Merge pull request #909 from nlgwcy/release-0.5

cherry-pick #610 to Release 0.5

Author: kmesh-bot

bpf/deserialization_to_bpf_map/deserialization_to_bpf_map.c

@@ -4,8 +4,21 @@
 #ifndef __DESERIALIZATION_TO_BPF_MAP_H__
 #define __DESERIALIZATION_TO_BPF_MAP_H__
 
+#include <stdbool.h>
+
 /* equal MAP_SIZE_OF_OUTTER_MAP */
-#define MAX_OUTTER_MAP_ENTRIES (8192)
+#define MAX_OUTTER_MAP_ENTRIES        (1 << 20)
+#define OUTTER_MAP_USAGE_HIGH_PERCENT (0.7)
+#define OUTTER_MAP_USAGE_LOW_PERCENT  (0.3)
+#define TASK_SIZE                     (512)
+
+// 32,768
+#define OUTTER_MAP_SCALEUP_STEP (1 << 15)
+// 8,192
+#define OUTTER_MAP_SCALEIN_STEP (1 << 13)
+
+#define ELASTIC_SLOTS_NUM                                                                                              \
+    ((OUTTER_MAP_SCALEUP_STEP > OUTTER_MAP_SCALEIN_STEP) ? OUTTER_MAP_SCALEUP_STEP : OUTTER_MAP_SCALEIN_STEP)
 
 struct element_list_node {
     void *elem;
@@ -20,6 +33,7 @@ void deserial_free_elem_list(struct element_list_node *head);
 int deserial_delete_elem(void *key, const void *msg_desciptor);
 
 int deserial_init();
-void deserial_uninit();
+void deserial_uninit(bool persist);
+int inner_map_mng_persist();
 
 #endif /* __DESERIALIZATION_TO_BPF_MAP_H__ */

bpf/deserialization_to_bpf_map/deserialization_to_bpf_map.h

@@ -17,7 +17,7 @@
 /* Ip(0.0.0.2 | ::2) used for control command, e.g. KmeshControl */
 #define CONTROL_CMD_IP 2
 
-#define MAP_SIZE_OF_OUTTER_MAP 8192
+#define MAP_SIZE_OF_OUTTER_MAP (1 << 20)
 
 #define BPF_DATA_MAX_LEN                                                                                               \
     192 /* this value should be                                                                                        \

bpf/include/bpf_common.h

@@ -29,16 +29,15 @@
 #define MAP_SIZE_OF_PER_CLUSTER      32
 #define MAP_SIZE_OF_PER_ENDPOINT     64
 
-#define MAP_SIZE_OF_MAX        8192
-#define MAP_SIZE_OF_OUTTER_MAP 8192
+#define MAP_SIZE_OF_MAX 8192
 
-#define MAP_SIZE_OF_LISTENER     BPF_MIN(MAP_SIZE_OF_MAX, MAP_SIZE_OF_PER_LISTENER)
-#define MAP_SIZE_OF_FILTER_CHAIN BPF_MIN(MAP_SIZE_OF_MAX, MAP_SIZE_OF_PER_FILTER_CHAIN *MAP_SIZE_OF_LISTENER)
-#define MAP_SIZE_OF_FILTER       BPF_MIN(MAP_SIZE_OF_MAX, MAP_SIZE_OF_PER_FILTER *MAP_SIZE_OF_FILTER_CHAIN)
-#define MAP_SIZE_OF_VIRTUAL_HOST BPF_MIN(MAP_SIZE_OF_MAX, MAP_SIZE_OF_PER_VIRTUAL_HOST *MAP_SIZE_OF_FILTER)
-#define MAP_SIZE_OF_ROUTE        BPF_MIN(MAP_SIZE_OF_MAX, MAP_SIZE_OF_PER_ROUTE *MAP_SIZE_OF_VIRTUAL_HOST)
-#define MAP_SIZE_OF_CLUSTER      BPF_MIN(MAP_SIZE_OF_MAX, MAP_SIZE_OF_PER_CLUSTER *MAP_SIZE_OF_ROUTE)
-#define MAP_SIZE_OF_ENDPOINT     BPF_MIN(MAP_SIZE_OF_MAX, MAP_SIZE_OF_PER_ENDPOINT *MAP_SIZE_OF_CLUSTER)
+#define MAP_SIZE_OF_LISTENER     (1 << 10)
+#define MAP_SIZE_OF_FILTER_CHAIN (MAP_SIZE_OF_PER_FILTER_CHAIN * MAP_SIZE_OF_LISTENER)
+#define MAP_SIZE_OF_FILTER       (MAP_SIZE_OF_PER_FILTER * MAP_SIZE_OF_FILTER_CHAIN)
+#define MAP_SIZE_OF_VIRTUAL_HOST (MAP_SIZE_OF_PER_VIRTUAL_HOST * MAP_SIZE_OF_FILTER)
+#define MAP_SIZE_OF_ROUTE        (1 << 14)
+#define MAP_SIZE_OF_CLUSTER      (1 << 14)
+#define MAP_SIZE_OF_ENDPOINT     (1 << 17)
 
