[DTrace-devel] [PATCH 01/14] Fix stack-skip counts for caller and stackdepth

[Nick Alcock]

On 19 Jun 2025, Kris Van Hees uttered the following:

> On Thu, Jun 19, 2025 at 12:20:22PM -0400, Kris Van Hees wrote:
>> On Thu, Jun 19, 2025 at 02:03:56PM +0100, Nick Alcock wrote:
>> > On 16 Jun 2025, Eugene Loh verbalised:
>> > 
>> > > On 6/13/25 10:33, Nick Alcock wrote:
>> > >
>> > >>> Note that we declare the skip count volatile.  The compiler might
>> > >>> optimize code that uses the STACK_SKIP value, but we will subsequently
>> > >>> perform relocations that adjust this value.
>> > >> ... why doesn't this apply to every other extern global variable in
>> > >> get_bvar()? They're all similarly relocated...
>> > >
>> > > Right.  There is potentially a broader problem.  But we simply do not have evidence of misbehavior in other cases.  Ruggedizing
>> > > other cases could be the subject of a different patch.
>> > 
>> > Aha, OK. I was just wondering if there was some extra reason.
>> > 
>> > > The problem in this case is that the compiler seems to assume &symbol!=0, which is reasonable except that we violate that behavior
>> > > for our relocation tricks.
>> > 
>> > I wonder where the code for that is... plenty of symbols have value
>> > zero.
>> > 
>> > But, really...
>> > 
>> > > Consider the C code:
>> > >
>> > >     uint64_t dt_bvar_stackdepth(const dt_dctx_t *dctx)
>> > >     {
>> > >         uint32_t          bufsiz = (uint32_t) (uint64_t) (&STKSIZ);
>> > >         char              *buf = dctx->mem + (uint64_t)(&STACK_OFF);
>> > 
>> > Hm...
>> > 
>> > extern uint64_t STACK_SKIP;
>> > 
>> > So we encode information about the stack size and skip value by encoding
>> > it in the *address* of the variable? Is there some reason we don't use
>> > its value? unlike the stack offset, we're *using* it as a value, not an
>> > address...
>> 
>> The issue seems to be (and perhaps this is a cross compiler problem) that e.g.
>> 
>> extern uint64_t PC;
>> 
>> and then code accessing the value of PC (e.g. foo(PC) as a call argument) will
>> yield:
>> 
>>   50:   18 02 00 00 00 00 00 00         lddw %r2,0
>>   58:   00 00 00 00 00 00 00 00 
>>                         50: R_BPF_INSN_64       PC
>>   60:   79 22 00 00 00 00 00 00         ldxdw %r2,[%r2+0]
>> 
>> which shows that it is interpreting PC as an address to a symbol, because it
>> loads the address of the symbol and then dereferences it with offset 0.  So,
>> we cannot plug in the value during relocation because the only value we can
>> put there would be an address where the vlaue can be found.  To get around
>> this, we "use" Tthe address as the entity to store the value in, knowing that
>> we *never* will interpret it as an address for these specific externs.
>
> Not a compiler error - since PC is extern uint64_t PC it *is* a variable and
> so it is present (and accessible) as an address in .data in the location where
> it is actually defined.  Since we never define it, we don't have a .data (which
> is fine because we only use this for constants known at link time) BUT the
> compiler of course is free to assume that we *do* have an address to the
> storage location for PC and thus that we get the value that way.  It does not
> know that the value is constant.  So, the trick I use is needed to make this
> work.

I'm just not sure why it's assuming that its value cannot be NULL, given
that you're not dereferencing it anywhere. I guess it's because it's a
symbol, but that's only a guess: I haven't managed to find the code that
implements this (?)optimization.

Anyway, this is not an objection to this patch, and a patch changing
this approach to something that didn't require us to throw volatiles
around and hope the compiler doesn't optimize things wrong would be a
separate patch anyway: you have my

Revieed-by: Nick Alcock <nick.alcock at oracle.com>