On 01/13/2016 03:04 AM, Will Deacon wrote:
On Tue, Jan 12, 2016 at 03:11:29PM +0900, AKASHI Takahiro wrote:
Will,
On 01/09/2016 12:53 AM, Will Deacon wrote:
On Fri, Jan 08, 2016 at 02:36:32PM +0900, AKASHI Takahiro wrote:
On 01/07/2016 11:56 PM, Richard Earnshaw (lists) wrote:
On 07/01/16 14:22, Will Deacon wrote:
On Thu, Dec 24, 2015 at 04:57:54PM +0900, AKASHI Takahiro wrote:
So I'd like to introduce a function prologue analyzer to determine
a size allocated by a function's prologue and deduce it from "Depth".
My implementation of this analyzer has been submitted to
linux-arm-kernel mailing list[1].
I borrowed some ideas from gdb's analyzer[2], especially a loop of
instruction decoding as well as stop of decoding at exiting a basic block,
but implemented my own simplified one because gdb version seems to do
a bit more than what we expect here.
Anyhow, since it is somewhat heuristic (and may not be maintainable for
a long term), could you review it from a broader viewpoint of toolchain,
please?
My main issue with this is that we cannot rely on the frame layout
generated by the compiler and there's little point in asking for
commitment here. Therefore, the heuristics will need updating as and
when we identify new frames that we can't handle. That's pretty fragile
and puts us on the back foot when faced with newer compilers. This might
be sustainable if we don't expect to encounter much variation, but even
that would require some sort of "buy-in" from the various toolchain
communities.
GCC already has an option (-fstack-usage) to determine the stack usage
on a per-function basis and produce a report at build time. Why can't
we use that to provide the information we need, rather than attempt to
compute it at runtime based on your analyser?
If -fstack-usage is not sufficient, understanding why might allow us to
propose a better option.
Can you not use the dwarf frame unwind data? That's always sufficient
to recover the CFA (canonical frame address - the value in SP when
executing the first instruction in a function). It seems to me it's
unlikely you're going to need something that's an exceedingly high
performance operation.
Thank you for your comment.
Yeah, but we need some utility routines to handle unwind data(.debug_frame).
In fact, some guy has already attempted to merge (part of) libunwind into
the kernel[1], but it was rejected by the kernel community (including Linus
if I correctly remember). It seems that they thought the code was still buggy.
The ARC guys seem to have sneaked something in for their architecture:
http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/tree/arch/arc/kernel/unwind.c
so it might not be impossible if we don't require all the bells and
whistles of libunwind.
Thanks. I didn't notice this code.
That is one of reasons that I wanted to implement my own analyzer.
I still don't understand why you can't use fstack-usage. Can you please
tell me why that doesn't work? Am I missing something?
I don't know how gcc calculates the usage here, but I guess it would be more
robust than my analyzer.
The issues, that come up to my mind, are
- -fstack-usage generates a separate output file, *.su and so we have to
manage them to be incorporated in the kernel binary.
That doesn't sound too bad to me. How much data are we talking about here?
This implies that (common) kernel makefiles might have to be a bit changed.
- more worse, what if kernel module case? We will have no way to let the kernel
know the stack usage without adding an extra step at loading.
We can easily add a new __init section to modules, which is a table
representing the module functions and their stack sizes (like we do
for other things like alternatives). We'd just then need to slurp this
information at load time and throw it into an rbtree or something.
I found another issue.
Let's think about 'dynamic storage' case like:
$ cat stack.c
extern long fooX(long a);
extern long fooY(long b[]);
long foo1(long a) {
if (a > 1) {
long b[a]; <== Here
return a + fooY(b);
} else {
return a + fooX(a);
}
}
Then, -fstack-usage returns 48 for foo1():
$ aarch64-linux-gnu-gcc -fno-omit-frame-pointer -fstack-usage main.c stack.c \
-pg -O2 -fasynchronous-unwind-tables
$ cat stack.su
stack.c:4:6:foo1 48 dynamic
This indicates that foo1() may use 48 bytes or more depending on a condition.
But in my case (ftrace-based stack tracer), I always expect 32 whether we're
backtracing from fooY() or from fooX() because my stack tracer estimates:
(stack pointer) = (callee's frame pointer) + (callee's stack usage)
(in my previous e-mail, '-(minus)' was wrong.)
where (callee's stack usage) is, as I described in my previous e-mail, a size of
memory which is initially allocated on a stack in a function prologue, and
should not
contain a size of dynamically allocate area.
Unfortunately, there are several places in the kernel where "b[a]"-like variable
definitions are used.
-Takahiro AKASHI
FYI,
(gdb) disas foo1
Dump of assembler code for function foo1:
0x0000000000400758 <+0>: stp x29, x30, [sp,#-32]!
0x000000000040075c <+4>: mov x29, sp
0x0000000000400760 <+8>: stp x19, x20, [sp,#16]
0x0000000000400764 <+12>: mov x19, x0
0x0000000000400768 <+16>: mov x0, x30
0x000000000040076c <+20>: bl 0x400540 <_mcount@plt>
0x0000000000400770 <+24>: cmp x19, #0x1
0x0000000000400774 <+28>: b.le 0x4007b0 <foo1+88>
0x0000000000400778 <+32>: lsl x0, x19, #3
0x000000000040077c <+36>: mov x20, sp
0x0000000000400780 <+40>: add x0, x0, #0x16
0x0000000000400784 <+44>: and x0, x0, #0xfffffffffffffff0
0x0000000000400788 <+48>: sub sp, sp, x0
0x000000000040078c <+52>: mov x0, sp
0x0000000000400790 <+56>: bl 0x400730 <fooY>
0x0000000000400794 <+60>: mov sp, x20
0x0000000000400798 <+64>: mov sp, x29
0x000000000040079c <+68>: add x0, x19, x0
0x00000000004007a0 <+72>: ldp x19, x20, [sp,#16]
0x00000000004007a4 <+76>: ldp x29, x30, [sp],#32
0x00000000004007a8 <+80>: ret
0x00000000004007ac <+84>: nop
0x00000000004007b0 <+88>: mov x0, x19
0x00000000004007b4 <+92>: bl 0x400708 <fooX>
0x00000000004007b8 <+96>: mov sp, x29
0x00000000004007bc <+100>: add x0, x19, x0
0x00000000004007c0 <+104>: ldp x19, x20, [sp,#16]
0x00000000004007c4 <+108>: ldp x29, x30, [sp],#32
0x00000000004007c8 <+112>: ret
End of assembler dump.
Will
