ureadahead patch

[Tim Gardner]

On 07/08/2010 01:46 PM, Chase Douglas wrote:
> On Thu, 2010-07-08 at 10:24 -0600, Tim Gardner wrote:
>> diff -u ureadahead-0.100.0/debian/changelog ureadahead-0.100.0/debian/changelog
>> --- ureadahead-0.100.0/debian/changelog
>> +++ ureadahead-0.100.0/debian/changelog
>> @@ -1,3 +1,9 @@
>> +ureadahead (0.100.0-5ubuntu1) maverick; urgency=low
>> +
>> +  *
>> +
>> + -- Tim Gardner<rtg at lochsa>   Thu, 08 Jul 2010 16:04:39 +0000
>> +
>>   ureadahead (0.100.0-5) maverick; urgency=low
>>
>>     * src/pack.c: Amend mount point detection logic to stat the mount point
>> only in patch2:
>> unchanged:
>> --- ureadahead-0.100.0.orig/src/trace.c
>> +++ ureadahead-0.100.0/src/trace.c
>> @@ -122,6 +122,7 @@
>>   	int                 old_open_exec_enabled = 0;
>>   	int                 old_uselib_enabled = 0;
>>   	int                 old_tracing_enabled = 0;
>> +	int                 old_buffer_size_kb = 0;
>>   	struct sigaction    act;
>>   	struct sigaction    old_sigterm;
>>   	struct sigaction    old_sigint;
>> @@ -165,7 +166,7 @@
>>
>>   		old_uselib_enabled = -1;
>>   	}
>> -	if (set_value (dfd, "buffer_size_kb", 128000, NULL)<  0)
>> +	if (set_value (dfd, "buffer_size_kb", 128000,&old_buffer_size_kb)<  0)
>>   		goto error;
>>   	if (set_value (dfd, "tracing_enabled",
>>   		       TRUE,&old_tracing_enabled)<  0)
>> @@ -217,6 +218,9 @@
>>   	if (set_value (dfd, "events/fs/do_sys_open/enable",
>>   		       old_sys_open_enabled, NULL)<  0)
>>   		goto error;
>> +	if (set_value (dfd, "buffer_size_kb",
>> +		       old_buffer_size_kb, NULL)<  0)
>> +		goto error;
>>
>>   	/* Be nicer */
>>   	if (nice (15))
>>
>
> I'm guessing that set_value copies what would be the result of "cat
> buffer_size_kb" into old_buffer_size_kb. Unfortunately, that would end
> up with:
>
> 7 (expanded: 1408)
>
> Or something like that. This output signifies that at first we only have
> 7 KB/cpu of buffer, but as soon as a function trace is started we
> allocate 1408 KB/cpu. When you manually change the value, you
> automatically get pushed to the expanded state. This is a one-way
> operation; we can't go back to the original state once we bump the value
> in ureadahead.
>
> I think that the best thing is to just set it back to 1 (a minimum
> value). If someone wants to do a trace, they can bump this value up. I'm
> wary of even going back to the expanded value because in the vast
> majority of cases we're just wasting memory.
>
> -- Chase
>
>

The first thing set_value() does is to perform atoi() on the buffer read 
from /sys/kernel/debug/tracing/buffer_size_kb, the first token of which 
appears to always be an integer. Isn't that the number we'd like to 
preserve?

rtg
-- 
Tim Gardner tim.gardner at canonical.com