| Message ID | 20240227014014.44855-4-niklas.soderlund+renesas@ragnatech.se (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Geert Uytterhoeven |
| Headers | show |
| Series | ravb: Align Rx descriptor setup and maintenance | expand |
On 27/02/2024 01:40, Niklas Söderlund wrote: > The RAVB device requires the SKB data to be aligned to 128 bytes. The > alignment is done by allocating a skb 128 bytes larger than the maximum > frame size supported by the device and adjusting the headroom to fit the > requirement. > > This code has been refactored a few times and small issues have been > added along the way. The issues are not harmful but prevents merging > parts of the Rx code which have been split in two implementations with > the addition of RZ/G2L support, a device that supports larger frame > sizes. > > This change removes the need for duplicated and somewhat inaccurate > hardware alignment constrains stored in the hardware information struct > by creating a helper to handle the allocation of a skb and alignment of > a skb data. > > For the R-Car class of devices the maximum frame size is 4K and each > descriptor is limited to 2K of data. The current implementation does not > support split descriptors, this limits the frame size to 2K. The > current hardware information however records the descriptor size just > under 2K due to bad understanding of the device when larger MTUs where > added. > > For the RZ/G2L device the maximum frame size is 8K and each descriptor > is limited to 4K of data. The current hardware information records this > correctly, but it gets the alignment constrains wrong as just aligns it > by 128, it does not extend it by 128 bytes to allow the full frame to be > stored. This works because the RZ/G2L device supports split descriptors > and allocates each skb to 8K and aligns each 4K descriptor in this > space. This change tidies up code which I re-wrote last week but I haven't sent patches yet... I was holding off sending further patches until we've completed gPTP testing, but maybe I can get an updated RFC series sent first then continue with testing. To get the best RX performance, we should allocate plain RX buffers using a page pool, then call build_skb() when a packet is received. Reducing RX buffer size from 8kB to 2kB for RZ/G2L gets me a 50% improvement in TCP RX throughput and a 200% improvement in UDP RX throughput. Using a page pool gets me a further 10% or so in both TCP & UDP RX throughput. So, if we merge this I'll be trying to remove ravb_alloc_skb() shortly anyway. So NACK on this for now, but let's co-ordinate the improvements we're both making so there aren't any merge conflicts for the maintainers to resolve. Thanks,
On 27/02/2024 01:40, Niklas Söderlund wrote: > The RAVB device requires the SKB data to be aligned to 128 bytes. The > alignment is done by allocating a skb 128 bytes larger than the maximum > frame size supported by the device and adjusting the headroom to fit the > requirement. > > This code has been refactored a few times and small issues have been > added along the way. The issues are not harmful but prevents merging > parts of the Rx code which have been split in two implementations with > the addition of RZ/G2L support, a