JP2014167818A - Data transfer device and data transfer method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform a determination of writing in a memory even if connected by a plurality of paths.SOLUTION: A data transfer device comprises: an acquisition part that acquires a transfer source address of a first memory serving as a transfer source of data, a transfer destination address of a second memory serving as a transfer destination of the data, and a transfer length of the data from a storage part; a reading part that acquires data on the first memory on the basis of the transfer source address; and a writing part that transfers the acquired data to a memory controller writing data in the second memory on the basis of the transfer destination address, and after the transfer is ended, determines whether a volume of the transferred data reaches the transfer length of the data, and as a result of the determination, when the volume of the transferred data reaches the transfer length of the data, transmits a reading request to the memory controller.

Description

  The present invention relates to a data transfer apparatus and a data transfer method.

  Recent storage devices (disk array systems and similar storage devices) are equipped with a cache memory in the storage controller in order to increase the speed of reading and writing data. The cache memory holds a part of data to be written to a final storage medium such as a hard disk drive (Hard Disk Drive, hereinafter referred to as “HDD”) or a solid state disk (hereinafter referred to as “SSD”) or frequently read data. This hides the media access latency and realizes a high-speed response to the host.

  For example, when the host writes data to the storage device, the storage controller in the storage device that processes the data from the host returns a write completion response to the host when the data is stored in the cache memory. Since the cache memory is much faster than the final storage medium, the host can immediately issue the next I / O request. On the other hand, the storage controller operates to write the data in the cache memory to the final storage medium at a certain timing asynchronously with the operation of the host.

  Storage controllers are often composed of a plurality of unit controllers in order to improve fault tolerance. For this reason, cache memories (hereinafter simply referred to as “memory”) are also distributed in a plurality of locations. Since data transfer is often performed between memories for processing efficiency and redundancy by duplication, a network path using internal bus technology is configured between unit controllers. For example, PCI Express (registered trademark) is used for the bus. Transfer between cache memories is performed using a data transfer controller called DMA (Direct Memory Access) according to instructions from a processor in which storage control software operates.

  In the case of PCI Express, a write request to the memory is a so-called posted access in which there is no completion response (Completion) from the target. In the memory controller of the storage controller, a buffer may be provided on the path from the network path to the memory controller for reasons of hardware data processing. Therefore, the write request source cannot confirm whether the data has been written to the memory through all the buffers on the path. For this reason, when writing to the memory via a network path or the like, the storage apparatus or the like issues a dummy memory read request on the same path in order to flush the buffer in the middle and to confirm the writing to the memory. In PCI Express, the memory read request is called non-posted access. Non-posted access on the same route does not reach the memory by overtaking posted access such as previously issued memory write requests, so if a response to this read request is returned, it will be issued first It can be assured that the write data related to the memory write request has been written to the final memory through the buffer.

  By the way, since the transfer bandwidth that can be realized by one path is limited, a plurality of network paths may be extended in order to secure the bandwidth. In order to efficiently transfer data between a plurality of memories, it is necessary to use this network path without any bias. As a method of eliminating the bias and using the path uniformly, a technique for distributing the load on the network path has been conventionally discussed.

  For example, Patent Document 1 discloses a method of calculating a hash value of a packet to be transferred and selecting a route according to the hash value. According to this method, it is possible to send a packet by distributing the load to a plurality of paths, so that network transfer can be efficiently performed.

JP 2008-263436 A

  However, when this method is applied to a storage apparatus that transfers data between memories between unit controllers connected by a bus such as PCI Express, and a similar data transfer system, several problems arise.

  The first problem is the dummy read problem described above. With PCI Express and other buses, it is guaranteed that non-posted access will not overtake posted access on the same route, but if it passes through a different route, it will be ahead of the posted access sent earlier. Non-posted accesses can reach the final memory. The purpose of the dummy read is to confirm the writing confirmation by using the fact that it does not overtake the previous write request, so it must be issued through the same route as the write request. However, since the method described in Patent Document 1 distributes the path according to the hash value of the packet, there is a problem that both request packets do not always pass through the same route. Since both packets are obviously different in format, the possibility of having the same hash value is extremely low. In the first place, there is no mechanism for issuing a dummy read when a write request is issued. Further, as a modification, Patent Document 1 describes that, instead of a hash value, another predetermined identifier may be used. However, which write packet is a dummy read packet and which route the write packet has taken You have to store a memory table. The number of packets flowing in memory-to-memory data transfer is enormous, hundreds of thousands to millions per second, and it takes time to search in the first place. In addition, such a process has a purpose different from that of load balancing which Patent Document 1 intends to solve in the first place.

  The second problem is that the transfer time greatly depends on the concentration of access to the destination memory. In Patent Document 1, it is assumed that the destination is a single device although the load is distributed by a plurality of paths (routes), and that data transfer is only performed through a network. This is also clear from the fact that a method for distributing the load by managing the degree of congestion of the route itself (usage status of the data buffer) is disclosed. On the other hand, in the case of this storage apparatus, access to a certain memory does not necessarily pass through the network path, but access by the same DMA in the storage controller as that memory also competes. Therefore, even if only the degree of congestion of paths (routes) is managed, the load cannot always be distributed significantly.

  The third problem is that the degree of memory access congestion is also caused by the memory allocation method (memory usage layout) of the control software. For example, if memory channel 1 is allocated a control information area with relatively few accesses, and memory channel 2 is allocated a large number of cache areas that are frequently accessed, packets are not allowed to pass through the network path at an equal rate. Access to the memory channel 2 is blocked, and as a result, subsequent packets of the memory channel 1 passing through the network path also stay. This problem cannot be solved only by a conventional network load distribution technique as shown in Patent Document 1.

The present invention has been made in view of the above problems.
That is, an object of the present invention is to provide a data transfer device for confirming writing to a memory without storing route information even without being connected through a plurality of paths and without requiring new intervention of control software, and It is to provide a data transfer method.

  Another object of the present invention is to provide a data transfer apparatus and a data transfer method that efficiently transfer without causing a bias in a path even when access to a specific memory from inside or outside a storage controller is concentrated. .

  Another object of the present invention is to provide a data transfer apparatus and a data transfer method that perform efficient transfer based on memory use information of higher-level control software.

  In the present invention, in order to solve at least one of the above problems, a transfer source address of a first memory as a data transfer source, a transfer destination address of a second memory as a data transfer destination, An acquisition unit that acquires a transfer length from a storage unit, a read unit that acquires data of the first memory based on the transfer source address, and a data transfer to the second memory based on the transfer destination address. The acquired data is transferred to a memory controller that performs writing, and when the transfer is completed, it is determined whether or not the amount of the transferred data has reached the transfer length of the data. A data transfer device is provided that includes a writing unit that transmits a read request to the memory controller when the amount of data reaches the data transfer length.

  According to the present invention, the internal bus technology is used in which a write response is not made even if the hardware does not store the write request issuance path for each packet or the control software does not perform new control for write confirmation. The write to the memory can be confirmed on the network path.

