The new way of ioctl()-compat_ioctl unlocked_ioctl

The ioctl() system call has long been out of favor among the kernel developers, who see it as a completely uncontrolled entry point into the kernel. Given the vast number of applications which expect ioctl() to be present, however, it will not go away anytime soon. So it is worth the trouble to ensure that ioctl() calls are performed quickly and correctly - and that they do not unnecessarily impact the rest of the system.

ioctl() is one of the remaining parts of the kernel which runs under the Big Kernel Lock (BKL). In the past, the usage of the BKL has made it possible for long-running ioctl() methods to create long latencies for unrelated processes. Recent changes, which have made BKL-covered code preemptible, have mitigated that problem somewhat. Even so, the desire to eventually get rid of the BKL altogether suggests that ioctl() should move out from under its protection.

Simply removing the lock_kernel() call before calling ioctl() methods is not an option, however. Each one of those methods must first be audited to see what other locking may be necessary for it to run safely outside of the BKL. That is a huge job, one which would be hard to do in a single "flag day" operation. So a migration path must be provided. As of 2.6.11, that path will exist.

The (by Michael s. Tsirkin) adds a new member to the file_operations structure:

    long (*unlocked_ioctl) (struct file *filp, unsigned int cmd, 
                            unsigned long arg);

If a driver or filesystem provides an unlocked_ioctl() method, it will be called in preference to the older ioctl(). The differences are that the inode argument is not provided (it's available as filp->f_dentry->d_inode) and the BKL is not taken prior to the call. All new code should be written with its own locking, and should use unlocked_ioctl(). Old code should be converted as time allows. For code which must run on multiple kernels, there is a new HAVE_UNLOCKED_IOCTL macro which can be tested to see if the newer method is available or not.

Michael's patch adds one other operation:

    long (*compat_ioctl) (struct file *filp, unsigned int cmd, 
                          unsigned long arg);

If this method exists, it will be called (without the BKL) whenever a 32-bit process calls ioctl() on a 64-bit system. It should then do whatever is required to convert the argument to native data types and carry out the request. If compat_ioctl() is not provided, the older conversion mechanism will be used, as before. The HAVE_COMPAT_IOCTL macro can be tested to see if this mechanism is available on any given kernel.

The compat_ioctl() method will probably filter down into a few subsystems. Andi Kleen has posted patches adding new compat_ioctl() methods to the block_device_operations and scsi_host_template structures, for example, though those patches have not been merged as of this writing.

---

( to post comments)

more on compat_ioctl

Posted Oct 23, 2005 14:19 UTC (Sun) by arnd (subscriber, #8866) []

There are a few noteworthy points about compat_ioctl:

• If you are writing a new device driver that needs ioctl methods (which some might argue you should not do in the first place), make sure the data structure are compatible between 32 and 64 bit, so unlocked_ioctl and compat_ioctl can point to the same function. In particular, data structures containing must not contain fields that have different sizes (e.g. 'void *' or 'long') or need padding (e.g. 'long long' after 'int') on 64 bit systems.
• As of 2.6.14, nobody has started converting the network layer to compat_ioctl, so the next person that needs new compatibility code for socket ioctls should add the infrastructure for that instead of adding on to fs/compat_ioctl.c.
• While the fs/compat_ioctl.c infrastructure still exists, it is valid for compat_ioctl methods to return -ENOIOCTLCMD for anything they don't know. This is particularly useful for block or tty devices that have a lot of ioctl numbers common to all drivers. The vfs layer first calls ->compat_ioctl and if that does not exist or returns -ENOIOCTLCMD, it scans the list of known conversions between 32 and 64 bit ioctls and if it finds a valid conversion, it enters the native 64 bit ->unlocked_ioctl/->ioctl path.