Er...no. It will lose this event. Depending on the event...the carnage
varies. We'll take a look at this.
This behavior is consistent with the Infiniband CM (see
drivers/infiniband/core/cm.c function cm_recv_handler()). But I think
we should at least log an error because a lost event will
Assuming minimal hard-coding of which methods are requests, a client would
drop
only about 1 MAD per method during start-up.
Is this only the new methods which are not hard coded ? Would this
invoke a timeout (and hopefully retry) ?
We can hard-code existing methods to avoid this problem. So
Is the only downside of a larger timeout that potentially more memory
accumulates (until the timeout occurs) before it is freed ?
This is the only one that I can think of. Can anyone think of others?
- Sean
___
openib-general mailing list
To clarify the motivation more, a question to answer is if we ignore iWarp
completely, does it make sense to provide a higher level communication manager
for IB. I believe that it does, especially for userspace applications. This
lets us leverage existing name services, ipoib, and provides an
Sean Hefty wrote:
I'd like to propose that the MAD layer detect duplicate requests. After a
request MAD has been handed to a client, its context would be maintained until
the user calls ib_free_recv_mad(), allowing duplicate requests to be
discarded.
{snip}
Finally, a way would need
I started thinking about the kill ib_uverbs_idr_mutex problem, and I
realized that there are actually some interesting issues there (as
described in the comment at the top of uverbs_cmd.c). In fact I ended
up coding the solution below. This passes some basic tests but it
could probably use
Sean Hefty wrote:
4. Modify umad to learn which requests generate responses, by examining
response
MADs. When a response is sent, umad would mark which method the response is
for
by flipping the R-bit. Based on the algorithm, this could result in losing
responses the first time
Tziporet - Bryan has confirmed that with the patches you've copied,
things should work correctly. We've been testing with our version, but
I really want to test on the OFED-1.0 version that you've built. Can you
send us a pointer to it?
How can you go from an RC6 that doesn't build to a 1.0
Hal Rosenstock wrote:
This brings up a concern. There doesn't seem to be a limit to the number of
received MADs that can be queued for a user. Should we have such a limit?
How are MADs counted ? Is a multisegment MAD 1 MAD or multiple MADs ? If
the latter, it seems problematic to limit
@@ -58,6 +58,8 @@ enum {
RDMA_USER_CM_CMD_GET_EVENT,
RDMA_USER_CM_CMD_GET_OPTION,
RDMA_USER_CM_CMD_SET_OPTION,
+ RDMA_USER_CM_CMD_JOIN_MCAST,
+ RDMA_USER_CM_CMD_LEAVE_MCAST,
RDMA_USER_CM_CMD_GET_DST_ATTR
};
I think this changes the exported ABI by changing the
Sean, is anyone looking at this? If not, given that Jack's approach does not
touch ABI or API, might it make sense to merge Jack's patch after all and use
that as a starting point? With current code in 2.6.17 large RMPPs often get
aborted because of the problem of the duplicates. On the other
I am planning to use RDMA CM for multicast functionality. It would be
great if you can point me to a simple multicast test program using
RDMA CM?
There is a userspace test program (mckey) that will be available, but has not
been posted yet. (A kernel test program would look fairly similar.) I
dev_addr-broadcast + 4/dev_addr-src_dev_addr + 4 may not be naturally
aligned,
so casting this pointer to structure type may cause compiler to generate
incorrect code.
Thanks - I'll update this.
- Sean
___
openib-general mailing list
mad.c calls ib_req_notify_cq on hotplug event in ib_mad_port_start, after QPs
are attached to a CQ. Since this function does not poll the CQ, if sufficient
number of MADs arrive at the QP before ib_req_notify_cq is called, RQ might get
empty and no completion events will ever be generated.
This
Hal Rosenstock wrote:
Note the MGRPs are MGIDs and switches are programmed with MLIDs and
these can be 1:1 or many:1 depending on the implementation. Most do not
do the many:1 but this is allowed by the spec. Also, note that switches
know nothing about the groups themselves (only MLIDs and which
Michael S. Tsirkin wrote:
These should eliminate any races with ipoib leaving,
then quickly re-joining a group as a result of an event.
