Re: [PATCH] trace: print alloca pointers as actual pointer values

Kris Van Hees <[email protected]>
Newsgroups dev.linux.lists.dtrace
Message-ID <[email protected]>
On Fri, Aug 29, 2025 at 11:20:47PM -0400, Eugene Loh wrote:
> On 8/29/25 14:46, Kris Van Hees wrote:
> 
> > Because alloca pointers are stored internally as ofssets into the
> 
> s/ofssets/offsets/
> 
> For the test, how about converting the trace() to printf("%x")? Makes the
> script a little simpler and the hex output might feel more intuitive to a
> human reader if/when it ever comes to that.

The reason for using trace is that it is the most basic action that calls
dt_cg_store_val().  That is what we are exercising here.

> Most of all, the test fails wherever I tried it.  I think the problem is
> that the BPF verifier doesn't like adding a map_value pointer and an
> unbounded value... even if we do not dereference the result!  So, in this
> patch, drop the DT_NF_REF checks, instead calling
> dt_cg_alloca_access_check() everytime.

Ah, tehe joy of fixing one problem only to uncover another.  Dropping the
DT_NF_REF checks breaks other tests.  I'll have to see what needs to be done
to ensure that all tests pass correctly with a variant of this patch.

> I wrote a longer explanation of this.  Here it is, in case it's helpful.
> 
> The BPF verifier complains something like this:
> 
>     [...]
>     r8 = *(u64 *)(r10 -8) R8_w=fp-88          r8 = dctx
>     r8 = *(u64 *)(r8 +72) R8_w=map_value(...) r8 = dctx->gvar
>     r8 = *(u64 *)(r8 + 0) R8_w=inv(id=0)      r8 = arr
>     r7 = *(u64 *)(r10 -8) R7_w=fp-88          r7 = dctx
>     r7 = *(u64 *)(r7 +40) R7_w=map_value(...) r7 = dctx->scratchmem
>     r8 += r7              math between map_value pointer and register with
> unbounded min value is not allowed
> 
> So what has happened is that we assign r8=arr (first gvar), whose value is
> small since it's an alloca().  Then, we go to the new dt_cg.c code:
> 
>     if ((dnp->dn_flags & DT_NF_REF) || (arg & DT_NF_REF))
>         dt_cg_alloca_access_check(dlp, drp, dnp->dn_reg, DT_ISIMM, size);
>     dt_cg_alloca_ptr(dlp, drp, dnp->dn_reg, dnp->dn_reg);
> 
> We do NOT call the dt_cg_alloca_access_check(), but drop directly into
> dt_cg_alloca_ptr().  It allocates %r7 and does something like this:
> 
>     dt_cg_access_dctx(%r7, dlp, drp, DCTX_SCRATCHMEM);
>     emit(dlp,  BPF_ALU64_REG(BPF_ADD, %r8, %r7));
> 
> Anyhow, that last add is forbidden, because %r7 is dctx->scratchmem and %r8
> is unbounded.
> 
> I guess the problem is that the dt_cg_alloca_access_check() is not being
> called, and the BPF verifier doesn't like that, even though we are not
> dereferencing the pointer.
> 
> Maybe drop the &DT_NF_REF conditions in the new code introduced by the
> patch?

I think that is more likely that we need to scalarize the alloca pointer if
it not a REF, so that the arithmetic will be allowed (the verifier will no
longer know it is a map pointer).

