This causes ip == save_stack_trace and also below for
save_stack_trace_tsk. Does this mean save_stack_trace() is included in
the trace? Looking at kernel/stacktrace.c, I think the library wants
to exclude itself from the trace, as it does '.skip = skipnr + 1' (and
'.skip = skipnr + (current == tsk)' for the _tsk variant).
If the arch-helper here is included, should this use _RET_IP_ instead?
Don't really know, I was inspired by arm64 which has:
void arch_stack_walk(stack_trace_consume_fn consume_entry, void *cookie,
struct task_struct *task, struct pt_regs *regs)
{
struct stackframe frame;
if (regs)
start_backtrace(&frame, regs->regs[29], regs->pc);
else if (task == current)
start_backtrace(&frame,
(unsigned long)__builtin_frame_address(0),
(unsigned long)arch_stack_walk);
else
start_backtrace(&frame, thread_saved_fp(task),
thread_saved_pc(task));
walk_stackframe(task, &frame, consume_entry, cookie);
}
But looking at x86 you may be right, so what should be done really ?
x86:
[ 2.843292] calling stack_trace_save:
[ 2.843705] test_func+0x6c/0x118
[ 2.844184] do_one_initcall+0x58/0x270
[ 2.844618] kernel_init_freeable+0x1da/0x23a
[ 2.845110] kernel_init+0xc/0x166
[ 2.845494] ret_from_fork+0x22/0x30
[ 2.867525] calling stack_trace_save_tsk:
[ 2.868017] test_func+0xa9/0x118
[ 2.868530] do_one_initcall+0x58/0x270
[ 2.869003] kernel_init_freeable+0x1da/0x23a
[ 2.869535] kernel_init+0xc/0x166
[ 2.869957] ret_from_fork+0x22/0x30
arm64:
[ 3.786911] calling stack_trace_save:
[ 3.787147] stack_trace_save+0x50/0x78
[ 3.787443] test_func+0x84/0x13c
[ 3.787738] do_one_initcall+0x5c/0x310
[ 3.788099] kernel_init_freeable+0x214/0x294
[ 3.788363] kernel_init+0x18/0x164
[ 3.788585] ret_from_fork+0x10/0x30
[ 3.803615] calling stack_trace_save_tsk:
[ 3.804266] stack_trace_save_tsk+0x9c/0x100
[ 3.804541] test_func+0xc4/0x13c
[ 3.804803] do_one_initcall+0x5c/0x310
[ 3.805031] kernel_init_freeable+0x214/0x294
[ 3.805284] kernel_init+0x18/0x164
[ 3.805505] ret_from_fork+0x10/0x30
+Cc arm64 folks.
So I think the arm64 version also has a bug, because I think a user of
<linux/stacktrace.h> really doesn't care about the library function
itself. And from reading kernel/stacktrace.c I think it wants to exclude
itself entirely.
It's a shame that <linux/stacktrace.h> isn't better documented, but I'm
pretty sure that including the library functions in the trace is not
useful.
For the ppc version, let's do what x86 does and start with the caller.
Thanks,
-- Marco
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
This causes ip == save_stack_trace and also below for
save_stack_trace_tsk. Does this mean save_stack_trace() is included in
the trace? Looking at kernel/stacktrace.c, I think the library wants
to exclude itself from the trace, as it does '.skip = skipnr + 1' (and
'.skip = skipnr + (current == tsk)' for the _tsk variant).
If the arch-helper here is included, should this use _RET_IP_ instead?
Don't really know, I was inspired by arm64 which has:
void arch_stack_walk(stack_trace_consume_fn consume_entry, void *cookie,
struct task_struct *task, struct pt_regs *regs)
{
struct stackframe frame;
if (regs)
start_backtrace(&frame, regs->regs[29], regs->pc);
else if (task == current)
start_backtrace(&frame,
(unsigned long)__builtin_frame_address(0),
(unsigned long)arch_stack_walk);
else
start_backtrace(&frame, thread_saved_fp(task),
thread_saved_pc(task));
walk_stackframe(task, &frame, consume_entry, cookie);
}
But looking at x86 you may be right, so what should be done really ?
x86:
[ 2.843292] calling stack_trace_save:
[ 2.843705] test_func+0x6c/0x118
[ 2.844184] do_one_initcall+0x58/0x270
[ 2.844618] kernel_init_freeable+0x1da/0x23a
[ 2.845110] kernel_init+0xc/0x166
[ 2.845494] ret_from_fork+0x22/0x30
[ 2.867525] calling stack_trace_save_tsk:
[ 2.868017] test_func+0xa9/0x118
[ 2.868530] do_one_initcall+0x58/0x270
[ 2.869003] kernel_init_freeable+0x1da/0x23a
[ 2.869535] kernel_init+0xc/0x166
[ 2.869957] ret_from_fork+0x22/0x30
arm64:
[ 3.786911] calling stack_trace_save:
[ 3.787147] stack_trace_save+0x50/0x78
[ 3.787443] test_func+0x84/0x13c
[ 3.787738] do_one_initcall+0x5c/0x310
[ 3.788099] kernel_init_freeable+0x214/0x294
[ 3.788363] kernel_init+0x18/0x164
[ 3.788585] ret_from_fork+0x10/0x30
[ 3.803615] calling stack_trace_save_tsk:
[ 3.804266] stack_trace_save_tsk+0x9c/0x100
[ 3.804541] test_func+0xc4/0x13c
[ 3.804803] do_one_initcall+0x5c/0x310
[ 3.805031] kernel_init_freeable+0x214/0x294
[ 3.805284] kernel_init+0x18/0x164
[ 3.805505] ret_from_fork+0x10/0x30
+Cc arm64 folks.
So I think the arm64 version also has a bug, because I think a user of
<linux/stacktrace.h> really doesn't care about the library function
itself. And from reading kernel/stacktrace.c I think it wants to exclude
itself entirely.
It's a shame that <linux/stacktrace.h> isn't better documented, but I'm
pretty sure that including the library functions in the trace is not
useful.
I agree this behaviour isn't desireable, and that the lack of
documentation is unfortunate.
It looks like GCC is happy to give us the function-entry-time FP if we use
__builtin_frame_address(1), and assuming clang is similarly happy we can do:
| diff --git a/arch/arm64/kernel/stacktrace.c b/arch/arm64/kernel/stacktrace.c
| index ad20981dfda4..5dfbf915eb7f 100644
| --- a/arch/arm64/kernel/stacktrace.c
| +++ b/arch/arm64/kernel/stacktrace.c
| @@ -203,8 +203,8 @@ void arch_stack_walk(stack_trace_consume_fn consume_entry, void *cookie,
| start_backtrace(&frame, regs->regs[29], regs->pc);
| else if (task == current)
| start_backtrace(&frame,
| - (unsigned long)__builtin_frame_address(0),
| - (unsigned long)arch_stack_walk);
| + (unsigned long)__builtin_frame_address(1),
| + (unsigned long)__builtin_return_address(0));
| else
| start_backtrace(&frame, thread_saved_fp(task),
| thread_saved_pc(task));
... such that arch_stack_walk() will try to avoid including itself in a
trace, and so the existing skipping should (w/ caveats below) skip
stack_trace_save() or stack_trace_save_tsk().
If that works for you, I can spin that as a patch, though we'll need to
check that doesn't introduce a new fencepost error elsewhere.
