(Or, if we crashed before we got to write out the channel_update).
It's a corner case, but one reported by @darosior and reproduced
on my test node (both with bad gossip_store due to previous iterations
of this patchset!).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Triggered by a previous variant of this PR, but a goo1d idea to simply
discard the store in general when we get a duplicate entry.
We crash trying to delete old ones, which means writing to the store.
But they should have already been deleted.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This clarifies things a fair bit: we simply add and remove from the
gossip_store directly.
Before this series: (--disable-developer, -Og)
store_load_msec:20669-20902(20822.2+/-82)
vsz_kb:439704-439712(439706+/-3.2)
listnodes_sec:0.890000-1.000000(0.92+/-0.04)
listchannels_sec:11.960000-13.380000(12.576+/-0.49)
routing_sec:3.070000-5.970000(4.814+/-1.2)
peer_write_all_sec:28.490000-30.580000(29.532+/-0.78)
After: (--disable-developer, -Og)
store_load_msec:19722-20124(19921.6+/-1.4e+02)
vsz_kb:288320
listnodes_sec:0.860000-0.980000(0.912+/-0.056)
listchannels_sec:10.790000-12.260000(11.65+/-0.5)
routing_sec:2.540000-4.950000(4.262+/-0.88)
peer_write_all_sec:17.570000-19.500000(18.048+/-0.73)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We have a problem: if we get halfway through writing the compacted store
and run out of disk space, we've already changed half the indexes.
This changes it so we do nothing until writing is finished: then we
iterate through and update indexes. It also weans us off broadcast
ordering, which we can now eliminated.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We didn't count some records before, so we could compare the two counters.
This is much simpler, and avoids reliance on bs.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This means we intercept the peer's gossip_timestamp_filter request
in the per-peer subdaemon itself. The rest of the semantics are fairly
simple however.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
(We don't increment the gossip_store version, since there are only a
few commits since the last time we did this).
This lets the reader simply filter messages; this is especially nice since
the channel_announcement timestamp is *derived*, not in the actual message.
This also creates a 'struct gossip_hdr' which makes the code a bit
clearer.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Keeping the uintmap ordering all the broadcastable messages is expensive:
130MB for the million-channels project. But now we delete obsolete entries
from the store, we can have the per-peer daemons simply read that sequentially
and stream the gossip itself.
This is the most primitive version, where all gossip is streamed;
successive patches will bring back proper handling of timestamp filtering
and initial_routing_sync.
We add a gossip_state field to track what's happening with our gossip
streaming: it's initialized in gossipd, and currently always set, but
once we handle timestamps the per-peer daemon may do it when the first
filter is sent.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
There's a corner case where otherwise a reader could see the header and
not the body of a message. It could handle that in various ways,
but simplest (and most efficient) is to avoid it happening.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
They already send *us* gossip messages, so they have to be distinct anyway.
Why make us both do extra work?
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We use the high bit of the length field: this way we can still check
that the checksums are valid on deleted fields.
Once this is done, serially reading the gossip_store file will result
in a complete, ordered, minimal gossip broadcast. Also, the horrible
corner case where we might try to delete things from the store during
load time is completely gone: we only load non-deleted things.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We're about to bump version again, and the code to upgrade it was
quite hairy (and buggy!). It's not worthwhile for such a
poorly-tested path: I will just add code to limit how much incoming
gossip we get to avoid flooding when we upgrade, however.
I also use a modern gossip_store version in our test_gossip_store_load
test, instead of relying on the upgrade path.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
They're really gossipd-internal, and we don't want per-peer daemons
to confuse them with normal updates.
I don't bump the gossip_store version; that's coming with another update
anyway.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Now we handle node_announcements properly, we have a failure case where we
try to move them when a channel is deleted while loading the store.
We're going to remove this soon, in favor of in-place delete, so
workaround this for now to avoid an assert() when we try to write to
the store while loading.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
When we first receive a channel_update, we write both the
channel_announcement and that channel_update to the store: we need
that first update so we can set the channel_announcement timestamp.
However, the channel_update can be replaced later. This means we can
have a channel_announcement, a node_update which relies on it, then
the channel_update later.
