dpaa_eth: undo merge on master branch

One of the merged patches was no longer needed and the rest
break a large set of patches already prepared for review.
The required patches are available on top of the large patch set.

Revert "dpaa_eth: Handle buffer pool depletion"
This reverts commit 1fd9018f9c6a10932e86285c256bfdf50da8c0c1.
Revert "dpaa_eth: Fix pool counters on error path"
This reverts commit e8a66453a9b71d4d5290bd4962cc6ec2fda1b848.
Revert "dpaa_eth: Switch compile-time default optimization"
This reverts commit 47a79e454a7f54b96b3133e139e707944375edca.
Revert "fmd,dpaa_eth: Selectively reenable FMan allocation algorithm"
This reverts commit 3cc98fd0b85bdb61f83672fdc4609b68f0c7a75f.
Revert "dpaa_eth: Fix endless loop in case of memory depletion"
This reverts commit b676b243d7ccdd4879ea8c5fb74a6d3e2a7f2e5e.
Revert "dpaa_eth: Set a fixed upper limit for recycleable buffer size"
This reverts commit 252b622b8058a309d2429a33c2e0de089be896ac.
Revert "dpaa_eth: Fix computation of cache padding"
This reverts commit a3b310dbe6d25132698a88618e7ef52bc6749177.
Revert "dpaa_eth: Fix Tx/TxConfirm initialization bug"
This reverts commit 429d9f28709eb7c916d4c6da4cccfae33906c510.

Change-Id: I22765f8406f43bb2c7d84fbf5fdb67c74697a008
Reviewed-on: http://git.am.freescale.net:8181/3305
Tested-by: Review Code-CDREVIEW <CDREVIEW@freescale.com>
Reviewed-by: Fleming Andrew-AFLEMING <AFLEMING@freescale.com>

4 files changed, 41 insertions, 104 deletions

diff --git a/drivers/net/ethernet/freescale/dpa/Kconfig b/drivers/net/ethernet/freescale/dpa/Kconfig
index aa0d1e6..801e8b2 100644
--- a/drivers/net/ethernet/freescale/dpa/Kconfig
+++ b/drivers/net/ethernet/freescale/dpa/Kconfig
@@ -28,7 +28,7 @@ config FSL_DPAA_ETH_SG_SUPPORT
 choice FSL_DPAA_ETH_OPTIMIZE
 	prompt "Optimization choices for the DPAA Ethernet driver"
 	depends on FSL_DPAA_ETH
-	default FSL_DPAA_ETH_OPTIMIZE_FOR_IPFWD
+	default FSL_DPAA_ETH_OPTIMIZE_FOR_TERM
 	---help---
 	  Compile-time switch between driver optimizations for forwarding use-cases and
 	  termination scenarios.
@@ -36,7 +36,6 @@ choice FSL_DPAA_ETH_OPTIMIZE
 	config FSL_DPAA_ETH_OPTIMIZE_FOR_IPFWD
 		bool "Optimize for forwarding"
 		select FSL_DPAA_TX_RECYCLE if FMAN_T4240
-		select FMAN_RESOURCE_ALLOCATION_ALGORITHM if (FMAN_P3040_P4080_P5020 || FMAN_P1023)
 		---help---
 		  Optimize the DPAA-Ethernet driver for IP/IPSec forwarding use-cases.
 		  This option disabled SG support in the DPAA Ethernet driver.
diff --git a/drivers/net/ethernet/freescale/dpa/dpaa_eth.c b/drivers/net/ethernet/freescale/dpa/dpaa_eth.c
index 55edca4..2d47c8a 100644
--- a/drivers/net/ethernet/freescale/dpa/dpaa_eth.c
+++ b/drivers/net/ethernet/freescale/dpa/dpaa_eth.c
@@ -130,17 +130,13 @@
 #define DPA_MAX_FD_OFFSET	((1 << 9) - 1)
 
