dma: mmp_pdma: support for getting residual bytes

In some of our drivers (e.g. UART) we may stop a running DMA before it finishes. So we need to know how many bytes have been transferred.

drivers/dma/mmp_pdma.c +0 -0

@@ -589,7 +638,13 @@ static int mmp_pdma_control(struct dma_chan *dchan, enum dma_ctrl_cmd cmd,
                mmp_pdma_free_desc_list(chan, &chan->chain_pending);

Quoted 3 lines

In some of our drivers (e.g. UART) we may stop a running DMA before it finishes. So we need to know how many bytes have been transferred.

Couple of comments below.

Quoted 12 lines

drivers/dma/mmp_pdma.c +4 -1

@@ -589,7 +638,13 @@ static int mmp_pdma_control(struct dma_chan *dchan, enum dma_ctrl_cmd cmd,
                mmp_pdma_free_desc_list(chan, &chan->chain_pending);
                mmp_pdma_free_desc_list(chan, &chan->chain_running);
                spin_unlock_irqrestore(&chan->desc_lock, flags);
-               chan->idle = true;
+               chan->status = DMA_SUCCESS;
+               chan->bytes_residue = 0;
+               break;
+       case DMA_PAUSE:

drivers/dma/mmp_pdma.c static int mmp_pdma_control(struct dma_chan *dchan, enum dma_ctrl_cmd cmd, · line 647

Quoted 3 lines

drivers/dma/mmp_pdma.c +3 -0

+               disable_chan(chan->phy);
+               chan->status = DMA_PAUSED;
+               chan->bytes_residue = mmp_pdma_get_bytes_residue(chan);

Does it mean user has to do DMA_PAUSE first to get more or less accurate residue?
Logically that sound correct, but in general we may allow user to get approximate residue value of on going transfer.

drivers/dma/mmp_pdma.c static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan, · line 697

Quoted 12 lines

drivers/dma/mmp_pdma.c +2 -1

                break;
        case DMA_SLAVE_CONFIG:
                if (cfg->direction == DMA_DEV_TO_MEM) {

@@ -637,7 +692,8 @@ static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan,
        unsigned long flags;
 
        spin_lock_irqsave(&chan->desc_lock, flags);
-       ret = dma_cookie_status(dchan, cookie, txstate);
+       ret = chan->status;
+       dma_set_residue(txstate, chan->bytes_residue);
        spin_unlock_irqrestore(&chan->desc_lock, flags);

Besides my patch which removes this spinlock I think the workflow should be something like

status = dma_cookie_status()
if status == DMA_SUCCESS or !txstate:
return status

dma_set_residue()
return status

Because there is no reason to return residue of successfully finished transfer. It should be 0.

--
With Best Regards,
Andy Shevchenko

2013/5/31 Andy Shevchenko <[email protected]>:

Quoted 19 lines · Xiang Wang

Quoted 3 lines

In some of our drivers (e.g. UART) we may stop a running DMA before it finishes. So we need to know how many bytes have been transferred.

Couple of comments below.

Quoted 12 lines

drivers/dma/mmp_pdma.c +4 -1

@@ -589,7 +638,13 @@ static int mmp_pdma_control(struct dma_chan *dchan, enum dma_ctrl_cmd cmd,
                mmp_pdma_free_desc_list(chan, &chan->chain_pending);
                mmp_pdma_free_desc_list(chan, &chan->chain_running);
                spin_unlock_irqrestore(&chan->desc_lock, flags);
-               chan->idle = true;
+               chan->status = DMA_SUCCESS;
+               chan->bytes_residue = 0;
+               break;
+       case DMA_PAUSE:

Quoted 8 lines

drivers/dma/mmp_pdma.c static int mmp_pdma_control(struct dma_chan *dchan, enum dma_ctrl_cmd cmd, · line 647

Quoted 3 lines

drivers/dma/mmp_pdma.c +3 -0

+               disable_chan(chan->phy);
+               chan->status = DMA_PAUSED;
+               chan->bytes_residue = mmp_pdma_get_bytes_residue(chan);

Does it mean user has to do DMA_PAUSE first to get more or less accurate residue?
Logically that sound correct, but in general we may allow user to get approximate residue value of on going transfer.

Your comment makes sense. But if the user is allowed to query the residue value in real-time, we cannot just return a saved value to him.

