merge: http://linuxtv.org/hg/~awalls/cx18-perf

From: Mauro Carvalho Chehab <mchehab@redhat.com>

Signed-off-by: Mauro Carvalho Chehab <mchehab@redhat.com>

9 files changed, 277 insertions, 133 deletions

diff --git a/linux/drivers/media/video/cx18/cx18-driver.c b/linux/drivers/media/video/cx18/cx18-driver.c
index 062129fb5..41997dd18 100644
--- a/linux/drivers/media/video/cx18/cx18-driver.c
+++ b/linux/drivers/media/video/cx18/cx18-driver.c
@@ -30,6 +30,7 @@
 #include "cx18-irq.h"
 #include "cx18-gpio.h"
 #include "cx18-firmware.h"
+#include "cx18-queue.h"
 #include "cx18-streams.h"
 #include "cx18-av-core.h"
 #include "cx18-scb.h"
@@ -546,6 +547,45 @@ done:
 	cx->card_i2c = cx->card->i2c;
 }
 
+static int __devinit cx18_create_in_workq(struct cx18 *cx)
+{
+	snprintf(cx->in_workq_name, sizeof(cx->in_workq_name), "%s-in",
+		 cx->v4l2_dev.name);
+	cx->in_work_queue = create_singlethread_workqueue(cx->in_workq_name);
+	if (cx->in_work_queue == NULL) {
+		CX18_ERR("Unable to create incoming mailbox handler thread\n");
+		return -ENOMEM;
+	}
+	return 0;
+}
+
+static int __devinit cx18_create_out_workq(struct cx18 *cx)
+{
+	snprintf(cx->out_workq_name, sizeof(cx->out_workq_name), "%s-out",
+		 cx->v4l2_dev.name);
+	cx->out_work_queue = create_workqueue(cx->out_workq_name);
+	if (cx->out_work_queue == NULL) {
+		CX18_ERR("Unable to create outgoing mailbox handler threads\n");
+		return -ENOMEM;
+	}
+	return 0;
+}
+
+static void __devinit cx18_init_in_work_orders(struct cx18 *cx)
+{
+	int i;
+	for (i = 0; i < CX18_MAX_IN_WORK_ORDERS; i++) {
+		cx->in_work_order[i].cx = cx;
+		cx->in_work_order[i].str = cx->epu_debug_str;
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 20)
+		INIT_WORK(&cx->in_work_order[i].work, cx18_in_work_handler);
+#else
+		INIT_WORK(&cx->in_work_order[i].work, cx18_in_work_handler,
+			  &cx->in_work_order[i].work);
+#endif
+	}
+}
+
 /* Precondition: the cx18 structure has been memset to 0. Only
    the dev and instance fields have been filled in.
    No assumptions on the card type may be made here (see cx18_init_struct2
@@ -553,7 +593,7 @@ done:
  */
 static int __devinit cx18_init_struct1(struct cx18 *cx)
 {
-	int i;
+	int ret;
 
 	cx->base_addr = pci_resource_start(cx->pci_dev, 0);
 
@@ -562,23 +602,18 @@ static int __devinit cx18_init_struct1(struct cx18 *cx)
 	mutex_init(&cx->epu2apu_mb_lock);
 	mutex_init(&cx->epu2cpu_mb_lock);
 
-	cx->work_queue = create_singlethread_workqueue(cx->v4l2_dev.name);
-	if (cx->work_queue == NULL) {
-		CX18_ERR("Unable to create work hander thread\n");
-		return -ENOMEM;
-	}
+	ret = cx18_create_out_workq(cx);
+	if (ret)
+		return ret;
 
-	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;
-#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 20)
-		INIT_WORK(&cx->epu_work_order[i].work, cx18_epu_work_handler);
-#else
-		INIT_WORK(&cx->epu_work_order[i].work, cx18_epu_work_handler,
-			  &cx->epu_work_order[i].work);
-#endif
+	ret = cx18_create_in_workq(cx);
+	if (ret) {
+		destroy_workqueue(cx->out_work_queue);
+		return ret;
 	}
 
+	cx18_init_in_work_orders(cx);
+
 	/* start counting open_id at 1 */
 	cx->open_id = 1;
 
@@ -764,17 +799,17 @@ static int __devinit cx18_probe(struct pci_dev *pci_dev,
 		retval = -ENODEV;
 		goto err;
 	}
-	if (cx18_init_struct1(cx)) {
-		retval = -ENOMEM;
+
+	retval = cx18_init_struct1(cx);
+	if (retval)
 		goto err;
-	}
 
 	CX18_DEBUG_INFO("base addr: 0x%08x\n", cx->base_addr);
 
 	/* PCI Device Setup */
 	retval = cx18_setup_pci(cx, pci_dev, pci_id);
 	if (retval != 0)
-		goto free_workqueue;
+		goto free_workqueues;
 
