| Message ID | 20200228180226.22986-1-michael@walle.cc |
|---|---|
| Headers | show |
| Series | AT8031 PHY timestamping support | expand |
On Fri, Feb 28, 2020 at 07:02:24PM +0100, Michael Walle wrote: > (1) The PHY doesn't support atomic reading of the (timestamp, > messageType, sequenceId) tuple. The workaround is to read the > timestamp again and check if it has changed. Actually, you'd have > to read the complete tuple again. This HW is broken by design :( > But if you're using a P2P clock with peer delay requests this whole > thing falls apart because of caveat (3). You'll often see messages like > received SYNC without timestamp > or > received PDELAY_RESP without timestamp > in linuxptp. Sometimes it working for some time and then it starts to > loosing packets. I suspect this depends on how the PDELAY messages are > interleaved with the SYNC message. If there is not enough time to until > the next event message is received either of these two messages won't > have a timestamp. And even the case where a Sync and a DelayResp arrive at nearly the same time will fail. > The PHY also supports appending the timestamp to the actual ethernet frame, > but this seems to only work when the PHY is connected via RGMII. I've never > get it to work with a SGMII connection. This is the way to go. I would try to get the vendor's help in making this work. Thanks, Richard
Hi Richard, Am 2020-02-28 19:15, schrieb Richard Cochran: > On Fri, Feb 28, 2020 at 07:02:24PM +0100, Michael Walle wrote: >> (1) The PHY doesn't support atomic reading of the (timestamp, >> messageType, sequenceId) tuple. The workaround is to read the >> timestamp again and check if it has changed. Actually, you'd have >> to read the complete tuple again. > > This HW is broken by design :( Yeah, I know. And actually I don't think I'll pursue this further. Like I said, I just wanted to my current work. Maybe it will be useful in the future who knows. >> But if you're using a P2P clock with peer delay requests this whole >> thing falls apart because of caveat (3). You'll often see messages >> like >> received SYNC without timestamp >> or >> received PDELAY_RESP without timestamp >> in linuxptp. Sometimes it working for some time and then it starts to >> loosing packets. I suspect this depends on how the PDELAY messages are >> interleaved with the SYNC message. If there is not enough time to >> until >> the next event message is received either of these two messages won't >> have a timestamp. > > And even the case where a Sync and a DelayResp arrive at nearly the > same time will fail. > >> The PHY also supports appending the timestamp to the actual ethernet >> frame, >> but this seems to only work when the PHY is connected via RGMII. I've >> never >> get it to work with a SGMII connection. > > This is the way to go. I would try to get the vendor's help in making > this work. Like I said, our FAE is pretty unresponsive. But I'll at least try to find out if my guess is correct (that it only works with RGMII). But even then, how should the outgoing timestamping work. There are two possibilities: (1) According to the datasheet, the PHY will attach the TX timestamp to the corresponding RX packet; whatever that means. Lets assume there is such a "corresponding packet", then we would be at the mercy of the peer to actually send such a packet, let alone in a timely manner. (2) Mixing both methods. Use attached timestamps for RX packets, read the timestamp via PHY registers for TX packets. Theoretically, we could control how the packets are send and make sure, we fetch the TX timestamp before sending another PTP packet. But well.. sounds really hacky to me. -michael
On Fri, Feb 28, 2020 at 08:43:05PM +0100, Michael Walle wrote: > > Yeah, I know. And actually I don't think I'll pursue this further. Like I > said, I just wanted to my current work. Maybe it will be useful in the > future who knows. I appreciate your publishing this work. It is a good starting place. Maybe the vendor will wake up and help this along. One can always hope. > Like I said, our FAE is pretty unresponsive. But I'll at least try to find > out if my guess is correct (that it only works with RGMII). But even then, > how should the outgoing timestamping work. There are two possibilities: > > (1) According to the datasheet, the PHY will attach the TX