 // rename map to avoid truncation when name length exceeds BPF_OBJ_NAME_LEN = 16
 #define map_of_listener       kmesh_listener

bpf/kmesh/ads/include/config.h

@@ -0,0 +1,69 @@
+---
+title: map-in-map management enhancement
+authors:
+- "@nlgwcy"
+reviewers:
+- "@hzxuzhonghu"
+- "@supercharge-xsy"
+- "@bitcoffeeiux"
+approvers:
+- "@robot"
+- TBD
+
+creation-date: 2024-07-20
+
+---
+
+## map-in-map management enhancement
+
+### Summary
+
+In ads mode, elastic scaling based on map-in-map records is supported to meet the traffic management requirements of large-scale clusters.
+
+### Motivation
+
+As mentioned in [optimizing_bpf_map_update_in_xDS_mode](https://github.com/kmesh-net/kmesh/blob/main/docs/proposal/optimizing_bpf_map_update_in_xDS_mode-en.md), to solve the problem of slow update of map-in-map records, Kmesh creates all records at a time during startup by exchanging space for time. This problem does not occur in small-scale cluster scenarios, however, when a large-scale cluster (for example, 5000 services and 100,000 pods) is supported, the size defined in the map-in-map table is very large, and the map of the `BPF_MAP_TYPE_ARRAY_OF_MAPS` type does not support `BPF_F_NO_PREALLOC`, which causes a great waste of memory. Elastic scaling of map-in-map records must be supported to meet the traffic management requirements of large-scale clusters.
+
+#### Goals
+
+- Supports traffic management in large-scale clusters.
+- Consider the configuration restoration scenario.
+
+### Proposal
+
+Kmesh manages the usage of map-in-map in user mode. To support elastic scaling, the management structure is extended as follows:
+
+```c
+struct inner_map_mng {
+    int inner_fd;
+    int outter_fd;
+    struct bpf_map_info inner_info;
+    struct bpf_map_info outter_info;
+    struct inner_map_stat inner_maps[MAX_OUTTER_MAP_ENTRIES];
+    int elastic_slots[OUTTER_MAP_ELASTIC_SIZE];
+    int used_cnt;           // real used count
+    int alloced_cnt;        // real alloced count
+    int max_alloced_idx;    // max alloced index, there may be holes.
+    int init;
+    sem_t fin_tasks;
+    int elastic_task_exit;  // elastic scaling thread exit flag
+};
+
+struct inner_map_stat {
+    int map_fd;
+    unsigned int used : 1;
+    unsigned int alloced : 1;
+    unsigned int resv : 30;
+};
+```
+
+Map-in-map scaling process:
+
+![map-in-map-elastic-process](pics/map-in-map-elastic-process.svg)
+
+The following is an example of map-in-map scale-in and scale-out:
+
+![map-in-map-elastic](pics/map-in-map-elastic.svg)
+
+
+

docs/proposal/map-in-map_management_enhancement-en.md

@@ -173,21 +173,23 @@ func StopMda() error {
 
 func (l *BpfLoader) Stop() {
 	var err error
-	if GetExitType() == Restart {
+	if GetExitType() == Restart && l.config.WdsEnabled() {
+		C.deserial_uninit(true)
 		log.Infof("kmesh restart, not clean bpf map and prog")
 		return
 	}
 
 	closeMap(l.versionMap)
 
 	if l.config.AdsEnabled() {
-		C.deserial_uninit()
+		C.deserial_uninit(false)
 		if err = l.obj.Detach(); err != nil {
 			CleanupBpfMap()
 			log.Errorf("failed detach when stop kmesh, err:%s", err)
 			return
 		}
 	} else if l.config.WdsEnabled() {
+		C.deserial_uninit(false)
 		if err = l.workloadObj.Detach(); err != nil {
 			CleanupBpfMap()
 			log.Errorf("failed detach when stop kmesh, err:%s", err)
@@ -311,6 +313,11 @@ func recoverVersionMap(pinPath string) *ebpf.Map {
 
 func closeMap(m *ebpf.Map) {
 	var err error
+
+	if m == nil {
+		return
+	}
+
 	err = m.Unpin()
 	if err != nil {
 		log.Errorf("Failed to unpin kmesh_version: %v", err)