The bigger problem here is that skipping is dodgy to begin with, and
this is still liable to break in some cases. One big concern is that
(especially with LTO) we cannot guarantee the compiler will not inline
or outline functions, causing the skipp value to be too large or too
small. That's liable to happen to callers, and in theory (though
unlikely in practice), portions of arch_stack_walk() or
stack_trace_save() could get outlined too.
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
Thanks,
Mark.
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
This causes ip == save_stack_trace and also below for
save_stack_trace_tsk. Does this mean save_stack_trace() is included in
the trace? Looking at kernel/stacktrace.c, I think the library wants
to exclude itself from the trace, as it does '.skip = skipnr + 1' (and
'.skip = skipnr + (current == tsk)' for the _tsk variant).
If the arch-helper here is included, should this use _RET_IP_ instead?
Don't really know, I was inspired by arm64 which has:
void arch_stack_walk(stack_trace_consume_fn consume_entry, void *cookie,
struct task_struct *task, struct pt_regs *regs)
{
struct stackframe frame;
if (regs)
start_backtrace(&frame, regs->regs[29], regs->pc);
else if (task == current)
start_backtrace(&frame,
(unsigned long)__builtin_frame_address(0),
(unsigned long)arch_stack_walk);
else
start_backtrace(&frame, thread_saved_fp(task),
thread_saved_pc(task));
walk_stackframe(task, &frame, consume_entry, cookie);
}
But looking at x86 you may be right, so what should be done really ?
x86:
[ 2.843292] calling stack_trace_save:
[ 2.843705] test_func+0x6c/0x118
[ 2.844184] do_one_initcall+0x58/0x270
[ 2.844618] kernel_init_freeable+0x1da/0x23a
[ 2.845110] kernel_init+0xc/0x166
[ 2.845494] ret_from_fork+0x22/0x30
[ 2.867525] calling stack_trace_save_tsk:
[ 2.868017] test_func+0xa9/0x118
[ 2.868530] do_one_initcall+0x58/0x270
[ 2.869003] kernel_init_freeable+0x1da/0x23a
[ 2.869535] kernel_init+0xc/0x166
[ 2.869957] ret_from_fork+0x22/0x30
arm64:
[ 3.786911] calling stack_trace_save:
[ 3.787147] stack_trace_save+0x50/0x78
[ 3.787443] test_func+0x84/0x13c
[ 3.787738] do_one_initcall+0x5c/0x310
[ 3.788099] kernel_init_freeable+0x214/0x294
[ 3.788363] kernel_init+0x18/0x164
[ 3.788585] ret_from_fork+0x10/0x30
[ 3.803615] calling stack_trace_save_tsk:
[ 3.804266] stack_trace_save_tsk+0x9c/0x100
[ 3.804541] test_func+0xc4/0x13c
[ 3.804803] do_one_initcall+0x5c/0x310
[ 3.805031] kernel_init_freeable+0x214/0x294
[ 3.805284] kernel_init+0x18/0x164
[ 3.805505] ret_from_fork+0x10/0x30
+Cc arm64 folks.
So I think the arm64 version also has a bug, because I think a user of
<linux/stacktrace.h> really doesn't care about the library function
itself. And from reading kernel/stacktrace.c I think it wants to exclude
itself entirely.
It's a shame that <linux/stacktrace.h> isn't better documented, but I'm
pretty sure that including the library functions in the trace is not
useful.
I agree this behaviour isn't desireable, and that the lack of
documentation is unfortunate.
It looks like GCC is happy to give us the function-entry-time FP if we use
__builtin_frame_address(1), and assuming clang is similarly happy we can do:
| diff --git a/arch/arm64/kernel/stacktrace.c b/arch/arm64/kernel/stacktrace.c
| index ad20981dfda4..5dfbf915eb7f 100644
| --- a/arch/arm64/kernel/stacktrace.c
| +++ b/arch/arm64/kernel/stacktrace.c
| @@ -203,8 +203,8 @@ void arch_stack_walk(stack_trace_consume_fn consume_entry, void *cookie,
| start_backtrace(&frame, regs->regs[29], regs->pc);
| else if (task == current)
| start_backtrace(&frame,
| - (unsigned long)__builtin_frame_address(0),
| - (unsigned long)arch_stack_walk);
| + (unsigned long)__builtin_frame_address(1),
| + (unsigned long)__builtin_return_address(0));
| else
| start_backtrace(&frame, thread_saved_fp(task),
| thread_saved_pc(task));
... such that arch_stack_walk() will try to avoid including itself in a
trace, and so the existing skipping should (w/ caveats below) skip
stack_trace_save() or stack_trace_save_tsk().
Thank you! Yes, that works.
If that works for you, I can spin that as a patch, though we'll need to
check that doesn't introduce a new fencepost error elsewhere.
The bigger problem here is that skipping is dodgy to begin with, and
this is still liable to break in some cases. One big concern is that
(especially with LTO) we cannot guarantee the compiler will not inline
or outline functions, causing the skipp value to be too large or too
small. That's liable to happen to callers, and in theory (though
unlikely in practice), portions of arch_stack_walk() or
stack_trace_save() could get outlined too.
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
Will LTO and friends respect 'noinline'? One thing I also noticed is
that tail calls would also cause the stack trace to appear somewhat
incomplete (for some of my tests I've disabled tail call
optimizations). Is there a way to also mark a function
non-tail-callable? But I'm also not sure if with all that we'd be
guaranteed the code we want, even though in practice it might.
Thanks,
-- Marco
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
From: Mark Rutland <mark.rutland@arm.com> Date: 2021-03-04 17:01:23
On Thu, Mar 04, 2021 at 04:30:34PM +0100, Marco Elver wrote:
On Thu, 4 Mar 2021 at 15:57, Mark Rutland [off-list ref] wrote:
quoted
[adding Mark Brown]
The bigger problem here is that skipping is dodgy to begin with, and
this is still liable to break in some cases. One big concern is that
(especially with LTO) we cannot guarantee the compiler will not inline
or outline functions, causing the skipp value to be too large or too
small. That's liable to happen to callers, and in theory (though
unlikely in practice), portions of arch_stack_walk() or
stack_trace_save() could get outlined too.
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
Will LTO and friends respect 'noinline'?
I hope so (and suspect we'd have more problems otherwise), but I don't
know whether they actually so.
I suspect even with 'noinline' the compiler is permitted to outline
portions of a function if it wanted to (and IIUC it could still make
specialized copies in the absence of 'noclone').
One thing I also noticed is that tail calls would also cause the stack
trace to appear somewhat incomplete (for some of my tests I've
disabled tail call optimizations).
I assume you mean for a chain A->B->C where B tail-calls C, you get a
trace A->C? ... or is A going missing too?
Is there a way to also mark a function non-tail-callable?
I think this can be bodged using __attribute__((optimize("$OPTIONS")))
on a caller to inhibit TCO (though IIRC GCC doesn't reliably support
function-local optimization options), but I don't expect there's any way
to mark a callee as not being tail-callable.
Accoding to the GCC documentation, GCC won't TCO noreturn functions, but
obviously that's not something we can use generally.
https://gcc.gnu.org/onlinedocs/gcc/Common-Function-Attributes.html#Common-Function-Attributes
But I'm also not sure if with all that we'd be guaranteed the code we
want, even though in practice it might.
True! I'd just like to be on the least dodgy ground we can be.
Thanks,
Mark.
