9.1. I/O Ports and I/O Memory

Every peripheral device is controlled by writing and reading its registers. Most of the time a device has several registers, and they are accessed at consecutive addresses, either in the memory address space or in the I/O address space.

At the hardware level, there is no conceptual difference between memory regions and I/O regions: both of them are accessed by asserting electrical signals on the address bus and control bus (i.e., the read and write signals)^[1] and by reading from or writing to the data bus.

^[1] Not all computer platforms use a read and a write signal; some have different means to address external circuits. The difference is irrelevant at software level, however, and we'll assume all have read and write to simplify the discussion.

While some CPU manufacturers implement a single address space in their chips, others decided that peripheral devices are different from memory and, therefore, deserve a separate address space. Some processors (most notably the x86 family) have separate read and write electrical lines for I/O ports and special CPU instructions to access ports.

Because peripheral devices are built to fit a peripheral bus, and the most popular I/O buses are modeled on the personal computer, even processors that do not have a separate address space for I/O ports must fake reading and writing I/O ports when accessing some peripheral devices, usually by means of external chipsets or extra circuitry in the CPU core. The latter solution is common within tiny processors meant for embedded use.

For the same reason, Linux implements the concept of I/O ports on all computer platforms it runs on, even on platforms where the CPU implements a single address space. The implementation of port access sometimes depends on the specific make and model of the host computer (because different models use different chipsets to map bus transactions into memory address space).

Even if the peripheral bus has a separate address space for I/O ports, not all devices map their registers to I/O ports. While use of I/O ports is common for ISA peripheral boards, most PCI devices map registers into a memory address region. This I/O memory approach is generally preferred, because it doesn't require the use of special-purpose processor instructions; CPU cores access memory much more efficiently, and the compiler has much more freedom in register allocation and addressing-mode selection when accessing memory.

9.1.1. I/O Registers and Conventional Memory

Despite the strong similarity between hardware registers and memory, a programmer accessing I/O registers must be careful to avoid being tricked by CPU (or compiler) optimizations that can modify the expected I/O behavior.

The main difference between I/O registers and RAM is that I/O operations have side effects, while memory operations have none: the only effect of a memory write is storing a value to a location, and a memory read returns the last value written there. Because memory access speed is so critical to CPU performance, the no-side-effects case has been optimized in several ways: values are cached and read/write instructions are reordered.

The compiler can cache data values into CPU registers without writing them to memory, and even if it stores them, both write and read operations can operate on cache memory without ever reaching physical RAM. Reordering can also happen both at the compiler level and at the hardware level: often a sequence of instructions can be executed more quickly if it is run in an order different from that which appears in the program text, for example, to prevent interlocks in the RISC pipeline. On CISC processors, operations that take a significant amount of time can be executed concurrently with other, quicker ones.

These optimizations are transparent and benign when applied to conventional memory (at least on uniprocessor systems), but they can be fatal to correct I/O operations, because they interfere with those "side effects" that are the main reason why a driver accesses I/O registers. The processor cannot anticipate a situation in which some other process (running on a separate processor, or something happening inside an I/O controller) depends on the order of memory access. The compiler or the CPU may just try to outsmart you and reorder the operations you request; the result can be strange errors that are very difficult to debug. Therefore, a driver must ensure that no caching is performed and no read or write reordering takes place when accessing registers.

The problem with hardware caching is the easiest to face: the underlying hardware is already configured (either automatically or by Linux initialization code) to disable any hardware cache when accessing I/O regions (whether they are memory or port regions).

The solution to compiler optimization and hardware reordering is to place a memory barrier between operations that must be visible to the hardware (or to another processor) in a particular order. Linux provides four macros to cover all possible ordering needs:

#include <linux/kernel.h>

void barrier(void)

This function tells the compiler to insert a memory barrier but has no effect on the hardware. Compiled code stores to memory all values that are currently modified and resident in CPU registers, and rereads them later when they are needed. A call to barrier prevents compiler optimizations across the barrier but leaves the hardware free to do its own reordering.

#include <asm/system.h>

void rmb(void);

void read_barrier_depends(void);

void wmb(void);

void mb(void);

These functions insert hardware memory barriers in the compiled instruction flow; their actual instantiation is platform dependent. An rmb (read memory barrier) guarantees that any reads appearing before the barrier are completed prior to the execution of any subsequent read. wmb guarantees ordering in write operations, and the mb instruction guarantees both. Each of these functions is a superset of barrier.

read_barrier_depends is a special, weaker form of read barrier. Whereas rmb prevents the reordering of all reads across the barrier, read_barrier_depends blocks only the reordering of reads that depend on data from other reads. The distinction is subtle, and it does not exist on all architectures. Unless you understand exactly what is going on, and you have a reason to believe that a full read barrier is exacting an excessive performance cost, you should probably stick to using rmb.

void smp_rmb(void);

void smp_read_barrier_depends(void);

void smp_wmb(void);

void smp_mb(void);

These versions of the barrier macros insert hardware barriers only when the kernel is compiled for SMP systems; otherwise, they all expand to a simple barrier call.

A typical usage of memory barriers in a device driver may have this sort of form:

writel(dev->registers.addr, io_destination_address);
writel(dev->registers.size, io_size);
writel(dev->registers.operation, DEV_READ);
wmb(  );
writel(dev->registers.control, DEV_GO);

In this case, it is important to be sure that all of the device registers controlling a particular operation have been properly set prior to telling it to begin. The memory barrier enforces the completion of the writes in the necessary order.

Because memory barriers affect performance, they should be used only where they are really needed. The different types of barriers can also have different performance characteristics, so it is worthwhile to use the most specific type possible. For example, on the x86 architecture, wmb( ) currently does nothing, since writes outside the processor are not reordered. Reads are reordered, however, so mb( ) is slower than wmb( ).

It is worth noting that most of the other kernel primitives dealing with synchronization, such as spinlock and atomic_t operations, also function as memory barriers. Also worthy of note is that some peripheral buses (such as the PCI bus) have caching issues of their own; we discuss those when we get to them in later chapters.

Some architectures allow the efficient combination of an assignment and a memory barrier. The kernel provides a few macros that perform this combination; in the default case, they are defined as follows:

#define set_mb(var, value)  do {var = value; mb(  );}  while 0
#define set_wmb(var, value) do {var = value; wmb(  );} while 0
#define set_rmb(var, value) do {var = value; rmb(  );} while 0

Where appropriate, <asm/system.h> defines these macros to use architecture-specific instructions that accomplish the task more quickly. Note that set_rmb is defined only by a small number of architectures. (The use of a do...while construct is a standard C idiom that causes the expanded macro to work as a normal C statement in all contexts.)