Patent Application
20080082759
The accompanying drawings, which are incorporated herein and form a part of the specification, illustrate the present invention and, together with the description, further serve to explain the principles of the invention and to enable a person skilled in the pertinent art to make and use the invention.
FIGS. 3C1-3C2 illustrate examples of maintaining memory coherency for an exclusive ownership request in a shared memory system according to an embodiment of the invention.
FIGS. 5C1-5C3 illustrate example operation of memory locks in test and set mode according to an embodiment of the invention.
FIG. 5C4 illustrates a flowchart showing steps performed by a node controller in test and set mode, according to an embodiment of the invention.
FIGS. 5D1-5D3 illustrate example operation of memory locks in sequencer mode according to an embodiment of the invention.
FIG. 5D4 illustrates a flowchart showing steps performed by a node controller in sequencer mode, according to an embodiment of the invention.
FIG. 5D5 illustrates a flowchart showing steps performed by a requester in sequencer mode, according to an embodiment of the invention.
The present invention will now be described with reference to the accompanying drawings. In the drawings, like reference numbers indicate identical or functionally similar elements. Additionally, the left-most digit(s) of a reference number identifies the drawing in which the reference number first appears.
Methods and systems for a multi-purpose adaptable memory are described herein. In particular, methods and systems for adapting the on-chip memory of a SOC to serve different purposes in different applications are described.
The present specification discloses one or more embodiments that incorporate the features of the invention. The disclosed embodiment(s) merely exemplify the invention. The scope of the invention is not limited to the disclosed embodiment(s). The invention is defined by the claims appended hereto.
References in the specification to “one embodiment,” “an embodiment,” “an example embodiment,” etc., indicate that the embodiment described may include a particular feature, structure, or characteristic, but every embodiment may not necessarily include the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the knowledge of one skilled in the art to effect such feature, structure, or characteristic in connection with other embodiments whether or not explicitly described.
Before describing embodiments of the present invention in detail, it is helpful to describe an example environment in which the invention may be implemented. In an example environment, embodiments of the invention are implemented on a multi-processor System On Chip (SOC) unit which includes on-chip memory and is coupled to external memory.
For clarity and ease of illustration, not all functional blocks of SOC 100a are illustrated in detail because these items are known to those of ordinary skill in the art. Furthermore, it is to be appreciated that similar or substitute functional blocks of SOC 100a may be used to implement embodiments of the invention without deviating from the scope of the present invention. Example configurations of OCM 108a and NC 104a according to embodiments of the invention are described in further detail below. Such example configurations are not limited to OCM 108a or NC 104a and may be used in alternative environments. Furthermore, the embodiments described herein may be adapted and modified, as would be apparent to persons skilled in the relevant art(s).
As shown in
In step 208, a request to access a memory location is received. The request includes an address of the memory location, whether it is a read or write request and the data to written if it is a write request.
In step 210, it is determined whether the address in the request received in step 210 corresponds to an on-chip memory location.
In step 212, if it is determined in step 210 that the address corresponds to an on-chip memory location, then the address is decoded. The address is typically decoded by an address decoder in the node controller. Decoding the address results in conversion of the address into electrical signals required to access the on-chip memory.
In step 214, based on the decoded address from step 212, the memory location is accessed and data is retrieved or stored, as specified by the instruction received in step 208.
In optional step 216, if the request received in step 208 was for reading data from memory, then the required data is supplied to the requesting service. In an embodiment, if the request received in step 208 was a write request, then a write confirmation may be provided to the requesting service upon completion of the write.
In step 218, if it is determined in step 210 that the address corresponds to an external memory location then the request is transferred to a memory controller. An address decoder in the memory controller decodes the address, accesses the external memory location and provides requested data or a write acknowledgement depending on the requirements of memory access request received in step 208.
Example embodiments describing a shared memory configuration for a node controller and on-chip memory are described below.
Processors 100a-n share a global address space. External memory 106a-n and/or caches of processors in each SOC are mapped into global address space. In embodiments of the invention, OCM 108a-n is not mapped into the global address space. In a shared memory system 300, maintaining cache coherence across memory shared by multiple processors in multiple SOCs is slow and has a significant processing overhead on processors. Embodiments of the invention described below offer a novel solution by utilizing OCM 108 via NC 104 as a cache directory to track location of shared data in system 300. As described below, each OCM 108a-n is used by its corresponding SOC 100a-n as a directory to track shared data in system 300. OCM 108a tracks shared data by storing entries for shared memory location. In an embodiment, each entry includes, but is not limited to, the following fields:
Field Valid (v) to indicate whether the a remote node has a valid copy of the data;
Field Dirty (d) to indicate whether the data sent to the remote node is modified by the remote node i.e. whether the remote node made an “exclusive ownership” request for the data;
Field Tag to identify the local memory location requested; and
Field Vector (vec) to identify the remote SOC that requested data from the memory location. For example, vec may include multiple bits where each bit corresponds to a SOC in system 300. In the example in
In the example presented below with reference to
When a remote SOC 100j issues a read request to local SOC 100a for memory location B in EM 106a, node controller 104a checks if an entry corresponding to memory location B with a dirty bit set exists in OCM 108a. If no such entry exists in OCM 108a, local memory 106a has valid data and node controller 104a provides data from memory 106a to requesting remote SOC 100j. If a dirty copy exists in, for example, SOC 100n, a corresponding dirty entry exists in OCM 108a with the n-th bit of the vector field in that entry set to 1. Node controller 104a issues a write-back command to the current owner SOC 100n to provide the most recent data of the location B. After SOC 100n returns the data, the node controller 104a writes the data back to local external memory 106a and also provides it to the requesting SOC 100j. The OCM 108a entry is modified such that the dirty bit is reset to 0 and the j-th and n-th bit in the vector are set to 1. This means that SOC 100j and SOC 100n have a shared, clean data copy of the memory location B. Their values are consistent with the memory copy in EM 106a.