 /*
- * Maximum size of a buffer for which recycling is allowed.
- * We need an upper limit such that forwarded skbs that get reallocated on Tx
- * aren't allowed to grow unboundedly. On the other hand, we need to make sure
- * that skbs allocated by us will not fail to be recycled due to their size.
- *
- * For a requested size, the kernel allocator provides the next power of two
- * sized block, which the stack will use as is, regardless of the actual size
- * it required; since we must acommodate at most 9.6K buffers (L2 maximum
- * supported frame size), set the recycling upper limit to 16K.
+ * Extra size of a buffer (beyond the size of the buffers that are seeded into
+ * the global pool) for which recycling is allowed.
+ * The value is arbitrary, but tries to reach a balance such that originating
+ * frames may get recycled, while forwarded skbs that get reallocated on Tx
+ * aren't allowed to grow unboundedly.
  */
-#define DPA_RECYCLE_MAX_SIZE	16384
+#define DPA_RECYCLE_EXTRA_SIZE	1024
 
 /* For MAC-based interfaces, we compute the tx needed headroom from the
  * associated Tx port's buffer layout settings.
@@ -390,14 +386,12 @@ static void dpaa_eth_seed_pool(struct dpa_bp *bp)
  * Add buffers/pages/skbuffs for Rx processing whenever bpool count falls below
  * REFILL_THRESHOLD.
  */
-static int dpaa_eth_refill_bpools(struct dpa_percpu_priv_s *percpu_priv)
+static void dpaa_eth_refill_bpools(struct dpa_percpu_priv_s *percpu_priv)
 {
 	int *countptr = percpu_priv->dpa_bp_count;
 	int count = *countptr;
 	const struct dpa_bp *dpa_bp = percpu_priv->dpa_bp;
-	int new_pages __maybe_unused;
 #ifndef CONFIG_FSL_DPAA_ETH_SG_SUPPORT
-
 	/* this function is called in softirq context;
 	 * no need to protect smp_processor_id() on RT kernel
 	 */
@@ -411,24 +405,10 @@ static int dpaa_eth_refill_bpools(struct dpa_percpu_priv_s *percpu_priv)
 	}
 #else
 	/* Add pages to the buffer pool */
-	while (count < CONFIG_FSL_DPAA_ETH_MAX_BUF_COUNT) {
-		new_pages = _dpa_bp_add_8_pages(dpa_bp);
-		if (unlikely(!new_pages)) {
-			/* Avoid looping forever if we've temporarily
-			 * run out of memory. We'll try again at the next
-			 * NAPI cycle.
-			 */
-			break;
-		}
-		count += new_pages;
-	}
+	while (count < CONFIG_FSL_DPAA_ETH_MAX_BUF_COUNT)
+		count += _dpa_bp_add_8_pages(dpa_bp);
 	*countptr = count;
-
-	if (*countptr < CONFIG_FSL_DPAA_ETH_MAX_BUF_COUNT)
-		return -ENOMEM;
 #endif
-
-	return 0;
 }
 
 static int dpa_make_shared_port_pool(struct dpa_bp *bp)
@@ -1291,9 +1271,10 @@ static int dpa_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
 static int dpa_process_one(struct dpa_percpu_priv_s *percpu_priv,
 		struct sk_buff *skb, struct dpa_bp *bp, const struct qm_fd *fd)
 {
-	dma_addr_t fd_addr = qm_fd_addr(fd);
-	unsigned long skb_addr = virt_to_phys(skb->head);
-	u32 pad = fd_addr - skb_addr;
+	dma_addr_t addr = qm_fd_addr(fd);
+	u32 addrlo = lower_32_bits(addr);
+	u32 skblo = lower_32_bits((unsigned long)skb->head);
+	u32 pad = (addrlo - skblo) & (PAGE_SIZE - 1);
 	unsigned int data_start;
 
