Fix and detect race condition in clienttrace.go #5700
Conversation
dmathieu
added
the
blocked: CLA
instrumentation/net/http/httptrace/otelhttptrace/constants_test.go
Outdated
Show resolved
Hide resolved
instrumentation/net/http/httptrace/otelhttptrace/clienttrace_test.go
Outdated
Show resolved
Hide resolved
| if !ct.useSpans { | ||
| if err != nil { | ||
| attrs = append(attrs, attribute.String(hook+".error", err.Error())) | ||
| } | ||
| if ct.root == nil { |
There was a problem hiding this comment.
Can the conditions of this be tested?
There was a problem hiding this comment.
I'm not very familiar with this codebase, I just tried to fix a race condition that I could repro locally.
What should I test instead of checking whether ct.root is nil?
In my latest revision, I moved the test for ct.root being nil to the top of the function and am successfully exiting early if it is nil.
I'm not sure that's right. I'm operating under the assumption that most functions that call events on ct.root cannot do useful work until ct.root is set and should just move on.
instrumentation/net/http/httptrace/otelhttptrace/clienttrace.go
Outdated
Show resolved
Hide resolved
| for i := 1; i < workers; i++ { | ||
| wg.Add(1) | ||
| go func() { | ||
| resp, err := client.Get("https://example.com") |
There was a problem hiding this comment.
This shouldn use httptest, not make an external network call.
There was a problem hiding this comment.
This still won't work without an active network / if example.com is unavailable.
gregorynisbet-google
force-pushed
the
2024-05-31-fix-race-condition
branch
2 times, most recently
from
52345f4 to
eb1dc63
Compare
|
Friendly ping. |
ct.root can be updated without holding the mutex on ct. Also, addEvent can be called on a nil ct.root. Fix both of these issues by moving the mutex acquisition logic so that ct.root is only touched while the mutex is held.
gregorynisbet-google
force-pushed
the
2024-05-31-fix-race-condition
branch
from
4994805 to
647b6f3
Compare
MrAlias
removed
the
blocked: CLA
|
|
||
| // TestNewClientParallelismWithoutSubspans tests running many Gets on a client simultaneously, | ||
| // which would trigger a race condition if root were not protected by a mutex. | ||
| func TestNewClientParallelismWithoutSubspans(t *testing.T) { |
There was a problem hiding this comment.
This test passes when run against main:
$ go test -race -count=20 .
ok go.opentelemetry.io/contrib/instrumentation/net/http/httptrace/otelhttptrace 5.443s
It does not seem to be testing the fix being applied here.
|
|
||
| var wg sync.WaitGroup | ||
|
|
||
| for i := 1; i < 10000; i++ { |
There was a problem hiding this comment.
This is a lot of goroutines being spawned here. What's the reasoning behind this? It is making this test quite resource intensive.
| if ct.root == nil { | ||
| return | ||
| } |
There was a problem hiding this comment.
It looks like this is what is breaking
ct.root can be updated without holding the mutex on ct. Also, addEvent can be called on a nil ct.root.
Fix both of these issues by moving the mutex acquisition logic so that ct.root is only touched while the mutex is held.