Page MenuHomeFreeBSD

Discourage the use of sysexits(3) in new code
Needs ReviewPublic

Authored by 0mp on Nov 11 2020, 12:18 PM.
Referenced Files
Unknown Object (File)
Sat, Jul 13, 2:32 AM
Unknown Object (File)
Tue, Jul 9, 12:19 AM
Unknown Object (File)
Tue, Jul 2, 5:31 AM
Unknown Object (File)
Fri, Jun 28, 5:15 PM
Unknown Object (File)
Wed, Jun 26, 1:39 PM
Unknown Object (File)
Tue, Jun 25, 9:20 PM
Unknown Object (File)
Jun 22 2024, 3:30 AM
Unknown Object (File)
Jun 16 2024, 5:30 PM


Group Reviewers

Commit message:

Discourage the use of sysexits(3) in new code

This commit removes an incorrect statement from the sysexits(3) manual page
about the documented exit values being encouraged by style(9). This has not
been true since 2008 (r186224).

In addition, a new section is added at the top of the manual to discourage
developers from using sysexits(3) in new code.  The usual problems people
have with sysexits(3) is that they are not portable, it's hard to pick
a good value in most situations and that good error messages are probably
better anyway.

Obtained from:	OpenBSD (partially)

Related links:

Diff Detail

rS FreeBSD src repository - subversion
Lint Passed
No Test Coverage
Build Status
Buildable 34742
Build 31798: arc lint + arc unit

Event Timeline

0mp requested review of this revision.Nov 11 2020, 12:18 PM
0mp added a subscriber: des.
  • Remove "preferable" from the manual page title

Exit statuses should implement a protocol between the calling and called process. Since only 8 bits (or 32 if the calling process uses waitid()) are available, there is not much flexibility. I think distinctions between different exit statuses should have a purpose, while most of the sysexits codes categorize errors without a clear purpose. If more flexibility is needed, a channel with more capacity should be used.

I think only EX_TEMPFAIL (75) has a clear and reasonably generic purpose, and I think it is fine to use it when the calling program expects this status, which requires the called program to be aware of the context to some degree. Alternatively, a temporary failure might be considered "false" (1) and a persistent failure "failure" (2), per the second below protocol. This would probably a more consistent way to write lockf(1) in 2020, but changing it now is probably not a good idea.

More portable conventions are as follows:

The two commonly used protocols are:

  • 0 = success, >0 = failure
  • 0 = true, 1 = false, >1 = failure

Additionally, there are common conventions when a program starts another:

  • 126 if the other program was found but could not be run
  • 127 if the other program could not be found
  • 128 plus the signal number if the process terminated abnormally (note: POSIX only specifies "greater than 128" and that kill -l exitstatus can be used to find the signal's short name)

I don't think sysexits(3) has ever been meant to be a replacement of English error messages for human interpretation.

I think only EX_TEMPFAIL (75) has a clear and reasonably generic purpose,

when composing D27161 , i found the only unambiguous error code to be EX_OSERR which i used whenever a syscall failed

yuripv added inline comments.

broken link?

emaste added inline comments.

Can remove "Do not use them" the "should be considered deprecated and not used in new code." is sufficient.

My opinion is that, whilst they are by no means completely unambiguous and fully descriptive, picking a value that's close to right so the user has some idea what the error was even in the absence of error messages is better than just using a hard-coded wholly-meaningless 1 everywhere, as will inevitably happen (and already does when not using sysexits) as a result of this.

Oops, I'm just seeing this.
We've been working on a duplicate on GitHub after a comment on a different PR asked "if this is no longer used that should be mentioned somewhere (including in the header itself).

Personally, I like sysexits and wish an improved version of that could be a standard interface. Cathedrals are beautiful. But I'm very happy to participate in improve consistency if we want to go the other way.

So I was aware of this and preferred the pull request. It documented the current policy that's been in place for a while.
I appreciate the views expressed here, but I think it is better to try to resolve this by getting a broader consensus than we can get in a phab review.
I kinda think the project went a little far in subscribing to bde's views on this, but I don't think we'd be well served making a big change to this policy (which has a lot of lurking emotion behind it) via only a phab request.
I do think there's some sympathy for documenting / standardizing at least some of these values, but also a recognition that some are overlapping and confusing and were just Eric's first guess at what to do in sendmail for different errors.

90% of the argument against sysexits is “Bruce didn't like them”. As he is no longer with us, I suggest we change our stance to:

  1. sysexits are permitted but not required
  2. if you do use them, use them consistently
  3. don't add them to existing code without good reason
  4. conversely, don't remove them from existing code without good reason
  5. mixed usage (within a single program) is a good reason (to either remove or apply them consistently, as you wish)