]> www.infradead.org Git - users/jedix/linux-maple.git/log
users/jedix/linux-maple.git
9 years agoRDS: IB: Make use of ARPOP_REQUEST instead of ARPOP_REPLY in bonding code
Santosh Shilimkar [Mon, 18 Apr 2016 14:09:39 +0000 (07:09 -0700)]
RDS: IB: Make use of ARPOP_REQUEST instead of ARPOP_REPLY in bonding code

Even though IPv4 ARP RFC allows for using either REQUEST or REPLY
for grat. arp, upstream code from 3.14 onwards have moved on to
use only REQUEST.

Relevant commit:
http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=56022a8fdd874c56bb61d8c82559e43044d1aa06

RDS active bonding gratuitous arp code needs to adopt to this change
to take advantage of the neighbor updates on UEK4. The current code
makes use  ARPOP_REPLY which needs to be changed to ARPOP_REQUEST.

Orabug: 23094704

Tested-by: Michael Nowak <michael.nowak@oracle.com>
Tested-by: Rose Wang <rose.wang@oracle.com>
Tested-by: Rafael Alejandro Peralez <rafael.peralez@oracle.com>
Acked-by: Mukesh Kacker <mukesh.kacker@oracle.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Reported-by: Mukesh Kacker <mukesh.kacker@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: IB: don't use the link-local address for ib transport
Santosh Shilimkar [Fri, 29 Apr 2016 21:50:55 +0000 (14:50 -0700)]
RDS: IB: don't use the link-local address for ib transport

Link-local address can't be used for IB failover and don't work
with IB stack. Even though the DB RDS usage has recommnded to not
use these addresses, we keep hitting issue because of accidental
usage of it because of missing application config or admin scripts
blindly doing rds-ping for each local address(s).

RDS TCP which doesn't support acitive active, there might be an
usecase so the current fix it limited for IB transport atm.

Example traces:
$ rds-ping -I 169.254.221.37 169.254.221.38
bind() failed, errno: 99 (Cannot assign requested address)

cosnole:
 RDS/IB: Link local address 169.254.221.37 NOT SUPPORTED
 RDS: rds_bind() could not find a transport for 169.254.221.37, load rds_tcp or rds_rdma?

Orabug: 23027670

Tested-by: Michael Nowak <michael.nowak@oracle.com>
Tested-by: Rose Wang <rose.wang@oracle.com>
Tested-by: Rafael Alejandro Peralez <rafael.peralez@oracle.com>
Acked-by: Mukesh Kacker <mukesh.kacker@oracle.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Reviewed-by: Wengang Wang <wen.gang.wang@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: IB: rebuild receive caches when needed
Santosh Shilimkar [Wed, 18 May 2016 17:44:56 +0000 (10:44 -0700)]
RDS: IB: rebuild receive caches when needed

RDS IB caches code leaks memory & it have been there from the
inception of cache code but we didn't noticed them since caches
are not teardown in normal operation paths. But now to support
features like variable fragment or connection destroy for ACL,
caches needs to be destroyed and rebuild if needed.

While freeing the caches is just fine, leaking memory while
doing that is bug and needs to be addressed. Thanks to Wengang
for spotting this stone age leak. Also the cache rebuild needs
to be done only when desired so patch optimises that part as
well.

Tested-by: Michael Nowak <michael.nowak@oracle.com>
Tested-by: Maria Rodriguez <maria.r.rodriguez@oracle.com>
Tested-by: Hong Liu <hong.x.liu@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agordma_cm: use cma_info() instead of cma_dbg()
Ajaykumar Hotchandani [Tue, 10 May 2016 23:58:02 +0000 (16:58 -0700)]
rdma_cm: use cma_info() instead of cma_dbg()

We want to have selected prints going into messages file when debug
is enabled.

Orabug: 22381123

Signed-off-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Wengang Wang <wen.gang.wang@oracle.com>
9 years agoOFED: indicate consistent vendor error
Ajaykumar Hotchandani [Tue, 10 May 2016 22:43:48 +0000 (15:43 -0700)]
OFED: indicate consistent vendor error

vendor error print should be consistent across protocols to avoid
any confusion.
Currently, it's decimal at some places and hex at some places.
This patch corrects that.

Orabug: 22381117

Suggested-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Acked-by: Wengang Wang <wen.gang.wang@oracle.com>
9 years agoRDS: Change number based conn-drop reasons to enum
Avinash Repaka [Tue, 17 May 2016 21:42:19 +0000 (14:42 -0700)]
RDS: Change number based conn-drop reasons to enum

This patch converts the number based connection-drop reasons to enums,
making it easy to grep the reasons and to develop new patches based on
these reasons.

Orabug: 23294707

Signed-off-by: Avinash Repaka <avinash.repaka@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: Move rds_rtd definitions from rds_rt_debug files to common files
Avinash Repaka [Wed, 18 May 2016 22:09:05 +0000 (15:09 -0700)]
RDS: Move rds_rtd definitions from rds_rt_debug files to common files

This patch moves rds_rtd definitions from rds_rtd_debug.h to rds.h and
rds_rt_debug_bitmap modparam definition from rds_rt_debug.c to af_rds.c.
The patch removes rds_rt_debug files since there isn't much content
in these files to be held separately.

Commit 'ib/rds: runtime debuggability enhancement' originally defined
rds_rtd definitions.

Orabug: 23294707

Signed-off-by: Avinash Repaka <avinash.repaka@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: Change the default value of rds_rt_debug_bitmap modparam to 0x488B
Avinash Repaka [Wed, 13 Apr 2016 00:48:42 +0000 (17:48 -0700)]
RDS: Change the default value of rds_rt_debug_bitmap modparam to 0x488B

This patch changes the default value of rds_rt_debug_bitmap module
parameter to 0x488B to enable RDS_RTD_ERR, RDS_RTD_ERR_EXT, RDS_RTD_CM,
RDS_RTD_ACT_BND, RDS_RTD_RCV, RDS_RTD_SND flags of rds_rtd.

Orabug: 23294707

Signed-off-by: Srinivas Eeda <srinivas.eeda@oracle.com>
Signed-off-by: Avinash Repaka <avinash.repaka@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: Replace rds_rtd printk with trace_printk
Avinash Repaka [Fri, 25 Mar 2016 19:58:37 +0000 (12:58 -0700)]
RDS: Replace rds_rtd printk with trace_printk

Forward rds_rtd prints to ftrace buffer by replacing
printk with trace_printk.

Orabug: 23294707

Signed-off-by: Srinivas Eeda <srinivas.eeda@oracle.com>
Signed-off-by: Avinash Repaka <avinash.repaka@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: IB: Print vendor error in recv completion error message
Avinash Repaka [Fri, 11 Mar 2016 22:09:08 +0000 (14:09 -0800)]
RDS: IB: Print vendor error in recv completion error message

This patch when applied, prints vendor error along with work
completion status in recv completion error message.

Orabug: 23294707

Signed-off-by: Avinash Repaka <avinash.repaka@oracle.com>
Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoIB/mlx4: Fix unaligned access in send_reply_to_slave
shamir rabinovitch [Wed, 18 May 2016 10:18:10 +0000 (06:18 -0400)]
IB/mlx4: Fix unaligned access in send_reply_to_slave

The problem is that the function 'send_reply_to_slave' gets the
'req_sa_mad' as a pointer whose address is only aliged to 4 bytes
but is 8 bytes in size.  This can result in unaligned access faults
on certain architectures.

Sowmini Varadhan pointed to this reply from Dave Miller that say
that memcpy should not be used to solve alignment issues:
https://lkml.org/lkml/2015/10/21/352

Optimization of memcpy to 'ldx' instruction can only happen if the
compiler knows that the size of the data we are copying is 8 bytes
and it assumes it is aligned to 8 bytes. If the compiler know the
type is not aligned to 8 it must not optimize the 8 byte copy.
Defining the data type as aligned to 4 forces the compiler to treat
all accesses as though they aren't aligned and avoids the 'ldx'
optimization.

Full credit for the idea goes to Jason Gunthorpe
<jgunthorpe@obsidianresearch.com>.

Orabug: 23311415

Signed-off-by: Shamir Rabinovitch <shamir.rabinovitch@oracle.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
9 years agords: schedule local connection activity in proper workqueue
Ajaykumar Hotchandani [Mon, 18 Apr 2016 22:59:26 +0000 (15:59 -0700)]
rds: schedule local connection activity in proper workqueue

While reconnect, local connection is scheduled on rds_wq; while it it
should have been scheduled rds_local_wq.
This patch corrects that.

Orabug: 23223537

Tested-by: Michael Nowak <michael.nowak@oracle.com>
Signed-off-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoIB/security: Restrict use of the write() interface
Jason Gunthorpe [Mon, 11 Apr 2016 01:13:13 +0000 (19:13 -0600)]
IB/security: Restrict use of the write() interface

The drivers/infiniband stack uses write() as a replacement for
bi-directional ioctl().  This is not safe. There are ways to
trigger write calls that result in the return structure that
is normally written to user space being shunted off to user
specified kernel memory instead.

For the immediate repair, detect and deny suspicious accesses to
the write API.

For long term, update the user space libraries and the kernel API
to something that doesn't present the same security vulnerabilities
(likely a structured ioctl() interface).

The impacted uAPI interfaces are generally only available if
hardware from drivers/infiniband is installed in the system.

Reported-by: Jann Horn <jann@thejh.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
[ Expanded check to all known write() entry points ]
Cc: stable@vger.kernel.org
Signed-off-by: Doug Ledford <dledford@redhat.com>
CVE-2016-4565
Orabug: 23276449

(cherry-pick from e6bd18f57aad1a2d1ef40e646d03ed0f2515c9e3)
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agonet/rds: Use max_mr from HCA caps than max_fmr
Yuval Shaia [Thu, 5 May 2016 07:47:48 +0000 (00:47 -0700)]
net/rds: Use max_mr from HCA caps than max_fmr

All HCA drivers seems to populate max_mr caps and few of them do both
max_mr and max_fmr.
Hence update RDS code to make use of max_mr.

Orabug: 23223564

Signed-off-by: Santosh Shilimkar <ssantosh@kernel.org>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Yuval Shaia <yuval.shaia@oracle.com>
9 years agoRDS: IB: disable ib_cache purging to avoid memory leak in reconnect path
Santosh Shilimkar [Tue, 17 May 2016 21:46:18 +0000 (14:46 -0700)]
RDS: IB: disable ib_cache purging to avoid memory leak in reconnect path

RDS IB caches don't work in reconnect path and if used can lead to
memory leaks. These leaks have been there for long time but we didn't
hit them since caches are not teardown in reconnect path. For different
frag rolling upgrade/downgrade support, its needed to work in reconnect
path but needs additional fixes.

Since the leak is blocking rest of the testing, temporary the cache
purging is disabled. It will be added back once fully fixed.

Orabug: 23275911

The change doesn't impact any of the existing RDS functionality.

Tested-by: Hong Liu <hong.x.liu@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: IB: avoid bit fields for i_frag_pages
Wengang Wang [Tue, 17 May 2016 16:21:20 +0000 (09:21 -0700)]
RDS: IB: avoid bit fields for i_frag_pages

i_frag_pages may need to store more than 1 page value for
higher fragments so bit field won't help.

Lets fix that.

Orabug: 23275911

Tested-by: Hong Liu <hong.x.liu@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
9 years agoRDS: TCP: Synchronize accept() and connect() paths on t_conn_lock.
Sowmini Varadhan [Tue, 3 May 2016 20:14:42 +0000 (13:14 -0700)]
RDS: TCP: Synchronize accept() and connect() paths on t_conn_lock.

Orabug 23228077

Backport of upstream commit bd7c5f983f31 ("RDS: TCP: Synchronize accept()
and connect() paths on t_conn_lock.")

An arbitration scheme for duelling SYNs is implemented as part of
commit 241b271952eb ("RDS-TCP: Reset tcp callbacks if re-using an
outgoing socket in rds_tcp_accept_one()") which ensures that both nodes
involved will arrive at the same arbitration decision. However, this
needs to be synchronized with an outgoing SYN to be generated by
rds_tcp_conn_connect(). This commit achieves the synchronization
through the t_conn_lock mutex in struct rds_tcp_connection.

