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>
Commit 1ae5c187 ("IB/uverbs: Don't store struct file * for event
files") changed the way that closed files are handled in the uverbs
code. However, after the conversion, is_closed flag is checked
incorrectly in ib_uverbs_async_handler(). As a result, no async
events are ever passed to applications.
Found by: Ronni Zimmerman <ronniz@mellanox.co.il>
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
nes_reg_user_mr() should fail if page_count becomes >= 1024 * 512
rather than just testing for strict >, because page_count is
essentially used as an index into an array with 1024 * 512 entries, so
allowing the loop to continue with page_count == 1024 * 512 means that
memory after the end of the array is corrupted. This leads to a crash
triggerable by a userspace application that requests registration of a
too-big region.
Also get rid of the call to pci_free_consistent() here to avoid
corrupting state with a double free, since the same memory will be
freed in the code jumped to at reg_user_mr_err.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
In 2.6.26, we added some support for send with invalidate work
requests, including a device capability flag to indicate whether a
device supports such requests. However, the support was incomplete:
the completion structure was not extended with a field for the key
contained in incoming send with invalidate requests.
Full support for memory management extensions (send with invalidate,
local invalidate, fast register through a send queue, etc) is planned
for 2.6.27. Since send with invalidate is not very useful by itself,
just remove the IB_DEVICE_SEND_W_INV bit before the 2.6.26 final
release; we will add an IB_DEVICE_MEM_MGT_EXTENSIONS bit in 2.6.27,
which makes things simpler for applications, since they will not have
quite as confusing an array of fine-grained bits to check.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
On a 64-bit architecture, if ib_umem_get() is called with a size value
that is so big that npages is negative when cast to int, then the
length of the page list passed to get_user_pages(), namely
min_t(int, npages, PAGE_SIZE / sizeof (struct page *))
will be negative, and get_user_pages() will immediately return 0 (at
least since 900cf086, "Be more robust about bad arguments in
get_user_pages()"). This leads to an infinite loop in ib_umem_get(),
since the code boils down to:
while (npages) {
ret = get_user_pages(...);
npages -= ret;
}
Fix this by taking the minimum as unsigned longs, so that the value of
npages is never truncated.
The impact of this bug isn't too severe, since the value of npages is
checked against RLIMIT_MEMLOCK, so a process would need to have an
astronomical limit or have CAP_IPC_LOCK to be able to trigger this,
and such a process could already cause lots of mischief. But it does
let buggy userspace code cause a kernel lock-up; for example I hit
this with code that passes a negative value into a memory registartion
function where it is promoted to a huge u64 value.
Cc: <stable@kernel.org>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
SM/SMA traps received by the ipath driver should be forwarded to the
SM if it is running on the host. The ib_ipath driver was incorrectly
replying with "bad method."
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The driver supports a few features (RNR NAK, port active event, SRQ
resize) that were not reported in the device capability flags. This
patch fixes that.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Gabriel C <nix.or.die@googlemail.com> pointed out that when the x86
bitops are updated to operate on unsigned long, the code in
sdma_abort_task() will produce warnings:
drivers/infiniband/hw/ipath/ipath_sdma.c: In function 'sdma_abort_task':
drivers/infiniband/hw/ipath/ipath_sdma.c:267: warning: passing argument 2 of 'constant_test_bit' from incompatible pointer type
and so on, because it uses test_bit() to operation on a u64 value
(returned by ipath_read_kref64() for a hardware register).
Fix up these warnings by converting the test_bit() operations to &ing
with appropriate symbolic defines of the bits within the hardware
register. This has the benign side-effect of making the code more
self-documenting as well.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland/infiniband:
IB/mad: Fix kernel crash when .process_mad() returns SUCCESS|CONSUMED
IPoIB: Test for NULL broadcast object in ipiob_mcast_join_finish()
MAINTAINERS: Add cxgb3 and iw_cxgb3 NIC and iWARP driver entries
IB/mlx4: Fix creation of kernel QP with max number of send s/g entries
IB/mthca: Fix max_sge value returned by query_device
RDMA/cxgb3: Fix uninitialized variable warning in iwch_post_send()
IB/mlx4: Fix uninitialized-var warning in mlx4_ib_post_send()
IB/ipath: Fix UC receive completion opcode for RDMA WRITE with immediate
IB/ipath: Fix printk format for ipath_sdma_status
If a low-level driver returns IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED,
handle_outgoing_dr_smp() doesn't clean up properly. The fix is to
kfree the local data and break, rather than falling through. This was
observed with the ipath driver, but could happen with any driver.
This fixes <https://bugs.openfabrics.org/show_bug.cgi?id=1027>.
