On Tue, 2019-02-26 at 14:34 -0800, Stephen Boyd wrote: > The common clk framework is lacking in ability to describe the clk > topology without specifying strings for every possible parent-child > link. There are a few drawbacks to the current approach: > > 1) String comparisons are used for everything, including describing > topologies that are 'local' to a single clock controller. > > 2) clk providers (e.g. i2c clk drivers) need to create globally unique > clk names to avoid collisions in the clk namespace, leading to awkward > name generation code in various clk drivers. > > 3) DT bindings may not fully describe the clk topology and linkages > between clk controllers because drivers can easily rely on globally unique > strings to describe connections between clks. > > This leads to confusing DT bindings, complicated clk name generation > code, and inefficient string comparisons during clk registration just so > that the clk framework can detect the topology of the clk tree. > Furthermore, some drivers call clk_get() and then __clk_get_name() to > extract the globally unique clk name just so they can specify the parent > of the clk they're registering. We have of_clk_parent_fill() but that > mostly only works for single clks registered from a DT node, which isn't > the norm. Let's simplify this all by introducing two new ways of > specifying clk parents. > > The first method is an array of pointers to clk_hw structures > corresponding to the parents at that index. This works for clks that are > registered when we have access to all the clk_hw pointers for the > parents. > > The second method is a mix of clk_hw pointers and strings of local and > global parent clk names. If the .name member of the map is set we'll > look for that clk by performing a DT based lookup of the device the clk > is registered with and the .name specified in the map. If that fails, > we'll fallback to the .fallback member and perform a global clk name > lookup like we've always done before. > > Using either one of these new methods is entirely optional. Existing > drivers will continue to work, and they can migrate to this new approach > as they see fit. Eventually, we'll want to get rid of the 'parent_names' > array in struct clk_init_data and use one of these new methods instead. > > Cc: Miquel Raynal <[email protected]> > Cc: Jerome Brunet <[email protected]> > Cc: Russell King <[email protected]> > Cc: Michael Turquette <[email protected]> > Cc: Jeffrey Hugo <[email protected]> > Cc: Chen-Yu Tsai <[email protected]> > Signed-off-by: Stephen Boyd <[email protected]> > --- > drivers/clk/clk.c | 260 ++++++++++++++++++++++++++--------- > include/linux/clk-provider.h | 19 +++ > 2 files changed, 217 insertions(+), 62 deletions(-)
Sorry for the delay. With the fix you sent to Jeffrey Tested by porting the aoclk controller of Amlogic g12a SoC. This allowed to test * hws only table * parent_data with a mix of hw pointers and fw_name (with different input controllers and also an input that is optional and never provided) Tested-by: Jerome Brunet <[email protected]> With the small comment below Reviewed-by Jerome Brunet <[email protected]> > > diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c > index 937b8d092d17..3d01e8c56400 100644 > --- a/drivers/clk/clk.c > +++ b/drivers/clk/clk.c > @@ -39,6 +39,13 @@ static LIST_HEAD(clk_notifier_list); > [...] > > +static int clk_cpy_name(const char *dst, const char *src, bool must_exist) > +{ > + if (!src) { > + if (must_exist) > + return -EINVAL; > + return 0; > + } > + > + dst = kstrdup_const(src, GFP_KERNEL); > + if (!dst) > + return -ENOMEM; > + > + return 0; > +} > + > +static int clk_core_populate_parent_map(struct clk_core *core) > +{ > + const struct clk_init_data *init = core->hw->init; > + u8 num_parents = init->num_parents; > + const char * const *parent_names = init->parent_names; > + const struct clk_hw **parent_hws = init->parent_hws; > + const struct clk_parent_data *parent_data = init->parent_data; > + int i, ret = 0; > + struct clk_parent_map *parents, *parent; > + > + if (!num_parents) > + return 0; > + > + /* > + * Avoid unnecessary string look-ups of clk_core's possible parents by > + * having a cache of names/clk_hw pointers to clk_core pointers. > + */ > + parents = kcalloc(num_parents, sizeof(*parents), GFP_KERNEL); > + core->parents = parents; > + if (!parents) > + return -ENOMEM; > + > + /* Copy everything over because it might be __initdata */ > + for (i = 0, parent = parents; i < num_parents; i++, parent++) { > + if (parent_names) { > + /* throw a WARN if any entries are NULL */ > + WARN(!parent_names[i], > + "%s: invalid NULL in %s's .parent_names\n", > + __func__, core->name); > + ret = clk_cpy_name(parent->name, parent_names[i], > + true); > + } else if (parent_data) { While testing, I mistakenly left both parent_names and parent_data. I was surprised that parent_data did not take precedence of parent_names. Maybe it should ? (... but I understand we are not supposed to provide both) > + parent->hw = parent_data[i].hw; > + ret = clk_cpy_name(parent->fw_name, > + parent_data[i].fw_name, false); > + if (!ret) > + ret = clk_cpy_name(parent->name, > + parent_data[i].name, > + false); > + } else if (parent_hws) { > + parent->hw = parent_hws[i]; > + } else { Maybe there should also some kinda of check to verify if more than one option (among hws, parent_data and parent_names) was provided and throw a warn ? Could be useful with drivers move away from parent_names ? > + ret = -EINVAL; > + WARN(1, "Must specify parents if num_parents > 0\n"); > + } > + > + if (ret) { > + do { > + kfree_const(parents[i].name); > + kfree_const(parents[i].fw_name); > + } while (--i >= 0); > + kfree(parents); > + > + return ret; > + } > + } > + > + return 0; > +} > + >
