Ai server hard disk backplane data interface management system

Abstract

The application relates to the technical field of computer interface management and data processing, in particular to an AI server hard disk backboard data interface management system. The system collects link state, protocol error, transaction time sequence and topology correlation data of a PCIe interface, establishes a protocol state shadow, generates an abnormal event fragment and a cause-effect fingerprint representing an abnormal source and a propagation path, constructs a candidate recovery control sequence according to the cause-effect fingerprint, simulates the influence of the sequence on an unfinished transaction, an adjacent interface and a device enumeration state, forms an expected recovery fingerprint, selects a recovery sequence meeting a transaction continuity constraint and having a small influence range to execute, and determines subsequent control according to the matching result of the state and the expected recovery fingerprint. The application can distinguish local link degradation, device response instability and shared path blockage, reduces transaction cross-border and enumeration context conflict, and maintains transaction continuity and state consistency in the recovery process.

Description

Technical Field

[0001] This invention relates to the field of computer interface management and data processing technology, specifically to an AI server hard drive backplane data interface management system. Background Technology

[0002] Artificial intelligence servers typically connect multiple high-speed solid-state drives (SSDs) via a hard drive backplane. Each SSD corresponds to a high-speed serial computer expansion bus interface and connects to the processor via a switching path or root port. Existing interface management schemes generally involve the operating system, bus driver, and backplane management program working together to continuously read link speed, link width, training status, correctable errors, uncorrectable errors, replay count, completion timeout, and device enumeration status. When the error count of a certain interface exceeds a preset value or the link status is interrupted, the management program usually executes pause command issuance, retraining the link, reducing the transmission rate, reducing the link width, functional level reset, or re-enumerating the device in a fixed order. Some schemes also combine hard drive command submission time, completion time, and queue depth to determine whether the device has timed out and restore the original mapping based on the bus address, device number, or disk bay number. The control basis of such schemes is mainly a single count value, a single status change, or a preset fault code. The recovery action lacks a unified correlation with the abnormal propagation path, the status of incomplete transactions, and the shared path relationship between adjacent interfaces. The recovery result is usually only judged by the interface coming back online or the device being re-identified.

[0003] Faced with situations where local link degradation, device response instability, and shared path blockage exhibit similar characteristics, management programs struggle to determine the origin and propagation range of anomalies. Consequently, they cannot determine whether a control sequence will cross incomplete transaction boundaries, alter the state of adjacent fault-free interfaces, or disrupt the device enumeration context before performing recovery. When link retraining or reset is triggered solely based on error counts, transient disturbances may be amplified into port interruptions. When device timeouts are directly re-enumerated, completed transactions still in the return process may enter new device contexts. When continuous degradation and reset are performed in a fixed order, the same invalid action may be repeatedly invoked, creating control oscillations. The core problem arising from this is that existing interface management lacks a control mechanism based on anomaly causality and verifies the scope of the recovery sequence's impact before actual execution. This results in recovery actions failing to simultaneously maintain transaction continuity, adjacent interface isolation, and enumeration state consistency. Summary of the Invention

[0004] The purpose of this invention is to provide an AI server hard disk backplane data interface management system, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The AI ​​server hard disk backplane data interface management system includes a processor and a memory. The memory stores instructions, which instruct the processor to: collect the link status, protocol errors, transaction timing and topology association data of each PCIe interface, and establish a protocol status shadow corresponding to each PCIe interface. Based on the protocol state shadow, generate abnormal event fragments and causal fingerprints representing the source and propagation path of the abnormality; Based on the causal fingerprint, candidate recovery control sequences are generated. The impact of each candidate recovery control sequence on incomplete transactions, adjacent interfaces, and device enumeration states is simulated in the protocol state shadow to form the corresponding expected recovery fingerprint. The candidate recovery control sequence that satisfies the transaction continuity constraint and has the smallest impact range is selected for execution, and the subsequent control sequence is determined based on the matching result between the actual state after execution and the expected recovery fingerprint.

[0006] Preferably, the protocol state shadow stores the link rate, link width, training status, correctable error, uncorrectable error, replay event, completion timeout, NVMe command submission time, command completion time, and number of incomplete commands according to the interface identifier, and merges data with the same transaction identifier, the same path identifier, or previous and subsequent state transition relationships in adjacent sampling periods into protocol events. The processor uses link state change events as time anchors to sort and associate protocol events before and after the time anchors, forming anomaly event fragments that include event subject, event type, occurrence order, duration range, and associated path.

[0007] Preferably, the processor establishes a topology propagation graph based on the logical connection relationship between the hard disk, the downstream PCIe interface, the PCIe switching path, the upstream PCIe interface, and the root port, projects the abnormal event fragments onto the topology propagation graph, and marks the propagation direction between adjacent nodes according to the order of event occurrence. Based on the event synchronization degree of multiple downlink interfaces under the same uplink path, the first occurrence of single interface errors, and the spread range of transaction completion delays, a causal fingerprint is generated, which includes the abnormal starting node, the set of propagation edges, event sequence constraints, and evidence consistency.

[0008] Preferably, the candidate recovery control sequence is formed by arranging at least two of the following according to control dependencies: queue command suppression, transaction silencing, function-level reset, link retraining, link rate degradation, link width degradation, and logical data path switching. The processor establishes a state transition copy for each candidate recovery control sequence, updates the incomplete transaction set, interface link status, adjacent interface shared path occupancy status and device enumeration context item by item in the state transition copy, and marks candidate recovery control sequences that cross transaction boundaries, cause irrelevant interface status changes or cause enumeration context conflicts as non-executable sequences.

[0009] Preferably, the processor sets up a pre-observation area, a core anomaly area, and a post-response area for each anomaly event segment, takes the earliest protocol error or completion delay in the pre-observation area as a candidate triggering event, takes the PCIe interface in the core anomaly area that undergoes a state change as a candidate starting interface, and layers the associated events of other interfaces in the post-response area according to topological distance. When the same event satisfies both transaction identifier association and path propagation association, its dual association marker is retained. When the two types of associations point to different starting interfaces, the corresponding abnormal event fragment is split into multiple sub-fragments to be determined and written into the protocol state shadow respectively.

[0010] Preferably, the processor constructs an interface event association matrix based on the topology propagation graph. The rows and columns of the interface event association matrix correspond to the PCIe interface, and the matrix elements record the event time difference, shared path level, and error type consistency. The set of events that occur only on a single downlink interface and do not spread to the shared node is encoded as a local link degradation fingerprint. The set of events that occur before the link error is encoded as a device response instability fingerprint. The set of events that occur when multiple downlink interfaces form a synchronization delay on the same uplink path and the error originates at the shared node is encoded as a shared path blocking fingerprint.