Is there a chance this will fix the crashes me and Or were seeing?
It shouldn't. The race that I was referring to only involved whether or not a
MAD is
Sean Hefty wrote:
Ipoib already checks for events that require rejoining multicast groups.
We just need to add code to handle (i.e. ignore) multicast group reset
notifications.
Roland,
Any issue committing this?
- Sean
___
openib-general mailing
Hal Rosenstock wrote:
The other issue is whether you trust the state of the network or not
when the SM comes up. That's sometimes a dangerous proposition.
I considered this, but I think there's a difference between trusting one of the
systems on the network, versus the network as a whole.
groups.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
___
openib-general mailing list
openib-general@openib.org
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
Extract the MGID used by ipoib for broadcast traffic from the device
address.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
This will be used to get the MCMemberRecord for the ipoib broadcast group.
--- svn3/gen2/trunk/src/linux-kernel/infiniband/include/rdma/ib_addr.h
2006-05-25 11:18
Add an API to allow retrieving an MCMemberRecord from the local cache
based on an MGID.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
This allows an existing MCMemberRecord to be used as a template for
creating other multicast groups.
--- svn3/gen2/trunk/src/linux-kernel/infiniband/include
Export a call to initialize an ib_ah_attr structure based on an
MCMemberRecord returned from a multicast join request.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
--- svn3/gen2/trunk/src/linux-kernel/infiniband/include/rdma/ib_sa.h
2006-06-06 15:21:05.0 -0700
+++ svn/gen2/trunk
Expose multicast abstraction through the CMA to userspace.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
--- svn3/gen2/trunk/src/linux-kernel/infiniband/include/rdma/rdma_user_cm.h
2006-06-06 16:53:46.0 -0700
+++ svn/gen2/trunk/src/linux-kernel/infiniband/include/rdma
Add IB multicast abstraction to the CMA.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
--- svn3/gen2/trunk/src/linux-kernel/infiniband/include/rdma/rdma_cm.h
2006-06-06 16:53:56.0 -0700
+++ svn/gen2/trunk/src/linux-kernel/infiniband/include/rdma/rdma_cm.h
2006-06-02 10:22
If this comment is directed at client reregister mechanism, you should
note that when this was brought up there was resistance to it based on
the recommendation (probably not a strong enough word for this) that SMs
be redundant in the subnet. There was a fair bit of anecdotal evidence
that this
The following patch series adds support for UD QPs to userspace through the RDMA
CM. UD QPs are referenced by an IP address, UDP port number. The RDMA CM
abstracts SIDR for Infiniband clients.
Roland,
Do you see any issues with this patch series or the related userspace changes?
There's a
Roland Dreier wrote:
I haven't looked too carefully yet.
What's the motivation? It seems strange to put an IB-only transport
into the RDMA CM -- iWARP can't handle datagrams, can it?
This allows using the address translation to locate the remote service. The
RDMA CM also provides an IP
Greg Lindahl wrote:
Isn't this a quality of implementation issue? It's hard to imagine a
SM author not realizing this is a good thing to do.
I don't know if any SM implementation actually does this today. I think that
all break all multicast groups.
If it was in the standard, how would
An SM, upon becoming the master, shall respect all existing communication in
the fabric, where possible.
To me, where possible doesn't sound like an appropriate language for a
compliance statement. Is there precedent for this in IB spec?
I was trying to express a concept, not formulate exact
Hal Rosenstock wrote:
2. There is lazy deletion of MC groups allowed so the reclamation may be
difficult.
I'm not familiar with the switch programming. Does the SM set the entire
MulticastForwardingTable for a switch every time a new group is created, or a
new member joins? If the SM loses
that
requires clients to rejoin a multicast group, the active members are moved
into an error state, and the clients are notified of a network reset error.
The group is then reset to force additional join requests to generate requests
to the SA.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Hal, can
Ipoib already checks for events that require rejoining multicast groups.
