[9fans] fossil riddle

[Richard Miller]

Who can guess what's happening here:

term% ls -ld src
d-rwxrwxr-x M 353 miller miller 0 Mar 27 17:15 src
term% ls -l src
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/9
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/ape
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/boot
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/cmd
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/games
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/lib9p
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libString
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libauth
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libauthsrv
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libavl
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libbin
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libbio
term% ls -l src/*
ls: src/9: 'src/9' does not exist
ls: src/ape: 'src/ape' does not exist
ls: src/boot: 'src/boot' does not exist
ls: src/cmd: 'src/cmd' does not exist
ls: src/games: 'src/games' does not exist
ls: src/lib9p: 'src/lib9p' does not exist
ls: src/libString: 'src/libString' does not exist
ls: src/libauth: 'src/libauth' does not exist
ls: src/libauthsrv: 'src/libauthsrv' does not exist
ls: src/libavl: 'src/libavl' does not exist
ls: src/libbin: 'src/libbin' does not exist
ls: src/libbio: 'src/libbio' does not exist

[Nemo]

I've seen that when the directory is ok but the file is not, fossil cannot
get to the data file although it can obtain the metadata for the file.
In fact, we still have a few files named "bug" which came from
mv a-file-in-such-shape to bug

Is that the case?

Post by Richard Miller
term% ls -ld src
d-rwxrwxr-x M 353 miller miller 0 Mar 27 17:15 src
term% ls -l src
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/9
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/ape
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/boot
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/cmd
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/games
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/lib9p
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libString
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libauth
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libauthsrv
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libavl
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libbin
d-rwxrwxr-x M 353 miller miller 0 Mar 27 09:31 src/libbio
term% ls -l src/*
ls: src/9: 'src/9' does not exist
ls: src/ape: 'src/ape' does not exist
ls: src/boot: 'src/boot' does not exist
ls: src/cmd: 'src/cmd' does not exist
ls: src/games: 'src/games' does not exist
ls: src/lib9p: 'src/lib9p' does not exist
ls: src/libString: 'src/libString' does not exist
ls: src/libauth: 'src/libauth' does not exist
ls: src/libauthsrv: 'src/libauthsrv' does not exist
ls: src/libavl: 'src/libavl' does not exist
ls: src/libbin: 'src/libbin' does not exist
ls: src/libbio: 'src/libbio' does not exist

[Richard Miller]

Post by Nemo
I've seen that when the directory is ok but the file is not, fossil cannot
get to the data file although it can obtain the metadata for the file.
..
Is that the case?

No, in this case the files are all actually fine.

[Richard Miller]

The riddle revealed:

term% ls -l ape
d-rwxrwxr-x M 1003 miller miller 0 Mar 27 09:31 ape/9src
d-rwxrwxr-x M 1003 miller miller 0 Mar 27 09:31 ape/cmd
--rw-rw-r-- M 1003 miller miller 573 Feb 28 2002 ape/config
d-rwxrwxr-x M 1003 miller miller 0 Mar 27 09:31 ape/lib
--rw-rw-r-- M 1003 miller miller 368 Feb 28 2002 ape/mkfile
term% ls -l ape/cmd
ls: ape/cmd: 'ape/cmd' does not exist
term% cd ape
term% ls -l cmd
ls: cmd: 'cmd' disk is full

It's actually not specific to fossil, but an artefact of
9P2000. If a walk(5) down a multi-level path fails at the
first level, the server can return Rerror with an informative
message. But if it partially succeeds, it must return Rwalk
with a list of qids for the successfully traversed directories.
The kernel can't tell why the walk didn't complete, so it
takes a guess at "does not exist".

Possible solutions (if anyone cares):

- Next time 9P is revised, add a status string to Rwalk.
- In the kernel, if an N-level walk fails at level K (1 < K <= N),
retry as two sub-walks (length K-1 and 1) to get an Rwalk for
the failing level. If there's a race condition and the retry
succeeds, that's OK too.
- Change the error message to something more truthful like
"can't walk to ..."