| Message ID | 20221024175823.145894-1-saproj@gmail.com (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [v5,net-next] net: ftmac100: support mtu > 1500 | expand |
Hello, On Mon, 2022-10-24 at 20:58 +0300, Sergei Antonov wrote: > The ftmac100 controller considers packets >1518 (1500 + Ethernet + FCS) > FTL (frame too long) and drops them. That is fine with mtu 1500 or less > and it saves CPU time. When DSA is present, mtu is bigger (for VLAN > tagging) and the controller's built-in behavior is not desired then. We > can make the controller deliver FTL packets to the driver by setting > FTMAC100_MACCR_RX_FTL. Then we have to check ftmac100_rxdes_frame_length() > (packet length sans FCS) on packets marked with FTMAC100_RXDES0_FTL flag. > > Check for mtu > 1500 in .ndo_open() and set FTMAC100_MACCR_RX_FTL to let > the driver FTL packets. Implement .ndo_change_mtu() and check for > mtu > 1500 to set/clear FTMAC100_MACCR_RX_FTL dynamically. > > Fixes: 8d77c036b57c ("net: add Faraday FTMAC100 10/100 Ethernet driver") For the records, Vladimir explicitly asked you to drop the 'Fixes' tag. Such tag makes little sense for a net-next commit, especially when referring to an old change - e.g. that did not enter mainline in this release cycle. Cheers, Paolo
Hi Sergei, On Mon, Oct 24, 2022 at 08:58:23PM +0300, Sergei Antonov wrote: > The ftmac100 controller considers packets >1518 (1500 + Ethernet + FCS) > FTL (frame too long) and drops them. That is fine with mtu 1500 or less > and it saves CPU time. When DSA is present, mtu is bigger (for VLAN > tagging) and the controller's built-in behavior is not desired then. We > can make the controller deliver FTL packets to the driver by setting > FTMAC100_MACCR_RX_FTL. Then we have to check ftmac100_rxdes_frame_length() > (packet length sans FCS) on packets marked with FTMAC100_RXDES0_FTL flag. > > Check for mtu > 1500 in .ndo_open() and set FTMAC100_MACCR_RX_FTL to let > the driver FTL packets. Implement .ndo_change_mtu() and check for > mtu > 1500 to set/clear FTMAC100_MACCR_RX_FTL dynamically. > > Fixes: 8d77c036b57c ("net: add Faraday FTMAC100 10/100 Ethernet driver") > Signed-off-by: Sergei Antonov <saproj@gmail.com> > --- I think it's clear there are problems in communication between us, so let me try differently. Does the attached series of 3 patches work for you? I only compile tested them. I tried to keep as much of your work and authorship as possible, the intention was to rewrite the justification in the commit message and to fix the things which your patches didn't do (as separate patches).
On Thu, 27 Oct 2022 at 14:35, Vladimir Oltean <olteanv@gmail.com> wrote: > > Hi Sergei, > > On Mon, Oct 24, 2022 at 08:58:23PM +0300, Sergei Antonov wrote: > > The ftmac100 controller considers packets >1518 (1500 + Ethernet + FCS) > > FTL (frame too long) and drops them. That is fine with mtu 1500 or less > > and it saves CPU time. When DSA is present, mtu is bigger (for VLAN > > tagging) and the controller's built-in behavior is not desired then. We > > can make the controller deliver FTL packets to the driver by setting > > FTMAC100_MACCR_RX_FTL. Then we have to check ftmac100_rxdes_frame_length() > > (packet length sans FCS) on packets marked with FTMAC100_RXDES0_FTL flag. > > > > Check for mtu > 1500 in .ndo_open() and set FTMAC100_MACCR_RX_FTL to let > > the driver FTL packets. Implement .ndo_change_mtu() and check for > > mtu > 1500 to set/clear FTMAC100_MACCR_RX_FTL dynamically. > > > > Fixes: 8d77c036b57c ("net: add Faraday FTMAC100 10/100 Ethernet driver") > > Signed-off-by: Sergei Antonov <saproj@gmail.com> > > --- > > I think it's clear there are problems in communication between us, so > let me try differently. > > Does the attached series of 3 patches work for you? I only compile > tested them. I have tested your patches. They fix the problem I have. If they can make it into mainline Linux - great. Thanks for your help! A remark on 0002-net-ftmac100-report-the-correct-maximum-MTU-of-1500.patch: I can not make a case for VLAN_ETH_HLEN because it includes 4 bytes from a switch and ftmac100 is not always used with a switch.
