[PATCH V2] ARM : unwinder : Prevent data abort due to stack overflow while executing unwind instructions Signed-off-by: Anurag Aggarwal <a.anurag at samsung.com>
Anurag Aggarwal
anurag19aggarwal at gmail.com
Sat Nov 30 10:15:06 EST 2013
>> Looks like you still need to move your S-o-B line. It needs to be at
>> the end of the commit message.
>>
>> On Fri, Nov 29, 2013 at 09:34:31AM +0000, Anurag Aggarwal wrote:
>>> While unwinding backtrace, stack overflow is possible. This stack overflow can sometimes
>>> lead to data abort in system if the area after stack is not mapped to physical memory.
>>>
>>> To prevent this problem from happening execute the instructions that can cause data
>>> abort in there seperate functions instead of unwind_exec_insn, where a check for there
>>> feasibility is made first.
>>
>> Minor nit, but please wrap the commit message lines to 72 chars or less.
>> This helps the patch message to be listed nicely in git log.
>>
>>
>> If you agree with the changes I suggest below, the second paragraph
>> could be reworded something like:
>>
>> --snip--
>>
>> To prevent this problem from happening, execute the instructions that
>> can cause a data abort in separate helper functions, where a check for
>> feasibility is made before reading each word from the stack.
>>
>> --snip--
>>
>>
>>> ---
>>> arch/arm/kernel/unwind.c | 197 ++++++++++++++++++++++++++++++++++------------
>>> 1 files changed, 148 insertions(+), 49 deletions(-)
>>>
>>> diff --git a/arch/arm/kernel/unwind.c b/arch/arm/kernel/unwind.c
>>> index 00df012..150e0fc 100644
>>> --- a/arch/arm/kernel/unwind.c
>>> +++ b/arch/arm/kernel/unwind.c
>>> @@ -49,6 +49,8 @@
>>> #include <asm/traps.h>
>>> #include <asm/unwind.h>
>>>
>>> +#define TOTAL_REGISTERS 16
>>> +
>>> /* Dummy functions to avoid linker complaints */
>>> void __aeabi_unwind_cpp_pr0(void)
>>> {
>>> @@ -66,7 +68,7 @@ void __aeabi_unwind_cpp_pr2(void)
>>> EXPORT_SYMBOL(__aeabi_unwind_cpp_pr2);
>>>
>>> struct unwind_ctrl_block {
>>> - unsigned long vrs[16]; /* virtual register set */
>>> + unsigned long vrs[TOTAL_REGISTERS]; /* virtual register set */
>>> const unsigned long *insn; /* pointer to the current instructions word */
>>> int entries; /* number of entries left to interpret */
>>> int byte; /* current byte number in the instructions word */
>>> @@ -235,6 +237,148 @@ static unsigned long unwind_get_byte(struct unwind_ctrl_block *ctrl)
>>> return ret;
>>> }
>>>
>>> +static int unwind_exec_insn_0x80(struct unwind_ctrl_block *ctrl,
>>> + unsigned long insn)
>>
>> Since these are now split out as named functions, it's useful to have
>> human readable names.
>>
>> Maybe something like:
>>
>> unwind_exec_pop_subset_r4_to_r13
>>
>> What do you think?
>>
>>> +{
>>> + unsigned long high, low;
>>> + unsigned long mask;
>>> + unsigned long *vsp = (unsigned long *)ctrl->vrs[SP];
>>> + int load_sp, reg = 4;
>>> +
>>> + low = ctrl->vrs[SP];
>>> + high = ALIGN(low, THREAD_SIZE);
>>
>> You can calculate low, high and high - low once and store them in
>> unwind_ctrl_block. No need to recalculate them every time.
>
> I don't think it is feasible to store high and low in unwind_ctrl_block,
> we will have to recalculate them every time in this case also as the
> value of sp is which change every time and depending on the value
> of sp the value of high and low will also change.
>
>
>>
>> Field names like "low" may be confusing though. "sp_low" etc. may be
>> clearer.
>>
>>> +
>>> + insn = (insn << 8) | unwind_get_byte(ctrl);
>>> + mask = insn & 0x0fff;
>>> + if (mask == 0) {
>>> + pr_warning("unwind: 'Refuse to unwind' instruction %04lx\n",
>>> + insn);
>>> + return -URC_FAILURE;
>>> + }
>>> +
>>> + /*
>>> + * Check whether there is enough space
>>> + * on stack to execute the instruction
>>> + * if not then return failure
>>> + */
>>> + if ((high - low) < TOTAL_REGISTERS) {
>>
>> I'm still not sure we are optimising something valuable by factoring out
>> this check.
>
> Even I am confused as to you why you are not sure. From the documentation
> and the original source code, it seems clear that only the last set of registers
> can create a stack overflow for these three instructions, so why waste cpu
> cycles in calculations that are unnecessary
>
>
>>
>> It's also unfortunate that the code describes how many registers to pop
>> twice -- once in the check, and once in the while loop that does the
>> popping. That kind of duplication brings risks of accidentally breaking
>> the code during future maintenance. Someone might not modify the two
>> pieces of code in a consistent way. Because stack overflow should never
>> normally occur anyway, a mistake like that could easily get missed
>> during testing.
