The current code in the RX flow uses two sg entries for each incoming
packet, the first one was for the IB headers and the second for the rest
of the data, that causes two dma map/unmap and two allocations, and few
more actions that were done at the data path.
Use only one linear skb on each incoming packet, for the data (IB
headers and payload), that reduces the packet processing in the
data-path (only one skb, no frags, the first frag was not used anyway,
less memory allocations) and the dma handling (only one dma map/unmap
over each incoming packet instead of two map/unmap per each incoming packet).
After commit 73d3fe6d1c ("gro: fix aggregation for skb using frag_list") from
Eric Dumazet, we will get full aggregation for large packets.
When running bandwidth tests before and after the (over the card's numa node),
using "netperf -H 1.1.1.3 -T -t TCP_STREAM", the results before are ~12Gbs before
and after ~16Gbs on my setup (Mellanox's ConnectX3).
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Commit a9c8ba5884 ("IPoIB: Fix usage of uninitialized multicast
objects") added a new flag MCAST_JOIN_STARTED, but was not very strict
in how it was used. We didn't always initialize the completion struct
before we set the flag, and we didn't always call complete on the
completion struct from all paths that complete it. And when we did
complete it, sometimes we continued to touch the mcast entry after
the completion, opening us up to possible use after free issues.
This made it less than totally effective, and certainly made its use
confusing. And in the flush function we would use the presence of this
flag to signal that we should wait on the completion struct, but we never
cleared this flag, ever.
In order to make things clearer and aid in resolving the rtnl deadlock
bug I've been chasing, I cleaned this up a bit.
1) Remove the MCAST_JOIN_STARTED flag entirely
2) Change MCAST_FLAG_BUSY so it now only means a join is in-flight
3) Test mcast->mc directly to see if we have completed
ib_sa_join_multicast (using IS_ERR_OR_NULL)
4) Make sure that before setting MCAST_FLAG_BUSY we always initialize
the mcast->done completion struct
5) Make sure that before calling complete(&mcast->done), we always clear
the MCAST_FLAG_BUSY bit
6) Take the mcast_mutex before we call ib_sa_multicast_join and also
take the mutex in our join callback. This forces
ib_sa_multicast_join to return and set mcast->mc before we process
the callback. This way, our callback can safely clear mcast->mc
if there is an error on the join and we will do the right thing as
a result in mcast_dev_flush.
7) Because we need the mutex to synchronize mcast->mc, we can no
longer call mcast_sendonly_join directly from mcast_send and
instead must add sendonly join processing to the mcast_join_task
8) Make MCAST_RUN mean that we have a working mcast subsystem, not that
we have a running task. We know when we need to reschedule our
join task thread and don't need a flag to tell us.
9) Add a helper for rescheduling the join task thread
A number of different races are resolved with these changes. These
races existed with the old MCAST_FLAG_BUSY usage, the
MCAST_JOIN_STARTED flag was an attempt to address them, and while it
helped, a determined effort could still trip things up.
One race looks something like this:
Thread 1 Thread 2
ib_sa_join_multicast (as part of running restart mcast task)
alloc member
call callback
ifconfig ib0 down
wait_for_completion
callback call completes
wait_for_completion in
mcast_dev_flush completes
mcast->mc is PTR_ERR_OR_NULL
so we skip ib_sa_leave_multicast
return from callback
return from ib_sa_join_multicast
set mcast->mc = return from ib_sa_multicast
We now have a permanently unbalanced join/leave issue that trips up the
refcounting in core/multicast.c
Another like this:
Thread 1 Thread 2 Thread 3
ib_sa_multicast_join
ifconfig ib0 down
priv->broadcast = NULL
join_complete
wait_for_completion
mcast->mc is not yet set, so don't clear
return from ib_sa_join_multicast and set mcast->mc
complete
return -EAGAIN (making mcast->mc invalid)
call ib_sa_multicast_leave
on invalid mcast->mc, hang
forever
By holding the mutex around ib_sa_multicast_join and taking the mutex
early in the callback, we force mcast->mc to be valid at the time we
run the callback. This allows us to clear mcast->mc if there is an
error and the join is going to fail. We do this before we complete
the mcast. In this way, mcast_dev_flush always sees consistent state
in regards to mcast->mc membership at the time that the
wait_for_completion() returns.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Various places in the IPoIB code had a deadlock related to flushing
the ipoib workqueue. Now that we have per device workqueues and a
specific flush workqueue, there is no longer a deadlock issue with
flushing the device specific workqueues and we can do so unilaterally.