  Further, according to the present invention, when there are a plurality of network paths, it is possible to select a path in consideration of the congestion status of the destination memory, and efficient transfer is possible.

  Further, according to the present invention, even with the same hardware, efficient data transfer can be performed according to the characteristics of the control software operating on the hardware.

1 is a block diagram showing a configuration of a storage device in an embodiment of the present invention. 1 is an external view showing a configuration of a storage device in an embodiment of the present invention. 2 is a block diagram illustrating a configuration of a storage controller of the storage apparatus according to the embodiment of the present invention. FIG. It is a block diagram which shows an example of the internal structure of ASIC mounted in the storage controller in embodiment of this invention. It is the figure which drew typically the starting method of the DMA controller in the execution form of this invention. It is the figure which showed an example of the format of the transfer instruction | indication parameter to the DMA controller in embodiment of this invention. It is a flowchart which shows the operation | movement procedure of DMA in embodiment of this invention. It is a figure which shows an example of the memory address map in embodiment of this invention. It is a ladder chart explaining the whole operation | movement of the data transfer control between memory in embodiment of this invention. It is a figure which shows the structural example of the write confirmation request | requirement in the modification of embodiment of this invention. It is a figure which shows the structural example of the write confirmation response in the modification of embodiment of this invention. It is a flowchart which shows the operation | movement procedure of DMA of the data transfer control between memory in the modification of embodiment of this invention. It is a block diagram which shows the internal structure of ASIC in the modification of embodiment of this invention. It is a block diagram which shows the internal structure of ASIC in Example 2 of this invention. It is a figure which shows the structural example of the table in ASIC in Example 2 of this invention. It is a flowchart explaining operation | movement of DMA in Example 2 of this invention. It is a figure which shows the structural example of the table with which the status block 400 in the modification of Example 2 of this invention is provided. It is the figure which showed an example of the format of the transfer instruction | indication parameter to the DMA controller in Example 3 of this invention. It is a figure which shows the structural example of the memory layout information setting table in Example 3 of this invention. It is the figure which showed typically the exchange of data between DMA and the internal data transfer path arbiter in Example 3 of this invention. It is the flowchart which showed the operation | movement of the internal data transfer path arbiter in Example 3 of this invention.

  The form for implementing this invention is demonstrated using Example 1, 2, and 3. FIG. In addition, the form which can be implemented by the partial change from the said Example is described as a modification within each Example.

<Example 1>
This embodiment will be described with reference to FIGS.
FIG. 1 is a block diagram showing the configuration of the storage system in this embodiment. The storage apparatus 10 includes a plurality of drive housings 51, 52,... And a storage controller housing 50. The storage device 10 is connected to the management terminal 31 through the management LAN 30, and device settings and the like can be executed through the management terminal 31. The management terminal 31 may be included in the storage device 10.

  The storage apparatus 10 is connected to the host computers 20A to 20C through the storage network 40. The host computers 20 </ b> A to 20 </ b> C are loaded with an operating system and applications or virtual machines that operate on the operating system, and issue I / O to the storage apparatus 10.

  A plurality of disk drives D00 to D0N, D10 to D1N are mounted on the drive housings 51, 52,. The disk drives D00 to D1N are HDDs, SSDs, or similar final storage media.

  The disk drives D00 to D1N are connected to the disk through the back-end interconnection mechanism 500 provided in the drive casings 51 and 52, the back-end connection cable 510, and the I / F cards 90A and 91B provided in the storage controller casing 50. It is connected to the array controller main packages 60A and 60B. The back-end interconnection mechanism 500 is, for example, a SAS (Serial Attached SCSI) expander, and the back-end connection cable 510 is, for example, a SAS cable. The I / F cards 90A and 91B are, for example, HBAs (Host Bus Adapters) equipped with a SAS protocol chip. With this connection, the storage controller can send and receive data to and from the disk.

  In the storage controller housing 50, main packages 60A and 60B of the disk array controller and MP (Micro Processor) packages 70A and 70B equipped with processors for controlling them, the drive connections 90A and 90B described above, and the host computers 20A to 20C. An I / F card is provided for connection.

  In the MP packages 70A and 70B, a storage control program is operating and connected to the disk array controller main packages 60A and 60B via control lines 71A to 72B in order to send instructions. On the other hand, the disk array main packages 60A and 60B are connected by a plurality of internal data transfer paths 101 and 102 in order to transfer data to each other.

  In the present embodiment, only two packages, A-system and B-system, are shown. However, three or more packages may be provided and connected to each other, or may be connected to each other through an I / F card. The bodies 50 may be joined together. For example, a part of the I / F cards 80A to 91B may be exchanged for a PCI Express storage controller inter-case connection I / F card. Further, these interconnections may be used like the internal data transfer paths 101 and 102.

  FIG. 2 is an external view shown to assist in understanding the storage apparatus in the present embodiment. The storage controller housing 50 is mounted at the bottom, and the drive housings 51 to 58 are shown thereon. Although not shown in the drawing, the back end connection cable and the cable for configuring the storage network 40 are wired on the back side.

FIG. 3 is a block diagram showing the internal configurations of the disk array controller main packages 60A and 60B and the MP packages 70A and 70B that control the main packages 60A and 60B in this embodiment, and the configurations of the I / F cards 80A to 91B connected thereto. FIG.
Since the A system and the B system are mutually targeted structures, the following description will be made using the A system as a representative.

An ASIC (Application Specific Integrated Circuit) 200A and memories M0, M1, M2, and M3 are mounted on the disk array controller main package 60A. Various functions are mounted on the ASIC 200A, and for example, DMA (Direct Memory Access) 210A to 212A for transferring data between memories is provided. The memories M0 to M3 are cache memories of the storage controller 50, and a DRAM (Dynamic Random Access Memory) or the like that is made nonvolatile by a battery or the like is used. The DRAM is often provided as a DIMM (Dual Inline Memory Module). The disk array controller main package 60A is connected to the I / F card group through the ASIC 200A. FE (Front-End) I / F cards 80A and 81A are used for connection to a host through the storage network 40. For example, a fiber channel protocol chip or the like is mounted and a plurality of host connection ports are provided.
On the other hand, BE (Back-End) I / F cards 90A and 91A are connected to the respective disks of the drive housings 51, 52.
The MP package 70A is provided with a processor chip set 75, processors 760 and 761, and local memories 770 and 771 in which programs running on these processors are stored. Depending on the type and generation of the processor, the local memories 770 and 771 may be connected to the chipset 75, or may be directly connected to the processors 760 and 761. Further, the chip set 75 may be included in the processors 760 and 761. The processors 760 and 761 are connected to the ASICs 200A and 200B of the main packages 60A and 60B through the control lines 71A and 71B, and control the hardware. An example of a technique used for the control lines 71A and 71B is PCI Express.

