The retransmission and F-RTO transmission currently happen inside
recovery state processing (tcp_fastretrans_alert) but before
congestion control.  This refactoring moves the logic after both
s.t. we can determine how much to send (cwnd) before deciding what to
send.
Signed-off-by: NYuchung Cheng <ycheng@google.com>
Signed-off-by: NNeal Cardwell <ncardwell@google.com>
Signed-off-by: NEric Dumazet <ncardwell@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Remove a write-only stack variable from unix_attach_fds(). This is a
left-over from the security fix in:

    commit 712f4aad
    Author: willy tarreau <w@1wt.eu>
    Date:   Sun Jan 10 07:54:56 2016 +0100

        unix: properly account for FDs passed over unix sockets
Signed-off-by: NDavid Herrmann <dh.herrmann@gmail.com>
Acked-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Allows userspace to have direct access to VRF table association
versus looking up master device and its table.
Signed-off-by: NDavid Ahern <dsa@cumulusnetworks.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

My recent patch to the Xen Project documents a protocol for 'dynamic
multicast control' in netif.h. This extends the previous multicast control
protocol to not require a shared ring reconnection to turn the feature off.
Instead the backend watches the "request-multicast-control" key in xenstore
and turns the feature off if the key value is written to zero.

This patch adds support for dynamic multicast control in xen-netback.
Signed-off-by: NPaul Durrant <paul.durrant@citrix.com>
Cc: Ian Campbell <ian.campbell@citrix.com>
Cc: Wei Liu <wei.liu2@citrix.com>
Acked-by: NWei Liu <wei.liu2@citrix.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Sriharsha Basavapatna says:

====================
be2net patch-set

v2 changes:
	Patch-4:	Changed a tab to space in be.h
	Patches-6,7,8:	Updated commit log summary line: benet --> be2net

Hi David,

The following patch set contains a few non-critical bug fixes. Please
consider applying this to the net-next tree. Thanks.

Patch-1 fixes be_set_phys_id() ethtool function to return an error code.
Patch-2 fixes a warning when some commands fail for VFs.
Patch-3 fixes be_vlan_rem_vid() to verify vlan being removed is in the list.
Patch-4 improves SRIOV queue distribution logic.
Patch-5 avoids running self test on VFs.
Patch-6 fixes error recovery in Lancer to clean up after moving to ready state.
Patch-7 adds retry logic to error recovery in case of recovery failures
Patch-8 fixes time interval used in eq delay computation routine
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Interrupt moderation parameters need to be recalculated only
after a time interval of 1 ms. Interval calculation is wrong
when there is a rollover of jiffies. Using recommended way of interval
calculation using jiffies to fix this.
Signed-off-by: NPadmanabh Ratnakar <padmanabh.ratnakar@broadcom.com>
Signed-off-by: NSriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Retry error recovery MAX_ERR_RECOVERY_RETRY_COUNT times in case of
failure during error recovery.
Signed-off-by: NPadmanabh Ratnakar <padmanabh.ratnakar@broadcom.com>
Signed-off-by: NSriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

After error is detected, wait for adapter to move to ready state
before destroying queues and cleanup of other resources. Also
skip performing any cleanup for non-Lancer chips and move debug
messages to correct routine.
Signed-off-by: NPadmanabh Ratnakar <padmanabh.ratnakar@broadcom.com>
Signed-off-by: NSriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The CMD_SUBSYSTEM_LOWLEVEL cmds need DEV_CFG Privilege to run
which VFs don't have by default.
Self-tests need to be issued only for PFs.
Signed-off-by: NSomnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: NSriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The SRIOV resource distribution logic for RX/TX queue counts is not optimal
when a small number of VFs are enabled. It does not take into account the
VF's EQ count while computing the queue counts. Because of this, the VF
gets a large number of queues, though it doesn't have sufficient EQs,
resulting in wasted queue resources. And the PF gets a smaller share of
queues though it has more EQs. Fix this by capping the VF queue count at
its EQ count.
Signed-off-by: NSriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The driver decrements its vlan count without checking if it is really
present in its list. This results in an invalid vlan count and impacts
subsequent vlan add/rem ops. The function be_vlan_rem_vid() should be
updated to fix this.
Signed-off-by: NSriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The driver currently logs the message "VF is not privileged to issue
opcode" by checking only the base_status field for UNAUTHORIZED_REQUEST.
Add check to look for INSUFFICIENT_PRIVILEGES in the additional status
field also as not all cmds fail with that base status.
Signed-off-by: NSuresh Reddy <suresh.reddy@broadcom.com>
Signed-off-by: NSriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

be_set_phys_id() returns 0 to ethtool when the command fails in the FW.

