The present invention is related to a synchronous bus system and method.
Buses are frequently used to transmit data between devices. Generally two types of buses are used, synchronous and asynchronous. In a synchronous system, the devices coupled to the bus operate synchronous to one another. Furthermore, the timing budget for data transmission, that is the time from outputting the data from the transmitting device to the time that the receiving device samples the data, is one clock cycle. As the complexity of computer systems has increased, it has become increasingly difficult to physically connect the devices close enough such that the time of flight across the connection plus the set up and hold time of the receiving device do not exceed the timing budget.
In an asynchronous system it is not necessary that the clocks of the receiving and sending devices are synchronous to one another. However, the receiving device has to include logic to wait a number of clock cycles before reading out the captured data and sampling the captured data in order to ensure that the data is stable.
The system and method of the present invention provides for a synchronous bus system. The sending device sends out data and a strobe during the first clock cycle. The receiving device latches the data based upon receipt of the strobe. The receiving device further samples the data latched a predetermined plurality of clock cycles after the first clock cycle. Thus, unlike prior art synchronous bus systems, the bus is not required to send, launch and sample the data within one clock cycle. An innovative reset process synchronizes both devices in such a manner that simple logic is implemented to latch and sample the received data. Thus, unlike asynchronous bus systems, the overhead required to ensure that the received launched data is stable before sampling is avoided. By expanding the time budget to multiple clock periods, clock skew sensitivity is reduced as the skew is distributed over the multiple clock period.
The objects, features, and advantages of the present invention will be apparent to one skilled in the art in view of the following detailed description in which:
a, 2b, 2c, 2d and 2e illustrate the signaling topology of one embodiment of the bus system of the present invention.
a, 9b, and 9c are flow charts illustrating one embodiment of process for transmitting data in accordance with the teachings of the present invention.
a, 10b, 10c, and 10d illustrate request and completion formats used in one embodiment of the system of the present invention.
An exemplary system which incorporates the teachings of the present invention is shown in FIG. 1. It is readily apparent that the present invention is applicable to a variety of systems and system configurations.
The signaling topology of the bus system of the present invention is illustrated in
In one embodiment, the bus is a 16 bit wide data bus, which carries commands, addresses, data and transaction ID information. In addition, two additional bits carry mask and other information for the data fields. In one embodiment, the function of the two additional bits varies according to the clock cycle. For example, the fields provide byte enables (mask information) identifying the bytes consisting of valid information and may alternately carry a command type or parity.
The bus is bi-directional between the sending and receiving devices. In the present embodiment, the bus transactions are full split transactions and consist of a request packet and a completion packet. The request packet initiates a transaction. Completion packets are used to return data, indicate that a transaction has completed on the destination device, and reallocate buffer sources between the source and destination device. All transactions can be classified as a read request or a write request. Read requests contain command address bit enables for non-fetchable reads, routing information and length for desired data. The read completion packet contains the status of the request, the data retrieved in response to the read request, and routing and transaction information to identify the corresponding request. A write request includes the write data in its packet. The write completion contains no data but indicates if the write is completed successfully. Each bus cycle (XCLK) is equivalent to the system host clock cycle. However, each bus cycle contains a “P” half cycle and “N” half cycle. The “P” half cycle occurs for example while XCLK clock is high. The “N” half cycle occurs while the XCLK clock is low thus the throughput is doubled by transmitting packets on each half cycle.
A packet of information consists of multiple 32 bit words. One word associated byte enables are sent over the bus each XCLK cycle. Each word is distributed between the positive and negative phase of the bus clock cycle with bits [31:16] set on the positive phase and bits [15:0] set on the negative phase. It is readily apparent that the bus is not limited to this packet structure and a variety of implementations may be used.
One key aspect of the high speed synchronous bus of the present invention is that the reset signal (XRST#) enables the synchronization of all devices connected to the bus. Once synchronized, the transmitting and receiving devices operate synchronously in accordance with prespecified timing protocols to synchronously transmit packets between devices over multiple clock cycles.