Signed-off-by: Doug Ledford <dledford@redhat.com>
During my recent work on the rtnl lock deadlock in the IPoIB driver, I
saw that even once I fixed the apparent races for a single device, as
soon as that device had any children, new races popped up. It turns
out that this is because no matter how well we protect against races
on a single device, the fact that all devices use the same workqueue,
and flush_workqueue() flushes *everything* from that workqueue means
that we would also have to prevent all races between different devices
(for instance, ipoib_mcast_restart_task on interface ib0 can race with
ipoib_mcast_flush_dev on interface ib0.8002, resulting in a deadlock on
the rtnl_lock).
There are several possible solutions to this problem:
Make carrier_on_task and mcast_restart_task try to take the rtnl for
some set period of time and if they fail, then bail. This runs the
real risk of dropping work on the floor, which can end up being its
own separate kind of deadlock.
Set some global flag in the driver that says some device is in the
middle of going down, letting all tasks know to bail. Again, this can
drop work on the floor.
Or the method this patch attempts to use, which is when we bring an
interface up, create a workqueue specifically for that interface, so
that when we take it back down, we are flushing only those tasks
associated with our interface. In addition, keep the global
workqueue, but now limit it to only flush tasks. In this way, the
flush tasks can always flush the device specific work queues without
having deadlock issues.
Signed-off-by: Doug Ledford <dledford@redhat.com>
This reverts commit 016d9fb25c.
The series of IPoIB bug fixes that went into 3.19-rc1 introduce
regressions, and after trying to sort things out, we decided to revert
to 3.18's IPoIB driver and get things right for 3.20.
Signed-off-by: Roland Dreier <roland@purestorage.com>
This reverts commit 5141861cd5.
The series of IPoIB bug fixes that went into 3.19-rc1 introduce
regressions, and after trying to sort things out, we decided to revert
to 3.18's IPoIB driver and get things right for 3.20.
Signed-off-by: Roland Dreier <roland@purestorage.com>
This reverts commit bb42a6dd02.
The series of IPoIB bug fixes that went into 3.19-rc1 introduce
regressions, and after trying to sort things out, we decided to revert
to 3.18's IPoIB driver and get things right for 3.20.
Signed-off-by: Roland Dreier <roland@purestorage.com>
This reverts commit ce347ab90e.
The series of IPoIB bug fixes that went into 3.19-rc1 introduce
regressions, and after trying to sort things out, we decided to revert
to 3.18's IPoIB driver and get things right for 3.20.
Signed-off-by: Roland Dreier <roland@purestorage.com>
Various places in the IPoIB code had a deadlock related to flushing
the ipoib workqueue. Now that we have per device workqueues and a
specific flush workqueue, there is no longer a deadlock issue with
flushing the device specific workqueues and we can do so unilaterally.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
We used to pass a flush variable to mcast_stop_thread to indicate if
we should flush the workqueue or not. This was due to some code
trying to flush a workqueue that it was currently running on which is
a no-no. Now that we have per-device work queues, and now that
ipoib_mcast_restart_task has taken the fact that it is queued on a
single thread workqueue with all of the ipoib_mcast_join_task's and
therefore has no need to stop the join task while it runs, we can do
away with the flush parameter and unilaterally flush always.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
During my recent work on the rtnl lock deadlock in the IPoIB driver, I
saw that even once I fixed the apparent races for a single device, as
soon as that device had any children, new races popped up. It turns
out that this is because no matter how well we protect against races
on a single device, the fact that all devices use the same workqueue,
and flush_workqueue() flushes *everything* from that workqueue, we can
have one device in the middle of a down and holding the rtnl lock and
another totally unrelated device needing to run mcast_restart_task,
which wants the rtnl lock and will loop trying to take it unless is
sees its own FLAG_ADMIN_UP flag go away. Because the unrelated
interface will never see its own ADMIN_UP flag drop, the interface
going down will deadlock trying to flush the queue. There are several
possible solutions to this problem:
Make carrier_on_task and mcast_restart_task try to take the rtnl for
some set period of time and if they fail, then bail. This runs the
real risk of dropping work on the floor, which can end up being its
own separate kind of deadlock.
Set some global flag in the driver that says some device is in the
middle of going down, letting all tasks know to bail. Again, this can
drop work on the floor. I suppose if our own ADMIN_UP flag doesn't go
away, then maybe after a few tries on the rtnl lock we can queue our
own task back up as a delayed work and return and avoid dropping work
on the floor that way. But I'm not 100% convinced that we won't cause
other problems.
