Increase IPoIB ring sizes to twice their original sizes (RX: 128->256,
TX: 64->128) to act as a shock absorber for high traffic peaks. With
the current settings, we have seen cases that there are many calls to
netif_stop_queue(), which causes degradation in throughput. Also,
larger receive buffer sizes help IPoIB in CM mode to avoid experiencing
RNR NAK conditions due to insufficient receive buffers at the SRQ.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Since IPoIB connected mode does not NETIF_F_SG, we only have one DMA
mapping per send, so we don't need a mapping[] array. Define a new
struct with a single u64 mapping member and use it for the CM tx_ring.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The IB spe. for SubnGet(NodeInfo) and query HCA says that the vendor
ID field should be the IEEE OUI assigned to the vendor. The ipath
driver was returning the PCI vendor ID instead. This will affect
applications which call ibv_query_device(). The old value was
0x001fc1 or 0x001077, the new value is 0x001175.
The vendor ID doesn't appear to be exported via /sys so that should
reduce possible compatibility issues. I'm only aware of Open MPI as a
major application which depends on this change, and they have made
necessary adjustments.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When the driver sets the MTU of the net device outside of its
change_mtu method, it should make use of dev_set_mtu() instead of
directly setting the mtu field of struct netdevice. Otherwise
functions registered to be called upon MTU change will not get called
(this is done through call_netdevice_notifiers() in dev_set_mtu()).
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Use of this lock is required to synchronize changes to the netdvice's
data structs. Also move the call to ipoib_flush_paths() after the
modification of the netdevice flags in set_mode().
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
ipoib_mcast_detach() does nothing except call ib_detach_mcast(), so just
use the core API in the one place that does a multicast group detach.
add/remove: 0/1 grow/shrink: 0/1 up/down: 0/-105 (-105)
function old new delta
ipoib_mcast_leave 357 319 -38
ipoib_mcast_detach 67 - -67
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The current code will set the Q_Key for any join of a non-sendonly
multicast group. The operation involves a modify QP operation, which
is fairly heavyweight, and is only really required after the join of
the broadcast group. Fix this by adding a parameter to ipoib_mcast_attach()
to control when the Q_Key is set.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
No need for a mutex around calls to ib_attach_mcast/ib_detach_mcast
since these operations are synchronized at the HW driver layer.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The IPOIB_MCAST_STARTED flag is not used at all since commit b3e2749b
("IPoIB: Don't drop multicast sends when they can be queued"), so
remove it.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Every caller of nes_post_cqp_request() passed it NES_CQP_REQUEST_RING_DOORBELL,
so just remove that parameter and always ring the doorbell.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Acked-by: Faisal Latif <flatif@neteffect.com>
cxgb3 does not currently report the page size capabilities, and
incorrectly reports them internally.
This version changes the bit-shifting to a static value (per Steve's
request).
Signed-off-by: Jon Mason <jon@opengridcomputing.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The patch tries to solve the problem of device going down and paths being
flushed on an SM change event. The method is to mark the paths as candidates for
refresh (by setting the new valid flag to 0), and wait for an ARP
probe a new path record query.
The solution requires a different and less intrusive handling of SM
change event. For that, the second argument of the flush function
changes its meaning from a boolean flag to a level. In most cases, SM
failover doesn't cause LID change so traffic won't stop. In the rare
cases of LID change, the remote host (the one that hadn't changed its
LID) will lose connectivity until paths are refreshed. This is no
worse than the current state. In fact, preventing the device from
going down saves packets that otherwise would be lost.
Signed-off-by: Moni Levy <monil@voltaire.com>
Signed-off-by: Moni Shoua <monis@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
This gives ehca an autogenerated modalias and therefore enables automatic loading.
Signed-off-by: Joachim Fenkes <fenkes@de.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add "ipoib_use_lro" module parameter to enable LRO and an
"ipoib_lro_max_aggr" module parameter to set the max number of packets
to be aggregated. Make LRO controllable and LRO statistics accessible
through ethtool.
Signed-off-by: Vladimir Sokolovsky <vlad@mellanox.co.il>
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Set IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK for IPoIB's UD QPs if
supported by the underlying device. This creates an improvement of up
to 39% in bandwidth when sending multicast packets with IPoIB, and an
improvment of 12% in cpu usage.
