aboutsummaryrefslogtreecommitdiffhomepage
path: root/src/core/lib/iomgr/ev_epollex_linux.cc
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/lib/iomgr/ev_epollex_linux.cc')
-rw-r--r--src/core/lib/iomgr/ev_epollex_linux.cc4
1 files changed, 4 insertions, 0 deletions
diff --git a/src/core/lib/iomgr/ev_epollex_linux.cc b/src/core/lib/iomgr/ev_epollex_linux.cc
index b082634af1..06a382c556 100644
--- a/src/core/lib/iomgr/ev_epollex_linux.cc
+++ b/src/core/lib/iomgr/ev_epollex_linux.cc
@@ -403,6 +403,10 @@ static void unref_by(grpc_fd* fd, int n) {
static void fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); }
static void fd_global_shutdown(void) {
+ // TODO(guantaol): We don't have a reasonable explanation about this
+ // lock()/unlock() pattern. It can be a valid barrier if there is at most one
+ // pending lock() at this point. Otherwise, there is still a possibility of
+ // use-after-free race. Need to reason about the code and/or clean it up.
gpr_mu_lock(&fd_freelist_mu);
gpr_mu_unlock(&fd_freelist_mu);
while (fd_freelist != nullptr) {