| Message ID | TY3P286MB2611C0FA24318AA397DB689B985A2@TY3P286MB2611.JPNP286.PROD.OUTLOOK.COM (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [v1] net: sfp: add quirks for ODI DFP-34X-2C2 | expand |
On Mon, Feb 26, 2024 at 09:23:46PM +0800, Shengyu Qu wrote: > ODI DFP-34X-2C2 is capable of 2500base-X, but incorrectly report its > capabilities in the EEPROM. > So use sfp_quirk_2500basex for this module to allow 2500Base-X mode. This was previously submitted by Sergio Palumbo, and comes in two different forms - an OEM version and non-OEM. There was extensive discussion about this, and the result is that I'm not accepting this quirk for this module. The reason is that the module _defaults_ to 1000base-X and requires manual reconfiguration by the user to operate at 2500base-X. Unfortunately, there is no way for the kernel to know whether that reconfiguration has occurred. The addition of this quirk makes the kernel select 2500base-X when the module is plugged in to a host that supports both 2500base-X and 1000base-X, resulting in the link with the module never coming up. (2500base-X is 1000base-X clocked 2.5x faster, and there is nothing in the line protocol that identifies this.) Consequently, adding this quirk makes modules in their default configuration not link with the host, and thus be inaccessible. Therefore, for the best user experience (in terms of having a working module when it turns up at the doorstep) the only option is to refuse this quirk. Now, what I really don't like is that I had a lengthy discussion over this with Sergio, and it seems from a mainline developer point of view that "oh, Sergio wasn't successful in getting this merged, someone else can have a go". It _isn't_ the person who is the problem - it is the principle and the confusion this will cause to users who receive modules in their default configuration (1000base-X), and then plug them in with this quirk in place, and the kernel selects 2500base-X resulting in no link and *no* access to the module.
Hi Russell, > On Mon, Feb 26, 2024 at 09:23:46PM +0800, Shengyu Qu wrote: >> ODI DFP-34X-2C2 is capable of 2500base-X, but incorrectly report its >> capabilities in the EEPROM. >> So use sfp_quirk_2500basex for this module to allow 2500Base-X mode. > This was previously submitted by Sergio Palumbo, and comes in two > different forms - an OEM version and non-OEM. There was extensive > discussion about this, and the result is that I'm not accepting this > quirk for this module. > > The reason is that the module _defaults_ to 1000base-X and requires > manual reconfiguration by the user to operate at 2500base-X. > Unfortunately, there is no way for the kernel to know whether that > reconfiguration has occurred. No, In the firmware of this stick, the speed rate is configured to auto negotiation rather than fixed 1000base-X. I tried 3 different versions of the firmware and gets the same result. Also, user can plug in and out the optic fiber for five times in 30 seconds to trigger a force factory reset.This function is also enabled by default. This would also reset the speed configuration to auto negotiation but still keeps LOID config. Best regards, Shengyu > The addition of this quirk makes the kernel select 2500base-X when > the module is plugged in to a host that supports both 2500base-X > and 1000base-X, resulting in the link with the module never coming > up. (2500base-X is 1000base-X clocked 2.5x faster, and there is > nothing in the line protocol that identifies this.) > > Consequently, adding this quirk makes modules in their default > configuration not link with the host, and thus be inaccessible. > > Therefore, for the best user experience (in terms of having a working > module when it turns up at the doorstep) the only option is to refuse > this quirk. > > Now, what I really don't like is that I had a lengthy discussion over > this with Sergio, and it seems from a mainline developer point of view > that "oh, Sergio wasn't successful in getting this merged, someone else > can have a go". > > It _isn't_ the person who is the problem - it is the principle and the > confusion this will cause to users who receive modules in their default > configuration (1000base-X), and then plug them in with this quirk in > place, and the kernel selects 2500base-X resulting in no link and *no* > access to the module. >
On Mon, Feb 26, 2024 at 10:16:46PM +0800, Shengyu Qu wrote: > Hi Russell, > > > On Mon, Feb 26, 2024 at 09:23:46PM +0800, Shengyu Qu wrote: > > > ODI DFP-34X-2C2 is capable of 2500base-X, but incorrectly report its > > > capabilities in the EEPROM. > > > So use sfp_quirk_2500basex for this module to allow 2500Base-X mode. > > This was previously submitted by Sergio Palumbo, and comes in two > > different forms - an OEM version and non-OEM. There was extensive > > discussion about this, and the result is that I'm not accepting this > > quirk for this module. > > > > The reason is that the module _defaults_ to 1000base-X and requires > > manual reconfiguration by the user to operate at 2500base-X. > > Unfortunately, there is no way for the kernel to know whether that > > reconfiguration has occurred. > No, In the firmware of this stick, the speed rate is configured to auto > negotiation rather than fixed 1000base-X. How does this "auto negotiation" work? I mean *exactly* how does it work? How does it know whether the host is operating at 1000base-X or 2500base-X? There is *no* inband protocol to allow this to be negotiated.
