diff mbox series

[3/4] gve: Rx Buffer Recycling

Message ID 20201103174651.590586-4-awogbemila@google.com
State Changes Requested
Delegated to: Netdev Driver Reviewers
Headers show
Series GVE Raw Addressing | expand

Checks

Context Check Description
jkicinski/cover_letter success Link
jkicinski/fixes_present success Link
jkicinski/patch_count success Link
jkicinski/tree_selection success Clearly marked for net-next
jkicinski/subject_prefix success Link
jkicinski/source_inline success Was 0 now: 0
jkicinski/verify_signedoff success Link
jkicinski/module_param success Was 0 now: 0
jkicinski/build_32bit success Errors and warnings before: 0 this patch: 0
jkicinski/kdoc success Errors and warnings before: 0 this patch: 0
jkicinski/verify_fixes success Link
jkicinski/checkpatch warning WARNING: line length of 81 exceeds 80 columns WARNING: line length of 84 exceeds 80 columns WARNING: line length of 87 exceeds 80 columns WARNING: line length of 89 exceeds 80 columns WARNING: line length of 93 exceeds 80 columns
jkicinski/build_allmodconfig_warn success Errors and warnings before: 0 this patch: 0
jkicinski/header_inline success Link
jkicinski/stable success Stable not CCed

Commit Message

David Awogbemila Nov. 3, 2020, 5:46 p.m. UTC
This patch lets the driver reuse buffers that have been freed by the
networking stack.

In the raw addressing case, this allows the driver avoid allocating new
buffers.
In the qpl case, the driver can avoid copies.

Signed-off-by: David Awogbemila <awogbemila@google.com>
---
 drivers/net/ethernet/google/gve/gve.h    |  10 +-
 drivers/net/ethernet/google/gve/gve_rx.c | 194 +++++++++++++++--------
 2 files changed, 129 insertions(+), 75 deletions(-)

Comments

Saeed Mahameed Nov. 4, 2020, 12:01 a.m. UTC | #1
On Tue, 2020-11-03 at 09:46 -0800, David Awogbemila wrote:
> This patch lets the driver reuse buffers that have been freed by the
> networking stack.
> 
> In the raw addressing case, this allows the driver avoid allocating
> new
> buffers.
> In the qpl case, the driver can avoid copies.
> 
> Signed-off-by: David Awogbemila <awogbemila@google.com>
> ---
>  drivers/net/ethernet/google/gve/gve.h    |  10 +-
>  drivers/net/ethernet/google/gve/gve_rx.c | 194 +++++++++++++++----
> ----

> +	if (len <= priv->rx_copybreak) {
> +		/* Just copy small packets */
> +		skb = gve_rx_copy(dev, napi, page_info, len);
> +		u64_stats_update_begin(&rx->statss);
> +		rx->rx_copied_pkt++;
> +		rx->rx_copybreak_pkt++;
> +		u64_stats_update_end(&rx->statss);
> +	} else {
> +		bool can_flip = gve_rx_can_flip_buffers(dev);
> +		int recycle = 0;
> +
> +		if (can_flip) {
> +			recycle = gve_rx_can_recycle_buffer(page_info-
> >page);
> +			if (recycle < 0) {
> +				gve_schedule_reset(priv);

How would a reset solve anything if your driver is handling pages with
"bad" refcount, i don't agree here that reset is the best course of
action, all you can do here is warn and leak the page ...
this is a critical driver bug and not something that user should
expect. 

> +	
> +		} else {
> +			/* It is possible that the networking stack has
> already
> +			 * finished processing all outstanding packets
> in the buffer
> +			 * and it can be reused.
> +			 * Flipping is unnecessary here - if the
> networking stack still
> +			 * owns half the page it is impossible to tell
> which half. Either
> +			 * the whole page is free or it needs to be
> replaced.
> +			 */
> +			int recycle =
> gve_rx_can_recycle_buffer(page_info->page);
> +
> +			if (recycle < 0) {
> +				gve_schedule_reset(priv);
> +				return false;

