A DSA master interface has upper network devices, each representing an
Ethernet switch port attached to it. Demultiplexing the source ports and
setting skb->dev accordingly is done through the catch-all ETH_P_XDSA
packet_type handler. Catch-all because DSA vendors have various header
implementations, which can be placed anywhere in the frame: before the
DMAC, before the EtherType, before the FCS, etc. So, the ETH_P_XDSA
handler acts like an rx_handler more than anything.

It is unlikely for the DSA master interface to have any other upper than
the DSA switch interfaces themselves. Only maybe a bridge upper*, but it
is very likely that the DSA master will have no 8021q upper. So
__netif_receive_skb_core() will try to untag the VLAN, despite the fact
that the DSA switch interface might have an 8021q upper. So the skb will
never reach that.

So far, this hasn't been a problem because most of the possible
placements of the DSA switch header mentioned in the first paragraph
will displace the VLAN header when the DSA master receives the frame, so
__netif_receive_skb_core() will not actually execute any VLAN-specific
code for it. This only becomes a problem when the DSA switch header does
not displace the VLAN header (for example with a tail tag).

What the patch does is it bypasses the untagging of the skb when there
is a DSA switch attached to this net device. So, DSA is the only
packet_type handler which requires seeing the VLAN header. Once skb->dev
will be changed, __netif_receive_skb_core() will be invoked again and
untagging, or delivery to an 8021q upper, will happen in the RX of the
DSA switch interface itself.

*see commit 9eb8eff0 ("net: bridge: allow enslaving some DSA master
network devices". This is actually the reason why I prefer keeping DSA
as a packet_type handler of ETH_P_XDSA rather than converting to an
rx_handler. Currently the rx_handler code doesn't support chaining, and
this is a problem because a DSA master might be bridged.
Signed-off-by: NVladimir Oltean <olteanv@gmail.com>
Reviewed-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

flush_all_backlogs() may cause deadlock on systems
running processes with FIFO scheduling policy.

The above is critical in -RT scenarios, where user-space
specifically ensure no network activity is scheduled on
the CPU running the mentioned FIFO process, but still get
stuck.

This commit tries to address the problem checking the
backlog status on the remote CPUs before scheduling the
flush operation. If the backlog is empty, we can skip it.

v1 -> v2:
 - explicitly clear flushed cpu mask - Eric
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

SOCK_QUEUE_SHRUNK is currently used by TCP as a temporary state
that remembers if some room has been made in the rtx queue
by an incoming ACK packet.

This is later used from tcp_check_space() before
considering to send EPOLLOUT.

Problem is: If we receive SACK packets, and no packet
is removed from RTX queue, we can send fresh packets, thus
moving them from write queue to rtx queue and eventually
empty the write queue.

This stall can happen if TCP_NOTSENT_LOWAT is used.

With this fix, we no longer risk stalling sends while holes
are repaired, and we can fully use socket sndbuf.

This also removes a cache line dirtying for typical RPC
workloads.

Fixes: c9bee3b7 ("tcp: TCP_NOTSENT_LOWAT socket option")
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: NNeal Cardwell <ncardwell@google.com>
Acked-by: NSoheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This comment is outdated and no longer reflects the actual implementation
of af_packet.c.

Reasons for the new comment:

1.

In af_packet.c, the function packet_snd first reserves a headroom of
length (dev->hard_header_len + dev->needed_headroom).
Then if the socket is a SOCK_DGRAM socket, it calls dev_hard_header,
which calls dev->header_ops->create, to create the link layer header.
If the socket is a SOCK_RAW socket, it "un-reserves" a headroom of
length (dev->hard_header_len), and checks if the user has provided a
header sized between (dev->min_header_len) and (dev->hard_header_len)
(in dev_validate_header).
This shows the developers of af_packet.c expect hard_header_len to
be consistent with header_ops.

2.

In af_packet.c, the function packet_sendmsg_spkt has a FIXME comment.
That comment states that prepending an LL header internally in a driver
is considered a bug. I believe this bug can be fixed by setting
hard_header_len to 0, making the internal header completely invisible
to af_packet.c (and requesting the headroom in needed_headroom instead).

3.

There is a commit for a WiFi driver:
commit 9454f7a8 ("mwifiex: set needed_headroom, not hard_header_len")
According to the discussion about it at:
  https://patchwork.kernel.org/patch/11407493/
The author tried to set the WiFi driver's hard_header_len to the Ethernet
header length, and request additional header space internally needed by
setting needed_headroom.
This means this usage is already adopted by driver developers.