Hi Russell, 在 2024/2/26 22:39, Russell King (Oracle) 写道: > On Mon, Feb 26, 2024 at 10:16:46PM +0800, Shengyu Qu wrote: >> Hi Russell, >> >>> On Mon, Feb 26, 2024 at 09:23:46PM +0800, Shengyu Qu wrote: >>>> ODI DFP-34X-2C2 is capable of 2500base-X, but incorrectly report its >>>> capabilities in the EEPROM. >>>> So use sfp_quirk_2500basex for this module to allow 2500Base-X mode. >>> This was previously submitted by Sergio Palumbo, and comes in two >>> different forms - an OEM version and non-OEM. There was extensive >>> discussion about this, and the result is that I'm not accepting this >>> quirk for this module. >>> >>> The reason is that the module _defaults_ to 1000base-X and requires >>> manual reconfiguration by the user to operate at 2500base-X. >>> Unfortunately, there is no way for the kernel to know whether that >>> reconfiguration has occurred. >> No, In the firmware of this stick, the speed rate is configured to auto >> negotiation rather than fixed 1000base-X. > > How does this "auto negotiation" work? > > I mean *exactly* how does it work? How does it know whether the host is > operating at 1000base-X or 2500base-X? > > There is *no* inband protocol to allow this to be negotiated. > Well, that seems some kind weird trick implemented in that chip's SDK (maybe hardware?). It would automatically detect the speed rate that host uses and switch to that rate. The system log of the stick shows that. Best regards, Shengyu
> Well, that seems some kind weird trick implemented in that chip's SDK (maybe > hardware?). It would automatically detect the speed rate that host uses and > switch to that rate. The system log of the stick shows that. Please could you show some examples from the system log. If this patch is accepted, these details need to be in the commit message, to explain how this actually works, when it should not actually work.... Andrew
Hi Andrew, When telnet connected to the stick(It's using busybox), "cat /proc/kmsg" would report something like "<4>change mode to 0", that 0 is the actual speed rate codename it changes to. You could check this [1] for more information. Sorry but I can't record my kmsg output now, since enable connection to my stick would make the network to disconnect for all users and my roommates are using it. Best regards, Shengyu [1] https://github.com/Anime4000/RTL960x/blob/main/Docs/FLASH_GETSET_INFO.md#lan_sds_mode 在 2024/2/26 23:06, Andrew Lunn 写道: >> Well, that seems some kind weird trick implemented in that chip's SDK (maybe >> hardware?). It would automatically detect the speed rate that host uses and >> switch to that rate. The system log of the stick shows that. > > Please could you show some examples from the system log. > > If this patch is accepted, these details need to be in the commit > message, to explain how this actually works, when it should not > actually work.... > > Andrew
