Peripheral Component Interconnect

Card voltage and keying

Typical PCI cards have either one or two key notches, depending on their signaling voltage. Cards requiring 3.3 volts have a notch 56.21 mm from the card backplate in pin positions 12 and 13; those requiring 5 volts have a notch 104.41 mm from the backplate in pin positions 50 and 51. This allows cards to be fitted only into slots with a voltage they support. "Universal cards" accepting either voltage have both key notches.

Connector pinout

The PCI connector is defined as having 62 contacts on each side of the edge connector, but two or four of them are replaced by key notches, so a card has 60 or 58 contacts on each side. Side A refers to the solder side and side B refers to the component side: if the card is held with the connector pointing down, a view of side A will have the backplate on the right, whereas a view of side B will have the backplate on the left. The pinout of B and A sides is as follows, looking down into the motherboard connector (pins A1 and B1 are closest to the backplate).^[16][18][19]

64-bit PCI extends this by an additional 32 contacts on each side which provide AD[63:32], C/BE[7:4]#, the PAR64 parity signal, and a number of power and ground pins.

Most lines are connected to each slot in parallel. The exceptions are:

• Each slot has its own REQ# output to, and GNT# input from the motherboard arbiter.
• Each slot has its own IDSEL line, usually connected to a specific AD line.
• TDO is daisy-chained to the following slot's TDI. Cards without JTAG support must connect TDI to TDO so as not to break the chain.
• PRSNT1# and PRSNT2# for each slot have their own pull-up resistors on the motherboard. The motherboard may (but does not have to) sense these pins to determine the presence of PCI cards and their power requirements.
• REQ64# and ACK64# are individually pulled up on 32-bit-only slots.
• The interrupt pins INTA# through INTD# are connected to all slots in different orders. (INTA# on one slot is INTB# on the next and INTC# on the one after that.)

Notes:

• IOPWR is +3.3 V or +5 V, depending on the backplane. The slots also have a ridge in one of two places which prevents insertion of cards that do not have the corresponding key notch, indicating support for that voltage standard. Universal cards have both key notches and use IOPWR to determine their I/O signal levels.
• The PCI SIG strongly encourages 3.3 V PCI signaling,^[16] requiring support for it since standard revision 2.3,^[18] but most PC motherboards use the 5 V variant. Thus, while many currently available PCI cards support both, and have two key notches to indicate that, there are still a large number of 5-V-only cards on the market.^{[needs update]}
• The M66EN pin is an additional ground on 5 V PCI buses found in most PC motherboards. Cards and motherboards that do not support 66 MHz operation also ground this pin. If all participants support 66 MHz operation, a pull-up resistor on the motherboard raises this signal high and 66 MHz operation is enabled. The pin is still connected to ground via coupling capacitors on each card to preserve its AC shielding function.^{[needs update]}
• The PCIXCAP pin is an additional ground on PCI buses and cards. If all cards and the motherboard support the PCI-X protocol, a pull-up resistor on the motherboard raises this signal high and PCI-X operation is enabled. The pin is still connected to ground via coupling capacitors on each card to preserve its AC shielding function.
• At least one of PRSNT1# and PRSNT2# must be grounded by the card. The combination chosen indicates the total power requirements of the card (25 W, 15 W, or 7.5 W).
• SBO# and SDONE are signals from a cache controller to the current target. They are not initiator outputs, but are colored that way because they are target inputs.
• PME# (19 A) – Power management event (optional) which is supported in PCI version 2.2 and higher. It is a 3.3 V, open drain, active low signal.^[20] PCI cards may use this signal to send and receive PME via the PCI socket directly, which eliminates the need for a special Wake-on-LAN cable.^[21]

Mixing of 32-bit and 64-bit PCI cards in different width slots

Most 32-bit PCI cards will function properly in 64-bit PCI-X slots, but the bus clock rate will be limited to the clock frequency of the slowest card, an inherent limitation of PCI's shared bus topology. For example, when a PCI 2.3, 66-MHz peripheral is installed into a PCI-X bus capable of 133 MHz, the entire bus backplane will be limited to 66 MHz. To get around this limitation, many motherboards have two or more PCI/PCI-X buses, with one bus intended for use with high-speed PCI-X peripherals, and the other bus intended for general-purpose peripherals.