 	/* map io memory */
 	CX18_DEBUG_INFO("attempting ioremap at 0x%08x len 0x%08x\n",
@@ -948,8 +983,9 @@ free_map:
 	cx18_iounmap(cx);
 free_mem:
 	release_mem_region(cx->base_addr, CX18_MEM_SIZE);
-free_workqueue:
-	destroy_workqueue(cx->work_queue);
+free_workqueues:
+	destroy_workqueue(cx->in_work_queue);
+	destroy_workqueue(cx->out_work_queue);
 err:
 	if (retval == 0)
 		retval = -ENODEV;
@@ -1059,11 +1095,19 @@ int cx18_init_on_first_open(struct cx18 *cx)
 }
 
 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 22)
-static void cx18_cancel_epu_work_orders(struct cx18 *cx)
+static void cx18_cancel_in_work_orders(struct cx18 *cx)
 {
 	int i;
-	for (i = 0; i < CX18_MAX_EPU_WORK_ORDERS; i++)
-		cancel_work_sync(&cx->epu_work_order[i].work);
+	for (i = 0; i < CX18_MAX_IN_WORK_ORDERS; i++)
+		cancel_work_sync(&cx->in_work_order[i].work);
+}
+
+static void cx18_cancel_out_work_orders(struct cx18 *cx)
+{
+	int i;
+	for (i = 0; i < CX18_MAX_STREAMS; i++)
+		if (&cx->streams[i].video_dev != NULL)
+			cancel_work_sync(&cx->streams[i].out_work_order);
 }
 
 #endif
@@ -1080,24 +1124,25 @@ static void cx18_remove(struct pci_dev *pci_dev)
 	if (atomic_read(&cx->tot_capturing) > 0)
 		cx18_stop_all_captures(cx);
 
-	/* Interrupts */
+	/* Stop interrupts that cause incoming work to be queued */
 	cx18_sw1_irq_disable(cx, IRQ_CPU_TO_EPU | IRQ_APU_TO_EPU);
-#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 22)
-	cx18_sw2_irq_disable(cx, IRQ_CPU_TO_EPU_ACK | IRQ_APU_TO_EPU_ACK);
-
-	cx18_halt_firmware(cx);
 
-	cx18_cancel_epu_work_orders(cx);
+	/* Incoming work can cause outgoing work, so clean up incoming first */
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 22)
+	cx18_cancel_in_work_orders(cx);
+	cx18_cancel_out_work_orders(cx);
 #else
+	flush_workqueue(cx->in_work_queue);
+	flush_workqueue(cx->out_work_queue);
+#endif
 
-	flush_workqueue(cx->work_queue);
-
+	/* Stop ack interrupts that may have been needed for work to finish */
 	cx18_sw2_irq_disable(cx, IRQ_CPU_TO_EPU_ACK | IRQ_APU_TO_EPU_ACK);
 
 	cx18_halt_firmware(cx);
-#endif
 
-	destroy_workqueue(cx->work_queue);
+	destroy_workqueue(cx->in_work_queue);
+	destroy_workqueue(cx->out_work_queue);
 
 	cx18_streams_cleanup(cx, 1);
 
diff --git a/linux/drivers/media/video/cx18/cx18-driver.h b/linux/drivers/media/video/cx18/cx18-driver.h
index ece4f281e..f89b82367 100644
--- a/linux/drivers/media/video/cx18/cx18-driver.h
+++ b/linux/drivers/media/video/cx18/cx18-driver.h
@@ -254,6 +254,7 @@ struct cx18_options {
 #define CX18_F_S_INTERNAL_USE	5	/* this stream is used internally (sliced VBI processing) */
 #define CX18_F_S_STREAMOFF	7	/* signal end of stream EOS */
 #define CX18_F_S_APPL_IO        8	/* this stream is used read/written by an application */
+#define CX18_F_S_STOPPING	9	/* telling the fw to stop capturing */
 
 /* per-cx18, i_flags */
 #define CX18_F_I_LOADED_FW		0 	/* Loaded firmware 1st time */
@@ -285,6 +286,7 @@ struct cx18_queue {
 	struct list_head list;
 	atomic_t buffers;
 	u32 bytesused;
+	spinlock_t lock;
 };
 
 struct cx18_dvb {
@@ -305,7 +307,7 @@ struct cx18_scb; /* forward reference */
 
 
 #define CX18_MAX_MDL_ACKS 2
-#define CX18_MAX_EPU_WORK_ORDERS (CX18_MAX_FW_MDLS_PER_STREAM + 7)
+#define CX18_MAX_IN_WORK_ORDERS (CX18_MAX_FW_MDLS_PER_STREAM + 7)
 /* CPU_DE_RELEASE_MDL can burst CX18_MAX_FW_MDLS_PER_STREAM orders in a group */
 
 #define CX18_F_EWO_MB_STALE_UPON_RECEIPT 0x1
@@ -313,7 +315,7 @@ struct cx18_scb; /* forward reference */
 #define CX18_F_EWO_MB_STALE \
 	     (CX18_F_EWO_MB_STALE_UPON_RECEIPT | CX18_F_EWO_MB_STALE_WHILE_PROC)
 
