Re: [PATCH 2/3] net: dsa: b53: mmap: register MDIO Mux bus controller

[Álvaro Fernández Rojas]

El vie, 17 mar 2023 a las 17:41, Florian Fainelli
(<f.fainelli@xxxxxxxxx>) escribió:
>
> On 3/17/23 09:23, Álvaro Fernández Rojas wrote:
> > El vie, 17 mar 2023 a las 15:29, Vladimir Oltean (<olteanv@xxxxxxxxx>) escribió:
> >>
> >> On Fri, Mar 17, 2023 at 03:17:12PM +0100, Álvaro Fernández Rojas wrote:
> >>>> The proposed solution is too radical for a problem that was not properly
> >>>> characterized yet, so this patch set has my temporary NACK.
> >>>
> >>> Forgive me, but why do you consider this solution too radical?
> >>
> >> Because it involves changing device tree bindings (stable ABI) in an
> >> incompatible way.
> >>
> >>>>
> >>>>> But maybe Florian or Jonas can give some more details about the issue...
> >>>>
> >>>> I think you also have the tools necessary to investigate this further.
> >>>> We need to know what resource belonging to the switch is it that the
> >>>> MDIO mux needs. Where is the earliest place you can add the call to
> >>>> b53_mmap_mdiomux_init() such that your board works reliably? Note that
> >>>> b53_switch_register() indirectly calls b53_setup(). By placing this
> >>>> function where you have, the entirety of b53_setup() has finished
> >>>> execution, and we don't know exactly what is it from there that is
> >>>> needed.
> >>>
> >>> In the following link you will find different bootlogs related to
> >>> different scenarios all of them with the same result: any attempt of
> >>> calling b53_mmap_mdiomux_init() earlier than b53_switch_register()
> >>> will either result in a kernel panic or a device hang:
> >>> https://gist.github.com/Noltari/b0bd6d5211160ac7bf349d998d21e7f7
> >>>
> >>> 1. before b53_switch_register():
> >>>
> >>> 2. before dsa_register_switch():
> >>>
> >>> 3. before b53_switch_init():
> >>
> >> Did you read what I said?
> >
> > Yes, but I didn't get your point, sorry for that.
> >
> >>
> >> | Note that b53_switch_register() indirectly calls b53_setup(). By placing
> >> | this function where you have, the entirety of b53_setup() has finished
> >> | execution, and we don't know exactly what is it from there that is
> >> | needed.
> >>
> >> Can you place the b53_mmap_mdiomux_init() in various places within
> >> b53_setup() to restrict the search further?
> >
> > I tried and these are the results:
> > https://gist.github.com/Noltari/d5bdba66b8f2e392c9e4c2759661d862
> >
> > All of them hang when dsa_tree_setup() is called for DSA tree 1
> > (external switch) without having completely setup DSA tree 0 (internal
> > switch):
> > [ 1.471345] b53-switch 10e00000.switch: found switch: BCM63xx, rev 0
> > [ 1.481099] bcm6368-enetsw 1000d800.ethernet: IRQ tx not found
> > [ 1.506752] bcm6368-enetsw 1000d800.ethernet: mtd mac 4c:60:de:86:52:12
> > [ 1.594365] bcm7038-wdt 1000005c.watchdog: Registered BCM7038 Watchdog
> > [ 1.612008] NET: Registered PF_INET6 protocol family
> > [ 1.645617] Segment Routing with IPv6
> > [ 1.649547] In-situ OAM (IOAM) with IPv6
> > [ 1.653948] NET: Registered PF_PACKET protocol family
> > [ 1.659984] 8021q: 802.1Q VLAN Support v1.8
> > [ 1.699193] b53-switch 10e00000.switch: found switch: BCM63xx, rev 0
> > [ 2.124257] bcm53xx 0.1:1e: found switch: BCM53125, rev 4
> > *** Device hang ***
> >
> > I don't know if there's a way to defer the probe of DSA tree 1 (the
> > external switch) until DSA tree 0 (the internal switch) is completely
> > setup, because that would probably be the only alternative way of
> > fixing this.
>
> Could you find out which part is hanging? It looks like there is a busy
> waiting operation that we never complete?

