V4L/DVB (9720): cx18: Major rewrite of interrupt handling for incoming mailbox processing

A major rewrite of interrupt handling for incoming mailbox processing, to split
the timing critical steps from the the deferrable steps as the sending XPU on
the CX23418 will time out and overwrite our incoming mailboxes rather quickly.
Setup a pool of work "order forms" for the irq handler to send jobs to the new
work handler routine which uses the kernel default work queue to do the
deferrable work.  Started optimizing some of the cx18-io calls as they are now
the low hanging fruit for recoving microseconds back from the timeline.
Future optimizations will get rid of mmio read retries, mmio stats logging, and
combine smaller functions in the irq path into the larger ones to save ~2 us
each.

Signed-off-by: Andy Walls <awalls@radix.net>
Signed-off-by: Mauro Carvalho Chehab <mchehab@redhat.com>

16 files changed, 408 insertions, 215 deletions

diff --git a/drivers/media/video/cx18/cx18-driver.c b/drivers/media/video/cx18/cx18-driver.c
index 8ef11d578b8..434cc7fdee3 100644
--- a/drivers/media/video/cx18/cx18-driver.c
+++ b/drivers/media/video/cx18/cx18-driver.c
@@ -440,6 +440,8 @@ done:
  */
 static int __devinit cx18_init_struct1(struct cx18 *cx)
 {
+	int i;
+
 	cx->base_addr = pci_resource_start(cx->dev, 0);
 
 	mutex_init(&cx->serialize_lock);
@@ -451,7 +453,11 @@ static int __devinit cx18_init_struct1(struct cx18 *cx)
 
 	spin_lock_init(&cx->lock);
 
-	INIT_WORK(&cx->work, cx18_work_handler);
+	for (i = 0; i < CX18_MAX_EPU_WORK_ORDERS; i++) {
+		cx->epu_work_order[i].cx = cx;
+		cx->epu_work_order[i].str = cx->epu_debug_str;
+		INIT_WORK(&cx->epu_work_order[i].work, cx18_epu_work_handler);
+	}
 
 	/* start counting open_id at 1 */
 	cx->open_id = 1;
diff --git a/drivers/media/video/cx18/cx18-driver.h b/drivers/media/video/cx18/cx18-driver.h
index ce76806759e..5ebf84b78ca 100644
--- a/drivers/media/video/cx18/cx18-driver.h
+++ b/drivers/media/video/cx18/cx18-driver.h
@@ -203,8 +203,6 @@ struct cx18_options {
 #define CX18_F_I_EOS			4 	/* End of encoder stream */
 #define CX18_F_I_RADIO_USER		5 	/* radio tuner is selected */
 #define CX18_F_I_ENC_PAUSED		13 	/* the encoder is paused */
-#define CX18_F_I_HAVE_WORK		15   	/* there is work to be done */
-#define CX18_F_I_WORK_HANDLER_DVB	18   	/* work to be done for DVB */
 #define CX18_F_I_INITED			21 	/* set after first open */
 #define CX18_F_I_FAILED			22 	/* set if first open failed */
 
@@ -247,6 +245,19 @@ struct cx18_dvb {
 struct cx18;	 /* forward reference */
 struct cx18_scb; /* forward reference */
 
+#define CX18_MAX_MDL_ACKS 2
+#define CX18_MAX_EPU_WORK_ORDERS 70 /* CPU_DE_RELEASE_MDL bursts 63 commands */
+
+struct cx18_epu_work_order {
+	struct work_struct work;
+	atomic_t pending;
+	struct cx18 *cx;
+	int rpu;
+	struct cx18_mailbox mb;
+	struct cx18_mdl_ack mdl_ack[CX18_MAX_MDL_ACKS];
+	char *str;
+};
+
 #define CX18_INVALID_TASK_HANDLE 0xffffffff
 
 struct cx18_stream {
@@ -388,7 +399,6 @@ struct cx18 {
 	struct mutex epu2apu_mb_lock; /* protect driver to chip mailbox in SCB*/
 	struct mutex epu2cpu_mb_lock; /* protect driver to chip mailbox in SCB*/
 
-
 	struct cx18_av_state av_state;
 
 	/* codec settings */
@@ -441,7 +451,8 @@ struct cx18 {
 	/* when the current DMA is finished this queue is woken up */
 	wait_queue_head_t dma_waitq;
 
-	struct work_struct work;
+	struct cx18_epu_work_order epu_work_order[CX18_MAX_EPU_WORK_ORDERS];
+	char epu_debug_str[256]; /* CX18_EPU_DEBUG is rare: use shared space */
 