-struct cx18_epu_work_order {
+struct cx18_in_work_order {
 	struct work_struct work;
 	atomic_t pending;
 	struct cx18 *cx;
@@ -337,7 +339,6 @@ struct cx18_stream {
 	unsigned mdl_offset;
 
 	u32 id;
-	struct mutex qlock; 	/* locks access to the queues */
 	unsigned long s_flags;	/* status flags, see above */
 	int dma;		/* can be PCI_DMA_TODEVICE,
 				   PCI_DMA_FROMDEVICE or
@@ -353,6 +354,8 @@ struct cx18_stream {
 	struct cx18_queue q_busy;	/* busy buffers - in use by firmware */
 	struct cx18_queue q_full;	/* full buffers - data for user apps */
 
+	struct work_struct out_work_order;
+
 	/* DVB / Digital Transport */
 	struct cx18_dvb dvb;
 };
@@ -568,10 +571,14 @@ struct cx18 {
 	u32 sw2_irq_mask;
 	u32 hw2_irq_mask;
 
-	struct workqueue_struct *work_queue;
-	struct cx18_epu_work_order epu_work_order[CX18_MAX_EPU_WORK_ORDERS];
+	struct workqueue_struct *in_work_queue;
+	char in_workq_name[11]; /* "cx18-NN-in" */
+	struct cx18_in_work_order in_work_order[CX18_MAX_IN_WORK_ORDERS];
 	char epu_debug_str[256]; /* CX18_EPU_DEBUG is rare: use shared space */
 
+	struct workqueue_struct *out_work_queue;
+	char out_workq_name[12]; /* "cx18-NN-out" */
+
 	/* i2c */
 	struct i2c_adapter i2c_adap[2];
 	struct i2c_algo_bit_data i2c_algo[2];
diff --git a/linux/drivers/media/video/cx18/cx18-dvb.c b/linux/drivers/media/video/cx18/cx18-dvb.c
index 3b86f57cd..e7285a109 100644
--- a/linux/drivers/media/video/cx18/cx18-dvb.c
+++ b/linux/drivers/media/video/cx18/cx18-dvb.c
@@ -23,6 +23,7 @@
 #include "cx18-version.h"
 #include "cx18-dvb.h"
 #include "cx18-io.h"
+#include "cx18-queue.h"
 #include "cx18-streams.h"
 #include "cx18-cards.h"
 #include "s5h1409.h"
diff --git a/linux/drivers/media/video/cx18/cx18-mailbox.c b/linux/drivers/media/video/cx18/cx18-mailbox.c
index 66964722f..36f4d3dd7 100644
--- a/linux/drivers/media/video/cx18/cx18-mailbox.c
+++ b/linux/drivers/media/video/cx18/cx18-mailbox.c
@@ -131,7 +131,7 @@ static void dump_mb(struct cx18 *cx, struct cx18_mailbox *mb, char *name)
  * Functions that run in a work_queue work handling context
  */
 
-static void epu_dma_done(struct cx18 *cx, struct cx18_epu_work_order *order)
+static void epu_dma_done(struct cx18 *cx, struct cx18_in_work_order *order)
 {
 	u32 handle, mdl_ack_count, id;
 	struct cx18_mailbox *mb;
@@ -191,29 +191,30 @@ static void epu_dma_done(struct cx18 *cx, struct cx18_epu_work_order *order)
 		if (buf == NULL) {
 			CX18_WARN("Could not find buf %d for stream %s\n",
 				  id, s->name);
-			/* Put as many buffers as possible back into fw use */
-			cx18_stream_load_fw_queue(s);
 			continue;
 		}
 
-		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_DEBUG_HI_DMA("%s recv bytesused = %d\n",
+				  s->name, buf->bytesused);
+
+		if (s->type != CX18_ENC_STREAM_TYPE_TS)
+			cx18_enqueue(s, buf, &s->q_full);
+		else {
+			if (s->dvb.enabled)
+				dvb_dmx_swfilter(&s->dvb.demux, buf->buf,
+						 buf->bytesused);
+			cx18_enqueue(s, buf, &s->q_free);
 		}
-		/* Put as many buffers as possible back into fw use */
-		cx18_stream_load_fw_queue(s);
-		/* Put back TS buffer, since it was removed from all queues */
-		if (s->type == CX18_ENC_STREAM_TYPE_TS)
-			cx18_stream_put_buf_fw(s, buf);
 	}
+	/* Put as many buffers as possible back into fw use */
+	cx18_stream_load_fw_queue(s);
+
 	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)
+static void epu_debug(struct cx18 *cx, struct cx18_in_work_order *order)
 {
 	char *p;
 	char *str = order->str;
@@ -224,7 +225,7 @@ static void epu_debug(struct cx18 *cx, struct cx18_epu_work_order *order)
 		CX18_INFO("FW version: %s\n", p - 1);
 }
 