Why I use a saved value (chan->bytes_residue)?
In current mmp pdma driver, a phy channel will be freed after the transmission finishes (chan->phy is set to NULL). So we cannot get the physical channel information after we call DMA_TERMINATE_ALL or DMA finishes itself.

That is to say, when the use queries the channel information at these points, the chan->phy is usually NULL.

Quoted 19 lines

drivers/dma/mmp_pdma.c static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan, · line 697

Quoted 8 lines

drivers/dma/mmp_pdma.c +2 -1

@@ -637,7 +692,8 @@ static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan,
        unsigned long flags;
 
        spin_lock_irqsave(&chan->desc_lock, flags);
-       ret = dma_cookie_status(dchan, cookie, txstate);
+       ret = chan->status;
+       dma_set_residue(txstate, chan->bytes_residue);
        spin_unlock_irqrestore(&chan->desc_lock, flags);

Besides my patch which removes this spinlock I think the workflow should be something like

status = dma_cookie_status()
if status == DMA_SUCCESS or !txstate:
return status

dma_set_residue()
return status

Quoted 2 lines

Because there is no reason to return residue of successfully finished transfer. It should be 0.

There is one exception from my point of view. When we are receiving data from peripheral devices, we usually start a DMA transaction with a target length of 4K for example. When a timed-out event occurs in peripheral device, it will notify DMA controller and DMA controller will send out a End of Receive interrupt (Marvell specific?).

In such situation, DMA status is also DMA_SUCCESS. But the residual bytes is not 0 and the user must query it.

Quoted 29 lines

2013/5/31 Andy Shevchenko <[email protected]>:

Quoted 19 lines · Xiang Wang

Quoted 3 lines

In some of our drivers (e.g. UART) we may stop a running DMA before it finishes. So we need to know how many bytes have been transferred.

Couple of comments below.

Quoted 12 lines

drivers/dma/mmp_pdma.c +4 -1

@@ -589,7 +638,13 @@ static int mmp_pdma_control(struct dma_chan *dchan, enum dma_ctrl_cmd cmd,
                mmp_pdma_free_desc_list(chan, &chan->chain_pending);
                mmp_pdma_free_desc_list(chan, &chan->chain_running);
                spin_unlock_irqrestore(&chan->desc_lock, flags);
-               chan->idle = true;
+               chan->status = DMA_SUCCESS;
+               chan->bytes_residue = 0;
+               break;
+       case DMA_PAUSE:

Quoted 8 lines

drivers/dma/mmp_pdma.c static int mmp_pdma_control(struct dma_chan *dchan, enum dma_ctrl_cmd cmd, · line 647

Quoted 3 lines

drivers/dma/mmp_pdma.c +3 -0

+               disable_chan(chan->phy);
+               chan->status = DMA_PAUSED;
+               chan->bytes_residue = mmp_pdma_get_bytes_residue(chan);

Does it mean user has to do DMA_PAUSE first to get more or less accurate residue?
Logically that sound correct, but in general we may allow user to get approximate residue value of on going transfer.

Quoted 3 lines

Your comment makes sense. But if the user is allowed to query the residue value in real-time, we cannot just return a saved value to him.

Right.

Quoted 5 lines

Why I use a saved value (chan->bytes_residue)?
In current mmp pdma driver, a phy channel will be freed after the transmission finishes (chan->phy is set to NULL). So we cannot get the physical channel information after we call DMA_TERMINATE_ALL or DMA finishes itself.

I don't see any contradiction to workflow.
So, If you call device_tx_status() when transfer is completed or aborted you will get 0 as a residue, which is correct.

Quoted 22 lines

That is to say, when the use queries the channel information at these points, the chan->phy is usually NULL.

Quoted 19 lines

drivers/dma/mmp_pdma.c static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan, · line 697

Quoted 8 lines

drivers/dma/mmp_pdma.c +2 -1

@@ -637,7 +692,8 @@ static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan,
        unsigned long flags;
 
        spin_lock_irqsave(&chan->desc_lock, flags);
-       ret = dma_cookie_status(dchan, cookie, txstate);
+       ret = chan->status;
+       dma_set_residue(txstate, chan->bytes_residue);
        spin_unlock_irqrestore(&chan->desc_lock, flags);

Besides my patch which removes this spinlock I think the workflow should be something like

status = dma_cookie_status()
if status == DMA_SUCCESS or !txstate:
return status

dma_set_residue()
return status

Quoted 8 lines

Quoted 2 lines

Because there is no reason to return residue of successfully finished transfer. It should be 0.

