thebestvibedev/tromcho.net
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix crashes on multiple OSD reconnects

Browse Source 
Identify clients by pointers instead of peer_fd as peer may be dropped
and reconnected between callbacks

Yeah maybe I need some Rust, but ... maybe in the future :)
This commit is contained in:
Vitaliy Filippov
2020-10-17 10:53:04 +00:00
parent 9350656af6
commit 776fe954a5
6 changed files with 258 additions and 240 deletions
Download Patch File Download Diff File Expand all files Collapse all files
msgr_send.cpp
+82 -77
View File
@@ -6,7 +6,7 @@
void osd_messenger_t::outbox_push(osd_op_t *cur_op)
{
assert(cur_op->peer_fd);
auto & cl = clients.at(cur_op->peer_fd);
osd_client_t *cl = clients.at(cur_op->peer_fd);
if (cur_op->op_type == OSD_OP_OUT)
{
clock_gettime(CLOCK_REALTIME, &cur_op->tv_begin);
@@ -15,12 +15,12 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
{
// Check that operation actually belongs to this client
bool found = false;
for (auto it = cl.received_ops.begin(); it != cl.received_ops.end(); it++)
for (auto it = cl->received_ops.begin(); it != cl->received_ops.end(); it++)
{
if (*it == cur_op)
{
found = true;
cl.received_ops.erase(it, it+1);
cl->received_ops.erase(it, it+1);
break;
}
}
@@ -30,85 +30,86 @@ void osd_messenger_t::outbox_push(osd_op_t *cur_op)
return;
}
}
cl.outbox.push_back(cur_op);
cl->outbox.push_back(cur_op);
if (!ringloop)
{
while (cl.write_op || cl.outbox.size())
while (cl->write_op || cl->outbox.size())
{
try_send(cl);
}
}
else if (cl.write_op || cl.outbox.size() > 1 || !try_send(cl))
else if (cl->write_op || cl->outbox.size() > 1 || !try_send(cl))
{
if (cl.write_state == 0)
if (cl->write_state == 0)
{
cl.write_state = CL_WRITE_READY;
cl->write_state = CL_WRITE_READY;
write_ready_clients.push_back(cur_op->peer_fd);
}
ringloop->wakeup();
}
}
bool osd_messenger_t::try_send(osd_client_t & cl)
bool osd_messenger_t::try_send(osd_client_t *cl)
{
int peer_fd = cl.peer_fd;
if (!cl.write_op)
int peer_fd = cl->peer_fd;
if (!cl->write_op)
{
// pick next command
cl.write_op = cl.outbox.front();
cl.outbox.pop_front();
cl.write_state = CL_WRITE_REPLY;
if (cl.write_op->op_type == OSD_OP_IN)
cl->write_op = cl->outbox.front();
cl->outbox.pop_front();
cl->write_state = CL_WRITE_REPLY;
if (cl->write_op->op_type == OSD_OP_IN)
{
// Measure execution latency
timespec tv_end;
clock_gettime(CLOCK_REALTIME, &tv_end);
stats.op_stat_count[cl.write_op->req.hdr.opcode]++;
if (!stats.op_stat_count[cl.write_op->req.hdr.opcode])
stats.op_stat_count[cl->write_op->req.hdr.opcode]++;
if (!stats.op_stat_count[cl->write_op->req.hdr.opcode])
{
stats.op_stat_count[cl.write_op->req.hdr.opcode]++;
stats.op_stat_sum[cl.write_op->req.hdr.opcode] = 0;
stats.op_stat_bytes[cl.write_op->req.hdr.opcode] = 0;
stats.op_stat_count[cl->write_op->req.hdr.opcode]++;
stats.op_stat_sum[cl->write_op->req.hdr.opcode] = 0;
stats.op_stat_bytes[cl->write_op->req.hdr.opcode] = 0;
}
stats.op_stat_sum[cl.write_op->req.hdr.opcode] += (
(tv_end.tv_sec - cl.write_op->tv_begin.tv_sec)*1000000 +
(tv_end.tv_nsec - cl.write_op->tv_begin.tv_nsec)/1000
stats.op_stat_sum[cl->write_op->req.hdr.opcode] += (
(tv_end.tv_sec - cl->write_op->tv_begin.tv_sec)*1000000 +
(tv_end.tv_nsec - cl->write_op->tv_begin.tv_nsec)/1000
);
if (cl.write_op->req.hdr.opcode == OSD_OP_READ ||
cl.write_op->req.hdr.opcode == OSD_OP_WRITE)
if (cl->write_op->req.hdr.opcode == OSD_OP_READ ||
cl->write_op->req.hdr.opcode == OSD_OP_WRITE)
{
stats.op_stat_bytes[cl.write_op->req.hdr.opcode] += cl.write_op->req.rw.len;
stats.op_stat_bytes[cl->write_op->req.hdr.opcode] += cl->write_op->req.rw.len;
}
else if (cl.write_op->req.hdr.opcode == OSD_OP_SEC_READ ||
cl.write_op->req.hdr.opcode == OSD_OP_SEC_WRITE ||
cl.write_op->req.hdr.opcode == OSD_OP_SEC_WRITE_STABLE)
else if (cl->write_op->req.hdr.opcode == OSD_OP_SEC_READ ||
cl->write_op->req.hdr.opcode == OSD_OP_SEC_WRITE ||
cl->write_op->req.hdr.opcode == OSD_OP_SEC_WRITE_STABLE)
{
stats.op_stat_bytes[cl.write_op->req.hdr.opcode] += cl.write_op->req.sec_rw.len;
