On Mon, Jun 8, 2020 at 2:51 PM Drew Fustini <drew@xxxxxxxxxxxxxxx> wrote:
> This patch causes pcs_parse_pinconf() to return -ENOTSUPP when no
> pinctrl_map is added. The current behavior is to return 0 when
> !PCS_HAS_PINCONF or !nconfs. Thus pcs_parse_one_pinctrl_entry()
> incorrectly assumes that a map was added and sets num_maps = 2.
>
> Analysis:
> =========
> The function pcs_parse_one_pinctrl_entry() calls pcs_parse_pinconf()
> if PCS_HAS_PINCONF is enabled. The function pcs_parse_pinconf()
> returns 0 to indicate there was no error and num_maps is then set to 2:
>
> 980 static int pcs_parse_one_pinctrl_entry(struct pcs_device *pcs,
> 981 struct device_node *np,
> 982 struct pinctrl_map **map,
> 983 unsigned *num_maps,
> 984 const char **pgnames)
> 985 {
> <snip>
> 1053 (*map)->type = PIN_MAP_TYPE_MUX_GROUP;
> 1054 (*map)->data.mux.group = np->name;
> 1055 (*map)->data.mux.function = np->name;
> 1056
> 1057 if (PCS_HAS_PINCONF && function) {
> 1058 res = pcs_parse_pinconf(pcs, np, function, map);
> 1059 if (res)
> 1060 goto free_pingroups;
> 1061 *num_maps = 2;
> 1062 } else {
> 1063 *num_maps = 1;
> 1064 }
>
> However, pcs_parse_pinconf() will also return 0 if !PCS_HAS_PINCONF or
> !nconfs. I believe these conditions should indicate that no map was
> added by returning -ENOTSUPP. Otherwise pcs_parse_one_pinctrl_entry()
> will set num_maps = 2 even though no maps were successfully added, as
> it does not reach "m++" on line 940:
>
> 895 static int pcs_parse_pinconf(struct pcs_device *pcs, struct device_node *np,
> 896 struct pcs_function *func,
> 897 struct pinctrl_map **map)
> 898
> 899 {
> 900 struct pinctrl_map *m = *map;
> <snip>
> 917 /* If pinconf isn't supported, don't parse properties in below. */
> 918 if (!PCS_HAS_PINCONF)
> 919 return 0;
> 920
> 921 /* cacluate how much properties are supported in current node */
> 922 for (i = 0; i < ARRAY_SIZE(prop2); i++) {
> 923 if (of_find_property(np, prop2[i].name, NULL))
> 924 nconfs++;
> 925 }
> 926 for (i = 0; i < ARRAY_SIZE(prop4); i++) {
> 927 if (of_find_property(np, prop4[i].name, NULL))
> 928 nconfs++;
> 929 }
> 930 if (!nconfs)
> 919 return 0;
> 932
> 933 func->conf = devm_kcalloc(pcs->dev,
> 934 nconfs, sizeof(struct pcs_conf_vals),
> 935 GFP_KERNEL);
> 936 if (!func->conf)
> 937 return -ENOMEM;
> 938 func->nconfs = nconfs;
> 939 conf = &(func->conf[0]);
> 940 m++;
>
> This situtation will cause a boot failure [0] on the BeagleBone Black
> (AM3358) when am33xx_pinmux node in arch/arm/boot/dts/am33xx-l4.dtsi
> has compatible = "pinconf-single" instead of "pinctrl-single".
>
> The patch fixes this issue by returning -ENOSUPP when !PCS_HAS_PINCONF
> or !nconfs, so that pcs_parse_one_pinctrl_entry() will know that no
> map was added.
>
> Logic is also added to pcs_parse_one_pinctrl_entry() to distinguish
> between -ENOSUPP and other errors. In the case of -ENOSUPP, num_maps
> is set to 1 as it is valid for pinconf to be enabled and a given pin
> group to not any pinconf properties.
>
> [0] https://lore.kernel.org/linux-omap/20200529175544.GA3766151@x1/
>
> Fixes: 9dddb4df90d1 ("pinctrl: single: support generic pinconf")
> Signed-off-by: Drew Fustini <drew@xxxxxxxxxxxxxxx>
Patch applied as non-critical (for-next) fix.
If there is no hurry let's merge it this way with lots of testing
along the way.
Yours,
Linus Walleij