[0011] Preferably, when generating the candidate recovery control sequence, the processor first determines the transaction quiescent boundary of each PCIe interface based on the set of incomplete transactions, and marks control actions that only change the command issuance state or logical path mapping as reversible actions, and marks control actions that cause link state renegotiation or device context reconstruction as state reconstruction actions. The candidate recovery control sequence is arranged according to the constraints that the reversible action precedes the state reconstruction action, the local interface action precedes the shared path action, and the transaction silent confirmation precedes the reset or retraining action. Each state reconstruction action is associated with a corresponding set of preceding transactions and a set of subsequent enumeration verification items.

[0012] Preferably, the processor retains the commit interface, target namespace, completion status, and data buffer occupancy flag for each incomplete transaction in the state transition copy, and sequentially calculates the transaction cross-boundary flag, adjacent interface linkage flag, enumeration context conflict flag, and control reversibility flag for each candidate recovery control sequence. The sequence to be executed is determined by first excluding sequences that have the transaction cross-boundary flag or the enumeration context conflict flag, then comparing the linkage range of adjacent interfaces, and finally comparing the number of irreversible control actions. The candidate recovery control sequences that were not selected and their corresponding state transition copies are kept as alternative branches.

[0013] Preferably, the processor maintains a fingerprint version and evidence source set for each causal fingerprint. When a newly added protocol event is consistent with the current anomaly starting node but the propagation edges are inconsistent, only the propagation edge set is updated. When the newly added protocol event causes a change in the candidate starting node, the order of the event reconstruction is constrained by the original candidate starting node and the new candidate starting node, respectively. If one set of event sequence constraints maintains continuous time anchors, consistent shared paths, and continuous transaction identifiers, while another set of event sequence constraints lacks at least two of these, then the corresponding fingerprint version of the other set is deleted. The processor inputs the retained fingerprint versions into the generation process of the candidate recovery control sequence, and merges the same control prefix generated by different fingerprint versions.

[0014] Preferably, after executing the sequence to be executed, the processor generates an actual recovery state sequence according to the data fields that are the same as the protocol state shadow, and compares it with the expected recovery fingerprint corresponding to the sequence to be executed event by event; When the actual recovery state sequence deviates from the expected recovery fingerprint at the predetermined control step, the subsequent state reconstruction action is stopped, the reverse control of the registered reversible action is executed, the candidate recovery control sequence that matches the current actual state prefix in the alternative branches is called, and the candidate recovery control sequence containing the same causal fingerprint and the same control prefix is ​​set as the disabled branch.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By collecting link status, protocol errors, transaction timing, and topology correlation data from each high-speed serial computer expansion bus interface, corresponding protocol status shadows are established. Abnormal event fragments are mapped to the logical connection relationships between hard disks, downlink interfaces, switching paths, uplink interfaces, and root ports, forming a causal fingerprint that includes the abnormal initiation node, propagation edges, and event sequence constraints. After simulating state transitions in the protocol status shadows, candidate recovery control sequences can pre-identify transaction cross-boundaries, adjacent interface linkages, and enumerated context conflicts, and select the sequence to be executed according to transaction continuity constraints and the scope of influence. Abnormal identification, sequence simulation, and recovery selection form a closed loop, enabling local link degradation, device response instability, and shared path blocking to be distinguished based on event propagation characteristics, and limiting recovery control to the interface and path range that matches the source of the abnormality.

[0016] 2. The protocol state shadow synchronously saves the set of incomplete transactions, interface link status, shared path occupancy status, and device enumeration context, enabling queue command suppression, transaction silencing, function-level reset, link retraining, link degradation, and logical path switching to be arranged according to control dependencies. After comparing the actual recovery state sequence with the expected recovery fingerprint event by event, subsequent state reconstruction actions can be stopped when the state deviates, reversible actions that have been registered can be revoked, and alternative control branches that match the current state prefix can be called. This processing method can reduce the continuous execution of invalid control sequences, block repeated calls of the same causal fingerprint and the same control prefix, maintain the correspondence between incomplete transaction records and device enumeration status, and reduce the linkage impact of the recovery process on fault-free interfaces under the shared path. Attached Figure Description

[0017] Figure 1 A flowchart illustrating the overall process of autonomous recovery of the hard drive backplane data interface for AI servers. Figure 2 This is a flowchart illustrating the process of associating protocol exception events and generating causal fingerprints. Detailed Implementation

[0018] In one embodiment, please refer to Figure 1A data interface management system for the hard disk backplane of an AI server is provided, including a processor and a memory. The memory stores instructions that enable the processor to collect link status, protocol errors, transaction timing, and topology association data of each PCIe interface, establish a protocol status shadow corresponding to each PCIe interface, generate abnormal event fragments and causal fingerprints representing the source and propagation path of the abnormality based on the protocol status shadows, generate candidate recovery control sequences based on the causal fingerprints, simulate the impact of each candidate recovery control sequence on incomplete transactions, adjacent interfaces, and device enumeration status in the protocol status shadows, form corresponding expected recovery fingerprints, select the candidate recovery control sequence that satisfies the transaction continuity constraint and has the smallest impact range for execution, and determine the subsequent control sequence based on the matching result between the actual state after execution and the expected recovery fingerprint.

[0019] In this embodiment, the processor runs an interface acquisition process, a state shadow maintenance process, a causal analysis process, and a recovery control process. The interface acquisition process receives state records generated by the root port, switching path, and downlink port using a unified time base, and encapsulates each state record into a data item containing an interface identifier, device identifier, acquisition time, event category, event value, and path identifier. The state shadow maintenance process writes the data item into the corresponding circular event area according to the interface identifier, and saves the current configuration space summary, incomplete transaction index, and device enumeration context. The causal analysis process only reads data items that have been time-aligned, and the recovery control process only executes control sequences that have been simulated by the shadow. Thus, acquisition, inference, and control are applied to specific data versions.

[0020] Specifically, the link status includes link active status, training status, negotiation rate, negotiation width, and power management status; the protocol errors include correctable errors, uncorrectable errors, replay events, completion timeouts, malformed transactions, and unsupported requests; the transaction timing includes command submission time, data transmission start time, completion return time, and transaction retry time; the topology association data includes the downlink interface corresponding to the hard disk, the switching path traversed, the shared uplink interface, and the root port; the processor uniformly represents the time of each source using a monotonically increasing internal timescale, and corrects short-term sampling disorder based on the continuity of status in adjacent sampling records on the same interface.

