Writing drivers

What is an EVL driver?

An EVL driver is a regular Linux character device driver which also implements a set of out-of-band I/O operations advertised by its file operation descriptor (struct file_operations). These out-of-band I/O requests are only available to EVL threads, since they run on the out-of-band execution stage. Applications running in user-space can start these I/O operations by issuing specific system calls to the EVL core.

Excerpt from include/linux/fs.h, as modified by Dovetail

struct file_operations {
	...
	ssize_t (*oob_read) (struct file *, char __user *, size_t);
	ssize_t (*oob_write) (struct file *, const char __user *, size_t);
	long (*oob_ioctl) (struct file *, unsigned int, unsigned long);
	__poll_t (*oob_poll) (struct file *, struct oob_poll_wait *);
	...
} __randomize_layout;

Dovetail is in charge of routing the system calls received from applications to the proper recipient, either the EVL core or the in-band kernel. As mentioned earlier when describing the everything-is-a-file mantra, only I/O transfer and control requests have to run from the out-of-band context (i.e. EVL’s real-time mode), creating and dismantling the underlying file from the regular in-band context is just fine. Therefore, the execution flow upon I/O system calls looks like this:

Alt text

Which translates as follows:

  • When an application issues the open(2) or close(2) system calls with a file descriptor referring to a file managed by an EVL driver, the request normally goes through the virtual filesystem switch (VFS) as usual, ending up into the .open() and .release() handlers (when the last reference is dropped) defined by the driver in its struct file_operations descriptor. The same goes for mmap(2), ioctl(2), read(2), write(2), and all other common file operations for character device drivers.

  • When an applications issues the oob_read(), oob_write() or oob_ioctl() system calls via the EVL library, Dovetail routes the call to the EVL core (instead of the VFS), which in turn fires the corresponding handlers defined by the driver’s struct file_operations descriptor: i.e. .oob_read(), .oob_write(), and .oob_ioctl(). Those handlers should use the EVL kernel API, or the main kernel services which are known to be usable from the out-of-band-context, exclusively. Failing to abide by this golden rule may lead to funky outcomes ranging from plain kernel crashes (lucky) to rampant memory corruption (unlucky).

Now, you may wonder: “what if an out-of-band operation is ongoing in the driver on a particular file, while I’m closing the last VFS-maintained reference to that file?” Well, with a properly written EVL driver, the close(2) call should block until the out-of-band operation finishes, at which point the .release() handler may proceed with dismantling the file. A simple rule for writing EVL drivers ensures this.

Last modified: Wed, 15 Jan 2020 18:02:54 CET