The new address list lock needs to handle the same device layering
issues that the _xmit_lock one does.
This integrates work done by Patrick McHardy.
Signed-off-by: David S. Miller <davem@davemloft.net>
alloc_netdev_mq() now allocates an array of netdev_queue
structures for TX, based upon the queue_count argument.
Furthermore, all accesses to the TX queues are now vectored
through the netdev_get_tx_queue() and netdev_for_each_tx_queue()
interfaces. This makes it easy to grep the tree for all
things that want to get to a TX queue of a net device.
Problem spots which are not really multiqueue aware yet, and
only work with one queue, can easily be spotted by grepping
for all netdev_get_tx_queue() calls that pass in a zero index.
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that we have a specific lock to protect the network
device unicast and multicast lists, remove extraneous
grabs of the TX lock in cases where the code only needs
address list protection.
Signed-off-by: David S. Miller <davem@davemloft.net>
Add netif_addr_{lock,unlock}{,_bh}() helpers.
Use them to protect operations that operate on or read
the network device unicast and multicast address lists.
Also use them in cases where the code simply wants to
block calls into the driver's ->set_rx_mode() and
->set_multicast_list() methods.
Signed-off-by: David S. Miller <davem@davemloft.net>
dev_set_promiscuity/allmulti might overflow.
Commit: "netdevice: Fix promiscuity and allmulti overflow" in net-next makes
dev_set_promiscuity/allmulti return error number if overflow happened.
In bond_alb and bond_main, we check all positive increment for promiscuity
and allmulti to get error return.
But there are still two problems left.
1. Some code path has no mechanism to signal errors upstream.
2. If there are multi slaves, it's hard to tell which slaves increment
promisc/allmulti successfully and which failed.
So I left these problems to be FIXME.
Fortunately, the overflow is very rare case.
Signed-off-by: Wang Chen <wangchen@cn.fujitsu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Accesses are mostly structured such that when there are multiple TX
queues the code transformations will be a little bit simpler.
Signed-off-by: David S. Miller <davem@davemloft.net>
Permit bonding to function rationally if max_bonds is set to
zero. This will load the module, but create no master devices (which can
be created via sysfs).
Requires some change to bond_create_sysfs; currently, the
netdev sysfs directory is determined from the first bonding device created,
but this is no longer possible. Instead, an interface from net/core is
created to create and destroy files in net_class.
Based on a patch submitted by Phil Oester <kernel@linuxaces.com>.
Modified by Jay Vosburgh to fix the sysfs issue mentioned above and to
update the documentation.
Signed-off-by: Phil Oester <kernel@linuxace.com>
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Support for sending multiple gratuitous ARPs during failovers
was added by commit:
commit 7893b2491a
Author: Moni Shoua <monis@voltaire.com>
Date: Sat May 17 21:10:12 2008 -0700
bonding: Send more than one gratuitous ARP when slave takes over
This change modifies that support to remove duplicated code,
add support for ARP monitor (the original only supported miimon), clear
the grat ARP counter in bond_close (lest a later "ifconfig up" immediately
start spewing ARPs), and add documentation for the module parameter.
Also updated driver version to 3.3.0.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
under active-backup mode and when there's actual new_active slave,
have bond_change_active_slave() call the networking core to deliver
NETDEV_BONDING_FAILOVER event such that the fail-over can be notable
by code outside of the bonding driver such as the RDMA stack and
monitoring tools.
As the correct context of locking appropriate for notifier calls is RTNL
and nothing else, bond->curr_slave_lock and bond->lock are unlocked and
later locked again. This is ensured by the rest of the code to be safe
under backup-mode AND when new_active is not NULL.
Jay Vosburgh modified the original patch for formatting and fixed a
compiler error.
Signed-off-by: Or Gerlitz <ogerlitz@voltaire.com>
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
simplified the code of bond_change_active_slave() such that under
active-backup mode there's one "if (new_active)" test and the rest
of the code only does extra checks on top of it. This removed an
unneeded "if (bond->send_grat_arp > 0)" check and avoid calling
bond_send_gratuitous_arp when there's no active slave.
