Merge branch 'Avoid-local_irq_save-and-use-napi_alloc_frag-where-possible'

Sebastian Andrzej says:

====================
Avoid local_irq_save() and use napi_alloc_frag() where possible

The first two patches remove local_irq_save() around
`netdev_alloc_cache' which does not work on -RT. Besides helping -RT it
whould benefit the users of the function since they can avoid disabling
interrupts and save a few cycles.
The remaining patches are from a time when I tried to remove
`netdev_alloc_cache' but then noticed that we still have non-NAPI
drivers using netdev_alloc_skb() and I dropped that idea. Using
napi_alloc_frag() over netdev_alloc_frag() would skip the not required
local_bh_disable() around the allocation.

v1…v2:
  - 1/7 + 2/7 use now "(in_irq() || irqs_disabled())" instead just
    "irqs_disabled()" to align with __dev_kfree_skb_any(). Pointed out
    by Eric Dumazet.

  - 6/7 has a typo less. Pointed out by Sergei Shtylyov.

  - 3/7 + 4/7 added acks from Ioana Radulescu.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.c

@@ -684,7 +684,7 @@ static void *bnx2x_frag_alloc(const struct bnx2x_fastpath *fp, gfp_t gfp_mask)
 		if (unlikely(gfpflags_allow_blocking(gfp_mask)))
 			return (void *)__get_free_page(gfp_mask);
 
-		return netdev_alloc_frag(fp->rx_frag_size);
+		return napi_alloc_frag(fp->rx_frag_size);
 	}
 
 	return kmalloc(fp->rx_buf_size + NET_SKB_PAD, gfp_mask);

drivers/net/ethernet/broadcom/tg3.c

@@ -6710,7 +6710,7 @@ static int tg3_alloc_rx_data(struct tg3 *tp, struct tg3_rx_prodring_set *tpr,
 	skb_size = SKB_DATA_ALIGN(data_size + TG3_RX_OFFSET(tp)) +
 		   SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 	if (skb_size <= PAGE_SIZE) {
-		data = netdev_alloc_frag(skb_size);
+		data = napi_alloc_frag(skb_size);
 		*frag_size = skb_size;
 	} else {
 		data = kmalloc(skb_size, GFP_ATOMIC);

drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c

@@ -555,7 +555,7 @@ static int build_sg_fd(struct dpaa2_eth_priv *priv,
 	/* Prepare the HW SGT structure */
 	sgt_buf_size = priv->tx_data_offset +
 		       sizeof(struct dpaa2_sg_entry) *  num_dma_bufs;
-	sgt_buf = netdev_alloc_frag(sgt_buf_size + DPAA2_ETH_TX_BUF_ALIGN);
+	sgt_buf = napi_alloc_frag(sgt_buf_size + DPAA2_ETH_TX_BUF_ALIGN);
 	if (unlikely(!sgt_buf)) {
 		err = -ENOMEM;
 		goto sgt_buf_alloc_failed;
@@ -997,13 +997,6 @@ static int seed_pool(struct dpaa2_eth_priv *priv, u16 bpid)
 	int i, j;
 	int new_count;
 
-	/* This is the lazy seeding of Rx buffer pools.
-	 * dpaa2_add_bufs() is also used on the Rx hotpath and calls
-	 * napi_alloc_frag(). The trouble with that is that it in turn ends up
-	 * calling this_cpu_ptr(), which mandates execution in atomic context.
-	 * Rather than splitting up the code, do a one-off preempt disable.
-	 */
-	preempt_disable();
 	for (j = 0; j < priv->num_channels; j++) {
 		for (i = 0; i < DPAA2_ETH_NUM_BUFS;
 		     i += DPAA2_ETH_BUFS_PER_CMD) {
@@ -1011,12 +1004,10 @@ static int seed_pool(struct dpaa2_eth_priv *priv, u16 bpid)
 			priv->channel[j]->buf_count += new_count;
 
 			if (new_count < DPAA2_ETH_BUFS_PER_CMD) {
-				preempt_enable();
 				return -ENOMEM;
 			}
 		}
 	}
-	preempt_enable();
 
 	return 0;
 }

drivers/net/ethernet/marvell/mvneta.c

@@ -1119,7 +1119,7 @@ static void mvneta_bm_update_mtu(struct mvneta_port *pp, int mtu)
 			SKB_DATA_ALIGN(MVNETA_RX_BUF_SIZE(bm_pool->pkt_size));
 