The rds_conn_state is checked in rds_tcp_conn_connect() after acquiring
the t_conn_lock mutex.  A SYN is sent out only if the RDS connection is
not already UP (an UP would indicate that rds_tcp_accept_one() has
completed 3WH, so no SYN needs to be generated).

Similarly, the rds_conn_state is checked in rds_tcp_accept_one() after
acquiring the t_conn_lock mutex. The only acceptable states (to
allow continuation of the arbitration logic) are UP (i.e., outgoing SYN
was SYN-ACKed by peer after it sent us the SYN) or CONNECTING (we sent
outgoing SYN before we saw incoming SYN).

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
9 years agoRDS:TCP: Synchronize rds_tcp_accept_one with rds_send_xmit when resetting t_sock
Sowmini Varadhan [Tue, 3 May 2016 18:55:08 +0000 (11:55 -0700)]
RDS:TCP: Synchronize rds_tcp_accept_one with rds_send_xmit when resetting t_sock

Orabug 23228077

Backport of upstream commit eb192840266f ("RDS:TCP: Synchronize
rds_tcp_accept_one with rds_send_xmit when resetting t_sock")

There is a race condition between rds_send_xmit -> rds_tcp_xmit
and the code that deals with resolution of duelling syns added
by commit 241b271952eb ("RDS-TCP: Reset tcp callbacks if re-using an
outgoing socket in rds_tcp_accept_one()").

Specifically, we may end up derefencing a null pointer in rds_send_xmit
if we have the interleaving sequence:
         rds_tcp_accept_one                  rds_send_xmit

                                           conn is RDS_CONN_UP, so
      invoke rds_tcp_xmit

                                           tc = conn->c_transport_data
      rds_tcp_restore_callbacks
          /* reset t_sock */
      null ptr deref from tc->t_sock

The race condition can be avoided without adding the overhead of
additional locking in the xmit path: have rds_tcp_accept_one wait
for rds_tcp_xmit threads to complete before resetting callbacks.
The synchronization can be done in the same manner as rds_conn_shutdown().
First set the rds_conn_state to something other than RDS_CONN_UP
(so that new threads cannot get into rds_tcp_xmit()), then wait for
RDS_IN_XMIT to be cleared in the conn->c_flags indicating that any
threads in rds_tcp_xmit are done.

Fixes: 241b271952eb ("RDS-TCP: Reset tcp callbacks if re-using an
outgoing socket in rds_tcp_accept_one()")
Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
9 years agoskbuff: Add pskb_extract() helper function
Sowmini Varadhan [Tue, 26 Apr 2016 10:51:46 +0000 (03:51 -0700)]
skbuff: Add pskb_extract() helper function

Orabug 23180876

Upstream commit 6fa01ccd8830 ("skbuff: Add pskb_extract() helper function")

A pattern of skb usage seen in modules such as RDS-TCP is to
extract `to_copy' bytes from the received TCP segment, starting
at some offset `off' into a new skb `clone'. This is done in
the ->data_ready callback, where the clone skb is queued up for rx on
the PF_RDS socket, while the parent TCP segment is returned unchanged
back to the TCP engine.

The existing code uses the sequence
        clone = skb_clone(..);
        pskb_pull(clone, off, ..);
        pskb_trim(clone, to_copy, ..);
with the intention of discarding the first `off' bytes. However,
skb_clone() + pskb_pull() implies pksb_expand_head(), which ends
up doing a redundant memcpy of bytes that will then get discarded
in __pskb_pull_tail().

To avoid this inefficiency, this commit adds pskb_extract() that
creates the clone, and memcpy's only the relevant header/frag/frag_list
to the start of `clone'. pskb_trim() is then invoked to trim clone
down to the requested to_copy bytes.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
9 years agoRDS: support individual receive trace reporting
Santosh Shilimkar [Thu, 21 Apr 2016 23:18:47 +0000 (16:18 -0700)]
RDS: support individual receive trace reporting

If application wants to get indvidual trace point, its easy
to support with existing infrastructure.

No change needed in API

Orabug: 23215779

Tested-by: Namrata Jampani <namrata.jampani@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoIB/ipoib: Add readout of statistics using ethtool
Hans Westgaard Ry [Wed, 13 Apr 2016 13:23:20 +0000 (15:23 +0200)]
IB/ipoib: Add readout of statistics using ethtool

Orabug: 21498734

IPoIB collects statistics of traffic including number of packets
sent/received, number of bytes transferred, and certain errors. This
patch makes these statistics available to be queried by ethtool.

Change-Id: Ic159815fe0cc08770cd4111ec1df117b7349c154
Signed-off-by: Hans Westgaard Ry <hans.westgaard.ry@oracle.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Tested-by: Yuval Shaia <yuval.shaia@oracle.com>
Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
9 years agoIB/ipoib: Add handling for sending of skb with many frags
Hans Westgaard Ry [Wed, 16 Mar 2016 13:01:02 +0000 (14:01 +0100)]
IB/ipoib: Add handling for sending of skb with many frags

Orabug: 21498734

IPoIB puts skb-fragments in SGEs adding 1 extra SGE when SG is enabled.
Current codepath assumes that the max number of SGEs a device supports
is at least MAX_SKB_FRAGS+1, there is no interaction with upper layers
to limit number of fragments in an skb if a device suports fewer SGEs.
The assumptions also lead to requesting a fixed number of SGEs when
IPoIB creates queue-pairs with SG enabled.

A fallback/slowpath is implemented using skb_linearize to
handle cases where the conversion would result in more sges than supported.

Change-Id: Ia81e69d7231987208ac298300fc5b9734f193a2d
Signed-off-by: Hans Westgaard Ry <hans.westgaard.ry@oracle.com>
Reviewed-by: Håkon Bugge <haakon.bugge@oracle.com>
Reviewed-by: Wei Lin Guay <wei.lin.guay@oracle.com>
9 years agoRevert "RDS: Make message size limit compliant with spec"
Chuck Anderson [Wed, 4 May 2016 17:18:17 +0000 (10:18 -0700)]
Revert "RDS: Make message size limit compliant with spec"

This reverts commit ef278157f938011ee08fb683311e5e31ffd29fdf.

Orabug: 23217242

From Avinash Repaka <avinash.repaka@oracle.com>:
This patch reverts the fix for Orabug: 22661521, since the fix assumes
that the memory region is always aligned on a page boundary, causing an
EMSGSIZE error when trying to register 1MB region that isn't 4KB aligned.
These issues were observed on kernel 4.1.12-39.el6uek tag.

Signed-off-by: Avinash Repaka <avinash.repaka@oracle.com>
Acked-by: Chuck Anderson <chuck.anderson@oracle.com>
9 years agoRDS: TCP: Remove unused constant
Sowmini Varadhan [Thu, 24 Mar 2016 19:13:57 +0000 (12:13 -0700)]
RDS: TCP: Remove unused constant

Orabug: 22993275

Upstream commit a3382e408b64 ("RDS: TCP: Remove unused constant")

RDS_TCP_DEFAULT_BUFSIZE has been unused since commit 1edd6a14d24f
("RDS-TCP: Do not bloat sndbuf/rcvbuf in rds_tcp_tune").

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: TCP: Add sysctl tunables for sndbuf/rcvbuf on rds-tcp socket
Sowmini Varadhan [Wed, 23 Mar 2016 23:37:26 +0000 (16:37 -0700)]
RDS: TCP: Add sysctl tunables for sndbuf/rcvbuf on rds-tcp socket

Orabug: 22993275

Upstream commit c6a58ffed536 ("RDS: TCP: Add sysctl tunables for
sndbuf/rcvbuf on rds-tcp socket")

Add per-net sysctl tunables to set the size of sndbuf and
rcvbuf on the kernel tcp socket.

The tunables are added at /proc/sys/net/rds/tcp/rds_tcp_sndbuf
and /proc/sys/net/rds/tcp/rds_tcp_rcvbuf.

These values must be set before accept() or connect(),
and there may be an arbitrary number of existing rds-tcp
sockets when the tunable is modified. To make sure that all
connections in the netns pick up the same value for the tunable,
we reset existing rds-tcp connections in the netns, so that
they can reconnect with the new parameters.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Reviewed-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: Make message size limit compliant with spec
Avinash Repaka [Mon, 29 Feb 2016 23:30:57 +0000 (15:30 -0800)]
RDS: Make message size limit compliant with spec

This patch limits RDS messages(both RDMA & non-RDMA) and RDS MR size
to RDS_MAX_MSG_SIZE(1MB) irrespective of underlying transport layer.

Orabug: 22661521

Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Avinash Repaka <avinash.repaka@oracle.com>
Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
9 years agoRDS: add flow control info to rds_info_rdma_connection
Wei Lin Guay [Wed, 27 Jan 2016 12:18:08 +0000 (13:18 +0100)]
RDS: add flow control info to rds_info_rdma_connection

Added per connection flow_ctl_post_credit and
flow_ctl_send_credit to rds-info. These info help
in debugging RDS IB flow control. The newly added
attributes are placed at the bottom of the data
structure to ensure backward compatibility.

Orabug: 22306628

Reviewed-by: Håkon Bugge <haakon.bugge@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Wei Lin Guay <wei.lin.guay@oracle.com>
9 years agoRDS: update IB flow control algorithm
Wei Lin Guay [Thu, 17 Dec 2015 08:34:33 +0000 (09:34 +0100)]
RDS: update IB flow control algorithm

The current algorithm that uses 16 as a hard-coded value
in rds_ib_advertise_credits() doesn't serve the purpose, as
post_recvs() are performed in bulk. Thus, the test
condition will always be true.

This patch moves rds_ib_advertise_credits() in to the
post_recvs() loop. Instead of updating the post_recv credits
after all the post_recvs() have completed, the post_recv
credit is being updated in log2 incremental manner.
The proposed exponential quadrupling algorithm serves as a
good compromise between early start of the peer and at the
same time reducing the amount of explicit ACKs. The credit
update explicit ACKs will be generated starting from 16,
256, 4096...etc.

The performance number below shows that this new flow
control algorithm has minimal impact performance even though
it requires additional explicit ACKs.

4 QPs, 32t, 16d -q 8448 -a 256 (rds-parameter)

HCAs flow_ctl no flow_ctl
Mellanox CX3 744K 742K
Oracle QDR M4 819K 831K

Orabug: 22306628

Suggested-by: Håkon Bugge <haakon.bugge@oracle.com>
Reviewed-by: Håkon Bugge <haakon.bugge@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Wei Lin Guay <wei.lin.guay@oracle.com>
9 years agoRDS: Add flow control in runtime debugging
Wei Lin Guay [Fri, 4 Dec 2015 15:50:13 +0000 (16:50 +0100)]
RDS: Add flow control in runtime debugging

Add RDS_RTD_FLOW_CNTRL feature flag in
rds_rt_debug_bitmap.

Orabug: 22306628

Reviewed-by: Håkon Bugge <haakon.bugge@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Wei Lin Guay <wei.lin.guay@oracle.com>
9 years agoRDS: fix IB transport flow control
Wei Lin Guay [Thu, 3 Dec 2015 08:23:59 +0000 (09:23 +0100)]
RDS: fix IB transport flow control

Orabug: 22306628

IB flow control is always disabled regardless of
rds_ib_sysctl_flow_control flag.
The issue is that the initial credit advertisement
annouces zero credits, because ib_recv_refill() has
not yet been called. An initial credit offering
of zero effectively disables flow control.

