T3 hardware doorbell FIFO overflows can cause application stalls due
to lost doorbell ring events. This has been seen when running large
NP IMB alltoall MPI jobs. The T3 hardware supports an xon/xoff-type
flow control mechanism to help avoid overflowing the HW doorbell FIFO.
This patch uses these interrupts to disable RDMA QP doorbell rings
when we near an overflow condition, and then turn them back on (and
ring all the active QP doorbells) when when the doorbell FIFO empties
out. In addition if an doorbell ring is dropped by the hardware, the
code will now recover.
Design:
cxgb3:
- enable these DB interrupts
- in the interrupt handler, schedule work tasks to call the ULPs event
handlers with the new events.
- ring all the qset txqs when an overflow is detected.
iw_cxgb3:
- disable db ringing on all active qps when we get the DB_FULL event
- enable db ringing on all active qps and ring all active dbs when we get
the DB_EMPTY event
- On DB_DROP event:
- disable db rings in the event handler
- delay-schedule a work task which rings and enables the dbs on
all active qps.
- in post_send and post_recv logic, don't ring the db if it's disabled.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Failure to rearm a CQ means the cxgb3 device is wedged, but we shouldn't
kill the whole system with a BUG_ON() if this happens.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland/infiniband:
RDMA/cm: Revert association of an RDMA device when binding to loopback
Revert the following change from commit 6f8372b6 ("RDMA/cm: fix
loopback address support")
The defined behavior of rdma_bind_addr is to associate an RDMA
device with an rdma_cm_id, as long as the user specified a non-
zero address. (ie they weren't just trying to reserve a port)
Currently, if the loopback address is passed to rdma_bind_addr,
no device is associated with the rdma_cm_id. Fix this.
It turns out that important apps such as Open MPI depend on
rdma_bind_addr() NOT associating any RDMA device when binding to a
loopback address. Open MPI is being updated to deal with this, but at
least until a new Open MPI release is available, maintain the previous
behavior: allow rdma_bind_addr() to succeed, but do not bind to a
device.
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Correct misspelled "CONFIG_IPv6" that was introduced in commit
d14714df ("IB/addr: Fix IPv6 routing lookup"). The config variable
should be all uppercase.
Signed-off-by: Robert P. J. Day <rpjday@crashcourse.ca>
[ This was my fault when I munged the original patch. - Roland ]
Signed-off-by: Roland Dreier <rolandd@cisco.com>
In mlx4_ib_post_recv(), we should check the queue for overflow using
recv_cq instead of send_cq (current code looks like a copy-and-paste
mistake).
Signed-off-by: Or Gerlitz <ogerlitz@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
As for memfree mthca hardware, ConnectX also requires SRQ WQE scatter
entries to be initialized with the invalid L_Key at SRQ creation time.
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fix the "ignoring return value of '...', declared with attribute
warn_unused_result" compiler warning in several users of the new kfifo
API.
It removes the __must_check attribute from kfifo_in() and
kfifo_in_locked() which must not necessary performed.
Fix the allocation bug in the nozomi driver file, by moving out the
kfifo_alloc from the interrupt handler into the probe function.
Fix the kfifo_out() and kfifo_out_locked() users to handle a unexpected
end of fifo.
Signed-off-by: Stefani Seibold <stefani@seibold.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
rename kfifo_put... into kfifo_in... to prevent miss use of old non in
kernel-tree drivers
ditto for kfifo_get... -> kfifo_out...
Improve the prototypes of kfifo_in and kfifo_out to make the kerneldoc
annotations more readable.
Add mini "howto porting to the new API" in kfifo.h
Signed-off-by: Stefani Seibold <stefani@seibold.net>
Acked-by: Greg Kroah-Hartman <gregkh@suse.de>
Acked-by: Mauro Carvalho Chehab <mchehab@redhat.com>
Acked-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Move the pointer to the spinlock out of struct kfifo. Most users in
tree do not actually use a spinlock, so the few exceptions now have to
call kfifo_{get,put}_locked, which takes an extra argument to a
spinlock.
Signed-off-by: Stefani Seibold <stefani@seibold.net>
Acked-by: Greg Kroah-Hartman <gregkh@suse.de>
Acked-by: Mauro Carvalho Chehab <mchehab@redhat.com>
Acked-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is a new generic kernel FIFO implementation.
The current kernel fifo API is not very widely used, because it has to
many constrains. Only 17 files in the current 2.6.31-rc5 used it.
FIFO's are like list's a very basic thing and a kfifo API which handles
the most use case would save a lot of development time and memory
resources.
