os/filestore: fix journal logger

[wjin]

counter type (value 10) shouldn't be decreased in that it may generate
negative value when monitoring.

Signed-off-by: Wei Jin wjin.cn@gmail.com

[liewegas]

@athanatos this was introduced by f54e563 ... it does look like the dec shouldn't be there; can you confirm?

[athanatos]

No, the decrements should be right. Those counters are supposed to be at any moment equal to throttle values, thus, we decrement them when we release the throttle. If you like, you could restructure it to a pair of _taken and _released throttles, but I think it would be harder to read the code. In the current code, it should never go negative. If you saw it underflow, that's a bug and you should find where we are taking throttle without incrementing the counters.

[wjin]

These two items (journal_ops/bytes) were used to count the journal total value according to description.

And from the document about perf counters: If bit 8 is set (counter), the reader may want to subtract off the previously read value to get the delta during the previous interval.
Here it conflicts with the document.

We may have two ways to fix it?

1. change it from value 10 to 2
2. If we want to track 'active journal data' (submitted to memory or committed to journal disk, but not applied to store), we may add another two items.

[wjin]

No, the decrements should be right. Those counters are supposed to be at any moment equal to throttle values, thus, we decrement them when we release the throttle.

Actually, the decrements are still there. Flush is called and I only remove the logger dec.

[liewegas]

the metric type should be changed to u64 then (not u64_counter).

[wjin]

Updated. As for dynamic throttling, I think people are care more about the active journal data instead of total data written to journal for monitoring.

[athanatos]

lgtm