This patch fixes the set_phys_id() to return -EIO in case the FW cmd fails.
Signed-off-by: NSuresh Reddy <suresh.reddy@broadcom.com>
Signed-off-by: NSriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Jiri Benc says:

====================
vxlan: cleanup headers and tx path

This is first part of vxlan cleanup. It makes vxlan.h better organized and
removes code duplication in the tx path. There's no change in functionality.

This is in preparation for VXLAN-GPE support. Other sets will follow with
cleanup of rx path, modifications of vxlan interface setup and
cleanup/modification of fdb handling. The goal of these patchsets is to
consolidate VXLAN extension handling (right now it's scattered all around
the code) and allow vxlan to operate in L3 mode (i.e. without inner Ethernet
headers).

The full picture (all 30 patches) can be seen at:
https://github.com/jbenc/linux-vxlan/commits/master

Changelog:
v2: added patch 5/6, fixing the udp sum problem spotted by Paolo Abeni
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

There's a lot of code duplication. Factor out the duplicate code to a new
function shared between IPv4 and IPv6 xmit path.
Signed-off-by: NJiri Benc <jbenc@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The flag for tx checksumming for tunneling over IPv4 and IPv6 is different.
Decide whether to do tx checksumming in vxlan_xmit_one and pass it on as
a separate flag. This will allow for tx path consolidation in the next
patch.

Unfortunately, gcc is not clever enough to see that udp_sum is always
initialized and gives an uninitialized variable warning. Set it to false to
silence the warning.
Signed-off-by: NJiri Benc <jbenc@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The code for output route lookup is duplicated for ndo_start_xmit and
ndo_fill_metadata_dst. Move it to a common function.
Signed-off-by: NJiri Benc <jbenc@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

RCO and GBP are VXLAN extensions, not specified in RFC 7348. Because of
that, they need to be explicitly enabled when creating vxlan interface. By
default, those extensions are not used and plain VXLAN header is sent and
received.

Reflect this in vxlan.h: first, the plain VXLAN header is defined. Following
it, RCO is documented and defined, and likewise for GBP.
Signed-off-by: NJiri Benc <jbenc@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

VNI_HASH_BITS and VNI_HASH_SIZE are defined twice. Remove the extra
definitions.
Signed-off-by: NJiri Benc <jbenc@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

include/net/vxlan.h is a kernel header, no need to prefix fixed size types
with double underscore.
Signed-off-by: NJiri Benc <jbenc@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When we acknowledge a FIN, it is not enough to ack the sequence number
and queue the skb into receive queue. We also have to call tcp_fin()
to properly update socket state and send proper poll() notifications.

It seems we also had the problem if we received a SYN packet with the
FIN flag set, but it does not seem an urgent issue, as no known
implementation can do that.

Fixes: 61d2bcae ("tcp: fastopen: accept data/FIN present in SYNACK message")
Signed-off-by: NEric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Parthasarathy Bhuvaragan says:

====================
tipc: cleanups, fixes & improvements for topology server

This series contains topology server cleanups, fixes and improvements.

Cleanups in #1-#4:
We remove duplicate data structures and aligin the rest of the code accordingly.

Fixes in #5-#8:
The bugs occur either during configuration or while running on SMP targets,
which are race conditions that pop up under different situations.

Improvements in #9-#10:
Updates to decrease timer usage and improve readability.

v2: Updated commit message in patch 6 based on feedback from
    Sergei Shtylyov sergei.shtylyov@cogentembedded.com
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Until now, tipc_rcv and tipc_send workqueues in server are allocated
with parameters WQ_UNBOUND & max_active = 1.
This parameters passed to this function makes it equivalent to
alloc_ordered_workqueue(). The later form is more explicit and
can inherit future ordered_workqueue changes.

In this commit we replace alloc_workqueue() with more readable
alloc_ordered_workqueue().
Acked-by: NYing Xue <ying.xue@windriver.com>
Reviewed-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Until now, we create timers even for the subscription requests
with timeout = TIPC_WAIT_FOREVER.
This can be improved by avoiding timer creation when the timeout
is set to TIPC_WAIT_FOREVER.