On Mon, Feb 26, 2024 at 10:51:36PM +0800, Shengyu Qu wrote: > Hi Russell, > > 在 2024/2/26 22:39, Russell King (Oracle) 写道: > > On Mon, Feb 26, 2024 at 10:16:46PM +0800, Shengyu Qu wrote: > > > Hi Russell, > > > > > > > On Mon, Feb 26, 2024 at 09:23:46PM +0800, Shengyu Qu wrote: > > > > > ODI DFP-34X-2C2 is capable of 2500base-X, but incorrectly report its > > > > > capabilities in the EEPROM. > > > > > So use sfp_quirk_2500basex for this module to allow 2500Base-X mode. > > > > This was previously submitted by Sergio Palumbo, and comes in two > > > > different forms - an OEM version and non-OEM. There was extensive > > > > discussion about this, and the result is that I'm not accepting this > > > > quirk for this module. > > > > > > > > The reason is that the module _defaults_ to 1000base-X and requires > > > > manual reconfiguration by the user to operate at 2500base-X. > > > > Unfortunately, there is no way for the kernel to know whether that > > > > reconfiguration has occurred. > > > No, In the firmware of this stick, the speed rate is configured to auto > > > negotiation rather than fixed 1000base-X. > > > > How does this "auto negotiation" work? > > > > I mean *exactly* how does it work? How does it know whether the host is > > operating at 1000base-X or 2500base-X? > > > > There is *no* inband protocol to allow this to be negotiated. > > > Well, that seems some kind weird trick implemented in that chip's SDK (maybe > hardware?). It would automatically detect the speed rate that host uses and > switch to that rate. The system log of the stick shows that. This sounds racy - between the SFP detecting the speed of the host and the kernel code reconfiguring the interface. More details please...
Hi Andrew, So I should describe in commit message that user should set the stick to 2.5g/automatic mode and include the link in previous mail? Is that enough? Best regards, Shengyu 在 2024/2/26 23:06, Andrew Lunn 写道: >> Well, that seems some kind weird trick implemented in that chip's SDK (maybe >> hardware?). It would automatically detect the speed rate that host uses and >> switch to that rate. The system log of the stick shows that. > > Please could you show some examples from the system log. > > If this patch is accepted, these details need to be in the commit > message, to explain how this actually works, when it should not > actually work.... > > Andrew
On Mon, Feb 26, 2024 at 11:20:47PM +0800, Shengyu Qu wrote: > Hi Andrew, > > When telnet connected to the stick(It's using busybox), "cat /proc/kmsg" > would report something like "<4>change mode to 0", that 0 is the actual > speed rate codename it changes to. You could check this [1] for more > information. > > Sorry but I can't record my kmsg output now, since enable connection to > my stick would make the network to disconnect for all users and my > roommates are using it. > > Best regards, > Shengyu > > [1] https://github.com/Anime4000/RTL960x/blob/main/Docs/FLASH_GETSET_INFO.md#lan_sds_mode I'm aware of that link, and sadly the information in that table can at best be described as "confused". For example, under "Behaviour" it mentions "1GbaseT/100baseT". How exactly does a SFP module exhibit baseT behaviour (which is twisted pair copper ethernet.) Hint: it can't and it doesn't. "Mode" being listed as "TP" is also misleading and wrong. The ethtool value of 0x20 is bit 5, which is 1000baseT_Full which has nothing to do with a SFP (and won't be reported as even a supported mode for this SFP.) Bit 41 would make sense (1000baseX), which is listed against mode 1. I'm afraid looking at that URL just adds to more confusion.