  FIG. 4 shows the configuration of the functional blocks inside the ASICs 200A and 200B described above. In the ASICs 200A and 200B, the DMAs 210, 211, and 212 described above, PCI Express cores 221 and 223 that control PCI Express used as an internal data transfer path (I-Path: Internal Path), and internal data transfer paths 101, 102 internal data transfer path arbiters 220, 222, memory controllers 230, 231, 232, 233, I / F control blocks 240, 241, 242, 243, processor ports 271, 272, a routing processor (RP) 260, 261, there is a switch block 250 that interconnects them.

  DMAs 210 to 212 (inter-memory data transfer control units) are transfer control blocks having a function of receiving an instruction from a processor and copying data for a specified transfer length from a certain memory address to another memory address. The DMAs 210 to 212 include a reading unit (RX) 2101 for reading data from the transfer source memory, a writing unit (TX) 2102 for writing data to the transfer destination memory, and a parameter fetch unit (PF) 2103 for fetching and analyzing transfer parameters. Consists of including. An instruction method related to data transfer and a transfer process activation method will be described later. In this embodiment, three DMAs are provided, but the number of DMAs may be two or less or four or more.

  The internal data transfer paths 101 and 102 are networks for transferring data between memories across packages, and PCI Express is used as described above. The PCI Express cores 221 and 223 are control blocks that perform basic transmission and reception operations of PCI Express, and perform hardware control, data link control, transaction protocol control, and the like of the PCI Express bus. The internal data transfer path arbiters 220 and 222 are mediation control blocks that arbitrate PCI Express network path use requests from a plurality of control blocks and use them by giving use authority according to a predetermined rule. For example, when the DMA 210 and the DMA 211 try to transfer data at the same time, arbitration processing is performed such that the DMA 210 is first given a usage right to perform transfer, and then the DMA 211 is given a right to use and perform transfer.

  The memory controllers 230 to 233 are connected to the memories M0 to M3 or M4 to M7, and based on requests from the DMAs 210 to 212 and other control blocks, the memories M0, M1, M2, M3 and the memories M4, M5, M6, M7. This is a control block that reads and writes data and transmits and receives data packets. The memory controllers 230 to 233 include an arbiter (ARB) for arbitrating the access order when accessed simultaneously from a plurality of control blocks. The memory controllers 230 to 233 are also called memory channels. A unique memory address space is allocated to the memories M0 to M7 connected to the memory controllers 230 to 233.

  The I / F control blocks 240 to 243 are blocks for connecting the previous I / F cards 80A to 81B and 90A to 91B and performing data transfer therewith. The processor ports 271 and 272 are connected to control lines 71A and 72A, and the processors 760 and 761 and the ASIC 200 exchange control information data, and the processors 760A, 761A, 760B, and 761B communicate with each other. It is a block that performs.

  The routing processors 260 and 261 are control blocks that perform distribution processing to the processors 760A to 761B in charge of controlling commands received from the protocol chip through the I / F control blocks 240 to 243.

  The ASIC 200 may have various other control blocks. For example, a hardware arithmetic control block such as data compression and data encryption, a controller dedicated to memory access from a processor, and the like can be considered.

  FIG. 5 is a diagram schematically showing a mechanism for the processors 760 and 761 to transfer instructions to the DMAs 210 to 213 in the ASIC 200 according to the present embodiment. In the following description, the procedure is the same regardless of which processor activates any DMA, and the case of the DMA 210 and the processor 760 will be described as a representative.

  When the processor 760 uses the DMA 210 to transfer data at a certain memory address to another memory address, the processor 760 builds a parameter 300 (transfer instruction parameter) on the local memory 770 for instructing the DMA 210 to transfer. The parameter 300 includes a transfer source memory address, a transfer destination memory address, a transfer length, and the like. Details will be described later. A plurality of parameters 300 are written in the local memory 770 continuously.

  The processor 760 continuously writes the parameter 300 to the local memory 770 as many as required for the target data transfer, and sets the start address of the local memory 770 in which the first parameter 300 is stored in the start trigger register 2 of the DMA 210. Set as. For example, in FIG. 5, since four parameters, parameter 0 to parameter 3, are constructed from the address 1000h address (hexadecimal) of the local memory 770, the processor 760 sets the value 1000h in the start trigger register 2 of the DMA 210.

  When a value is set in the register 2, the DMA 210 reads and analyzes the parameter 300 from the address of the local memory 770 indicated by the value, and performs data transfer designated by the parameter 300. For example, in FIG. 5, since the value 1000h (hexadecimal) is set in the activation trigger register 2, the DMA 210 operates to read the parameter 300 from the address 1000h of the local memory 770. Details of the operation of the DMA 210 will be described later.

  FIG. 6 is a diagram illustrating a format example of the parameter 300 used by the processor 760 to issue a data transfer instruction to the DMA 210 in the present embodiment. In the parameter ID field 3001, a unique number set for each parameter 300 is designated. In the transfer source address field 3002, the start address of the memory serving as the transfer source is designated. In the transfer destination address field 3003, the start address of the memory to be the transfer destination is designated. The transfer length field 3004 specifies the data length of the data to be transferred. A chain number field 3005 is a field for specifying grouping of parameters 300. The chain number field 3005 is designated in order from 1 when there are other parameters 300 to be grouped, and 0 is designated as the last parameter 300 of the grouping. That is, the chain number 3005 increases in order from 1, and the parameter group up to the chain number 3005 is regarded as one transfer group (the chain number 3005 is 0 when there is no chain). When the chain number 3005 is other than 0, the DMA 210 does not end the processing of the parameter 300, and performs a chain operation for continuously performing data transfer instructed to the subsequent parameter 300 group. In the path number field 3006, it is specified which number of the plurality of internal data transfer paths 101 and 102 is to be used. In FIG. 6, the number of the path number field 3006 is one, but a plurality of internal data transfer paths 101 and 102 may be provided so that they can be individually specified at the time of data reading or data writing. Note that the configuration of the parameter 300 shown in FIG. 6 is an example, and there may be another field necessary for other control.

  FIG. 7 is a flowchart showing the data transfer operation of the DMA 210 in this embodiment. The DMA transfer operation of the present embodiment, in which writing is automatically determined every time the writing destination memory surface is switched and every time a series of data transfer is completed, will be described with reference to FIG.

  The DMA 210 starts its operation in response to a start instruction from the processor 760. In this embodiment, the operation is started when a value is set in the trigger register 2 from the processor 760, but the DMA 210 may be activated in response to signal transmission or message transmission via the signal line. When a value is set in the activation trigger register 2, the parameter fetch unit 2103 of the DMA 210 refers to the address of the local memory 770 indicated by the set value, fetches one of the parameters 300, and analyzes the setting content (S8000). ). For example, the start address of the transfer source memory is acquired from the transfer source address field 3002, the start address of the transfer destination memory is acquired from the transfer destination address field 3003, and the transfer length (total data transfer amount) is acquired from the transfer length field 3004. Once the DMA 210 is activated, the next fetch of the parameter 300 is prefetched (prefetched) in the background (S8005).

