Problem: FS Consistency after a crash

File System Inconsistencies

• lost some file data: tolerable

• inconsistent metadata: not tolerable
• metadata could include: inode, directory, bitmap for resouce management

fsck: File System Checker

• check & restore consistency

• takes long time to run

• can’t always restore consistency, may need human intervention

File Buffer Cache

• caches data in memory

• write back to disk periodically

• hold data and metadata

• fsync syscall: force to flush to disks

Consistency in FFS

• update metadata synchronously

Soft Updates: A Solution to the Metadata Update Problem in File Systems

Ordering Constraints in FS

• no dangling pointers: initilize before pointing to an object

• don’t reuse until old pointers be removed

• never reset old pointers until new pointers persists (example: rename file)

Approach

• write modifications to buffer cache

• track dependencies between blocks

• write blocks asynchronously in a safe order (blocks with no dependencies should be written first ideally)

Challenges: circular dependencies

• if dependencies are tracked at block level, there are often circular dependencies

Dependencies

• Create files: (ordering constraint 1)
• create a inode
• update directory entry to points to it

• Delete files: (ordering constrains 2)
• remove directory entry
• update inode

Softupdate Solutions

• fine-grained dependency tracking: track at pointer level

• undo/redo
• temporarily undo updates in memory
• write back
• redo updates in memory

Overhead

• cpu overhead: extra work such as undo / redo and mataining dependencies

• memory overhead: track dependencies data structures

Optimizations

• write many blocks at once (when no dependencies)

• skip write that are undone

Evaluation

• Conventional FFS: reliable, poor performance

• FFS async write (No-Order): good performance, no consistency guarentee

• Softupdate

• Write Ahead Logging

• result: performance approaches no-order, better recovery

Summary

• metadata consistency and ordering constraints

• softupdates:
• fine-grained dependency tracking
• undo / redo

SplitFS: Reducing Software Overhead in File Systems for Persistent Memory

Persistent Memory

• attached via memory bus (faster than PCIe)

• released by Intel at around 2019

PM Property

• persistent

• 1 / 2 - 1 / 3 bandwidth of DRAM

• 2 - 4x slower than DRAM

• load / store interface

• larger capacity than DRAM, smaller than disks

• byte addressable

Challenges: caching

• need to force write through
• non-temporal move MOVNT on x86
• flush CLFLUSH on x86
• fence instructions to prevent reordering

SplitFS

• NOVA: in kernel FS

• Others: in user FS

• SplitFS: hybrid approach
• data operations: in user space
• meta data operations: in kernel space
• optimize common case (data operations)

• goal: handle unmodified apps

• use staging files for consistency

• use relink to avoid copy

Summary

• hybrid (user/kernel) FS for persistent memory

• optimize for data operations

• staging files & relink