Signed-off-by: Ron Livne <ronli@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add support for handling the IB_QP_CREATE_MULTICAST_BLOCK_LOOPBACK
flag by using the per-multicast group loopback blocking feature of
mlx4 hardware.
Signed-off-by: Ron Livne <ronli@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
- Add a new rdma ctl command called RDMA_GET_MIB to the cxgb3 low
level driver to obtain the protocol mib from the rnic hardware.
- Add new iw_cxgb3 provider method to get the MIB from the low level
driver.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
This patch adds a sysfs attribute group called "proto_stats" under
/sys/class/infiniband/$device/ and populates this group with protocol
statistics if they exist for a given device. Currently, only iWARP
stats are defined, but the code is designed to allow InfiniBand
protocol stats if they become available. These stats are per-device
and more importantly -not- per port.
Details:
- Add union rdma_protocol_stats in ib_verbs.h. This union allows
defining transport-specific stats. Currently only iwarp stats are
defined.
- Add struct iw_protocol_stats to define the current set of iwarp
protocol stats.
- Add new ib_device method called get_proto_stats() to return protocol
statistics.
- Add logic in core/sysfs.c to create iwarp protocol stats attributes
if the device is an RNIC and has a get_proto_stats() method.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
For devices that don't support SRQs, ipoib_cm_post_receive_nonsrq() is
called from both ipoib_cm_handle_rx_wc() and ipoib_cm_nonsrq_init_rx(),
and these two callers are not synchronized against each other.
However, ipoib_cm_post_receive_nonsrq() always reuses the same receive
work request and scatter list structures, so multiple callers can end
up stepping on each other, which leads to posting garbled work
requests.
Fix this by having the caller pass in the ib_recv_wr and ib_sge
structures to use, and allocating new local structures in
ipoib_cm_nonsrq_init_rx().
Based on a patch by Pradeep Satyanarayana <pradeep@us.ibm.com> and
David Wilder <dwilder@us.ibm.com>, with debugging help from Hoang-Nam
Nguyen <hnguyen@de.ibm.com>.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The members struct iwch_rnic_attributes.vendor_id and .vendor_part_id
are write-only, so we might as well get rid of them.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
- set fw_ver
- set hw_ver
- set max_qp_wr to something reasonable
- set max_cqe to something reasonable
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
During corner case testing, we noticed that some versions of ehca do
not properly transition to interrupt done in special load situations.
This can be resolved by periodically triggering EOI through H_EOI, if
EQEs are pending.
Signed-off-by: Stefan Roscher <stefan.roscher@de.ibm.com>
Acked-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Commit 65adfa91 ("IB/mlx4: Fix RESET to RESET and RESET to ERROR
transitions") added some extra code to handle a QP state transition
from RESET to ERROR. However, the latest 1.2.1 version of the IB spec
has clarified that this transition is actually not allowed, so we can
remove this extra code again.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Commit b18aad71 ("IB/mthca: Fix RESET to ERROR transition") added some
extra code to handle a QP state transition from RESET to ERROR.
However, the latest 1.2.1 version of the IB spec has clarified that
this transition is actually not allowed, so we can remove this extra
code again.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
I was reviewing the QP state transition diagram in the IB 1.2.1 spec
and the code for qp_state_table[], and noticed that the code allows a
QP to be modified from IB_QPS_RESET to IB_QPS_ERR whereas the notes
for figure 124 (pg 457) specifically says that this transition isn't
allowed. This is a clarification from earlier versions of the IB
spec, which were ambiguous in this area and suggested that the RESET
to ERR transition was allowed.
Fix up the qp_state_table[] to make RESET->ERR not allowed.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
ConnectX HCAs support the IB_MGMT_CLASS_CONG_MGMT management class, so
process MADs of this class through the MAD_IFC firmware command.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
ConnectX returns the max message size it supports through the
QUERY_DEV_CAP firmware command. When modifying a QP to RTR, the max
message size for the QP must be specified. This value must not exceed
the value declared through QUERY_DEV_CAP. The current code ignores
the max allowed size and unconditionally sets the value to 2^31. This
patch sets all QPs to the max value allowed as returned from firmware.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
- set IB_DEVICE_MEM_MGT_EXTENSIONS capability bit if fw supports it.
- set max_fast_reg_page_list_len device attribute.
- add iwch_alloc_fast_reg_mr function.
