PCI Bus

空灵静世

PCI Bus

The Peripheral Component Interconnect (PCI) bus is the standard IO bus on recent computers in general, and PCs in particular. There's a lot more good information out there about it than I could even pretend to write, so here are some references.

References

• PCI - Webopedia Definition and Links (these guys give lots of good explanations of concepts, not much really PCI-specific)
• Adaptec, PCI, 64-Bit and 66-MHz Benefits (good concept-level description)
• Adaptec, PCI Bridges
• pci_sokos.htm (full of terrific low-level details)
• Rubini, Alessandro, and Jonathon Corbet, Linux Device Drivers, 2nd Edition Chapter 15, O'Reilly and Associates, 2001 (Linux-centric -- surprise, surprise. Does a good job covering PCI).
• PCI Express Whitepaper from PCI-SIG

Just as a brief note, it was developed by Intel in 1993 to replace the various busses which had been in use on both PCs and Macintoshes. To Intel's credit, it is a remarkably architecture-neutral bus. A very brief description would be that it is a 32-bit, 33MHz bus with multiplexed address and data, and very nice capabilities for autoconfiguration ("Plug and Play"). It also supports both old, 5 volt devices and newer, 3.3 volt devices.

There are many extensions to PCI. Best known is that it has simply been extended to 64 bits and 66 MHz. In addition, there is a variant called PC-104+, which is a 32-bit PCI bus in a highly shock and vibration resistant packaging. PCI-X is a backward-compatible extension to PCI, with PCI-X itself running at 266MHz and PCI-X 2.0 at 533 MHz. This latter also defines a 16 bit interface for space-constrained applications, and a new bus mastering protocol (PCI SIG likes to call this peer-to-peer) that looks a lot like messaging.

All transfers on the PCI bus are "burst" transfers. What this means is that once a device obtains the bus to perform a transfer, it is able to hang on to the bus indefinitely, and keep sending more data every bus cycle (there's actually a time in the bus controller which will take control back after some configurable time period, to keep transfers from being too long. The longer the tranfers are the better the throughput, but this can cause unacceptable delays for other devices).

Autoconfiguration

One of the nicest features of PCI is its support for autoconfiguration. In addition to every device having an address on the PCI bus, every card has its own address determined by which slot it is plugged into. This is referred to as the card's configuration space, and can be queried (and parts of it can be written) by the CPU. This normally occurs at boot time; it may be performed by the BIOS prior to starting the boot loader, or it may be performed by the OS as it boots.

Here's a picture of the configuration space for a PCI device (taken from the Rubini page above):

The most important parts of the configuration space (IMHO) are:

Vendor and Device ID The Vendor ID is a 16 bit number, assigned by the PCI SIG. You can look this number up in a database to find out who built the card. The device ID is another 16 bit number, assigned by the vendor. You can look this up in a database to find out the device model number. Put them together and you can know what kind of device you're going to be talking to, so you can run the right device driver.

Class Code This is a 24 bit number, assigned by I-don't-know-who, which identifies what kind of device is on the card. The difference between this and the vendor/device id fields is that this will specify something like "serial port" while the vendor and device ID fields will say "Bob's Card Shop Model XY-Zowie." You can run the device based on its class code, but to take advantage of any extra features (like the fact it might be an 8-port card instead of a single-port card) requires the vendor and device IDs. Base Registers Up to six base registers can be specified, for the devices located on the card. If you have fewer than six logical devices you will actually use fewer than these; if you have more, you will have to get into some ugly hacks (for instance, on an eight port serial card I have, six of the ports' base addresses are specified in the base addresses, while two are at fixed offsets from the first two of the six). Unlike the vendor and device ID fields, and the class codes, the base register addresses are read/write. PCI Interrupt Handling

As with many aspects of the PCI bus, one of the challenges is to design the interrupt handling so that it can be mapped into the interrupt scheme expected by the CPU. The basic solution is the same as for other aspects of the bus: chips called ``bridges'' are used for the translation.

PCI uses four pins, called INTA-INTD, for interrupt requests. When an interrupt is required, the proper pin is asserted. A card which only has a single interrupt will normally use INTA; a card with two (they exist! Particularly cards with more than one logical device) will use INTA and INTB, and so forth.

The bus wiring determines how the requested interrupt is presented to the bridge chip; the standard doesn't specify how this routing should be performed. I've come across one source that says the routing for current PCs is:

	Slot 1	Slot 2	Slot 3	Slot 4	Slot 5
INTA	PIRQ1	PIRQ2	PIRQ3	PIRQ4	PIRQ4
INTB	PIRQ2	PIRQ3	PIRQ4	PIRQ1	PIRQ1
INTC	PIRQ3	PIRQ4	PIRQ1	PIRQ2	PIRQ2
INTD	PIRQ4	PIRQ1	PIRQ2	PIRQ3	PIRQ3

(where the PIRQ# is the interrupt as presented to the bridge chip). So if in fact all the devices are using INTA, they will be routed to different pins (except for cards 4 and 5). Notice that this is how they are wired, not how they have to be wired; it would be entirely possible for a bus to route all 20 interrupts from these five devices to different inputs on the bridge. On a PC, the BIOS programs the bridge to route its PIRQ inputs to Intel IRQ requests in an emulated pair of 8259s.

