慶應義塾大学
2013年度 春学期

システム・ソフトウェア
System Software / Operating Systems

2013年度春学期 火曜日1時限
科目コード: 60730
開講場所:SFC
授業形態:講義
担当: Rodney Van Meter
E-mail: rdv@sfc.keio.ac.jp

第9回 6月11日 ファイルシステム
Lecture 9, June 11: File Systems

Outline

Passing This Class

  1. So far, no one has submitted a project proposal via SFS, um, because there was no homework place to do so! That has been fixed. You have one week to submit them! Without a completed project, you will not pass this class!
  2. You are responsible for submitting five programming homeworks and several reports on readings.

Quoting from the first lecture:

Your grade will be determined as follows:

This year, heavier emphasis will be placed on the term project than on homeworks. You should expect your project to take 40-60 hours. There will be five homework assignments over the course of the semester, and readings almost every week. At least one of the programming assignments will involve parallel programming.

Summarizing the currently-assigned work:

Project:

* Proposal: (overdue, get them in this week!)
* final report & interview due: July 16-26

Programming:
HW1/L1. hello, world
HW2/L2. system call & assembly
HW3/L3. fork() + Berkeley parallel programming exercise (important!)
L4. (none)
L5. (none)
L6. (none)
L7. (none)
HW4/L8. study socket definition
L9. (none)
L10. TBD
L11. TBD
L12. TBD
L13. TBD
L14. TBD

Required Reading:
HW1/L1. Levin & Redell (reading & writing papers)
HW2/L2. Lampson, hints (system design)
L3. (none)
HW4+5/L4. Scherer, scalable sync queues (example of recent work
coupling OS principles to language design and broader systems design)
L5. (none)
HW?/L6. Raymond, Cathedral & Bazaar (software development philosophy)
L7. (none)
L8. (none)
L9. (none)
L10. TBD
L11. TBD
L12. TBD
L13. TBD
L14. TBD

Recommended Textbook Reading:
(various things, see the week-by-week readings)

Recommended Paper & Website Reading:
L1. (none)
L2. (none)
L3. (none)
L4. (none)
L5. Larus, spending Moore's dividend
L7. several things on page replacement
L8. several things on the Morris worm, Ken Thompson's hack, Yoshifuji on Linux IPv6
L9. Sweeney, XFS
L10. TBD
L11. TBD
L12. TBD
L13. TBD
L14. parallel debugging from SOSP 2009
Submit a proposal for your final report, including the topic and the list of papers you will read and review.

Final Project

Requirements, as discussed in class:

Contents of the final paper:

Examples of top systems conferences:

What's a File?

A regular file (the only kind we will deal with for the moment) is a collection of data stored on some non-volatile storage managed by the operating system.

The most common model of file today is the simple byte stream, but originally file I/O involved a much more hardware-oriented view and/or more sophisticated, database-oriented system services. There have been numerous types of basic files:

Using Files

File Access APIs

The file access APIs most commonly used today were developed on Unix, but they are far from the only ones. Even on Unix, there are several varieties, some system calls and some library routines, all oriented toward C:

On OpenVMS, there were several block and record-oriented services:

One extremely important feature of read() and write() is that they allow arbitrary alignment of the application's buffer. In the VMS block-oriented calls, applications had to align write data on 512-byte boundaries.

Regular files represent non-volatile data stored on disk. Using the standard APIs, however, the data may not be committed to disk when the write() call completes; the data may still be buffered in the system somewhere (e.g., in a special place in kernel memory). Again our concept of relativity, (相対性理論) comes into play: the information we have has not yet flowed to its final resting place. The data is guaranteed to be committed to disk once the close() is complete, or once a sync() call is complete. Note that data does not necessarily all land on disk in the order in which you wrote it! If you care, you should sync() every time you need guaranteed ordering of the writes. Also note that the semantics allow close() or sync() to fail even after the write() has succeeded!

File Session APIs

...What's missing from the above list?

Most file systems use the concept of a file session, spanning the time you first ask to access the file to the time you are done with it. The way you ask to access a file is via the open(), which requires a file name.

Most file APIs use the concept of a file pointer. The file pointer represents the current position in the file where the application is reading or writing. For regular files, it can be adjusted to the beginning, end, or any arbitrary position in the file. For byte stream files, the pointer is just an integer representing the byte offset; for record-oriented files, it is a more complex structure. The file pointer is part of the kernel's structure that it uses to track which files a process is accessing. The pointer assists the kernel in managing the reading and writing of file data.

File Management APIs

What's a File System?

Key Services

We have discussed individual files, but what's a file system? A file system is the overall structure that holds the user files (and sometimes other things, which we won't discuss today). A file system provides a number of services:

We will talk about space management next week.

File Naming: Directories, Devices and Mount Points

Current operating systems all store individual files in directories, or folders. A directory provides the mapping from a name to the actual file. When we speak of the files that are referenced by the directory, we say that a directory contains those files, but it's not literally true; the directory actually contains only pointers to the files.

Directories can contain other directories, known as subdirectories, creating a hierarchical namespace, or a directory tree. A complete path name may look like /home/rdv/keio/file1.

In Unix terminology, the normal mapping from name to file is a hard link. A regular file can have more than one hard link, or more than one name. A file with more than one hard link is not really deleted until the last link is deleted. Files can also have soft links, which are just name to name mappings that are held in the file system, but which do not participate in the actual management of the file. If the file is deleted but the soft link is not, the soft link is referred to as a dangling reference. One reason for the existence of soft links is to allow linking to directories without violating the requirement that each directory has a single parent. Another is to allow linking across partition boundaries or mount points.

In a Unix system, there is a single root to the directory tree. Applications and users only rarely have to know on which disk their data is stored. System managers can expand parts of the directory tree by mounting other file systems in any place in the tree.

In many other operating systems, the devices are explicitly named. On Windows, they have names such as C:\RDV\KEIO\file1, where the colon separates the device and the directory. On VMS, it could be SYS$HOME:[rdv.keio]file1.

Note that a name for a file is non-volatile, but not permanent; files can be renamed by users and applications, or the system manager may change the mount point and hence the full path name to a file. This behavior creates problems for long-term tracking of data, and numerous research systems (including Plan 9) have attempted to address this need.

Some file systems support case-sensitive file names, others do not. You have probably also noticed that sometimes non-ASCII file names are not printed properly. Most file systems originally assumed ASCII file names, and non-ASCII names are a problem because the character sets are not self-describing. NTFS solves this problem by storing all names in Unicode.

Security

All operating systems operate with some sort of security model, and the most important feature of that model is its plan for file protections. Multi-user OSes, by definition, understand users; the user is usually the basis of decisions about permissions. Unix and VMS both make a distinction between the user, the user's group, and "other" users.

Permissions, on Unix and its descendants, consist of read, write, and execute. Execute is used for both programs, and for directories.

An important, independent concept is that of an Access Control List, or ACL. An ACL matches specific permissions to either specific users, or to users holding a particular identifier or access token, which may be encrypted, and may be transferable from one user or process to another.

Other Metadata

The file system uses several forms of metadata, or data about the data, to manage files. We have seen two of the most important types of metadata already: names, and security information. But there are other types, which may or may not be supported:

Some of these attributes are preserved across backup and restore, or transfer to another system via ftp. Others are not, and sometimes problems occur.

Way back at the beginning, I mentioned file forks but didn't discuss them. Forks were originally developed for the Macintosh, to hold file icons. NTFS has a similar feature called file streams. These forks really blur the boundary between system metadata and user data, and typically are not preserved when files move between systems.

Other File Types

Homework, Etc.

Submit your project proposal. Older homeworks that had not been added to SFS are now there.

Next Lecture

Next lecture:

第10回 6月18日 ファイル・システムの実装
Lecture 10, June 18: File System Implementations

Readings for next week and followup for this week:

その他 Additional Information