On Wed, Oct 11, 2023 at 12:10:25PM -0700, Darrick J. Wong wrote: > I'm pretty sure I've seen some NVME SSDs where you can issue devicewide > DISCARDs and slowly watch the namespace utilization go down over tens of > minutes; and reads will only eventually start returning zeroes. Well, the second part is broken. The first part is fine, and I've briefly consulted with a firmware team implementing such a feature. It just needs to make sure to return zeroes right after the return of the discard even if the blocks aren't erased yet, including after a powerfail. (anyone who knows the XFS truncate / hole punch code will have a vague idea of how that could work). > However, that's orthogonal to this patch -- if the device doesn't > support discard, _scratch_mkfs won't zero the entire disk to remove old > dead superblocks that might have been written by previous tests. After > we shatter the primary super, the xfs_repair scanning code can still > trip over those old supers and break the golden output. True. I have to admit I stopped reading the patch after the unmap description. I'll take another look.