IB flow control is only enabled if both active and
passive connections have set the rds_ib_sysctl
flow_control flag. E.g,

Conn. A (on),  Conn. B (on)  = enable
Conn. A (off), Conn. B (on)  = disable
Conn. A (on),  Conn. B (off) = disable
Conn. A (off), Conn. B (off) = disable

Reviewed-by: Håkon Bugge <haakon.bugge@oracle.com>
Reviewed-by: Knut Omang <knut.omang@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Signed-off-by: Wei Lin Guay <wei.lin.guay@oracle.com>
9 years agomlx4_core: scale_profile should work without params set to 0
Mukesh Kacker [Wed, 13 Apr 2016 01:53:37 +0000 (18:53 -0700)]
mlx4_core: scale_profile should work without params set to 0

The "scale_profile" parameter is to be used to do scaling of
hca params without requiring specific tuning setting for each
one of them individually and yet allowing manual setting of
variables.

In UEK2, the module params were default zero and initialized
later from a "driver default" or a scale_profile dictated value.

In UEK4 (derived from Mellanox OFED 2.4) module params were
pre-initialized to default values and are not zero and have
to be forced to 0 for dynamic scaling to be activated.
This defeats the purpose of having a single parameter to achieve
scaling and not requiring setting individual parameters (while
retaining ability to revert to driver defaults).

The changes here (re)introduce a separate static instance
containing default parameters separate from module parameters
which are pre-initialized to zero for parameters that can scale
dynamically and to default values for others. The zero module
parameters are later initialized to either a scale_profile
governed value or driver defaults.

Orabug: 23078816

Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
Reviewed-by: Wengang Wang <wen.gang.wang@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agomlx4_core: bump default scaled value of num of cqs and srqs
Mukesh Kacker [Tue, 29 Mar 2016 02:12:50 +0000 (19:12 -0700)]
mlx4_core: bump default scaled value of num of cqs and srqs

Bump up the number of cqs and srqs when scale_profile parameter
is set or log_num_cq/log_num_srq are zero implying a scaled
default profile.

This allows higher scalability needed for cloud based platforms
running higher number of processes per hca.

If the individual parameters log_num_srq and/or log_num_cq are
set, they take precedence and scale_profile is ignored.

(Fixes for style problems with quoted strings split across
lines and and block comment format are also piggybacked
in this commit!)

Orabug: 23078966

Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
Reviewed-by: Wengang Wang <wen.gang.wang@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years ago[PATCH 2/2] Avoid redundant call to rds_bind_lookup() in recv path.
Sowmini Varadhan [Tue, 12 Apr 2016 01:49:12 +0000 (18:49 -0700)]
[PATCH 2/2] Avoid redundant call to rds_bind_lookup() in recv path.

Orabug 20930687

When RAC tries to scale RDS-TCP, they are hitting bottlenecks
due to inefficiencies in rds_bind_lookup. Each call to
rds_bind_lookup results in an irqsave/irqrestore sequence, and
when the list of RDS sockets is large, we end up having IRQs
suppressed for long intervals. This trigger flow-control assertions
and causes TX queue watchdog hangs in the sender. The current
implementation makes this even worse, by superfluously calling
rds_bind_lookup(). This patch set takes the first step to solving
this problem by avoiding one of the redundant calls to rds_bind_lookup.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: TOS fixes in failure paths when RDS-TCP and RDS-RDMA are run together
Sowmini Varadhan [Tue, 12 Apr 2016 00:17:37 +0000 (17:17 -0700)]
RDS: TOS fixes in failure paths when RDS-TCP and RDS-RDMA are run together

Orabug 20930687

When errors such as connection hangs or failures are encountered
over RDS-TCP,  the sending RDS, in an attempt at HA, will try to
reconnect, and trip up on all sorts of data structures intended
for ToS support. The ToS feature is currently only supported for
RDS-IB, and  unplanned/untested usage of these data
structures by RDS-TCP causes deadlocks and panics.

Until we properly design, support, and test the ToS feature for
RDS-TCP, such paths should not be wandered into. Thus this patchset
adds defensive checks to ignore rs_tos settings in rds_sendmsg() for
TCP transports, and prevents the sending of ToS heartbeat pings

Until we properly design, support, and test the ToS feature for
RDS-TCP, such paths should not be wandered into. Thus this patchset
adds defensive checks to ignore rs_tos settings in rds_sendmsg() for
TCP transports, and prevents the sending of ToS heartbeat pings
in rds_send_hb() for TCP transport.

For reference, the deadlock that can be encountered in the
hb ping path is:

 [<ffffffff814846f4>] do_tcp_setsockopt+0x244/0x710  <-- wants sock_lock
 [<ffffffff81119ead>] ? __alloc_pages_nodemask+0x12d/0x230
 [<ffffffff8125cf4d>] ? rb_insert_color+0x9d/0x160
 [<ffffffff8101cd43>] ? native_sched_clock+0x13/0x80
 [<ffffffff8101c369>] ? sched_clock+0x9/0x10
 [<ffffffff810e25a9>] ? trace_clock_local+0x9/0x10
 [<ffffffff810e8d97>] ? rb_reserve_next_event+0x67/0x480
 [<ffffffff81292a1f>] ? __alloc_and_insert_iova_range+0x17f/0x1f0
 [<ffffffff81117f52>] ? get_page_from_freelist+0x1e2/0x550
 [<ffffffff81484c1a>] tcp_setsockopt+0x2a/0x30
 [<ffffffff8142e724>] sock_common_setsockopt+0x14/0x20
 [<ffffffffa00de5cd>] rds_tcp_xmit_prepare+0x5d/0x70 [rds_tcp]
 [<ffffffffa0749b35>] rds_send_xmit+0xe5/0x860 [rds]
 [<ffffffffa074a37d>] rds_send_hb+0xcd/0x130 [rds]
 [<ffffffffa0747a1b>] rds_recv_local+0x20b/0x330 [rds]
 [<ffffffffa074851d>] rds_recv_incoming+0x7d/0x290 [rds]
 [<ffffffff8101cd43>] ? native_sched_clock+0x13/0x80
 [<ffffffff8101c369>] ? sched_clock+0x9/0x10
 [<ffffffffa00de2c6>] rds_tcp_data_recv+0x316/0x440 [rds_tcp]
 [<ffffffff8148642a>] tcp_read_sock+0xda/0x230
 [<ffffffffa00ddfb0>] ? rds_tcp_recv+0x60/0x60 [rds_tcp]
 [<ffffffff81430fb0>] ? sock_get_timestamp+0xc0/0xc0
 [<ffffffffa00dde6d>] rds_tcp_read_sock+0x4d/0x60 [rds_tcp]
 [<ffffffffa00ddefa>] rds_tcp_data_ready+0x7a/0xd0 [rds_tcp]
 [<ffffffff8148df2e>] tcp_data_queue+0x2fe/0xac0
 [<ffffffff814916a9>] tcp_rcv_established+0x349/0x740
 [<ffffffff81499bb5>] tcp_v4_do_rcv+0x125/0x1f0
 [<ffffffff8149b3e7>] tcp_v4_rcv+0x597/0x830   <---- holds sock_lock
 [<ffffffff8148b56e>] ? __tcp_ack_snd_check+0x5e/0xa0
 [<ffffffff814916cd>] ? tcp_rcv_established+0x36d/0x740
 [<ffffffff814774ed>] ip_local_deliver_finish+0xdd/0x2a0
 [<ffffffff81477730>] ip_local_deliver+0x80/0x90
 [<ffffffff81476d75>] ip_rcv_finish+0x105/0x370

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agords: rds-stress show all zeros after few minutes
shamir rabinovitch [Wed, 16 Mar 2016 13:57:19 +0000 (09:57 -0400)]
rds: rds-stress show all zeros after few minutes

Issue can be seen on platforms that use 8K and above page size
while rds fragment size is 4K. On those platforms single page is
shared between 2 or more rds fragments. Each fragment has it's own
offeset and rds cong map code need to take this offset to account.
Not taking this offset to account lead to reading the data fragment
as congestion map fragment and hang of the rds transmit due to far
cong map corruption.

Orabug: 23045970

Reviewed-by: Wengang Wang <wen.gang.wang@oracle.com>
Reviewed-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Tested-by: Anand Bibhuti <anand.bibhuti@oracle.com>
Signed-off-by: shamir rabinovitch <shamir.rabinovitch@oracle.com>
9 years agoRDS: Fix the atomicity for congestion map update
Wengang Wang [Thu, 7 Apr 2016 07:29:43 +0000 (15:29 +0800)]
RDS: Fix the atomicity for congestion map update

Orabug: 23022620 (UEK4)
QuickRef: 22118109 (UEK2)

Two different threads with different rds sockets may be in
rds_recv_rcvbuf_delta() via receive path. If their ports
both map to the same word in the congestion map, then
using non-atomic ops to update it could cause the map to
be incorrect. Lets use atomics to avoid such an issue.

Full credit to Wengang <wen.gang.wang@oracle.com> for
finding the issue, analysing it and also pointing out
to offending code with spin lock based fix.

Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: IB: Run rds_fmr_flush WQ closer to ib_device
Wei Lin Guay [Wed, 18 Nov 2015 11:23:08 +0000 (12:23 +0100)]
RDS: IB: Run rds_fmr_flush WQ closer to ib_device

Orabug 22269408

rds_fmr_flush workqueue is calling ib_unmap_fmr
to invalidate a list of FMRs. Today, this workqueue
can be scheduled at any CPUs. In a NUMA-aware system,
schedule this workqueue to run on a CPU core closer to
ib_device can improve performance. As for now, we use
"sequential-low" policy. This policy selects two lower
cpu cores closer to HCA. In a non-NUMA aware system,
schedule rds_fmr_flush workqueue in a fixed cpu core
improves performance.

The mapping of cpu to the rds_fmr_flush workqueue
can be enabled/disabled via  sysctl and it is enable
by default. To disable the feature, use below sysctl.

rds_ib_sysctl_disable_unmap_fmr_cpu = 1

Putting down some of the rds-stress performance number
comparing default and sequential-low policy in a NUMA
system with Oracle M4 QDR and Mellanox CX3.

rds-stress 4 conns, 32 threads, 16 depths, RDMA write
and unidirectional (higher is better).

(Oracle M4 QDR)
default : 645591 IOPS
sequential-low : 806196 IOPS

(Mellanox CX3)
default : 473836 IOPS
sequential-low : 544187 IOPS

Reviewed-by: Håkon Bugge <haakon.bugge@oracle.com>
Reviewed-by: Knut Omang <knut.omang@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Wei Lin Guay <wei.lin.guay@oracle.com>
9 years agoRDS: IB: support larger frag size up to 16KB
Santosh Shilimkar [Wed, 21 Oct 2015 23:47:28 +0000 (16:47 -0700)]
RDS: IB: support larger frag size up to 16KB

Infiniband (IB) transport supports larger message size
than RDS_FRAG_SIZE, which is usually in 4KB PAGE_SIZE.
Nevertheless, RDS always fragments each payload into
RDS_FRAG_SIZE before hands it over to the underlying
IB HCA.

One of the important message size required for database
is 8448 (8K + 256B control message) for BCOPY. This RDS
message, even with IB transport, will generate three
IB work requests (WR)  with each having its own RDS header.
This series of patches improve RDS performance by allowing
IB transport to send/receive RDS message with a larger
RDS_FRAG_SIZE (Ideally, using a single WR).

In order to maintain the backward compatibility and
interoperability between various RDS versions, and at
the same time to support various FRAG_SIZE, the IB
fragment size is negotiated per connection.
Although IB is capable of supporting 4GB of message size,
currently we limit the IB RDS_FRAG_SIZE up to 16KB due to
two reasons:-
 1. This is needed for current 8448 RDS message size usecase.
 2. Minizing the size for each receive queue entry in order
    to optimal memory usage.

In term of implementation, The 'dp_reserved2' field of
'struct rds_ib_connect_private' now carries information about
supported IB fragment size. Since we are just
using the IB connection private data and a reserved field,
the protocol version is not bumped up. Furthermore, the feature
is enabled only for RDS_PROTOCOL_v4.1 and above (future).

To keep thing simpler for user, a module parameter
'rds_ib_max_frag' is provided. Without module parameter,
the default PAGE_SIZE frag will be used. During the connection
establishment, the smallest fragment size will be
chosen. If the fragment size is 0, it means RDS module
doesn't support large fragment size and the default
RDS_FRAG_SIZE will be used.