 	(*percpu_priv->dpa_bp_count)--;
@@ -1964,7 +1945,8 @@ static int skb_to_contig_fd(struct dpa_priv_s *priv,
 	 * - there's enough room in the buffer pool
 	 */
 	if (likely(skb_is_recycleable(skb, dpa_bp->size) &&
-		   (skb_end_pointer(skb) - skb->head <= DPA_RECYCLE_MAX_SIZE) &&
+		   (skb_end_pointer(skb) - skb->head <=
+			dpa_bp->size + DPA_RECYCLE_EXTRA_SIZE) &&
 		   (*percpu_priv->dpa_bp_count < dpa_bp->target_count))) {
 		/* Compute the minimum necessary fd offset */
 		offset = dpa_bp->size - skb->len - skb_tailroom(skb);
@@ -2199,7 +2181,6 @@ ingress_rx_error_dqrr(struct qman_portal		*portal,
 	struct net_device		*net_dev;
 	struct dpa_priv_s		*priv;
 	struct dpa_percpu_priv_s	*percpu_priv;
-	int err;
 
 	net_dev = ((struct dpa_fq *)fq)->net_dev;
 	priv = netdev_priv(net_dev);
@@ -2211,16 +2192,7 @@ ingress_rx_error_dqrr(struct qman_portal		*portal,
 		return qman_cb_dqrr_stop;
 	}
 
-	err = dpaa_eth_refill_bpools(percpu_priv);
-	if (err) {
-		/* Unable to refill the buffer pool due to insufficient
-		 * system memory. Just release the frame back into the pool,
-		 * otherwise we'll soon end up with an empty buffer pool.
-		 */
-		dpa_fd_release(net_dev, &dq->fd);
-		return qman_cb_dqrr_consume;
-	}
-
+	dpaa_eth_refill_bpools(percpu_priv);
 	_dpa_rx_error(net_dev, priv, percpu_priv, &dq->fd, fq->fqid);
 
 	return qman_cb_dqrr_consume;
@@ -2368,7 +2340,6 @@ ingress_rx_default_dqrr(struct qman_portal		*portal,
 	struct net_device		*net_dev;
 	struct dpa_priv_s		*priv;
 	struct dpa_percpu_priv_s	*percpu_priv;
-	int err;
 
 	net_dev = ((struct dpa_fq *)fq)->net_dev;
 	priv = netdev_priv(net_dev);
@@ -2385,16 +2356,7 @@ ingress_rx_default_dqrr(struct qman_portal		*portal,
 	}
 
 	/* Vale of plenty: make sure we didn't run out of buffers */
-	err = dpaa_eth_refill_bpools(percpu_priv);
-	if (err) {
-		/* Unable to refill the buffer pool due to insufficient
-		 * system memory. Just release the frame back into the pool,
-		 * otherwise we'll soon end up with an empty buffer pool.
-		 */
-		dpa_fd_release(net_dev, &dq->fd);
-		return qman_cb_dqrr_consume;
-	}
-
+	dpaa_eth_refill_bpools(percpu_priv);
 	_dpa_rx(net_dev, priv, percpu_priv, &dq->fd, fq->fqid);
 
 	return qman_cb_dqrr_consume;
@@ -3591,16 +3553,12 @@ static void dpa_setup_egress(struct dpa_priv_s *priv,
 	/* Allocate frame queues to all available CPUs no matter the number of
 	 * queues specified in device tree.
 	 */
-	for (i = 0, ptr = &fq->list; i < DPAA_ETH_TX_QUEUES; i++) {
+	for (i = 0; i < DPAA_ETH_TX_QUEUES; i++) {
 		iter = list_entry(ptr, struct dpa_fq, list);
 		priv->egress_fqs[i] = &iter->fq_base;
 