As illustrated in
An illustrative timing diagram for the reset process for a 2 clock cycle timing budget is shown in FIG. 3. Each device connected to the bus sees the XRST# deasserted on the same generating XCLK clock signal. Each component starts its synchronous strobes signal running a predetermined number of clock cycles (e.g. three clock cycles) after observing an XRST# deassert. Although a three clock cycle is specified in the present embodiment, the number of predetermined cycles can vary so long as all devices start their synchronous strobe signal on the same cycle. With reference to
The system and timing relationships can be defined in a variety of ways. However, in the present embodiment the rising clock edge that samples XRST# deassertion is referred to the odd cycle and the first data strobe is started from an even clock edge. The earliest even clock edge that starts the strobe signals is the second even clock edge after the XRST# deassertion is sampled. In the present embodiment which implements a two clock cycle timing budget, the sampling, for reception of data, always selects the capture element (e.g. flip-flop) that contains data that was launched two clock cycles earlier. For example, in a three clock cycle mode, the selection would select that which was launched three clock cycles earlier. The multiplexor identifies the odd clock when XRST# deasserts. Since it is defined that the first strobe is always sent on an even clock, the capture flops and sampling multiplexors remain synchronized.
As described earlier, the distance between devices is longer than typical synchronous bus systems as the timing budget has been expanded to span multiple clock cycles. Furthermore, greater data throughput using fewer pins is achieved in part by launching data on both the even and odd numbered clock cycles. The capture mechanism at the receiver, which enables this capability as well as expansion of the timing budget, is shown in FIG. 4. Data is received via one of two capture flops 405 or 410. The flop enable is controlled by a third flop 415, which causes the enabled flop to toggle between capture flops 405 and 410, as driven by the positive data strobe signal (P_STB#). Thus, data that is launched on an even clock is captured by the even capture flop 410. Data that is captured on an odd clock is always captured by the odd capture flop 405. The present circuit, illustrated in
Referring again to
Once the data is processed through the sampling multiplexor, the data is input to combinatorial logic and into a sampling flip-flop 440. This is subsequently output into other circuitry of the device. It should be noted that the circuitry 430 shows a number of flip-flops which cause a delay sufficient to provide adequate initialization for valid sampling of data. The delay path synchronizes the sampling multiplexor 420 to the launched data. The delay can be varied according to the configuration implemented. Preferably, as shown in
A timing circuit showing the timing of exemplary packet transmissions is shown in FIG. 5. Referring to
At time T37 the sender device asserts HRTS# to indicate its request to send. At time T37, XRTS# (not shown) was not observed asserted, so the sending device knows that it has won arbitration of the bus. The sender asserts XADS# at time T38 to frame the packet information indicated as 1P, 1N, 2P, 2N.
At the receiving end, the receiver device observes (captures) HRTS# asserted at time T38. This is the time shifted HRTS# signal asserted at time T37. The receiver knows to expect XADS# during the next clock (T39). The present embodiment utilizes a distributed arbiter. Thus, if the sender in this example did not have high priority, XADS# would have been sent two clocks after HRTS# instead of one clock after HRTS#. Each device knows its priority. By convention, the high priority device will send its data one clock earlier than the low priority device (assuming the low priority device was not already requesting). Therefore, the low priority device must wait an additional clock when it asserts its request in order to guarantee the high priority device has observed the request. At clock T39, the receiver samples HRTS# from the capture FLOP that captured it. Data is then sampled starting at time T39 from the respective flops.
The processes for resetting the system to operate in a synchronous matter and transmission of data are illustrated by the simplified flow diagrams
At step 630, data transmission is initiated on a clock edge of an even clock cycle, which coincides with the issuance of the data strobes on the even clock cycle. Preferably, the system waits a predetermined number of clock cycles, such as 64 clock cycles, before initiating data transmission such that sufficient time is given for initialization of circuitry.