Upto ~10+ % improvement seen with Orion and ~9+ % with RDBMS
update queries.

Orabug: 21894138

Reviwed-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: IB: add frag size to per connection info
Santosh Shilimkar [Sat, 21 Nov 2015 22:16:48 +0000 (14:16 -0800)]
RDS: IB: add frag size to per connection info

rds-toools (rds-info) needs tobe updated to display per IB connection
fragment info. Have a patch for it and will get that merged accordingly.

Orabug: 21894138
Reviwed-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: IB: log the endpoint rds connection role
Santosh Shilimkar [Fri, 20 Nov 2015 21:25:40 +0000 (13:25 -0800)]
RDS: IB: log the endpoint rds connection role

Useful to know the active and passive end points in a
RDS conection.

Orabug: 21894138
Reviewed-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: IB: purge receive frag cache on connection shutdown
Santosh Shilimkar [Mon, 21 Mar 2016 06:24:32 +0000 (23:24 -0700)]
RDS: IB: purge receive frag cache on connection shutdown

RDS IB connections can be formed with different fragment size across
reconnect and hence the current frag cache needs to be purged to
avoid stale frag usages.

Orabug: 21894138
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: IB: use i_frag_sz for cache stat updates
Santosh Shilimkar [Mon, 21 Mar 2016 10:49:30 +0000 (03:49 -0700)]
RDS: IB: use i_frag_sz for cache stat updates

Updates the receive cache statistics based on ic->i_frag_sz

Orabug: 21894138
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: IB: scale rds_ib_allocation based on fragment size
Santosh Shilimkar [Wed, 4 Nov 2015 21:42:39 +0000 (13:42 -0800)]
RDS: IB: scale rds_ib_allocation based on fragment size

The 'rds_ib_sysctl_max_recv_allocation' allocation is used to manage
and allocate the size of IB receive queue entry (RQE) for each IB
connection. However, it relies on the hardcoded RDS_FRAG_SIZE.

Lets make it scalable based on supported fragment sizes for different
IB connection. Each connection can allocate different RQE size
depending on the per connection fragment_size.

Orabug: 21894138
Reviewed-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: IB: make fragment size (RDS_FRAG_SIZE) dynamic
Santosh Shilimkar [Wed, 21 Oct 2015 23:47:28 +0000 (16:47 -0700)]
RDS: IB: make fragment size (RDS_FRAG_SIZE) dynamic

IB fabric is capable of fragment 4GB of data payload into
send_first, send_middle and send_last. Nevertheless,
RDS fragments each payload into PAGE_SIZE, which is usually
4KB. This patch makes the RDS_FRAG_SIZE for RDS IB transport
dynamic.

In the preperation for subsequent patch(es), this patch
adds per connection peer negotiation to determine the
supported fragment size for IB transport.

Orabug: 21894138
Reviewed-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Signed-off-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: log the IP address as well on bind failure
Santosh Shilimkar [Wed, 4 Nov 2015 21:42:39 +0000 (13:42 -0800)]
RDS: log the IP address as well on bind failure

It's useful to know the IP address when RDS fails to bind a
connection. Thus, adding it to the error message.

Orabug: 21894138
Reviewed-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: fix the sg allocation based on actual message size
Wei Lin Guay [Fri, 20 Nov 2015 23:14:48 +0000 (15:14 -0800)]
RDS: fix the sg allocation based on actual message size

Fix an issue where only PAGE_SIZE bytes are allocated per
scatter-gather entry (SGE) regardless of the actual message
size: Furthermore, use buddy memory allocation technique to
allocate/free memmory (if possible) to reduce SGE.

Orabug: 21894138
Signed-off-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: make congestion code independent of PAGE_SIZE
Santosh Shilimkar [Mon, 16 Nov 2015 21:28:11 +0000 (13:28 -0800)]
RDS: make congestion code independent of PAGE_SIZE

RDS congestion map code is designed with base assumption of
4K page size. The map update as well transport code assumes
it that way. Ofcourse it breaks when transport like IB starts
supporting larger fragments than 4K.

To overcome this limitation without too many changes to the core
congestion map update logic, define indepedent RDS_CONG_PAGE_SIZE
and use it.

While at it we also move rds_message_map_pages() whose sole
purpose it to map congestion pages to congestion code.

Orabug: 21894138
Reviwed-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: Back out OoO send status fix since it causes the regression
Santosh Shilimkar [Tue, 16 Feb 2016 18:24:31 +0000 (10:24 -0800)]
RDS: Back out OoO send status fix since it causes the regression

With the DB build, the crash was observed which was boiled down to
this change on UEK2. Proactively we back this out on UEK4 as well
till the issue gets addresssed.

<0>------------[ cut here ]------------
<2>kernel BUG at net/rds/send.c:511!
<0>invalid opcode: 0000 [#1] SMP
<4>CPU 197
<4>Modules linked in: oracleacfs(P)(U) oracleadvm(P)(U) oracleoks(P)(U)
ipmi_poweroff ipmi_devintf ipmi_si ipmi_msghandler bonding acpi_cpufreq
freq_table mperf rds_rdma rds ib_sdp ib_ipoib rdma_ucm ib_ucm ib_uverbs
ib_umad rdma_cm ib_cm iw_cm ib_addr ipv6 mlx4_ib ib_sa ib_mad ib_core
mlx4_core ext3 jbd fuse ghes hed wmi i2c_i801 iTCO_wdt
iTCO_vendor_support
igb i2c_algo_bit i2c_core ixgbe hwmon dca sg ext4 mbcache jbd2 sd_mod
crc_t10dif megaraid_sas dm_mirror dm_region_hash dm_log dm_mod [last
unloaded: scsi_wait_scan]
<4>
<4>Pid: 85252, comm: oracle_85252_tp Tainted: P

[...]

<0>Call Trace:
<4> [<ffffffffa0340cab>] rds_send_drop_to+0xcb/0x470 [rds]
<4> [<ffffffffa033a19e>] rds_release+0x8e/0x110 [rds]
<4> [<ffffffff8142a369>] sock_release+0x29/0x90
<4> [<ffffffff8142a3e7>] sock_close+0x17/0x30
<4> [<ffffffff8116f2fe>] __fput+0xbe/0x240
<4> [<ffffffff8116f4a5>] fput+0x25/0x30
<4> [<ffffffff8116b3e3>] filp_close+0x63/0x90
<4> [<ffffffff8116b4c7>] sys_close+0xb7/0x120
<4> [<ffffffff81516762>] system_call_fastpath+0x16/0x1b
<0>Code: 00 75 1f 65 8b 14 25 58 c3 00 00 48 63 d2 48 c7 c0 80 26 01 00
48 03
04 d5 40 0e 99 81 48 83 40 68 01 5b 41 5c c9 c3 0f 0b eb fe <0f> 0b eb
fe 0f
1f 40 00 55 48 89 e5 48 83 ec 40 48 89 5d d8 4c
<1>RIP  [<ffffffffa033e5e8>] rds_send_sndbuf_remove+0x68/0x70 [rds]
<4> RSP <ffff8825b0577dd8>

Revert "RDS: Fix out-of-order RDS_CMSG_RDMA_SEND_STATUS"

This reverts commit 5631f1a303104d41f6ded0d603011d6c172b8644.

Orabug: 21894138
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agonet/mlx4_core: Modify default value of log_rdmarc_per_qp to be consistent with HW...
Yuval Shaia [Thu, 17 Sep 2015 09:52:44 +0000 (02:52 -0700)]
net/mlx4_core: Modify default value of log_rdmarc_per_qp to be consistent with HW capability

This value is used to calculate max_qp_dest_rdma.
Default value of 4 brings us to 16 while HW supports 128
(max_requester_per_qp)
Although this value can be changed by module param it is best that default
will be optimized

Orabag: 19883194

Signed-off-by: Yuval Shaia <yuval.shaia@oracle.com>
Reviewed-by: Wengang Wang <wen.gang.wang@oracle.com>
9 years agoFixed vnic issue after saturn reset
Pradeep Gopanapalli [Tue, 1 Mar 2016 01:45:59 +0000 (01:45 +0000)]
Fixed vnic issue after saturn reset

After saturn reboots uVnic will not recover
The issue happens When a host connected to NM3 via infiniband switch.
The fact here is xve driver sets OPER_DOWN when it losses HEART_BEAT and
there is no way for it set this state back .
Fixing this by disjoining Multicast group once HeartBeat is lost and
joining back once saturn comes back

Orabug: 22862488

Reported-by: Ye Jin <ye.jin@oracle.com>
Signed-off-by: Pradeep Gopanapalli <pradeep.gopanapalli@oracle.com>
Reviewed-by: sajid zia <szia@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
9 years agouvnic issues
Pradeep Gopanapalli [Thu, 11 Feb 2016 23:29:39 +0000 (23:29 +0000)]
uvnic issues

When there is Heartbeat loss because of Link event on  EDR fabric, uvnic
recovery fails(some times).
This is due to the fact that XVE_OPER_UP state would have cleared and
xve driver has to check this once multicast group membership is retained

While running MAXq test we are running into soft crash around
skb_try_coalesce .Since xve driver allocates PAGE_SIZE it has to use
full PAGE_SIZE for truesize of skb.

xve driver doesn't allocate enough tailroom in skbs, so IP/TCP
stacks need to reallocate skb head to pull IP/TCP headers.
this .

xve allocates some resources which are not needed by uVnic
functionality.
Fix this by using cm_supported flag.

Orabug: 22862488

Reported-by: ye jin <ye.jin@oracle.com>
Signed-off-by: Pradeep Gopanapalli <pradeep.gopanapalli@oracle.com>
Reviewed-by: sajid zia <szia@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
9 years agoFixed wrongly checked return type Added Debug print
Pradeep Gopanapalli [Tue, 9 Feb 2016 23:15:19 +0000 (23:15 +0000)]
Fixed wrongly checked return type Added Debug print

In xs_post_recv instead of checking for -ENOMEM xscore driver is checking
 ENOMEM, fixed this by checking proper return type

Orabug: 22862488

Reported-by: Haakon Bugge <haakon.bugge@oracle.com>
Signed-off-by: Pradeep Gopanapalli <pradeep.gopanapalli@oracle.com>
Reviewed-by: sajid zia <szia@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
9 years agoRDS/IB: VRPC DELAY / OSS RECONNECT CAUSES 5 MINUTE STALL ON PORT FAILURE
Venkat Venkatsubra [Tue, 1 Mar 2016 22:27:28 +0000 (14:27 -0800)]
RDS/IB: VRPC DELAY / OSS RECONNECT CAUSES 5 MINUTE STALL ON PORT FAILURE

This problem occurs when the user gets notified of a successful
rdma write + bcopy message completion but the peer application
does not receive the bcopy message. This happens during a port down/up test.

What seems to happen is the rdma write succeeds but the bcopy message fails.

RDS should not be returning successful completion status to the user
in this case.

When RDS does a rdma followed by a bcopy message the user notification is
supposed to be implemented by method #3 below.

/* If the user asked for a completion notification on this
 * message, we can implement three different semantics:
 *  1.  Notify when we received the ACK on the RDS message
 *      that was queued with the RDMA. This provides reliable
 *      notification of RDMA status at the expense of a one-way
 *      packet delay.
 *  2.  Notify when the IB stack gives us the completion event for
 *      the RDMA operation.
 *  3.  Notify when the IB stack gives us the completion event for
 *      the accompanying RDS messages.
 * Here, we implement approach #3. To implement approach #2,
 * we would need to take an event for the rdma WR. To implement #1,
 * don't call rds_rdma_send_complete at all, and fall back to the notify
 * handling in the ACK processing code.

But unfortunately the user gets notified earlier to knowing the bcopy
send status. Right after rdma write completes the user gets notified
even though the subsequent bcopy eventually fails.

The fix is to delay signaling completions of rdma op till the
bcopy send completes.

