Prev: net: Micrel KSZ8841/2 PCI Ethernet driver
Next: [PATCH] perf: Fix implicit declaration of getline in util.c
From: Hitoshi Mitake on 16 Jan 2010 08:10
On 2010年01月13日 18:52, Peter Zijlstra wrote:
> On Thu, 2010-01-07 at 19:39 +0900, Hitoshi Mitake wrote:
>> There are a lot of lock instances with same names (e.g. port_lock).
>> This patch series add __FILE__ and __LINE__ to lockdep_map,
>> and these will be used for trace lock events.
>>
>> Example use from perf lock map:
>>
>> | 0xffffea0004c992b8: __pte_lockptr(page) (src: include/linux/mm.h, line: 952)
>> | 0xffffea0004b112b8: __pte_lockptr(page) (src: include/linux/mm.h, line: 952)
>> | 0xffffea0004a3f2b8: __pte_lockptr(page) (src: include/linux/mm.h, line: 952)
>> | 0xffffea0004cd5228: __pte_lockptr(page) (src: include/linux/mm.h, line: 952)
>> | 0xffff8800b91e2b28:&sb->s_type->i_lock_key (src: fs/inode.c, line: 166)
>> | 0xffff8800bb9d7ae0: key (src: kernel/wait.c, line: 16)
>> | 0xffff8800aa07dae0:&dentry->d_lock (src: fs/dcache.c, line: 944)
>> | 0xffff8800b07fbae0:&dentry->d_lock (src: fs/dcache.c, line: 944)
>> | 0xffff8800b07f3ae0:&dentry->d_lock (src: fs/dcache.c, line: 944)
>> | 0xffff8800bf15fae0:&sighand->siglock (src: kernel/fork.c, line: 1490)
>> | 0xffff8800b90f7ae0:&dentry->d_lock (src: fs/dcache.c, line: 944)
>> | ...
>>
>> (This output of perf lock map is produced by my local version,
>> I'll send this later.)
>>
>> And sadly, as Peter Zijlstra predicted, this produces certain overhead.
>>
>> Before appling this series:
>> | % sudo ./perf lock rec perf bench sched messaging
>> | # Running sched/messaging benchmark...
>> | # 20 sender and receiver processes per group
>> | # 10 groups == 400 processes run
>> |
>> | Total time: 3.834 [sec]
>> After:
>> sudo ./perf lock rec perf bench sched messaging
>> | # Running sched/messaging benchmark...
>> | # 20 sender and receiver processes per group
>> | # 10 groups == 400 processes run
>> |
>> | Total time: 5.415 [sec]
>> | [ perf record: Woken up 0 times to write data ]
>> | [ perf record: Captured and wrote 53.512 MB perf.data (~2337993 samples) ]
>>
>> But raw exec of perf bench sched messaging is this:
>> | % perf bench sched messaging
>> | # Running sched/messaging benchmark...
>> | # 20 sender and receiver processes per group
>> | # 10 groups == 400 processes run
>> |
>> | Total time: 0.498 [sec]
>>
>> Tracing lock events already produces amount of overhead.
>> I think the overhead produced by this series is not a fatal problem,
>> radically optimization is required...
>
> Right, these patches look OK, for the tracing overhead, you could
> possibly hash the file:line into a u64 and reduce the tracepoint size,
> that should improve the situation I tihnk, because I seem to remember
> the only thing that really matters for speed is the size of things.
>
>

Thanks for your opinion, Peter.
I'll work on reducing size of events later. Hashing is a good idea.
I think indexing is also way to reduce size.

And I want lockdep_map to have another thing, type of lock.
For example, mutex and spinlock has completely different acquired time
and attributes,
so I want to separate these things. If lockdep_map has member
to express type, things will be easy.

How do you think?
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
 | 
Pages: 1
Prev: net: Micrel KSZ8841/2 PCI Ethernet driver
Next: [PATCH] perf: Fix implicit declaration of getline in util.c