Jay Vosburgh made minor coding style changes to the orignal patch.
Signed-off-by: Or Gerlitz <ogerlitz@voltaire.com>
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Add a "follow" selection for fail_over_mac. This option
causes the MAC address to move from slave to slave as the active
slave changes. This is in addition to the existing fail_over_mac option
that causes the bond's MAC address to change during failover.
This new option is useful for devices that cannot tolerate
multiple ports using the same MAC address simultaneously, either
because it confuses them or incurs a performance penalty (as is the
case with some LPAR-aware multiport devices). Because the MAC of the
bond itself does not change, the "follow" option is slightly more
reliable during failover and doesn't change the MAC of the bond during
operation.
This patch requires a previous ARP monitor change to properly
handle RTNL during failovers.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Refactor ARP monitor for active-backup mode. The motivation for
this is to take care of locking issues in a clear manner (particularly to
correctly handle RTNL vs. the bonding locks). Currently, the a-b ARP
monitor does not hold RTNL at all, but future changes will require RTNL
during ARP monitor failovers.
Rather than using conditional locking, this patch instead breaks
up the ARP monitor into three discrete steps: inspection, commit changes,
and probe. The inspection phase marks slaves that require link state
changes. The commit phase is only called if inspection detects that
changes are needed, and is called with RTNL. Lastly, the probe phase
issues the ARP probes that the inspection phase uses to determine link
state.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
With IPoIB, reception of gratuitous ARP by neighboring hosts
is essential for a successful change of slaves in case of failure.
Otherwise, they won't learn about the HW address change and need
to wait a long time until the neighboring system gives up and sends
an ARP request to learn the new HW address. This patch decreases
the chance for a lost of a gratuitous ARP packet by sending it more
than once. The number retries is configurable and can be set with a
module param.
Signed-off-by: Moni Shoua <monis@voltaire.com>
Acked-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Some places iterate over the checked list right after the check
itself, so even if the list is empty, the list_for_each_xxx
iterator will make everything right by himself.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Acked-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Many places either do not modify the list under the list_for_each_xxx,
or break out of the loop as soon as the first element is removed.
Thus, this _safe iteration just occupies some unneeded .text space
and requires an additional variable.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Acked-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
While we're fixing the bond_create, I hope it's OK to polish it
a bit after the fixes.
The third argument is NULL at the first caller and is ignored by
the second one, so remove it.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Acked-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Remove bond_has_ip and all references to it. With this change,
the ARP monitor will always send ARP probes if the master is up and has
at least one slave. If the bond has an IP address, it is used in the
ARP probe; if not, the probes are sent with all zeros in the sender's
IP address (which is consistent with an RFC 2131 4.4.1 duplicate address
probe).
This is useful for cases when bonding itself is hidden underneath
a layer of virtual devices, e.g., with Xen.
Change suggested by Tsutomu Fujii <t-fujii@nb.jp.nec.com>, who
included a one-line patch that only affected active-backup mode.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Convert bonding to use msecs_to_jiffies instead of doing the
math. For the ARP monitor, there was an underflow problem that could
result in an infinite loop. The miimon already had that worked around,
but this is cleaner.
Originally by Nicolas de Pesloüan <nicolas.2p.debian@free.fr>
Jay Vosburgh corrected a math error in the original; Nicolas' original
commit message is:
When setting arp_interval parameter to a very low value, delta_in_ticks
for next arp might become 0, causing an infinite loop.
See http://bugzilla.kernel.org/show_bug.cgi?id=10680
Same problem for miimon parameter already fixed, but fix might be
enhanced, by using msecs_to_jiffies() function.
Signed-off-by: Nicolas de Pesloüan <nicolas.2p.debian@free.fr>
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
If device already exists named bonding_masters, then fail. This is a wierd
corner case only a QA group could love.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As part of:
commit c2edacf80e
Author: Jay Vosburgh <fubar@us.ibm.com>
Date: Mon Jul 9 10:42:47 2007 -0700
bonding / ipv6: no addrconf for slaves separately from master
two steps were rearranged in the enslavement process: netdev_set_master
is now before the call to dev_open to open the slave.
