[PATCHv3] arm: ftrace: Adds support for CONFIG_DYNAMIC_FTRACE_WITH_REGS

[Abel Vesa]

On Fri, Feb 10, 2017 at 11:36:12AM +0100, Jean-Jacques Hiblot wrote:
> 2017-02-09 17:29 GMT+01:00 Russell King - ARM Linux <linux at armlinux.org.uk>:
> > On Tue, Feb 07, 2017 at 10:57:55PM +0000, Abel Vesa wrote:
> >> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_REGS
> >> +
> >> +.macro __ftrace_regs_caller
> >> +
> >> +     add     ip, sp, #4      @ move in IP the value of SP as it was
> >> +                             @ before the push {lr} of the mcount mechanism
> >> +     stmdb   sp!, {ip,lr,pc}
> >> +     stmdb   sp!, {r0-r11,lr}
> >> +
> >> +     @ stack content at this point:
> >> +     @ 0  4          44    48   52       56   60   64
> >> +     @ R0 | R1 | ... | R11 | LR | SP + 4 | LR | PC | previous LR |
> >
> > How important is this to be close to "struct pt_regs" ?  Do we care about
> > r12 being "wrong" ?  The other issue is that pt_regs is actually 72
> > bytes in size, not 68 bytes.  So, does that mean we end up inappropriately
> > leaking some of the kernel stack to userspace through ftrace?
> >
> > It's possible to save all the registers like this if we need to provide
> > a complete picture of the register set at function entry:
> >
> >         str     ip, [sp, #-16]!
> >         add     ip, sp, #20
> >         stmia   sp, {ip, lr, pc}
> >         stmdb   sp!, {r0 - r11}
> >
> > However, is that even correct - don't we want pt_regs' LR and PC to be
> > related to the function call itself?  The "previous LR" as you describe
> > it is where the called function (the one that is being traced) will
> > return to.  The current LR at this point is the address within the
> > traced function.  So actually I think this is more strictly correct, if
> > I'm understanding the intention here correctly:
> >
> >         str     ip, [sp, #S_IP - PT_REGS_SIZE]! @ save current IP
> >         ldr     ip, [sp, #PT_REGS_SIZE - S_IP]  @ get LR at traced function entry
> >         str     lr, [sp, #S_PC - S_IP]          @ save current LR as PC
> >         str     ip, [sp, #S_LR - S_IP]          @ save traced function return
> >         add     ip, sp, #PT_REGS_SIZE - S_IP + 4
> >         str     ip, [sp, #S_SP - SP_IP]         @ save stack pointer at function entry
> >         stmdb   sp!, {r0 - r11}
> >         @ clear CPSR and old_r0 words
> >         mov     r3, #0
> >         str     r3, [sp, #S_PSR]
> >         str     r3, [sp, #S_OLD_R0]
> >
> > However, that has the side effect of misaligning the stack (the stack
> > needs to be aligned to 8 bytes).  So, if we decide we don't care about
> > the saved LR value (except as a mechanism to preserve it across the
> > call into the ftrace code):
> >
> >         str     ip, [sp, #S_IP - PT_REGS_SIZE + 4]!
> >         str     lr, [sp, #S_PC - S_IP]
> >         ldr     lr, [sp, #PT_REGS_SIZE - 4 - S_IP]
> >         add     ip, sp, #PT_REGS_SIZE - S_IP
> >         stmib   sp, {ip, lr}
> >         stmdb   sp!, {r0 - r11}
> >         @ clear CPSR and old_r0 words
> >         mov     r3, #0
> >         str     r3, [sp, #S_PSR]
> >         str     r3, [sp, #S_OLD_R0]
> >
> > and the return would be:
> >
> >         ldmia   sp, {r0 - pc}
> >
> > That all said - maybe someone from the ftrace community can comment on
> > how much of pt_regs is actually necessary here?
> 
> I would suggest the following:
> r0-r11: filled with current values.
> r12 :  the value of r12 doesn't matter (Intra-procedure call scratch
> reg), we can either save it or not.
> r13 - sp: the value as it was when the instrumented function was
> entered. in the mcount case, it's the current sp value - 4, otherwise
> it'f sp -4
> r14 - lr: the value as it was when the instrumented function was
> entered. first element in stack or available in frame depending on
> GCC's version (mcount vs __gnu_mcount_nc)
> r15 - pc : the address after the modified instruction (value of lr
> when the ftrace caller is entered)
> 
So basically you're saying: save all regs, r0 through r15, except r12.
Based on that, I think it's easier to save all of them and then restore
all of them except r12. Plus, you have to take into consideration all 
the possible users of ftrace with regs. At some point some implementation
of ftrace_regs_call will probably need the value from r12. 
> I don't think we need CSPR and ORIG_r0.
I think we do. As I said before, because PT_REGS is 72 and some function
might (in the future) make use of CSPR or ORIG_r0, to ensure there is no
stack corruption taking place, we have to provide whole pt_regs, that is
72 (including CSPR and ORIG_r0). Plus, the stack alignment thing Russell
mentioned would be fixed.

The only problem I don't have a solution for at this point is OLD_LR (or previous LR
as it is called in this patch). If we take the approach described earlier,
we need to add to pt_regs the OLD_LR which I really don't like because it is 
breaking the whole purpose of pt_regs (it should only contain one copy of each reg, 
though it already has the ORIG_r0 in it) and will also break the stack alignment.
> 
> >
> > --
> > RMK's Patch system: http://www.armlinux.org.uk/developer/patches/
> > FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
> > according to speedtest.net.