Linux system calls table

System Calls¶

Lecture objectives:¶

• Linux system calls implementation
• VDSO and virtual syscalls
• Accessing user space from system calls

Linux system calls implementation¶

At a high level system calls are «services» offered by the kernel to user applications and they resemble library APIs in that they are described as a function call with a name, parameters and return value.

However, on a closer look, we can see that system calls are actually not function calls, but specific assembly instructions (architecture and kernel specific) that do the following:

• setup information to identify the system call and its parameters
• trigger a kernel mode switch
• retrieve the result of the system call

In Linux, system calls are identified by numbers and the parameters for system calls are machine word sized (32 or 64 bit). There can be a maximum of 6 system call parameters. Both the system call number and the parameters are stored in certain registers.

For example, on 32bit x86 architecture, the system call identifier is stored in the EAX register, while parameters in registers EBX, ECX, EDX, ESI, EDI, EBP.

System libraries (e.g. libc) offers functions that implement the actual system calls in order to make it easier for applications to use them.

When a user to kernel mode transition occurs, the execution flow is interrupted and it is transfered to a kernel entry point. This is similar with how interrupts and exception are handled (in fact on some architectures this transition happens as a result of an exception).

The system call entry point will save registers (which contains values from user space, including system call number and system call parameters) on stack and then it will continue with executing the system call dispatcher.

During the user — kernel mode transition the stack is also switched from ther user stack to the kernel stack. This is explained in more details in the interrupts lecture.

The purpose of the system call dispatcher is to verify the system call number and run the kernel function associated with the system call.

To demonstrate the system call flow we are going to use the virtual machine setup, attach gdb to a running kernel, add a breakpoint to the dup2 system call and inspect the state.

In summary, this is what happens during a system call:

• The application is setting up the system call number and parameters and it issues a trap instruction
• The execution mode switches from user to kernel; the CPU switches to a kernel stack; the user stack and the return address to user space is saved on the kernel stack
• The kernel entry point saves registers on the kernel stack
• The system call dispatcher identifies the system call function and runs it
• The user space registers are restored and execution is switched back to user (e.g. calling IRET)
• The user space application resumes

System call table¶

The system call table is what the system call dispatcher uses to map system call numbers to kernel functions:

System call parameters handling¶

Handling system call parameters is tricky. Since these values are setup by user space, the kernel can not assume correctness and must always verify them throughly.

Pointers have a few important special cases that must be checked:

• Never allow pointers to kernel-space
• Check for invalid pointers

Since system calls are executed in kernel mode, they have access to kernel space and if pointers are not properly checked user applications might get read or write access to kernel space.

For example, lets consider the case where such a check is not made for the read or write system calls. If the user passes a kernel-space pointer to a write system call then it can get access to kernel data by later reading the file. If it passes a kernel-space pointer to a read system call then it can corrupt kernel memory.

Likewise, if a pointer passed by the application is invalid (e.g. unmapped, read-only for cases where it is used for writing), it could «crash» the kernel. There two approaches that could be used:

• Check the pointer against the user address space before using it, or
• Avoid checking the pointer and rely on the MMU to detect when the pointer is invalid and use the page fault handler to determine that the pointer was invalid

Although it sounds tempting, the second approach is not that easy to implement. The page fault handler uses the fault address (the address that was accessed), the faulting address (the address of the instruction that did the access) and information from the user address space to determine the cause:

• Copy on write, demand paging, swapping: both the fault and faulting addresses are in user space; the fault address is valid (checked against the user address space)
• Invalid pointer used in system call: the faulting address is in kernel space; the fault address is in user space and it is invalid
• Kernel bug (kernel accesses invalid pointer): same as above

But in the last two cases we don’t have enough information to determine the cause of the fault.

In order to solve this issue Linux uses special APIs (e.g copy_to_user() ) to accesses user space that are specially crafted:

• The exact instructions that access user space are recorded in a table (exception table)
• When a page fault occurs the faulting address is checked against this table

Although the fault handling case may be more costly overall depending on the address space vs exception table size, and it is more complex, it is optimized for the common case and that is why it is preferred and used in Linux.

Cost	Pointer checks	Fault handling
Valid address	address space search	negligible
Invalid address	address space search	exception table search

Virtual Dynamic Shared Object (VDSO)¶

The VDSO mechanism was born out of the necessity of optimizing the system call implementation, in a way that does not impact libc with having to track the CPU capabilities in conjunction with the kernel version.

For example: x86 has two ways of issuing system calls: int 0x80 and sysenter. The later is significantly faster so it should be used when available. However, it is only available for processors newer than Pentium II and only for kernel versions greater than 2.6.

With VDSO the system call interface is decided by the kernel:

• a stream of instructions to issue the system call is generated by the kernel in a special memory area (formatted as an ELF shared object)
• that memory area is mapped towards the end of the user address space
• libc searches for VDSO and if present will use it to issue the system call

An interesting development of the VDSO are the virtual system calls (vsyscalls) which run directly from user space. These vsyscalls are also part of VDSO and they are accessing data from the VDSO page that is either static or modified by the kernel in a separate read-write map of the VDSO page. Examples of system calls that can be implemented as vsyscalls are: getpid or gettimeofday.

