop(u32 chanid)
 {
 	chan_tab_t	*ctp;
-	volatile au1x_dma_chan_t *cp;
+	au1x_dma_chan_t *cp;
 	int halt_timeout = 0;
 
 	ctp = *((chan_tab_t **)chanid);
@@ -667,6 +757,7 @@
 	cp->ddma_stat |= (DDMA_STAT_DB | DDMA_STAT_V);
 	au_sync();
 }
+EXPORT_SYMBOL(au1xxx_dbdma_stop);
 
 /* Start using the current descriptor pointer.  If the dbdma encounters
  * a not valid descriptor, it will stop.  In this case, we can just
@@ -676,17 +767,17 @@
 au1xxx_dbdma_start(u32 chanid)
 {
 	chan_tab_t	*ctp;
-	volatile au1x_dma_chan_t *cp;
+	au1x_dma_chan_t *cp;
 
 	ctp = *((chan_tab_t **)chanid);
-
 	cp = ctp->chan_ptr;
 	cp->ddma_desptr = virt_to_phys(ctp->cur_ptr);
 	cp->ddma_cfg |= DDMA_CFG_EN;	/* Enable channel */
 	au_sync();
-	cp->ddma_dbell = 0xffffffff;	/* Make it go */
+	cp->ddma_dbell = 0;
 	au_sync();
 }
+EXPORT_SYMBOL(au1xxx_dbdma_start);
 
 void
 au1xxx_dbdma_reset(u32 chanid)
@@ -705,15 +796,21 @@
 
 	do {
 		dp->dscr_cmd0 &= ~DSCR_CMD0_V;
+		/* reset our SW status -- this is used to determine
+		 * if a descriptor is in use by upper level SW. Since
+		 * posting can reset 'V' bit.
+		 */
+		dp->sw_status = 0;
 		dp = phys_to_virt(DSCR_GET_NXTPTR(dp->dscr_nxtptr));
 	} while (dp != ctp->chan_desc_base);
 }
+EXPORT_SYMBOL(au1xxx_dbdma_reset);
 
 u32
 au1xxx_get_dma_residue(u32 chanid)
 {
 	chan_tab_t	*ctp;
-	volatile au1x_dma_chan_t *cp;
+	au1x_dma_chan_t *cp;
 	u32		rv;
 
 	ctp = *((chan_tab_t **)chanid);
@@ -726,6 +823,7 @@
 
 	return rv;
 }
+EXPORT_SYMBOL(au1xxx_get_dma_residue);
 
 void
 au1xxx_dbdma_chan_free(u32 chanid)
@@ -739,35 +837,35 @@
 
 	au1xxx_dbdma_stop(chanid);
 
-	if (ctp->chan_desc_base != NULL)
-		kfree(ctp->chan_desc_base);
-	
+	kfree((void *)ctp->chan_desc_base);
+
 	stp->dev_flags &= ~DEV_FLAGS_INUSE;
 	dtp->dev_flags &= ~DEV_FLAGS_INUSE;
 	chan_tab_ptr[ctp->chan_index] = NULL;
 
 	kfree(ctp);
 }
+EXPORT_SYMBOL(au1xxx_dbdma_chan_free);
 
-static void
+static irqreturn_t
 dbdma_interrupt(int irq, void *dev_id, struct pt_regs *regs)
 {
-	u32	intstat;
-	u32	chan_index;
+	u32 intstat;
+	u32 chan_index;
 	chan_tab_t		*ctp;
 	au1x_ddma_desc_t	*dp;
-	volatile au1x_dma_chan_t *cp;
+	au1x_dma_chan_t *cp;
 
 	intstat = dbdma_gptr->ddma_intstat;
 	au_sync();
 	chan_index = au_ffs(intstat) - 1;
 
 	ctp = chan_tab_ptr[chan_index];
 	cp = ctp->chan_ptr;
 	dp = ctp->cur_ptr;
 
 	/* Reset interrupt.
-	*/
+	 */
 	cp->ddma_irq = 0;
 	au_sync();
 
@@ -775,18 +875,28 @@
 		(ctp->chan_callback)(irq, ctp->chan_callparam, regs);
 
 	ctp->cur_ptr = phys_to_virt(DSCR_GET_NXTPTR(dp->dscr_nxtptr));
-	
+
+	return IRQ_RETVAL(1);
 }
 