Signed-off-by: Dave Olson <dave.olson@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
We saw a kernel oops in our regression testing when a multicast "join
finish" occurred just after the interface was -- this is
<https://bugs.openfabrics.org/show_bug.cgi?id=1040>. The test
randomly causes the HCA physical port to go down then up.
The cause of this is that ipoib_mcast_join_finish() processing happen
just after ipoib_mcast_dev_flush() was invoked (in which case the
broadcast pointer is NULL). This patch tests for and handles the case
where priv->broadcast is NULL.
Cc: <stable@kernel.org>
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When creating a kernel QP where the consumer asked for a send queue
with lots of scatter/gater entries, set_kernel_sq_size() incorrectly
returned an error if the send queue stride is larger than the
hardware's maximum send work request descriptor size. This is not a
problem; the only issue is to make sure that the actual descriptors
used do not overflow the maximum descriptor size, so check this instead.
Clamp the returned max_send_sge value to be no bigger than what
query_device returns for the max_sge to avoid confusing hapless users,
even if the hardware is capable of handling a few more s/g entries.
This bug caused NFS/RDMA mounts to fail when the server adapter used
the mlx4 driver.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
There is a race from when a device is created with device_create() and
then the drvdata is set with a call to dev_set_drvdata() in which a
sysfs file could be open, yet the drvdata will be NULL, causing all
sorts of bad things to happen.
This patch fixes the problem by using the new function,
device_create_drvdata().
Cc: Kay Sievers <kay.sievers@vrfy.org>
Reviewed-by: Roland Dreier <rolandd@cisco.com>
Cc: Sean Hefty <sean.hefty@intel.com>
Cc: Hal Rosenstock <hal.rosenstock@gmail.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
The mthca driver returns the maximum number of scatter/gather entries
returned by the firmware as the max_sge value when device properties
are queried. However, the firmware also reports a limit on the
maximum descriptor size allowed, and because mthca takes into account
the worst case send request overhead when checking whether to allow a
QP to be created, the largest number of scatter/gather entries that
can be used with mthca may be limited by the maximum descriptor size
rather than just by the actual s/g entry limit.
This means that applications cannot actually create QPs with
max_send_sge equal to the limit returned by ib_query_device(). Fix
this by checking if the maximum descriptor size imposes a lower limit
and if so returning that lower limit.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/cxgb3/iwch_qp.c: In function 'iwch_post_send':
drivers/infiniband/hw/cxgb3/iwch_qp.c:232: warning: 't3_wr_flit_cnt' may be used uninitialized in this function
This is what akpm describes as "the dopey
gcc-doesn't-know-that-foo(&var)-writes-to-var problem."
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
drivers/infiniband/hw/mlx4/qp.c: In function 'mlx4_ib_post_send':
drivers/infiniband/hw/mlx4/qp.c:1460: warning: 'seglen' may be used uninitialized in this function
This is the dopey gcc-doesn't-know-that-foo(&var)-writes-to-var problem.
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When I fixed the RC receive completion opcode in 2bfc8e9e ("IB/ipath:
Return the correct opcode for RDMA WRITE with immediate"), I forgot to
fix UC, which had the same problem for RDMA write with immediate
returning the wrong opcode.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Commit f018c7e1 ("IB/ipath: Change ipath_devdata.ipath_sdma_status to be
unsigned long") changed ipath_sdma_status to be unsigned long, but left
a few debug messages that printed it out with a %016llx format, which
generates the warnings
drivers/infiniband/hw/ipath/ipath_sdma.c:348: warning: format '%016llx' expects type 'long long unsigned int', but argument 3 has type 'long unsigned int'
drivers/infiniband/hw/ipath/ipath_sdma.c:618: warning: format '%016llx' expects type 'long long unsigned int', but argument 3 has type 'long unsigned int'
Fix this by changing the format used to print out the value to %08lx
(8 hex digits are now sufficient, because the highest bit used is 31).
Warnings reported by Randy Dunlap <randy.dunlap@oracle.com>.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
cxio_flush_sq() was failing to wrap around the software send queue
causing garbage completion entries on a flush operation.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Andrew Morton <akpm@linux-foundation.org> pointed out that bitops
should take an unsigned long * arg. However, the ipath driver was
doing bitops on struct ipath_devdata.ipath_sdma_status, which is u64.
Change this member to unsigned long to avoid tons of warnings when x86
fixes the bitops to take unsigned long * instead of void *.
Also, change the IPATH_SDMA_RUNNING and IPATH_SDMA_SHUTDOWN bit
numbers to 30 and 31 (instead of 62 and 63) so that we're not setting
another booby trap for someone who tries to make ipath work on a
32-bit architecture.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The official reason is "with the presence of pid namespaces in the
kernel using pid_t-s inside one is no longer safe."
