A narrow window for race condition still exist between
multicast join thread and *dev_flush workers.
A kernel crash caused by prolong erratic link state changes
was observed (most likely a faulty cabling):
[167275.656270] BUG: unable to handle kernel NULL pointer dereference at
0000000000000020
[167275.665973] IP: [<ffffffffa05f8f2e>] ipoib_mcast_join+0xae/0x1d0 [ib_ipoib]
[167275.674443] PGD 0
[167275.677373] Oops: 0000 [#1] SMP
...
[167275.977530] Call Trace:
[167275.982225] [<ffffffffa05f92f0>] ? ipoib_mcast_free+0x200/0x200 [ib_ipoib]
[167275.992024] [<ffffffffa05fa1b7>] ipoib_mcast_join_task+0x2a7/0x490
[ib_ipoib]
[167276.002149] [<ffffffff8109d5fb>] process_one_work+0x17b/0x470
[167276.010754] [<ffffffff8109e3cb>] worker_thread+0x11b/0x400
[167276.019088] [<ffffffff8109e2b0>] ? rescuer_thread+0x400/0x400
[167276.027737] [<ffffffff810a5aef>] kthread+0xcf/0xe0
Here was a hit spot:
ipoib_mcast_join() {
..............
rec.qkey = priv->broadcast->mcmember.qkey;
^^^^^^^
.....
}
Proposed patch should prevent multicast join task to continue
if link state change is detected.
Signed-off-by: Alex Estrin <alex.estrin@intel.com>
Changes from v4:
- as suggested by Doug Ledford, optimized spinlock usage,
i.e. ipoib_mcast_join() is called with lock held.
Changes from v3:
- sync with priv->lock before flag check.
Chages from v2:
- Move check for OPER_UP flag state to mcast_join() to
ensure no event worker is in progress.
- minor style fixes.
Changes from v1:
- No need to lock again if error detected.
Signed-off-by: Doug Ledford <dledford@redhat.com>
We are seeing this warning: at net/core/skbuff.c:4174
and before commit a44878d100 ("IB/ipoib: Use one linear skb in RX flow")
skb truesize was not being set when ipoib was using just one skb.
Removing this line avoids the warning when running tcp tests like iperf.
Fixes: a44878d100 ("IB/ipoib: Use one linear skb in RX flow")
Signed-off-by: Carol L Soto <clsoto@linux.vnet.ibm.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
ipoib_mcast_restart_task calls ipoib_mcast_remove_list with the
parameter mcast->dev. That mcast is a temporary (used as an iterator)
variable that may be uninitialized.
There is no need to send the variable dev to the function, as each mcast
has its dev as a member in the mcast struct.
This causes the next panic:
RIP: 0010: ipoib_mcast_leave+0x6d/0xf0 [ib_ipoib]
RSP: 0018: EFLAGS: 00010246
RAX: f0201 RBX: 24e00 RCX: 00000
....
....
Stack:
Call Trace:
ipoib_mcast_remove_list+0x3a/0x70 [ib_ipoib]
ipoib_mcast_restart_task+0x3bb/0x520 [ib_ipoib]
process_one_work+0x164/0x470
worker_thread+0x11d/0x420
...
Fixes: 5a0e81f6f4 ('IB/IPoIB: factor out common multicast list removal code')
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Reported-by: Doron Tsur <doront@mellanox.com>
Reviewed-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Code cleanup to move multicast specific code that checks for
a sendonly join to ipoib_multicast.c. This allows the removal
of the export of __ipoib_mcast_find().
Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Code cleanup to remove multicast specific code from ipoib_main.c
The removal of a list of multicast groups occurs in three places.
Create a new function ipoib_mcast_remove_list(). Use this new
function in ipoib_main.c too.
That in turn allows the dropping of two functions that were
exported from ipoib_multicast.c for expiration of mc groups.
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Instead, use the cached copy of the attributes present on the device.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This adds an abstraction that allows ULPs to simply pass a completion
object and completion callback with each submitted WR and let the RDMA
core handle the nitty gritty details of how to handle completion
interrupts and poll the CQ.
