new proposal for locality loadbalance

docs/proposal/kmesh_support_localityLB.md

@@ -37,64 +37,47 @@ What is locality strict mode? In locality strict mode, the LB (load balancing) a
 
 ### Design Details
 
-Maintain a prio map in both user space and kernel space, where the key is a combination of service ID and rank, and the value consists of a count and a UID list. In user space, compute the rank for each backend corresponding to a service and store it in a BPF map named KmeshPrio. In kernel space, iterate through the ranks using the service ID to query KmeshPrio and perform locality load balancing on the UID list.
+In both user space and kernel space, add a layer of priority rank between the existing services and endpoints. The new serviceKey will consist of the service ID, rank, and index, with the value being a list of UIDs.
+
+In user space, add a set of logic to maintain the priority relationship between services and backends. When adding or updating a workload, calculate the priority of the workload relative to the current kmesh process locality, store the backend in the corresponding map, and update the relationship in the kernel BPF map. Additionally, keep track of the number of endpoints for different priorities associated with each service.
+
+In kernel space, traverse the ranks based on the service ID, and perform locality load balancing by querying the backends according to the number of endpoints.
 
 ![locality_lb_pic](pics/locality_lb.svg)
 
 #### Control plane（user-space）
-1. The user-space kmesh process maintains locality, clusterId, and network information, which is updated when xDS receives information about the workload currently on this node. [handleWorkload]
-2. The user-space kmesh also maintains routingPreference, which is updated when xDS receives information about the service currently on this node. [storeServiceData]
-3. The user-space kmesh maintains a map, `backendsByService`, storing all services present within the current kmesh process along with their respective backends, which will participate in the locality rank calculation.
-4. Once both sets of information are maintained, the locality rank is calculated and the locality load balancing policy is updated in the BPF map `KmeshPrio` via the method `[updateLocalityLBPrio]`:
-  - `updateLocalityLBPrio`: Iterates through and compares the locality information of the kmesh process with every backend associated with the service, matching according to the routing rules defined in the service's load balance scope. For example, if the service is configured with a region scope, the region of the kmesh process is compared with the region of the target backend, incrementing the rank value if they match. After comparing each backend, store the UID of that backend in the `UidList` of `PrioValue` corresponding to its rank value and service ID as the key. Each rank corresponds to a set of backends for that priority level.
-  - In `handleWorkload`, when handling workloads with service information, first execute the existing logic `handleDataWithService` to handle the service, then use `storeServiceBackend` to store the backend in `backendsByService`, and finally call `updateLocalityLBPrio`.
-  - In `handleService`, call `updateLocalityLBPrio`.
-  - In `removeWorkloadFromLocalityLBPrio`, perform a full update (since removing a workload requires updating the entire map, as the current `prio map` stores the backend UID list in an array form, making selective deletion of map elements impractical).
+1. User space needs to maintain locality, clusterId, and network information, which is updated when the xDS receives information about the workloads currently on this node.
+2. User space also needs to maintain routingPreference, which is updated when the xDS receives information about the services currently on this node.
+3. User space maintains a map called workloadsByServiceAndPrio that stores all backends corresponding to different priorities for all services present on the current kmesh process. When these backends are first added, they are placed into an array with a priority of 7. The program then periodically checks the backends in this priority level and performs locality rank calculations. Based on these calculations, pods are assigned to endpoints corresponding to their ranks. When a backend is deleted, the deleteEndpointRecords function is called to remove the corresponding endpoint.
+4. After both pieces of information are maintained, store the backends in the corresponding endpoints based on the workloadsByServiceAndPrio in user space:
+    updateLocalityLBPrio: Iterate through and compare the locality information of the kmesh process with every backend associated with the service, matching according to the routing rules defined in the service's load balance scope. For example, if the service is configured with a region scope, the region of the kmesh process is compared with the region of the target backend, and the rank value is incremented by one if they match. After comparing each backend, store the UID of that backend into the endpoint corresponding to its rank value.
 
 #### Data plane（kernel-space）
-When a request is made for a service, the BPF program queries the backend UID list from the prio map using the request's service ID and iterates through the ranks. 
+When a request is made to a service, the BPF program traverses the priority ranks based on the requested service ID and count, querying the backend IDs from the endpoint map.
 - In strict mode, it selects backends that match all scopes.
 - In failover mode, it traverses the backend UID array starting from the highest rank and decrementing step by step. 
-- From the queried backend UIDs, a backend is randomly selected to serve the request.
+- From the queried prio, a backend is randomly selected to serve the request.
 