Or the method this patch attempts to use, which is when we bring an
interface up, create a workqueue specifically for that interface, so
that when we take it back down, we are flushing only those tasks
associated with our interface. In addition, keep the global
workqueue, but now limit it to only flush tasks. In this way, the
flush tasks can always flush the device specific work queues without
having deadlock issues.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Commit a9c8ba5884 ("IPoIB: Fix usage of uninitialized multicast
objects") added a new flag MCAST_JOIN_STARTED, but was not very strict
in how it was used. We didn't always initialize the completion struct
before we set the flag, and we didn't always call complete on the
completion struct from all paths that complete it. This made it less
than totally effective, and certainly made its use confusing. And in
the flush function we would use the presence of this flag to signal
that we should wait on the completion struct, but we never cleared
this flag, ever. This is further muddied by the fact that we overload
the MCAST_FLAG_BUSY flag to mean two different things: we have a join
in flight, and we have succeeded in getting an ib_sa_join_multicast.
In order to make things clearer and aid in resolving the rtnl deadlock
bug I've been chasing, I cleaned this up a bit.
1) Remove the MCAST_JOIN_STARTED flag entirely
2) Un-overload MCAST_FLAG_BUSY so it now only means a join is in-flight
3) Test on mcast->mc directly to see if we have completed
ib_sa_join_multicast (using IS_ERR_OR_NULL)
4) Make sure that before setting MCAST_FLAG_BUSY we always initialize
the mcast->done completion struct
5) Make sure that before calling complete(&mcast->done), we always clear
the MCAST_FLAG_BUSY bit
6) Take the mcast_mutex before we call ib_sa_multicast_join and also
take the mutex in our join callback. This forces
ib_sa_multicast_join to return and set mcast->mc before we process
the callback. This way, our callback can safely clear mcast->mc
if there is an error on the join and we will do the right thing as
a result in mcast_dev_flush.
7) Because we need the mutex to synchronize mcast->mc, we can no
longer call mcast_sendonly_join directly from mcast_send and
instead must add sendonly join processing to the mcast_join_task
A number of different races are resolved with these changes. These
races existed with the old MCAST_FLAG_BUSY usage, the
MCAST_JOIN_STARTED flag was an attempt to address them, and while it
helped, a determined effort could still trip things up.
One race looks something like this:
Thread 1 Thread 2
ib_sa_join_multicast (as part of running restart mcast task)
alloc member
call callback
ifconfig ib0 down
wait_for_completion
callback call completes
wait_for_completion in
mcast_dev_flush completes
mcast->mc is PTR_ERR_OR_NULL
so we skip ib_sa_leave_multicast
return from callback
return from ib_sa_join_multicast
set mcast->mc = return from ib_sa_multicast
We now have a permanently unbalanced join/leave issue that trips up the
refcounting in core/multicast.c
Another like this:
Thread 1 Thread 2 Thread 3
ib_sa_multicast_join
ifconfig ib0 down
priv->broadcast = NULL
join_complete
wait_for_completion
mcast->mc is not yet set, so don't clear
return from ib_sa_join_multicast and set mcast->mc
complete
return -EAGAIN (making mcast->mc invalid)
call ib_sa_multicast_leave
on invalid mcast->mc, hang
forever
By holding the mutex around ib_sa_multicast_join and taking the mutex
early in the callback, we force mcast->mc to be valid at the time we
run the callback. This allows us to clear mcast->mc if there is an
error and the join is going to fail. We do this before we complete
the mcast. In this way, mcast_dev_flush always sees consistent state
in regards to mcast->mc membership at the time that the
wait_for_completion() returns.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The error flow of ipoib_ib_dev_open() invokes ipoib_ib_dev_stop() with
workqueue flushing enabled, which deadlocks if the open procedure
itself was called by a worker thread.
Fix this by adding a flush enabled flag to ipoib_ib_dev_open() and set
it accordingly from the locations where such a call is made.
