led_class: storing a value can act but return -EINVAL
When I store something into the brightness sysfs attribute of an LED, it will accept the value but return -EINVAL:
johannes:/sys/class/leds/pmu-front-led# echo 255 > brightness bash: echo: write error: Invalid argument
(yet the LED turns on)
This happens because the store callback doesn't consume all the input.
There are two possible ways to handle this:
a) accept anything that begins with a valid number.
b) reject anything that isn't *only* a number
The following patch implements a), for b) you'd have to make the if statement have ' && after-buf == size' instead of this patch.
I don't know which approach is generally preferred, but acting and then returning an error value doesn't seem nice.
Maybe b) should be implemented instead to stop people from storing things like '0 hahaha stupid kernel ignores this' into the attribute :)
johannes
drivers/leds/led-class.c +1 -1
@@ -45,7 +45,7 @@ static ssize_t led_brightness_store(struunsigned long state =simple_strtoul (buf, &after,10 );if (after - buf >0 ) { - ret = after - buf; + ret = size;led_set_brightness (led_cdev, state); }
Hi!
Quoted 9 lines
When I store something into the brightness sysfs attribute of an LED, it will accept the value but return
-EINVAL:johannes:/sys/class/leds/pmu-front-led# echo 255 > brightness bash: echo: write error: Invalid argument(yet the LED turns on)
This happens because the store callback doesn't consume all the input.
Well, I'd argue current behaviour is okay... can you strace it? It should accept the number (return 3) then return -EINVAL.
Quoted 3 lines
There are two possible ways to handle this:
a) accept anything that begins with a valid number.
b) reject anything that isn't *only* a number
c) accept anything that is number, ignore newlines.
a) is just way too ugly...
Pavel
Signature · 1 line
Thanks, Sharp!
Hi,
Quoted 2 lines
Well, I'd argue current behaviour is okay... can you strace it? It should accept the number (return 3) then return
-EINVAL.
That's exactly what happens.
Which is totally bogus, because userspace will think that the setting didn't succeed. Or application authors will ignore the return value assuming that it always succeeded. Or read the value back to see if it succeeded. All icky, when we can well have a good return value.
Quoted 5 lines
Quoted 3 lines
There are two possible ways to handle this:
a) accept anything that begins with a valid number.
b) reject anything that isn't *only* a numberc) accept anything that is number, ignore newlines.
Which is kinda hard to implement.
Quoted 1 line
a) is just way too ugly...
Well, I'd argue that it doesn't matter much since sysfs values are by definition a single value per file, so you'll already know that putting multiple "values" in is bogus.
johannes
Quoted 5 lines · Johannes Berg
Hi,
Quoted 2 lines
Well, I'd argue current behaviour is okay... can you strace it? It should accept the number (return 3) then return
-EINVAL.
Quoted 12 lines
That's exactly what happens.
Which is totally bogus, because userspace will think that the setting didn't succeed. Or application authors will ignore the return value assuming that it always succeeded. Or read the value back to see if it succeeded. All icky, when we can well have a good return value.
Quoted 5 lines
Quoted 3 lines
There are two possible ways to handle this:
a) accept anything that begins with a valid number.
b) reject anything that isn't *only* a numberc) accept anything that is number, ignore newlines.
echo 255> brightness works, returns success. echo 255 > brightness works but then returns -EINVAL.
So we currently do b, quite strictly. Its the trailing space thats the problem. It also shouldn't have altered the brightness value if it ends up returning -EINVAL.
I've looked around other implementations and it would appear we should accept an optional space. Most sysfs attributes seem to handle this differently, each with its own "bugs".
I've some fixes in mind both for the led and backlight classes which I'll post once I've done a little more testing. I'd be interested to know the official view on what the attributes should/shouldn't accept is.
Cheers,
Richard
Quoted 1 line · Richard Purdie
echo 255> brightness works, returns success.
??
For me (bash) that doesn't do anything useful.
Were you looking for "echo -n 255 > brightness"?
Quoted 5 lines
echo 255 > brightness works but then returns
-EINVAL.So we currently do b, quite strictly. Its the trailing space thats the problem. It also shouldn't have altered the brightness value if it ends up returning
-EINVAL.
Yes, but you do change the actual value, which IMHO you shouldn't when it will return -EINVAL. I should have said
b) reject anything that isn't *only* a number and take no action
instead.
Quoted 3 lines
I've looked around other implementations and it would appear we should accept an optional space. Most sysfs attributes seem to handle this differently, each with its own "bugs".
Yeah, unfortunately that is true. Maybe there should've been helper functions like when you have a sysfs-int attribute that is set directly without get/set calls. I'd suggest looking at that code.
Quoted 4 lines
I've some fixes in mind both for the led and backlight classes which I'll post once I've done a little more testing. I'd be interested to know the official view on what the attributes should/shouldn't accept is.
I have a question about the backlight class: I'm writing a patch currently to control the *keyboard* backlight on powerbooks, is that appropriate for the backlight class (setting the fbdev callback to reject everything)?
johannes
Hi!
Quoted 4 lines
Quoted 2 lines
Well, I'd argue current behaviour is okay... can you strace it? It should accept the number (return 3) then return
-EINVAL.That's exactly what happens.
And that's pretty much exactly okay.
You got success return when you wrote "255" there, then you got -EINVAL when you tried to put newline there. Expected behaviour, I'd say.
...I can see it looks ugly when you do echo manually, but you simply should not do that.
Quoted 4 lines
Which is totally bogus, because userspace will think that the setting didn't succeed. Or application authors will ignore the return value assuming that it always succeeded. Or read the value back to see if it succeeded. All icky, when we can well have a good return value.
Well, they got what they deserve, that \n does not belong there. And they _did_ get success report -- "3" was returned. That's how unix works, I'd say. Educate userspace authors that adding \n is bad idea.
Pavel
Signature · 1 line
Thanks for all the (sleeping) penguins.