PRD (physical region descriptor) pre-fetch methods for DMA (direct memory access) units

1. Field of the Invention

The disclosure relates generally to DMA (Directly Memory Access) management, and, more particularly to PRD (Physical Region Descriptor) pre-fetch methods for DMA units.

2. Description of the Related Art

In computer architecture, DMA allows specific hardware to independently access the system memory without the use of a CPU (Central Processing Unit). DMA transactions copy memory regions between devices. At the same time, the CPU can be scheduled for other tasks, improving system performance.

PRD entries are stored in a PRD table in the system memory. A PRD entry defines information such as starting address and size of a specific memory block in the memory. Before the DMA operates, the DMA unit usually reads a PRD entry from the PRD table, thus to obtain the starting address and the size of a memory block to be accessed according to the PRD entry. Then, the DMA unit performs access operations to the memory block corresponding to the PRD entry, that is, to write data to the memory block or read data from the memory block. FIG. 1 is a schematic diagram illustrating a conventional DMA unit. As shown in FIG. 1 , the DMA unit 300 comprises an interface A 310 , an interface B 320 , and a cache memory 350 . The interface A 310 and the interface B 320 can be used to access the buses A and B, respectively. The cache memory 350 can be used to store the PRD entries pre-fetched by the interface A via the bus A. The DAM unit 300 further comprises a queue A 331 such as FIFO (First In First Out) queue, and a queue B 332 . For DMA out transaction, the interface A 310 reads data from the memory 340 via the bus A, and stores the data to the queue A 331 . The interface B 320 reads data from the queue A 331 , and writes the data to the bus B, thus to transfer the data to a corresponding device, such as SATA or USB device. For DMA in transactions, the interface B 320 reads data from the bus B, and writes the data to the queue B 332 . The interface A 310 reads data from the queue B 332 , and writes the data to the bus A, thus to write the data to the memory 340 . In some cases, few masters are on the bus B, but more masters are on the bus A. For providing best data throughput, the interface B 320 is not expected to be idle, that is data transactions is expected to be performed continuously. In other words, the queue A 331 is not expected to be empty for DMA out transactions, and the queue B 332 is not expected to be full for DMA in transactions.

Usually, the DMA unit 300 uses a “Scatter-Gather” mechanism to reduce the copy steps for data. The “Scatter-Gather” mechanism allows the DMA unit 300 to transfer data to several memory blocks defined by corresponding PRD entries in one data transaction. In other words, the DMA unit 300 can first collect several DMA requests, and then perform DMA transactions correspondingly. In the “Scatter-Gather” mechanism, a PRD pre-fetch mechanism can improve the performance and throughput for the DMA unit. FIG. 2 is a flowchart of a PRD pre-fetch method for a conventional DMA unit. First, in step S 210 , the interface A 310 reads a PRD entry from a PRD table, and stores the PRD entry to the cache memory 350 of the DMA unit. In step S 220 , it is determined whether the PRD table is at the end thereof, that is, to determine whether the current PRD entry is the last entry in the PRD table. If not, in step S 230 , it is determined whether the cache memory 350 is full. If the cache memory 350 is not full (No in step S 230 ), the procedure returns to step S 210 , continuing to read another PRD entry from the PRD table, and store the PRD entry to the cache memory 350 . If the current PRD entry is the last entry in the PRD table (Yes in step S 220 ) or the cache memory 350 is full (Yes in step S 230 ), in step S 240 , a PRD entry is read from the cache memory 350 , and in step S 250 , a data transaction is performed according to the PRD entry. In step S 260 , it is determined whether the cache memory 350 is empty. If the cache memory 350 is not empty (No in step S 260 ), the procedure returns to step S 240 to read another PRD entry from the cache memory 350 , and in step S 250 , perform a data transaction accordingly. If the cache memory 350 is empty (Yes in step S 260 ), in step S 270 , it is determined whether the PRD table is at the end thereof. If not, the procedure returns to step S 210 to read a PRD entry from the PRD table. If so, the procedure is complete.

For the above described conventional mechanism, when the DMA unit 300 is triggered to start, the DMA unit 300 will pre-fetch PRD entries from the PRD table and store the pre-fetched PRD entries to the cache memory 350 of the DMA unit 300 until the cache memory 350 is full. If the cache memory 350 is full, the DMA unit 300 performs DMA transactions according to the PRD entries within the cache memory 350 until the last PRD entry in the cache memory 350 is finished. Then, the DMA unit 300 will loop back to pre-fetch PRD entries until the last PRD entry in the PRD table is pre-fetched.

For a DMA out transaction in FIG. 2 , it is likely that the queue A 331 will go empty when the DMA unit 300 pre-fetches the PRD entries. For a DMA in transaction, it is likely that the queue B 332 will go full when the DMA unit 300 pre-fetches the PRD entries. Specifically, in the DMA out transaction, the DMA transaction corresponding to the last PRD entry is finished on the bus A, while the next DMA transaction will not begin until the fetching of the PRD entries is finished. If the fetching of the PRD entries takes too much time and the interface B continues to read data from the queue A 331 , the queue A 331 may be underflow. In the DMA in transaction, the DMA transaction corresponding to the last PRD entry is finished on the bus A, while the next DMA transaction will not begin until the fetching of the PRD entries is finished. If the fetching of the PRD entries takes too much time and the interface B 320 continues to write data to the queue B 332 , the queue B 332 may be overflow. The above two cases cause the interface B 320 to go into an idle state to prevent the queue A 331 in underflow or queue B 332 in overflow, eventually leading the performance of data throughput of the DMA unit 300 to descend.

PRD pre-fetch methods for DMA units are provided.

In an embodiment of a PRD pre-fetch method for a DMA unit, when a DMA request is received, a data storage state of a queue is detected to determine whether to pre-fetch a PRD entry, wherein the queue is used to store data corresponding to the DMA request. If it is determined to pre-fetch the PRD entry, at least one PRD entry is read from a PRD table, and stored to a cache memory. If it is determined not to pre-fetch the PRD entry, a PRD entry is read from the cache memory, and a DMA transaction is performed according to the PRD entry.

In an embodiment of a PRD pre-fetch method for a DMA unit, when a DMA out transaction for a memory is performed, it is determined whether a first queue has a first predefined size of available space, wherein the first queue is used to store data corresponding to the DMA out transaction. If the first queue does not have the first predefined size of available space, at least one PRD entry is read from a first PRD table, and stored to a first cache memory. When a DMA in transaction for the memory is performed, it is determined whether a second queue has a second predefined size of data, wherein the second queue is used to store data corresponding to the DMA in transaction. If the second queue does not have the second predefined size of data, at least one PRD entry is read from a second PRD table, and stored to a second cache memory.

In an embodiment of a PRD pre-fetch method for a DMA unit, when a DMA out transaction for a memory is performed, it is determined whether a first queue is full or nearly full, wherein the first queue is used to store data corresponding to the DMA out transaction. If the first queue is full or nearly full, at least one PRD entry is read from a first PRD table, and stored to a first cache memory. When a DMA in transaction for the memory is performed, it is determined whether a second queue is empty or nearly empty, wherein the second queue is used to store data corresponding to the DMA in transaction. If the second queue is empty or nearly empty, at least one PRD entry is read from a second PRD table, and stored to a second cache memory.

PRD pre-fetch methods for DMA units may take the form of a program code embodied in a tangible media. When the program code is loaded into and executed by a machine, the machine becomes an apparatus for practicing the disclosed method.

The invention will become more fully understood by referring to the following detailed description with reference to the accompanying drawings, wherein:

FIG. 1 is a schematic diagram illustrating a conventional DMA unit;

FIG. 2 is a flowchart of a PRD pre-fetch method for a conventional DMA unit;

FIG. 3 is a flowchart of an embodiment of a PRD pre-fetch method for a DMA unit according to the invention;

FIG. 4 is a flowchart of an embodiment of a PRD pre-fetch method for a DMA unit for DMA out transaction according to the invention; and

FIG. 5 is a flowchart of an embodiment of a PRD pre-fetch method for a DMA unit for DMA in transaction according to the invention.

PRD pre-fetch methods for DMA units are provided. The PRD pre-fetch methods are suitable for use in a DMA unit comprising similar architecture as shown in FIG. 1 . The DMA unit comprises a first interface (interface A) and a second interface (interface B) used to access a first bus (bus A) and a second bus (bus B), respectively. The DMA unit couples to a memory via the first bus. The DMA unit comprises a first queue (queue A) and a second queue (queue B). For DMA out transactions, the first interface reads data from the memory via the first bus, and stores the data to the first queue. The second interface reads data from the first queue, and writes the data to the second bus. For DMA in transactions, the second interface reads data from the second bus, and writes the data to the second queue. The first interface reads data from the second queue, and writes the data to the first bus, thus to write the data to the memory. It is noted that the DMA unit comprises at least one cache memory for storing PRD entries pre-fetched from at least one PRD table.

FIG. 3 is a flowchart of an embodiment of a PRD pre-fetch method for a DMA unit according to the invention.

