On 12/7/23 07:46, Christoph Hellwig wrote:
On Wed, Nov 29, 2023 at 05:33:09PM -0800, Bart Van Assche wrote:Revert commit 7b12e49669c9 ("fs: remove fs.f_write_hint") to enable testing write hint support with fio and direct I/O.To enable testing seems like a pretty bad argument for bloating struct file. I'd much prefer to not restore it, but if you want to do so please write a convincing commit log.
Hi Christoph, I have submitted a pull request for fio such that my tests can be run even if F_SET_FILE_RW_HINT is not supported (see also https://github.com/axboe/fio/pull/1682). The only other application that I found that uses F_SET_FILE_RW_HINT is Ceph. Do we want to make the Ceph code work again that uses F_SET_FILE_RW_HINT? I think this code cannot be converted to F_SET_RW_HINT. From the Ceph source code: ---------------------------------------------------------------------- int KernelDevice::choose_fd(bool buffered, int write_hint) const { #if defined(F_SET_FILE_RW_HINT) if (!enable_wrt) write_hint = WRITE_LIFE_NOT_SET; #else // Without WRITE_LIFE capabilities, only one file is used. // And rocksdb sets this value also to > 0, so we need to catch this // here instead of trusting rocksdb to set write_hint. write_hint = WRITE_LIFE_NOT_SET; #endif return buffered ? fd_buffereds[write_hint] : fd_directs[write_hint]; } ---------------------------------------------------------------------- Thanks, Bart.