-static void epu_cmd(struct cx18 *cx, struct cx18_epu_work_order *order)
+static void epu_cmd(struct cx18 *cx, struct cx18_in_work_order *order)
 {
 	switch (order->rpu) {
 	case CPU:
@@ -253,24 +254,24 @@ static void epu_cmd(struct cx18 *cx, struct cx18_epu_work_order *order)
 }
 
 static
-void free_epu_work_order(struct cx18 *cx, struct cx18_epu_work_order *order)
+void free_in_work_order(struct cx18 *cx, struct cx18_in_work_order *order)
 {
 	atomic_set(&order->pending, 0);
 }
 
 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 20)
-void cx18_epu_work_handler(struct work_struct *work)
+void cx18_in_work_handler(struct work_struct *work)
 {
-	struct cx18_epu_work_order *order =
-			container_of(work, struct cx18_epu_work_order, work);
+	struct cx18_in_work_order *order =
+			container_of(work, struct cx18_in_work_order, work);
 #else
-void cx18_epu_work_handler(void *arg)
+void cx18_in_work_handler(void *arg)
 {
-	struct cx18_epu_work_order *order = arg;
+	struct cx18_in_work_order *order = arg;
 #endif
 	struct cx18 *cx = order->cx;
 	epu_cmd(cx, order);
-	free_epu_work_order(cx, order);
+	free_in_work_order(cx, order);
 }
 
 
@@ -278,7 +279,7 @@ void cx18_epu_work_handler(void *arg)
  * Functions that run in an interrupt handling context
  */
 
-static void mb_ack_irq(struct cx18 *cx, struct cx18_epu_work_order *order)
+static void mb_ack_irq(struct cx18 *cx, struct cx18_in_work_order *order)
 {
 	struct cx18_mailbox __iomem *ack_mb;
 	u32 ack_irq, req;
@@ -314,7 +315,7 @@ static void mb_ack_irq(struct cx18 *cx, struct cx18_epu_work_order *order)
 	return;
 }
 
-static int epu_dma_done_irq(struct cx18 *cx, struct cx18_epu_work_order *order)
+static int epu_dma_done_irq(struct cx18 *cx, struct cx18_in_work_order *order)
 {
 	u32 handle, mdl_ack_offset, mdl_ack_count;
 	struct cx18_mailbox *mb;
@@ -340,7 +341,7 @@ static int epu_dma_done_irq(struct cx18 *cx, struct cx18_epu_work_order *order)
 }
 
 static
-int epu_debug_irq(struct cx18 *cx, struct cx18_epu_work_order *order)
+int epu_debug_irq(struct cx18 *cx, struct cx18_in_work_order *order)
 {
 	u32 str_offset;
 	char *str = order->str;
@@ -361,7 +362,7 @@ int epu_debug_irq(struct cx18 *cx, struct cx18_epu_work_order *order)
 }
 
 static inline
-int epu_cmd_irq(struct cx18 *cx, struct cx18_epu_work_order *order)
+int epu_cmd_irq(struct cx18 *cx, struct cx18_in_work_order *order)
 {
 	int ret = -1;
 
@@ -393,12 +394,12 @@ int epu_cmd_irq(struct cx18 *cx, struct cx18_epu_work_order *order)
 }
 
 static inline
-struct cx18_epu_work_order *alloc_epu_work_order_irq(struct cx18 *cx)
+struct cx18_in_work_order *alloc_in_work_order_irq(struct cx18 *cx)
 {
 	int i;
-	struct cx18_epu_work_order *order = NULL;
+	struct cx18_in_work_order *order = NULL;
 
-	for (i = 0; i < CX18_MAX_EPU_WORK_ORDERS; i++) {
+	for (i = 0; i < CX18_MAX_IN_WORK_ORDERS; i++) {
 		/*
 		 * We only need "pending" atomic to inspect its contents,
 		 * and need not do a check and set because:
@@ -407,8 +408,8 @@ struct cx18_epu_work_order *alloc_epu_work_order_irq(struct cx18 *cx)
 		 * 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];
+		if (atomic_read(&cx->in_work_order[i].pending) == 0) {
+			order = &cx->in_work_order[i];
 			atomic_set(&order->pending, 1);
 			break;
 		}
@@ -420,7 +421,7 @@ 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;
+	struct cx18_in_work_order *order;
 	int submit;
 
 	switch (rpu) {
@@ -434,7 +435,7 @@ void cx18_api_epu_cmd_irq(struct cx18 *cx, int rpu)
 		return;
 	}
 
-	order = alloc_epu_work_order_irq(cx);
+	order = alloc_in_work_order_irq(cx);
 	if (order == NULL) {
 		CX18_WARN("Unable to find blank work order form to schedule "
 			  "incoming mailbox command processing\n");
@@ -467,7 +468,7 @@ void cx18_api_epu_cmd_irq(struct cx18 *cx, int rpu)
 	 */
 	submit = epu_cmd_irq(cx, order);
 	if (submit > 0) {
-		queue_work(cx->work_queue, &order->work);
+		queue_work(cx->in_work_queue, &order->work);
 	}
 }
 
@@ -484,9 +485,10 @@ static int cx18_api_call(struct cx18 *cx, u32 cmd, int args, u32 data[])
 	u32 __iomem *xpu_state;
 	wait_queue_head_t *waitq;
 	struct mutex *mb_lock;
-	long int timeout, ret;
+	unsigned long int t0, timeout, ret;
 	int i;
 	char argstr[MAX_MB_ARGUMENTS*11+1];
+	DEFINE_WAIT(w);
 
 	if (info == NULL) {
 		CX18_WARN("unknown cmd %x\n", cmd);
@@ -568,25 +570,49 @@ static int cx18_api_call(struct cx18 *cx, u32 cmd, int args, u32 data[])
 
 	CX18_DEBUG_HI_IRQ("sending interrupt SW1: %x to send %s\n",
 			  irq, info->name);
+
+	/* So we don't miss the wakeup, prepare to wait before notifying fw */
+	prepare_to_wait(waitq, &w, TASK_UNINTERRUPTIBLE);
 	cx18_write_reg_expect(cx, irq, SW1_INT_SET, irq, irq);
 
-	ret = wait_event_timeout(
-		       *waitq,
-		       cx18_readl(cx, &mb->ack) == cx18_readl(cx, &mb->request),
-		       timeout);
+	t0 = jiffies;
+	ack = cx18_readl(cx, &mb->ack);
+	if (ack != req) {
+		schedule_timeout(timeout);
+		ret = jiffies - t0;
+		ack = cx18_readl(cx, &mb->ack);
+	} else {
+		ret = jiffies - t0;
+	}
 
-	if (ret == 0) {
-		/* Timed out */
+	finish_wait(waitq, &w);
+
+	if (req != ack) {
 		mutex_unlock(mb_lock);
-		CX18_DEBUG_WARN("sending %s timed out waiting %d msecs for RPU "
-				"acknowledgement\n",
-				info->name, jiffies_to_msecs(timeout));
+		if (ret >= timeout) {
+			/* Timed out */
+			CX18_DEBUG_WARN("sending %s timed out waiting %d msecs "
+					"for RPU acknowledgement\n",
+					info->name, jiffies_to_msecs(ret));
+		} else {
+			CX18_DEBUG_WARN("woken up before mailbox ack was ready "
+					"after submitting %s to RPU.  only "
+					"waited %d msecs on req %u but awakened"
+					" with unmatched ack %u\n",
+					info->name,
+					jiffies_to_msecs(ret),
+					req, ack);
+		}
 		return -EINVAL;
 	}
 
-	if (ret != timeout)
+	if (ret >= timeout)
+		CX18_DEBUG_WARN("failed to be awakened upon RPU acknowledgment "
+				"sending %s; timed out waiting %d msecs\n",
+				info->name, jiffies_to_msecs(ret));
+	else
 		CX18_DEBUG_HI_API("waited %u msecs for %s to be acked\n",
-				  jiffies_to_msecs(timeout-ret), info->name);
+				  jiffies_to_msecs(ret), info->name);
 
 	/* Collect data returned by the XPU */
 	for (i = 0; i < MAX_MB_ARGUMENTS; i++)
diff --git a/linux/drivers/media/video/cx18/cx18-mailbox.h b/linux/drivers/media/video/cx18/cx18-mailbox.h
index a667f1ae4..4bd0ba57f 100644
--- a/linux/drivers/media/video/cx18/cx18-mailbox.h
+++ b/linux/drivers/media/video/cx18/cx18-mailbox.h
@@ -96,9 +96,9 @@ int cx18_api_func(void *priv, u32 cmd, int in, int out,
 void cx18_api_epu_cmd_irq(struct cx18 *cx, int rpu);
 
 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 20)
-void cx18_epu_work_handler(struct work_struct *work);
+void cx18_in_work_handler(struct work_struct *work);
 #else
-void cx18_epu_work_handler(void *arg);
+void cx18_in_work_handler(void *arg);
 #endif
 
 #endif
diff --git a/linux/drivers/media/video/cx18/cx18-queue.c b/linux/drivers/media/video/cx18/cx18-queue.c
index 3046b8e74..fa1ed7897 100644
--- a/linux/drivers/media/video/cx18/cx18-queue.c
+++ b/linux/drivers/media/video/cx18/cx18-queue.c
@@ -23,8 +23,8 @@
  */
 
 #include "cx18-driver.h"
-#include "cx18-streams.h"
 #include "cx18-queue.h"
+#include "cx18-streams.h"
 #include "cx18-scb.h"
 
 void cx18_buf_swap(struct cx18_buffer *buf)
@@ -53,13 +53,13 @@ struct cx18_queue *_cx18_enqueue(struct cx18_stream *s, struct cx18_buffer *buf,
 		buf->skipped = 0;
 	}
 
-	mutex_lock(&s->qlock);
-
 	/* q_busy is restricted to a max buffer count imposed by firmware */
 	if (q == &s->q_busy &&
 	    atomic_read(&q->buffers) >= CX18_MAX_FW_MDLS_PER_STREAM)
 		q = &s->q_free;
 