I think this are the reasons why kfifo is not in use:
- The API is to simple, important functions are missing
- A fifo can be only allocated dynamically
- There is a requirement of a spinlock whether you need it or not
- There is no support for data records inside a fifo
So I decided to extend the kfifo in a more generic way without blowing up
the API to much. The new API has the following benefits:
- Generic usage: For kernel internal use and/or device driver.
- Provide an API for the most use case.
- Slim API: The whole API provides 25 functions.
- Linux style habit.
- DECLARE_KFIFO, DEFINE_KFIFO and INIT_KFIFO Macros
- Direct copy_to_user from the fifo and copy_from_user into the fifo.
- The kfifo itself is an in place member of the using data structure, this save an
indirection access and does not waste the kernel allocator.
- Lockless access: if only one reader and one writer is active on the fifo,
which is the common use case, no additional locking is necessary.
- Remove spinlock - give the user the freedom of choice what kind of locking to use if
one is required.
- Ability to handle records. Three type of records are supported:
- Variable length records between 0-255 bytes, with a record size
field of 1 bytes.
- Variable length records between 0-65535 bytes, with a record size
field of 2 bytes.
- Fixed size records, which no record size field.
- Preserve memory resource.
- Performance!
- Easy to use!
This patch:
Since most users want to have the kfifo as part of another object,
reorganize the code to allow including struct kfifo in another data
structure. This requires changing the kfifo_alloc and kfifo_init
prototypes so that we pass an existing kfifo pointer into them. This
patch changes the implementation and all existing users.
[akpm@linux-foundation.org: fix warning]
Signed-off-by: Stefani Seibold <stefani@seibold.net>
Acked-by: Greg Kroah-Hartman <gregkh@suse.de>
Acked-by: Mauro Carvalho Chehab <mchehab@redhat.com>
Acked-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs-2.6: (38 commits)
direct I/O fallback sync simplification
ocfs: stop using do_sync_mapping_range
cleanup blockdev_direct_IO locking
make generic_acl slightly more generic
sanitize xattr handler prototypes
libfs: move EXPORT_SYMBOL for d_alloc_name
vfs: force reval of target when following LAST_BIND symlinks (try #7)
ima: limit imbalance msg
Untangling ima mess, part 3: kill dead code in ima
Untangling ima mess, part 2: deal with counters
Untangling ima mess, part 1: alloc_file()
O_TRUNC open shouldn't fail after file truncation
ima: call ima_inode_free ima_inode_free
IMA: clean up the IMA counts updating code
ima: only insert at inode creation time
ima: valid return code from ima_inode_alloc
fs: move get_empty_filp() deffinition to internal.h
Sanitize exec_permission_lite()
Kill cached_lookup() and real_lookup()
Kill path_lookup_open()
...
Trivial conflicts in fs/direct-io.c
Always set bad_wr when an immediate error is detected. Return ENOMEM
for queue full instead of EINVAL to match other drivers.
Signed-off-by: Frank Zago <fzago@systemfabricworks.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu: (34 commits)
m68k: rename global variable vmalloc_end to m68k_vmalloc_end
percpu: add missing per_cpu_ptr_to_phys() definition for UP
percpu: Fix kdump failure if booted with percpu_alloc=page
percpu: make misc percpu symbols unique
percpu: make percpu symbols in ia64 unique
percpu: make percpu symbols in powerpc unique
percpu: make percpu symbols in x86 unique
percpu: make percpu symbols in xen unique
percpu: make percpu symbols in cpufreq unique
percpu: make percpu symbols in oprofile unique
percpu: make percpu symbols in tracer unique
percpu: make percpu symbols under kernel/ and mm/ unique
percpu: remove some sparse warnings
percpu: make alloc_percpu() handle array types
vmalloc: fix use of non-existent percpu variable in put_cpu_var()
this_cpu: Use this_cpu_xx in trace_functions_graph.c
this_cpu: Use this_cpu_xx for ftrace
this_cpu: Use this_cpu_xx in nmi handling
this_cpu: Use this_cpu operations in RCU
this_cpu: Use this_cpu ops for VM statistics
...
