On Mon, 19 Nov 2012 05:22:25 +0000 Dan Williams [off-list ref] wrote:

quoted

On 11/18/12 5:06 PM, "NeilBrown" [off-list ref] wrote:

quoted

Hi Dan,
could you comment on this please?  Would it make sense to arrange for
errors
to propagate up?  Or should we arrange to do a software-fallback in the
dma
engine is a problem?  What sort of things can cause error here anyway?

Propagating up is missing reliable "dma abort" operation.

In these cases the engine failed to complete due to hardware hang / driver
bug, or has hit a memory error (uncorrectable even with software
fallback).  This originally should have been using async_tx_quiesce()
which also does the panic.

The engines that I have worked with have either lacked support for
aborting, or were otherwise unable to recover from a hardware hang.
However, for engines that do support error recovery they should be able to
hide the failure from the upper layers.

So maybe I could:

diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
index ac09fa4..ffbf0ca 100644
--- a/drivers/md/raid5.c
+++ b/drivers/md/raid5.c

@@ -3268,7 +3268,7 @@ static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
 	/* done submitting copies, wait for them to complete */
 	if (tx) {
 		async_tx_ack(tx);
-		dma_wait_for_async_tx(tx);
+		async_tx_quiesce(&tx);
 	}
 }
 

and then the panic would be somebody else's problem?

I note that handle_stripe_expansion has:

 		async_tx_ack(tx);
		dma_wait_for_async_tx(tx);

while async_tx_quiesce() has:

		if (dma_wait_for_async_tx(*tx) == DMA_ERROR)
			panic("DMA_ERROR waiting for transaction\n");
		async_tx_ack(*tx);


i.e. the same two functions called in the reverse order.  Is the order
important?  Is handle_stripe_expansion wrong?   Should the patch I apply
actually be:

diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
index ac09fa4..e51d903 100644
--- a/drivers/md/raid5.c
+++ b/drivers/md/raid5.c

@@ -3266,10 +3266,7 @@ static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
 
 		}
 	/* done submitting copies, wait for them to complete */
-	if (tx) {
-		async_tx_ack(tx);
-		dma_wait_for_async_tx(tx);
-	}
+	async_tx_quiesce(&tx);
 }