exception.txt
上传用户:lgb322
上传日期:2013-02-24
资源大小:30529k
文件大小:12k
- Kernel level exception handling in Linux 2.1.8
- Commentary by Joerg Pommnitz <joerg@raleigh.ibm.com>
- When a process runs in kernel mode, it often has to access user
- mode memory whose address has been passed by an untrusted program.
- To protect itself the kernel has to verify this address.
- In older versions of Linux this was done with the
- int verify_area(int type, const void * addr, unsigned long size)
- function.
- This function verified that the memory area starting at address
- addr and of size size was accessible for the operation specified
- in type (read or write). To do this, verify_read had to look up the
- virtual memory area (vma) that contained the address addr. In the
- normal case (correctly working program), this test was successful.
- It only failed for a few buggy programs. In some kernel profiling
- tests, this normally unneeded verification used up a considerable
- amount of time.
- To overcome this situation, Linus decided to let the virtual memory
- hardware present in every Linux-capable CPU handle this test.
- How does this work?
- Whenever the kernel tries to access an address that is currently not
- accessible, the CPU generates a page fault exception and calls the
- page fault handler
- void do_page_fault(struct pt_regs *regs, unsigned long error_code)
- in arch/i386/mm/fault.c. The parameters on the stack are set up by
- the low level assembly glue in arch/i386/kernel/entry.S. The parameter
- regs is a pointer to the saved registers on the stack, error_code
- contains a reason code for the exception.
- do_page_fault first obtains the unaccessible address from the CPU
- control register CR2. If the address is within the virtual address
- space of the process, the fault probably occurred, because the page
- was not swapped in, write protected or something similar. However,
- we are interested in the other case: the address is not valid, there
- is no vma that contains this address. In this case, the kernel jumps
- to the bad_area label.
- There it uses the address of the instruction that caused the exception
- (i.e. regs->eip) to find an address where the execution can continue
- (fixup). If this search is successful, the fault handler modifies the
- return address (again regs->eip) and returns. The execution will
- continue at the address in fixup.
- Where does fixup point to?
- Since we jump to the contents of fixup, fixup obviously points
- to executable code. This code is hidden inside the user access macros.
- I have picked the get_user macro defined in include/asm/uaccess.h as an
- example. The definition is somewhat hard to follow, so let's peek at
- the code generated by the preprocessor and the compiler. I selected
- the get_user call in drivers/char/console.c for a detailed examination.
- The original code in console.c line 1405:
- get_user(c, buf);
- The preprocessor output (edited to become somewhat readable):
- (
- {
- long __gu_err = - 14 , __gu_val = 0;
- const __typeof__(*( ( buf ) )) *__gu_addr = ((buf));
- if (((((0 + current_set[0])->tss.segment) == 0x18 ) ||
- (((sizeof(*(buf))) <= 0xC0000000UL) &&
- ((unsigned long)(__gu_addr ) <= 0xC0000000UL - (sizeof(*(buf)))))))
- do {
- __gu_err = 0;
- switch ((sizeof(*(buf)))) {
- case 1:
- __asm__ __volatile__(
- "1: mov" "b" " %2,%" "b" "1n"
- "2:n"
- ".section .fixup,"ax"n"
- "3: movl %3,%0n"
- " xor" "b" " %" "b" "1,%" "b" "1n"
- " jmp 2bn"
- ".section __ex_table,"a"n"
- " .align 4n"
- " .long 1b,3bn"