Same.
David Awogbemila Nov. 6, 2020, 8:16 p.m. UTC | #2
On Tue, Nov 3, 2020 at 4:01 PM Saeed Mahameed <saeed@kernel.org> wrote:
>
> On Tue, 2020-11-03 at 09:46 -0800, David Awogbemila wrote:
> > This patch lets the driver reuse buffers that have been freed by the
> > networking stack.
> >
> > In the raw addressing case, this allows the driver avoid allocating
> > new
> > buffers.
> > In the qpl case, the driver can avoid copies.
> >
> > Signed-off-by: David Awogbemila <awogbemila@google.com>
> > ---
> >  drivers/net/ethernet/google/gve/gve.h    |  10 +-
> >  drivers/net/ethernet/google/gve/gve_rx.c | 194 +++++++++++++++----
> > ----
>
> > +     if (len <= priv->rx_copybreak) {
> > +             /* Just copy small packets */
> > +             skb = gve_rx_copy(dev, napi, page_info, len);
> > +             u64_stats_update_begin(&rx->statss);
> > +             rx->rx_copied_pkt++;
> > +             rx->rx_copybreak_pkt++;
> > +             u64_stats_update_end(&rx->statss);
> > +     } else {
> > +             bool can_flip = gve_rx_can_flip_buffers(dev);
> > +             int recycle = 0;
> > +
> > +             if (can_flip) {
> > +                     recycle = gve_rx_can_recycle_buffer(page_info-
> > >page);
> > +                     if (recycle < 0) {
> > +                             gve_schedule_reset(priv);
>
> How would a reset solve anything if your driver is handling pages with
> "bad" refcount, i don't agree here that reset is the best course of
> action, all you can do here is warn and leak the page ...
> this is a critical driver bug and not something that user should
> expect.
>
Thanks for pointing this out. For the raw addressing case it would be
fine to warn and leak the page, but for the qpl (non raw addressing)
case, the driver pre-allocates a set of pages which it registers with
the NIC so we'd want to avoid leaking pages - a reset will help here
because we'd allocate a new set of pages and register them with the
NIC.
I will handle both cases separately:
- in raw addressing mode, just warn and leak the page
- in qpl mode, schedule a reset.


> > +
> > +             } else {
> > +                     /* It is possible that the networking stack has
> > already
> > +                      * finished processing all outstanding packets
> > in the buffer
> > +                      * and it can be reused.
> > +                      * Flipping is unnecessary here - if the
> > networking stack still
> > +                      * owns half the page it is impossible to tell
> > which half. Either
> > +                      * the whole page is free or it needs to be
> > replaced.
> > +                      */
> > +                     int recycle =
> > gve_rx_can_recycle_buffer(page_info->page);
> > +
> > +                     if (recycle < 0) {
> > +                             gve_schedule_reset(priv);
> > +                             return false;
>
> Same.
>
>
diff mbox series

Patch

diff --git a/drivers/net/ethernet/google/gve/gve.h b/drivers/net/ethernet/google/gve/gve.h
index 5c48f9e2d04c..d1a2a526a250 100644
--- a/drivers/net/ethernet/google/gve/gve.h
+++ b/drivers/net/ethernet/google/gve/gve.h
@@ -50,6 +50,7 @@  struct gve_rx_slot_page_info {
 	struct page *page;
 	void *page_address;
 	u32 page_offset; /* offset to write to in page */
+	bool can_flip;
 };
 
 /* A list of pages registered with the device during setup and used by a queue
@@ -503,15 +504,6 @@  static inline enum dma_data_direction gve_qpl_dma_dir(struct gve_priv *priv,
 		return DMA_FROM_DEVICE;
 }
 
-/* Returns true if the max mtu allows page recycling */
-static inline bool gve_can_recycle_pages(struct net_device *dev)
-{
-	/* We can't recycle the pages if we can't fit a packet into half a
-	 * page.
-	 */
-	return dev->max_mtu <= PAGE_SIZE / 2;
-}
-
 /* buffers */
 int gve_alloc_page(struct gve_priv *priv, struct device *dev,
 		   struct page **page, dma_addr_t *dma,
diff --git a/drivers/net/ethernet/google/gve/gve_rx.c b/drivers/net/ethernet/google/gve/gve_rx.c
index 6d1ca6985eec..2e9d5c6fc411 100644
--- a/drivers/net/ethernet/google/gve/gve_rx.c
+++ b/drivers/net/ethernet/google/gve/gve_rx.c
@@ -270,8 +270,7 @@  static enum pkt_hash_types gve_rss_type(__be16 pkt_flags)
 	return PKT_HASH_TYPE_L2;
 }
 
-static struct sk_buff *gve_rx_copy(struct gve_rx_ring *rx,
-				   struct net_device *dev,
+static struct sk_buff *gve_rx_copy(struct net_device *dev,
 				   struct napi_struct *napi,
 				   struct gve_rx_slot_page_info *page_info,
 				   u16 len)
@@ -289,10 +288,6 @@  static struct sk_buff *gve_rx_copy(struct gve_rx_ring *rx,
 
 	skb->protocol = eth_type_trans(skb, dev);
 
-	u64_stats_update_begin(&rx->statss);
-	rx->rx_copied_pkt++;
-	u64_stats_update_end(&rx->statss);
-
 	return skb;
 }
 
