seaweedfs

Chris Lu d745e6e41d Fix masterclient vidmap race condition (#7412 ) * fallback to check master * clean up * parsing * refactor * handle parse error * return error * avoid dup lookup * use batch key * dedup lookup logic * address comments * errors.Join(lookupErrors...) * add a comment * Fix: Critical data race in MasterClient vidMap Fixes a critical data race where resetVidMap() was writing to the vidMap pointer while other methods were reading it concurrently without synchronization. Changes: - Removed embedded vidMap from MasterClient - Added vidMapLock (sync.RWMutex) to protect vidMap pointer access - Created safe accessor methods (GetLocations, GetDataCenter, etc.) - Updated all direct vidMap accesses to use thread-safe methods - Updated resetVidMap() to acquire write lock during pointer swap The vidMap already has internal locking for its operations, but this fix protects the vidMap pointer itself from concurrent read/write races. Verified with: go test -race ./weed/wdclient/... Impact: - Prevents potential panics from concurrent pointer access - No performance impact - uses RWMutex for read-heavy workloads - Maintains backward compatibility through wrapper methods fmt * Fix: Critical data race in MasterClient vidMap Fixes a critical data race where resetVidMap() was writing to the vidMap pointer while other methods were reading it concurrently without synchronization. Changes: - Removed embedded vidMap from MasterClient struct - Added vidMapLock (sync.RWMutex) to protect vidMap pointer access - Created minimal public accessor methods for external packages: GetLocations, GetLocationsClone, GetVidLocations * LookupFileId, LookupVolumeServerUrl * GetDataCenter - Internal code directly locks and accesses vidMap (no extra indirection) - Updated resetVidMap() to acquire write lock during pointer swap - Updated shell/commands.go to use GetDataCenter() method Design Philosophy: - vidMap already has internal locking for its map operations - This fix specifically protects the vidMap pointer from concurrent access - Public methods for external callers, direct locking for internal use - Minimizes wrapper overhead while maintaining thread safety Verified with: go test -race ./weed/wdclient/... (passes) Impact: - Prevents potential panics/crashes from data races - Minimal performance impact (RWMutex for read-heavy workload) - Maintains full backward compatibility * fix more concurrent access * reduce lock scope * Optimize vidMap locking for better concurrency Improved locking strategy based on the understanding that: - vidMapLock protects the vidMap pointer from concurrent swaps - vidMap has internal locks that protect its data structures Changes: 1. Read operations: Grab pointer with RLock, release immediately, then operate - Reduces lock hold time - Allows resetVidMap to proceed sooner - Methods: GetLocations, GetLocationsClone, GetVidLocations, LookupVolumeServerUrl, GetDataCenter 2. Write operations: Changed from Lock() to RLock() - RLock prevents pointer swap during operation - Allows concurrent readers and other writers (serialized by vidMap's lock) - Methods: addLocation, deleteLocation, addEcLocation, deleteEcLocation Benefits: - Significantly reduced lock contention - Better concurrent performance under load - Still prevents all race conditions Verified with: go test -race ./weed/wdclient/... (passes) * Further reduce lock contention in LookupVolumeIdsWithFallback Optimized two loops that were holding RLock for extended periods: Before: - Held RLock during entire loop iteration - Included string parsing and cache lookups - Could block resetVidMap for significant time with large batches After: - Grab vidMap pointer with brief RLock - Release lock immediately - Perform all loop operations on local pointer Impact: - First loop: Cache check on initial volumeIds - Second loop: Double-check after singleflight wait Benefits: - Minimal lock hold time (just pointer copy) - resetVidMap no longer blocked by long loops - Better concurrent performance with large volume ID lists - Still thread-safe (vidMap methods have internal locks) Verified with: go test -race ./weed/wdclient/... (passes) * Add clarifying comments to vidMap helper functions Added inline documentation to each helper function (addLocation, deleteLocation, addEcLocation, deleteEcLocation) explaining the two-level locking strategy: - RLock on vidMapLock prevents resetVidMap from swapping the pointer - vidMap has internal locks that