The issue was found on BCM5718 which has two NCSI channels in one
package: C0 and C1. C0 is in link-up state while C1 is in link-down
state. C0 is chosen as active channel until unplugging and plugging
C0's cable: On unplugging C0's cable, LSC (Link State Change) AEN
packet received on C0 to report link-down event. After that, C1 is
chosen as active channel. LSC AEN for link-up event is lost on C0
when plugging C0's cable back. We lose the network even C0 is usable.
This resolves the issue by recording the (hot) channel that was ever
chosen as active one. The hot channel is chosen to be active one
if none of available channels in link-up state. With this, C0 is still
the active one after unplugging C0's cable. LSC AEN packet received
on C0 when plugging its cable back.
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The issue was found on BCM5718 which has two NCSI channels in one
package: C0 and C1. Both of them are connected to different LANs,
means they are in link-up state and C0 is chosen as the active one
until resetting BCM5718 happens as below.
Resetting BCM5718 results in LSC (Link State Change) AEN packet
received on C0, meaning LSC AEN is missed on C1. When LSC AEN packet
received on C0 to report link-down, it fails over to C1 because C1
is in link-up state as software can see. However, C1 is in link-down
state in hardware. It means the link state is out of synchronization
between hardware and software, resulting in inappropriate channel (C1)
selected as active one.
This resolves the issue by sending separate GLS (Get Link Status)
commands to all channels in the package before trying to do failover.
The last link states of all channels in the package are retrieved.
With it, C0 (not C1) is selected as active one as expected.
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are several if/else statements in the state machine implemented
by switch/case in ncsi_suspend_channel() to avoid duplicated code. It
makes the code a bit hard to be understood.
This drops if/else statements in ncsi_suspend_channel() to improve the
code readability as Joel Stanley suggested. Also, it becomes easy to
add more states in the state machine without affecting current code.
No logical changes introduced by this.
Suggested-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This introduces ncsi_stop_dev(), as counterpart to ncsi_start_dev(),
to stop the NCSI device so that it can be reenabled in future. This
API should be called when the network device driver is going to
shutdown the device. There are 3 things done in the function: Stop
the channel monitoring; Reset channels to inactive state; Report
NCSI link down.
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
The original NCSI channel monitoring was implemented based on a
backoff algorithm: the GLS response should be received in the
specified interval. Otherwise, the channel is regarded as dead
and failover should be taken if current channel is an active one.
There are several problems in the implementation: (A) On BCM5718,
we found when the IID (Instance ID) in the GLS command packet
changes from 255 to 1, the response corresponding to IID#1 never
comes in. It means we cannot make the unfair judgement that the
channel is dead when one response is missed. (B) The code's
readability should be improved. (C) We should do failover when
current channel is active one and the channel monitoring should
be marked as disabled before doing failover.
This reworks the channel monitoring to address all above issues.
The fields for channel monitoring is put into separate struct
and the state of channel monitoring is predefined. The channel
is regarded alive if the network controller responses to one of
two GLS commands or both of them in 5 seconds.
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is only one NCSI request property for now: the response for
the sent command need drive the workqueue or not. So we had one
field (@driven) for the purpose. We lost the flexibility to extend
NCSI request properties.
This replaces @driven with @flags and @req_flags in NCSI request
and NCSI command argument struct. Each bit of the newly introduced
field can be used for one property. No functional changes introduced.
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
The NCSI request index (struct ncsi_request::id) is put into instance
ID (IID) field while sending NCSI command packet. It was designed the
available IDs are given in round-robin fashion. @ndp->request_id was
introduced to represent the next available ID, but it has been used
as number of successively allocated IDs. It breaks the round-robin
design. Besides, we shouldn't put 0 to NCSI command packet's IID
field, meaning ID#0 should be reserved according section 6.3.1.1
in NCSI spec (v1.1.0).
This fixes above two issues. With it applied, the available IDs will
be assigned in round-robin fashion and ID#0 won't be assigned.
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
We needn't send CIS (Clear Initial State) command to the NCSI
reserved channel (0x1f) in the enumeration. We shouldn't receive
a valid response from CIS on NCSI channel 0x1f.
