Re: [PATCH v9 2/5] firmware: psci: Read and use vendor reset types

From: Lorenzo Pieralisi
Date: Fri Mar 14 2025 - 07:19:56 EST


On Mon, Mar 03, 2025 at 01:08:31PM -0800, Elliot Berman wrote:
> From: Elliot Berman <elliot.berman@xxxxxxxxxxxxxxxx>
>
> SoC vendors have different types of resets and are controlled through
> various registers. For instance, Qualcomm chipsets can reboot to a
> "download mode" that allows a RAM dump to be collected. Another example
> is they also support writing a cookie that can be read by bootloader
> during next boot. PSCI offers a mechanism, SYSTEM_RESET2, for these
> vendor reset types to be implemented without requiring drivers for every
> register/cookie.
>
> Add support in PSCI to statically map reboot mode commands from
> userspace to a vendor reset and cookie value using the device tree.

I have managed to discuss a little bit this patchset over the last
few days and I think we have defined a plan going forward.

A point that was raised is:

https://man7.org/linux/man-pages/man2/reboot.2.html

LINUX_REBOOT_CMD_RESTART2 *arg command, what is it supposed to
represent ?

Is it the mode the system should reboot into OR it is the
actual command to be issued (which is what this patchset
implements) ?

LINUX_REBOOT_CMD_RESTART "..a default restart..."

It is unclear what "default" means. We wonder whether the
reboot_mode variable was introduced to _define_ that "default".

So, in short, my aim is trying to decouple reboot_mode from the
LINUX_REBOOT_CMD_RESTART2 *arg command.

I believe that adding a sysfs interface to reboot-mode driver
infrastructure would be useful, so that the commands would
be exposed to userspace and userspace can set the *arg command
specifically to issue a given reset/mode.

I wonder why this is not already in place for eg syscon-reboot-mode
resets, how does user space issue a command in those systems if the
available commands aren't exposed to userspace ?

Is there a kernel entity exposing those "modes" to userspace, somehow ?

> A separate initcall is needed to parse the devicetree, instead of using
> psci_dt_init because mm isn't sufficiently set up to allocate memory.
>
> Reboot mode framework is close but doesn't quite fit with the
> design and requirements for PSCI SYSTEM_RESET2. Some of these issues can
> be solved but doesn't seem reasonable in sum:
> 1. reboot mode registers against the reboot_notifier_list, which is too
> early to call SYSTEM_RESET2. PSCI would need to remember the reset
> type from the reboot-mode framework callback and use it
> psci_sys_reset.
> 2. reboot mode assumes only one cookie/parameter is described in the
> device tree. SYSTEM_RESET2 uses 2: one for the type and one for
> cookie.

This can be changed and I think it should, so that the reboot modes
are exposed to user space and PSCI can use that.

> 3. psci cpuidle driver already registers a driver against the
> arm,psci-1.0 compatible. Refactoring would be needed to have both a
> cpuidle and reboot-mode driver.
>
> Signed-off-by: Elliot Berman <elliot.berman@xxxxxxxxxxxxxxxx>
> ---
> drivers/firmware/psci/psci.c | 105 +++++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 105 insertions(+)
>
> diff --git a/drivers/firmware/psci/psci.c b/drivers/firmware/psci/psci.c
> index a1ebbe9b73b136218e9d9f9b8daa7756b3ab2fbe..6f8c47deaec0225f26704e1f3bcad52603127a85 100644
> --- a/drivers/firmware/psci/psci.c
> +++ b/drivers/firmware/psci/psci.c
> @@ -80,6 +80,14 @@ static u32 psci_cpu_suspend_feature;
> static bool psci_system_reset2_supported;
> static bool psci_system_off2_hibernate_supported;
>
> +struct psci_reset_param {
> + const char *mode;
> + u32 reset_type;
> + u32 cookie;
> +};
> +static struct psci_reset_param *psci_reset_params __ro_after_init;
> +static size_t num_psci_reset_params __ro_after_init;
> +
> static inline bool psci_has_ext_power_state(void)
> {
> return psci_cpu_suspend_feature &
> @@ -306,9 +314,39 @@ static int get_set_conduit_method(const struct device_node *np)
> return 0;
> }
>
> +static int psci_vendor_system_reset2(const char *cmd)
> +{
> + unsigned long ret;
> + size_t i;
> +
> + for (i = 0; i < num_psci_reset_params; i++) {
> + if (!strcmp(psci_reset_params[i].mode, cmd)) {
> + ret = invoke_psci_fn(PSCI_FN_NATIVE(1_1, SYSTEM_RESET2),
> + psci_reset_params[i].reset_type,
> + psci_reset_params[i].cookie, 0);
> + /*
> + * if vendor reset fails, log it and fall back to
> + * architecture reset types

That's not what the code does.