 	/* i2c */
 	struct i2c_adapter i2c_adap[2];
diff --git a/drivers/media/video/cx18/cx18-dvb.c b/drivers/media/video/cx18/cx18-dvb.c
index 4845f732ef4..b74253b1377 100644
--- a/drivers/media/video/cx18/cx18-dvb.c
+++ b/drivers/media/video/cx18/cx18-dvb.c
@@ -23,8 +23,6 @@
 #include "cx18-dvb.h"
 #include "cx18-io.h"
 #include "cx18-streams.h"
-#include "cx18-queue.h"
-#include "cx18-scb.h"
 #include "cx18-cards.h"
 #include "s5h1409.h"
 #include "mxl5005s.h"
@@ -305,26 +303,3 @@ static int dvb_register(struct cx18_stream *stream)
 
 	return ret;
 }
-
-void cx18_dvb_work_handler(struct cx18 *cx)
-{
-	struct cx18_buffer *buf;
-	struct cx18_stream *s = &cx->streams[CX18_ENC_STREAM_TYPE_TS];
-
-	while ((buf = cx18_dequeue(s, &s->q_full)) != NULL) {
-		if (s->dvb.enabled)
-			dvb_dmx_swfilter(&s->dvb.demux, buf->buf,
-					 buf->bytesused);
-
-		cx18_buf_sync_for_device(s, buf);
-		cx18_enqueue(s, buf, &s->q_free);
-
-		if (s->handle == CX18_INVALID_TASK_HANDLE ||
-		    !test_bit(CX18_F_S_STREAMING, &s->s_flags))
-			continue;
-
-		cx18_vapi(cx, CX18_CPU_DE_SET_MDL, 5, s->handle,
-		       (void __iomem *)&cx->scb->cpu_mdl[buf->id] - cx->enc_mem,
-		       1, buf->id, s->buf_size);
-	}
-}
diff --git a/drivers/media/video/cx18/cx18-dvb.h b/drivers/media/video/cx18/cx18-dvb.h
index bbdcefc87f2..bf8d8f6f545 100644
--- a/drivers/media/video/cx18/cx18-dvb.h
+++ b/drivers/media/video/cx18/cx18-dvb.h
@@ -23,4 +23,3 @@
 
 int cx18_dvb_register(struct cx18_stream *stream);
 void cx18_dvb_unregister(struct cx18_stream *stream);
-void cx18_dvb_work_handler(struct cx18 *cx);
diff --git a/drivers/media/video/cx18/cx18-firmware.c b/drivers/media/video/cx18/cx18-firmware.c
index d9c5f55ab17..fdea4b75204 100644
--- a/drivers/media/video/cx18/cx18-firmware.c
+++ b/drivers/media/video/cx18/cx18-firmware.c
@@ -121,6 +121,7 @@ static int load_cpu_fw_direct(const char *fn, u8 __iomem *mem, struct cx18 *cx)
 			if (cx18_raw_readl(cx, dst) != *src) {
 				CX18_ERR("Mismatch at offset %x\n", i);
 				release_firmware(fw);
+				cx18_setup_page(cx, 0);
 				return -EIO;
 			}
 			dst++;
@@ -131,6 +132,7 @@ static int load_cpu_fw_direct(const char *fn, u8 __iomem *mem, struct cx18 *cx)
 		CX18_INFO("loaded %s firmware (%zd bytes)\n", fn, fw->size);
 	size = fw->size;
 	release_firmware(fw);
+	cx18_setup_page(cx, SCB_OFFSET);
 	return size;
 }
 
@@ -150,6 +152,7 @@ static int load_apu_fw_direct(const char *fn, u8 __iomem *dst, struct cx18 *cx,
 	if (request_firmware(&fw, fn, &cx->dev->dev)) {
 		CX18_ERR("unable to open firmware %s\n", fn);
 		CX18_ERR("did you put the firmware in the hotplug firmware directory?\n");
+		cx18_setup_page(cx, 0);
 		return -ENOMEM;
 	}
 
@@ -185,6 +188,7 @@ static int load_apu_fw_direct(const char *fn, u8 __iomem *dst, struct cx18 *cx,
 					CX18_ERR("Mismatch at offset %x\n",
 						 offset + j);
 					release_firmware(fw);
+					cx18_setup_page(cx, 0);
 					return -EIO;
 				}
 			}
@@ -196,6 +200,7 @@ static int load_apu_fw_direct(const char *fn, u8 __iomem *dst, struct cx18 *cx,
 				fn, apu_version, fw->size);
 	size = fw->size;
 	release_firmware(fw);
+	cx18_setup_page(cx, 0);
 	return size;
 }
 
diff --git a/drivers/media/video/cx18/cx18-io.c b/drivers/media/video/cx18/cx18-io.c
index 3c6485fceea..c67694f63d0 100644
--- a/drivers/media/video/cx18/cx18-io.c
+++ b/drivers/media/video/cx18/cx18-io.c
@@ -166,41 +166,6 @@ u8 cx18_readb_retry(struct cx18 *cx, const void __iomem *addr)
 	return val;
 }
 
