16 Dec
2017
16 Dec
'17
10:11 p.m.
On Sat, Dec 16, 2017, at 05:50, Doug McIlroy wrote:
Like UTF-8, a variable-length time would be something
normal
programs aren't supposed to see--it would be a format for
external media (i.e. file systems). Times in inodes would be
variable-length. Times returned by stat() and time() would
be full length. Only the kernel needs to know the encoding.
One may note that inodes already use a variable-length
encoding for the list of physical block addresses, so there
is nothing radical here.
I admit, though, that if the kernel can tell "old" from "new"
inodes, then they could have different time encodings rather
than one variable-length encoding.
I actually thought about this some more since making my previous post,
though I hadn't connected it to your post since I'd thought we were
still talking about the userspace interface.
One way the kernel could tell old from new inodes is by reserving the
high bit of one of the current 32-bit fields (either time, or size since
IIRC an existing inode can't actually hold a 2GB file) as a flag. My
initial thought was to simply encode in place (stealing some bits from
atime for the flag and high bits of ctime, and exploiting the
relationship that ctime<=mtime<=atime)
Or a whole "extended inode", that occupies two (or more) inodes on disk,
either adjacent to each other or by stealing a whole field as an inode
pointer.
Speaking of the address list, it looks like there's an unused byte in
the on-disk address list, which is 40 bytes for 13*3=39 bytes of
addresses
(The above is for V7 - V6 inodes are smaller and have less unused space)
Making this robust to being mounted by old kernels (and preventing old
kernels from creating files that would appear to new kernels to have
extended inodes) is the hard part, and I haven't thought any about what
the format of an extended inode would be.