This is necessary in order to support > 1 VF/PF in a VM for software
that uses the node guid as a discriminator, such as librdmacm.
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Remove the error returns for IB ports from mlx4_ib_add,
mlx4_INIT_PORT_wrapper, and mlx4_CLOSE_PORT_wrapper.

Currently, SRIOV is supported only for devices for which the
link layer is IB on all ports; RoCE support will be added later.
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

1. Allow only master to change node description.
2. Prevent AH leakage in send mads.
3. Take device part number from PCI structure, so that guests see the
   VF part number (and not the PF part number).
4. Place the device revision ID into caps structure at startup.
5. SET_PORT in update_gids_task needs to go through wrapper on master.
6. In mlx4_ib_event(), PORT_MGMT_EVENT needs be handled in a work
   queue on the master, since it propagates events to slaves using
   GEN_EQE.
7. Do not support FMR on slaves.
8. Add spinlock to slave_event(), since it is called both in interrupt
   context and in process context (due to 6 above, and also if
   smp_snoop is used).  This fix was found and implemented by Saeed
   Mahameed <saeedm@mellanox.com>
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

This directory is added only for the master -- slaves do not have it.

The sysfs iov directory is used to manage and examine the port P_Key
and guid paravirtualization.

Under iov/ports, the administrator may examine the gid and P_Key tables
as they are present in the device (and as are seen in the "network
view" presented to the SM).

Under the iov/<pci slot number> directories, the admin may map the
index numbers in the physical tables (as under iov/ports) to the
paravirtualized index numbers that guests see.

For example, if the administrator, for port 1 on guest 2 maps physical
pkey index 10 to virtual index 1, then that guest, whenever it uses
its pkey index 1, will actually be using the real pkey index 10.

Based on patch from Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

P_Key change and guid change events are not of interest to all slaves,
but only to those slaves which "see" the table slots whose contents
have change.

For example, if the guid at port 1, index 5 has changed in the PPF, we
wish to propagate the gid-change event only to the function which has
that guid index mapped to its port/guid table (in this case it is
slave #5). Other functions should not get the event, since the event
does not affect them.

Similarly with P_Keys -- P_Key change events are forwarded only to
slaves which have that P_Key index mapped to their virtual P_Key table.
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

For IB ports, we paravirtualize the GUID at index 0 on slaves.  The
GUID at index 0 seen by a slave is the actual GUID occupying the GUID
table at the slave-id index.

The driver, by default, requests at startup time that subnet manager
populate its entire guid table with GUIDs. These guids are then mapped
(paravirtualized) to the slaves, and appear for each slave as its GUID
at index 0.

Until each slave has such a guid, its port status is DOWN.

The guid table is cached to support special QP paravirtualization, and
event propagation to slaves on guid change (we test to see if the guid
really changed before propagating an event to the slave).

To support this caching, add capability to __mlx4_ib_query_gid() to
obtain the network view (i.e., physical view) gid at index X, not just
the host (paravirtualized) view.

Based on a patch from Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

In CM para-virtualization:

1. Incoming requests are steered to the correct vHCA according to the
   embedded GID.
2. Communication IDs on outgoing requests are replaced by a globally
   unique ID, generated by the PPF, since there is no synchronization
   of ID generation between guests (and so these IDs are not
   guaranteed to be globally unique).  The guest's comm ID is stored,
   and is returned to the response MAD when it arrives.
Signed-off-by: NAmir Vadai <amirv@mellanox.co.il>
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

MCG paravirtualization support includes:
- Creating multicast groups by VFs, and keeping accounting of them
- Leaving multicast groups by VFs
- Updating SM only with real changes in the overall picture of MCGs status
- Creation of MGID=0 groups (let SM choose MGID)

Note that the MCG module maintains its own internal MCG object
reference counts.  The reason for this is that the IB core is used to
track only the multicast groups joins generated by the PF it runs
over.  The PF IB core layer is unaware of slaves, so it cannot be used
to keep track of MCG joins they generate.
Signed-off-by: NOren Duer <oren@mellanox.co.il>
Signed-off-by: NEli Cohen <eli@mellanox.com>
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

The MAD_IFC firmware command fulfills two functions.

First, it is used in the QP0/QP1 MAD-handling flow to obtain
information from the FW (for answering queries), and for setting
variables in the HCA (MAD SET packets).

For this, MAD_IFC should provide the FW (physical) view of the data.
This is the view that OpenSM needs.  We call this the "network view".