We just need to add code to handle (i.e. ignore) multicast group reset
notifications.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Ignoring the callback is a simple fix. I didn't try to see what it would
take to have
Roland Dreier wrote:
Looks like the same crash mst saw related to the multicast module
being unloaded and then having sa call back into it. One small clue:
esi: f38a5bec edi: f38a5bf4 ebp: fffc esp: f599be60
ebp is -4, which is -EINTR. So this may be a callback from sa_query's
Hal Rosenstock wrote:
This
leads to a race where NonMembers and SendOnlyNonMembers will fail to re-join
until one of the FullMembers joins.
Might also be true with joins (not creates) from FullMembers too. I
would presume in such cases, the join would be retried. SendOnlyMembers
(at least
I might be missing your point but UD is unreliable so the sends can be
dropped. The delay/retry is to make sure the join does occur,
This is different than a dropped request or reply. In this case, the receiver
gets a reply, but it will be a failure from the SA to join the group. For
example, a
Does anyone know if the following multicast configurations have been tested?
1. Receiving messages on the same port that they were sent, but on a different
QP.
2. Receiving messages on multiple QPs on the same port.
- Sean
___
openib-general mailing
I think it's fine. Should I queue it for 2.6.18 too?
That probably makes sense. I'll send a couple of svn revs that should be safe
to pull into 2.6.18 after committing this.
- Sean
___
openib-general mailing list
openib-general@openib.org
The following patch series adds support for UD QPs to userspace through the RDMA
CM. UD QPs are referenced by an IP address, UDP port number. The RDMA CM
abstracts SIDR for Infiniband clients.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
A subsequent patch series will add multicast handling
Record remote QP information returned from SIDR. Expose attributes through
a new API. This functionality is similar to the ib_cm_init_qp_attr()
routine that exists for RC QPs.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Index: core/cm.c
messages to
IB CM SIDR REQ messages.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Index: core/cma.c
===
--- core/cma.c (revision 7758)
+++ core/cma.c (working copy)
@@ -66,6 +66,7 @@ static DEFINE_MUTEX(lock);
static struct
Export the ib_copy_ah_attr_to_user() routine to allow copy ib_ah_attr
to userspace to support UD QPs.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Index: core/uverbs_marshall.c
===
--- core/uverbs_marshall.c (revision 7758
Export the RDMA CM's support of UD QPs to the userspace library.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
My intent is to bump the ABI version only once. The multicast patches
will not increment the ABI.
Index: core/ucma.c
Adds some helper functions to simplify using UD QPs.
Add new routines: ibv_init_ah_from_wc() and ibv_create_ah_from_wc()
to simplify UD QP communication.
Expose ibv_copy_ah_attr_from_kern to retrieve ibv_ah_attr from kernel for
a UD QP.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Index
Add support for UD QPs to the RDMA CM library, along with a goofy test program.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Index: include/rdma/rdma_cma_ib.h
===
--- include/rdma/rdma_cma_ib.h (revision 7743)
+++ include/rdma
I'm running into an issue trying to build libmthca.
During the ./configure step, I get:
checking size of long... configure: error: cannot compute sizeof (long), 77
Has anyone else run into this?
- Sean
___
openib-general mailing list
Sean Hefty wrote:
I'm running into an issue trying to build libmthca.
During the ./configure step, I get:
checking size of long... configure: error: cannot compute sizeof (long), 77
Has anyone else run into this?
Rebooting my system and rebuilding made this error go away.
- Sean
I just hit this too today. Inspecting the config log file revealed that
it could find libibverbs.so. I ran ldconfig, then reran autogen and
configure and it worked. Try that...
Thanks - I'll try that next time.
___
openib-general mailing list
Boyd R. Faulkner wrote:
The code currently in place seems to expect there to be a null element at the
end of the dev_list to trigger the end of the loop. ibv_get_device_list
does not provide such an entry, but the number of entries is
available. This patch retrieves that number and loops
Sean Hefty wrote:
Modify the CM to maintain the necessary information needed to send
to a UD QP after a user has performed SIDR. Expose the remote QPN,
remote QKey, and address handle attributes through the
ib_cm_init_qp_attr() routine, so that the information is available
from userspace
Eitan Zahavi wrote:
Well, this is a very old topic well discussed years ago. All credits to
Ashok Raj which you know better then me. The argument against the idea
for the SM to be the keeper of these registrations goes as follows:
Yes - and I still don't understand why this isn't a personal
Hal Rosenstock wrote:
I believe that this still works. Ipoib should leave all multicast groups, then
rejoin when an event occurs. As long as no other clients join the ipoib groups,
this should work.
