DMA Controller Having a Plurality of DMA Channels Each Having Multiple Register Sets Storing Different Information Controlling Respective Data Transfer

This application is related to U.S. application Ser. No. 08/189,139, now abandoned, entitled "Dual Bus Concurrent Multi-Channel Direct Memory Access Controller and Method", and U.S. application Ser. No. 08/189,131, now abandoned, entitled "Direct Memory Access Channel Architecture and Method for Reception of Network Information", each of which is assigned to the assignee of the present application and filed concurrently herewith.

The present invention pertains to the field of data transfer in a digital computer system. More particularly, this invention relates to direct memory access in a digital computer.

In a computer system, direct memory access (DMA) is used to transfer data between input/output (I/O) devices and the main memory of the system. In some systems, DMA is also used to move data from one location in main memory to another. DMA is achieved through a DMA controller which manages the transfer, thereby alleviating the central processing unit (CPU) of the task. That is, the DMA controller is capable of receiving data from I/O devices and issuing the requisite write commands to store the data in main memory and is also capable of issuing the requisite read commands to read data from main memory and then transferring it to a particular I/O device. The DMA controller therefore allows the CPU to perform other tasks while the DMA controller is transferring the data.

One common way of implementing DMA is shown in FIG. 1A. A block diagram of a computer system is shown with a system bus 110 connected to a CPU 115, a DMA controller 120, a main memory (which may be random access memory--RAM) 125, a Read Only Memory (ROM) 127, and two I/O devices 130 and 131 through bus interface logic 132 and 133. The DMA controller 120 is also coupled to the I/O devices 130 and 131 over control lines 121a, 121b, 121c and 121d. System bus 110 is shared by the DMA controller 120 and the CPU 115. Therefore, CPU 115 may be forced to sit idle while DMA controller 120 is using system bus 110 to transfer data between main memory 125 and the I/O devices 130 and 131.

FIG. 1A illustrates a typical prior art fly-by DMA scheme. In a fly-by DMA scheme the DMA controller 120 manages the data transfer, however the data does not pass through the DMA controller. The DMA controller 120 issues the read and write commands necessary to transfer the data between the I/O device and main memory. The controller 120 does not actually "see" any of the data being transferred, the data passes by on system bus 110 on its way to main memory 125 or I/O device 130 or 131. In some prior art systems a buffer may be situated between the system bus 110 and each I/O device 130 and 131, or a buffer may be contained within main memory 125. This buffer may act as an intermediary between the system bus 110 and the particular I/O device (e.g. as a transition from 16-bit to 32-bit), or may temporarily store the data before it is written to its destination (either main memory 125 or an I/O device 130 or 131). This buffer, however, is not part of the DMA controller; the data is transferred to the buffer after being transferred over the system bus 110.

A typical prior art fly-by DMA controller contains a single register set, shown in FIG. 1B. The register 140 consists of a counter 142, an address register 144, and may have a control register 146. The counter 142 is programmed by the CPU 115 prior to a DMA transfer with the amount of data to be transferred, usually a number of bytes. The address register 144 is programmed by the CPU 115 prior to a DMA transfer with the beginning location for the transfer (either where the first data byte to be read from main memory is located or where the first byte is to be written in main memory). The control register 146 may be included and if so contains necessary control information pertaining to a transfer. The DMA transfer begins after counter 142 and address register 144 are programmed. As data is transferred by the DMA controller 120, the counter 142 is decremented so that it contains the amount of data remaining to be transferred and the address register 144 is updated to hold the memory address of the next byte to be read or written.

Another prior art implementation of a DMA controller is a linked list 150 shown in FIG. 1C. The linked list 150 is stored in main memory 125 and may contain multiple control structures 156; the linked list 150 is programmed by the CPU 115 with multiple counts 152 and multiple addresses 154. The linked list 150 allows CPU 115 to program multiple transfers, thereby allowing DMA controller 120 to make multiple DMA transfers without stopping to be reprogrammed by CPU 115. However, the use of a linked list 150 has significant drawbacks. Among the most significant of these drawbacks is the increased control and complexity required in the DMA controller. Due to the nature of a linked list, the location of the count 152 and address register 154 for the current transfer will change after each transfer. The DMA controller 120 therefore requires additional control logic, and therefore additional hardware complexity, to keep track of these constantly changing locations, and to access their values in main memory 125.

In view of limitations of known systems and methods, one of the objectives of the present invention is to provide a direct memory access (DMA) controller and method to reduce CPU interrupt latency restrictions.

Another objective of the present invention is to provide a DMA controller and method to efficiently support arbitrary-length spatially scattered transfers.

Another objective of the present invention is to reduce the hardware complexity of a DMA controller without significantly degrading system performance.

These and other objects of the invention are provided for by a multiple register set DMA controller channel architecture and method. The DMA controller is connected to the CPU bus of a computer system through a bus interface. The DMA controller is also connected to an I/O bus, which is coupled to one or more I/O devices (e.g., a hard disk drive or an ethernet network). Multiple channels, each corresponding to a particular I/O device, are contained within the DMA controller and are connected to both the bus interface and the I/O bus.

The DMA controller controls DMA transfers between the I/O devices and the main memory of the system. Each of the channels within the DMA controller contains multiple register sets and a first in first out (FIFO) buffer; each register set comprises a count register, an address register, and a command/status register. The count registers store the amount of data to be transferred for a particular DMA transfer, either from the I/O device to main memory or from main memory to the I/O device. The address registers store the starting location for the transfer in main memory; i.e., the address of either the first byte to be read from or written to memory. Each command/status register stores control information for its register set and the FIFO buffers store a portion of the data being transferred between the two buses.

Prior to a DMA transfer, the computer software programs the count and address registers of one register set with the requisite information for the transfer. If a second transfer is required, the software may also program a second register set at this time with the count and address information for the second transfer. Once a register set is programmed, and when the previously programmed transfer is completed, the register set will become the active register set and will attempt to begin its transfer. The DMA channel's transfer of data will depend on what other channels are requesting access to the bus this DMA channel is requesting access to, and the ability of the DMA channel's corresponding I/O device to accept or provide data.

The present invention is illustrated by way of example and not limitation in the figures of the accompanying drawings, in which like references indicate similar elements and in which:

FIG. 1A is a block diagram of a prior art implementation of direct memory access (DMA) in a computer system;

FIG. 1B is a diagram of a prior art single register implementation of a DMA controller;

FIG. 1C is a diagram of a prior art linked list data structure of a DMA controller;

FIG. 2A is a block diagram of a general computer system utilizing the DMA controller of the present invention;

FIG. 2B is a diagram of the DMA controller of the preferred embodiment of the present invention;

FIG. 3 is a diagram of an example single DMA channel within the direct memory access controller of the preferred embodiment of the present invention;

FIG. 4 is a diagram of the registers in the direct memory access controller of the preferred embodiment of the present invention;

FIG. 5 is a diagram of the CPU bus interface of the preferred embodiment of the present invention;

FIG. 6 is a diagram of the internal CPU bus arbiter of the preferred embodiment of the present invention;

FIG. 7A is a diagram of the I/O bus arbiter of the preferred embodiment of the present invention; and

FIG. 7B is a diagram of an example I/O state machine of the preferred embodiment of the present invention.