Many 64-bit PCI-X cards are designed to work in 32-bit mode if inserted in shorter 32-bit connectors, with some loss of performance.^[22][23] An example of this is the Adaptec 29160 64-bit SCSI interface card.^[24] However, some 64-bit PCI-X cards do not work in standard 32-bit PCI slots.^{[25][unreliable source?]}

Installing a 64-bit PCI-X card in a 32-bit slot will leave the 64-bit portion of the card edge connector not connected and overhanging. This requires that there be no motherboard components positioned so as to mechanically obstruct the overhanging portion of the card edge connector.

Mini PCI

Mini PCI was added to PCI version 2.2 for use in laptops and some routers;^{[citation needed]} it uses a 32-bit, 33 MHz bus with powered connections (3.3 V only; 5 V is limited to 100 mA) and support for bus mastering and DMA. The standard size for Mini PCI cards is approximately a quarter of their full-sized counterparts. There is no access to the card from outside the case, unlike desktop PCI cards with brackets carrying connectors. This limits the kinds of functions a Mini PCI card can perform.

Many Mini PCI devices were developed such as Wi-Fi, Fast Ethernet, Bluetooth, modems (often Winmodems), sound cards, cryptographic accelerators, SCSI, IDE–ATA, SATA controllers and combination cards. Mini PCI cards can be used with regular PCI-equipped hardware, using Mini PCI-to-PCI converters. Mini PCI has been superseded by the much narrower PCI Express Mini Card.

PCI address spaces

PCI has three address spaces: memory, I/O address, and configuration.

Memory addresses are 32 bits (optionally 64 bits) in size, support caching and can be burst transactions.

I/O addresses are for compatibility with the Intel x86 architecture's I/O port address space. Although the PCI bus specification allows burst transactions in any address space, most devices only support it for memory addresses and not I/O.

Finally, PCI configuration space provides access to 256 bytes of special configuration registers per PCI device. Each PCI slot gets its own configuration space address range. The registers are used to configure devices' memory and I/O address ranges they should respond to from transaction initiators. When a computer is first turned on, all PCI devices respond only to their configuration space accesses. The computer's BIOS scans for devices and assigns Memory and I/O address ranges to them.

If an address is not claimed by any device, the transaction initiator's address phase will time out, causing the initiator to abort the operation. In case of reads, it is customary to supply all-ones for the read data value (0xFFFFFFFF) in this case. PCI devices therefore generally attempt to avoid using the all-ones value in important status registers, so that such an error can be easily detected by software.

PCI command codes

There are 16 possible 4-bit command codes, and 12 of them are assigned. With the exception of the unique dual address cycle, the least significant bit of the command code indicates whether the following data phases are a read (data sent from target to initiator) or a write (data sent from an initiator to target). PCI targets must examine the command code as well as the address and not respond to address phases that specify an unsupported command code.

The commands that refer to cache lines depend on the PCI configuration space cache line size register being set up properly; they may not be used until that has been done.

0000: Interrupt Acknowledge

This is a special form of read cycle implicitly addressed to the interrupt controller, which returns an interrupt vector. The 32-bit address field is ignored. One possible implementation is to generate an interrupt acknowledge cycle on an ISA bus using a PCI/ISA bus bridge. This command is for IBM PC compatibility; if there is no Intel 8259 style interrupt controller on the PCI bus, this cycle need never be used.

0001: Special Cycle

This cycle is a special broadcast write of system events that PCI cards may be interested in. The address field of a special cycle is ignored, but it is followed by a data phase containing a payload message. The currently defined messages announce that the processor is stopping for some reason (e.g., to save power). No device ever responds to this cycle; it is always terminated with a master abort after leaving the data on the bus for at least 4 cycles.

