Change stack_skip to 3
PATCH 1/1 Change stack_skip to 3
From: Eugene Loh <[email protected]>
For most probes, bpf_get_stack() strips off artificial BPF JIT stack frames. In a few cases -- like fprobes and rawtp -- we have to skip the first few frames manually. Typically, we should skip 3 frames (trampoline, clause, and the bpf_get_stack call itself). Further, our BPF functions like
get_bvar_caller()
get_bvar_stackdepth()
get_bvar_ucaller()
get_bvar_ustackdepth()
skip an additional frame to account for that call.
Change stack_skip to 3 for the dt_fbt_fprobe and dt_rawtp providers.
This raises the thorny question of testing. Until recently, the stack_skip value we had been using -- 4, not 3 -- worked since it was chosen empirically: there was a bug in the kernel that added an extra frame. But now, the kernel situation is muddled. The bug has been fixed in recent kernels, but only for x86. See
https://github.com/torvalds/linux/commit/6d08340d1e354787d6c65a8c3cdd4d41ffb8a5ed
Revert "perf/x86: Always store regs->ip in perf_callchain_kernel()"
A broader fix, for non-x86, is also expected. See the thread that starts at:
https://lore.kernel.org/all/[email protected]/
So, for now, XFAIL tests on "unfixed" (older) kernels.
Orabug: 38776929
--- libdtrace/dt_prov_fbt.c | 2 +- libdtrace/dt_prov_rawtp.c | 2 +- test/unittest/funcs/stack/tst.asgn_dvar.x | 1 + test/unittest/funcs/stack/tst.asgn_gvar.x | 1 + test/unittest/funcs/stack/tst.asgn_lvar.x | 1 + test/unittest/funcs/stack/tst.asgn_tvar.x | 1 + test/unittest/funcs/stack/tst.ref_addrs.x | 1 + test/unittest/printf/tst.stack.x | 1 + test/unittest/stack/tst.stack.x | 1 + test/utils/stack_skip_test.sh | 33 +++++++++++++++++++++++ 10 files changed, 42 insertions(+), 2 deletions(-) create mode 120000 test/unittest/funcs/stack/tst.asgn_dvar.x create mode 120000 test/unittest/funcs/stack/tst.asgn_gvar.x create mode 120000 test/unittest/funcs/stack/tst.asgn_lvar.x create mode 120000 test/unittest/funcs/stack/tst.asgn_tvar.x create mode 120000 test/unittest/funcs/stack/tst.ref_addrs.x create mode 120000 test/unittest/printf/tst.stack.x create mode 120000 test/unittest/stack/tst.stack.x create mode 100755 test/utils/stack_skip_test.sh
libdtrace/dt_prov_fbt.c +1 -1
@@ -736,7 +736,7 @@ reject:dt_provimpl_t dt_fbt_fprobe = { .name = prvname, .prog_type =BPF_PROG_TYPE_TRACING , - .stack_skip =4 , + .stack_skip =3 , .populate = &populate, .provide = &provide, .load_prog = &fprobe_prog_load,
libdtrace/dt_prov_rawtp.c +1 -1
@@ -296,7 +296,7 @@ use_alt:dt_provimpl_t dt_rawtp = { .name = prvname, .prog_type =BPF_PROG_TYPE_RAW_TRACEPOINT , - .stack_skip =4 , + .stack_skip =3 , .populate = &populate, .load_prog = &dt_bpf_prog_load, .trampoline = &trampoline,
test/unittest/funcs/stack/tst.asgn_dvar.x new file +1 -0
@@ -0,0 +1 @@ +../../../utils/stack_skip_test.sh
test/unittest/funcs/stack/tst.asgn_gvar.x new file +1 -0
@@ -0,0 +1 @@ +../../../utils/stack_skip_test.sh
test/unittest/funcs/stack/tst.asgn_lvar.x new file +1 -0
@@ -0,0 +1 @@ +../../../utils/stack_skip_test.sh
test/unittest/funcs/stack/tst.asgn_tvar.x new file +1 -0
@@ -0,0 +1 @@ +../../../utils/stack_skip_test.sh
test/unittest/funcs/stack/tst.ref_addrs.x new file +1 -0
@@ -0,0 +1 @@ +../../../utils/stack_skip_test.sh
test/unittest/printf/tst.stack.x new file +1 -0
@@ -0,0 +1 @@ +../../utils/stack_skip_test.sh
test/unittest/stack/tst.stack.x new file +1 -0
@@ -0,0 +1 @@ +../../utils/stack_skip_test.sh
test/utils/stack_skip_test.sh new file +33 -0
@@ -0,0 +1,33 @@ +#!/bin/sh + +# Older kernels had a bug in which the leaf frame was replicated. +# +# This bug was fixed on x86_64. See +# https://github.com/torvalds/linux/commit/6d08340d1e354787d6c65a8c3cdd4d41ffb8a5ed +# Revert "perf/x86: Always store regs->ip in perf_callchain_kernel()" +# +# A broader fix, for non-x86, is also expected. See the thread that starts at: +# https://lore.kernel.org/all/[email protected]/ +# +# Use this as a .x file to mark tests as "XFAIL" for older kernels with the +# superfluous frame (and therefore exposing a BPF JIT frame, appearing as an +# untranslated 0xffff address). On older kernels, it is also possible for fbt +# to be implemented with kprobes rather than fprobes; in this case, the .x +# issues no concerns and the tests are expected to PASS. + +$dtrace -c test/triggers/periodic_output -qn ' +fbt::hrtimer_nanosleep:entry +{ + stack(1); + exit(0); +} + +ERROR +{ + exit(1); +}' | grep -q 0xffff +if [ $? -eq 0 ]; then + echo stack appears to have superfluous frame + exit 1 +fi +exit 0
Signature · 1 line
2.47.3
On 21 Jan 2026, eugene loh told this:
Quoted 24 lines
From: Eugene Loh <[email protected]>
For most probes,
bpf_get_stack()strips off artificial BPF JIT stack frames. In a few cases -- like fprobes and rawtp -- we have to skip the first few frames manually. Typically, we should skip 3 frames (trampoline, clause, and thebpf_get_stackcall itself). Further, our BPF functions like
get_bvar_caller()
get_bvar_stackdepth()
get_bvar_ucaller()
get_bvar_ustackdepth()
skip an additional frame to account for that call.Change
stack_skipto 3 for thedt_fbt_fprobeanddt_rawtpproviders.This raises the thorny question of testing. Until recently, the
stack_skipvalue we had been using -- 4, not 3 -- worked since it was chosen empirically: there was a bug in the kernel that added an extra frame. But now, the kernel situation is muddled. The bug has been fixed in recent kernels, but only for x86. Seehttps://github.com/torvalds/linux/commit/6d08340d1e354787d6c65a8c3cdd4d41ffb8a5ed
Revert "perf/x86: Always storeregs->ipinperf_callchain_kernel()"
Quoted 8 lines
A broader fix, for non-x86, is also expected. See the thread that starts at:
https://lore.kernel.org/all/[email protected]/So, for now, XFAIL tests on "unfixed" (older) kernels.
Orabug: 38776929
Signed-off-by Eugene Loh <[email protected]>
(verified that it fixes these tests and doesn't break older kernels, which XFAIL with the appropriate message.)
Signature · 1 line
NULL && (void)