The transmission process will now be described with reference to FIG. 7. At step 700 the transmitting device simultaneously launches a strobe and data to the receiving device. At step 701, the strobe and data are received at the receiving device. At step 702, if the strobe was sent on an even clock the data is captured by the even flops; if the strobe was sent on an odd clock, the data is captured by the odd flops. At step 703, data is sampled at the receiver two clocks after launch from the sending device. Thus, data is sampled by the even flop if launched on even clock cycle and sampled by the odd flop if launched on an odd clock cycle. As mentioned above, once the circuitry in both devices are synchronized, the receiver circuitry simply toggles between even flops and odd flops. Thus, a process of operation for synchronous bus transmission across multiple clock cycles in which the sending and receiving devices receive clock signals at the same frequency is described.
Although not required for operation of the high speed synchronous system as described above, the effectiveness of the system is further enhanced using the embedded flow control method and apparatus described below.
In particular, bus overhead is decreased by distributing flow control to the devices coupled to the bus and embedding flow control data into the packets. Each device has at least one tracker device or circuit that tracks the flow of data and bus requests inbound and outbound onto the bus. At initialization, each tracker is provided information regarding the buffer capacities of the other coupled devices. During the process of transmission of packets, the tracker accesses predetermined bits of each packet to determine the states of the queues (i.e., how full/empty) and controls the flow of packets between devices. Thus flow control is embedded in the packet protocol.
In the present embodiment, flow control between two devices is described. However, it is contemplated that the structure can be expanded to support flow control between multiple pairs of devices by replication of trackers. A simplified block diagram of the flow control portion of the system is illustrated in FIG. 8. Referring to
The memory controller 805 includes request queue tracker logic 822, data queue tracker logic 832, outbound request queue 824, outbound data buffer 826, inbound request queue 828 and inbound data queue 830. Also shown is interface/control logic 834 which provides supporting logic for interfacing with the memory 802 and processor 803, performing the memory operations with memory 802 and processor 803, as well as providing the request packets and confirmation packets that are described below.
For purposes of simplification of explanation, the data communicated between the memory 802, processor 803 and the memory controller 805 is shown to be transmitted through the interface/control logic 834; however, it is contemplated that data may be transmitted directly between the queues and the memory 802 and processor 803. The request queue tracker logic 822 and data queue tracker logic 832 respectively track how full the respective queues 824, 852 and 826, 856 are, such that once queue is full, the tracker prevents a packet from being generated and placed in the queues 824, 826.
In the present embodiment, the tracker 822, 832 functions as a counter to maintain counts of available queue space. The interface/control logic 834 operates in conjunction with the tracker 822, 832 to issue the corresponding control signals/data to processor 803 and memory 802 to permit/prevent outbound packet generation and placement in the corresponding queues. Inbound request queue 828 and inbound data queue 830 respectively receive inbound requests and confirmation packets (and associated data) from the bus bridge 810. In one embodiment, the write data and read data is separately queued and tracked. In one embodiment, the request queue maintains both read and write requests, but the tracker permits only a predetermined maximum number of read requests and a predetermined number of write requests regardless of the number of available slots in the queue.
In one embodiment, the tracker logic 822 is configured to permit only two read requests and six write requests in an eight deep queue. This is desirable so that the one type of request, e.g., write request, does not prevent the queuing of read requests when the number of requests exceeds the size of a queue. Thus in the current example, if six write requests are currently queued and the device wishes to queue a seventh write request, the tracker will not permit it even though the queue has the capacity to receive two more requests. (those that are preallocated per read requests). If the queue currently has six write requests and the device wishes to issue a read request, the tracker will permit the read request to be queued.
