[DTrace-devel] [PATCH 16/16] dis: add support for variables used in trampoline code
Kris Van Hees
kris.van.hees at oracle.com
Fri Jan 19 15:16:03 UTC 2024
On Thu, Jan 18, 2024 at 11:43:41PM -0500, Eugene Loh via DTrace-devel wrote:
> Reviewed-by: Eugene Loh <eugene.loh at oracle.com>
> with comments...
>
> On 1/12/24 19:24, Kris Van Hees via DTrace-devel wrote:
> > diff --git a/libdtrace/dt_dis.c b/libdtrace/dt_dis.c
> > @@ -90,6 +90,11 @@ dt_dis_varname_off(const dtrace_difo_t *dp, uint_t off, uint_t scope, uint_t add
> > *
> > * For variables, the sequence of instructions we are looking for is:
> > * insn code dst src offset imm
> > + * -1: ld dst %r9 DCTX_*VARS 00000000
> > + * 0: ld xxx dst var_offset 00000000
> > + * 0: st dst src var_offset 00000000
> > + * 0: add dst 0 0 var_offset
> > + * -or-
> > * -2: ld dst %fp DT_STK_DCTX 00000000
> > * -1: ld dst dst DCTX_*VARS 00000000
> > * 0: ld xxx dst var_offset 00000000
> > @@ -103,6 +108,9 @@ dt_dis_varname_off(const dtrace_difo_t *dp, uint_t off, uint_t scope, uint_t add
> > *
> > * For string constants, the sequence of instructions we are looking for is:
> > * insn code dst src offset imm
> > + * -1: ld dst %r9 DCTX_STRTAB 00000000
> > + * 0: add dst 0 0 var_offset
> > + * -or-
> > * -2: ld dst %fp DT_STK_DCTX 00000000
> > * -1: ld dst dst DCTX_STRTAB 00000000
> > * 0: add dst 0 0 var_offset
>
> Judging comments is a subjective matter, but the comments here IMHO can be
> "factored" apart. Specifically, this whole disasm stuff is tied intimately
> to what we do in cg, and so specific references to what we're doing in cg
> strikes me not only as fair but even as helpful. And referring to
> dt_cg_access_dctx() explicitly helps (me) understand what's going on and
> separates out something that is common to both vars and strings. Listing
> instructions -2 and -1 twice (once for vars and once for strings) in the
> comments is not only redundant but also does not follow what happens in the
> actual code. Also, since the in-code comments refer to trampolines, so it
> makes sense for the comment block to follow suit.
>
> In all, here is my suggestion for the comment block:
>
> /*
> * Check if we are loading from the gvar, lvar, or strtab BPF map.
> *
> * If we are loading a gvar or lvar, we want to report the variable name.
> * If we are loading a string constant, we want to report its value.
> *
> * For variables, the current instruction (instruction 0) is one of:
> * insn code dst src offset imm
> * 0: ld xxx dst var_offset 00000000
> * 0: st dst src var_offset 00000000
> * 0: add dst 0 0 var_offset
> * representing:
> * - load by value
> * - store by value
> * - access by reference
> *
> * For string constants, the current instruction is:
> * 0: add dst 0 0 var_offset
> *
> * In either case, the preceding instructions are generated by
> * dt_cg_access_dctx() and are:
> * if not trampoline:
> * insn code dst src offset imm
> * -2: ld dst %fp DT_STK_DCTX 00000000
> * -1: ld dst dst DCTX_xxx 00000000
> * if trampoline:
> * -1: ld dst %r9 DCTX_xxx 00000000
> * where DCTX_xxx is DCTX_LVARS, DCTX_GVARS, or DCTX_STRTAB.
> *
> * In all cases, if a relocation symbol name is provided, report that.
> */
Sure. Thanks.
> Also...
>
> > @@ -125,13 +133,28 @@ dt_dis_refname(const dtrace_difo_t *dp, const struct bpf_insn *in, uint_t addr,
> > goto out;
> > }
> > - /* make sure in[-2] and in[-1] exist */
> > - if (addr < 2)
> > + /* make sure in[-1] exists and is a ld instruction */
> > + if (addr < 1 || in[-1].code != ldcode || in[-1].imm != 0)
> > goto out;
> > - /* get dst and scope */
> > + /* get dst register */
> > dst = in[-1].dst_reg;
> > + /* trampoline code uses %r9, regular code needs to load dctx */
> > + if (in[-1].src_reg != BPF_REG_9) {
> > + /* make sure in[-2] exists and is a ld instruction */
> > + if (addr < 2 || in[-2].code != ldcode || in[-2].imm != 0)
> > + goto out;
> > +
> > + /* check preceding instructions */
> > + if (in[-2].dst_reg != dst ||
> > + in[-2].src_reg != BPF_REG_FP ||
> > + in[-2].off != DT_STK_DCTX ||
> > + in[-1].code != ldcode ||
>
> We already checked that in[-1].code==ldcode.
Ah yes, thank you.
> > + in[-1].src_reg != dst)
> > + goto out;
> > + }
> > +
> > if (in[-1].off == DCTX_GVARS)
> > scope = DIFV_SCOPE_GLOBAL;
> > else if (in[-1].off == DCTX_LVARS)
> > @@ -139,17 +162,6 @@ dt_dis_refname(const dtrace_difo_t *dp, const struct bpf_insn *in, uint_t addr,
> > else if (in[-1].off != DCTX_STRTAB)
> > goto out;
> > - /* check preceding instructions */
> > - if (in[-2].code != ldcode ||
> > - in[-2].dst_reg != dst ||
> > - in[-2].src_reg != BPF_REG_FP ||
> > - in[-2].off != DT_STK_DCTX ||
> > - in[-2].imm != 0 ||
> > - in[-1].code != ldcode ||
> > - in[-1].src_reg != dst ||
> > - in[-1].imm != 0)
> > - goto out;
> > -
> > /* check the current instruction and read var_offset */
> > if (in[-1].off == DCTX_STRTAB) {
> > if (in->dst_reg == dst &&
>
> _______________________________________________
> DTrace-devel mailing list
> DTrace-devel at oss.oracle.com
> https://oss.oracle.com/mailman/listinfo/dtrace-devel
More information about the DTrace-devel
mailing list