-void cx18_memcpy_fromio(struct cx18 *cx, void *to,
-			const void __iomem *from, unsigned int len)
-{
-	const u8 __iomem *src = from;
-	u8 *dst = to;
-
-	/* Align reads on the CX23418's addresses */
-	if ((len > 0) && ((unsigned long) src & 1)) {
-		*dst = cx18_readb(cx, src);
-		len--;
-		dst++;
-		src++;
-	}
-	if ((len > 1) && ((unsigned long) src & 2)) {
-		*((u16 *)dst) = cx18_raw_readw(cx, src);
-		len -= 2;
-		dst += 2;
-		src += 2;
-	}
-	while (len > 3) {
-		*((u32 *)dst) = cx18_raw_readl(cx, src);
-		len -= 4;
-		dst += 4;
-		src += 4;
-	}
-	if (len > 1) {
-		*((u16 *)dst) = cx18_raw_readw(cx, src);
-		len -= 2;
-		dst += 2;
-		src += 2;
-	}
-	if (len > 0)
-		*dst = cx18_readb(cx, src);
-}
-
 void cx18_memset_io(struct cx18 *cx, void __iomem *addr, int val, size_t count)
 {
 	u8 __iomem *dst = addr;
diff --git a/drivers/media/video/cx18/cx18-io.h b/drivers/media/video/cx18/cx18-io.h
index 4486b73faf5..fdc2bcc92fc 100644
--- a/drivers/media/video/cx18/cx18-io.h
+++ b/drivers/media/video/cx18/cx18-io.h
@@ -249,8 +249,13 @@ static inline u32 cx18_write_sync(struct cx18 *cx, u32 val, void __iomem *addr)
 }
 
 
+static inline
 void cx18_memcpy_fromio(struct cx18 *cx, void *to,
-			const void __iomem *from, unsigned int len);
+			const void __iomem *from, unsigned int len)
+{
+	memcpy_fromio(to, from, len);
+}
+
 void cx18_memset_io(struct cx18 *cx, void __iomem *addr, int val, size_t count);
 
 
diff --git a/drivers/media/video/cx18/cx18-irq.c b/drivers/media/video/cx18/cx18-irq.c
index 4912b3c8e6a..bc36a6b8f77 100644
--- a/drivers/media/video/cx18/cx18-irq.c
+++ b/drivers/media/video/cx18/cx18-irq.c
@@ -21,121 +21,9 @@
 
 #include "cx18-driver.h"
 #include "cx18-io.h"
-#include "cx18-firmware.h"
-#include "cx18-fileops.h"
-#include "cx18-queue.h"
 #include "cx18-irq.h"
-#include "cx18-ioctl.h"
 #include "cx18-mailbox.h"
-#include "cx18-vbi.h"
 #include "cx18-scb.h"
-#include "cx18-dvb.h"
-
-void cx18_work_handler(struct work_struct *work)
-{
-	struct cx18 *cx = container_of(work, struct cx18, work);
-	if (test_and_clear_bit(CX18_F_I_WORK_HANDLER_DVB, &cx->i_flags))
-		cx18_dvb_work_handler(cx);
-}
-
-static void epu_dma_done(struct cx18 *cx, struct cx18_mailbox *mb, int rpu)
-{
-	u32 handle = mb->args[0];
-	struct cx18_stream *s = NULL;
-	struct cx18_buffer *buf;
-	u32 off;
-	int i;
-	int id;
-
-	for (i = 0; i < CX18_MAX_STREAMS; i++) {
-		s = &cx->streams[i];
-		if ((handle == s->handle) && (s->dvb.enabled))
-			break;
-		if (s->v4l2dev && handle == s->handle)
-			break;
-	}
-	if (i == CX18_MAX_STREAMS) {
-		CX18_WARN("Got DMA done notification for unknown/inactive"
-			  " handle %d\n", handle);
-		mb->error = CXERR_NOT_OPEN;
-		mb->cmd = 0;
-		cx18_mb_ack(cx, mb, rpu);
-		return;
-	}
-
-	off = mb->args[1];
-	if (mb->args[2] != 1)
-		CX18_WARN("Ack struct = %d for %s\n",
-			mb->args[2], s->name);
-	id = cx18_read_enc(cx, off);
-	buf = cx18_queue_get_buf_irq(s, id, cx18_read_enc(cx, off + 4));
-	CX18_DEBUG_HI_DMA("DMA DONE for %s (buffer %d)\n", s->name, id);
-	if (buf) {
-		cx18_buf_sync_for_cpu(s, buf);
-		if (s->type == CX18_ENC_STREAM_TYPE_TS && s->dvb.enabled) {
-			CX18_DEBUG_HI_DMA("TS recv bytesused = %d\n",
-					buf->bytesused);
-
-			set_bit(CX18_F_I_WORK_HANDLER_DVB, &cx->i_flags);
-			set_bit(CX18_F_I_HAVE_WORK, &cx->i_flags);
-		} else
-			set_bit(CX18_F_B_NEED_BUF_SWAP, &buf->b_flags);
-	} else {
-		CX18_WARN("Could not find buf %d for stream %s\n",
-				cx18_read_enc(cx, off), s->name);
-	}
-	mb->error = 0;
-	mb->cmd = 0;
-	cx18_mb_ack(cx, mb, rpu);
-	wake_up(&cx->dma_waitq);
-	if (s->id != -1)
-		wake_up(&s->waitq);
-}
-
-static void epu_debug(struct cx18 *cx, struct cx18_mailbox *mb, int rpu)
-{
-	char str[256] = { 0 };
-	char *p;
-
-	if (mb->args[1]) {
-		cx18_setup_page(cx, mb->args[1]);
-		cx18_memcpy_fromio(cx, str, cx->enc_mem + mb->args[1], 252);
-		str[252] = 0;
-	}
-	cx18_mb_ack(cx, mb, rpu);
-	CX18_DEBUG_INFO("%x %s\n", mb->args[0], str);
-	p = strchr(str, '.');
-	if (!test_bit(CX18_F_I_LOADED_FW, &cx->i_flags) && p && p > str)
-		CX18_INFO("FW version: %s\n", p - 1);
-}
-
-static void epu_cmd(struct cx18 *cx, u32 sw1)
-{
-	struct cx18_mailbox mb;
-
-	if (sw1 & IRQ_CPU_TO_EPU) {
-		cx18_memcpy_fromio(cx, &mb, &cx->scb->cpu2epu_mb, sizeof(mb));
-		mb.error = 0;
-
-		switch (mb.cmd) {
-		case CX18_EPU_DMA_DONE:
-			epu_dma_done(cx, &mb, CPU);
-			break;
-		case CX18_EPU_DEBUG:
-			epu_debug(cx, &mb, CPU);
-			break;
-		default:
-			CX18_WARN("Unknown CPU_TO_EPU mailbox command %#08x\n",
-				  mb.cmd);
-			break;
-		}
-	}
-
-	if (sw1 & IRQ_APU_TO_EPU) {
-		cx18_memcpy_fromio(cx, &mb, &cx->scb->apu2epu_mb, sizeof(mb));
-		CX18_WARN("Unknown APU_TO_EPU mailbox command %#08x\n", mb.cmd);
-	}
-}
 
 static void xpu_ack(struct cx18 *cx, u32 sw2)
 {
@@ -145,6 +33,14 @@ static void xpu_ack(struct cx18 *cx, u32 sw2)
 		wake_up(&cx->mb_apu_waitq);
 }
 
+static void epu_cmd(struct cx18 *cx, u32 sw1)
+{
+	if (sw1 & IRQ_CPU_TO_EPU)
+		cx18_api_epu_cmd_irq(cx, CPU);
+	if (sw1 & IRQ_APU_TO_EPU)
+		cx18_api_epu_cmd_irq(cx, APU);
+}
+
 irqreturn_t cx18_irq_handler(int irq, void *dev_id)
 {
 	struct cx18 *cx = (struct cx18 *)dev_id;
@@ -170,6 +66,13 @@ irqreturn_t cx18_irq_handler(int irq, void *dev_id)
 		CX18_DEBUG_HI_IRQ("received interrupts "
 				  "SW1: %x  SW2: %x  HW2: %x\n", sw1, sw2, hw2);
 
+	/*
+	 * SW1 responses have to happen first.  The sending XPU times out the
+	 * incoming mailboxes on us rather rapidly.
+	 */
+	if (sw1)
+		epu_cmd(cx, sw1);
+
 	/* To do: interrupt-based I2C handling
 	if (hw2 & (HW2_I2C1_INT|HW2_I2C2_INT)) {
 	}
@@ -178,11 +81,5 @@ irqreturn_t cx18_irq_handler(int irq, void *dev_id)
 	if (sw2)
 		xpu_ack(cx, sw2);
 
-	if (sw1)
-		epu_cmd(cx, sw1);
-
-	if (test_and_clear_bit(CX18_F_I_HAVE_WORK, &cx->i_flags))
-		schedule_work(&cx->work);
-
 	return (sw1 || sw2 || hw2) ? IRQ_HANDLED : IRQ_NONE;
 }
diff --git a/drivers/media/video/cx18/cx18-irq.h b/drivers/media/video/cx18/cx18-irq.h
index 6f3ec896376..cd8f548b7ee 100644
--- a/drivers/media/video/cx18/cx18-irq.h
+++ b/drivers/media/video/cx18/cx18-irq.h
@@ -32,5 +32,3 @@
 #define SW2_INT_ENABLE_PCI              0xc7315c
 
 irqreturn_t cx18_irq_handler(int irq, void *dev_id);
-
-void cx18_work_handler(struct work_struct *work);
diff --git a/drivers/media/video/cx18/cx18-mailbox.c b/drivers/media/video/cx18/cx18-mailbox.c
index 35f7188d4d3..3d210d2ffda 100644
--- a/drivers/media/video/cx18/cx18-mailbox.c
+++ b/drivers/media/video/cx18/cx18-mailbox.c
@@ -26,6 +26,10 @@
 #include "cx18-scb.h"
 #include "cx18-irq.h"
 #include "cx18-mailbox.h"
+#include "cx18-queue.h"
+#include "cx18-streams.h"
+
+static const char *rpu_str[] = { "APU", "CPU", "EPU", "HPU" };
 