The bus bridge 810 is similarly configured with a request queue tracker 850, data queue tracker 860, outbound request queue 852, inbound request queue 854, outbound data queue 856, inbound data queue 858 and interface/control logic 882. The queue tracking functions are performed similar to that described above. Trackers 850, 860 maintain counts of information stored in the queues 854, 828, and 858, 830, respectively, and prevent the generation of packets when one of the queues is full. Interface/control logic 882 not described in detail herein represents the logic used to communicate with the bus 820 and generate the request and confirmation packets as described below.
a and 9b are simplified flow charts respectively illustrating the flow control process for requests and data. Although the two processes are described separately and flow control can be initiated using either one or both processes, it is preferred that both processes are used concurrently to control flow control as shown in
Turning back to the present embodiment, if an inbound PCI (write) request, for example, is attempted from bus 820, the request will be retried until the inbound tracker 850 indicates that the inbound queue in device 805 has room for the write request. The same occurs for outbound transactions. If an inbound request queue were to accept a transaction for which there is no room in the receiving inbound queue, a deadlock can occur even though no packet is sent, until there is room in the receiving queue.
Referring to
At step 905, if a completion packet is received, the request tracker decrements the request buffer count, step 910, as receipt of a completion packet is indicative that the request has been processed and is no longer in the buffer. At step 915, if a request packet is to be sent, at step 920, the request buffer count is incremented and it is determined whether the count exceeds the predetermined maximum, step 925. If the count does not exceed the predetermined maximum, then the receiving buffer in the device has the capacity to receive the request and the request packet is prepared for transmission and subsequently sent out over the bus, step 940. If the count exceeds the predetermined maximum, then the available capacity of the buffer cannot accept the request packet and the request packet tracker prevents the request packet from being created or enqueued and causes the transmission process at the initiating bus to be retried, step 935.
It should be noted that
A very similar process is performed to control flow control with respect to data contained in the packet. A request packet is a determined size which fits in a predetermined amount of space. However, the amount of data is variable. Thus for data buffers, a length field in the packet is accessed to determine the amount of buffer space needed. A similar process is then performed to determine when data to be queued would cause the capacity of the data queue to be exceeded. The tracker will not allow the capacity of the data buffer to be exceeded. For example, if a device on the bus 820 wants to write 16 DWORDS (16×4 bytes), but the tracker indicates only room for 8, the control logic 882 will only accept eight DWORDS. The device (not shown) on the bus 820 must retry a write for the remaining DWORDS until the tracker indicates room for them. Alternately, control logic 882 will be configured such that the logic will not allow the generation of packets unless all data can be placed in the queue.
Referring to
As noted earlier, it is preferable that the flow control process takes into account available request buffer space and available data buffer space. If either buffer is full and cannot receive data, the request is not processed. This is illustrated by the flow chart of
Thus the flow control is embedded into the packet protocol. Illustrative packets are shown in
For example, when a read request is pushed into the memory controller's outbound transaction queue, TP[1:0] is 00 to indicate a request with no data and RCOM[4:0] is 0 to indicate that the request is to use a read queue slot. The packet is formed and placed in the queue and the outbound read queue tracker therefore is decremented by one. When the completion packet corresponding to the read request is sent back by the PXB, TP[1:0] is [1:x], where x is 1 if the data returned and 0 if no data was return. CCOM[4:0] is 0 to indicate this is a completion for a read request. The outbound read queue tracker therefore increments the count by one. It follows that when a read completion is popped from the memory controller inbound transaction queue, the outbound read queue tracker is incremented by one. Similar operations occur with respect to the bus bridge.
When a write is to be performed, the request is pushed into the device's outbound transaction queue. TP[1:0] is 01 to indicate a request with data and RCOM[4:0] is 1 to indicate the request is using a write queue slot. The output write request queue tracker is incremented by 1. When the completion for a write request is sent back, TP[1:0] is 10 to indicate a completion with no data. CCOM[4:0] is 1 to indicate a completion for a write request. When a write completion is popped from the device's inbound transaction queue, the outbound write queue tracker is incremented by 1. As noted above, when a transaction queue tracker decrements to zero, transactions of that type can no longer be pushed into the transaction queue Preferably, the requesting device will retry any additional actions of this type.