When a local processor of SOC 100a issues a read request, entries in OCM 108a are checked. If no remote dirty copy exists, local memory 106a has a valid copy and node controller 104a provides the data. If a remote dirty copy exists, node controller 104a issues a write-back request to the current SOC owner of the requested memory location, updates local memory 106a by the returned data and modifies the corresponding OCM 108a entry by resetting the dirty bit to 0.
When remote SOC 100j issues a write request to local SOC 100a for memory location B in EM 106a, node controller 104a ensures that SOC 100j has one and only one copy of memory location B in system 300 to ensure data coherence. Node controller 104a checks if an OCM 108a entry exists. If no OCM 108a entry exists for location B, node controller 104a provides a data copy from local memory EM 106a to the requesting SOC 100j. If an OCM 108a directory entry exists and the dirty bit is not set, it implies clean copies in the shared state may exist in system 300. In this case, an invalidation signal is sent by node controller 104a to those SOCs whose corresponding bit in the vec entry has a value of 1. These SOCs will invalidate their locally cached copy of memory location B and send acknowledgement to local SOC 100a after completion of the invalidation. Local SOC 100a is responsible for collecting all invalidation acknowledgements. After all remote shared copies are invalidated, local SOC controller 104a provides a data copy from the local memory EM 106a to the requesting SOC 100j.
If an OCM 108a entry exists and the dirty bit is set in the entry, it implies a dirty copy exists in system 300. The SOC that currently holds a dirty copy is specified by the bit whose value is 1 in the vector field of the entry. For example, if the n-th bit is 1, the SOC 100n is the current owner and has a dirty copy of the location B. In this case, node controller 104a issues a flush command to SOC 100n. Upon receiving the flush command, SOC 100n sends back its data copy and invalidates its copy. Subsequently, node controller 104a forwards the received data copy from SOC 100n to requesting SOC 100j. In this case, an OCM 108a entry is updated such that only SOC 100j has a copy in the dirty state, that is, the dirty bit is set and only the j-th bit in the vector is set in the corresponding OCM 108a entry. This means SOC 100j has the exclusive copy in system 300.
When a local processor of SOC 100a issues a write request, entries in OCM 108a are checked. If no entry or no remote dirty copy exists, the local memory 106a has a valid copy and node controller 104a provides the required data. If a remote dirty copy exists, node controller 104a issues a flush request to the current owner SOC. If remote shared copies exist, node controller 104a issues invalidations to the remote SOCs who have a copy. SOC 100a now has the exclusive copy and no other SOC in system 300 has a data copy of the accessed memory location B in EM 106a.
FIG. 3C1 illustrates an example of maintaining memory coherency on an exclusive ownership request in shared memory system 300 according to an embodiment of the invention.
As shown in FIG. 3C1, SOC 100j sends a request for exclusive ownership of data at a memory location B in EM 106a. Exclusive ownership allows SOC 100j to retrieve the data from memory location B and modify it. Therefore the copy at memory location B will become obsolete. The request may originate from a processor in SOC 100j, for example processor PT. The request may be sent by a message from NC 104j to NC 104a via interconnect 302. In response to receiving a request from node controller 104j in SOC 100j for exclusive ownership of memory location B, node controller 104a in SOC 100a sends data from memory location B to node controller 104j via interconnect 302. To maintain coherence, node controller 104a tracks the location of data corresponding to memory location B by storing an entry 306 in OCM 108a.
In the present example shown in FIG. 3C1, v is set as 1 to indicate that the entry is valid, d is set as 1 since exclusive ownership was requested and the copy in memory location B of SOC 100a will not be the latest copy, tag is set as B since data from memory location B was requested. In the above example, since the request originated from SOC 100j, the jth bit of the vec field is set to 1. In an alternate embodiment, the processor PT of SOC 100j may be indicated in the vec field as the originator of the request by setting the Tth bit of the vec field to 1. If a process, thread or processor in SOC 100a wants to access data that was at memory B, the: requester recalls the data by requesting it from SOC 100j as shown in FIG. 3C2 and described below.
