On Mon, 2007-10-29 at 16:18 +0100, Mike Galbraith wrote:On Mon, 2007-10-29 at 15:43 +0100, Mike Galbraith wrote:On Mon, 2007-10-29 at 13:39 +0000, Anton Altaparmakov wrote:Hi Mike,
Thanks for the files. That is really odd. And you are sure this just
works with 2.6.22.10 on the exact same file? Have you run "chkdsk /
f /v /x" on the NTFS volume from Windows?
Yes, 2.6.22.10 can md5sum that file just fine, did it several times. I
haven't run chkdsk.
I now have fun chkdsk, it didn't gripe, and the error is still present.
Not being very good at walking away from unsolved mysteries, I chased it
down. The problem is that...
commit[a32ea1e1f925399e0d81ca3f7394a44a6dafa12c] Fix read/truncate race
...calls ntfs_readpage() for a zero i_size inode, which it isn't
accustomed to.
Below is the hammer which made my box a happy camper again.
diff --git a/fs/ntfs/aops.c b/fs/ntfs/aops.c
index 6e5c253..ddab5a3 100644
--- a/fs/ntfs/aops.c
+++ b/fs/ntfs/aops.c
@@ -401,7 +401,7 @@ static int ntfs_readpage(struct file *file, struct page *page)
MFT_RECORD *mrec;
unsigned long flags;
u32 attr_len;
- int err = 0;
+ int err = 0, once = 0;
retry_readpage:
BUG_ON(!PageLocked(page));
@@ -414,6 +414,18 @@ retry_readpage:
return 0;
}
vi = page->mapping->host;
+ /*
+ * If we've been called to read a zero sized inode, zero and bail.
+ */
+ if (!once) {
+ loff_t i_size = i_size_read(vi);
+
+ once++;
+ if (!i_size) {
+ zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
+ goto done;
+ }
+ }
ni = NTFS_I(vi);
/*
* Only $DATA attributes can be encrypted and only unnamed $DATA