| Message ID | ef59d6fd3b2201b912d5eaa7f7a037d8f9adb744.1636561068.git.geert+renesas@glider.be (mailing list archive) |
|---|---|
| State | Under Review |
| Delegated to: | Geert Uytterhoeven |
| Headers | show |
| Series | [PATCH/RFC] of: Shrink struct of_device_id | expand |
On 10/11/2021 17.23, 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... > > 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 <rasmus.villemoes@prevas.dk> wrote: > On 10/11/2021 17.23, 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... > > > > 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 wasn't aware of that. I reworked some code which used multiple of_find_compatible_node() calls before, and noticed the end result had grown a lot due to the sheer size of of_device_id ("[PATCH] soc: renesas: Consolidate product register handling", https://lore.kernel.org/all/057721f46c7499de4133135488f0f3da7fb39265.1636570669.git.geert+renesas@glider.be). > 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
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 > > 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(-) > > diff --git a/include/linux/mod_devicetable.h b/include/linux/mod_devicetable.h > index ae2e75d15b219920..2bb2558d52d30d2b 100644 > --- a/include/linux/mod_devicetable.h > +++ b/include/linux/mod_devicetable.h > @@ -266,9 +266,9 @@ struct sdw_device_id { > * Struct used for matching a device > */ > struct of_device_id { > - char name[32]; > - char type[32]; > - char compatible[128]; > + char name[24]; > + char type[24]; > + char compatible[48]; > const void *data; > }; > >
diff --git a/include/linux/mod_devicetable.h b/include/linux/mod_devicetable.h index ae2e75d15b219920..2bb2558d52d30d2b 100644 --- a/include/linux/mod_devicetable.h +++ b/include/linux/mod_devicetable.h @@ -266,9 +266,9 @@ struct sdw_device_id { * Struct used for matching a device */ struct of_device_id { - char name[32]; - char type[32]; - char compatible[128]; + char name[24]; + char type[24]; + char compatible[48]; const void *data; };
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(-)