FIG. 3C2 illustrates retrieving data corresponding to memory location B in EM 106a that was requested for exclusive ownership by NC 104j in SOC 100j in the example above.
In the example shown in FIG. 3C2, processor PK requests data at memory location B. Node controller 100a looks up entry 306, corresponding to tag B, in OCM 108a. Since, in the previous example shown in FIG. 3C1, exclusive ownership of data at memory location B was requested by processor PT of SOC 100j, NC 104a sends a request to NC 104j to send data corresponding to memory location B in EM 106a. Node controller 108j accesses the data corresponding to memory location B, for example, from the processor caches (not shown), and sends the data to NC 104a. Since node controller 108a recalled the data from SOC 100j, it again updates entry 306 in OCM 108a as shown in FIG. 3C2.
In the present example shown in FIG. 3C2, v is set as 0 to indicate that the entry is no longer valid in SOC 100j, d is set to 0, if NC 104a requested exclusive ownership. If NC 104a requested data for read only, then v is maintained as 1 to indicate that SOC 100j still has a valid copy. However, d is set to 0 to indicate that SOC 100j no longer has exclusive ownership of the requested data i.e. shared copies of the same data exist in the processor caches in SOC 100a and SOC 100j. Tag is maintained as B.
It is to be appreciated that the numbering of events shown in
In step 314, a request is received from a remote SOC. The request may be received by a local node controller coupled to the remote node controller of the remote SOC. The request may identify the memory location of the local SOC to be accessed along with identifying the remote SOC requesting data. The request may also identify whether the remote SOC wants to read the memory location or have exclusive ownership of the data at a memory location.
In step 316, it is determined whether the local SOC has valid data in the requested memory location. Validity of data is determined by checking an entry, if one exists for the requested memory location, in local on-chip memory. If no entry exists in local on-chip memory for the requested memory location, then valid data in local memory is assumed. If an entry exists, then the local SOC checks if the dirty bit is set. If the dirty bit is set then the local SOC does not have valid data in its local memory.
In step 318, if it is determined in step 316 that the local node does not have valid data in the requested memory location, then the local SOC sends a request to the SOC having the valid data to return the valid data to the local SOC. The local SOC may determine the current SOC having the valid data by examining a vec entry in an on-chip memory as described above.
In step 320, the data received in step 318 is sent to the remote SOC requesting data.
In step 322, if it is determined in step 316 that valid data corresponding to the memory location requested in step 314 is present in local memory, then it is determined from the request received in step 314 whether the remote node has requested exclusive ownership.
In step 324, an entry is stored in the local OCM corresponding to the request by the remote SOC. For example, if the remote SOC requests exclusive ownership an entry is made (as in entry 306 in FIG. 3C1) where the dirty bit is set to 1. If the remote SOC has a read only request then an entry (as in entry 304 in
In step 326, the data from the requested memory location is sent to the remote SOC. The data may be sent by the local node controller to the remote node controller via bus 302 as described above.
In step 332, a remote SOC sends a request to a local SOC for data from a memory location. The request may be received by a local node controller of the local SOC that is coupled to the remote node controller of the remote SOC. The request may identify the memory location of the local SOC to be accessed along with identifying the remote SOC making the request. The request may also identify whether the remote SOC desires to only read the memory location of the local SOC or have exclusive ownership of the memory location. In an embodiment, the local SOC and remote SOC share a global memory address space and the remote SOC uses typical load/store instructions to access a memory location in the local SOC.
In step 334, the local node controller receives the request sent by the remote SOC in step 332. The request may be received and prioritized by the local node controller in the local SOC, according to the order in which the request was received.
In step 336, the local node controller determines whether it has valid data in the memory location for which the request was received in step 334. Validity of data is determined by checking an entry, if one exists for the requested memory location, in local OCM. If an entry exists in the local OCM for the requested memory location and, for example, the dirty bit in the entry is set to 1, then the local SOC does not have valid data in the memory location. If the dirty bit in the entry is set to 0 then the local SOC has valid data in the requested memory location. If no entry exists in local on-chip memory for the requested memory location, then it is assumed that valid data is present in the requested memory location.
In step 338, if the local node controller determines that the local SOC does not have valid data in the requested memory location, then the local NC, based on a vector field in the entry in a local OCM, requests the data from the SOC having the valid data.
In step 340, the data received by the local SOC from the SOC in step 338 is sent by the local SOC to the remote SOC. The local SOC also updates the local memory with the received data and updates the entry in local OCM to indicate that the local SOC has valid data which is not dirty.
In step 342, the local NC stores an entry in the local OCM corresponding to the request. For example, if the remote SOC requested exclusive ownership then an entry 306 as in FIG. 3C1 may be made where the dirty bit is set to 1. If the remote SOC made a read only request then an entry 304 as shown in
In step 344, the local NC sends the requested data to the remote NC. The data may be sent via bus 302 as described above.
