On 25.07.2011 01:01, Andi Kleen wrote:
quoted
I wasn't clear enough on that: We only track read errors, here. Ans
error correction can only happen on the read path. So if the write
attempt fails, we can't go into a loop.
Not in a loop, but you trigger more IO errors, which can be nasty
if the IO error logging triggers more IO (pretty common because
syslogd calls fsync). And then your code does even more IO, floods
more etc.etc. And the user will be unhappy if their
console gets flooded.
Okay, I see your point now. Thanks for pointing that out.
We've have a similar problems in the past with readahead causing
error flooding.
Any time where an error can cause more IO you have to be extremly
careful.
Right now this seems rather risky to me.
Hum. This brings up a lot of questions. Would you consider throttling an
appropriate solution to prevent error flooding? What would you use as a
base? A per device counter (which might be misleading if there are more
layers below)? A per filesystem counter (which might need
configurability)? Should those "counters" regenerate over time? Any
other approaches?
-Jan