The call trace was the following:
[<ffffffff81095bc4>] ? flush_workqueue+0x54/0x80
[<ffffffffa056c657>] ? ipoib_ib_dev_stop+0x447/0x650 [ib_ipoib]
[<ffffffffa056cc34>] ? ipoib_ib_dev_open+0x284/0x430 [ib_ipoib]
[<ffffffffa05674a8>] ? ipoib_open+0x78/0x1d0 [ib_ipoib]
[<ffffffffa05697b8>] ? ipoib_pkey_open+0x38/0x40 [ib_ipoib]
[<ffffffffa056cf3c>] ? __ipoib_ib_dev_flush+0x15c/0x2c0 [ib_ipoib]
[<ffffffffa056ce56>] ? __ipoib_ib_dev_flush+0x76/0x2c0 [ib_ipoib]
[<ffffffffa056d0a0>] ? ipoib_ib_dev_flush_heavy+0x0/0x20 [ib_ipoib]
[<ffffffffa056d0ba>] ? ipoib_ib_dev_flush_heavy+0x1a/0x20 [ib_ipoib]
[<ffffffff81094d20>] ? worker_thread+0x170/0x2a0
[<ffffffff8109b2a0>] ? autoremove_wake_function+0x0/0x40
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Acked-by: Alex Estrin <alex.estrin@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The current code use a dedicated polling logic to determine when the P_Key
assigned to the ipoib device is present in HCA port table and act accordingly.
Move to use the code which acts upon getting PKEY_CHANGE event to handle this
task and remove the P_Key polling logic/thread as they add extra complexity
which isn't needed.
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Acked-by: Alex Estrin <alex.estrin@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The driver should avoid calling ib_sa_free_multicast on the mcast->mc
object until it finishes its initialization state. Otherwise we can
crash when ipoib_mcast_dev_flush() attempts to use the uninitialized
multicast object.
Instead, only call wait_for_completion() for multicast entries that
started the join process, meaning that ib_sa_join_multicast() finished.
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
When ipoib interface is going down it takes all of its children with
it, under mutex.
For each child, dev_change_flags() is called. That function calls
ipoib_stop() via the ndo, and causes flush of the workqueue.
Sometimes in the workqueue an __ipoib_dev_flush work() is waiting and
when invoked tries to get the same mutex, which leads to a deadlock,
as seen below.
The solution is to switch to rw-sem instead of mutex.
The deadlock:
[11028.165303] [<ffffffff812b0977>] ? vgacon_scroll+0x107/0x2e0
[11028.171844] [<ffffffff814eaac5>] schedule_timeout+0x215/0x2e0
[11028.178465] [<ffffffff8105a5c3>] ? perf_event_task_sched_out+0x33/0x80
[11028.185962] [<ffffffff814ea743>] wait_for_common+0x123/0x180
[11028.192491] [<ffffffff8105fa40>] ? default_wake_function+0x0/0x20
[11028.199504] [<ffffffff814ea85d>] wait_for_completion+0x1d/0x20
[11028.206224] [<ffffffff8108b4f1>] flush_cpu_workqueue+0x61/0x90
[11028.212948] [<ffffffff8108b5a0>] ? wq_barrier_func+0x0/0x20
[11028.219375] [<ffffffff8108bfc4>] flush_workqueue+0x54/0x80
[11028.225712] [<ffffffffa05a0576>] ipoib_mcast_stop_thread+0x66/0x90 [ib_ipoib]
[11028.233988] [<ffffffffa059ccea>] ipoib_ib_dev_down+0x6a/0x100 [ib_ipoib]
[11028.241678] [<ffffffffa059849a>] ipoib_stop+0x8a/0x140 [ib_ipoib]
[11028.248692] [<ffffffff8142adf1>] dev_close+0x71/0xc0
[11028.254447] [<ffffffff8142a631>] dev_change_flags+0xa1/0x1d0
[11028.261062] [<ffffffffa059851b>] ipoib_stop+0x10b/0x140 [ib_ipoib]
[11028.268172] [<ffffffff8142adf1>] dev_close+0x71/0xc0
[11028.273922] [<ffffffff8142a631>] dev_change_flags+0xa1/0x1d0
[11028.280452] [<ffffffff8148f20b>] devinet_ioctl+0x5eb/0x6a0
[11028.286786] [<ffffffff814903b8>] inet_ioctl+0x88/0xa0
[11028.292633] [<ffffffff8141591a>] sock_ioctl+0x7a/0x280
[11028.298576] [<ffffffff81189012>] vfs_ioctl+0x22/0xa0
[11028.304326] [<ffffffff81140540>] ? unmap_region+0x110/0x130
[11028.310756] [<ffffffff811891b4>] do_vfs_ioctl+0x84/0x580
[11028.316897] [<ffffffff81189731>] sys_ioctl+0x81/0xa0
and
11028.017533] [<ffffffff8105a5c3>] ? perf_event_task_sched_out+0x33/0x80
[11028.025030] [<ffffffff8100bb8e>] ? apic_timer_interrupt+0xe/0x20
[11028.031945] [<ffffffff814eb2ae>] __mutex_lock_slowpath+0x13e/0x180
[11028.039053] [<ffffffff814eb14b>] mutex_lock+0x2b/0x50