In step 352, a request for reading data from or writing data to a local memory location is received from a local processor.
In step 354, it is determined whether the local SOC has valid data or exclusive ownership in the local memory location. Validity of data is determined by checking an entry, if one exists for the requested data, in local on-chip memory. In an embodiment, the local node controller checks the local on-chip memory. If an entry exists in the local OCM for the requested memory location and, for example, the dirty bit in the entry is set to 1, then the local SOC does not have valid data in the memory location. If the dirty bit in the entry is set to 0 then the local SOC has valid data in the requested memory location. If no entry exists in local on-chip memory for the requested memory location, then validity of data in local memory is assumed.
In step 356, if it is determined in step 354 that the local SOC does not have valid data in the memory location, then local OCM memory is checked, preferably by the local NC, for an entry corresponding to the memory location specified by the request received in step 352. In an example, as shown in FIG. 3C2 and described above, the vec field of entry 306 is used to determine the remote SOC that has exclusive ownership of the requested memory location.
In step 358, a request is sent to the remote SOC determined in step 356 to send the data.
In step 360, in response to the request sent in step 358, the local SOC receives the data from the remote SOC. The local SOC will update the corresponding entry in the local OCM to indicate that the local SOC has valid data in the requested memory location.
In step 362, the data received from the remote SOC is sent to the processor that requested the data.
In step 364, if it is determined in step 354 that the local SOC does have valid data in the memory location, then the memory location is accessed and data is sent to the processor requesting data. In one example the local NC may access the memory location and in another example, the local NC may use a local memory controller to access the data.
Example embodiments describing a cluster configuration for a node controller and on-chip memory are described below.
System 400 includes n SOC units 100a-n interconnected via network 402. In an embodiment, network 402 is a Local Area Network (LAN). The type of network and number of SOC units n is arbitrary. In an embodiment, SOCs 100a-n are connected to network 402 via node controllers 104a-n. In an embodiment of the invention, communication between SOC units 100a-n is via messaging. For example, in a “pull” model embodiment, a process, thread or processor in a source SOC unit sends a signal to a destination SOC unit indicating presence of a message in an allocated memory of the source SOC. The destination SOC, in response to receiving the signal, retrieves or “pulls” the message from the source SOC memory. In a “push” model embodiment, the source SOC transmits or “pushes” a message directly into an allocated memory location of the destination SOC. In an embodiment, the source SOC and destination SOC are part of a Redundant Array of Independent Disks (RAID) system and the message from the source SOC to the destination SOC is a “heartbeat” notification that indicates active operation of the source SQC to the destination SOC. Pull and push model messaging are described in further detail below.
Field “E” indicates an error;
Field “R” indicates a request for an acknowledgement. For example, the source SOC sending the message sets this bit to solicit an acknowledgement from the destination SOC after processing the message;
Field “A” indicates acknowledgement. The destination SOC sets this bit when returning an acknowledgement message to the source SOC that sent the original message;
Field “RSVD” is a reserved field for implementation specific usages;
Field “Status” indicates the outcome of processing the message. For example, a message might request the destination SOC to perform certain operations. The status field indicates whether operation was carried out and any associated information from the performed operation;
Field “INTR” indicates the type of interrupt, if any, that should be generated in the destination SOC and/or the priority of the message;
Field “RID” includes data to uniquely identify a message;
Field “SID” indicates the source SOC unit that generated the message; and
Field “Software Dependent Context” is for message payload. For example, the message payload may include data that is to be processed by the receiving SOC.
It is to be appreciated that the content and fields of a message are arbitrary and may change according to design requirements. In an embodiment, all or part of OCM 108a and 108j may be used as a mailbox for sending or receiving messages. In another embodiment, a specific portion of any memory in SOC 100a or SOC 100j may be used as a mailbox. For example, portions of EM 106a and EM 106j may be allocated as a mailbox in lieu of OCM 108a and 108j respectively. Allocating a specific portion of EM 106 may also allow for isolation of program code from potential viruses in messages.
In step 412, a signal is received from a remote SOC. The signal indicates presence of a message in the remote SOC. The signal may also indicate the memory location of the message in the remote SOC that contains the message.
In optional step 414, the local SOC requests the message from the remote SOC if it is does not know the memory location in the remote SOC to access or does not have access to the memory location in the remote SOC that contains the message. In an embodiment, the local node controller requests the remote node controller for the message.
In step 416, the local SOC retrieves the message from the remote SOC. In an embodiment, the local node controller of the local SOC retrieves the message from a memory location allocated in an on-chip memory of the remote SOC. The memory location may be part of a global address space used for messaging. Alternately, the local SOC might be pre-programmed with the memory location in the remote SOC that contains the message. If optional step 414 is implemented, then in step 416, the local node controller receives the message from the remote SOC instead of retrieving it.