In the following detailed description of the present invention numerous specific details are set forth in order to provide a thorough understanding of the present invention. However, it will be obvious to one skilled in the art that the present invention may be practiced without these specific details. In other instances, well known methods, procedures, components, and circuits have not been described in detail as not to unnecessarily obscure aspects of the present invention.

Some portions of the detailed descriptions which follow are presented in terms of algorithms and symbolic representations of operations on data bits within a computer memory. These algorithmic descriptions and representations are the means used by those skilled in the data processing arts to most effectively convey the substance of their work to others skilled in the art. An algorithm is here, and generally, conceived to be a self-consistent sequence of steps leading to a desired result. The steps are those requiring physical manipulations of physical quantifies. Usually, though not necessarily, these quantities take the form of electrical or magnetic signals capable of being stored, transferred, combined, compared, and otherwise manipulated. It has proven convenient at times, principally for reasons of common usage, to refer to these signals as bits, values, elements, symbols, characters, terms, numbers, or the like.

It should be borne in mind, however, that all of these and similar terms are to be associated with the appropriate physical quantifies and are merely convenient labels applied to these quantities. Unless specifically stated otherwise as apparent from the following discussions, it is appreciated that throughout the present invention, discussions utilizing terms such as "processing" or "computing" or "calculating" or "determining" or "displaying" or the like, refer to the action and processes of a computer system, or similar electronic computing device, that manipulates and transforms data represented as physical (electronic) quantities within the computer system's registers and memories into other data similarly represented as physical quantifies within the computer system memories or registers or other such information storage, transmission or display devices.

One embodiment of the preferred embodiment of the present invention is a dual register set direct memory access (DMA) controller. However, it should be realized that the present invention is not limited to such a controller and applies equally to other multiple register set DMA controllers, such as triple register set DMA controller.

FIG. 2A shows a block diagram of a general computer system utilizing the direct memory access (DMA) controller 218 of the present invention. In the preferred embodiment of the present invention the DMA controller 218 is contained within a peripheral subsystem controller (PSC) 220. The PSC 220 controls the access between the central processing unit (CPU) bus 210 and the input/output (I/O) bus 214. The DMA controller 218 of the present invention controls the DMA transfer portion of this access and part of the non-DMA transfers. The PSC 220 will not be discussed further, except as it relates to the DMA controller 218 of the present invention.

The computer system contains two buses: the CPU bus 210 and the I/O bus 214. The CPU bus 210 of the preferred embodiment is a subset of a Motorola 68040 bus; the 68040 bus comprises a 32 bit data bus, a 32 bit address bus, and a control bus. The I/O bus 214 of the preferred embodiment comprises a 16 bit data bus, a 5 bit address bus, and a control bus. However, it should be readily apparent to those of ordinary skill in the art that the present invention may be utilized with any of a wide variety of commercially available busses.

Multiple I/O controllers are coupled to the I/O bus 214; for example, a SCSI controller 228 connected to a hard drive and an Ethernet controller 230 connected to a network may be connected to the I/O bus 214. The CPU bus 210 is coupled to multiple standard computer components, namely the CPU 222, a main memory (such as random access memory--RAM) 224, and a read only memory (ROM) 226. Multiple other components (not shown) may also be connected, such as a display device for displaying information to the computer user, an alphanumeric input device including alphanumeric and function keys for communicating information and command selections to the CPU, a cursor control device for communicating user input information and command selections to the CPU, or a stylus and tablet for communicating information representative of handwriting to the CPU. Alternatively, devices such as the alphanumeric input device, the cursor control device, and the stylus and tablet may be connected to the CPU bus 210 through a specialized port on the PSC 220. The DMA controller 218 is coupled to both the CPU bus 210 and the I/O bus 214. The DMA controller 218 manages the transfer of data between the two buses 210 and 214.

The multiple register set DMA controller 218 is shown in more detail in FIG. 2B. The controller 218 is shown connected to the CPU bus 210 on one side and the I/O bus 214 on the other side. The I/O bus 214 is shown connected to multiple I/O controllers 237. These controllers 237 are further connected to multiple I/O devices (not shown). This connection between I/O controllers and I/O devices is well known in the art and will not be described further.

The CPU bus 210 is connected to the controller 218 through CPU bus interface 235. Interface 235 is coupled to the I/O multiplexer 252 through bus 260, the internal CPU bus arbiter 248, the internal CPU bus multiplexer 249 through internal CPU bus 280 and control line 243, the DMA transaction description multiplexer 246 over bus 245, the CPU I/O interface state machine 257, and the state machines 258. The interface 235 permits access between the CPU bus 210 and the I/O bus 214 either through the DMA channels 244 or directly (thereby essentially transforming the DMA controller into a bridge between the buses 210 and 214).

The DMA controller 218 contains multiple DMA controller channels 244. These channels are shown as a single block in FIG. 2B so as not to unnecessarily clutter the figure. The multiple register set architecture of the channels 244 is described in more detail below.

Each channel 244 contains a first in-first out (FIFO) buffer which provides for data input from and output to both the CPU bus 210 and the I/O bus 214. Each channel 244 of the DMA controller 218 is connected to the internal CPU bus multiplexer 249, the DMA data multiplexer 247, and the DMA transaction description multiplexer 246. Multiplexers 246, 247 and 249 allow only one channel 244 access to the interface 235 at any given moment.

Internal CPU bus multiplexer 249 receives as input data from I/O bus 214, bus 269, and DMA data multiplexer 247. Multiplexer 249 is controlled by control line 243 from bus interface 235. Interface 235 determines which input to multiplexer 249 should be given access to the internal CPU bus 280, and thus the CPU bus 210. When interface 235 is acting as a bus master, i.e., performing a DMA transfer, on CPU bus 210, multiplexer 249 will allow the data on bus 268 of the appropriate channel (as determined by multiplexer 247, discussed below) access to the internal CPU bus 280. When interface 235 is acting as a bus slave, i.e., the values in its channels' registers are being read from or written to, multiplexer 249 will allow one of the buses 269, described in more detail below, to have access to the internal CPU bus 280. The bus 269 which should be given access to internal CPU bus 280 is determined by interface 235; that is, interface 235 will select the proper bus 269 based on decoding the address presented to interface 235 for the slave access. Additionally, when the DMA controller 218 is acting as a bridge for CPU 222 to directly access I/O bus 214, multiplexer 249 will allow I/O bus 214 to have access to internal CPU bus 280.

The internal CPU bus arbiter 248 controls DMA data multiplexer 247 and DMA transaction description multiplexer 246. Multiplexer 247 receives as input a bus 268 from each channel 244, and is controlled by control line 278. The internal CPU bus arbiter 248 determines which channel 244 should be given access to the interface 235, as discussed in more detail below. After making that determination, CPU bus arbiter 248 asserts the proper signal over control line 278, thereby causing multiplexer 247 to allow the bus 268 from the channel 244 which won the arbitration to be input to multiplexer 249.

Multiplexer 246 receives as input from each DMA channel 244 a bus 269 and transaction description information 242. Similar to multiplexer 247, multiplexer 246 is controlled by internal CPU bus arbiter 248, and thus allows the bus 269 from the DMA channel 244 which won the arbitration access to bus 245. Transaction description information 242 is output from each channel 244 and contains information such as the size of the buffer 310 within that particular channel and whether the channel is currently programmed for a read or a write (as determined by the DIR bit of the channel's active register set, discussed in more detail below).