In the present embodiment, data buffer management is handled a little differently; however, it is contemplated that data buffer management can be handled the same way as requests. The TP[1:0], RCOM[4:0] and LEN[7:0] fields in the request packet header are used to allocate data buffers by the data buffer trackers. The TP[1:0], CCOM[4:0] and LEN[7:0] fields in the completion packet header are used to deallocate data buffers by the data buffer trackers.
For example, when a read is pushed into the memory controller outbound transaction queue, e.g. by the processor, TP[1:0] is 00 to indicate a request with no data and RCOM[0] is 0 to indicate the request is using a read queue slot. The outbound read data buffer tracker is decremented by LEN where LEN indicates data size, in the present embodiment, the number of DWORDS being requested.
When the completion packet for the read is sent back by the bus bridge, TP[1:0] is [1:x] where x is 1 if data is returned and 0 if no data was returned. CCOM[4:0] is 0 to indicate that the packet is a completion packet for a read. When a read completion is popped from the memory controller inbound transaction queue, the outbound read data buffer is incremented by LEN.
When a write packet is pushed into the memory controller outbound transaction queue, e.g. by the coupled processor, TP[1:0] is 01 to indicate a request with data and RCOM[4:0] is 1 to indicate the request is using a write queue slot. The outbound write data buffer tracker is decremented by LEN where LEN indicates the number of DWORDS being written. The value in the LEN field of the write request packet and the associated completion packet are always equal even if the write was not successful at the other bus.
When the completion packet for the write is sent back by the PXB, TP[1:0] is 10 to indicate a completion with no data. CCOM[0] is 1 to indicate that the packet is a completion packet for a write request. When the write completion is received by the outbound write data buffer tracker, the count is incremented by LEN. Normally, requests and completions leave a transaction queue in the same order as entered. This is necessary to preserve proper transaction ordering, i.e., the order of occurrence on one bus is the same as the order on the receiving bus. However, a write completion contains no data, hence, no ordering requirement. Therefore, it is preferred that the completion packet is sent directly to the tracker.
When a data buffer tracker decrements to zero or has insufficient data buffers for a particular request, that request cannot be pushed into the transaction queue. The data buffer tracker's bus interface will therefore retry any additional transactions of that type. Similar logic is used to support write packets issued by the bus bridge.
A simplified example of the embedded flow control process is illustrated below. For purposes of discussion, the example is simplified and does not take into account other configuration parameters such as those related to prefetching. In addition, the below example and the discussion that follows discusses the flow control mechanism in the context of a device, such as a memory controller, coupled through the high speed bus of the present invention to a PCI bus bridge expander that transfers the data to 2 32 bit PCI busses or 1 64 bit PCI bus.
Certain transactions demand a fixed number of DWORDS to transfer. For example, a line write command (PCI MWI) must transfer a full line. If a line consists of 8 DWORDS and less than 8 DWORDS of buffering is available, the transaction must be retried. A normal write burst, however, could result in a portion of the number of DWORDS being accepted and the remainder being retried. For example, Memory Read Line (MRL) transaction would be retried unless buffer space corresponding to a full line of DWORDS is available.
As noted above, the bus bridge is preferably configured to route packets for dual 32 bit operating modes and single 64 bit operating modes. In dual 32 bit mode the ‘a’ and ‘b’ transaction queues operate independently on their respective buses. The only interaction occurs at the high speed bus interface where one or the other set of queues send or receive on the high speed bus between the bus bridge and the memory controller.
In single 64 bit mode the outbound transaction queues are paired up to appear as a single outbound queue and the inbound transaction queues are paired up to appear as a single inbound transaction queue. Effectively, the 64 bit PCI bus interface has twice the queue depth of each of the dual 32 bit PCI interfaces. Thus, queue tracking is configurable to track a pair of inbound/outbound queues as well as a single set of queues.