In step 418, the local SOC stores the retrieved message. In an embodiment, the local node controller stores the message in-local on-chip memory.
In step 422, the remote SOC sends a signal indicating presence of a message in its on-chip memory. The signal may indicate the memory location containing the message.
In optional step 424, the remote SOC receives a request from a local SOC for the message.
In step 426, if in step 424 the remote SOC received a request from a local SOC for the message, the remote SOC sends the message to the local SOC. Alternatively, the remote SOC allows the local SOC access to the remote on-chip memory location containing the message.
In step 432, the memory location in a local SOC where a message is to be written is determined. The memory location in the local SOC that is to be used may be pre-programmed into the remote SOC or the local SOC may have previously indicated the memory location to be used for messaging. Alternatively, the memory location to be written in the local SOC may be part of a global address space used for messaging.
In step 434, the remote SOC writes the message to the memory location determined in step 432. Alternatively, the remote node controller in the remote SOC sends the message to the local node controller in the local SOC.
Example embodiments describing a memory lock configuration for a node controller and on-chip memory are described below.
A memory lock is a synchronization mechanism for enforcing limits on access to a memory location in an environment where multiple “requesters” (such as SOC units, threads, processes and/or processors) attempt to access the same memory location at proximately the same time. Locks are one way of enforcing concurrency control policies. Typically, requesters cooperate with each other by using an arbitration protocol that involves passing messages amongst requesters to decide which requester acquires the lock to access the memory location. However, when multiple requesters attempt to access the same memory location concurrently, typical methods of arbitration can be time consuming and increase system overhead. Embodiments of the invention described below overcome this deficiency by using node controller 104 in conjunction with OCM 108 to implement memory locks. The memory locks can be implemented in a stand alone SOC 100, shared memory systems such as system 300 or a cluster configuration as in system 400.
Test-and-set locks are used to atomically write to a memory location. Test-and-set locks involve setting a value in a lock entry to acquire a lock, but first performing some test (such as, testing the current set value in a lock as being equal to 0 or 1). If the test fails, the lock cannot be acquired or set and hence the memory location corresponding to the lock cannot be accessed by the requester. The requester has to again “test” the lock in the future to determine whether it can be acquired. Example embodiments using test-and-set mode are described below.
In sequencer mode, a ticket is assigned to each requester in the order in which a requester requests access to write a memory location. The lock entry corresponding to the requested memory location stores the current ticket and the next ticket. In the sequencer mode, the requester with the current ticket number acquires a lock and gains access to the corresponding memory location. While a memory location is occupied by a current ticket holder, the sequencer provides unique next tickets to subsequent requesters. The next ticket numbers are typically issued in sequential order thereby forming a virtual First-In-First-Out (FIFO) queue of requesters. When a current ticket holder accessing the memory location releases the lock, the current ticket +1 becomes the subsequent current ticket value. The requester with the subsequent current ticket value can now acquire the lock and gain access to modify the memory location.
Using node controller 104 to control access to memory locations in EM 106 removes the arbitration load from processors 110 and substantially eliminates bus traffic due to arbitration. Since OCM 108 is on-chip, accessing OCM 108 to look up a lock and determine whether to grant access to a corresponding memory location is fast and efficient.
Field “Mode” indicates the mode of operation to be used. For example, if mode is set as 0, then “test-and-set” lock mode is used. If mode is set as 1, the “sequencer” lock mode is used. The test-and-set lock mode and sequencer lock mode are further described below;
Field “S” or “Set” is applicable only when the field mode is in test-and-set lock mode. When S is 0, the lock is free. If a requester attempts to access a memory location and the corresponding lock entry's S bit indicates that it is free, then S is set to 1 and the requester is granted access. Setting the S bit to 1 is functionally equivalent to acquiring the lock. After the lock is acquired, the corresponding memory location can be accessed. Accessing a free lock entry returns the current value of 0 and the Software Dependent Context field (described below) as they were last saved. After the lock has been acquired, the Owner Identification (OID) field (described below) is set to indicate the requester that acquired the lock;
If, upon accessing a lock, a S value of 1 is returned to the requester by the NC, then the requester indicated by the OID field currently has the lock. Accordingly, the memory location corresponding to the lock cannot be accessed. The requester has to try again to acquire the lock;
Field “R” or “Recovery” bit is typically set during recovery operations. When the R bit is set as 0, the lock can be accessed normally. During recovery, when the R bit is set to 1, locks that are currently owned by inoperative or “dead” requesters are re-claimed. For example, a processor that acquires a lock might malfunction and become inoperative before it can release an acquired lock. When the R bit is set to 1, the lock is to be reset for recovery. With the R bit set to 1, recovery software can force release of a lock and re-initialize the corresponding memory location that the lock protects. During recovery, with the R bit set to 1 by a requester initiating the recovery process, the semantics of other fields are overruled. The OID is set to the ID of the requester selected to perform recovery. After a lock has been recovered, it can function normally;
