BLKDISCARD... off by -4kiB @ >=4GiB?

BLKDISCARD... off by -4kiB @ >=4GiB?

[Maciej Żenczykowski]

On a personal machine with 6.6.13-200.fc39.x86_64 and a 2280 pcie nvme
SSD: Samsung MZ-VKW1T00 SSD
  blkdiscard -v /dev/nvme0n1
seems to issue (per strace):
  ioctl(3, BLKDISCARD, [0, 1024209543168])
but looking at the resulting block device (xxd -a < /dev/nvme0n1)...
it looks like exactly the last 4kiB of every 4GiB is not zero'ed.

Note: this is in spite of running it more than once.
(lsblk -D claims DISC-ALN 0, DISC-GRAN 512B, DISC-MAX 2T, DISC-ZERO 0)

So it seems like there's a bug either in the ssd firmware or in the kernel...?

The issue (at the end of the first 4GiB) goes away if I:
  blkdiscard -v -o 0 -l $[2<<30] /dev/nvme0n1
  blkdiscard -v -o $[2<<30] -l $[2<<30] /dev/nvme0n1

or (for full drive):
  blkdiscard -v -p $[2<<30] /dev/nvme0n1

Looking (but not too deeply) at the kernel code, I don't see any
obvious red herring, but maybe something should just limit to 2GiB?

Although __blkdev_issue_discard's:
  bio->bi_iter.bi_size = req_sects << 9;
seems odd considering
  struct bvec_iter {
    sector_t bi_sector; /* device address in 512 byte sectors */
    unsigned int bi_size; /* residual I/O count */

which should imply 4GiB would need to be stored as a u32 bi_size of 0...
But I do networking, so maybe I'm missing some block/fs convention...
Or maybe this is already splitting into smaller pieces, and then
coallescing, or who knows what...

--
Maciej Żenczykowski, Kernel Networking Developer @ Google