Cc: Willem de Bruijn <willemdebruijn.kernel@gmail.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Brian Norris <briannorris@chromium.org>
Cc: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: NXie He <xie.he.0141@gmail.com>
Acked-by: NWillem de Bruijn <willemb@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

That is needed to let the subflows announce promptly when new
space is available in the receive buffer.

tcp_cleanup_rbuf() is currently a static function, drop the
scope modifier and add a declaration in the TCP header.
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Update the scheduler to less trivial heuristic: cache
the last used subflow, and try to send on it a reasonably
long burst of data.

When the burst or the subflow send space is exhausted, pick
the subflow with the lower ratio between write space and
send buffer - that is, the subflow with the greater relative
amount of free space.

v1 -> v2:
 - fix 32 bit build breakage due to 64bits div
 - fix checkpath issues (uint64_t -> u64)
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Currently the 'backup' attribute of local endpoint
is ignored. Let's use it for the MP_JOIN handshake
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

So that can be accessed easily from the subflow creation
helper. No functional change intended.
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add a bunch of MPTCP mibs related to MPTCP OoO data
processing.
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

There is no need to use the tcp_read_sock(), we can
simply drop the skb. Additionally try to look at the
next buffer for in order data.

This both simplifies the code and avoid unneeded indirect
calls.
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add an RB-tree to cope with OoO (at MPTCP level) data.
__mptcp_move_skb() insert into the RB tree "future"
data, eventually coalescing skb as allowed by the
MPTCP DSN.

To simplify sequence accounting, move the DSN inside
the cb.

After successfully enqueuing in sequence data, check
if we can use any data from the RB tree.

Additionally move the data_fin check after spooling
data from the OoO tree, otherwise we could miss shutdown
events.

The RB tree code is copied as verbatim as possible
from tcp_data_queue_ofo(), with a few simplifications
due to the fact that MPTCP doesn't need to cope with
sacks. All bugs here are added by me.
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Factor-out existing code, will be re-used by the
next patch.
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Let the msk sendbuf track the size of the larger subflow's
send window, so that we ensure mptcp_sendmsg() does not
exceed MPTCP-level send window.

The update is performed just before try to send any data.
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This is a prerequisite to allow receiving data from multiple
subflows without re-injection.

Instead of dropping the OoO - "future" data in
subflow_check_data_avail(), call into __mptcp_move_skbs()
and let the msk drop that.

To avoid code duplication factor out the mptcp_subflow_discard_data()
helper.

Note that __mptcp_move_skbs() can now find multiple subflows
with data avail (comprising to-be-discarded data), so must
update the byte counter incrementally.

v1 -> v2:
 - fix checkpatch issues (unsigned -> unsigned int)
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This simplify mptcp_subflow_data_available() and will
made follow-up patches simpler.

Additionally remove the unneeded checks on subflow copied_seq:
we always whole skbs out of subflows.
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Currently, when checking for the 'msk is writable' condition, we
look at the individual subflows write space.
That works well while we send data via a single subflow, but will
not as soon as we will enable concurrent xmit on multiple subflows.

With this change msk becomes writable when the following conditions
hold:
- the socket has some free write space
- there is at least a subflow with write free space

Additionally we need to set the NOSPACE bit on all subflows
before blocking.
Signed-off-by: NPaolo Abeni <pabeni@redhat.com>
Reviewed-by: NMat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When setting up a client connection, a second ref is accidentally obtained
on the connection bundle (we get one when allocating the conn and a second
one when adding the conn to the bundle).

Fix it to only use the ref obtained by rxrpc_alloc_client_connection() and
not to add a second when adding the candidate conn to the bundle.

