Required reading: readi, writei, fileread, filewrite, create, and dirlink, and code related to these calls in fs.c, bio.c, ide.c, file.c, and sysfile.c
The next 3 lectures are about file systems:
Users desire to store their data durably, so that data survives when the user turns off their computer. The primary media for doing so are: magnetic disks, flash memory, and tapes. We focus on magnetic disks (e.g., through the IDE interface in xv6).
To allow users to remember where they stored a file, they can assign a symbolic name to a file, which appears in a directory.
The data in a file can be organized in a structured way or not. The structured variant is often called a database. UNIX uses the unstructured variant: files are streams of bytes. Any particular structure is likely to be useful to only a small class of applications, and other applications will have to work hard to fit their data into one of the pre-defined structures. Besides, if you want structure, you can easily write a user-mode library program that imposes that format on any file. The end-to-end argument in action. (Databases have special requirements and support an important class of applications, and thus have a specialized plan.)
The API for a minimal file system consists of: open, read, write, seek, close, and stat. Dup duplicates a file descriptor. For example:
fd = open("x/y", O_RDWR);
read (fd, buf, 100);
write (fd, buf, 512);
close (fd)
There's a couple of different things going on here.
Maintaining the file offset behind the read/write interface is an interesting design decision. The alternative is that the state of a read operation should be maintained by the process doing the reading (i.e., that the pointer should be passed as an argument to read). This argument is compelling in view of the UNIX fork() semantics, which clones a process which shares the file descriptors of its parent.
With offsets in the file descriptor, a read by the parent of a shared file descriptor (e.g., stdin) changes the read pointer seen by the child. This isn't always desirable: for example, consider a data file, in which the program seeks around and reads various records. If we fork(), the child and parent might interfere. On the other hand, the alternative (no offset in FD) would make it difficult to get "(echo one; echo two) > x" right. Easy to implement separate-offsets if kernel provides shared-offsets (re-open file, mostly), but not the other way around.
The file API turned out to be quite a useful abstraction. Unix uses it for many things that aren't just files on local disk, e.g. pipes, devices in /dev, network storage, etc. Plan9 took this further, and a few of those ideas came back to Linux, like the /proc filesystem.
Unix API doesn't specify that the effects of write are immediately on the disk before a write returns. It is up to the implementation of the file system within certain bounds. Choices include (that aren't non-exclusive):
What makes filesystems hard/interesting? At one level, it's much like any other part of an OS: trying to achieve good performance and provide useful functionality for lots of different applications with very different workload characteristics. However, more specifically:
A design issue is the semantics of a file system operation that requires multiple disk writes. In particular, what happens if the logical update requires writing multiple disks blocks and the power fails during the update? For example, to create a new file, requires allocating an inode (which requires updating the list of free inodes on disk), writing a directory entry to record the allocated i-node under the name of the new file (which may require allocating a new block and updating the directory inode). If the power fails during the operation, the list of free inodes and blocks may be inconsistent with the blocks and inodes in use. Again this is up to implementation of the file system to keep on disk data structures consistent:
Another design issue is the semantics are of concurrent writes to the same data item. What is the order of two updates that happen at the same time? For example, two processes open the same file and write to it. Modern Unix operating systems allow the application to lock a file to get exclusive access. If file locking is not used and if the file descriptor is shared, then the bytes of the two writes will get into the file in some order (this happens often for log files). If the file descriptor is not shared, the end result is not defined. For example, one write may overwrite the other one (e.g., if they are writing to the same part of the file.) Lots of other examples with directories, names, etc.
An implementation issue is performance, because writing to magnetic disk is relatively expensive compared to computing. Three primary ways to improve performance are: careful file system layout and data structure design that induces few seeks (locality, btrees, logging, etc), an in-memory cache of frequently-accessed blocks (or even prefetching), and overlap I/O with computation so that file operations don't have to wait until their completion and so that that the disk driver has more data to write, which allows disk scheduling. (We will talk about performance in detail later.)
xv6 implements a minimal Unix file system interface. xv6 doesn't pay attention to file system layout. It overlaps computation and I/O, but doesn't do any disk scheduling. Its cache is write-through, which simplifies keeping on disk datastructures consistent, but is bad for performance.
On disk files are represented by an inode (struct dinode in fs.h), and blocks. Small files have up to 12 block addresses in their inode; large files use files the last address in the inode as a disk address for a block with 128 disk addresses (512/4). The size of a file is thus limited to 12 * 512 + 128*512 bytes. What would you change to support larger files? (Ans: e.g., double indirect blocks.)
Directories are files with a bit of structure to them. The file contains of records of the type struct dirent. The entry contains the name for a file (or directory) and its corresponding inode number. How many files can appear in a directory?
In memory files are represented by struct inode in fsvar.h. What is the role of the additional fields in struct inode? Why don't we embed dinode in the directory? (Links from multiple directories to the same file.)
Kernel allows users to look at the inode state using stat() or fstat(), which returns a struct stat, from stat.h.
What is xv6's disk layout? Who determines how many inodes, blocks, etc. (mkfs.c). How does xv6 keep track of free blocks and inodes? See balloc()/bfree() and ialloc()/ifree(). Is this layout a good one for performance? What are other options?
Let's look at:
echo > a echo x > a echo < a > b rm a mkdir b
What file systems operations do these shell commands invoke?
1rst command. Let's walk through sys_open, create, etc, skipping directory stuff, and look at the number of disk writes. study ialloc; layout of inodes.
buffer cache caches blocks containing inodes. what replacement policy does the buffer cache implement? what is B_BUSY? B_VALID? why does bget() panic if there's no free blocks? could we sleep? if the block wasn't in the cache, read it from disk--are read's synchronous or asynchronous? why does the ide driver need a queue?
what does ilock(ip) after ialloc()? how does ilock() implement locks? why not use a spinlock to lock an inode? why have locks? Atomic read and update. For example, allocating an inode: read block containing inode, mark it allocated, and write it back. This operation must be atomic. Are writes write-through?
Returning to create, why is the update of the inode before the update of the directory? Can we make the disk write asynchronous? Is there any reason to wait in ide_rw() if we're writing a block?
2nd command: write a block. bmap allocates block. is data written to disk? must it?
3rd command. read. does the read cause a disk read to happen?
4th command. unlink. why are the old block written?