In detail there is a new ib_cqe structure which just contains the
completion callback, and which can be used to get at the containing
object using container_of. It is pointed to by the WR and WC as an
alternative to the wr_id field, similar to how many ULPs already use
the field to store a pointer using casts.
A driver using the new completion callbacks allocates it's CQs using
the new ib_create_cq API, which in addition to the number of CQEs and
the completion vectors also takes a mode on how we poll for CQEs.
Three modes are available: direct for drivers that never take CQ
interrupts and just poll for them, softirq to poll from softirq context
using the to be renamed blk-iopoll infrastructure which takes care of
rearming and budgeting, or a workqueue for consumer who want to be
called from user context.
Thanks a lot to Sagi Grimberg who helped reviewing the API, wrote
the current version of the workqueue code because my two previous
attempts sucked too much and converted the iSER initiator to the new
API.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Adding an ability to query the IB cache by a netdev and get the
attributes of a GID. These parameters are necessary in order to
successfully resolve the required GID (when the netdevice is known)
and get the Ethernet L2 attributes from a GID.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-By: Devesh Sharma <devesh.sharma@avagotech.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
When we leave the multicast group on expiration of a neighbor we
do not free the mcast structure. This results in a memory leak
that causes ib_dealloc_pd to fail and print a WARN_ON message
and backtrace.
Fixes: bd99b2e05c (IB/ipoib: Expire sendonly multicast joins)
Signed-off-by: Christoph Lameter <cl@linux.com>
Tested-by: Sagi Grimberg <sagig@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This patch split up struct ib_send_wr so that all non-trivial verbs
use their own structure which embedds struct ib_send_wr. This dramaticly
shrinks the size of a WR for most common operations:
sizeof(struct ib_send_wr) (old): 96
sizeof(struct ib_send_wr): 48
sizeof(struct ib_rdma_wr): 64
sizeof(struct ib_atomic_wr): 96
sizeof(struct ib_ud_wr): 88
sizeof(struct ib_fast_reg_wr): 88
sizeof(struct ib_bind_mw_wr): 96
sizeof(struct ib_sig_handover_wr): 80
And with Sagi's pending MR rework the fast registration WR will also be
down to a reasonable size:
sizeof(struct ib_fastreg_wr): 64
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com> [srp, srpt]
Reviewed-by: Chuck Lever <chuck.lever@oracle.com> [sunrpc]
Tested-by: Haggai Eran <haggaie@mellanox.com>
Tested-by: Sagi Grimberg <sagig@mellanox.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
When performing sendonly joins, we queue the packets that trigger
the join until the join completes. This may take on the order of
hundreds of milliseconds. It is easy to have many more than three
packets come in during that time. Expand the maximum queue depth
in order to try and prevent dropped packets during the time it
takes to join the multicast group.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Since IPoIB should, as much as possible, emulate how multicast
sends work on Ethernet for regular TCP/IP apps, there should be
no requirement to subscribe to a multicast group before your
sends are properly sent. However, due to the difference in how
multicast is handled on InfiniBand, we must join the appropriate
multicast group before we can send to it. Previously we tried
not to trigger the auto-create feature of the subnet manager when
doing this because we didn't have tracking of these sendonly
groups and the auto-creation might never get undone. The previous
patch added timing to these sendonly joins and allows us to
leave them after a reasonable idle expiration time. So supply
all of the information needed to auto-create group.
Signed-off-by: Doug Ledford <dledford@redhat.com>
On neighbor expiration, check to see if the neighbor was actually a
sendonly multicast join, and if so, leave the multicast group as we
expire the neighbor.
Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
We expect send only joins to fail, it just means there are no listeners
for the group. The correct thing to do is silently drop the packet
at source.
Eg avahi will full join 224.0.0.251 which causes a send only IGMP packet
to 224.0.0.22, and then a warning level kmessage like this:
ib0: sendonly multicast join failed for ff12:401b:ffff:0000:0000:0000:0000:0016, status -22
If there is no IP router listening to IGMP.
Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Even though we don't expect the group to be created by the SM we
sill need to provide all the parameters to force the SM to validate
they are correct.
Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The majority of callers never check the return value, and even if they
did, they can't do anything about a failure.
All possible failure cases represent a bug in the caller, so just
WARN_ON inside the function instead.
This fixes a few random errors:
net/rd/iw.c infinite loops while it fails. (racing with EBUSY?)
This also lays the ground work to get rid of error return from the
drivers. Most drivers do not error, the few that do are broken since
it cannot be handled.
Since uverbs can legitimately make use of EBUSY, open code the
check.
Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Reviewed-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The pd now has a local_dma_lkey member which completely replaces
ib_get_dma_mr, use it instead.
Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Now that there are no ib_cm clients using the compare_data feature for
matching IB CM requests' private data, remove the compare_data parameter of
ib_cm_listen and remove the code implementing the feature.
Signed-off-by: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Implement the get_net_device_by_port_pkey_ip callback that returns network
device to ib_core according to connection parameters. Check the ipoib
device and iterate over all child devices to look for a match.
For each IPoIB device we iterate through all upper devices when searching
for a matching IP, in order to support bonding.
Signed-off-by: Guy Shapiro <guysh@mellanox.com>
Signed-off-by: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Yotam Kenneth <yotamke@mellanox.com>
Signed-off-by: Shachar Raindel <raindel@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
An ib_client callback that is called with the lists_rwsem locked only for
read is protected from changes to the IB client lists, but not from
ib_unregister_device() freeing its client data. This is because
ib_unregister_device() will remove the device from the device list with
lists_rwsem locked for write, but perform the rest of the cleanup,
including the call to remove() without that lock.
Mark client data that is undergoing de-registration with a new going_down
flag in the client data context. Lock the client data list with lists_rwsem
for write in addition to using the spinlock, so that functions calling the
callback would be able to lock only lists_rwsem for read and let callbacks
sleep.
Since ib_unregister_client() now marks the client data context, no need for
remove() to search the context again, so pass the client data directly to
remove() callbacks.
Reviewed-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
If the above is turned off then ipoib_cm_dev_init unconditionally
returns ENOSYS, and the newly added error handling in
0b3957 prevents ipoib from coming up at all:
kernel: mlx4_0: ipoib_transport_dev_init failed
kernel: mlx4_0: failed to initialize port 1 (ret = -12)
Fixes: 0b39578bcd (IB/ipoib: Use dedicated workqueues per interface)
Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
When switching between modes (datagram / connected) change the MTU
accordingly.
datagram mode up to 4K, connected mode up to (64K - 0x10).
Signed-off-by: ELi Cohen <eli@mellanox.com>
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
By default, IPoIB-CM driver uses 64k MTU. Larger MTU gives better
performance.
This MTU plus overhead puts the memory allocation for IP based packets at
32 4k pages (order 5), which have to be contiguous.
When the system memory under pressure, it was observed that allocating 128k
contiguous physical memory is difficult and causes serious errors (such as
system becomes unusable).
This enhancement resolve the issue by removing the physically contiguous
memory requirement using Scatter/Gather feature that exists in Linux stack.
With this fix Scatter-Gather will be supported also in connected mode.
This change reverts some of the change made in commit e112373fd6
("IPoIB/cm: Reduce connected mode TX object size").
The ability to use SG in IPoIB CM is possible because the coupling
between NETIF_F_SG and NETIF_F_CSUM was removed in commit
ec5f061564 ("net: Kill link between CSUM and SG features.")
Signed-off-by: Yuval Shaia <yuval.shaia@oracle.com>
Acked-by: Christian Marie <christian@ponies.io>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Error values of ib_query_port() and ib_query_device() weren't propagated
correctly. Because of that, ipoib_add_port() could return NULL value,
which escaped the IS_ERR() check in ipoib_add_one() and we crashed.
