Re: [BUG] general protection fault in jffs2_xattr_delete_inode

On Thu, 20 Aug 2026 03:28:04 +0900, Jaeyoung Chung wrote:
> We have not analyzed the root cause, so we do not have a proposed fix
> to offer.

f->inocache is never initialised, and jffs2_new_inode() has two error paths
that evict the inode before it gets assigned.

Your log says it outright.  RSI is the ic argument:

  RSI: 6464646464646464
  RBX: 646464646464648c
  1f: 48 8d 5e 28   lea 0x28(%rsi),%rbx

0x64 is 'd', and th_churn() does memset(p + 1, 'a' + (idx & 7), 3998) with
idx 3.  So ic is not a stale pointer to anything, it is the pathname bytes
that thread left lying in the page the inode was later carved from.  With
CONFIG_JFFS2_FS_XATTR the 0x28 is pino_nlink, so the faulting access is the
ic->pino_nlink of xattr.c:602.

Nothing initialises that field.  jffs2_inode_cachep has a constructor, and
jffs2_i_init_once() only does mutex_init(), target = NULL and
inode_init_once().  jffs2_alloc_inode() allocates with plain GFP_KERNEL, so
no zeroing.  jffs2_init_inode_info() sets seven fields and inocache is not
among them.  inocache also sits before vfs_inode in the struct, so
inode_init_always() does not reach it.

On this path f->inocache is assigned only by jffs2_do_new_inode(),
write.c:35.  Both of these return before that, and both then do
make_bad_inode() and iput():

  fs.c:459  jffs2_init_acl_pre() fails
  fs.c:466  jffs2_do_new_inode() fails in jffs2_alloc_inode_cache(),
            write.c:28

Either is an allocation failure, so fail-nth reaches both.  The inode is
unhashed, so inode_generic_drop() makes iput() evict it, which is the
evict() -> jffs2_do_clear_inode() -> jffs2_xattr_delete_inode() in your
trace.  The mdelay() before new_inode() gives the churn threads a window to
leave their bytes where the inode is then carved from.

mtd/next has the same jffs2_init_inode_info(), so:

--- a/fs/jffs2/os-linux.h
+++ b/fs/jffs2/os-linux.h
@@ -55,6 +55,7 @@ static inline void jffs2_init_inode_info(struct jffs2_inode_info *f)
 	f->metadata = NULL;
 	f->dents = NULL;
 	f->target = NULL;
+	f->inocache = NULL;
 	f->flags = 0;
 	f->usercompr = 0;
 }

Both callers assign inocache immediately afterwards, jffs2_do_read_inode()
at readinode.c:1335 as its first act and jffs2_do_new_inode() at write.c:35,
and jffs2_xattr_delete_inode() returns early on a NULL ic, so the read path
is unaffected.

I have not run your reproducer, so please confirm it on your setup before
this goes out as a patch.

Thanks,
Zhan Xusheng

______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/