The outbound transaction queues are treated in a similar manner to the inbound transaction queues. If an outbound transaction from the high speed bus interface enters the ‘a’ outbound queue (OutQa), the next outbound transaction will enter the ‘b’ outbound queue (OutQb) and so forth. At the bus bridge interface, logic (e.g., a state machine) toggles between OutQa and OutQb. Starting at OutQa, the first outbound transaction is attempted on the bus coupled to the bus bridge (e.g., a PCI bus). If the transaction completes, it is popped from OutQa and the completion packet is pushed into whichever inbound queue the queue pointer currently is pointing. Next, the transaction at the top of OutQb is attempted. If every outbound transaction completes on first attempt, the outbound queue pointer keeps toggling with each completed transaction.
If a read transaction at the top of the outbound queue is retried, it is moved into the corresponding read request queue RRQ (a or b) and the outbound queue pointer toggles to the other queue. If a write transaction at the top of the outbound queue is retried, it is preferred that the queue pointer does not toggle. A retried write must succeed before the outbound queue pointer will toggle to the opposite queue. However, between attempts to complete the write at the top of the current queue, any reads in either RRQ may also be attempted. Once the current outbound write succeeds it is popped from the queue and a completion packet is inserted into the current inbound queue. The outbound queue pointer will then toggle to the opposite queue even if an uncompleted read remains in the RRQ.
In summary, the outbound queue pointer toggles to the opposite queue as soon as a transaction is popped from the current queue. A retried write is not popped until it succeeds. A retried read is popped from the outbound queue and pushed into the RRQ. A read in a RRQ can be attempted at any time because its ordering requirements were met at the time it was popped from the outbound queue. (Note that outbound reads in one RRQ can pass outbound reads in the other RRQ in a 64 bit PCI mode.)
In 32 bit mode, an outbound transaction is routed from the high speed bus to either outbound queue ‘a’ or ‘b’ depending upon the packet's destination identification (Destination ID). Multiplexors select the next outbound request or a previously retired read as discussed in the previous section. Preferably a separate multiplexor is used for 64 bit PCI mode. When the bus bridge initiates a PCI transaction in 64 bit mode, a multiplexor selects the command and address bits from either outbound queue ‘a’ or outbound queue ‘b’.
Inbound transactions can address more than 32 bits so both inbound queues support dual address cycle (DAC) decode in 32 bit mode and 64 bit mode. The inbound request queues have separate latch enables for upper and lower 32 bits of address. In 32 bit mode, the low order address is latched in address latch ‘a’ or address latch ‘b’ for PCI bus ‘a’ or ‘b’ respectively. The inbound request queue latches the low order address prior to the next PCI clock in preparation for the arrival of the high order address of a DAC. If the inbound transaction is a single address cycle transaction, zeros must be loaded into the high order address field of the inbound request queues.
In 64 bit mode, the inbound transaction can be initiated by either a 32 bit PCI master or 64 bit PCI master. DAC is required to be asserted on C/B[3:0] in packets by 32 bit and 64 bit PCI masters (e.g., memory controller) addressing above 4 GB because it is unknown to the master at this time if the target is 64 bit capable or not. A 64 bit PCI master is not required to drive the high order address bits to zero for addresses below 4 GB. If REQ64# is asserted with FRAME# and the PXB decodes DAC on C/B[3:0] during the first address cycle, it can immediately decode the full address. If C/B[3:0] does not indicate DAC, the PXB must force the high order address to all zeros before decoding the address.
As noted previously, it is preferred that the data buffers exist as separate structures from the transaction or request queues. The data for PCI transactions is stored in a separate queue structure from the transaction queues. This data queue structure is referred to as the data buffers or the data queues. Separate queues are needed for data because the transactions and completions in the transaction queues do not always get retired in the same order that they entered the transaction queues. For example, write transactions may pass read transactions in the same direction. Also, PCI delayed reads get retired in the order that the PCI masters return for their data which is not necessarily the same order that the read requests or read data were received.