[0021] The protocol state shadow is not a simple copy of the real interface's state. Instead, it stores a traceable state sequence with the interface as the primary key and attaches an event set to each state version to form that state version. When the processor receives a new data item, it copies the current valid version and updates the corresponding fields of the data item to the copy. After the data item completes time consistency verification and path consistency verification, the copy is submitted as the new valid version. Copy that fails the verification is kept in the temporary storage area and waits for subsequent data to be completed, thus avoiding misaligned events from directly changing the state basis on which the anomaly judgment is based.

[0022] In this embodiment, an abnormal event segment consists of multiple events with a continuous temporal relationship on the same interface or the same path. The processor selects the segment boundary from changes in link state, changes in error category, or sudden changes in transaction delay. Within the segment boundary, the events are sorted according to their time. Events related to the same command identifier, the same transaction tag, or the same device enumeration instance are horizontally associated. The causal fingerprint records at least the candidate abnormal starting interface, the abnormal starting event, the propagation direction, the set of interfaces reached by the propagation, the event sequence constraints, and the evidence source corresponding to each judgment conclusion.

[0023] To ensure that event association has a computable criterion, the processor calculates an event association value for any two events to be associated: ; in, This represents the association value between event a and event b. Indicates the temporal proximity of two events. Indicates the degree of overlap between the paths of two events. Indicates the consistency of transaction identifiers. Indicates the compatibility of state transitions for event categories. , , and They are respectively represented as , , and The corresponding non-negative weights sum to 1.

[0024] In a preferred embodiment, if event a is a continuous increase in the number of replays on a certain downlink interface, and event b is a subsequent increase in the completion delay on the same path, and the interval between the two events is within the same observation window, then it can be set as follows: , , , When the four weights are set to 0.3, 0.3, 0.1, and 0.3 respectively, we get... The processor merges the two events into the same abnormal event fragment; if the two events come from different uplink paths and there is no transaction identifier association, the path overlap and transaction identifier consistency are both set to 0, the association value is lower than the fragment merging condition, and the processor keeps the two events independent of each other.

[0025] When the processor generates candidate recovery control sequences, it does not directly map fault codes to fixed actions. Instead, it establishes a control constraint graph based on the abnormal initiation interface, propagation edge, and affected transaction set in the causal fingerprint. The nodes in the control constraint graph represent queue command suppression, transaction silencing, functional level reset, link retraining, link rate degradation, link width degradation, or logical data path switching. The directed edges between nodes indicate that the previous control action must be completed before the next control action. The processor performs a directed acyclic expansion on the control constraint graph to obtain multiple candidate recovery control sequences that satisfy control dependencies.

[0026] The transaction continuity constraint is jointly determined by whether the incomplete transaction still has a unique device context, whether the data buffer is in a verifiable state, whether the completion return can belong to the original commit interface, and whether the control action spans device enumeration generations. The processor establishes a state transition copy for each candidate recovery control sequence, copies the current version of the protocol state shadow as the initial state, and executes the logical state transitions corresponding to the candidate recovery control sequences in the copy without sending control commands to the real interface. After each state transition, the incomplete transaction set, interface link state, adjacent interface shared path occupancy state, and device enumeration context are updated.

[0027] Preferably, the expected recovery fingerprint consists of key state changes generated in the state transition copy of the candidate recovery control sequence. Each key state change includes at least a control action identifier, the expected occurrence time interval, the target interface, the expected link state, the expected change in the number of incomplete transactions, the expected adjacent interface state, and the expected enumeration context version. The processor compresses continuous intermediate states that do not affect control judgment into interval records, and retains states that change transaction ownership or device context as independent records so that the execution phase can compare the real state at the event granularity.

[0028] When selecting candidate recovery control sequences, the processor first excludes sequences that violate transaction continuity constraints, then compares the number of interfaces affected by each sequence, the number of shared paths changed, the number of irreversible control actions, and the number of enumerated contexts that need to be rebuilt. If the comparison results are the same, the sequence with fewer control actions is selected. The minimum impact range is not determined by a single numerical condition, but by comparing the transaction boundary, the topology diffusion range, the device context change range, and the ordered conditions formed by control reversibility in sequence.

[0029] During the execution phase, the recovery control process issues control commands item by item according to the selected sequence. After each control action is completed, the corresponding actual state is read from the interface acquisition process, written into an independent recovery execution area, and compared with the expected state after the same control action in the expected recovery fingerprint. If they match, the next control action is executed. If they do not match, the state reconstruction action that has not yet been executed is paused, and the current actual state is used as the new state transition starting point for selecting alternative control sequences or generating subsequent control sequences.

[0030] In one embodiment, the protocol state shadow stores link rate, link width, training status, correctable errors, uncorrectable errors, replay events, completion timeouts, NVMe command submission time, command completion time, and number of incomplete commands according to the interface identifier. It also merges data with the same transaction identifier, the same path identifier, or a state transition relationship within adjacent sampling periods into protocol events. The processor uses the link state change event as a time anchor point to sort and associate the protocol events before and after the time anchor point, forming an abnormal event fragment containing the event subject, event type, occurrence order, duration interval, and associated path.

[0031] In this embodiment, the protocol state shadow of each interface adopts a storage format in which state pages, event pages, and transaction pages reference each other. The state page stores the link and error state that can be directly read at a certain moment, the event page stores the original record that caused the state change and its time sequence, and the transaction page stores the command that has not yet been completed and its data buffer state. The state page points to the event page that formed the current state through the event index, and the event page points to the relevant transaction page through the transaction index. Thus, even after the abnormal segment is formed, the command and data path corresponding to each state change can still be determined in reverse.

[0032] Data merging between adjacent sampling periods is performed using an event closure approach. The processor backtracks from the time anchor point until no data with transaction identifier, path identifier, or state transition association with the anchor point appears in two consecutive sampling periods. Then it extends backward from the time anchor point until no more associated events are added. The backtracking interval and the extended interval together form the initial fragment. Subsequently, data that only has time proximity but no transaction, path, or state transition association is deleted, so that the abnormal event fragment maintains the minimum event set required for causal analysis.

