Avoid infinite loop in renameat(2)

[kib]

Summary

rename(2): do not allow to rename root vnode of the mounted filesystem

Check for tdvp being vp_crossmp.  This cannot happen for the normal
rename cases, but could if the target path specified by the syscall
points to the nullfs mount over the regular file.  In this case namei()
cannot step over crossmp, and keep it in ni_dvp.

Since crossmp VOP_GETWRITEMOUNT() returns NULL mp, we retry the locking
dance since the belief is that NULL return is transient.

PR:     295826


renameat(2): when retrying, check for pending signals

The vn_start_write() call there is already interruptible.  Check for
user signals before restarting due to ERELOOKUP, or after failed
vn_start_write().  Note that vn_start_write(V_XSLEEP | V_PCATCH)
does not check for signals if not sleeping.

PR:     295826

[kib]

Also check for pending signals after vn_start_write()

[kib]

Tested by: pho

[markj]

sys/kern/vfs_syscalls.c

3862

Don't you need to call vn_finished_write(mp) before returning?

[kib]

Do not leak mnt_writecount.

[markj]

sys/kern/vfs_syscalls.c

3867

Or maybe the check should come before vn_start_write()?

[kib]

Check for signals before taking the write ref.