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
From: Marco Elver <elver@google.com> Date: 2021-03-04 17:27:16
On Thu, Mar 04, 2021 at 04:59PM +0000, Mark Rutland wrote:
On Thu, Mar 04, 2021 at 04:30:34PM +0100, Marco Elver wrote:
quoted
On Thu, 4 Mar 2021 at 15:57, Mark Rutland [off-list ref] wrote:
quoted
[adding Mark Brown]
The bigger problem here is that skipping is dodgy to begin with, and
this is still liable to break in some cases. One big concern is that
(especially with LTO) we cannot guarantee the compiler will not inline
or outline functions, causing the skipp value to be too large or too
small. That's liable to happen to callers, and in theory (though
unlikely in practice), portions of arch_stack_walk() or
stack_trace_save() could get outlined too.
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
Will LTO and friends respect 'noinline'?
I hope so (and suspect we'd have more problems otherwise), but I don't
know whether they actually so.
I suspect even with 'noinline' the compiler is permitted to outline
portions of a function if it wanted to (and IIUC it could still make
specialized copies in the absence of 'noclone').
quoted
One thing I also noticed is that tail calls would also cause the stack
trace to appear somewhat incomplete (for some of my tests I've
disabled tail call optimizations).
I assume you mean for a chain A->B->C where B tail-calls C, you get a
trace A->C? ... or is A going missing too?
Correct, it's just the A->C outcome.
quoted
Is there a way to also mark a function non-tail-callable?
I think this can be bodged using __attribute__((optimize("$OPTIONS")))
on a caller to inhibit TCO (though IIRC GCC doesn't reliably support
function-local optimization options), but I don't expect there's any way
to mark a callee as not being tail-callable.
I don't think this is reliable. It'd be
__attribute__((optimize("-fno-optimize-sibling-calls"))), but doesn't
work if applied to the function we do not want to tail-call-optimize,
but would have to be applied to the function that does the tail-calling.
So it's a bit backwards, even if it worked.
Perhaps we can ask the toolchain folks to help add such an attribute. Or
maybe the feature already exists somewhere, but hidden.
+Cc linux-toolchains@vger.kernel.org
quoted
But I'm also not sure if with all that we'd be guaranteed the code we
want, even though in practice it might.
True! I'd just like to be on the least dodgy ground we can be.
It's been dodgy for a while, and I'd welcome any low-cost fixes to make
it less dodgy in the short-term at least. :-)
Thanks,
-- Marco
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
From: Nick Desaulniers <ndesaulniers@google.com> Date: 2021-03-04 17:56:43
On Thu, Mar 4, 2021 at 9:42 AM Marco Elver [off-list ref] wrote:
On Thu, Mar 04, 2021 at 04:59PM +0000, Mark Rutland wrote:
quoted
On Thu, Mar 04, 2021 at 04:30:34PM +0100, Marco Elver wrote:
quoted
On Thu, 4 Mar 2021 at 15:57, Mark Rutland [off-list ref] wrote:
quoted
[adding Mark Brown]
The bigger problem here is that skipping is dodgy to begin with, and
this is still liable to break in some cases. One big concern is that
(especially with LTO) we cannot guarantee the compiler will not inline
or outline functions, causing the skipp value to be too large or too
small. That's liable to happen to callers, and in theory (though
unlikely in practice), portions of arch_stack_walk() or
stack_trace_save() could get outlined too.
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
Will LTO and friends respect 'noinline'?
I hope so (and suspect we'd have more problems otherwise), but I don't
know whether they actually so.
I suspect even with 'noinline' the compiler is permitted to outline
portions of a function if it wanted to (and IIUC it could still make
specialized copies in the absence of 'noclone').
quoted
One thing I also noticed is that tail calls would also cause the stack
trace to appear somewhat incomplete (for some of my tests I've
disabled tail call optimizations).
I assume you mean for a chain A->B->C where B tail-calls C, you get a
trace A->C? ... or is A going missing too?
Correct, it's just the A->C outcome.
quoted
quoted
Is there a way to also mark a function non-tail-callable?
I think this can be bodged using __attribute__((optimize("$OPTIONS")))
on a caller to inhibit TCO (though IIRC GCC doesn't reliably support
function-local optimization options), but I don't expect there's any way
to mark a callee as not being tail-callable.
I don't think this is reliable. It'd be
__attribute__((optimize("-fno-optimize-sibling-calls"))), but doesn't
work if applied to the function we do not want to tail-call-optimize,
but would have to be applied to the function that does the tail-calling.
So it's a bit backwards, even if it worked.
include/linux/compiler.h:246:
prevent_tail_call_optimization
commit a9a3ed1eff36 ("x86: Fix early boot crash on gcc-10, third try")
Perhaps we can ask the toolchain folks to help add such an attribute. Or
maybe the feature already exists somewhere, but hidden.
+Cc linux-toolchains@vger.kernel.org
quoted
quoted
But I'm also not sure if with all that we'd be guaranteed the code we
want, even though in practice it might.
True! I'd just like to be on the least dodgy ground we can be.
It's been dodgy for a while, and I'd welcome any low-cost fixes to make
it less dodgy in the short-term at least. :-)
Thanks,
-- Marco
On Thu, Mar 04, 2021 at 09:54:44AM -0800, Nick Desaulniers wrote:
On Thu, Mar 4, 2021 at 9:42 AM Marco Elver [off-list ref] wrote:
include/linux/compiler.h:246:
prevent_tail_call_optimization
commit a9a3ed1eff36 ("x86: Fix early boot crash on gcc-10, third try")
That is much heavier than needed (an mb()). You can just put an empty
inline asm after a call before a return, and that call cannot be
optimised to a sibling call: (the end of a function is an implicit
return:)
Instead of:
void g(void);
void f(int x)
if (x)
g();
}
Do:
void g(void);
void f(int x)
if (x)
g();
asm("");
}
This costs no extra instructions, and certainly not something as heavy
as an mb()! It works without the "if" as well, of course, but with it
it is a more interesting example of a tail call.
Segher
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Le 04/03/2021 à 20:24, Segher Boessenkool a écrit :
On Thu, Mar 04, 2021 at 09:54:44AM -0800, Nick Desaulniers wrote:
quoted
On Thu, Mar 4, 2021 at 9:42 AM Marco Elver [off-list ref] wrote:
include/linux/compiler.h:246:
prevent_tail_call_optimization
commit a9a3ed1eff36 ("x86: Fix early boot crash on gcc-10, third try")
That is much heavier than needed (an mb()). You can just put an empty
inline asm after a call before a return, and that call cannot be
optimised to a sibling call: (the end of a function is an implicit
return:)
Instead of:
void g(void);
void f(int x)
if (x)
g();
}
Do:
void g(void);
void f(int x)
if (x)
g();
asm("");
}
This costs no extra instructions, and certainly not something as heavy
as an mb()! It works without the "if" as well, of course, but with it
it is a more interesting example of a tail call.
In the commit mentionned at the top, it is said:
The next attempt to prevent compilers from tail-call optimizing
the last function call cpu_startup_entry(), ... , was to add an empty asm("").
This current solution was short and sweet, and reportedly, is supported
by both compilers but we didn't get very far this time: future (LTO?)
optimization passes could potentially eliminate this, which leads us
to the third attempt: having an actual memory barrier there which the
compiler cannot ignore or move around etc.
Christophe
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
That is much heavier than needed (an mb()). You can just put an empty
inline asm after a call before a return, and that call cannot be
optimised to a sibling call: (the end of a function is an implicit
return:)
In the commit mentionned at the top, it is said:
The next attempt to prevent compilers from tail-call optimizing
the last function call cpu_startup_entry(), ... , was to add an empty
asm("").