[0033] Preferably, the event subject is jointly represented by the interface identifier, device identifier, and enumeration context version. The event type is divided according to the link layer state change, transaction layer exception, device response exception, and enumeration state change. The occurrence order is represented by a partial order relationship under the same time base. The duration interval is defined by the time of the first associated event and the last associated event. The associated path records the logical nodes passed between the event subject and the root port. When two events have the same time or the collection order is uncertain, the processor establishes a partial order edge based on the prerequisites of the protocol state transition without forcibly assigning an absolute order.

[0034] In a preferred embodiment, if the training state changes from normal to recovery state, and at the same time the replay events increase and the command completion time is delayed, the processor uses the change in training state as a time anchor point and adds the replay events in the previous observation interval and the completion delay in the next observation interval to the same initial segment; if another interface has an independent correctable error but does not share a path and has no transaction association at the same time, the event will not enter the current segment, but will be reserved as another segment to be analyzed.

[0035] Further, please refer to Figure 2 The processor establishes a topology propagation graph based on the logical connection relationship between the hard disk, downlink PCIe interface, PCIe switching path, uplink PCIe interface and root port. It projects the abnormal event fragments onto the topology propagation graph and marks the propagation direction between adjacent nodes according to the order of event occurrence. Based on the event synchronization degree of multiple downlink interfaces under the same uplink path, the first occurrence of single interface errors and the diffusion range of transaction completion delay, it generates the causal fingerprint, which includes the abnormal starting node, the propagation edge set, event sequence constraints and evidence consistency.

[0036] The nodes in the topology propagation graph are divided into device nodes, downlink interface nodes, switching path nodes, uplink interface nodes, and root port nodes. The edges between nodes are established based on the actual enumeration relationship and routing relationship. The processor attaches the node to each abnormal event, projects the events in the same event segment to the corresponding nodes according to the partial order relationship, and then establishes candidate propagation edges for the sequential events on adjacent nodes. The direction of the candidate propagation edges is from the node where the earlier event occurs to the node where the later event occurs. If the sequential relationship between the events of two nodes is not determined, the candidate propagation edge remains undirected and waits for subsequent evidence.

[0037] The processor calculates the initial confidence value for the candidate anomaly initiation node: ; in, This represents the credibility value of node v as the starting node of the anomaly. This indicates the order in which the earliest abnormal event occurred on node v. This represents the consistency of events propagating from node v to neighboring nodes. This represents the path reachability between node v and subsequent completion delays. This indicates the degree to which node v is interpreted by other candidate starting nodes. , , and They are respectively represented as , , and The corresponding non-negative coefficient.

[0038] For example, if a downlink interface first experiences continuous replay, followed by a completion delay on the same path, and the shared uplink interface does not experience an anomaly before the downlink interface, the pre-emptiveness can be set to 1, the diffusion consistency to 0.8, the path reachability to 1, and the degree of interpretation by other nodes to 0.1, with coefficients of 0.35, 0.25, 0.25, and 0.15 respectively. The corresponding initial confidence value is 0.785. If the shared uplink interface experiences flow control stagnation synchronously before multiple downlink interfaces, the pre-emptiveness and diffusion consistency of the shared uplink interface increase, and the processor adjusts the starting node from a single downlink interface to the shared uplink interface accordingly.

[0039] Evidence consistency is formed by time sequence evidence, topological path evidence, transaction identifier evidence, and state transition evidence. The processor does not directly consider the absence of a certain type of evidence as negation, but records the missing type and its conditions for completion. When subsequent sampling obtains the completion information of the same transaction or the new error of the same path, the source of evidence is added to the cause-effect fingerprint and the propagation edge set is updated. The fingerprint version used for execution control is kept read-only so that the judgment basis used at that time can be traced back when the deviation occurs.

[0040] In one embodiment, the candidate recovery control sequence is formed by arranging at least two of the following according to control dependencies: queue command suppression, transaction silencing, function-level reset, link retraining, link rate degradation, link width degradation, and logical data path switching. The processor establishes a state transition copy for each candidate recovery control sequence, updates the set of incomplete transactions, interface link status, adjacent interface shared path occupancy status, and device enumeration context item by item in the state transition copy, and marks candidate recovery control sequences that cross transaction boundaries, cause irrelevant interface status changes, or cause enumeration context conflicts as non-executable sequences.

[0041] Specifically, queue command suppression means stopping the submission of new commands to the target interface while preserving the completion channel of already submitted commands; transaction silencing means waiting for or converging submitted transactions until the boundary where the link state can be safely changed is reached; function-level reset means resetting the device context corresponding to the target function; link retraining means renegotiating the link state; link rate degradation and link width degradation respectively change the negotiated transmission level and the number of effective channels; logical data path switching means mapping subsequent migrated access to other pre-existing logical paths; and the executable prerequisites and state change results of each control action are saved in the control rule table.

[0042] When the processor builds a state transition copy, it calls the corresponding state transition function for each control action. The state transition function reads the current interface state, the set of incomplete transactions, and the enumeration context from the copy, and generates the next state and new constraints. For example, queue command suppression changes the permission state of new commands to prohibit but does not delete incomplete transactions; transaction silencing generates a silencing flag after incomplete transactions meet the completion, cancellation, or isolation conditions; function-level reset generates a new function context version and puts the old version of transactions into the set to be verified; and link retraining generates the expected training process and negotiation results without directly assuming training success.

[0043] The comprehensive constraint value of the candidate recovery control sequence is calculated according to the following formula: ; in, Indicates candidate recovery control sequence The comprehensive constraint value, Indicates cross-boundary items in transactions. Indicates the number of adjacent interfaces affected. This indicates the number of device enumeration contexts that have been modified. This indicates the number of irreversible control actions. , , and A positive coefficient set according to constraint priority.

[0044] In this embodiment, the transaction cross-boundary item is set to 1 if it exists and 0 if it does not exist, and the corresponding coefficient is greater than the sum of the coefficients of the other three items, so that any sequence that crosses the safety boundary of an incomplete transaction will not be selected because the other items are smaller; for example, if the transaction cross-boundary item of sequence A is 1, the number of adjacent interfaces is 0, the number of enumeration contexts is 0, and the number of irreversible actions is 1, and the transaction cross-boundary item of sequence B is 0, the number of adjacent interfaces is 1, the number of enumeration contexts is 0, and the number of irreversible actions is 1, when the transaction cross-boundary item coefficient is dominant, the processor retains sequence B and marks sequence A as non-executable.

