2 Commits (0d64cbde5756eaa1c7ee62116e112b7ebd198bbe)

Author	SHA1	Message	Date
George Joseph	be77b845d9	res_pjsip_pubsub: Correctly implement persisted subscriptions ... This patch fixes 2 original issues and more that those 2 exposed. * When we send a NOTIFY, and the client either doesn't respond or responds with a non OK, pjproject only calls our pubsub_on_evsub_state callback, no others. Since pubsub_on_evsub_state (which does the sub_tree cleanup) does not expect to be called back without the other callbacks being called first, it just returns leaving the sub_tree orphaned. Now pubsub_on_evsub_state checks the event for PJSIP_EVENT_TSX_STATE which is what pjproject will set to tell us that it was the transaction that timed out or failed and not the subscription itself timing our or being terminated by the client. If is TSX_STATE, pubsub_on_evsub_state now does the proper cleanup regardless of the state of the subscription. * When a client renews a subscription, we don't update the persisted subscription with the new expires timestamp. This causes subscription_persistence_recreate to prune the subscription if/when asterisk restarts. Now, pubsub_on_rx_refresh calls subscription_persistence_update to apply the new expires timestamp. This exposed other issues however... * When creating a dialog from rdata (which sub_persistence_recreate does from the packet buffer) there must NOT be a tag on the To header (which there will be when a client refreshes a subscription). If there is one, pjsip_dlg_create_uas will fail. To address this, subscription_persistence_update now accepts a flag that indicates that the original packet buffer must not be updated. New subscribes don't set the flag and renews do. This makes sure that when the rdata is recreated on asterisk startup, it's done from the original subscribe packet which won't have the tag on To. * When creating a dialog from rdata, we were setting the dialog's remote (SUBSCRIBE) cseq to be the same as the local (NOTIFY) cseq. When the client tried to resubscribe after a restart with the correct cseq, we'd reject the request with an Invalid CSeq error. * The acts of creating a dialog and evsub by themselves when recreating a subscription does NOT restart pjproject's subscription timer. The result was that even if we did correctly recreate the subscription, we never removed it if the client happened to go away or send a non-OK response to a NOTIFY. However, there is no pjproject function exposed to just set the timer on an evsub that wasn't created by an incoming subscribe request. To address this, we create our own timer using ast_sip_schedule_task. This timer is used only for re-establishing subscriptions after a restart. An earlier approach was to add support for setting pjproject's timer (via a pjproject patch) and while that patch is still included here, we don't use that call at the moment. While addressing these issues, additional debugging was added and some existing messages made more useful. A few formatting changes were also made to 'pjsip show scheduled tasks' to make displaying the subscription timers a little more friendly. ASTERISK-26696 ASTERISK-26756 Change-Id: I8c605fc1e3923f466a74db087d5ab6f90abce68e	9 years ago
George Joseph	9740277713	res_pjsip: Add serialized scheduler (res_pjsip/pjsip_scheduler.c) ... There are several places that do scheduled tasks or periodic housecleaning, each with its own implementation: * res_pjsip_keepalive has a thread that sends keepalives. * pjsip_distributor has a thread that cleans up expired unidentified requests. * res_pjsip_registrar_expire has a thread that cleans up expired contacts. * res_pjsip_pubsub uses ast_sched directly and then calls ast_sip_push_task. * res_pjsip_sdp_rtp also uses ast_sched to send keepalives. There are also places where we should be doing scheduled work but aren't. A good example are the places we have sorcery observers to start registration or qualify. These don't work when changes are made to a backend database without a pjsip reload. We need to check periodically. As a first step to solving these issues, a new ast_sip_sched facility has been created. ast_sip_sched wraps ast_sched but only uses ast_sched as a scheduled queue. When a task is ready to run, ast_sip_task_pusk is called for it. This ensures that the task is executed in a PJLIB registered thread and doesn't hold up the ast_sched thread so it can immediately continue processing the queue. The serializer used by ast_sip_sched is one of your choosing or a random one from the res_pjsip pool if you don't choose one. Another feature is the ability to automatically clean up the task_data when the task expires (if ever). If it's an ao2 object, it will be dereferenced, if it's a malloc'd object it will be freed. This is selectable when the task is scheduled. Even if you choose to not auto dereference an ao2 task data object, the scheduler itself maintains a reference to it while the task is under it's control. This prevents the data from disappearing out from under the task. There are two scheduling models. AST_SIP_SCHED_TASK_PERIODIC specifies that the invocations of the task occur at the specific interval. That is, every "interval" milliseconds, regardless of how long the task takes. If the task takes longer than the interval, it will be scheduled at the next available multiple of interval. For exmaple: If the task has an interval of 60 secs and the task takes 70 secs (it better not), the next invocation will happen at 120 seconds. AST_SIP_SCHED_TASK_DELAY specifies that the next invocation of the task should start "interval" milliseconds after the current invocation has finished. Also, the same ast_sched facility for fixed or variable intervals exists. The task's return code in conjunction with the AST_SIP_SCHED_TASK_FIXED or AST_SIP_SCHED_TASK_VARIABLE flags controls the next invocation start time. One res_pjsip.h housekeeping change was made. The pjsip header files were added to the top. There have been a few cases lately where I've needed res_pjsip.h just for ast_sip calls and had compiles fail spectacularly because I didn't add the pjsip header files to my source even though I never referenced any pjsip calls. Finally, a few new convenience APIs were added to astobj2 to make things a little easier in the scheduler. ao2_ref_and_lock() calls ao2_ref() and ao2_lock() in one go. ao2_unlock_and_unref() does the reverse. A few macros were also copied from res_phoneprov because I got tired of having to duplicate the same hash, sort and compare functions over and over again. The AO2_STRING_FIELD_(HASH|SORT|CMP)_FN macros will insert functions suitable for aor_container_alloc into your source. This facility can be used immediately for the situations where we already have a thread that wakes up periodically or do some scheduled work. For the registration and qualify issues, additional sorcery and schema changes would need to be made so that we can easily detect changed objects on a periodic basis without having to pull the entire database back to check. I'm thinking of a last-updated timestamp on the rows but more on this later. Change-Id: I7af6ad2b2d896ea68e478aa1ae201d6dd016ba1c	10 years ago