> + */
> + pr_err("failed to perform reset \"%s\": %ld\n", cmd,
> + (long)ret);
> + return 0;
> + }
> + }
> +
> + return -ENOENT;
> +}
> +
> static int psci_sys_reset(struct notifier_block *nb, unsigned long action,
> void *data)
> {
> + /*
> + * try to do the vendor system_reset2
> + * If there wasn't a matching command, fall back to architectural resets
> + */
> + if (data && !psci_vendor_system_reset2(data))
> + return NOTIFY_DONE;
> +
> if ((reboot_mode == REBOOT_WARM || reboot_mode == REBOOT_SOFT) &&
> psci_system_reset2_supported) {
> /*
> @@ -795,6 +833,73 @@ static const struct of_device_id psci_of_match[] __initconst = {
> {},
> };
>
> +#define REBOOT_PREFIX "mode-"
> +
> +static int __init psci_init_system_reset2_modes(void)
> +{
> + const size_t len = strlen(REBOOT_PREFIX);
> + struct psci_reset_param *param;
> + struct device_node *psci_np __free(device_node) = NULL;
> + struct device_node *np __free(device_node) = NULL;
> + struct property *prop;
> + size_t count = 0;
> + u32 magic[2];
> + int num;
> +
> + if (!psci_system_reset2_supported)
> + return 0;
> +
> + psci_np = of_find_matching_node(NULL, psci_of_match);
> + if (!psci_np)
> + return 0;
> +
> + np = of_find_node_by_name(psci_np, "reset-types");
> + if (!np)
> + return 0;

Related to my initial question above. If LINUX_REBOOT_CMD_RESTART2 *arg command,
is the actual reset to be issued, should we add a default mode "cold"
and, if SYSTEM_RESET2 is supported, a "warm" reset mode too ?

It all boils down to what *arg represents - adding "cold" and "warm"
modes would remove the dependency on reboot_mode for resets issued
through LINUX_REBOOT_CMD_RESTART2, the question is whether this
is the correct thing to do.

Comments very welcome.

Thanks,
Lorenzo

> +
> + for_each_property_of_node(np, prop) {
> + if (strncmp(prop->name, REBOOT_PREFIX, len))
> + continue;
> + num = of_property_count_u32_elems(np, prop->name);
> + if (num != 1 && num != 2)
> + continue;
> +
> + count++;
> + }
> +
> + param = psci_reset_params =
> + kcalloc(count, sizeof(*psci_reset_params), GFP_KERNEL);
> + if (!psci_reset_params)
> + return -ENOMEM;
> +
> + for_each_property_of_node(np, prop) {
> + if (strncmp(prop->name, REBOOT_PREFIX, len))
> + continue;
> +
> + num = of_property_read_variable_u32_array(np, prop->name, magic,
> + 1, ARRAY_SIZE(magic));
> + if (num < 0) {
> + pr_warn("Failed to parse vendor reboot mode %s\n",
> + param->mode);
> + kfree_const(param->mode);
> + continue;
> + }
> +
> + param->mode = kstrdup_const(prop->name + len, GFP_KERNEL);
> + if (!param->mode)
> + continue;
> +
> + /* Force reset type to be in vendor space */
> + param->reset_type = PSCI_1_1_RESET_TYPE_VENDOR_START | magic[0];
> + param->cookie = num > 1 ? magic[1] : 0;
> + param++;
> + num_psci_reset_params++;
> + }
> +
> + return 0;
> +}
> +arch_initcall(psci_init_system_reset2_modes);
> +
> int __init psci_dt_init(void)
> {
> struct device_node *np;
>
> --
> 2.34.1
>