Lecture: crash consistency

All File Systems Are Not Created Equal: On the Complexity of Crafting Crash-Consistent Applications, OSDI 2014

Discussion lead: Sowmya Janakiram Dharanipragada, Helga Gudmundsdottir, and Helgi Sigurbjarnarson

Questions

• Consider the following update protocol for overwriting adjacent blocks i and j (from old data a to new data b) atomically in a file foo:
  • i) Create a log file using the creat(/log) system call.
  • ii) Write the current contents of blocks i and j in foo using write(/log, “i, j, a_i, a_j”).
  • iii) Replace the contents of blocks i and j in foo: pwrite(/foo, i, “b_i, b_j”).
  • iv) Delete the log via unlink(/log).

  A recovery function would (on startup) check for the existence of /log and continue from iii) if it is well formed. Ideally, this protocol results in either: both blocks i and j are updated, or neither is updated. Describe all the other possible states under the weak APM model. How can you fix it?

• What is the difference between stdout and fsync in the APM in terms of ordering? Why do their definitions make sense?

• What kind of vulnerabilities does ALICE focus on detecting? Give one example of a bug that ALICE can detect and one that it cannot detect.

• Provide a list of questions you would like to discuss in class. Feel free to provide any comments on the paper and related topics (e.g., which parts you like and which parts you find confusing).