device that supports larger frame > sizes. > > This change removes the need for duplicated and somewhat inaccurate > hardware alignment constrains stored in the hardware information struct > by creating a helper to handle the allocation of a skb and alignment of > a skb data. > > For the R-Car class of devices the maximum frame size is 4K and each > descriptor is limited to 2K of data. The current implementation does not > support split descriptors, this limits the frame size to 2K. The > current hardware information however records the descriptor size just > under 2K due to bad understanding of the device when larger MTUs where > added. > > For the RZ/G2L device the maximum frame size is 8K and each descriptor > is limited to 4K of data. The current hardware information records this > correctly, but it gets the alignment constrains wrong as just aligns it > by 128, it does not extend it by 128 bytes to allow the full frame to be > stored. This works because the RZ/G2L device supports split descriptors > and allocates each skb to 8K and aligns each 4K descriptor in this > space. > > Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se> After some discussion with Niklas on IRC, I'm dropping my NACK so that this can hopefully get in to v6.9. I'll have to re-do some of my work, but it was unlikely that would be ready to go in for v6.9 anyway. So, here's some review... > --- > drivers/net/ethernet/renesas/ravb.h | 1 - > drivers/net/ethernet/renesas/ravb_main.c | 41 +++++++++++++----------- > 2 files changed, 22 insertions(+), 20 deletions(-) > > diff --git a/drivers/net/ethernet/renesas/ravb.h b/drivers/net/ethernet/renesas/ravb.h > index 7f9e8b2c012a..751bb29cd488 100644 > --- a/drivers/net/ethernet/renesas/ravb.h > +++ b/drivers/net/ethernet/renesas/ravb.h > @@ -1057,7 +1057,6 @@ struct ravb_hw_info { > netdev_features_t net_hw_features; > netdev_features_t net_features; > int stats_len; > - size_t max_rx_len; > u32 tccr_mask; > u32 rx_max_frame_size; > unsigned aligned_tx: 1; > diff --git a/drivers/net/ethernet/renesas/ravb_main.c b/drivers/net/ethernet/renesas/ravb_main.c > index 3c59e2c317c7..6e39d498936f 100644 > --- a/drivers/net/ethernet/renesas/ravb_main.c > +++ b/drivers/net/ethernet/renesas/ravb_main.c > @@ -113,12 +113,21 @@ static void ravb_set_rate_rcar(struct net_device *ndev) > } > } > > -static void ravb_set_buffer_align(struct sk_buff *skb) > +static struct sk_buff * > +ravb_alloc_skb(struct net_device *ndev, const struct ravb_hw_info *info) This function should take an extra `gfp_t gfp_mask` argument since it is called from two contexts: RX ring initialization where we want regular allocation, and RX ring refill where we need atomic allocation. > { > - u32 reserve = (unsigned long)skb->data & (RAVB_ALIGN - 1); > + struct sk_buff *skb; > + u32 reserve; > > + skb = netdev_alloc_skb(ndev, info->rx_max_frame_size + RAVB_ALIGN - 1); Call __netdev_alloc_skb() instead with the gfp_mask argument. > + if (!skb) > + return NULL; > + > + reserve = (unsigned long)skb->data & (RAVB_ALIGN - 1); > if (reserve) > skb_reserve(skb, RAVB_ALIGN - reserve); > + > + return skb; > } > > /* Get MAC address from the MAC address registers > @@ -251,7 +260,7 @@ static void ravb_rx_ring_free_gbeth(struct