-static void
-au1xxx_dbdma_init(void)
+static void au1xxx_dbdma_init(void)
 {
+	int irq_nr;
+
 	dbdma_gptr->ddma_config = 0;
 	dbdma_gptr->ddma_throttle = 0;
 	dbdma_gptr->ddma_inten = 0xffff;
 	au_sync();
 
-	if (request_irq(AU1550_DDMA_INT, dbdma_interrupt, SA_INTERRUPT,
+#if defined(CONFIG_SOC_AU1550)
+	irq_nr = AU1550_DDMA_INT;
+#elif defined(CONFIG_SOC_AU1200)
+	irq_nr = AU1200_DDMA_INT;
+#else
+	#error Unknown Au1x00 SOC
+#endif
+
+	if (request_irq(irq_nr, dbdma_interrupt, SA_INTERRUPT,
 			"Au1xxx dbdma", (void *)dbdma_gptr))
 		printk("Can't get 1550 dbdma irq");
 }
@@ -797,7 +907,8 @@
 	chan_tab_t		*ctp;
 	au1x_ddma_desc_t	*dp;
 	dbdev_tab_t		*stp, *dtp;
-	volatile au1x_dma_chan_t *cp;
+	au1x_dma_chan_t *cp;
+		u32			i = 0;
 
 	ctp = *((chan_tab_t **)chanid);
 	stp = ctp->chan_src;
@@ -809,7 +920,7 @@
 	printk("desc base %x, get %x, put %x, cur %x\n",
 		(u32)(ctp->chan_desc_base), (u32)(ctp->get_ptr),
 		(u32)(ctp->put_ptr), (u32)(ctp->cur_ptr));
-	
+
 	printk("dbdma chan %x\n", (u32)cp);
 	printk("cfg %08x, desptr %08x, statptr %08x\n",
 		cp->ddma_cfg, cp->ddma_desptr, cp->ddma_statptr);
@@ -822,28 +933,65 @@
 	dp = ctp->chan_desc_base;
 
 	do {
-		printk("dp %08x, cmd0 %08x, cmd1 %08x\n",
-			(u32)dp, dp->dscr_cmd0, dp->dscr_cmd1);
-		printk("src0 %08x, src1 %08x, dest0 %08x\n",
-			dp->dscr_source0, dp->dscr_source1, dp->dscr_dest0);
-		printk("dest1 %08x, stat %08x, nxtptr %08x\n",
-			dp->dscr_dest1, dp->dscr_stat, dp->dscr_nxtptr);
+                printk("Dp[%d]= %08x, cmd0 %08x, cmd1 %08x\n",
+                        i++, (u32)dp, dp->dscr_cmd0, dp->dscr_cmd1);
+                printk("src0 %08x, src1 %08x, dest0 %08x, dest1 %08x\n",
+                        dp->dscr_source0, dp->dscr_source1, dp->dscr_dest0, dp->dscr_dest1);
+                printk("stat %08x, nxtptr %08x\n",
+                        dp->dscr_stat, dp->dscr_nxtptr);
 		dp = phys_to_virt(DSCR_GET_NXTPTR(dp->dscr_nxtptr));
 	} while (dp != ctp->chan_desc_base);
 }
+EXPORT_SYMBOL(au1xxx_dbdma_dump);
 
+/* Put a descriptor into the DMA ring.
+ * This updates the source/destination pointers and byte count.
+ */
+u32
+au1xxx_dbdma_put_dscr(u32 chanid, au1x_ddma_desc_t *dscr )
+{
+	chan_tab_t *ctp;
+	au1x_ddma_desc_t *dp;
+	u32 nbytes=0;
 
-EXPORT_SYMBOL(au1xxx_dbdma_dump);
-EXPORT_SYMBOL(au1xxx_dbdma_put_source);
-EXPORT_SYMBOL(au1xxx_dbdma_put_dest);
-EXPORT_SYMBOL(au1xxx_dbdma_ring_alloc);
-EXPORT_SYMBOL(au1xxx_dbdma_start);
-EXPORT_SYMBOL(au1xxx_dbdma_get_dest);
-EXPORT_SYMBOL(au1xxx_dbdma_chan_alloc);
-EXPORT_SYMBOL(au1xxx_get_dma_residue);
-EXPORT_SYMBOL(au1xxx_dbdma_set_devwidth);
-EXPORT_SYMBOL(au1xxx_dbdma_chan_free);
-EXPORT_SYMBOL(au1xxx_dbdma_reset);
+	/* I guess we could check this to be within the
+	* range of the table......
+	*/
+	ctp = *((chan_tab_t **)chanid);
 