  Based on the analysis result of the parameter fetch unit 2103, the reading unit 2101 of the DMA 210 specifies which memory controller to access from the transfer source address (S8010). For example, assume that memory addresses are allocated as shown in FIG. When the transfer source address is 23000, since the address is between addresses 20000h-2FFFFh of the memory M1, the reading unit 2101 determines that it is only necessary to access the memory controller 231 to which the memory M1 is connected. Subsequently, the reading unit 2101 issues a memory read request designating the address to be accessed and the data length to be read to the corresponding memory controller 231 (S8020). The memory read request is issued directly to the memory controller when the target memory controller is in the same ASIC 200, and is issued through the internal data transfer paths 101 and 102 when the target memory controller is another ASIC 200. The transfer length that the DMA 210 can transfer at one time is relatively small. For this reason, the data of the transfer length specified by the parameter 300 may be transferred in multiple times. When the transfer length specified by the parameter 300 is larger than the transfer length that the DMA 210 can transfer at one time, the maximum transfer length that can be transferred by the DMA 210 is specified as the data length. When the transfer length specified in the parameter 300 is less than or equal to the transfer length that can be transferred at one time by the DMA 210, the transfer length specified in the parameter 300 is specified in the read request as the data length. For example, if the transfer length that the DMA 210 can transfer at one time is 2048 bytes, 2048 bytes if the transfer length specified in the parameter is 4096 bytes, 512 bytes if 512 bytes are specified in the read request. Will be.

  The memory controllers 230 to 233 read from the memories M0 to M7 the data having the transfer length specified in the read request specified by the address specified in the read request, and return a read response (Completion) together with the read data to the reading unit 2101. . The read response is directly transmitted in the case of the DMA 210 in the same ASIC 200, and is transmitted through the internal data transfer paths 101 and 102 in the case of another ASIC 200. The reading unit 2101 of the DMA 210 receives a read response together with the data read from the memory controllers 230 to 233 (S8030), and stores the received data in a buffer inside the DMA 210.

  When the data is stored in the buffer in the DMA 210, the writing unit 2102 of the DMA 210 specifies which memory controller 230 to 233 should be accessed from the transfer destination address based on the analysis result of the parameter fetch unit 2103 (S8040). For example, as shown in FIG. 8, when a memory address is allocated and the transfer destination address is 51000h, the writing unit 2102 determines that it is only necessary to access the memory controller 230 to which the memory M4 is connected.

  If the memory controller identified as the access destination differs from the memory controller identified in the previous S8040 for the second and subsequent parameters 300 (S8050: Yes), the writing unit 2102 A dummy read request is issued (S8060). Dummy read requests that do not overtake previously issued memory write requests, such as memory read requests with a transfer length of 0 (ie, all previously issued memory write requests have been written to memory) Requests to be processed later). The dummy read request is issued directly to the memory controller when the target memory controller to which the dummy read request is transmitted is in the same ASIC 200, and the internal data transfer paths 101 and 102 when the target memory controller is another ASIC 200. Forwarded through. If there is no portion (buffer or the like) where data stays on the path from the DMA 210 to the memory controllers 230 to 233 in the same ASIC 200, the memory controller specified in this S8040 is the memory specified in the previous S8040. Even if it is different from the controller (S8050: Yes), if the previous memory controller is in the same ASIC 200, it may be mounted so as not to issue a dummy read request. When the read response to the dummy read request is returned, the writing unit 2102 determines whether there is an error (S8070). If there is no error (S8070: No), the writing unit 2102 issues a memory write request to the new access destination memory controller together with the data stored in the buffer in the DMA 210 (S8080).

In S8050, when the access destination memory controller is the same as the previous one (S8050: No) and when the parameter 300 read in S8000 is the first one, it is not necessary to perform the processing of S8060 to S8070, and only S8080 Is executed.
In S8070, if there is an error in the memory read response (S8070: Yes), the writing unit 2102, the reading unit 2101, and the parameter fetching unit 2103 have the parameter 300 and the subsequent parameters chained to the parameter 300 (that is, All the data transfer processing is interrupted for a series of parameters 300 in which the value set in the chain number column 3005 is continuously incremented from the parameter 300 (S8300), and the status of the end due to the error is set to the processor 760, 761 is notified.

  As described above, since the transfer length that the DMA 210 can transfer at one time is often small, the writing unit 2102 accumulates the previous transfer amount and the current transfer amount when the transfer ends, and the transfer specified in the parameter. It is determined whether it has been achieved for a long time (S8090). When the designated transfer length has been reached (S8090: Yes), the writing unit 2102 sends a dummy read request to the memory controller finally specified in S8040 to write the data so far into the memory. Confirm (S8100). The writing unit 2102 receives a read response from the memory controller in response to the dummy read request, determines whether there is an error (S8110), and if there is an error, proceeds to S8300 and interrupts the transfer as described above. The reading unit 2101 and writing unit 2102 notify the parameter fetch unit 2103 that the designated transfer length has been reached and waits for the next instruction. The parameter fetch unit 2103 determines whether the next parameter is chained (S8120). If the next parameter is not chained (S8120: No), the parameter fetch unit 2103 notifies the processor 760, 761 of the transfer end status (S8130), and ends the data transfer process.

  In S8090, when the cumulative transfer amount does not reach the transfer length specified in the parameter 300 (S8090: No), the writing unit 2102 and the reading unit 2101 reduce the remaining transfer length by the current transfer (S8200). The original address and the transfer destination address are shifted by the current transfer length (S8210) and repeated until the transfer length specified in the parameter 300 is reached (the process returns to S8010).

In S8100, when the parameter 300 is chained (S8100: Yes), the data transfer process is performed using the next parameter 300 (return to S8000).
As described above, the DMA 210 of this embodiment performs data transfer processing.

  In the actual control implementation, the parameter fetch unit 2103, the reading unit 2101, and the writing unit 2102 can operate simultaneously in parallel, and therefore each process (S8000 to S8130) may be performed in parallel in time. In this embodiment, the parameter fetch unit 2103, the reading unit 2101, and the writing unit 2102 are provided and the roles are shared as described above. However, other configurations may be used as long as the DMA performs the same function.

  FIG. 9 is a ladder chart showing an overall image of the inter-memory data transfer control in this embodiment, including the DMA transfer control of FIG. The vertical direction in the figure shows the passage of time, and the horizontal direction in the figure shows the control and data flow between resources. The respective resources are local memories 770 and 771, processors 760 and 761, memories M0 to M7, and DMAs 210 to 213. Since the operations are the same if the resources are the same, the numbers are omitted. Note that memory #x indicates a transfer source memory, and memory # 0 and memory # 1 indicate transfer destination memories, each of which is an arbitrary one of memory controllers 230 to 233 to which M0 to M7 are connected. The transfer source and the transfer destination may overlap. An overview of data transfer control will be described with reference to FIG.

