io-threads, till a few weeks back, used to channel all operations on a particular inode over the same thread so that there are no race conditions due to writes over-taking each other.

Recent changes in iothreads have ensured that this race condition is hit very easily. io-threads fires new threads to handle requests when it sees that the queue length is increasing beyond a certain point. Queue length increases can happen due to a slow disk or due to O_SYNC or O_DIRECT writes and in other cases also. The logic has also changed to allow multiple threads to serve fops for the same file. The race condition occurs in posix, where the lseek-readv or lseek-writev sequence is not a critical section, such that lseek of thread1 can be followed by lseek of thread2, resulting in a write on offset of thread2 by both threads. This results in a data corruption.