+	spin_lock(&q->lock);
+
 	if (to_front)
 		list_add(&buf->list, &q->list); /* LIFO */
 	else
@@ -67,7 +67,7 @@ struct cx18_queue *_cx18_enqueue(struct cx18_stream *s, struct cx18_buffer *buf,
 	q->bytesused += buf->bytesused - buf->readpos;
 	atomic_inc(&q->buffers);
 
-	mutex_unlock(&s->qlock);
+	spin_unlock(&q->lock);
 	return q;
 }
 
@@ -75,7 +75,7 @@ struct cx18_buffer *cx18_dequeue(struct cx18_stream *s, struct cx18_queue *q)
 {
 	struct cx18_buffer *buf = NULL;
 
-	mutex_lock(&s->qlock);
+	spin_lock(&q->lock);
 	if (!list_empty(&q->list)) {
 		buf = list_first_entry(&q->list, struct cx18_buffer, list);
 		list_del_init(&buf->list);
@@ -83,7 +83,7 @@ struct cx18_buffer *cx18_dequeue(struct cx18_stream *s, struct cx18_queue *q)
 		buf->skipped = 0;
 		atomic_dec(&q->buffers);
 	}
-	mutex_unlock(&s->qlock);
+	spin_unlock(&q->lock);
 	return buf;
 }
 
@@ -94,9 +94,23 @@ struct cx18_buffer *cx18_queue_get_buf(struct cx18_stream *s, u32 id,
 	struct cx18_buffer *buf;
 	struct cx18_buffer *tmp;
 	struct cx18_buffer *ret = NULL;
-
-	mutex_lock(&s->qlock);
+	LIST_HEAD(sweep_up);
+
+	/*
+	 * We don't have to acquire multiple q locks here, because we are
+	 * serialized by the single threaded work handler.
+	 * Buffers from the firmware will thus remain in order as
+	 * they are moved from q_busy to q_full or to the dvb ring buffer.
+	 */
+	spin_lock(&s->q_busy.lock);
 	list_for_each_entry_safe(buf, tmp, &s->q_busy.list, list) {
+		/*
+		 * We should find what the firmware told us is done,
+		 * right at the front of the queue.  If we don't, we likely have
+		 * missed a buffer done message from the firmware.
+		 * Once we skip a buffer repeatedly, relative to the size of
+		 * q_busy, we have high confidence we've missed it.
+		 */
 		if (buf->id != id) {
 			buf->skipped++;
 			if (buf->skipped >= atomic_read(&s->q_busy.buffers)-1) {
@@ -105,38 +119,41 @@ struct cx18_buffer *cx18_queue_get_buf(struct cx18_stream *s, u32 id,
 					  "times - it must have dropped out of "
 					  "rotation\n", s->name, buf->id,
 					  buf->skipped);
-				/* move it to q_free */
-				list_move_tail(&buf->list, &s->q_free.list);
-				buf->bytesused = buf->readpos = buf->b_flags =
-					buf->skipped = 0;
+				/* Sweep it up to put it back into rotation */
+				list_move_tail(&buf->list, &sweep_up);
 				atomic_dec(&s->q_busy.buffers);
-				atomic_inc(&s->q_free.buffers);
 			}
 			continue;
 		}
-
-		buf->bytesused = bytesused;
-		/* Sync the buffer before we release the qlock */
-		cx18_buf_sync_for_cpu(s, buf);
-		if (s->type == CX18_ENC_STREAM_TYPE_TS) {
-			/*
-			 * TS doesn't use q_full.  As we pull the buffer off of
-			 * the queue here, the caller will have to put it back.
-			 */
-			list_del_init(&buf->list);
-		} else {
-			/* Move buffer from q_busy to q_full */
-			list_move_tail(&buf->list, &s->q_full.list);
-			set_bit(CX18_F_B_NEED_BUF_SWAP, &buf->b_flags);
-			s->q_full.bytesused += buf->bytesused;
-			atomic_inc(&s->q_full.buffers);
-		}
+		/*
+		 * We pull the desired buffer off of the queue here.  Something
+		 * will have to put it back on a queue later.
+		 */
+		list_del_init(&buf->list);
 		atomic_dec(&s->q_busy.buffers);
-
 		ret = buf;
 		break;
 	}
-	mutex_unlock(&s->qlock);
+	spin_unlock(&s->q_busy.lock);
+
+	/*
+	 * We found the buffer for which we were looking.  Get it ready for
+	 * the caller to put on q_full or in the dvb ring buffer.
+	 */
+	if (ret != NULL) {
+		ret->bytesused = bytesused;
+		ret->skipped = 0;
+		/* readpos and b_flags were 0'ed when the buf went on q_busy */
+		cx18_buf_sync_for_cpu(s, ret);
+		if (s->type != CX18_ENC_STREAM_TYPE_TS)
+			set_bit(CX18_F_B_NEED_BUF_SWAP, &ret->b_flags);
+	}
+
+	/* Put any buffers the firmware is ignoring back into normal rotation */
+	list_for_each_entry_safe(buf, tmp, &sweep_up, list) {
+		list_del_init(&buf->list);
+		cx18_enqueue(s, buf, &s->q_free);
+	}
 	return ret;
 }
 