Fixes: 245500d8 ("rxrpc: Rewrite the client connection manager")
Signed-off-by: NDavid Howells <dhowells@redhat.com>

When the network namespace exits, rxrpc_clean_up_local_conns() needs to
unbundle each client connection it evicts.  Fix it to do this.

kernel BUG at net/rxrpc/conn_object.c:481!
RIP: 0010:rxrpc_destroy_all_connections.cold+0x11/0x13 net/rxrpc/conn_object.c:481
Call Trace:
 rxrpc_exit_net+0x1a4/0x2e0 net/rxrpc/net_ns.c:119
 ops_exit_list+0xb0/0x160 net/core/net_namespace.c:186
 cleanup_net+0x4ea/0xa00 net/core/net_namespace.c:603
 process_one_work+0x94c/0x1670 kernel/workqueue.c:2269
 worker_thread+0x64c/0x1120 kernel/workqueue.c:2415
 kthread+0x3b5/0x4a0 kernel/kthread.c:292
 ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:294

Fixes: 245500d8 ("rxrpc: Rewrite the client connection manager")
Reported-by: syzbot+52071f826a617b9c76ed@syzkaller.appspotmail.com
Signed-off-by: NDavid Howells <dhowells@redhat.com>

The alloc_error field in the rxrpc_bundle struct should be signed as it has
negative error codes assigned to it.  Checks directly on it may then fail,
and may produce a warning like this:

	net/rxrpc/conn_client.c:662 rxrpc_wait_for_channel()
	warn: 'bundle->alloc_error' is unsigned

Fixes: 245500d8 ("rxrpc: Rewrite the client connection manager")
Reported-by Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: NDavid Howells <dhowells@redhat.com>

Fix an error-handling goto in rxrpc_connect_call() whereby it will jump to
free the bundle it failed to allocate.

Fixes: 245500d8 ("rxrpc: Rewrite the client connection manager")
Reported-by: Nkernel test robot <lkp@intel.com>
Reported-by: NDan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: NDavid Howells <dhowells@redhat.com>

This reverts commit 314f76d7.

Citing that commit message, the call graph was:

    dsa_slave_vlan_rx_add_vid   dsa_port_setup_8021q_tagging
                |                        |
                |                        |
                |          +-------------+
                |          |
                v          v
               dsa_port_vid_add      dsa_slave_port_obj_add
                      |                         |
                      +-------+         +-------+
                              |         |
                              v         v
                           dsa_port_vlan_add

Now that tag_8021q has its own ops structure, it no longer relies on
dsa_port_vid_add, and therefore on the dsa_switch_ops to install its
VLANs.

So dsa_port_vid_add now only has one single caller. So we can simplify
the call graph to what it was before, aka:

        dsa_slave_vlan_rx_add_vid     dsa_slave_port_obj_add
                      |                         |
                      +-------+         +-------+
                              |         |
                              v         v
                           dsa_port_vlan_add
Signed-off-by: NVladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

While working on another tag_8021q driver implementation, some things
became apparent:

- It is not mandatory for a DSA driver to offload the tag_8021q VLANs by
  using the VLAN table per se. For example, it can add custom TCAM rules
  that simply encapsulate RX traffic, and redirect & decapsulate rules
  for TX traffic. For such a driver, it makes no sense to receive the
  tag_8021q configuration through the same callback as it receives the
  VLAN configuration from the bridge and the 8021q modules.

- Currently, sja1105 (the only tag_8021q user) sets a
  priv->expect_dsa_8021q variable to distinguish between the bridge
  calling, and tag_8021q calling. That can be improved, to say the
  least.

- The crosschip bridging operations are, in fact, stateful already. The
  list of crosschip_links must be kept by the caller and passed to the
  relevant tag_8021q functions.

So it would be nice if the tag_8021q configuration was more
self-contained. This patch attempts to do that.

Create a struct dsa_8021q_context which encapsulates a struct
dsa_switch, and has 2 function pointers for adding and deleting a VLAN.
These will replace the previous channel to the driver, which was through
the .port_vlan_add and .port_vlan_del callbacks of dsa_switch_ops.

Also put the list of crosschip_links into this dsa_8021q_context.
Drivers that don't support cross-chip bridging can simply omit to
initialize this list, as long as they dont call any cross-chip function.

The sja1105_vlan_add and sja1105_vlan_del functions are refactored into
a smaller sja1105_vlan_add_one, which now has 2 entry points:
- sja1105_vlan_add, from struct dsa_switch_ops
- sja1105_dsa_8021q_vlan_add, from the tag_8021q ops
But even this change is fairly trivial. It just reflects the fact that
for sja1105, the VLANs from these 2 channels end up in the same hardware
table. However that is not necessarily true in the general sense (and
that's the reason for making this change).