- add iwch_alloc_fastreg_pbl
- add iwch_free_fastreg_pbl
- adjust the WQ depth for kernel mode work queues to account for
fastreg possibly taking 2 WR slots.
- add fastreg_mr work request support.
- add local_inv work request support.
- add send_with_inv and send_with_se_inv work request support.
- removed useless duplicate enums/defines for TPT/MW/MR stuff.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
This patch adds support for the IB "base memory management extension"
(BMME) and the equivalent iWARP operations (which the iWARP verbs
mandates all devices must implement). The new operations are:
- Allocate an ib_mr for use in fast register work requests.
- Allocate/free a physical buffer lists for use in fast register work
requests. This allows device drivers to allocate this memory as
needed for use in posting send requests (eg via dma_alloc_coherent).
- New send queue work requests:
* send with remote invalidate
* fast register memory region
* local invalidate memory region
* RDMA read with invalidate local memory region (iWARP only)
Consumer interface details:
- A new device capability flag IB_DEVICE_MEM_MGT_EXTENSIONS is added
to indicate device support for these features.
- New send work request opcodes IB_WR_FAST_REG_MR, IB_WR_LOCAL_INV,
IB_WR_RDMA_READ_WITH_INV are added.
- A new consumer API function, ib_alloc_mr() is added to allocate
fast register memory regions.
- New consumer API functions, ib_alloc_fast_reg_page_list() and
ib_free_fast_reg_page_list() are added to allocate and free
device-specific memory for fast registration page lists.
- A new consumer API function, ib_update_fast_reg_key(), is added to
allow the key portion of the R_Key and L_Key of a fast registration
MR to be updated. Consumers call this if desired before posting
a IB_WR_FAST_REG_MR work request.
Consumers can use this as follows:
- MR is allocated with ib_alloc_mr().
- Page list memory is allocated with ib_alloc_fast_reg_page_list().
- MR R_Key/L_Key "key" field is updated with ib_update_fast_reg_key().
- MR made VALID and bound to a specific page list via
ib_post_send(IB_WR_FAST_REG_MR)
- MR made INVALID via ib_post_send(IB_WR_LOCAL_INV),
ib_post_send(IB_WR_RDMA_READ_WITH_INV) or an incoming send with
invalidate operation.
- MR is deallocated with ib_dereg_mr()
- page lists dealloced via ib_free_fast_reg_page_list().
Applications can allocate a fast register MR once, and then can
repeatedly bind the MR to different physical block lists (PBLs) via
posting work requests to a send queue (SQ). For each outstanding
MR-to-PBL binding in the SQ pipe, a fast_reg_page_list needs to be
allocated (the fast_reg_page_list is owned by the low-level driver
from the consumer posting a work request until the request completes).
Thus pipelining can be achieved while still allowing device-specific
page_list processing.
The 32-bit fast register memory key/STag is composed of a 24-bit index
and an 8-bit key. The application can change the key each time it
fast registers thus allowing more control over the peer's use of the
key/STag (ie it can effectively be changed each time the rkey is
rebound to a page list).
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The connected mode implementation in the IPoIB driver has a large
overhead in the way SKBs are handled in the receive flow. It usually
allocates an SKB with as big as was used in the currently received SKB
and moves unused fragments from the old SKB to the new one. This
involves a loop on all the remaining fragments and incurs overhead on
the CPU. This patch, for small SKBs, allocates an SKB just large
enough to contain the received data and copies to it the data from the
received SKB. The newly allocated SKB is passed to the stack and the
old SKB is reposted.
When running netperf, UDP small messages, without this pach I get:
UDP UNIDIRECTIONAL SEND TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
14.4.3.178 (14.4.3.178) port 0 AF_INET
Socket Message Elapsed Messages
Size Size Time Okay Errors Throughput
bytes bytes secs # # 10^6bits/sec
114688 128 10.00 5142034 0 526.31
114688 10.00 1130489 115.71
With this patch I get both send and receive at ~315 mbps.