Fix up trivial (famous last words) global per-cpu naming conflicts in
arch/x86/kvm/svm.c
mm/slab.c
* git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi-misc-2.6: (222 commits)
[SCSI] zfcp: Remove flag ZFCP_STATUS_FSFREQ_TMFUNCNOTSUPP
[SCSI] zfcp: Activate fc4s attributes for zfcp in FC transport class
[SCSI] zfcp: Block scsi_eh thread for rport state BLOCKED
[SCSI] zfcp: Update FSF error reporting
[SCSI] zfcp: Improve ELS ADISC handling
[SCSI] zfcp: Simplify handling of ct and els requests
[SCSI] zfcp: Remove ZFCP_DID_MASK
[SCSI] zfcp: Move WKA port to zfcp FC code
[SCSI] zfcp: Use common code definitions for FC CT structs
[SCSI] zfcp: Use common code definitions for FC ELS structs
[SCSI] zfcp: Update FCP protocol related code
[SCSI] zfcp: Dont fail SCSI commands when transitioning to blocked fc_rport
[SCSI] zfcp: Assign scheduled work to driver queue
[SCSI] zfcp: Remove STATUS_COMMON_REMOVE flag as it is not required anymore
[SCSI] zfcp: Implement module unloading
[SCSI] zfcp: Merge trace code for fsf requests in one function
[SCSI] zfcp: Access ports and units with container_of in sysfs code
[SCSI] zfcp: Remove suspend callback
[SCSI] zfcp: Remove global config_mutex
[SCSI] zfcp: Replace local reference counting with common kref
...
When the remote node's ethernet address changes, the connection keeps
trying to connect using the old address. The connection wil continue
failing until the driver is unloaded and loaded again (eiter reboot or
rmmod). Fix this by checking that the NIC has the correct address
before starting a connection.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
A FIN that is received during an MPA start up sequence causes a
timeout in iwcm.c. The connection has not been completely closed so
the iwcm code is waiting for resources to be cleaned up. This closes
the connection so everything cleans up correctly.
Signed-off-by: Don Wood <donald.e.wood@intel.com>
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
We fail when creating many qps as kmap() fails for sq_vbase.
Fix this by doing kunmap() as soon as we are done with sq_vbase.
We do kunmap() in one of the locations below:
(1) nes_destroy_qp()
(2) nes_accept()
(3) nes_connect_event
We keep a flag to avoid multiple calls to kunmap().
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
STags are generated randomly but the driver does not correctly prevent
a zero STag. Using STag zero is privileged and causes a user space
application to fail. This change prevents the driver from trying to
allocate a zero STag.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
While running a Xansation test, an active side node crashed. The
problem started on the passive side, which generated an STtag that was
0. The passive side sent a TERMINATE instead of an MPA REJECT msg.
The active side, receives TERMINATE and sends connect_err() and set
the cm_node state to CLOSED. The passive side sends FIN + ACK after
TERMINATE. Active side ends up in handle_ack_pkt() and send_reset().
send_reset() consumes 1 cm_node's ref_count. Because the cm_node is
in CLOSED state, which means that cm_node will be destroyed after
completion of the connect_err() indication, CM will crash after
send_reset().
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When the listener is destroyed for a loopback connection, the listener
node gets a reset event. This causes a crash as the listener is not
expecting a reset event. Code review of cm_event_reset() during
debugging showed the cm_id ref count is incremented after calling its
event handler and not before.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
While running IMP_EXT's window test, we saw a crash in nes_accept().
Here is the sequence of what happened:
(1) In MVAPICH2, connect request is received for port #0.
FIX: Add a nes_connect() check to make sure local or remote tcp port
is not 0.
(2) Remote node's (passive) TCP stack sends a reset when it gets a
connect request because of port = 0. Active side set the connect
error to IW_CM_EVENT_STATUS_REJECTED when it received the RST from
remote node.
FIX: The corect error code is -ECONNRESET.
(3) Wrong error code of IW_CM_EVENT_STATUS_REJECTED causes the core to
destroy its listener ports. Here there are connections that may
have sent an MPA request up and waiting for accept or reject. But
the listener and its cm_nodes have been freed already causing the
crash noticed.
FIX: The cm_node is freed only if its state is not
NES_CM_STATE_MPAREQ_RCVD. If cm_node's state is
NES_CM_STATE_MPAREQ_RCVD then its new state is set to
NES_CM_STATE_LISTENER_DESTROYED and it is not freed. When
nes_accept() or nes_reject() is received, its state is checked
for NES_CM_STATE_LISTENER_DESTROYED and in this case the cm_node
is freed and error is returned.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
During testing of REJECT connection error handling, we saw that the
cm_id resources are not released. When the retransmit timer expires,
we need to send a reset message to remote node before issuing the
ABORTED event.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
During Xansation testing, we saw that error handling of MPA frame
msg/response is not handled properly.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The ORD size needs updating as we are supporting more inbound READ
resources per connection.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Change MAX_CM_BUFFER for MPA frames to be conformant to RFC 5044:
we need 512 + 20 instead of 512.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The size argument to ioremap_nocache should be the size of desired
information, not the pointer to it.