[11028.044910] [<ffffffffa059f7e7>] __ipoib_ib_dev_flush+0x37/0x210 [ib_ipoib]
[11028.052894] [<ffffffffa059fa00>] ? ipoib_ib_dev_flush_light+0x0/0x20 [ib_ipoib]
[11028.061363] [<ffffffffa059fa17>] ipoib_ib_dev_flush_light+0x17/0x20 [ib_ipoib]
[11028.069738] [<ffffffff8108b120>] worker_thread+0x170/0x2a0
[11028.076068] [<ffffffff81090990>] ? autoremove_wake_function+0x0/0x40
[11028.083374] [<ffffffff8108afb0>] ? worker_thread+0x0/0x2a0
[11028.089709] [<ffffffff81090626>] kthread+0x96/0xa0
[11028.095266] [<ffffffff8100c0ca>] child_rip+0xa/0x20
[11028.100921] [<ffffffff81090590>] ? kthread+0x0/0xa0
[11028.106573] [<ffffffff8100c0c0>] ? child_rip+0x0/0x20
[11028.112423] INFO: task ifconfig:23640 blocked for more than 120 seconds.
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Implement version info as well as report firmware version and bus info
of the underlying IB HW device.
Signed-off-by: Yan Burman <yanb@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The hash function introduced in commit b63b70d877 ("IPoIB: Use a
private hash table for path lookup in xmit path") was designd to use
the 3 octets of the IPoIB HW address that holds the remote QPN.
However, this currently isn't the case on little-endian machines,
because the the code there uses the flags part (octet[0]) and not the
last octet of the QPN (octet[3]). Fix this.
The fix caused a checkpatch warning on line over 80 characters, to
solve that changed the name of the temp variable that holds the daddr.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
With the new netlink support in commit 862096a8bb ("IB/ipoib: Add more
rtnl_link_ops callbacks") we need ipoib_set_mode() to be available even
if connected mode isn't built. Move the function from ipoib_cm.c to
ipoib_main.c (and make a few CM-related macros available unconditonally).
This fixes the build error
drivers/built-in.o: In function 'ipoib_changelink':
ipoib_netlink.c:(.text+0x6a5fc9): undefined reference to 'ipoib_set_mode'
ipoib_netlink.c:(.text+0x6a5fe3): undefined reference to 'ipoib_set_mode'
when CONFIG_INFINIBAND_IPOIB_CM isn't set.
Reported-by: Randy Dunlap <rdunlap@xenotime.net>
Reported-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Add the rtnl_link_ops changelink and fill_info callbacks, through
which the admin can now set/get the driver mode, etc policies.
Maintain the proprietary sysfs entries only for legacy childs.
For child devices, set dev->iflink to point to the parent
device ifindex, such that user space tools can now correctly
show the uplink relation as done for vlan, macvlan, etc
devices. Pointed out by Patrick McHardy <kaber@trash.net>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
drivers/net/team/team.c
drivers/net/usb/qmi_wwan.c
net/batman-adv/bat_iv_ogm.c
net/ipv4/fib_frontend.c
net/ipv4/route.c
net/l2tp/l2tp_netlink.c
The team, fib_frontend, route, and l2tp_netlink conflicts were simply
overlapping changes.
qmi_wwan and bat_iv_ogm were of the "use HEAD" variety.
With help from Antonio Quartulli.
Signed-off-by: David S. Miller <davem@davemloft.net>
Add rtnl_link_ops to IPoIB, with the first usage being child device
create/delete through them. Childs devices are now either legacy ones,
created/deleted through the ipoib sysfs entries, or RTNL ones.
Adding support for RTNL childs involved refactoring of ipoib_vlan_add
which is now used by both the sysfs and the link_ops code.
Also, added ndo_uninit entry to support calling unregister_netdevice_queue
from the rtnl dellink entry. This required removal of calls to
ipoib_dev_cleanup from the driver in flows which use unregister_netdevice,
since the networking core will invoke ipoib_uninit which does exactly that.
Signed-off-by: Erez Shitrit <erezsh@mellanox.co.il>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Lockdep points out a circular locking dependency betwwen the ipoib
device priv spinlock (priv->lock) and the neighbour table rwlock
(ntbl->rwlock).