Are you saying that IPoIB handles the event ? Does the multicast module
cooperate (in terms
Eitan Zahavi wrote:
Leonid just sent an example for a race that might happen if the SM is to
be the maintainer of the data.
The race Leonid mentioned is a client sending a request when the SM is down.
That request will fail, so there's no data for the SM to maintain for that node.
That's a
Eitan Zahavi wrote:
[EZ] The race is happening when the SM received the request and
responded but the other SMs or the file system did not fully stored that
registration and the SM crashed.
If the client received a response that the join was successful, then I consider
that an SM issue.
The
Mainly nits...
Steve Wise wrote:
-static int copy_addr(struct rdma_dev_addr *dev_addr, struct net_device *dev,
+int copy_addr(struct rdma_dev_addr *dev_addr, struct net_device *dev,
unsigned char *dst_dev_addr)
Might want to rename this to something like rdma_copy_addr if
Hal Rosenstock wrote:
Send-only joins is another case. These are full member joins (JoinState
1) to groups which are not yet created so they fail.
I see the problem, and checked in a fix. I forgot to record the last join
operation that was initiated, so that it could be failed on an error.
I think that there are some issues that would need to be worked out, but in
general I'm in favor of trying to do something here.
Currently, this is not something that can be implemented by ULP on top of
CMA, because returning error from REQ will result in reject rather than REQ
drop.
A generic
This approach would affect all ULPs, however. For example, no SDP imlementation
that I know of retries after a REJ - so this approach won't be interoperable.
And AFAIK SDP spec already interprets reject as connection refused.
There's no provision I cansee in SDP spec for retries on specific
reject
Michael S. Tsirkin wrote:
I'm still looking at isolating this failure. I'd like to understand the new
code better, however. What prevents ipoib_mcast_leave and later
ipoib_mcast_free from being called on an mcast that has an outstanding query?
We used to have a completion to signal that but
Hal Rosenstock wrote:
Is client reregister handled properly by the multicast module ?
Can you clarify what you mean by this? Are you asking about re-sending join
requests based on some event?
- Sean
___
openib-general mailing list
On Tue, 2006-05-30 at 17:33, Sean Hefty wrote:
Hal Rosenstock wrote:
Is client reregister handled properly by the multicast module ?
Can you clarify what you mean by this? Are you asking about re-sending join
requests based on some event?
Yes; when the SM sends a Set PortInfo
branch, which will bring the ucm up to date.
---
Convert semaphore in ib_ucm_file to a real mutex.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
diff --git a/drivers/infiniband/core/ucm.c b/drivers/infiniband/core/ucm.c
index fefc9b6..67caf36 100644
--- a/drivers/infiniband/core/ucm.c
+++ b
Michael S. Tsirkin wrote:
Fix CMA for loopback configurations: in cma_bind_loopback, make sure sa query is
performed from an active port.
Thanks! - committed in 7502.
- Sean
___
openib-general mailing list
openib-general@openib.org
.
Signed-off-by: Ali Ayoub [EMAIL PROTECTED]
Signed-off-by: Michael S. Tsirkin [EMAIL PROTECTED]
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
Index: openib_gen2/drivers/infiniband/core/cma.c
===
--- openib_gen2.orig/drivers/infiniband
Destroy_workqueue already does flush_workqueue.
Signed-off-by: Michael S. Tsirkin [EMAIL PROTECTED]
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
diff --git a/drivers/infiniband/core/cm.c b/drivers/infiniband/core/cm.c
index 490fd03..1c7463b 100644
--- a/drivers/infiniband/core/cm.c
+++ b
Michael S. Tsirkin wrote:
What modules were unloaded and being unloaded when this occurred?