timestamp to > the corresponding RX packet; whatever that means. Lets assume there > is such a "corresponding packet", then we would be at the mercy of the > peer to actually send such a packet, let alone in a timely manner. I see. Mysterious. For a Sync frame, I can't think of any "corresponding RX packet". > (2) Mixing both methods. Use attached timestamps for RX packets, read the > timestamp via PHY registers for TX packets. Theoretically, we could > control how the packets are send and make sure, we fetch the TX > timestamp before sending another PTP packet. But well.. sounds really > hacky to me. It would not be that bad. Some of the MAC cards can only buffer one Tx time stamp, and for this reason, the ptp4l program always fetches the time stamp immediately after sending. Thanks, Richard
I've put some additional time into figuring out how the PHY works since my last RFC. My previous comments on the interrupt handling was caused by my broken setup. This patchset is the current state of my work for adding PHY timestamping support. I just wanted to post this to the mailinglist before I never do it. Maybe its a starting point for other people. That being said, I wouldn't mind comments ;) Also I like to share my findings about the PHY. The PHY has three major caveats which IMHO makes it really hard to work properly. (1) The PHY doesn't support atomic reading of the (timestamp, messageType, sequenceId) tuple. The workaround is to read the timestamp again and check if it has changed. Actually, you'd have to read the complete tuple again. (2) The PHY generates an interrupt on every PTP packet not only on event messages. Thus the interrupt handler may read the capture registers and then determine that nothing has changed. This also means we have to remember the last read timestamp. I make the assumption that the timestamp is unique. (3) The biggest drawback is that the PHY only provide one set of RX and TX capture registers. It is possible that the timestamp will change when another PTP event message is received while we are still reading the timestamp of the previously received packet. Example A The driver basically works when there is low PTP traffic. Eg. the following works pretty good. ptp4l -m -2 -i eth0 -f ptp.cfg -s ptp.cfg: [global] tx_timestamp_timeout 100 Example B But if you're using a P2P clock with peer delay requests this whole thing falls apart because of caveat (3). You'll often see messages like received SYNC without timestamp or received PDELAY_RESP without timestamp in linuxptp. Sometimes it working for some time and then it starts to loosing packets. I suspect this depends on how the PDELAY messages are interleaved with the SYNC message. If there is not enough time to until the next event message is received either of these two messages won't have a timestamp. ptp4l -m -f gPTP.cfg -i eth0 -s gPTP.cfg is the one from stock linuxptp with tx_timestamp_timeout set to 100. The PHY also supports appending the timestamp to the actual ethernet frame, but this seems to only work when the PHY is connected via RGMII. I've never get it to work with a SGMII connection. The first patch might actually be useful outside of this series. See also https://lore.kernel.org/netdev/bd47f8e1ebc04fa98856ed8d89b91419@walle.cc/ Changes since RFC v1: net: phy: let the driver register its own IRQ handler - fixed mistake for calling phy_request_interrupt(). Thanks Heiner. - removed phy_drv_interrupt(). just calling phy_mac_interrupt() net: phy: at803x: add PTP support for AR8031 - moved rereading the timestamp out of at8031_read_ts() to at8031_get_rxts()/at8031_get_txts(). - call phy_mac_interrupt() instead of phy_drv_interrupt() Please note that Heiner suggested that I should use handle_interrupt() instead of registering my own handler. I've still included my old patch here because the discussion is still ongoing and the proposed fix is only working for this use case. See also https://lore.kernel.org/netdev/2e4371354d84231abf3a63deae1a0d04@walle.cc/ -michael Michael Walle (2): net: phy: let the driver register its own IRQ handler net: phy: at803x: add PTP support for AR8031 drivers/net/phy/Kconfig | 17 + drivers/net/phy/at803x.c | 855 ++++++++++++++++++++++++++++++++++- drivers/net/phy/phy_device.c | 6 +- include/linux/phy.h | 1 + 4 files changed, 852 insertions(+), 27 deletions(-)