 
 #### data struct
-1. prio map
-
-workload.h
+1. workload.h
 ```
 typedef struct {
-    __u32 service_id; // service id
-    __u32 rank; // rank
-} prio_key;
-typedef struct {
-    __u32 count; // count of current prio
-    __u32 uid_list[MAP_SIZE_OF_PRIO]; // workload_uid to backend
-} prio_value;
-
-struct {
-    __uint(type, BPF_MAP_TYPE_HASH);
-    __uint(key_size, sizeof(prio_key));
-    __uint(value_size, sizeof(prio_value));
-    __uint(max_entries, MAP_SIZE_OF_SERVICE*MAX_PRIO);
-    __uint(map_flags, BPF_F_NO_PREALLOC);
-} map_of_prio SEC(".maps");
-
+    __u32 service_id;    // service id
+    __u32 prio;          // prio means rank, 6 means match all, and 0 means match nothing
+    __u32 backend_index; // if endpoint_count = 3, then backend_index = 0/1/2
+} endpoint_key;
 ```
 
 2. service
 
 workload.h
 ```
 typedef struct {
-    __u32 endpoint_count;               // endpoint count of current service
+    __u32 prio_endpoint_count[MAX_PRIO_COUNT+1];// endpoint count of current service with prio, prio from 6->0, 7是未被分配的workload暂存区
     __u32 lb_policy;                    // load balancing algorithm, currently supports random algorithm, locality loadbalance Failover/strict mode
-    __u32 lb_strict_index;
-    __u32 service_port[MAX_PORT_COUNT]; // service_port[i] and target_port[i] are a pair, i starts from 0 and max value
-                                        // is MAX_PORT_COUNT-1
-    __u32 target_port[MAX_PORT_COUNT];
-    struct ip_addr wp_addr;
-    __u32 waypoint_port;
+    __u32 lb_strict_index;              // for failover strict mode
+    ...
 } service_value;
 ```
 
@@ -110,45 +93,38 @@ typedef enum {
 } lb_policy_t;
 ```
 
-4. prio
-
-priority.go
+4. locality_cache.go
 ```
-const (
-	MaxPrio       = 7
-	MaxSizeOfPrio = 1000
-)
-
-type PrioKey struct {
-	ServiceId uint32
-	Rank      uint32
+type localityInfo struct {
+    region    string // init from workload.GetLocality().GetRegion()
+    zone      string // init from workload.GetLocality().GetZone()
+    subZone   string // init from workload.GetLocality().GetSubZone()
+    nodeName  string // init from os.Getenv("NODE_NAME"), workload.GetNode()
+    clusterId string // init from workload.GetClusterId()
+    network   string // workload.GetNetwork()
+    mask      uint32 // mask
 }
-type PrioValue struct {
-	Count   uint32                // count of current prio
-	UidList [MaxSizeOfPrio]uint32 // workload_uid to backend
+
+type LocalityCache interface {
+    SetLocality(nodeName, clusterId, network string, locality *workloadapi.Locality)
+    SetRoutingPreference(s []workloadapi.LoadBalancing_Scope)
+    CanLocalityLB() bool
+    GetWorkloadsByServiceAndPrio(string, uint32) map[*workloadapi.Workload]struct{}
+    AddOrUpdateWorkload(workload *workloadapi.Workload, proc *Processor)
+    DeleteWorkload(workload *workloadapi.Workload, proc *Processor)
+    updateLocalityLBPrio(map[*workloadapi.Workload]struct{}, uint32, *Processor) bool
 }
-...
-```
 
-5. add new locality info to kmesh process
-```
-type Processor struct {
-	ack *service_discovery_v3.DeltaDiscoveryRequest
-	req *service_discovery_v3.DeltaDiscoveryRequest
-
-	hashName *HashName
-	// workloads indexer, svc key -> workload id
-	endpointsByService      map[string]map[string]struct{}
-	backendsByService       map[string]map[string]*workloadapi.Workload
-	bpf                     *bpf.Cache
-	nodeName                string                            // init from env("NODE_NAME")
-	clusterId               string                            // init from workload
-	network                 string                            // init from workload
-	routingPreference       []workloadapi.LoadBalancing_Scope // init from service
-	locality                workloadapi.Locality              // the locality of node, init from workload
-	WorkloadCache           cache.WorkloadCache
-	ServiceCache            cache.ServiceCache
-	is_locality_ok          bool
-	is_routingPreference_ok bool
+type localityCache struct {
+    mutex                     sync.RWMutex
+    LbPolicy                  uint32
+    localityInfo              localityInfo
+    LbStrictIndex             uint32 // for failover strict mode
+    isLocalityInfoSet         bool
+    RoutingPreference         []workloadapi.LoadBalancing_Scope
+    isRoutingPreferenceSet    bool
+    workloadsByServiceAndPrio map[string]map[uint32]map[*workloadapi.Workload]struct{} // prio 0->6 , prio 7(bpf.MaxPrioNum) represents workloads need to be LB
+    workloadBelongsToService  map[*workloadapi.Workload]map[uint32]string
+    workloadBelongsToPrio     map[*workloadapi.Workload]map[uint32]uint32 // index response to service
 }
 ```