 #define API_FAST (1 << 2) /* Short timeout */
 #define API_SLOW (1 << 3) /* Additional 300ms timeout */
@@ -92,12 +96,149 @@ static const struct cx18_api_info *find_api_info(u32 cmd)
 	return NULL;
 }
 
-long cx18_mb_ack(struct cx18 *cx, const struct cx18_mailbox *mb, int rpu)
+static void dump_mb(struct cx18 *cx, struct cx18_mailbox *mb, char *name)
+{
+	char argstr[MAX_MB_ARGUMENTS*11+1];
+	char *p;
+	int i;
+
+	if (!(cx18_debug & CX18_DBGFLG_API))
+		return;
+
+	for (i = 0, p = argstr; i < MAX_MB_ARGUMENTS; i++, p += 11) {
+		/* kernel snprintf() appends '\0' always */
+		snprintf(p, 12, " %#010x", mb->args[i]);
+	}
+	CX18_DEBUG_API("%s: req %#010x ack %#010x cmd %#010x err %#010x args%s"
+		"\n", name, mb->request, mb->ack, mb->cmd, mb->error, argstr);
+}
+
+
+/*
+ * Functions that run in a work_queue work handling context
+ */
+
+static void epu_dma_done(struct cx18 *cx, struct cx18_epu_work_order *order)
+{
+	u32 handle, mdl_ack_count;
+	struct cx18_mailbox *mb;
+	struct cx18_mdl_ack *mdl_ack;
+	struct cx18_stream *s;
+	struct cx18_buffer *buf;
+	int i;
+
+	mb = &order->mb;
+	handle = mb->args[0];
+	s = cx18_handle_to_stream(cx, handle);
+
+	if (s == NULL) {
+		CX18_WARN("Got DMA done notification for unknown/inactive"
+			  " handle %d\n", handle);
+		return;
+	}
+
+	mdl_ack_count = mb->args[2];
+	mdl_ack = order->mdl_ack;
+	for (i = 0; i < mdl_ack_count; i++, mdl_ack++) {
+		buf = cx18_queue_get_buf(s, mdl_ack->id, mdl_ack->data_used);
+		CX18_DEBUG_HI_DMA("DMA DONE for %s (buffer %d)\n", s->name,
+				  mdl_ack->id);
+		if (buf == NULL) {
+			CX18_WARN("Could not find buf %d for stream %s\n",
+				  mdl_ack->id, s->name);
+			continue;
+		}
+
+		cx18_buf_sync_for_cpu(s, buf);
+		if (s->type == CX18_ENC_STREAM_TYPE_TS && s->dvb.enabled) {
+			CX18_DEBUG_HI_DMA("TS recv bytesused = %d\n",
+					  buf->bytesused);
+
+			dvb_dmx_swfilter(&s->dvb.demux, buf->buf,
+					 buf->bytesused);
+
+			cx18_buf_sync_for_device(s, buf);
+
+			if (s->handle != CX18_INVALID_TASK_HANDLE &&
+			    test_bit(CX18_F_S_STREAMING, &s->s_flags))
+				cx18_vapi(cx,
+				       CX18_CPU_DE_SET_MDL, 5, s->handle,
+				       (void __iomem *)
+				       &cx->scb->cpu_mdl[buf->id] - cx->enc_mem,
+				       1, buf->id, s->buf_size);
+		} else
+			set_bit(CX18_F_B_NEED_BUF_SWAP, &buf->b_flags);
+	}
+	wake_up(&cx->dma_waitq);
+	if (s->id != -1)
+		wake_up(&s->waitq);
+}
+
+static void epu_debug(struct cx18 *cx, struct cx18_epu_work_order *order)
+{
+	char *p;
+	char *str = order->str;
+
+	CX18_DEBUG_INFO("%x %s\n", order->mb.args[0], str);
+	p = strchr(str, '.');
+	if (!test_bit(CX18_F_I_LOADED_FW, &cx->i_flags) && p && p > str)
+		CX18_INFO("FW version: %s\n", p - 1);
+}
+
+static void epu_cmd(struct cx18 *cx, struct cx18_epu_work_order *order)
+{
+	switch (order->rpu) {
+	case CPU:
+	{
+		switch (order->mb.cmd) {
+		case CX18_EPU_DMA_DONE:
+			epu_dma_done(cx, order);
+			break;
+		case CX18_EPU_DEBUG:
+			epu_debug(cx, order);
+			break;
+		default:
+			CX18_WARN("Unknown CPU to EPU mailbox command %#0x\n",
+				  order->mb.cmd);
+			break;
+		}
+		break;
+	}
+	case APU:
+		CX18_WARN("Unknown APU to EPU mailbox command %#0x\n",
+			  order->mb.cmd);
+		break;
+	default:
+		break;
+	}
+}
+
+static
+void free_epu_work_order(struct cx18 *cx, struct cx18_epu_work_order *order)
+{
+	atomic_set(&order->pending, 0);
+}
+
+void cx18_epu_work_handler(struct work_struct *work)
+{
+	struct cx18_epu_work_order *order =
+			container_of(work, struct cx18_epu_work_order, work);
+	struct cx18 *cx = order->cx;
+	epu_cmd(cx, order);
+	free_epu_work_order(cx, order);
+}
+
+
+/*
+ * Functions that run in an interrupt handling context
+ */
+
+static void mb_ack_irq(struct cx18 *cx, const struct cx18_epu_work_order *order)
 {
 	struct cx18_mailbox __iomem *ack_mb;
-	u32 ack_irq;
+	u32 ack_irq, req;
 