0010: I/O Read

This performs a read from I/O space. All 32 bits of the read address are provided, so that a device may (for compatibility reasons) implement less than 4 bytes worth of I/O registers. If the byte enables request data not within the address range supported by the PCI device (e.g. a 4-byte read from a device which only supports 2 bytes of I/O address space), it must be terminated with a target abort. Multiple data cycles are permitted, using linear (simple incrementing) burst ordering.

The PCI standard is discouraging the use of I/O space in new devices, preferring that as much as possible be done through main memory mapping.

0011: I/O Write

This performs a write to I/O space.

010x: Reserved

A PCI device must not respond to an address cycle with these command codes.

0110: Memory Read

This performs a read cycle from memory space. Because the smallest memory space a PCI device is permitted to implement is 16 bytes,^{[18][16]: §6.5.2.1} the two least significant bits of the address are not needed during the address phase; equivalent information will arrive during the data phases in the form of byte select signals. They instead specify the order in which burst data must be returned.^{[18][16]: §3.2.2.2} If a device does not support the requested order, it must provide the first word and then disconnect.

If a memory space is marked as "prefetchable", then the target device must ignore the byte-select signals on a memory read and always return 32 valid bits.

0111: Memory Write

This operates similarly to a memory read. The byte select signals are more important in a write, as unselected bytes must not be written to memory.

Generally, PCI writes are faster than PCI reads, because a device may buffer the incoming write data and release the bus faster. For a read, it must delay the data phase until the data has been fetched.

100x: Reserved

A PCI device must not respond to an address cycle with these command codes.

1010: Configuration Read

This is similar to an I/O read, but reads from PCI configuration space. A device must respond only if the low 11 bits of the address specify a function and register that it implements, and if the special IDSEL signal is asserted. It must ignore the high 21 bits. Burst reads (using linear incrementing) are permitted in PCI configuration space.

Unlike I/O space, standard PCI configuration registers are defined so that reads never disturb the state of the device. It is possible for a device to have configuration space registers beyond the standard 64 bytes which have read side effects, but this is rare.^[31]

Configuration space accesses often have a few cycles of delay to allow the IDSEL lines to stabilize, which makes them slower than other forms of access. Also, a configuration space access requires a multi-step operation rather than a single machine instruction. Thus, it is best to avoid them during routine operation of a PCI device.

1011: Configuration Write

This operates analogously to a configuration read.

1100: Memory Read Multiple

This command is identical to a generic memory read, but includes the hint that a long read burst will continue beyond the end of the current cache line, and the target should internally prefetch a large amount of data. A target is always permitted to consider this a synonym for a generic memory read.

1101: Dual Address Cycle

When accessing a memory address that requires more than 32 bits to represent, the address phase begins with this command and the low 32 bits of the address, followed by a second cycle with the actual command and the high 32 bits of the address. PCI targets that do not support 64-bit addressing may simply treat this as another reserved command code and not respond to it. This command code may only be used with a non-zero high-order address word; it is forbidden to use this cycle if not necessary.

1110: Memory Read Line

This command is identical to a generic memory read, but includes the hint that the read will continue to the end of the cache line. A target is always permitted to consider this a synonym for a generic memory read.

1111: Memory Write and Invalidate

This command is identical to a generic memory write, but comes with the guarantee that one or more whole cache lines will be written, with all byte selects enabled. This is an optimization for write-back caches snooping the bus. Normally, a write-back cache holding dirty data must interrupt the write operation long enough to write its own dirty data first. If the write is performed using this command, the data to be written back is guaranteed to be irrelevant, and may simply be invalidated in the write-back cache.

This optimization only affects the snooping cache, and makes no difference to the target, which may treat this as a synonym for the memory write command.

Delayed transactions

Devices unable to meet those timing restrictions must use a combination of posted writes (for memory writes) and delayed transactions (for other writes and all reads). In a delayed transaction, the target records the transaction (including the write data) internally and aborts (asserts STOP# rather than TRDY#) the first data phase. The initiator must retry exactly the same transaction later. In the interim, the target internally performs the transaction and waits for the retried transaction. When the retried transaction is seen, the buffered result is delivered.