In this commit, we introduce a check to creates timers only
when timeout != TIPC_WAIT_FOREVER.
Acked-by: NYing Xue <ying.xue@windriver.com>
Reviewed-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Until now, during subscription creation the mod_time() &
tipc_subscrb_get() are called after releasing the subscriber
spin lock.

In a SMP system when performing a subscription creation, if the
subscription timeout occurs simultaneously (the timer is
scheduled to run on another CPU) then the timer thread
might decrement the subscribers refcount before the create
thread increments the refcount.

This can be simulated by creating subscription with timeout=0 and
sometimes the timeout occurs before the create request is complete.
This leads to the following message:
[30.702949] BUG: spinlock bad magic on CPU#1, kworker/u8:3/87
[30.703834] general protection fault: 0000 [#1] SMP
[30.704826] CPU: 1 PID: 87 Comm: kworker/u8:3 Not tainted 4.4.0-rc8+ #18
[30.704826] Workqueue: tipc_rcv tipc_recv_work [tipc]
[30.704826] task: ffff88003f878600 ti: ffff88003fae0000 task.ti: ffff88003fae0000
[30.704826] RIP: 0010:[<ffffffff8109196c>]  [<ffffffff8109196c>] spin_dump+0x5c/0xe0
[...]
[30.704826] Call Trace:
[30.704826]  [<ffffffff81091a16>] spin_bug+0x26/0x30
[30.704826]  [<ffffffff81091b75>] do_raw_spin_lock+0xe5/0x120
[30.704826]  [<ffffffff81684439>] _raw_spin_lock_bh+0x19/0x20
[30.704826]  [<ffffffffa0096f10>] tipc_subscrb_rcv_cb+0x1d0/0x330 [tipc]
[30.704826]  [<ffffffffa00a37b1>] tipc_receive_from_sock+0xc1/0x150 [tipc]
[30.704826]  [<ffffffffa00a31df>] tipc_recv_work+0x3f/0x80 [tipc]
[30.704826]  [<ffffffff8106a739>] process_one_work+0x149/0x3c0
[30.704826]  [<ffffffff8106aa16>] worker_thread+0x66/0x460
[30.704826]  [<ffffffff8106a9b0>] ? process_one_work+0x3c0/0x3c0
[30.704826]  [<ffffffff8106a9b0>] ? process_one_work+0x3c0/0x3c0
[30.704826]  [<ffffffff8107029d>] kthread+0xed/0x110
[30.704826]  [<ffffffff810701b0>] ? kthread_create_on_node+0x190/0x190
[30.704826]  [<ffffffff81684bdf>] ret_from_fork+0x3f/0x70

In this commit,
1. we remove the check for the return code for mod_timer()
2. we protect tipc_subscrb_get() using the subscriber spin lock.
   We increment the subscriber's refcount as soon as we add the
   subscription to subscriber's subscription list.
Acked-by: NYing Xue <ying.xue@windriver.com>
Reviewed-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Until now, while creating a subscription the subscriber lock
protects only the subscribers subscription list and not the
nametable. The call to tipc_nametbl_subscribe() is outside
the lock. However, at subscription timeout and cancel both
the subscribers subscription list and the nametable are
protected by the subscriber lock.

This asymmetric locking mechanism leads to the following problem:
In a SMP system, the timer can be fire on another core before
the create request is complete.
When the timer thread calls tipc_nametbl_unsubscribe() before create
thread calls tipc_nametbl_subscribe(), we get a nullptr exception.

This can be simulated by creating subscription with timeout=0 and
sometimes the timeout occurs before the create request is complete.

