The change to osv::run() (and friends) to reuse the calling thread rather
than start a new one to run a new shared-object introduced a new problem:

The new shared object may use pthread_key_create() to add destructors to
run when closing threads. There is (currently) a global list of these
destructors and once added they never get removed (even if the object is
unloaded). The danger is that osv::run() runs code which adds such a
destructors and allocates data (with pthread_setspecific()) and then the
program is unloaded - but the thread will only exit (and the destructors
run) later - and crash when trying to run the destructor pointing to
already-deleted code.

Previously we didn't have this problem, because osv::run() had a new thread,
which exited (and ran the destructors) *before* the reference counter on
the application dropped to 0 and the shared object was unloaded.

A hacky workaround added in this patch is to run in
application::start_and_join - used by osv::run, application::run_and_join
and osv_execve - a new function which runs all these destructors. This
assumes the thread starting these run functions did not have any thread-
specific data (it would be deleted too!). It's a hack, but should be good
enough for use cases like osv_execve() which anyway uses a fresh thread
for running the application.

Signed-off-by: Nadav Har'El <[email protected]>
---
 libc/pthread.hh |  4 ++++
 core/app.cc     |  9 +++++++++
 libc/pthread.cc | 38 ++++++++++++++++++++++++++------------
 3 files changed, 39 insertions(+), 12 deletions(-)

diff --git a/libc/pthread.hh b/libc/pthread.hh
index 53e9db9..1a4d02c 100644
--- a/libc/pthread.hh
+++ b/libc/pthread.hh
@@ -20,6 +20,10 @@ extern "C" {
 
 #ifdef __cplusplus
 }
+
+namespace pthread_private {
+void run_tsd_dtors();
+}
 #endif
 
 #endif /* PTHREAD_HH_ */
diff --git a/core/app.cc b/core/app.cc
index 56fd18b..ad3145f 100644
--- a/core/app.cc
+++ b/core/app.cc
@@ -17,6 +17,7 @@
 #include <algorithm>
 #include <boost/range/algorithm/transform.hpp>
 #include <osv/wait_record.hh>
+#include "libc/pthread.hh"
 
 using namespace boost::range;
 
@@ -251,6 +252,14 @@ void application::start_and_join(waiter* setup_waiter)
     }
     _thread = pthread_self(); // may be null if the caller is not a pthread.
     main();
+    // FIXME: run_tsd_dtors() is a hack - If the new program registered a
+    // destructor via pthread_key_create() and this thread keeps living with
+    // data for this key, the destructor function, part of the new program,
+    // may be called after the program is unloaded - and crash. Let's run
+    // the destructors now. This is wrong if the calling program has its own
+    // thread-local data. It is fine if the thread was created specially for
+    // running start_and_join (or run_and_join() or osv::run()).
+    pthread_private::run_tsd_dtors();
     sched::thread::current()->set_name(original_name);
     sched::thread::current()->set_app_runtime(original_app);
     original_app.reset();
diff --git a/libc/pthread.cc b/libc/pthread.cc
index ac7b3f0..62db224 100644
--- a/libc/pthread.cc
+++ b/libc/pthread.cc
@@ -41,27 +41,41 @@ namespace pthread_private {
     __thread pthread_t current_pthread;
     __thread int cancel_state = PTHREAD_CANCEL_ENABLE;
 
+    // NOTE: currently, the list of keys and destructor for each is global,
+    // not per shared object or ELF namespace. So if a shared object uses
+    // pthread_key_create() but doesn't call pthread_key_delete() before
+    // exiting, the key will be leaked. This is relatively harmless (beyond
+    // running out of keys) unless the shared object is unloaded before the
+    // thread exits and the destructors are run.
+    // As a *hack* you can call run_tsd_dtors() before unloading the object,
+    // but this will run all dtors, not just those belonging to the unloaded
+    // object, so this is only useful on a thread especially created for
+    // running the object.
     __attribute__ ((init_priority ((int)init_prio::pthread))) mutex 
tsd_key_mutex;
     __attribute__ ((init_priority ((int)init_prio::pthread))) std::vector<bool>
                                           tsd_used_keys(tsd_nkeys);
     __attribute__ ((init_priority ((int)init_prio::pthread)))
                   std::vector<void (*)(void*)> tsd_dtor(tsd_nkeys);
 
+    void run_tsd_dtors() {
+        bool done = false;
+        for (unsigned iter = 0; !done && iter < PTHREAD_DESTRUCTOR_ITERATIONS; 
++iter) {
+            done = true;
+            for (unsigned i = 0; i < tsd_nkeys; ++i) {
+                if (tsd[i] && tsd_dtor[i]) {
+                    void *val = tsd[i];
+                    tsd[i] = nullptr;
+                    tsd_dtor[i](val);
+                    done = false;
+                }
+            }
+        }
+    }
+
     void __attribute__((constructor)) pthread_register_tsd_dtor_notifier()
     {
         sched::thread::register_exit_notifier([] {
-            bool done = false;
-            for (unsigned iter = 0; !done && iter < 
PTHREAD_DESTRUCTOR_ITERATIONS; ++iter) {
-                done = true;
-                for (unsigned i = 0; i < tsd_nkeys; ++i) {
-                    if (tsd[i] && tsd_dtor[i]) {
-                        void *val = tsd[i];
-                        tsd[i] = nullptr;
-                        tsd_dtor[i](val);
-                        done = false;
-                    }
-                }
-            }
+            run_tsd_dtors();
         });
     }
 
-- 
2.7.4

-- 
You received this message because you are subscribed to the Google Groups "OSv 
Development" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
For more options, visit https://groups.google.com/d/optout.

Reply via email to