A device may be the target of other transactions while completing one delayed transaction; it must remember the transaction type, address, byte selects and (if a write) data value, and only complete the correct transaction.

If the target has a limit on the number of delayed transactions that it can record internally (simple targets may impose a limit of 1), it will force those transactions to retry without recording them. They will be dealt with when the current delayed transaction is completed. If two initiators attempt the same transaction, a delayed transaction begun by one may have its result delivered to the other; this is harmless.

A target abandons a delayed transaction when a retry succeeds in delivering the buffered result, the bus is reset, or when 2¹⁵=32768 clock cycles (approximately 1 ms) elapse without seeing a retry. The latter should never happen in normal operation, but it prevents a deadlock of the whole bus if one initiator is reset or malfunctions.

Posted writes

Generally, when a bus bridge sees a transaction on one bus that must be forwarded to the other, the original transaction must wait until the forwarded transaction completes before a result is ready. One notable exception occurs in the case of memory writes. Here, the bridge may record the write data internally (if it has room) and signal completion of the write before the forwarded write has completed. Or, indeed, before it has begun. Such "sent but not yet arrived" writes are referred to as "posted writes", by analogy with a postal mail message. Although they offer great opportunity for performance gains, the rules governing what is permissible are somewhat intricate.^[34]

Combining, merging, and collapsing

The PCI standard permits bus bridges to convert multiple bus transactions into one larger transaction under certain situations. This can improve the efficiency of the PCI bus.

Signal timing

All PCI bus signals are sampled on the rising edge of the clock. Signals nominally change on the falling edge of the clock, giving each PCI device approximately one half a clock cycle to decide how to respond to the signals it observed on the rising edge, and one half a clock cycle to transmit its response to the other device.

The PCI bus requires that every time the device driving a PCI bus signal changes, one turnaround cycle must elapse between the time the one device stops driving the signal and the other device starts. Without this, there might be a period when both devices were driving the signal, which would interfere with bus operation.

The combination of this turnaround cycle and the requirement to drive a control line high for one cycle before ceasing to drive it means that each of the main control lines must be high for a minimum of two cycles when changing owners. The PCI bus protocol is designed so this is rarely a limitation; only in a few special cases (notably fast back-to-back transactions) is it necessary to insert additional delay to meet this requirement.

Arbitration

Any device on a PCI bus that is capable of acting as a bus master may initiate a transaction with any other device. To ensure that only one transaction is initiated at a time, each master must first wait for a bus grant signal, GNT#, from an arbiter located on the motherboard. Each device has a separate request line REQ# that requests the bus, but the arbiter may "park" the bus grant signal at any device if there are no current requests.

The arbiter may remove GNT# at any time. A device that loses GNT# may complete its current transaction, but may not start one (by asserting FRAME#) unless it observes GNT# asserted the cycle before it begins.

The arbiter may also provide GNT# at any time, including during another master's transaction. During a transaction, either FRAME# or IRDY# or both are asserted; when both are deasserted, the bus is idle. A device may initiate a transaction at any time that GNT# is asserted and the bus is idle.

Address phase

A PCI bus transaction begins with an address phase. The initiator (usually a chipset), seeing that it has GNT# and the bus is idle, drives the target address onto the AD[31:0] lines, the associated command (e.g. memory read, or I/O write) on the C/BE[3:0]# lines, and pulls FRAME# low.

Each other device examines the address and command and decides whether to respond as the target by asserting DEVSEL#. A device must respond by asserting DEVSEL# within 3 cycles. Devices that promise to respond within 1 or 2 cycles are said to have "fast DEVSEL" or "medium DEVSEL", respectively. (Actually, the time to respond is 2.5 cycles, since PCI devices must transmit all signals half a cycle early so that they can be received three cycles later.)

