- The Previous Chapter
- This Chapter
- The Next Chapter
- Introduction
- Two Types of Local Link Traffic
- Transaction Layer Packet Routing Basics
- Applying Routing Mechanisms
- Plug-And-Play Configuration of Routing Options
Plug-And-Play Configuration of Routing Options
PCI-compatible configuration space and PCI Express extended configuration space are covered in detail in the Part 6. For reference, the programming of three sets of configuration space registers related to routing is summarized here.
Routing Configuration Is PCI-Compatible
PCI Express supports the basic 256 byte PCI configuration space common to all compatible devices, including the Type 0 and Type 1 PCI configuration space header formats used by non-bridge and switch/bridge devices, respectively. Devices may implement basic PCI-equivalent functionality with no change to drivers or Operating System software.
Two Configuration Space Header Formats: Type 0, Type 1
PCI Express endpoint devices support a single PCI Express link and use the Type 0 (non-bridge) format header. Switch/bridge devices support multiple links, and implement a Type 1 format header for each link interface. Figure 3-15 on page 136 illustrates a PCI Express topology and the use of configuration space Type 0 and Type 1 header formats.
Figure 3-15. PCI Express Devices And Type 0 And Type 1 Header Use
Routing Registers Are Located in Configuration Header
As with PCI, registers associated with transaction routing are located in the first 64 bytes (16 DW) of configuration space (referred to in PCI Express as the PCI 2.3 compatible header area). The three sets of registers of principal interest are:
-
Base Address Registers (BARs) found in Type 0 and Type 1 headers.
-
Three sets of Base/Limit Register pairs supported in the Type 1 header of switch/bridge devices.
-
Three Bus Number Registers, also found in Type 1 headers of bridge/devices.
Figure 3-16 on page 137 illustrates the Type 0 and Type 1 PCI Express Configuration Space header formats. Key routing registers are indicated.
Figure 3-16. PCI Express Configuration Space Type 0 and Type 1 Headers
Base Address Registers (BARs): Type 0, 1 Headers
General
The first of the configuration space registers related to routing are the Base Address Registers (BARs) These are marked “<1” in Figure 3-16 on page 137, and are implemented by all devices which require system memory, IO, or memory mapped IO (MMIO) addresses allocated to them as targets. The location and use of BARs is compatible with PCI and PCI-X. As shown in Figure 3-16 on page 137, a Type 0 configuration space header has 6 BARs available for the device designer (at DW 4-9), while a Type 1 header has only two BARs (at DW 4-5).
After discovering device resource requirements, system software programs each BAR with start address for a range of addresses the device may respond to as a completer (target). Set up of BARs involves several things:
-
The device designer uses a BAR to hard-code a request for an allocation of one block of prefetchable or non-prefetchable memory, or of IO addresses in the system memory or IO map. A pair of adjacent BARs are concatenated if a 64-bit memory request is being made.
-
Hard-coded bits in the BAR include an indication of the request type, the size of the request, and whether the target device may be considered prefetchable (memory requests only).
During enumeration, all PCI-compatible devices are discovered and the BARs are examined by system software to decode the request. Once the system memory and IO maps are established, software programs upper bits in implemented BARs with the start address for the block allocated to the target.
BAR Setup Example One: 1MB, Prefetchable Memory Request
Figure 3-17 depicts the basic steps in setting up a BAR which is being used to track a 1 MB block of prefetchable addresses for a device residing in the system memory map. In the diagram, the BAR is shown at three points in the configuration process:
-
The uninitialized BAR in Figure 3-17 is as it looks after power-up or a reset. While the designer has tied lower bits to indicate the request type and size, there is no requirement about how the upper bits (which are read-write) must come up in a BAR, so these bits are indicated with XXXXX. System software will first write all 1's to the BAR to set all read-write bits = 1. Of course, the hard-coded lower bits are not affected by the configuration write.
-
The second view of the BAR shown in Figure 3-17 is as it looks after configuration software has performed the write of all 1's to it. The next step in configuration is a read of the BAR to check the request. Table 3-8 on page 140 summarizes the results of this configuration read.
