]> git.ipfire.org Git - thirdparty/asterisk.git/commitdiff
bundled_pjproject: Avoid deadlock between transport and transaction master
authorStanislav Abramenkov <stas.abramenkov@gmail.com>
Tue, 1 Jul 2025 09:05:48 +0000 (12:05 +0300)
committergithub-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Thu, 3 Jul 2025 14:34:13 +0000 (14:34 +0000)
Backport patch from upstream
* Avoid deadlock between transport and transaction
https://github.com/pjsip/pjproject/commit/edde06f261ac

Issue described in
https://github.com/pjsip/pjproject/issues/4442

third-party/pjproject/patches/0010-Avoid_deadlock_between_transport_and_transaction.patch [new file with mode: 0644]

diff --git a/third-party/pjproject/patches/0010-Avoid_deadlock_between_transport_and_transaction.patch b/third-party/pjproject/patches/0010-Avoid_deadlock_between_transport_and_transaction.patch
new file mode 100644 (file)
index 0000000..e23af92
--- /dev/null
@@ -0,0 +1,158 @@
+From edde06f261ac807a89a6086b7f03460867675f95 Mon Sep 17 00:00:00 2001
+From: Nanang Izzuddin <nanang@teluu.com>
+Date: Tue, 1 Jul 2025 15:13:36 +0700
+Subject: [PATCH] Avoid deadlock between transport and transaction (#4453)
+
+---
+ pjsip/include/pjsip/sip_transaction.h |   1 +
+ pjsip/src/pjsip/sip_transaction.c     | 101 ++++++++++++++++++++++----
+ 2 files changed, 88 insertions(+), 14 deletions(-)
+
+diff --git a/pjsip/include/pjsip/sip_transaction.h b/pjsip/include/pjsip/sip_transaction.h
+index 72d4bc81c5..5f28b2d69c 100644
+--- a/pjsip/include/pjsip/sip_transaction.h
++++ b/pjsip/include/pjsip/sip_transaction.h
+@@ -141,6 +141,7 @@ struct pjsip_transaction
+     int                         retransmit_count;/**< Retransmission count. */
+     pj_timer_entry              retransmit_timer;/**< Retransmit timer.     */
+     pj_timer_entry              timeout_timer;  /**< Timeout timer.         */
++    pj_timer_entry              misc_timer;     /**< Miscellaneous timer.   */
+     /** Module specific data. */
+     void                       *mod_data[PJSIP_MAX_MODULE];
+diff --git a/pjsip/src/pjsip/sip_transaction.c b/pjsip/src/pjsip/sip_transaction.c
+index 4335f11ff5..31dbaaab6e 100644
+--- a/pjsip/src/pjsip/sip_transaction.c
++++ b/pjsip/src/pjsip/sip_transaction.c
+@@ -140,6 +140,7 @@ static int max_retrans_count = -1;
+ #define TRANSPORT_ERR_TIMER     3
+ #define TRANSPORT_DISC_TIMER    4
+ #define TERMINATE_TIMER         5
++#define TRANSPORT_CB_TIMER      6
+ /* Flags for tsx_set_state() */
+ enum
+@@ -2265,23 +2266,21 @@ static void send_msg_callback( pjsip_send_state *send_state,
+ }
+-/* Transport callback. */
+-static void transport_callback(void *token, pjsip_tx_data *tdata,
+-                               pj_ssize_t sent)
+-{
+-    pjsip_transaction *tsx = (pjsip_transaction*) token;
++/* Transport callback parameter. */
++struct tp_cb_param {
++    pjsip_transaction* tsx;
++    pjsip_tx_data* tdata;
++    pj_ssize_t sent;
++};
+-    /* Check if the transaction layer has been shutdown. */
+-    if (mod_tsx_layer.mod.id < 0)
+-        return;
+-    /* In other circumstances, locking tsx->grp_lock AFTER transport mutex
+-     * will introduce deadlock if another thread is currently sending a
+-     * SIP message to the transport. But this should be safe as there should
+-     * be no way this callback could be called while another thread is
+-     * sending a message.
+-     */
++/* Transport callback actual implementation. */
++static void transport_callback_impl(pjsip_transaction *tsx,
