mirror of
https://github.com/Telecominfraproject/OpenCellular.git
synced 2026-01-14 16:46:23 +00:00
32b1e3add7
AP and EC consoles may generate a lot of bursty traffic, and cr50 UART
console to USB processing is very slow: when characters become
available, a hooks task callback is invoked, which retrieves received
characters one at a time and queues them up to the appropriate USB
transmit queue.
This patch speeds up things as follows:
- increases the seize of USB transmit queues for AP and EC console
channels to 512 bytes. Experiments supported by code
instrumentation has shown that even this is not enough to avoid
underruns, but this is a good compromise between memory use and
performance, these sizes could be revisited later,
- raises UART RX interrupt priority from level 1 to 0
- moving bytes from UART TX FIFO to USB queue happens on the
interrupt context when UART TX interrupt is asserted
- as many characters as possible are read from the UART first,
before queuing function is called, and the entire received batch
is passed to the queuing function.
It has to be mentioned here that presently batch processing is not
necessarily much more efficient, because queuing function becomes
more complicated when multiple objects are passed to it, this will
have to be dealt with in a separate patch.
There is still a lot of room for improvement:
- functions used to queue up data are very generic, dedicated code
could help a lot.
- UART drivers should have methods for collecting all bytes
available in receive FIFO in one invocation,
- USB side of things (dequeuing data and passing it to the
controller.
BRANCH=cr50, cr50mp
BUG=b:38448364
TEST=ran 'chargen' application on both AP and EC to flood the console
channels and observed the flow of characters on the host site, it
is pretty smooth with occasional hiccups, especially when TPM is
active, before this patch it was impossible to have both stream
up, both were garbled.
- Verified that new account can be created and user logged in on
restarts while chargen is running, i.e. TPM task gets enough
processing bandwidth.
- When EC is reset, there seem to be no lost characters on the
console (it used to cause some garbled console output before this
patch). The below output was collected on Coral:
> reboot
Rebooting!
--- UART initialized after reboot ---
[Reset cause: soft]
[Image: RO, coral_v1.1.8363+2cc945d5a 2018-05-15 17:41:57 ...
[0.003605 init buttons]
[0.003826 Inits done]
[0.004094 tablet mode disabled
]
[0.008272 found batt:SMP]
[0.022278 SW 0x01]
[0.042247 hash start 0x00040000 0x00021994]
[0.045823 Battery FET: reg 0x0018 mask 0x0018 disc 0x0000]
[0.071136 kblight registered]
[0.071544 PB init-on]
[0.071818 USB charge p0 m0]
[0.073670 ID/SKU ADC 4 = 1309 mV]
[0.075630 ID/SKU ADC 3 = 852 mV]
[0.076077 SKU ID: 71]
[0.076335 Motion Sensor Count = 3]
[0.083594 PD comm enabled]
...
- did not test bitbang programming mode, it is in line for
reworking for speeding up as well.
Change-Id: Ic9f3972f585dd1976169965c2a2422253aeac87a
Signed-off-by: Vadim Bendebury <vbendeb@chromium.org>
Reviewed-on: https://chromium-review.googlesource.com/1016037
Reviewed-by: Randall Spangler <rspangler@chromium.org>
Reviewed-by: Mary Ruthven <mruthven@chromium.org>