-#endif /* defined(CONFIG_SOC_AU1550) || defined(CONFIG_SOC_AU1200) */
+	/* We should have multiple callers for a particular channel,
+	* an interrupt doesn't affect this pointer nor the descriptor,
+	* so no locking should be needed.
+	*/
+	dp = ctp->put_ptr;
+
+	/* If the descriptor is valid, we are way ahead of the DMA
+	* engine, so just return an error condition.
+	*/
+	if (dp->dscr_cmd0 & DSCR_CMD0_V)
+		return 0;
 
+	/* Load up buffer addresses and byte count.
+	*/
+	dp->dscr_dest0 = dscr->dscr_dest0;
+	dp->dscr_source0 = dscr->dscr_source0;
+	dp->dscr_dest1 = dscr->dscr_dest1;
+	dp->dscr_source1 = dscr->dscr_source1;
+	dp->dscr_cmd1 = dscr->dscr_cmd1;
+	nbytes = dscr->dscr_cmd1;
+	/* Allow the caller to specifiy if an interrupt is generated */
+	dp->dscr_cmd0 &= ~DSCR_CMD0_IE;
+	dp->dscr_cmd0 |= dscr->dscr_cmd0 | DSCR_CMD0_V;
+	ctp->chan_ptr->ddma_dbell = 0;
+
+	/* Get next descriptor pointer.
+	*/
+	ctp->put_ptr = phys_to_virt(DSCR_GET_NXTPTR(dp->dscr_nxtptr));
+
+	/* return something not zero.
+	*/
+	return nbytes;
+}
+EXPORT_SYMBOL(au1xxx_dbdma_put_dscr);
+
+#endif /* defined(CONFIG_SOC_AU1550) || defined(CONFIG_SOC_AU1200) */
--- linux-old/drivers/sound/au1550_psc.c	2006-05-09 14:49:35.000000000 +0200
+++ linux/drivers/sound/au1550_psc.c	2006-05-09 18:19:01.000000000 +0200
@@ -82,6 +83,11 @@
 
 #define err(format, arg...) printk(KERN_ERR PFX ": " format "\n" , ## arg)
 #define info(format, arg...) printk(KERN_INFO PFX ": " format "\n" , ## arg)
+#ifdef AU1000_VERBOSE_DEBUG
+#define dbg(format, arg...) printk(KERN_DEBUG PFX ": " format "\n" , ## arg)
+#else
+#define dbg(format, arg...)
+#endif /* AU1000_VERBOSE_DEBUG */
 
 /* Boot options
  * 0 = no VRA, 1 = use VRA if codec supports it
@@ -127,7 +129,7 @@
 		unsigned        fragshift;
 		void		*nextIn;
 		void		*nextOut;
-		int		count;
+		volatile int	count;
 		unsigned        total_bytes;
 		unsigned        error;
 		wait_queue_head_t wait;
@@ -198,7 +200,7 @@
 	struct au1550_state *s = (struct au1550_state *)codec->private_data;
 	unsigned long   flags;
 	u32             cmd, val;
-	u16             data;
+	u16             data=0;
 	int             i;
 
 	spin_lock_irqsave(&s->lock, flags);
@@ -210,26 +212,21 @@
 			break;
 	}
 	if (i == POLL_COUNT)
-		err("rdcodec: codec cmd pending expired!");
+		err("rdcodec: codec cmd pending expired! %i %i", val, addr);
+
+	val = au_readl(PSC_AC97EVNT);
+	if (val & PSC_AC97EVNT_CD) {
+		err("rdcodec: command done is set! %i %i", val, addr);
+		au_readl(PSC_AC97CDC);  // manual says: you must read CDC[DATA] before resetting EVNT[CD], if command was read; and we don't know here what the last command was
+		au_sync();
+		au_writel(PSC_AC97EVNT_CD, PSC_AC97EVNT);
+	}
 
 	cmd = (u32)PSC_AC97CDC_INDX(addr);
 	cmd |= PSC_AC97CDC_RD;	/* read command */
 	au_writel(cmd, PSC_AC97CDC);
 	au_sync();
 