Hi Russell, Sadly I don't have more information. I think we could only get that from Realtek guys(This stick is using RTL9601D). What we can get now is all inside that github repo. Realtek didn't release SDK source code to public for this chip. :( Best regards, Shengyu 在 2024/2/26 23:42, Russell King (Oracle) 写道: > On Mon, Feb 26, 2024 at 10:51:36PM +0800, Shengyu Qu wrote: >> Hi Russell, >> >> 在 2024/2/26 22:39, Russell King (Oracle) 写道: >>> On Mon, Feb 26, 2024 at 10:16:46PM +0800, Shengyu Qu wrote: >>>> Hi Russell, >>>> >>>>> On Mon, Feb 26, 2024 at 09:23:46PM +0800, Shengyu Qu wrote: >>>>>> ODI DFP-34X-2C2 is capable of 2500base-X, but incorrectly report its >>>>>> capabilities in the EEPROM. >>>>>> So use sfp_quirk_2500basex for this module to allow 2500Base-X mode. >>>>> This was previously submitted by Sergio Palumbo, and comes in two >>>>> different forms - an OEM version and non-OEM. There was extensive >>>>> discussion about this, and the result is that I'm not accepting this >>>>> quirk for this module. >>>>> >>>>> The reason is that the module _defaults_ to 1000base-X and requires >>>>> manual reconfiguration by the user to operate at 2500base-X. >>>>> Unfortunately, there is no way for the kernel to know whether that >>>>> reconfiguration has occurred. >>>> No, In the firmware of this stick, the speed rate is configured to auto >>>> negotiation rather than fixed 1000base-X. >>> >>> How does this "auto negotiation" work? >>> >>> I mean *exactly* how does it work? How does it know whether the host is >>> operating at 1000base-X or 2500base-X? >>> >>> There is *no* inband protocol to allow this to be negotiated. >>> >> Well, that seems some kind weird trick implemented in that chip's SDK (maybe >> hardware?). It would automatically detect the speed rate that host uses and >> switch to that rate. The system log of the stick shows that. > > This sounds racy - between the SFP detecting the speed of the host and > the kernel code reconfiguring the interface. More details please... >
On Mon, Feb 26, 2024 at 11:42:49PM +0800, Shengyu Qu wrote: > Hi Andrew, > > So I should describe in commit message that user should set the stick to > 2.5g/automatic mode and include the link in previous mail? Is that enough? As Russell pointed out, this sounds racy. We need lots of details in the commit message to convince us this is safe, and is not going to cause a regression. If it does break and cause a regression, having those details will also help us decide what to do, other than just revert the change. The further the devices get from well defined standardised behaviour, the more details are needed. Andrew
Hi Russell, I got contact with someone that has access to this chip's datasheet. He tolds me that according to the datasheet, this chip "supports a rate adaptor feature, which is a Realtek proprietary SERDES mechanism." " When the chip's SERDES operates in Rate adaptor mode, the SERDES speed is fixed at 2.5Gbps and the Ethernet speed cannot be higher than the SERDES speed. The SOC need not follow the Ethernet link speed, and cannot change the SERDES speed in this mode. There is a data flow control mechanism to ensure correct data transmission." Is it enough to put this into the commit information? Best regards, Shengyu 在 2024/2/26 23:42, Russell King (Oracle) 写道: > On Mon, Feb 26, 2024 at 10:51:36PM +0800, Shengyu Qu wrote: >> Hi Russell, >> >> 在 2024/2/26 22:39, Russell King (Oracle) 写道: >>> On Mon, Feb 26, 2024 at 10:16:46PM +0800, Shengyu Qu wrote: >>>> Hi Russell, >>>> >>>>> On Mon, Feb 26, 2024 at 09:23:46PM +0800, Shengyu Qu wrote: >>>>>> ODI DFP-34X-2C2 is capable of 2500base-X, but incorrectly report its >>>>>> capabilities in the EEPROM. >>>>>> So use sfp_quirk_2500basex for this module to allow 2500Base-X mode. >>>>> This was previously