On Monday 14 September 2015 06:19 PM, Russell King - ARM Linux wrote:

On Mon, Sep 14, 2015 at 06:03:40PM +0530, Vaibhav Hiremath wrote:

quoted

On Monday 14 September 2015 04:20 PM, Russell King - ARM Linux wrote:

quoted

On Mon, Sep 14, 2015 at 03:45:43PM +0530, Vaibhav Hiremath wrote:

quoted

Came across below lines in the datasheet,

========= Copy-n-paste from datasheet============

All SDH interfaces share the same clock which is enabled when any of the SDH
clock enables are
set (from PMUA_SDH1_CLK_RES_CTRL, PMUA_SDH2_CLK_RES_CTRL,
PMUA_SDH3_CLK_RES_CTRL, PMUA_SDH4_CLK_RES_CTRL,
PMUA_SDH5_CLK_RES_CTRL), with clock source select and divider ratio
controlled by
PMUA_SDH1_CLK_RES_CTRL.

==================================================


And I can confirm that after disabling AXI interface clock for all the
SDH modules (1-5) I see I get an abort.

This clearly explains/justifies/proves that the existing code is
working as expected. I have eMMC mounted on the board, which makes
clock to always stay ON on SDH3.

So there is an OR gate implemented inside which takes input from
SDHx_AXI_EN and feeds back to all SDHx instances. Don't ask me why it
has been designed that way :)

And I did some experiment as well, so what I have observed is,
SDH_AXI_CLOCK is required to generate card detection, without that I do
not see card detection working.

What that means is that if DT configures the interface to use its
internal card detection, the AXI clock must never shut off when entering
runtime-PM.

Yes, exactly.
Its clock driver which is doing this.

static struct mmp_param_gate_clk apmu_gate_clks[] = {
         /* The gate clocks has mux parent. */
         {PXA1928_CLK_SDH0, "sdh0_clk", "sdh_div", CLK_SET_RATE_PARENT,
PXA1928_CLK_SDH0 * 4, 0x1b, 0x1b, 0x0, 0, &sdh0_lock},
};

Here the mask and enable_val both are set to 0x1b, which means it
shuts off both peripheral and AXI clock both.

*Sigh*.

So, rather than represent the hardware in DT by telling the SDHCI code
that you have two independent clocks, you've chosen to do the brain
dead thing, and combine the two clocks *which are logically different*
to suit the Linux implementation, thereby leaking implementation details
into DT, and creating a compatibility headache through doing so.

Stop doing this.  Just stop it.  It's broken, it's wrong, and it's down
right idiotic.

DT should _always_ describe the bloody hardware, not the implementation.

I suggest that you have only one way to solve this on this platform.

1. Declare new clocks from your PXA clock driver which expose the AXI
    and peripheral clock separately.
2. Add code (if not already there) to the SDHCI crap-pile to claim and
    appropriately manage these two clocks, falling back to the existing
    but broken method with existing DT.
3. Change DT to provide the new clocks to SDHCI.

That's about the only way I can see to ensure that an old DT file would
continue to work with a new kernel, given this fsckup.

sdhci-pxav3 driver already handles both, in terms of 'core' and 'io'
clock. And it does disable both the clocks in runtime_pm_suspend()

So I have to find a way not to disable AXI clock here.

And yes, both clocks should be defined separately.

Looping Rob here as well.
Rob please add if I am missing anything here.