protect the actual map mutations - This design provides optimal concurrency The comments make it clear why RLock (not Lock) is correct and intentional, preventing future confusion about the locking strategy. * Improve encapsulation: Add shallowClone() method to vidMap Added a shallowClone() method to vidMap to improve encapsulation and prevent MasterClient from directly accessing vidMap's internal fields. Changes: 1. Added vidMap.shallowClone() in vid_map.go - Encapsulates the shallow copy logic within vidMap - Makes vidMap responsible for its own state representation - Documented that caller is responsible for thread safety 2. Simplified resetVidMap() in masterclient.go - Uses tail := mc.vidMap.shallowClone() instead of manual field access - Cleaner, more maintainable code - Better adherence to encapsulation principles Benefits: - Improved code organization and maintainability - vidMap internals are now properly encapsulated - Easier to modify vidMap structure in the future - More self-documenting code Verified with: go test -race ./weed/wdclient/... (passes) * Optimize locking: Reduce lock acquisitions and use helper methods Two optimizations to further reduce lock contention and improve code consistency: 1. LookupFileIdWithFallback: Eliminated redundant lock acquisition - Before: Two separate locks to get vidMap and dataCenter - After: Single lock gets both values together - Benefit: 50% reduction in lock/unlock overhead for this hot path 2. KeepConnected: Use GetDataCenter() helper for consistency - Before: Manual lock/unlock to access DataCenter field - After: Use existing GetDataCenter() helper method - Benefit: Better encapsulation and code consistency Impact: - Reduced lock contention in high-traffic lookup path - More consistent use of accessor methods throughout codebase - Cleaner, more maintainable code Verified with: go test -race ./weed/wdclient/... (passes) * Refactor: Extract common locking patterns into helper methods Eliminated code duplication by introducing two helper methods that encapsulate the common locking patterns used throughout MasterClient: 1. getStableVidMap() - For read operations - Acquires lock, gets pointer, releases immediately - Returns stable snapshot for thread-safe reads - Used by: GetLocations, GetLocationsClone, GetVidLocations, LookupFileId, LookupVolumeServerUrl, GetDataCenter 2. withCurrentVidMap(f func(vm vidMap)) - For write operations - Holds RLock during callback execution - Prevents pointer swap while allowing concurrent operations - Used by: addLocation, deleteLocation, addEcLocation, deleteEcLocation Benefits: - Reduced code duplication (eliminated 48 lines of repetitive locking code) - Centralized locking logic makes it easier to understand and maintain - Self-documenting pattern through named helper methods - Easier to modify locking strategy in the future (single point of change) - Improved readability - accessor methods are now one-liners Code size reduction: ~40% fewer lines for accessor/helper methods Verified with: go test -race ./weed/wdclient/... (passes) consistent * Fix cache pointer race condition with atomic.Pointer Use atomic.Pointer for vidMap cache field to prevent data races during cache trimming in resetVidMap. This addresses the race condition where concurrent GetLocations calls could read the cache pointer while resetVidMap is modifying it during cache chain trimming. Changes: - Changed cache field from vidMap to atomic.Pointer[vidMap] - Updated all cache access to use Load() and Store() atomic operations - Updated shallowClone, GetLocations, deleteLocation, deleteEcLocation - Updated resetVidMap to use atomic operations for cache trimming Merge: Resolve conflict in deleteEcLocation - keep atomic.Pointer and fix bug Resolved merge conflict by combining: 1. Atomic pointer access pattern (from HEAD): cache.Load() 2. Correct method call (from fix): deleteEcLocation (not deleteLocation) Resolution: - Before (HEAD): cachedMap.deleteLocation() - WRONG, reintroduced bug - Before (fix): vc.cache.deleteEcLocation() - RIGHT method, old pattern - After (merged): cachedMap.deleteEcLocation() - RIGHT method, new pattern This preserves both improvements: ✓ Thread-safe atomic.Pointer access pattern ✓ Correct recursive call to deleteEcLocation Verified with: go test -race ./weed/wdclient/... (passes) * Update vid_map.go * remove shallow clone * simplify	3 months ago
..
request_id.go	Context-based logging with request ID (#6899)	7 months ago