stats.op_stat_bytes[cl->write_op->req.hdr.opcode] += cl->write_op->req.sec_rw.len;
}
cl.send_list.push_back(cl.write_op->reply.buf, OSD_PACKET_SIZE);
if (cl.write_op->req.hdr.opcode == OSD_OP_READ ||
cl.write_op->req.hdr.opcode == OSD_OP_SEC_READ ||
cl.write_op->req.hdr.opcode == OSD_OP_SEC_LIST ||
cl.write_op->req.hdr.opcode == OSD_OP_SHOW_CONFIG)
cl->send_list.push_back(cl->write_op->reply.buf, OSD_PACKET_SIZE);
if (cl->write_op->req.hdr.opcode == OSD_OP_READ ||
cl->write_op->req.hdr.opcode == OSD_OP_SEC_READ ||
cl->write_op->req.hdr.opcode == OSD_OP_SEC_LIST ||
cl->write_op->req.hdr.opcode == OSD_OP_SHOW_CONFIG)
{
cl.send_list.append(cl.write_op->iov);
cl->send_list.append(cl->write_op->iov);
}
}
else
{
cl.send_list.push_back(cl.write_op->req.buf, OSD_PACKET_SIZE);
if (cl.write_op->req.hdr.opcode == OSD_OP_WRITE ||
cl.write_op->req.hdr.opcode == OSD_OP_SEC_WRITE ||
cl.write_op->req.hdr.opcode == OSD_OP_SEC_WRITE_STABLE ||
cl.write_op->req.hdr.opcode == OSD_OP_SEC_STABILIZE ||
cl.write_op->req.hdr.opcode == OSD_OP_SEC_ROLLBACK)
cl->send_list.push_back(cl->write_op->req.buf, OSD_PACKET_SIZE);
if (cl->write_op->req.hdr.opcode == OSD_OP_WRITE ||
cl->write_op->req.hdr.opcode == OSD_OP_SEC_WRITE ||
cl->write_op->req.hdr.opcode == OSD_OP_SEC_WRITE_STABLE ||
cl->write_op->req.hdr.opcode == OSD_OP_SEC_STABILIZE ||
cl->write_op->req.hdr.opcode == OSD_OP_SEC_ROLLBACK)
{
cl.send_list.append(cl.write_op->iov);
cl->send_list.append(cl->write_op->iov);
}
}
}
cl.write_msg.msg_iov = cl.send_list.get_iovec();
cl.write_msg.msg_iovlen = cl.send_list.get_size();
cl->write_msg.msg_iov = cl->send_list.get_iovec();
cl->write_msg.msg_iovlen = cl->send_list.get_size();
cl->refs++;
if (ringloop && !use_sync_send_recv)
{
io_uring_sqe* sqe = ringloop->get_sqe();
@@ -117,17 +118,17 @@ bool osd_messenger_t::try_send(osd_client_t & cl)
return false;
}
ring_data_t* data = ((ring_data_t*)sqe->user_data);
data->callback = [this, peer_fd](ring_data_t *data) { handle_send(data->res, peer_fd); };
my_uring_prep_sendmsg(sqe, peer_fd, &cl.write_msg, 0);
data->callback = [this, cl](ring_data_t *data) { handle_send(data->res, cl); };
my_uring_prep_sendmsg(sqe, peer_fd, &cl->write_msg, 0);
}
else
{
int result = sendmsg(peer_fd, &cl.write_msg, MSG_NOSIGNAL);
int result = sendmsg(peer_fd, &cl->write_msg, MSG_NOSIGNAL);
if (result < 0)
{
result = -errno;
}
handle_send(result, peer_fd);
handle_send(result, cl);
}
return true;
}
@@ -146,41 +147,45 @@ void osd_messenger_t::send_replies()
write_ready_clients.clear();
}
void osd_messenger_t::handle_send(int result, int peer_fd)
void osd_messenger_t::handle_send(int result, osd_client_t *cl)
{
auto cl_it = clients.find(peer_fd);
if (cl_it != clients.end())
cl->refs--;
if (cl->peer_state == PEER_STOPPED)
{
auto & cl = cl_it->second;
if (result < 0 && result != -EAGAIN)
if (!cl->refs)
{
// this is a client socket, so don't panic. just disconnect it
printf("Client %d socket write error: %d (%s). Disconnecting client\n", peer_fd, -result, strerror(-result));
stop_client(peer_fd);
return;
delete cl;
}
if (result >= 0)
return;
}
if (result < 0 && result != -EAGAIN)
{
// this is a client socket, so don't panic. just disconnect it
printf("Client %d socket write error: %d (%s). Disconnecting client\n", cl->peer_fd, -result, strerror(-result));
stop_client(cl->peer_fd);
return;
}
if (result >= 0)
{
cl->send_list.eat(result);
if (cl->send_list.done >= cl->send_list.count)
{
cl.send_list.eat(result);
if (cl.send_list.done >= cl.send_list.count)
// Done
cl->send_list.reset();
if (cl->write_op->op_type == OSD_OP_IN)
{
// Done
cl.send_list.reset();
if (cl.write_op->op_type == OSD_OP_IN)
{
delete cl.write_op;
}
else
{
cl.sent_ops[cl.write_op->req.hdr.id] = cl.write_op;
}
cl.write_op = NULL;
cl.write_state = cl.outbox.size() > 0 ? CL_WRITE_READY : 0;
delete cl->write_op;
}
}
if (cl.write_state != 0)
{
write_ready_clients.push_back(peer_fd);
else
{
cl->sent_ops[cl->write_op->req.hdr.id] = cl->write_op;
}
cl->write_op = NULL;
cl->write_state = cl->outbox.size() > 0 ? CL_WRITE_READY : 0;
}
}
if (cl->write_state != 0)
{
write_ready_clients.push_back(cl->peer_fd);
}
}