This current solution was short and sweet, and reportedly, is supported
by both compilers but we didn't get very far this time: future (LTO?)
optimization passes could potentially eliminate this,
This is simply not true. A volatile inline asm (like this is, all
asm without outputs are) is always run on the reel machine exactly like
on the abstract machine. LTO can not eliminate it, not more than any
other optimisation can. The compiler makes no assumption about the
constents of the template of an asm, empty or not.
If you are really scared the compiler violates the rules of GCC inline
asm and thinks it knows what "" means, you can write
asm(";#");
(that is a comment on all supported archs).
which leads us
to the third attempt: having an actual memory barrier there which the
compiler cannot ignore or move around etc.
Why would it not be allowed to delete this, and delete some other asm?
And the compiler *can* move around asm like this. But the point is,
it has to stay in order with other side effects, so there cannot be a
sibling call here, the call has to remain: any call contains a sequence
point, so side effects cannot be reordered over it, so the call (being
before the asm) cannot be transformed to a tail call.
Segher
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
From: Mark Rutland <mark.rutland@arm.com> Date: 2021-03-04 18:03:24
On Thu, Mar 04, 2021 at 06:25:33PM +0100, Marco Elver wrote:
On Thu, Mar 04, 2021 at 04:59PM +0000, Mark Rutland wrote:
quoted
On Thu, Mar 04, 2021 at 04:30:34PM +0100, Marco Elver wrote:
quoted
On Thu, 4 Mar 2021 at 15:57, Mark Rutland [off-list ref] wrote:
quoted
[adding Mark Brown]
The bigger problem here is that skipping is dodgy to begin with, and
this is still liable to break in some cases. One big concern is that
(especially with LTO) we cannot guarantee the compiler will not inline
or outline functions, causing the skipp value to be too large or too
small. That's liable to happen to callers, and in theory (though
unlikely in practice), portions of arch_stack_walk() or
stack_trace_save() could get outlined too.
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
Will LTO and friends respect 'noinline'?
I hope so (and suspect we'd have more problems otherwise), but I don't
know whether they actually so.
I suspect even with 'noinline' the compiler is permitted to outline
portions of a function if it wanted to (and IIUC it could still make
specialized copies in the absence of 'noclone').
quoted
One thing I also noticed is that tail calls would also cause the stack
trace to appear somewhat incomplete (for some of my tests I've
disabled tail call optimizations).
I assume you mean for a chain A->B->C where B tail-calls C, you get a
trace A->C? ... or is A going missing too?
Correct, it's just the A->C outcome.
I'd assumed that those cases were benign, e.g. for livepatching what
matters is what can be returned to, so B disappearing from the trace
isn't a problem there.
Is the concern debugability, or is there a functional issue you have in
mind?
quoted
quoted
Is there a way to also mark a function non-tail-callable?
I think this can be bodged using __attribute__((optimize("$OPTIONS")))
on a caller to inhibit TCO (though IIRC GCC doesn't reliably support
function-local optimization options), but I don't expect there's any way
to mark a callee as not being tail-callable.
I don't think this is reliable. It'd be
__attribute__((optimize("-fno-optimize-sibling-calls"))), but doesn't
work if applied to the function we do not want to tail-call-optimize,
but would have to be applied to the function that does the tail-calling.
Yup; that's what I meant then I said you could do that on the caller but
not the callee.
I don't follow why you'd want to put this on the callee, though, so I
think I'm missing something. Considering a set of functions in different
compilation units:
A->B->C->D->E->F->G->H->I->J->K
... if K were marked in this way, and J was compiled with visibility of
this, J would stick around, but J's callers might not, and so the a
trace might see:
A->J->K
... do you just care about the final caller, i.e. you just need
certainty that J will be in the trace?
If so, we can somewhat bodge that by having K have an __always_inline
wrapper which has a barrier() or similar after the real call to K, so
the call couldn't be TCO'd.
Otherwise I'd expect we'd probably need to disable TCO generally.
Perhaps we can ask the toolchain folks to help add such an attribute. Or
maybe the feature already exists somewhere, but hidden.
+Cc linux-toolchains@vger.kernel.org
quoted
quoted
But I'm also not sure if with all that we'd be guaranteed the code we
want, even though in practice it might.
True! I'd just like to be on the least dodgy ground we can be.
It's been dodgy for a while, and I'd welcome any low-cost fixes to make
it less dodgy in the short-term at least. :-)
From: Marco Elver <elver@google.com> Date: 2021-03-04 18:24:59
On Thu, 4 Mar 2021 at 19:02, Mark Rutland [off-list ref] wrote:
On Thu, Mar 04, 2021 at 06:25:33PM +0100, Marco Elver wrote:
quoted
On Thu, Mar 04, 2021 at 04:59PM +0000, Mark Rutland wrote:
quoted
On Thu, Mar 04, 2021 at 04:30:34PM +0100, Marco Elver wrote:
quoted
On Thu, 4 Mar 2021 at 15:57, Mark Rutland [off-list ref] wrote:
quoted
[adding Mark Brown]
The bigger problem here is that skipping is dodgy to begin with, and
this is still liable to break in some cases. One big concern is that
(especially with LTO) we cannot guarantee the compiler will not inline
or outline functions, causing the skipp value to be too large or too
small. That's liable to happen to callers, and in theory (though
unlikely in practice), portions of arch_stack_walk() or
stack_trace_save() could get outlined too.
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
Will LTO and friends respect 'noinline'?
I hope so (and suspect we'd have more problems otherwise), but I don't
know whether they actually so.
I suspect even with 'noinline' the compiler is permitted to outline
portions of a function if it wanted to (and IIUC it could still make
specialized copies in the absence of 'noclone').
quoted
One thing I also noticed is that tail calls would also cause the stack
trace to appear somewhat incomplete (for some of my tests I've
disabled tail call optimizations).
I assume you mean for a chain A->B->C where B tail-calls C, you get a
trace A->C? ... or is A going missing too?
Correct, it's just the A->C outcome.
I'd assumed that those cases were benign, e.g. for livepatching what
matters is what can be returned to, so B disappearing from the trace
isn't a problem there.
Is the concern debugability, or is there a functional issue you have in
mind?
For me, it's just been debuggability, and reliable test cases.
quoted
quoted
quoted
Is there a way to also mark a function non-tail-callable?
I think this can be bodged using __attribute__((optimize("$OPTIONS")))
on a caller to inhibit TCO (though IIRC GCC doesn't reliably support
function-local optimization options), but I don't expect there's any way
to mark a callee as not being tail-callable.
I don't think this is reliable. It'd be
__attribute__((optimize("-fno-optimize-sibling-calls"))), but doesn't
work if applied to the function we do not want to tail-call-optimize,
but would have to be applied to the function that does the tail-calling.
Yup; that's what I meant then I said you could do that on the caller but
not the callee.
I don't follow why you'd want to put this on the callee, though, so I
think I'm missing something. Considering a set of functions in different
compilation units:
A->B->C->D->E->F->G->H->I->J->K
I was having this problem with KCSAN, where the compiler would
tail-call-optimize __tsan_X instrumentation. This would mean that
KCSAN runtime functions ended up in the trace, but the function where
the access happened would not. However, I don't care about the runtime
functions, and instead want to see the function where the access
happened. In that case, I'd like to just mark __tsan_X and any other
kcsan instrumentation functions as do-not-tail-call-optimize, which
would solve the problem.
