Umm. What does that matter wrt what I said?
What does that matter wrt your bogus argument about EIO (which we do
_wrong_ right now, exactly because we return EIO too eagerly, instead of
returning the partial data we got)?
vfs_readdir() itself would not change AT ALL. The change I talked about
was in the caller. Which you should realize, since I actually _called_
vfs_readdir() in that example code.
Here's a patch to get the _semantics_ I talked about, to make the thing
more obvious.
But the actual _change_ I talked about would be to try to avoid using
"buf.error" entirely, and just make all the (many, I know) filesystem
readdir() functions return the callback value instead of 0. Which would
mean that most of the users would be able to drop their "buf.error" use
entirely, along with the ugly
if (error >= 0)
error = buf.error;
that I have in this patch.
(Side note: the "if (error >= 0)" approach is what old_readdir() already
does, so that function actually gets the EIO case _correct_: it's only
the newer sys_getdents() that is buggy)
Linus
---
fs/readdir.c | 6 ++----
1 files changed, 2 insertions(+), 4 deletions(-)
diff --git a/fs/readdir.c b/fs/readdir.c
index 4e026e5..e4ded16 100644
--- a/fs/readdir.c
+++ b/fs/readdir.c
@@ -205,9 +205,8 @@ asmlinkage long sys_getdents(unsigned int fd, struct linux_dirent __user * diren
buf.error = 0;
error = vfs_readdir(file, filldir, &buf);
- if (error < 0)
- goto out_putf;
- error = buf.error;
+ if (error >= 0)
+ error = buf.error;
lastdirent = buf.previous;
if (lastdirent) {
if (put_user(file->f_pos, &lastdirent->d_off))
@@ -216,7 +215,6 @@ asmlinkage long sys_getdents(unsigned int fd, struct linux_dirent __user * diren
error = count - buf.count;
}
-out_putf:
fput(file);
out:
return error;
--
