Hello,

I was looking at the performance of faddr2line and noticed that it spends
most of its time performing two tasks:
- (1) Making redundant calls to readelf and addr2line for each address
  (e.g., it makes 7 readelf calls and 2 addr2line calls when converting 1
address, and it makes 252 readelf calls and 51 addr2line calls when
converting 50 addresses); and
- (2) Calculating a symbol’s size by unnecessarily looping through every
  symbol in a particular section.

This patch series consists of the following parts:
- Patches 1-2 reduce the total number of readelf calls to one,
- Patches 3-4 make minor changes in preparation for the following patches,
- Patches 5-6 reduce the total number of addr2line calls to one, and
- Patch 7 modifies the symbol size calculation to only check two symbols.

I evaluated the performance of the patch series by running faddr2line with
a standard kernel image (X86_64 defconfig with CONFIG_DEBUG_INFO enabled).
When converting only 1 address, the patch series gives a negligible
speedup. When converting 50 addresses, however, it gives a 15x speedup.

Cheers,

Brian

Brian Johannesmeyer (7):
  scripts/faddr2line: Reduce number of readelf calls to three
  scripts/faddr2line: Combine three readelf calls into one
  scripts/faddr2line: Check vmlinux only once
  scripts/faddr2line: Pass --addresses argument to addr2line
  scripts/faddr2line: Invoke addr2line as a single long-running process
  scripts/faddr2line: Remove call to addr2line from find_dir_prefix()
  scripts/faddr2line: Check only two symbols when calculating symbol
    size

 scripts/faddr2line | 110 ++++++++++++++++++++++++++++++++++-----------
 1 file changed, 84 insertions(+), 26 deletions(-)

-- 
2.34.1

On Mon, Mar 11, 2024 at 09:40:12PM +0100, Brian Johannesmeyer wrote:
> Hello,
> 
> I was looking at the performance of faddr2line and noticed that it spends
> most of its time performing two tasks:
> - (1) Making redundant calls to readelf and addr2line for each address
>   (e.g., it makes 7 readelf calls and 2 addr2line calls when converting 1
> address, and it makes 252 readelf calls and 51 addr2line calls when
> converting 50 addresses); and
> - (2) Calculating a symbol’s size by unnecessarily looping through every
>   symbol in a particular section.
> 
> This patch series consists of the following parts:
> - Patches 1-2 reduce the total number of readelf calls to one,
> - Patches 3-4 make minor changes in preparation for the following patches,
> - Patches 5-6 reduce the total number of addr2line calls to one, and
> - Patch 7 modifies the symbol size calculation to only check two symbols.
> 
> I evaluated the performance of the patch series by running faddr2line with
> a standard kernel image (X86_64 defconfig with CONFIG_DEBUG_INFO enabled).
> When converting only 1 address, the patch series gives a negligible
> speedup. When converting 50 addresses, however, it gives a 15x speedup.

This looks really nice.

If testing runs ok, I'll go ahead and queue it up.

Thanks!

-- 
Josh