The solution today is that when you compile a kernel with KCSAN, every
instrumented TU is compiled with -fno-optimize-sibling-calls. The
better solution would be to just mark KCSAN runtime functions somehow,
but permit tail calling other things. Although, I probably still want
to see the full trace, and would decide that having
-fno-optimize-sibling-calls is a small price to pay in a
debug-only-kernel to get complete traces.
... if K were marked in this way, and J was compiled with visibility of
this, J would stick around, but J's callers might not, and so the a
trace might see:
A->J->K
... do you just care about the final caller, i.e. you just need
certainty that J will be in the trace?
Yes. But maybe it's a special problem that only sanitizers have.
If so, we can somewhat bodge that by having K have an __always_inline
wrapper which has a barrier() or similar after the real call to K, so
the call couldn't be TCO'd.
Otherwise I'd expect we'd probably need to disable TCO generally.
From: Mark Rutland <mark.rutland@arm.com> Date: 2021-03-04 18:53:39
On Thu, Mar 04, 2021 at 07:22:53PM +0100, Marco Elver wrote:
On Thu, 4 Mar 2021 at 19:02, Mark Rutland [off-list ref] wrote:
quoted
On Thu, Mar 04, 2021 at 06:25:33PM +0100, Marco Elver wrote:
quoted
On Thu, Mar 04, 2021 at 04:59PM +0000, Mark Rutland wrote:
quoted
On Thu, Mar 04, 2021 at 04:30:34PM +0100, Marco Elver wrote:
quoted
On Thu, 4 Mar 2021 at 15:57, Mark Rutland [off-list ref] wrote:
quoted
[adding Mark Brown]
The bigger problem here is that skipping is dodgy to begin with, and
this is still liable to break in some cases. One big concern is that
(especially with LTO) we cannot guarantee the compiler will not inline
or outline functions, causing the skipp value to be too large or too
small. That's liable to happen to callers, and in theory (though
unlikely in practice), portions of arch_stack_walk() or
stack_trace_save() could get outlined too.
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
Will LTO and friends respect 'noinline'?
I hope so (and suspect we'd have more problems otherwise), but I don't
know whether they actually so.
I suspect even with 'noinline' the compiler is permitted to outline
portions of a function if it wanted to (and IIUC it could still make
specialized copies in the absence of 'noclone').
quoted
One thing I also noticed is that tail calls would also cause the stack
trace to appear somewhat incomplete (for some of my tests I've
disabled tail call optimizations).
I assume you mean for a chain A->B->C where B tail-calls C, you get a
trace A->C? ... or is A going missing too?
Correct, it's just the A->C outcome.
I'd assumed that those cases were benign, e.g. for livepatching what
matters is what can be returned to, so B disappearing from the trace
isn't a problem there.
Is the concern debugability, or is there a functional issue you have in
mind?
For me, it's just been debuggability, and reliable test cases.
quoted
quoted
quoted
quoted
Is there a way to also mark a function non-tail-callable?
I think this can be bodged using __attribute__((optimize("$OPTIONS")))
on a caller to inhibit TCO (though IIRC GCC doesn't reliably support
function-local optimization options), but I don't expect there's any way
to mark a callee as not being tail-callable.
I don't think this is reliable. It'd be
__attribute__((optimize("-fno-optimize-sibling-calls"))), but doesn't
work if applied to the function we do not want to tail-call-optimize,
but would have to be applied to the function that does the tail-calling.
Yup; that's what I meant then I said you could do that on the caller but
not the callee.
I don't follow why you'd want to put this on the callee, though, so I
think I'm missing something. Considering a set of functions in different
compilation units:
A->B->C->D->E->F->G->H->I->J->K
I was having this problem with KCSAN, where the compiler would
tail-call-optimize __tsan_X instrumentation.
Those are compiler-generated calls, right? When those are generated the
compilation unit (and whatever it has included) might not have provided
a prototype anyway, and the compiler has special knowledge of the
functions, so it feels like the compiler would need to inhibit TCO here
for this to be robust. For their intended usage subjecting them to TCO
doesn't seem to make sense AFAICT.
I suspect that compilers have some way of handling that; otherwise I'd
expect to have heard stories of mcount/fentry calls getting TCO'd and
causing problems. So maybe there's an easy fix there?
This would mean that KCSAN runtime functions ended up in the trace,
but the function where the access happened would not. However, I don't
care about the runtime functions, and instead want to see the function
where the access happened. In that case, I'd like to just mark
__tsan_X and any other kcsan instrumentation functions as
do-not-tail-call-optimize, which would solve the problem.
I understand why we don't want to TCO these calls, but given the calls
are implicitly generated, I strongly suspect it's better to fix the
implicit call generation to not be TCO'd to begin with.
The solution today is that when you compile a kernel with KCSAN, every
instrumented TU is compiled with -fno-optimize-sibling-calls. The
better solution would be to just mark KCSAN runtime functions somehow,
but permit tail calling other things. Although, I probably still want
to see the full trace, and would decide that having
-fno-optimize-sibling-calls is a small price to pay in a
debug-only-kernel to get complete traces.
quoted
... if K were marked in this way, and J was compiled with visibility of
this, J would stick around, but J's callers might not, and so the a
trace might see:
A->J->K
... do you just care about the final caller, i.e. you just need
certainty that J will be in the trace?
Yes. But maybe it's a special problem that only sanitizers have.
I reckon for basically any instrumentation we don't want calls to be
TCO'd, though I'm not immediately sure of cases beyond sanitizers and
mcount/fentry.
Thanks,
Mark.
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
From: Marco Elver <elver@google.com> Date: 2021-03-04 19:03:30
On Thu, 4 Mar 2021 at 19:51, Mark Rutland [off-list ref] wrote:
On Thu, Mar 04, 2021 at 07:22:53PM +0100, Marco Elver wrote:
quoted
On Thu, 4 Mar 2021 at 19:02, Mark Rutland [off-list ref] wrote:
quoted
On Thu, Mar 04, 2021 at 06:25:33PM +0100, Marco Elver wrote:
quoted
On Thu, Mar 04, 2021 at 04:59PM +0000, Mark Rutland wrote:
quoted
On Thu, Mar 04, 2021 at 04:30:34PM +0100, Marco Elver wrote:
quoted
On Thu, 4 Mar 2021 at 15:57, Mark Rutland [off-list ref] wrote:
quoted
[adding Mark Brown]
The bigger problem here is that skipping is dodgy to begin with, and
this is still liable to break in some cases. One big concern is that
(especially with LTO) we cannot guarantee the compiler will not inline
or outline functions, causing the skipp value to be too large or too
small. That's liable to happen to callers, and in theory (though
unlikely in practice), portions of arch_stack_walk() or
stack_trace_save() could get outlined too.
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
Will LTO and friends respect 'noinline'?
I hope so (and suspect we'd have more problems otherwise), but I don't
know whether they actually so.
I suspect even with 'noinline' the compiler is permitted to outline
portions of a function if it wanted to (and IIUC it could still make
specialized copies in the absence of 'noclone').
quoted
One thing I also noticed is that tail calls would also cause the stack
trace to appear somewhat incomplete (for some of my tests I've
disabled tail call optimizations).
I assume you mean for a chain A->B->C where B tail-calls C, you get a
trace A->C? ... or is A going missing too?
Correct, it's just the A->C outcome.
I'd assumed that those cases were benign, e.g. for livepatching what
matters is what can be returned to, so B disappearing from the trace
isn't a problem there.
