libbinder: fix buffer free race
Well, so the race is:
- client sends a large transaction (buffer A)
- server processes result
- server sends reply (1)
- client gets reply
- client sends another large transaction (buffer B)
- transaction fails, not enough space
- server frees buffer A (2)
This CL moves (2) to happen before (1). We set the Parcel size
to 0, which has the effect of freeing data, before the destructor
runs.
Test: binderLibTest
Test: binderLibTest --gtest_filter="*Garg*" --gtest_repeat=1000 --gtest_break_on_failure
Fixes: 238777741
Change-Id: Ic223a98c55904bb3f77ca13729cdf24a992cef1e
diff --git a/libs/binder/IPCThreadState.cpp b/libs/binder/IPCThreadState.cpp
index b50cfb3..bfcf39a 100644
--- a/libs/binder/IPCThreadState.cpp
+++ b/libs/binder/IPCThreadState.cpp
@@ -1318,6 +1318,13 @@
LOG_ONEWAY("Sending reply to %d!", mCallingPid);
if (error < NO_ERROR) reply.setError(error);
+ // b/238777741: clear buffer before we send the reply.
+ // Otherwise, there is a race where the client may
+ // receive the reply and send another transaction
+ // here and the space used by this transaction won't
+ // be freed for the client.
+ buffer.setDataSize(0);
+
constexpr uint32_t kForwardReplyFlags = TF_CLEAR_BUF;
sendReply(reply, (tr.flags & kForwardReplyFlags));
} else {
diff --git a/libs/binder/tests/binderLibTest.cpp b/libs/binder/tests/binderLibTest.cpp
index 5de08bd..6e1c8ac 100644
--- a/libs/binder/tests/binderLibTest.cpp
+++ b/libs/binder/tests/binderLibTest.cpp
@@ -1161,8 +1161,7 @@
// see ProcessState.cpp BINDER_VM_SIZE = 1MB.
// This value is not exposed, but some code in the framework relies on being able to use
// buffers near the cap size.
-// TODO(b/238777741): why do larger values, like 300K fail sometimes
-constexpr size_t kSizeBytesAlmostFull = 100'000;
+constexpr size_t kSizeBytesAlmostFull = 950'000;
constexpr size_t kSizeBytesOverFull = 1'050'000;
TEST_F(BinderLibTest, GargantuanVectorSent) {