A device must latch the address on the first cycle; the initiator is required to remove the address and command from the bus on the following cycle, even before receiving a DEVSEL# response. The additional time is available only for interpreting the address and command after it is captured.

On the fifth cycle of the address phase (or earlier if all other devices have medium DEVSEL or faster), a catch-all "subtractive decoding" is allowed for some address ranges. This is commonly used by an ISA bus bridge for addresses within its range (24 bits for memory and 16 bits for I/O).

On the sixth cycle, if there has been no response, the initiator may abort the transaction by deasserting FRAME#. This is known as master abort termination and it is customary for PCI bus bridges to return all-ones data (0xFFFFFFFF) in this case. PCI devices, therefore, are generally designed to avoid using the all-ones value in important status registers, so that such an error can be easily detected by software.

              _  0_  1_  2_  3_  4_  5_  6_
        CLK _/ \_/ \_/ \_/ \_/ \_/ \_/ \_/
            ___
       GNT#    \___/XXXXXXXXXXXXXXXXXXXXXXX
            _______
     FRAME#        \_______________________
                    ___ ___
   AD[31:0] -------<___X___>--------------- (Low, then high bits)
                    ___ ___ _______________
 C/BE[3:0]# -------<___X___X_______________ (DAC, then actual command)
            ___________________________
    DEVSEL#                \___\___\___\___
                         Fast Med Slow
              _   _   _   _   _   _   _   _
        CLK _/ \_/ \_/ \_/ \_/ \_/ \_/ \_/
                 0   1   2   3   4   5   6

Data phases

After the address phase (specifically, beginning with the cycle that DEVSEL# goes low) comes a burst of one or more data phases. In all cases, the initiator drives active-low byte select signals on the C/BE[3:0]# lines, but the data on the AD[31:0] may be driven by the initiator (in case of writes) or target (in case of reads).

During data phases, the C/BE[3:0]# lines are interpreted as active-low byte enables. In case of a write, the asserted signals indicate which of the four bytes on the AD bus are to be written to the addressed location. In the case of a read, they indicate which bytes the initiator is interested in. For reads, it is always legal to ignore the byte-enable signals and simply return all 32 bits; cacheable memory resources are required to always return 32 valid bits. The byte enables are mainly useful for I/O space accesses where reads have side effects.

A data phase with all four C/BE# lines deasserted is explicitly permitted by the PCI standard, and must have no effect on the target other than to advance the address in the burst access in progress.

The data phase continues until both parties are ready to complete the transfer and continue to the next data phase. The initiator asserts IRDY# (initiator ready) when it no longer needs to wait, while the target asserts TRDY# (target ready). Whichever side is providing the data must drive it on the AD bus before asserting its ready signal.

Once one of the participants asserts its ready signal, it may not become unready or otherwise alter its control signals until the end of the data phase. The data recipient must latch the AD bus each cycle until it sees both IRDY# and TRDY# asserted, which marks the end of the current data phase and indicates that the just-latched data is the word to be transferred.

To maintain full burst speed, the data sender then has half a clock cycle after seeing both IRDY# and TRDY# asserted to drive the next word onto the AD bus.

             0_  1_  2_  3_  4_  5_  6_  7_  8_  9_
        CLK _/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \_/
                ___         _______     ___ ___ ___
   AD[31:0] ---<___XXXXXXXXX_______XXXXX___X___X___ (If a write)
                ___             ___ _______ ___ ___
   AD[31:0] ---<___>~~~<XXXXXXXX___X_______X___X___ (If a read)
                ___ _______________ _______ ___ ___
 C/BE[3:0]# ---<___X_______________X_______X___X___ (Must always be valid)
            _______________      |  ___  |   |   |
      IRDY#              x \_______/ x \___________
            ___________________  |       |   |   |
      TRDY#              x   x \___________________
            ___________          |       |   |   |
    DEVSEL#            \___________________________
            ___                  |       |   |   |
     FRAME#    \___________________________________
              _   _   _   _   _  |_   _  |_  |_  |_
        CLK _/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \_/
             0   1   2   3   4   5   6   7   8   9