Field “Next Ticket” stores a value that is utilized only in sequencer lock mode. Upon receiving a request to access a memory location, for example in EM 106, NC 104 looks up the lock corresponding to the requested memory location in OCM 108. If the lock is occupied, then NC 104 returns the current value in the Next Ticket field of the lock to the requester. NC 104 then increments the value in the Next Ticket field by one and returns that value to a subsequent requester when the lock is occupied. Thus, requesters waiting to access a memory location effectively insert themselves in a queue with an order of access defined by the assigned ticket number. If the lock is free when a requester requests access, then the current ticket value equals the next ticket value and the requester is assigned the current ticket value and granted access to the corresponding memory location. In an embodiment, Next Ticket is a 6-bit unsigned number. In another embodiment, the Next Ticket value may be a random alphanumeric code generated by NC 104. The node controller will then keep track of the sequence in which alphanumeric codes were generated and grant access in that order;
Field “Current Ticket” stores a value that is utilized only in sequencer lock mode. A requester having the current ticket number is allowed to acquire the lock and access the corresponding memory location. To release the lock, the requester stores a new value that is equivalent to the current ticket +1 in the Current Ticket field of the lock. This effectively allows the requester that is next-in-line to acquire the lock and access the memory location; and
Field “Software Dependent Context” is a 96-bit field that can be read from or written to by a requester that has acquired the corresponding lock. The context field may contain pointers to structures being protected by the lock. When a requester acquires a lock, it also gets a pointer to the memory location corresponding to the lock in, for example, external memory. System recovery often requires rebuilding or relocating data tables. The context field can be used to avoid significant performance problems with distributed table management.
Processes may also use the context field to implement efficient hardware mailboxes as described above. For exchanging data that does not exhibit temporal locality, the use of protected hardware mailboxes avoids cache coherence overhead.
FIGS. 5C1-5C3 illustrate example operation of memory locks in test-and-set mode according to an embodiment of the invention. In the examples presented in FIGS. 5C1-5C3, the mode bit is set as 0 (for test-and-set mode) and the next ticket and current ticket fields are not applicable (NA), since the lock is in test-and-set mode. The R bit is 0 indicating that the lock is not in recovery mode. In this example of the test-and-set mode, a value of S=0 indicates the lock is free and S=1 indicates that the lock is acquired.
As shown in FIG. 5C1, NC 104 receives a request from a requester R1 for access to modify a memory location X in EM 106. In response to receiving the request, NC 104 checks a lock entry 500a stored in memory location Y of OCM 108 that corresponds to memory location X. In lock entry 500a, S is 0 indicating that the lock can be acquired and the corresponding memory location can be accessed. Accordingly, NC 104 grants access to requester R1 to modify memory location X. Subsequently, either NC 104 or requester R1 modifies lock entry 500a to lock entry 500b. In lock entry 500b, S is set to 1 and OID is set to R1 to indicate the requester R1 has acquired the lock in memory location Y.
In FIG. 5C2, NC 104 receives a request from requester R2 for access to modify memory location X in EM 106. In response to receiving the request, NC 104, checks the lock entry 500b stored in memory location Y corresponding to memory location X. In lock entry 500b, S is 1 and OID is R1 indicating that the lock is currently owned by requester R1. Thus, the corresponding memory location X cannot be modified by requester R2. Accordingly, NC 104 rejects the request from requester R2 to modify memory location X. Requester R2 will have to try again to gain access to modify memory location X.
In FIG. 5C3, requester R1 releases memory location X. NC 104 sets S as 0 thereby changing lock entry 500b to lock entry 500c. Since the lock is released, memory location X in EM 106 is again available for modification. Upon again receiving a request from requester R2 to modify memory location X, NC 104 checks lock entry 500c. In lock entry 500c, since S is 0, the lock can be acquired. Accordingly, NC 104 grants access to requester R2 to modify memory location X. Either NC 104 or requester R2 modifies lock entry 500c to lock entry 500d by setting S to 1 and OID to R2. Lock entry 500d now indicates that the lock in memory location Y of OCM 108 is acquired and requester R2 has current rights to modify memory location X in EM 106.
FIG. 5C4 illustrates a flowchart 510 showing steps performed by a node controller in test-and-set mode, according to an embodiment of the invention.
In step 512, a request to modify a first memory location is received. In an example, a node controller receives the request from a requester that may be a SOC, a processor, a thread or process.
In step 514, an entry in a second memory location corresponding to the first memory location is checked to determine whether the first memory location is available for modification. In an example, a node controller looks up a table for the second memory location corresponding to the first memory location. The second memory location may be in an on chip memory. The data in the second memory location may be a lock entry with a field that indicates whether the first memory location is available for modification.
In step 516, if it is determined in step 514 that the first memory location is not available for modification then the requester is denied access to modify the first memory location. The first memory location is indicated as being locked.
