Message-ID: <20221121051940.GA9293@pengutronix.de>
Date:   Mon, 21 Nov 2022 06:19:40 +0100
From:   Oleksij Rempel <o.rempel@...gutronix.de>
To:     Devid Antonio Filoni <devid.filoni@...uetechnologies.com>
Cc:     Kurt Van Dijck <dev.kurt@...dijck-laurijssen.be>,
        kbuild test robot <lkp@...el.com>,
        Maxime Jayat <maxime.jayat@...ile-devices.fr>,
        Oliver Hartkopp <socketcan@...tkopp.net>,
        David Jander <david@...tonic.nl>, linux-kernel@...r.kernel.org,
        Oleksij Rempel <linux@...pel-privat.de>,
        netdev@...r.kernel.org, Marc Kleine-Budde <mkl@...gutronix.de>,
        kernel@...gutronix.de, Robin van der Gracht <robin@...tonic.nl>,
        Jakub Kicinski <kuba@...nel.org>,
        Paolo Abeni <pabeni@...hat.com>,
        "David S. Miller" <davem@...emloft.net>, linux-can@...r.kernel.org
Subject: Re: [PATCH RESEND] can: j1939: do not wait 250ms if the same addr
 was already claimed

On Sun, Nov 20, 2022 at 08:18:32PM +0100, Devid Antonio Filoni wrote:
> On Sun, 2022-11-20 at 09:45 +0100, Oleksij Rempel wrote:
> > On Sun, Nov 20, 2022 at 01:11:52AM +0100, Devid Antonio Filoni wrote:
> > > On Sat, 2022-11-19 at 11:12 +0100, Oleksij Rempel wrote:
> > > > On Fri, Nov 18, 2022 at 04:12:40PM +0100, Devid Antonio Filoni wrote:
> > > > > Hi Oleksij,
> > > > > 
> > > > > honestly I would apply proposed patch because it is the easier solution
> > > > > and makes the driver compliant with the standard for the following
> > > > > reasons:
> > > > > - on the first claim, the kernel will wait 250 ms as stated by the
> > > > > standard
> > > > > + on successive claims with the same name, the kernel will not wait
> > > > > 250ms, this implies:
> > > > >   - it will not wait after sending the address-claimed message when the
> > > > > claimed address has been spoofed, but the standard does not explicitly
> > > > > states what to do in this case (see previous emails in this thread), so
> > > > > it would be up to the application developer to decide how to manage the
> > > > > conflict
> > > > >   - it will not wait after sending the address-claimed message when a
> > > > > request for address-claimed message has been received as stated by the
> > > > > standard
> > > > 
> > > > Standard says:
> > > > 1. No CF _shall_ begin, or resume, transmission on the network until 250 ms
> > > >    after it has successfully claimed an address (Figure 4).
> > > > 2. This does not apply when responding to a request for address claimed.
> > > > 
> > > > With current patch state: 1. is implemented and working as expected, 2.
> > > > is not implemented.
> > > > With this patch: 1. is partially broken and 2. is partially faking
> > > > needed behavior.
> > > > 
> > > > It will not wait if remote ECU which rebooted for some reasons. With this patch
> > > > we are breaking one case of the standard in favor to fake compatibility to the
> > > > other case. We should avoid waiting only based on presence of RfAC not based
> > > > on the old_addr == new_addr.
> > > 
> > > I'm sorry, I don't think I understood the point about reboot ("It will
> > > not wait if remote ECU which rebooted for some reasons"). If another ECU
> > > rebooted, then *it* will have to perform the claim procedure again
> > > waiting 250 ms before beginning the transmission. Your ECU doesn't have
> > > to check if the other ECUs respected the 250 ms wait.
> > 
> > With proposed patch:
> > - local application which is sending to the remote NAME, will start or continue
> >   communication with ECU which should stay silent.
> 
> And this is not forbidden by the standard, the standard states that the
> remote ECU shall not start or continue the communication but it can
> *receive* messages.
> For example, what would you do if:
> - during the 250 ms wait, another ECU sends a request-for-address-
> claimed message meant to the address you're claiming?
> From "4.5.3 Other requirements for initialization":
> A CF shall respond to a request-for-address-claimed message when the
> destination address is the same as the CF's address and shall transmit
> its response to the Global address (255).
> - during the 250 ms wait another ECU sends a normal message (non
> address-claim related) using the SA you're currently claiming?
> 
> > - local application which was manually or automatically restarted (see
> >   application watchdogs), will bypass address claim procedure
> >   completion and start sending without 250ms delay.
> 
> Then the application will be violating the standard, you're right,
> however please note that, as per driver implementation, each time the
> socket is closed and opened again (if bound with a name) you have to
> send the address-claimed message again.
> The standard also states how to treat this kind of violations on the
> remote ECU side.
> 
> > 
> > > Also, the ISO11783-5 standard, with "Figure 6 (Resolving address
> > > contention between two self-configurable-address CF)" of "4.5.4.2 -
> > > Address-claim prioritization", shows that:
> > > - ECU1 claims the address (time: 0 ms)
> > > - ECU2 claims the same address (time: 0+x ms)
> > > - ECU1 NAME has the higher priority, so ECU1 sends again the address
> > > claimed message as soon as it received the address-claim from ECU2
> > > (time: 0+x+y ms)
> > > - ECU1 starts normal transmission (time: 250 ms)
> > > With current implementation, the ECU1 would start the transmission at
> > > time 0+x+y+250 ms, with proposed patch it would not.
> > 
> > You are right, this should be fixed.
> > But proposed patch closes one issues and opens another, with this patch it will
> > be enough to send at least two address claimed messages to bypass the delay.
> 
> No, because the timer associated with the first claim *is not stopped*.
> 
> > 
> > > Same is showed in "Figure 7 (Resolving address contention between a non-
> > > configurable address CF and a self-configurable address CF)", the ECU
> > > waits again 250 ms only when claiming a different address.
> > 
> > Ack
> > 
> > > Also, as previously discussed in this thread, the standard states in
> > > 4.4.4.3 - Address violation:
> > > If a CF receives a message, other than the address-claimed message,
> > > which uses the CF's own SA,
> > > then the CF:
> > > - shall send the address-claim message to the Global address;
> > > - shall activate a diagnostic trouble code with SPN = 2000+SA and FMI =
> > > 31
> > > It is not *explicitly* stated that you have to wait 250 ms after the
> > > address-claim message has been sent.
> > 
> > There is no need to explicitly state it. The requirement is clearly described
> > in the 4.5.2.d part 1 with clearly defined exception in  4.5.2.d part 2.
> > If something is not explicitly stated, the stated requirement has always
> > priority.
> > 
> > > Please note that the 250 ms wait is  mentioned only in "4.5 - Network
> > > initialization"
> > 
> > OK, we need to refer to the wording used in a specifications, in
> > general:
> > Shall – Shall is used to designate a mandatory requirement.
> > Should – Should is used for requirements that are considered good and are
> >          recommended, but are not absolutely mandatory.
> > May – May is used to for requirements that are optional.
> > 
> > If a requirement with strong wording as "shall" is not strong enough for
> > you and you are suing words as ".. mentioned only in .." then even a
> > statistical analysis of this spec will have no meaning. In all
> > cases we can just invalidate all arguments by using: it is only X or Y. 
> > 
> > > while above statements come from "4.4 - Network-management procedures".
> > > Also in this case, the proposed patch is still standard compliant.
> > 
> > If we remove 4.5.2.d from the spec, then yes.
> > 
> > > So I'm sorry but I have to disagree with you, there are many things
> > > broken in the current implementation because it is forcing the 250 wait
> > > to all cases but it should not.
> > 
> > If we remove 4.5.2.d from the spec, then yes. Every construction is
> > logical if we adopt input variables to the construction.
> 
> From "4.4.4.3 - Address violation":
> - *shall send the address-claim message* to the Global address
> From "4.5.2 Address claim requirements":
> - No CF shall begin, or resume, transmission on the network until 250 ms
> after it has successfully *claimed an address*, except when responding
> to a request for address-claimed.
> 
> Do you see any difference?
> With your interpretation of the standard, then above 4.5.2.d sentence
> shall be:
> - No CF shall begin, or resume, transmission on the network until 250 ms
> after it has successfully *sent the address-claim message*, except when
> responding to a request for address-claimed.
> 
> I think "it has successfully claimed an address" is valid for the whole
> claim procedure and not for the address-claimed message only.
> 
> Please note that the ECU shall send the address-claim message also when
> it receives a request for a matching NAME ("4.4.3.2 NAME management (NM)
> message"). This does not mean that is claiming again the address.
> 
> > 
> > > > Without words 2. part should be implemented without breaking 1.
> > > > 
> > > > > Otherwise you will have to keep track of above cases and decide if the
> > > > > wait is needed or not, but this is hard do accomplish because is the
> > > > > application in charge of sending the address-claimed message, so you
> > > > > would have to decide how much to keep track of the request for address-
> > > > > claimed message thus adding more complexity to the code of the driver.
> > > > 
> > > > Current kernel already tracks all claims on the bus and knows all registered
> > > > NAMEs. I do not see increased complicity in this case.
> > > 
> > > The kernel tracks the claims but it does *not track* incoming requests
> > > for address-claimed message, it would have to and it would have to
> > 
> > yes
> > 
> > > allow the application to answer to it *within a defined time window*.
> > 
> > yes.
> > 