Note that the internal CPU bus 280 is distinct from the CPU bus 210; the internal CPU bus 280 is contained wholly within the DMA controller 218 and can access the CPU bus 210 only through the bus interface 235.

Each channel 244 is also connected to the I/O multiplexer 252, which allows only one channel 244 access to the I/O bus 214 at any given moment. The I/O bus arbiter 250 determines which channel 244 should be given access to the I/O bus 214, as discussed in more detail below. Thus, these connections allow data to be transferred between the FIFO of each channel 244 and both the CPU bus 210 and the I/O bus 214.

Each channel 244 is also connected to an I/O state machine 258. Each I/O state machine 258 corresponds to one of the I/O controllers 237. Each state machine 258 is connected to and drives its corresponding I/O controller 237, and therefore also drives the I/O device connected to its corresponding I/O controller 237. The interaction between the state machines 258 and the channels 244 is described in more detail below.

Each channel 244 is also connected to a 32-bit bus 269 which is connected to multiplexer 249. Each bus 269 carries the address in main memory 224 which the data in FIFO 310 will be written to during a DMA read operation or read from during a DMA write operation for its corresponding channel. Bus 269 may also carry the values within the control register 366, the command/status registers 316 and 326, or the values in the count registers 314 and 324 of a particular channel, as shown in FIG. 3.

An example channel 244 is shown in more detail in FIG. 3.

When the DMA controller 218 is performing a read operation from an I/O controller (i.e., a transfer from an I/O device via an I/O controller to main memory), the DMA channel 244 corresponding to the I/O device, using the values in its active register set, first requests access to the I/O bus 214 from arbiter 250. The request is eventually granted, and the channel 244 retrieves a portion of the data from the I/O controller 237 over the I/O bus 214 and stores this portion in the channel's FIFO buffer. When the FIFO buffer is full (or no more data needs to be transferred), the channel 244 requests, and is eventually granted, access to the internal CPU bus 280 and bus 245. The bus interface 235 then requests, and is eventually granted, access to the CPU bus 210 of the computer system. The interface 235 will then transfer the contents of the DMA channel's FIFO buffer to main memory 224 over the CPU bus 210 of the computer system. After this portion of the transfer is complete, the DMA channel 244 will again request access to the I/O bus 214 if the overall transfer programmed by the CPU is not complete.

When the DMA controller 218 is performing a write operation to an I/O controller (i.e., a transfer from main memory to an I/O device via an I/O controller), the DMA controller 218 works in reverse of the read operation above. First, the DMA channel 244 requests access to the internal CPU bus 280; once the interface 235 is granted access to the CPU bus 210 of the computer system the channel 244 will transfer a portion of the data in main memory 224 into its FIFO buffer over bus 260. After the transfer of this portion is complete, the channel 244 requests access to the I/O bus 214 from the I/O bus arbiter 250. Once access is granted, the channel 244 transfers the data in its FIFO buffer to the I/O controller. Then, if additional data to be transferred remains in main memory 224, the channel will repeat the process.

When the channel 244 has completed the transfer that the currently active register set is programmed for, the channel 244 will switch register sets, thereby making the second register set the active register set. If the second register set has been programmed and is ready to begin a transfer, it will attempt to begin its transfer, following the steps described above.

The use of a multiple register set architecture, as described above, allows the DMA controller 218 to chain two transfers together. That is, each register set can be programmed for a transfer and the transfer for the second register set will begin immediately following the completion of the first register set's transfer. Furthermore, if the first register's transfer can be serviced quickly enough, the first register set can be programmed for a third transfer, which will begin immediately upon completion of the second register set's transfer. Thus, an infinite number of transfers could be chained together if the completed transfer can be serviced before the completion of the current transfer.

FIG. 5 shows CPU bus interface 235 in more detail. Interface 235 is the interface between the CPU bus 210 and the DMA controller 218. Interface 235 is connected to the internal CPU bus arbiter 248 over control lines 282 and 283. This connection and the function of the arbiter 248 are described in more detail below.

Interface 235 is connected to CPU bus 210 and controls the DMA controller's requests for access to CPU bus 210. Interface 235 contains an address decoder (not shown) which decodes the input address from bus 210. The address decoder is needed in the preferred embodiment in order to provide access to the proper I/O controller and the register sets. It should be noted that in the preferred embodiment, address locations in main memory 224 which are programmed in the DMA channels 244 must be physical locations; the interface 235 does not translate addresses when transferring data from the channels 244 to main memory 224.

When interface 235 receives a request for the CPU bus 210 from the internal CPU bus arbiter 248, interface 235 will request access to the CPU bus 210 from the CPU bus arbiter (not shown). One such arbiter is described in more detail in copending U.S. patent application No. 08/189,138, now U.S. Pat. No. 5,546,547, entitled "Memory Bus Arbiter for a Computer System Having a DSP Co-processor", invented by Michael J. Bowes and Farid A. Yazdy, assigned to the assignee of the present invention and filed concurrently herewith. However, access to the CPU bus 210 may also be controlled in a well known manner using a conventional CPU bus arbiter.

Interface 235 has three control lines 510, 512, and 514 which carry the signals controlling whether interface 235 will be permitted to act as a bus master on the CPU bus 210. When interface 235 requests access to the bus 210 it does so by asserting a bus request signal over control line 510. Interface 235 must then wait to access the bus 210 until it is not receiving a bus busy signal and it receives a bus grant signal. The interface 235 receives a bus busy signal over control line 514 whenever another device on the bus 210 is using the bus. When the CPU bus arbiter grants the DMA controller access to the CPU bus 210, the CPU bus arbiter will send a bus granted signal to interface 235 over control line 512; interface 235 will then assert a bus busy signal over control line 514, thereby notifying other devices on bus 210 that the bus is in use. After being granted access to the CPU bus 210 and completing the access, interface 235 will assert an "access done" signal to the internal CPU bus arbiter 248 and the DMA channel 244 over control line 283; the DMA channel may then proceed with its data transfer, as described below.

Several additional control signals are also shown connected to interface 235. These signals are used in controlling transfers to and from the interface 235 over the CPU bus 210. The use of these signals in transferring data between memory and other devices is well known in the art and will not be discussed further. Interface 235 also receives two 32-bit bus lines 245 and 280 from the multiplexers 246 and 249. The multiplexers 246 and 249 are coupled to each DMA channel 244 as described above.

Interface 235 also outputs a bus 260 to multiplexer 252 of the DMA controller 218. Bus 260 is used to transfer data from the CPU bus 210 directly to the I/O bus 214. When the CPU 222 is given access to the I/O bus 214, as discussed below, data will flow from the interface 235 directly to the I/O bus 214 over bus 260, thereby bypassing the DMA channels 244. Similarly, data being returned to the CPU bus 210 when the CPU 222 is given access to the I/O bus 214 will flow from the I/O bus 214 through multiplexer 249 and internal bus 280 back to interface 235. Thus, data can be read directly from the I/O devices, bypassing the DMA channels 244.