@@ -148,7 +165,7 @@ static void cx18_queue_flush(struct cx18_stream *s, struct cx18_queue *q)
 	if (q == &s->q_free)
 		return;
 
-	mutex_lock(&s->qlock);
+	spin_lock(&q->lock);
 	while (!list_empty(&q->list)) {
 		buf = list_first_entry(&q->list, struct cx18_buffer, list);
 		list_move_tail(&buf->list, &s->q_free.list);
@@ -156,7 +173,7 @@ static void cx18_queue_flush(struct cx18_stream *s, struct cx18_queue *q)
 		atomic_inc(&s->q_free.buffers);
 	}
 	cx18_queue_init(q);
-	mutex_unlock(&s->qlock);
+	spin_unlock(&q->lock);
 }
 
 void cx18_flush_queues(struct cx18_stream *s)
diff --git a/linux/drivers/media/video/cx18/cx18-streams.c b/linux/drivers/media/video/cx18/cx18-streams.c
index 7c1db9c18..142302ba5 100644
--- a/linux/drivers/media/video/cx18/cx18-streams.c
+++ b/linux/drivers/media/video/cx18/cx18-streams.c
@@ -116,12 +116,21 @@ static void cx18_stream_init(struct cx18 *cx, int type)
 	s->buffers = cx->stream_buffers[type];
 	s->buf_size = cx->stream_buf_size[type];
 
-	mutex_init(&s->qlock);
 	init_waitqueue_head(&s->waitq);
 	s->id = -1;
+	spin_lock_init(&s->q_free.lock);
 	cx18_queue_init(&s->q_free);
+	spin_lock_init(&s->q_busy.lock);
 	cx18_queue_init(&s->q_busy);
+	spin_lock_init(&s->q_full.lock);
 	cx18_queue_init(&s->q_full);
+
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 20)
+	INIT_WORK(&s->out_work_order, cx18_out_work_handler);
+#else
+	INIT_WORK(&s->out_work_order, cx18_out_work_handler,
+		  &s->out_work_order);
+#endif
 }
 
 static int cx18_prep_dev(struct cx18 *cx, int type)
@@ -431,14 +440,16 @@ static void cx18_vbi_setup(struct cx18_stream *s)
 	cx18_api(cx, CX18_CPU_SET_RAW_VBI_PARAM, 6, data);
 }
 
-struct cx18_queue *cx18_stream_put_buf_fw(struct cx18_stream *s,
-					  struct cx18_buffer *buf)
+static
+struct cx18_queue *_cx18_stream_put_buf_fw(struct cx18_stream *s,
+					   struct cx18_buffer *buf)
 {
 	struct cx18 *cx = s->cx;
 	struct cx18_queue *q;
 
 	/* Don't give it to the firmware, if we're not running a capture */
 	if (s->handle == CX18_INVALID_TASK_HANDLE ||
+	    test_bit(CX18_F_S_STOPPING, &s->s_flags) ||
 	    !test_bit(CX18_F_S_STREAMING, &s->s_flags))
 		return cx18_enqueue(s, buf, &s->q_free);
 
@@ -453,7 +464,8 @@ struct cx18_queue *cx18_stream_put_buf_fw(struct cx18_stream *s,
 	return q;
 }
 
-void cx18_stream_load_fw_queue(struct cx18_stream *s)
+static
+void _cx18_stream_load_fw_queue(struct cx18_stream *s)
 {
 	struct cx18_queue *q;
 	struct cx18_buffer *buf;
@@ -467,11 +479,25 @@ void cx18_stream_load_fw_queue(struct cx18_stream *s)
 		buf = cx18_dequeue(s, &s->q_free);
 		if (buf == NULL)
 			break;
-		q = cx18_stream_put_buf_fw(s, buf);
+		q = _cx18_stream_put_buf_fw(s, buf);
 	} while (atomic_read(&s->q_busy.buffers) < CX18_MAX_FW_MDLS_PER_STREAM
 		 && q == &s->q_busy);
 }
 
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 20)
+void cx18_out_work_handler(struct work_struct *work)
+{
+#else
+void cx18_out_work_handler(void *arg)
+{
+	struct work_struct *work = arg;
+#endif
+	struct cx18_stream *s =
+			 container_of(work, struct cx18_stream, out_work_order);
+
+	_cx18_stream_load_fw_queue(s);
+}
+
 int cx18_start_v4l2_encode_stream(struct cx18_stream *s)
 {
 	u32 data[MAX_MB_ARGUMENTS];
@@ -600,19 +626,20 @@ int cx18_start_v4l2_encode_stream(struct cx18_stream *s)
 
 	/* Init all the cpu_mdls for this stream */
 	cx18_flush_queues(s);
-	mutex_lock(&s->qlock);
+	spin_lock(&s->q_free.lock);
 	list_for_each_entry(buf, &s->q_free.list, list) {
 		cx18_writel(cx, buf->dma_handle,
 					&cx->scb->cpu_mdl[buf->id].paddr);
 		cx18_writel(cx, s->buf_size, &cx->scb->cpu_mdl[buf->id].length);
 	}
-	mutex_unlock(&s->qlock);
-	cx18_stream_load_fw_queue(s);
+	spin_unlock(&s->q_free.lock);
+	_cx18_stream_load_fw_queue(s);
 
