On 06/28/2016 06:39 PM, Christoph Hellwig wrote: > On Tue, Jun 28, 2016 at 04:56:30PM +0300, Boaz Harrosh wrote: >> Actually with O_APPEND each write request should write a different region >> of the file, there are no overlapping writes. And no issue of which version of the >> write came last and got its data written. > > You have one fd for multiple threads or processes (it doesn't matter if > you're using O_APPEND or not), and all of them write to it. > Yes so? but they do not write to the same (overlapping) region of the file, each thread usually writes to his own record. > i_size is only updated once the write finishes, so having multiple > concurrent writes will mean multiple records go into the same regions. > Now to be fair in current XFS writes beyond i_size will always take > the lock exclusively, so for this case we will not get concurrent > writes and thus data corruption anyway. Exactly that I understand. And it must be so. > But if you have a cycling > log that gets overwritten (say a database journal) we're back to > square one. > No! In this "cycling log" case the application (the DB) has an Head and a Tail pointers each thread grabs the next available record and writes to it. The IO is not overlapping, each thread writes to his own record, and even if they write at the same time they do not over-write each other. As long as they properly sync on the Head pointer the write itself can happen in parallel. This is not a good example and will work perfectly well with the old (current) DAX code. (Even if such records where 4k aligned) >> I still don't see how an application can use the fact that two writers >> will not give them mixed records. And surly it does not work on a shared >> FS. So I was really wondering if you know of any such app > > If it doesn't work for two threads using the same fd on a shared fs > the fs is broken. What works? the above cycling log, sure it will work, also on current dax code. I wish you could write a test to demonstrate this bug. Sorry for my slowness but I don't see it. I do not see how the fact that there is only a single memcpy in progress can help an application. Yes sure isize update must be synced, which it is in current code. Thanks Christoph, I've taken too much of your time, I guess the QA will need to bang every possible application and see what breaks when multiple parallel writers are allowed on a single file. So far for whatever I use in a VM it all works just the same. Boaz _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