The semantic match that finds this problem is as follows:
(http://coccinelle.lip6.fr/)
// <smpl>
@expression@
expression *x;
@@
x =
<+...
*sizeof(x)
...+>// </smpl>
Signed-off-by: Julia Lawall <julia@diku.dk>
Acked-by: Chien Tung <chien.tin.tung@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Update copyright from Intel-NE, Inc. to Intel Corporation. Use proper
branding string in Kconfig and simplify description.
Signed-off-by: Chien Tung <chien.tin.tung@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add a check to nes_create_cq() to return -EINVAL if creating a CQ with
depth > max_cqe (32766).
Signed-off-by: Chien Tung <chien.tin.tung@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add IB_SINGAL_ALL_WR support as an iWARP extension. If set, make sure
all WR for the QP are signalled. Consolidate flags used in nesqp
structure.
Signed-off-by: Chien Tung <chien.tin.tung@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add additional PHY uC status check in case PHY firmware is not running
properly with heartbeat. Add a hard PHY reset if uC status is 0x0
after initial reset.
Signed-off-by: Chien Tung <chien.tin.tung@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Always set bad_wr when an immediate error is detected. Do not report
success if an error occurred.
Signed-off-by: Frank Zago <fzago@systemfabricworks.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Always set bad_wr when an immediate error is detected.
Signed-off-by: Frank Zago <fzago@systemfabricworks.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add support for IB_WR_SEND_WITH_INV, IB_WR_RDMA_READ_WITH_INV
and IB_WR_LOCAL_INV.
Signed-off-by: Chien Tung <chien.tin.tung@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
On error, set bad_wr in nes_post_recv(). Stop processing ib_wr queue
when an error is detected.
Signed-off-by: Frank Zago <fzago@systemfabricworks.com>
Signed-off-by: Chien Tung <chien.tin.tung@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
On error, set bad_wr in nes_post_send(). Stop processing ib_wr queue
when an error is detected.
Signed-off-by: Frank Zago <fzago@systemfabricworks.com>
Signed-off-by: Chien Tung <chien.tin.tung@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The ibmebus_free_irq() function, which might sleep, was called with
interrupts disabled. To fix this, make sure that no interrupts are
running by killing the interrupt tasklet. Also lock the
shca_list_lock to protect against the poll_eqs_timer running
concurrently.
Signed-off-by: Alexander Schmidt <alexs@linux.vnet.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Use bitmap_weight() instead of finding all set bits in bitmap by hand.
Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Cc: Ralph Campbell <infinipath@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
IPoIB can miss a change in destination GID under some conditions. The
problem is caused when ipoib_neigh->dgid contains a stale address.
The fix is to set ipoib_neigh->dgid to zero in ipoib_neigh_alloc().
This can happen when a system using bonding on its IPoIB interfaces
has switched its active interface from interface A to B and back to A.
The system that fails over will not correctly processes the 2nd
address change, as described below.
When an address has changed neighbor->ha is updated with the new
address. Each neighbor has an associated ipoib_neigh.
ipoib_neigh->dgid also holds a copy of the remote node's hardware
address. When an address changes neighbor->ha is updated by the
network layer (arp code) with the new address. IPoIB detects this
change in ipoib_start_xmit() by comparing neighbor->ha with
ipoib_neigh->dgid. The bug is that ipoib_neigh->dgid may already
contain the new address (A) thus the change from B to A is missed by
ipoib. Here is the sequence of events:
ipoib_neigh->dgid = A and neighbor->ha = A
The address is switched to B (the first switch)
neighbor->ha = B
The change is seen in ipoib_start_xmit() -- neighbor->ha !=
ipoib_neigh->dgid so ipoib_neigh is released, and a new one is
allocated.
The allocator may return the same chunk of memory that was just
released, therefore ipoib_neigh->dgid still contains A at this point.
ipoib_neigh->dgid should be updated in neigh_add_path(), but if the
following conditions are true dgid is not updated:
1) __path_find() returns a path
2) path->ah is NULL
The remote system now switches from address B to A, neighbor->ha is
updated to A.
Now we have again : ipoib_neigh->dgid = A and neighbor->ha = A
Since the addresses are the same ipoib won't process the change in
address. Fix this by zeroing out the dgid field when allocating a new
struct ipoib_neigh.
Signed-off-by: David Wilder <dwilder@us.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