After many tests I was able to find the part that was hanging the device.
It turns out that if the MDIO bus controller is registered soon
enough, b53_phy_read16 will be called for the RGMII port on the
internal switch:
[ 4.042698] b53-switch 10e00000.switch: b53_phy_read16: ds=81fede80
phy_read16=00000000 addr=4 reg=2
It turns out that the device is hanging on the following line of
b53_phy_read16():
    b53_read16(priv, B53_PORT_MII_PAGE(addr), reg * 2, &value);
Maybe it's not safe to access B53_PORT_MII_PAGE() on MMAP switches?

Only in one specific image in which I had a lot of debugging this
access didn't hang, but it just returned 0:
[ 5.129715] b53_mmap_write16: dev=83547680 page=0 reg=3c val=100
[ 5.135914] b53_mmap_write16: dev=83547680 page=0 reg=3c val=100 done!
[ 5.143721] b53-switch 10e00000.switch: b53_phy_read16: ds=83547580
phy_read16=00000000 addr=4 reg=2
[ 5.153204] b53-switch 10e00000.switch: b53_phy_read16: ds=83547580
phy_read16=00000000 addr=4 reg=2 val=0
[ 5.163171] b53-switch 10e00000.switch: b53_phy_read16: ds=83547580
phy_read16=00000000 addr=4 reg=3
[ 5.172560] b53-switch 10e00000.switch: b53_phy_read16: ds=83547580
phy_read16=00000000 addr=4 reg=3 val=0
[ 5.218764] b53-switch 10e00000.switch: Using legacy PHYLIB callbacks.
Please migrate to PHYLINK!

However, if I implement b53_mmap_phy_read16() and
b53_mmap_phy_write16() in MMAP it seems to solve the issue and the
device doesn't hang anymore:
[ 2.783407] b53-switch 10e00000.switch: found switch: BCM63xx, rev 0
[ 2.951877] b53-switch 10e00000.switch: b53_phy_read16: addr=4 reg=2
[ 2.958393] b53_mmap_phy_read16: dev=836f6580 phy_id=4 loc=2
[ 2.964367] b53_mmap_phy_read16: dev=836f6580 phy_id=4 loc=2 val=362
[ 2.970923] b53-switch 10e00000.switch: b53_phy_read16: addr=4 reg=3
[ 2.977420] b53_mmap_phy_read16: dev=836f6580 phy_id=4 loc=3
[ 2.983315] b53_mmap_phy_read16: dev=836f6580 phy_id=4 loc=3 val=5e80
[ 3.026253] b53-switch 10e00000.switch: Using legacy PHYLIB callbacks.
Please migrate to PHYLINK!
[ 3.072584] b53-switch 10e00000.switch: Configured port 8 for internal
[ 3.082850] DSA: tree 0 setup

However, what I did is just replicating mdio-mux-bcm6368 source code
in MMAP (for the internal PHY only):
static int b53_mmap_phy_read16(struct b53_device *dev, int phy_id, int
loc, u16 *val)
{
        uint32_t reg;

        b53_mmap_write32(dev, 0, REG_MDIOC, 0);

        reg = REG_MDIOC_RD_MASK |
        (phy_id << REG_MDIOC_PHYID_SHIFT) |
        (loc << REG_MDIOC_REG_SHIFT);

        b53_mmap_write32(dev, 0, REG_MDIOC, reg);
        udelay(50);
        b53_mmap_read16(dev, 0, REG_MDIOD, val);

        return 0;
}

static int b53_mmap_phy_write16(struct b53_device *dev, int phy_id,
int loc, u16 val)
{
        uint32_t reg;

        b53_mmap_write32(dev, 0, REG_MDIOC, 0);

        reg = REG_MDIOC_WR_MASK |
        (phy_id << REG_MDIOC_PHYID_SHIFT) |
        (loc << REG_MDIOC_REG_SHIFT) |
        val;

        b53_mmap_write32(dev, 0, REG_MDIOC, reg);
        udelay(50);

        return 0;
}

Is it safe to add those functions in MMAP or is there a way of forcing
the use of mdio-mux-bcm6368 for those PHY accesses?

>
> DSA should be perfectly capable of dealing with disjoint trees being
> cascaded to one another, as this is entirely within how the framework is
> designed.
>
> What I suspect might be happening is a "double programming" effect,
> similar or identical to what was described in this commit:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=b8c6cd1d316f3b01ae578d8e29179f6396c0eaa2
>
> using the MDIO mux would properly isolate the pseudo PHYs of the switch
> such that a given MDIO write does not end up programming *both* the
> internal and external switches. It could also be a completely different
> problem.
> --
> Florian
>

--
Álvaro