Re: [PATCH 0/5] kstrdup optimization

[Andrzej Hajda]

On 01/14/2015 12:37 AM, Andrew Morton wrote:
> On Mon, 12 Jan 2015 10:18:38 +0100 Andrzej Hajda <a.hajda@xxxxxxxxxxx> wrote:
>
>> Hi,
>>
>> kstrdup if often used to duplicate strings where neither source neither
>> destination will be ever modified. In such case we can just reuse the source
>> instead of duplicating it. The problem is that we must be sure that
>> the source is non-modifiable and its life-time is long enough.
>>
>> I suspect the good candidates for such strings are strings located in kernel
>> .rodata section, they cannot be modifed because the section is read-only and
>> their life-time is equal to kernel life-time.
>>
>> This small patchset proposes alternative version of kstrdup - kstrdup_const,
>> which returns source string if it is located in .rodata otherwise it fallbacks
>> to kstrdup.
>> To verify if the source is in .rodata function checks if the address is between
>> sentinels __start_rodata, __end_rodata. I guess it should work with all
>> architectures.
>>
>> The main patch is accompanied by four patches constifying kstrdup for cases
>> where situtation described above happens frequently.
>>
>> As I have tested the patchset on mobile platform (exynos4210-trats) it saves
>> 3272 string allocations. Since minimal allocation is 32 or 64 bytes depending
>> on Kconfig options the patchset saves respectively about 100KB or 200KB of memory.
> That's a lot of memory.  I wonder where it's all going to.  sysfs,
> probably?

Stats from tested platform.
By caller:
  2260 __kernfs_new_node
    631 clk_register+0xc8/0x1b8
    318 clk_register+0x34/0x1b8
      51 kmem_cache_create
      12 alloc_vfsmnt

By string (with count >= 5):
    883 power
    876 subsystem
    135 parameters
    132 device
     61 iommu_group
     44 sclk_mpll
     42 aclk100
     41 driver
     36 sclk_vpll
     35 none
     34 sclk_epll
     34 aclk160
     32 sclk_hdmi24m
     31 xxti
     31 xusbxti
     31 sclk_usbphy0
     30 sclk_hdmiphy
     28 bdi
     28 aclk133
     14 sclk_apll
     14 aclk200
      9 module
      9 fin_pll
      5 div_core2
   


>
> What the heck does (the cheerily undocumented) KERNFS_STATIC_NAME do
> and can we remove it if this patchset is in place?
>
>

The only call path when this flag is set starts from
sysfs_add_file_mode_ns function.
But I guess this function can be called also for non-const names.

Regards
Andrzej


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/