The bigzaplock was used to prevent changes to all the channels during
masterspan processing while conference parameters were changing. The
master span processing also took a global lock out on the chan_lock to
prevent new channels from being registered or unregistered during
processing. Instead of grabbing both those locks, just changing the
semantics of chan_lock to those of bigzaplock seems to fit the bill and
removes another lock from the driver (and saves ~10 ns on a 2.40 GHz
Xeon in prcocess_masterspan)
chan_lock is a candidate for conversion to RCU but usage in conferencing
(bigzaplock type usage) needs to be audited carefully before conversion.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Kinsey Moore <kmoore@digium.com>
Acked-by: Russ Meyerriecks <rmeyerriecks@digium.com>
Acked-by: Tzafrir Cohen <tzafrir.cohen@xorcom.com>
Review: https://reviewboard.asterisk.org/r/940/
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9414 a0bf4364-ded3-4de4-8d8a-66a801d63aff
Very slight performance increase when timertest is running in the
background. Measured at ~10 ns improvement over 10K samples.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Kinsey Moore <kmoore@digium.com>
Acked-by: Russ Meyerriecks <rmeyerriecks@digium.com>
Acked-by: Tzafrir Cohen <tzafrir.cohen@xorcom.com>
Review: https://reviewboard.asterisk.org/r/940/
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9413 a0bf4364-ded3-4de4-8d8a-66a801d63aff
The check for DEFINE_SPINLOCK was spread throughout the source tree. If
not defined we can just define it in inlucde/dahdi/kernel.h. Now
include/dahdi/kernel.h is the only place that references
SPIN_LOCK_UNLOCKED (which breaks lockdep checking if DEFINE_SPINLOCK is
otherwise defined in the kernel).
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Kinsey Moore <kmoore@digium.com>
Acked-by: Russ Meyerriecks <rmeyerriecks@digium.com>
Acked-by: Tzafrir Cohen <tzafrir.cohen@xorcom.com>
Review: https://reviewboard.asterisk.org/r/940/
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9411 a0bf4364-ded3-4de4-8d8a-66a801d63aff
Now 'make -j 5' works the way it's supposed to on multi core machines.
On one test machine build times went from ~33 seconds to ~11 seconds.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Kinsey Moore <kmoore@digium.com>
Acked-by: Russ Meyerriecks <rmeyerriecks@digium.com>
Acked-by: Tzafrir Cohen <tzafrir.cohen@xorcom.com>
Review: https://reviewboard.asterisk.org/r/940/
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9409 a0bf4364-ded3-4de4-8d8a-66a801d63aff
'[[' is a bash construct specifically, yet #!/bin/sh is at the top of the
file.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Kinsey Moore <kmoore@digium.com>
Acked-by: Russ Meyerriecks <rmeyerriecks@digium.com>
Acked-by: Tzafrir Cohen <tzafrir.cohen@xorcom.com>
Review: https://reviewboard.asterisk.org/r/940/
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9408 a0bf4364-ded3-4de4-8d8a-66a801d63aff
When no hardware spans are configured DAHDI will use a kernel timer in order
provide timing for conferences. This is what dahdi_dummy historically was used
for. When kernel timers are used to provide timing DAHDI can both potentially
slow the rate at which time timer runs and also needs to account for how many
milliseconds of audio are processed for each 'tick' of process_masterspan.
The result is that if you are only using DAHDI for app_meetme, you can change
DAHDI_CHUNKSIZE to 40 (5ms) or 80 (10ms) bytes to reduce overhead since user
space is dealing with chunks of 20ms by default anyway.
NOTE: If you set this, you may still need to comment out the board drivers in
drivers/dahdi/Kbuild since they all do not support operating with a
DAHDI_CHUNKSIZE != 8 currently.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Kinsey Moore <kmoore@digium.com>
Acked-by: Russ Meyerriecks <rmeyerriecks@digium.com>
Acked-by: Tzafrir Cohen <tzafrir.cohen@xorcom.com>
Review: https://reviewboard.asterisk.org/r/940/
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9407 a0bf4364-ded3-4de4-8d8a-66a801d63aff
dahdi_receive, dahdi_transmit, and dahdi_ec_span are mostly called from
interrupt context anyway, so we can save a few cycles by not saving and
restoring the interrupt flags for every channel.
On one 2.40GHz Xeon test machine, for a span with 24 channels w/o echocan
enabled with ~10000 samples:
Function Avg Before Avg After
======================================
dahdi_receive 2.109 us 1.547 us
dahdi_transmit 3.203 us 2.766 us
dahdi_ec_span 0.416 us 0.454 us
NOTE: The time went up slightly on dahdi_ec_span since I did not have
software echocan enabled and this change calls local_irq_save regardless
in dahdi_ec_span. The slight increase in processing time in this case
is overshadowed by the savings in dahdi_receive and dahdi_transmit. If
echocan was enabled on all the channels there would be a time savings
in that dahdi_ec_span too.
When dahdi_receive/dahdi_transmit are called every millisecond (when
DAHDI_CHUNKSIZE == 8) this saves ~0.1% CPU time for each span.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Kinsey Moore <kmoore@digium.com>
Acked-by: Russ Meyerriecks <rmeyerriecks@digium.com>
Acked-by: Tzafrir Cohen <tzafrir.cohen@xorcom.com>
Review: https://reviewboard.asterisk.org/r/940/
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9406 a0bf4364-ded3-4de4-8d8a-66a801d63aff
The voicebus interface constantly uses the same buffers for moving data
to/from the card. A fixed pool has less overhead in this case than
constantly mapping/unmapping the kmem_cache_alloced buffers.