- ".text" : "=r"(__gu_err), "=q" (__gu_val): "m"((*(struct __large_struct *)
- ( __gu_addr )) ), "i"(- 14 ), "0"( __gu_err )) ;
- break;
- case 2:
- __asm__ __volatile__(
- "1: mov" "w" " %2,%" "w" "1n"
- "2:n"
- ".section .fixup,"ax"n"
- "3: movl %3,%0n"
- " xor" "w" " %" "w" "1,%" "w" "1n"
- " jmp 2bn"
- ".section __ex_table,"a"n"
- " .align 4n"
- " .long 1b,3bn"
- ".text" : "=r"(__gu_err), "=r" (__gu_val) : "m"((*(struct __large_struct *)
- ( __gu_addr )) ), "i"(- 14 ), "0"( __gu_err ));
- break;
- case 4:
- __asm__ __volatile__(
- "1: mov" "l" " %2,%" "" "1n"
- "2:n"
- ".section .fixup,"ax"n"
- "3: movl %3,%0n"
- " xor" "l" " %" "" "1,%" "" "1n"
- " jmp 2bn"
- ".section __ex_table,"a"n"
- " .align 4n" " .long 1b,3bn"
- ".text" : "=r"(__gu_err), "=r" (__gu_val) : "m"((*(struct __large_struct *)
- ( __gu_addr )) ), "i"(- 14 ), "0"(__gu_err));
- break;
- default:
- (__gu_val) = __get_user_bad();
- }
- } while (0) ;
- ((c)) = (__typeof__(*((buf))))__gu_val;
- __gu_err;
- }
- );
- WOW! Black GCC/assembly magic. This is impossible to follow, so let's
- see what code gcc generates:
- > xorl %edx,%edx
- > movl current_set,%eax
- > cmpl $24,788(%eax)
- > je .L1424
- > cmpl $-1073741825,64(%esp)
- > ja .L1423
- > .L1424:
- > movl %edx,%eax
- > movl 64(%esp),%ebx
- > #APP
- > 1: movb (%ebx),%dl /* this is the actual user access */
- > 2:
- > .section .fixup,"ax"
- > 3: movl $-14,%eax
- > xorb %dl,%dl
- > jmp 2b
- > .section __ex_table,"a"
- > .align 4
- > .long 1b,3b
- > .text
- > #NO_APP
- > .L1423:
- > movzbl %dl,%esi
- The optimizer does a good job and gives us something we can actually
- understand. Can we? The actual user access is quite obvious. Thanks
- to the unified address space we can just access the address in user
- memory. But what does the .section stuff do?????
- To understand this we have to look at the final kernel:
- > objdump --section-headers vmlinux
- >
- > vmlinux: file format elf32-i386
- >
- > Sections:
- > Idx Name Size VMA LMA File off Algn
- > 0 .text 00098f40 c0100000 c0100000 00001000 2**4
- > CONTENTS, ALLOC, LOAD, READONLY, CODE
- > 1 .fixup 000016bc c0198f40 c0198f40 00099f40 2**0
- > CONTENTS, ALLOC, LOAD, READONLY, CODE
- > 2 .rodata 0000f127 c019a5fc c019a5fc 0009b5fc 2**2
- > CONTENTS, ALLOC, LOAD, READONLY, DATA
- > 3 __ex_table 000015c0 c01a9724 c01a9724 000aa724 2**2
- > CONTENTS, ALLOC, LOAD, READONLY, DATA
- > 4 .data 0000ea58 c01abcf0 c01abcf0 000abcf0 2**4
- > CONTENTS, ALLOC, LOAD, DATA
- > 5 .bss 00018e21 c01ba748 c01ba748 000ba748 2**2
- > ALLOC
- > 6 .comment 00000ec4 00000000 00000000 000ba748 2**0
- > CONTENTS, READONLY
- > 7 .note 00001068 00000ec4 00000ec4 000bb60c 2**0
- > CONTENTS, READONLY
- There are obviously 2 non standard ELF sections in the generated object
- file. But first we want to find out what happened to our code in the
- final kernel executable:
- > objdump --disassemble --section=.text vmlinux
- >
- > c017e785 <do_con_write+c1> xorl %edx,%edx
- > c017e787 <do_con_write+c3> movl 0xc01c7bec,%eax
- > c017e78c <do_con_write+c8> cmpl $0x18,0x314(%eax)
- > c017e793 <do_con_write+cf> je c017e79f <do_con_write+db>
- > c017e795 <do_con_write+d1> cmpl $0xbfffffff,0x40(%esp,1)
- > c017e79d <do_con_write+d9> ja c017e7a7 <do_con_write+e3>
- > c017e79f <do_con_write+db> movl %edx,%eax
- > c017e7a1 <do_con_write+dd> movl 0x40(%esp,1),%ebx
- > c017e7a5 <do_con_write+e1> movb (%ebx),%dl
- > c017e7a7 <do_con_write+e3> movzbl %dl,%esi
- The whole user memory access is reduced to 10 x86 machine instructions.