Each DMA controller channel 244 corresponds to a single I/O controller 237. For example, DMA controller channel(1) 244 may correspond to I/O controller(a) 237, which may be a SCSI controller, and DMA controller channel(2) 244 may correspond to I/O controller(b) 237 which may be a network controller such as an Ethernet controller. Thus, whenever data is DMA transferred through DMA controller channel(1), it will be transferred to or from SCSI controller(a). Similarly, each DMA controller channel 244 corresponds to a particular I/O state machine 258 (the same state machine 258 which corresponds to the channel's corresponding I/O controller 237). It should be noted that the number of DMA channels 244 and I/O controllers 237 need not be identical--multiple channels 244 may exist for which no corresponding I/O controller 237 is connected to I/O bus 214. It should also be noted that although each DMA channel 244 corresponds to a particular I/O state machine 258, each state machine 258 may correspond to multiple DMA channels 244. Thus, for example, the ethernet receive channel and the ethernet write channel may both correspond to the same ethernet I/O state machine 258. Each controller 237 may have an I/O device connected to it, e.g. a hard disk drive attached to SCSI controller 228 shown in FIG. 2A. These connections between I/O controllers and I/O devices are well known to those of ordinary skill in the art and need not be discussed further.

FIFO Buffer

FIG. 3 shows an example DMA channel in more detail. In the preferred embodiment of the present invention, the FIFO buffer 310 is a pseudo-FIFO buffer. That is, data is transferred through the buffer 310 in a first in-first out manner, however all data is removed from the buffer before any additional data is added. For example, if the buffer 310 is 16 bytes, then during a transfer 16 bytes will be loaded into the buffer and then removed in a first in-first out manner. Additional data will not, however, be added to the buffer 310 until all 16 bytes have been removed. One such buffer is described in more detail in copending U.S. patent application No. 08/189,141, still pending, entitled "Pseudo First-In-First-Out (FIFO) Buffer", invented by Brian A. Childers and Michael J. Bowes, assigned to the assignee of the present invention and filed concurrently herewith. However, it should be apparent to those of ordinary skill in the art that a wide range of FIFO-type buffers could be used as the FIFO buffer 310.

The use of a separate FIFO 310 in each channel 244 serves several important functions. Among these functions is that it allows for concurrent transfers. That is, one channel 244 may be transferring the contents of its FIFO over the CPU bus 210 while at the same lime a second channel 244 is loading its FIFO from an I/O controller over I/O bus 214. Furthermore, the FIFO insulates the CPU bus 210 from the slow access times of the I/O devices on the I/O bus 214. Data is transferred to or from an I/O controller from or to the FIFO of a channel 244 then, when access to the CPU bus is granted, the FIFO's contents are burst over the CPU bus 210, thereby minimizing the impact of I/O transfers on the CPU bus 210 bandwidth.

The FIFO 310 of each DMA controller channel 244 may be of varying size. One embodiment of the present invention utilizes FIFOs of either 4 or 16 bytes, i.e., equivalent to either a longword or a quad longword. The size of the buffer is determined giving weight to several considerations, including the bandwidth of the given I/O device and the latency requirements of the device.

Each channel 244 contains only a single FIFO 310 which is shared by the two register sets. Only one register set is active (i.e., transferring data) at any given moment, as is described in more detail below. Thus, the single FIFO 310 can be shared by the two register sets without conflict.

Each channel 244 is coupled to two 32-bit buses 260 and 268 which are connected to interface 235, as described above. Each bus 260 is input to FIFO 310 from interface 235; bus 260 carries data from the CPU bus 210 which is loaded into FIFO 310 during a DMA write operation. Similarly, bus 268 carries data from FIFO 310 to CPU bus 210 (through multiplexers 247 and 249, and interface 235) during a DMA read operation or a slave access. Note that bus 260 also carries data to the register sets and bus 269 carries data from the register sets during slave accesses, described in more detail below.

Each FIFO 310 is also coupled to two 16-bit data buses 262 and 264 which are connected to I/O bus 214. Data bus 262 is an input bus which carries data from an I/O controller 237 to FIFO 310 over I/O bus 214 during a read operation. Similarly, data bus 264 carries data from FIFO 310 to an I/O controller 237, as determined by multiplexer 252, during a write operation. Multiplexer 252 is controlled by I/O bus arbiter 250 and will allow the data from only one channel 244 to access I/O bus 214 at any given time.

Write FIFO Control

A set of control lines used for the internal CPU bus arbitration and for the I/O bus arbitration are also coupled to each DMA channel 244. These control lines are input to and output from the Write FIFO Control block 320 (these lines are also connected to the Read FIFO Control block 330 and Read Flush block 350, as described in more detail below). The control lines include a CPU bus request signal line 270, a CPU bus grant signal line 271, a CPU access done signal line 334, an I/O bus request signal line 274, an I/O bus busy signal line 344, an I/O bus grant signal line 340, and an I/O access done signal line 348. The interaction of these control lines with the other components of the DMA channel 244 is described in more detail below.

Write block 320 also receives input on command/status bus 362 and control bus 364. Buses 362 and 364 are input from control register 366 and command/status registers 316 and 326 respectively. The inputs on buses 362 and 364 provide Write block 320 with the information from registers 366, 316 and 326 necessary to perform the write transfer. The information transferred over these buses is described in more detail below.

The CPU bus request signal line 270 is output from the Write block 320 to the internal CPU bus arbiter 248 and is asserted when the channel 244 has been programmed and is ready to transfer data from main memory 224 to the channel FIFO 310. The Write block 320 will assert the CPU bus request signal line 270 when the ENABLED bit of the command/status register 316 or 326, discussed in more detail below, of the currently active register set has been set, the DIR bit of the command/status register 316 or 326 indicates a write, the channel is not PAUSEd, and the value in the count register 314 or 324 of the currently active register set is greater than zero. When the channel is granted access to the internal CPU bus 280, the Write block 320 will receive a signal over the CPU bus grant signal line 271. The interface 235, described above, will then request access to the CPU bus 210, and upon receiving access will transfer the data from main memory 224 to FIFO 310, filling FIFO 310. Upon completion of the transfer from main memory 224 to FIFO 310, Write block 320 receives a CPU access done signal over signal line 334 from arbiter 248. This informs the Write block 320 that the transfer is complete and that both the transfer from FIFO 310 to I/O controller 237 should commence and the address register should be updated.

In an alternate embodiment, Write block 320 may receive a CPU access done signal directly from bus interface 235. This would relieve arbiter 248 of the task of sending the access done signal, however it would also require arbiter 248 to send a signal to interface 235 indicating which channel is performing the DMA transfer. The interface 235 would need to know which channel is performing the DMA transfer in order to know which channel to directly send the CPU access done signal to.

After FIFO 310 has been filled with data from memory 224, Write block 320 asserts a write increment signal over control line 371. Control line 371 is ORd together with a read increment signal on control line 373 to form an address increment signal on control line 375. The address increment signal is input as a control signal into adder 370, which receives as data inputs the addresses stored in the two address registers 312 and 322. Only one of the two addresses from registers 312 and 322 will be input to adder 370 at any one time, i.e., only the address of the currently active register set will be updated at any one time. This is shown by multiplexer 386, which determines which register's address will be input into adder 370 based on the currently active register set, as stored in control register 366.

The example channel shown in FIG. 3 also has as input to adder 370 a fixed value of 16. Thus, when adder 370 is activated by an address increment signal over control line 375, the adder 370 will increment the address in the currently active register set by 16 and store this new address in the address register of the currently active register set. It should be noted that in the example of FIG. 3 the FIFO buffer 310 is 16 bytes; thus, each transfer of a portion of the entire data set to be transferred from main memory 224 to an I/O controller is done in 16-byte increments. Therefore, by incrementing the address in the address register by 16, the address is incremented to the location of the next 16-byte portion of data to be transferred from main memory 224 into FIFO 310. It should also be noted that the preferred embodiment of the example of the present invention shown in FIG. 3 only utilizes bits 4-31 of the 32-bit register; i.e., the four least significant bits are hardwired to zero. These four bits are hardwired to zero because transfers are made in 16-byte blocks; thus there is no need to store a memory address that is not quad-longword (16-byte) aligned. It will be apparent to those of ordinary skill in the art that this also requires the data being transferred to start in a memory location which is quad-longword aligned, i.e., each of the last four bits of each memory address must be zero.