-
The third view of the BAR shown in Figure 3-17 on page 139 is as it looks after configuration software has performed another configuration write (Type 0) to program the start address for the block. In this example, the device start address is 2GB, so bit 31 is written = 1 (231 = 2GB) and all other upper bits are written = 0's.
Figure 3-17. 32-Bit Prefetchable Memory BAR Set Up
At this point the configuration of the BAR is complete. Once software enables memory address decoding in the PCI command register, the device will claim memory transactions in the range 2GB to 2GB+1MB.
Table 3-8. Results Of Reading The BAR after Writing All “1s” To It
BAR Bits |
Meaning |
---|---|
0 |
Read back as a “0”, indicating a memory request |
2:1 |
Read back as 00b indicating the target only supports a 32 bit address decoder |
3 |
Read back as a “1”, indicating request is for prefetchable memory |
19:4 |
All read back as “0”, used to help indicate the size of the request (also see bit 20) |
31:20 |
All read back as “1” because software has not yet programmed the upper bits with a start address for the block. Note that because bit 20 was the first bit (above bit 3) to read back as written (=1); this indicates the memory request size is 1MB (220 = 1MB). |
BAR Setup Example Two: 64-Bit, 64MB Memory Request
Figure 3-18 on page 141 depicts the basic steps in setting up a pair of BARs being used to track a 64 MB block of prefetchable addresses for a device residing in the system memory map. In the diagram, the BARs are shown at three points in the configuration process:
-
The uninitialized BARs are as they look after power-up or a reset. The designer has hard-coded lower bits of the lower BAR to indicate the request type and size; the upper BAR bits are all read-write. System software will first write all 1's to both BARs to set all read-write bits = 1. Of course, the hard-coded bits in the lower BAR are unaffected by the configuration write.
-
The second view of the BARs in Figure 3-18 on page 141 shows them as they look after configuration software has performed the write of all 1's to both. The next step in configuration is a read of the BARs to check the request. Table 3-9 on page 142 summarizes the results of this configuration read.
-
The third view of the BAR pair Figure 3-18 on page 141 indicates conditions after configuration software has performed two configuration writes (Type 0) to program the two halves of the 64 bit start address for the block. In this example, the device start address is 16GB, so bit 1 of the Upper BAR (address bit 33 in the BAR pair) is written = 1 (233 = 16GB); all other read-write bits in both BARs are written = 0's.
Figure 3-18. 64-Bit Prefetchable Memory BAR Set Up
At this point the configuration of the BAR pair is complete. Once software enables memory address decoding in the PCI command register, the device will claim memory transactions in the range 16GB to 16GB+64MB.
Table 3-9. Results Of Reading The BAR Pair after Writing All “1s” To Both
BAR |
BAR Bits |
Meaning |
---|---|---|
Lower |
0 |
Read back as a “0”, indicating a memory request |
Lower |
2:1 |
Read back as 10 b indicating the target supports a 64 bit address decoder, and that the first BAR is concatenated with the next |
Lower |
3 |
Read back as a “1”, indicating request is for prefetchable memory |
Lower |
25:4 |
All read back as “0”, used to help indicate the size of the request (also see bit 26) |
Lower |
31:26 |
All read back as “1” because software has not yet programmed the upper bits with a start address for the block. Note that because bit 26 was the first bit (above bit 3) to read back as written (=1); this indicates the memory request size is 64MB (226 = 64MB). |
Upper |
31:0 |
All read back as “1”. These bits will be used as the upper 32 bits of the 64-bit start address programmed by system software. |
BAR Setup Example Three: 256-Byte IO Request
Figure 3-19 on page 143 depicts the basic steps in setting up a BAR which is being used to track a 256 byte block of IO addresses for a legacy PCI Express device residing in the system IO map. In the diagram, the BAR is shown at three points in the configuration process:
-
The uninitialized BAR in Figure 3-19 is as it looks after power-up or a reset. System software first writes all 1's to the BAR to set all read-write bits = 1. Of course, the hard-coded bits are unaffected by the configuration write.
-
The second view of the BAR shown in Figure 3-19 on page 143 is as it looks after configuration software has performed the write of all 1's to it. The next step in configuration is a read of the BAR to check the request. Table 3-10 on page 144 summarizes the results of this configuration read.