  Based on the parameter format shown in FIG. 6, the processors 760 and 761 create a group of parameters 300 on the local memory by the method shown in FIG. 5 (R7000). The processors 760 and 761 subsequently set a trigger register 2 by setting the address of the local memory that created the DMA parameter 300 in the trigger register 2 (R7100). The DMA refers to the trigger register 2 and fetches the first parameter 300 of the group of parameters 300 from the local memory (R7200). Further, the next parameter 300 is prefetched at an appropriate timing as a background task (R7210). The DMA analyzes the parameter 300, reads out data from the memory #n based on the result (R7300 to R73xx), and writes data to the memory # 1 in parallel (R7400 to R74xx). When the write destination memory surface is switched from the memory # 0 to the memory # 1, the DMA issues a dummy read request for confirming the previous writing (R7500). Then, while continuing the data read (R7600 to R76xx), the data writing to the memory # 1 is continued (R7700 to R77xx). When the data transfer for the transfer length specified in the parameter 300 is completed, a dummy read request is issued to the final memory surface to perform a write confirmation process (R7800). If the subsequent parameter 300 is not chained (the chain number 3005 of the next parameter 300 is 0), the transfer status is returned to the processors 760 and 761 through the local memory (R7900).

(Effect of the first embodiment)
According to the first embodiment, write requests to the same memory controllers 230 to 233 are continuously transmitted without confirming writing, and when write requests are transmitted to memory controllers 230 to 233 having different write destinations. Can send a dummy read request to the previous memory controllers 230-233. Since requests to the same memory controller 230 to 233 are transmitted through the same path, a write request issued by a non-posted access is transmitted after a write request by a posted access, whereby a previously issued write The writing of data related to the request can be confirmed. Therefore, even when a protocol that does not make a write completion response is used in the internal data transfer paths 101 and 102, the write to the memory can be confirmed without storing the write request issuance path for each packet. it can. In addition, a path storage area for each packet is not necessary, and the mounting can be downsized. In this embodiment, the DMAs 210 to 212 can automatically transmit a dummy read request at the timing when the write destination memories M0 to M7 are switched and the timing when the writing is completed, so that the management terminal 31 and the host computer 20 Since there is no need to perform additional processing such as issuing a dummy read request by the control software to be executed, the performance of the entire system can be increased.

(Modification 1 of the first embodiment)
In the present embodiment, a dummy read request is issued to the memories M0 to M7 in order to confirm writing, but it may be modified so that a read request is issued to the registers of the memory controllers 230 to 233 instead. It is assumed that the read request for the register is also non-posted access. In the case of this modification, since the memories M0 to M7 are not accessed to confirm the writing, it is possible to confirm the writing at a higher speed, and there is an additional effect of improving the memory utilization efficiency. The registers of the memory controllers 230 to 233 may have a function of recording the number of write requests. In this case, the DMAs 210 to 212 can determine whether transfer is normal or abnormal by comparing the difference between the value read from the register and the number of issued write requests. Therefore, there is an additional effect that the reliability of writing determination can be further improved. The register may be provided for each resource that accesses the DMA or other memory, or may be provided for each access area. In addition to the write request, the memory controllers 230 to 233 may include a register that holds the number of read requests and other statistical information.

(Modification 2 of the first embodiment)
In this embodiment, a dummy read request (non-posted access) is issued in order to confirm writing to the memory. Instead, a dedicated instruction for posting access that does not overtake the preceding write request (write confirmation request) It may be modified so that it is defined and issued. For example, when a TLP (Transaction Layer Packet) in the format shown in FIG. 10 is defined as a write confirmation request and this write confirmation request is transmitted, the memory controllers 230 to 233 flush the data in the buffer to the memory, A response by a posted access such as 11 (write confirmation response) is returned. The sequence in this case is as shown in FIG. In the case of this modification, unlike the non-posted access, the DMAs 210 to 212 do not have to secure a buffer for receiving response data (read response) and wait. Thereby, there is an additional effect that the parallelism of the operation increases and the performance can be improved. In ordinary PCI Express, an undefined request (TLP) is discarded. Therefore, the definition is implemented in the PCI cores 221 and 223 so as to accept the above request and response.

(Modification 3 of the first embodiment)
In this embodiment, a dummy read request is issued to the memory in order to confirm the writing to the memory. However, a dedicated signal for a request for confirming the writing is provided on the mounting in the ASIC 200 and the paths of the internal data transfer paths 101 and 102. A line may be provided. FIG. 13 shows this implementation example. When it is necessary to confirm writing, the DMAs 210 to 212 instruct the signal controller 280 to send signals to the memory controllers 230 to 233 that need to confirm writing. The signal includes, for example, information such as the number of write transfers from the DMAs 210 to 212, and the memory controllers 230 to 233 return a confirmation completion signal when the requested number of transfers arrives. The DMAs 210 to 212 can confirm that the writing has been confirmed through the signal controller 280. In the case of this modification, since the write confirmation request and response do not pass through the data transfer path, there is an additional effect that the transfer efficiency is improved without affecting the data transfer.

(Modification 4 of the first embodiment)
In the present embodiment, the write is confirmed by the point that the memory controllers 230 to 233 are switched, but the memory controller 230 to 233 may be used instead of the memory controller 230 to 233. In this case, it is realized by changing the address range for each DIMM instead of the address range for each of the memory controllers 230 to 233 so as to determine whether or not the access destination DIMM is different from the previous time in S8050 of FIG. Is possible. A normal DIMM is composed of a plurality of memory chips and is managed in units of banks. Therefore, similarly, the point at which the bank in the DIMM is switched may be used as a trigger for confirming the writing. In this case, the address range for each bank is managed.

(Modification 5 of the first embodiment)
In the present embodiment, the write confirmation process is performed in response to the switching of the memory surface. However, it may be performed periodically every predetermined number of times before the memory surface is switched. In the case of this modification, when writing data with a very large transfer length, the writing can be confirmed periodically, so the reliability of data writing can be improved and the write confirmation process is performed in detail. In addition, there is an additional effect that recovery processing when an error occurs becomes easy. Further, the transfer parameter 300 may allow the processors 760 and 761 to explicitly specify the necessity and frequency of the write confirmation process. Alternatively, two types of activation trigger registers for DMA 210 to 212 are provided, and when activated by one trigger register, operates with a write confirmation process, and when activated by the other trigger register, the write confirmation process is not performed. You may do it. In this case, unnecessary write confirmation operations can be reduced, so that there is an additional effect of improving processing performance.

(Modification 6 of the first embodiment)
In this embodiment, the DMAs 210 to 212 read the parameter 300 and execute data transfer between the memories according to the instruction, but transfer the small data written in the buffer or register to the memory while executing the write confirmation process. Another DMA may be provided. Further, since the data transfer length and transfer source address can be known from the register size and address, the construction of the parameter 300 may be omitted and the transfer destination address may be set in the trigger register. Further, when the data transfer is normally completed, the transmission completion notification may be omitted. In the case of this modification, since the overhead of parameter creation and transfer status check in data transfer can be reduced, there is an additional effect of improving processing performance even in data transfer in small data units.
In addition, you may combine said modification 1-6 as needed.