In step 518, if it is determined in step 514 that the first memory location is available for modification then the requester is granted access to modify the first memory location. If there were multiple requesters in step 512 for the first memory location, then the node controller grants access to the first requester to modify the first memory location.
In step 520, the entry in the second memory location is modified to indicate that the first memory location is locked. The ID of the requester that has current access to modify the first memory location is also stored.
In optional step 522, if there were multiple requesters in step 512, then the remaining requesters are denied access to modify the first memory location.
FIGS. 5D1-5D3 illustrate example operation of memory locks in sequencer mode according to an embodiment of the invention. In the examples presented in FIGS. 5D1-5D3, the “Mode” field is set as 1 and the set field “S” is not applicable (NA), since test-and-set lock mode is not in use. The R bit is set as 0 to indicate that the lock is not in recovery mode.
In FIG. 5D1, NC 104 receives a request from a requester R1 to modify a memory location X in EM 106. The example shown in FIG. 5D1 is an initial startup condition where R1 is the first requester to request access to modify memory location X and there are no prior requesters. Due to initial conditions, when NC 104 checks lock 500e in memory location Y of OCM 108, the current ticket, next ticket and OID fields are Not Applicable (NA). In response to receiving the request, NC 104 grants access to requester R1 to modify memory location X and assigns R1 current ticket value of, for example, “A”. NC 104 or requester R1 updates lock entry 500e to lock entry 500f that indicates the current ticket as A, next ticket as A+1 and OID as R1 to indicate that requester R1 has current ownership of the lock in memory location Y of OCM 108 and rights to modify memory location X in EM 106.
In FIG. 5D2, NC 104 receives a request from requester R2 to modify memory location X in EM 106. In response to receiving the request, NC 104, checks the lock entry 500f stored in memory location Y corresponding to memory location X. In lock entry 500f, current ticket is A, and OID is R1 indicating that the lock is currently owned by requester R1. NC 104 denies access to requester R2 and grants R2 the next ticket value of A+1. When requester R1 with ticket A releases the lock, requester R2 with ticket A+1 will be granted access to modify memory location X. NC 104 or requester R2 changes lock entry 500f to lock entry 500g to indicate the next ticket as A+2. The next requester after R2 will be granted ticket A+2. Thus a queue of requesters is maintained while memory location X is unavailable for modification.
In FIG. 5D3, requester R1 releases memory location X. Since R2 is in next in the queue and has ticket A+1, R2 is granted access to memory location X. Lock entry 500g is updated to entry 500h that indicates the current ticket as A+1 and OID as R2, the current owner of the lock. If requester R2 does not take ownership of the memory location X, for example, because it no longer needs to modify memory location X or if R2 malfunctions, then ownership is granted to the next requester in queue.
It is to be appreciated that in the event that a requester fails to release a memory lock, the lock can be re-acquired by setting R as 1 and initiating recovery mode.
FIG. 5D4 illustrates a flowchart 530 showing steps performed by a node controller in sequencer mode, according to an embodiment of the invention.
In step 532, a request to modify a first memory location in a first memory is received. In an example, a node controller receives the request from a requester that may be a SOC, a processor, a thread or a process. The first memory may be a RAM memory external to the SOC. The node controller may receive multiple requests from multiple requesters for access to modify the first memory location.
In step 534, an entry in a second memory location corresponding to the first memory location is checked to determine whether the first memory location is available for modification. In an example, to determine the corresponding second memory location, a node controller looks up a table that includes the second memory location corresponding to the first memory location. The second memory location may be in a second memory such as, for example, on chip memory. Alternatively, the second memory location may be in the first memory. The data in the second memory location may be a lock entry as described above. The data in the lock entry, such as the OID field of the lock entry may indicate whether there is a current owner of the first memory location. Alternatively, the current ticket and next ticket field may indicate that there are requesters in queue ahead of the current requester from step 532.
In step 536, if it is determined in step 534 that the first memory location is not available for modification then, the requester from step 532 is assigned a ticket number. The requester is thereby inserted in a queue and will access the first memory location according to order of the assigned ticket number.
In step 538, the entry in the second memory location corresponding to the first memory location is updated to change the next ticket number.
In step 540, if it is determined in step 536 that the first memory location is available for modification then, the requester from step 532 is granted access to the memory location. The requester from step 532 is also assigned a current ticket value from a current ticket field in the lock. The requester may use the current ticket value to release the lock on the first memory location after it has completed modification of the first memory location.
In step 542, the entry in the second memory location is updated to reflect the requester from step 532 as the current owner of the lock and the current ticket holder.
FIG. 5D5 illustrates a flowchart 550 showing steps performed by a requester requesting access to modify a memory location, according to an embodiment of the invention.
In step 552, a request to modify a memory location is sent. The request may be sent to a remote node controller controlling access to the memory location. In an example, the requester may be a SOC, a processor, a thread or a process.
In step 554, it is determined whether access is received to modify the memory location.