There is one exception from my point of view. When we are receiving data from peripheral devices, we usually start a DMA transaction with a target length of 4K for example. When a timed-out event occurs in peripheral device, it will notify DMA controller and DMA controller will send out a End of Receive interrupt (Marvell specific?).

Might be your hardware specifics, in our case we have got a timeout interrupt from uart controller.

Quoted 2 lines

In such situation, DMA status is also DMA_SUCCESS. But the residual bytes is not 0 and the user must query it.

Which sounds wrong approach.

P.S. take a look at drivers/tty/serial/8250/8250_dma.c

--
With Best Regards,
Andy Shevchenko

2013/6/3 Andy Shevchenko <[email protected]>:

Quoted 6 lines · Xiang Wang

Quoted 5 lines

2013/5/31 Andy Shevchenko <[email protected]>:

Quoted 4 lines · Xiang Wang

Quoted 3 lines

In some of our drivers (e.g. UART) we may stop a running DMA before it finishes. So we need to know how many bytes have been transferred.

Quoted 46 lines

Quoted 5 lines

Why I use a saved value (chan->bytes_residue)?
In current mmp pdma driver, a phy channel will be freed after the transmission finishes (chan->phy is set to NULL). So we cannot get the physical channel information after we call DMA_TERMINATE_ALL or DMA finishes itself.

I don't see any contradiction to workflow.
So, If you call device_tx_status() when transfer is completed or aborted you will get 0 as a residue, which is correct.

Quoted 19 lines

Quoted 19 lines

drivers/dma/mmp_pdma.c static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan, · line 697

Quoted 8 lines

drivers/dma/mmp_pdma.c +2 -1

@@ -637,7 +692,8 @@ static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan,
        unsigned long flags;
 
        spin_lock_irqsave(&chan->desc_lock, flags);
-       ret = dma_cookie_status(dchan, cookie, txstate);
+       ret = chan->status;
+       dma_set_residue(txstate, chan->bytes_residue);
        spin_unlock_irqrestore(&chan->desc_lock, flags);

Besides my patch which removes this spinlock I think the workflow should be something like

status = dma_cookie_status()
if status == DMA_SUCCESS or !txstate:
return status

dma_set_residue()
return status

Quoted 8 lines

Quoted 2 lines

Because there is no reason to return residue of successfully finished transfer. It should be 0.

There is one exception from my point of view. When we are receiving data from peripheral devices, we usually start a DMA transaction with a target length of 4K for example. When a timed-out event occurs in peripheral device, it will notify DMA controller and DMA controller will send out a End of Receive interrupt (Marvell specific?).

Might be your hardware specifics, in our case we have got a timeout interrupt from uart controller.

Quoted 2 lines

In such situation, DMA status is also DMA_SUCCESS. But the residual bytes is not 0 and the user must query it.

Which sounds wrong approach.

Quoted 1 line

P.S. take a look at drivers/tty/serial/8250/8250_dma.c

When peripheral device (e.g. UART) is using DMA controller, we should have 2 solutions to deal with trailing bytes:

1. Timeout interrupt from UART controller.
In this case, we usually pause DMA and read out data from DMA buffer in uart irq handler.

2. DMA controller handles trailing bytes for us.
This is the case I mentioned in my previous email. "When a timed-out event occurs in peripheral device, it will notify DMA controller and DMA controller will send out a End of Receive interrupt" I think we should know how many residual bytes in this case even the DMA status is DMA_SUCCESS.

Quoted 7 lines

2013/6/3 Andy Shevchenko <[email protected]>:

Quoted 6 lines · Xiang Wang

Quoted 5 lines

2013/5/31 Andy Shevchenko <[email protected]>:

Quoted 4 lines · Xiang Wang

Quoted 3 lines

In some of our drivers (e.g. UART) we may stop a running DMA before it finishes. So we need to know how many bytes have been transferred.

[]

Quoted 50 lines

Quoted 41 lines

Quoted 5 lines

Why I use a saved value (chan->bytes_residue)?
In current mmp pdma driver, a phy channel will be freed after the transmission finishes (chan->phy is set to NULL). So we cannot get the physical channel information after we call DMA_TERMINATE_ALL or DMA finishes itself.