<Example 2>
A second embodiment will be described with reference to FIGS.
The basic configuration of the apparatus in the second embodiment is the storage apparatus shown in FIG. 1, and the configuration of the storage controller is the same as that in FIG. 3, but the functions of the ASIC 200 are different. In the following description, portions different from the first embodiment will be mainly described, and description of similar portions will be omitted. Since the ASIC 200A and the ASIC 200B have the same configuration, the alphabetical character at the end is omitted in the following description.

  FIG. 14 shows functional blocks of the ASIC 200 in the storage controller of the second embodiment. The ASIC 200 is provided with a DMA status block 400 for managing the status of the DMAs 210 to 212. The DMAs 210 to 212 are provided with a state management unit (SM) 2104 for managing its own state and acquiring information of other DMAs 210 to 212.

  The DMA status block 400 has a table 401 indicating the transfer status of each of the DMAs 210 to 212. Details of the table 401 will be described later with reference to FIG. The state management unit 2104 of each DMA 210-212 accesses this DMA status block 400 through the switch block 250, records its own information, acquires information of other DMAs 210-212, and uses it to determine the path selection. To do.

  FIG. 15 shows an example of the table 401 in the ASIC 200 of the second embodiment. The table 401 includes a number (DMA number) 4010 of each DMA 210 to 212, a current state 4020, a memory controller number 4030 of a transfer source, a memory controller number 4040 of a transfer destination, an internal data transfer path number 4050 in use, a request type. There are fields such as 4060.

  The DMA number 4010 is a field in which a number for identifying the DMAs 210 to 212 in the ASICs 200A and 200B is stored. The current state 4020 is a field indicating whether each DMA is transferring data (BUSY) or stopped (IDLE). The transfer source memory controller number 4030 is a number for identifying the transfer source memory controllers 230 to 233 specified by the parameter analysis in the DMAs 210 to 212. Similarly, the transfer destination memory controller number 4040 is a number for identifying the transfer destination memory controllers 230 to 233 specified by the parameter analysis in the DMAs 210 to 212. The internal data transfer path number 4050 is a number for identifying the internal data transfer paths 101 and 102 used when data is transferred through the internal data transfer paths 101 and 102. The request type 4060 indicates whether it is used for reading (RX), used for writing (TX), or used for both (TX / RX). In this embodiment, there is one field indicating the transfer destination memory controller number 4040. However, when the DMA has a function of transferring data from one transfer source to a plurality of transfer destinations simultaneously, the table 401 has a plurality of fields corresponding to the number. Transfer destination memory controller number field.

  FIG. 16 is an example of a flowchart when the DMA 210 in this embodiment performs a data transfer operation. The DMAs 211 and 212 perform the same operation as the DMA 210. The parameter fetch unit 2103 of the DMA 210 analyzes the parameter 300 by the same method as shown in FIG. 7, and specifies the transfer source and transfer destination memory controllers 230 to 233 (S8010, S8040). As a result, if it is necessary to access the memory controllers 230 to 233 of the transfer source or the transfer destination using the internal data transfer paths 101 and 102 (that is, when accessing the memories M4 to M7 of another package), the DMA 210 The state management unit 2104 accesses the DMA status block 400 and determines which internal data transfer path 101 or 102 should be used. Specifically, the state management unit 2104 refers to the table 401 and totals the number of records for each request type 4060 for each of the internal data transfer paths 101 and 102. When the access using the internal data transfer paths 101 and 102 is read (that is, when only the transfer source memory controllers 230 to 233 correspond to the memories M4 to M7 of another package) (S9020: Yes), the state management unit 2104 The number of records in which “RX” is included in the request type 4060 is counted for each of the internal data transfer paths 101 and 102 (S9030). On the other hand, when the access using the internal data transfer paths 101 and 102 is writing (that is, when only the transfer destination memory controllers 230 to 233 correspond to the memories M4 to M7 of another package) (S9040: Yes), the state management unit The 2104 counts the number of records in which “TX” is included in the request type 4060 for each of the internal data transfer paths 101 and 102 (S9050). When access using the internal data transfer paths 101 and 102 is both reading and writing (that is, when both the transfer source and transfer destination memory controllers 230 to 233 correspond to memories M4 to M7 in different packages) The management unit 2104 counts the number of records for each of the internal data transfer paths 101 and 102 (S9060). The state management unit 2104 selects the internal data transfer paths 101 and 102 with the smallest number of records counted for the request type 4060 (S9070). Then, the state management unit 2104 updates the record information corresponding to the own DMA 210 such as the memory controllers 230 to 233 to be used and the internal data transfer pass number to be used. For example, when the number of the own DMA 210 is “0” and data is transferred from the memory controller 0 of the same package as the DMA 210 to the memory controller 5 of another package using the internal data transfer path 0 (101), the state management unit 2104 Is “BUSY” in the status 4020, “0” in the transfer source memory controller number 4030, “5” in the transfer destination memory controller number 4040, and “0” in the internal data path number 4050 for the record with the DMA number 4010 “0”. ”And“ TX ”are set in the request type 4060, respectively.

  When the data transfer according to the parameter 300 is completed (S8020 to S8030 or S8050 to S8110 in FIG. 7), the state management unit 2104 of the DMA 210 updates the state of the table 401 again (S9090). That is, the state 4020 corresponding to the own DMA 210 is set to “IDLE”, and the transfer source memory controller number 4030, the transfer destination memory controller number 4040, the internal data transfer path number 4050, and the request type 4060 are cleared.

(Effect of Example 2)
According to the second embodiment, when there are a plurality of internal data transfer paths 101 and 102, the internal data transfer path used for data transfer is determined based on the usage status of the internal data transfer paths 101 and 102 by other resources (DMA). 101 and 102 can be selected, and efficient transfer becomes possible. Conventionally, there is a method of selecting the network path transfer buffer by checking the availability of the network path, but when a plurality of DMAs 210 to 212 see the availability with a slight time difference, all the DMAs 210 to 212 select the same path. There is a possibility that efficient transfer cannot be performed. In the second embodiment, the DMAs 210 to 212 referring to the table 401 reflect the selection result in the table 401. Since the next DMA 210 to 212 can determine the internal data transfer paths 101 and 102 used for data transfer based on the selection result of the previous DMA 210 to 212, all the DMAs 210 to 212 have the same internal data transfer path 101, 102. The possibility of selecting 102 is reduced, and more efficient path selection becomes possible.

(Modification 1 in Example 2)
In the second embodiment, each of the DMAs 210 to 212 determines the path selection by looking at the information of the other DMAs 210 to 212. However, a determination logic is provided in the status block 400 so that the DMAs 210 to 212 inquire the status block 400. May be. In this case, the status block 400 is provided with a path selection unit, and the DMAs 210 to 212 transmit requests including the transfer source and transfer destination memory controllers 230 to 233 to the path selection unit, and the path selection unit responds to this request. , The processing of S9000 to S9070 in FIG. 16 is performed, and the selected internal data transfer paths 101 and 102 are responded. In the case of the first modification, the path selection logic can be consolidated in one place, so that the mounting efficiency may be improved. In addition, the update timing of the record in the table 401 can be adjusted by the status block 400, so writing during data reading, etc. This has the additional effect of making it easier to prevent unauthorized information.