The reason that send performance actually slows down is as follows:
When using this patch, the overhead of the CPU for handling RX packets
is dramatically reduced. As a result, we do not experience RNR NAK
messages from the receiver which cause the connection to be closed and
reopened again; when the patch is not used, the receiver cannot handle
the packets fast enough so there is less time to post new buffers and
hence the mentioned RNR NACKs. So what happens is that the
application *thinks* it posted a certain number of packets for
transmission but these packets are flushed and do not really get
transmitted. Since the connection gets opened and closed many times,
each time netperf gets the CPU time that otherwise would have been
given to IPoIB to actually transmit the packets. This can be verified
when looking at the port counters -- the output of ifconfig and the
oputput of netperf (this is for the case without the patch):
tx packets
==========
port counter: 1,543,996
ifconfig: 1,581,426
netperf: 5,142,034
rx packets
==========
netperf 1,1304,089
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
The idea is that for QPs with fixed size work requests (eg selective
signaling QPs), before stamping the WQE, we read the value of the DS
field, which gives the effective size of the descriptor as used in the
previous post. Then we stamp only that area, since the rest of the
descriptor is already stamped.
When initializing the send queue buffer, make sure the DS field is
initialized to the max descriptor size so that the subsequent stamping
will be done on the entire descriptor area.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
This patch solves a race that occurs after an event occurs that causes
the SA query module to flush its SM address handle (AH). When SM AH
becomes invalid and needs an update it is handled by the global
workqueue. On the other hand this event is also handled in the IPoIB
driver by queuing work in the ipoib_workqueue that does multicast
joins. Although queuing is in the right order, it is done to 2
different workqueues and so there is no guarantee that the first to be
queued is the first to be executed.
This causes a problem because IPoIB may end up sending an request to
the old SM, which will take a long time to time out (since the old SM
is gone); this leads to a much longer than necessary interruption in
multicast traffer.
The patch sets the SA query module's SM AH to NULL when the event
occurs, and until update_sm_ah() is done, any request that needs sm_ah
fails with -EAGAIN return status.
For consumers, the patch doesn't make things worse. Before the patch,
MADs are sent to the wrong SM so the request gets lost. Consumers can
be improved if they examine the return code and respond to EAGAIN
properly but even without an improvement the situation is not getting
worse.
Signed-off-by: Moni Levy <monil@voltaire.com>
Signed-off-by: Moni Shoua <monis@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The license text for several files references a third software license
that was inadvertently copied in. Update the license to what was
intended. This update was based on a request from HP.
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Remove an explicit memset(..., 0, ...) of a 'listener' structure
allocated with kzalloc().
Signed-off-by: Christophe Jaillet <christophe.jaillet@wanadoo.fr>
Acked-by: Faisal Latif <faisal@neteffect.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The SRP initiator is currently using ib_find_cached_pkey() and
ib_get_cached_gid() in situations where the uncached ib_find_pkey()
and ib_query_gid() functions serve just as well: sleeping is allowed
and performance is not an issue. Since we want to eliminate the
cached operations in the long term, convert SRP to use the uncached
variants.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Remove explicit lock_kernel() calls and document why the code is safe.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Jonathan Corbet <corbet@lwn.net>
The change to iwch_provider.c in commit f4e91eb4 ("IB: convert struct
class_device to struct device") undid the fix done in commit 7f049f2f
("RDMA/cxgb3: Hold rtnl_lock() around ethtool get_drvinfo call"). It
removed the calls to rtnl_lock() that serialized the iw_cxgb3 ethtool
ops calls into the cxgb3 driver. This locking is needed to avoid
messing up the internal state of the cxgb3 driver.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Remove explicit lock_kernel() calls and document why the code is safe.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Jonathan Corbet <corbet@lwn.net>
Current memfree FW has a bug which in some cases, assumes that ICM
pages passed to it are cleared. This patch uses __GFP_ZERO to
allocate all ICM pages passed to the FW. Once firmware with a fix is
released, we can make the workaround conditional on firmware version.
This fixes the bug reported by Arthur Kepner <akepner@sgi.com> here:
http://lists.openfabrics.org/pipermail/general/2008-May/050026.html
Cc: <stable@kernel.org>
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
[ Rewritten to be a one-liner using __GFP_ZERO instead of vmap()ing
ICM memory and memset()ing it to 0. - Roland ]
Signed-off-by: Roland Dreier <rolandd@cisco.com>
All of the open() functions which don't need the BKL on their face may
still depend on its acquisition to serialize opens against driver
initialization. So make those functions acquire then release the BKL to be
on the safe side.
Signed-off-by: Jonathan Corbet <corbet@lwn.net>
This documents the fact that somebody looked at the relevant open()
functions and concluded that, due to their trivial nature, no locking was
needed.
Signed-off-by: Jonathan Corbet <corbet@lwn.net>
