mkfs.ext4 vs. e2fsck discard oddities

I've been testing Lukas' last 2 patches for e2fsck discard, and noticed something a little odd.

If I make a 512M file, loopback mount it, and mkfs.ext4 it with discard, it uses about 17M at that point. If I then run fsstress on it with a known seed, then run e2fsck -E discard on it, it uses about 52M.

If I repeat the above test telling mkfs.ext4 NOT to discard, I'm left with about 94M after the discarding e2fsck.

So it seems that perhaps e2fsck is not discarding everything that it could; after a discarding fsck, we should be left with the same (minimal) nr. of blocks "in use" no?

I guess that's better than discarding _more_ than it should though. ;)

(I suppose it is possible that this is the underlying filesytem being selective about which discards it accepts, but it behaves the same way on ext4 and xfs backing filesystems)

-Eric

FWIW, sequence of events here, tested with and without "-K" on mkfs.ext4:

dd if=/dev/zero of=fsfile bs=1M count=512
losetup /dev/loop0 fsfile
mkfs.ext4 -F /dev/loop0&>/dev/null
mount /dev/loop0 mnt/
/root/git/xfstests/ltp/fsstress -s 1 -d mnt/ -n 2000 -p 4 umount mnt/
e2fsck/e2fsck.static -fy -E discard /dev/loop0> fsck1.out || exit du -hc fsfile
losetup -d /dev/loop0

Quoted 8 lines · Eric Sandeen

I've been testing Lukas' last 2 patches for e2fsck discard, and noticed something a little odd.

If I make a 512M file, loopback mount it, and mkfs.ext4 it with discard, it uses about 17M at that point. If I then run fsstress on it with a known seed, then run e2fsck -E discard on it, it uses about 52M.

If I repeat the above test telling mkfs.ext4 NOT to discard, I'm left with about 94M after the discarding e2fsck.

So it seems that perhaps e2fsck is not discarding everything that it could; after a discarding fsck, we should be left with the same (minimal) nr. of blocks "in use" no?

The reason is (as I commented in the patch #2) that we will not discard BLOCK_UNINIT groups. We use BLOCK_UNINIT as a optimization measure to skip groups which are likely to be non-provisioned, because we have never written there anything since the mkfs.

If you create file system without discard, then obviously nothing is discarded, image is fully provisioned and e2fsck discard *only* initialized groups. So you'll end up with the bigger image, in case that your image was not sparse.

I hope that makes sense.

Actually I want to make the same optimization for fitrim. We discussed it with Ted and Phillip (see the discussion under [RESEND] [PATCH 2/2 v2] e2fsck: Do not forget to discard last block group. They did seem to be convinced by that, however I think it is right thing to do for the reasons I gave in that thread.

Thanks!
-Lukas

Quoted 20 lines

I guess that's better than discarding _more_ than it should though. ;)

(I suppose it is possible that this is the underlying filesytem being selective about which discards it accepts, but it behaves the same way on ext4 and xfs backing filesystems)

-Eric

FWIW, sequence of events here, tested with and without "-K" on mkfs.ext4:

dd if=/dev/zero of=fsfile bs=1M count=512
losetup /dev/loop0 fsfile
mkfs.ext4 -F /dev/loop0&>/dev/null
mount /dev/loop0 mnt/
/root/git/xfstests/ltp/fsstress -s 1 -d mnt/ -n 2000 -p 4 umount mnt/
e2fsck/e2fsck.static -fy -E discard /dev/loop0> fsck1.out || exit du -hc fsfile
losetup -d /dev/loop0

The reason is (as I commented in the patch #2) that we will not discard BLOCK_UNINIT groups. We use BLOCK_UNINIT as a optimization measure to skip groups which are likely to be non-provisioned, because we have never written there anything since the mkfs.

If you create file system without discard, then obviously nothing is discarded, image is fully provisioned and e2fsck discard *only* initialized groups. So you'll end up with the bigger image, in case that your image was not sparse.

Quoted 10 lines · Lukas Czerner

The reason is (as I commented in the patch #2) that we will not discard BLOCK_UNINIT groups. We use BLOCK_UNINIT as a optimization measure to skip groups which are likely to be non-provisioned, because we have never written there anything since the mkfs.

If you create file system without discard, then obviously nothing is discarded, image is fully provisioned and e2fsck discard *only* initialized groups. So you'll end up with the bigger image, in case that your image was not sparse.

i still think it makes sense to have an option where we discard everything
including BLOCK_UNINIT blocks. Mke2fs doesn't discard blocks by default
because of a fear of crappy SSD drives, and while that fear may be overstated, assuming that all of the unused blocks will *always* have been
discarded at mkfs time isn't necessarily a good thing to assume. I'll grant
that it might be a fine default, but there needs to be *some* way to discard everything that's unused….

-- Ted

actually mke2fs does discard block by default. It has been like that since the beginning. Back then we only had '-K' argument to 'keep' blocks and do not attempt to discard. Nowadays user can do '-E nodiscard', but it is users choice.

Quoted 5 lines · Lukas Czerner

actually mke2fs does discard block by default. It has been like that since the beginning. Back then we only had '-K' argument to 'keep' blocks and do not attempt to discard. Nowadays user can do '-E nodiscard', but it is users choice.

Ah, you're right. The defaults had changed back and forth a couple of times over time and I had lost track of how things had been settled for mke2fs (which is different from e2fsck). At least at one point it was _not_ the default, and in fact the man page was out of sync with the behavior of the mke2fs.

The point remains the same, though, if the file system was created with mke2fs -E nodiscard, how do you undo that decision if there's no way to force the discard of BLOCK_UNINIT blocks?

- Ted

Well, it is not default right ? So the user should better know what is he doing. Moreover it is not like it is end of the world when we do not provide that option, since SSD's will handle over provisioning to some extent even without slowdown, and as for thin-provisioned devices you should know why you're overriding defaults and what it means for you.

Anyway, if people really want this another option to discard all the block groups including those UNINIT ones, I guess I can not resist that :). '-E discard_all' maybe ?

Quoted 9 lines · Lukas Czerner

Well, it is not default right ? So the user should better know what is he doing. Moreover it is not like it is end of the world when we do not provide that option, since SSD's will handle over provisioning to some extent even without slowdown, and as for thin-provisioned devices you should know why you're overriding defaults and what it means for you.

Anyway, if people really want this another option to discard all the block groups including those UNINIT ones, I guess I can not resist that :). '-E discard_all' maybe ?

I think the option is a little more generic than discard. The uninit groups are not discarded because they are not checked in the first place. A bad group descriptor checksum will force the group to be checked, and thus discarded as well. I think what is needed is an option to trigger the same
thing: force all groups to be checked, even if they are uninit and have good descriptor checksums. Maybe -E thorough?

Why would we try to check UNINIT groups with valid descriptor checksums ? I think that this problem will be solved with BLOCK_DISCARDED flag as we discussed with Ted in another thread. No need to have yet another option so it is win-win :)