- The instructions bracketed in the .section directives are no longer
- in the normal execution path. They are located in a different section
- of the executable file:
- > objdump --disassemble --section=.fixup vmlinux
- >
- > c0199ff5 <.fixup+10b5> movl $0xfffffff2,%eax
- > c0199ffa <.fixup+10ba> xorb %dl,%dl
- > c0199ffc <.fixup+10bc> jmp c017e7a7 <do_con_write+e3>
- And finally:
- > objdump --full-contents --section=__ex_table vmlinux
- >
- > c01aa7c4 93c017c0 e09f19c0 97c017c0 99c017c0 ................
- > c01aa7d4 f6c217c0 e99f19c0 a5e717c0 f59f19c0 ................
- > c01aa7e4 080a18c0 01a019c0 0a0a18c0 04a019c0 ................
- or in human readable byte order:
- > c01aa7c4 c017c093 c0199fe0 c017c097 c017c099 ................
- > c01aa7d4 c017c2f6 c0199fe9 c017e7a5 c0199ff5 ................
- ^^^^^^^^^^^^^^^^^
- this is the interesting part!
- > c01aa7e4 c0180a08 c019a001 c0180a0a c019a004 ................
- What happened? The assembly directives
- .section .fixup,"ax"
- .section __ex_table,"a"
- told the assembler to move the following code to the specified
- sections in the ELF object file. So the instructions
- 3: movl $-14,%eax
- xorb %dl,%dl
- jmp 2b
- ended up in the .fixup section of the object file and the addresses
- .long 1b,3b
- ended up in the __ex_table section of the object file. 1b and 3b
- are local labels. The local label 1b (1b stands for next label 1
- backward) is the address of the instruction that might fault, i.e.
- in our case the address of the label 1 is c017e7a5:
- the original assembly code: > 1: movb (%ebx),%dl
- and linked in vmlinux : > c017e7a5 <do_con_write+e1> movb (%ebx),%dl
- The local label 3 (backwards again) is the address of the code to handle
- the fault, in our case the actual value is c0199ff5:
- the original assembly code: > 3: movl $-14,%eax
- and linked in vmlinux : > c0199ff5 <.fixup+10b5> movl $0xfffffff2,%eax
- The assembly code
- > .section __ex_table,"a"
- > .align 4
- > .long 1b,3b
- becomes the value pair
- > c01aa7d4 c017c2f6 c0199fe9 c017e7a5 c0199ff5 ................
- ^this is ^this is
- 1b 3b
- c017e7a5,c0199ff5 in the exception table of the kernel.
- So, what actually happens if a fault from kernel mode with no suitable
- vma occurs?
- 1.) access to invalid address:
- > c017e7a5 <do_con_write+e1> movb (%ebx),%dl
- 2.) MMU generates exception
- 3.) CPU calls do_page_fault
- 4.) do page fault calls search_exception_table (regs->eip == c017e7a5);
- 5.) search_exception_table looks up the address c017e7a5 in the
- exception table (i.e. the contents of the ELF section __ex_table)
- and returns the address of the associated fault handle code c0199ff5.
- 6.) do_page_fault modifies its own return address to point to the fault
- handle code and returns.
- 7.) execution continues in the fault handling code.
- 8.) 8a) EAX becomes -EFAULT (== -14)
- 8b) DL becomes zero (the value we "read" from user space)
- 8c) execution continues at local label 2 (address of the
- instruction immediately after the faulting user access).
- The steps 8a to 8c in a certain way emulate the faulting instruction.
- That's it, mostly. If you look at our example, you might ask why
- we set EAX to -EFAULT in the exception handler code. Well, the
- get_user macro actually returns a value: 0, if the user access was
- successful, -EFAULT on failure. Our original code did not test this
- return value, however the inline assembly code in get_user tries to
- return -EFAULT. GCC selected EAX to return this value.
- NOTE:
- Due to the way that the exception table is built and needs to be ordered,
- only use exceptions for code in the .text section. Any other section
- will cause the exception table to not be sorted correctly, and the
- exceptions will fail.