In an alternate embodiment, the FIFO buffer 310 may store only 4 bytes rather than 16 bytes. Thus, the adder 370 would receive a value of four as input, rather than 16, and therefore would increment the address register by 4 rather than 16 after data has been transferred into FIFO 310.

In the preferred embodiment, after updating the address register write block 320 begins to transfer the data in FIFO 310 to its corresponding I/O controller 237. Write block 320 asserts a signal over the I/O bus request signal line 274 to the I/O bus arbiter 250. I/O bus request signal line 274 is also input to the I/O state machine 258 corresponding to this channel 244. Write block 320 will continue to receive a bus busy signal from arbiter 250 over I/O bus busy signal line 344 while the I/O bus is occupied by another transfer. When the arbiter 250 does grant the channel 244 access to the I/O bus 214, the arbiter 250 will send an I/O bus granted signal to both the Write block 320 and the I/O state machine 258 over I/O bus granted signal line 340. The I/O state machine 258 will then transfer two bytes (the size of the I/O bus 214) from FIFO 310 to the I/O controller 237. It should be noted that some I/O devices are 8-bit devices rather than 16-bit devices, and thus only one byte would be transferred rather than two bytes.

Upon completion of the transfer of these two bytes, the I/O state machine 258 asserts an I/O access done signal over signal line 348 to Write block 320. This informs Write block 320 to update the count register and proceed with the transfer. To update the count register, Write block 320 asserts a write decrement signal over control line 372. Control line 372 is ORd together with a read decrement signal on control line 374 to form a count decrement signal on control line 376. The count decrement signal is input as a control signal into adder 380, which receives as data inputs the count values stored in the two count registers 314 and 324. Only one of the two count values from registers 314 and 324 will be input to adder 380 at any one time, i.e., only the count value of the currently active register set will be updated at any one time. This is shown by multiplexer 387, which determines which register's count value will be input into adder 380 based on the currently active register set, as stored in control register 366.

The example channel shown in FIG. 3 also has as input to adder 380 a fixed value of -2. Thus, when adder 380 is activated by a count decrement signal over control line 376 the adder 370 will decrement the value in the currently active register set by 2 and store this new count value in the count register of the currently active register set.

In an alternate embodiment, the FIFO buffer 310 may store only 4 bytes rather than 16 bytes, and the I/O data lines 262 and 264 may be only 8 bits rather than 16 bits. Thus, the adder 370 would receive a value of -1 as input, rather than -2, and therefore would decrement the count register by 1 rather than 2 after one byte has been transferred into FIFO 310.

After updating the appropriate count register, the write control block 320 will do one of three things, depending on whether the programmed transfer has been completed. If the programmed transfer has been completed, then write control block 320 will go idle and send a switch register set signal to combinatorial logic 360. This signal will cause the active register set to be switched; the active register byte will be updated in control register 366 and the other register set will become the active register set. If the newly activated register set is enabled, which is discussed in more detail below, then either the write control block 320 will repeat the above described process, or the read control block 330, as described below, will perform the transfer, depending on whether the newly activated register set is programmed for a DMA write or read, respectively. If, however, the programmed transfer has not been completed, as determined by the value in the count register 314 or 324 of the active register set, control block 320 will follow one of two courses of action. If any data remains in the FIFO buffer 310, then control block 320 will repeat the above process starting with requesting access to the I/O bus and will transfer another two bytes from the FIFO 310 to the I/O controller 237. However, if the contents of the FIFO 310 have been fully transferred to the I/O controller 237, then the write control block 320 will again request access to the internal CPU bus 280 and will repeat the above process to transfer another 16 bytes into the FIFO 310.

The Write control block 320 may also be connected to the FIFO buffer 310. This connection would allow write control 320 to send any required control signals to FIFO buffer 310. These control signals will not be described further as the present invention may use any of a wide variety of FIFO buffers.

Read FIFO Control

The set of control lines used for the I/O bus arbitration, discussed above, are also coupled to the Read FIFO Control block 330. The control lines include the I/O bus request signal line 274, the I/O bus busy signal line 344, the I/O bus grant signal line 340, and the I/O access done signal line 348.

Read block 330 also receives input on command/status bus 362 and control bus 364. Buses 362 and 364 are input from control register 366 and command/status registers 316 and 326 respectively. The inputs on buses 362 and 364 provide Read block 330 with the information from registers 366, 316 and 326 necessary to perform the read transfer. The information transferred over these buses is described in more detail below.

The I/O bus request signal line 274 is output from the Read block 330 to the I/O bus arbiter 250 and is asserted when the channel 244 has been programmed and is ready to transfer data from an I/O controller 237 to its FIFO 310. I/O bus request signal line 274 is also input to the I/O state machine 258 corresponding to this channel 244. As discussed above with the write control block 320, the read block 330 knows to assert the signal line 274 when the ENABLED bit of the command/status register 316 or 326 of the currently active register set has been set, the DIR bit of the command/status register 316 or 326 indicates a read, the channel is not PAUSEd, and the value in the count register 314 or 324 of the currently active register set is greater than zero.

The Read block 330 will receive a bus busy signal from arbiter 250 over I/O bus busy signal line 344 while the I/O bus is occupied by another transfer. When the arbiter 250 does grant the channel 244 access to the I/O bus 214, the arbiter 250 will send an I/O bus granted signal to the Read block 330 and the I/O state machine 258 over I/O bus granted signal line 340. The I/O bus granted signal combined with the I/O bus busy signal being inactive allows the I/O state machine 258 to transfer two bytes (the size of the I/O bus 214) from I/O controller 237 to FIFO 310. As noted above, a particular I/O device may be an 8-bit rather than 16-bit device, and thus only one byte would be transferred rather than two bytes.

Upon completion of the transfer of these two bytes, the I/O state machine 258 asserts an I/O access done signal over signal line 348 to Read block 330. This informs Read block 330 to update the count register and proceed with the transfer. To update the count register, Read block 330 asserts a read decrement signal over control line 374. Control line 374 is ORd together with a write decrement signal on control line 374 to form a count decrement signal on control line 376. The count decrement signal is input as a control signal into adder 380, and functions as described above.

To proceed with the transfer, read control 330 will follow one of two courses of action. If additional data remains to be transferred from the I/O controller to main memory 224 (i.e., the TERMCNT bit of the active command/status register is not set) and the FIFO 310 is not full, then read control block 330 will repeat the above process to transfer another two bytes into FIFO 310. If, however, FIFO 310 is full or the transfer is complete (i.e., TERMCNT is set), then read block 330 asserts a start flush signal over control line 331 to the read flush block 350, which proceeds to transfer the data from FIFO 310 to main memory 224, as described in more detail below.