(Modification 2 in Example 2)
In the second embodiment, the selection of the internal data transfer paths 101 and 102 by the DMAs 210 to 212 is determined to determine the path selection. In addition to this, the buffer availability of the internal data transfer paths 101 and 102 and the PCI status The determination may be made in consideration of the remaining number of Express transfer credits. For example, when the internal data transfer path 0 (101) is rarely used but there is no space in the corresponding buffer, it is possible to determine to select the internal data transfer path 1 (102). As an example, the internal data transfer path arbiters 220 and 222 have a function of monitoring the buffer amount used in the internal data transfer paths 101 and 102 and a function of monitoring the transfer credit remaining number in the PCI Express cores 221 and 223. This can be realized. In the second modification, the status block 400 stores a table 402 as shown in FIG. In the example of FIG. 17, more detailed buffer credits such as headers and data, that is, posted header buffer credits (PH_CREDITS; Posted Header buffer CREDITS), posted data buffer credits (PD_CREDITS; Posted Data buffer CREDITS), Non-Posted Header Buffer Credits (NPH_CREDITS), Non-Posted Data Buffer Credits (NPD_CREDITS; Non-Posted Data Buffer CREDITS), Completion Response Header Buffer Credits (CPLH_CREDITS; ComPLetion Header Buffer CREDITS) And completion response data buffer credits (CPLD_CREDITS; ComPLetion Data buffer CREDITS). In the case of a posted access such as a memory write request, the remaining number of PH_CREDITS or PD_CREDITS is referred to, and in the case of a non-posted access such as a memory read request, the remaining number of NPH_CREDITS or NPD_CREDITS is referred to. That's fine. In the case of this modification, since the congestion status of the internal data transfer paths 101 and 102 can be taken into account, there is an additional effect that path selection with higher accuracy is possible.

(Modification 3 in Example 2)
In the second embodiment, the determination is made by checking the selection status of the internal data transfer paths 101 and 102 by the DMAs 210 to 212. In addition to this, the usage status of the transfer source / destination memory controller numbers 4030 and 4040 in the table 401 is used. You may make it judge in consideration of. Assume that the memory controller 231 is accessed through the internal data transfer paths 101 and 102. For example, if the internal data transfer path 0 (101) has a smaller number of DMAs, but the access destinations are all the same memory controller 230, the buffer of the memory controller 230 overflows and the subsequent data on the internal data transfer path 0 Packets may be stuck. Therefore, there may be a case where the internal data transfer path 1 (102) is selected even if the number of selected DMAs is large. In the third modification, the state management unit 2104 can be realized by limiting the internal data transfer paths 101 and 102 to, for example, a DMA having a destination memory controller of 231 and performing the counts shown in S9030, S9050, and S9060. In the case of the third modification, in addition to the congestion status of the internal data transfer paths 101 and 102, the congestion status of the access-destination memory controllers 230 to 233 can be taken into account, so that there is an additional effect that path selection with higher accuracy is possible. .

(Modification 4 in Example 2)
In the second embodiment, the determination is made by checking the selection status of the internal data transfer paths 101 and 102 by the DMAs 210 to 212. In addition to this, the transfer length, the number of transfer blocks, the number of remaining transfer blocks, etc. More detailed information may be added to the table 401 and used for determination. In the case of this modification, the accuracy in selecting a path can be increased, and there is an additional effect that enables more efficient determination.

  In addition, you may combine the modifications 1-4 of said Example 2 as needed.

<Example 3>
A third embodiment will be described with reference to FIGS.
The basic configuration of the apparatus in the third embodiment is the storage apparatus shown in FIG. 1, and the configuration of the storage controller is the same as in FIG. 3, but the format of the parameter 300 used by the processors 760 and 761 for the DMA 210 to 212 transfer instructions, and The difference is that the processors 760 and 761 can notify / set the use layout information of the memories M0 to M7 to the ASIC 200. In the following description, only the parts different from the first and second embodiments will be mainly described, and the description of the same parts will be omitted.
FIG. 18 shows an example of the format of the parameter 300 created by the processors 760 and 761 in the third embodiment in order to issue a transfer instruction to the DMAs 210 to 212. In the parameter 300 of the third embodiment, in addition to the fields 3001 to 3006 included in the parameter 300 of the previous embodiment, a field (I / O type) 3007 for designating the I / O type and processing priority of the transfer concerned. There is. The I / O type field 3007 is an 8-bit field, with the upper 4 bits indicating the I / O type and the lower 4 bits indicating the priority. For example, as the type of I / O, random I / O from the host is 1000b (binary number), sequential I / O from the host is 0100b, and transfer such as a synchronization function between volumes provided in the storage is 0110b to 0111b Etc. shall be set for each type. On the other hand, the priority is 1000b for the highest-priority transfer, 0001b for the lowest-priority transfer, and the like.

  FIG. 19 shows a memory layout information setting table 800 in the third embodiment. This table is provided in a setting register for the internal data transfer path arbiters 220 and 222 in the ASIC 200. The information in the memory layout information setting table 800 is only required to be used by the internal data transfer path arbiters 220 and 222, and may be arranged in another place as long as the same effect can be obtained. For example, the memory layout information setting table 800 can be arranged in the local memories 770 and 771 and the memories M0 to M7.

  In the memory layout information setting table 800, a memory range used by the DMAs 210 to 212 for memory transfer is specified. In general, the entire area of the memories M0 to M7 cannot be used as a cache memory area of the storage. Some of the memories M0 to M7 are a control information storage area of a protocol chip, a cache management information storage area of the processors 760 and 761, and the like. Used. Therefore, by designating other areas in this table 800, which memory controllers 230 to 233 are accessed by the ASIC 200, in particular, the internal data transfer path arbiters 220 and 222 for controlling the data passing through the internal data transfer paths 101 and 102. You can tell if more occur.

  The configuration of the memory layout information table 800 will be described. In the memory number field 8010, the memory number is recorded. In the third embodiment, the memory number 8010 is the memories M0 to M7. A start address field 8020 is a start address of an effective access range by the DMAs 210 to 212. An end address field 8030 is an end address of an effective access range by the DMAs 210 to 212. The ratio field 8040 is a field for setting a frequency ratio at which the ASIC 200 accesses each of the memory controllers 230 to 233. The ratio is a value corresponding to the size of the address range specified by the processors 760 and 761. In the third embodiment, the number of spaces is shown for each address space. In the example of FIG. 19, the address space 1000h (addresses 0000h to 0FFFh) is used as a unit, and both the memory M2 and the memory M3 provide a space of 10000h (16), so the memory M0 provides a space of 5000h. It is 5 from that. The counter field 8050 is an operation counter used by the arbiter to manage the change of the arbitration ratio, and the initial value is set to the same value as the ratio field 8040.

