On 9/18/2024 5:31 PM, Christoph Hellwig wrote: > On Wed, Sep 18, 2024 at 01:42:51PM +0530, Kanchan Joshi wrote: >> Would you prefer a new queue attribute (say nr_streams) that tells that? > > No. For one because using the same file descriptors as the one used > to set the hind actually makes it usable - finding the block device > does not. And second as told about half a dozend time for this scheme > to actually work on a regular file the file system actually needs the > arbiter, as it can work on top of multiple block devices, consumes > streams, might export streams even if the underlying devices don't and > so on. > FS managed/created hints is a different topic altogether, and honestly that is not the scope of this series. That needs to be thought at per-FS level due to different data/meta layouts. This scope of this series is to enable application-managed hints passing through the file system. FS only needs to pass what it receives. No active decision making (since application is doing that). Whether it works fine or not - is application's problem. But due to the simplicity it scales across filesystems. This is for the class of applications that know about their data and have decided to be in control. Regardless, since placement-hints are not getting the reception I imagined, I will backtrack.