[0045] The shared path occupancy status of adjacent interfaces in the state transition replica is updated together based on the current incomplete transactions and expected control actions. When a candidate sequence contains a shared uplink interface reset or an action that will change the shared switching path, the processor searches along the topology propagation graph for all interfaces that depend on the shared node and marks the link state of the interface in the replica as potentially changing. If the causal fingerprint does not include the corresponding interface in the affected range, the candidate sequence is marked as causing an irrelevant interface state change.

[0046] The device enumeration context is represented by the enumeration version, device identifier, interface path, namespace set, and command attribution table. When a candidate control action triggers re-enumeration, the processor creates a new context to be committed in the copy without overwriting the original context, and checks whether the incomplete transactions in the old context have a transferable attribution relationship. If the old transaction cannot be uniquely attributed to the original device or the device to be committed, an enumeration context conflict flag is formed and the subsequent simulation of the candidate sequence is terminated.

[0047] In a preferred embodiment, the processor sets up a pre-observation area, a core anomaly area, and a post-response area for each anomaly event segment. The earliest protocol error or completion delay in the pre-observation area is taken as a candidate triggering event, the PCIe interface in the core anomaly area where a state change occurs is taken as a candidate starting interface, and the associated events of other interfaces in the post-response area are layered according to topological distance. When the same event satisfies both transaction identifier association and path propagation association, its dual association mark is retained. When the two types of association point to different starting interfaces, the corresponding anomaly event segment is split into multiple sub-segments to be determined and written into the protocol state shadow respectively.

[0048] The pre-observation area extends forward from the starting point of the core anomaly area to the location where the continuity of related events is interrupted. The core anomaly area covers the period during which the link state or transaction state changes abruptly. The post-response area extends backward from the ending point of the core anomaly area to the location where propagation events no longer increase. The processor saves the event sets and boundary formation basis of the three intervals respectively, so that when new data arrives, the interval boundaries can be locally adjusted without rebuilding the entire historical fragment.

[0049] The determination of candidate triggering events is based on the earliest time and is subject to the interpretability of the event. If the earliest protocol error can be explained by the existing device response anomaly, the processor continues to search for the previous associated event. If the earliest completion delay occurs simultaneously on multiple interfaces sharing the same uplink path, the processor will take the first event on the shared path as the candidate triggering event. The candidate starting interface in the core anomaly zone is determined based on whether the link state change occurred before propagation and whether there is a direct transaction association.

[0050] Events in the post-response area are layered according to the shortest topological distance from the candidate starting interface. The layer with a distance of 1 contains the switching path or device node directly adjacent to the candidate starting interface. As the distance increases, the uplink interface and other downlink interfaces are covered in turn. The processor establishes a partial order according to the event timing within the same layer and establishes propagation constraints according to the topological reachability between different layers to avoid interpreting events that do not have path connections as the same abnormal propagation based solely on the time sequence.

[0051] When the transaction identifier association points to the device node and the path propagation association points to the shared uplink node, the processor copies the original abnormal event fragment to form two sub-fragments to be determined. Sub-fragment A retains the events related to the device transaction and the path evidence that can explain the events. Sub-fragment B retains the synchronization events on the shared path and the downlink interface events affected by them. The two sub-fragments calculate the initial confidence value and establish a causal fingerprint, until subsequent evidence causes one of the sub-fragments to lose the necessary continuity conditions.

[0052] Furthermore, the processor constructs an interface event association matrix based on the topology propagation graph. The rows and columns of the interface event association matrix correspond to PCIe interfaces, and the matrix elements record the event time difference, shared path level, and error type consistency. The set of events that occur continuously only on a single downlink interface and do not spread to the shared node is encoded as a local link degradation fingerprint. The set of events that occur before the link error is encoded as a device response instability fingerprint. The set of events where multiple downlink interfaces form a synchronization delay on the same uplink path and the error originates at the shared node is encoded as a shared path blocking fingerprint.

[0053] Each element in the interface event association matrix uses a structured record rather than a single numerical value. The time difference field records the minimum directed time interval between related events of two interfaces, the shared path level field records the topology node level at which the two interfaces first converge, and the error type consistency field records whether the errors of the two interfaces belong to the same protocol category. The processor identifies whether the anomaly is propagated from a single interface, from a shared node to multiple interfaces, or caused by a device response anomaly that triggers a subsequent reaction at the link layer through the row and column patterns in the matrix.

[0054] The processor calculates the confidence level of candidate propagation edges: ; in, This represents the reliability of the propagation edge from node x to node y. Indicates the consistency of the direction of events. Indicates the compatibility of shared paths. Indicates the coherence between error types or transaction states. , and They are respectively represented as , and The corresponding non-negative weights sum to 1.

[0055] For example, if a downlink interface node first experiences a replay increase, and the directly adjacent switching path node subsequently recovers, and both are located on the same path and their event categories have a sequential relationship, then the direction consistency, path compatibility, and sequential connection can all be set to 1, making the propagation edge reliability 1. If the two interface events occur synchronously, but the first convergence node is located at a higher layer and no shared node's first event is found, then the shared path compatibility is reduced, and the processor does not directly establish a propagation edge from one downlink interface to another interface, but instead retains the common upstream node as the starting point for judgment.

[0056] The fingerprint of local link degradation includes a combination of continuous errors of a single downlink interface, asynchronous anomalies of adjacent shared nodes, and other interface transactions maintaining independent timing. The fingerprint of device response instability includes a combination of command completion anomalies preceding link errors, continuous backlog of transactions on the same device, and link state changes occurring after transaction anomalies. The fingerprint of shared path blocking includes a combination of synchronization delays of multiple downlink interfaces, pre-emptive anomalies of common uplink nodes, and propagation direction from shared nodes to downlink nodes.

[0057] In one embodiment, when generating a candidate recovery control sequence, the processor first determines the transaction quiescent boundary of each PCIe interface based on the set of incomplete transactions, and marks control actions that only change the command issuance state or logical path mapping as reversible actions, and marks control actions that cause link state renegotiation or device context reconstruction as state reconstruction actions. The candidate recovery control sequence is arranged according to the constraints that reversible actions precede state reconstruction actions, local interface actions precede shared path actions, and transaction quiescent confirmation precedes reset or retraining actions, and each state reconstruction action is associated with a corresponding set of preceding transactions and a set of following enumeration verification items.

[0058] The transaction silent boundary is established separately for each interface. The processor reads the command type, data direction, completion status and buffer occupancy flag of the incomplete transaction, deletes the transaction that has returned completion information from the boundary set, marks the transaction that can be confirmed for cancellation by the driver layer and has not committed data as a terminateable transaction, and marks the transaction that has entered data transmission but has not yet been completed as a transaction to be converged. Transaction silent confirmation is only generated when there are no transactions to be converged that cannot be confirmed in the boundary set.

