Re: Memory corruption due to word sharing

[David Sterba <dsterba@suse.cz>]

On Wed, Feb 01, 2012 at 04:19:18PM +0100, Jan Kara wrote:
> We actually spotted this race in practice in btrfs on structure
> fs/btrfs/ctree.h:struct btrfs_block_rsv where spinlock content got
> corrupted due to update of following bitfield and there seem to be other
> places in kernel where this could happen.

Here's the list of structures where a bitfield is shared with spinlock
or atomic/kref within an 8B word, generated from 3.3-rc2:

spinlock+bitfield:

Struct: struct ak4113; Field: init
Struct: struct ak4114; Field: init
Struct: struct ak4117; Field: init
Struct: struct btrfs_block_rsv; Field: full
Struct: struct cm109_dev; Field: buzzer_pending
Struct: struct pch_udc_dev; Field: active
Struct: struct rds_iw_device; Field: dma_local_lkey
Struct: struct sierra_intf_private; Field: suspended
Struct: struct sm501_gpio; Field: registered
Struct: struct unix_sock; Field: gc_candidate
Struct: struct usb_anchor; Field: poisoned
Struct: struct usb_wwan_intf_private; Field: suspended

atomic/kref+bitfield:

Struct: struct dlm_lock_resource; Field: migration_pending
Struct: struct extent_map; Field: in_tree
Struct: struct kobject; Field: state_initialized
Struct: struct page; Field: inuse
Struct: struct rds_ib_connection; Field: i_flowctl
Struct: struct rds_iw_connection; Field: i_flowctl
Struct: struct sctp_transport; Field: dead
Struct: struct transaction_s; Field: t_synchronous_commit
Struct: struct xfs_ioend; Field: io_isasync


Not all listed structs are necessarily subject to the bug. There may be
another mechanism preventing concurrent access to the bitfield and
spinlock/atomic, or the bitfield is modified from a single cpu, or is
not used. But all of them need to be reviewed of course.


david