Is the concern debugability, or is there a functional issue you have in
mind?
For me, it's just been debuggability, and reliable test cases.
quoted
quoted
quoted
quoted
Is there a way to also mark a function non-tail-callable?
I think this can be bodged using __attribute__((optimize("$OPTIONS")))
on a caller to inhibit TCO (though IIRC GCC doesn't reliably support
function-local optimization options), but I don't expect there's any way
to mark a callee as not being tail-callable.
I don't think this is reliable. It'd be
__attribute__((optimize("-fno-optimize-sibling-calls"))), but doesn't
work if applied to the function we do not want to tail-call-optimize,
but would have to be applied to the function that does the tail-calling.
Yup; that's what I meant then I said you could do that on the caller but
not the callee.
I don't follow why you'd want to put this on the callee, though, so I
think I'm missing something. Considering a set of functions in different
compilation units:
A->B->C->D->E->F->G->H->I->J->K
I was having this problem with KCSAN, where the compiler would
tail-call-optimize __tsan_X instrumentation.
Those are compiler-generated calls, right? When those are generated the
compilation unit (and whatever it has included) might not have provided
a prototype anyway, and the compiler has special knowledge of the
functions, so it feels like the compiler would need to inhibit TCO here
for this to be robust. For their intended usage subjecting them to TCO
doesn't seem to make sense AFAICT.
I suspect that compilers have some way of handling that; otherwise I'd
expect to have heard stories of mcount/fentry calls getting TCO'd and
causing problems. So maybe there's an easy fix there?
I agree, the compiler builtins should be handled by the compiler
directly, perhaps that was a bad example. But we also have "explicit
instrumentation", e.g. everything that's in <linux/instrumented.h>.
quoted
This would mean that KCSAN runtime functions ended up in the trace,
but the function where the access happened would not. However, I don't
care about the runtime functions, and instead want to see the function
where the access happened. In that case, I'd like to just mark
__tsan_X and any other kcsan instrumentation functions as
do-not-tail-call-optimize, which would solve the problem.
I understand why we don't want to TCO these calls, but given the calls
are implicitly generated, I strongly suspect it's better to fix the
implicit call generation to not be TCO'd to begin with.
quoted
The solution today is that when you compile a kernel with KCSAN, every
instrumented TU is compiled with -fno-optimize-sibling-calls. The
better solution would be to just mark KCSAN runtime functions somehow,
but permit tail calling other things. Although, I probably still want
to see the full trace, and would decide that having
-fno-optimize-sibling-calls is a small price to pay in a
debug-only-kernel to get complete traces.
quoted
... if K were marked in this way, and J was compiled with visibility of
this, J would stick around, but J's callers might not, and so the a
trace might see:
A->J->K
... do you just care about the final caller, i.e. you just need
certainty that J will be in the trace?
Yes. But maybe it's a special problem that only sanitizers have.
I reckon for basically any instrumentation we don't want calls to be
TCO'd, though I'm not immediately sure of cases beyond sanitizers and
mcount/fentry.
Thinking about this more, I think it's all debugging tools. E.g.
lockdep, if you lock/unlock at the end of a function, you might tail
call into lockdep. If the compiler applies TCO, and lockdep determines
there's a bug and then shows a trace, you'll have no idea where the
actual bug is. The kernel has lots of debugging facilities that add
instrumentation in this way. So perhaps it's a general debugging-tool
problem (rather than just sanitizers).
Thanks,
-- Marco
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
From: Mark Rutland <mark.rutland@arm.com> Date: 2021-03-05 12:06:17
On Thu, Mar 04, 2021 at 08:01:29PM +0100, Marco Elver wrote:
On Thu, 4 Mar 2021 at 19:51, Mark Rutland [off-list ref] wrote:
quoted
On Thu, Mar 04, 2021 at 07:22:53PM +0100, Marco Elver wrote:
quoted
quoted
I was having this problem with KCSAN, where the compiler would
tail-call-optimize __tsan_X instrumentation.
Those are compiler-generated calls, right? When those are generated the
compilation unit (and whatever it has included) might not have provided
a prototype anyway, and the compiler has special knowledge of the
functions, so it feels like the compiler would need to inhibit TCO here
for this to be robust. For their intended usage subjecting them to TCO
doesn't seem to make sense AFAICT.
I suspect that compilers have some way of handling that; otherwise I'd
expect to have heard stories of mcount/fentry calls getting TCO'd and
causing problems. So maybe there's an easy fix there?
I agree, the compiler builtins should be handled by the compiler
directly, perhaps that was a bad example. But we also have "explicit
instrumentation", e.g. everything that's in <linux/instrumented.h>.
True -- I agree for those we want similar, and can see a case for a
no-tco-calls-to-me attribute on functions as with noreturn.
Maybe for now it's worth adding prevent_tail_call_optimization() to the
instrument_*() call wrappers in <linux/instrumented.h>? As those are
__always_inline, that should keep the function they get inlined in
around. Though we probably want to see if we can replace the mb() in
prevent_tail_call_optimization() with something that doesn't require a
real CPU barrier.
[...]
quoted
I reckon for basically any instrumentation we don't want calls to be
TCO'd, though I'm not immediately sure of cases beyond sanitizers and
mcount/fentry.
Thinking about this more, I think it's all debugging tools. E.g.
lockdep, if you lock/unlock at the end of a function, you might tail
call into lockdep. If the compiler applies TCO, and lockdep determines
there's a bug and then shows a trace, you'll have no idea where the
actual bug is. The kernel has lots of debugging facilities that add
instrumentation in this way. So perhaps it's a general debugging-tool
problem (rather than just sanitizers).
Hi!
On Thu, Mar 04, 2021 at 02:57:30PM +0000, Mark Rutland wrote:
It looks like GCC is happy to give us the function-entry-time FP if we use
__builtin_frame_address(1),
From the GCC manual:
Calling this function with a nonzero argument can have
unpredictable effects, including crashing the calling program. As
a result, calls that are considered unsafe are diagnosed when the
'-Wframe-address' option is in effect. Such calls should only be
made in debugging situations.
It *does* warn (the warning is in -Wall btw), on both powerpc and
aarch64. Furthermore, using this builtin causes lousy code (it forces
the use of a frame pointer, which we normally try very hard to optimise
away, for good reason).
And, that warning is not an idle warning. Non-zero arguments to
__builtin_frame_address can crash the program. It won't on simpler
functions, but there is no real definition of what a simpler function
*is*. It is meant for debugging, not for production use (this is also
why no one has bothered to make it faster).
On Power it should work, but on pretty much any other arch it won't.
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
You cannot get such a guarantee, other than not letting the compiler
see into the routine at all, like with assembler code (not inline asm,
real assembler code).
The real way forward is to bite the bullet and to no longer pretend you
can do a full backtrace from just the stack contents. You cannot.
Segher
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
From: Mark Rutland <mark.rutland@arm.com> Date: 2021-03-09 16:06:32
On Thu, Mar 04, 2021 at 03:54:48PM -0600, Segher Boessenkool wrote:
Hi!
Hi Segher,
On Thu, Mar 04, 2021 at 02:57:30PM +0000, Mark Rutland wrote:
quoted
It looks like GCC is happy to give us the function-entry-time FP if we use
__builtin_frame_address(1),
From the GCC manual:
Calling this function with a nonzero argument can have
unpredictable effects, including crashing the calling program. As
a result, calls that are considered unsafe are diagnosed when the
'-Wframe-address' option is in effect. Such calls should only be
made in debugging situations.