 	/* Fill entire long pool */
-	num = hwbm_pool_add(hwbm_pool, hwbm_pool->size, GFP_ATOMIC);
+	num = hwbm_pool_add(hwbm_pool, hwbm_pool->size);
 	if (num != hwbm_pool->size) {
 		WARN(1, "pool %d: %d of %d allocated\n",
 		     bm_pool->id, num, hwbm_pool->size);

drivers/net/ethernet/marvell/mvneta_bm.c

@@ -190,7 +190,7 @@ struct mvneta_bm_pool *mvneta_bm_pool_use(struct mvneta_bm *priv, u8 pool_id,
 			SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 		hwbm_pool->construct = mvneta_bm_construct;
 		hwbm_pool->priv = new_pool;
-		spin_lock_init(&hwbm_pool->lock);
+		mutex_init(&hwbm_pool->buf_lock);
 
 		/* Create new pool */
 		err = mvneta_bm_pool_create(priv, new_pool);
@@ -201,7 +201,7 @@ struct mvneta_bm_pool *mvneta_bm_pool_use(struct mvneta_bm *priv, u8 pool_id,
 		}
 
 		/* Allocate buffers for this pool */
-		num = hwbm_pool_add(hwbm_pool, hwbm_pool->size, GFP_ATOMIC);
+		num = hwbm_pool_add(hwbm_pool, hwbm_pool->size);
 		if (num != hwbm_pool->size) {
 			WARN(1, "pool %d: %d of %d allocated\n",
 			     new_pool->id, num, hwbm_pool->size);

include/net/hwbm.h

@@ -12,18 +12,18 @@ struct hwbm_pool {
 	/* constructor called during alocation */
 	int (*construct)(struct hwbm_pool *bm_pool, void *buf);
 	/* protect acces to the buffer counter*/
-	spinlock_t lock;
+	struct mutex buf_lock;
 	/* private data */
 	void *priv;
 };
 #ifdef CONFIG_HWBM
 void hwbm_buf_free(struct hwbm_pool *bm_pool, void *buf);
 int hwbm_pool_refill(struct hwbm_pool *bm_pool, gfp_t gfp);
-int hwbm_pool_add(struct hwbm_pool *bm_pool, unsigned int buf_num, gfp_t gfp);
+int hwbm_pool_add(struct hwbm_pool *bm_pool, unsigned int buf_num);
 #else
 void hwbm_buf_free(struct hwbm_pool *bm_pool, void *buf) {}
 int hwbm_pool_refill(struct hwbm_pool *bm_pool, gfp_t gfp) { return 0; }
-int hwbm_pool_add(struct hwbm_pool *bm_pool, unsigned int buf_num, gfp_t gfp)
+int hwbm_pool_add(struct hwbm_pool *bm_pool, unsigned int buf_num)
 { return 0; }
 #endif /* CONFIG_HWBM */
 #endif /* _HWBM_H */

net/core/hwbm.c

@@ -43,34 +43,33 @@ int hwbm_pool_refill(struct hwbm_pool *bm_pool, gfp_t gfp)
 }
 EXPORT_SYMBOL_GPL(hwbm_pool_refill);
 
-int hwbm_pool_add(struct hwbm_pool *bm_pool, unsigned int buf_num, gfp_t gfp)
+int hwbm_pool_add(struct hwbm_pool *bm_pool, unsigned int buf_num)
 {
 	int err, i;
-	unsigned long flags;
 
-	spin_lock_irqsave(&bm_pool->lock, flags);
+	mutex_lock(&bm_pool->buf_lock);
 	if (bm_pool->buf_num == bm_pool->size) {
 		pr_warn("pool already filled\n");
-		spin_unlock_irqrestore(&bm_pool->lock, flags);
+		mutex_unlock(&bm_pool->buf_lock);
 		return bm_pool->buf_num;
 	}
 
 	if (buf_num + bm_pool->buf_num > bm_pool->size) {
 		pr_warn("cannot allocate %d buffers for pool\n",
 			buf_num);
-		spin_unlock_irqrestore(&bm_pool->lock, flags);
+		mutex_unlock(&bm_pool->buf_lock);
 		return 0;
 	}
 
 	if ((buf_num + bm_pool->buf_num) < bm_pool->buf_num) {
 		pr_warn("Adding %d buffers to the %d current buffers will overflow\n",
 			buf_num,  bm_pool->buf_num);
-		spin_unlock_irqrestore(&bm_pool->lock, flags);
+		mutex_unlock(&bm_pool->buf_lock);
 		return 0;
 	}
 