net_device *ndev, int q) > le32_to_cpu(desc->dptr))) > dma_unmap_single(ndev->dev.parent, > le32_to_cpu(desc->dptr), > - GBETH_RX_BUFF_MAX, > + priv->info->rx_max_frame_size, > DMA_FROM_DEVICE); > } > ring_size = sizeof(struct ravb_rx_desc) * (priv->num_rx_ring[q] + 1); > @@ -276,7 +285,7 @@ static void ravb_rx_ring_free_rcar(struct net_device *ndev, int q) > le32_to_cpu(desc->dptr))) > dma_unmap_single(ndev->dev.parent, > le32_to_cpu(desc->dptr), > - RX_BUF_SZ, > + priv->info->rx_max_frame_size, > DMA_FROM_DEVICE); > } > ring_size = sizeof(struct ravb_ex_rx_desc) * > @@ -342,7 +351,7 @@ static void ravb_rx_ring_format_gbeth(struct net_device *ndev, int q) > rx_desc = &priv->rx_ring[q].desc[i]; > rx_desc->ds_cc = cpu_to_le16(GBETH_RX_DESC_DATA_SIZE); > dma_addr = dma_map_single(ndev->dev.parent, priv->rx_skb[q][i]->data, > - GBETH_RX_BUFF_MAX, > + priv->info->rx_max_frame_size, > DMA_FROM_DEVICE); > /* We just set the data size to 0 for a failed mapping which > * should prevent DMA from happening... > @@ -372,7 +381,7 @@ static void ravb_rx_ring_format_rcar(struct net_device *ndev, int q) > rx_desc = &priv->rx_ring[q].ex_desc[i]; > rx_desc->ds_cc = cpu_to_le16(RX_BUF_SZ); > dma_addr = dma_map_single(ndev->dev.parent, priv->rx_skb[q][i]->data, > - RX_BUF_SZ, > + priv->info->rx_max_frame_size, > DMA_FROM_DEVICE); > /* We just set the data size to 0 for a failed mapping which > * should prevent DMA from happening... > @@ -476,10 +485,9 @@ static int ravb_ring_init(struct net_device *ndev, int q) > goto error; > > for (i = 0; i < priv->num_rx_ring[q]; i++) { > - skb = __netdev_alloc_skb(ndev, info->max_rx_len, GFP_KERNEL); > + skb = ravb_alloc_skb(ndev, info); Add GFP_KERNEL as the gfp_mask argument. > if (!skb) > goto error; > - ravb_set_buffer_align(skb); > priv->rx_skb[q][i] = skb; > } > > @@ -805,7 +813,8 @@ static struct sk_buff *ravb_get_skb_gbeth(struct net_device *ndev, int entry, > skb = priv->rx_skb[RAVB_BE][entry]; > priv->rx_skb[RAVB_BE][entry] = NULL; > dma_unmap_single(ndev->dev.parent, le32_to_cpu(desc->dptr), > - ALIGN(GBETH_RX_BUFF_MAX, 16), DMA_FROM_DEVICE); > + ALIGN(priv->info->rx_max_frame_size, 16), > + DMA_FROM_DEVICE); > > return skb; > } > @@ -912,13 +921,12 @@ static bool ravb_rx_gbeth(struct net_device *ndev, int *quota, int q) > desc->ds_cc = cpu_to_le16(GBETH_RX_DESC_DATA_SIZE); > > if (!priv->rx_skb[q][entry]) { > - skb = netdev_alloc_skb(ndev, info->max_rx_len); > + skb = ravb_alloc_skb(ndev, info); Add GFP_ATOMIC as the gfp_mask argument. > if (!skb) > break; > - ravb_set_buffer_align(skb); > dma_addr = dma_map_single(ndev->dev.parent, > skb->data, > - GBETH_RX_BUFF_MAX, > + priv->info->rx_max_frame_size, > DMA_FROM_DEVICE); > skb_checksum_none_assert(skb); > /* We just set the data size to 0 for a failed mapping > @@ -992,7 +1000,7 @@ static bool ravb_rx_rcar(struct net_device *ndev, int *quota, int q) > skb = priv->rx_skb[q][entry]; > priv->rx_skb[q][entry] = NULL; > dma_unmap_single(ndev->dev.parent, le32_to_cpu(desc->dptr), > - RX_BUF_SZ, > + priv->info->rx_max_frame_size, > DMA_FROM_DEVICE); > get_ts &= (q == RAVB_NC) ? > RAVB_RXTSTAMP_TYPE_V2_L2_EVENT : > @@ -1028,10 +1036,9 @@ static bool ravb_rx_rcar(struct net_device *ndev, int *quota, int