Currently struct of_device_id is 196 (32-bit) or 200 (64-bit) bytes
large. It contains fixed-size strings for a name, a type, and a
compatible value, but the first two are barely used.
OF device ID tables contain multiple entries, plus an empty sentinel
entry.
Statistics for my current kernel source tree:
- 4487 tables with 16836 entries (3367200 bytes)
- 176 names (average 6.7 max 23 chars)
- 66 types (average 5.1 max 21 chars)
- 12192 compatible values (average 18.0 max 45 chars)
Taking into account the minimum needed size to store the strings, only
6.9% of the allocated space is used...
Reduce kernel size by reducing the sizes of the fixed strings by one
half.
This reduces the size of an ARM multi_v7_defconfig kernel by ca. 400
KiB. For a typical kernel supporting a single board, you can expect to
save 50-100 KiB.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
---
Notes:
- While gcc complains if the non-NUL characters in a string do not fit
in the available space, it does not complain if there is no space to
store the string's NUL-terminator. However, that should be caught
during testing, as the affected entry won't ever match. The kernel
won't crash, as such strings will still be terminated by the
sentinel in the table.
- We could save even more by converting the strings from fixed-size
arrays to pointers, at the expense of making it harder to mark
entries __init. Given most drivers support binding and unbinding
and thus cannot use __init for of_device_id tables, perhaps that's
the way to go?
Thanks for your comments!
---
include/linux/mod_devicetable.h | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
Currently struct of_device_id is 196 (32-bit) or 200 (64-bit) bytes
large. It contains fixed-size strings for a name, a type, and a
compatible value, but the first two are barely used.
OF device ID tables contain multiple entries, plus an empty sentinel
entry.
Statistics for my current kernel source tree:
- 4487 tables with 16836 entries (3367200 bytes)
- 176 names (average 6.7 max 23 chars)
- 66 types (average 5.1 max 21 chars)
- 12192 compatible values (average 18.0 max 45 chars)
Taking into account the minimum needed size to store the strings, only
6.9% of the allocated space is used...
Reduce kernel size by reducing the sizes of the fixed strings by one
half.
Tried something like this 2.5 years ago:
https://lore.kernel.org/lkml/20190425203101.9403-1-linux@rasmusvillemoes.dk/
I think that there might be some not-in-tree code that relies on the
existing layout. I considered adding a CONFIG_ knob, either for these
sizes in particular, or more generally a def_bool y "CONFIG_LEGACY"
which embedded folks that build the entire distro from source and don't
have any legacy things can turn off, and then get more sensible defaults
all around.
Including fx in the TCP stack where some CVE fix required changing some
parameter, but the kernel itself couldn't ship a sane default because
no-regressions, so userspace had to learn to set yet another sysctl
properly.
Rasmus
Hi Rasmus,
On Wed, Nov 10, 2021 at 5:51 PM Rasmus Villemoes
[off-list ref] wrote:
On 10/11/2021 17.23, Geert Uytterhoeven wrote:
quoted
Currently struct of_device_id is 196 (32-bit) or 200 (64-bit) bytes
large. It contains fixed-size strings for a name, a type, and a
compatible value, but the first two are barely used.
OF device ID tables contain multiple entries, plus an empty sentinel
entry.
Statistics for my current kernel source tree:
- 4487 tables with 16836 entries (3367200 bytes)
- 176 names (average 6.7 max 23 chars)
- 66 types (average 5.1 max 21 chars)
- 12192 compatible values (average 18.0 max 45 chars)
Taking into account the minimum needed size to store the strings, only
6.9% of the allocated space is used...
Reduce kernel size by reducing the sizes of the fixed strings by one
half.
I think that there might be some not-in-tree code that relies on the
existing layout. I considered adding a CONFIG_ knob, either for these
sizes in particular, or more generally a def_bool y "CONFIG_LEGACY"
which embedded folks that build the entire distro from source and don't
have any legacy things can turn off, and then get more sensible defaults
all around.
Most of that should have been gone since the #ifdef KERNEL was removed
from include/linux/mod_devicetable.h in commit 6543becf26fff612
("mod/file2alias: make modalias generation safe for cross compiling").
Of course you can never know for sure...
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
From: Frank Rowand <hidden> Date: 2021-11-11 04:06:37
Hi Geert,
On 11/10/21 11:23 AM, Geert Uytterhoeven wrote:
Currently struct of_device_id is 196 (32-bit) or 200 (64-bit) bytes
large. It contains fixed-size strings for a name, a type, and a
compatible value, but the first two are barely used.
OF device ID tables contain multiple entries, plus an empty sentinel
entry.
Statistics for my current kernel source tree:
- 4487 tables with 16836 entries (3367200 bytes)
- 176 names (average 6.7 max 23 chars)
- 66 types (average 5.1 max 21 chars)
- 12192 compatible values (average 18.0 max 45 chars)
Taking into account the minimum needed size to store the strings, only
6.9% of the allocated space is used...
I like the idea of using less memory (and thank you for the above data!),
but I do not like the implementation, which reduces the size (of name at
least - I didn't check each field) to less than what the standard allows.
I have an idea of another way to accomplish the same goal, but I need to
dig a bit to make sure I'm not shooting myself in the foot.
-Frank
quoted hunk
Reduce kernel size by reducing the sizes of the fixed strings by one
half.
This reduces the size of an ARM multi_v7_defconfig kernel by ca. 400
KiB. For a typical kernel supporting a single board, you can expect to
save 50-100 KiB.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
---
Notes:
- While gcc complains if the non-NUL characters in a string do not fit
in the available space, it does not complain if there is no space to
store the string's NUL-terminator. However, that should be caught
during testing, as the affected entry won't ever match. The kernel
won't crash, as such strings will still be terminated by the
sentinel in the table.
- We could save even more by converting the strings from fixed-size
arrays to pointers, at the expense of making it harder to mark
entries __init. Given most drivers support binding and unbinding
and thus cannot use __init for of_device_id tables, perhaps that's
the way to go?
Thanks for your comments!
---
include/linux/mod_devicetable.h | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)