On EF10, the firmware will initiate a queue flush in certain
error cases.  We need to accept that flush events might appear
at any time after a queue has been initialised, not just when
we try to flush them.

We can handle Falcon-architecture in just the same way.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

On EF10, event queue initialisation requires an MCDI request which
may return failure.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Also, since we handle all DMA errors in the same way, merge
RESET_TYPE_(RX|TX)_DESC_FETCH into RESET_TYPE_DMA_ERROR.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

On Falcon we implement MAC filtering requested by the stack using the
MAC wrapper's single unicast filter and multicast hash filter.  Siena
is very similar, though MAC configuration is mediated by the MC.

Since MCDI operations may sleep, reconfiguration is deferred from
ndo_set_rx_mode to a work item.  However, it still updates the private
variables describing the filter state synchronously.  Contrary to
comments, the later use of these variables is not protected using the
address lock, resulting in race conditions.

Move the state update to a new function
efx_farch_filter_sync_rx_mode() and make the Falcon-arch MAC
configuration functions call that, so that its use is consistently
serialised by the mac_lock.

Invert and rename the promiscuous flag to the more accurate
unicast_filter, and comment that both this and multicast_hash are
not used on EF10.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

MAC filters inserted on request from the stack (ndo_set_rx_mode)
should allow manual steering but not removal.  Currently we have a
special case for Siena's all-multicast and all-unicast MAC filters,
but on EF10 we need to allow for steering of precise MAC filters as
well.

The EFX_FILTER_FLAG_RX_STACK flag changes the behaviour of replacement
and removal requests:

- Replacement *of* a filter with this flag never clears the flag but
  does change steering and saved priority
- Replacement *by* a filter with this flag only sets the flag but does
  not change steering
- Removal with priority < EFX_FILTER_PRI_REQUIRED really resets RX
  steering and saved priority

This could support precise MAC filtering on Siena in future.

As a side-benefit, the default MAC filters are hidden from ethtool
until they are steered.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Move the special case for removal of default filters from
efx_farch_filter_table_clear_entry() into a wrapper function,
efx_farch_filter_table_remove().  Move the existence and priority
checks into the latter and use it where appropriate.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Aside from accelerated RFS, there is almost nothing that can be shared
between the filter table implementations for the Falcon architecture
and EF10.

Move the few shared functions into efx.c and rx.c and the rest into
farch.c.  Introduce efx_nic_type operations for the implementation and
inline wrapper functions that call these.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

The workarounds that currently use EFX_WORKAROUND_ALWAYS are in
Falcon-specific or Falcon-arch-specific code, so get rid of the
conditions altogether.  Add/move comments as appropriate.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Add efx_nic_type operations for the many efx_nic functions that need
to be implemented different on EF10.  For now, change most of the
existing efx_nic_*() functions into inline wrappers.  As a later step,
we may be able to improve branch prediction for operations used on the
fast path by copying the pointers into each queue/channel structure.

Move the Falcon/Siena implementations to new file farch.c and rename
the functions and static data to use a prefix of 'efx_farch_'.

Move efx_may_push_tx_desc() to nic.h, as the EF10 TX code will also
use it.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Currently efx_stop_datapath() will try to flush our DMA queues (if DMA
is enabled), then finalise software and hardware state for each queue.
However, for EF10 we must ask the MC to finalise each queue, which
implicitly starts flushing it, and then wait for the flush events.
We therefore need to delegate more of this to the NIC type.

Combine all the hardware operations into a new NIC-type operation
efx_nic_type::fini_dmaq, and call this before tearing down the
software state and buffers for all the DMA queues.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

rx_queue::enabled guards refill, so rename it to reflect that.  Clear
it at the start of the queue teardown process rather than waiting for
the RX queue to be flushed.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

We unconditionally acknowledge legacy interrupts just before disabling
them.  This workaround is needed on Falcon A1 but probably not on
later chips where the legacy interrupt mechanism is different.  It was
also originally done after the IRQ handler was removed, not before.
Restore the original behaviour for Falcon A1 only by doing this
acknowledgement in the efx_nic_type::fini operation.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

There are many problems with the current efx_stop_interrupts() and
efx_start_interrupts():

1. On Siena, it is unsafe to disable the master IRQ enable bit
(DRV_INT_EN_KER) while any IRQ sources are enabled.

2. On EF10 there is no master IRQ enable bit, so we cannot expect to
defer IRQs without tearing down event queues.  (Though I don't think
we will need to keep any event queues around while the device is down,
as we do for VFDI on Siena.)

3. synchronize_irq() only waits for a running IRQ handler to finish,
not for any propagation through IRQ controllers.  Therefore an IRQ may
still be received and handled after efx_stop_interrupts() returns.
IRQ handlers can then race with channel reallocation.