q) > desc->ds_cc = cpu_to_le16(RX_BUF_SZ); > > if (!priv->rx_skb[q][entry]) { > - skb = netdev_alloc_skb(ndev, info->max_rx_len); > + skb = ravb_alloc_skb(ndev, info); Add GFP_ATOMIC as the gfp_mask argument. > if (!skb) > break; /* Better luck next round. */ > - ravb_set_buffer_align(skb); > dma_addr = dma_map_single(ndev->dev.parent, skb->data, > le16_to_cpu(desc->ds_cc), > DMA_FROM_DEVICE); > @@ -2682,7 +2689,6 @@ static const struct ravb_hw_info ravb_gen3_hw_info = { > .net_hw_features = NETIF_F_RXCSUM, > .net_features = NETIF_F_RXCSUM, > .stats_len = ARRAY_SIZE(ravb_gstrings_stats), > - .max_rx_len = RX_BUF_SZ + RAVB_ALIGN - 1, > .tccr_mask = TCCR_TSRQ0 | TCCR_TSRQ1 | TCCR_TSRQ2 | TCCR_TSRQ3, > .rx_max_frame_size = SZ_2K, > .internal_delay = 1, > @@ -2708,7 +2714,6 @@ static const struct ravb_hw_info ravb_gen2_hw_info = { > .net_hw_features = NETIF_F_RXCSUM, > .net_features = NETIF_F_RXCSUM, > .stats_len = ARRAY_SIZE(ravb_gstrings_stats), > - .max_rx_len = RX_BUF_SZ + RAVB_ALIGN - 1, > .tccr_mask = TCCR_TSRQ0 | TCCR_TSRQ1 | TCCR_TSRQ2 | TCCR_TSRQ3, > .rx_max_frame_size = SZ_2K, > .aligned_tx = 1, > @@ -2731,7 +2736,6 @@ static const struct ravb_hw_info ravb_rzv2m_hw_info = { > .net_hw_features = NETIF_F_RXCSUM, > .net_features = NETIF_F_RXCSUM, > .stats_len = ARRAY_SIZE(ravb_gstrings_stats), > - .max_rx_len = RX_BUF_SZ + RAVB_ALIGN - 1, > .tccr_mask = TCCR_TSRQ0 | TCCR_TSRQ1 | TCCR_TSRQ2 | TCCR_TSRQ3, > .rx_max_frame_size = SZ_2K, > .multi_irqs = 1, > @@ -2756,7 +2760,6 @@ static const struct ravb_hw_info gbeth_hw_info = { > .net_hw_features = NETIF_F_RXCSUM | NETIF_F_HW_CSUM, > .net_features = NETIF_F_RXCSUM | NETIF_F_HW_CSUM, > .stats_len = ARRAY_SIZE(ravb_gstrings_stats_gbeth), > - .max_rx_len = ALIGN(GBETH_RX_BUFF_MAX, RAVB_ALIGN), > .tccr_mask = TCCR_TSRQ0, > .rx_max_frame_size = SZ_8K, > .aligned_tx = 1, Looks ok other than the above comments. I'll try to do some testing tomorrow. Thanks,
diff --git a/drivers/net/ethernet/renesas/ravb.h b/drivers/net/ethernet/renesas/ravb.h index 7f9e8b2c012a..751bb29cd488 100644 --- a/drivers/net/ethernet/renesas/ravb.h +++ b/drivers/net/ethernet/renesas/ravb.h @@ -1057,7 +1057,6 @@ struct ravb_hw_info { netdev_features_t net_hw_features; netdev_features_t net_features; int stats_len; - size_t max_rx_len; u32 tccr_mask; u32 rx_max_frame_size; unsigned aligned_tx: 1; diff --git a/drivers/net/ethernet/renesas/ravb_main.c b/drivers/net/ethernet/renesas/ravb_main.c index 3c59e2c317c7..6e39d498936f 100644 --- a/drivers/net/ethernet/renesas/ravb_main.c +++ b/drivers/net/ethernet/renesas/ravb_main.c @@ -113,12 +113,21 @@ static void ravb_set_rate_rcar(struct net_device *ndev) } } -static void ravb_set_buffer_align(struct sk_buff *skb) +static struct sk_buff * +ravb_alloc_skb(struct net_device *ndev, const struct ravb_hw_info *info) { - u32 reserve = (unsigned long)skb->data & (RAVB_ALIGN - 1); + struct sk_buff *skb; + u32 reserve; + skb = netdev_alloc_skb(ndev, info->rx_max_frame_size + RAVB_ALIGN - 1); + if (!skb) + return NULL; + + reserve = (unsigned long)skb->data & (RAVB_ALIGN - 1); if (reserve) skb_reserve(skb, RAVB_ALIGN - reserve); + + return skb; } /* Get MAC address from the MAC address registers @@ -251,7 +260,7 @@ static void