In step 556, if access was denied to modify the memory location in step 554, then a ticket is received. The ticket may be received from the node controller. The ticket may indicate the position of the requester in a queue to access the memory location. The requester waits in queue for its turn to modify the memory location.
In step 558, the requester receives access according to the ticket number received in step 556. The requester can now modify the memory location.
In step 560, the requester releases the lock thereby allowing the next requester in queue to access the memory location. In an example, the requester transmits its ticket number to the node controller which in turn releases the lock and grants access to the next requester in queue.
In step 562, if access is received to modify the memory location in step 554, then the requester modifies the memory location. The requester then releases the lock in step 560 as described above.
In the sequencer mode of operation, if a controller receives multiple requests to modify the same memory location, where the requests are closely spaced in time, the sequencer orders the requests in order of arrival and then assigns tickets in that order. In an alternate embodiment, the controller assigns tickets in order of a priority level of a request. The priority level may be a positive number sent by the requester along with the request. Higher priority numbers may indicate higher priority. In yet another embodiment, requesters are ordered using a combination of arrival time and priority level. For example, if arrival times and priority levels are each assigned a numeric weight, then a linear combination of numeric weights may be used to order requesters and assign tickets.
In an embodiment, node controller 104 is under the control of software. In this embodiment, node controller 104 only supports mutual exclusion of a memory location. It is up to the software for associating a lock to a memory location that needs to be protected. The software knows that the hardware exports n memory locations that can act as locks. The software then allocates and associates a particular lock for data objects to be protected. In this embodiment, node controller 104 simply performs the needed operations for supporting the semantics of test-and-Set or sequencer lock modes. This simplifies the hardware for node controller 104. The hardware does not need to do the bookkeeping for the association between a lock and the memory location being protected by the lock.
With reference to
In the above pseudo-code, the Test-and-Set command causes node controller 104 to return the present value of the Set field in the entry stored in memory location Y of OCM 108. The code loops until the lock is free i.e. the value returned by node controller 104 is 0 indicating the lock is free for requester R to acquire. The Test-and-Set command also causes node controller 104 to change the Set field of the lock in memory location Y back to 1. Change of the Owner Identification field to indicate requester R as the current owner is done only if the present value in the Set field is 0. Requester R accesses and modifies memory location X. The Unlock command changes the Set field back to 0 to allow another requester to access memory location X.
With reference to
In the example pseudo-code above, the Get_Ticket( ) command causes node controller 104 to return a next ticket number from an entry in memory location Y of OCM 108. The ticket is assigned to requester R and stored in the variable “ticket”. The Read_Current_Ticket( ) command causes node controller 104 to return the current ticket value stored in memory location Y in OCM 108. Requester R keeps checking the current ticket value stored in memory location Y until the current ticket value in memory location Y equals the ticket value assigned to requester R. After gaining access to memory location X and modifying it, requester R issues an Increment_Ticket( ) command to increment the current ticket field in memory location X and allow the next requester in line to acquire the lock in memory location Y and access memory X.
In the embodiments shown in
In this document, the terms “computer program medium” and “computer usable medium” are used to generally refer to media such as a removable storage unit, a hard disk installed in hard disk drive, and signals (i.e., electronic, electromagnetic, optical, or other types of signals capable of being received by a communications interface). These computer program products are means for providing software to a computer system. The invention, in an embodiment, is directed to such computer program products.
In an embodiment where aspects of the present invention are implemented using software, the software may be stored in a computer program product and loaded into a computer system (e.g., a reader or host) using a removable storage drive, hard drive, or communications interface. The control logic (software), when executed by a processor, causes the processor to perform the functions of the invention as described herein. Still further, a sensor may execute computer readable instructions to collect data. Still further, a near field transceiver may execute computer readable instructions to communicate with sensors and/or tags.
According to an example embodiment, a reader may execute computer-readable instructions to read tags, as described above. Furthermore, in an embodiment, a tag may execute computer-readable instructions to respond to a reader transmitted signal, as further described elsewhere herein.
While various embodiments of the present invention have been described above, it should be understood that they have been presented by way of example only, and not limitation. It will be apparent to persons skilled in the relevant art that various changes in form and detail can be made therein without departing from the spirit and scope of the invention. Thus, the breadth and scope of the present invention should not be limited by any of the above-described exemplary embodiments, but should be defined only in accordance With the following claims and their equivalents.
This application is related to commonly owned U.S. patent application Ser. No. ______ , titled “Memory Management in a Shared Memory System,” filed on the same day herewith (Attorney Docket No. 1875.9040000), and U.S. patent application Ser. No. ______ , titled “Communication in a Cluster System,” filed on the same day herewith (Attorney Docket No. 1875.9040001), and U.S. patent application Ser. No. ______ , titled “Hardware Memory Locks,” filed on the same day herewith (Attorney Docket No. 1875.9040002), each of which is incorporated herein by reference in its entirety.