The following is the oops:
[57.569661] BUG: unable to handle kernel NULL pointer dereference at (null)
[57.577498] IP: [<ffffffffa02135aa>] tipc_nametbl_unsubscribe+0x8a/0x120 [tipc]
[57.584820] PGD 0
[57.586834] Oops: 0002 [#1] SMP
[57.685506] CPU: 14 PID: 10077 Comm: kworker/u40:1 Tainted: P OENX 3.12.48-52.27.1.     9688.1.PTF-default #1
[57.703637] Workqueue: tipc_rcv tipc_recv_work [tipc]
[57.708697] task: ffff88064c7f00c0 ti: ffff880629ef4000 task.ti: ffff880629ef4000
[57.716181] RIP: 0010:[<ffffffffa02135aa>]  [<ffffffffa02135aa>] tipc_nametbl_unsubscribe+0x8a/   0x120 [tipc]
[...]
[57.812327] Call Trace:
[57.814806]  [<ffffffffa0211c77>] tipc_subscrp_delete+0x37/0x90 [tipc]
[57.821357]  [<ffffffffa0211e2f>] tipc_subscrp_timeout+0x3f/0x70 [tipc]
[57.827982]  [<ffffffff810618c1>] call_timer_fn+0x31/0x100
[57.833490]  [<ffffffff81062709>] run_timer_softirq+0x1f9/0x2b0
[57.839414]  [<ffffffff8105a795>] __do_softirq+0xe5/0x230
[57.844827]  [<ffffffff81520d1c>] call_softirq+0x1c/0x30
[57.850150]  [<ffffffff81004665>] do_softirq+0x55/0x90
[57.855285]  [<ffffffff8105aa35>] irq_exit+0x95/0xa0
[57.860290]  [<ffffffff815215b5>] smp_apic_timer_interrupt+0x45/0x60
[57.866644]  [<ffffffff8152005d>] apic_timer_interrupt+0x6d/0x80
[57.872686]  [<ffffffffa02121c5>] tipc_subscrb_rcv_cb+0x2a5/0x3f0 [tipc]
[57.879425]  [<ffffffffa021c65f>] tipc_receive_from_sock+0x9f/0x100 [tipc]
[57.886324]  [<ffffffffa021c826>] tipc_recv_work+0x26/0x60 [tipc]
[57.892463]  [<ffffffff8106fb22>] process_one_work+0x172/0x420
[57.898309]  [<ffffffff8107079a>] worker_thread+0x11a/0x3c0
[57.903871]  [<ffffffff81077114>] kthread+0xb4/0xc0
[57.908751]  [<ffffffff8151f318>] ret_from_fork+0x58/0x90

In this commit, we do the following at subscription creation:
1. set the subscription's subscriber pointer before performing
   tipc_nametbl_subscribe(), as this value is required further in
   the call chain ex: by tipc_subscrp_send_event().
2. move tipc_nametbl_subscribe() under the scope of subscriber lock
Acked-by: NYing Xue <ying.xue@windriver.com>
Reviewed-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Until now, the subscribers endianness for a subscription
create/cancel request is determined as:
    swap = !(s->filter & (TIPC_SUB_PORTS | TIPC_SUB_SERVICE))
The checks are performed only for port/service subscriptions.

The swap calculation is incorrect if the filter in the subscription
cancellation request is set to TIPC_SUB_CANCEL (it's a malformed
cancel request, as the corresponding subscription create filter
is missing).
Thus, the check if the request is for cancellation fails and the
request is treated as a subscription create request. The
subscription creation fails as the request is illegal, which
terminates this connection.

In this commit we determine the endianness by including
TIPC_SUB_CANCEL, which will set swap correctly and the
request is processed as a cancellation request.
Acked-by: NYing Xue <ying.xue@windriver.com>
Reviewed-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In 'commit 7fe8097c ("tipc: fix nullpointer bug when subscribing
to events")', we terminate the connection if the subscription
creation fails.
In the same commit, the subscription creation result was based on
the value of subscription pointer (set in the function) instead of
the return code.

Unfortunately, the same function also handles subscription
cancellation request. For a subscription cancellation request,
the subscription pointer cannot be set. Thus the connection is
terminated during cancellation request.

In this commit, we move the subcription cancel check outside
of tipc_subscrp_create(). Hence,
- tipc_subscrp_create() will create a subscripton
- tipc_subscrb_rcv_cb() will subscribe or cancel a subscription.

Fixes: 'commit 7fe8097c ("tipc: fix nullpointer bug when subscribing to events")'
Acked-by: NYing Xue <ying.xue@windriver.com>
Reviewed-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In this commit, we split tipc_subscrp_create() into two:
1. tipc_subscrp_create() creates a subscription
2. A new function tipc_subscrp_subscribe() adds the
   subscription to the subscriber subscription list,
   activates the subscription timer and subscribes to
   the nametable updates.

In future commits, the purpose of tipc_subscrb_rcv_cb() will
be to either subscribe or cancel a subscription.

