This is the v2 of the patchset from July:
https://lists.openwall.net/netdev/2019/07/07/81
Changes:
- Adapted the taprio offload patch to work by specifying "flags 2" to
the iproute2-next tc. At the moment I don't clearly understand whether
the full offload and the txtime assist ("flags 1") are mutually
exclusive or not (i.e. whether a "flags 3" mode should be rejected,
which it currently isn't).
- Added reference counting to the taprio offload structure. Maybe the
function names and placement could have been better though. As for the
other complaint (cycle time calculation) it got fixed in the taprio
parser in the meantime.
- Converted sja1105 to use the hardware PTP registers, and save/restore
the PTP time across resets.
- Made the DSA callback for ndo_setup_tc a bit more generic, but I don't
know whether it fulfills expectations. Drivers still can't do blocking
operations in its execution context.
- Added a state machine for starting/stopping the scheduler based on the
last command run on the PTP clock.
For those who want to follow along with the hardware implementation, the
manual is here:
https://www.nxp.com/docs/en/user-guide/UM10944.pdf
Original cover letter:
Using Vinicius Costa Gomes' configuration interface for 802.1Qbv (later
resent by Voon Weifeng for the stmmac driver), I am submitting for
review a draft implementation of this offload for a DSA switch.
I don't want to insist too much on the hardware specifics of SJA1105
which isn't otherwise very compliant to the IEEE spec.
In order to be able to test with Vedang Patel's iproute2 patch for
taprio offload (https://www.spinics.net/lists/netdev/msg573072.html)
I had to actually revert the txtime-assist branch as it had changed the
iproute2 interface.
In terms of impact for DSA drivers, I would like to point out that:
- Maybe somebody should pre-populate qopt->cycle_time in case the user
does not provide one. Otherwise each driver needs to iterate over the
GCL once, just to set the cycle time (right now stmmac does as well).
- Configuring the switch over SPI cannot apparently be done from this
ndo_setup_tc callback because it runs in atomic context. I also have
some downstream patches to offload tc clsact matchall with mirred
action, but in that case it looks like the atomic context restriction
does not apply.
- I had to copy the struct tc_taprio_qopt_offload to driver private
memory because a static config needs to be constructed every time a
change takes place, and there are up to 4 switch ports that may take a
TAS configuration. I have created a private
tc_taprio_qopt_offload_copy() helper for this - I don't know whether
it's of any help in the general case.
There is more to be done however. The TAS needs to be integrated with
the PTP driver. This is because with a PTP clock source, the base time
is written dynamically to the PTPSCHTM (PTP schedule time) register and
must be a time in the future. Then the "real" base time of each port's
TAS config can be offset by at most ~50 ms (the DELTA field from the
Schedule Entry Points Table) relative to PTPSCHTM.
Because base times in the past are completely ignored by this hardware,
we need to decide if it's ok behaviorally for a driver to "roll" a past
base time into the immediate future by incrementally adding the cycle
time (so the phase doesn't change). If it is, then decide by how long in
the future it is ok to do so. Or alternatively, is it preferable if the
driver errors out if the user-supplied base time is in the past and the
hardware doesn't like it? But even then, there might be fringe cases
when the base time becomes a past PTP time right as the driver tries to
apply the config.
Also applying a tc-taprio offload to a second SJA1105 switch port will
inevitably need to roll the first port's (now past) base time into an
equivalent future time.
All of this is going to be complicated even further by the fact that
resetting the switch (to apply the tc-taprio offload) makes it reset its
PTP time.
Vinicius Costa Gomes (1):
taprio: Add support for hardware offloading
Vladimir Oltean (14):
net: dsa: sja1105: Change the PTP command access pattern
net: dsa: sja1105: Get rid of global declaration of struct
ptp_clock_info
net: dsa: sja1105: Switch to hardware operations for PTP
net: dsa: sja1105: Implement the .gettimex64 system call for PTP
net: dsa: sja1105: Restore PTP time after switch reset
net: dsa: sja1105: Disallow management xmit during switch reset
net: dsa: sja1105: Move PTP data to its own private structure
net: dsa: sja1105: Advertise the 8 TX queues
net: dsa: Pass ndo_setup_tc slave callback to drivers
net: dsa: sja1105: Add static config tables for scheduling
net: dsa: sja1105: Configure the Time-Aware Scheduler via tc-taprio
offload
net: dsa: sja1105: Make HOSTPRIO a kernel config
net: dsa: sja1105: Make the PTP command read-write
net: dsa: sja1105: Implement state machine for TAS with PTP clock
source
drivers/net/dsa/sja1105/Kconfig | 17 +
drivers/net/dsa/sja1105/Makefile | 4 +
drivers/net/dsa/sja1105/sja1105.h | 36 +-
.../net/dsa/sja1105/sja1105_dynamic_config.c | 8 +
drivers/net/dsa/sja1105/sja1105_main.c | 94 +-
drivers/net/dsa/sja1105/sja1105_ptp.c | 345 ++++---
drivers/net/dsa/sja1105/sja1105_ptp.h | 103 ++-
drivers/net/dsa/sja1105/sja1105_spi.c | 58 +-
.../net/dsa/sja1105/sja1105_static_config.c | 167 ++++
.../net/dsa/sja1105/sja1105_static_config.h | 48 +-
drivers/net/dsa/sja1105/sja1105_tas.c | 851 ++++++++++++++++++
drivers/net/dsa/sja1105/sja1105_tas.h | 74 ++
include/linux/netdevice.h | 1 +
include/net/dsa.h | 3 +
include/net/pkt_sched.h | 33 +
include/uapi/linux/pkt_sched.h | 3 +-
net/dsa/slave.c | 12 +-
net/dsa/tag_sja1105.c | 3 +-
net/sched/sch_taprio.c | 246 ++++-
19 files changed, 1883 insertions(+), 223 deletions(-)
create mode 100644 drivers/net/dsa/sja1105/sja1105_tas.c
create mode 100644 drivers/net/dsa/sja1105/sja1105_tas.h
--
2.17.1