I don't see any contradiction to workflow.
So, If you call device_tx_status() when transfer is completed or aborted you will get 0 as a residue, which is correct.

Quoted 19 lines

Quoted 19 lines

drivers/dma/mmp_pdma.c static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan, · line 697

Quoted 8 lines

drivers/dma/mmp_pdma.c +2 -1

@@ -637,7 +692,8 @@ static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan,
        unsigned long flags;
 
        spin_lock_irqsave(&chan->desc_lock, flags);
-       ret = dma_cookie_status(dchan, cookie, txstate);
+       ret = chan->status;
+       dma_set_residue(txstate, chan->bytes_residue);
        spin_unlock_irqrestore(&chan->desc_lock, flags);

Besides my patch which removes this spinlock I think the workflow should be something like

status = dma_cookie_status()
if status == DMA_SUCCESS or !txstate:
return status

dma_set_residue()
return status

Quoted 8 lines

Quoted 2 lines

Because there is no reason to return residue of successfully finished transfer. It should be 0.

There is one exception from my point of view. When we are receiving data from peripheral devices, we usually start a DMA transaction with a target length of 4K for example. When a timed-out event occurs in peripheral device, it will notify DMA controller and DMA controller will send out a End of Receive interrupt (Marvell specific?).

Might be your hardware specifics, in our case we have got a timeout interrupt from uart controller.

Quoted 6 lines

Quoted 2 lines

In such situation, DMA status is also DMA_SUCCESS. But the residual bytes is not 0 and the user must query it.

Which sounds wrong approach.

P.S. take a look at drivers/tty/serial/8250/8250_dma.c

When peripheral device (e.g. UART) is using DMA controller, we should have 2 solutions to deal with trailing bytes:

Quoted 3 lines

1. Timeout interrupt from UART controller.
In this case, we usually pause DMA and read out data from DMA buffer in uart irq handler.

Hmm... When UART timeout occurs you have trailing bytes in the UART FIFO. Correct? What do you mean under "DMA buffer" ?
I think you have to read bytes from UART FIFO.

Quoted 7 lines

2. DMA controller handles trailing bytes for us.
This is the case I mentioned in my previous email. "When a timed-out event occurs in peripheral device, it will notify DMA controller and DMA controller will send out a End of Receive interrupt" I think we should know how many residual bytes in this case even the DMA status is DMA_SUCCESS.

I'm still not convinced that you have implement such algorithm. Anyway, if I got it correctly you have something like "Receive FIFO Occupancy Register (FOR)" in UART. When you got timeout interrupt, you may get residue and value from FOR, sum them, program DMA to transfer the trailing bytes.
What did I miss?

--
With Best Regards,
Andy Shevchenko

2013/6/6 Andy Shevchenko <[email protected]>:

Quoted 1 line · Xiang Wang

Quoted 7 lines

Quoted 6 lines

2. DMA controller handles trailing bytes for us.
This is the case I mentioned in my previous email. "When a timed-out event occurs in peripheral device, it will notify DMA controller and DMA controller will send out a End of Receive interrupt" I think we should know how many residual bytes in this case even the DMA status is DMA_SUCCESS.

Quoted 6 lines

I'm still not convinced that you have implement such algorithm. Anyway, if I got it correctly you have something like "Receive FIFO Occupancy Register (FOR)" in UART. When you got timeout interrupt, you may get residue and value from FOR, sum them, program DMA to transfer the trailing bytes.
What did I miss?

This case is another working mode of peripheral device (e.g. UART). If we configure UART to work in this mode, the hardware(DMA controller and UART controller) will deal with the trailing bytes for us. The workflow is likely to be:
a. When timeout occurs in UART, it notifies DMA controller to move the trailing bytes in UART FIFO. (software transparent)
b. After DMA finishes moving all bytes in UART FIFO, it will send out an End Of Receive(EOR) interrupt.
So in this case, software sees an EOR interrupt from DMA controller instead of an Timeout Interrupt from UART controller.

From: Xiang Wang <[email protected]>

In some of our drivers (e.g. UART) we may stop a running DMA before it finishes. So we need to know how many bytes have been transferred.

