[PATCH 6/9] of: add support for 'dynamic' DT property

[Frank Rowand]

Hi Matt, Greg,

On 10/8/21 1:51 PM, Frank Rowand wrote:
> On 10/6/21 7:09 PM, Zev Weiss wrote:
>> Nodes marked with this (boolean) property will have a writable status
>> sysfs file, which can be used to toggle them between "okay" and
>> "reserved", effectively hot-plugging them.  Note that this will only
>> be effective for devices on busses that register for OF reconfig
>> notifications (currently spi, i2c, and platform), and only if
>> CONFIG_OF_DYNAMIC is enabled.
>>
>> Signed-off-by: Zev Weiss <zev at bewilderbeest.net>
>> ---
>>  drivers/of/kobj.c | 69 +++++++++++++++++++++++++++++++++++++++++++++++
>>  1 file changed, 69 insertions(+)
>>
>> diff --git a/drivers/of/kobj.c b/drivers/of/kobj.c
>> index 378cb421aae1..141ae23f3130 100644
>> --- a/drivers/of/kobj.c
>> +++ b/drivers/of/kobj.c
>> @@ -36,6 +36,69 @@ static ssize_t of_node_property_read(struct file *filp, struct kobject *kobj,
>>      return memory_read_from_buffer(buf, count, &offset, pp->value, pp->length);
>>  }
>>  
>> +static ssize_t of_node_status_write(struct file *filp, struct kobject *kobj,
>> +                                    struct bin_attribute *bin_attr, char *buf,
>> +                                    loff_t offset, size_t count)
>> +{
>> +    int rc;
>> +    char *newstatus;
>> +    struct property **deadprev;
>> +    struct property *newprop = NULL;
>> +    struct property *oldprop = container_of(bin_attr, struct property, attr);
>> +    struct device_node *np = container_of(kobj, struct device_node, kobj);
>> +
>> +    if (WARN_ON_ONCE(strcmp(oldprop->name, "status")))
>> +            return -EIO;
>> +
>> +    if (offset)
>> +            return -EINVAL;
>> +
>> +    if (sysfs_buf_streq(buf, count, "okay") || sysfs_buf_streq(buf, count, "ok"))
>> +            newstatus = "okay";
>> +    else if (sysfs_buf_streq(buf, count, "reserved"))
>> +            newstatus = "reserved";
>> +    else if (sysfs_buf_streq(buf, count, "disabled"))
>> +            newstatus = "disabled";
>> +    else
>> +            return -EINVAL;
>> +
>> +    if (!strcmp(newstatus, oldprop->value))
>> +            return count;
>> +
> 
> If the general approach of this patch set is the correct way to provide the desired
> functionality (I'm still pondering that), then a version of the following code

After pondering, this approach does not appear workable to me.

If we allow one property to be writable via sysfs we open the door for any property to
be writable from sysfs.  This will likely lead to a desire to modify more than one
related property as a single transaction (so that the changes occur as a single
transaction, under a single lock event, with a single notification after all
of the properties are modified).  This is not meant to be an exhaustive list of
the issues that have already been thought through in the context of overlays
(though not all of the issues have been addressed with overlays, at least many
of them, such as one transaction to apply an entire overlay, have been.)

I don't want to make this a long missive, but will at least note the next
issue that popped up in my pondering, which is complications from modifying
the same items in a devicetree via different methods, such as both writing
to sysfs and applying/removing overlays.  If the problems in the previous
paragraph are not sufficient to prevent the sysfs approach then I can
elaborate further on these additional issues.

So another approach is needed.  I have no yet thought this through, but I
have an alternative.  First, change the new property name from "dynamic"
to something more descriptive like "ownership_shifts_between_os_and_others"
(yes, my suggestions is way too verbose and needs to be word smithed, but
the point is to clearly state the underlying action that occurs), then
define the result of this variable to be driver specific, where the
driver is required upon probe to instantiate the device in a manner
that does not impact the other user(s) of the underlying hardware
and to use a driver specific method to transfer control of the
hardware between the os and the other user(s).  I propose the method
would be via a device specific file (or set of files) in sysfs (Greg's
input invited on the use of sysfs in this manner - if I recall correctly
this is the current preferred mechanism).

-Frank


> probably belongs in drivers/of/dynamic.c so that it is easier to maintain and keep
> consistent with other dynamic devicetree updates.  If you look at the code there
> that touches deadprops (eg __of_changeset_entry_apply()) you will notice that the
> locking issues are more extensive than what is implemented here.
> 
> I'm still thinking about how this interacts with other forms of dynamic devicetree
> changes (eg drivers/of/dynamic.c and also overlays).
> 
>> +    /*
>> +     * of_update_property_self() doesn't free replaced properties, so
>> +     * rifle through deadprops first to see if there's an equivalent old
>> +     * status property we can reuse instead of allocating a new one.
>> +     */
>> +    mutex_lock(&of_mutex);
>> +    for (deadprev = &np->deadprops; *deadprev; deadprev = &(*deadprev)->next) {
>> +            struct property *deadprop = *deadprev;
>> +            if (!strcmp(deadprop->name, "status") &&
>> +                deadprop->length == strlen(newstatus) + 1 &&
>> +                !strcmp(deadprop->value, newstatus)) {
>> +                    *deadprev = deadprop->next;
>> +                    deadprop->next = NULL;
>> +                    newprop = deadprop;
>> +                    break;
>> +            }
>> +    }
>> +    mutex_unlock(&of_mutex);
>> +
>> +    if (!newprop) {
>> +            newprop = kzalloc(sizeof(*newprop), GFP_KERNEL);
>> +            if (!newprop)
>> +                    return -ENOMEM;
>> +
>> +            newprop->name = oldprop->name;
>> +            newprop->value = newstatus;
>> +            newprop->length = strlen(newstatus) + 1;
>> +    }
>> +
>> +    rc = of_update_property_self(np, newprop, true);
> 
> -Frank
> 
>> +
>> +    return rc ? rc : count;
>> +}
>> +
>>  /* always return newly allocated name, caller must free after use */
>>  static const char *safe_name(struct kobject *kobj, const char *orig_name)
>>  {
>> @@ -79,6 +142,12 @@ int __of_add_property_sysfs(struct device_node *np, struct property *pp)
>>      pp->attr.size = secure ? 0 : pp->length;
>>      pp->attr.read = of_node_property_read;
>>  
>> +    if (!strcmp(pp->name, "status") && of_property_read_bool(np, "dynamic")) {
>> +            pp->attr.attr.mode |= 0200;
>> +            pp->attr.write = of_node_status_write;
>> +            pp->attr.growable = true;
>> +    }
>> +
>>      rc = sysfs_create_bin_file(&np->kobj, &pp->attr);
>>      WARN(rc, "error adding attribute %s to node %pOF\n", pp->name, np);
>>      return rc;
>>