-	/* now wait for the data
-	*/
-	for (i = 0; i < POLL_COUNT; i++) {
-		val = au_readl(PSC_AC97STAT);
-		au_sync();
-		if (!(val & PSC_AC97STAT_CP))
-			break;
-	}
-	if (i == POLL_COUNT) {
-		err("rdcodec: read poll expired!");
-		return 0;
-	}
-
 	/* wait for command done?
 	*/
 	for (i = 0; i < POLL_COUNT; i++) {
@@ -239,8 +236,8 @@
 			break;
 	}
 	if (i == POLL_COUNT) {
-		err("rdcodec: read cmdwait expired!");
-		return 0;
+		err("rdcodec: read cmdwait expired! %i %i", val, addr);
+		goto rcodec_out;
 	}
 
 	data = au_readl(PSC_AC97CDC) & 0xffff;
@@ -251,8 +248,11 @@
 	au_writel(PSC_AC97EVNT_CD, PSC_AC97EVNT);
 	au_sync();
 
+  rcodec_out:
 	spin_unlock_irqrestore(&s->lock, flags);
 
+	dbg("rdcodec %x -> %x", addr, data);
+
 	return data;
 }
 
@@ -274,7 +274,15 @@
 			break;
 	}
 	if (i == POLL_COUNT)
-		err("wrcodec: codec cmd pending expired!");
+		err("wrcodec: codec cmd pending expired! %i %i", val, addr);
+
+	val = au_readl(PSC_AC97EVNT);
+	if (val & PSC_AC97EVNT_CD) {
+		err("wrcodec: command done is set! %i %i", val, addr);
+		au_readl(PSC_AC97CDC); // manual says: you must read CDC[DATA] before resetting EVNT[CD], if command was read; and we don't know here what the last command was
+		au_sync();
+		au_writel(PSC_AC97EVNT_CD, PSC_AC97EVNT);
+	}
 
 	cmd = (u32)PSC_AC97CDC_INDX(addr);
 	cmd |= (u32)data;
@@ -282,22 +290,13 @@
 	au_sync();
 
 	for (i = 0; i < POLL_COUNT; i++) {
-		val = au_readl(PSC_AC97STAT);
-		au_sync();
-		if (!(val & PSC_AC97STAT_CP))
-			break;
-	}
-	if (i == POLL_COUNT)
-		err("wrcodec: codec cmd pending expired!");
-
-	for (i = 0; i < POLL_COUNT; i++) {
 		val = au_readl(PSC_AC97EVNT);
 		au_sync();
 		if (val & PSC_AC97EVNT_CD)
 			break;
 	}
 	if (i == POLL_COUNT)
-		err("wrcodec: read cmdwait expired!");
+		err("wrcodec: read cmdwait expired! %i %i", val, addr);
 
 	/* Clear command done event.
 	*/
@@ -305,6 +304,8 @@
 	au_sync();
 
 	spin_unlock_irqrestore(&s->lock, flags);
+
+	dbg("wrcodec %x -> %x done", data, addr);
 }
 
 static void
@@ -392,7 +393,7 @@
 	adc_rate = rdcodec(s->codec, AC97_PCM_LR_ADC_RATE);
 
 #ifdef AU1000_VERBOSE_DEBUG
-	dbg(__FUNCTION__ ": set to %d Hz", adc_rate);
+	dbg("%s : set to %d Hz", __FUNCTION__, adc_rate);
 #endif
 
 	/* some codec's don't allow unequal DAC and ADC rates, in which case
@@ -452,7 +453,7 @@
 	dac_rate = rdcodec(s->codec, AC97_PCM_FRONT_DAC_RATE);
 
 #ifdef AU1000_VERBOSE_DEBUG
-	dbg(__FUNCTION__ ": set to %d Hz", dac_rate);
+	dbg("%s : set to %d Hz", __FUNCTION__, dac_rate);
 #endif
 
 	/* some codec's don't allow unequal DAC and ADC rates, in which case
@@ -489,6 +490,9 @@
 	au1xxx_dbdma_reset(db->dmanr);
 
 	db->stopped = 1;
+	db->count = 0;
+	db->dma_qcount = 0;
+	db->nextIn = db->nextOut = db->rawbuf;
 
 	spin_unlock_irqrestore(&s->lock, flags);
 }
@@ -518,6 +522,9 @@
 	au1xxx_dbdma_reset(db->dmanr);
 
 	db->stopped = 1;
+	db->count = 0;
+	db->dma_qcount = 0;
+	db->nextIn = db->nextOut = db->rawbuf;
 
 	spin_unlock_irqrestore(&s->lock, flags);
 }
@@ -609,7 +616,8 @@
 
 	spin_lock_irqsave(&s->lock, flags);
 
-	set_xmit_slots(db->num_channels);
+	au1xxx_dbdma_reset(db->dmanr);
+
 	au_writel(PSC_AC97PCR_TC, PSC_AC97PCR);
 	au_sync();
 	au_writel(PSC_AC97PCR_TS, PSC_AC97PCR);
@@ -634,17 +642,20 @@
 
 	spin_lock_irqsave(&s->lock, flags);
 
+	au1xxx_dbdma_reset(db->dmanr);
+
 	/* Put two buffers on the ring to get things started.
 	*/
-	for (i=0; i<2; i++) {
-		au1xxx_dbdma_put_dest(db->dmanr, db->nextIn, db->dma_fragsize);
+	for (i=0; i<NUM_DBDMA_DESCRIPTORS; i++) {
+		if ( au1xxx_dbdma_put_dest(db->dmanr, db->nextIn, db->dma_fragsize) ) {
 
-		db->nextIn += db->dma_fragsize;
-		if (db->nextIn >= db->rawbuf + db->dmasize)
-			db->nextIn -= db->dmasize;
+			db->nextIn += db->dma_fragsize;
+			if (db->nextIn >= db->rawbuf + db->dmasize)
+				db->nextIn -= db->dmasize;
+		} else
+			info("Cannot put dest %i", i);
 	}
 
-	set_recv_slots(db->num_channels);
 	au1xxx_dbdma_start(db->dmanr);
 	au_writel(PSC_AC97PCR_RC, PSC_AC97PCR);
 	au_sync();
@@ -665,9 +676,12 @@
 	if (!db->rawbuf) {
 		db->ready = db->mapped = 0;
 		db->buforder = 5;	/* 32 * PAGE_SIZE */
+		bufs = PAGE_SIZE << db->buforder;
 		db->rawbuf = kmalloc((PAGE_SIZE << db->buforder), GFP_KERNEL);
 		if (!db->rawbuf)
 			return -ENOMEM;
+	} else {
+		bufs = PAGE_SIZE << db->buforder;
 	}
 
 	db->cnt_factor = 1;
@@ -686,7 +700,6 @@
 					2 : db->num_channels);
 
 	user_bytes_per_sec = rate * db->user_bytes_per_sample;
-	bufs = PAGE_SIZE << db->buforder;
 	if (db->ossfragshift) {
 		if ((1000 << db->ossfragshift) < user_bytes_per_sec)
 			db->fragshift = ld2(user_bytes_per_sec/1000);
@@ -731,6 +744,7 @@
 static int
 prog_dmabuf_adc(struct au1550_state *s)
 {
+	dbg("prog_dmabuf_adc s%i c%i", s->dma_adc.sample_size, s->dma_adc.num_channels);
 	stop_adc(s);
 	return prog_dmabuf(s, &s->dma_adc);
 
@@ -739,6 +753,8 @@
 static int
 prog_dmabuf_dac(struct au1550_state *s)
 {
+	dbg("prog_dmabuf_dac s%i c%i", s->dma_dac.sample_size, s->dma_dac.num_channels);
+
 	stop_dac(s);
 	return prog_dmabuf(s, &s->dma_dac);
 }
@@ -752,20 +768,20 @@
 {
 	struct au1550_state *s = (struct au1550_state *) dev_id;
 	struct dmabuf  *db = &s->dma_dac;
+#ifdef AU1000_VERBOSE_DEBUG
 	u32	ac97c_stat;
+#endif
 
-	ac97c_stat = au_readl(PSC_AC97STAT);
 #ifdef AU1000_VERBOSE_DEBUG
-	if (ac97c_stat & (AC97C_XU | AC97C_XO | AC97C_TE))
-		dbg("AC97C status = 0x%08x", ac97c_stat);
+//	if (ac97c_stat & (AC97C_XU | AC97C_XO | AC97C_TE))
+//		dbg("AC97C status = 0x%08x", ac97c_stat);
 #endif
 	db->dma_qcount--;
 
-	if (db->count >= db->fragsize) {
-		if (au1xxx_dbdma_put_source(db->dmanr, db->nextOut,
-							db->fragsize) == 0) {
-			err("qcount < 2 and no ring room!");
-		}
+	// put source buffers as long as we have data and free dma descr
+	while ( (db->count >= db->fragsize) &&
+			(au1xxx_dbdma_put_source(db->dmanr, db->nextOut,
+									 db->fragsize) ) ) {
 		db->nextOut += db->fragsize;
 		if (db->nextOut >= db->rawbuf + db->dmasize)
 			db->nextOut -= db->dmasize;
@@ -785,30 +799,44 @@
 	struct	dmabuf  *dp = &s->dma_adc;
 	u32	obytes;
 	char	*obuf;
+#ifdef AU1000_VERBOSE_DEBUG
+	u32	ac97c_stat;
+#endif
 
-	/* Pull the buffer from the dma queue.
-	*/
-	au1xxx_dbdma_get_dest(dp->dmanr, (void *)(&obuf), &obytes);
-
-	if ((dp->count + obytes) > dp->dmasize) {
-		/* Overrun. Stop ADC and log the error
-		*/
-		stop_adc(s);
-		dp->error++;
-		err("adc overrun");
-		return;
-	}
+#ifdef AU1000_VERBOSE_DEBUG
+//	if (ac97c_stat & (AC97C_XU | AC97C_XO | AC97C_TE))
+//		dbg("AC97C status = 0x%08x", ac97c_stat);
+#endif
 
-	/* Put a new empty buffer on the destination DMA.
+	/* Pull completed buffer(s) from the dma queue.
 	*/
-	au1xxx_dbdma_put_dest(dp->dmanr, dp->nextIn, dp->dma_fragsize);
+	int cnt=0;
+	while ( au1xxx_dbdma_get_dest(dp->dmanr, (void *)(&obuf), &obytes) ) {
+		if ((dp->count + obytes) > dp->dmasize) {
+			/* Overrun. Stop ADC and log the error
+			 */
+			stop_adc(s);
+			dp->error++;
+			err("adc overrun");
+			break;
+		}
 
-	dp->nextIn += dp->dma_fragsize;
-	if (dp->nextIn >= dp->rawbuf + dp->dmasize)
-		dp->nextIn -= dp->dmasize;
+		dp->count += obytes;
+		dp->total_bytes += obytes;
 
-	dp->count += obytes;
-	dp->total_bytes += obytes;
+		/* Put new empty buffer(s) on the destination DMA.
+		 */
+		while ( au1xxx_dbdma_put_dest(dp->dmanr, dp->nextIn, dp->dma_fragsize) ) {
+			dp->nextIn += dp->dma_fragsize;
+			if (dp->nextIn >= dp->rawbuf + dp->dmasize)
+				dp->nextIn -= dp->dmasize;
+		}
+
+		if ( ++cnt>3 ) {
+			// get max 4 buffers at once here
+			break;
+		}
+	}
 
 	/* wake up anybody listening
 	*/
@@ -1078,13 +1104,10 @@
 		/* wait for samples in ADC dma buffer
 		*/
 		do {
-			if (db->stopped)
+		    if (db->stopped)
 				start_adc(s);
-			spin_lock_irqsave(&s->lock, flags);
+			set_current_state(TASK_INTERRUPTIBLE);
 			avail = db->count;
-			if (avail <= 0)
-				__set_current_state(TASK_INTERRUPTIBLE);
-			spin_unlock_irqrestore(&s->lock, flags);
 			if (avail <= 0) {
 				if (file->f_flags & O_NONBLOCK) {
 					if (!ret)
@@ -1163,11 +1189,8 @@
 		/* wait for space in playback buffer
 		*/
 		do {
-			spin_lock_irqsave(&s->lock, flags);
+			set_current_state(TASK_INTERRUPTIBLE);
 			avail = (int) db->dmasize - db->count;
-			if (avail <= 0)
-				__set_current_state(TASK_INTERRUPTIBLE);
-			spin_unlock_irqrestore(&s->lock, flags);
 			if (avail <= 0) {
 				if (file->f_flags & O_NONBLOCK) {
 					if (!ret)
@@ -1190,6 +1214,7 @@
 		if ((cnt = copy_dmabuf_user(db, (char *) buffer,
 					    count > avail ?
 					    avail : count, 0)) < 0) {
+			err("copy_dmabuf_user error %i", cnt);
 			if (!ret)
 				ret = -EFAULT;
 			goto out;
@@ -1213,6 +1228,7 @@
 				err("qcount < 2 and no ring room!");
 			}
 			db->nextOut += db->fragsize;
+			db->count -= db->fragsize;
 			if (db->nextOut >= db->rawbuf + db->dmasize)
 				db->nextOut -= db->dmasize;
 			db->total_bytes += db->dma_fragsize;
@@ -1244,13 +1269,20 @@
 	unsigned int    mask = 0;
 
