On Wed, Sep 01, 2021 at 04:05:59PM +0200, Vasily Gorbik wrote: > current_stack_pointer() simply returns current value of %r15. If > current_stack_pointer() caller allocates stack (which is the case in > unwind code) %r15 points to a stack frame allocated for callees, meaning > current_stack_pointer() caller (e.g. stack_trace_save) will end up in > the stacktrace. This is not expected by stack_trace_save*() callers and > causes problems. > > current_frame_address() on the other hand returns function stack frame > address, which matches %r15 upon function invocation. Using it in > get_stack_pointer() makes it more aligned with x86 implementation > (according to BACKTRACE_SELF_TEST output) and meets stack_trace_save*() > caller's expectations, notably KCSAN. > > Also make sure unwind_start is always inlined. > > Reported-by: Nathan Chancellor <nathan@xxxxxxxxxx> > Suggested-by: Marco Elver <elver@xxxxxxxxxx> > Signed-off-by: Vasily Gorbik <gor@xxxxxxxxxxxxx> Sorry for the late response and I see that this has already been applied but I took this for a spin and all of the tests pass with clang-14 in QEMU. Thank you for the quick fix so that we can get this turned on in CI :) [ 10.362073] ok 1 - test_basic [ 13.870386] ok 2 - test_concurrent_races [ 17.379643] ok 3 - test_novalue_change [ 17.393315] ok 4 - test_novalue_change_exception [ 17.409815] ok 5 - test_unknown_origin [ 20.914289] ok 6 - test_write_write_assume_atomic [ 20.982545] ok 7 - test_write_write_struct [ 21.106135] ok 8 - test_write_write_struct_part [ 24.622205] ok 9 - test_read_atomic_write_atomic [ 24.662048] ok 10 - test_read_plain_atomic_write [ 24.775291] ok 11 - test_read_plain_atomic_rmw [ 28.294457] ok 12 - test_zero_size_access [ 31.829529] ok 13 - test_data_race [ 31.867174] ok 14 - test_assert_exclusive_writer [ 31.929184] ok 15 - test_assert_exclusive_access [ 35.446281] ok 16 - test_assert_exclusive_access_writer [ 35.540228] ok 17 - test_assert_exclusive_bits_change [ 39.052271] ok 18 - test_assert_exclusive_bits_nochange [ 39.097020] ok 19 - test_assert_exclusive_writer_scoped [ 39.152914] ok 20 - test_assert_exclusive_access_scoped [ 42.675158] ok 21 - test_jiffies_noreport [ 46.192453] ok 22 - test_seqlock_noreport [ 49.712712] ok 23 - test_atomic_builtins [ 49.746428] ok 24 - test_1bit_value_change [ 49.753316] ok 1 - kcsan Tested-by: Nathan Chancellor <nathan@xxxxxxxxxx> > --- > arch/s390/include/asm/stacktrace.h | 20 ++++++++++---------- > arch/s390/include/asm/unwind.h | 8 ++++---- > 2 files changed, 14 insertions(+), 14 deletions(-) > > diff --git a/arch/s390/include/asm/stacktrace.h b/arch/s390/include/asm/stacktrace.h > index 3d8a4b94c620..22c41d7fd95c 100644 > --- a/arch/s390/include/asm/stacktrace.h > +++ b/arch/s390/include/asm/stacktrace.h > @@ -34,16 +34,6 @@ static inline bool on_stack(struct stack_info *info, > return addr >= info->begin && addr + len <= info->end; > } > > -static __always_inline unsigned long get_stack_pointer(struct task_struct *task, > - struct pt_regs *regs) > -{ > - if (regs) > - return (unsigned long) kernel_stack_pointer(regs); > - if (task == current) > - return current_stack_pointer(); > - return (unsigned long) task->thread.ksp; > -} > - > /* > * Stack layout of a C stack frame. > */ > @@ -74,6 +64,16 @@ struct stack_frame { > ((unsigned long)__builtin_frame_address(0) - \ > offsetof(struct stack_frame, back_chain)) > > +static __always_inline unsigned long get_stack_pointer(struct task_struct *task, > + struct pt_regs *regs) > +{ > + if (regs) > + return (unsigned long) kernel_stack_pointer(regs); > + if (task == current) > + return current_frame_address(); > + return (unsigned long) task->thread.ksp; > +} > + > /* > * To keep this simple mark register 2-6 as being changed (volatile) > * by the called function, even though register 6 is saved/nonvolatile. > diff --git a/arch/s390/include/asm/unwind.h b/arch/s390/include/asm/unwind.h > index de9006b0cfeb..5ebf534ef753 100644 > --- a/arch/s390/include/asm/unwind.h > +++ b/arch/s390/include/asm/unwind.h > @@ -55,10 +55,10 @@ static inline bool unwind_error(struct unwind_state *state) > return state->error; > } > > -static inline void unwind_start(struct unwind_state *state, > - struct task_struct *task, > - struct pt_regs *regs, > - unsigned long first_frame) > +static __always_inline void unwind_start(struct unwind_state *state, > + struct task_struct *task, > + struct pt_regs *regs, > + unsigned long first_frame) > { > task = task ?: current; > first_frame = first_frame ?: get_stack_pointer(task, regs); > -- > 2.25.4