This patch updates the error cases and unwind process at the
end of bond_enslave to match the new order. Without this patch, it is
possible for the enslavement to fail, but leave the slave with IFF_SLAVE
set in its flags.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
The sysfs layer has an internal protection, that ensures, that
all the process sitting inside ->sore/->show callback exits
before the appropriate entry is unregistered (the calltraces
are rather big, but I can provide them if required).
On the other hand, bonding takes rtnl_lock in
a) the bonding_store_bonds, i.e. in ->store callback,
b) module exit before calling the sysfs unregister routines.
Thus, the classical AB-BA deadlock may occur. To reproduce run
# while :; do modprobe bonding; rmmod bonding; done
and
# while :; do echo '+bond%d' > /sys/class/net/bonding_masters ; done
in parallel.
The fix is to move the bond_destroy_sysfs out of the rtnl_lock,
but _before_ bond_free_all to make sure no bonding devices exist
after module unload.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Acked-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Fixed an error unwind in bonding_store_bonds that didn't release
the locks it held, and consolidated unwinds into a common block at the
end of the function. Bug reported by Pavel Emelyanov <xemul@openvz.org>,
who provided a different fix.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
If the call to bond_create_sysfs_entry in bond_create fails, the
proper rollback is to call unregister_netdevice, not free_netdev.
Otherwise - kernel BUG at net/core/dev.c:4057!
Checked with artificial failures injected into bond_create_sysfs_entry.
Pavel's original patch modified by Jay Vosburgh to move code around
for clarity (remove goto-hopping within the unwind block).
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
Use proc_create()/proc_create_data() to make sure that ->proc_fops and ->data
be setup before gluing PDE to main tree.
Signed-off-by: Denis V. Lunev <den@openvz.org>
Cc: Jeff Garzik <jgarzik@pobox.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
original_mtu is only used if we end up with a non-NULL
dev, and it is assigned in all such cases, but GCC can't
see that.
Signed-off-by: David S. Miller <davem@davemloft.net>
For bonding interfaces any attempt to read the sysfs directory contents after
module removal results in an oops. The fix is to release sysfs attributes
for the interfaces upon module unload.
Signed-off-by: Libor Pechacek <lpechacek@suse.cz>
Acked-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Fix two compiler warnings that are new with recent versions of gcc
(apparently 4.2 and up). One is fixed by refactoring; this change was
supplied by Stephen Hemminger. The other was fixed by labelling the
variable as uninitialized_var() after confirming via inspection that it
cannot actually be used uninitialized.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
The 802.3ad state machine lock can be acquired in both softirq and
not softirq context, but was not held at _bh to prevent a deadlock (which
could occur if a LACPDU arrived and was processed while the lock was
held).
Corrected this, now hold the state machine lock at _bh to prevent
deadlock.
Bug reported by Todd Fleisher <todd@fleish.org>.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Introduce per-net_device inlines: dev_net(), dev_net_set().
Without CONFIG_NET_NS, no namespace other than &init_net exists.
Let's explicitly define them to help compiler optimizations.
Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
There are some place, that calculate the ARP header length. These
calculations are correct, but
a) some operate with "magic" constants,
b) enlarge the code length (sometimes at the cost of coding style),
c) are not informative from the first glance.
The proposal is to introduce a helper, that includes all the good
sides of these calculations.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
ip_fib_init is kept enabled. It is already namespace-aware.
Signed-off-by: Denis V. Lunev <den@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Update bonding to version 3.2.4.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The ARP monitor functions currently acquire RTNL when performing
failover operations, but do so incorrectly (out of order). This causes
various warnings from might_sleep.