@@ -338,22 +333,90 @@  static void gve_rx_flip_buff(struct gve_rx_slot_page_info *page_info,
 {
 	u64 addr = be64_to_cpu(data_ring->addr);
 
+	/* "flip" to other packet buffer on this page */
 	page_info->page_offset ^= PAGE_SIZE / 2;
 	addr ^= PAGE_SIZE / 2;
 	data_ring->addr = cpu_to_be64(addr);
 }
 
+static bool gve_rx_can_flip_buffers(struct net_device *netdev)
+{
+#if PAGE_SIZE == 4096
+	/* We can't flip a buffer if we can't fit a packet
+	 * into half a page.
+	 */
+	return netdev->mtu + GVE_RX_PAD + ETH_HLEN <= PAGE_SIZE / 2;
+#else
+	/* PAGE_SIZE != 4096 - don't try to reuse */
+	return false;
+#endif
+}
+
+static int gve_rx_can_recycle_buffer(struct page *page)
+{
+	int pagecount = page_count(page);
+
+	/* This page is not being used by any SKBs - reuse */
+	if (pagecount == 1)
+		return 1;
+	/* This page is still being used by an SKB - we can't reuse */
+	else if (pagecount >= 2)
+		return 0;
+	WARN(pagecount < 1, "Pagecount should never be < 1");
+	return -1;
+}
+
 static struct sk_buff *
 gve_rx_raw_addressing(struct device *dev, struct net_device *netdev,
 		      struct gve_rx_slot_page_info *page_info, u16 len,
 		      struct napi_struct *napi,
-		      struct gve_rx_data_slot *data_slot)
+		      struct gve_rx_data_slot *data_slot, bool can_flip)
 {
-	struct sk_buff *skb = gve_rx_add_frags(napi, page_info, len);
+	struct sk_buff *skb;
 
+	skb = gve_rx_add_frags(napi, page_info, len);
 	if (!skb)
 		return NULL;
 
+	/* Optimistically stop the kernel from freeing the page by increasing
+	 * the page bias. We will check the refcount in refill to determine if
+	 * we need to alloc a new page.
+	 */
+	get_page(page_info->page);
+	page_info->can_flip = can_flip;
+
+	return skb;
+}
+
+static struct sk_buff *
+gve_rx_qpl(struct device *dev, struct net_device *netdev,
+	   struct gve_rx_ring *rx, struct gve_rx_slot_page_info *page_info,
+	   u16 len, struct napi_struct *napi,
+	   struct gve_rx_data_slot *data_slot, bool recycle)
+{
+	struct sk_buff *skb;
+
+	/* if raw_addressing mode is not enabled gvnic can only receive into
+	 * registered segments. If the buffer can't be recycled, our only
+	 * choice is to copy the data out of it so that we can return it to the
+	 * device.
+	 */
+	if (recycle) {
+		skb = gve_rx_add_frags(napi, page_info, len);
+		/* No point in recycling if we didn't get the skb */
+		if (skb) {
+			/* Make sure the networking stack can't free the page */
+			get_page(page_info->page);
+			gve_rx_flip_buff(page_info, data_slot);
+		}
+	} else {
+		skb = gve_rx_copy(netdev, napi, page_info, len);
+		if (skb) {
+			u64_stats_update_begin(&rx->statss);
+			rx->rx_copied_pkt++;
+			u64_stats_update_end(&rx->statss);
+		}
+	}
 	return skb;
 }
 
@@ -367,7 +430,6 @@  static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
 	struct gve_rx_data_slot *data_slot;
 	struct sk_buff *skb = NULL;
 	dma_addr_t page_bus;
-	int pagecount;
 	u16 len;
 
 	/* drop this packet */
@@ -388,64 +450,36 @@  static bool gve_rx(struct gve_rx_ring *rx, struct gve_rx_desc *rx_desc,
 	dma_sync_single_for_cpu(&priv->pdev->dev, page_bus,
 				PAGE_SIZE, DMA_FROM_DEVICE);
 