The rest of the patch is mostly plain refactoring of "ds" -> "ctx". The
dsa_8021q_context structure needs to be propagated because adding a VLAN
is now done through the ops function pointers inside of it.
Signed-off-by: NVladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

There is no point in calling dsa_port_setup_8021q_tagging for each
individual port. Additionally, it will become more difficult to do that
when we'll have a context structure to tag_8021q (next patch). So
refactor this now.
Signed-off-by: NVladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Each MDB entry is encoded in a nested netlink attribute called
'MDBA_MDB_ENTRY'. In turn, this attribute contains another nested
attributed called 'MDBA_MDB_ENTRY_INFO', which encodes a single port
group entry within the MDB entry.

The cited commit added the ability to restart a dump from a specific
port group entry. However, on failure to add a port group entry to the
dump the entire MDB entry (stored in 'nest2') is removed, resulting in
missing port group entries.

Fix this by finalizing the MDB entry with the partial list of already
encoded port group entries.

Fixes: 5205e919 ("net: bridge: mcast: add support for src list and filter mode dumping")
Signed-off-by: NIdo Schimmel <idosch@nvidia.com>
Acked-by: NNikolay Aleksandrov <nikolay@nvidia.com>
Reviewed-by: NJiri Pirko <jiri@nvidia.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The variable err is being initialized with a value that is never read and
it is being updated later with a new value. The initialization is redundant
and can be removed.  Also re-order variable declarations in reverse
Christmas tree ordering.

Addresses-Coverity: ("Unused value")
Signed-off-by: NColin Ian King <colin.king@canonical.com>
Reviewed-by: NDavid Ahern <dsahern@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

There are 6 types of workers which exist per smc connection. 3 of them
are used for listen and handshake processing, another 2 are used for
close and abort processing and 1 is the tx worker that moves calls to
sleeping functions into a worker.
To prevent flooding of the system work queue when many connections are
opened or closed at the same time (some pattern uperf implements), move
those workers to one of 3 smc-specific work queues. Two work queues are
module-global and used for handshake and close workers. The third work
queue is defined per link group and used by the tx workers that may
sleep waiting for resources of this link group.
And in smc_llc_enqueue() queue the llc_event_work work to the system
prio work queue because its critical that this work is started fast.
Reviewed-by: NUrsula Braun <ubraun@linux.ibm.com>
Signed-off-by: NKarsten Graul <kgraul@linux.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When the netlink messages to be sent to the userspace
are too big for a single netlink message, send them in
chunks using the netlink_dump infrastructure. Modify the
smc diag dump code so that it can signal to the netlink_dump
infrastructure that it needs to send more data.
Signed-off-by: NGuvenc Gulce <guvenc@linux.ibm.com>
Signed-off-by: NKarsten Graul <kgraul@linux.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

smc_lgr_cleanup_early() schedules the free worker with delay. DMB
unregistering occurs in this delayed worker increasing the risk
to reach the SMCD SBA limit without need. Terminate the
linkgroup immediately, since termination means early DMB unregistering.

For SMCD the global smc_server_lgr_pending lock is given up early.
A linkgroup to be given up with smc_lgr_cleanup_early() may already
contain more than one connection. Using __smc_lgr_terminate() in
smc_lgr_cleanup_early() covers this.

And consolidate smc_ism_put_vlan() and smc_put_device() into smc_lgr_free()
only.
Signed-off-by: NUrsula Braun <ubraun@linux.ibm.com>
Signed-off-by: NKarsten Graul <kgraul@linux.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

smc_listen_work() contains already an smc_listen_decline() exit.
Use this exit for smc_listen_rdma_finish() problems as well.
Signed-off-by: NUrsula Braun <ubraun@linux.ibm.com>
Signed-off-by: NKarsten Graul <kgraul@linux.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Move check whether peer can be reached into smc_pnet_find_ism_by_pnetid().
Thus searching continues for another ism device, if check fails.
Signed-off-by: NUrsula Braun <ubraun@linux.ibm.com>
Signed-off-by: NKarsten Graul <kgraul@linux.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

smc_clc_send_accept() and smc_clc_send_confirm() are quite similar.
Move common code into a separate function smc_clc_send_confirm_accept().
And introduce separate SMCD and SMCR struct definitions for CLC accept
resp. confirm.
No functional change.
Signed-off-by: NUrsula Braun <ubraun@linux.ibm.com>
Signed-off-by: NKarsten Graul <kgraul@linux.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Reduce stack size for smc_listen_work() and smc_clc_send_proposal()
by dynamic allocation of the CLC buffer to be received or sent.
Signed-off-by: NUrsula Braun <ubraun@linux.ibm.com>
Signed-off-by: NKarsten Graul <kgraul@linux.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Field names "srv_first_contact" and "cln_first_contact" are misleading,
since they apply to both, server and client. Rename them to
"first_contact_peer" and "first_contact_local".
Rename "ism_gid" by the more precise name "ism_peer_gid".
Rename version constant "SMC_CLC_V1" into "SMC_V1".
No functional change.
Signed-off-by: NUrsula Braun <ubraun@linux.ibm.com>
Signed-off-by: NKarsten Graul <kgraul@linux.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