• «System calls» that run directly from user space, part of the VDSO
• Static data (e.g. getpid())
• Dynamic data update by the kernel a in RW map of the VDSO (e.g. gettimeofday(), time(), )

Accessing user space from system calls¶

As we mentioned earlier, user space must be accessed with special APIs ( get_user() , put_user() , copy_from_user() , copy_to_user() ) that check wether the pointer is in user space and also handle the fault if the pointer is invalid. In case of invalid pointers they return a non zero value.

Let’s examine the simplest API, get_user, as implemented for x86:

The implementation uses inline assembly, that allows inserting ASM sequences in C code and also handles access to / from variables in the ASM code.

Based on the type size of the x variable, one of __get_user_1, __get_user_2 or __get_user_4 will be called. Also, before executing the assembly call, ptr will be moved to the first register EAX while after the completion of assembly part the value of EAX will be moved to __ret_gu and the EDX register will be moved to __val_gu.

It is equivalent to the following pseudo code:

The __get_user_1 implementation for x86 is the following:

The first two statements check the pointer (which is stored in EDX) with the addr_limit field of the current task (process) descriptor to make sure that we don’t have a pointer to kernel space.

Then, SMAP is disabled, to allow access to user from kernel, and the access to user space is done with the instruction at the 1: label. EAX is then zeroed to mark success, SMAP is enabled, and the call returns.

The movzbl instruction is the one that does the access to user space and its address is captured with the 1: label and stored in a special section:

For each address that accesses user space we have an entry in the exception table, that is made up of: the faulting address(from), where to jump to in case of a fault, and a handler function (that implements the jump logic). All of these addresses are stored on 32bit in relative format to the exception table, so that they work for both 32 and 64 bit kernels.

All of the exception table entries are then collected in the __ex_table section by the linker script:

The section is guarded with __start___ex_table and __stop___ex_table symbols, so that it is easy to find the data from C code. This table is accessed by the fault handler:

All it does is to set the return address to the one in the to field of the exception table entry which, in case of the get_user exception table entry, is bad_get_user which return -EFAULT to the caller.

© Copyright The kernel development community.

Источник

yamnikov-oleg / calling_conventions.md

Source: man syscall

Architecture calling conventions

Every architecture has its own way of invoking and passing arguments to the kernel. The details for various architectures are listed in the two tables below.

The first table lists the instruction used to transition to kernel mode, (which might not be the fastest or best way to transition to the kernel, so you might have to refer to vdso(7) ), the register used to indicate the system call number, and the register used to return the system call result.

arch/ABI	instruction	syscall #	retval	Notes
arm/OABI	swi NR	—	a1	NR is syscall #
arm/EABI	swi 0x0	r7	r0
arm64	svc #0	x8	x0
blackfin	excpt 0x0	P0	R0
i386	int $0x80	eax	eax
ia64	break 0x100000	r15	r8	See below
mips	syscall	v0	v0	See below
parisc	ble 0x100(%sr2, %r0)	r20	r28
s390	svc 0	r1	r2	See below
s390x	svc 0	r1	r2	See below
sparc/32	t 0x10	g1	o0
sparc/64	t 0x6d	g1	o0
x86_64	syscall	rax	rax	See below
x32	syscall	rax	rax	See below

For s390 and s390x, NR (the system call number) may be passed directly with «svc NR» if it is less than 256.

The x32 ABI uses the same instruction as the x86_64 ABI and is used on the same processors. To differentiate between them, the bit mask __X32_SYSCALL_BIT is bitwise-ORed into the system call number for system calls under the x32 ABI.

On a few architectures, a register is used to indicate simple boolean failure of the system call: ia64 uses r10 for this purpose, and mips uses a3.

The second table shows the registers used to pass the system call arguments.

arch/ABI	arg1	arg2	arg3	arg4	arg5	arg6	arg7	Notes
arm/OABI	a1	a2	a3	a4	v1	v2	v3
arm/EABI	r0	r1	r2	r3	r4	r5	r6
arm64	x0	x1	x2	x3	x4	x5	—
blackfin	R0	R1	R2	R3	R4	R5	—
i386	ebx	ecx	edx	esi	edi	ebp	—
ia64	out0	out1	out2	out3	out4	out5	—
mips/o32	a0	a1	a2	a3	—	—	—	See below
mips/n32,64	a0	a1	a2	a3	a4	a5	—
parisc	r26	r25	r24	r23	r22	r21	—
s390	r2	r3	r4	r5	r6	r7	—
s390x	r2	r3	r4	r5	r6	r7	—
sparc/32	o0	o1	o2	o3	o4	o5	—
sparc/64	o0	o1	o2	o3	o4	o5	—
x86_64	rdi	rsi	rdx	r10	r8	r9	—
x32	rdi	rsi	rdx	r10	r8	r9	—

The mips/o32 system call convention passes arguments 5 through 8 on the user stack.

Note that these tables don’t cover the entire calling convention — some architectures may indiscriminately clobber other registers not listed here.

Review: cb4c4e8 on 2 Dec 2015.

32-bit system call numbers and entry vectors

Источник