Retrying revoke()

There is only one problem: Linux does not support revoke(), and every attempt to add it over the years has ended in failure. The functionality behind revoke() turns out to be quite difficult to implement in a safe way. The latest attempt at a revoke() implementation may well come to a similar conclusion; there is not even a proof-of-concept patch to evaluate, after all. But, since the developer behind it is Al Viro, one assumes that its chances of success are mildly better than average.

Not every file or device will support revoke(); in some cases, it may still prove too hard to do properly. With Al's proposal, in cases where revocation is supported, there would be a new structure associated with the relevant device (or other) structure:

    struct revokable {
	atomic_t in_use;		// number of threads in methods,
	spinlock_t lock;
	hlist_head list;
	struct completion *c;
	void (*kick)(struct revokable *);
    };

The in_use field is charged with tracking how many threads are actively executing in the file_operations methods associated with this object. Performing this tracking would require changing every method call site throughout the kernel to call a couple of helper functions and check for a revoked file. So a call that currently looks like:

    ret = file->f_op->read(...);

Would be turned into something like:

    if (start_using(file)) {
	ret = file->f_op->read(...);
	stop_using(file);
    } else {
	ret = -EIO;  /* File revoked */
    }

The start_using() and stop_using() helper functions increment and decrement the in_use counter. If that counter is negative, though, access is being revoked and start_using() will return false; in such cases, the file_operations method should not be called and an appropriate error code should be returned. Naturally, the details of these helper functions are a bit more complex than this; see Al's posting for a more complete story. As Al notes, there are quite a few call sites for file_operations methods in the kernel, so this particular change would be relatively intrusive.

The purpose of the kick() callback is to instruct the object's driver that access is being revoked and any outstanding I/O operations should be brought to an end. Processes waiting on I/O should return with an error code and the I/O canceled. After the kick() call, the number of threads running within the object's file_operations should quickly drop to zero.

When open() is called on an object that supports revocation, the associated file structure will gain a pointer to a structure like:

    struct revoke {
	struct file *file;
	struct revokable *revokable;
	struct hlist_node list;
	bool closing;
	struct completion *c;
    };

The list field is used to track all open files associated with a given revocable object. As the last step in an open() implementation, the make_revokable() helper will be called to allocate the revoke structure and attach it to the list in the object's revokable structure.

With this infrastructure in place, an implementation of revoke() becomes possible. The steps, roughly, are these:

1. Mark the object as being revoked by subtracting a large number from its in_use counter, turning that counter negative. That will prevent any further calls to the object's file_operations methods.

2. If in_use indicates that threads are currently running in the object's file_operations, call kick() to encourage them all to finish and wait until they all complete.

3. For each open file, call the release() method to close that file, and remove the file from the list.

There is, of course, one other thorny little problem: what do to about processes that have used mmap() to map the object into their address space. One possibility is to forcibly unmap the memory, tearing down the associated page tables and marking the virtual memory area (VMA) structure accordingly; the process would then most likely receive a SIGSEGV signal if it attempted to access that address space. That approach is secure, but also risks causing programs to crash unexpectedly. In cases where device memory has been mapped, a better solution might be to just cause all accesses to return 0xff (extended out to the correct width for the specific access). Proper handling of mmap() in this situation is an open question, and one apparently without precedent in the current implementations of revoke() in other systems — revoke() on BSD systems works only on devices without mapped memory.

There is a fair gap between an RFC posting with a clever idea and an actual, working implementation; it may well be that this approach to revoke() will, like its predecessors, run aground in the real world. But the lack of a working revoke() has been seen as a shortcoming in Linux for many years; it would be nice to finally get this functionality into place. So, just maybe, things will work out this time around.

-rwsr-sr-x 1 root root 14256 Mar  3  2012 /usr/bin/X