Since IQD devices complete (most of) their transmissions synchronously,
they don't offer TX completion IRQs and have no HW coalescing controls.
But we can fake the easy parts in SW, and give the user some control wrt
to how often the TX NAPI code should be triggered to process the TX
completions.
Having per-queue controls can in particular help the dedicated mcast
queue, as it likely benefits from different fine-tuning than what the
ucast queues need.
CC: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Count the number of TX doorbells we issue to the qdio layer.
Also count the number of actual frames in a TX buffer, and then
use this data along with the byte count during TX completion.
We'll make additional use of the frame count in a subsequent patch.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Versions are meaningless for an in-kernel driver.
Instead use the UTS_RELEASE that is set by ethtool_get_drvinfo().
Cc: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds support for SOF_TIMESTAMPING_TX_SOFTWARE.
No support for non-IQD devices, since they orphan the skb in their xmit
path.
To play nice with TX bulking, set the timestamp when the buffer that
contains the skb(s) is actually flushed out to HW.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For ucast traffic, qeth_iqd_select_queue() falls back to
netdev_pick_tx(). This will potentially use skb_tx_hash() to distribute
the flow over all active TX queues - so txq 0 is a valid selection, and
qeth_iqd_select_queue() needs to check for this and put it on some other
queue. As a result, the distribution for ucast flows is unbalanced and
hits QETH_IQD_MIN_UCAST_TXQ heavier than the other queues.
Open-coding a custom variant of skb_tx_hash() isn't an option, since
netdev_pick_tx() also gives us eg. access to XPS. But we can pull a
little trick: add a single TC class that excludes the mcast txq, and
thus encourage skb_tx_hash() to not pick the mcast txq.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Similar to the support for z/VM NICs, but we need to take extra care
about the dedicated mcast queue:
1. netdev_pick_tx() is unaware of this limitation and might select the
mcast txq. Catch this.
2. require at least _two_ TX queues - one for ucast, one for mcast.
3. when reducing the number of TX queues, there's a potential race
where netdev_cap_txqueue() over-rules the selected txq index and
falls back to index 0. This would place ucast traffic on the mcast
queue, and result in TX errors.
So for IQD, reject a reduction while the interface is running.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add support for ETHTOOL_SCHANNELS to change the count of active
TX queues.
Since all TX queue structs are pre-allocated and -registered, we just
need to trivially adjust dev->real_num_tx_queues.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Implement the ethtool hooks for the ETHTOOL_RX_COPYBREAK tunable.
The copybreak is stored into netdev_priv, so that we automatically go
back to the default value if the netdev is re-allocated.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Depending on a packet's type, the RX path needs to access fields in the
packet headers and thus requires a minimum packet length.
Enforce this length when building the skb.
On the other hand a single runt packet is no reason to drop the whole
RX buffer. So just skip it, and continue processing on the next packet.
Fixes: 4a71df5004 ("qeth: new qeth device driver")
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Where available, use the fine-grained counters in rtnl_link_stats64 to
indicate different RX error causes. For drop reasons, use driver-private
ethtool counters.
In particular this patch allows us to keep track of driver-side drops due
to unknown/unsupported HW descriptor format.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Due to their large MTU and potentially low utilization of TX buffers,
IQD devices in particular require fast TX recycling. This makes them
a prime candidate for a TX NAPI path in qeth.
qeth_tx_poll() uses the recently introduced qdio_inspect_queue() helper
to poll the TX queue for completed buffers. To avoid hogging the CPU for
too long, we yield to the stack after completing an entire queue's worth
of buffers.
While IQD is expected to transfer its buffers synchronously (and thus
doesn't support TX interrupts), a timer covers for the odd case where a
TX buffer doesn't complete synchronously. Currently this timer should
only ever fire for
(1) the mcast queue,
(2) the occasional race, where the NAPI poll code observes an update to
queue->used_buffers while the TX doorbell hasn't been issued yet.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Current xmit code only stops the txq after attempting to fill an
IO buffer that hasn't been TX-completed yet. In many-connection
scenarios, this can result in frequent rejected TX attempts, requeuing
of skbs with NETDEV_TX_BUSY and extra overhead.
Now that we have a proper 1-to-1 relation between stack-side txqs and
our HW Queues, overhaul the stop/wake logic so that the xmit code
stops the txq as needed.
Given that we might map multiple skbs into a single buffer, it's crucial
to ensure that the queue always provides an _entirely_ empty IO buffer.
Otherwise large skbs (eg TSO) might not fit into the last available
buffer. So whenever qeth_do_send_packet() first utilizes an _empty_
buffer, it updates & checks the used_buffers count.
This now ensures that an skb passed to qeth_xmit() can always be mapped
into an IO buffer, so remove all of the -EBUSY roll-back handling in the
TX path. We preserve the minimal safety-checks ("Is this IO buffer
really available?"), just in case some nasty future bug ever attempts to
corrupt an in-use buffer.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
qeth has been supporting multiple HW Output Queues for a long time. But
rather than exposing those queues to the stack, it uses its own queue
selection logic in .ndo_start_xmit... with all the drawbacks that
entails.
Start off by switching IQD devices over to a proper mqs net_device,
and converting all the netdev_queue management code.
One oddity with IQD devices is the requirement to place all mcast
traffic on the _highest_ established HW queue. Doing so via
.ndo_select_queue seems straight-forward - but that won't work if only
some of the HW queues are active
(ie. when dev->real_num_tx_queues < dev->num_tx_queues), since
netdev_cap_txqueue() will not allow us to put skbs on the higher queues.
To make this work, we
1. let .ndo_select_queue() map all mcast traffic to netdev_queue 0, and
2. later re-map the netdev_queue and HW queue indices in
.ndo_start_xmit and the TX completion handler.
With this patch we default to a fixed set of 1 ucast and 1 mcast queue.
Support for dynamic reconfiguration is added at a later time.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Implement a trivial callback that exposes the queue sizes.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Accumulate per-TX queue statistics, and increase their size to 64 bit.
Don't bother with enabling/disabling the statistics, the overhead is
negligible.
Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Julian Wiedmann