The CNode member nSendBytes is incremented under the node’s lock cs_vSend. However, RecordBytesSent is not. An RPC thread that acquires the cs_vSend lock puts the message handler thread on hold. While the thread is on hold, getnettotals returns “stale” values or values that don’t add up.
This can be fixed by making cs_vSend a “write lock” for the total bytes sent in connman.
The CNode member nSendBytes is incremented under the node's lock
cs_vSend. However, RecordBytesSent is not. An RPC thread that acquires
the cs_vSend lock puts the message handler thread on hold. While the
thread is on hold, getnettotals returns "stale" values or values that
don't add up.
This can be fixed by making cs_vSend a "write lock" for the total bytes
sent in connman.
After this commit, both calls to RecordBytesSent are done under the
LOCK(pnode->cs_vSend);
The following sections might be updated with supplementary metadata relevant to reviewers and maintainers.
Reviewers, this pull request conflicts with the following ones:
If you consider this pull request important, please also help to review the conflicting pull requests. Ideally, start with the one that should be merged first.
~ACK faae6ca9e8a3f67c737c03e122fd315cef5e66d7~
~Reviewed, ran tests.~
~This is a nice readability improvement too 📖~
Changed my mind, see below.
getnettotals returns “stale” values or values that don’t add up
Can you explain this a bit? What do the values not add up to? As far as I can tell, getnettotals simply returns the totalbytessent and totalbytesreceived at some point in time. There are not consistency guarantees with subsequent calls to the getpeerinfo values.
I’ve changed my mind on this one. getnettotals and getpeerinfo are two separate rpcs, we shouldn’t try and make them sync up. nTotalBytesSent is separate from a node’s nSendBytes. Looking at the analogous structures on the receiving side, nTotalBytesRecv is clearly separate from nRecvBytes.
However, I think a PR like this is needed to bring consistency between the two calls to RecordBytesSent (as it stands before this PR, one is under cs_vSend and the other is not). I can see a few more improvements available too.
Concept ACK.
What is the motivation to have the lock when calling got it.RecordBytesSent?
I’ve changed my mind on this one. getnettotals and getpeerinfo are two separate rpcs, we shouldn’t try and make them sync up. nTotalBytesSent is separate from a node’s nSendBytes. Looking at the analogous structures on the receiving side, nTotalBytesRecv is clearly separate from nRecvBytes.
Why are they separate? Let’s assume we connect to a peer, send 100 bytes, disconnect. Send 100 bytes to another peer, then the total bytes sent should simply equal 200 bytes.
Why are they separate? Let’s assume we connect to a peer, send 100 bytes, disconnect. Send 100 bytes to another peer, then the total bytes sent should simply equal 200 bytes.
Maybe I’m completely off-base here, I’m not trying to make a big statement.
Isn’t it impossible to keep these RPCs synced? They are two separate calls, something might happen in between them.
getnettotals, note the value of totalbytesrecvgetpeerinfo, see that the sum of bytesrecv for each node is 100 bytes more than totalbytesrecv from above
Ok, I see. The reason I created this pull was that a call to
0 net_totals_before = self.nodes[0].getnettotals()
1 peer_info = self.nodes[0].getpeerinfo()
2 net_totals_after = self.nodes[0].getnettotals()
and then asserting that before <= sum <= after would sometimes fail.
I haven’t looked into #19673 so I am wondering if #19673 also fixes that issue.
^
Woah that shouldn’t happen. Will investigate.
a call to
net_totals_before = self.nodes[0].getnettotals() peer_info = self.nodes[0].getpeerinfo() net_totals_after = self.nodes[0].getnettotals()and then asserting that before <= sum <= after would sometimes fail.`
What is sum here? Are you saying that getnettotals.totalbytessent can go down?
I don’t understand the problem you’re trying to solve.
Are you saying that getnettotals.totalbytessent can go down?
No, it is returning the value it returned in a previous call. But getpeerinfo already returned the new correct value.
I agree with Wlad’s assessment in that issue:
these are meant to be overall statistics, it doesn’t seem super important that the net totals are immediately up to date.
Ok, then the test is wrong and should be removed:
0diff --git a/test/functional/rpc_net.py b/test/functional/rpc_net.py
1index 03c858c694..afb67891af 100755
2--- a/test/functional/rpc_net.py
3+++ b/test/functional/rpc_net.py
4@@ -104,29 +104,15 @@ class NetTest(BitcoinTestFramework):
5
6 def test_getnettotals(self):
7 self.log.info("Test getnettotals")
8- # getnettotals totalbytesrecv and totalbytessent should be
9- # consistent with getpeerinfo. Since the RPC calls are not atomic,
10- # and messages might have been recvd or sent between RPC calls, call
11- # getnettotals before and after and verify that the returned values
12- # from getpeerinfo are bounded by those values.
13- net_totals_before = self.nodes[0].getnettotals()
14 peer_info = self.nodes[0].getpeerinfo()
15- net_totals_after = self.nodes[0].getnettotals()
16- assert_equal(len(peer_info), 2)
17- peers_recv = sum([peer['bytesrecv'] for peer in peer_info])
18- peers_sent = sum([peer['bytessent'] for peer in peer_info])
19-
20- assert_greater_than_or_equal(peers_recv, net_totals_before['totalbytesrecv'])
21- assert_greater_than_or_equal(net_totals_after['totalbytesrecv'], peers_recv)
22- assert_greater_than_or_equal(peers_sent, net_totals_before['totalbytessent'])
23- assert_greater_than_or_equal(net_totals_after['totalbytessent'], peers_sent)
24+ net_totals = self.nodes[0].getnettotals()
25
26 # test getnettotals and getpeerinfo by doing a ping
27 # the bytes sent/received should change
28 # note ping and pong are 32 bytes each
29 self.nodes[0].ping()
30- self.wait_until(lambda: (self.nodes[0].getnettotals()['totalbytessent'] >= net_totals_after['totalbytessent'] + 32 * 2), timeout=1)
31- self.wait_until(lambda: (self.nodes[0].getnettotals()['totalbytesrecv'] >= net_totals_after['totalbytesrecv'] + 32 * 2), timeout=1)
32+ self.wait_until(lambda: (self.nodes[0].getnettotals()['totalbytessent'] >= net_totals['totalbytessent'] + 32 * 2), timeout=1)
33+ self.wait_until(lambda: (self.nodes[0].getnettotals()['totalbytesrecv'] >= net_totals['totalbytesrecv'] + 32 * 2), timeout=1)
34
35 peer_info_after_ping = self.nodes[0].getpeerinfo()
36 for before, after in zip(peer_info, peer_info_after_ping):