vfs: stop using SAVESTART for rename

[mjg]

Summary

As far as I can say ni_startdir is of no significance for rename. The target buf is used, but that one is covered with SAVENAME.

ufs rename asserts that SAVESTART is passed, but the assert dates back to the original BSD source import and I don't see it to be of any use here. msdosfs had the same assert copy-pasted.

The goal is to eradicate the SAVESTART flag.

This is 2 commits squashed into one review:

commit f32b40a60e76332951e6dabc9977bd7a76e06d48 (HEAD -> kekeacz)
Author: Mateusz Guzik <[e-mail]>
Date:   Mon Mar 7 13:50:45 2022 +0100

    vfs: stop using SAVESTART for rename
    
    ni_startdir has never reached rename routines anyway. Do pass SAVENAME
    instead as the name is used for relookup (if needed).

commit 4b965c66044dfb3a3849db5e5fd5397408c9e244
Author: Mateusz Guzik <[e-mail]>
Date:   Mon Mar 7 16:34:52 2022 +0100

    vfs: make relookup take an additional argument
    
    instead of looking at SAVESTART
    
    This is a step towards removing the flag.

Test Plan

will ask pho

ran several rename tests from stress2, no issues

[mjg]

hm, ext2 is using relookup which looks at SAVESTART. I may need to extend the patch, but the ufs and msdosfs remarks remain.

[mjg]

• added relookup changes

[mjg]

sys/kern/vfs_lookup.c

1439

relookup would preferably drop this feature altogether -- interested parties can just vref on their own, but that's for later

[mckusick]

This change looks reasonable to me.

sys/kern/vfs_lookup.c

1484

This panic should be more descriptive and should have the correct function name.

[markj]

sys/kern/vfs_syscalls.c

3680 ↗

(On Diff #103590)

Looks like zfs_freebsd_rename() asserts SAVESTART is set, for some reason. Doesn't it need to be patched as well?

[mjg]

sys/kern/vfs_syscalls.c

3680 ↗

(On Diff #103590)

it asserts SAVENAME *and/or* SAVESTART, so it wont fire. I'm going to patch it when removing the flag to begin with.

[mjg]

sys/kern/vfs_syscalls.c

3680 ↗

(On Diff #103590)

maybe i should note that SAVESTART implies SAVENAME

[markj]

sys/fs/ext2fs/ext2_vnops.c

913

I'm having trouble understanding this. Without the change, kern_renameat() passes SAVESTART, so tdvp gets ref'ed in relookup(). Where does that ref get released?

sys/kern/vfs_lookup.c

1482

Does this comment still make sense?

sys/kern/vfs_syscalls.c

3680 ↗

(On Diff #103590)

Oops, I see.

[mjg]

sys/fs/ext2fs/ext2_vnops.c

913

i did not track all the releaes, i did verify the code de facto expects SAVESTART and counts line up as long as it is "forced"