PCI Overview

The PCI bus supports several address spaces: memory, IO, and configuration. All PCI devices must support mandatory configuration space registers. Some devices may also present IO mapped and/or memory mapped resources. Before devices on the bus can be used, they must be configured. Basically, configuration will assign PCI IO and/or memory address ranges to each device and then enable that device. All PCI devices have a unique address in configuration space. This address is comprised of a bus number, a device number, and a function number. Special devices called bridges are used to connect two PCI busses together. The PCI standard supports up to 255 busses with each bus having up to 32 devices and each device having up to 8 functions.

The environment in which a platform operates will dictate if and how eCos should configure devices on the PCI bus. If the platform acts as a host on a single PCI bus, then devices may be configured individually from the relevant device driver. If the platform is not the primary host, such as a PCI card plugged into a PC, configuration of PCI devices may be left to the PC BIOS. If PCI-PCI bridges are involved, configuration of all devices is best done all at once early in the boot process. This is because all devices on the secondary side of a bridge must be evaluated for their IO and memory space requirements before the bridge can be configured.

Initializing the bus

The PCI bus needs to be initialized before it can be used. This only needs to be done once - some HALs may do it as part of the platform initialization procedure, other HALs may leave it to the application or device drivers to do it. The following function will do the initialization only once, so it's safe to call from multiple drivers:

void cyg_pci_init( void );

Scanning for devices

After the bus has been initialized, it is possible to scan it for devices. This is done using the function:

cyg_bool cyg_pci_find_next(  cyg_pci_device_id cur_devid, 
	                     cyg_pci_device_id *next_devid );

It will scan the bus for devices starting at cur_devid. If a device is found, its devid is stored in next_devid and the function returns true.

The pci1 test's outer loop looks like:

    cyg_pci_init();
    if (cyg_pci_find_next(CYG_PCI_NULL_DEVID, &devid)) {
        do {
             <use devid>
        } while (cyg_pci_find_next(devid, &devid));
    }

What happens is that the bus gets initialized and a scan is started. CYG_PCI_NULL_DEVID causes cyg_pci_find_next() to restart its scan. If the bus does not contain any devices, the first call to cyg_pci_find_next() will return false.

If the call returns true, a loop is entered where the found devid is used. After devid processing has completed, the next device on the bus is searched for; cyg_pci_find_next() continues its scan from the current devid. The loop terminates when no more devices are found on the bus.

This is the generic way of scanning the bus, enumerating all the devices on the bus. But if the application is looking for a device of a given device class (e.g., a SCSI controller), or a specific vendor device, these functions simplify the task a bit:

cyg_bool cyg_pci_find_class(  cyg_uint32 dev_class,
                              cyg_pci_device_id *devid );
cyg_bool cyg_pci_find_device(  cyg_uint16 vendor, cyg_uint16 device,
                               cyg_pci_device_id *devid );

They work just like cyg_pci_find_next(), but only return true when the dev_class or vendor/device qualifiers match those of a device on the bus. The devid serves as both an input and an output operand: the scan starts at the given device, and if a device is found devid is updated with the value for the found device.

The <cyg/io/pci_cfg.h> header file (included by pci.h) contains definitions for PCI class, vendor and device codes which can be used as arguments to the find functions. The list of vendor and device codes is not complete: add new codes as necessary. If possible also register the codes at the PCI Code List (http://www.yourvote.com/pci) which is where the eCos definitions are generated from.

Generic config information

When a valid device ID (devid) is found using one of the above functions, the associated device can be queried and controlled using the functions:

void cyg_pci_get_device_info (  cyg_pci_device_id devid, 
                                cyg_pci_device *dev_info );
void cyg_pci_set_device_info (  cyg_pci_device_id devid, 
                                cyg_pci_device *dev_info );

The cyg_pci_device structure (defined in pci.h) primarily holds information as described by the PCI specification [1]. The pci1 test prints out some of this information:

            // Get device info
            cyg_pci_get_device_info(devid, &dev_info);
            diag_printf("\n Command   0x%04x, Status 0x%04x\n",
                        dev_info.command, dev_info.status);

The command register can also be written to, controlling (among other things) whether the device responds to IO and memory access from the bus.

Specific config information

The above functions only allow access to generic PCI config registers. A device can have extra config registers not specified by the PCI specification. These can be accessed with these functions:

void cyg_pci_read_config_uint8(  cyg_pci_device_id devid,
                                 cyg_uint8 offset, cyg_uint8 *val);
void cyg_pci_read_config_uint16(  cyg_pci_device_id devid,
                                  cyg_uint8 offset, cyg_uint16 *val);
void cyg_pci_read_config_uint32(  cyg_pci_device_id devid,
                                  cyg_uint8 offset, cyg_uint32 *val);
void cyg_pci_write_config_uint8(  cyg_pci_device_id devid,
                                  cyg_uint8 offset, cyg_uint8 val);
void cyg_pci_write_config_uint16(  cyg_pci_device_id devid,
                                   cyg_uint8 offset, cyg_uint16 val);
void cyg_pci_write_config_uint32(  cyg_pci_device_id devid,
                                   cyg_uint8 offset, cyg_uint32 val);

The write functions should only be used for device-specific config registers since using them on generic registers may invalidate the contents of a previously fetched cyg_pci_device structure.