[0059] The silent boundary state of a transaction is determined by the following formula: ; in, This indicates the transaction quiescent boundary state of interface m. This represents the set of incomplete transactions for interface m. This represents the boundary acknowledgment value of transaction r. It is 1 when the transaction has been completed, confirmed as cancelled, or isolated and retains unique ownership. It is 0 when the transaction may still change the data buffer or device context. The product of an empty set is defined as 1.

[0060] For example, if an interface has 3 incomplete transactions, 2 of which have returned completion information and the other has been confirmed as cancelled, then all 3 boundary confirmation values ​​are 1, the transaction silent boundary state is 1, and the processor allows subsequent links to be retrained and enter the candidate sequence; if one of the write transactions is still in the data commit state, then its boundary confirmation value is 0, the transaction silent boundary state is 0, and the processor only allows reversible actions that do not change the ownership of the transaction, such as queue command suppression or logical path switching.

[0061] Before execution, reversible actions register reverse control and its applicable conditions. The reverse control for queue command suppression is to restore command permission. The reverse control for logical data path switching is to restore the original logical mapping. State reconstruction actions do not register simple reverse control, but save the enumeration context, link state and transaction set snapshot before the action. The processor only allows the state reconstruction action to enter the execution sequence after the transaction is silently confirmed and the corresponding preceding transaction set is fixed.

[0062] Local interface actions include control actions that only apply to abnormal downlink interfaces or corresponding device functions. Shared path actions include control actions that change the state of uplink interfaces, switch paths, or multiple downlink interfaces that depend on each other. The processor establishes the action scope based on the affected range in the causal fingerprint. Candidate recovery control sequences are only allowed to be added to shared path actions when local scope actions cannot form an executable branch or the causal fingerprint clearly points to a shared node.

[0063] The post-enumeration verification items are bound to each state reconstruction action, including at least device identifier consistency, namespace set consistency, interface path validity, old transaction ownership uniqueness, and new context version continuity. The state transition copy generates the expected verification results item by item after simulating the state reconstruction action. In the actual execution phase, the new device enumeration context is only submitted after all the corresponding verification items have obtained definite results.

[0064] In one embodiment, the processor retains the commit interface, target namespace, completion status, and data buffer occupancy flag for each incomplete transaction in the state transition copy. For each candidate recovery control sequence, the processor sequentially calculates the transaction cross-boundary flag, adjacent interface linkage flag, enumeration context conflict flag, and control reversibility flag. The processor determines the sequence to be executed in the following order: first, excluding sequences with transaction cross-boundary flags or enumeration context conflict flags; then, comparing the linkage range of adjacent interfaces; and finally, comparing the number of irreversible control actions. The unselected candidate recovery control sequences and their corresponding state transition copies are kept as alternative branches.

[0065] The transaction cross-boundary flag is generated when any state reconstruction action in the candidate sequence changes the ownership of a transaction that still has a data buffer occupancy flag. The adjacent interface linkage flag is generated when the candidate sequence changes the state of an interface outside the scope of the causal fingerprint. The enumeration context conflict flag is generated when the same device identifier or namespace is simultaneously associated with two different context versions. The control reversibility flag records whether each action in the candidate sequence has registered reverse control and its execution prerequisites.

[0066] The processor uses a hierarchical filtering approach instead of directly weighting all the tags. Candidate sequences with transaction cross-boundary tags or enumeration context conflict tags are immediately added to the non-executable set. The remaining sequences are arranged in ascending order of the number of interfaces covered by the adjacent interface linkage tags. If the number of interfaces is the same, the number of state reconstruction actions is compared. If the number of state reconstruction actions is still the same, the total number of control actions is compared to keep the priority of the selection logic consistent with the transaction continuity constraints.

[0067] The alternative branch saves the candidate recovery control sequence, state transition copy, corresponding causal fingerprint version, expected recovery fingerprint, and the reason for exclusion or not being selected. The processor will not delete the candidate sequence simply because it is currently ranked lower. Instead, when the actual execution state deviates from the expectation, it will re-examine whether its initial state is compatible with the current actual state prefix. The alternative branch will only be closed when the initial state is no longer reachable or the corresponding causal fingerprint version is refuted by evidence.

[0068] In a preferred embodiment, the processor maintains a fingerprint version and a set of evidence sources for each causal fingerprint. When a new protocol event is consistent with the current anomaly starting node but the propagation edges are inconsistent, only the propagation edge set is updated. When a new protocol event causes a change in the candidate starting node, the event sequence constraints are reconstructed with the original candidate starting node and the new candidate starting node, respectively. If one set of event sequence constraints maintains continuous time anchors, consistent shared paths, and continuous transaction identifiers, while another set of event sequence constraints is missing at least two of these, then the corresponding fingerprint version of the other set is deleted. The processor inputs the retained fingerprint versions into the candidate recovery control sequence generation process and merges the same control prefixes generated by different fingerprint versions.

[0069] The fingerprint version is represented by a parent version identifier, an incremental event set, an abnormal starting node, a propagation edge set, event sequence constraints, and a set of evidence sources. When a new protocol event only supplements the existing propagation path, the processor copies the current version and replaces the propagation edge set, while the parent version remains unchanged. When a new protocol event negates the existing starting node, the processor establishes two sub-versions with the original starting node and the new starting node as the root, respectively. Only one copy of the original event data is saved and referenced by the two sub-versions.

[0070] The processor calculates the continuity value for the fingerprint version: ; in, This represents the continuity value of the k-th fingerprint version. Indicates the continuous state of the time anchor point. Indicates the shared path is consistent. This indicates the continuous state of the transaction identifier. It is set to 1 when all three states meet the corresponding conditions, and 0 when they do not. , and It is a positive coefficient of the state.

[0071] For example, if version A maintains continuous time anchors, consistent shared paths, and continuous transaction identifiers, all three states are set to 1. If version B only maintains continuous time anchors and is missing the other two items, version B is deleted. If version B is missing one item but new events can still fill in the missing evidence, the processor retains version B and records the type to be filled in the evidence source set to avoid prematurely fixing the abnormal starting node before the collection delay has been eliminated.

[0072] Candidate recovery control sequences generated by different fingerprint versions may have the same queue command suppression and transaction silencing prefix. The processor determines whether the prefixes are the same by using the control action identifier, target interface and previous state. Only one execution node is retained for the same prefix, and the subsequent link retraining, function-level reset or shared path action of each version is connected at the branch position, so that the prefix execution result can be reused by multiple subsequent branches.