It *does* warn (the warning is in -Wall btw), on both powerpc and
aarch64. Furthermore, using this builtin causes lousy code (it forces
the use of a frame pointer, which we normally try very hard to optimise
away, for good reason).
And, that warning is not an idle warning. Non-zero arguments to
__builtin_frame_address can crash the program. It won't on simpler
functions, but there is no real definition of what a simpler function
*is*. It is meant for debugging, not for production use (this is also
why no one has bothered to make it faster).
On Power it should work, but on pretty much any other arch it won't.
I understand this is true generally, and cannot be relied upon in
portable code. However as you hint here for Power, I believe that on
arm64 __builtin_frame_address(1) shouldn't crash the program due to the
way frame records work on arm64, but I'll go check with some local
compiler folk. I agree that __builtin_frame_address(2) and beyond
certainly can, e.g. by NULL dereference and similar.
For context, why do you think this would work on power specifically? I
wonder if our rationale is similar.
Are you aware of anything in particular that breaks using
__builtin_frame_address(1) in non-portable code, or is this just a
general sentiment of this not being a supported use-case?
quoted
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
You cannot get such a guarantee, other than not letting the compiler
see into the routine at all, like with assembler code (not inline asm,
real assembler code).
If we cannot reliably ensure this then I'm happy to go write an assembly
trampoline to snapshot the state at a function call boundary (where our
procedure call standard mandates the state of the LR, FP, and frame
records pointed to by the FP). This'll require reworking a reasonable
amount of code cross-architecture, so I'll need to get some more
concrete justification (e.g. examples of things that can go wrong in
practice).
The real way forward is to bite the bullet and to no longer pretend you
can do a full backtrace from just the stack contents. You cannot.
I think what you mean here is that there's no reliable way to handle the
current/leaf function, right? If so I do agree.
Beyond that I believe that arm64's frame records should be sufficient.
Thanks,
Mark.
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Hi!
On Tue, Mar 09, 2021 at 04:05:23PM +0000, Mark Rutland wrote:
On Thu, Mar 04, 2021 at 03:54:48PM -0600, Segher Boessenkool wrote:
quoted
On Thu, Mar 04, 2021 at 02:57:30PM +0000, Mark Rutland wrote:
quoted
It looks like GCC is happy to give us the function-entry-time FP if we use
__builtin_frame_address(1),
From the GCC manual:
Calling this function with a nonzero argument can have
unpredictable effects, including crashing the calling program. As
a result, calls that are considered unsafe are diagnosed when the
'-Wframe-address' option is in effect. Such calls should only be
made in debugging situations.
It *does* warn (the warning is in -Wall btw), on both powerpc and
aarch64. Furthermore, using this builtin causes lousy code (it forces
the use of a frame pointer, which we normally try very hard to optimise
away, for good reason).
And, that warning is not an idle warning. Non-zero arguments to
__builtin_frame_address can crash the program. It won't on simpler
functions, but there is no real definition of what a simpler function
*is*. It is meant for debugging, not for production use (this is also
why no one has bothered to make it faster).
On Power it should work, but on pretty much any other arch it won't.
I understand this is true generally, and cannot be relied upon in
portable code. However as you hint here for Power, I believe that on
arm64 __builtin_frame_address(1) shouldn't crash the program due to the
way frame records work on arm64, but I'll go check with some local
compiler folk. I agree that __builtin_frame_address(2) and beyond
certainly can, e.g. by NULL dereference and similar.
I still do not know the aarch64 ABI well enough. If only I had time!
For context, why do you think this would work on power specifically? I
wonder if our rationale is similar.
On most 64-bit Power ABIs all stack frames are connected together as a
linked list (which is updated atomically, importantly). This makes it
possible to always find all previous stack frames.
Are you aware of anything in particular that breaks using
__builtin_frame_address(1) in non-portable code, or is this just a
general sentiment of this not being a supported use-case?
It is not supported, and trying to do it anyway can crash: it can use
random stack contents as pointer! Not really "random" of course, but
where it thinks to find a pointer into the previous frame, which is not
something it can rely on (unless the ABI guarantees it somehow).
See gcc.gnu.org/PR60109 for example.
quoted
quoted
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
You cannot get such a guarantee, other than not letting the compiler
see into the routine at all, like with assembler code (not inline asm,
real assembler code).
If we cannot reliably ensure this then I'm happy to go write an assembly
trampoline to snapshot the state at a function call boundary (where our
procedure call standard mandates the state of the LR, FP, and frame
records pointed to by the FP).
Is the frame pointer required?!
This'll require reworking a reasonable
amount of code cross-architecture, so I'll need to get some more
concrete justification (e.g. examples of things that can go wrong in
practice).
Say you have a function that does dynamic stack allocation, then there
is usually no way to find the previous stack frame (without function-
specific knowledge). So __builtin_frame_address cannot work (it knows
nothing about frames further up).
Dynamic stack allocation (alloca, or variable length automatic arrays)
is just the most common and most convenient example; it is not the only
case you have problems here.
quoted
The real way forward is to bite the bullet and to no longer pretend you
can do a full backtrace from just the stack contents. You cannot.
I think what you mean here is that there's no reliable way to handle the
current/leaf function, right? If so I do agree.
No, I meant what I said.
There is the separate issue that you do not know where the return
address (etc.) is stored in a function that has not yet done a call
itself, sure. You cannot assume anything the ABI does not tell you you
can depend on.
Beyond that I believe that arm64's frame records should be sufficient.
Do you have a simple linked list connecting all frames? The aarch64 GCC
port does not define anything special here (DYNAMIC_CHAIN_ADDRESS), so
the default will be used: every frame pointer has to point to the
previous one, no exceptions whatsoever.
Segher
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
From: Mark Rutland <mark.rutland@arm.com> Date: 2021-03-10 11:34:39
On Tue, Mar 09, 2021 at 04:05:32PM -0600, Segher Boessenkool wrote:
Hi!
On Tue, Mar 09, 2021 at 04:05:23PM +0000, Mark Rutland wrote:
quoted
On Thu, Mar 04, 2021 at 03:54:48PM -0600, Segher Boessenkool wrote:
quoted
On Thu, Mar 04, 2021 at 02:57:30PM +0000, Mark Rutland wrote:
quoted
It looks like GCC is happy to give us the function-entry-time FP if we use
__builtin_frame_address(1),
From the GCC manual:
Calling this function with a nonzero argument can have
unpredictable effects, including crashing the calling program. As
a result, calls that are considered unsafe are diagnosed when the
'-Wframe-address' option is in effect. Such calls should only be
made in debugging situations.
It *does* warn (the warning is in -Wall btw), on both powerpc and
aarch64. Furthermore, using this builtin causes lousy code (it forces
the use of a frame pointer, which we normally try very hard to optimise
away, for good reason).
And, that warning is not an idle warning. Non-zero arguments to
__builtin_frame_address can crash the program. It won't on simpler
functions, but there is no real definition of what a simpler function
*is*. It is meant for debugging, not for production use (this is also
why no one has bothered to make it faster).
On Power it should work, but on pretty much any other arch it won't.
I understand this is true generally, and cannot be relied upon in
portable code. However as you hint here for Power, I believe that on
arm64 __builtin_frame_address(1) shouldn't crash the program due to the
way frame records work on arm64, but I'll go check with some local
compiler folk. I agree that __builtin_frame_address(2) and beyond
certainly can, e.g. by NULL dereference and similar.