 	if (file->f_mode & FMODE_WRITE) {
-		if (!s->dma_dac.ready)
+		if (!s->dma_dac.ready) {
+			err("poll: dma_dac not ready");
 			return 0;
+		}
 		poll_wait(file, &s->dma_dac.wait, wait);
 	}
 	if (file->f_mode & FMODE_READ) {
-		if (!s->dma_adc.ready)
+		if (!s->dma_adc.ready) {
+			err("poll: dma_adc not ready");
 			return 0;
+		}
+		if ( s->dma_adc.stopped ) {
+ 			start_adc(s);
+		}
 		poll_wait(file, &s->dma_adc.wait, wait);
 	}
 
@@ -1388,7 +1420,7 @@
 			break;
 	}
 	if (count < sizeof(ioctl_str) / sizeof(ioctl_str[0]))
-		dbg("ioctl %s, arg=0x%lx", ioctl_str[count].str, arg);
+		dbg("ioctl %s(%x), arg=0x%lx", ioctl_str[count].str, cmd, arg);
 	else
 		dbg("ioctl 0x%x unknown, arg=0x%lx", cmd, arg);
 #endif
@@ -1456,12 +1488,14 @@
 		if (file->f_mode & FMODE_READ) {
 			stop_adc(s);
 			s->dma_adc.num_channels = val ? 2 : 1;
+			set_recv_slots(s->dma_adc.num_channels);
 			if ((ret = prog_dmabuf_adc(s)))
 				return ret;
 		}
 		if (file->f_mode & FMODE_WRITE) {
 			stop_dac(s);
 			s->dma_dac.num_channels = val ? 2 : 1;
+			set_xmit_slots(s->dma_dac.num_channels);
 			if (s->codec_ext_caps & AC97_EXT_DACS) {
 				/* disable surround and center/lfe in AC'97
 				*/
@@ -1486,6 +1520,7 @@
 					return -EINVAL;
 				stop_adc(s);
 				s->dma_adc.num_channels = val;
+				set_recv_slots(s->dma_adc.num_channels);
 				if ((ret = prog_dmabuf_adc(s)))
 					return ret;
 			}
@@ -1543,6 +1578,7 @@
 				}
 
 				s->dma_dac.num_channels = val;
+				set_xmit_slots(s->dma_dac.num_channels);
 				if ((ret = prog_dmabuf_dac(s)))
 					return ret;
 			}
@@ -1799,6 +1835,7 @@
 		return -EINVAL;
 	}
 
+	info("mixdev_ioctl(%i)", cmd);
 	return mixdev_ioctl(s->codec, cmd, arg);
 }
 
@@ -1813,9 +1850,9 @@
 
 #ifdef AU1000_VERBOSE_DEBUG
 	if (file->f_flags & O_NONBLOCK)
-		dbg(__FUNCTION__ ": non-blocking");
+		info("%s : non-blocking", __FUNCTION__);
 	else
-		dbg(__FUNCTION__ ": blocking");
+		info("%s : blocking", __FUNCTION__);
 #endif
 
 	file->private_data = s;
@@ -1846,6 +1883,7 @@
 		s->dma_adc.num_channels = 1;
 		s->dma_adc.sample_size = 8;
 		set_adc_rate(s, 8000);
+		set_recv_slots(s->dma_adc.num_channels);
 		if ((minor & 0xf) == SND_DEV_DSP16)
 			s->dma_adc.sample_size = 16;
 	}
@@ -1856,22 +1894,31 @@
 		s->dma_dac.num_channels = 1;
 		s->dma_dac.sample_size = 8;
 		set_dac_rate(s, 8000);
+		set_xmit_slots(s->dma_dac.num_channels);
 		if ((minor & 0xf) == SND_DEV_DSP16)
 			s->dma_dac.sample_size = 16;
 	}
 
