Tag Management for Non-Posted Transactions - 1.0 English

The requester side of the core maintains the state of all pending Non-Posted transactions (memory reads, I/O reads and writes, configuration reads and writes, Atomic Operations) initiated by the user application, so that the completions returned by the targets can be matched against the corresponding requests. The state of each outstanding transaction is held in a Split Completion Table in the requester side of the interface, which has a capacity of up to 768 Non-Posted transactions. The returning Completions are matched with the pending requests using an 5-/8-/10-bit tag. There are two options for management of these tags:

Internal Tag Management

Internal Tag Management is enabled when the Enable Client Tag option is not set in the AMD Vivado™ IDE, which is the default setting for the core. In this mode, logic within the core is responsible for allocating the tag for each Non-Posted request initiated from the requester side. The core maintains a list of free tags and assigns one of them to each request when the user logic initiates a Non-Posted transaction, and communicates the assigned tag value to the user logic through the output pcie_rq_tag0[9:0] and pcie_rq_tag1[9:0] . The value on this bus is valid when the core asserts pcie_rq_tag_vld0 and pcie_rq_tag_vld1. Use of pcie_rq_tag_vld0 and pcie_rq_tag_vld1 are orthogonal to whether the Straddle option is enabled. The integrated block can use either the pcie_rq_tag_vld0 or pcie_rq_tag_vld1 port to showcase the valid tags. The user logic must copy the tag so that any Completions delivered by the core in response to the request can be matched to the request.

In this mode, logic within the core checks for the Split Completion Table full condition, and backpressures a Non-Posted request from the user logic (using s_axis_rq_tready) if the total number of Non-Posted requests currently outstanding has reached its limit.

External Tag Management

External Tag Management is enabled when the Enable Client Tag option is set in the AMD Vivado™ IDE. In this mode, the user logic is responsible for allocating the tag for each Non-Posted request initiated from the requester side. The user logic must choose the tag value without conflicting with the tags of all other Non-Posted transactions outstanding at that time, and must communicate this chosen tag value to the core within the request descriptor. The core still maintains the outstanding requests in its Split Completion Table and matches the incoming Completions to the requests, but does not perform any checks for the uniqueness of the tags, or for the Split Completion Table full condition.

When internal tag management is in use, the core asserts pcie_rq_tag_vld for one cycle for each Non-Posted request, after it has placed its allocated tag on pcie_rq_tag. When straddle option is enabled, the core are provide up to two allocated tags in the same cycle on this interface. The states of the signals pcie_rq_tag_vld and pcie_rq_tag must be interpreted as follows:

• Assertion of pcie_rq_tag_vld0 in any cycle indicates that the core has placed an allocated tag on pcie_rq_tag0[9:0].
• Simultaneous assertion of pcie_rq_tag_vld0 and pcie_rq_tag_vld1 in the same cycle indicates that the core has placed two allocated tags, the first on pcie_rq_tag0[9:0] and the second on pcie_rq_tag1[9:0]. The tag on pcie_rq_tag0[9:0] corresponds to an earlier request sent by the user logic and the tag on pcie_rq_tag1[9:0] corresponds to a later request.
• pcie_rq_tag_vld1 is never asserted when pcie_rq_tag_vld0 is not asserted. That is, when there is only one tag to communicate in any cycle, it is always communicated on pcie_rq_tag0[9:0].
• When straddle is not in use, only a single tag can be communicated in any cycle, and pcie_rq_tag_vld1 is never asserted.

There can be a delay of several cycles between the transfer of the request on the s_axis_rq_tdata bus and the assertion of pcie_rq_tag_vld by the core to provide the allocated tag for the request. The user logic are, meanwhile, continue to send new requests. The tags for requests are communicated on the pcie_rq_tag bus in FIFO order, so it is easy for the user logic to associate the tag value with the request it transferred.