Signed-off-by: Amir Vadai <amirv@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Persuant to Liran's comments on node_type on linux-rdma
mailing list:
In an effort to reform the RDMA core and ULPs to minimize use of
node_type in struct ib_device, an additional bit is added to
struct ib_device for is_switch (IB switch). This is needed
to be initialized by any IB switch device driver. This is a
NEW requirement on such device drivers which are all
"out of tree".
In addition, an ib_switch helper was added to ib_verbs.h
based on the is_switch device bit rather than node_type
(although those should be consistent).
The RDMA core (MAD, SMI, agent, sa_query, multicast, sysfs)
as well as (IPoIB and SRP) ULPs are updated where
appropriate to use this new helper. In some cases,
the helper is now used under the covers of using
rdma_[start end]_port rather than the open coding
previously used.
Reviewed-by: Sean Hefty <sean.hefty@intel.com>
Reviewed-By: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Tested-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Hal Rosenstock <hal@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Currently, ib_create_cq uses cqe and comp_vecotr instead
of the extendible ib_cq_init_attr struct.
Earlier patches already changed the vendors to work with
ib_cq_init_attr. This patch changes the consumers too.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Avoid that sparse complains about ipoib_neigh_hash_init(). This
patch does not change any functionality. See also patch "IPoIB:
Fix memory leak in the neigh table deletion flow" (commit ID
66172c0993).
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Use raw management helpers to reform IB-ulp ipoib.
Signed-off-by: Michael Wang <yun.wang@profitbricks.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Tested-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Sean Hefty <sean.hefty@intel.com>
Reviewed-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Tested-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
See also patch "IPoIB/cm: Add connected mode support for devices
without SRQs" (commit ID 68e995a295). Detected by smatch.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Pradeep Satyanarayana <pradeeps@linux.vnet.ibm.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Currently, iflink of the parent interface was always accessed, even
when interface didn't have a parent and hence we crashed there.
Handle the interface types properly: for a child interface, return
the ifindex of the parent, for parent interface, return its ifindex.
For child devices, make sure to set the parent pointer prior to
invoking register_netdevice(), this allows the new ndo to be called
by the stack immediately after the child device is registered.
Fixes: 5aa7add8f1 ('infiniband/ipoib: implement ndo_get_iflink')
Reported-by: Honggang Li <honli@redhat.com>
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Honggang Li <honli@redhat.com>
Reviewed-By: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>+
Acked-by: Nicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
After Doug Ledford's changes there is no need in that bit, it's
semantic becomes subset of the IPOIB_FLAG_OPER_UP bit.
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Whenever there is no path->ah to the destination, keep only defined
number of skb's. Otherwise there are cases that the driver can keep
infinite list of skb's.
For example, when one device want to send unicast arp to the destination,
and from some reason the SM doesn't respond, the driver currently keeps
all the skb's. If that unicast arp traffic stopped, all these skb's
are kept by the path object till the interface is down.
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
As the result of a completion error the QP can moved to SQE state by
the hardware. Since it's not the Error state, there are no flushes
and hence the driver doesn't know about that.
The fix creates a task that after completion with error which is not a
flush tracks the QP state and if it is in SQE state moves it back to RTS.
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Update the cached broadcast record in the priv object after every new
join of this broadcast domain group.
These values are needed for the port configuration (MTU size) and to
all the new multicast (non-broadcast) join requests initial parameters.
For example, SM starts with 2K MTU for all the fabric, and after that it
restarts (or handover to new SM) with new port configuration of 4K MTU.
Without using the new values, the driver will keep its old configuration
of 2K and will not apply the new configuration of 4K.
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
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>
We needed the mcast_mutex when we had to prevent the join completion
callback from having the value it stored in mcast->mc overwritten
by a delayed return from ib_sa_join_multicast. By storing the return
of ib_sa_join_multicast in an intermediate variable, we prevent a
delayed return from ib_sa_join_multicast overwriting the valid
contents of mcast->mc, and we no longer need a mutex to force the
join callback to run after the return of ib_sa_join_multicast. This
allows us to do away with the mutex entirely and protect our critical
sections with a just a spinlock instead. This is highly desirable
as there were some places where we couldn't use a mutex because the
code was not allowed to sleep, and so we were currently using a mix
of mutex and spinlock to protect what we needed to protect. Now we
only have a spin lock and the locking complexity is greatly reduced.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Allow the ipoib layer to attempt to join all outstanding multicast
groups at once. The ib_sa layer will serialize multiple attempts to
join the same group, but will process attempts to join different groups
in parallel. Take advantage of that.