It should also be noted that Read control 330 will also assert a start flush signal over control line 331 in response to the FLUSH bit of control register 366 being set. Thus, the software may force the DMA channel to transfer the contents of the FIFO 310 to main memory by setting the FLUSH bit

Read Flush Block

The set of control lines used for the internal CPU bus arbitration, described above, are also coupled to fie Read flush block 350. The control lines include the CPU bus request signal line 270, the CPU bus grant signal line 271, and the CPU access done signal line 334.

The Flush block 350 controls the transfer of data in the FIFO 310 to main memory 224 during a read operation. Flush block 350 begins this transfer, as discussed below, upon receipt of a start flush signal from the Read FIFO control block 330 over control line 331.

The CPU bus request signal line 270 is output from the Flush block 350 to the internal CPU bus arbiter 248 and is asserted when the channel 244 has been programmed to and is ready to transfer data from the channel FIFO 310 to main memory 224. When the channel is granted access to the internal CPU bus 280, the Flush block 350 will receive a signal over the CPU bus grant signal line 271. The interface 235, described above, will then request access to the CPU bus 210 and, upon receiving access, transfer the entire contents of FIFO 310 to main memory 224. Upon completion of the transfer from FIFO 310 to main memory 224, Flush block 350 receives a CPU access done signal over signal line 334 from internal CPU bus arbiter 248; which informs Flush block 350 that the transfer is complete.

After receiving the CPU access done signal, Flush block 350 asserts a Read increment signal over control line 373. Control line 373 is ORd together with the write increment signal, discussed above, on control line 371 to form an address increment signal on control line 375. The address increment signal is input as a control signal into adder 370 and increments the address register of the active register set, as described above.

After incrementing the address register, Flush block 350 asserts a flush done signal over control line 332, which is carried to Read FIFO control block 330. This signal informs Read control block 330 that the transfer from FIFO 310 to main memory 224 has been completed. Flush block 350 will then go idle and wait for another start flush signal from control block 330.

Upon receiving the flush done signal over control line 332, Read FIFO control block 330 will do one of two things, depending on whether the programmed transfer has been completed. If the programmed transfer has not been completed, as indicated by the TERMCNT bit of the active register set not having been set, then control block 330 will repeat the above described process to transfer another portion of the data. If, however, the programmed transfer has been completed, then read control block 330 will go idle and send a switch register set signal to combinatorial logic 360. This signal will cause the active register set to be switched; the active register byte will be updated in control register 366 and the other register set will become the active register set. If the newly activated register set is enabled, which is discussed in more detail below, then either the read control block 330 will repeat the above described process, or the write control block 320, as described above, will perform the transfer, depending on whether the newly activated register set is programmed for a DMA read or write, respectively.

Register Sets

Address registers 312 and 322 and count registers 314 and 324 are programmed by the CPU prior to a data transfer. Address registers 312 and 322 are programmed with the address in main memory 224 which the DMA controller is to transfer data to or read data from. Count registers 314 and 324 are programmed with the number of bytes the DMA controller is to transfer. Command/status registers 316 and 326 are referred to as "command registers" when they are being written to and "status registers" when they are being read from. The information stored in these registers is described below.

The address registers 312 and 322 are updated after each transfer of data to or from the FIFO 310, as discussed above. The updated addresses indicate the next location in main memory 224 to write to or read from. Similarly, the count registers 314 and 324 are updated after each transfer of data to main memory 224 or to an I/O controller 237. The updated count values indicate the number of bytes remaining to be transferred.

When the DMA channel 244 is configured for writing, the count register 314 or 324 indicates how many bytes remain to be transferred to the I/O controller as the transfer proceeds. As discussed above, the count register is updated after the data is transferred to the FIFO to or from the I/O controller. The count registers will not indicate the number of bytes which have been transferred to the DMA channel's FIFO (e.g., some bytes may be in the FIFO waiting to be transferred to the I/O controller). When the DMA channel 244 is configured for reading, the count register 314 or 324 indicates how many bytes have been transferred from the I/O controller to the channel's FIFO. As described above, the count register is updated after data is transferred from the I/O controller to the FIFO. The count registers will not indicate the number of bytes which have been actually transferred to main memory 224.

It should be noted that in the preferred embodiment of the present invention the 32-bit count register is actually hardwired to contain only 17 bits of programmable counter. Seventeen bits allows up to 128K-1 bytes to be transferred in a single DMA operation. In the preferred embodiment, the largest transfers are SCSI transfers, which are limited to 64K bytes. Implementing only 17 of the possible 32 bits reduces gate count and cost. However, it should be readily apparent to those of ordinary skill in the art that the full 32 bits of the count registers 314 and 324 could be utilized if an I/O controller capable of transferring more than 64K bytes in a single transfer was coupled to the I/O bus.

It should also be noted that in the preferred embodiment of the present invention the address registers 312 and 322 contain the physical location of the next transfer, not the logical location. The multiple register set DMA controller 218 of the preferred embodiment of the present invention does not contain the circuitry necessary to decode logical addressing, including virtual memory. It should therefore be apparent to those of ordinary skill in the art that the physical location requirement demands that the address space in memory where the transfer is to take place must not cross a physical page boundary unless the pages affected are locked down in a single physical contiguous block. Furthermore, even for transfers within a single page, that page must remain locked during the DMA transfer. It should also be apparent, however, that the multiple register set DMA controller of the present invention could accommodate logical addresses if a proper decoding means was added.

The CPU can perform a slave access to the DMA controller 218, thereby allowing the software to program the controller 218 for a transfer. That is, during a slave access the software can write directly to (or read from) the register sets in any given channel. A series of select signals (not shown) from interface 235 is connected through a set of multiplexers to allow access to each of the registers in controller 218. During a slave access, interface 235 will decode the address received over CPU bus 210 and assert the proper select signals based on this address, thereby allowing software to access the desired registers. During a slave read, the values contained within the requested registers are returned to interface 235 over bus 269 (through multiplexer 249). During a slave write, the values to be written into the specified registers are input from CPU bus 210, through interface 235, to the registers over bus 260.

The ability to perform slave accesses to the DMA controller 218 is required to program the register sets for transfers. In addition, the ability is particularly valuable when the software needs to check on the status of a programmed transfer, or whether a PAUSE has been completed, as indicated by the FROZEN bit, discussed below.

In the case of programming the register sets, upon completing the programming the software will set the ENABLED bit of the programmed register set's command/status register 316 or 326. As described above, this indicates to either the read control block 330 or write control block 320 to begin a transfer, depending on the setting of the DIR bit.

FIG. 4 shows the registers used in the preferred embodiment of the present invention. A single bus error interrupt enable register (BERRIE) 410 and a single interrupt status register (ISR) 420 contain control information for all DMA channels 244 in the DMA controller 218. A channel control register 430 contains control information for a particular channel; there is an individual control register 430 for each channel 244 in controller 218. Two register sets, register set(0) 440 and register set(1) 450, also contain control information for each channel 244. Each register set 440 and 450 comprises three registers: memory address registers 312 and 322, count registers 314 and 324, and command/status registers 316 and 326.

A single register, BERRIE 410, is used in the preferred embodiment of the present invention to enable or disable the bus error interrupts for all the channels of the DMA controller 218. BERRIE 410 is an 8-bit register, however only bit 0 is used; bits 7-1 are undefined and reserved for future use. Bit 0 of BERRIE 410 acts as the bus error global interrupt enable. This bit is set by software to "1" to enable, or "0" to disable, the bus error interrupts for all the DMA channels 244.