This continues the address cycle illustrated above, assuming a single address cycle with medium DEVSEL, so the target responds in time for clock 3. However, at that time, neither side is ready to transfer data. For clock 4, the initiator is ready, but the target is not. On clock 5, both are ready, and a data transfer takes place (as indicated by the vertical lines). For clock 6, the target is ready to transfer, but the initiator is not. On clock 7, the initiator becomes ready, and data is transferred. For clocks 8 and 9, both sides remain ready to transfer data, and data is transferred at the maximum possible rate (32 bits per clock cycle).

In case of a read, clock 2 is reserved for turning around the AD bus, so the target is not permitted to drive data on the bus even if it is capable of fast DEVSEL.

Ending transactions

Either side may request that a burst end after the current data phase. Simple PCI devices that do not support multi-word bursts will always request this immediately. Even devices that do support bursts will have some limit on the maximum length they can support, such as the end of their addressable memory.

Burst addressing

For memory space accesses, the words in a burst may be accessed in several orders. The unnecessary low-order address bits AD[1:0] are used to convey the initiator's requested order. A target that does not support a particular order must terminate the burst after the first word. Some of these orders depend on the cache line size, which is configurable on all PCI devices.

If the starting offset within the cache line is zero, all of these modes reduce to the same order.

Cache line toggle and cache line wrap modes are two forms of critical-word-first cache line fetching. Toggle mode XORs the supplied address with an incrementing counter. This is the native order for Intel 486 and Pentium processors. It has the advantage that it is not necessary to know the cache line size to implement it.

PCI version 2.1 obsoleted toggle mode and added the cache line wrap mode,^[32]: 2 where fetching proceeds linearly, wrapping around at the end of each cache line. When one cache line is completely fetched, fetching jumps to the starting offset in the next cache line.

Most PCI devices only support a limited range of typical cache line sizes; if the cache line size is programmed to an unexpected value, they force single-word access.

PCI also supports burst access to I/O and configuration space, but only linear mode is supported. (This is rarely used, and may be buggy in some devices; they may not support it, but not properly force single-word access either.)

Transaction examples

This is the highest-possible speed four-word write burst, terminated by the master:

             0_  1_  2_  3_  4_  5_  6_  7_
        CLK _/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \
                ___ ___ ___ ___ ___
   AD[31:0] ---<___X___X___X___X___>---<___>
                ___ ___ ___ ___ ___
 C/BE[3:0]# ---<___X___X___X___X___>---<___>
                     |   |   |   |  ___
      IRDY# ^^^^^^^^\______________/   ^^^^^
                     |   |   |   |  ___
      TRDY# ^^^^^^^^\______________/   ^^^^^
                     |   |   |   |  ___
    DEVSEL# ^^^^^^^^\______________/   ^^^^^
            ___      |   |   |  ___
     FRAME#    \_______________/ | ^^^^\____
              _   _  |_  |_  |_  |_   _   _
        CLK _/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \
             0   1   2   3   4   5   6   7

On clock edge 1, the initiator starts a transaction by driving an address, command, and asserting FRAME#. The other signals are idle (indicated by ^^^), pulled high by the motherboard's pull-up resistors. That might be their turnaround cycle. On cycle 2, the target asserts both DEVSEL# and TRDY#. As the initiator is also ready, a data transfer occurs. This repeats for three more cycles, but before the last one (clock edge 5), the master deasserts FRAME#, indicating that this is the end. On clock edge 6, the AD bus and FRAME# are undriven (turnaround cycle) and the other control lines are driven high for 1 cycle. On clock edge 7, another initiator can start a different transaction. This is also the turnaround cycle for the other control lines.

