Hi Doug,

在 2016/5/12 1:37, Doug Anderson 写道:

Hi,

On Tue, May 10, 2016 at 12:31 PM, David Wu [off-list ref] wrote:

quoted

  static void rk3x_i2c_adapt_div(struct rk3x_i2c *i2c, unsigned long clk_rate)
  {
         struct i2c_timings *t = &i2c->t;
         struct rk3x_i2c_calced_timings calc;
         u64 t_low_ns, t_high_ns;
+       u32 val;
         int ret;

-       ret = rk3x_i2c_calc_divs(clk_rate, t, &calc);
+       ret = i2c->soc_data->calc_timings(clk_rate, t, &calc);
         WARN_ONCE(ret != 0, "Could not reach SCL freq %u", t->bus_freq_hz);

-       clk_enable(i2c->clk);
+       clk_enable(i2c->pclk);
+
         i2c_writel(i2c, (calc.div_high << 16) | (calc.div_low & 0xffff),
                    REG_CLKDIV);
-       clk_disable(i2c->clk);
+
+       val = i2c_readl(i2c, REG_CON);
+       val &= ~REG_CON_TUNING_MASK;
+       val |= calc.tuning;
+       i2c_writel(i2c, val, REG_CON);

Another subtle bug here.  You need to be holding the spinlock here
since you're doing a read-modify-write of a register that is also
touched by the interrupt handler.  We never needed it before because
the previous register update wasn't touched by anyone else and it was
a single atomic write.

Also: technically if we are midway through a transfer when all this
happens then there will be a very short period of time when the two
timing-related registers won't match with each other.  I have no idea
how much that would matter, but in the very least it seems wise to
minimize the time where they mismatch.  So I'd probably write:

        spin_lock_irqsave(&i2c->lock, flags);
        val = i2c_readl(i2c, REG_CON);
        val &= ~REG_CON_TUNING_MASK;
        val |= calc.tuning;
        i2c_writel(i2c, val, REG_CON);
        i2c_writel(i2c, (calc.div_high << 16) | (calc.div_low & 0xffff),
                   REG_CLKDIV);
        spin_unlock_irqrestore(&i2c->lock, flags);

...if we really end up with on a system with a dynamically changing
clock that uses the new-style timing and we see real problems, we can
always try to come up with a way to avoid any problems.  Sound OK?

Good, add spin_lock is very necessary for atomic write here, thanks for 
your advice.

Otherwise, I think things look good to me.  Caesar's comments would
also be good to fix.


-Doug