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This defines NCSI_RESERVED_CHANNEL as the reserved NCSI channel
ID (0x1f). No logical changes introduced.
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
xchg() is used to set NCSI channel's state in order for consistent
access to the state. xchg()'s return value should be used. Otherwise,
one build warning will be raised (with -Wunused-value) as below message
indicates. It is reported by ia64-linux-gcc (GCC) 4.9.0.
net/ncsi/ncsi-manage.c: In function 'ncsi_channel_monitor':
arch/ia64/include/uapi/asm/cmpxchg.h:56:2: warning: value computed is \
not used [-Wunused-value]
((__typeof__(*(ptr))) __xchg((unsigned long) (x), (ptr), sizeof(*(ptr))))
^
net/ncsi/ncsi-manage.c:202:3: note: in expansion of macro 'xchg'
xchg(&nc->state, NCSI_CHANNEL_INACTIVE);
This removes the atomic access to NCSI channel's state avoid the above
build warning. We have to hold the channel's lock when its state is readed
or updated. No functional changes introduced.
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Reviewed-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
gcc-4.9 and higher warn about the newly added NSCI code:
net/ncsi/ncsi-manage.c: In function 'ncsi_process_next_channel':
net/ncsi/ncsi-manage.c:1003:2: error: 'old_state' may be used uninitialized in this function [-Werror=maybe-uninitialized]
The warning is a false positive and therefore harmless, but it would be good to
avoid it anyway. I have determined that the barrier in the spin_unlock_irqsave()
is what confuses gcc to the point that it cannot track whether the variable
was unused or not.
This rearranges the code in a way that makes it obvious to gcc that old_state
is always initialized at the time of use, functionally this should not
change anything.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This manages NCSI packages and channels:
* The available packages and channels are enumerated in the first
time of calling ncsi_start_dev(). The channels' capabilities are
probed in the meanwhile. The NCSI network topology won't change
until the NCSI device is destroyed.
* There in a queue in every NCSI device. The element in the queue,
channel, is waiting for configuration (bringup) or suspending
(teardown). The channel's state (inactive/active) indicates the
futher action (configuration or suspending) will be applied on the
channel. Another channel's state (invisible) means the requested
action is being applied.
* The hardware arbitration will be enabled if all available packages
and channels support it. All available channels try to provide
service when hardware arbitration is enabled. Otherwise, one channel
is selected as the active one at once.
* When channel is in active state, meaning it's providing service, a
timer started to retrieve the channe's link status. If the channel's
link status fails to be updated in the determined period, the channel
is going to be reconfigured. It's the error handling implementation
as defined in NCSI spec.
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Acked-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
NCSI spec (DSP0222) defines several objects: package, channel, mode,
filter, version and statistics etc. This introduces the data structs
to represent those objects and implement functions to manage them.
Also, this introduces CONFIG_NET_NCSI for the newly implemented NCSI
stack.
* The user (e.g. netdev driver) dereference NCSI device by
"struct ncsi_dev", which is embedded to "struct ncsi_dev_priv".
The later one is used by NCSI stack internally.
* Every NCSI device can have multiple packages simultaneously, up
to 8 packages. It's represented by "struct ncsi_package" and
identified by 3-bits ID.
* Every NCSI package can have multiple channels, up to 32. It's
represented by "struct ncsi_channel" and identified by 5-bits ID.
* Every NCSI channel has version, statistics, various modes and
filters. They are represented by "struct ncsi_channel_version",
"struct ncsi_channel_stats", "struct ncsi_channel_mode" and
"struct ncsi_channel_filter" separately.
* Apart from AEN (Asynchronous Event Notification), the NCSI stack
works in terms of command and response. This introduces "struct
ncsi_req" to represent a complete NCSI transaction made of NCSI
request and response.
link: https://www.dmtf.org/sites/default/files/standards/documents/DSP0222_1.1.0.pdf
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Acked-by: Joel Stanley <joel@jms.id.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Gavin Shan
Arnd Bergmann