-	switch (rpu) {
+	switch (order->rpu) {
 	case APU:
 		ack_irq = IRQ_EPU_TO_APU_ACK;
 		ack_mb = &cx->scb->apu2epu_mb;
@@ -108,16 +249,175 @@ long cx18_mb_ack(struct cx18 *cx, const struct cx18_mailbox *mb, int rpu)
 		break;
 	default:
 		CX18_WARN("Unhandled RPU (%d) for command %x ack\n",
-			  rpu, mb->cmd);
-		return -EINVAL;
+			  order->rpu, order->mb.cmd);
+		return;
 	}
 
-	cx18_setup_page(cx, SCB_OFFSET);
-	cx18_write_sync(cx, mb->request, &ack_mb->ack);
+	req = order->mb.request;
+	/* Don't ack if the RPU has gotten impatient and timed us out */
+	if (req != cx18_readl(cx, &ack_mb->request) ||
+	    req == cx18_readl(cx, &ack_mb->ack))
+		return;
+	cx18_writel(cx, req, &ack_mb->ack);
 	cx18_write_reg_expect(cx, ack_irq, SW2_INT_SET, ack_irq, ack_irq);
-	return 0;
+	return;
+}
+
+static int epu_dma_done_irq(struct cx18 *cx, struct cx18_epu_work_order *order,
+			    int stale)
+{
+	u32 handle, mdl_ack_offset, mdl_ack_count;
+	struct cx18_mailbox *mb;
+
+	mb = &order->mb;
+	handle = mb->args[0];
+	mdl_ack_offset = mb->args[1];
+	mdl_ack_count = mb->args[2];
+
+	if (handle == CX18_INVALID_TASK_HANDLE ||
+	    mdl_ack_count == 0 || mdl_ack_count > CX18_MAX_MDL_ACKS) {
+		if (!stale)
+			mb_ack_irq(cx, order);
+		return -1;
+	}
+
+	cx18_memcpy_fromio(cx, order->mdl_ack, cx->enc_mem + mdl_ack_offset,
+			   sizeof(struct cx18_mdl_ack) * mdl_ack_count);
+	if (!stale)
+		mb_ack_irq(cx, order);
+	return 1;
+}
+
+static
+int epu_debug_irq(struct cx18 *cx, struct cx18_epu_work_order *order, int stale)
+{
+	u32 str_offset;
+	char *str = order->str;
+
+	str[0] = '\0';
+	str_offset = order->mb.args[1];
+	if (str_offset) {
+		cx18_setup_page(cx, str_offset);
+		cx18_memcpy_fromio(cx, str, cx->enc_mem + str_offset, 252);
+		str[252] = '\0';
+		cx18_setup_page(cx, SCB_OFFSET);
+	}
+
+	if (!stale)
+		mb_ack_irq(cx, order);
+
+	return str_offset ? 1 : 0;
+}
+
+static inline
+int epu_cmd_irq(struct cx18 *cx, struct cx18_epu_work_order *order, int stale)
+{
+	int ret = -1;
+
+	switch (order->rpu) {
+	case CPU:
+	{
+		switch (order->mb.cmd) {
+		case CX18_EPU_DMA_DONE:
+			ret = epu_dma_done_irq(cx, order, stale);
+			break;
+		case CX18_EPU_DEBUG:
+			ret = epu_debug_irq(cx, order, stale);
+			break;
+		default:
+			CX18_WARN("Unknown CPU to EPU mailbox command %#0x\n",
+				  order->mb.cmd);
+			break;
+		}
+		break;
+	}
+	case APU:
+		CX18_WARN("Unknown APU to EPU mailbox command %#0x\n",
+			  order->mb.cmd);
+		break;
+	default:
+		break;
+	}
+	return ret;
+}
+
+static inline
+struct cx18_epu_work_order *alloc_epu_work_order_irq(struct cx18 *cx)
+{
+	int i;
+	struct cx18_epu_work_order *order = NULL;
+
+	for (i = 0; i < CX18_MAX_EPU_WORK_ORDERS; i++) {
+		/*
+		 * We only need "pending" atomic to inspect its contents,
+		 * and need not do a check and set because:
+		 * 1. Any work handler thread only clears "pending" and only
+		 * on one, particular work order at a time, per handler thread.
+		 * 2. "pending" is only set here, and we're serialized because
+		 * we're called in an IRQ handler context.
+		 */
+		if (atomic_read(&cx->epu_work_order[i].pending) == 0) {
+			order = &cx->epu_work_order[i];
+			atomic_set(&order->pending, 1);
+			break;
+		}
+	}
+	return order;
 }
 