On Thu, Oct 27, 2022 at 07:59:11PM +0300, Sergei Antonov wrote: > On Thu, 27 Oct 2022 at 14:35, Vladimir Oltean <olteanv@gmail.com> wrote: > > Does the attached series of 3 patches work for you? I only compile > > tested them. > > I have tested your patches. They fix the problem I have. If they can > make it into mainline Linux - great. Thanks for your help! Do you mind submitting these patches yourself, to get a better understanding of the process? You only need to make sure that you preserve the "From:" field (the authorship), and that below the existing Signed-off-by line, you also add yours (to make it clear that the patches authored by me were not submitted by me). Like this: | From: Vladimir Oltean <vladimir.oltean@nxp.com> | | bla bla | | Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> <- same as author | Signed-off-by: Sergei Antonov <saproj@gmail.com> <- patch carried by X | ...etc | when patch is merged, the netdev maintainer adds his own sign off at | the end to indicate that the patch went through his own hands I would do the same if I was the one submitting the series; I would add my sign-off to patch 3/3, which has your authorship. > A remark on 0002-net-ftmac100-report-the-correct-maximum-MTU-of-1500.patch: > I can not make a case for VLAN_ETH_HLEN because it includes 4 bytes > from a switch and ftmac100 is not always used with a switch. Why do you think that? What VLAN are you talking about? 802.1Q or 802.1ad? What VLAN ID? Where does it come from, where do you see it? VLAN_ETH_HLEN in patch 2 has nothing to do with a switch. I tried to preserve the functionality of the driver as best I could. It accepts skb->len on TX up to 1518, and it drops in hardware packets with a skb->len larger than 1514 on RX (this includes L2 header, and is measured before the eth_type_trans() call; the latter consumes ETH_HLEN bytes and thus, skb->len becomes 1500). A VLAN in the real sense (ip link add link eth0 name eth0.100 type vlan id 100) does not increase the MTU of eth0 specifically because the 802.1Q header is part of the L2 overhead, and not part of the L2 payload. So this is why drivers could have a valid reason to subtract the VLAN header length from the maximum packet size (which is total L2 length). There's no way, really, to reconcile the fact that the driver can transmit VLAN-tagged frames with an L2 payload length of 1500 bytes but it cannot receive them (or at least I think that it can't, based on what you said; maybe the hardware is smart and makes an exception for the VLAN header on RX, letting packets with a size of up to 1522 bytes length + CRC to be accepted?). In any case, 1500 is the MTU value that the driver supports as of patch 2 (given the definition of MTU as L2 payload length) and I wanted to make that absolutely clear by bringing in sync what is reported with what is supported, prior to making any other change to the max_mtu.
On Thu, 27 Oct 2022 at 21:54, Vladimir Oltean <olteanv@gmail.com> wrote: > > On Thu, Oct 27, 2022 at 07:59:11PM +0300, Sergei Antonov wrote: > > On Thu, 27 Oct 2022 at 14:35, Vladimir Oltean <olteanv@gmail.com> wrote: > > > Does the attached series of 3 patches work for you? I only compile > > > tested them. > > > > I have tested your patches. They fix the problem I have. If they can > > make it into mainline Linux - great. Thanks for your help! > > Do you mind submitting these patches yourself, to get a better > understanding of the process? You only need to make sure that you > preserve the "From:" field (the authorship), and that below the existing > Signed-off-by line, you also add yours (to make it clear that the > patches authored by me were not submitted by me). Like this: > > | From: Vladimir Oltean <vladimir.oltean@nxp.com> > | > | bla bla > | > | Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com> <- same as author > | Signed-off-by: Sergei Antonov <saproj@gmail.com> <- patch carried by X > | ...etc > | when patch is merged, the netdev maintainer adds his own sign off at > | the end to indicate that the patch went through his own hands > > I would do the same if I was the one submitting the series; I would add > my sign-off to patch 3/3, which has your authorship. OK. I will do it. > > A remark on 0002-net-ftmac100-report-the-correct-maximum-MTU-of-1500.patch: > > I can not make a case for VLAN_ETH_HLEN because it includes 4 bytes > > from a switch and ftmac100 is not always used with a switch. > > Why do you think that? What VLAN are you talking about? 802.1Q or > 802.1ad? What VLAN ID? Where does it come from, where do you see it? Patch 2 contains this change: - netdev->max_mtu = MAX_PKT_SIZE; + netdev->max_mtu = MAX_PKT_SIZE - VLAN_ETH_HLEN; VLAN_ETH_HLEN is equal to 18 which is 6+6+4+2 It includes 4 bytes of 802.1Q. > VLAN_ETH_HLEN in patch 2 has nothing to do with a switch. OK then.