In the normal path, ie neigbour garbage collection task, the neigh
table rwlock is taken first and then if the neighbour needs to be
deleted, priv->lock is taken.
However in some error paths, such as in ipoib_cm_handle_tx_wc(),
priv->lock is taken first and then ipoib_neigh_free routine is called
which in turn takes the neighbour table ntbl->rwlock.
The solution is to get rid the neigh table rwlock completely and use
only priv->lock.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
If the neighbours hash table is empty when unloading the module, then
ipoib_flush_neighs(), the cleanup routine, isn't called and the
memory used for the hash table itself leaked.
To fix this, ipoib_flush_neighs() is allways called, and another
completion object is added to signal when the table is freed.
Once invoked, ipoib_flush_neighs() flushes all the neighbours (if
there are any), calls the the hash table RCU free routine, which now
signals completion of the deletion process, and waits for the last
neighbour to be freed.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Dave Miller <davem@davemloft.net> provided a detailed description of
why the way IPoIB is using neighbours for its own ipoib_neigh struct
is buggy:
Any time an ipoib_neigh is changed, a sequence like the following is made:
spin_lock_irqsave(&priv->lock, flags);
/*
* It's safe to call ipoib_put_ah() inside
* priv->lock here, because we know that
* path->ah will always hold one more reference,
* so ipoib_put_ah() will never do more than
* decrement the ref count.
*/
if (neigh->ah)
ipoib_put_ah(neigh->ah);
list_del(&neigh->list);
ipoib_neigh_free(dev, neigh);
spin_unlock_irqrestore(&priv->lock, flags);
ipoib_path_lookup(skb, n, dev);
This doesn't work, because you're leaving a stale pointer to the freed up
ipoib_neigh in the special neigh->ha pointer cookie. Yes, it even fails
with all the locking done to protect _changes_ to *ipoib_neigh(n), and
with the code in ipoib_neigh_free() that NULLs out the pointer.
The core issue is that read side calls to *to_ipoib_neigh(n) are not
being synchronized at all, they are performed without any locking. So
whether we hold the lock or not when making changes to *ipoib_neigh(n)
you still can have threads see references to freed up ipoib_neigh
objects.
cpu 1 cpu 2
n = *ipoib_neigh()
*ipoib_neigh() = NULL
kfree(n)
n->foo == OOPS
[..]
Perhaps the ipoib code can have a private path database it manages
entirely itself, which holds all the necessary information and is
looked up by some generic key which is available easily at transmit
time and does not involve generic neighbour entries.
See <http://marc.info/?l=linux-rdma&m=132812793105624&w=2> and
<http://marc.info/?l=linux-rdma&w=2&r=1&s=allows+references+to+freed+memory&q=b>
for the full discussion.
This patch aims to solve the race conditions found in the IPoIB driver.
The patch removes the connection between the core networking neighbour
structure and the ipoib_neigh structure. In addition to avoiding the
race described above, it allows us to handle SKBs carrying IP packets
that don't have any associated neighbour.
We add an ipoib_neigh hash table with N buckets where the key is the
destination hardware address. The ipoib_neigh is fetched from the
hash table and instead of the stashed location in the neighbour
structure. The hash table uses both RCU and reference counting to
guarantee that no ipoib_neigh instance is ever deleted while in use.
Fetching the ipoib_neigh structure instance from the hash also makes
the special code in ipoib_start_xmit that handles remote and local
bonding failover redundant.
Aged ipoib_neigh instances are deleted by a garbage collection task
that runs every M seconds and deletes every ipoib_neigh instance that
was idle for at least 2*M seconds. The deletion is safe since the
ipoib_neigh instances are protected using RCU and reference count
mechanisms.
The number of buckets (N) and frequency of running the GC thread (M),
are taken from the exported arb_tbl.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Commit a0417fa3a1 ("net: Make qdisc_skb_cb upper size bound
explicit.") made it possible for a netdev driver to use skb->cb
between its header_ops.create method and its .ndo_start_xmit
method. Use this in ipoib_hard_header() to stash away the LL address
(GID + QPN), instead of the "ipoib_pseudoheader" hack. This allows
IPoIB to stop lying about its hard_header_len, which will let us fix
the L2 check for GRO.