The ARP monitor isn't supported for any of the bonding modes
that actually require RTNL, so it is safe to not hold RTNL when
failing over in the ARP monitor.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
I've seen reports of invalid stats in /proc/net/dev for bonding
interfaces, and found it's a pretty easy problem to reproduce. Since
the current code zeros the bonding stats when a read is requested and a
pointer to that data is returned to the caller we cannot guarantee that
the caller has completely accessed the data before a successive call to
request the stats zeroes the stats again.
This patch creates a new stack variable to keep track of the updated
stats and copies the data from that variable into the bonding stats
structure. This ensures that the value for any of the bonding stats
should not incorrectly return zero for any of the bonding statistics.
This does use more stack space and require an extra memcpy, but it seems
like a fair trade-off for consistently correct bonding statistics.
Signed-off-by: Andy Gospodarek <andy@greyhouse.net>
Signed-off-by: Chris Snook <csnook@redhat.com>
Acked-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix the "are we creating a duplicate" check to not compare
the name if the name is NULL (meaning that the system should select
a name). Bug reported by Benny Amorsen <benny+usenet@amorsen.dk>.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch eliminates a problem (reported by lockdep) in the
bond_set_multicast_list function. It first reduces the locking on
bond->lock to a simple read_lock, and second, adds netif_tx locking
around the bonding mc_list manipulations that occur outside of the
set_multicast_list function.
The original problem was related to IPv6 addrconf activity.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
My last fix (commit ece95f7fef)
didn't handle one case correctly. This resolves that, and it will now
correctly parse parameters with arbitrary white space, and either text
names or mode values.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Needed to propagate it down to the ip_route_output_flow.
Signed-off-by: Denis V. Lunev <den@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Change bond_mii_monitor to not hold any locks when calling rtnl_unlock,
as rtnl_unlock can sleep (when acquring another mutex in netdev_run_todo).
Bug reported by Makito SHIOKAWA <mshiokawa@miraclelinux.com>, who
included a different patch.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Fix the handling of rtnl and the bonding_rwsem to always be acquired
in a consistent order (rtnl, then bonding_rwsem).
The existing code sometimes acquired them in this order, and sometimes
in the opposite order, which opens a window for deadlock between ifenslave
and sysfs.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
A recent change to add an additional hash policy modified
bond_parse_parm, but it now does not correctly match parameters passed in
via sysfs.
Rewrote bond_parse_parm to handle (a) parameter matches that
are substrings of one another and (b) user input with whitespace (e.g.,
sysfs input often has a trailing newline).
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Add a call to bond_release_all in the bonding netdev event
handler for the master. This releases the slaves for the case of, e.g.,
"echo -bond0 > /sys/class/net/bonding_masters", which otherwise will spin
forever waiting for references to be released.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
alb_fasten_mac_swap (actually rlb_teach_disabled_mac_on_primary)
requries RTNL and no other locks. This could cause dev_set_promiscuity
and/or dev_set_mac_address to be called with improper locking.
Changed callers to hold only RTNL during calls to alb_fasten_mac_swap
or functions calling it. Updated header comments in affected functions to
reflect proper reality of locking requirements.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Move an ASSERT_RTNL down to where we should hold only RTNL;
the existing check produces spurious warnings because we hold additional
locks at _bh, tripping a debug warning in spin_lock_mutex().
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Fix the functions that store the primary and active slave
options via sysfs to hold the correct locks in the correct order.
The bond_change_active_slave and bond_select_active_slave
functions both require rtnl, bond->lock for read and curr_slave_lock for
write_bh, and no other locks. This is so that the lower level
mode-specific functions (notably for balance-alb mode) can release locks
down to just rtnl in order to call, e.g., dev_set_mac_address with the
locks it expects (rtnl only).
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Andy Gospodarek <andy@greyhouse.net>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
Fixes a race condition in module unload. Without this change,
workqueue events may fire while bonding data structures are partially
freed but before bond_close() is invoked by unregister_netdevice().
Update version to 3.2.3.
Signed-off-by: Jay Vosburgh <fubar@us.ibm.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
David S. Miller