Fix masterclient vidmap race condition (#7412 )

* fallback to check master

* clean up

* parsing

* refactor

* handle parse error

* return error

* avoid dup lookup

* use batch key

* dedup lookup logic

* address comments

* errors.Join(lookupErrors...)

* add a comment

* Fix: Critical data race in MasterClient vidMap

Fixes a critical data race where resetVidMap() was writing to the vidMap
pointer while other methods were reading it concurrently without synchronization.

Changes:
- Removed embedded *vidMap from MasterClient
- Added vidMapLock (sync.RWMutex) to protect vidMap pointer access
- Created safe accessor methods (GetLocations, GetDataCenter, etc.)
- Updated all direct vidMap accesses to use thread-safe methods
- Updated resetVidMap() to acquire write lock during pointer swap

The vidMap already has internal locking for its operations, but this fix
protects the vidMap pointer itself from concurrent read/write races.

Verified with: go test -race ./weed/wdclient/...

Impact:
- Prevents potential panics from concurrent pointer access
- No performance impact - uses RWMutex for read-heavy workloads
- Maintains backward compatibility through wrapper methods

* fmt

* Fix: Critical data race in MasterClient vidMap

Fixes a critical data race where resetVidMap() was writing to the vidMap
pointer while other methods were reading it concurrently without synchronization.

Changes:
- Removed embedded *vidMap from MasterClient struct
- Added vidMapLock (sync.RWMutex) to protect vidMap pointer access
- Created minimal public accessor methods for external packages:
  * GetLocations, GetLocationsClone, GetVidLocations
  * LookupFileId, LookupVolumeServerUrl
  * GetDataCenter
- Internal code directly locks and accesses vidMap (no extra indirection)
- Updated resetVidMap() to acquire write lock during pointer swap
- Updated shell/commands.go to use GetDataCenter() method

Design Philosophy:
- vidMap already has internal locking for its map operations
- This fix specifically protects the vidMap *pointer* from concurrent access
- Public methods for external callers, direct locking for internal use
- Minimizes wrapper overhead while maintaining thread safety

Verified with: go test -race ./weed/wdclient/... (passes)

Impact:
- Prevents potential panics/crashes from data races
- Minimal performance impact (RWMutex for read-heavy workload)
- Maintains full backward compatibility

* fix more concurrent access

* reduce lock scope

* Optimize vidMap locking for better concurrency

Improved locking strategy based on the understanding that:
- vidMapLock protects the vidMap pointer from concurrent swaps
- vidMap has internal locks that protect its data structures

Changes:
1. Read operations: Grab pointer with RLock, release immediately, then operate
   - Reduces lock hold time
   - Allows resetVidMap to proceed sooner
   - Methods: GetLocations, GetLocationsClone, GetVidLocations,
     LookupVolumeServerUrl, GetDataCenter

2. Write operations: Changed from Lock() to RLock()
   - RLock prevents pointer swap during operation
   - Allows concurrent readers and other writers (serialized by vidMap's lock)
   - Methods: addLocation, deleteLocation, addEcLocation, deleteEcLocation

Benefits:
- Significantly reduced lock contention
- Better concurrent performance under load
- Still prevents all race conditions

Verified with: go test -race ./weed/wdclient/... (passes)

* Further reduce lock contention in LookupVolumeIdsWithFallback

Optimized two loops that were holding RLock for extended periods:

Before:
- Held RLock during entire loop iteration
- Included string parsing and cache lookups
- Could block resetVidMap for significant time with large batches

After:
- Grab vidMap pointer with brief RLock
- Release lock immediately
- Perform all loop operations on local pointer

Impact:
- First loop: Cache check on initial volumeIds
- Second loop: Double-check after singleflight wait

Benefits:
- Minimal lock hold time (just pointer copy)
- resetVidMap no longer blocked by long loops
- Better concurrent performance with large volume ID lists
- Still thread-safe (vidMap methods have internal locks)

Verified with: go test -race ./weed/wdclient/... (passes)

* Add clarifying comments to vidMap helper functions

Added inline documentation to each helper function (addLocation, deleteLocation,
addEcLocation, deleteEcLocation) explaining the two-level locking strategy:

- RLock on vidMapLock prevents resetVidMap from swapping the pointer
- vidMap has internal locks that protect the actual map mutations
- This design provides optimal concurrency

The comments make it clear why RLock (not Lock) is correct and intentional,
preventing future confusion about the locking strategy.