In dual 32 bit PCI mode when an inbound PCI write transaction enters InQa, the data that follows the address and command on the PCI bus will enter the PW Data 1 inbound data queue. When the associated write packet is sent over the F16 bus, the packet header containing the write command and address will be pulled from the InQa transaction queue and the write data will be pulled from the PW Data 1/DRPLY Data 1 inbound data queue. Likewise, an inbound PCI write on PCI Bus ‘b’ pushes the command and address into InQb and the associated data that follows on the PCI bus is pushed into PW Data 2 inbound data queue.
In dual 32 bit PCI mode, an outbound 32 bit PCI read to PCI bus ‘a’ is pulled from OutQa or RRQa when the read succeeds on the PCI bus and a Read Completion is pushed into the InQa inbound transaction queue. The associated read data enters the PW Data 1/DRPLY Data 1 inbound data queue. When the Completion packet is sent over the F16 bus, the packet header containing the read completion identifier will be pulled from the top of the InQa transaction queue and the read data will be pulled from the PW Data 1/DRPLY Data 1 inbound data queue.
Each 32 bit PCI port can have two inbound PCI reads outstanding. An inbound PCI read on PCI port a is pushed into InQa if there is a slot available in the PXB inbound queue for a read and there are inbound read data buffers available in the PXB and MIOC. At this time the inbound delayed read completion tracker is loaded with the command and address fields of the inbound read so that it can identify the PCI master requesting the read. A transaction identifier unique to this inbound transaction is also loaded into the inbound delayed read completion tracker so that the read completion can be identified when it arrives in the OutQa. When the inbound read completes on the P6 bus, a delayed read completion (DRC) packet containing the read data will arrive to the bus bridge over the high speed bus. The DRC translation header containing the inbound read identifier will be pushed into OutQa. The read data that follows in the packet will be pushed into DRC Data 1 data queue or DRC 2 data queue depending upon which DRC data queue was assigned to this inbound read. When the PCI master returns for its data (it will be continuously retired until the data arrives) it will receive the data from DRC Data 1 or DRC Data 2 data queue if the associated inbound read completion has been popped from the top of the OutQa transaction queue and marked the inbound read as complete in the inbound delayed read completion tracker.
In 64 bit PCI mode, the two sets of data buffer queues are paired similar to the transaction queue in 64 bit PCI mode. An inbound write will result in data being alternately pushed into PW Data 1 and PW Data 2 data queues. The data queues are 32 bits wide (DWord). If data is received 64 bits at a time from a 64 bit PCI master and the data queue pointer is pointing at PW Data 1 queue, the first DWord is pushed into PW Data 1 data queue and the next DWord is pushed into PW Data 2 date queue. Additional DWORDS alternate between the two inbound data queues.
The DRC data queues and write data queues are paired and interleaved in a similar fashion.
The innovative packet format described above in addition to embedding flow control information, also provides at least one field referred to herein as the transaction identification (TID) field, that can be used in a variety of ways. The field is preferably configurable, depending upon the application. The advantage is that the sending device, i.e., the device issuing a request packet, can store predetermined data in this field, e.g., a transaction identifier or other identifier. The control logic of the receiving device, after processing the request and preparing the completion packet, simply copies the contents of the field into the completion packet for transmission back to the initial sending device. Thus, the configuration can be such that the field contents is meaningful only to the sending device as the receiving device simply copies the contents and sends it back. Furthermore, as the packet is not limited to specific data, the field can be used for a variety of purposes. Furthermore, as the receiving device simply copies the contents into the completion packet, the contents remain undisturbed.
