Re: [PATCH 0/7] arm: Add livepatch support

[Abel Vesa]

On Wed, Dec 07, 2016 at 09:38:07AM +0800, zhouchengming wrote:
> On 2016/12/7 1:06, Abel Vesa wrote:
> >This is just an idea I've been trying out for a while now.
> >
> >Just in case somebody wants to play with it, this applies to linux-arm/for-next.
> >
> >Also please note that this was only tested in qemu, but I will do some testing
> >on some real hardware in the following days.
> >
> >FWICT, on this arch the compiler always generates a function prologue somewhere
> >between these lines:
> >
> >e1a0c00d        mov     ip, sp
> >e92ddff0        push    {r4-r9, sl, fp, ip, lr, pc}
> >e24cb004        sub     fp, ip, #4
> >e24dd064        sub     sp, sp, #100    ; 0x64<--- local variables
> >e52de004        push    {lr}            ; (str lr, [sp, #-4]!)
> >ebf9c2c9        bl      80110364<__gnu_mcount_nc>
> >....
> >
> >Every function that follows this pattern (the number of registers pushed and the
> >sp subtraction for the local variables being the only acceptable exception) can
> >be patched with this mechanism. IIRC, only the inline functions and notrace
> >functions do not follow this pattern.
> >
> >Considering that the function is livepatchable, when the time comes to call
> >ftrace_call, the ftrace_regs_caller is called instead.
> >
> >Because this arch didn't have a ftrace with regs implementation, the
> >ftrace_regs_caller was added.
> >
> >This new function adds the regs saving/restoring part, plus the part necessary
> >for the livepatch mechanism to work. After the regs are saved and the r3 is set
> >to contain the sp's value, we're keeping the old pc into r10 in order to be
> >checked later against the new pc.
> >
> >Next, the r1 and r0 are set for the ftrace_func, then, the ftrace_stub is called
> >and the klp_ftrace_handler overwrites the old pc with the new one.
> >
> >Here comes the tricky part. We're checking if the pc is still the old one, if it
> >is we jump the whole livepatching and go ahead with restoring the saved regs.
> >
> >If the pc is modified, it means we're livepatching current function and we need
> >to pop all regs from r1 through r12, jump over the next two regs saved on stack
> >(we're not interested in those since we're trying to get the same regs context
> >as it was at the point the function-to-be-patched was called) and put the new pc
> >into r11.
> >
> >Since r12 contains the sp from when the function just got branched to, we need
> >to set the sp back to that.
> >
> >Then we need to put the new pc on stack so that when we're popping r11 through
> >pc, we will actually jump to the first instruction from the new function.
> >
> >We don't need to worry about the returning phase since the epilogue of the new
> >function will take care of that and from there on everything goes back to
> >normal.
> >
> >The whole advantage of this over adding compiler support is that we're not
> >introducing nops at the beginning of the function. As a matter of fact, we're
> >not changing anything between an image with livepatch and an image without it
> >(except the ftrace_regs_call addition and the livepatch necessary code).
> >
> >As for the implementation of the ftrace_regs_caller, I still think there might
> >be some unsafe stack handling since I'm getting some build warnings. Those are
> >due to pushing/popping of a list of regs in which the sp resides. I'll try to
> >get around those in a next iteration (if necessary), but first I would like to
> >hear some opinions about this work and if it's worth going forward.
> >
> 
> Hi, so your idea is that when the pc is modified, we undo the work of the prologue
> of the old function, and then jump to the first instruction of the new function.
> But I doubt if we can really undo the work of the prologue correctly ? I don't know
> about arm, but gcc on arm64 may do some tricky things in prologue. So is there any
> chance we may restore a wrong context for the new function ?
> 
> Thanks.
>
I forgot to mention that this is actually taking advantage of how this arch deals 
with function calling. This mechanism might not be appliable to any other arch 
AFAIK. On arm 32bit, as long as the mcount prologue looks like in the shown 
example, the function is livepatchable. I will come back with a new version of
this patch today (latest tomorrow) with comments as Russel and Steven mentioned.
I hope that will clarify why this is working on arm 32bit.
> >Everything else should be pretty straightforward, so I'll skip explaining that.
> >
> >Abel Vesa (7):
> >   arm: Add livepatch arch specific code
> >   arm: ftrace: Add call modify mechanism
> >   arm: module: Add apply_relocate_add
> >   arm: Add ftrace with regs support
> >   arm: ftrace: Add ARCH_SUPPORTS_FTRACE_OPS for ftrace with regs
> >   arm: Add livepatch to build if CONFIG_LIVEPATCH
> >   arm: Add livepatch necessary arch selects into Kconfig
> >
> >  MAINTAINERS                      |  3 +++
> >  arch/arm/Kconfig                 |  4 ++++
> >  arch/arm/include/asm/ftrace.h    |  4 ++++
> >  arch/arm/include/asm/livepatch.h | 46 +++++++++++++++++++++++++++++++++++++
> >  arch/arm/kernel/Makefile         |  1 +
> >  arch/arm/kernel/entry-ftrace.S   | 49 ++++++++++++++++++++++++++++++++++++++++
> >  arch/arm/kernel/ftrace.c         | 21 +++++++++++++++++
> >  arch/arm/kernel/livepatch.c      | 43 +++++++++++++++++++++++++++++++++++
> >  arch/arm/kernel/module.c         |  9 ++++++++
> >  9 files changed, 180 insertions(+)
> >  create mode 100644 arch/arm/include/asm/livepatch.h
> >  create mode 100644 arch/arm/kernel/livepatch.c
> >
> 
>