Allocating memory

A PCI device ignores all IO and memory access from the PCI bus until it has been activated. Activation cannot happen until after device configuration. Configuration means telling the device where it should map its IO and memory resources. This is done with one of the following functions::

cyg_bool cyg_pci_configure_device( cyg_pci_device *dev_info );
cyg_bool cyg_pci_configure_bus( cyg_uint8 bus, cyg_uint8 *next_bus );

The cyg_pci_configure_device handles all IO and memory regions that need configuration on non-bridge devices. On platforms with multiple busses connected by bridges, the cyg_pci_configure_bus function should be used. It will recursively configure all devices on the given bus and all subordinate busses. cyg_pci_configure_bus will use cyg_pci_configure_device to configure individual non-bridge devices.

Each region is represented in the PCI device's config space by BARs (Base Address Registers) and is handled individually according to type using these functions:

cyg_bool cyg_pci_allocate_memory(  cyg_pci_device *dev_info,
                                   cyg_uint32 bar, 
                                   CYG_PCI_ADDRESS64 *base );
cyg_bool cyg_pci_allocate_io(  cyg_pci_device *dev_info,
                               cyg_uint32 bar, 
                               CYG_PCI_ADDRESS32 *base );

The memory bases (in two distinct address spaces) are increased as memory regions are allocated to devices. Allocation will fail (the function returns false) if the base exceeds the limits of the address space (IO is 1MB, memory is 2^32 or 2^64 bytes).

These functions can also be called directly by the application/driver if necessary, but this should not be necessary.

The bases are initialized with default values provided by the HAL. It is possible for an application to override these using the following functions:

void cyg_pci_set_memory_base(  CYG_PCI_ADDRESS64 base );
void cyg_pci_set_io_base( CYG_PCI_ADDRESS32 base );

When a device has been configured, the cyg_pci_device structure will contain the physical address in the CPU's address space where the device's memory regions can be accessed.

This information is provided in base_map[] - there is a 32 bit word for each of the device's BARs. For 32 bit PCI memory regions, each 32 bit word will be an actual pointer that can be used immediately by the driver: the memory space will normally be linearly addressable by the CPU.

However, for 64 bit PCI memory regions, some (or all) of the region may be outside of the CPUs address space. In this case the driver will need to know how to access the region in segments. This functionality may be adopted by the eCos HAL if deemed useful in the future. The 2GB available on many systems should suffice though.

Interrupts

A device may generate interrupts. The HAL vector associated with a given device on the bus is platform specific. This function allows a driver to find the actual interrupt vector for a given device:

cyg_bool cyg_pci_translate_interrupt(  cyg_pci_device *dev_info,
                                       CYG_ADDRWORD *vec );

If the function returns false, no interrupts will be generated by the device. If it returns true, the CYG_ADDRWORD pointed to by vec is updated with the HAL interrupt vector the device will be using. This is how the function is used in the pci1 test:

            if (cyg_pci_translate_interrupt(&dev_info, &irq))
                diag_printf(" Wired to HAL vector %d\n", irq);
            else
                diag_printf(" Does not generate interrupts.\n");

The application/drive should attach an interrupt handler to a device's interrupt before activating the device.

Activating a device

When the device has been allocated memory space it can be activated. This is not done by the library since a driver may have to initialize more state on the device before it can be safely activated.

Activating the device is done by enabling flags in its command word. As an example, see the pci1 test which can be configured to enable the devices it finds. This allows these to be accessed from GDB (if a breakpoint is set on cyg_test_exit):

#ifdef ENABLE_PCI_DEVICES
      {
          cyg_uint16 cmd;

          // Don't use cyg_pci_set_device_info since it clears
          // some of the fields we want to print out below.
          cyg_pci_read_config_uint16(dev_info.devid,
                                     CYG_PCI_CFG_COMMAND, &cmd);
          cmd |= CYG_PCI_CFG_COMMAND_IO|CYG_PCI_CFG_COMMAND_MEMORY;
          cyg_pci_write_config_uint16(dev_info.devid,
                                      CYG_PCI_CFG_COMMAND, cmd);
      }
      diag_printf(" **** Device IO and MEM access enabled\n");
#endif

Links

See these links for more information about PCI:

1. http://www.pcisig.com/ - information on the PCI specifications

2. http://www.yourvote.com/pci/ - list of vendor and device IDs

3. http://www.picmg.org/ - PCI Industrial Computer Manufacturers Group