> > > But keep in mind that there are other cases when the 250 ms wait is wrong
> > > or it is not explicitly stated by the standard.
> > 
> > If it is not stated in the standard how can we decide if it is wrong?
> And how can we decide if it is right? :)
> 
> > And if strongly worded statements have no value just because it is
> > stated only one time, how proper standard should look like? 
> See above.
> 
> > 
> > > > IMHO, only missing part i a user space interface. Some thing like "ip n"
> > > > will do.
> > > > 
> > > > > Another solution is to let the driver send the address-claimed message
> > > > > waiting or without waiting 250 ms for successive messages depending on
> > > > > the case.
> > > > 
> > > > You can send "address-claimed message" in any time you wont. Kernel will
> > > > just not resolve the NAME to address until 1. part of the spec will
> > > > apply. Do not forget, the NAME cache is used for local _and_ remote
> > > > names. You can trick out local system, not remote.
> > > > 
> > > > Even if you implement "smart" logic in user space and will know better
> > > > then kernel, that this application is responding to RfAC. You will newer
> > > > know if address-claimed message of remote system is a response to RfAC.
> > > > 
> > > > From this perspective, I do not know, how allowing the user space break
> > > > the rules will help to solve the problem?
> > > 
> > > I think you did not understand this last proposal: since the driver is
> > > already implementing part of the standard, then it might as well send
> > > the address-claimed message when needed and wait 250 ms or not depending
> > > on the case.
> > 
> > Let's try following test:
> > j1939acd -r 80 -c /tmp/1122334455667788.jacd 11223344556677 vcan0 &
> > while(true); do testj1939 -s8 vcan0:0x80 :0x90,0x12300; done
> > 
> > And start candump with delta time stamps:
> > :~ candump -t d vcan0                                                 
> >  (000.000000)  vcan0  18EAFFFE   [3]  00 EE 00               
> >  (000.002437)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF <---- no 250ms delay
> >  (000.011458)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.011964)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.011712)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012585)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012891)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012082)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012604)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012357)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012790)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012765)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012483)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012680)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012144)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> > ... snip ...
> >  (000.012592)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012515)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.013183)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012653)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.011886)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012836)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.009494)  vcan0  18EEFF80   [8]  77 66 55 44 33 22 11 00 <---- SA 0x80 address claimed 
> >  (000.003362)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF <---- next packet from SA 0x80 3 usecs after previous. No 250ms delay.
> >  (000.012351)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012983)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012602)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012594)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.012348)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> >  (000.011922)  vcan0  19239080   [8]  01 23 45 67 89 AB CD EF
> > 
> > As you can see, the j1939 stack do not forcing application to use NAMEs and
> > do not preventing sending any message withing 250ms delay. The only thing
> > what has the 250 timer is NAME to address resolution which should be fixed in
> > respect of 4.5.2.d without breaking every thing else.
> 
> Yes this is clear, this is working because the socket used by testj1939
> is not bound to any name.
> 
> Just to clarify: are you suggesting to applications developer to use one
> socket (bound with the name) to manage the address-claim and another one
> (bound without the name) for other transmissions? If so, then why that
> code exists in the driver?
> Honestly I would consider this proposal really bad since this would
> allow to completely violate the standard. I really hope you agree with
> me about this.

Hm... you are right.

Please add to your patch code comments with standard snippets and
clarification why it should be so. Commit comment will be often
overseen.

Regards,
Oleksij
-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |

Powered by blists - more mailing lists