The API is simplified: do not expose the user code to several
queues. A "connection queue" is now an internal concept.
The user processes connections using the single function
lsquic_engine_process_conns(). When this function is called,
only those connections are processed that need to be processed.
A connection needs to be processed when:
1. New incoming packets have been fed to the connection.
2. User wants to read from a stream that is readable.
3. User wants to write to a stream that is writeable.
4. There are buffered packets that can be sent out. (This
means that the user wrote to a stream outside of the
lsquic library callback.)
5. A control frame (such as BLOCKED) needs to be sent out.
6. A stream needs to be serviced or delayed stream needs to
be created.
7. An alarm rings.
8. Pacer timer expires.
To achieve this, the library places the connections into two
priority queues (min heaps):
1. Tickable Queue; and
2. Advisory Tick Time queue (ATTQ).
Each time lsquic_engine_process_conns() is called, the Tickable
Queue is emptied. After the connections have been ticked, they are
queried again: if a connection is not being closed, it is placed
either in the Tickable Queue if it is ready to be ticked again or
it is placed in the Advisory Tick Time Queue. It is assumed that
a connection always has at least one timer set (the idle alarm).
The connections in the Tickable Queue are arranged in the least
recently ticked order. This lets connections that have been quiet
longer to get their packets scheduled first.
This change means that the library no longer needs to be ticked
periodically. The user code can query the library when is the
next tick event and schedule it exactly. When connections are
processed, only the tickable connections are processed, not *all*
the connections. When there are no tick events, it means that no
timer event is necessary -- only the file descriptor READ event
is active.
The following are improvements and simplifications that have
been triggered:
- Queue of connections with incoming packets is gone.
- "Pending Read/Write Events" Queue is gone (along with its
history and progress checks). This queue has become the
Tickable Queue.
- The connection hash no longer needs to track the connection
insertion order.
- [OPTIMIZATION] Merge series of ACKs if possible
Parsed single-range ACK frames (that is the majority of frames) are
saved in the connection and their processing is deferred until the
connection is ticked. If several ACKs come in a series between
adjacent ticks, we check whether the latest ACK is a strict superset
of the saved ACK. If it is, the older ACK is not processed.
If ACK frames can be merged, they are merged and only one of them is
either processed or saved.
- [OPTIMIZATION] Speed up ACK verification by simplifying send history.
Never generate a gap in the sent packet number sequence. This reduces
the send history to a single number instead of potentially a series of
packet ranges and thereby speeds up ACK verification.
By default, detecting a gap in the send history is not fatal: only a
single warning is generated per connection. The connection can continue
to operate even if the ACK verification code is not able to detect some
inconsistencies.
- [OPTIMIZATION] Rearrange the lsquic_send_ctl struct
The first part of struct lsquic_send_ctl now consists of members that
are used in lsquic_send_ctl_got_ack() (in the absense of packet loss,
which is the normal case). To speed up reads and writes, we no longer
try to save space by using 8- and 16-bit integers. Use regular integer
width for everything.
- [OPTIMIZATION] Cache size of sent packet.
- [OPTIMIZATION] Keep track of the largest ACKed in packet_out
Instead of parsing our own ACK frames when packet has been acked,
use the value saved in the packet_out structure when the ACK frame
was generated.
- [OPTIMIZATION] Take RTT sampling conditional out of ACK loop
- [OPTIMIZATION] ACK processing: only call clock_gettime() if needed
- [OPTIMIZATION] Several code-level optimizations to ACK processing.
- Fix: http_client: fix -I flag; switch assert() to abort()
- [API Change] lsquic_engine_connect() returns pointer to the connection
object.
- [API Change] Add lsquic_conn_get_engine() to get engine object from
connection object.
- [API Change] Add lsquic_conn_status() to query connection status.
- [API Change] Add add lsquic_conn_set_ctx().
- [API Change] Add new timestamp format, e.g. 2017-03-21 13:43:46.671345
- [OPTIMIZATION] Process handshake STREAM frames as soon as packet
arrives.
- [OPTIMIZATION] Do not compile expensive send controller sanity check
by default.
