pstore: Solve lockdep warning by moving inode locks
authorKees Cook <keescook@chromium.org>
Thu, 27 Apr 2017 22:53:21 +0000 (15:53 -0700)
committerKees Cook <keescook@chromium.org>
Fri, 28 Apr 2017 03:35:34 +0000 (20:35 -0700)
Lockdep complains about a possible deadlock between mount and unlink
(which is technically impossible), but fixing this improves possible
future multiple-backend support, and keeps locking in the right order.

The lockdep warning could be triggered by unlinking a file in the
pstore filesystem:

  -> #1 (&sb->s_type->i_mutex_key#14){++++++}:
         lock_acquire+0xc9/0x220
         down_write+0x3f/0x70
         pstore_mkfile+0x1f4/0x460
         pstore_get_records+0x17a/0x320
         pstore_fill_super+0xa4/0xc0
         mount_single+0x89/0xb0
         pstore_mount+0x13/0x20
         mount_fs+0xf/0x90
         vfs_kern_mount+0x66/0x170
         do_mount+0x190/0xd50
         SyS_mount+0x90/0xd0
         entry_SYSCALL_64_fastpath+0x1c/0xb1

  -> #0 (&psinfo->read_mutex){+.+.+.}:
         __lock_acquire+0x1ac0/0x1bb0
         lock_acquire+0xc9/0x220
         __mutex_lock+0x6e/0x990
         mutex_lock_nested+0x16/0x20
         pstore_unlink+0x3f/0xa0
         vfs_unlink+0xb5/0x190
         do_unlinkat+0x24c/0x2a0
         SyS_unlinkat+0x16/0x30
         entry_SYSCALL_64_fastpath+0x1c/0xb1

  Possible unsafe locking scenario:

        CPU0                    CPU1
        ----                    ----
   lock(&sb->s_type->i_mutex_key#14);
                                lock(&psinfo->read_mutex);
                                lock(&sb->s_type->i_mutex_key#14);
   lock(&psinfo->read_mutex);

Reported-by: Marta Lofstedt <marta.lofstedt@intel.com>
Reported-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Kees Cook <keescook@chromium.org>
Acked-by: Namhyung Kim <namhyung@kernel.org>
fs/pstore/inode.c
fs/pstore/internal.h
fs/pstore/platform.c

index 06504b69575bfb742aa5fc9779c86a1e9aeda73c..792a4e5f9226d16fefa0b6fd47061a76c79416ba 100644 (file)
@@ -311,9 +311,8 @@ bool pstore_is_mounted(void)
  * Load it up with "size" bytes of data from "buf".
  * Set the mtime & ctime to the date that this record was originally stored.
  */
-int pstore_mkfile(struct pstore_record *record)
+int pstore_mkfile(struct dentry *root, struct pstore_record *record)
 {
-       struct dentry           *root = pstore_sb->s_root;
        struct dentry           *dentry;
        struct inode            *inode;
        int                     rc = 0;
@@ -322,6 +321,8 @@ int pstore_mkfile(struct pstore_record *record)
        unsigned long           flags;
        size_t                  size = record->size + record->ecc_notice_size;
 
+       WARN_ON(!inode_is_locked(d_inode(root)));
+
        spin_lock_irqsave(&allpstore_lock, flags);
        list_for_each_entry(pos, &allpstore, list) {
                if (pos->record->type == record->type &&
@@ -336,7 +337,7 @@ int pstore_mkfile(struct pstore_record *record)
                return rc;
 
        rc = -ENOMEM;
-       inode = pstore_get_inode(pstore_sb);
+       inode = pstore_get_inode(root->d_sb);
        if (!inode)
                goto fail;
        inode->i_mode = S_IFREG | 0444;
@@ -394,11 +395,9 @@ int pstore_mkfile(struct pstore_record *record)
                break;
        }
 
-       inode_lock(d_inode(root));
-
        dentry = d_alloc_name(root, name);
        if (!dentry)
-               goto fail_lockedalloc;
+               goto fail_private;
 
        inode->i_size = private->total_size = size;
 
@@ -413,12 +412,9 @@ int pstore_mkfile(struct pstore_record *record)
        list_add(&private->list, &allpstore);
        spin_unlock_irqrestore(&allpstore_lock, flags);
 
-       inode_unlock(d_inode(root));
-
        return 0;
 
-fail_lockedalloc:
-       inode_unlock(d_inode(root));
+fail_private:
        free_pstore_private(private);
 fail_alloc:
        iput(inode);
@@ -427,6 +423,27 @@ fail:
        return rc;
 }
 
+/*
+ * Read all the records from the persistent store. Create
+ * files in our filesystem.  Don't warn about -EEXIST errors
+ * when we are re-scanning the backing store looking to add new
+ * error records.
+ */
+void pstore_get_records(int quiet)
+{
+       struct pstore_info *psi = psinfo;
+       struct dentry *root;
+
+       if (!psi || !pstore_sb)
+               return;
+
+       root = pstore_sb->s_root;
+
+       inode_lock(d_inode(root));
+       pstore_get_backend_records(psi, root, quiet);
+       inode_unlock(d_inode(root));
+}
+
 static int pstore_fill_super(struct super_block *sb, void *data, int silent)
 {
        struct inode *inode;
index af1df5a36d860bba6989744169821c3d5c71c807..c416e653dc4f59d34a261ee2a617d82850067587 100644 (file)
@@ -25,7 +25,10 @@ extern struct pstore_info *psinfo;
 
 extern void    pstore_set_kmsg_bytes(int);
 extern void    pstore_get_records(int);
-extern int     pstore_mkfile(struct pstore_record *record);
+extern void    pstore_get_backend_records(struct pstore_info *psi,
+                                          struct dentry *root, int quiet);
+extern int     pstore_mkfile(struct dentry *root,
+                             struct pstore_record *record);
 extern bool    pstore_is_mounted(void);
 
 #endif
index 43b3ca5e045ffc5b605338fdbb6aff0e2bbf4521..d468eec9b8a6ba9aa574055c0e76c615e2ede9d0 100644 (file)
@@ -810,17 +810,17 @@ static void decompress_record(struct pstore_record *record)
 }
 
 /*
- * Read all the records from the persistent store. Create
+ * Read all the records from one persistent store backend. Create
  * files in our filesystem.  Don't warn about -EEXIST errors
  * when we are re-scanning the backing store looking to add new
  * error records.
  */
-void pstore_get_records(int quiet)
+void pstore_get_backend_records(struct pstore_info *psi,
+                               struct dentry *root, int quiet)
 {
-       struct pstore_info *psi = psinfo;
        int failed = 0;
 
-       if (!psi)
+       if (!psi || !root)
                return;
 
        mutex_lock(&psi->read_mutex);
@@ -850,7 +850,7 @@ void pstore_get_records(int quiet)
                        break;
 
                decompress_record(record);
-               rc = pstore_mkfile(record);
+               rc = pstore_mkfile(root, record);
                if (rc) {
                        /* pstore_mkfile() did not take record, so free it. */
                        kfree(record->buf);