Orabug: 22847528

Acked-by: Rama Nichanamatlu <rama.nichanamatlu@oracle.com>
Signed-off-by: Venkat Venkatsubra <venkat.x.venkatsubra@oracle.com>
9 years agords: add infrastructure to find more details for reconnect failure
Ajaykumar Hotchandani [Fri, 4 Mar 2016 03:23:05 +0000 (19:23 -0800)]
rds: add infrastructure to find more details for reconnect failure

This patch adds run-time support to debug scenarios where reconnect is
not successful for certain time.
We add two sysctl variables for start time and end time. These are
number of seconds after reconnect was initiated.

Orabug: 22631108

Signed-off-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Wengang Wang <wen.gang.wang@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Brian Maly <brian.maly@oracle.com>
9 years agords: find connection drop reason
Ajaykumar Hotchandani [Fri, 4 Mar 2016 03:18:28 +0000 (19:18 -0800)]
rds: find connection drop reason

This patch attempts to find connection drop details.

Rational for adding this type of patch is, there are too many
places from where connection can get dropped.
And, in some cases, we don't have any idea of the source of
connection drop. This is especially painful for issues which
are reproducible in customer environment only.

Idea here is, we have tracker variable which keeps latest value
of connection drop source.
We can fetch that tracker variable as per our need.

Orabug: 22631108

Signed-off-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Acked-by: Wengang Wang <wen.gang.wang@oracle.com>
Signed-off-by: Brian Maly <brian.maly@oracle.com>
9 years agoRDS: Add interface for receive MSG latency trace
Santosh Shilimkar [Fri, 11 Dec 2015 20:01:56 +0000 (12:01 -0800)]
RDS: Add interface for receive MSG latency trace

Socket option to tap receive path latency.
SO_RDS: SO_RDS_MSG_RXPATH_LATENCY
with parameter,
struct rds_rx_trace_so {
u8 rx_traces;
        u8 rx_trace_pos[RDS_MSG_RX_DGRAM_TRACE_MAX];
}

CMSG:
RDS_CMSG_RXPATH_LATENCY(recvmsg)
Returns rds message latencies in various stages of receive
path in nS. Its set per socket using SO_RDS_MSG_RXPATH_LATENCY
socket option. Legitimate points are defined in
enum rds_message_rxpath_latency. More points can be added in
future.

CSMG format:
struct rds_cmsg_rx_trace {
        u8 rx_traces;
        u8 rx_trace_pos[RDS_MSG_RX_DGRAM_TRACE_MAX];
        u64 rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX];
}

Receive MSG trace points: RDS message Receive Path Latency points
enum rds_message_rxpath_latency {
RDS_MSG_RX_HDR_TO_DGRAM_START = 0,
RDS_MSG_RX_DGRAM_REASSEMBLE,
RDS_MSG_RX_DGRAM_DELIVERED,
RDS_MSG_RX_DGRAM_TRACE_MAX
}

Tested-by: Namrata Jampani <namrata.jampani@oracle.com>
Reviewed-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Reviewed-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Orabug: 22630180
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoIB/mlx4: Replace kfree with kvfree in mlx4_ib_destroy_srq
Wengang Wang [Thu, 17 Dec 2015 02:54:15 +0000 (10:54 +0800)]
IB/mlx4: Replace kfree with kvfree in mlx4_ib_destroy_srq

Upstream commit 0ef2f05c7e02ff99c0b5b583d7dee2cd12b053f2 uses vmalloc for
WR buffers when needed and uses kvfree to free the buffers. It missed
changing kfree to kvfree in mlx4_ib_destroy_srq().

Reported-by: Matthew Finaly <matt@Mellanox.com>
Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Backport from upstream: df4176677fdf7ac0c5748083eb7a5b269fb1e156

Orabug: 22487409
Reviewed-by: Wengang Wang <wen.gang.wang@oracle.com>
Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Yuval Shaia <yuval.shaia@oracle.com>
9 years agoib_core: Add udata argument to alloc_shpd()
Mukesh Kacker [Tue, 22 Sep 2015 08:31:31 +0000 (01:31 -0700)]
ib_core: Add udata argument to alloc_shpd()

ib_core: Add udata argument to alloc_shpd()

Add udata argument to shared pd interface alloc_shpd()
consistent with evolution of other similar ib_core
interfaces so providers that wish to support it can use it.

For providers (like current Mellanox driver code) that
do not expect user user data, we assert a warning.

Orabug: 21884873

Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
9 years agoRDS: establish connection for legitimate remote RDMA message
Santosh Shilimkar [Thu, 29 Oct 2015 16:24:46 +0000 (09:24 -0700)]
RDS: establish connection for legitimate remote RDMA message

The first message to a remote node should prompt a new
connection even if it is RDMA operation via CMSG. So that
means before CMSG parsing, the connection needs to be
established. Commit 3d6e0fed8edc ("rds_rdma: rds_sendmsg
should return EAGAIN if connection not setup")' tried to
address that issue as part of bug 20232581.

But it inadvertently broke the QoS policy evaluation. Basically
QoS has opposite requirement where it needs information from
CMSG to evaluate if the message is legitimate to be sent over
the wire. It basically needs to know how the total payload
which should include the actual payload and additional rdma
bytes. It then evaluates total payload with the systems QoS
thresholds to determine if the message is legitimate to be
sent.

Patch addresses these two opposite requirement by fetching
only the rdma bytes information for QoS evaluation and let
the full CMSG parsing happen after the connection is
initiated.  Since the connection establishment is asynchronous,
we make sure the map failure because of unavailable
connection reach to the user by appropriate error code.

Orabug: 22139696

Reviewed-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agords: remove the _reuse_ rds ib pool statistics
Wengang Wang [Tue, 10 Nov 2015 02:43:38 +0000 (10:43 +0800)]
rds: remove the _reuse_ rds ib pool statistics

Orabug: 22124214
fix a regress introduced by ceb99ba579a769f4e02375a3d52e36f44ae5f27f

The above commit introduced the two new statistics to rds_ib_statistics.

       uint64_t        s_ib_rdma_mr_1m_pool_reuse;
       uint64_t        s_ib_rdma_mr_8k_pool_reuse;

But didn't have rds-info changed accordingly thus rds-info gets shifted stats.

this removes these two stats.

Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
9 years agoRDS: Add support for per socket SO_TIMESTAMP for incoming messages
Santosh Shilimkar [Sat, 5 Sep 2015 00:01:06 +0000 (17:01 -0700)]
RDS: Add support for per socket SO_TIMESTAMP for incoming messages

The SO_TIMESTAMP generates time stamp for each incoming RDS message
using the wall time. Result is returned via recv_msg() in a control
message as timeval (usec resolution).

User app can enable it by using SO_TIMESTAMP setsocketopt() at
SOL_SOCKET level. CMSG data of cmsg type SO_TIMESTAMP contains the
time stamp in struct timeval format.

Orabug: 22190837

Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Reviewed-by: Shamir Rabinovitch <shamir.rabinovitch@oracle.com>
Tested-by: Namrata Jampani <namrata.jampani@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: Fix out-of-order RDS_CMSG_RDMA_SEND_STATUS
Wei Lin Guay [Wed, 11 Nov 2015 17:31:14 +0000 (09:31 -0800)]
RDS: Fix out-of-order RDS_CMSG_RDMA_SEND_STATUS

Orabug: 22126982

If the RDS user application requests notification of RDMA send
completions, there is a possibility that RDS_CMSG_RDMA_SEND_STATUS
will be delivered out-of-order.

This can happen if RDS drops sending ACK after it received an explicit
ACK. In this case, the rds message ended up in reverse order in the
list.

Reviewed-by: Håkon Bugge <haakon.bugge@oracle.com>
Reviewed-by: Knut Omang <knut.omang@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Reviewed-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Signed-off-by: Wei Lin Guay <wei.lin.guay@oracle.com>
9 years agoMerge branch 'topic/uek-4.1/ofed' of git://ca-git.us.oracle.com/linux-uek-ofed into...
Santosh Shilimkar [Tue, 10 Nov 2015 02:17:26 +0000 (18:17 -0800)]
Merge branch 'topic/uek-4.1/ofed' of git://ca-git.us.oracle.com/linux-uek-ofed into topic/uek-4.1/ofed

9 years agoMerge branch 'topic/uek-4.1/ofed.ovn' into topic/uek-4.1/ofed
Qing Huang [Tue, 10 Nov 2015 00:23:23 +0000 (16:23 -0800)]
Merge branch 'topic/uek-4.1/ofed.ovn' into topic/uek-4.1/ofed

* topic/uek-4.1/ofed.ovn:
  Integrate Uvnic functionality into uek-4.1 Revision 8008
  1) S_IRWXU causing kernel soft crash changing to 0644 WARNING: CPU: 0 PID: 20907 at fs/sysfs/group.c:61 create_files+0x171/0x180() Oct 12 21:43:14 ovn87-180 kernel: [252606.588541] Attribute vhba_default_scsi_timeout: Invalid permissions 0700 [Rev 8008]
  1) Support vnic for EDR based platform(uVnic) 2) Supported Types now Type 0 - XSMP_XCM_OVN - Xsigo VP780/OSDN standalone Chassis, (add pvi) Type 1 - XSMP_XCM_NOUPLINK - EDR Without uplink (add public-network) Type 2 - XSMP_XCM_UPLINK -EDR with uplink (add public-network <with -if> 3) Intelligence in driver to support all the modes 4) Added Code for printing Multicast LID [Revision 8008] 5) removed style errors

9 years agoIntegrate Uvnic functionality into uek-4.1 Revision 8008
Pradeep Gopanapalli [Sat, 7 Nov 2015 02:14:12 +0000 (18:14 -0800)]
Integrate Uvnic functionality into uek-4.1 Revision 8008

Reviewed-by: Sajid Zia <sajid.zia@oracle.com>
Signed-off-by: Pradeep Gopanapalli <pradeep.gopanapalli@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
9 years ago1) S_IRWXU causing kernel soft crash changing to 0644 WARNING: CPU: 0 PID: 20907...
Pradeep Gopanapalli [Sat, 7 Nov 2015 02:11:33 +0000 (18:11 -0800)]
1) S_IRWXU causing kernel soft crash changing to 0644 WARNING: CPU: 0 PID: 20907 at fs/sysfs/group.c:61 create_files+0x171/0x180() Oct 12 21:43:14 ovn87-180 kernel: [252606.588541] Attribute vhba_default_scsi_timeout: Invalid permissions 0700 [Rev 8008]

Reviewed-by: Sajid Zia <sajid.zia@oracle.com>
Signed-off-by: Pradeep Gopanapalli <pradeep.gopanapalli@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
9 years ago1) Support vnic for EDR based platform(uVnic) 2) Supported Types now Type 0 - XSMP_XC...
Pradeep Gopanapalli [Thu, 5 Nov 2015 02:58:15 +0000 (18:58 -0800)]
1) Support vnic for EDR based platform(uVnic) 2) Supported Types now Type 0 - XSMP_XCM_OVN - Xsigo VP780/OSDN standalone Chassis, (add pvi) Type 1 - XSMP_XCM_NOUPLINK - EDR Without uplink (add public-network) Type 2 - XSMP_XCM_UPLINK -EDR with uplink (add public-network <with -if> 3) Intelligence in driver to support all the modes 4) Added Code for printing Multicast LID [Revision 8008] 5) removed style errors

Reviewed-by: Sajid Zia <sajid.zia@oracle.com>
Signed-off-by: Pradeep Gopanapalli <pradeep.gopanapalli@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
9 years agoIB/mlx4: Use vmalloc for WR buffers when needed
Wengang Wang [Thu, 29 Oct 2015 06:32:45 +0000 (14:32 +0800)]
IB/mlx4: Use vmalloc for WR buffers when needed

Orabug: 22025570

There are several hits that WR buffer allocation(kmalloc) failed.
It failed at order 3 and/or 4 contigous pages allocation. At the same time
there are actually 100MB+ free memory but well fragmented.
So try vmalloc when kmalloc failed.

Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
Acked-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agonet/rds: start rdma listening after ib/iw initialization is done
Qing Huang [Tue, 6 Oct 2015 22:32:22 +0000 (15:32 -0700)]
net/rds: start rdma listening after ib/iw initialization is done

This prevents RDS from handling incoming rdma packets before RDS
completes initializing its recv/send components.