ravb_rx_ring_free_gbeth(struct net_device *ndev, int q) le32_to_cpu(desc->dptr))) dma_unmap_single(ndev->dev.parent, le32_to_cpu(desc->dptr), - GBETH_RX_BUFF_MAX, + priv->info->rx_max_frame_size, DMA_FROM_DEVICE); } ring_size = sizeof(struct ravb_rx_desc) * (priv->num_rx_ring[q] + 1); @@ -276,7 +285,7 @@ static void ravb_rx_ring_free_rcar(struct net_device *ndev, int q) le32_to_cpu(desc->dptr))) dma_unmap_single(ndev->dev.parent, le32_to_cpu(desc->dptr), - RX_BUF_SZ, + priv->info->rx_max_frame_size, DMA_FROM_DEVICE); } ring_size = sizeof(struct ravb_ex_rx_desc) * @@ -342,7 +351,7 @@ static void ravb_rx_ring_format_gbeth(struct net_device *ndev, int q) rx_desc = &priv->rx_ring[q].desc[i]; rx_desc->ds_cc = cpu_to_le16(GBETH_RX_DESC_DATA_SIZE); dma_addr = dma_map_single(ndev->dev.parent, priv->rx_skb[q][i]->data, - GBETH_RX_BUFF_MAX, + priv->info->rx_max_frame_size, DMA_FROM_DEVICE); /* We just set the data size to 0 for a failed mapping which * should prevent DMA from happening... @@ -372,7 +381,7 @@ static void ravb_rx_ring_format_rcar(struct net_device *ndev, int q) rx_desc = &priv->rx_ring[q].ex_desc[i]; rx_desc->ds_cc = cpu_to_le16(RX_BUF_SZ); dma_addr = dma_map_single(ndev->dev.parent, priv->rx_skb[q][i]->data, - RX_BUF_SZ, + priv->info->rx_max_frame_size, DMA_FROM_DEVICE); /* We just set the data size to 0 for a failed mapping which * should prevent DMA from happening... @@ -476,10 +485,9 @@ static int ravb_ring_init(struct net_device *ndev, int q) goto error; for (i = 0; i < priv->num_rx_ring[q]; i++) { - skb = __netdev_alloc_skb(ndev, info->max_rx_len, GFP_KERNEL); + skb = ravb_alloc_skb(ndev, info); if (!skb) goto error; - ravb_set_buffer_align(skb); priv->rx_skb[q][i] = skb; } @@ -805,7 +813,8 @@ static struct sk_buff *ravb_get_skb_gbeth(struct net_device *ndev, int entry, skb = priv->rx_skb[RAVB_BE][entry]; priv->rx_skb[RAVB_BE][entry] = NULL; dma_unmap_single(ndev->dev.parent, le32_to_cpu(desc->dptr), - ALIGN(GBETH_RX_BUFF_MAX, 16), DMA_FROM_DEVICE); + ALIGN(priv->info->rx_max_frame_size, 16), + DMA_FROM_DEVICE); return skb; } @@ -912,13 +921,12 @@ static bool ravb_rx_gbeth(struct net_device *ndev, int *quota, int q) desc->ds_cc = cpu_to_le16(GBETH_RX_DESC_DATA_SIZE); if (!priv->rx_skb[q][entry]) { - skb = netdev_alloc_skb(ndev, info->max_rx_len); + skb = ravb_alloc_skb(ndev, info); if (!skb) break; - ravb_set_buffer_align(skb); dma_addr = dma_map_single(ndev->dev.parent, skb->data, - GBETH_RX_BUFF_MAX, + priv->info->rx_max_frame_size, DMA_FROM_DEVICE); skb_checksum_none_assert(skb); /* We just set the data size to 0 for a failed mapping @@ -992,7 +1000,7 @@ static bool ravb_rx_rcar(struct net_device *ndev, int *quota, int q) skb = priv->rx_skb[q][entry]; priv->rx_skb[q][entry] = NULL; dma_unmap_single(ndev->dev.parent, le32_to_cpu(desc->dptr), - RX_BUF_SZ, + priv->info->rx_max_frame_size, DMA_FROM_DEVICE); get_ts &= (q == RAVB_NC) ? RAVB_RXTSTAMP_TYPE_V2_L2_EVENT : @@ -1028,10 +1036,9 @@ static bool ravb_rx_rcar(struct net_device *ndev, int *quota, int q) desc->ds_cc = cpu_to_le16(RX_BUF_SZ); if (!priv->rx_skb[q][entry]) { - skb = netdev_alloc_skb(ndev, info->max_rx_len); + skb = ravb_alloc_skb(ndev, info); if (!skb) break; /* Better luck next round. */ - ravb_set_buffer_align(skb); dma_addr = dma_map_single(ndev->dev.parent, skb->data, le16_to_cpu(desc->ds_cc), DMA_FROM_DEVICE); @@ -2682,7 +2689,6 @@ static const struct ravb_hw_info ravb_gen3_hw_info = { .net_hw_features = NETIF_F_RXCSUM, .net_features = NETIF_F_RXCSUM, .stats_len = ARRAY_SIZE(ravb_gstrings_stats), - .max_rx_len = RX_BUF_SZ + RAVB_ALIGN - 1, .tccr_mask = TCCR_TSRQ0 | TCCR_TSRQ1 | TCCR_TSRQ2 | TCCR_TSRQ3, .rx_max_frame_size = SZ_2K, .internal_delay = 1, @@ -2708,7 +2714,6 @@ static const struct ravb_hw_info ravb_gen2_hw_info = { .net_hw_features = NETIF_F_RXCSUM, .net_features = NETIF_F_RXCSUM, .stats_len = ARRAY_SIZE(ravb_gstrings_stats), - .max_rx_len = RX_BUF_SZ + RAVB_ALIGN - 1, .tccr_mask = TCCR_TSRQ0 | TCCR_TSRQ1 | TCCR_TSRQ2 | TCCR_TSRQ3, .rx_max_frame_size = SZ_2K, .aligned_tx = 1, @@ -2731,7 +2736,6 @@ static const struct ravb_hw_info ravb_rzv2m_hw_info = { .net_hw_features = NETIF_F_RXCSUM, .net_features = NETIF_F_RXCSUM, .stats_len = ARRAY_SIZE(ravb_gstrings_stats), - .max_rx_len = RX_BUF_SZ + RAVB_ALIGN - 1, .tccr_mask = TCCR_TSRQ0 | TCCR_TSRQ1 | TCCR_TSRQ2 | TCCR_TSRQ3, .rx_max_frame_size = SZ_2K, .multi_irqs = 1, @@ -2756,7 +2760,6 @@ static const struct ravb_hw_info gbeth_hw_info = { .net_hw_features = NETIF_F_RXCSUM | NETIF_F_HW_CSUM, .net_features = NETIF_F_RXCSUM | NETIF_F_HW_CSUM, .stats_len = ARRAY_SIZE(ravb_gstrings_stats_gbeth), - .max_rx_len = ALIGN(GBETH_RX_BUFF_MAX, RAVB_ALIGN), .tccr_mask = TCCR_TSRQ0, .rx_max_frame_size = SZ_8K, .aligned_tx = 1,
The RAVB device requires the SKB data to be aligned to 128 bytes. The alignment is done by allocating a skb 128 bytes larger than the maximum frame size supported by the device and adjusting the headroom to fit the requirement. This code has been refactored a few times and small issues have been added along the way. The issues are not harmful but prevents merging parts of the Rx code which have been split in two implementations with the addition of RZ/G2L support, a device that supports larger frame sizes. This change removes the need for duplicated and somewhat inaccurate hardware alignment constrains stored in the hardware information struct by creating a helper to handle the allocation of a skb and alignment of a skb data. For the R-Car class of devices the maximum frame size is 4K and each descriptor is limited to 2K of data. The current implementation does not support split descriptors, this limits the frame size to 2K. The current hardware information however records the descriptor size just under 2K due to bad understanding of the device when larger MTUs where added. For the RZ/G2L device the maximum frame size is 8K and each descriptor is limited to 4K of data. The current hardware information records this correctly, but it gets the alignment constrains wrong as just aligns it by 128, it does not extend it by 128 bytes to allow the full frame to be stored. This works because the RZ/G2L device supports split descriptors and allocates each skb to 8K and aligns each 4K descriptor in this space. Signed-off-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se> --- drivers/net/ethernet/renesas/ravb.h | 1 - drivers/net/ethernet/renesas/ravb_main.c | 41 +++++++++++++----------- 2 files changed, 22 insertions(+), 20 deletions(-)