Signed-off-by: Roland Dreier <roland@purestorage.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This allows us to move duplicated code in <asm/atomic.h>
(atomic_inc_not_zero() for now) to <linux/atomic.h>
Signed-off-by: Arun Sharma <asharma@fb.com>
Reviewed-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: Mike Frysinger <vapier@gentoo.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
As a first step in moving from LRO to GRO, revert commit af40da894e
("IPoIB: add LRO support"). Also eliminate the ethtool set_flags
callback which isn't needed anymore. Finally, we need to include
<linux/sched.h> directly to get the declaration of restart_syscall()
(which used to be included implicitly through <linux/inet_lro.h>).
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: Vladimir Sokolovsky <vlad@mellanox.co.il>
Signed-off-by: Or Gerlitz <ogerlitz@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Child devices were created without any offload features set, fix this by
moving the code that computes the features into generic function which is
now called through non-child and child device creation.
Signed-off-by: Or Gerlitz <ogerlitz@voltaire.com>
-- v1 has a bug where the 'result' flag in ipoib_vlan_add may be used uninitialized
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Currently, IPoIB is an LLTX driver that uses its own IRQ-disabling
tx_lock. Not only do we want to get rid of LLTX, this actually causes
problems because of the skb_orphan() done with this tx_lock held: some
skb destructors expect to be run with interrupts enabled.
The simplest fix for this is to get rid of the driver-private tx_lock
and stop using LLTX. We kill off priv->tx_lock and use
netif_tx_lock[_bh]() instead; the patch to do this is a tiny bit
tricky because we need to update places that take priv->lock inside
the tx_lock to disable IRQs, rather than relying on tx_lock having
already disabled IRQs.
Also, there are a couple of places where we need to disable BHs to
make sure we have a consistent context to call netif_tx_lock() (since
we no longer can use _irqsave() variants), and we also have to change
ipoib_send_comp_handler() to call drain_tx_cq() through a timer rather
than directly, because ipoib_send_comp_handler() runs in interrupt
context and drain_tx_cq() must run in BH context so it can call
netif_tx_lock().
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Taking rtnl_lock in ipoib_mcast_join_complete() causes a deadlock with
ipoib_stop(). We avoid it by scheduling the piece of code that takes
the lock on ipoib_workqueue instead of executing it directly. This
works because we only flush the ipoib_workqueue with the RTNL not held.
The deadlock happens because ipoib_stop() calls ipoib_ib_dev_down()
which calls ipoib_mcast_dev_flush(), which calls ipoib_mcast_free(),
which calls ipoib_mcast_leave(). The latter calls
ib_sa_free_multicast(), and this waits until the multicast completion
handler finishes. This handler is ipoib_mcast_join_complete(), which
waits for the rtnl_lock(), which was already taken by ipoib_stop().
This bug was introduced in commit a77a57a1 ("IPoIB: Fix deadlock on
RTNL in ipoib_stop()").
Signed-off-by: Yossi Etigin <yosefe@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Increase IPoIB ring sizes to twice their original sizes (RX: 128->256,
TX: 64->128) to act as a shock absorber for high traffic peaks. With
the current settings, we have seen cases that there are many calls to
netif_stop_queue(), which causes degradation in throughput. Also,
larger receive buffer sizes help IPoIB in CM mode to avoid experiencing
RNR NAK conditions due to insufficient receive buffers at the SRQ.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Since IPoIB connected mode does not NETIF_F_SG, we only have one DMA
mapping per send, so we don't need a mapping[] array. Define a new
struct with a single u64 mapping member and use it for the CM tx_ring.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
ipoib_mcast_detach() does nothing except call ib_detach_mcast(), so just
use the core API in the one place that does a multicast group detach.
add/remove: 0/1 grow/shrink: 0/1 up/down: 0/-105 (-105)
function old new delta
ipoib_mcast_leave 357 319 -38
ipoib_mcast_detach 67 - -67
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The current code will set the Q_Key for any join of a non-sendonly
multicast group. The operation involves a modify QP operation, which
is fairly heavyweight, and is only really required after the join of
the broadcast group. Fix this by adding a parameter to ipoib_mcast_attach()
to control when the Q_Key is set.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
No need for a mutex around calls to ib_attach_mcast/ib_detach_mcast
since these operations are synchronized at the HW driver layer.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The IPOIB_MCAST_STARTED flag is not used at all since commit b3e2749b
("IPoIB: Don't drop multicast sends when they can be queued"), so
remove it.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The patch tries to solve the problem of device going down and paths being
flushed on an SM change event. The method is to mark the paths as candidates for
refresh (by setting the new valid flag to 0), and wait for an ARP
probe a new path record query.