In the second case, MAD_IFC is used by various verbs to obtain data
regarding the local HCA (e.g., ib_query_device()).  We call this the
"host view".

This data needs to be paravirtualized.

MAD_IFC therefore needs a wrapper function, and also needs another
flag indicating whether it should provide the network view (when it is
called by ib_process_mad in special-qp packet handling), or the host
view (when it is called while implementing a verb).

There are currently 2 flag parameters in mlx4_MAD_IFC already:
ignore_bkey and ignore_mkey.  These two parameters are replaced by a
single "mad_ifc_flags" parameter, with different bits set for each
flag.  A third flag is added: "network-view/host-view".
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Special QPs are paravirtualized.

vHCAs are not given direct access to QP0/1. Rather, these QPs are
operated by a special context hosted by the PF, which mediates access
to/from vHCAs.  This is done by opening a "tunnel" per vHCA port per
QP0/1. A tunnel comprises a pair of UD QPs: a "Tunnel QP" in the
PF-context and a "Proxy QP" in the vHCA.  All vHCA MAD traffic must
pass through the corresponding tunnel.  vHCA QPs cannot be assigned to
VL15 and are denied of the well-known QKey.

Outgoing messages are "de-multiplexed" (i.e., directed to the wire via
the real special QP).

Incoming messages are "multiplexed" (i.e. steered by the PPF to the
correct VF or to the PF)

QP0 access is restricted to the PF vHCA. VF vHCAs also have (virtual)
QP0s, but they never receive any SMPs and all SMPs sent are discarded.
QP1 traffic is allowed for all vHCAs, but special care is required to
bridge the gap between the host and network views.

Specifically:
- Transaction IDs are mapped to guarantee uniqueness among vHCAs
- CM para-virtualization
  o   Incoming requests are steered to the correct vHCA according to the embedded GID
  o   Local communication IDs are mapped to ensure uniqueness among vHCAs
  (see the patch that adds CM paravirtualization.)
- Multicast para-virtualization
  o   The PF context aggregates membership state from all vHCAs
  o   The SA is contacted only when the aggregate membership changes
  o   If the aggregate does not change, the PF context will provide the
      requesting vHCA with the proper response.
  (see the patch that adds multicast group paravirtualization)

Incoming MADs are steered according to:
- the DGID If a GRH is present
- the mapped transaction ID for response MADs
- the embedded GID in CM requests
- the remote communication ID in other CM messages
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

This requires:

1. Replacing the paravirtualized P_Key index (inserted by the guest)
   with the real P_Key index.

2. For UD QPs, placing the guest's true source GID index in the
   address path structure mgid field, and setting the ud_force_mgid
   bit so that the mgid is taken from the QP context and not from the
   WQE when posting sends.

3. For UC and RC QPs, placing the guest's true source GID index in the
   address path structure mgid field.

4. For tunnel and proxy QPs, setting the Q_Key value reserved for that
   proxy/tunnel pair.

Since not all the above adjustments occur in all the QP transitions,
the QP transitions require separate wrapper functions.

Secondly, initialize the P_Key virtualization table to its default
values: Master virtualized table is 1-1 with the real P_Key table,
guest virtualized table has P_Key index 0 mapped to the real P_Key
index 0, and all the other P_Key indices mapped to the reserved
(invalid) P_Key at index 127.

Finally, add logic in smp_snoop for maintaining the phys_P_Key_cache.
and generating events on the master only if a P_Key actually changed.
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Allocate SR-IOV paravirtualization resources and MAD demuxing contexts
on the master.

This has two parts.  The first part is to initialize the structures to
contain the contexts.  This is done at master startup time in
mlx4_ib_init_sriov().

The second part is to actually create the tunneling resources required
on the master to support a slave.  This is performed the master
detects that a slave has started up (MLX4_DEV_EVENT_SLAVE_INIT event
generated when a slave initializes its comm channel).

For the master, there is no such startup event, so it creates its own
tunneling resources when it starts up.  In addition, the master also
creates the real special QPs.  The ib_core layer on the master causes
creation of proxy special QPs, since the master is also
paravirtualized at the ib_core layer.
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

1. Introduce the basic SR-IOV parvirtualization context objects for
   multiplexing and demultiplexing MADs.
2. Introduce support for the new proxy and tunnel QP types.

