On Sat, Jun 22, 2019 at 08:10:29PM +0200, Michael Olbrich wrote:
> On Sat, Jun 22, 2019 at 05:53:18PM +0100, Russell King - ARM Linux admin wrote:
> > Old code:
> >
> > - while (!(ret = readl_relaxed(sdma->regs + SDMA_H_INTR) & 1)) {
> > - if (timeout-- <= 0)
> > - break;
> > - udelay(1);
> > - }
> >
> > So, while bit 0 is _clear_ the loop continues to poll.
> >
> >
> > New code:
> >
> > + ret = readl_relaxed_poll_timeout_atomic(sdma->regs + SDMA_H_STATSTOP,
> > + reg, !(reg & 1), 1, 500);
> >
> > Doesn't really tell us what the termination condition is (because of
> > the obfuscation taking away the details), but if we dig into the
> > macro maze:
> >
> > #define readl_relaxed_poll_timeout_atomic(addr, val, cond, delay_us, timeout_us) \
> > readx_poll_timeout_atomic(readl_relaxed, addr, val, cond, delay_us, timeout_us)
> >
> > #define readx_poll_timeout_atomic(op, addr, val, cond, delay_us, timeout_us) \
> > ({ \
> > u64 __timeout_us = (timeout_us); \
> > unsigned long __delay_us = (delay_us); \
> > ktime_t __timeout = ktime_add_us(ktime_get(), __timeout_us); \
> > for (;;) { \
> > (val) = op(addr); \
> > if (cond) \
> > break; \
> >
> > "cond" is passed in to here unmodified, so this becomes:
> >
> > for (;;) {
> > reg = readl_relaxed(sdma->regs + SDMA_H_STATSTOP);
> > if (!(reg & 1))
> > break;
> >
> > So, if bit 0 of this register is clear, we terminate the loop.
> >
> > Seems to me like this is a great illustration why using a helper
> > _introduces_ bugs, because it hides the detail about what the exit
> > condition for the embedded loop actually is, and leads to this kind
> > of error.
> >
> > In any case, the conversion is obviously incorrect.
> >
> > I occasionally see the "Timeout waiting for CH0 ready" error during
> > boot on a cbi4, which, given the above, means that we did end up
> > seeing bit 1 set (so according to the old code, we waited
> > successfully.)
>
> The old code was polling SDMA_H_INTR so it waited for the bit to be set.
> The new code (as documented in the commit message) polls SDMA_H_STATSTOP
> instead.
> I believe this register is called SDMAARM_STOP_STAT in the reference
> manual. And the documentation states: "Reading this register yields the
> current state of the HE[i] bits".
> And from the documentation of the SDMA "DONE" instruction:
> "Clear HE bit for the current channel, send an interrupt to the Arm
> platform for the current channel and reschedule."
>
> My interpretation of this is, that waiting for the bit in SDMA_H_STATSTOP
> to become zero has the same effect as waiting for the bit in SDMA_H_INTR to
> be set. Or am I missing something?
So, why do all my iMX6 platforms now randomly spit out:
"imx-sdma 20ec000.sdma: Timeout waiting for CH0 ready"
at boot, whereas they didn't used to with older kernels? Maybe channel
0 does not clear the HE[0] bit?
The documentation explicitly states that for initialisation, the
following is required:
• Set bit 0 of the SDMA_HSTART register to set HE[0] and allow Channel 0
to run (assumes EO[0] and DO[0] were both set in previous step). This
will cause SDMA toload the program RAM and channel contexts configured
previously.
• Wait for Channel 0 to finish running. This is indicated by HI[0]=1 in
the SDMA_SDMA_INTR register, or by optional interrupt to the ARM platform.
So, is there a way for a HI bit to be set without clearing the HE bit?
Yes, via the NOTIFY command:
55.5.2.35 NOTIFY (Notify to ARM platform)
Operation:
if (jjj & 4 == 0)
{
if (jjj&2 == 2)
HE[CCR] ← 0
if (jjj&1== 1)
HI[CCR] ← 1
}
else if (jjj == 4)
EP[CCR] ← 0
else
So, if jjj is 001 binary, the HE bit can remain set while the HI bit
is cleared. Maybe the firmware uses this rather than a DONE instruction
when performing the initialisation functions, which means your idea of
going against what is specified in the manual, and using HE[0] instead
of HI[0] is on _very_ shakey ground.
Given that I'm seeing the same issue on _four_ iMX6 platforms here,
I think it's pretty much obvious that your assumptions here are
false.
> Michael
>
> > Looking at the date of the commit, this is almost a three year old
> > bug.
> >
> > --
> > RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
> > FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
> > According to speedtest.net: 11.9Mbps down 500kbps up
> >
>
> --
> Pengutronix e.K. | |
> Industrial Linux Solutions | http://www.pengutronix.de/ |
> Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
> Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
>
--
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
According to speedtest.net: 11.9Mbps down 500kbps up