I think this can be seen in oops.
I'm guessing that ipoib and ib_multicast were unloaded, and the crash occurred
unloading ib_sa. Is this correct?
- Sean
Michael S. Tsirkin wrote:
Ali here says he's just bringing the ib0 up, and then unloads the module.
Linux is sending broadcasts out all the time, so ...
Were ipoib and ib_multicast already loaded, and he was just doing an ifconfig?
Did he unload both ipoib and ib_multicast, or just ipoib?
-
Michael S. Tsirkin wrote:
The last trunk build causes kernel oops on 2.6.16 while restarting the driver.
(the previous build -rev 7422- works fine)
Note that ipoib moved to ib_multicast in rev 7401.
May 24 16:00:40 sw037 kernel: Modules linked in: ib_sa ib_uverbs ib_umad
ib_mthca ib_mad
Sean Hefty wrote:
Reviewing the code, the multicast module should cancel all SA queries
and wait for them to complete before unloading. (Even if it didn't
perform the cancel, it should still wait for any outstanding SA query to
complete.)
As a thought, is there any chance this crash
Eric Barton wrote:
I just tested your patch and checked that it prevents the double
DISCONNECT event callback (it does :).
Thanks for testing this. I've committed this patch to svn.
- Sean
___
openib-general mailing list
openib-general@openib.org
Jack Morgenstein wrote:
Prevent opening multiple RMPP MAD transaction sessions at responder side
with the same TID, GID/LID, class.
Could happen if RMPP requests are retried while response is in progress.
My preference for handling this is to detect and discard duplicate requests, and
verify
Roland Dreier wrote:
Sean Thanks for testing this. I've committed this patch to svn.
Should this be merged into what I have queued for 2.6.18?
I think so. I was going to send another update later today that included the
patches that Michael wanted for SDP support as well. (I didn't
keshetti mahesh wrote:
i need to develop a kernel utility capable of RDMA read/write operations
i have seen example utilities under
svn/gen2/utils/src/linux-kernel/infiniband/util/ tree.
where can i find the documentation related to them?
There's no documentation for those utilities - just
Thanks both. I'll solve this by adding htonl/ntohl, only on the
VAPI side
Since VAPI is wanting the data in host order, while openib uses network order,
it makes more sense to me to do the swapping on the openib side.
- Sean
___
openib-general mailing
Fabian Tillier wrote:
It doesn't matter what VAPI wants - it's the application that matters.
If the application is using the immediate data for flags, you don't
need any swapping on the OpenIB side of things, and you can avoid the
swap altogether. While this makes the VAPI implementation less
Fabian Tillier wrote:
If you swap the flag constants in little endian systems so they're
always in network order - something the compiler can do for you - then
you're checking for a bit, and it is safe to treat the value in
network order always. The flag constant will be different in little
amit byron wrote:
o to make multiple connections using ib_send_cm_req()
i would have make connection requests using
ib_send_cm_req() with different service id. is this
sufficient, or i missed something? i can use the
same port number, correct?
You need one service ID per
Sean Hefty wrote:
Convert ipoib to make use of the new multicast module interface.
I've committed this patch to svn.
- Sean
___
openib-general mailing list
openib-general@openib.org
http://openib.org/mailman/listinfo/openib-general
To unsubscribe
while reviewing this i noticed that this documentation error:
ib_send_cm_sidr_rep - Sends a service ID resolution ***request*** to the
remote node.
which you can fix when committing this patch.
Thanks - I'll update the documentation as part of this patch.
- Sean
amit byron wrote:
o where can i find sample code to do rdma write? how to
setup rdma write read between two infiniband nodes?
There are some kernel tests in svn/gen2/utils/src/linux-kernel/infiniband/util.
See krping for RDMA read/write.
There are usermode tests in
Eric Barton wrote:
I'm using the rdam_cm API. I've seen it call my CM callback with
RDMA_CM_EVENT_DISCONNECTED twice. Is this a bug?
I would consider this a bug.
