On 20-08-25 14:07:29, Philipp Zabel wrote:
> On Tue, 2020-08-25 at 14:24 +0300, Abel Vesa wrote:
> [...]
> > > > +static int imx_blk_ctrl_reset_set(struct reset_controller_dev *rcdev,
> > > > + unsigned long id, bool assert)
> > > > +{
> > > > + struct imx_blk_ctrl_drvdata *drvdata = container_of(rcdev,
> > > > + struct imx_blk_ctrl_drvdata, rcdev);
> > > > + unsigned int offset = drvdata->rst_hws[id].offset;
> > > > + unsigned int shift = drvdata->rst_hws[id].shift;
> > > > + unsigned int mask = drvdata->rst_hws[id].mask;
> > > > + void __iomem *reg_addr = drvdata->base + offset;
> > > > + unsigned long flags;
> > > > + unsigned int asserted_before = 0, asserted_after = 0;
> > > > + u32 reg;
> > > > + int i;
> > > > +
> > > > + spin_lock_irqsave(&drvdata->lock, flags);
> > > > +
> > > > + for (i = 0; i < drvdata->rcdev.nr_resets; i++)
> > > > + if (drvdata->rst_hws[i].asserted)
> > > > + asserted_before++;
> > > > +
> > > > + if (asserted_before == 0 && assert)
> > > > + pm_runtime_get(rcdev->dev);
> > >
> > > Shouldn't that be pm_runtime_get_sync() ?
> > >
> > > I would do that unconditionally before locking drvdata->lock and then
> > > drop unnecessary refcounts afterwards.
> > >
> >
> > I thought we already discussed this on the last's version thread.
>
> This is about something different. pm_runtime_get() just queues the
> device to be enabled at a later point, but I presume you want to have it
> enabled before writing to its registers. (The question here is can you
> write to the registers, and have the device update its internal state,
> while the power domain is disabled?)
> Either way, if you want the reset to be asserted after the function
> returns (as is required by the reset API), as I understand it, you have
> to make sure that the power domain is activated before the function
> returns.
> Therefore pm_runtime_get_sync() is required instead of pm_runtime_get(),
> and that must be called outside of the spin locked section. My
> suggestion would be:
>
> if (assert)
> pm_runtime_get_sync();
> spin_lock_irqsave();
> /* ... */
> spin_unlock_irqrestore();
> if (assert && asserted_before)
> pm_runtime_put();
>
On a second thought this doesn't work because, for the first assertion,
the runtime put will never be called, if the asserted_before does not count
the current assertion. If it counts the current assertion, then every assertion
will end with runtime put. None of these options work here.
How about the following:
if (assert && !test_and_set_bit(1, &drvdata->rst_hws[id].asserted))
pm_runtime_get_sync(rcdev->dev);
spin_lock_irqsave(&drvdata->lock, flags);
reg = readl(reg_addr);
if (assert)
writel(reg & ~(mask << shift), reg_addr);
else
writel(reg | (mask << shift), reg_addr);
spin_unlock_irqrestore(&drvdata->lock, flags);
if (!assert && test_and_clear_bit(1, &drvdata->rst_hws[id].asserted))
pm_runtime_put(rcdev->dev);
This would only call the get_sync/put once for each reset bit.
> unless the following might be an issue:
>
> > > > +
> > > > + if (assert) {
> > > > + reg = readl(reg_addr);
> > > > + writel(reg & ~(mask << shift), reg_addr);
> > > > + drvdata->rst_hws[id].asserted = true;
> > > > + } else {
> > > > + reg = readl(reg_addr);
> > > > + writel(reg | (mask << shift), reg_addr);
>
> Could this cause problems if the power domain is already disabled? If
> so, it would be best to either temporarily enable power, or to skip the
> register writes if asserted_before == 0 && !assert.
>
> > > > + drvdata->rst_hws[id].asserted = false;
> > > > + }
> > > > +
> > > > + for (i = 0; i < drvdata->rcdev.nr_resets; i++)
> > > > + if (drvdata->rst_hws[i].asserted)
> > > > + asserted_after++;
> > > > +
> > > > + if (asserted_before == 1 && asserted_after == 0)
> > > > + pm_runtime_put(rcdev->dev);
> > > > +
> > > > + spin_unlock_irqrestore(&drvdata->lock, flags);
> > > > +
> > > > + return 0;
> > > > +}
>
> regards
> Philipp
