[jserv]

Since 6.7, simplefs_kill_sb() read sb->s_fs_info to release the external journal device, but the VFS calls kill_sb even when fill_super failed, where that pointer is NULL (or already freed by fill_super's error path). This is the unmount crash described at sysprog21/rv32emu#827 (comment). kill_sb also dropped the journal device before put_super had finished using it. The journal device is now released at the end of simplefs_put_super(), as 5.10 to 6.5 already did, and kill_sb is just kill_block_super(); fill_super clears s_fs_info when it frees sbi.

The same route also had a handful of error-path bugs, fixed here: simplefs_get_dev_journal() returned NULL where the caller checks IS_ERR(), a double iput() after d_make_root() failure, a double brelse() of the superblock buffer, path references leaked or read after path_put(), and a bogus return value from option parsing.

The branch also carries the two fixes needed to get CI green, which were already failing on master. simplefs_write() returned 0 once the write position hit the maximum file size, and a caller that retries zero-byte writes spins forever; the CI image now ships uutils dd, which does, so test_too_large_file hung until the 6 hour job limit. It now returns -EFBIG as write(2) specifies and the test expects the failure. The coding_style job also flagged three clang-format deviations in mkfs.c and inode.c that predate this branch, reformatted in their own commit.

The teardown change was verified only by compiling fs.c and super.c against the 6.17 kernel headers in Lima; the module was not loaded or mounted and make check was not run. The 7.3+ branch (fs_bdev_file_release() in put_super) and the kernels below 6.17 were not built. Reviewed independently by Codex, Agy and Claude. Not addressed: mounting read-only should fail up front instead of surfacing later write errors, and journal_path pointing at a non-block file is still silently ignored.