readdir returns d_ino==-1 in offset==0 mode

[mpartel]

(Background: @McBane87 and I are trying to upgrade bindfs to FUSE 3.)

Always passing 0 as the offset to the readdir filler function ("mode 1" in fuse.h) causes clients to always see d_ino == -1 when they readdir (test program).

• The passthrough example exhibits this problem, but passthrough_fh and passthrough_ll don't.
• use_ino is set.
• It doesn't matter whether FUSE_FILL_DIR_PLUS is passed.

Possible fix

On this line, remove off && .

Then passthrough passes through inode numbers if I just change it to (always) pass FUSE_FILL_DIR_PLUS.

I don't know the code well enough to tell whether that change is safe. It doesn't break any tests at least.

[Nikratio]

Thanks for the report!

Looks this check has been there ever since the first introduction of readdirplus support in 2014.

Looking at the code, I am not sure why there is a condition on off here. On the other hand, I also do not see how always passing 0 is expected to work. If you look further down, then setting off == 0 bypasses all sorts of buffer extension logic. I think the check on off is meant to check if this is the first call to the function, and calling the function repeatedly with off==0 means things will not work right.

[mpartel]

off == 0 is documented here and this codepath seems to handle it correctly, by adding the entry to a linked list, which is later copied into the buffer. Also this is in agreement with the documentation that off should not be 0 even in the first call if the 2nd mode is desired.

The first if-else in fill_dir_plus seems entirely unrelated to the filling strategy. It doesn't even pass dh anywhere. So I'm pretty convinced now that the off check can be removed.

[Nikratio]

Makes sense - pull requests welcome :-). This would be a great opportunity to add a unit test for this mode of operation, so that it doesn't break again in the future.

[jpandre]

I seem to have fallen in the same trap under the same conditions.
But IMHO the fix is only partial : the st_mode is not set, which causes the d_type not to be returned in readdir().

Moreover, the st_ino is not taken from the "struct stat" provided by the filler when use_ino is set, resulting of d_ino returned as -1.

The patch below fixes that, but the conditions might not be correct, maybe st_mode should be filled unconditionally.

--- a/lib/fuse.c 2021-03-02 21:40:51.112958271 +0100
+++ b/lib/fuse.c 2021-03-04 15:50:52.294095236 +0100
@@ -3576,6 +3576,8 @@
return 1;
}
}

• } else if (statp && f->conf.use_ino) {

•   e.attr = *statp;
  

  } else {
  e.attr.st_ino = FUSE_UNKNOWN_INO;
  if (!f->conf.use_ino && f->conf.readdir_ino) {

[jpandre]

The same patch again with proper fencing. Sorry about it.

--- a/lib/fuse.c	2021-03-02 21:40:51.112958271 +0100
+++ b/lib/fuse.c	2021-03-04 15:50:52.294095236 +0100
@@ -3576,6 +3576,8 @@
 				return 1;
 			}
 		}
+	} else if (statp && f->conf.use_ino) {
+		e.attr = *statp;
 	} else {
 		e.attr.st_ino = FUSE_UNKNOWN_INO;
 		if (!f->conf.use_ino && f->conf.readdir_ino) {