We don't need to call rds_rdma_listen_stop() if rds_rdma_listen_init()
didn't succeed.

We only need to call rds_ib_exit() if rds_ib_init() succeeds but
other parts fail. The same applies to rds_iw_init()/rds_iw_exit().
So we need to change error handling sequence accordingly.

Jump to ib/iw error handling path when we get an err code from
rds_rdma_listen_init().

Orabug: 21684447

Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Reviewed-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoMerge branch 'topic/uek-4.1/ofed.rds-p2' into topic/uek-4.1/ofed
Mukesh Kacker [Thu, 22 Oct 2015 09:44:31 +0000 (02:44 -0700)]
Merge branch 'topic/uek-4.1/ofed.rds-p2' into topic/uek-4.1/ofed

* topic/uek-4.1/ofed.rds-p2:
  net/rds: start rdma listening after ib/iw initialization is done

9 years agonet/rds: start rdma listening after ib/iw initialization is done
Qing Huang [Tue, 6 Oct 2015 22:32:22 +0000 (15:32 -0700)]
net/rds: start rdma listening after ib/iw initialization is done

This prevents RDS from handling incoming rdma packets before RDS
completes initializing its recv/send components.

We don't need to call rds_rdma_listen_stop() if rds_rdma_listen_init()
didn't succeed.

We only need to call rds_ib_exit() if rds_ib_init() succeeds but
other parts fail. The same applies to rds_iw_init()/rds_iw_exit().
So we need to change error handling sequence accordingly.

Jump to ib/iw error handling path when we get an err code from
rds_rdma_listen_init().

Orabug: 21684447

Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Reviewed-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoMerge branch 'topic/uek-4.1/ofed' of git://ca-git.us.oracle.com/linux-uek-ofed into...
Santosh Shilimkar [Wed, 21 Oct 2015 18:15:14 +0000 (11:15 -0700)]
Merge branch 'topic/uek-4.1/ofed' of git://ca-git.us.oracle.com/linux-uek-ofed into topic/uek-4.1/ofed

* 'topic/uek-4.1/ofed' of git://ca-git.us.oracle.com/linux-uek-ofed:
  RDS-TCP: Reset tcp callbacks if re-using an outgoing socket in rds_tcp_accept_one()
  RDS: Invoke ->laddr_check() in rds_bind() for explicitly bound transports.
  RDS: rds_conn_lookup() should factor in the struct net for a match
  RDS: Use a single TCP socket for both send and receive.
  RDS-TCP: Do not bloat sndbuf/rcvbuf in rds_tcp_tune
  RDS-TCP: Set up MSG_MORE and MSG_SENDPAGE_NOTLAST as appropriate in rds_tcp_
  Revert "rds_rdma: rds_sendmsg should return EAGAIN if connection not setup"
  rds: make sure base connection is up on both sides
  rds_ib/iw: fixed big endianness conversion issue for dp->dp_ack_seq
  RDS: fix race condition when sending a message on unbound socket.
  RDS: verify the underlying transport exists before creating a connection
  mlx4: indicate memory resource exhaustion
  IB/mlx4: Use correct order of variables in log message
  mlx4_core: Introduce restrictions for PD update

9 years agoMerge branch 'topic/uek-4.1/ofed.rds-p2' into topic/uek-4.1/ofed
Mukesh Kacker [Wed, 21 Oct 2015 16:11:46 +0000 (09:11 -0700)]
Merge branch 'topic/uek-4.1/ofed.rds-p2' into topic/uek-4.1/ofed

* topic/uek-4.1/ofed.rds-p2:
  RDS-TCP: Reset tcp callbacks if re-using an outgoing socket in rds_tcp_accept_one()
  RDS: Invoke ->laddr_check() in rds_bind() for explicitly bound transports.
  RDS: rds_conn_lookup() should factor in the struct net for a match
  RDS: Use a single TCP socket for both send and receive.
  RDS-TCP: Do not bloat sndbuf/rcvbuf in rds_tcp_tune
  RDS-TCP: Set up MSG_MORE and MSG_SENDPAGE_NOTLAST as appropriate in rds_tcp_
  Revert "rds_rdma: rds_sendmsg should return EAGAIN if connection not setup"
  rds: make sure base connection is up on both sides
  rds_ib/iw: fixed big endianness conversion issue for dp->dp_ack_seq
  RDS: fix race condition when sending a message on unbound socket.
  RDS: verify the underlying transport exists before creating a connection

9 years agoRDS-TCP: Reset tcp callbacks if re-using an outgoing socket in rds_tcp_accept_one()
Sowmini Varadhan [Thu, 15 Oct 2015 02:11:54 +0000 (22:11 -0400)]
RDS-TCP: Reset tcp callbacks if re-using an outgoing socket in rds_tcp_accept_one()

Orabug: 22012202

Backport of upstream commit 241b271952eb ("RDS-TCP: Reset tcp callbacks
if re-using an outgoing socket in rds_tcp_accept_one()")

Consider the following "duelling syn" sequence between two peers A and B:
             A B
             SYN1     -->
                  <-- SYN2
             SYN2ACK  -->

Note that the SYN/ACK has already been sent out by TCP before
rds_tcp_accept_one() gets invoked as part of callbacks.

If the inet_addr(A) is numerically less than inet_addr(B),
the arbitration scheme in rds_tcp_accept_one() will prefer the
TCP connection triggered by SYN1, and will send a CLOSE for the
SYN2 (just after the SYN2ACK was sent).

Since B also follows the same arbitration scheme, it will send the SYN-ACK
for SYN1 that will set up a healthy ESTABLISHED connection on both sides.
B will also get a  CLOSE for SYN2, which should result in the cleanup
of the TCP state machine for SYN2, but it should not trigger any
stale RDS-TCP callbacks (such as ->writespace, ->state_change etc),
that would disrupt the progress of the SYN2 based RDS-TCP  connection.

Thus the arbitration scheme in rds_tcp_accept_one() should restore
rds_tcp callbacks for the winner before setting them up for the
new accept socket, and also make sure that conn->c_outgoing
is set to 0 so that we do not trigger any reconnect attempts on the
passive side of the tcp socket in the future, in conformance with
commit c82ac7e69efe ("net/rds: RDS-TCP: only initiate reconnect attempt
on outgoing TCP socket.")

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
9 years agoRDS: Invoke ->laddr_check() in rds_bind() for explicitly bound transports.
Sowmini Varadhan [Thu, 15 Oct 2015 00:37:47 +0000 (20:37 -0400)]
RDS: Invoke ->laddr_check() in rds_bind() for explicitly bound transports.

Orabug: 22012202

Backport of upstream commit 486798001b92 ("RDS: Invoke ->laddr_check()
in rds_bind() for explicitly bound transports.")

The IP address passed to rds_bind() should be vetted by the
transport's ->laddr_check() for a previously bound transport.
This needs to be done to avoid cases where, for example,
the application has asked for an IB transport,
but the IP address passed to bind is only usable on
ethernet interfaces.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoRDS: rds_conn_lookup() should factor in the struct net for a match
Sowmini Varadhan [Wed, 14 Oct 2015 19:52:23 +0000 (15:52 -0400)]
RDS: rds_conn_lookup() should factor in the struct net for a match

Orabug: 22012202

Backport of upstream commit 8f384c0177a0 ("RDS: rds_conn_lookup() should
factor in the struct net for a match")

Only return a conn if the rds_conn_net(conn) matches the struct
net passed to rds_conn_lookup().

Fixes: 467fa15356ac ("RDS-TCP: Support multiple RDS-TCP listen endpoints,
           one per netns.")

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoRDS: Use a single TCP socket for both send and receive.
Sowmini Varadhan [Wed, 14 Oct 2015 14:29:31 +0000 (10:29 -0400)]
RDS: Use a single TCP socket for both send and receive.

Orabug: 22012202

Backport of upstream commit 3b20fc389705 ("RDS: Use a single TCP
socket for both send and receive.");

Commit f711a6ae062c ("net/rds: RDS-TCP: Always create a new rds_sock
for an incoming connection.") modified rds-tcp so that an incoming SYN
would ignore an existing "client" TCP connection which had the local
port set to the transient port.  The motivation for ignoring the existing
"client" connection in f711a6ae was to avoid race conditions and an
endless duel of reconnect attempts triggered by a restart/abort of one
of the nodes in the TCP connection.

However, having separate sockets for active and passive sides
is avoidable, and the simpler model of a single TCP socket for
both send and receives of all RDS connections associated with
that tcp socket makes for easier observability. We avoid the race
conditions from f711a6ae by attempting reconnects in rds_conn_shutdown
if, and only if, the (new) c_outgoing bit is set for RDS_TRANS_TCP.
The c_outgoing bit is initialized in __rds_conn_create().

A side-effect of re-using the client rds_connection for an incoming
SYN is the potential of encountering duelling SYNs, i.e., we
have an outgoing RDS_CONN_CONNECTING socket when we get the incoming
SYN. The logic to arbitrate this criss-crossing SYN exchange in
rds_tcp_accept_one() has been modified to emulate the BGP state
machine: the smaller IP address should back off from the connection
attempt.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoRDS-TCP: Do not bloat sndbuf/rcvbuf in rds_tcp_tune
Sowmini Varadhan [Wed, 14 Oct 2015 11:23:20 +0000 (07:23 -0400)]
RDS-TCP: Do not bloat sndbuf/rcvbuf in rds_tcp_tune

Orabug: 22012202

Backport of upstream commit 1edd6a14d24f ("RDS-TCP: Do not bloat
sndbuf/rcvbuf in rds_tcp_tune")

Using the value of RDS_TCP_DEFAULT_BUFSIZE (128K)
clobbers efficient use of TSO because it inflates the size_goal
that is computed in tcp_sendmsg/tcp_sendpage and skews packet
latency, and the default values for these parameters actually
results in significantly better performance.

In request-response tests using rds-stress with a packet size of
100K with 16 threads (test parameters -q 100000 -a 256 -t16 -d16)
between a single pair of IP addresses achieves a throughput of
6-8 Gbps. Without this patch, throughput maxes at 2-3 Gbps under
equivalent conditions on these platforms.

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
9 years agoRDS-TCP: Set up MSG_MORE and MSG_SENDPAGE_NOTLAST as appropriate in rds_tcp_
Sowmini Varadhan [Wed, 14 Oct 2015 01:03:04 +0000 (21:03 -0400)]
RDS-TCP: Set up MSG_MORE and MSG_SENDPAGE_NOTLAST as appropriate in rds_tcp_

Backport of upstream commit 76b29ef120f5 ("RDS-TCP: Set up MSG_MORE and
MSG_SENDPAGE_NOTLAST as appropriate in rds_tcp_xmit")

For the same reasons as commit 2f5338442425 ("tcp: allow splice() to
build full TSO packets") and commit 35f9c09fe9c7 ("tcp: tcp_sendpages()
should call tcp_push() once"), rds_tcp_xmit may have multiple pages to
send, so use the MSG_MORE and MSG_SENDPAGE_NOTLAST as hints to
tcp_sendpage()

Orabug: 22012202

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoRevert "rds_rdma: rds_sendmsg should return EAGAIN if connection not setup"
Rama Nichanamatlu [Tue, 20 Oct 2015 13:16:25 +0000 (06:16 -0700)]
Revert "rds_rdma: rds_sendmsg should return EAGAIN if connection not setup"

This reverts commit 3d6e0fed8edc2f5d5439bee22c2fa153096c77ea.

Reverted because fix has bug that affects rds qos threshold
settings.

Orabug: 21664735

Signed-off-by: Rama Nichanamatlu <rama.nichanamatlu@oracle.com>
Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agords: make sure base connection is up on both sides
Ajaykumar Hotchandani [Tue, 5 May 2015 03:09:42 +0000 (20:09 -0700)]
rds: make sure base connection is up on both sides

Current RDS active side requires zero lane path records for establishing
non-zero lane connection. For this reason, active side makes sure to
have zero lane connection up before establishing non-zero lane
connection. Passive side does not require to fetch path records, so it
does not have this check.

