[pulseaudio-discuss] [PATCH RFCv2 00/27] memory optimizations
Peter Meerwald
pmeerw at pmeerw.net
Tue Oct 28 12:46:13 PDT 2014
this patch series aims to save memory allocations and some system calls
related to PA's client/server protocol implementation
v2 incorporates David's and Alexander's suggestions; no functional change to
the first 17 patches; v2 then adds some cleanup and two optimizations
patches 1 to 5 ('tagstruct:') introduce a new tagstruct type _APPENDED
which can hold tagstruct data up to a certain size; tagstructs are now
kept in a specific free-list -- this typically replaces two malloc()/free()s
with one flist push()/pop()
patches 6 to 8 ('packet:') make packets fixed-size (typically); packets are
kept in a specific free-list -- this replaces one malloc()/free() with one
flist push()/pop()
patches 9 to 14 ('pstream:') allows to send tagstructs directly to a pstream
without encapsulation in a packet -- this saves one flist push()/pop()
patches 15 and 16 ('pstream') often save a read() call by reading more than
just the descriptor (up to 40 bytes, e.g. description (20 bytes) + shm
info (16 bytes)); the idea is similar to b4342845d, "Optimize write
of smaller packages", but for read -- this trades some extra memcpy() for
a read()
patch 17 ('iochannel') fixes a strange behaviour in iochannel/mainloop that
deleted the input_event with every read which caused a rebuild of the pollfds
for every read()!
v2 material:
patches 18 to 20 ('queue', 'pstream') aim to combine two write items into one
minibuffer by peeking ahead in the send queue
patch 21 inlines pa_run_once() as this function came out high in profiling
patch 25 ('mainloop') only clears the wakeup pipe when poll() indicates that
the pipe is readable; if the only ready file descriptor is the wakeup pipe,
searching io_events can be avoided
patch 26 and 27 ('flish') removes the volatile annotation and makes flist_elem attributes
non-atomic -- needed?
with these patches typical playback (i.e. after setup) runs without any malloc()/free()
thanks to the use of free-lists; the number of memory management operations is reduced
no benchmarking yet, on i7 64-bit 'paplay --latency-msec=10 48KHz.wav" improved from
around 12% CPU to around 10% CPU for me; I plan to benchmark this on ARM soonish
Peter Meerwald (27):
tagstruct: Distinguish pa_tagstruct_new() use cases
tagstruct: Replace dynamic flag with type
tagstruct: Get rid of pa_tagstruct_free_data()
tagstruct: Add type _APPENDED
tagstruct: Use flist to potentially save calls to malloc()/free()
packet: Hide internals of pa_packet, introduce pa_packet_data()
packet: Make pa_packet_new() create fixed-size packets
packet: Introduce pa_packet_new_data() to copy data into a newly
created packet
packet: Use flist to save calls to malloc()/free()
pstream: Unionize item_info
pstream: Add pa_pstream_send_tagstruct()
pstream: #define PA_PSTREAM_SHM_SIZE
pstream: Duplicate assignment, write.data is always NULL
pstream: Only reset memchunk if it has been used
pstream: Split up do_read()
pstream: Use small minibuffer to combine several read()s if possible
iochannel: Fix channel enable
queue: Add pa_queue_peek() function
pstream: Add helper functions reset_descriptor(), shm_descriptor()
pstream: Peek into next item on send queue to see if it can be put
into minibuffer together with current item
once: Inline functions
rtpoll: typo
rtpoll: Fix condition for DEBUG_TIMING output
rtpoll: Drop extra wait_op argument to pa_rtpoll_run()
mainloop: Clear wakeup pipe only when necessary
flist: Don't use atomic operations to manipulate ptr, next
flist: Don't make flist volatile
src/modules/alsa/alsa-sink.c | 2 +-
src/modules/alsa/alsa-source.c | 2 +-
src/modules/module-card-restore.c | 4 +-
src/modules/module-combine-sink.c | 2 +-
src/modules/module-device-manager.c | 12 +-
src/modules/module-device-restore.c | 16 +-
src/modules/module-esound-sink.c | 2 +-
src/modules/module-null-sink.c | 2 +-
src/modules/module-null-source.c | 2 +-
src/modules/module-pipe-sink.c | 2 +-
src/modules/module-pipe-source.c | 2 +-
src/modules/module-sine-source.c | 2 +-
src/modules/module-stream-restore.c | 12 +-
src/modules/module-tunnel.c | 54 +-
src/modules/oss/module-oss.c | 2 +-
src/modules/raop/module-raop-sink.c | 2 +-
src/pulse/context.c | 26 +-
src/pulse/ext-device-manager.c | 14 +-
src/pulse/ext-device-restore.c | 10 +-
src/pulse/ext-stream-restore.c | 10 +-
src/pulse/introspect.c | 82 +--
src/pulse/mainloop.c | 24 +-
src/pulse/scache.c | 10 +-
src/pulse/stream.c | 22 +-
src/pulse/subscribe.c | 2 +-
src/pulsecore/flist.c | 14 +-
src/pulsecore/flist.h | 2 +-
src/pulsecore/iochannel.c | 35 +-
src/pulsecore/once.c | 18 +-
src/pulsecore/once.h | 27 +-
src/pulsecore/packet.c | 55 +-
src/pulsecore/packet.h | 20 +-
src/pulsecore/pdispatch.c | 9 +-
src/pulsecore/protocol-native.c | 94 ++--
src/pulsecore/pstream-util.c | 33 +-
src/pulsecore/pstream-util.h | 2 -
src/pulsecore/pstream.c | 728 +++++++++++++++++----------
src/pulsecore/pstream.h | 2 +
src/pulsecore/queue.c | 11 +
src/pulsecore/queue.h | 3 +
src/pulsecore/rtpoll.c | 14 +-
src/pulsecore/rtpoll.h | 7 +-
src/pulsecore/tagstruct.c | 67 ++-
src/pulsecore/tagstruct.h | 4 +-
src/tests/rtpoll-test.c | 4 +-
src/tests/srbchannel-test.c | 21 +-
47 files changed, 919 insertions(+), 616 deletions(-)
--
1.9.1
More information about the pulseaudio-discuss
mailing list