| Message ID | 20200709081142.30861-1-etienne.duble@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | bcmgenet: fix DMA buffer management | expand |
On Mon, 20 Jul 2020 09:29:31 +0200 (CEST) etienne.duble@imag.fr (Etienne Dublé) wrote: > Hi Jason, > Thanks for testing and for the additional fix. > In my case (with dhcp followed by 3 TFTP transfers), the board booted fine 20 times in a row with only my fix applied. Previously it was booting once in 10 trials or so. > The driver resets tx_index in bcmgenet_gmac_eth_send() if it is higher than 256, and the same occurs for rx_index in bcmgenet_gmac_free_pkt(), so I suppose the issue you faced only occurs when the boundary case is reached between two commands? > Anyway, thanks for the update. FWIW I also have an update. My Raspberry Pi 4 was consistently failing to boot from the network. Network packet capture was full of truncated, duplicate and otherwise abnormal packets. Most interestingly, many truncated packets had the same length as a packet transmitted exactly 256 frames before. Matthias has rebuilt U-Boot with these two patches added: - https://patchwork.ozlabs.org/project/uboot/patch/20200717133200.136257-1-jason.wessel@windriver.com/ - https://patchwork.ozlabs.org/project/uboot/patch/20200717133200.136257-2-jason.wessel@windriver.com/ With these changes, the boot process is rock-solid, and I'm not seeing any abnormal packets in the capture any longer. Therefore let me add my Tested-by: Petr Tesarik <ptesarik@suse.com> Cheers, Petr T > Cheers > Etienne > > > > De: "Jason Wessel" <jason.wessel at windriver.com> > ?: "Etienne DUBLE" <etienne.duble at gmail.com>, "joe hershberger" <joe.hershberger at ni.com> > Cc: u-boot at lists.denx.de, "ETIENNE DUBLE" <etienne.duble at imag.fr> > Envoy?: Jeudi 16 Juillet 2020 18:02:11 > Objet: Re: [PATCH] bcmgenet: fix DMA buffer management > > On 7/16/20 7:02 AM, Jason Wessel wrote: > > On 7/9/20 3:11 AM, etienne.duble at gmail.com wrote: > >> From: Etienne Dubl? <etienne.duble at imag.fr> > >> > >> This commit fixes a serious issue occuring when several network > >> commands are run on a raspberry pi 4 board: for instance a "dhcp" > >> command and then one or several "tftp" commands. In this case, > >> packet recv callbacks were called several times on the same packets, > >> and send function was failing most of the time. > >> > >> note: if the boot procedure is made of a single network > >> command, the issue is not visible. > >> > >> The issue is related to management of the packet ring buffers > >> (producer / consumer) and DMA. > >> Each time a packet is received, the ethernet device stores it > >> in the buffer and increments an index called RDMA_PROD_INDEX. > >> Each time the driver outputs a received packet, it increments > >> another index called RDMA_CONS_INDEX. > >> > >> Between each pair of network commands, as part of the driver > >> 'start' function, previous code tried to reset both RDMA_CONS_INDEX > >> and RDMA_PROD_INDEX to 0. But RDMA_PROD_INDEX cannot be written from > >> driver side, thus its value was actually not updated, and only > >> RDMA_CONS_INDEX was reset to 0. This was resulting in a major > >> synchronization issue between the driver and the device. Most > >> visible bahavior was that the driver seemed to receive again the > >> packets from the previous commands (e.g. DHCP response packets > >> "received" again when performing the first TFTP command). > >> > >> This fix consists in setting RDMA_CONS_INDEX to the same > >> value as RDMA_PROD_INDEX, when resetting the driver. > >> > >> The same kind of fix was needed on the TX side, and a few variables > >> had to be reset accordingly (c_index, tx_index, rx_index). > > > > > > While there is some kind