So move the "this applies to a pending announcement" check lower, where
gossip_store can use it too. Has a nice side-effect of avoiding
one lookup of the node id.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
It turns out that we don't look at type when we return 0, but gcc isn't
quite smart enough for that. Initializing to -1 is good practice anyway
for the failure path.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Fix a path where tal_free is called on an uninitialized variable
If the first `goto bad_total` executes, then that path has
uninitialized `short_route` but bad_total passes through to `out`
whose first call is tal_free(short_route).
This was noticed by a maybe-uninitialized heuristic on gcc 7.4.0:
gossipd/routing.c: In function ‘find_shorter_route’:
gossipd/routing.c:1096:2: error: ‘short_route’ may be used
uninitialized in this function [-Werror=maybe-uninitialized]
tal_free(short_route);
Reported-by: @ZmnSCPxj <https://github.com/ElementsProject/lightning/pull/2674#issuecomment-495617253>
Signed-off-by: William Casarin <jb55@jb55.com>
Each destructor2 costs 40 bytes, and struct chan is only 120 bytes. So
this drops our memory usage quite a bit:
MCP bench results change:
-vsz_kb:580004-580016(580006+/-4.8)
+vsz_kb:533148
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This has two effects: most importantly, it avoids the problem where
lightningd creates a 800MB JSON blob in response to listchannels,
which causes OOM on the Raspberry Pi (our previous max allocation was
832MB). This is because lightning-cli can start draining the JSON
while we're filling the buffer, so we end up with a max allocation of
68MB.
But despite being less efficient (multiple queries to gossipd), it
actually speeds things up due to the parallelism:
MCP with -O3 -flto before vs after:
-listchannels_sec:8.980000-9.330000(9.206+/-0.14)
+listchannels_sec:7.500000-7.830000(7.656+/-0.11)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We now have a test blockchain for MCP which has the correct channels,
so this is not needed.
Also fix a benchmark script bug where 'mv "$DIR"/log
"$DIR"/log.old.$$' would fail if you log didn't exist from a previous run.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Instead of reading the store ourselves, we can just send them an
offset. This saves gossipd a lot of work, putting it where it belongs
(in the daemon responsible for the specific peer).
MCP bench results:
store_load_msec:28509-31001(29206.6+/-9.4e+02)
vsz_kb:580004-580016(580006+/-4.8)
store_rewrite_sec:11.640000-12.730000(11.908+/-0.41)
listnodes_sec:1.790000-1.880000(1.83+/-0.032)
listchannels_sec:21.180000-21.950000(21.476+/-0.27)
routing_sec:2.210000-11.160000(7.126+/-3.1)
peer_write_all_sec:36.270000-41.200000(38.168+/-1.9)
Signficant savings in streaming gossip:
-peer_write_all_sec:48.160000-51.480000(49.608+/-1.1)
+peer_write_all_sec:35.780000-37.980000(36.43+/-0.81)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We need to store the channel capacity for channel_announcement: hand it
in directly rather than having the gossip_store code do a lookup.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Save some overhead, plus gets us ready for giving subdaemons direct
store access. This is the first time we *upgrade* the gossip_store,
rather than just discarding.
The downside is that we need to add an extra message after each
channel_announcement, containing the channel capacity.
After:
store_load_msec:28337-30288(28975+/-7.4e+02)
vsz_kb:582304-582316(582306+/-4.8)
store_rewrite_sec:11.240000-11.800000(11.55+/-0.21)
listnodes_sec:1.800000-1.880000(1.84+/-0.028)
listchannels_sec:22.690000-26.260000(23.878+/-1.3)
routing_sec:2.280000-9.570000(6.842+/-2.8)
peer_write_all_sec:48.160000-51.480000(49.608+/-1.1)
Differences:
-vsz_kb:582320
+vsz_kb:582316
-listnodes_sec:2.100000-2.170000(2.118+/-0.026)
+listnodes_sec:1.800000-1.880000(1.84+/-0.028)
-peer_write_all_sec:51.600000-52.550000(52.188+/-0.34)
+peer_write_all_sec:48.160000-51.480000(49.608+/-1.1)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We know exactly how many there will be, so allocate an entire array up-front.
