> -----Original Message----- > From: Frantisek Rysanek <frantisek.rysa...@post.cz> > Sent: Friday, May 26, 2023 9:01 AM > To: linuxptp-devel@lists.sourceforge.net > Subject: [Linuxptp-devel] Side note: i225 / igc time sync and TSN > capabilities... > > Dear gents, > > this is just a slightly off topic gratuitous side note. > > In the context of fumbling in the "igb" driver in the other thread, I > also took a look inside driver "igc" for comparison = to see what the > i225 had to offer. > Unsurprisingly, the register map related to time sync features seems > quite similar to that of the i210. The source code of igc_ptp.c is > pretty clean. The datasheet for the i225 is not public yet and I > don't have it, so I cannot comment further. > > Interestingly to me, the TIMADJ register is now mentioned, but only > to set some novel flag/bit in that register during port init/reset... > For stepwise time adjustment, the SYSTIM registers get written > directly :-) >
Hi, I didn't work directly on the i350 code, but I recall historically that the TIMADJ registers were poorly designed and didn't map very well onto the actual adjtime callback. Assuming the i350 is anything like i210, here's the description of TIMADJ from i210's datasheet: Dynamic update of SYSTIM registers can be done by using the TIMADJ registers by the following flow. It can also be done by adjusting the INC_TIME as described later in this section. Adjusting the time by TIMADJ are meant to be used only when the time difference between the master and the slave are small enough (at least smaller than one 8th of the time between consecutive SYNC cycles). If this assumption is incorrect, than this process might take longer time than the SYNC cycle to take effect. If this is an issue, software might need to set the SYSTIM by direct access as described in the “Initial Setting” phase: — Write the Tadjust value and its Sign to the TIMADJ register (the Sign bit indicates if the Tadjust value should be added or subtracted) — Following the write access to the TIMADJ register, the hardware repeats the following two steps at each 8 nsec clock as long as the Tadjust > zero. • SYSTIM = SYSTIM + INC_TIME +/- 1 nsec. Add or subtract 1 nsec is defined by TIMADJ.Sign (while 0bmeans Add and 1b means Subtract) • Tadjust = Tadjust - 1 nsec • Note that the SYSTIM timer is incremented monotonically at all times. When updating the SYSTIM by the TIMADJ and concurrent non-zero TIMINCA, the SYSTIM is incremented each clock by steps in the range of 6.5ns up to 9.5ns units. — As shown above, the time adjustment might take multiple clocks. Software might write a new value to the TIMADJ register before the hardware completed the previous adjustment. In such a case, the new value written by software, overrides the above equation. If such a race is not desired, the software could check that the previous adjustment is completed by one of the following methods: • Wait enough time before accessing the TIMADJ register which guarantees that the previous update procedure is completed. • Poll the matched TSICR.TADJ flag which is set by the hardware each time the update procedure is completed. • Enable the TADJ interrupt by setting the TADJ flag in the TSIM register and enable timesync interrupts by setting the Time_Sync flag in the IMS register. The TADJ interrupt indicates that the hardware completed the adjustment procedure. This method is unlikely to be used in nominal operation since the expected adjustments are in the sub s range. TL;DR; the TIMADJ basically just amounts to a +1 or -1 to the increment value beyond the range of the INC_TIME register. It doesn't actually do an immediate atomic adjustment. The delay makes it not as useful, and the fact that you can't enable it "permanently" makes it not useful for performing frequency adjustment as we'd have to keep re-applying the value in order to keep the +1 for longer periods. We could apply the adjustment and then poll until it completes which might take time and delay the adjtime operation, or we could set a timer and block new adjtime operations until the thing completes... Thanks, Jake _______________________________________________ Linuxptp-devel mailing list Linuxptp-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/linuxptp-devel
