Jens Axboe

@theknowledge That's definitely not right, you probably wanted to have used bisect skip for those. @vincentkfu that seems more reasonable So I wonder if this is really a problem elsewhere....

We're talking about the in-kernel shipped ntfs driver? It doesn't set ->direct_IO at all, so IO should fail with EINVAL and fio should print its usual blurb about that.

That's not a bad idea, but it should be added separately and take the two counts. I'd be happy to take a patch that does that, as I do think...

Yes, that's exactly right, it should be made streaming somehow. We already have the infrastructure, in fact we stream the entries in at init time. We just stream all of...

Indeed, the same kind of chunking could be done on blktrace replay. It should be trivial to do, since the input is an already sorted blktrace trace.

That sounds useful, please do feel free to work on that change.

Just to confirm what @sitsofe said - I'm totally fine with having windows binaries generated in a predictable fashion so that folks can use them. I don't have access to...

> If I'm not wrong, the Windows binaries are build today already, right? That's why they can be grabbed from the Appveyor artifacts, as stated in the README. Right, it's...

How many files are being written in parallel? There's work in progress for supporting more efficient buffered writes, but in terms of the current situation, io_uring will punt the write...

I haven't looked at your benchmark yet, was referring to the file system in the kernel locking the file for write. But in general, NEVER share a ring if you...