In step S 410 , it is determined whether the DMA unit performs a DMA out transaction or DMA in transaction, that is, to determine whether a received DMA request is a DMA out transaction or a DMA in transaction. If a DMA out transaction is performed, in step S 420 , it is determined whether the first queue is full or nearly full. In other words, it is determined whether to pre-fetch a PRD entry by detecting the data storage state of the first queue. If the first queue is full or nearly full (Yes in step S 420 ), representing allowance of the fetching of PRD entry, in step S 430 , at least one PRD entry is read from a first PRD table, and stored to a first cache memory. If the first queue is not full or nearly full (No in step S 420 ), in step S 440 , a PRD entry is obtained from the first cache memory, and a DMA transaction is performed according to the obtained PRD entry. The procedure then returns to step S 420 . If a DMA in transaction is performed, in step S 450 , it is determined whether the second queue is empty or nearly empty. In other words, it is determined whether to pre-fetch a PRD entry by detecting the data storage state of the second queue. If the second queue is not empty or nearly empty (No in step S 450 ), in step S 460 , a PRD entry is obtained from a second cache memory, and a DMA transaction is performed according to the obtained PRD entry. The procedure then returns to step S 450 . If the second queue is empty or nearly empty (Yes in step S 450 ), in step S 470 , at least one PRD entry is read from a second PRD table, and stored to the second cache memory. The procedure then returns to step S 450 . It is understood that the first cache memory and the second cache memory may be parts of the cache memory of the DMA unit in some embodiments.

Details for DMA out transaction and DMA in transaction are discussed follow.

FIG. 4 is a flowchart of an embodiment of a PRD pre-fetch method for a DMA unit for DMA out transaction according to the invention.

When the DMA unit performs a DMA out transaction, in step S 502 , a PRD entry is read from a PRD table, and the fetched entry is stored to the cache memory. In step S 504 , it is determined whether the first queue has a first predefined size of available space. It is noted that step S 504 is used to determine whether the first queue is full or nearly full. The first predefined size can be set as the burst length supported by the second interface, or any other value according to various requirements. Usually, the burst length supported by the second interface is greater than the size of memory block corresponding to a single PRD entry. If the first queue has the first predefined size of available space (Yes in step S 504 ), the procedure goes to step S 510 . If the first queue does not have the first predefined size of available space (No in step S 504 ), in step S 506 , it is determined whether the fetched PRD entry in step S 502 is the last entry of the PRD table. If so (Yes in step S 506 ), the procedure goes to step S 510 . If not (No in step S 506 ), in step S 508 , it is determined whether the cache memory is full. If the cache memory is not full (No in step S 508 ), the procedure returns to step S 502 . If the cache memory is full (Yes in step S 508 ), in step S 510 , it is determined whether a DMA transaction corresponding to a current PRD entry is finished, that is, to determine whether the first interface has read the data in the corresponding memory block according to the current PRD entry. If the DMA transaction corresponding to the current PRD entry is not finished (No in step S 510 ), the procedure goes to step S 518 . If the DMA transaction corresponding to the current PRD entry is finished (Yes in step S 510 ), in step S 512 , it is determined whether the current PRD entry is the last entry of the PRD table. If so (Yes in step S 512 ), the procedure is complete. If the current PRD entry is not the last entry of the PRD table (No in step S 512 ), in step S 514 , it is determined whether the cache memory is empty. If the cache memory is empty (Yes in step S 514 ), the procedure returns to step S 502 . If the cache memory is not empty (No in step S 514 ), in step S 516 , a PRD entry is popped from the cache memory. Then, in step S 518 , it is determined whether the first queue has the first predefined size of available space. If the first queue has the first predefined size of available space (Yes in step S 518 ), in step S 520 , a DMA transaction is performed according to the popped PRD entry, and the procedure returns to step S 510 . If the first queue does not have the first predefined size of available space (No in step S 518 ), in step S 522 , it is determined whether the cache memory is full. If the cache memory is not full (No in step S 522 ), the procedure returns to step S 502 . If the cache memory is full (Yes in step S 522 ), in step S 524 , it is determined whether the first queue has the first predefined size of available space. If the first queue does not have the first predefined size of available space (No in step S 524 ), step S 524 repeats until the first queue has the first predefined size of available space, that is, to wait till the second interface transfers the data on the first queue to the second bus. If the first queue has the first predefined size of available space (Yes in step S 524 ), in step S 520 , a DMA transaction is performed according to the popped PRD entry, that is, the first interface reads the data in the memory block corresponding to the popped entry via the first bus. The procedure then returns to step S 510 . At this time, the current PRD entry in step S 510 becomes the popped PRD entry from the cache memory in step S 516 . It is understood that, if no PRD entry is popped from the cache memory to become a new current PRD entry when step S 510 performs, it is determined the DMA transaction corresponding to the current PRD entry is finished. If no new current PRD entry is obtained when step S 512 executes, it is determined the current PRD entry is not the last PRD entry in the PRD table.