>>
>> Removing this whole if block and just doing a simple check each time a
>> register is loaded...
>>
>>
>>> + unsigned long mask_copy = mask;
>>> + int required_stack = 0;
>>> +
>>> + while (mask_copy) {
>>> + if (mask_copy & 1)
>>> + required_stack++;
>>> + mask_copy >>= 1;
>>> + }
>>> +
>>> + if ((high - low) < required_stack)
>>> + return -URC_FAILURE;
>>> + }
>>> +
>>> + load_sp = mask & (1 << (13 - 4));
>>> + while (mask) {
>>> + if (mask & 1)
>>> + ctrl->vrs[reg] = *vsp++;
>>
>> ... like this ...
>>
>> if (mask & 1) {
>> if (vsp >= high)
>> return -URC_FAILURE;
>>
>> ctrl->vrs[reg] = *vsp++;
>>
>> ... feels simpler.
>>
>> Better though, that can be factored out as a separate function:
Regarding this
>> static int unwind_pop_reg(struct unwind_ctrl_block *ctrl, unsigned long **vsp,
>> unsigned int reg)
>> {
>> if (*vsp >= ctrl->high)
>> return -URC_FAILURE;
>>
>> ctrl->vrs[reg] = *(*vsp)++;
>> }
>>
>> Then, in the insn helper, just do:
>>
>> if (mask & 1) {
>> if (unwind_pop_reg(ctrl, &vsp, reg))
>> return -URC_FAILURE;
>>
>>
>> And you can reuse that in each of the insn helpers.
>>
>> This is likely to be a bit less efficient, but I think it will reduce
>> the amount of code and make things a bit cleaner.
I agree that the code is simpler but it adds a lot of overhead as the
check that is
being made is not required every time.
This has been my main focus, to avoid overhead of checking stack
pointer every time
even though it is not needed.
The design of the instruction seems such to me that only the last of
registers will create
the problem,
>>> + mask >>= 1;
>>> + reg++;
>>> + }
>>> + if (!load_sp)
>>> + ctrl->vrs[SP] = (unsigned long)vsp;
>>> +
>>> + pr_debug("%s: fp = %08lx sp = %08lx lr = %08lx pc = %08lx\n", __func__,
>>> + ctrl->vrs[FP], ctrl->vrs[SP], ctrl->vrs[LR], ctrl->vrs[PC]);
>>> +
>>> + return URC_OK;
>>> +}
>>> +
>>> +static int unwind_exec_insn_0xa0(struct unwind_ctrl_block *ctrl,
>>
>> Maybe call this unwind_exec_pop_r4_to_rN?
>>
>>> + unsigned long insn)
>>> +{
>>> + unsigned long high, low;
>>> + unsigned long *vsp = (unsigned long *)ctrl->vrs[SP];
>>> + int reg;
>>> +
>>> + low = ctrl->vrs[SP];
>>> + high = ALIGN(low, THREAD_SIZE);
>>> +
>>> + /*
>>> + * Check whether there is enough space
>>> + * on stack to execute the instruction
>>> + * if not then return failure
>>> + */
>>> + if ((high-low) < TOTAL_REGISTERS) {
>>> + int required_stack;
>>> +
>>> + required_stack = insn & 7;
>>> + required_stack += (insn & 0x80) ? 1 : 0;
>>> +
>>> + if ((high-low) < required_stack)
>>> + return -URC_FAILURE;
>>> + }
>>> +
>>> + /* pop R4-R[4+bbb] */
>>> + for (reg = 4; reg <= 4 + (insn & 7); reg++)
>>> + ctrl->vrs[reg] = *vsp++;
>>> + if (insn & 0x80)
>>> + ctrl->vrs[14] = *vsp++;
>>> + ctrl->vrs[SP] = (unsigned long)vsp;
>>> +
>>> + pr_debug("%s: fp = %08lx sp = %08lx lr = %08lx pc = %08lx\n", __func__,
>>> + ctrl->vrs[FP], ctrl->vrs[SP], ctrl->vrs[LR], ctrl->vrs[PC]);
>>> +
>>> + return URC_OK;
>>> +}
>>> +
>>> +static int unwind_exec_insn_0xb1(struct unwind_ctrl_block *ctrl,
>>
>> Maybe call this unwind_exec_pop_subset_r0_to_r3?