submitted by Sergio Palumbo, and comes in two >>>>> different forms - an OEM version and non-OEM. There was extensive >>>>> discussion about this, and the result is that I'm not accepting this >>>>> quirk for this module. >>>>> >>>>> The reason is that the module _defaults_ to 1000base-X and requires >>>>> manual reconfiguration by the user to operate at 2500base-X. >>>>> Unfortunately, there is no way for the kernel to know whether that >>>>> reconfiguration has occurred. >>>> No, In the firmware of this stick, the speed rate is configured to auto >>>> negotiation rather than fixed 1000base-X. >>> >>> How does this "auto negotiation" work? >>> >>> I mean *exactly* how does it work? How does it know whether the host is >>> operating at 1000base-X or 2500base-X? >>> >>> There is *no* inband protocol to allow this to be negotiated. >>> >> Well, that seems some kind weird trick implemented in that chip's SDK (maybe >> hardware?). It would automatically detect the speed rate that host uses and >> switch to that rate. The system log of the stick shows that. > > This sounds racy - between the SFP detecting the speed of the host and > the kernel code reconfiguring the interface. More details please... >
On Wed, Feb 28, 2024 at 06:42:55PM +0800, Shengyu Qu wrote: > Hi Russell, > > I got contact with someone that has access to this chip's datasheet. > He tolds me that according to the datasheet, this chip > "supports a rate adaptor feature, which is a Realtek proprietary SERDES > mechanism." > " When the chip's SERDES operates in Rate adaptor mode, the SERDES speed > is fixed at 2.5Gbps and the Ethernet speed cannot be higher than the > SERDES speed. > The SOC need not follow the Ethernet link speed, and cannot change the > SERDES speed in this mode. There is a data flow control mechanism to > ensure correct data transmission." > > Is it enough to put this into the commit information? I think you need to read that again and understand it. Specifically "the SERDES speed is fixed at 2.5Gbps". When operating in this mode (rate adapter mode) this would mean the SERDES (which is what talks to the host) is fixed at 2500base-X and is fundamentally incompatible and incapable of linking with a host operating at 1000base-X. "Rate adapter" is a term that is used by manufacturers to mean that a e.g. PHY is able to operate at several different speeds on its media side while operating at a specific speed on its host side. E.g. a PHY that presents at 10GBASE-R host interface but supports media at 10M, 100M, 1G, 2.5G and 5G speeds. Typically, it controls the rate at which the host sends packets by sending PAUSE frames to the host. This feature is irrelevant to the abilities of the module to link at either 1000base-X or 2500base-X with the host. Thanks.
diff --git a/drivers/net/phy/sfp.c b/drivers/net/phy/sfp.c index f75c9eb3958e..2021cb4ff2f6 100644 --- a/drivers/net/phy/sfp.c +++ b/drivers/net/phy/sfp.c @@ -495,6 +495,13 @@ static const struct sfp_quirk sfp_quirks[] = { // 2500MBd NRZ in their EEPROM SFP_QUIRK_M("Lantech", "8330-262D-E", sfp_quirk_2500basex), + // ODI DFP-34X-2C2 can operate at 2500base-X, but incorrectly report 1300MBd + // NRZ in the EEPROM. + // Besides, In early batches, vendor id is set to OEM, but that is fixed in + // newer batches. + SFP_QUIRK_M("ODI", "DFP-34X-2C2", sfp_quirk_2500basex), + SFP_QUIRK_M("OEM", "DFP-34X-2C2", sfp_quirk_2500basex), + SFP_QUIRK_M("UBNT", "UF-INSTANT", sfp_quirk_ubnt_uf_instant), // Walsun HXSX-ATR[CI]-1 don't identify as copper, and use the
ODI DFP-34X-2C2 is capable of 2500base-X, but incorrectly report its capabilities in the EEPROM. So use sfp_quirk_2500basex for this module to allow 2500Base-X mode. Signed-off-by: Shengyu Qu <wiagn233@outlook.com> --- drivers/net/phy/sfp.c | 7 +++++++ 1 file changed, 7 insertions(+)