Signed-off-by Xiang Wang <[email protected]>

---
 drivers/dma/mmp_pdma.c |   77 +++++++++++++++++++++++++++++++++++++++++++-----
 1 files changed, 69 insertions(+), 8 deletions(-)

drivers/dma/mmp_pdma.c +2 -1

index 76a8dcf..0c623cc 100644
@@ -104,7 +104,8 @@ struct mmp_pdma_chan {
 	spinlock_t desc_lock;		/* Descriptor list lock */
 	struct list_head chain_pending;	/* Link descriptors queue for pending */
 	struct list_head chain_running;	/* Link descriptors queue for running */
-	bool idle;			/* channel statue machine */
+	enum dma_status status; /* channel status machine */
+	u32 bytes_residue;
 
 	struct dma_pool *desc_pool;	/* Descriptors pool */
 };

drivers/dma/mmp_pdma.c +8 -0

+	 * stored value for safe.
+	 */
+	if (!chan || !chan->phy)
+		return chan->bytes_residue;
+
+	if (!list_empty(&chan->chain_running))
+		list_chain = &chan->chain_running;
+	else

drivers/dma/mmp_pdma.c +18 -2

+	case DMA_MEM_TO_DEV:
+		reg = (chan->phy->idx << 4) + DSADR;
+		cur_pos = readl_relaxed(chan->phy->base + reg);
+		orig_pos = desc_sw->desc.dsadr;
+		break;
+	case DMA_MEM_TO_MEM:
+		reg = (chan->phy->idx << 4) + DTADR;
+		cur_pos = readl_relaxed(chan->phy->base + reg);
+		orig_pos = desc_sw->desc.dtadr;
+		break;
+	default:
+		return 0;
+	}
+
+	residue = (u32)(desc_sw->desc.dcmd & DCMD_LENGTH) + orig_pos - cur_pos;
+	return residue;
 }
 
 

@@ -381,7 +430,7 @@ static int mmp_pdma_alloc_chan_resources(struct dma_chan *dchan)
 		chan->phy->vchan = NULL;
 		chan->phy = NULL;
 	}
-	chan->idle = true;
+	chan->status = DMA_SUCCESS;
 	chan->dev_addr = 0;
 	return 1;
 }

@@ -409,7 +458,7 @@ static void mmp_pdma_free_chan_resources(struct dma_chan *dchan)
 
 	dma_pool_destroy(chan->desc_pool);
 	chan->desc_pool = NULL;
-	chan->idle = true;
+	chan->status = DMA_SUCCESS;
 	chan->dev_addr = 0;
 	if (chan->phy) {
 		chan->phy->vchan = NULL;

drivers/dma/mmp_pdma.c +9 -2

+	case DMA_PAUSE:
+		disable_chan(chan->phy);
+		chan->status = DMA_PAUSED;
+		chan->bytes_residue = mmp_pdma_get_bytes_residue(chan);
 		break;
 	case DMA_SLAVE_CONFIG:
 		if (cfg->direction == DMA_DEV_TO_MEM) {

@@ -637,7 +692,8 @@ static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan,
 	unsigned long flags;
 
 	spin_lock_irqsave(&chan->desc_lock, flags);
-	ret = dma_cookie_status(dchan, cookie, txstate);
+	ret = chan->status;
+	dma_set_residue(txstate, chan->bytes_residue);
 	spin_unlock_irqrestore(&chan->desc_lock, flags);
 
 	return ret;

@@ -684,6 +740,8 @@ static void dma_do_tasklet(unsigned long data)
 		dev_dbg(chan->dev, "completed_cookie=%d\n", cookie);
 	}
 
+	/* update bytes_residue so that the user can query later on */
+	chan->bytes_residue = mmp_pdma_get_bytes_residue(chan);
 	/*
 	 * move the descriptors to a temporary list so we can drop the lock
 	 * during the entire cleanup operation

@@ -691,7 +749,7 @@ static void dma_do_tasklet(unsigned long data)
 	list_splice_tail_init(&chan->chain_running, &chain_cleanup);
 
 	/* the hardware is now idle and ready for more */
-	chan->idle = true;
+	chan->status = DMA_SUCCESS;
 
 	/* Start any pending transactions automatically */
 	start_pending_queue(chan);

Quoted 28 lines

From: Xiang Wang <[email protected]>

In some of our drivers (e.g. UART) we may stop a running DMA before it finishes. So we need to know how many bytes have been transferred.

