Applying Routing Mechanisms
Once configuration of the system routing strategy is complete and transactions are enabled, PCI Express devices decode inbound TLP headers and use corresponding fields in configuration space Base Address Registers, Base/Limit registers, and Bus Number registers to apply address, ID, and implicit routing to the packet. Note that there are actually two levels of decision: the device first determines if the packet targets an internal location; if not, and the device is a switch, it will evaluate the packet to see if it should be forwarded out of an egress port. A third possibility is that the packet has been received in error or is malformed; in this case, it will be handled as a receive error. There are a number of cases when this may happen, and a number of ways it may be handled. Refer to “PCI Express Error Checking Mechanisms” on page 356 for a description of error checking and handling. The following sections describe the basic features of each routing mechanism; we will assume no errors are encountered.
PCI Express transactions using address routing reference the same system memory and IO maps that PCI and PCIX transactions do. Address routing is used to transfer data to or from memory, memory mapped IO, or IO locations. Memory transaction requests may carry either 32 bit addresses using the 3DW TLP header format, or 64 bit addresses using the 4DW TLP header format. IO transaction requests are restricted to 32 bits of address using the 3DW TLP header format, and should only target legacy devices.
Memory and IO Address Maps
Figure 3-6 on page 122 depicts generic system memory and IO maps. Note that the size of the system memory map is a function of the range of addresses that devices are capable of generating (often dictated by the CPU address bus). As in PCI and PCI-X, PCI Express permits either 32 bit or 64 bit memory addressing. The size of the system IO map is limited to 32 bits (4GB), although in many systems only the lower 16 bits (64KB) are used.
Figure 3-6. Generic System Memory And IO Address Maps
Key TLP Header Fields in Address Routing
If the Type field in a received TLP indicates address routing is to be used, then the Address Fields in the header are used to performing the routing check. There are two cases: 32-bit addresses and 64-bit addresses.
TLPs with 3DW, 32-Bit Address
For IO or a 32-bit memory requests, only 32 bits of address are contained in the header. Devices targeted with these TLPs will reside below the 4GB memory or IO address boundary. Figure 3-7 on page 123 depicts this case.
Figure 3-7. 3DW TLP Header Address Routing Fields
TLPs With 4DW, 64-Bit Address
For 64-bit memory requests, 64 bits of address are contained in the header. Devices targeted with these TLPs will reside above the 4GB memory boundary. Figure 3-8 on page 124 shows this case.
Figure 3-8. 4DW TLP Header Address Routing Fields
An Endpoint Checks an Address-Routed TLP
If the Type field in a received TLP indicates address routing is to be used, then an endpoint device simply checks the address in the packet header against each of its implemented BARs in its Type 0 configuration space header. As it has only one link interface, it will either claim the packet or reject it. Figure 3-9 on page 125 illustrates this case.
Figure 3-9. Endpoint Checks Routing Of An Inbound TLP Using Address Routing
A Switch Receives an Address Routed TLP: Two Checks
If the Type field in a received TLP indicates address routing is to be used, then a switch first checks to see if it is the intended completer. It compares the header address against target addresses programmed in its two BARs. If the address falls within the range, it consumes the packet. This case is indicated by (1) in Figure 3-10 on page 126. If the header address field does not match a range programmed in a BAR, it then checks the Type 1 configuration space header for each downstream link. It checks the non-prefetchable memory (MMIO) and prefetchable Base/Limit registers if the transaction targets memory, or the I/O Base and Limt registers if the transaction targets I/O address space. This check is indicated by (2) in Figure 3-10 on page 126.
Figure 3-10. Switch Checks Routing Of An Inbound TLP Using Address Routing
Other Notes About Switch Address-Routing
The following notes also apply to switch address routing:
If the address-routed packet address falls in the range of one of its secondary bridge interface Base/Limit register sets, it will forward the packet downstream.
If the address-routed packet was moving downstream (was received on the primary interface) and it does not map to any BAR or downstream link Base/Limit registers, it will be handled as an unsupported request on the primary link.
Upstream address-routed packets are always forwarded to the upstream link if they do not target an internal location or another downstream link.
ID routing is based on the logical position (Bus Number, Device Number, Function Number) of a device function within the PCI bus topology. ID routing is compatible with routing methods used in the PCI and PCIX protocols when performing Type 0 or Type 1 configuration transactions. In PCI Express, it is also used for routing completions and may be used in message routing as well.
ID Bus Number, Device Number, Function Number Limits
PCI Express supports the same basic topology limits as PCI and PCI-X:
A maximum of 256 busses/links in a system. This includes busses created by bridges to other PCI-compatible protocols such as PCI, PCI-X, AGP, etc.
A maximum of 32 devices per bus/link. Of course, While a PCI(X) bus or the internal bus of a switch may host more than one downstream bridge interface, external PCI Express links are always point-to-point with only two devices per link. The downstream device on an external link is device 0.
A maximum of 8 internal functions per device.
A significant difference in PCI Express over PCI is the provision for extending the amount of configuration space per function from 256 bytes to 4KB. Refer to the “Configuration Overview” on page 711 for a detailed description of the compatible and extended areas of PCI Express configuration space.
Key TLP Header Fields in ID Routing
If the Type field in a received TLP indicates ID routing is to be used, then the ID fields in the header are used to perform the routing check. There are two cases: ID routing with a 3DW header and ID routing with a 4DW header.
3DW TLP, ID Routing
Figure 3-11 on page 128 illustrates a TLP using ID routing and the 3DW header.