+void cx18_api_epu_cmd_irq(struct cx18 *cx, int rpu)
+{
+	struct cx18_mailbox __iomem *mb;
+	struct cx18_mailbox *order_mb;
+	struct cx18_epu_work_order *order;
+	int stale = 0;
+	int submit;
+
+	switch (rpu) {
+	case CPU:
+		mb = &cx->scb->cpu2epu_mb;
+		break;
+	case APU:
+		mb = &cx->scb->apu2epu_mb;
+		break;
+	default:
+		return;
+	}
+
+	order = alloc_epu_work_order_irq(cx);
+	if (order == NULL) {
+		CX18_WARN("Unable to find blank work order form to schedule "
+			  "incoming mailbox command processing\n");
+		return;
+	}
+
+	order->rpu = rpu;
+	order_mb = &order->mb;
+	cx18_memcpy_fromio(cx, order_mb, mb, sizeof(struct cx18_mailbox));
+
+	if (order_mb->request == order_mb->ack) {
+		CX18_WARN("Possibly falling behind: %s self-ack'ed our incoming"
+			  " %s to EPU mailbox (sequence no. %u)\n",
+			  rpu_str[rpu], rpu_str[rpu], order_mb->request);
+		dump_mb(cx, order_mb, "incoming");
+		stale = 1;
+	}
+
+	/*
+	 * Individual EPU command processing is responsible for ack-ing
+	 * a non-stale mailbox as soon as possible
+	 */
+	submit = epu_cmd_irq(cx, order, stale);
+	if (submit > 0) {
+		schedule_work(&order->work);
+	}
+}
+
+
+/*
+ * Functions called from a non-interrupt, non work_queue context
+ */
+
 static void cx18_api_log_ack_delay(struct cx18 *cx, int msecs)
 {
 	if (msecs > CX18_MAX_MB_ACK_DELAY)
@@ -167,8 +467,6 @@ static int cx18_api_call(struct cx18 *cx, u32 cmd, int args, u32 data[])
 	}
 
 	mutex_lock(mb_lock);
-	cx18_setup_page(cx, SCB_OFFSET);
-
 	/*
 	 * Wait for an in-use mailbox to complete
 	 *
diff --git a/drivers/media/video/cx18/cx18-mailbox.h b/drivers/media/video/cx18/cx18-mailbox.h
index 54758f32db1..002c0526afa 100644
--- a/drivers/media/video/cx18/cx18-mailbox.h
+++ b/drivers/media/video/cx18/cx18-mailbox.h
@@ -37,6 +37,17 @@
 
 struct cx18;
 
+/*
+ * This structure is used by CPU to provide completed buffers information
+ * Its structure is dictrated by the layout of the SCB, required by the
+ * firmware, but its defintion needs to be here, instead of in cx18-scb.h,
+ * for mailbox work order scheduling
+ */
+struct cx18_mdl_ack {
+    u32 id;        /* ID of a completed MDL */
+    u32 data_used; /* Total data filled in the MDL for buffer 'id' */
+};
+
 /* The cx18_mailbox struct is the mailbox structure which is used for passing
    messages between processors */
 struct cx18_mailbox {
@@ -73,6 +84,9 @@ int cx18_vapi_result(struct cx18 *cx, u32 data[MAX_MB_ARGUMENTS], u32 cmd,
 int cx18_vapi(struct cx18 *cx, u32 cmd, int args, ...);
 int cx18_api_func(void *priv, u32 cmd, int in, int out,
 		u32 data[CX2341X_MBOX_MAX_DATA]);
-long cx18_mb_ack(struct cx18 *cx, const struct cx18_mailbox *mb, int rpu);
+
+void cx18_api_epu_cmd_irq(struct cx18 *cx, int rpu);
+
+void cx18_epu_work_handler(struct work_struct *work);
 
 #endif
diff --git a/drivers/media/video/cx18/cx18-queue.c b/drivers/media/video/cx18/cx18-queue.c
index 174682c2582..c5a81aed61b 100644
--- a/drivers/media/video/cx18/cx18-queue.c
+++ b/drivers/media/video/cx18/cx18-queue.c
@@ -75,30 +75,37 @@ struct cx18_buffer *cx18_dequeue(struct cx18_stream *s, struct cx18_queue *q)
 	return buf;
 }
 
-struct cx18_buffer *cx18_queue_get_buf_irq(struct cx18_stream *s, u32 id,
+struct cx18_buffer *cx18_queue_get_buf(struct cx18_stream *s, u32 id,
 	u32 bytesused)
 {
 	struct cx18 *cx = s->cx;
 	struct list_head *p;
+	unsigned long flags = 0;
 
-	spin_lock(&s->qlock);
+	spin_lock_irqsave(&s->qlock, flags);
 	list_for_each(p, &s->q_free.list) {
 		struct cx18_buffer *buf =
 			list_entry(p, struct cx18_buffer, list);
 
