If you have several operating systems on your disks, then each uses one or more disk partitions. A disagreement on where these partitions are may have catastrophic consequences.
The MBR contains a partition table describing where the (primary) partitions are. There are 4 table entries, for 4 primary partitions, and each looks like
(where CHS stands for Cylinder/Head/Sector).struct partition { char active; /* 0x80: bootable, 0: not bootable */ char begin[3]; /* CHS for first sector */ char type; char end[3]; /* CHS for last sector */ int start; /* 32 bit sector number (counting from 0) */ int length; /* 32 bit number of sectors */ };
This information is redundant: the location of a partition is given both by the 24-bit begin and end fields, and by the 32-bit start and length fields.
Linux only uses the start and length fields, and can therefore handle partitions of not more than 2^32 sectors, that is, partitions of at most 2 TiB. That is thirty times larger than the disks available today, so maybe it will be enough for the next six years or so. (So, partitions can be very large, but there is a serious restriction in that a file in an ext2 filesystem on hardware with 32-bit integers cannot be larger than 2 GiB.)
DOS uses the begin and end fields, and uses the BIOS INT13 call to access the disk, and can therefore only handle disks of not more than 8.4 GB, even with a translating BIOS. (Partitions cannot be larger than 2.1 GB because of restrictions of the FAT16 file system.) The same holds for Windows 3.11 and WfWG and Windows NT 3.*.
Windows 95 has support for the Extended INT13 interface, and uses special partition types (c, e, f instead of b, 6, 5) to indicate that a partition should be accessed in this way. When these partition types are used, the begin and end fields contain dummy information (1023/255/63). Windows 95 OSR2 introduces the FAT32 file system (partition type b or c), that allows partitions of size at most 2 TiB.
What is this nonsense you get from fdisk about `overlapping' partitions, when in fact nothing is wrong? Well - there is something `wrong': if you look at the begin and end fields of such partitions, as DOS does, they overlap. (And that cannot be corrected, because these fields cannot store cylinder numbers above 1024 - there will always be `overlap' as soon as you have more than 1024 cylinders.) However, if you look at the start and length fields, as Linux does, and as Windows 95 does in the case of partitions with partition type c, e or f, then all is well. So, ignore these warnings when cfdisk is satisfied and you have a Linux-only disk. Be careful when the disk is shared with DOS. Use the commands cfdisk -Ps /dev/hdx and cfdisk -Pt /dev/hdx to look at the partition table of /dev/hdx.
Many old IBM PS/2 systems used disks with a defect map written to the end of the disk. (Bit 0x20 in the control word of the disk parameter table is set.) Therefore, FDISK would not use the last cylinder. Just to be sure, the BIOS often already reports the size of the disk as one cylinder smaller than reality, and that may mean that two cylinders are lost. Newer BIOSes have several disk size reporting functions, where internally one calls the other. When both subtract 1 for this reserved cylinder and also FDISK does so, then one may lose three cylinders. These days all of this is irrelevant, but this may provide an explanation if one observes that different utilities have slightly different opinions about the disk size.
A well-known claim says that partitions should start and end at cylinder boundaries.
Since "disk geometry" is something without objective existence, different operating systems will invent different geometries for the same disk. One often sees a translated geometry like */255/63 used by one and an untranslated geometry like */16/63 used by another OS. Thus, it may be impossible to align partitions to cylinder boundaries according to each of the the various ideas about the size of a cylinder that one's systems have. Also different Linux kernels may assign different geometries to the same disk. Also, enabling or disabling the BIOS of a SCSI card may change the fake geometry of the connected SCSI disks.
Fortunately, for Linux there is no alignment requirement at all. (Except that some semi-broken installation software likes to be very sure that all is OK; thus, it may be impossible to install RedHat 7.1 on a disk with unaligned partitions because DiskDruid is unhappy.)
People report that it is easy to create nonaligned partitions in Windows NT, without any noticeable bad effects.
But MSDOS 6.22 has an alignment requirement. Extended partition sectors that are not on a cylinder boundary are ignored by its FDISK. The system itself is happy with any alignment, but interprets relative starting addresses as if relative to an aligned address: The starting address of a logical partition is given relative not to the address of the extended partition sector that describes it, but relative to the start of the cylinder that contains that sector. (So, it is not surprising that also PartitionMagic requires alignment.)
What is the definition of alignment? MSDOS 6.22 FDISK will do the following: 1. If the first sector of the cylinder is a partition table sector, then the rest of the track is unused, and the partition starts with the the next track. This applies to sector 0 (the MBR) and the partition table sectors preceding logical partitions. 2. Otherwise, the partition starts at the first sector of the cylinder. Also the extended partition starts at a cylinder boundary. The cfdisk man page says that old versions of DOS did not align partitions.
Use of partition type 85 for the extended partition makes it invisible to DOS, making sure that only Linux will look inside.
As an aside: on a Sparc, the boot partition must start on a cylinder boundary (but there is no requirement on the end).