-	if (PAGE_SIZE == 4096) {
-		if (len <= priv->rx_copybreak) {
-			/* Just copy small packets */
-			skb = gve_rx_copy(rx, dev, napi, page_info, len);
-			u64_stats_update_begin(&rx->statss);
-			rx->rx_copybreak_pkt++;
-			u64_stats_update_end(&rx->statss);
-			goto have_skb;
+	if (len <= priv->rx_copybreak) {
+		/* Just copy small packets */
+		skb = gve_rx_copy(dev, napi, page_info, len);
+		u64_stats_update_begin(&rx->statss);
+		rx->rx_copied_pkt++;
+		rx->rx_copybreak_pkt++;
+		u64_stats_update_end(&rx->statss);
+	} else {
+		bool can_flip = gve_rx_can_flip_buffers(dev);
+		int recycle = 0;
+
+		if (can_flip) {
+			recycle = gve_rx_can_recycle_buffer(page_info->page);
+			if (recycle < 0) {
+				gve_schedule_reset(priv);
+				return false;
+			}
 		}
 		if (rx->data.raw_addressing) {
 			skb = gve_rx_raw_addressing(&priv->pdev->dev, dev,
 						    page_info, len, napi,
-						     data_slot);
-			goto have_skb;
-		}
-		if (unlikely(!gve_can_recycle_pages(dev))) {
-			skb = gve_rx_copy(rx, dev, napi, page_info, len);
-			goto have_skb;
-		}
-		pagecount = page_count(page_info->page);
-		if (pagecount == 1) {
-			/* No part of this page is used by any SKBs; we attach
-			 * the page fragment to a new SKB and pass it up the
-			 * stack.
-			 */
-			skb = gve_rx_add_frags(napi, page_info, len);
-			if (!skb) {
-				u64_stats_update_begin(&rx->statss);
-				rx->rx_skb_alloc_fail++;
-				u64_stats_update_end(&rx->statss);
-				return false;
-			}
-			/* Make sure the kernel stack can't release the page */
-			get_page(page_info->page);
-			/* "flip" to other packet buffer on this page */
-			gve_rx_flip_buff(page_info, &rx->data.data_ring[idx]);
-		} else if (pagecount >= 2) {
-			/* We have previously passed the other half of this
-			 * page up the stack, but it has not yet been freed.
-			 */
-			skb = gve_rx_copy(rx, dev, napi, page_info, len);
+						    data_slot,
+						    can_flip && recycle);
 		} else {
-			WARN(pagecount < 1, "Pagecount should never be < 1");
-			return false;
+			skb = gve_rx_qpl(&priv->pdev->dev, dev, rx,
+					 page_info, len, napi, data_slot,
+					 can_flip && recycle);
 		}
-	} else {
-		if (rx->data.raw_addressing)
-			skb = gve_rx_raw_addressing(&priv->pdev->dev, dev,
-						    page_info, len, napi,
-						    data_slot);
-		else
-			skb = gve_rx_copy(rx, dev, napi, page_info, len);
 	}
 
-have_skb:
-	/* We didn't manage to allocate an skb but we haven't had any
-	 * reset worthy failures.
-	 */
 	if (!skb) {
 		u64_stats_update_begin(&rx->statss);
 		rx->rx_skb_alloc_fail++;
@@ -498,16 +532,44 @@  static bool gve_rx_refill_buffers(struct gve_priv *priv, struct gve_rx_ring *rx)
 
 	while (empty || ((fill_cnt & rx->mask) != (rx->cnt & rx->mask))) {
 		struct gve_rx_slot_page_info *page_info;
-		struct device *dev = &priv->pdev->dev;
-		struct gve_rx_data_slot *data_slot;
 		u32 idx = fill_cnt & rx->mask;
 
 		page_info = &rx->data.page_info[idx];
-		data_slot = &rx->data.data_ring[idx];
-		gve_rx_free_buffer(dev, page_info, data_slot);
-		page_info->page = NULL;
-		if (gve_rx_alloc_buffer(priv, dev, page_info, data_slot, rx))
-			break;
+		if (page_info->can_flip) {
+			/* The other half of the page is free because it was
+			 * free when we processed the descriptor. Flip to it.
+			 */
+			struct gve_rx_data_slot *data_slot =
+						&rx->data.data_ring[idx];
+
+			gve_rx_flip_buff(page_info, data_slot);
+			page_info->can_flip = false;
+		} else {
+			/* It is possible that the networking stack has already
+			 * finished processing all outstanding packets in the buffer
+			 * and it can be reused.
+			 * Flipping is unnecessary here - if the networking stack still
+			 * owns half the page it is impossible to tell which half. Either
+			 * the whole page is free or it needs to be replaced.
+			 */
+			int recycle = gve_rx_can_recycle_buffer(page_info->page);
+
+			if (recycle < 0) {
+				gve_schedule_reset(priv);
+				return false;
+			}
+			if (!recycle) {
+				/* We can't reuse the buffer - alloc a new one*/
+				struct gve_rx_data_slot *data_slot =
+						&rx->data.data_ring[idx];
+				struct device *dev = &priv->pdev->dev;
+
+				gve_rx_free_buffer(dev, page_info, data_slot);
+				page_info->page = NULL;
+				if (gve_rx_alloc_buffer(priv, dev, page_info, data_slot, rx))
+					break;
+			}
+		}
 		empty = false;
 		fill_cnt++;
 	}