libwebsockets
Lightweight C library for HTML5 websockets
struct lws_sequencer introduction

Often a single network action like a client GET is just part of a larger series of actions, perhaps involving different connections.

Since lws operates inside an event loop, if the outer sequencing doesn't, it can be awkward to synchronize these steps with what's happening on the network with a particular connection on the event loop thread.

lws_sequencer

struct lws_sequencer provides a generic way to stage multi-step operations from inside the event loop. Because it participates in the event loop similar to a wsi, it always operates from the service thread context and can access structures that share the service thread without locking. It can also provide its own higher-level timeout handling.

Naturally you can have many of them running in the same event loop operating independently.

Sequencers themselves bind to a pt (per-thread) service thread, by default there's only one of these and it's the same as saying they bind to an lws_context. The sequencer callback may create wsi which in turn are bound to a vhost, but the sequencer itself is above all that.

Sequencer timeouts

The sequencer additionally maintains its own second-resolution timeout checked by lws for the step being sequenced... this is independent of any lws wsi timeouts which tend to be set and reset for very short-term timeout protection inside one transaction.

The sequencer timeout operates separately and above any wsi timeout, and is typically only reset by the sequencer callback when it receives an event indicating a step completed or failed, or it sets up the next sequence step.

If the sequencer timeout expires, then the sequencer receives a queued LWSSEQ_TIMED_OUT message informing it, and it can take corrective action or schedule a retry of the step. This message is queued and sent normally under the service thread context and in order of receipt.

Unlike lws timeouts which force the wsi to close, the sequencer timeout only sends the message. This allows the timeout to be used to, eg, wait out a retry cooloff period and then start the retry when the LWSSEQ_TIMED_OUT is received, according to the state of the sequencer.

Creating an <tt>struct lws_sequencer</tt>

typedef struct lws_seq_info {
struct lws_context *context; /* lws_context for seq */
int tsi; /* thread service idx */
size_t user_size; /* size of user alloc */
void **puser; /* place ptr to user */
lws_seq_event_cb cb; /* seq callback */
const char *name; /* seq name */
const lws_retry_bo_t *retry; /* retry policy */
} lws_seq_info_t;
struct lws_sequencer *
lws_sequencer_create(lws_seq_info_t *info);

When created, in lws the sequencer objects are bound to a 'per-thread', which is by default the same as to say bound to the lws_context. You can tag them with an opaque user data pointer, and they are also bound to a user-specified callback which handles sequencer events

typedef int (*lws_seq_event_cb)(struct lws_sequencer *seq, void *user_data,
lws_seq_events_t event, void *data);

struct lws_sequencer objects are private to lws and opaque to the user. A small set of apis lets you perform operations on the pointer returned by the create api.

Queueing events on a sequencer

Each sequencer object can be passed "events", which are held on a per-sequencer queue and handled strictly in the order they arrived on subsequent event loops. LWSSEQ_CREATED and LWSSEQ_DESTROYED events are produced by lws reflecting the sequencer's lifecycle, but otherwise the event indexes have a user-defined meaning and are queued on the sequencer by user code for eventual consumption by user code in the sequencer callback.

Pending events are removed from the sequencer queues and sent to the sequencer callback from inside the event loop at a rate of one per event loop wait.

Destroying sequencers

struct lws_sequencer objects are cleaned up during context destruction if they are still around.

Normally the sequencer callback receives a queued message that informs it that it's either failed at the current step, or succeeded and that was the last step, and requests that it should be destroyed by returning LWSSEQ_RET_DESTROY from the sequencer callback.

Lifecycle considerations

Sequencers may spawn additional assets like client wsi as part of the sequenced actions... the lifecycle of the sequencer and the assets overlap but do not necessarily depend on each other... that is a wsi created by the sequencer may outlive the sequencer.

It's important therefore to detach assets from the sequencer and the sequencer from the assets when each step is over and the asset is "out of scope" for the sequencer. It doesn't necessarily mean closing the assets, just making sure pointers are invalidated. For example, if a client wsi held a pointer to the sequencer as its .user_data, when the wsi is out of scope for the sequencer it can set it to NULL, eg, lws_set_wsi_user(wsi, NULL);.

Under some conditions wsi may want to hang around a bit to see if there is a subsequent client wsi transaction they can be reused on. They will clean themselves up when they time out.

Watching wsi lifecycle from a sequencer

When a sequencer is creating a wsi as part of its sequence, it will be very interested in lifecycle events. At client wsi creation time, the sequencer callback can set info->seq to itself in order to receive lifecycle messages about its wsi.

message meaning
LWSSEQ_WSI_CONNECTED The wsi has become connected
LWSSEQ_WSI_CONN_FAIL The wsi has failed to connect
LWSSEQ_WSI_CONN_CLOSE The wsi had been connected, but has now closed

By receiving these, the sequencer can understand when it should attempt reconnections or that it cannot progress the sequence.

When dealing with wsi that were created by the sequencer, they may close at any time, eg, be closed by the remote peer or an intermediary. The LWSSEQ_WSI_CONN_CLOSE message may have been queued but since they are strictly handled in the order they arrived, before it was handled an earlier message may want to cause some api to be called on the now-free()-d wsi. To detect this situation safely, there is a sequencer api lws_sequencer_check_wsi() which peeks the message buffer and returns nonzero if it later contains an LWSSEQ_WSI_CONN_CLOSE already.