Hi Geert,

On 9/19/18 3:16 PM, Geert Uytterhoeven wrote:

Hi Eric,

On Wed, Sep 19, 2018 at 2:09 PM Auger Eric [off-list ref] wrote:

quoted

On 9/17/18 6:39 PM, Geert Uytterhoeven wrote:

quoted

In some SoCs multiple hardware blocks may share a reset control.
The existing reset control API for shared resets will only assert such a
reset when the drivers for all hardware blocks agree.
The existing exclusive reset control API still allows to assert such a
reset, but that impacts all other hardware blocks sharing the reset.

Sometimes a driver needs to reset a specific hardware block, and be 100%
sure it has no impact on other hardware blocks.  This is e.g. the case
for virtualization with device pass-through, where the host wants to
reset any exported device before and after exporting it for use by the
guest, for isolation.

Hence a new flag for dedicated resets is added to the internal methods,
with a new public reset_control_get_dedicated() method, to obtain an
exclusive handle to a reset that is dedicated to one specific hardware
block.

This supports both DT-based and lookup-based reset controls.

Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
---
v4:
  - New.

Notes:
  - Dedicated lookup-based reset controls were not tested,
  - Several internal functions now take 3 boolean flags, and should
    probably be converted to take a bitmask instead,
  - I think __device_reset() should call __reset_control_get() with
    dedicated=true.  However, that will impact existing users,

why should it?

device_reset{,_optional}() are supposed to reset the passed device.
If the reset is not dedicated, doing so will reset other devices, too.

ok, that's not obvious too me but I am not familiar enough with the API
and existing callers.

quoted

quoted

--- a/drivers/reset/core.c
+++ b/drivers/reset/core.c

@@ -459,9 +459,38 @@ static void __reset_control_put_internal(struct reset_control *rstc)
      kref_put(&rstc->refcnt, __reset_control_release);
 }

+static bool __of_reset_is_dedicated(const struct device_node *node,
+                                 const struct of_phandle_args args)
+{
+     struct of_phandle_args args2;
+     struct device_node *node2;
+     int index, ret;
+
+     for_each_node_with_property(node2, "resets") {
+             if (node == node2)
+                     continue;
+
+             for (index = 0; ; index++) {
+                     ret = of_parse_phandle_with_args(node2, "resets",
+                                                      "#reset-cells", index,
+                                                      &args2);
+                     if (ret)
+                             break;
+
+                     if (args2.np == args.np &&
+                         args2.args_count == args.args_count &&
+                         !memcmp(args2.args, args.args,
+                                 args.args_count * sizeof(args.args[0])))
+                             return false;

You need to call of_node_put(args2.np) (see of_parse_phandle_with_args
kernel doc)

Thanks, nice catch!

quoted

Isn't it sufficient to check device_node handles are equal?

That would make it work with #reset-cells == 0 only.
If #reset-cells > 0, the reset line specifier includes extra arguments.

On the Renesas SoCs I'm using, there's a single reset controller, so
args.np is always the same.  The actual reset line is specified by
args.args[0].  See the "resets" properties in e.g.
https://git.kernel.org/pub/scm/linux/kernel/git/geert/renesas-drivers.git/tree/arch/arm64/boot/dts/renesas/r8a7795.dtsi

OK get it now. Thank you for the explanations.

Best Regards

Eric

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds