On Tue, May 03, 2022 at 05:04:06PM +0200, Miklos Szeredi wrote:

On Tue, 3 May 2022 at 16:53, Greg KH [off-list ref] wrote:

quoted

On Tue, May 03, 2022 at 05:39:46PM +0300, Amir Goldstein wrote:

quoted

quoted

It should be noted that while this API mandates text keys,
it does not mandate text values, so for example, sb iostats could be
exported as text or as binary struct, or as individual text/binary records or
all of the above.

Ugh, no, that would be a total mess.  Don't go exporting random binary
structs depending on the file, that's going to be completely
unmaintainable.  As it is, this is going to be hard enough with random
text fields.

As for this format, it needs to be required to be documented in
Documentation/ABI/ for each entry and key type so that we have a chance
of knowing what is going on and tracking how things are working and
validating stuff.

My preference would be a single text value for each key.

Yes!  That's the only sane way to maintain apis, and is why we do that
for sysfs.

If the key isn't present, there's no value, so userspace "knows" this
automatically and parsing this is trivial.

Contents of ":mnt:info" contradicts that, but mountinfo has a long
established, well documented format, and nothing prevents exporting
individual attributes with separate names as well (the getvalues(2)
patch did exactly that).

I understand, for "legacy" things like this, that's fine, but don't add
new fields or change them over time please, that way just gets us back
to the nightmare of preserving /proc/ file apis.

thanks,

greg k-h