-		if (buf->id != id)
+		if (buf->id != id) {
+			CX18_DEBUG_HI_DMA("Skipping buffer %d searching for %d "
+					  "in stream %s q_free\n", buf->id, id,
+					  s->name);
 			continue;
+		}
 
 		buf->bytesused = bytesused;
-		atomic_dec(&s->q_free.buffers);
-		atomic_inc(&s->q_full.buffers);
-		s->q_full.bytesused += buf->bytesused;
-		list_move_tail(&buf->list, &s->q_full.list);
+		if (s->type != CX18_ENC_STREAM_TYPE_TS) {
+			atomic_dec(&s->q_free.buffers);
+			atomic_inc(&s->q_full.buffers);
+			s->q_full.bytesused += buf->bytesused;
+			list_move_tail(&buf->list, &s->q_full.list);
+		}
 
-		spin_unlock(&s->qlock);
+		spin_unlock_irqrestore(&s->qlock, flags);
 		return buf;
 	}
-	spin_unlock(&s->qlock);
+	spin_unlock_irqrestore(&s->qlock, flags);
 	CX18_ERR("Cannot find buffer %d for stream %s\n", id, s->name);
 	return NULL;
 }
diff --git a/drivers/media/video/cx18/cx18-queue.h b/drivers/media/video/cx18/cx18-queue.h
index 7f93bb13c09..92edb5e3a62 100644
--- a/drivers/media/video/cx18/cx18-queue.h
+++ b/drivers/media/video/cx18/cx18-queue.h
@@ -46,7 +46,7 @@ void cx18_queue_init(struct cx18_queue *q);
 void cx18_enqueue(struct cx18_stream *s, struct cx18_buffer *buf,
 	struct cx18_queue *q);
 struct cx18_buffer *cx18_dequeue(struct cx18_stream *s, struct cx18_queue *q);
-struct cx18_buffer *cx18_queue_get_buf_irq(struct cx18_stream *s, u32 id,
+struct cx18_buffer *cx18_queue_get_buf(struct cx18_stream *s, u32 id,
 	u32 bytesused);
 void cx18_flush_queues(struct cx18_stream *s);
 
diff --git a/drivers/media/video/cx18/cx18-scb.h b/drivers/media/video/cx18/cx18-scb.h
index 594713bbed6..29866f02f01 100644
--- a/drivers/media/video/cx18/cx18-scb.h
+++ b/drivers/media/video/cx18/cx18-scb.h
@@ -85,12 +85,6 @@ struct cx18_mdl {
     u32 length; /* Length of the buffer segment */
 };
 
-/* This structure is used by CPU to provide completed buffers information */
-struct cx18_mdl_ack {
-    u32 id;        /* ID of a completed MDL */
-    u32 data_used; /* Total data filled in the MDL for buffer 'id' */
-};
-
 struct cx18_scb {
 	/* These fields form the System Control Block which is used at boot time
 	   for localizing the IPC data as well as the code positions for all
@@ -276,7 +270,7 @@ struct cx18_scb {
 	struct cx18_mailbox  hpu2epu_mb;
 	struct cx18_mailbox  ppu2epu_mb;
 
-	struct cx18_mdl_ack  cpu_mdl_ack[CX18_MAX_STREAMS][2];
+	struct cx18_mdl_ack  cpu_mdl_ack[CX18_MAX_STREAMS][CX18_MAX_MDL_ACKS];
 	struct cx18_mdl      cpu_mdl[1];
 };
 
diff --git a/drivers/media/video/cx18/cx18-streams.c b/drivers/media/video/cx18/cx18-streams.c
index c7d431f4923..e6d808f7cc8 100644
--- a/drivers/media/video/cx18/cx18-streams.c
+++ b/drivers/media/video/cx18/cx18-streams.c
@@ -595,3 +595,21 @@ u32 cx18_find_handle(struct cx18 *cx)
 	}
 	return CX18_INVALID_TASK_HANDLE;
 }
+
+struct cx18_stream *cx18_handle_to_stream(struct cx18 *cx, u32 handle)
+{
+	int i;
+	struct cx18_stream *s;
+
+	if (handle == CX18_INVALID_TASK_HANDLE)
+		return NULL;
+
+	for (i = 0; i < CX18_MAX_STREAMS; i++) {
+		s = &cx->streams[i];
+		if (s->handle != handle)
+			continue;
+		if (s->v4l2dev || s->dvb.enabled)
+			return s;
+	}
+	return NULL;
+}
diff --git a/drivers/media/video/cx18/cx18-streams.h b/drivers/media/video/cx18/cx18-streams.h
index f327e947b24..3fd578d0e6c 100644
--- a/drivers/media/video/cx18/cx18-streams.h
+++ b/drivers/media/video/cx18/cx18-streams.h
@@ -22,6 +22,7 @@
  */
 
 u32 cx18_find_handle(struct cx18 *cx);
+struct cx18_stream *cx18_handle_to_stream(struct cx18 *cx, u32 handle);
 int cx18_streams_setup(struct cx18 *cx);
 int cx18_streams_register(struct cx18 *cx);
 void cx18_streams_cleanup(struct cx18 *cx, int unregister);