On 05/23/2011 06:58 PM, Simon Tian wrote:
Hi,
If call wait_for_complete() and release() every time, the It will seem
like just a synchronized mean..Ha ha,
So I think the callback function should pass the return value in, and
the completion should be released as soon as possible. This will make
the performance better.
So I modified some code:
librbd.hpp:
typedef void (*callback_t)(completion_t cb, void *arg, ssize_t rval);
librbd.cc:
void AioCompletion::complete_block(AioBlockCompletion
*block_completion, ssize_t r)
{
dout(10)<< "AioCompletion::complete_block this="<< (void *)this<<
" complete_cb="<< (void *)complete_cb<< dendl;
lock.Lock();
if (rval>= 0) {
if (r< 0&& r != -EEXIST)
rval = r;
else if (r> 0)
rval += r;
}
assert(pending_count);
//put_unlock();
int count = --pending_count;
if (!count) {
done = true;
cond.Signal();
if (complete_cb)
{
complete_cb(rbd_comp, complete_arg, rval);
}
lock.Unlock();
librbd::RBD::AioCompletion * comp_temp =
(librbd::RBD::AioCompletion *)(rbd_comp);
comp_temp->release();
}
else{
lock.Unlock();
}
}
I'm not sure how much this will help. Can you measure the difference
once it's working?
I apply this, when aio reading, segment fault comes:
=======================================
Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x410a3940 (LWP 8913)]
0x00007f2c622a5b54 in librbd::AioBlockCompletion::complete
(this=0x7f2c5c006ea0, r=0) at librbd.cc:1231
1231 memcpy(buf + buf_bl_pos, data_bl.c_str() + bl_ofs, extent_len);
(gdb) bt
#0 0x00007f2c622a5b54 in librbd::AioBlockCompletion::complete
(this=0x7f2c5c006ea0, r=0) at librbd.cc:1231
#1 0x00007f2c622a5e66 in librbd::rados_aio_sparse_read_cb
(c=0x7f2c5c008520, arg=0x7f2c5c006ea0) at librbd.cc:1342
#2 0x00007f2c61e7d832 in
librados::RadosClient::C_aio_sparse_read_Ack::finish
(this=0x7f2c5c0060b0, r=0)
at librados.cc:470
#3 0x00007f2c61ea62af in Objecter::handle_osd_op_reply
(this=0x612bd0, m=0x614420) at osdc/Objecter.cc:801
#4 0x00007f2c61e58922 in librados::RadosClient::_dispatch
(this=0x60ebf0, m=0x614420) at librados.cc:751
#5 0x00007f2c61e589e6 in librados::RadosClient::ms_dispatch
(this=0x60ebf0, m=0x614420) at librados.cc:717
#6 0x00007f2c61e9bd77 in Messenger::ms_deliver_dispatch
(this=0x6102c0, m=0x614420) at msg/Messenger.h:98
#7 0x00007f2c61e82907 in SimpleMessenger::dispatch_entry
(this=0x6102c0) at msg/SimpleMessenger.cc:352
#8 0x00007f2c61e805f5 in SimpleMessenger::DispatchThread::entry
(this=0x610748) at ./msg/SimpleMessenger.h:533
#9 0x00007f2c61e93582 in Thread::_entry_func (arg=0x610748) at
common/Thread.h:41
#10 0x000000312be064a7 in start_thread () from /lib64/libpthread.so.0
#11 0x000000312b6d3c2d in clone () from /lib64/libc.so.6
(gdb) p buf
$6 = 0x7f2c583fe010<Address 0x7f2c583fe010 out of bounds>
(gdb)
=======================================
I guess releasing is too early for reading, is it right? Can anyone help me?
Your change shouldn't cause this crash -
librbd::AioCompletion::block_complete isn't called until the end of
librbd::AioBlockCompletion::complete.
If you're using the C++ interface, how are you initializing the
bufferlist you pass to Image::aio_read?
If you're using the C interface, make sure your beffer has enough space
for the length of your read.
Josh
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