To fix this:

a. Introduce a software IRQ enable flag.  So long as this is clear,
IRQ handlers will only acknowledge IRQs and not touch the channel
structures.

b. Define a new struct efx_msi_context as the context for MSIs.  This
is never reallocated and is sufficient to find the software enable
flag and the channel structure.  It also includes the channel/IRQ
name, which was previously separated out as it must also not be
reallocated.

c. Split efx_{start,stop}_interrupts() into
efx_{,soft_}_{enable,disable}_interrupts().  The 'soft' functions
don't touch the hardware master enable flag (if it exists) and don't
reinitialise or tear down channels with the keep_eventq flag set.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

The EF10 architecture has a very different register layout from
previous controllers, so we'll use separate files for the two sets of
register definitions.  Use 'farch' as an abbreviation for
Falcon-architecture.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

On EF10, the firmware is in charge of allocating buffer table entries.
Change struct efx_special_buffer to use a struct efx_buffer member,
so that it can be used with efx_nic_{alloc,free}_buffer() in that
case.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Most call sites for efx_nic_alloc_buffer() are part of the probe or
reconfiguration paths and can allocate with GFP_KERNEL.  A few others
should use GFP_NOIO (I think).  Only one is in atomic context and
must use the current GFP_ATOMIC.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

The lifetime of an irq_cpu_rmap is odd: we have to allocate it before
installing IRQ handlers and free it before removing the IRQ handlers.
As a result of this asymmetry, it was omitted from some failure paths.

On another failure path, we could try to remove IRQ handlers we
had not yet installed.

Move the irq_cpu_rmap allocation and freeing alongside IRQ handler
installation and removal, in efx_nic_{init,fini}_interrupts().
Count the number of IRQ handlers successfully installed and only
remove those on the failure path.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

This will be useful for shortcutting some software packet parsing.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

PCI legacy interrupts are level-triggered, and we cannot mask them up
on an isolated device.  Instead, disable the IRQ at the controller
until we have recovered.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Reduce the number of calls required to alloc
a zeroed block of memory.

Trivially reduces overall object size.

Other changes around these removals
o Neaten call argument alignment
o Remove an unnecessary OOM message after dma_alloc_coherent failure
o Remove unnecessary gfp_t stack variable
Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Using TX push when notifying the NIC of multiple new descriptors in
the ring will very occasionally cause the TX DMA engine to re-use an
old descriptor.  This can result in a duplicated or partly duplicated
packet (new headers with old data), or an IOMMU page fault.  This does
not happen when the pushed descriptor is the only one written.

TX push also provides little latency benefit when a packet requires
more than one descriptor.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Enable RX DMA scattering iff an RX buffer large enough for the current
MTU will not fit into a single page and the NIC supports DMA
scattering for kernel-mode RX queues.

On Falcon and Siena, the RX_USR_BUF_SIZE field is used as the DMA
limit for both all RX queues with scatter enabled.  Set it to 1824,
matching what Onload uses now.

Maintain a statistic for frames truncated due to lack of descriptors
(rx_nodesc_trunc).  This is distinct from rx_frm_trunc which may be
incremented when scattering is disabled and implies an over-length
frame.

Whenever an MTU change causes scattering to be turned on or off,
update filters that point to the PF queues, but leave others
unchanged, as VF drivers assume scattering is off.

Add n_frags parameters to various functions, and make them iterate:
- efx_rx_packet()
- efx_recycle_rx_buffers()
- efx_rx_mk_skb()
- efx_rx_deliver()

Make efx_handle_rx_event() responsible for updating
efx_rx_queue::removed_count.

Change the RX pipeline state to a starting ring index and number of
fragments, and make __efx_rx_packet() responsible for clearing it.

Based on earlier versions by David Riddoch and Jon Cooper.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

We sometimes hit a "failed to flush" timeout on some TX queues, but the
flushes have completed and the flush completion events seem to go missing.
In this case, we can check the TX_DESC_PTR_TBL register and drain the
queues if the flushes had finished.

[bwh: Minor fixes to coding style]
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Currently we initialise the newly allocated buffer to all-1s, which is
important for event queues but not for descriptor queues.  And since
we also do that in efx_nic_init_eventq(), it is completely pointless
to do it here.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

efx_writed_table() uses a step of 16 bytes but efx_readd_table() uses
a step of 4 bytes.  Why are they different?

Firstly, register access is asymmetric:

- The EVQ_RPTR table and RX_INDIRECTION_TBL can (or must?) be written
  as dwords even though they have a step size of 16 bytes, unlike
  most other CSRs.
- In general, a read of any width is valid for registers, so long as
  it does not cross register boundaries.  There is also no latching
  behaviour in the BIU, contrary to rumour.