>>
>>> + unsigned long insn)
>>> +{
>>> + unsigned long high, low;
>>> + unsigned long mask = unwind_get_byte(ctrl);
>>> + unsigned long *vsp = (unsigned long *)ctrl->vrs[SP];
>>> + int reg = 0;
>>> +
>>> + if (mask == 0 || mask & 0xf0) {
>>> + pr_warning("unwind: Spare encoding %04lx\n",
>>> + (insn << 8) | mask);
>>> + return -URC_FAILURE;
>>> + }
>>> +
>>> + low = ctrl->vrs[SP];
>>> + high = ALIGN(low, THREAD_SIZE);
>>> +
>>> + /*
>>> + * Check whether there is enough space
>>> + * on stack to execute the instruction
>>> + * if not then return failure
>>> + */
>>> + if ((high-low) < TOTAL_REGISTERS) {
>>> + unsigned long mask_copy = mask;
>>> + int required_stack = 0;
>>> +
>>> + while (mask_copy) {
>>> + if (mask_copy & 1)
>>> + required_stack++;
>>> + mask_copy >>= 1;
>>> + }
>>> + if ((high-low) < required_stack)
>>> + return -URC_FAILURE;
>>> + }
>>> +
>>> + /* pop R0-R3 according to mask */
>>> + while (mask) {
>>> + if (mask & 1)
>>> + ctrl->vrs[reg] = *vsp++;
>>> + mask >>= 1;
>>> + reg++;
>>> + }
>>> + ctrl->vrs[SP] = (unsigned long)vsp;
>>> +
>>> + pr_debug("%s: fp = %08lx sp = %08lx lr = %08lx pc = %08lx\n", __func__,
>>> + ctrl->vrs[FP], ctrl->vrs[SP], ctrl->vrs[LR], ctrl->vrs[PC]);
>>> +
>>> + return URC_OK;
>>> +}
>>> +
>>> /*
>>> * Execute the current unwind instruction.
>>> */
>>> @@ -249,65 +393,20 @@ static int unwind_exec_insn(struct unwind_ctrl_block *ctrl)
>>> else if ((insn & 0xc0) == 0x40)
>>> ctrl->vrs[SP] -= ((insn & 0x3f) << 2) + 4;
>>> else if ((insn & 0xf0) == 0x80) {
>>> - unsigned long mask;
>>> - unsigned long *vsp = (unsigned long *)ctrl->vrs[SP];
>>> - int load_sp, reg = 4;
>>> -
>>> - insn = (insn << 8) | unwind_get_byte(ctrl);
>>> - mask = insn & 0x0fff;
>>> - if (mask == 0) {
>>> - pr_warning("unwind: 'Refuse to unwind' instruction %04lx\n",
>>> - insn);
>>> - return -URC_FAILURE;
>>> - }
>>> -
>>> - /* pop R4-R15 according to mask */
>>> - load_sp = mask & (1 << (13 - 4));
>>> - while (mask) {
>>> - if (mask & 1)
>>> - ctrl->vrs[reg] = *vsp++;
>>> - mask >>= 1;
>>> - reg++;
>>> - }
>>> - if (!load_sp)
>>> - ctrl->vrs[SP] = (unsigned long)vsp;
>>> + return unwind_exec_insn_0x80(ctrl, insn);
>>> } else if ((insn & 0xf0) == 0x90 &&
>>> (insn & 0x0d) != 0x0d)
>>> ctrl->vrs[SP] = ctrl->vrs[insn & 0x0f];
>>> else if ((insn & 0xf0) == 0xa0) {
>>> - unsigned long *vsp = (unsigned long *)ctrl->vrs[SP];
>>> - int reg;
>>> -
>>> - /* pop R4-R[4+bbb] */
>>> - for (reg = 4; reg <= 4 + (insn & 7); reg++)
>>> - ctrl->vrs[reg] = *vsp++;
>>> - if (insn & 0x80)
>>> - ctrl->vrs[14] = *vsp++;
>>> - ctrl->vrs[SP] = (unsigned long)vsp;
>>> + return unwind_exec_insn_0xa0(ctrl, insn);
>>
>> There's an extra space before "return", here.
>>
>>> } else if (insn == 0xb0) {
>>> if (ctrl->vrs[PC] == 0)
>>> ctrl->vrs[PC] = ctrl->vrs[LR];
>>> /* no further processing */
>>> ctrl->entries = 0;
>>> } else if (insn == 0xb1) {
>>> - unsigned long mask = unwind_get_byte(ctrl);
>>> - unsigned long *vsp = (unsigned long *)ctrl->vrs[SP];
>>> - int reg = 0;
>>>
>>> - if (mask == 0 || mask & 0xf0) {
>>> - pr_warning("unwind: Spare encoding %04lx\n",
>>> - (insn << 8) | mask);
>>> - return -URC_FAILURE;
>>> - }
>>> -
>>> - /* pop R0-R3 according to mask */
>>> - while (mask) {
>>> - if (mask & 1)
>>> - ctrl->vrs[reg] = *vsp++;
>>> - mask >>= 1;
>>> - reg++;
>>> - }
>>> - ctrl->vrs[SP] = (unsigned long)vsp;
>>> + return unwind_exec_insn_0xb1(ctrl, insn);
>>
>> And here.
>>
>>> } else if (insn == 0xb2) {
>>> unsigned long uleb128 = unwind_get_byte(ctrl);
>>>
>>> --
>>> 1.7.0.4
>>>
>>>
>
>
>
> --
> Anurag Aggarwal
--
Anurag Aggarwal
More information about the linux-arm-kernel
mailing list