diff --git a/drivers/net/ethernet/faraday/ftmac100.c b/drivers/net/ethernet/faraday/ftmac100.c index d95d78230828..f276d54bcd85 100644 --- a/drivers/net/ethernet/faraday/ftmac100.c +++ b/drivers/net/ethernet/faraday/ftmac100.c @@ -159,6 +159,7 @@ static void ftmac100_set_mac(struct ftmac100 *priv, const unsigned char *mac) static int ftmac100_start_hw(struct ftmac100 *priv) { struct net_device *netdev = priv->netdev; + unsigned int maccr; if (ftmac100_reset(priv)) return -EIO; @@ -175,7 +176,20 @@ static int ftmac100_start_hw(struct ftmac100 *priv) ftmac100_set_mac(priv, netdev->dev_addr); - iowrite32(MACCR_ENABLE_ALL, priv->base + FTMAC100_OFFSET_MACCR); + maccr = MACCR_ENABLE_ALL; + + /* We have to set FTMAC100_MACCR_RX_FTL in case MTU > 1500 + * and do extra length check in ftmac100_rx_packet_error(). + * Otherwise the controller silently drops these packets. + * + * When the MTU of the interface is standard 1500, rely on + * the controller's functionality to drop too long packets + * and save some CPU time. + */ + if (netdev->mtu > 1500) + maccr |= FTMAC100_MACCR_RX_FTL; + + iowrite32(maccr, priv->base + FTMAC100_OFFSET_MACCR); return 0; } @@ -337,9 +351,18 @@ static bool ftmac100_rx_packet_error(struct ftmac100 *priv, error = true; } - if (unlikely(ftmac100_rxdes_frame_too_long(rxdes))) { + /* If the frame-too-long flag FTMAC100_RXDES0_FTL is set, check + * if ftmac100_rxdes_frame_length(rxdes) exceeds the currently + * set MTU plus ETH_HLEN. FCS is not included here. + * The controller would set FTMAC100_RXDES0_FTL for all incoming + * frames longer than 1518 (includeing FCS) in the presence of + * FTMAC100_MACCR_RX_FTL in the MAC Control Register. + */ + if (unlikely(ftmac100_rxdes_frame_too_long(rxdes) && + ftmac100_rxdes_frame_length(rxdes) > netdev->mtu + ETH_HLEN)) { if (net_ratelimit()) - netdev_info(netdev, "rx frame too long\n"); + netdev_info(netdev, "rx frame too long (%u)\n", + ftmac100_rxdes_frame_length(rxdes)); netdev->stats.rx_length_errors++; error = true; @@ -1037,6 +1060,28 @@ static int ftmac100_do_ioctl(struct net_device *netdev, struct ifreq *ifr, int c return generic_mii_ioctl(&priv->mii, data, cmd, NULL); } +static int ftmac100_change_mtu(struct net_device *netdev, int new_mtu) +{ + struct ftmac100 *priv = netdev_priv(netdev); + unsigned int maccr; + + maccr = ioread32(priv->base + FTMAC100_OFFSET_MACCR); + if (new_mtu <= 1500) { + /* Let the controller drop incoming packets greater + * than 1518 (that is 1500 + 14 Ethernet + 4 FCS). + */ + maccr &= ~FTMAC100_MACCR_RX_FTL; + } else { + /* process FTL packets in the driver */ + maccr |= FTMAC100_MACCR_RX_FTL; + } + iowrite32(maccr, priv->base + FTMAC100_OFFSET_MACCR); + + netdev->mtu = new_mtu; + netdev_info(netdev, "changed mtu to %d\n", new_mtu); + return 0; +} + static const struct net_device_ops ftmac100_netdev_ops = { .ndo_open = ftmac100_open, .ndo_stop = ftmac100_stop, @@ -1044,6 +1089,7 @@ static const struct net_device_ops ftmac100_netdev_ops = { .ndo_set_mac_address = eth_mac_addr, .ndo_validate_addr = eth_validate_addr, .ndo_eth_ioctl = ftmac100_do_ioctl, + .ndo_change_mtu = ftmac100_change_mtu, }; /******************************************************************************
The ftmac100 controller considers packets >1518 (1500 + Ethernet + FCS) FTL (frame too long) and drops them. That is fine with mtu 1500 or less and it saves CPU time. When DSA is present, mtu is bigger (for VLAN tagging) and the controller's built-in behavior is not desired then. We can make the controller deliver FTL packets to the driver by setting FTMAC100_MACCR_RX_FTL. Then we have to check ftmac100_rxdes_frame_length() (packet length sans FCS) on packets marked with FTMAC100_RXDES0_FTL flag. Check for mtu > 1500 in .ndo_open() and set FTMAC100_MACCR_RX_FTL to let the driver FTL packets. Implement .ndo_change_mtu() and check for mtu > 1500 to set/clear FTMAC100_MACCR_RX_FTL dynamically. Fixes: 8d77c036b57c ("net: add Faraday FTMAC100 10/100 Ethernet driver") Signed-off-by: Sergei Antonov <saproj@gmail.com> --- v5: * Handle ndo_change_mtu(). * Make description and code comments correct (hopefully). v4: * Set FTMAC100_MACCR_RX_FTL depending on the "mtu > 1500" condition. * DSA tagging seems unrelated to the issue - updated description and a code comment accordingly. v3: * Corrected the explanation of the problem: datasheet is correct. * Rewrote the code to use the currently set mtu to handle DSA frames. v2: * Typos in description fixed. drivers/net/ethernet/faraday/ftmac100.c | 52 +++++++++++++++++++++++-- 1 file changed, 49 insertions(+), 3 deletions(-)