The equivalent read burst takes one more cycle, because the target must wait 1 cycle for the AD bus to turn around before it may assert TRDY#:

             0_  1_  2_  3_  4_  5_  6_  7_  8_
        CLK _/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \
                ___     ___ ___ ___ ___
   AD[31:0] ---<___>---<___X___X___X___>---<___>
                ___ _______ ___ ___ ___
 C/BE[3:0]# ---<___X_______X___X___X___>---<___>
            ___          |   |   |   |  ___
      IRDY#    ^^^^\___________________/   ^^^^^
            ___    _____ |   |   |   |  ___
      TRDY#    ^^^^     \______________/   ^^^^^
            ___          |   |   |   |  ___
    DEVSEL#    ^^^^\___________________/   ^^^^^
            ___          |   |   |  ___
     FRAME#    \___________________/ | ^^^^\____
              _   _   _  |_  |_  |_  |_   _   _
        CLK _/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \
             0   1   2   3   4   5   6   7   8

A high-speed burst terminated by the target will have an extra cycle at the end:

             0_  1_  2_  3_  4_  5_  6_  7_  8_
        CLK _/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \
                ___     ___ ___ ___ ___
   AD[31:0] ---<___>---<___X___X___X___XXXX>----
                ___ _______ ___ ___ ___ ___
 C/BE[3:0]# ---<___X_______X___X___X___X___>----
                         |   |   |   |      ___
      IRDY# ^^^^^^^\_______________________/
                   _____ |   |   |   |  _______
      TRDY# ^^^^^^^     \______________/
                   ________________  |      ___
      STOP# ^^^^^^^      |   |   | \_______/
                         |   |   |   |      ___
    DEVSEL# ^^^^^^^\_______________________/
            ___          |   |   |   |  ___
     FRAME#    \_______________________/   ^^^^
              _   _   _  |_  |_  |_  |_   _   _
        CLK _/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \_/ \
             0   1   2   3   4   5   6   7   8

On clock edge 6, the target indicates that it wants to stop (with data), but the initiator is already holding IRDY# low, so there is a fifth data phase (clock edge 7), during which no data is transferred.

Parity

The PCI bus detects parity errors, but does not attempt to correct them by retrying operations; it is purely a failure indication. Due to this, there is no need to detect the parity error before it has happened, and the PCI bus actually detects it a few cycles later. During a data phase, whichever device is driving the AD[31:0] lines computes even parity over them and the C/BE[3:0]# lines, and sends that out the PAR line one cycle later. All access rules and turnaround cycles for the AD bus apply to the PAR line, just one cycle later. The device listening on the AD bus checks the received parity and asserts the PERR# (parity error) line one cycle after that. This generally generates a processor interrupt, and the processor can search the PCI bus for the device which detected the error.

The PERR# line is only used during data phases, once a target has been selected. If a parity error is detected during an address phase (or the data phase of a Special Cycle), the devices which observe it assert the SERR# (System error) line.

Even when some bytes are masked by the C/BE# lines and not in use, they must still have some defined value, and this value must be used to compute the parity.

Fast back-to-back transactions

Due to the need for a turnaround cycle between different devices driving PCI bus signals, in general it is necessary to have an idle cycle between PCI bus transactions. However, in some circumstances it is permitted to skip this idle cycle, going directly from the final cycle of one transfer (IRDY# asserted, FRAME# deasserted) to the first cycle of the next (FRAME# asserted, IRDY# deasserted).

An initiator may only perform back-to-back transactions when:

• they are by the same initiator (or there would be no time to turn around the C/BE# and FRAME# lines),
• the first transaction was a write (so there is no need to turn around the AD bus), and
• the initiator still has permission (from its GNT# input) to use the PCI bus.

Additional timing constraints may come from the need to turn around are the target control lines, particularly DEVSEL#. The target deasserts DEVSEL#, driving it high, in the cycle following the final data phase, which in the case of back-to-back transactions is the first cycle of the address phase. The second cycle of the address phase is then reserved for DEVSEL# turnaround, so if the target is different from the prior one, it must not assert DEVSEL# until the third cycle (medium DEVSEL speed).

One case where this problem cannot arise is if the initiator knows somehow (presumably because the addresses share sufficient high-order bits) that the second transfer is addressed to the same target as the prior one. In that case, it may perform back-to-back transactions. All PCI targets must support this.