This patch introduces the objects required by the master for managing
QP paravirtualization for guests.

struct mlx4_ib_sriov is created by the master only.
It is a container for the following:

1. All the info required by the PPF to multiplex and de-multiplex MADs
   (including those from the PF). (struct mlx4_ib_demux_ctx demux)
2. All the info required to manage alias GUIDs (i.e., the GUID at
   index 0 that each guest perceives.  In fact, this is not the GUID
   which is actually at index 0, but is, in fact, the GUID which is at
   index[<VF number>] in the physical table.
3. structures which are used to manage CM paravirtualization
4. structures for managing the real special QPs when running in SR-IOV
   mode.  The real SQPs are controlled by the PPF in this case.  All
   SQPs created and controlled by the ib core layer are proxy SQP.

struct mlx4_ib_demux_ctx contains the information per port needed
to manage paravirtualization:

1. All multicast paravirt info
2. All tunnel-qp paravirt info for the port.
3. GUID-table and GUID-prefix for the port
4. work queues.

struct mlx4_ib_demux_pv_ctx contains all the info for managing the
paravirtualized QPs for one slave/port.

struct mlx4_ib_demux_pv_qp contains the info need to run an individual
QP (either tunnel qp or real SQP).

Note:  We made use of the 2 most significant bits in enum
mlx4_ib_qp_flags (based on enum ib_qp_create_flags in ib_verbs.h).
We need these bits in the low-level driver for internal purposes.
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

When P_Key tables potentially contain both full and partial membership
copies for the same P_Key, we need a function to find the index for an
exact (16-bit) P_Key.

This is necessary when the master forwards QP1 MADs sent by guests.
If the guest has sent the MAD with a limited membership P_Key, we need
to to forward the MAD using the same limited membership P_Key.  Since
the master may have both the limited and the full member P_Keys in its
table, we must make sure to retrieve the limited membership P_Key in
this case.
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NOr Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Extend the cached and non-cached P_Key table lookups to handle limited
and full membership of the same P_Key to co-exist in the P_Key table.

This is necessary for SR-IOV, to allow for some guests would to have
the full membership P_Key in their virtual P_Key table, while other
guests on the same physical HCA would have the limited one.
To support this, we need both the limited and full membership P_Keys
to be present in the master's (hypervisor physical port) P_Key table.

The algorithm for handling P_Key tables which contain both the limited
and the full membership versions of the same P_Key works as follows:

When scanning the P_Key table for a 15-bit P_Key:

A. If there is a full member version of that P_Key anywhere in the
    table, return its index (even if a limited-member version of the
    P_Key exists earlier in the table).

B. If the full member version is not in the table, but the
   limited-member version is in the table, return the index of the
   limited P_Key.
Signed-off-by: NLiran Liss <liranl@mellanox.com>
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NOr Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Commit 3236b2d4 ("IB/qib: MADs with misset M_Keys should return
failure") introduced a return code assignment that unfortunately
introduced an unconditional exit for the routine due to missing braces.

This patch adds the braces to correct the original patch.
Reviewed-by: NDean Luick <dean.luick@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Fix CQE expansion of unsignaled WQE -- don't expand the CQE when the
WQE index of the completed CQE matches with last pending WQE (tail) in
the queue.
Signed-off-by: NParav Pandit <parav.pandit@emulex.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

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: NShlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: NOr Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: NRoland 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: NShlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: NOr Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

spatch with a semantic match is used to found this.
(http://coccinelle.lip6.fr/)
Signed-off-by: NWei Yongjun <yongjun_wei@trendmicro.com.cn>
Acked-by: NSteve Wise <swise@opengridcomputing.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Unlike other parts of the mlx4_ib code, the function build_mlx_header()
doesn't check if the iboe netdev of the given port is valid before
dereferencing it, which can cause a crash if the ethernet interface
has already been taken down.

Fix this by checking for a valid netdev pointer before using it to get
the port MAC address.
Signed-off-by: NKleber Sacilotto de Souza <klebers@linux.vnet.ibm.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Avoid a crash caused by the scmnd->scsi_done(scmnd) call in
srp_process_rsp() being invoked with scsi_done == NULL.  This can
happen if a reply is received during or after a command abort.
Reported-by: NJoseph Glanville <joseph.glanville@orionvm.com.au>
Reference: http://marc.info/?l=linux-rdma&m=134314367801595
Cc: <stable@vger.kernel.org>
Acked-by: NDavid Dillow <dillowda@ornl.gov>
Signed-off-by: NBart Van Assche <bvanassche@acm.org>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Convert a 0 error return code to a negative one, as returned elsewhere
in the function.