of problem with the driver, because I too > > have observed a problem with multiple requests timing out or failing, > > this patch makes the problem much worse. I was only able to complete > > a single tftp request. > > > > In my case I am using a static IP address and serverip. > > > > Also your patch was missing the sign-off line. Please consider > > running your patches through scripts/checkpatch.pl. > > > > Cheers, > > Jason. > > > >> --- > >> drivers/net/bcmgenet.c | 15 +++++++-------- > >> 1 file changed, 7 insertions(+), 8 deletions(-) > >> > >> diff --git a/drivers/net/bcmgenet.c b/drivers/net/bcmgenet.c > >> index 11b6148ab6..a4facfd63f 100644 > >> --- a/drivers/net/bcmgenet.c > >> +++ b/drivers/net/bcmgenet.c > >> @@ -378,8 +378,6 @@ static void rx_descs_init(struct bcmgenet_eth_priv *priv) > >> u32 len_stat, i; > >> void *desc_base = priv->rx_desc_base; > >> > >> - priv->c_index = 0; > >> - > >> len_stat = (RX_BUF_LENGTH << DMA_BUFLENGTH_SHIFT) | DMA_OWN; > >> > >> for (i = 0; i < RX_DESCS; i++) { > >> @@ -403,8 +401,10 @@ static void rx_ring_init(struct bcmgenet_eth_priv *priv) > >> writel(RX_DESCS * DMA_DESC_SIZE / 4 - 1, > >> priv->mac_reg + RDMA_RING_REG_BASE + DMA_END_ADDR); > >> > >> - writel(0x0, priv->mac_reg + RDMA_PROD_INDEX); > >> - writel(0x0, priv->mac_reg + RDMA_CONS_INDEX); > >> + /* cannot init RDMA_PROD_INDEX to 0, so align RDMA_CONS_INDEX on it instead */ > >> + priv->c_index = readl(priv->mac_reg + RDMA_PROD_INDEX); > >> + writel(priv->c_index, priv->mac_reg + RDMA_CONS_INDEX); > >> + priv->rx_index = priv->c_index; > > > printf("before RX_IDX: 0x%x\n", priv->rx_index); > > I added a printf() like above for the RX and TX to see what is going on when > I try and transfer a kernel Image file the second time. > > > U-Boot> tftp ${loadaddr} bootfs/Image > before RX_IDX: 0x0 > before TX_IDX: 0x0 > Using ethernet at 7d580000 device > Filename 'bootfs/Image'. > Load address: 0x80000 > Loading: ## Warning: gatewayip needed but not set > ################################################## 16.8 MiB > 6.1 MiB/s > done > Bytes transferred = 17615360 (10cca00 hex) > U-Boot> tftp ${loadaddr} bootfs/Image > before RX_IDX: 0xe4 > before TX_IDX: 0x2ee3 > Using ethernet at 7d580000 device > Filename 'bootfs/Image'. > Load address: 0x80000 > Loading: ## Warning: gatewayip needed but not set > > > > The TX_IDX is now 0x2ee3 which is definitely not going to work. > > According to the driver file there are only 256 (0xFF) slots, > which is why it hangs, with your change. > > Jason. > > >> writel((RX_DESCS << DMA_RING_SIZE_SHIFT) | RX_BUF_LENGTH, > >> priv->mac_reg + RDMA_RING_REG_BASE + DMA_RING_BUF_SIZE); > >> writel(DMA_FC_THRESH_VALUE, priv->mac_reg + RDMA_XON_XOFF_THRESH); > >> @@ -421,8 +421,9 @@ static void tx_ring_init(struct bcmgenet_eth_priv *priv) > >> writel(0x0, priv->mac_reg + TDMA_WRITE_PTR); > >> writel(TX_DESCS * DMA_DESC_SIZE / 4 - 1, > >> priv->mac_reg + TDMA_RING_REG_BASE + DMA_END_ADDR); > >> - writel(0x0, priv->mac_reg + TDMA_PROD_INDEX); > >> - writel(0x0, priv->mac_reg + TDMA_CONS_INDEX); > >> + /* cannot init TDMA_CONS_INDEX to 0, so align TDMA_PROD_INDEX on it instead */ > >> + priv->tx_index = readl(priv->mac_reg + TDMA_CONS_INDEX); > >> + writel(priv->tx_index, priv->mac_reg + TDMA_PROD_INDEX); > >> writel(0x1, priv->mac_reg + TDMA_RING_REG_BASE + DMA_MBUF_DONE_THRESH); > >> writel(0x0, priv->mac_reg + TDMA_FLOW_PERIOD); > >> writel((TX_DESCS << DMA_RING_SIZE_SHIFT) | RX_BUF_LENGTH, > >> @@ -469,8 +470,6 @@ static int bcmgenet_gmac_eth_start(struct udevice *dev) > >> > >> priv->tx_desc_base = priv->mac_reg + GENET_TX_OFF; > >> priv->rx_desc_base = priv->mac_reg + GENET_RX_OFF; > >> - priv->tx_index = 0x0; > >> - priv->rx_index = 0x0; > >> > >> bcmgenet_umac_reset(priv); > >> > >>