It is also possible for the target to keep track of the requirements. If it never does fast DEVSEL, they are met trivially. If it does, it must wait until medium DEVSEL time unless:

• the current transaction was preceded by an idle cycle (is not back-to-back), or
• the prior transaction was to the same target, or
• the current transaction began with a double address cycle.

Targets that have this ability indicate it by a special bit in a PCI configuration register, and if all targets on a bus have it, all initiators may use back-to-back transfers freely.

A subtractive decoding bus bridge must know to expect this extra delay in the event of back-to-back cycles, to advertise back-to-back support.

64-bit PCI

Starting from revision 2.1,^{[clarification needed]} the PCI specification includes optional 64-bit support. This is provided via an extended connector which provides the 64-bit bus extensions AD[63:32], C/BE[7:4]#, and PAR64, and a number of additional power and ground pins. The 64-bit PCI connector can be distinguished from a 32-bit connector by the additional 64-bit segment.

Memory transactions between 64-bit devices may use all 64 bits to double the data transfer rate. Non-memory transactions (including configuration and I/O space accesses) may not use the 64-bit extension. During a 64-bit burst, burst addressing works just as in a 32-bit transfer, but the address is incremented twice per data phase. The starting address must be 64-bit aligned; i.e., AD2 must be 0. The data corresponding to the intervening addresses (with AD2 = 1) is carried on the upper half of the AD bus.

To initiate a 64-bit transaction, the initiator drives the starting address on the AD bus and asserts REQ64# at the same time as FRAME#. If the selected target can support a 64-bit transfer for this transaction, it replies by asserting ACK64# at the same time as DEVSEL#. A target may decide on a per-transaction basis whether to allow a 64-bit transfer.

If REQ64# is asserted during the address phase, the initiator also drives the high 32 bits of the address and a copy of the bus command on the high half of the bus. If the address requires 64 bits, a dual address cycle is still required, but the high half of the bus carries the upper half of the address and the final command code during both address phase cycles; this allows a 64-bit target to see the entire address and begin responding earlier.

If the initiator sees DEVSEL# asserted without ACK64#, it performs 32-bit data phases. The data which would have been transferred on the upper half of the bus during the first data phase is instead transferred during the second data phase. Typically, the initiator drives all 64 bits of data before seeing DEVSEL#. If ACK64# is missing, it may cease driving the upper half of the data bus.

The REQ64# and ACK64# lines are held asserted for the entire transaction, save the last data phase, and deasserted at the same time as FRAME# and DEVSEL#, respectively.

The PAR64 line operates just like the PAR line, but provides even parity over AD[63:32] and C/BE[7:4]#. It is only valid for address phases if REQ64# is asserted. PAR64 is only valid for data phases if both REQ64# and ACK64# are asserted.

Cache snooping (obsolete)

PCI originally included optional support for write-back cache coherence. This required support by cacheable memory targets, which would listen to two pins from the cache on the bus, SDONE (snoop done) and SBO# (snoop backoff).^[35]

Because this was rarely implemented in practice, it was deleted from revision 2.2 of the PCI specification,^[16][36] and the pins re-used for SMBus access in revision 2.3.^[18]

The cache would watch all memory accesses, without asserting DEVSEL#. If it noticed an access that might be cached, it would drive SDONE low (snoop not done). A coherence-supporting target would avoid completing a data phase (asserting TRDY#) until it observed SDONE high.

In the case of a write to data that was clean in the cache, the cache would only have to invalidate its copy and would assert SDONE as soon as this was established. However, if the cache contained dirty data, the cache would have to write it back before the access could proceed. so it would assert SBO# when raising SDONE. This would signal the active target to assert STOP# rather than TRDY#, causing the initiator to disconnect and retry the operation later. In the meantime, the cache would arbitrate for the bus and write its data back to memory.

Targets supporting cache coherency are also required to terminate bursts before they cross cache lines.