But the reason I fix this right now is the following:
About a month ago (when 2.6.25 was not yet released) there still was a
one last caller of a to-be-deprecated-soon function find_pid() - the
kill_proc() function, which in turn was only used by nfs callback
code.
During the last merge window, this last caller was finally eliminated
by some NFS patch(es) and I was about to finally kill this kill_proc()
and find_pid(), but found, that I was late and the kill_proc is now
called from the ipath driver since commit 58411d1c ("IB/ipath: Head of
Line blocking vs forward progress of user apps").
So here's a patch that fixes this code to use struct pid * and (!)
the kill_pid routine.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
If an out of sequence RDMA read response middle or last packet is
received, we should only resend the RDMA read request on the first
out of sequence packet and drop subsequent out of sequence packets
otherwise, we get "too many retries".
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The send DMA hardware queue voided a number of prior assumptions about
when a send is complete which led to completions being generated out of
order. There were also a number of locking issues when switching the QP
to the error or reset states, and we implement the IB_QPS_SQD state.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When errors are detected in RC, the QP should transition to the
IB_QPS_ERR state, not the IB_QPS_SQE state. Also, when the error is on
the responder side, the receive work completion error was incorrect
(remote vs. local).
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fix some bugs with the max_aggr module parameter added with LRO support:
- The module parameter value ignored and not actually used to set
lro_mgr.max_aggr.
- MODULE_PARM_DESC had a typo "_mro_" instead of "_lro_" so it didn't
end up describing the actual module parameter.
- The nes_lro_max_aggr variable was declared as unsigned, but the
module_param line said "int" instead of "uint" for the type.
- The default value for the parameter was stuck in the permissions
field of module_param, which led to nonsensical permissions for the
file under /sys/module/iw_nes/param.
- The parameter was used in only one file but defined in another, which
led to the variable being global for no good reason. Move everything
related to the parameter to the file nes_hw.c where it is actually
used.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
This is necessary because, in a multicore environment, a race between
uverbs async handler and destroy QP could occur.
Signed-off-by: Stefan Roscher <stefan.roscher at de.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
What's fixed:
in ipath_cancel_sends()
We need to unconditionally set ABORTING. So, swap the tests
so the set_bit() isn't shadowed by the &&.
If we've disarmed the piobufs, then we need to unconditionally
set DISARMED. So, move it out from the overly protective if
at the bottom.
in sdma_abort_task()
Abort_task was written knowing that the SDMA engine would always
be reset (and restarted) on error. A recent change broke that
fundamental assumption by taking the restart portion and making
it conditional on a link status change. But, SDMA can go boom
without a link status change in some conditions.
Signed-off-by: John Gregor <john.gregor@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Now that we always use PIO for vl15 on 7220, we could get stuck forever
if we happened to run out of PIO buffers from the verbs code, because
the setup code wouldn't run; the interrupt was also ignored if SDMA was
supported. We also have to reduce the pio update threshold if we have
fewer kernel buffers than the existing threshold.
Clean up the initialization a bit to get ordering safer and more
sensible, and use the existing ipath_chg_kernavail call to do init,
rather than doing it separately.
Drop unnecessary clearing of pio buffer on pio parity error.
Drop incorrect updating of pioavailshadow when exitting freeze mode
(software state may not match chip state if buffer has been allocated
and not yet written).
If we couldn't get a kernel buffer for a while, make sure we are
in sync with hardware, mainly to handle the exitting freeze case.
Signed-off-by: Dave Olson <dave.olson@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The loop in ipath_kreceive() that processes packets increments the
loop-index 'i' once too often, because the exit condition does not
depend on it, and is checked after the increment. By adding a check for
!last to the iterator in the for loop, we correct that in a way that is
not so likely to be re-broken by changes in the loop body.
Signed-off-by: Michael Albaugh <micheal.albaugh@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
This patch fixes a bug in the RC responder which generates a completion
entry with the wrong opcode when an RDMA WRITE with immediate is received.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The semantics of cancel_sends changed, but the code using it was missed.
Don't leave sends and pioavail updates disabled, and add a comment as to
why the force update is needed.
Signed-off-by: Dave Olson <dave.olson@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
If a send work request has immediate errors and is not put on the
send queue, we shouldn't update any of the QP state.
The increment of the SSN wasn't obeying this.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
We warn about prototype chips, but the function that checks for
support is also called as a result of a get_portinfo request, which
can clutter the logs.
Restrict warning to only appear during initialization.
Signed-off-by: Michael Albaugh <michael.albaugh@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