The ISR 420 contains 256 bits in the preferred embodiment of the present invention. The first 32 bits are used by the preferred embodiment and the remaining 224 bits are undefined and reserved for future expansion. Each of the 32 bits used in ISR 420 corresponds to an individual DMA channel 244. If the bit for a particular register is set to "1" then the corresponding DMA channel 244 is requesting an interrupt, as determined by the setting of the CIRQ bit of the channel's control register is set. A DMA channel could be requesting an interrupt for two reasons: either the completion of a register set's DMA transfer or a bus error (BERR). The BERRIE 410 bit allows the software to deactivate the bus error interrupt, thereby making the only possible interrupt a DMA transfer complete interrupt. Once an interrupt is requested, the DMA interrupt handler may read the control register 430 of the interrupting channel to determine which of the two reasons listed above is the cause of the interrupt and respond accordingly. Note that if the BERRIE 410 bit is disabled then the interrupt handler will know that the interrupt is due to the completion of a register set's DMA transfer.

The channel control register 430 is a 16-bit register in the preferred embodiment of the present invention. The information contained in control register 430 is shown in Table 1.

Each command/status register 316 or 326 is a 16-bit register in the preferred embodiment of the present invention. The information contained within each command/status register is shown in Table 2.

The combinatorial logic 360 receives inputs from the CPU bus 260, the control register 366, the command/status registers 316 and 326, and the write and read control blocks 320 and 330. The combinatorial logic 360 utilizes these inputs to calculate new values for the control register 366 and command/status registers 316 and 326, described above. The values stored in control register 366 and command/status registers 316 and 326 are also made available over buses 362 and 364, respectively, to the write and read control blocks 320 and 330. These outputs are utilized by the write and read control blocks 320 and 330, to determine whether to begin, and proceed with a transfer.

Combinatorial logic 360 also receives as input the values in the count registers 314 and 324, and will set the TERMCNT bit of the active register set's command/status register when the value in the register set's count register is decremented to zero. Thus, this may be used by read control block 330 and write control block 320 to determine that the programmed transfer is complete.

Returning to FIG. 2B, the internal CPU bus arbiter 248 controls which DMA channel will have access to the CPU bus interface 235, and thus can write the data in its FIFO buffer 310 to main memory 224 or read data into its FIFO buffer 310 from main memory 224. The arbiter 248 is connected to DMA data multiplexer 247 and DMA transaction description multiplexer 246 through control line 278, described above.

The internal CPU bus arbiter 248 is shown in more detail in FIG. 6. Arbiter 248 is connected to the bus interface 235 over control lines 282 and 283. The CPU access request signal over line 282 is given by arbiter 248 to interface 235 to inform interface 235 that the DMA controller has a channel ready to transfer data to or from main memory 224. This signal informs interface 235 to request access to the CPU bus 210 in order to transfer the data. Access to the CPU bus 210 is determined by the CPU bus arbiter (not shown), as discussed above. Arbiter 248 will leave access to the internal CPU bus 280 with the channel 244 which requested access to the CPU bus 210 until the transfer is complete.

When the transfer is completed, interface 235 sends an access done signal to arbiter 248 over control line 283, thereby informing arbiter 248 that the requested transfer has been completed. Arbiter 248 therefore begins arbitration again to determine which DMA controller channel 244 will be allowed access to the internal CPU bus 280.

Channel 244 asserts a CPU bus request signal over control line 270 to arbiter 248 whenever the channel 244 has data in its FIFO buffer ready to be transferred to main memory 224 or is ready to transfer data from main memory 224 to its FIFO buffer. Bus arbiter 248 will assert a CPU bus grant signal over control line 271 to channel 244 when the arbiter 248 determines the channel can have access to the internal CPU bus 280, and therefore the CPU bus 210.

Arbitration for the internal CPU bus 280 is carried out following a predetermined fixed priority scheme. That is, the arbiter 248 is preprogrammed with a specific priority order--high priority channels are given access to the internal CPU bus 280 before low priority channels. Which channel is given a higher priority is determined by the system designer, taking into account the ability of the corresponding I/O controller to tolerate long waiting periods, in an attempt to keep data flowing to and from the most I/O devices given limited bus bandwidth. Depending on the particular I/O device, this tolerance may be determined by taking into account the maximum latency and the average bandwidth. That is, the latency of the system bus, as seen by the PSC, and the devices' tolerance for latency (as determined by their individual data rates and the amount of FIFOing their individual I/O device controllers have) would be considered. In addition, the average data bandwidths of the different I/O devices may be considered.

In the preferred embodiment of the present invention, the sound I/O DMA channels are have the highest priority because they are the least tolerant to waiting long periods for access to the CPU bus. The SCSI controller is the lowest priority because it is the least affected by long waiting periods.

It should also be noted that the arbiter for the CPU bus (not shown) is designed to allocate sufficient bandwidth to the DMA controller 218 so that all channels will be able to gain access to the CPU bus 210. That is, higher priority DMA channels will not be able to starve out lower priority DMA channels because of insufficient access to the CPU bus 210.

One such arbitration scheme is described in copending U.S. patent application No. 08/189,138, now U.S. Pat. No. 5,546,547, "Memory Bus Arbiter for a Computer System Having a DSP Co-processor", invented by Michael J. Bowes and Farid A. Yazdy, assigned to the assignee of the present invention and filed concurrently herewith. That application describes an arbitration scheme implemented for a memory bus shared by four bus masters: the CPU, the present invention DMA controller, a NuBus controller and a digital signal processor for carrying out real-time signal processing. Of course, the present invention DMA controller may be implemented in less complicated systems in which the CPU bus constraints are less critical than the arrangement described in that application.

Returning to FIG. 2B, the I/O bus arbiter 250 controls whether the DMA controller 218 or the CPU 222 will have access to the I/O controllers 237 at any particular moment. If the arbiter 250 determines that the DMA controller 218 will have access to the I/O controllers 237 at a particular moment then the arbiter further determines which DMA controller channel 244 will be given access to the I/O controllers 237.

Arbiter 250 controls which channel 244, or CPU 222, will have access through the I/O multiplexer 252. Multiplexer 252 receives control signals from arbiter 250 over I/O multiplexer control line 284. Multiplexer 252 receives data bus 264 and CPU output bus 260 as input. The inputs to multiplexer 252 are multiplexed onto I/O bus 214.

Arbiter 250 receives as input CPU I/O access request control line 288 and DMA I/O bus request lines 274 as shown in more detail in FIG. 7A. CPU I/O access request line 288 carries a signal from the CPU I/O interface state machine 257 that the CPU 222 is requesting access to the I/O bus 214, described in more detail below. I/O bus request line 274 informs the arbiter 250 that the DMA channel 244 is ready for a transfer to or from an I/O controller 237, as discussed above. The I/O bus request line 274 is also input to the corresponding I/O state machine 258, which allows the state machine 258 to prepare for a data transfer. The device ready signal informs arbiter 250 over device ready line 514 that the I/O controller 237 which corresponds to this state machine 258 is ready for data transfer; the determination of this readiness is discussed in more detail below.