Figure 3-11. 3DW TLP Header ID Routing Fields
4DW TLP, ID Routing
Figure 3-12 on page 129 illustrates a TLP using ID routing and the 4DW header.
Figure 3-12. 4DW TLP Header ID Routing Fields
An Endpoint Checks an ID-Routed TLP
If the Type field in a received TLP indicates ID routing is to be used, then an endpoint device simply checks the ID field in the packet header against its own Bus Number, Device Number, and Function Number(s). In PCI Express, each device “captures” (and remembers) its own Bus Number and Device Number contained in TLP header bytes 8-9 each time a configuration write (Type 0) is detected on its primary link. At reset, all bus and device numbers in the system revert to 0, so a device will not respond to transactions other than configuration cycles until at least one configuration write cycle (Type 0) has been performed. Note that the PCI Express protocol does not define a configuration space location where the device function is required to store the captured Bus Number and Device Number information, only that it must do it.
Once again, as it has only one link interface, an endpoint will either claim an ID-routed packet or reject it. Figure 3-11 on page 128 illustrates this case.
A Switch Receives an ID-Routed TLP: Two Checks
If the Type field in a received TLP indicates ID routing is to be used, then a switch first checks to see if it is the intended completer. It compares the header ID field against its own Bus Number, Device Number, and Function Number(s). This is indicated by (1) in Figure 3-13 on page 131. As in the case of an endpoint, a switch captures its own Bus Number and Device number each time a configuration write (Type 0) is detected on i's primary link interface. If the header ID agrees with the ID of the switch, it consumes the packet. If the ID field does not match i's own, it then checks the Secondary-Subordinate Bus Number registers in the configuration space for each downstream link. This check is indicated by (2) in Figure 3-13 on page 131.
Figure 3-13. Switch Checks Routing Of An Inbound TLP Using ID Routing
Other Notes About Switch ID Routing
If the ID-routed packet matches the range of one of its secondary bridge interface Secondary-Subordinate registers, it will forward the packet downstream.
If the ID-routed packet was moving downstream (was received on the primary interface) and it does not map to any downstream interface, it will be handled as an unsupported request on the primary link.
Upstream ID-routed packets are always forwarded to the upstream link if they do not target an internal location or another downstream link.
Implicit routing is based on the intrinsic knowledge PCI Express devices are required to have concerning upstream and downstream traffic and the existence of a single PCI Express Root Complex at the top of the PCI Express topology. This awareness allows limited routing of packets without the need to assign and include addresses with certain message packets. Because the Root Complex generally implements power management and interrupt controllers, as well as system error handling, it is either the source or recipient of most PCI Express messages.
Only Messages May Use Implicit Routing
With the elimination of many sideband signals in the PCI Express protocol, alternate methods are required to inform the host system when devices need service with respect to interrupts, errors, power management, etc. PCI Express addresses this by defining a number of special TLPs which may be used as virtual wires in conveying sideband events. Message groups currently defined include:
INTx legacy interrupt signaling
Locked Transaction support
Hot Plug signaling
Slot Power Limit messages
Messages May Also Use Address or ID Routing
In systems where all or some of this event traffic should target the system memory map or a logical location in the PCI bus topology, address routing and ID routing may be used in place of implicit routing. If address or ID routing is chosen for a message, then the routing mechanisms just described are applied in the same way as they would for other posted write packets.
Routing Sub-Field in Header Indicates Routing Method
As a message TLP moves between PCI Express devices, packet header fields indicate both that it is a message, and whether it should be routed using address, ID, or implicitly.
Key TLP Header Fields in Implicit Routing
If the Type field in a received message TLP indicates implicit routing is to be used, then the routing sub-field in the header is also used to determine the message destination when the routing check is performed. Figure 3-14 on page 133 illustrates a message TLP using implicit routing.
Figure 3-14. 4DW Message TLP Header Implicit Routing Fields
Message Type Field Summary
Table 3-7 on page 134 summarizes the use of the TLP header Type field when a message is being sent. As shown, the upper two bits of the 5 bit Type field indicate the packet is a message, and the lower three bits are the routing sub-field which specify the routing method to apply. Note that the 4DW header is always used with message TLPs, regardless of the routing option selected.
Table 3-7. Message Request Header Type Field Usage
Type Field Bits
Defines the type of transaction:
10b = Message Transaction
Message Routing Subfield R[2:0], used to select message routing:
An Endpoint Checks a TLP Routed Implicitly
If the Type field in a received message TLP indicates implicit routing is to be used, then an endpoint device simply checks that the routing sub-field is appropriate for it. For example, an endpoint may accept a broadcast message or a message which terminates at the receiver; it won't accept messages which implicitly target the Root Complex.
A Switch Receives a TLP Routed Implicitly
If the Type field in a received message TLP indicates implicit routing is to be used, then a switch device simply considers the ingress port it arrived on and whether the routing sub-field code is appropriate for it. Some examples:
The upstream link interface of a switch may legitimately receive a broadcast message routed implicitly from the Root Complex. If it does, it will forward it intact onto all downstream links. It should not see an implicitly routed broadcast message arrive on a downstream ingress port, and will handle this as a malformed TLP.
The switch may accept messages indicating implicit routing to the root complex on secondary links; it will forward all of these upstream because it “knows” the location of the Root Complex is on its primary side. It would not accept messages routed implicitly to the Root Complex if they arrived on the primary link receive interface.
If the implicitly-routed message arrives on either upstream or downstream ingress ports, the switch may consume the packet if routing indicates it should terminate at receiver.
If messages are routed using address or ID methods, a switch will simply perform normal address checks in deciding whether to accept or forward it.