A simplified version of the semantic match that finds this problem is
as follows: (http://coccinelle.lip6.fr/)

// <smpl>
@@
identifier ret;
expression e,e1,e2,e3,e4,x;
@@

(
if (\(ret != 0\|ret < 0\) || ...) { ... return ...; }
|
ret = 0
)
... when != ret = e1
*x = \(kmalloc\|kzalloc\|kcalloc\|devm_kzalloc\|ioremap\|ioremap_nocache\|devm_ioremap\|devm_ioremap_nocache\)(...);
... when != x = e2
    when != ret = e3
*if (x == NULL || ...)
{
  ... when != ret = e4
*  return ret;
}
// </smpl>
Signed-off-by: NJulia Lawall <Julia.Lawall@lip6.fr>
Acked-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Correct spelling typos in comments in drivers/infiniband.
Signed-off-by: NMasanari Iida <standby24x7@gmail.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Commit b63b70d8 ("IPoIB: Use a private hash table for path lookup
in xmit path") introduced a bug where in ipoib_neigh_free() (which is
called from a few errors flows in the driver), rcu_dereference() is
invoked with the wrong pointer object, which results in a crash.
Signed-off-by: NShlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: NOr Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Commit b63b70d8 ("IPoIB: Use a private hash table for path lookup
in xmit path") introduced a bug where in ipoib_cm_destroy_tx() a CM
object is moved between lists without any supported locking.  Under a
stress test, this eventually leads to list corruption and a crash.

Previously when this routine was called, callers were taking the
device priv lock.  Currently this function is called from the RCU
callback associated with neighbour deletion.  Fix the race by taking
the same lock we used to before.
Signed-off-by: NShlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: NOr Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

It is possible for asynchronous RDMA_CM_EVENT_ESTABLISHED events to be
generated with ctx->uid == 0, because ucma_set_event_context() copies
ctx->uid to the event structure outside of ctx->file->mut.  This leads
to a crash in the userspace library, since it gets a bogus event.

Fix this by taking the mutex a bit earlier in ucma_event_handler.
Signed-off-by: NTatyana Nikolova <Tatyana.E.Nikolova@intel.com>
Signed-off-by: NSean Hefty <Sean.Hefty@intel.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

If CONFIG_VLAN_8021Q is not set, then vlan_dev_real_dev() just goes BUG(),
so we shouldn't call it unless we're actually dealing with a VLAN netdev.
Reported-by: NFengguang Wu <fengguang.wu@intel.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

The sm_lock spinlock is taken in the process context by
mlx4_ib_modify_device, and in the interrupt context by update_sm_ah,
so we need to take that spinlock with irqsave, and release it with
irqrestore.

Lockdeps reports this as follows:

    [ INFO: inconsistent lock state ]
    3.5.0+ #20 Not tainted
    inconsistent {HARDIRQ-ON-W} -> {IN-HARDIRQ-W} usage.
    swapper/0/0 [HC1[1]:SC0[0]:HE0:SE1] takes:
    (&(&ibdev->sm_lock)->rlock){?.+...}, at: [<ffffffffa028af1d>] update_sm_ah+0xad/0x100 [mlx4_ib]
    {HARDIRQ-ON-W} state was registered at:
      [<ffffffff810b84a0>] mark_irqflags+0x120/0x190
      [<ffffffff810b9ce7>] __lock_acquire+0x307/0x4c0
      [<ffffffff810b9f51>] lock_acquire+0xb1/0x150
      [<ffffffff815523b1>] _raw_spin_lock+0x41/0x50
      [<ffffffffa028d563>] mlx4_ib_modify_device+0x63/0x240 [mlx4_ib]
      [<ffffffffa026d1fc>] ib_modify_device+0x1c/0x20 [ib_core]
      [<ffffffffa026c353>] set_node_desc+0x83/0xc0 [ib_core]
      [<ffffffff8136a150>] dev_attr_store+0x20/0x30
      [<ffffffff81201fd6>] sysfs_write_file+0xe6/0x170
      [<ffffffff8118da38>] vfs_write+0xc8/0x190
      [<ffffffff8118dc01>] sys_write+0x51/0x90
      [<ffffffff8155b869>] system_call_fastpath+0x16/0x1b