-listnodes_sec:2.540000-2.610000(2.584+/-0.029)
+listnodes_sec:2.100000-2.170000(2.118+/-0.026)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Entries we've already loaded expect to exist in the store. We could go
back and remove them all, but instead just truncate at the known-good
point.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Some tests require dev support, but the rest can run. We simplify
the gossip_store output so it's the same in non-dev mode too.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Now we can benchmark, and remove 500 bytes per node.
MCP results from 5 runs, min-max(mean +/- stddev):
store_load_msec:35093-37907(36146+/-1.1e+03)
vsz_kb:555168
store_rewrite_sec:12.120000-13.750000(12.7+/-0.6)
listnodes_sec:1.270000-1.370000(1.322+/-0.039)
listchannels_sec:29.770000-31.600000(30.82+/-0.64)
routing_sec:0.00
peer_write_all_sec:63.630000-67.850000(65.432+/-1.7)
MCP notable changes from pre-Dijkstra (>1 stddev):
-vsz_kb:577456
+vsz_kb:555168
-routing_sec:60.70
+routing_sec:12.04
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Do it inside the can_reach() function, which is less optimal for BFG
which does 20 ops on the same channel, but fine for Dijkstra.
This does have a measurable cost, so we might want to use
non-cryptographic fuzz in future:
$ gossipd/test/run-bench-find_route 100000 100:
Before:
100 (100 succeeded) routes in 100000 nodes in 97346 msec (973461784 nanoseconds per route)
After:
100 (100 succeeded) routes in 100000 nodes in 113381 msec (1133813412 nanoseconds per route)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
If a route is too long, we try to bias Dijkstra towards choosing a
shorter route by adding a per-hop cost. We do a naive "shortest path"
pass, then using that cost as a ceiling on per-hop cost, we do a
binary search.
There are some subtleties: we use risk rather than total as our
counter field (we normally bias this by 1 anyway, so it's easy to make
that a variable), and we set riskfactor to a mimimal value once we're
iterating. It's good enough to get a solution, we don't need to do a
2-dimensional search on riskfactor and riskbias.
Of course, this is extremely slow if we hit it on our benchmark,
though it doesn't happen in a more realistic network:
$ gossipd/test/run-bench-find_route 100000 100:
Before:
100 (79 succeeded) routes in 100000 nodes in 25341 msec (253412314 nanoseconds per route)
After:
100 (100 succeeded) routes in 100000 nodes in 97346 msec (973461784 nanoseconds per route)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Our uintmap can be a little slow with all the reallocation, so leave
NULL entries and walk to find the first one. Since we don't clean
them up, keep a cache of where the min non-all-NULL value is in the
heap.
It's clearer benefit on really large tests, so here's 1M nodes:
Comparison using gossipd/test/run-bench-find_route 1000000 10:
Before:
10 (10 succeeded) routes in 1000000 nodes in 91995 msec (9199532898 nanoseconds per route)
After:
10 (10 succeeded) routes in 1000000 nodes in 20605 msec (2060539287 nanoseconds per route)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Use a uintmap as our minheap.
Note that Dijkstra can give overlength routes, so some checks are disabled.
Comparison using gossipd/test/run-bench-find_route 100000 10:
Before:
10 (10 succeeded) routes in 100000 nodes in 120087 msec (12008708402 nanoseconds per route)
After:
10 (10 succeeded) routes in 100000 nodes in 2269 msec (226925462 nanoseconds per route)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This is a weakness with Dijkstra, so write an explicit unit test that
we can find a short enough (but more expensive) route.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
When we compact the store, we need to adjust the broadast index for
peers so they know where they're up to.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This requires some trickiness when we want to re-add unannounced channels
to the store after compaction, so we extract a common "copy_message" to
transfer from old store to new.
MCP results from 5 runs, min-max(mean +/- stddev):
store_load_msec:36034-37853(37109.8+/-5.9e+02)
vsz_kb:577456
store_rewrite_sec:12.490000-13.250000(12.862+/-0.27)
listnodes_sec:1.250000-1.480000(1.364+/-0.09)
listchannels_sec:30.820000-31.480000(31.068+/-0.24)
routing_sec:26.940000-27.990000(27.616+/-0.39)
peer_write_all_sec:65.690000-68.600000(66.698+/-0.99)
MCP notable changes from previous patch (>1 stddev):
-vsz_kb:1202316
+vsz_kb:577456
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Rusty Russell
William Casarin
darosior
trueptolemy