[0073] Furthermore, after executing the sequence to be executed, the processor generates an actual recovery state sequence according to the data field that is the same as the protocol state shadow, and compares it with the expected recovery fingerprint corresponding to the sequence to be executed event by event. When the actual recovery state sequence deviates from the expected recovery fingerprint at a predetermined control step, the processor stops executing subsequent state reconstruction actions, executes reverse control of the registered reversible actions, calls the candidate recovery control sequence in the alternative branches that matches the current actual state prefix, and sets the candidate recovery control sequence containing the same causal fingerprint and the same control prefix as the disabled branch.

[0074] Each record in the actual recovery state sequence includes a control action identifier, action completion time, target interface, link status, number of incomplete transactions, adjacent interface status, device enumeration context version, and verification result. The processor converts the collected raw state into the same field order and value range as the expected recovery fingerprint. It uses allowed interval comparison for the time field, equality comparison for discrete states, and transaction identifier and attribution comparison for transaction sets.

[0075] The deviation between the expected recovered fingerprint and the actual recovered state sequence is calculated using the following formula: ; in, This represents the deviation value after the h-th control step. Indicates differences in the target interface link status. This indicates the difference in the set of incomplete transactions. Indicates the difference in state between adjacent interfaces. Indicates differences in the context of enumeration. , , and The difference is a non-negative weight.

[0076] In this embodiment, the difference is set to 0 when the discrete states are completely consistent and 1 when they are inconsistent. The difference of the transaction set is normalized to between 0 and 1 according to the number of missing transactions, extra transactions and changes in ownership. For example, if the link state and the adjacent interface state are as expected, but there is one transaction written to a new enumeration context, the enumeration context difference is set to 1. The processor regards the deviation as an unstoppable state, stops the subsequent reset or retraining action and enters the reverse control process.

[0077] Reverse control is invoked in the reverse order of reversible actions. The logic path switch executed later is restored first, and the queue command executed earlier is suppressed and then released. Before each reverse control, the processor checks whether its applicable conditions are still valid. If the actual state has changed irreversibly due to the state reconstruction action, the corresponding reverse control is not executed. Instead, the interface isolation is maintained and an alternative branch that can be continued from the current state is selected.

[0078] The candidate branch matching is based on the control prefix, interface state, and enumeration context version of the current actual state sequence. The processor searches for branches with the same executed control actions and whose initial state of the state transition copy can cover the current state. If multiple branches match at the same time, the branches are determined in the order of transaction continuity, adjacent interface linkage range, and number of state reconstruction actions. After a successful match, execution continues from the first control action that has not yet been executed in the branch.

[0079] When a candidate recovery control sequence containing the same causal fingerprint and the same control prefix is ​​set as a disabled branch, the disabled record simultaneously saves the control step that caused the deviation, the actual deviation field, and the corresponding state version. If subsequent new evidence forms a new causal fingerprint version or changes the preceding state corresponding to the same control prefix, the processor can create a new candidate branch, but it may not directly re-enable the original branch with the same preceding state and the same deviation position.

[0080] In a preferred embodiment, when a single downlink interface experiences continuous replay and is accompanied by a link recovery state, the processor reads the events on the interface and its shared path from the protocol state shadow, forms a local link degradation candidate fingerprint, and generates a candidate recovery control sequence that includes queue command suppression, transaction silencing, link retraining, and link rate degradation when the training result does not meet expectations. The state transition copy retains all incomplete transactions of the target interface and verifies that adjacent interfaces do not change state. In actual execution, each control action is compared with the corresponding expected recovery fingerprint after completion.

[0081] In this embodiment, if the transaction quiz boundary is not established, the processor only executes queue command suppression and continues to receive existing completed transactions without performing link retraining. When all transactions to be converged are completed or isolated to a unique device context, the processor generates a transaction quiz confirmation and performs link retraining. If the actual link recovers to the expected state and the states of adjacent interfaces remain unchanged, the queue command suppression is lifted. If the link training result deviates from the expectation but the device enumeration context remains unchanged, the link rate degradation control is called from the alternative branch, and the subsequent expected recovery fingerprint is regenerated.

[0082] In another preferred embodiment, when a command completion exception on the same hard drive occurs before a link error and the exception does not propagate along the shared path, the processor forms a device response instability fingerprint. The candidate recovery control sequence includes queue command suppression, transaction silencing, functional-level reset, and post-enumeration verification. The state transition copy establishes a new device context version for the functional-level reset and writes the incomplete transactions in the old context into the set to be verified. The new context is only submitted when the device identifier, namespace set, and transaction attribution are all compatible with expectations.

[0083] In this embodiment, if the completion information returned after a functional-level reset carries the transaction identifier corresponding to the old context, the processor matches it with the set to be verified and writes it back to the old transaction record, and does not allow the completion information to change the command state in the new context; if the returned information cannot uniquely match the old transaction or the new transaction, an enumeration context conflict flag is generated, the interface is kept isolated and the submission of new device contexts is stopped, so that the device ownership in the recovery execution area, protocol state shadow and transaction page remains consistent.

[0084] In another preferred embodiment, when multiple downlink interfaces experience completion delays within the same time interval, and the common uplink node experiences an abnormal state before each downlink interface, the processor forms a shared path blocking fingerprint. The candidate recovery control sequence first executes queue commands to suppress the affected interfaces and establishes transaction quiescent boundaries. Then, based on the state transition copy, it determines whether the shared path action will affect interfaces outside the causal fingerprint range. Only when all interfaces that depend on the shared node are included in the affected range and the transaction continuity constraint is met, is the state reconstruction action corresponding to the shared path allowed to be executed.

[0085] In this embodiment, the processor maintains boundary confirmation values ​​for the set of incomplete transactions for each downlink interface. It does not substitute the silencing of one interface for the silencing of other interfaces. When some interfaces still have write transactions whose ownership cannot be confirmed, the shared path action remains unexecutable. The processor can switch logical data paths or maintain command suppression for silenced interfaces and continue to receive completion returns for non-silenced interfaces. After all necessary interfaces meet the transaction silencing conditions, the comprehensive constraint value of the candidate sequence is recalculated.