    ...
    *** DEADLOCK ***

    1 lock held by swapper/0/0:

    stack backtrace:
    Pid: 0, comm: swapper/0 Not tainted 3.5.0+ #20
    Call Trace:
    <IRQ>  [<ffffffff810b7bea>] print_usage_bug+0x18a/0x190
    [<ffffffff810b7370>] ? print_irq_inversion_bug+0x210/0x210
    [<ffffffff810b7fb2>] mark_lock_irq+0xf2/0x280
    [<ffffffff810b8290>] mark_lock+0x150/0x240
    [<ffffffff810b84ef>] mark_irqflags+0x16f/0x190
    [<ffffffff810b9ce7>] __lock_acquire+0x307/0x4c0
    [<ffffffffa028af1d>] ? update_sm_ah+0xad/0x100 [mlx4_ib]
    [<ffffffff810b9f51>] lock_acquire+0xb1/0x150
    [<ffffffffa028af1d>] ? update_sm_ah+0xad/0x100 [mlx4_ib]
    [<ffffffff815523b1>] _raw_spin_lock+0x41/0x50
    [<ffffffffa028af1d>] ? update_sm_ah+0xad/0x100 [mlx4_ib]
    [<ffffffffa026b2fa>] ? ib_create_ah+0x1a/0x40 [ib_core]
    [<ffffffffa028af1d>] update_sm_ah+0xad/0x100 [mlx4_ib]
    [<ffffffff810c27c3>] ? is_module_address+0x23/0x30
    [<ffffffffa028b05b>] handle_port_mgmt_change_event+0xeb/0x150 [mlx4_ib]
    [<ffffffffa028c177>] mlx4_ib_event+0x117/0x160 [mlx4_ib]
    [<ffffffff81552501>] ? _raw_spin_lock_irqsave+0x61/0x70
    [<ffffffffa022718c>] mlx4_dispatch_event+0x6c/0x90 [mlx4_core]
    [<ffffffffa0221b40>] mlx4_eq_int+0x500/0x950 [mlx4_core]

Reported by: Or Gerlitz <ogerlitz@mellanox.com>
Tested-by: NBart Van Assche <bvanassche@acm.org>
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NRoland 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: NShlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: NOr Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Commit 36a8f01c ("IB/qib: Add congestion control agent
implementation") tries to store the value 1984 in a u8, which leads to
truncation.  Fix this by making the member big enough.

This bug was detected by a smatch warning.
Reported-by: NDan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: NRamkrishna Vepa <ramkrishna.vepa@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Suggested by scripts/coccinelle/api/memdup_user.cocci.
Reported-by: NFengguang Wu <fengguang.wu@intel.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

It looks like one check was accidentally duplicated, and the other 3
checks were left out.  This was detected by scripts/coccinelle/tests/doubletest.cocci:

    drivers/infiniband/hw/ocrdma/ocrdma_verbs.c:895:6-54: duplicated argument to && or ||
Reported-by: NFengguang Wu <fengguang.wu@intel.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Suggested by scripts/coccinelle/api/ptr_ret.cocci.
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Elminate some simple_strto* usage.

checkpatch also noted pr_ conversations, which have been done as
recommended.  The pr_fmt() define is used to shorten line length.

Other multi-line string warnings are also elmininated.
Reviewed-by: NDean Luick <dean.luick@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Add a congestion control agent in the driver that handles gets and
sets from the congestion control manager in the fabric for the
Performance Scale Messaging (PSM) library.
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Profiling has shown that sdma_lock is proving a bottleneck for
performance. The situations include:
 - RDMA reads when krcvqs > 1
 - post sends from multiple threads

For RDMA read the current global qib_wq mechanism runs on all CPUs
and contends for the sdma_lock when multiple RMDA read requests are
fielded on differenct CPUs. For post sends, the direct call to
qib_do_send() from multiple threads causes the contention.

Since the sdma mechanism is per port, this fix converts the existing
workqueue to a per port single thread workqueue to reduce the lock
contention in the RDMA read case, and for any other case where the QP
is scheduled via the workqueue mechanism from more than 1 CPU.

For the post send case, This patch modifies the post send code to test
for a non empty sdma engine.  If the sdma is not idle the (now single
thread) workqueue will be used to trigger the send engine instead of
the direct call to qib_do_send().
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

There is a cut-and-paste typo in the function qib_pci_slot_reset()
where it prints that the "link_reset" function is called rather than
the "slot_reset" function.  This makes the message misleading.
Signed-off-by: NBetty Dall <betty.dall@hp.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>

Enable callers of mlx4_assign_eq to supply a pointer to cpu_rmap.
If supplied, the assigned IRQ is tracked using rmap infrastructure.
Signed-off-by: NAmir Vadai <amirv@mellanox.com>
Signed-off-by: NOr Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Commit af061a64 ("IB/qib: Use RCU for qpn lookup") introduced sparse
warnings.

This patch corrects those issues.
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NRoland Dreier <roland@purestorage.com>