The solution requires a different and less intrusive handling of SM
change event. For that, the second argument of the flush function
changes its meaning from a boolean flag to a level. In most cases, SM
failover doesn't cause LID change so traffic won't stop. In the rare
cases of LID change, the remote host (the one that hadn't changed its
LID) will lose connectivity until paths are refreshed. This is no
worse than the current state. In fact, preventing the device from
going down saves packets that otherwise would be lost.
Signed-off-by: Moni Levy <monil@voltaire.com>
Signed-off-by: Moni Shoua <monis@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add "ipoib_use_lro" module parameter to enable LRO and an
"ipoib_lro_max_aggr" module parameter to set the max number of packets
to be aggregated. Make LRO controllable and LRO statistics accessible
through ethtool.
Signed-off-by: Vladimir Sokolovsky <vlad@mellanox.co.il>
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The connected mode implementation in the IPoIB driver has a large
overhead in the way SKBs are handled in the receive flow. It usually
allocates an SKB with as big as was used in the currently received SKB
and moves unused fragments from the old SKB to the new one. This
involves a loop on all the remaining fragments and incurs overhead on
the CPU. This patch, for small SKBs, allocates an SKB just large
enough to contain the received data and copies to it the data from the
received SKB. The newly allocated SKB is passed to the stack and the
old SKB is reposted.
When running netperf, UDP small messages, without this pach I get:
UDP UNIDIRECTIONAL SEND TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
14.4.3.178 (14.4.3.178) port 0 AF_INET
Socket Message Elapsed Messages
Size Size Time Okay Errors Throughput
bytes bytes secs # # 10^6bits/sec
114688 128 10.00 5142034 0 526.31
114688 10.00 1130489 115.71
With this patch I get both send and receive at ~315 mbps.
The reason that send performance actually slows down is as follows:
When using this patch, the overhead of the CPU for handling RX packets
is dramatically reduced. As a result, we do not experience RNR NAK
messages from the receiver which cause the connection to be closed and
reopened again; when the patch is not used, the receiver cannot handle
the packets fast enough so there is less time to post new buffers and
hence the mentioned RNR NACKs. So what happens is that the
application *thinks* it posted a certain number of packets for
transmission but these packets are flushed and do not really get
transmitted. Since the connection gets opened and closed many times,
each time netperf gets the CPU time that otherwise would have been
given to IPoIB to actually transmit the packets. This can be verified
when looking at the port counters -- the output of ifconfig and the
oputput of netperf (this is for the case without the patch):
tx packets
==========
port counter: 1,543,996
ifconfig: 1,581,426
netperf: 5,142,034
rx packets
==========
netperf 1,1304,089
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Commit f56bcd80 ("IPoIB: Use separate CQ for UD send completions")
introduced a bug where the transmit queue could get stopped and never
woken up. The problem is that send completions are only polled at the
end of the xmit function, so if the send queue fills up and the xmit
path stops the queue, then there is no way for send completions to
ever get polled, and so the transmit queue stays stopped forever.
Fix this by arming the send CQ just before posting the last send
request that fills the send queue. Then, when the completion event
handler is called, drain the send CQ. Since it is possible that not
enough send completions are in the CQ, verify that the the net queue
has been woken up after draining the send CQ, and if not arm a timer
and drain again at the timer function.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Use a dedicated CQ for UD send completions. Also, do not arm the UD
send CQ, which reduces the number of interrupts generated. This patch
farther reduces overhead by not calling poll CQ for every posted send
WR -- it does polls only when there 16 or more outstanding work requests.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
This patch enables IPoIB to use 4K UD messages (when the underlying
device and fabrics support a 4K MTU) by using two scatter buffers when
PAGE_SIZE is less than or equal to thhe HCA IB MTU size. The first
buffer is for IPoIB header + GRH header, and the second buffer is the
IPoIB payload, which is 4K-4.
Signed-off-by: Shirley Ma <xma@us.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
This can be used to tune at run time the parameters controlling the
event (interrupt) generation rate and thus reduce the overhead
incurred by handling interrupts resulting in better throughput. Since
IPoIB uses a single CQ for both RX and TX, RX is chosen to dictate
configuration for both RX and TX.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Just add the infrastructure so we can add functionality later.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
For HCAs that support TCP segmentation offload (IB_DEVICE_UD_TSO), set
NETIF_F_TSO and use HW LSO to offload TCP segmentation.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
