Organizing kernel messages [LWN.net]

Organizing kernel messages

Posted Oct 8, 2010 19:39 UTC (Fri) by roelofs (guest, #2599)
In reply to: Organizing kernel messages by giraffedata
Parent article: Organizing kernel messages

Message manual technology goes back to a time when disk space and memory were so expensive that the OS couldn't afford to display more than a code; paper and operator time were cheaper.

Memory is still expensive, it's just that now we measure the cost in performance (think "cache miss"). A paragraph of useful description for each of 75,000 printks? No thanks... (Not that most developers are willing or even necessarily capable of writing a paragraph of useful documentation.)

And, of course, it need not stop there: a wiki could provide not only the paragraph-length descriptions but also user experiences, e.g., "I also saw that in this other situation that seemed unrelated, but because of foo bug and baz fallthrough, it triggered."

And with detailed error messages, I think searching the code or the web for the text would be as easy as searching for an arbitrary message identifier.

Um...have you ever tried? Finding copies of the same message is trivial--millions of logs are exposed on the Net. Finding the one-in-10000 web page where somebody quotes the message and actually describes the fix (or at least the underlying problem) is damn near impossible. (Kernel errors might be more amenable to it than gcc/binutils errors, but I wouldn't bet on it. Even the number of forums containing orphaned what-does-this-mean / how-do-I-fix-this? postings--or postings with useless / outdated / incorrect responses--is ridiculously high.)

Greg

to post comments

Organizing kernel messages

Posted Oct 9, 2010 16:28 UTC (Sat) by giraffedata (guest, #1954) [Link]

Memory is still expensive, it's just that now we measure the cost in performance (think "cache miss"). A paragraph of useful description for each of 75,000 printks? No thanks

I don't see how long error messages would cause appreciable cache misses or other slowdowns. I note that printks rarely execute and string literals are stored in a separate page from the rest of the stuff the program is likely to be accessing.

I hope there aren't people scrimping on messages (in any program) because they think they cost too much performance.

And with detailed error messages, I think searching the code or the web for the text would be as easy as searching for an arbitrary message identifier.
Um...have you ever tried? Finding copies of the same message is trivial--millions of logs are exposed on the Net. Finding the one-in-10000 web page where somebody quotes the message and actually describes the fix

I agree that searching the web for message text is a poor way to diagnose a problem for the reasons you give (so I'd prefer that the program just tell me what the problem is), but how is searching the web for a message number any better? I've done both.

Note that I switched context in my original post, addressing two separate advantages of message numbers discussed in the article: 1) message manual ala VMS, and 2) Google.