Re: [PATCH 0/9] some fixes for bugs spotted by valgrind

[Sergei Trofimovich <slyfox@gentoo.org>]

[-- Attachment #1: Type: text/plain, Size: 1317 bytes --]

On Thu, 02 Jun 2011 13:17:55 -0700
Andi Kleen <andi@firstfloor.org> wrote:

> Sergei Trofimovich <slyfox@gentoo.org> writes:
> >
> > Am I too paranoid about the issue?
> 
> It sounds weird, because if the kernel would really checksum
> mutexes on disk you would have a lot of on disk
> format incompatibility between different kernel versions
> (e.g. between lockdep and normal kernels or kernels
> running on different architectures)
>
> If it would really happen (no opinion on that) it would
> be a serious bug.

Oh, I don't think things are so bad.

In order it to be a problem superblock loading would have to
be loaded exactly the same way as it's stored, but it isn't.
At least super copies (baked into btrfs_fs_info) are read
to separate data block (buffer_hear) and then copied properly
(in open_ctree) to super_copy/super_for_commit:

        bh = btrfs_read_dev_super(fs_devices->latest_bdev);
        if (!bh) {
                err = -EINVAL;
                goto fail_alloc;
        }

        memcpy(&fs_info->super_copy, bh->b_data, sizeof(fs_info->super_copy));
        memcpy(&fs_info->super_for_commit, &fs_info->super_copy,
               sizeof(fs_info->super_for_commit));
        brelse(bh);

But the way superblocks are written look racy.

-- 

  Sergei

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 198 bytes --]