> > scratchmem area, they were printed as small integers.  They are
> > now printed as actual pointer values into kernel space.
> > 
> > Signed-off-by: Kris Van Hees <[email protected]>
> > ---
> >   libdtrace/dt_cg.c                          | 19 +++++++++--------
> >   test/unittest/actions/trace/tst.alloca.d   | 24 ++++++++++++++++++++++
> >   test/unittest/actions/trace/tst.alloca.r   |  1 +
> >   test/unittest/actions/trace/tst.alloca.r.p | 11 ++++++++++
> >   4 files changed, 46 insertions(+), 9 deletions(-)
> >   create mode 100644 test/unittest/actions/trace/tst.alloca.d
> >   create mode 100644 test/unittest/actions/trace/tst.alloca.r
> >   create mode 100755 test/unittest/actions/trace/tst.alloca.r.p
> > 
> > diff --git a/libdtrace/dt_cg.c b/libdtrace/dt_cg.c
> > index cd9e7f4e9..7af3dd44b 100644
> > --- a/libdtrace/dt_cg.c
> > +++ b/libdtrace/dt_cg.c
> > @@ -1687,16 +1687,17 @@ dt_cg_store_val(dt_pcb_t *pcb, dt_node_t *dnp, dtrace_actkind_t kind,
> >   		align = vtype.dtdt_align;
> >   		/*
> > -		 * A DEREF of a REF node does not get resolved in dt_cg_node()
> > -		 * because the ref node already holds the pointer.  But for
> > -		 * alloca pointers, that will be the offset into scratchmem so
> > -		 * we still need to turn it into a real pointer here.
> > +		 * Alloca pointers are stored as an offset into scratchmem, so
> > +		 * they need to be converted into real pointers before we go on.
> > +		 * If the alloca pointer is a REF or ref-by-value is requested,
> > +		 * we need to do bounds checking before turning the alloca
> > +		 * pointer into a real pointer.
> >   		 */
> > -		if (dnp->dn_kind == DT_NODE_OP1 &&
> > -		    dnp->dn_op == DT_TOK_DEREF && (dnp->dn_flags & DT_NF_REF) &&
> > -		    (dnp->dn_child->dn_flags & DT_NF_ALLOCA)) {
> > -			dt_cg_alloca_access_check(dlp, drp, dnp->dn_reg,
> > -						  DT_ISIMM, size);
> > +		if (dnp->dn_flags & DT_NF_ALLOCA) {
> > +			if ((dnp->dn_flags & DT_NF_REF) || (arg & DT_NF_REF))
> > +				dt_cg_alloca_access_check(dlp, drp, dnp->dn_reg,
> > +							  DT_ISIMM, size);
> > +
> >   			dt_cg_alloca_ptr(dlp, drp, dnp->dn_reg, dnp->dn_reg);
> >   			not_null = 1;
> >   		}
> > diff --git a/test/unittest/actions/trace/tst.alloca.d b/test/unittest/actions/trace/tst.alloca.d
> > new file mode 100644
> > index 000000000..d2ff5152d
> > --- /dev/null
> > +++ b/test/unittest/actions/trace/tst.alloca.d
> > @@ -0,0 +1,24 @@
> > +#pragma D option quiet
> > +
> > +BEGIN
> > +{
> > +	arr = (int *)alloca(5 * sizeof(int));
> > +	idx = 4;
> > +	arr[0] = 1;
> > +	arr[1] = 22;
> > +	arr[2] = 333;
> > +	arr[3] = 4444;
> > +	arr[4] = 55555;
> > +	trace(arr);
> > +	trace(" ");
> > +	trace(*arr);
> > +	trace(" ");
> > +	trace(arr + 2);
> > +	trace(" ");
> > +	trace(*(arr + 2));
> > +	trace(" ");
> > +	trace(arr + idx);
> > +	trace(" ");
> > +	trace(*(arr + idx));
> > +	exit(0);
> > +}
> > diff --git a/test/unittest/actions/trace/tst.alloca.r b/test/unittest/actions/trace/tst.alloca.r
> > new file mode 100644
> > index 000000000..e9bbf2f5d
> > --- /dev/null
> > +++ b/test/unittest/actions/trace/tst.alloca.r
> > @@ -0,0 +1 @@
> > +OK 1 OK 333 OK 55555
> > diff --git a/test/unittest/actions/trace/tst.alloca.r.p b/test/unittest/actions/trace/tst.alloca.r.p
> > new file mode 100755
> > index 000000000..8515861ad
> > --- /dev/null
> > +++ b/test/unittest/actions/trace/tst.alloca.r.p
> > @@ -0,0 +1,11 @@
> > +#!/usr/bin/gawk -f
> > +
> > +{
> > +	$1 = $1 > 0x7fffffff ? "OK" : "BAD";
> > +	$3 = $3 > 0x7fffffff ? "OK" : "BAD";
> > +	$5 = $5 > 0x7fffffff ? "OK" : "BAD";
> > +}
> > +
> > +{
> > +	print;
> > +}
lmpx.com only provides a reader for public news (NNTP) servers. It is not affiliated with the servers or forums shown here and is not responsible for the content of articles, which is written by their respective authors.