* Improve encapsulation: Add shallowClone() method to vidMap

Added a shallowClone() method to vidMap to improve encapsulation and prevent
MasterClient from directly accessing vidMap's internal fields.

Changes:
1. Added vidMap.shallowClone() in vid_map.go
   - Encapsulates the shallow copy logic within vidMap
   - Makes vidMap responsible for its own state representation
   - Documented that caller is responsible for thread safety

2. Simplified resetVidMap() in masterclient.go
   - Uses tail := mc.vidMap.shallowClone() instead of manual field access
   - Cleaner, more maintainable code
   - Better adherence to encapsulation principles

Benefits:
- Improved code organization and maintainability
- vidMap internals are now properly encapsulated
- Easier to modify vidMap structure in the future
- More self-documenting code

Verified with: go test -race ./weed/wdclient/... (passes)

* Optimize locking: Reduce lock acquisitions and use helper methods

Two optimizations to further reduce lock contention and improve code consistency:

1. LookupFileIdWithFallback: Eliminated redundant lock acquisition
   - Before: Two separate locks to get vidMap and dataCenter
   - After: Single lock gets both values together
   - Benefit: 50% reduction in lock/unlock overhead for this hot path

2. KeepConnected: Use GetDataCenter() helper for consistency
   - Before: Manual lock/unlock to access DataCenter field
   - After: Use existing GetDataCenter() helper method
   - Benefit: Better encapsulation and code consistency

Impact:
- Reduced lock contention in high-traffic lookup path
- More consistent use of accessor methods throughout codebase
- Cleaner, more maintainable code

Verified with: go test -race ./weed/wdclient/... (passes)

* Refactor: Extract common locking patterns into helper methods

Eliminated code duplication by introducing two helper methods that encapsulate
the common locking patterns used throughout MasterClient:

1. getStableVidMap() - For read operations
   - Acquires lock, gets pointer, releases immediately
   - Returns stable snapshot for thread-safe reads
   - Used by: GetLocations, GetLocationsClone, GetVidLocations,
     LookupFileId, LookupVolumeServerUrl, GetDataCenter

2. withCurrentVidMap(f func(vm *vidMap)) - For write operations
   - Holds RLock during callback execution
   - Prevents pointer swap while allowing concurrent operations
   - Used by: addLocation, deleteLocation, addEcLocation, deleteEcLocation

Benefits:
- Reduced code duplication (eliminated 48 lines of repetitive locking code)
- Centralized locking logic makes it easier to understand and maintain
- Self-documenting pattern through named helper methods
- Easier to modify locking strategy in the future (single point of change)
- Improved readability - accessor methods are now one-liners

Code size reduction: ~40% fewer lines for accessor/helper methods

Verified with: go test -race ./weed/wdclient/... (passes)

* consistent

* Fix cache pointer race condition with atomic.Pointer

Use atomic.Pointer for vidMap cache field to prevent data races
during cache trimming in resetVidMap. This addresses the race condition
where concurrent GetLocations calls could read the cache pointer while
resetVidMap is modifying it during cache chain trimming.

Changes:
- Changed cache field from *vidMap to atomic.Pointer[vidMap]
- Updated all cache access to use Load() and Store() atomic operations
- Updated shallowClone, GetLocations, deleteLocation, deleteEcLocation
- Updated resetVidMap to use atomic operations for cache trimming

* Merge: Resolve conflict in deleteEcLocation - keep atomic.Pointer and fix bug

Resolved merge conflict by combining:
1. Atomic pointer access pattern (from HEAD): cache.Load()
2. Correct method call (from fix): deleteEcLocation (not deleteLocation)

Resolution:
- Before (HEAD): cachedMap.deleteLocation() - WRONG, reintroduced bug
- Before (fix): vc.cache.deleteEcLocation() - RIGHT method, old pattern
- After (merged): cachedMap.deleteEcLocation() - RIGHT method, new pattern

This preserves both improvements:
✓ Thread-safe atomic.Pointer access pattern
✓ Correct recursive call to deleteEcLocation

Verified with: go test -race ./weed/wdclient/... (passes)

* Update vid_map.go

* remove shallow clone

* simplify

3 months ago

request_id.go

Context-based logging with request ID (#6899)

7 months ago