Arbiter 250 will determine which channel, of those fully ready for transfer, is entitled to access to the I/O bus 214. A channel is considered fully ready for transfer if two conditions are met: first, arbiter 250 must receive an I/O bus request signal from the channel 244, and second, arbiter 250 must receive a device ready signal over line 514 from the corresponding I/O state machine 258. Thus, if a channel is fully ready both the channel 244 is ready for transfer and the corresponding I/O controller 237 is ready, so that once access to the I/O bus 214 is granted the transfer can begin. Note that even if a controller is fully ready, the I/O state machine 258 will not attempt to access the I/O bus 214 because it will be receiving a bus busy signal over bus busy line 344. The state machine 258 will not access the I/O bus 214 until it is not receiving a busy signal over line 344 and it receives an I/O bus grant signal over control line 340, described below.

After a channel is fully ready it will eventually be granted access to the I/O bus 214 by the arbiter 250. When access is granted, arbiter 250 sends an I/O bus grant signal over bus grant line 340 to both the channel 244 and the state machine 258. The bus grant signal informs the channel 244 that transfer is about to begin and that it can deactivate its bus request line. The bus grant signal also informs the state machine 258 that the bus is available for its use. State machine 258 will respond to the bus grant signal by asserting a state machine busy signal over state machine busy line 512. The state machine busy signal informs the arbiter 250 that the state machine 258 has control of the I/O bus 214 and that nothing else should try to access bus 214.

FIG. 7B shows an example I/O state machine 258 of the present invention. The control lines connected to the SCSI I/O controller 237 are shown. These lines carry the control signals to perform the "handshake" between the DMA controller and the corresponding I/O controller 237, thereby effecting a data transfer. The I/O state machines 258 transfer the two bytes of data between the I/O controllers and the FIFO 310, as discussed above. The I/O state machines operate in a conventional manner to transfer data between the DMA channels and their corresponding I/O controller. These handshake signals and the transferring of 16 bits of data between the I/O controllers and a data buffer or register are well-known to those of ordinary skill in the art and will not be described further. The remaining I/O state machines 258 are not shown as they differ from the SCSI state machine in FIG. 7B only in the manner in which they perform the handshake with their corresponding controllers.

Arbiter 250 determines access fights to the I/O controllers 237 using a two-step process. First, arbiter 250 determines whether the CPU 222 or the DMA controller 218 is entitled to access. Second, if the DMA controller 218 is entitled to access then arbiter 250 determines which DMA controller channel 244 is entitled to access.

Whether the CPU 222 or the DMA controller 218 is entitled to access is determined using a three stage round-robin scheme: CPU, DMA, DMA. That is, the CPU 222 will be given access to the I/O bus 214, then the DMA controller 218 will be given access twice, followed by the CPU 222 being given access, etc. Note that this round-robin scheme is applied only if both the DMA controller 218 and the CPU 222 request access to the I/O bus 214; therefore, the DMA controller 218 would be given continuous access to the I/O bus 214 if the CPU 222 never requested access. This round-robin scheme assures that both the DMA controller 218 and the CPU 222 will be given access to the I/O bus 214; neither is allowed to starve off the other, and the CPU 222 still has reasonably quick direct access to the I/O controllers. Note also that the DMA controller 218 receives twice the access that the CPU 222 receives because one objective of the present invention is to allow the CPU 222 to perform non-I/O tasks and leave the I/O tasks to the DMA controller 218.

If arbiter 250 determines that DMA controller 218 is entitled to access to the I/O bus 214 then the arbiter determines which DMA controller channel 244 is entitled to access. The channels 244 which will participate in any given arbitration are those which are fully ready, i.e., both ready for transfer themselves and whose corresponding I/O controller is ready for transfer, as described above. Arbiter 250 follows a fixed priority scheme, as described above utilized by the internal CPU bus arbiter 248. In the preferred embodiment of the present invention the fixed priority followed by arbiter 250 is the same as the fixed priority followed by arbiter 248. It should be noted, however, that this similarity is not required; the system designer may utilize different priorities depending on the circumstances.

As discussed above, if both the DMA controller 218 and the CPU 222 are requesting access to the I/O bus 214 then the I/O bus arbiter 250 of the preferred embodiment will follow a CPU/DMA/DMA round robin scheme. Thus, one out of every three accesses to the I/O bus 214 will be made by the CPU 222. For a CPU access the DMA controller essentially becomes a bridge between the CPU bus 210 and the I/O bus 214, bypassing the multiple channels described above. This bridge functionality allows the CPU to either set up an I/O controller for a DMA transfer or perform a direct access to an I/O controller when needed.

When the CPU 222 has access to the I/O bus 214, and the PSC is therefore acting as a bridge, the DMA channels 244 are bypassed. Thus, CPU I/O interface state machine 257 is used to augment control of the CPU access to the I/O bus 214. A series of select signals also connects interface 235 and each state machine 258. The interface 235 asserts the proper select signal based on the decoded value received from CPU 222. That is, when CPU 222 is requesting direct access to the I/O bus 214, the address CPU 222 specifies is decoded to determine the I/O controller that CPU 222 is requesting access to, and interface 235 asserts the appropriate select signal. Note that additional state machines (not shown) may exist for additional system resources which only operate in slave mode, i.e., devices which would not be participating in a DMA transfer.

A CPU I/O request signal is also asserted to CPU I/O interface state machine 257 from interface 235 when the CPU 222 is requesting access to the I/O bus 214. Upon receipt of the CPU I/O request signal, CPU I/O interface 257 asserts a signal over CPU I/O access request line 288 to arbiter 250. When arbiter 250 grants I/O bus access to the CPU 222, as described above, arbiter 250 will assert a CPU I/O request signal 346 to all the I/O state machines 258 and CPU I/O interface state machine 257. Each state machine 258 will combine this signal 346 with its select line input from interface 235. If both the signal line 346 and the select line from interface 235 are set, and the I/O bus busy signal on line 344 is deasserted, then the I/O state machine will assert its state machine busy line 512 and will perform the desired transfer of data with its corresponding I/O controller. CPU I/O interface state machine 257 receives as input the I/O bus busy signal 344 and the CPU I/O request signal 346 from arbiter 250. When interface 257 observes CPU I/O request signal 346 asserted and I/O bus busy signal 344 deasserted, it knows the I/O state machine is about to make the transfer. Arbiter 250 then asserts I/O bus busy line 344 in response to the state machine asserting its state machine busy line 512. Upon completion of the transfer, the state machine will release (deassert) its state machine busy line 512, which causes I/O arbiter 250 to deassert I/O bus busy line 344. This second deassertion of the I/O bus busy line 344 indicates to CPU I/O interface state machine 257 that the access has been completed. In response, the CPU I/O interface state machine 257 will send a CPU I/O access done signal to interface 235, thereby informing CPU 222 that the access to the I/O bus has been completed.

In an alternate embodiment of the present invention a second I/O bus is coupled to the DMA controller 218 for sound I/O. The DMA controller 218 has two channels for sound I/O--a sound in channel and a sound out channel. The operation of the two sound channels and the second I/O bus is essentially identical to the above described preferred embodiment, except that the sound channels are not involved in the arbitration for the I/O bus 214. The sound channels do participate in arbitration for the internal CPU bus 280, however they are predetermined to have the highest priority. Thus, whenever sound is input or output, it will not have to wait for access to the I/O bus 214 and it will be guaranteed the highest priority for the CPU bus 210. The high priority and separate bus are utilized because of the exceptional intolerance to waiting that sound exhibits.

The preferred embodiment of the present invention, a multiple register set direct memory access channel architecture and method, is thus described. While the present invention has been described in particular embodiments, it should be appreciated that the present invention should not be construed as limited by such embodiments, but rather construed according to the below claims.