 	/* begin_capture */
 	if (cx18_vapi(cx, CX18_CPU_CAPTURE_START, 1, s->handle)) {
 		CX18_DEBUG_WARN("Error starting capture!\n");
 		/* Ensure we're really not capturing before releasing MDLs */
+		set_bit(CX18_F_S_STOPPING, &s->s_flags);
 		if (s->type == CX18_ENC_STREAM_TYPE_MPG)
 			cx18_vapi(cx, CX18_CPU_CAPTURE_STOP, 2, s->handle, 1);
 		else
@@ -622,6 +649,7 @@ int cx18_start_v4l2_encode_stream(struct cx18_stream *s)
 		cx18_vapi(cx, CX18_CPU_DE_RELEASE_MDL, 1, s->handle);
 		cx18_vapi(cx, CX18_DESTROY_TASK, 1, s->handle);
 		s->handle = CX18_INVALID_TASK_HANDLE;
+		clear_bit(CX18_F_S_STOPPING, &s->s_flags);
 		if (atomic_read(&cx->tot_capturing) == 0) {
 			set_bit(CX18_F_I_EOS, &cx->i_flags);
 			cx18_write_reg(cx, 5, CX18_DSP0_INTERRUPT_MASK);
@@ -666,6 +694,7 @@ int cx18_stop_v4l2_encode_stream(struct cx18_stream *s, int gop_end)
 	if (atomic_read(&cx->tot_capturing) == 0)
 		return 0;
 
+	set_bit(CX18_F_S_STOPPING, &s->s_flags);
 	if (s->type == CX18_ENC_STREAM_TYPE_MPG)
 		cx18_vapi(cx, CX18_CPU_CAPTURE_STOP, 2, s->handle, !gop_end);
 	else
@@ -716,6 +745,7 @@ int cx18_stop_v4l2_encode_stream(struct cx18_stream *s, int gop_end)
 
 	cx18_vapi(cx, CX18_DESTROY_TASK, 1, s->handle);
 	s->handle = CX18_INVALID_TASK_HANDLE;
+	clear_bit(CX18_F_S_STOPPING, &s->s_flags);
 
 	if (atomic_read(&cx->tot_capturing) > 0)
 		return 0;
diff --git a/linux/drivers/media/video/cx18/cx18-streams.h b/linux/drivers/media/video/cx18/cx18-streams.h
index 420e0a172..b51a014b5 100644
--- a/linux/drivers/media/video/cx18/cx18-streams.h
+++ b/linux/drivers/media/video/cx18/cx18-streams.h
@@ -28,10 +28,28 @@ int cx18_streams_setup(struct cx18 *cx);
 int cx18_streams_register(struct cx18 *cx);
 void cx18_streams_cleanup(struct cx18 *cx, int unregister);
 
+/* Related to submission of buffers to firmware */
+static inline void cx18_stream_load_fw_queue(struct cx18_stream *s)
+{
+	struct cx18 *cx = s->cx;
+	queue_work(cx->out_work_queue, &s->out_work_order);
+}
+
+static inline void cx18_stream_put_buf_fw(struct cx18_stream *s,
+					  struct cx18_buffer *buf)
+{
+	/* Put buf on q_free; the out work handler will move buf(s) to q_busy */
+	cx18_enqueue(s, buf, &s->q_free);
+	cx18_stream_load_fw_queue(s);
+}
+
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 20)
+void cx18_out_work_handler(struct work_struct *work);
+#else
+void cx18_out_work_handler(void *arg);
+#endif
+
 /* Capture related */
-void cx18_stream_load_fw_queue(struct cx18_stream *s);
-struct cx18_queue *cx18_stream_put_buf_fw(struct cx18_stream *s,
-					  struct cx18_buffer *buf);
 int cx18_start_v4l2_encode_stream(struct cx18_stream *s);
 int cx18_stop_v4l2_encode_stream(struct cx18_stream *s, int gop_end);
 
diff --git a/linux/drivers/media/video/cx18/cx18-version.h b/linux/drivers/media/video/cx18/cx18-version.h
index bd9bd44da..45494b094 100644
--- a/linux/drivers/media/video/cx18/cx18-version.h
+++ b/linux/drivers/media/video/cx18/cx18-version.h
@@ -24,7 +24,7 @@
 
 #define CX18_DRIVER_NAME "cx18"
 #define CX18_DRIVER_VERSION_MAJOR 1
-#define CX18_DRIVER_VERSION_MINOR 1
+#define CX18_DRIVER_VERSION_MINOR 2
 #define CX18_DRIVER_VERSION_PATCHLEVEL 0
 
 #define CX18_VERSION __stringify(CX18_DRIVER_VERSION_MAJOR) "." __stringify(CX18_DRIVER_VERSION_MINOR) "." __stringify(CX18_DRIVER_VERSION_PATCHLEVEL)