-		if (list_is_last(ptr, head)) {
+		if (list_is_last(ptr, head))
 			ptr = &fq->list;
-			continue;
-		}
-
-		ptr = ptr->next;
 	}
 }
 
diff --git a/drivers/net/ethernet/freescale/dpa/dpaa_eth_sg.c b/drivers/net/ethernet/freescale/dpa/dpaa_eth_sg.c
index f93c346..0f0f726 100644
--- a/drivers/net/ethernet/freescale/dpa/dpaa_eth_sg.c
+++ b/drivers/net/ethernet/freescale/dpa/dpaa_eth_sg.c
@@ -45,6 +45,22 @@
 #ifdef CONFIG_FSL_DPAA_ETH_SG_SUPPORT
 #define DPA_SGT_MAX_ENTRIES 16 /* maximum number of entries in SG Table */
 
+/*
+ * It does not return a page as you get the page from the fd,
+ * this is only for refcounting and DMA unmapping
+ */
+static inline void dpa_bp_removed_one_page(struct dpa_bp *dpa_bp,
+					   dma_addr_t dma_addr)
+{
+	int *count_ptr;
+
+	count_ptr = __this_cpu_ptr(dpa_bp->percpu_count);
+	(*count_ptr)--;
+
+	dma_unmap_single(dpa_bp->dev, dma_addr, dpa_bp->size,
+		DMA_BIDIRECTIONAL);
+}
+
 /* DMA map and add a page into the bpool */
 static void dpa_bp_add_page(struct dpa_bp *dpa_bp, unsigned long vaddr)
 {
@@ -112,9 +128,7 @@ release_bufs:
 	return i;
 
 bail_out:
-	net_err_ratelimited("dpa_bp_add_8_pages() failed\n");
-	WARN_ONCE(1, "Memory allocation failure on Rx\n");
-
+	dev_err(dpa_bp->dev, "dpa_bp_add_8_pages() failed\n");
 	bm_buffer_set64(&bmb[i], 0);
 	/*
 	 * Avoid releasing a completely null buffer; bman_release() requires
@@ -315,7 +329,6 @@ static struct sk_buff *__hot sg_fd_to_skb(const struct dpa_priv_s *priv,
 	int i;
 	const t_FmPrsResult *parse_results;
 	struct sk_buff *skb = NULL;
-	int *count_ptr;
 
 	vaddr = phys_to_virt(addr);
 #ifdef CONFIG_FSL_DPAA_1588
@@ -336,33 +349,21 @@ static struct sk_buff *__hot sg_fd_to_skb(const struct dpa_priv_s *priv,
 
 		dpa_bp = dpa_bpid2pool(sgt[i].bpid);
 		BUG_ON(IS_ERR(dpa_bp));
-		count_ptr = __this_cpu_ptr(dpa_bp->percpu_count);
 
 		sg_addr = qm_sg_addr(&sgt[i]);
 		sg_vaddr = phys_to_virt(sg_addr);
 
-		if (i == 0) {
-			/* Tentatively access the first buffer, but don't unmap
-			 * it until we're certain the skb allocation succeeds.
-			 */
-			dma_sync_single_for_cpu(dpa_bp->dev, sg_addr,
-				dpa_bp->size, DMA_BIDIRECTIONAL);
+		dpa_bp_removed_one_page(dpa_bp, sg_addr);
 
+		if (i == 0) {
 			/* This is the first S/G entry, so build the skb
 			 * around its data buffer
 			 */
 			skb = build_skb(sg_vaddr,
 					dpa_bp->size + DPA_SKB_TAILROOM);
 			if (unlikely(!skb))
-				/* dpa_fd_release() will put the current frame
-				 * back into the pool. DMA mapping status has
-				 * not changed, nor have the pool counts.
-				 */
 				return NULL;
 