This has possibility of connection having non-ideal path records in
following scenario:
- Host1 had PORT_UP event.
- Lane0 and Lane6 connection went down.
- Lane0 connection came up.
- Host1 sent connection request for Lane6.
- Host2 had PORT_UP event.
- Lane0 and Lane6 connections went down.
- Host2 sent DREQ for Lane0.
- Since Lane6 connection is not up, it does not require to do anything.
- Host2 received connection request from host1 having old path records
  for Lane6.
- Lane6 connection got established on old path records.

Following are impacts of having connections with non-ideal path records:
- minor performance hit because of extra hop with ISL path
- in port failure scenario, it impacts connections which are not related
  to that port.

With this patch we make sure that base connection is up on passive side
as well before allowing to establish connection.

(This is port of UEK2 commit 7ab7ef255a)

Orabug: 21675157

Tested-by: Michael Nowak <michael.nowak@oracle.com>
Signed-off-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Qing Huang <qing.huang@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agords_ib/iw: fixed big endianness conversion issue for dp->dp_ack_seq
Qing Huang [Tue, 13 Oct 2015 00:23:50 +0000 (17:23 -0700)]
rds_ib/iw: fixed big endianness conversion issue for dp->dp_ack_seq

dp->dp_ack_seq is used in big endian format. We need to do the
big endianness conversion when we assign a value in host format
to it.

(This patch is ported from UEK2)

Orabug: 21684819

Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Reviewed-by: Wengang Wang <wen.gang.wang@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoRDS: fix race condition when sending a message on unbound socket.
Quentin Casasnovas [Mon, 19 Oct 2015 21:22:27 +0000 (14:22 -0700)]
RDS: fix race condition when sending a message on unbound socket.

Sasha's found a NULL pointer dereference in the RDS connection code when
sending a message to an apparently unbound socket.  The problem is caused
by the code checking if the socket is bound in rds_sendmsg(), which checks
the rs_bound_addr field without taking a lock on the socket.  This opens a
race where rs_bound_addr is temporarily set but where the transport is not
in rds_bind(), leading to a NULL pointer dereference when trying to
dereference 'trans' in __rds_conn_create().

Vegard wrote a reproducer for this issue, so kindly ask him to share if
you're interested.

I cannot reproduce the NULL pointer dereference using Vegard's reproducer
with this patch, whereas I could without.

Complete earlier incomplete fix to CVE-2015-6937:

  74e98eb08588 ("RDS: verify the underlying transport exists before creating a connection")

Signed-off-by: Quentin Casasnovas <quentin.casasnovas@oracle.com>
Reviewed-by: Vegard Nossum <vegard.nossum@oracle.com>
Reviewed-by: Sasha Levin <sasha.levin@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Cc: Vegard Nossum <vegard.nossum@oracle.com>
Cc: Sasha Levin <sasha.levin@oracle.com>
Cc: Chien Yen <chien.yen@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: stable@vger.kernel.org
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: verify the underlying transport exists before creating a connection
Sasha Levin [Tue, 8 Sep 2015 14:53:40 +0000 (10:53 -0400)]
RDS: verify the underlying transport exists before creating a connection

There was no verification that an underlying transport exists when creating
a connection, this would cause dereferencing a NULL ptr.

It might happen on sockets that weren't properly bound before attempting to
send a message, which will cause a NULL ptr deref:

[135546.047719] kasan: GPF could be caused by NULL-ptr deref or user memory accessgeneral protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC KASAN
[135546.051270] Modules linked in:
[135546.051781] CPU: 4 PID: 15650 Comm: trinity-c4 Not tainted 4.2.0-next-20150902-sasha-00041-gbaa1222-dirty #2527
[135546.053217] task: ffff8800835bc000 ti: ffff8800bc708000 task.ti: ffff8800bc708000
[135546.054291] RIP: __rds_conn_create (net/rds/connection.c:194)
[135546.055666] RSP: 0018:ffff8800bc70fab0  EFLAGS: 00010202
[135546.056457] RAX: dffffc0000000000 RBX: 0000000000000f2c RCX: ffff8800835bc000
[135546.057494] RDX: 0000000000000007 RSI: ffff8800835bccd8 RDI: 0000000000000038
[135546.058530] RBP: ffff8800bc70fb18 R08: 0000000000000001 R09: 0000000000000000
[135546.059556] R10: ffffed014d7a3a23 R11: ffffed014d7a3a21 R12: 0000000000000000
[135546.060614] R13: 0000000000000001 R14: ffff8801ec3d0000 R15: 0000000000000000
[135546.061668] FS:  00007faad4ffb700(0000) GS:ffff880252000000(0000) knlGS:0000000000000000
[135546.062836] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[135546.063682] CR2: 000000000000846a CR3: 000000009d137000 CR4: 00000000000006a0
[135546.064723] Stack:
[135546.065048]  ffffffffafe2055c ffffffffafe23fc1 ffffed00493097bf ffff8801ec3d0008
[135546.066247]  0000000000000000 00000000000000d0 0000000000000000 ac194a24c0586342
[135546.067438]  1ffff100178e1f78 ffff880320581b00 ffff8800bc70fdd0 ffff880320581b00
[135546.068629] Call Trace:
[135546.069028] ? __rds_conn_create (include/linux/rcupdate.h:856 net/rds/connection.c:134)
[135546.069989] ? rds_message_copy_from_user (net/rds/message.c:298)
[135546.071021] rds_conn_create_outgoing (net/rds/connection.c:278)
[135546.071981] rds_sendmsg (net/rds/send.c:1058)
[135546.072858] ? perf_trace_lock (include/trace/events/lock.h:38)
[135546.073744] ? lockdep_init (kernel/locking/lockdep.c:3298)
[135546.074577] ? rds_send_drop_to (net/rds/send.c:976)
[135546.075508] ? __might_fault (./arch/x86/include/asm/current.h:14 mm/memory.c:3795)
[135546.076349] ? __might_fault (mm/memory.c:3795)
[135546.077179] ? rds_send_drop_to (net/rds/send.c:976)
[135546.078114] sock_sendmsg (net/socket.c:611 net/socket.c:620)
[135546.078856] SYSC_sendto (net/socket.c:1657)
[135546.079596] ? SYSC_connect (net/socket.c:1628)
[135546.080510] ? trace_dump_stack (kernel/trace/trace.c:1926)
[135546.081397] ? ring_buffer_unlock_commit (kernel/trace/ring_buffer.c:2479 kernel/trace/ring_buffer.c:2558 kernel/trace/ring_buffer.c:2674)
[135546.082390] ? trace_buffer_unlock_commit (kernel/trace/trace.c:1749)
[135546.083410] ? trace_event_raw_event_sys_enter (include/trace/events/syscalls.h:16)
[135546.084481] ? do_audit_syscall_entry (include/trace/events/syscalls.h:16)
[135546.085438] ? trace_buffer_unlock_commit (kernel/trace/trace.c:1749)
[135546.085515] rds_ib_laddr_check(): addr 36.74.25.172 ret -99 node type -1

Orabug: 22010933

Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Wengang Wang <wen.gang.wang@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
(cherry picked from commit 74e98eb085889b0d2d4908f59f6e00026063014f)

Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoMerge branch 'topic/uek-4.1/ofed.mlnx2.4-p3.orclFixes' into topic/uek-4.1/ofed
Mukesh Kacker [Wed, 21 Oct 2015 15:09:05 +0000 (08:09 -0700)]
Merge branch 'topic/uek-4.1/ofed.mlnx2.4-p3.orclFixes' into topic/uek-4.1/ofed

* topic/uek-4.1/ofed.mlnx2.4-p3.orclFixes:
  mlx4: indicate memory resource exhaustion
  IB/mlx4: Use correct order of variables in log message
  mlx4_core: Introduce restrictions for PD update

9 years agomlx4: indicate memory resource exhaustion
Ajaykumar Hotchandani [Wed, 14 Oct 2015 23:30:26 +0000 (16:30 -0700)]
mlx4: indicate memory resource exhaustion

Change here provides details about pid, which resource got exhausted and
current limits for that resource.

These details are requested by database folks to attempt recovery.

This is port from UEK2 commit f3774780e30

Orabug: 21549767

Signed-off-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: Guangyu Sun <guangyu.sun@oracle.com>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoIB/mlx4: Use correct order of variables in log message
Wengang Wang [Thu, 15 Oct 2015 01:04:35 +0000 (09:04 +0800)]
IB/mlx4: Use correct order of variables in log message

There was a incorrect order of variables in mlx4_log().
This fix corrects that mistake.

Orabug: 21906781

Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
Acked-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agomlx4_core: Introduce restrictions for PD update
Ajaykumar Hotchandani [Wed, 14 Oct 2015 23:38:11 +0000 (16:38 -0700)]
mlx4_core: Introduce restrictions for PD update

From 2.31.5350 firmware onwards,
- RDS with RDMA data transfer stopped working.
- Mellanox has introduced limitations related to PD updates.
  These imposed limitations are inline with PRM.

This patch makes driver in sync with these imposed limitations.
Mellanox R&D has approved this patch.

It's been tested on both old firmware (2.11.1280) and new firmware.

Mellanox case number is 179121.
OraBug: 22022389

v2: Change subject as per suggestion from Yuval

Tested-by: Pierre Orzechowski <pierre.e.orzechowski@oracle.com>
Tested-by: Kushagra Misra <kushagra.misra@oracle.com>
Signed-off-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoRDS: fix race condition when sending a message on unbound socket.
Quentin Casasnovas [Mon, 19 Oct 2015 21:22:27 +0000 (14:22 -0700)]
RDS: fix race condition when sending a message on unbound socket.

Sasha's found a NULL pointer dereference in the RDS connection code when
sending a message to an apparently unbound socket.  The problem is caused
by the code checking if the socket is bound in rds_sendmsg(), which checks
the rs_bound_addr field without taking a lock on the socket.  This opens a
race where rs_bound_addr is temporarily set but where the transport is not
in rds_bind(), leading to a NULL pointer dereference when trying to
dereference 'trans' in __rds_conn_create().

Vegard wrote a reproducer for this issue, so kindly ask him to share if
you're interested.

I cannot reproduce the NULL pointer dereference using Vegard's reproducer
with this patch, whereas I could without.

Complete earlier incomplete fix to CVE-2015-6937:

  74e98eb08588 ("RDS: verify the underlying transport exists before creating a connection")

Signed-off-by: Quentin Casasnovas <quentin.casasnovas@oracle.com>
Reviewed-by: Vegard Nossum <vegard.nossum@oracle.com>
Reviewed-by: Sasha Levin <sasha.levin@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Cc: Vegard Nossum <vegard.nossum@oracle.com>
Cc: Sasha Levin <sasha.levin@oracle.com>
Cc: Chien Yen <chien.yen@oracle.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: stable@vger.kernel.org
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoMerge branch 'topic/uek-4.1/ofed' of git://ca-git.us.oracle.com/linux-uek-ofed into...
Santosh Shilimkar [Tue, 13 Oct 2015 17:10:40 +0000 (10:10 -0700)]
Merge branch 'topic/uek-4.1/ofed' of git://ca-git.us.oracle.com/linux-uek-ofed into topic/uek-4.1/ofed

* 'topic/uek-4.1/ofed' of git://ca-git.us.oracle.com/linux-uek-ofed:
  RDS/IB: print string constants in more places
  ib/rds: runtime debuggability enhancement

9 years agoRDS: make send_batch_count tunable effective
Santosh Shilimkar [Thu, 8 Oct 2015 22:59:14 +0000 (15:59 -0700)]
RDS: make send_batch_count tunable effective

The send_batch_count tunable is stale and code relies on
hard-coded batch count value. Its a nice feature and lets you
tune the system based on different HCAs. TCP transport as well
have different characteristics and tunable can be useful.

There is no change in default behavior with this patch.