Saves about ~2.608 us for each run of vb_isr on a 2.40 GHz Xeon test
machine.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Russ Meyerriecks <rmeyerriecks@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9402 a0bf4364-ded3-4de4-8d8a-66a801d63aff
This change was made in revision 8176.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Russ Meyerriecks <rmeyerriecks@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9401 a0bf4364-ded3-4de4-8d8a-66a801d63aff
The registers on the device are already accessed with readl/writel and
the readchunk and writechunk are mapped into coherent DMA region. The
contents of those buffers should not be changing in the middle of any
transmit/receive prep call.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Russ Meyerriecks <rmeyerriecks@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9400 a0bf4364-ded3-4de4-8d8a-66a801d63aff
Normally you can see RBS bit states in dahdi_tool, but you might also
want them logged to dmesg for troubleshooting.
(issue #18025)
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Russ Meyerriecks <rmeyerriecks@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9399 a0bf4364-ded3-4de4-8d8a-66a801d63aff
The voicebus library by default configures the PCI interface to poll the
descriptor ring for available buffers. There are some platforms like
the Intel SG3420P motherboard where this does not appear to be
sufficient. Writing to the transmit demand poll register resolves this
problem on these troublesome platforms. DAHDI-700 DAHDI-702.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9397 a0bf4364-ded3-4de4-8d8a-66a801d63aff
If building within a git repository search the last log message for a
'git-svn-id'. If found, the commit has a corresponding svn revision
number and we will use the SVN-xxx-rxxx revision form. Otherwise use the
output of 'git describe --long --always --tags --dirty=M' as the
version.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Tzafrir Cohen <tzafrir.cohen@xorcom.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9396 a0bf4364-ded3-4de4-8d8a-66a801d63aff
Before CONFIG_DAHDI_PPP can be on by default, some more work needs to be
done to ensure that the ppp_generic module is not always loaded and
that all channels do not carry around all the PPP members unnecessarily.
(issue #17990)
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
Acked-by: Kevin P. Fleming <kpfleming@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9392 a0bf4364-ded3-4de4-8d8a-66a801d63aff
This module parameter will allow patgen/pattest to be used only on FXO
ports. *ALL* FXO ports will be placed in digital loopback mode when
set.
The current intent is for this to be removed as an optional module
parameter when there is a channel by channel representation in sysfs.
Otherwise, a new IOCTL would have to be defined and a tool written in
order to support this. DAHDI-696.
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9391 a0bf4364-ded3-4de4-8d8a-66a801d63aff
If we know that an open file is associated with a channel (pseudo or
real) we can just update the file->f_ops pointer so that we can short
circuit some of the checks in read/write/poll. Trades sizeof(file_operations)
bytes for less function call overhead in the "hot" path.
On a 2.4 GHz Xeon, saves around 150 ns on each read / write.
Review: https://reviewboard.asterisk.org/r/905/
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9386 a0bf4364-ded3-4de4-8d8a-66a801d63aff
Since we have the pseudo channels on their own list, we don't need to
step through the chans array when we just want to find the pseudo
channels. Likewise, all the real channels will have be on a span. Also
removes references to the global chans array in process_masterspan.
Review: https://reviewboard.asterisk.org/r/905/
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9384 a0bf4364-ded3-4de4-8d8a-66a801d63aff
Since pseudo channels are without spans and we would like to use the
span list to enumerate through all the channels, keeping the psuedo
channels on their own list is required. I believe this is a more
natural choice than making a dummy spans for pseudos since pseudo
channels should *really* just be an implementation detail that the user
shouldn't care about.
Review: https://reviewboard.asterisk.org/r/905/
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9383 a0bf4364-ded3-4de4-8d8a-66a801d63aff
This change also involves setting the DAHDI_FLAGBIT_CLOSE bit on the
channel before calling the 'close' callback in span ops. It appears
that only the dynamic spans may delete the channels on close, but that
code path doesn't check the DAHDI_FLAGBIT_CLOSE bit anyway.
Review: https://reviewboard.asterisk.org/r/905/
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9373 a0bf4364-ded3-4de4-8d8a-66a801d63aff
If someone was still using this...feel free to add it back in with a
comment about why it needs to be in the mainline. It looked better to
remove it than edit it to make sure it worked.
Review: https://reviewboard.asterisk.org/r/905/
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9370 a0bf4364-ded3-4de4-8d8a-66a801d63aff
This ioctl is called either via the "/dev/dahdi/channel" file or via a
/dev/dahdi/channo file. In either case, either the minor number of the
device file will match up with a channel, or the "private_data" member
on the channel will be set, so the unit is redundant.
Review: https://reviewboard.asterisk.org/r/905/
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9369 a0bf4364-ded3-4de4-8d8a-66a801d63aff
I believe this makes it a little more clear what is happening. Also could
provide a potential speedup if the span doesn't need to be constantly
dereferenced (although an optimizer might make this moot). Regardless, this
change doesn't hurt.
Review: https://reviewboard.asterisk.org/r/905/
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9362 a0bf4364-ded3-4de4-8d8a-66a801d63aff
In __dahdi_process_getaudio_chunk and __dahdi_process_putaudio_chunk we
can streamline the operation slightly by saving a constant pointer to
the conference channel instead of constantly dereferencing the master
channel and indexing into the chan array.
Review: https://reviewboard.asterisk.org/r/905/
Signed-off-by: Shaun Ruffell <sruffell@digium.com>
git-svn-id: http://svn.asterisk.org/svn/dahdi/linux/trunk@9361 a0bf4364-ded3-4de4-8d8a-66a801d63aff
