CALL

A near CALL pushes IP, which already points to the instruction after the CALL, and jumps to the target: SP goes down by 2 and [SP] holds the return offset. A far CALL pushes CS first and then IP, four bytes in all, and loads both CS and IP. The procedure later executes RET (or RETF), which pops that address and comes back. Flags are not touched.

There is no short form: a near direct call always reaches the whole segment. The indirect forms take the address from a register or memory, for tables of procedures and callbacks.

MOV     DX, OFFSET MSG1
CALL    PRINT1        ; near CALL: pushes the return IP
MOV     DX, OFFSET MSG2
MOV     BX, OFFSET PRINT1
CALL    BX            ; the same procedure, address in a register
 
PRINT1  PROC    NEAR  ; print the $-string at DS:DX
MOV     AH, 09h
INT     21h
RET
PRINT1  ENDP
MSG1    DB      "First call", 13, 10, "$"
MSG2    DB      "Second call", 13, 10, "$"

Pitfalls

The return address sits on top of the stack, so the procedure must pop exactly what it pushed before RET; one stray PUSH and RET jumps to whatever value that was. A procedure changes registers unless it saves them, so document what it keeps. Near and far must match: a near RET after a far CALL leaves CS on the stack. A .COM program cannot CALL FAR PTR a label of its own, because a .COM file has no relocations; a CALL followed directly by RET can be replaced with a plain JMP.

Later CPUs

From the 386, 32-bit code pushes 4-byte return offsets (EIP). The 16-bit forms are unchanged.