  FIG. 20 is a diagram schematically showing data exchange between the DMA 210 and the internal data transfer path arbiter 220 in the present embodiment. Since the exchange of data between the DMAs 210 to 213 and the internal data transfer path arbiters 220 and 221 is the same, the DMA 210 and the internal data transfer path arbiter 220 will be described here. Between the DMA 210 and the internal data transfer path arbiter 220, a REQ signal line 601 for requesting a path use right from the DMA 210 to the internal data transfer path arbiter 220, a REQ_MEM_NO signal line 602 for transmitting an access destination memory controller number at the time of request, and an internal data transfer path arbiter An ACT signal line 603 for transmitting permission of a path use right from 220 to the DMA 210, and various signal lines 60n for communicating a signal for transmission / reception and synchronization of both data, an error status signal, and the like are provided. The internal data transfer path arbiter 220 can communicate with all the DMAs 210 to 213 and other resources of the in-controller ASIC 200 using the path via the switch block 250 using these signal line groups.

  FIG. 21 is a flowchart showing the operation of the internal data transfer path arbiters 220 and 222 in this embodiment. Hereinafter, the internal data transfer path arbiter is simply referred to as an arbiter. Since the arbiter 220 and the arbiter 222 perform the same operation, only the arbiter 220 will be described here.

  When the DMAs 210 to 212 are activated by the same procedure as in FIG. 5, the parameters 300 are analyzed in the same manner as in the procedure of FIG. 7, and the access destination memory controllers 230 to 233 are specified. When accessing the access destination memory controllers 230 to 233 via the internal data transfer paths 101 and 102, the DMAs 210 to 212 receive a request signal for acquiring the path use right together with a signal indicating the access destination memory controller number and the REQ signal line 601 and REQ_MEM_NO. This is issued to the arbiter 220 through the signal line 602.

When the arbiter 220 receives a request signal from the DMAs 210 to 220 (S9000), the arbitration operation for the usage right is started in order.
When the received request signal is one (S9010: Yes), the arbiter 220 gives the right to use the request signal (S9100), and returns the ACT signal through the ACT signal 603. When the DMA receives the ACT signal, the DMA transfers data to and from the memory controllers 230 to 233. When the DMA ends, the DMA REQ signal on the REQ signal line 601 is withdrawn (the signal is negated), and the usage right is released. The arbiter 220 waits for the usage right to be released (S9110) and waits for a request signal again.

  If a plurality of request signals are received at the same time or if there are a plurality of requests waiting (S9010: No), the arbiter 220 indicates the access destination memory controller number received through the REQ_MEM_NO signal line 602 together with the REQ signal. Check (S9020). Then, referring to the memory layout information setting table 800, the value of the counter field 8050 corresponding to each memory controller number is compared (S9030). Then, the memory controller number having the largest value in the counter field 8050 is selected from the memory controllers 230 to 233 requested as the access destination, and the path use right is given to the DMA 210 to 212 that designates the number (S9040). ). When a plurality of DMAs 210 to 212 are applicable, the arbiter 220 arbitrates the usage right in round robin. Then, the counter value corresponding to the memory controller number is decremented by 1 and the priority level is decremented by 1 (S9050). When all the counter values become 0 (S9060: Yes), the arbiter 220 refers to the ratio field 8040 and resets the counter value of each memory so as to match the value of the ratio field 8040 (S9070). Then, it waits for the given path usage right to be released (S9080), and again enters the signal waiting state.

(Effect of Example 3)
According to the third embodiment, by using the memory resource usage information set by the processors 760 and 761, that is, the control software, the internal data transfer paths 101 and 102 can be used more efficiently than the arbitration using only hardware. Arbitration can be performed, and the data transfer of the internal data transfer paths 101 and 102 can be made efficient.

  Although the three embodiments and the modifications associated therewith have been described above, this embodiment is for explaining an example of means for solving at least one of the problems listed in the present invention. This is not a limitation. Further, the above embodiments can be applied individually or in combination.

  The present invention can be applied to one or a plurality of storage apparatuses described in the first to third embodiments, and is not limited to PCI Express, but uses one or a plurality of networks that do not respond to the completion of writing. Therefore, the present invention can be widely applied to any data transfer device connected in this manner.

  Although the present embodiment has been described above, the above embodiment is intended to facilitate understanding of the present invention and is not intended to limit the present invention. The present invention can be changed and improved without departing from the gist thereof, and the present invention includes equivalents thereof.

Claims (14)

An acquisition unit that acquires from a storage unit a transfer source address of a first memory that is a data transfer source, a transfer destination address of a second memory that is a data transfer destination, and a data transfer length;
A read unit for acquiring data of the first memory based on the transfer source address;
Based on the transfer destination address, the acquired data is transferred to a memory controller that writes data to the second memory,
When the transfer is completed, it is determined whether the amount of the transferred data has reached the transfer length of the data,
As a result of the determination, if the amount of transferred data has reached the transfer length of the data, a writing unit that transmits a read request to the memory controller;
A data transfer device comprising: The data transfer device according to claim 1, wherein the data transfer device receives a response from the memory controller and notifies the processor of the end of the data transfer. The data transfer apparatus according to claim 2, wherein the data transfer apparatus starts the data transfer process based on an instruction from the processor. The data transfer device according to claim 3, wherein the data transfer device notifies the processor of the end of the data transfer when there is no error in the response from the memory controller. 5. The data transfer device according to claim 4, wherein if the amount of the transferred data does not reach the data transfer length as a result of the determination, the data transfer device performs data transfer until the data transfer length is reached. . The writing unit transmits the read request to the memory controller after transmitting the write request,
6. The memory controller according to claim 1, wherein the memory controller writes the transferred data to the second memory based on the write request, and performs processing based on the read request after the writing is completed. The data transfer device according to item. The data transfer device according to claim 6, wherein the writing unit transmits the write request and the read request to the memory controller via the same bus. The data transfer device according to claim 6, wherein the writing unit accesses the register included in the memory controller by transmitting the read request. The data transfer device according to claim 8, wherein the register records the number of write requests. The data transfer device
Based on the number of write requests recorded in the register acquired by the read request and the number of write requests transmitted from the writing unit,
The data transfer device according to claim 9, wherein it is determined whether or not the data transfer has been normally performed. A data transfer method for a data transfer device for transferring data from a first memory to a second memory, comprising:
Obtaining a transfer source address of the first memory, a transfer destination address of the second memory, and a data transfer length from a storage unit;
Obtaining data in the first memory based on the transfer source address;
Transferring the acquired data to a memory controller that writes data to the second memory based on the transfer destination address;
When the transfer is finished, determining whether the amount of transferred data has reached the transfer length of the data;
As a result of the determination, if the amount of transferred data has reached the transfer length of the data, sending a read request to the memory controller;
A data transfer method. The data transfer method according to claim 11, further comprising a step of receiving a response from the memory controller and notifying a processor of completion of data transfer. The data transfer method according to claim 12, further comprising a step of starting the data transfer process based on an instruction from the processor. The data transfer device transmits the read request after transmitting the write request to the memory controller,
The memory controller writes the transferred data to the second memory based on the write request, and performs processing based on the read request after the writing is completed. The data transfer method according to item.

Images