-
The third view of the BAR shown Figure 3-19 on page 143 is as it looks after configuration software has performed another configuration write (Type 0) to program the start address for the IO block. In this example, the device start address is 16KB, so bit 14 is written = 1 (214 = 16KB); all other upper bits are written = 0's.
Figure 3-19. IO BAR Set Up
At this point the configuration of the IO BAR is complete. Once software enables IO address decoding in the PCI command register, the device will claim IO transactions in the range 16KB to 16KB+256.
Table 3-10. Results Of Reading The IO BAR after Writing All “1s” To It
BAR Bits |
Meaning |
---|---|
0 |
Read back as a “1”, indicating an IO request |
1 |
Reserved. Tied low and read back as “0”. |
7:2 |
All read back as “0”, used to help indicate the size of the request (also see bit 8) |
31:8 |
All read back as “1” because software has not yet programmed the upper bits with a start address for the block. Note that because bit 8 was the first bit (above bit 1) to read back as written (=1); this indicates the IO request size is 256 bytes (28 = 256). |
Base/Limit Registers, Type 1 Header Only
General
The second set of configuration registers related to routing are also found in Type 1 configuration headers and used when forwarding address-routed TLPs. Marked “<2” in Figure 3-16 on page 137, these are the three sets of Base/Limit registers programmed in each bridge interface to enable a switch/bridge to claim and forward address-routed TLPs to a secondary bus. Three sets of Base/Limit Registers are needed because transactions are handled differently (e.g. prefetching, write-posting, etc.) in the prefetchable memory, non-prefetchable memory (MMIO), and IO address domains. The Base Register in each pair establishes the start address for the community of downstream devices and the Limit Register defines the upper address for that group of devices. The three sets of Base/Limit Registers include:
-
Prefetchable Memory Base and Limit Registers
-
Non-Prefetchable Memory Base and Limit Register
-
I/O Base and Limit Registers
Prefetchable Memory Base/Limit Registers
The Prefetchable Memory Base/Limit registers are located at DW 9 and Prefetchable Memory Base/Limit Upper registers at DW 10-11 within the header 1. These registers track all downstream prefetchable memory devices. Either 32 bit or 64 bit addressing can be supported by these registers. If the Upper Registers are not implemented, only 32 bits of memory addressing is available, and the TLP headers mapping to this space will be the 3DW format. If the Upper registers and system software maps the device above the 4GB boundary, TLPs accessing the device will carry the 4DW header format. In the example shown in Figure 3-20 on page 145, a 6GB prefetchable address range is being set up for the secondary link of a switch.
Figure 3-20. 6GB, 64-Bit Prefetchable Memory Base/Limit Register Set Up
Register programming in the example shown in Figure 3-20 on page 145 is summarized in Table 3-11.
Table 3-11. 6 GB, 64-Bit Prefetchable Base/Limit Register Setup
Register |
Value |
Use |
---|---|---|
Prefetchable Memory Base |
8001h |
Upper 3 nibbles (800h) are used to provide most significant 3 digits of the 32-bit Base Address for Prefetchable Memory behind this switch. The lower 5 digits of the address are assumed to be 00000h. The least significant nibble of this register value (1h) indicates that a 64 bit address decoder is supported and that the Upper Base/Limit Registers are also used. |
Prefetchable Memory Limit |
FFF1h |
Upper 3 nibbles (FFFh) are used to provide most significant 3 digits of the 32-bit Limit Address for Prefetchable Memory behind this switch. The lower 5 digits of the address are assumed to be FFFFFh. The least significant nibble of this register value (1h) indicates that a 64 bit address decoder is supported and that the Upper Base/Limit Registers are also used. |
Prefetchable Memory Base Upper 32 Bits |
00000001h |
Upper 32 bits of the 64-bit Base address for Prefetchable Memory behind this switch. |
Prefetchable Memory Limit Upper 32 Bits |
00000002h |
Upper 32 bits of the 64-bit Limit address for Prefetchable Memory behind this switch. |
Non-Prefetchable Memory Base/Limit Registers
Non-Prefetchable Memory Base/Limit (at DW 8). These registers are used to track all downstream non-prefetchable memory (memory mapped IO) devices. Non-prefetchable memory devices are limited to 32 bit addressing; TLPs targeting them always use the 3DW header format.