I still do not know the aarch64 ABI well enough. If only I had time!
quoted
For context, why do you think this would work on power specifically? I
wonder if our rationale is similar.
On most 64-bit Power ABIs all stack frames are connected together as a
linked list (which is updated atomically, importantly). This makes it
possible to always find all previous stack frames.
We have something similar on arm64, where the kernel depends on being
built with a frame pointer following the AAPCS frame pointer rules.
Every stack frame contains a "frame record" *somewhere* within that
stack frame, and the frame records are chained together as a linked
list. The frame pointer points at the most recent frame record (and this
is what __builtin_frame_address(0) returns).
The records themselves are basically:
struct record {
struct record *next;
unsigned long ret_addr;
};
At function call boundaries, we know that the FP is the caller's record
(or NULL for the first function), and the LR is the address the current
function should return to. Within a function with a stack frame, we can
access that function's record and the `next` field (equivalent to the FP
at the time of entry to the function) is what __builtin_frame_address(1)
should return.
quoted
Are you aware of anything in particular that breaks using
__builtin_frame_address(1) in non-portable code, or is this just a
general sentiment of this not being a supported use-case?
It is not supported, and trying to do it anyway can crash: it can use
random stack contents as pointer! Not really "random" of course, but
where it thinks to find a pointer into the previous frame, which is not
something it can rely on (unless the ABI guarantees it somehow).
See gcc.gnu.org/PR60109 for example.
Sure; I see that being true generally (and Ramana noted that on 32-bit
arm a frame pointer wasn't mandated), but I think in this case we have a
stronger target (and configuration) specific guarantee.
quoted
quoted
quoted
Unless we can get some strong guarantees from compiler folk such that we
can guarantee a specific function acts boundary for unwinding (and
doesn't itself get split, etc), the only reliable way I can think to
solve this requires an assembly trampoline. Whatever we do is liable to
need some invasive rework.
You cannot get such a guarantee, other than not letting the compiler
see into the routine at all, like with assembler code (not inline asm,
real assembler code).
If we cannot reliably ensure this then I'm happy to go write an assembly
trampoline to snapshot the state at a function call boundary (where our
procedure call standard mandates the state of the LR, FP, and frame
records pointed to by the FP).
Is the frame pointer required?!
The arm64 Linux port mandates frame pointers for kernel code. It is
generally possible to build code without frame pointers (e.g. userspace),
but doing that for kernel code would be a bug.
quoted
This'll require reworking a reasonable
amount of code cross-architecture, so I'll need to get some more
concrete justification (e.g. examples of things that can go wrong in
practice).
Say you have a function that does dynamic stack allocation, then there
is usually no way to find the previous stack frame (without function-
specific knowledge). So __builtin_frame_address cannot work (it knows
nothing about frames further up).
Dynamic stack allocation (alloca, or variable length automatic arrays)
is just the most common and most convenient example; it is not the only
case you have problems here.
I agree with those as general concerns, but I don't think that affects
arm64's frame records, since their location within a stack frame is
immaterial given the chaining.
quoted
quoted
The real way forward is to bite the bullet and to no longer pretend you
can do a full backtrace from just the stack contents. You cannot.
I think what you mean here is that there's no reliable way to handle the
current/leaf function, right? If so I do agree.
No, I meant what I said.
There is the separate issue that you do not know where the return
address (etc.) is stored in a function that has not yet done a call
itself, sure. You cannot assume anything the ABI does not tell you you
can depend on.
This is in the frame record per the AAPCS.
quoted
Beyond that I believe that arm64's frame records should be sufficient.
Do you have a simple linked list connecting all frames?
Hi!
On Wed, Mar 10, 2021 at 11:32:20AM +0000, Mark Rutland wrote:
On Tue, Mar 09, 2021 at 04:05:32PM -0600, Segher Boessenkool wrote:
quoted
On Tue, Mar 09, 2021 at 04:05:23PM +0000, Mark Rutland wrote:
quoted
On Thu, Mar 04, 2021 at 03:54:48PM -0600, Segher Boessenkool wrote:
quoted
On Thu, Mar 04, 2021 at 02:57:30PM +0000, Mark Rutland wrote:
quoted
It looks like GCC is happy to give us the function-entry-time FP if we use
__builtin_frame_address(1),
From the GCC manual:
Calling this function with a nonzero argument can have
unpredictable effects, including crashing the calling program. As
a result, calls that are considered unsafe are diagnosed when the
'-Wframe-address' option is in effect. Such calls should only be
made in debugging situations.
It *does* warn (the warning is in -Wall btw), on both powerpc and
aarch64. Furthermore, using this builtin causes lousy code (it forces
the use of a frame pointer, which we normally try very hard to optimise
away, for good reason).
And, that warning is not an idle warning. Non-zero arguments to
__builtin_frame_address can crash the program. It won't on simpler
functions, but there is no real definition of what a simpler function
*is*. It is meant for debugging, not for production use (this is also
why no one has bothered to make it faster).
On Power it should work, but on pretty much any other arch it won't.
I understand this is true generally, and cannot be relied upon in
portable code. However as you hint here for Power, I believe that on
arm64 __builtin_frame_address(1) shouldn't crash the program due to the
way frame records work on arm64, but I'll go check with some local
compiler folk. I agree that __builtin_frame_address(2) and beyond
certainly can, e.g. by NULL dereference and similar.
I still do not know the aarch64 ABI well enough. If only I had time!
quoted
For context, why do you think this would work on power specifically? I
wonder if our rationale is similar.
On most 64-bit Power ABIs all stack frames are connected together as a
linked list (which is updated atomically, importantly). This makes it
possible to always find all previous stack frames.
We have something similar on arm64, where the kernel depends on being
built with a frame pointer following the AAPCS frame pointer rules.
The huge difference is on Power this is about the stack itself: you do
not need a frame pointer at all for it (there is no specific register
named as frame pointer, even).
Every stack frame contains a "frame record" *somewhere* within that
stack frame, and the frame records are chained together as a linked
list. The frame pointer points at the most recent frame record (and this
is what __builtin_frame_address(0) returns).
quoted
See gcc.gnu.org/PR60109 for example.
Sure; I see that being true generally (and Ramana noted that on 32-bit
arm a frame pointer wasn't mandated), but I think in this case we have a
stronger target (and configuration) specific guarantee.
It sounds like it, yes. You need to have a frame pointer in the ABI,
with pretty strong rules, and have everything follow those rules.
quoted
Is the frame pointer required?!
The arm64 Linux port mandates frame pointers for kernel code. It is
generally possible to build code without frame pointers (e.g. userspace),
but doing that for kernel code would be a bug.
I see. And it even is less expensive to do this than on most machines,
because of register pair load/store instructions :-)
quoted
quoted
quoted
The real way forward is to bite the bullet and to no longer pretend you
can do a full backtrace from just the stack contents. You cannot.
I think what you mean here is that there's no reliable way to handle the
current/leaf function, right? If so I do agree.
No, I meant what I said.
There is the separate issue that you do not know where the return
address (etc.) is stored in a function that has not yet done a call
itself, sure. You cannot assume anything the ABI does not tell you you
can depend on.
This is in the frame record per the AAPCS.
But you do not know where in the function it will store that. It often
can be optimised by the compiler to only store the LR and FP on paths
where a call will happen later, and there is no way (without DWARF info
or similar) to know whether that has happened yet or not.
This is a well-known problem of course. For the current function you
cannot know in general if there is an activation frame yet or not.
Segher
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel