http://www.linuxsa.org.au/pipermail/linuxsa/2002-February/038793.html[OT] raw-disks && cooked-disks ?Glen Turner glen.turner at aarnet.edu.auThu Feb 21 17:33:19 CST 2002
On Thu, 21 Feb 2002, Wilkinson,Alex wrote: > What I don't know and find confusing is the difference between mkfs on either a > raw disk or a cooked disk. Traditional UNIXen have - a raw device. Byte 0 is the first byte on the disk. - a cooked device, better thought of as a 'formatted' device. Byte 0 is the first byte on the device that is available to user data. Think of a tape drive with a compulsory tape label. Writing to byte 0 of the raw tape device will overwrite the first byte of the label, but writing to byte 0 of the cooked tape device will write to the first byte after the label (that is, the first byte of user data). So we use the cooked device name as input to applications like 'tar', as we are writing user data. But we use the raw device name as input to 'label', as we are writing the tape's system data. For disks the distinction gets fuzzier. After all you access disks by filename, not by byte offset. But you can see how 'mkfs', with its need to write superblocks and i-nodes and all sorts of other system stuff, needs the raw device name. Linux broke with the UNIXen tradition and did away with raw devices. After all, as we noticed they aren't really needed for using disks and Linux was a IBM PC based operating system without an old-fashioned compulsory-label tape drive in sight. So when you use mkfs on Linux you give the name of the cooked device. It turns out that not needing raw devices is only sort-of-true. User-space filesystems need access to the raw disk device, just like kernel filesystems have. And databases like Oracle work fastest when they use their own filesystem. So later Linux did add raw devices, but in a completely different way to all other UNIXen. Ouch. > Why does mkfs only work /dev/rdsk/... on IRIX but on others it can > work with /dev/dsk/...? Because this is such a common error and the distinction between raw devices and cooked devices is so minor for disks that the program should really accept either name and format the disk. Anything else is just pedantic. > I know that raw disks are character devices and cooked disks are block > devices and that by running mkfs on a disk we are creating the > a fs with a specfic block size and layout etc... Whoa. Devices are classified into character or block. Think of it as: Character devices can only be accessed serially (read character 1, read character 2). With Block devices any block can be retrieved (give me block 10 now, without reading block 1,2,3,...,7,8,9 first). *Access methods* in UNIX are cooked or raw. In traditional UNIXen the access method is indicated by which device you use. This isn't the only alternative: Linux uses an ioctl() directive to set the access method; on IBM mainframes you link in differing I/O libraries; VMS has a whole raft of access methods, you select the one you want when you OPEN the file. > But I would have thought you could only create a fs on a character disk ? You can create a filesystem on blocked or character devices. Blocked devices such as disks are by far the most common in these times. But newer tape technologies are formatted with a filesystem and accessed by filename just like disks are. The Linux kernel has really good filesystem support. Almost anything that can be named can be turned into a filesystem. smbfs and NFS are filesystems that send bytes over a network rather than access disk. davfs can be used to access web sites. Linux itself has a filesystem that is used to access the kernel's confguration and counters (procfs, which is usually mounted at /proc). So the filesystem these days isn't really thought of as being restricted to any device type. |