- [OPTIMIZATION] Add fast path to gquic_be_gen_reg_pkt_header.
- [OPTIMIZATION] Only make squeeze function call if necessary.
- [OPTIMIZATION] Speed up Q039 ACK frame parsing.
- [OPTIMIZATION] Fit most used elements of packet_out into first 64 bytes.
- [OPTIMIZATION] Keep track of scheduled bytes instead of calculating.
- [OPTIMIZATION] Prefetch next unacked packet when processing ACK.
- [OPTIMIZATION] Leverage fact that ACK ranges and unacked list are.
ordered.
- [OPTIMIZATION] Reduce function pointer use for STREAM frame generation
- Fix: reset incoming streams that arrive after we send GOAWAY.
- Fix: delay client on_new_conn() call until connection is fully set up.
- Fixes to buffered packets logic: splitting, STREAM frame elision.
- Fix: do not dispatch on_write callback if no packets are available.
- Fix WINDOW_UPDATE send and resend logic.
- Fix STREAM frame extension code.
- Fix: Drop unflushed data when stream is reset.
- Switch to tracking CWND using bytes rather than packets.
- Fix TCP friendly adjustment in cubic.
- Fix: do not generate invalid STOP_WAITING frames during high packet
loss.
- Pacer fixes.
- [API Change] Sendfile-like functionality is gone. The stream no
longer opens files and deals with file descriptors. (Among other
things, this makes the code more portable.) Three writing functions
are provided:
lsquic_stream_write
lsquic_stream_writev
lsquic_stream_writef (NEW)
lsquic_stream_writef() is given an abstract reader that has function
pointers for size() and read() functions which the user can implement.
This is the most flexible way. lsquic_stream_write() and
lsquic_stream_writev() are now both implemented as wrappers around
lsquic_stream_writef().
- [OPTIMIZATION] When writing to stream, be it within or without the
on_write() callback, place data directly into packet buffer,
bypassing auxiliary data structures. This reduces amount of memory
required, for the amount of data that can be written is limited
by the congestion window.
To support writes outside the on_write() callback, we keep N
outgoing packet buffers per connection which can be written to
by any stream. One half of these are reserved for the highest
priority stream(s), the other half for all other streams. This way,
low-priority streams cannot write instead of high-priority streams
and, on the other hand, low-priority streams get a chance to send
their packets out.
The algorithm is as follows:
- When user writes to stream outside of the callback:
- If this is the highest priority stream, place it onto the
reserved N/2 queue or fail.
(The actual size of this queue is dynamic -- MAX(N/2, CWND) --
rather than N/2, allowing high-priority streams to write as
much as can be sent.)
- If the stream is not the highest priority, try to place the
data onto the reserved N/2 queue or fail.
- When tick occurs *and* more packets can be scheduled:
- Transfer packets from the high N/2 queue to the scheduled
queue.
- If more scheduling is allowed:
- Call on_write callbacks for highest-priority streams,
placing resulting packets directly onto the scheduled queue.
- If more scheduling is allowed:
- Transfer packets from the low N/2 queue to the scheduled
queue.
- If more scheduling is allowed:
- Call on_write callbacks for non-highest-priority streams,
placing resulting packets directly onto the scheduled queue
The number N is currently 20, but it could be varied based on
resource usage.
- If stream is created due to incoming headers, make headers readable
from on_new.
- Outgoing packets are no longer marked non-writeable to prevent placing
more than one STREAM frame from the same stream into a single packet.
This property is maintained via code flow and an explicit check.
Packets for stream data are allocated using a special function.
- STREAM frame elision is cheaper, as we only perform it if a reset
stream has outgoing packets referencing it.
- lsquic_packet_out_t is smaller, as stream_rec elements are now
inside a union.
- Do not send RST_STREAM when stream is closed for reading
- Raise maximum header size from 4K to 64K
- Check header name and value lengths against maximum imposed by HPACK
- Fix NULL dereference in stream flow controller
- Hide handshake implementation behind a set of function pointers
- Use monotonically increasing clock
- Make sure that retx delay is not larger than the max of 60 seconds