Signed-off-by Xiang Wang <[email protected]>

---
 drivers/dma/mmp_pdma.c |   77 +++++++++++++++++++++++++++++++++++++++++++-----
 1 files changed, 69 insertions(+), 8 deletions(-)

drivers/dma/mmp_pdma.c +2 -1

index 76a8dcf..0c623cc 100644
@@ -104,7 +104,8 @@ struct mmp_pdma_chan {
      spinlock_t desc_lock;           /* Descriptor list lock */
      struct list_head chain_pending; /* Link descriptors queue for pending */
      struct list_head chain_running; /* Link descriptors queue for running */
-     bool idle;                      /* channel statue machine */
+     enum dma_status status; /* channel status machine */
+     u32 bytes_residue;
 
      struct dma_pool *desc_pool;     /* Descriptors pool */
 };

@@ -270,7 +271,7 @@ static void start_pending_queue(struct mmp_pdma_chan *chan)
      struct mmp_pdma_desc_sw *desc;
 

drivers/dma/mmp_pdma.c static void start_pending_queue(struct mmp_pdma_chan *chan) · line 275

Quoted 4 lines

drivers/dma/mmp_pdma.c +1 -1

      /* still in running, irq will start the pending list */
-     if (!chan->idle) {
+     if (chan->status == DMA_IN_PROGRESS) {
              dev_dbg(chan->dev, "DMA controller still busy\n");

why are you not reading the residue and filling it here as well

Quoted 4 lines

drivers/dma/mmp_pdma.c +0 -0

              return;
      }

drivers/dma/mmp_pdma.c static void start_pending_queue(struct mmp_pdma_chan *chan) · line 316

Quoted 10 lines

drivers/dma/mmp_pdma.c +6 -1

@@ -307,7 +308,55 @@ static void start_pending_queue(struct mmp_pdma_chan *chan)
       */
      set_desc(chan->phy, desc->async_tx.phys);
      enable_chan(chan->phy);
-     chan->idle = false;
+     chan->status = DMA_IN_PROGRESS;
+     chan->bytes_residue = 0;
+}
+
+/*
+ * Get the number of bytes untransferred by DMA.

                              ^^^^^^^^^^^^^

The word which might fit better is pending

drivers/dma/mmp_pdma.c static void start_pending_queue(struct mmp_pdma_chan *chan) · line 325

Quoted 9 lines

drivers/dma/mmp_pdma.c +9 -0

+ */
+static u32 mmp_pdma_get_bytes_residue(struct mmp_pdma_chan *chan)
+{
+     u32 reg, orig_pos, cur_pos, residue = 0;
+     struct mmp_pdma_desc_sw *desc_sw;
+     struct list_head *list_chain = NULL;
+
+     /*
+      * When a phy channel is unavailable, maybe its destroied, return last

typo ^^^^^^^^^

Quoted 11 lines

drivers/dma/mmp_pdma.c +11 -0

+      * stored value for safe.
+      */
+     if (!chan || !chan->phy)
+             return chan->bytes_residue;
+
+     if (!list_empty(&chan->chain_running))
+             list_chain = &chan->chain_running;
+     else
+             return 0;
+
+     desc_sw = list_first_entry(list_chain, struct mmp_pdma_desc_sw, node);

drivers/dma/mmp_pdma.c static void start_pending_queue(struct mmp_pdma_chan *chan) · line 343

Quoted 7 lines

drivers/dma/mmp_pdma.c +7 -0

+
+     switch (chan->dir) {
+     case DMA_DEV_TO_MEM:
+             reg = (chan->phy->idx << 4) + DTADR;
+             cur_pos = readl_relaxed(chan->phy->base + reg);
+             orig_pos = desc_sw->desc.dtadr;
+             break;

empty line here and below pls

Quoted 31 lines

drivers/dma/mmp_pdma.c +17 -1

+     case DMA_MEM_TO_DEV:
+             reg = (chan->phy->idx << 4) + DSADR;
+             cur_pos = readl_relaxed(chan->phy->base + reg);
+             orig_pos = desc_sw->desc.dsadr;
+             break;
+     case DMA_MEM_TO_MEM:
+             reg = (chan->phy->idx << 4) + DTADR;
+             cur_pos = readl_relaxed(chan->phy->base + reg);
+             orig_pos = desc_sw->desc.dtadr;
+             break;
+     default:
+             return 0;
+     }
+
+     residue = (u32)(desc_sw->desc.dcmd & DCMD_LENGTH) + orig_pos - cur_pos;
+     return residue;
 }
 
