Fran?ois,

Thanks for your detailed review. I have a few comments/questions below
on specific topics.

On Sat, 3 Nov 2012 12:53:21 +0100, Francois Romieu wrote:

quoted

+	if (rxq->descs == NULL) {
+		netdev_err(pp->dev,
+			   "rxQ=%d: Can't allocate %d bytes for %d RX descr\n",
+			   rxq->id, rxq->size * MVNETA_DESC_ALIGNED_SIZE,
+			   rxq->size);
+		return -ENOMEM;
+	}
+
+	BUG_ON(rxq->descs !=
+	       PTR_ALIGN(rxq->descs, MVNETA_CPU_D_CACHE_LINE_SIZE));

There is no reason to crash.

Well, there is a reason: the hardware will not work properly if
rxq->descs is not aligned on a MVNETA_CPU_D_CACHE_LINE_SIZE boundary.
So one solution is to over-allocated to guarantee the alignment, but
since practically speaking MVNETA_CPU_D_CACHE_LINE_SIZE=32 and
dma_alloc_coherent() returns things that seem at least 32 bytes
aligned, it sounded overkill to include more code to fix a problem that
doesn't exist. This BUG_ON() is here solely for the purpose of noisily
letting the user know if this implicit assumption on the alignment of
dma_alloc_coherent() allocated buffer changes in the future. I can turn
this into an error if you prefer:

	if (rxq->descs != PTR_ALIGN(rxq->descs, MVNETA_CPU_D_CACHE_LINE_SIZE)) {
		netdev_err(pp->dev, "improper buffer alignement assumption, driver needs fixing\n");
		dma_free_coherent(...);
		return -EINVAL;
	}

(...]

quoted

+static int mvneta_txq_init(struct mvneta_port *pp,
+			   struct mvneta_tx_queue *txq)
+{
+	txq->size = pp->tx_ring_size;
+
+	/* Allocate memory for TX descriptors */
+	txq->descs = dma_alloc_coherent(pp->dev->dev.parent,
+					txq->size * MVNETA_DESC_ALIGNED_SIZE,
+					&txq->descs_phys,
+					DMA_BIDIRECTIONAL);

					&txq->descs_phys, DMA_BIDIRECTIONAL);

Aaah, thanks for pointing this one! It should have been GFP_KERNEL, not
DMA_BIDIRECTIONAL here.

quoted

+	if (txq->descs == NULL) {
+		netdev_err(pp->dev,
+			   "txQ=%d: Can't allocate %d bytes for %d TX descr\n",
+			   txq->id, txq->size * MVNETA_DESC_ALIGNED_SIZE,
+			   txq->size);
+		return -ENOMEM;
+	}
+
+	/* Make sure descriptor address is cache line size aligned  */
+	BUG_ON(txq->descs !=
+	       PTR_ALIGN(txq->descs, MVNETA_CPU_D_CACHE_LINE_SIZE));

There is no reason to crash.

Same as above :-)

[...]

quoted

+static int mvneta_setup_rxqs(struct mvneta_port *pp)
+{
+	int queue;
+
+	for (queue = 0; queue < rxq_number; queue++) {
+		int err = mvneta_rxq_init(pp, &pp->rxqs[queue]);
+		if (err) {
+			netdev_err(pp->dev,
+				   "%s: can't create RxQ rxq=%d\n",

			netdev_err(pp->dev, "%s: can't create RxQ rxq=%d\n",

quoted

+				   __func__, queue);
+			mvneta_cleanup_rxqs(pp);
+			return -ENODEV;

mvneta_rxq_init should return a proper error code and it should be
propagated (option: break instead of return and mvneta_setup_rxqs scoped
err variable)

Besides turning the "return -ENODEV;" into "return err;", I don't see
what is the other problem here? mvneta_rxq_init() properly returns
-ENOMEM where there is a memory allocation failure, and
mvneta_cleanup_rxqs() properly cleans up *all* initialized rxqs. Is
there something I'm missing?

Thanks again for your review!

Thomas
-- 
Thomas Petazzoni, Free Electrons
Kernel, drivers, real-time and embedded Linux
development, consulting, training and support.
http://free-electrons.com