Orabug: 22010933

Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Wengang Wang <wen.gang.wang@oracle.com>
Reported-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: make use of kfree_rcu() and avoid the call_rcu() chain
Santosh Shilimkar [Thu, 8 Oct 2015 23:26:32 +0000 (16:26 -0700)]
RDS: make use of kfree_rcu() and avoid the call_rcu() chain

call_rcu() chains are expensive and its use in rds_ib_remove_ipaddr()
is just to kfree() the rds_ib_ipaddr. Chains make use of  high-latency
rcu_barrier() in modules which can be avoided.

Makes use of kfree_rcu() which is exactly meant for such use

Orabug: 22010933

Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Wengang Wang <wen.gang.wang@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoRDS: verify the underlying transport exists before creating a connection
Sasha Levin [Tue, 8 Sep 2015 14:53:40 +0000 (10:53 -0400)]
RDS: verify the underlying transport exists before creating a connection

There was no verification that an underlying transport exists when creating
a connection, this would cause dereferencing a NULL ptr.

It might happen on sockets that weren't properly bound before attempting to
send a message, which will cause a NULL ptr deref:

[135546.047719] kasan: GPF could be caused by NULL-ptr deref or user memory accessgeneral protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC KASAN
[135546.051270] Modules linked in:
[135546.051781] CPU: 4 PID: 15650 Comm: trinity-c4 Not tainted 4.2.0-next-20150902-sasha-00041-gbaa1222-dirty #2527
[135546.053217] task: ffff8800835bc000 ti: ffff8800bc708000 task.ti: ffff8800bc708000
[135546.054291] RIP: __rds_conn_create (net/rds/connection.c:194)
[135546.055666] RSP: 0018:ffff8800bc70fab0  EFLAGS: 00010202
[135546.056457] RAX: dffffc0000000000 RBX: 0000000000000f2c RCX: ffff8800835bc000
[135546.057494] RDX: 0000000000000007 RSI: ffff8800835bccd8 RDI: 0000000000000038
[135546.058530] RBP: ffff8800bc70fb18 R08: 0000000000000001 R09: 0000000000000000
[135546.059556] R10: ffffed014d7a3a23 R11: ffffed014d7a3a21 R12: 0000000000000000
[135546.060614] R13: 0000000000000001 R14: ffff8801ec3d0000 R15: 0000000000000000
[135546.061668] FS:  00007faad4ffb700(0000) GS:ffff880252000000(0000) knlGS:0000000000000000
[135546.062836] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[135546.063682] CR2: 000000000000846a CR3: 000000009d137000 CR4: 00000000000006a0
[135546.064723] Stack:
[135546.065048]  ffffffffafe2055c ffffffffafe23fc1 ffffed00493097bf ffff8801ec3d0008
[135546.066247]  0000000000000000 00000000000000d0 0000000000000000 ac194a24c0586342
[135546.067438]  1ffff100178e1f78 ffff880320581b00 ffff8800bc70fdd0 ffff880320581b00
[135546.068629] Call Trace:
[135546.069028] ? __rds_conn_create (include/linux/rcupdate.h:856 net/rds/connection.c:134)
[135546.069989] ? rds_message_copy_from_user (net/rds/message.c:298)
[135546.071021] rds_conn_create_outgoing (net/rds/connection.c:278)
[135546.071981] rds_sendmsg (net/rds/send.c:1058)
[135546.072858] ? perf_trace_lock (include/trace/events/lock.h:38)
[135546.073744] ? lockdep_init (kernel/locking/lockdep.c:3298)
[135546.074577] ? rds_send_drop_to (net/rds/send.c:976)
[135546.075508] ? __might_fault (./arch/x86/include/asm/current.h:14 mm/memory.c:3795)
[135546.076349] ? __might_fault (mm/memory.c:3795)
[135546.077179] ? rds_send_drop_to (net/rds/send.c:976)
[135546.078114] sock_sendmsg (net/socket.c:611 net/socket.c:620)
[135546.078856] SYSC_sendto (net/socket.c:1657)
[135546.079596] ? SYSC_connect (net/socket.c:1628)
[135546.080510] ? trace_dump_stack (kernel/trace/trace.c:1926)
[135546.081397] ? ring_buffer_unlock_commit (kernel/trace/ring_buffer.c:2479 kernel/trace/ring_buffer.c:2558 kernel/trace/ring_buffer.c:2674)
[135546.082390] ? trace_buffer_unlock_commit (kernel/trace/trace.c:1749)
[135546.083410] ? trace_event_raw_event_sys_enter (include/trace/events/syscalls.h:16)
[135546.084481] ? do_audit_syscall_entry (include/trace/events/syscalls.h:16)
[135546.085438] ? trace_buffer_unlock_commit (kernel/trace/trace.c:1749)
[135546.085515] rds_ib_laddr_check(): addr 36.74.25.172 ret -99 node type -1

Orabug: 22010933

Acked-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Acked-by: Wengang Wang <wen.gang.wang@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
(cherry picked from commit 74e98eb085889b0d2d4908f59f6e00026063014f)

Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
9 years agoMerge branch 'topic/uek-4.1/ofed.rds-p2' into topic/uek-4.1/ofed
Mukesh Kacker [Tue, 13 Oct 2015 15:44:42 +0000 (08:44 -0700)]
Merge branch 'topic/uek-4.1/ofed.rds-p2' into topic/uek-4.1/ofed

* topic/uek-4.1/ofed.rds-p2:
  RDS/IB: print string constants in more places
  ib/rds: runtime debuggability enhancement

9 years agoRDS/IB: print string constants in more places
Zach Brown [Tue, 3 Aug 2010 20:52:47 +0000 (13:52 -0700)]
RDS/IB: print string constants in more places

This prints the constant identifier for work completion status and rdma
cm event types, like we already do for IB event types.

A core string array helper is added that each string type uses.

Note: The following is the original commit in uek prior to the big uek2
ofed blob patch.

    commit 59f740a6aeb2cde2f79fe0df38262d4c1ef35cd8

Orabug 21314268

Reviewed-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
Signed-off-by: Zach Brown <zach.brown@oracle.com>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoib/rds: runtime debuggability enhancement
Qing Huang [Wed, 15 Jul 2015 01:36:43 +0000 (18:36 -0700)]
ib/rds: runtime debuggability enhancement

This patch provides the ability to dynamically turn on or off various
types of debug/diag prints inside the RDS module.

The run-time debug prints are controlled by a rds module parameter,
rds_rt_debug_bitmap.

Here is the definition for different bits. We have implemented feature
related bits, such as Connection Management, Active Bonding, Error prints,
Send, Recv.

in net/rds/rds_rt_debug.h
...
enum {
        /* bit 0 ~ 19 are feature related bits */
        RDS_RTD_ERR                     = 1 << 0,       /* 0x1    */
        RDS_RTD_ERR_EXT                 = 1 << 1,       /* 0x2    */

        RDS_RTD_CM                      = 1 << 3,       /* 0x8    */
        RDS_RTD_CM_EXT                  = 1 << 4,       /* 0x10   */
        RDS_RTD_CM_EXT_P                = 1 << 5,       /* 0x20   */

        RDS_RTD_ACT_BND                 = 1 << 7,       /* 0x80   */
        RDS_RTD_ACT_BND_EXT             = 1 << 8,       /* 0x100  */

        RDS_RTD_RCV                     = 1 << 11,      /* 0x800  */
        RDS_RTD_RCV_EXT                 = 1 << 12,      /* 0x1000 */

        RDS_RTD_SND                     = 1 << 14,      /* 0x4000 */
        RDS_RTD_SND_EXT                 = 1 << 15,      /* 0x8000 */
...

In general, *EXTRA bits mean that you will get extra information but
possible flood prints as well. But every bit can be controlled by users
so users can decide how much information they want to see/collect. The
current embedded printk level used for this patch is KERN_INFO. Most
likely all the msgs will only go to /var/log/messages without showing up
on console if we use the default settings for /proc/sys/kernel/printk and
/etc/rsyslog.conf in ol6 environment.

E.g if we want to turn on RDS_RTD_ERR and RDS_RTD_CM bits. What we can
do is

echo 0x9 > /sys/module/rds/parameters/rds_rt_debug_bitmap

To turn on RDS_RTD_ERR(0x1), RDS_RTD_CM(0x8), and RDS_RTD_RCV(0x800) bits

echo 0x809 > /sys/module/rds/parameters/rds_rt_debug_bitmap

Performance penalty: with all the debug flag bits set to 0, there
should be no porformance impact in a kernel with this patch.

Orabug 21314268

Reviewed-by: Ajaykumar Hotchandani <ajaykumar.hotchandani@oracle.com>
Reviewed-by: Shamir Rabinovitch <shamir.rabinovitch@oracle.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: Qing Huang <qing.huang@oracle.com>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agouek-rpm: Enable config for OVN xsigo drivers
Mukesh Kacker [Mon, 9 Feb 2015 02:54:50 +0000 (18:54 -0800)]
uek-rpm: Enable config for OVN xsigo drivers

Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoMerge branch 'topic/uek-4.1/ofed.xsigo' into topic/uek-4.1/ofed
Mukesh Kacker [Wed, 7 Oct 2015 11:40:32 +0000 (04:40 -0700)]
Merge branch 'topic/uek-4.1/ofed.xsigo' into topic/uek-4.1/ofed

* topic/uek-4.1/ofed.xsigo:
  Add Oracle virtual Networking Drivers for uek4 kernel

9 years agoAdd Oracle virtual Networking Drivers for uek4 kernel
Pradeep Gopanapalli [Wed, 23 Sep 2015 01:56:41 +0000 (18:56 -0700)]
Add Oracle virtual Networking Drivers for uek4 kernel

This commit adds 4 kernel modules: xscore, xsvnic, xve
and xsvhba developed by Xsigo (acquired by Oracle) and used in the Oracle
virtual networking (OVN) products which provide provide virtual network and
storage adapter devices on the servers dynamically at runtime.

The heart of OVN product is the Fabric Interconnect (FI).
Hosts and IO modules connect to the FI using Infiniband fabric.
IO modules can be N/W card or/and FC card.

The "xscore" module is responsible for doing FI topology discovery
and establishing the connection with FI. It is involved in retrieving
virtual device management commands such as INSTALL, DELETE, etc.
This module provides wrapper for IB framework API's which will be used
by its client  modules "xsvnic", "xsvhba" and "xve".

The "xve" module supprots the Xsigo Virtual Ethernet(XVE) protocol.
The "xsvnic" module supports the Xsigo vNIC functinality. These modules
interface between kernel networking stack and the "xscore" module.
On the egress side, it processes the N/W packet sends it to "xscore"
module which is then wrapped into a IB packet.

On the ingress side, "xscore" receives the N/W packet which is
encapsulated inside IB packet and transfers it to "xsvnic" or "xve".
The modules "xsvnic"/"xve" process this packet and send it to the
kernel networking stack. The "xsvnic" interacts with N/W card gateway
connected to the FI whereas, "xve" interacts with another host in the
same IB fabric.

The "xsvhba" module support for the Xsigo virtual HBA allowing SAN
Connectivity. The "xsvhba" module interfaces with SCSI layer. It
communicates with the FC card gateway connected to the FI. It is
responsible for accepting/transporting the SCSI commands from/to
the specified SCSI target. The "xsvhba" module uses "xscore" to
wrap(unwrap) the commands in a IB packet and transmit(receive) it.

Signed-off-by: Pradeep Gopanapalli <pradeep.gopanapalli@oracle.com>
Signed-off-by: Mukesh Kacker <mukesh.kacker@oracle.com>
9 years agoMerge branch 'topic/uek-4.1/ofed.mlnx2.4-p3.orclFixes' into topic/uek-4.1/ofed
Mukesh Kacker [Tue, 6 Oct 2015 14:36:37 +0000 (07:36 -0700)]
Merge branch 'topic/uek-4.1/ofed.mlnx2.4-p3.orclFixes' into topic/uek-4.1/ofed

* topic/uek-4.1/ofed.mlnx2.4-p3.orclFixes:
  ib_sdp/cma: readd SDP support to cma_save_net_info