Claims

1. An AI server hard disk backplane data interface management system, characterized in that, Includes a processor and a memory, wherein the memory stores instructions, and the instructions instruct the processor to: collect link status, protocol errors, transaction timing and topology association data of each PCIe interface, and establish a protocol status shadow corresponding to each PCIe interface; Based on the protocol state shadow, generate abnormal event fragments and causal fingerprints representing the source and propagation path of the abnormality; Based on the causal fingerprint, candidate recovery control sequences are generated. The impact of each candidate recovery control sequence on incomplete transactions, adjacent interfaces, and device enumeration states is simulated in the protocol state shadow to form the corresponding expected recovery fingerprint. The candidate recovery control sequence that satisfies the transaction continuity constraint and has the smallest impact range is selected for execution, and the subsequent control sequence is determined based on the matching result between the actual state after execution and the expected recovery fingerprint.

2. The AI ​​server hard disk backplane data interface management system according to claim 1, characterized in that, The protocol state shadow stores the link rate, link width, training status, correctable error, uncorrectable error, replay event, completion timeout, NVMe command submission time, command completion time, and number of incomplete commands according to the interface identifier. It also merges data with the same transaction identifier, the same path identifier, or previous and subsequent state transition relationships in adjacent sampling periods into protocol events. The processor uses link state change events as time anchors to sort and associate protocol events before and after the time anchors, forming anomaly event fragments that include event subject, event type, occurrence order, duration range, and associated path.

3. The AI ​​server hard disk backplane data interface management system according to claim 1, characterized in that, The processor establishes a topology propagation graph based on the logical connection relationship between the hard disk, downlink PCIe interface, PCIe switching path, uplink PCIe interface and root port, projects the abnormal event fragments onto the topology propagation graph, and marks the propagation direction between adjacent nodes according to the order of event occurrence. Based on the event synchronization degree of multiple downlink interfaces under the same uplink path, the first occurrence of single interface errors, and the spread range of transaction completion delays, a causal fingerprint is generated, which includes the abnormal starting node, the set of propagation edges, event sequence constraints, and evidence consistency.

4. The AI ​​server hard disk backplane data interface management system according to claim 1, characterized in that, The candidate recovery control sequence is formed by arranging at least two of the following according to control dependencies: queue command suppression, transaction silencing, function-level reset, link retraining, link rate degradation, link width degradation, and logical data path switching. The processor establishes a state transition copy for each candidate recovery control sequence, updates the incomplete transaction set, interface link status, adjacent interface shared path occupancy status and device enumeration context item by item in the state transition copy, and marks candidate recovery control sequences that cross transaction boundaries, cause irrelevant interface status changes or cause enumeration context conflicts as non-executable sequences.

5. The AI ​​server hard disk backplane data interface management system according to claim 2, characterized in that, The processor sets up a pre-observation area, a core anomaly area, and a post-response area for each anomaly event segment. The earliest protocol error or completion delay in the pre-observation area is taken as a candidate triggering event. The PCIe interface in the core anomaly area where a state change occurs is taken as a candidate starting interface. The associated events of other interfaces in the post-response area are layered according to topological distance. When the same event satisfies both transaction identifier association and path propagation association, its dual association marker is retained. When the two types of associations point to different starting interfaces, the corresponding abnormal event fragment is split into multiple sub-fragments to be determined and written into the protocol state shadow respectively.

6. The AI ​​server hard disk backplane data interface management system according to claim 3, characterized in that, The processor constructs an interface event association matrix based on the topology propagation graph. The rows and columns of the interface event association matrix correspond to the PCIe interface, and the matrix elements record the event time difference, shared path level, and error type consistency. The set of events that occur only on a single downlink interface and do not spread to the shared node is encoded as a local link degradation fingerprint. The set of events that occur before the link error is encoded as a device response instability fingerprint. The set of events that occur when multiple downlink interfaces form a synchronization delay on the same uplink path and the error originates at the shared node is encoded as a shared path blocking fingerprint.

7. The AI ​​server hard disk backplane data interface management system according to claim 4, characterized in that, When generating the candidate recovery control sequence, the processor first determines the transaction quiescent boundary of each PCIe interface based on the set of incomplete transactions, and marks control actions that only change the command issuance state or logical path mapping as reversible actions, and marks control actions that cause link state renegotiation or device context reconstruction as state reconstruction actions. The candidate recovery control sequence is arranged according to the constraints that the reversible action precedes the state reconstruction action, the local interface action precedes the shared path action, and the transaction silent confirmation precedes the reset or retraining action. Each state reconstruction action is associated with a corresponding set of preceding transactions and a set of subsequent enumeration verification items.

8. The AI ​​server hard disk backplane data interface management system according to claim 4, characterized in that, The processor retains the commit interface, target namespace, completion status, and data buffer occupancy flag for each incomplete transaction in the state transition copy, and sequentially calculates the transaction cross-boundary flag, adjacent interface linkage flag, enumeration context conflict flag, and control reversibility flag for each candidate recovery control sequence. The sequence to be executed is determined by first excluding sequences that have the transaction cross-boundary flag or the enumeration context conflict flag, then comparing the linkage range of adjacent interfaces, and finally comparing the number of irreversible control actions. The candidate recovery control sequences that were not selected and their corresponding state transition copies are kept as alternative branches.

9. The AI ​​server hard disk backplane data interface management system according to claim 1, characterized in that, The processor maintains the fingerprint version and evidence source set for each causal fingerprint. When a new protocol event is consistent with the current abnormal starting node but the propagation edges are inconsistent, only the propagation edge set is updated. When the newly added protocol event causes a change in the candidate starting node, the order of the event reconstruction is constrained by the original candidate starting node and the new candidate starting node, respectively. If one set of event sequence constraints maintains continuous time anchors, consistent shared paths, and continuous transaction identifiers, while another set of event sequence constraints lacks at least two of these, then the corresponding fingerprint version of the other set is deleted. The processor inputs the retained fingerprint versions into the generation process of the candidate recovery control sequence, and merges the same control prefix generated by different fingerprint versions.

10. The AI ​​server hard disk backplane data interface management system according to claim 8, characterized in that, After executing the sequence to be executed, the processor generates an actual recovery state sequence according to the data fields that are the same as the protocol state shadow, and compares it with the expected recovery fingerprint corresponding to the sequence to be executed event by event. When the actual recovery state sequence deviates from the expected recovery fingerprint at the predetermined control step, the subsequent state reconstruction action is stopped, the reverse control of the registered reversible action is executed, the candidate recovery control sequence that matches the current actual state prefix in the alternative branches is called, and the candidate recovery control sequence containing the same causal fingerprint and the same control prefix is ​​set as the disabled branch.

Images