 

@@ -381,7 +430,7 @@ static int mmp_pdma_alloc_chan_resources(struct dma_chan *dchan)
              chan->phy->vchan = NULL;
              chan->phy = NULL;
      }
-     chan->idle = true;
+     chan->status = DMA_SUCCESS;
      chan->dev_addr = 0;
      return 1;
 }

@@ -409,7 +458,7 @@ static void mmp_pdma_free_chan_resources(struct dma_chan *dchan)
 

Quoted 8 lines

drivers/dma/mmp_pdma.c +1 -1

      dma_pool_destroy(chan->desc_pool);
      chan->desc_pool = NULL;
-     chan->idle = true;
+     chan->status = DMA_SUCCESS;
      chan->dev_addr = 0;
      if (chan->phy) {
              chan->phy->vchan = NULL;

drivers/dma/mmp_pdma.c static int mmp_pdma_control(struct dma_chan *dchan, enum dma_ctrl_cmd cmd, · line 643

Quoted 7 lines

drivers/dma/mmp_pdma.c +3 -1

@@ -589,7 +638,13 @@ static int mmp_pdma_control(struct dma_chan *dchan, enum dma_ctrl_cmd cmd,
              mmp_pdma_free_desc_list(chan, &chan->chain_pending);
              mmp_pdma_free_desc_list(chan, &chan->chain_running);
              spin_unlock_irqrestore(&chan->desc_lock, flags);
-             chan->idle = true;
+             chan->status = DMA_SUCCESS;
+             chan->bytes_residue = 0;
+             break;

ditto

Quoted 39 lines

drivers/dma/mmp_pdma.c +9 -2

+     case DMA_PAUSE:
+             disable_chan(chan->phy);
+             chan->status = DMA_PAUSED;
+             chan->bytes_residue = mmp_pdma_get_bytes_residue(chan);
              break;
      case DMA_SLAVE_CONFIG:
              if (cfg->direction == DMA_DEV_TO_MEM) {

@@ -637,7 +692,8 @@ static enum dma_status mmp_pdma_tx_status(struct dma_chan *dchan,
      unsigned long flags;
 
      spin_lock_irqsave(&chan->desc_lock, flags);
-     ret = dma_cookie_status(dchan, cookie, txstate);
+     ret = chan->status;
+     dma_set_residue(txstate, chan->bytes_residue);
      spin_unlock_irqrestore(&chan->desc_lock, flags);
 
      return ret;

@@ -684,6 +740,8 @@ static void dma_do_tasklet(unsigned long data)
              dev_dbg(chan->dev, "completed_cookie=%d\n", cookie);
      }
 
+     /* update bytes_residue so that the user can query later on */
+     chan->bytes_residue = mmp_pdma_get_bytes_residue(chan);
      /*
       * move the descriptors to a temporary list so we can drop the lock
       * during the entire cleanup operation

@@ -691,7 +749,7 @@ static void dma_do_tasklet(unsigned long data)
      list_splice_tail_init(&chan->chain_running, &chain_cleanup);
 
      /* the hardware is now idle and ready for more */
-     chan->idle = true;
+     chan->status = DMA_SUCCESS;
 
      /* Start any pending transactions automatically */
      start_pending_queue(chan);

@@ -750,6 +808,9 @@ static int mmp_pdma_chan_init(struct mmp_pdma_device *pdev,
      INIT_LIST_HEAD(&chan->chain_pending);
      INIT_LIST_HEAD(&chan->chain_running);
 

drivers/dma/mmp_pdma.c static int mmp_pdma_chan_init(struct mmp_pdma_device *pdev, · line 816

Quoted 6 lines

drivers/dma/mmp_pdma.c +3 -0

+     chan->status = DMA_SUCCESS;
+     chan->bytes_residue = 0;
+
      /* register virt channel to dma engine */
      list_add_tail(&chan->chan.device_node,
                      &pdev->device.channels);

Okay I went thru the thread briefly. I understand why you are returning the reside and nice usage model you have here.
What i odnt understand is that the reason for limiting the dma driver to a certain usage model. Yes this is what you need a nd what you have done. But nothing prevents you from reading the reside and returning on the fly. All the basic elements are alreayd in this patch. SO lets do the right thing now :)

--
~Vinod