Hi
when reviewing some i_size problem, I searched the kernel for i_size usage
(the variable should really be written with i_size_write and read with
i_size_read).
Properly locked direct use of "i_size" inside memory management or
filesystems may not be a problem, but there are many problems in general
code outside mm.
The misuses are:
SOUND/SOUND_FIRMWARE.C:l = filp->f_path.dentry->d_inode->i_size;
KERNEL/RELAY.C:buf->dentry->d_inode->i_size = buf->early_bytes;
KERNEL/RELAY.C:buf->dentry->d_inode->i_size += buf->chan->subbuf_size
-buf->padding[old_subbuf];
DRIVERS/USB/CORE/INODE.C:dev->usbfs_dentry->d_inode->i_size = i_size;
DRIVERS/MTD/DEVICES/BLOCK2MTD.C:dev->mtd.size =
dev->blkdev->bd_inode->i_size & PAGE_MASK;
DRIVERS/MD/MD.C: many reads of i_size
DRIVERS/BLOCK/NBD.C: many writes to i_size without i_size_write
DRIVERS/BLOCK/DRBD/DRBD_INT.H: return bdev ? bdev->bd_inode->i_size >> 9 : 0;
DRIVERS/BLOCK/DRBD/DRBD_WRAPPERS.H: mdev->this_bdev->bd_inode->i_size =
(loff_t)size << 9;
BLOCK/BLK-CORE.C:printk(KERN_INFO "%s: rw=%ld, want=%Lu, limit=%Lu\n",
bdevname(bio->bi_bdev, b),
bio->bi_rw,
(unsigned long long)bio->bi_sector + bio_sectors(bio),
(long long)(bio->bi_bdev->bd_inode->i_size >> 9));
maxsector = bio->bi_bdev->bd_inode->i_size >> 9;
BLOCK/COMPAT_IOCTL.C: size = bdev->bd_inode->i_size;
return compat_put_u64(arg, bdev->bd_inode->i_size);
BLOCK/IOCTL.C: if (start + len > (bdev->bd_inode->i_size >> 9))
size = bdev->bd_inode->i_size;
return put_u64(arg, bdev->bd_inode->i_size);
The problem with this code is that if you read i_size without i_size_read
and the size wraps around 32 bits, for example from 0xffffffff to
0x100000000 , there is a possibility on 32-bit machines to read an invalid
value (either 0 or 0x1ffffffff). Similarly, if you write i_size without
i_size_write, the readers can see intermediate invalid values.
The original problem that caused this investigation is the question, how a
block device driver can change the size of its device. Normal method (used
in a few drivers, including dm), consists of
mutex_lock(&inode->i_mutex);
i_size_write(inode, new_size);
mutex_unlock(&inode->i_mutex);
This is deadlock-prone, because i_mutex is also held on fsync path.
Therefore, this deadlock happens: fsync takes i_mutex and issues I/Os,
block device driver wants to change its size, so it waits on i_mutex,
the I/Os wait until the device driver did its internal maintenance and
changed the inode size. The driver doesn't change the size until fsync
finished.
Jens, as a block maintainer, please think about it and propose some
specification how to clean this up. Also a clean verifiable rule regarding
i_size should be specified and the code should be fixed to conform to the
rule: maybe we could rename i_size to __i_size and ban its using.
Mikulas
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