FIG. 5 is a flowchart of an embodiment of a PRD pre-fetch method for a DMA unit for DMA in transaction according to the invention.

When the DMA unit performs a DMA in transaction, in step S 602 , a PRD entry is read from a PRD table, and the fetched entry is stored to the cache memory. In step S 604 , it is determined whether the second queue has a second predefined size of data. It is noted that step S 604 is used to determine whether the second queue is empty or nearly empty. The second predefined size can be set as the burst length supported by the first interface, or any other value according to various requirements. Usually, the burst length supported by the first interface is greater than the size of memory block corresponding to single PRD entry. If the second queue has the second predefined size of data (Yes in step S 604 ), the procedure goes to step S 610 . If the second queue does not have the second predefined size of data (No in step S 604 ), in step S 606 , it is determined whether the fetched PRD entry is the last entry of the PRD table. If so (Yes in step S 606 ), the procedure goes to step S 610 . If not (No in step S 606 ), in step S 608 , it is determined whether the cache memory is full. If the cache memory is not full (No in step S 608 ), the procedure returns to step S 602 . If the cache memory is full (Yes in step S 608 ), in step S 610 , it is determined whether a DMA transaction corresponding to a current PRD entry is finished, that is, to determine whether the first interface has written data to the corresponding memory block according to the current PRD entry. If the DMA transaction corresponding to the current PRD entry is not finished (No in step S 610 ), the procedure goes to step S 618 . If the DMA transaction corresponding to the current PRD entry is finished (Yes in step S 610 ), in step S 612 , it is determined whether the current PRD entry is the last entry of the PRD table. If so (Yes in step S 612 ), representing no more PRD entry can be read from the PRD table, the procedure is complete. If the current PRD entry is not the last entry of the PRD table (No in step S 612 ), in step S 614 , it is determined whether the cache memory is empty. If the cache memory is empty (Yes in step S 614 ), the procedure returns to step S 602 , and another PRD entry is read. If the cache memory is not empty (No in step S 614 ), in step S 616 , a PRD entry is popped from the cache memory. Then, in step S 618 , it is determined whether the second queue has the second predefined size of data. If the second queue has the second predefined size of data (Yes in step S 618 ), in step S 620 , a DMA transaction is performed according to the popped PRD entry, that is the first interface writes the data from the second queue to the memory block corresponding to the popped PRD entry via the first bus, and the procedure returns to step S 610 . If the second queue does not have the second predefined size of data (No in step S 618 ), in step S 622 , it is determined whether the cache memory is full. If the cache memory is not full (No in step S 622 ), the procedure returns to step S 602 . If the cache memory is full (Yes in step S 622 ), in step S 624 , it is determined whether the second queue has the second predefined size of data. If the second queue does not have the second predefined size of data (No in step S 624 ), step S 624 repeats until the second queue has the second predefined size of data, that is, to wait till the second interface stores the data transferred by the second bus to the second queue. If the second queue has the second predefined size of data (Yes in step S 624 ), in step S 620 , a DMA transaction is performed according to the popped PRD entry, and the procedure returns to step S 610 .

In this embodiment, for DMA out transactions, the pre-fetching of PRD entries is performed when the first queue is full or nearly full. For DMA in transactions, the pre-fetching of PRD entries is performed when the second queue is empty or nearly empty. The PRD pre-fetch methods of the embodiments change the timing for pre-fetching PRD entries to prevent the queues of the DMA unit in underflow or overflow, improving the performance of data throughput of the DMA unit.

PRD pre-fetch methods for DMA units, or certain aspects or portions thereof, may take the form of a program code (i.e., executable instructions) embodied in tangible media, such as products, floppy diskettes, CD-ROMS, hard drives, or any other machine-readable storage medium, wherein, when the program code is loaded into and executed by a machine, such as a computer, the machine thereby becomes an apparatus for practicing the methods. The methods may also be embodied in the form of a program code transmitted over some transmission medium, such as electrical wiring or cabling, through fiber optics, or via any other form of transmission, wherein, when the program code is received and loaded into and executed by a machine, such as a computer, the machine becomes an apparatus for practicing the disclosed methods. When implemented on a general-purpose processor, the program code combines with the processor to provide a unique apparatus that operates analogously to application specific logic circuits.

While the invention has been described by way of example and in terms of preferred embodiment, it is to be understood that the invention is not limited thereto. Those who are skilled in this technology can still make various alterations and modifications without departing from the scope and spirit of this invention. Therefore, the scope of the present invention shall be defined and protected by the following claims and their equivalents.