This process is described generally with reference to FIG. 11. At step 1105, the sending device forms a request packet. The request packet includes the transaction ID field which is used to store requesting device data. At step 1110, the request packet is issued and at step 1115, the receiving device receives the packet and forms a reply packet, step 1120. The receiving device simply copies the TID field into the reply packet for subsequent access by the sending device. Thus, the contents of the TID are not required to be interpreted by the receiving device as a simple copy operation is all that is required. At step 1125, the reply packet, including the copied contents of the TID field, is sent back to the requesting device.
In the present embodiment, the field is used for a deferred outbound read (processor to PCI) transaction. A deferred transaction is a split transaction where the read is split into the initial read request followed at a later time by a deferred reply. The requested data is returned by the deferred reply. Thus, the device and transaction ID of the read requester is put into the TID field. When the completion packet with the read data is sent, the TID is copied from the request packet to the completion packet. When the completion reaches the top of the inbound request queue, a deferred reply is sent to the requesting processor. The deferred reply copies the completion TID into the deferred reply where it is used to address the processor that initiated the original read.
The invention has been described in conjunction with the preferred embodiment. It is evident that the numerous alternatives, modifications, variations, and uses will be apparent to those skilled in the art in light of the foregoing description.
This is a continuation of application Ser. No. 08/934,640, now U.S. Pat. No. 6,088,370, filed Sep. 22, 1997. This is also related to U.S. Pat. No. 6,266,778 B1, filed Aug. 25, 1999.
Number | Name | Date | Kind |
---|---|---|---|
3764920 | Galcik et al. | Oct 1973 | A |
4503499 | Mason | Mar 1985 | A |
4719621 | May | Jan 1988 | A |
4922486 | Lidinsky et al. | May 1990 | A |
4941089 | Fischer | Jul 1990 | A |
5101347 | Balakrishnan | Mar 1992 | A |
5193090 | Filipiak et al. | Mar 1993 | A |
5210858 | Jensen et al. | May 1993 | A |
5257258 | Birman | Oct 1993 | A |
5325492 | Bonevento et al. | Jun 1994 | A |
5361252 | Sallberg et al. | Nov 1994 | A |
5404171 | Golstein et al. | Apr 1995 | A |
5448708 | Ward | Sep 1995 | A |
5467464 | Oprescu | Nov 1995 | A |
5471587 | Fernando | Nov 1995 | A |
5491799 | Kreuzenztein et al. | Feb 1996 | A |
5499338 | Gercekci | Mar 1996 | A |
5541919 | Yong et al. | Jul 1996 | A |
5548733 | Sarangdhar | Aug 1996 | A |
5574862 | Marianetti, II | Nov 1996 | A |
5625779 | Solomon | Apr 1997 | A |
5634015 | Chang et al. | May 1997 | A |
5657457 | Gaskins | Aug 1997 | A |
5659718 | Osman | Aug 1997 | A |
5668971 | Neufeld | Sep 1997 | A |
5671441 | Glassen et al. | Sep 1997 | A |
5715438 | Silha | Feb 1998 | A |
5729760 | Poisner | Mar 1998 | A |
5751969 | Kapoor | May 1998 | A |
5758166 | Ajanovic | May 1998 | A |
5764961 | Bhat | Jun 1998 | A |
5768545 | Solomon | Jun 1998 | A |
5768546 | Kwon | Jun 1998 | A |
5768550 | Dean et al. | Jun 1998 | A |
5771356 | Leger et al. | Jun 1998 | A |
5784579 | Pawlowski | Jul 1998 | A |
5802278 | Isfeld et al. | Sep 1998 | A |
5862338 | Walker et al. | Jan 1999 | A |
5894567 | Dodd et al. | Apr 1999 | A |
5905766 | Nguyen | May 1999 | A |
6044474 | Klein | Mar 2000 | A |
6330650 | Toda et al. | Dec 2001 | B1 |
6347351 | Osborne et al. | Feb 2002 | B1 |
Number | Date | Country | |
---|---|---|---|
Parent | 08934640 | Sep 1997 | US |
Child | 09471445 | US |