 	if (file->f_mode & FMODE_READ) {
-		if ((ret = prog_dmabuf_adc(s)))
+		if ((ret = prog_dmabuf_adc(s))) {
+			err("prog_dmabuf_adc failed");
+			up(&s->open_sem);
 			return ret;
+		}
+
 	}
 	if (file->f_mode & FMODE_WRITE) {
-		if ((ret = prog_dmabuf_dac(s)))
+		if ((ret = prog_dmabuf_dac(s))) {
+			err("prog_dmabuf_dac failed");
+			up(&s->open_sem);
 			return ret;
+		}
 	}
 
 	s->open_mode |= file->f_mode & (FMODE_READ | FMODE_WRITE);
 	up(&s->open_sem);
 	init_MUTEX(&s->sem);
+
 	return 0;
 }
 
@@ -1954,12 +2001,13 @@
 		err("AC'97 ports in use");
 	}
 
+
 	/* Allocate the DMA Channels
 	*/
 	if ((s->dma_dac.dmanr = au1xxx_dbdma_chan_alloc(DBDMA_MEM_CHAN,
 	    DBDMA_AC97_TX_CHAN, dac_dma_interrupt, (void *)s)) == 0) {
 		err("Can't get DAC DMA");
-		goto err_dma1;
+		goto err_dev1;
 	}
 	au1xxx_dbdma_set_devwidth(s->dma_dac.dmanr, 16);
 	if (au1xxx_dbdma_ring_alloc(s->dma_dac.dmanr,
@@ -1968,10 +2016,11 @@
 		goto err_dma1;
 	}
 
+
 	if ((s->dma_adc.dmanr = au1xxx_dbdma_chan_alloc(DBDMA_AC97_RX_CHAN,
-	    DBDMA_MEM_CHAN, adc_dma_interrupt, (void *)s)) == 0) {
+													DBDMA_MEM_CHAN, adc_dma_interrupt, (void *)s)) == 0) {
 		err("Can't get ADC DMA");
-		goto err_dma2;
+		goto err_dma1;
 	}
 	au1xxx_dbdma_set_devwidth(s->dma_adc.dmanr, 16);
 	if (au1xxx_dbdma_ring_alloc(s->dma_adc.dmanr,
@@ -1980,7 +2029,7 @@
 
 #ifdef AU1550_DEBUG
 	/* intialize the debug proc device */
-	s->ps = create_proc_read_entry(AU1000_MODULE_NAME, 0, NULL,
+	s->ps = create_proc_read_entry(AU1550_MODULE_NAME, 0, NULL,
 				       proc_au1550_dump, NULL);
 #endif /* AU1550_DEBUG */
 
@@ -2102,11 +2150,11 @@
 	unregister_sound_mixer(s->codec->dev_mixer);
  err_dev2:
 	unregister_sound_dsp(s->dev_audio);
- err_dev1:
-	au1xxx_dbdma_chan_free(s->dma_adc.dmanr);
  err_dma2:
-	au1xxx_dbdma_chan_free(s->dma_dac.dmanr);
+	au1xxx_dbdma_chan_free(s->dma_adc.dmanr);
  err_dma1:
+	au1xxx_dbdma_chan_free(s->dma_dac.dmanr);
+ err_dev1:
 	release_region(PHYSADDR(AC97_PSC_SEL), 0x30);
 
 	ac97_release_codec(s->codec);
@@ -2125,11 +2173,11 @@
 		remove_proc_entry(AU1000_MODULE_NAME, NULL);
 #endif /* AU1000_DEBUG */
 	synchronize_irq();
+	unregister_sound_dsp(s->dev_audio);
+	unregister_sound_mixer(s->codec->dev_mixer);
 	au1xxx_dbdma_chan_free(s->dma_adc.dmanr);
 	au1xxx_dbdma_chan_free(s->dma_dac.dmanr);
 	release_region(PHYSADDR(AC97_PSC_SEL), 0x30);
-	unregister_sound_dsp(s->dev_audio);
-	unregister_sound_mixer(s->codec->dev_mixer);
 	ac97_release_codec(s->codec);
 }