 	for (i = 0; i < buf_num; i++) {
-		err = hwbm_pool_refill(bm_pool, gfp);
+		err = hwbm_pool_refill(bm_pool, GFP_KERNEL);
 		if (err < 0)
 			break;
 	}
@@ -79,7 +78,7 @@ int hwbm_pool_add(struct hwbm_pool *bm_pool, unsigned int buf_num, gfp_t gfp)
 	bm_pool->buf_num += i;
 
 	pr_debug("hwpm pool: %d of %d buffers added\n", i, buf_num);
-	spin_unlock_irqrestore(&bm_pool->lock, flags);
+	mutex_unlock(&bm_pool->buf_lock);
 
 	return i;
 }

net/core/skbuff.c

@@ -366,19 +366,21 @@ struct napi_alloc_cache {
 static DEFINE_PER_CPU(struct page_frag_cache, netdev_alloc_cache);
 static DEFINE_PER_CPU(struct napi_alloc_cache, napi_alloc_cache);
 
-static void *__netdev_alloc_frag(unsigned int fragsz, gfp_t gfp_mask)
+static void *__napi_alloc_frag(unsigned int fragsz, gfp_t gfp_mask)
 {
-	struct page_frag_cache *nc;
-	unsigned long flags;
-	void *data;
+	struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
 
-	local_irq_save(flags);
-	nc = this_cpu_ptr(&netdev_alloc_cache);
-	data = page_frag_alloc(nc, fragsz, gfp_mask);
-	local_irq_restore(flags);
-	return data;
+	return page_frag_alloc(&nc->page, fragsz, gfp_mask);
 }
 
+void *napi_alloc_frag(unsigned int fragsz)
+{
+	fragsz = SKB_DATA_ALIGN(fragsz);
+
+	return __napi_alloc_frag(fragsz, GFP_ATOMIC);
+}
+EXPORT_SYMBOL(napi_alloc_frag);
+
 /**
  * netdev_alloc_frag - allocate a page fragment
  * @fragsz: fragment size
@@ -388,26 +390,21 @@ static void *__netdev_alloc_frag(unsigned int fragsz, gfp_t gfp_mask)
  */
 void *netdev_alloc_frag(unsigned int fragsz)
 {
-	fragsz = SKB_DATA_ALIGN(fragsz);
-
-	return __netdev_alloc_frag(fragsz, GFP_ATOMIC);
-}
-EXPORT_SYMBOL(netdev_alloc_frag);
-
-static void *__napi_alloc_frag(unsigned int fragsz, gfp_t gfp_mask)
-{
-	struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
-
-	return page_frag_alloc(&nc->page, fragsz, gfp_mask);
-}
+	struct page_frag_cache *nc;
+	void *data;
 
-void *napi_alloc_frag(unsigned int fragsz)
-{
 	fragsz = SKB_DATA_ALIGN(fragsz);
-
-	return __napi_alloc_frag(fragsz, GFP_ATOMIC);
+	if (in_irq() || irqs_disabled()) {
+		nc = this_cpu_ptr(&netdev_alloc_cache);
+		data = page_frag_alloc(nc, fragsz, GFP_ATOMIC);
+	} else {
+		local_bh_disable();
+		data = __napi_alloc_frag(fragsz, GFP_ATOMIC);
+		local_bh_enable();
+	}
+	return data;
 }
-EXPORT_SYMBOL(napi_alloc_frag);
+EXPORT_SYMBOL(netdev_alloc_frag);
 
 /**
  *	__netdev_alloc_skb - allocate an skbuff for rx on a specific device
@@ -426,7 +423,6 @@ struct sk_buff *__netdev_alloc_skb(struct net_device *dev, unsigned int len,
 				   gfp_t gfp_mask)
 {
 	struct page_frag_cache *nc;
-	unsigned long flags;
 	struct sk_buff *skb;
 	bool pfmemalloc;
 	void *data;
@@ -447,13 +443,17 @@ struct sk_buff *__netdev_alloc_skb(struct net_device *dev, unsigned int len,
 	if (sk_memalloc_socks())
 		gfp_mask |= __GFP_MEMALLOC;
 
-	local_irq_save(flags);
-
+	if (in_irq() || irqs_disabled()) {
 		nc = this_cpu_ptr(&netdev_alloc_cache);
 		data = page_frag_alloc(nc, len, gfp_mask);
 		pfmemalloc = nc->pfmemalloc;
-
-	local_irq_restore(flags);
+	} else {
+		local_bh_disable();
+		nc = this_cpu_ptr(&napi_alloc_cache.page);
+		data = page_frag_alloc(nc, len, gfp_mask);
+		pfmemalloc = nc->pfmemalloc;
+		local_bh_enable();
+	}
 
 	if (unlikely(!data))
 		return NULL;