We write to the EVQ_RPTR table with efx_writed_table() but never read
it back as it's write-only.  We write to the RX_INDIRECTION_TBL with
efx_writed_table(), but only read it back for the register dump, where
we use efx_reado_table() as for any other table with step size of 16.

We read MC_TREG_SMEM with efx_readd_table() for the register dump, but
normally read and write it with efx_readd() and efx_writed() using
offsets calculated in bytes.

Since these functions are trivial and have few callers, it's clearer
to open-code them at the call sites.  While we're at it, update the
comments on the BIU behaviour again.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Receiving pause frames can block TX queue flushes.  Earlier changes
work around this by reconfiguring the MAC during flushes for VFs, but
during flushes for the PF we would only change the fc_disable counter.
Unless the MAC is reconfigured for some other reason during the flush
(which I would not expect to happen) this had no effect at all.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

There are now standard functions for dealing with little-endian bit
arrays, so use them instead of our own implementations.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NTakuya Yoshikawa <yoshikawa.takuya@oss.ntt.co.jp>
Cc: David Miller <davem@davemloft.net>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

TSO header buffers contain a control structure immediately followed by
the packet headers, and are kept on a free list when not in use.  This
complicates buffer management and tends to result in cache read misses
when we recycle such buffers (particularly if DMA-coherent memory
requires caches to be disabled).

Replace the free list with a simple mapping by descriptor index.  We
know that there is always a payload descriptor between any two
descriptors with TSO header buffers, so we can allocate only one
such buffer for each two descriptors.

While we're at it, use a standard error code for allocation failure,
not -1.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Add a flags field to struct efx_tx_buffer, replacing the
continuation and map_single booleans.

Since a single descriptor cannot be both a TSO header and the last
descriptor for an skb, unionise efx_tx_buffer::{skb,tsoh} and add
flags for validity of these fields.

Clear all flags in free buffers (whereas previously the continuation
flag would be set).
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Currently VF queues and drivers may remain active during this test.
This could cause memory corruption or spurious test failures.
Therefore we reset the port/function before running these tests on
Siena.

On Falcon this doesn't work: we have to do some additional
initialisation before some blocks will work again.  So refactor the
reset/register-test sequence into an efx_nic_type method so
efx_selftest() doesn't have to consider such quirks.

In the process, fix another minor bug: Siena does not have an
'invisible' reset and the self-test currently fails to push the PHY
configuration after resetting.  Passing RESET_TYPE_ALL to
efx_reset_{down,up}() fixes this.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Generate a test event on each event queue whenever the interface is
brought up, then after 1 second check that we have either handled a
test event or handled another IRQ for each event queue.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Cleanup in preparation for doing an event test on ifup.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

RX and TX completions on the same event queue are generally not associated
with the same flows.  The inclusion of TX completions in the adaptive IRQ
score is more of a source of noise rather than useful feedback.  Therefore,
do not include them in the score, and adjust the default threshold scores
down.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

On the SFC9000 family, each port has 1024 Virtual Interfaces (VIs),
each with an RX queue, a TX queue, an event queue and a mailbox
register.  These may be assigned to up to 127 SR-IOV virtual functions
per port, with up to 64 VIs per VF.

We allocate an extra channel (IRQ and event queue only) to receive
requests from VF drivers.

There is a per-port limit of 4 concurrent RX queue flushes, and queue
flushes may be initiated by the MC in response to a Function Level
Reset (FLR) of a VF.  Therefore, when SR-IOV is in use, we submit all
flush requests via the MC.

The RSS indirection table is shared with VFs, so the number of RX
queues used in the PF is limited to the number of VIs per VF.

This is almost entirely the work of Steve Hodgson, formerly
shodgson@solarflare.com.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Each port has a block of 64-bit SRAM that is divided between buffer
table and descriptor cache regions at initialisation time.  Currently
we use a fixed allocation, but it needs to be changed to support
larger numbers of queues.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

The TX DMA engine issues upstream read requests when there is room in
the TX FIFO for the completion. However, the fetches for the rest of
the packet might be delayed by any back pressure.  Since a flush must
wait for an EOP, the entire flush may be delayed by back pressure.

Mitigate this by disabling flow control before the flushes are
started.  Since PF and VF flushes run in parallel introduce
fc_disable, a reference count of the number of flushes outstanding.

The same principle could be applied to Falcon, but that
would bring with it its own testing.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>

For SR-IOV we will need to send events to event queues that belong to
VFs serviced by other drivers.  Change the parameters of
efx_generate_event() to allow this and declare it extern.

While we're at it, remove the existing declaration under the wrong
name efx_nic_generate_event().
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>