++                                    pjsip_tx_data* tdata,
++                                    pj_ssize_t sent)
++{
+     pj_grp_lock_acquire(tsx->grp_lock);
++
+     tsx->transport_flag &= ~(TSX_HAS_PENDING_TRANSPORT);
+     if (sent > 0 || tsx->role == PJSIP_ROLE_UAS) {
+@@ -2299,6 +2298,7 @@ static void transport_callback(void *token, pjsip_tx_data *tdata,
+             tsx_set_state( tsx, PJSIP_TSX_STATE_DESTROYED,
+                            PJSIP_EVENT_UNKNOWN, NULL, 0 );
+             pj_grp_lock_release(tsx->grp_lock);
++            pj_grp_lock_dec_ref(tsx->grp_lock);
+             return;
+         }
+@@ -2354,6 +2354,79 @@ static void transport_callback(void *token, pjsip_tx_data *tdata,
+ }
++/* Timer callback for transport callback.
++ * This is currently only used to avoid deadlock due to inversed locking order
++ * between transport and transaction.
++ */
++static void tsx_misc_timer_callback(pj_timer_heap_t *theap,
++                                    pj_timer_entry *entry)
++{
++    PJ_UNUSED_ARG(theap);
++
++    if (entry->id == TRANSPORT_CB_TIMER) {
++        struct tp_cb_param* param = (struct tp_cb_param*)entry->user_data;
++
++        /* Check if the transaction layer has been shutdown. */
++        if (mod_tsx_layer.mod.id >= 0) {
++            /* Call transport callback implementation */
++            transport_callback_impl(param->tsx, param->tdata, param->sent);
++        }
++
++        /* Release tdata */
++        pjsip_tx_data_dec_ref(param->tdata);
++    }
++}
++
++
++/* Transport callback. */
++static void transport_callback(void *token, pjsip_tx_data *tdata,
++                               pj_ssize_t sent)
++{
++    pjsip_transaction *tsx = (pjsip_transaction*) token;
++    pj_status_t status;
++
++    /* Check if the transaction layer has been shutdown. */
++    if (mod_tsx_layer.mod.id < 0)
++        return;
++
++    /* In other circumstances, locking tsx->grp_lock AFTER transport mutex
++     * will introduce deadlock if another thread is currently sending a
++     * SIP message to the transport. But this should be safe as there should
++     * be no way this callback could be called while another thread is
++     * sending a message.
++     */
++    // Deadlock does happen, see #4453.
++    // So now, to avoid deadlock, we'll try to acquire the group lock first,
++    // and if it fails, we'll schedule the processing via timer.
++    status = pj_grp_lock_tryacquire(tsx->grp_lock);
++    if (status != PJ_SUCCESS) {
++        pj_time_val delay = { 0, 0 };
++        struct tp_cb_param *param = NULL;
++
++        lock_timer(tsx);
++        tsx_cancel_timer(tsx, &tsx->misc_timer);
++
++        /* Increment tdata ref count to avoid premature destruction.
++         * Note that tsx ref count is already handled by tsx_schedule_timer().
++         */
++        pjsip_tx_data_add_ref(tdata);
++
++        param = PJ_POOL_ZALLOC_T(tsx->pool, struct tp_cb_param);
++        param->sent = sent;
++        param->tdata = tdata;
++        param->tsx = tsx;
++        pj_timer_entry_init(&tsx->misc_timer, TIMER_INACTIVE, param,
++                            &tsx_misc_timer_callback);
++        tsx_schedule_timer(tsx, &tsx->misc_timer, &delay, TRANSPORT_CB_TIMER);
++        unlock_timer(tsx);
++        return;
++    }
++
++    transport_callback_impl(tsx, tdata, sent);
++    pj_grp_lock_release(tsx->grp_lock);
++}
++
++
+ /*
+  * Callback when transport state changes.
+  */