SMC starts a separate tcp_listen worker for every SMC socket in
state SMC_LISTEN, and can accept an incoming connection request only,
if this worker is really running and waiting in kernel_accept(). But
the number of running workers is limited.
This patch reworks the listening SMC code and starts a tcp_listen worker
after the SYN-ACK handshake on the internal clc-socket only.
Suggested-by: NKarsten Graul <kgraul@linux.ibm.com>
Signed-off-by: NUrsula Braun <ubraun@linux.ibm.com>
Reviewed-by: NGuvenc Gulce <guvenc@linux.ibm.com>
Signed-off-by: NKarsten Graul <kgraul@linux.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This commit adds a new TCP feature to reflect the tos value received in
SYN, and send it out on the SYN-ACK, and eventually set the tos value of
the established socket with this reflected tos value. This provides a
way to set the traffic class/QoS level for all traffic in the same
connection to be the same as the incoming SYN request. It could be
useful in data centers to provide equivalent QoS according to the
incoming request.
This feature is guarded by /proc/sys/net/ipv4/tcp_reflect_tos, and is by
default turned off.
Signed-off-by: NWei Wang <weiwan@google.com>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This commit adds tos as a new passed in parameter to
ip_build_and_send_pkt() which will be used in the later commit.
This is a pure restructure and does not have any functional change.
Signed-off-by: NWei Wang <weiwan@google.com>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

A new field is added to the request sock to record the TOS value
received on the listening socket during 3WHS:
When not under syn flood, it is recording the TOS value sent in SYN.
When under syn flood, it is recording the TOS value sent in the ACK.
This is a preparation patch in order to do TOS reflection in the later
commit.
Signed-off-by: NWei Wang <weiwan@google.com>
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

netpoll needs to traverse dev->napi_list under RCU, make
sure it uses the right iterator and that removal from this
list is handled safely.
Signed-off-by: NJakub Kicinski <kuba@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

To RCUify napi->dev_list we need to replace list_del_init()
with list_del_rcu(). There is no _init() version for RCU for
obvious reasons. Up until now netif_napi_del() was idempotent
so to make sure it remains such add a bit which is set when
NAPI is listed, and cleared when it removed. Since we don't
expect multiple calls to netif_napi_add() to be correct,
add a warning on that side.

Now that napi_hash_add / napi_hash_del are only called by
napi_add / del we can actually steal its bit. We just need
to make sure hash node is initialized correctly.
Signed-off-by: NJakub Kicinski <kuba@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We allow drivers to call napi_hash_del() before calling
netif_napi_del() to batch RCU grace periods. This makes
the API asymmetric and leaks internal implementation details.
Soon we will want the grace period to protect more than just
the NAPI hash table.

Restructure the API and have drivers call a new function -
__netif_napi_del() if they want to take care of RCU waits.

Note that only core was checking the return status from
napi_hash_del() so the new helper does not report if the
NAPI was actually deleted.

Some notes on driver oddness:
 - veth observed the grace period before calling netif_napi_del()
   but that should not matter
 - myri10ge observed normal RCU flavor
 - bnx2x and enic did not actually observe the grace period
   (unless they did so implicitly)
 - virtio_net and enic only unhashed Rx NAPIs

The last two points seem to indicate that the calls to
napi_hash_del() were a left over rather than an optimization.
Regardless, it's easy enough to correct them.

This patch may introduce extra synchronize_net() calls for
interfaces which set NAPI_STATE_NO_BUSY_POLL and depend on
free_netdev() to call netif_napi_del(). This seems inevitable
since we want to use RCU for netpoll dev->napi_list traversal,
and almost no drivers set IFF_DISABLE_NETPOLL.
Signed-off-by: NJakub Kicinski <kuba@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>