When the device requests its interrupt, the bridge responds with an Interrupt Acknowledge (INTA) bus cycle; the card responds with an interrupt vector. This vector is an eight-bit number loaded into a device configuration register by the BIOS or the OS at boot time.

I haven't been able to find a specification of the arbitration that decides which device wins when multiple devices attempt to interrupt simultaneously.

PCI Commands

There are a total of 16 possible commands on a PCI cycle. They're in the following table:

Command	Command Type
0000	Interrupt Acknowledge
0001	Special Cycle
0010	I/O Read
0011	I/O Write
0100	reserved
0101	reserved
0110	Memory Read
0111	Memory Write
1000	reserved
1001	reserved
1010	Configuration Read
1011	Configuration Write
1100	Multiple Memory Read
1101	Dual Address Cycle
1110	Memory-Read Line
1111	Memory Write and Invalidate

Here are some notes on the different transfer types (taken almost verbatim from pci_sokos.htm).

Interrupt Acknowledge (0000) The interrupt controller automatically recognizes and reacts to the INTA (interrupt acknowledge) command. In the data phase, it transfers the interrupt vector to the AD lines. Special Cycle (0001)

AD15-AD0
0x0000	Processor Shutdown
0x0001	Processor Halt
0x0002	x86 Specific Code
0x0003 to 0xFFFF	Reserved

I/O Read (0010) and I/O Write (0011) Input/Output device read or write operation. The AD lines contain a byte address (AD0 and AD1 must be decoded). PCI I/O ports may be 8 or 16 bits. PCI allows 32 bits of address space. On IBM compatible machines, the Intel CPU is limited to 16 bits of I/O space, which is further limited by some ISA cards that may also be installed in the machine (many ISA cards only decode the lower 10 bits of address space, and thus mirror themselves throughout the 16 bit I/O space). This limit assumes that the machine supports ISA or EISA slots in addition to PCI slots. The PCI configuration space may also be accessed through I/O ports 0x0CF8 (Address) and 0x0CFC (Data). The address port must be written first. Memory Read (0110) and Memory Write (0111) A read or write to the system memory space. The AD lines contain a doubleword address. AD0 and AD1 do not need to be decoded. The Byte Enable lines (C/BE) indicate which bytes are valid. Configuration Read (1010) and Configuration Write (1011) A read or write to the PCI device configuration space, which is 256 bytes in length. It is accessed in doubleword units. AD0 and AD1 contain 0, AD2-7 contain the doubleword address, AD8-10 are used for selecting the addressed unit a the malfunction unit, and the remaining AD lines are not used. Multiple Memory Read (1100) This is an extension of the memory read bus cycle. It is used to read large blocks of memory without caching, which is beneficial for long sequential memory accesses. Dual Address Cycle (1101) Two address cycles are necessary when a 64 bit address is used, but only a 32 bit physical address exists. The least significant portion of the address is placed on the AD lines first, followed by the most significant 32 bits. The second address cycle also contains the command for the type of transfer (I/O, Memory, etc). The PCI bus supports a 64 bit I/O address space, although this is not available on Intel based PCs due to limitations of the CPU. Memory-Read Line (1110) This cycle is used to read in more than two 32 bit data blocks, typically up to the end of a cache line. It is more effecient than normal memory read bursts for a long series of sequential memory accesses. Memory Write and Invalidate (1111) This indicates that a minimum of one cache line is to be transferred. This allows main memory to be updated, saving a cache write-back cycle. PCI Express

PCI Express (a bus formerly known as 3GIO) is the successor to PCI. In fact, while the hardware is very different from PCI's the software model is unchanged. If a manufacturer moves a device from a PCI implementation to a PCI Express implementation (and makes no other changes), the old drivers will all continue to work.

Physical Layer

The first thing to notice about PCI Express is that it isn't a bus. Instead, each PCI express slot is independently connected to a switch. The communication between devices and the bus controller is very reliable, with CRC being used to check for errors in transmission.

Devices are connected to PCI Express with two differential signal pairs: one for transmitting and one for receiving. Signals are transmitted across these signal pairs at 2.5 Gb/s/direction (and they're anticipating scaling to 10 Gb/s/direction later). Notice that this means the data transfers across a single signal pair is already substantially faster than across PCI (32 bits * 33 MHz = 1 Gb/s, half-duplex), with later scaling.

But there's more. What I just described is what PCI-SIG calls a "lane." Where PCI has a 120-pin connector, a one-lane PCI Express device uses a 36 pin connector. The specification allows a device to use more than one lane; adding another lane adds another 2.5 GB/s/direction. You can get up to an 8x PCI Express connector before it takes more pins than PCI did; 16x (used for graphics) and 32x connectors have been defined.

The last thing to comment on is that the architecture automatically supports slower devices: if you have a (say) 2x board, it'll plug into an 8x connector just fine. It just won't use all the 8x pins, and will only run at 2x speed. But you can't plug a wider card into a narrower slot.

Last modified: Fri Apr 22 09:28:45 MDT 2005
原文链接：http://www.cs.nmsu.edu/~pfeiffer/classes/473/notes/pci.html