On Thu, May 06, 2021 at 08:18:36AM +0100, Christoph Hellwig wrote:

On Wed, May 05, 2021 at 02:27:11PM -0700, Darrick J. Wong wrote:

quoted

Er... what specifically does "data" mean?  File data, or just the dirent
blocks?

I think this is only true if discard_zeroes_data == 1, right?  The last
I looked, ext4 was calling REQ_OP_DISCARD, not REQ_OP_WRITE_ZEROES.

Also, there are some SSDs that "implement" discard as nop, which means
that the old contents can still be read by re-reading the LBAs.  What
about those?

Not just some, but most at least for corner cases.  ATA TRIM, SCSI UNMAP
and NVMe Deallocate all explicitly allow for keeping some of the old
data, and devices make use of that when the discard requests does not
map to their internal granularities.

Heh, so that's a "stable" behavior. :)

quoted

(Also wondering if this is where FS_SECRM_FL files should get their
freed file blocks erased with REQ_OP_SECURE_ERASE...)

Only implemented for mmc..

<shrug> If the wording got tweaked to "...not readable via LBA interface
after delete" then you could also REQ_OP_WRITE_ZEROES, which would work
on a broader range of hardware.

--D