In order to make this happen, change the mcast_join_thread to loop
through all needed joins, sending a join request for each one that we
still need to join. There are a few special cases we handle though:
1) Don't attempt to join anything but the broadcast group until the join
of the broadcast group has succeeded.
2) No longer restart the join task at the end of completion handling.
If we completed successfully, we are done. The join task now needs kicked
either by mcast_send or mcast_restart_task or mcast_start_thread, but
should not need started anytime else except when scheduling a backoff
attempt to rejoin.
3) No longer use separate join/completion routines for regular and
sendonly joins, pass them all through the same routine and just do the
right thing based on the SENDONLY join flag.
4) Only try to join a SENDONLY join twice, then drop the packets and
quit trying. We leave the mcast group in the list so that if we get a
new packet, all that we have to do is queue up the packet and restart
the join task and it will automatically try to join twice and then
either send or flush the queue again.
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>
We blindly assume that we can just take the rtnl lock and that will
prevent races with downing this interface. Unfortunately, that's not
the case. In ipoib_mcast_stop_thread() we will call flush_workqueue()
in an attempt to clear out all remaining instances of ipoib_join_task.
But, since this task is put on the same workqueue as the join task,
the flush_workqueue waits on this thread too. But this thread is
deadlocked on the rtnl lock. The better thing here is to use trylock
and loop on that until we either get the lock or we see that
FLAG_OPER_UP has been cleared, in which case we don't need to do
anything anyway and we just return.
While investigating which flag should be used, FLAG_ADMIN_UP or
FLAG_OPER_UP, it was determined that FLAG_OPER_UP was the more
appropriate flag to use. However, there was a mix of these two flags in
use in the existing code. So while we check for that flag here as part
of this race fix, also cleanup the two places that had used the less
appropriate flag for their tests.
Signed-off-by: Doug Ledford <dledford@redhat.com>
The ipoib_mcast_flush_dev routine is called with the rtnl_lock held and
needs to keep it held. It also needs to call flush_workqueue() to flush
out any outstanding work. In the past, we've had to try and make sure
that we didn't flush out any outstanding join completions because they
also wanted to grab rtnl_lock() and that would deadlock. It turns out
that the only thing in the join completion handler that needs this lock
can be safely moved to our carrier_on_task, thereby reducing the
potential for the join completion code and the flush code to deadlock
against each other.
Signed-off-by: Doug Ledford <dledford@redhat.com>
In preparation for using per device work queues, we need to move the
start of the neighbor thread task to after ipoib_ib_dev_init and move
the destruction of the neighbor task to before ipoib_ib_dev_cleanup.
Otherwise we will end up freeing our workqueue with work possibly
still on it.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Create a an ipoib_flush_ah and ipoib_stop_ah routines to use at
appropriate times to flush out all remaining ah entries before we shut
the device down.
Because neighbors and mcast entries can each have a reference on any
given ah, we must make sure to free all of those first before our ah
will actually have a 0 refcount and be able to be reaped.
This factoring is needed in preparation for having per-device work
queues. The original per-device workqueue code resulted in the following
error message:
<ibdev>: ib_dealloc_pd failed
That error was tracked down to this issue. With the changes to which
workqueues were flushed when, there were no flushes of the per device
workqueue after the last ah's were freed, resulting in an attempt to
dealloc the pd with outstanding resources still allocated. This code
puts the explicit flushes in the needed places to avoid that problem.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Don't use dev->iflink anymore.
CC: Roland Dreier <roland@kernel.org>
Signed-off-by: Nicolas Dichtel <nicolas.dichtel@6wind.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