-			dma_unmap_single(dpa_bp->dev, sg_addr, dpa_bp->size,
-				DMA_BIDIRECTIONAL);
-
 			/* In the case of a SG frame, FMan stores the Internal
 			 * Context in the buffer containing the sgt.
 			 * Inspect the parse results before anything else.
@@ -379,8 +380,6 @@ static struct sk_buff *__hot sg_fd_to_skb(const struct dpa_priv_s *priv,
 				dpa_get_rx_extra_headroom());
 			skb_put(skb, sgt[i].length);
 		} else {
-			dma_unmap_single(dpa_bp->dev, sg_addr, dpa_bp->size,
-				DMA_BIDIRECTIONAL);
 			/*
 			 * Not the first S/G entry; all data from buffer will
 			 * be added in an skb fragment; fragment index is offset
@@ -397,8 +396,6 @@ static struct sk_buff *__hot sg_fd_to_skb(const struct dpa_priv_s *priv,
 			skb_add_rx_frag(skb, i - 1, page, frag_offset, frag_len,
 					frag_len);
 		}
-		/* Update the pool count for the current {cpu x bpool} */
-		(*count_ptr)--;
 
 		if (sgt[i].final)
 			break;
@@ -424,7 +421,6 @@ void __hot _dpa_rx(struct net_device *net_dev,
 	unsigned int skb_len;
 	struct rtnl_link_stats64 *percpu_stats = &percpu_priv->stats;
 	int use_gro = net_dev->features & NETIF_F_GRO;
-	int *count_ptr;
 
 	if (unlikely(fd_status & FM_FD_STAT_ERRORS) != 0) {
 		if (netif_msg_hw(priv) && net_ratelimit())
@@ -436,13 +432,7 @@ void __hot _dpa_rx(struct net_device *net_dev,
 	}
 
 	dpa_bp = dpa_bpid2pool(fd->bpid);
-	count_ptr = __this_cpu_ptr(dpa_bp->percpu_count);
-	/* Prepare to read from the buffer, but don't unmap it until
-	 * we know the skb allocation succeeded. At this point we already
-	 * own the buffer - i.e. FMan won't access it anymore.
-	 */
-	dma_sync_single_for_cpu(dpa_bp->dev, addr, dpa_bp->size,
-		DMA_BIDIRECTIONAL);
+	dpa_bp_removed_one_page(dpa_bp, addr);
 
 	/* prefetch the first 64 bytes of the frame or the SGT start */
 	prefetch(phys_to_virt(addr) + dpa_fd_offset(fd));
@@ -455,18 +445,8 @@ void __hot _dpa_rx(struct net_device *net_dev,
 		/* The only FD types that we may receive are contig and S/G */
 		BUG();
 	if (unlikely(!skb))
-		/* We haven't yet touched the DMA mapping or the pool count;
-		 * dpa_fd_release() will just put the buffer back in the pool
-		 */
 		goto _release_frame;
 
-	/* Account for either the contig buffer or the SGT buffer (depending on
-	 * which case we were in) having been removed from the pool.
-	 * Also, permanently unmap the buffer.
-	 */
-	(*count_ptr)--;
-	dma_unmap_single(dpa_bp->dev, addr, dpa_bp->size, DMA_BIDIRECTIONAL);
-
 	skb->protocol = eth_type_trans(skb, net_dev);
 
 	/* IP Reassembled frames are allowed to be larger than MTU */
diff --git a/drivers/net/ethernet/freescale/fman/Kconfig b/drivers/net/ethernet/freescale/fman/Kconfig
index ec7449f..a6d52e5 100644
--- a/drivers/net/ethernet/freescale/fman/Kconfig
+++ b/drivers/net/ethernet/freescale/fman/Kconfig
@@ -34,7 +34,7 @@ endmenu
 
 config FMAN_RESOURCE_ALLOCATION_ALGORITHM
 	bool "Enable FMan dynamic resource allocation algorithm"
-	default y if ((FMAN_P3040_P4080_P5020 || FMAN_P1023) && FSL_DPAA_ETH_OPTIMIZE_FOR_IPFWD)
+	default n
 	---help---
 		Enables algorithm for dynamic resource allocation