ieee1394: rawiso: requeue packet for transmission after skipped cycle

As it seems, some host controllers have issues that can cause them to
skip cycles now and then when using large packets. I suspect that this
is due to DMA not succeeding in time. If the transmit fifo can't contain
more than one packet (big packets), the DMA should provide a new packet
each cycle (125us). I am under the impression that my current PCI
express test system can't guarantee this.

In any case, the patch tries to provide a workaround as follows:
The DMA program descriptors are modified such that when an error occurs,
the DMA engine retries the descriptor the next cycle instead of
stalling. This way no data is lost. The side effect of this is that
packets are sent with one cycle delay. This however might not be that
much of a problem for certain protocols (e.g. AM824). If they use
padding packets for e.g. rate matching they can drop one of those to
resync the streams.

The amount of skips between two userspace wakeups is counted. This
number is then propagated to userspace through the upper 16 bits of the
'dropped' parameter. This allows unmodified userspace applications due
to the following:
1) libraw simply passes this dropped parameter to the user application
2) the meaning of the dropped parameter is: if it's nonzero, something
bad has happened. The actual value of the parameter at this moment does
not have a specific meaning.

A libraw client can then retrieve the number of skipped cycles and
account for them if needed.

Signed-off-by: Pieter Palmers <pieterp@joow.be>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>

1 files changed, 34 insertions, 0 deletions

diff --git a/drivers/ieee1394/ohci1394.c b/drivers/ieee1394/ohci1394.c
index 0690469fcecf..e509e13cb7a7 100644
--- a/drivers/ieee1394/ohci1394.c
+++ b/drivers/ieee1394/ohci1394.c
@@ -1723,6 +1723,8 @@ struct ohci_iso_xmit {
 	struct dma_prog_region prog;
 	struct ohci1394_iso_tasklet task;
 	int task_active;
+	int last_cycle;
+	atomic_t skips;
 
 	u32 ContextControlSet;
 	u32 ContextControlClear;
@@ -1759,6 +1761,8 @@ static int ohci_iso_xmit_init(struct hpsb_iso *iso)
 	iso->hostdata = xmit;
 	xmit->ohci = iso->host->hostdata;
 	xmit->task_active = 0;
+	xmit->last_cycle = -1;
+	atomic_set(&iso->skips, 0);
 
 	dma_prog_region_init(&xmit->prog);
 
@@ -1856,6 +1860,26 @@ static void ohci_iso_xmit_task(unsigned long data)
 		/* parse cycle */
 		cycle = le32_to_cpu(cmd->output_last.status) & 0x1FFF;
 
+		if (xmit->last_cycle > -1) {
+			int cycle_diff = cycle - xmit->last_cycle;
+			int skip;
+
+			/* unwrap */
+			if (cycle_diff < 0) {
+				cycle_diff += 8000;
+				if (cycle_diff < 0)
+					PRINT(KERN_ERR, "bogus cycle diff %d\n",
+					      cycle_diff);
+			}
+
+			skip = cycle_diff - 1;
+			if (skip > 0) {
+				DBGMSG("skipped %d cycles without packet loss", skip);
+				atomic_add(skip, &iso->skips);
+			}
+		}
+		xmit->last_cycle = cycle;
+
 		/* tell the subsystem the packet has gone out */
 		hpsb_iso_packet_sent(iso, cycle, event != 0x11);
 
@@ -1943,6 +1967,16 @@ static int ohci_iso_xmit_queue(struct hpsb_iso *iso, struct hpsb_iso_packet_info
 	prev->output_last.branchAddress = cpu_to_le32(
 		dma_prog_region_offset_to_bus(&xmit->prog, sizeof(struct iso_xmit_cmd) * next_i) | 3);
 
+	/*
+	 * Link the skip address to this descriptor itself. This causes a
+	 * context to skip a cycle whenever lost cycles or FIFO overruns occur,
+	 * without dropping the data at that point the application should then
+	 * decide whether this is an error condition or not. Some protocols
+	 * can deal with this by dropping some rate-matching padding packets.
+	 */
+	next->output_more_immediate.branchAddress =
+			prev->output_last.branchAddress;
+
 	/* disable interrupt, unless required by the IRQ interval */
 	if (prev_i % iso->irq_interval) {
 		prev->output_last.control &= cpu_to_le32(~(3 << 20)); /* no interrupt */