Register programming in the example shown in Figure 3-21 on page 147 is summarized in Table 3-12.
Figure 3-21. 2MB, 32-Bit Non-Prefetchable Base/Limit Register Set Up
Table 3-12. 2MB, 32-Bit Non-Prefetchable Base/Limit Register Setup
Register |
Value |
Use |
---|---|---|
Memory Base (Non-Prefetchable) |
1210h |
Upper 3 nibbles (121h) are used to provide most significant 3 digits of the 32-bit Base Address for Non-Prefetchable Memory behind this switch. The lower 5 digits of the address are assumed to be 00000h. The least significant nibble of this register value (0h) is reserved and should be set = 0. |
Memory Limit (Non-Prefetchable) |
1220h |
Upper 3 nibbles (122h) are used to provide most significant 3 digits of the 32-bit Limit Address for Prefetchable Memory behind this switch. The lower 5 digits of the address are assumed to be FFFFFh. The least significant nibble of this register value (0h) is reserved and should be set = 0. |
IO Base/Limit Registers
IO Base/Limit (at DW 7) and IO Base/Limit Upper registers (at DW 12). These registers are used to track all downstream IO target devices. If the Upper Registers are used, then IO address space may be extended to a full 32 bits (4GB). If they are not implemented, then IO address space is limited to 16 bits (64KB). In either case, TLPs targeting these IO devices always carry the 3DW header format.
Register programming in the example shown in Figure 3-22 on page 149 is summarized in Table 3-13 on page 150.
Figure 3-22. IO Base/Limit Register Set Up
Table 3-13. 256 Byte IO Base/Limit Register Setup
Register |
Value |
Use |
---|---|---|
IO Base |
21h |
Upper nibble (2h) specifies the most significant hex digit of the 32 bit IO Base address (the lower digits are 000h) The lower nibble (1h) indicates that the device supports 32 bit IO behind the bridge interface. This also means the device implements the Upper IO Base/Limit register set, and those registers will be concatenated with Base/Limit. |
IO Limit |
41h |
Upper nibble (4h) specifies the most significant hex digit of the 32 bit IO Limit address (the lower digits are FFFh). The lower nibble (1h) indicates that the device supports 32 bit IO behind the bridge interface. This also means the device implements the Upper IO Base/Limit register set, and those registers will be concatenated with Base/Limit. |
IO Base Upper 16 Bits |
0000h |
Upper 16 bits of the 32-bit Base address for IO behind this switch. |
IO Limit Upper 16 Bits |
0000h |
Upper 16 bits of the 32-bit Limit address for IO behind this switch. |
Bus Number Registers, Type 1 Header Only
The third set of configuration registers related to routing are used when forwarding ID-routed TLPs, including configuration cycles and completions and optionally messages. These are marked “<3” in Figure 3-16 on page 137. As in PCI, a switch/bridge interface requires three registers: Primary Bus Number, Secondary Bus Number, and Subordinate bus number. The function of these registers is summarized here.
Primary Bus Number
The Primary Bus Number register contains the bus (link) number to which the upstream side of a bridge (switch) is connected. In PCI Express, the primary bus is the one in the direction of the Root Complex and host processor.
Secondary Bus Number
The Secondary Bus Number register contains the bus (link) number to which the downstream side of a bridge (switch) is connected.
Subordinate Bus Number
The Subordinate Bus Number register contains the highest bus (link) number on the downstream side of a bridge (switch). The Subordinate and Secondary Bus Number registers will contain the same value unless there is another bridge (switch) on the secondary side.
A Switch Is a Two-Level Bridge Structure
Because PCI does not natively support bridges with multiple downstream ports, PCI Express switch devices appear logically as two-level PCI bridge structures, consisting of a single bridge to the primary link and an internal PCI bus which hosts one or more virtual bridges to secondary interfaces. Each bridge interface has an independent Type 1 format configuration header with its own sets of Base/Limit Registers and Bus Number Registers. Figure 3-23 on page 152 illustrates the bus numbering associated with the external links and internal bus of a switch. Note that the secondary bus on the primary link interface is the internal virtual bus, and that the primary interface of all downstream link interfaces connect to the internal bus logically.
Figure 3-23. Bus Number Registers In A Switch