Do not update inode stats in real time - collect during dump

[mykaul]

Do not update inode stats in real time - collect during dump

[mykaul]

I was not suggesting to remove all counters. Only those that are relevant for dump time only.

[jkroonza]

Plus unless it's locked it's wrong anyway as it could be raced unless it's atomic_t types (which does a memory bus lock at time of read-update-store), and those you don't increment and decrement using ++ and/or -- operators.

[mykaul]

Plus unless it's locked it's wrong anyway as it could be raced unless it's atomic_t types (which does a memory bus lock at time of read-update-store), and those you don't increment and decrement using ++ and/or -- operators.

But it done under lock. See

glusterfs/libglusterfs/src/inode.c

Line 2431 in 929d71b

ret = pthread_mutex_trylock(&itable->lock);

[jkroonza]

I think we do need the LRU inode entries in table kept as a counter because we also have LRU limit, which this counter is checked against to determine if we need to invalidate some entries. So in my opinion rather sack the extra cycle or two on insert/remove rather than re-counting every time we need to know how many entries there are (if we have 2m table size that's 32MB of memory being traversed minimum, trashing page caches on the CPU as we go, and that's without traversing the individual chains).