There is no functional change in this commit.
Acked-by: NYing Xue <ying.xue@windriver.com>
Reviewed-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Until now, struct tipc_subscriber has duplicate fields for
type, upper and lower (as member of struct tipc_name_seq) at:
1. as member seq in struct tipc_subscription
2. as member seq in struct tipc_subscr, which is contained
   in struct tipc_event
The former structure contains the type, upper and lower
values in network byte order and the later contains the
intact copy of the request.
The struct tipc_subscription contains a field swap to
determine if request needs network byte order conversion.
Thus by using swap, we can convert the request when
required instead of duplicating it.

In this commit,
1. we remove the references to these elements as members of
   struct tipc_subscription and replace them with elements
   from struct tipc_subscr.
2. provide new functions to convert the user request into
   network byte order.
Acked-by: NYing Xue <ying.xue@windriver.com>
Reviewed-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Until now, struct tipc_subscription has duplicate timeout and filter
attributes present:
1. directly as members of struct tipc_subscription
2. in struct tipc_subscr, which is contained in struct tipc_event

In this commit, we remove the references to these elements as
members of struct tipc_subscription and replace them with elements
from struct tipc_subscr.
Acked-by: NYing Xue <ying.xue@windriver.com>
Reviewed-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Until now, during subscription creation we set sub->timeout by
converting the timeout request value in milliseconds to jiffies.
This is followed by setting the timeout value in the timer if
sub->timeout != TIPC_WAIT_FOREVER.

For a subscription create request with a timeout value of
TIPC_WAIT_FOREVER, msecs_to_jiffies(TIPC_WAIT_FOREVER)
returns MAX_JIFFY_OFFSET (0xfffffffe). This is not equal to
TIPC_WAIT_FOREVER (0xffffffff).

In this commit, we remove this check.
Acked-by: NYing Xue <ying.xue@windriver.com>
Reviewed-by: NJon Maloy <jon.maloy@ericsson.com>
Signed-off-by: NParthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

netdev_dbg() will add bond device name, it will be helpful if we print
slave device name.
Signed-off-by: NZhang Shengju <zhangshengju@cmss.chinamobile.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Chipsets with BCM4707 / BCM53018 ID require special handling at a few
places in the code. It's likely there will be more IDs to check in the
future. To simplify it add this trivial helper.
Signed-off-by: NRafał Miłecki <zajec5@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Alexei Starovoitov says:

====================
bpf: introduce per-cpu maps

We've started to use bpf to trace every packet and atomic add
instruction (event JITed) started to show up in perf profile.
The solution is to do per-cpu counters.
For PERCPU_(HASH|ARRAY) map the existing bpf_map_lookup() helper
returns per-cpu area which bpf programs can use to store and
increment the counters. The BPF_MAP_LOOKUP_ELEM syscall command
returns areas from all cpus and user process aggregates the counters.
The usage example is in patch 6. The api turned out to be very
easy to use from bpf program and from user space.
Long term we were discussing to add 'bounded loop' instruction,
so bpf programs can do aggregation within the program which may
help some use cases. Right now user space aggregation of
per-cpu counters fits the best.

This patch set is new approach for per-cpu hash and array maps.
I've reused the map tests written by Martin and Ming, but
implementation and api is new. Old discussion here:
http://thread.gmane.org/gmane.linux.kernel/2123800/focus=2126435
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

A sanity test for BPF_MAP_TYPE_PERCPU_ARRAY
Signed-off-by: NMing Lei <tom.leiming@gmail.com>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

A sanity test for BPF_MAP_TYPE_PERCPU_HASH.
Signed-off-by: NMartin KaFai Lau <kafai@fb.com>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The functions bpf_map_lookup_elem(map, key, value) and
bpf_map_update_elem(map, key, value, flags) need to get/set
values from all-cpus for per-cpu hash and array maps,
so that user space can aggregate/update them as necessary.

Example of single counter aggregation in user space:
  unsigned int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
  long values[nr_cpus];
  long value = 0;

  bpf_lookup_elem(fd, key, values);
  for (i = 0; i < nr_cpus; i++)
    value += values[i];

The user space must provide round_up(value_size, 8) * nr_cpus
array to get/set values, since kernel will use 'long' copy
of per-cpu values to try to copy good counters atomically.
It's a best-effort, since bpf programs and user space are racing
to access the same memory.
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Primary use case is a histogram array of latency
where bpf program computes the latency of block requests or other
events and stores histogram of latency into array of 64 elements.
All cpus are constantly running, so normal increment is not accurate,
bpf_xadd causes cache ping-pong and this per-cpu approach allows
fastest collision-free counters.
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>