diff --git a/drivers/net/bcmgenet.c b/drivers/net/bcmgenet.c index 11b6148ab6..a4facfd63f 100644 --- a/drivers/net/bcmgenet.c +++ b/drivers/net/bcmgenet.c @@ -378,8 +378,6 @@ static void rx_descs_init(struct bcmgenet_eth_priv *priv) u32 len_stat, i; void *desc_base = priv->rx_desc_base; - priv->c_index = 0; - len_stat = (RX_BUF_LENGTH << DMA_BUFLENGTH_SHIFT) | DMA_OWN; for (i = 0; i < RX_DESCS; i++) { @@ -403,8 +401,10 @@ static void rx_ring_init(struct bcmgenet_eth_priv *priv) writel(RX_DESCS * DMA_DESC_SIZE / 4 - 1, priv->mac_reg + RDMA_RING_REG_BASE + DMA_END_ADDR); - writel(0x0, priv->mac_reg + RDMA_PROD_INDEX); - writel(0x0, priv->mac_reg + RDMA_CONS_INDEX); + /* cannot init RDMA_PROD_INDEX to 0, so align RDMA_CONS_INDEX on it instead */ + priv->c_index = readl(priv->mac_reg + RDMA_PROD_INDEX); + writel(priv->c_index, priv->mac_reg + RDMA_CONS_INDEX); + priv->rx_index = priv->c_index; writel((RX_DESCS << DMA_RING_SIZE_SHIFT) | RX_BUF_LENGTH, priv->mac_reg + RDMA_RING_REG_BASE + DMA_RING_BUF_SIZE); writel(DMA_FC_THRESH_VALUE, priv->mac_reg + RDMA_XON_XOFF_THRESH); @@ -421,8 +421,9 @@ static void tx_ring_init(struct bcmgenet_eth_priv *priv) writel(0x0, priv->mac_reg + TDMA_WRITE_PTR); writel(TX_DESCS * DMA_DESC_SIZE / 4 - 1, priv->mac_reg + TDMA_RING_REG_BASE + DMA_END_ADDR); - writel(0x0, priv->mac_reg + TDMA_PROD_INDEX); - writel(0x0, priv->mac_reg + TDMA_CONS_INDEX); + /* cannot init TDMA_CONS_INDEX to 0, so align TDMA_PROD_INDEX on it instead */ + priv->tx_index = readl(priv->mac_reg + TDMA_CONS_INDEX); + writel(priv->tx_index, priv->mac_reg + TDMA_PROD_INDEX); writel(0x1, priv->mac_reg + TDMA_RING_REG_BASE + DMA_MBUF_DONE_THRESH); writel(0x0, priv->mac_reg + TDMA_FLOW_PERIOD); writel((TX_DESCS << DMA_RING_SIZE_SHIFT) | RX_BUF_LENGTH, @@ -469,8 +470,6 @@ static int bcmgenet_gmac_eth_start(struct udevice *dev) priv->tx_desc_base = priv->mac_reg + GENET_TX_OFF; priv->rx_desc_base = priv->mac_reg + GENET_RX_OFF; - priv->tx_index = 0x0; - priv->rx_index = 0x0; bcmgenet_umac_reset(priv);
From: Etienne Dubl? <etienne.duble at imag.fr> This commit fixes a serious issue occuring when several network commands are run on a raspberry pi 4 board: for instance a "dhcp" command and then one or several "tftp" commands. In this case, packet recv callbacks were called several times on the same packets, and send function was failing most of the time. note: if the boot procedure is made of a single network command, the issue is not visible. The issue is related to management of the packet ring buffers (producer / consumer) and DMA. Each time a packet is received, the ethernet device stores it in the buffer and increments an index called RDMA_PROD_INDEX. Each time the driver outputs a received packet, it increments another index called RDMA_CONS_INDEX. Between each pair of network commands, as part of the driver 'start' function, previous code tried to reset both RDMA_CONS_INDEX and RDMA_PROD_INDEX to 0. But RDMA_PROD_INDEX cannot be written from driver side, thus its value was actually not updated, and only RDMA_CONS_INDEX was reset to 0. This was resulting in a major synchronization issue between the driver and the device. Most visible bahavior was that the driver seemed to receive again the packets from the previous commands (e.g. DHCP response packets "received" again when performing the first TFTP command). This fix consists in setting RDMA_CONS_INDEX to the same value as RDMA_PROD_INDEX, when resetting the driver. The same kind of fix was needed on the TX side, and a few variables had to be reset accordingly (c_index, tx_index, rx_index). --- drivers/net/bcmgenet.c | 15 +++++++-------- 1 file changed, 7 insertions(+), 8 deletions(-)