The problem is that the underlying IB CM is reporting two events: DREQ_RECEIVED,
followed by DREP_RECEIVED. The RDMA CM reports
the krping test module make use of rdma_cm* apis. will
the module work with ib_cm* api?
You would have to adapt the module to use
the ib_cm APIs. The posting of the work requests is the same, however.
should i using rdma_cm module or ib_cm module,
which is the standard?
Both are
Eric
Can you try this patch and let me know if it fixes your problem?
- Sean
---
Prevent generating duplicated DISCONNECT events.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Index: cma.c
===
--- cma.c (revision 7362
Thanks! Committed with only minor adjustment to spacing.
- Sean
___
openib-general mailing list
openib-general@openib.org
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
Sean Thanks! Committed with only minor adjustment to spacing. -
Should I add that commit to what I have queued for 2.6.18?
It shouldn't hurt, but isn't strictly needed. The changes are for SDP, and IPv6
support still requires more work. My personal vote would be yes, with the hope
that
The pkey is provided into a SIDR REQ in two places, once as a parameter,
and again in the path record. Remove the pkey as a parameter and always
use that given in the path record.
This change has no practical effect on ABI functionality.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Index
Expose a new call to initialize address handle attributes from a work
completion. This functionality is duplicated by both verbs and the CM.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Index: include/rdma/ib_verbs.h
calls: resolve_addr, resolve_route, and
connect. A server calls: listen and accept. Connect and accept
correspond to SIDR REQ / SIDR REP, respectively.
This patch introduces a new protocol for SIDR that is the same as that
used by the CMA for connection REQs.
Signed-off-by: Sean Hefty [EMAIL
Remove the pkey from the API for SIDR REQ. The pkey is provided
in the path record.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Index: include/infiniband/cm.h
===
--- include/infiniband/cm.h (revision 7020)
+++ include
And a patch to support UD QPs from userspace through the librdmacm.
Included in the patch is a test program.
Signed-off-by: Sean Hefty [EMAIL PROTECTED]
---
Index: include/rdma/rdma_cma_ib.h
===
--- include/rdma/rdma_cma_ib.h
Sean Hefty wrote:
+int ib_init_ah_from_path(struct ib_device *device, u8 port_num,
+struct ib_sa_path_rec *rec, struct ib_ah_attr *ah_attr)
+{
+ int ret;
+ u16 gid_index;
+
+ memset(ah_attr, 0, sizeof *ah_attr);
+ ah_attr-dlid = be16_to_cpu(rec
Any reason not to apply this?
Looks fine to apply be me.
- Sean
___
openib-general mailing list
openib-general@openib.org
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
Dotan Barak wrote:
Added checks to the return values of all of the functions that may fail
(in order to add this test to the regression system).
Thanks - applied with one minor change.
+ int rc;
Changed 'rc' to 'ret' to match the rest of the code.
- Sean
OK, I just tested and this works for me. Here's the SDP patch to do what you
described. The code actually got cleaner now: its convenient to get
different events on active versus passive side - previously I had
to check a flag to figure out what does ESTABLISHED mean.
I committed the CMA patch.
Please post generic questions to the openib
mailing list.
i have started working over infiniband recently
i want to develop a sample utility that would perform simple RDMA (read/write)
operations
There are some test applications that can be
used as a base. Are you wanting a
Michael S. Tsirkin wrote:
Sean, CMA currently does not support IPv6 addresses at all. Is that right?
This is correct. At best, there's some code in places to handle it.
However, while I don't have immediate need to make real IPv6 addressing to work,
some applications (notably Java) always
Ian Brown wrote:
Are there Etherenet NICS in the market (or will there be soon such nics)
such NICs which are RDMA nics (RDMA enabled NICs)?
And in case the answer is positive - does linux kernel has support
for such nics?
The Linux kernel does not have support for RNICs at this time, but
Thanks! - applied.
- Sean
___
openib-general mailing list
openib-general@openib.org
http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
OK. May we discuss the design/API for now?
Sounds good.
My understanding is an IPv4 socket should only listen on IPv4 requests,
while IPv6 socket should listen on both IPv4 and IPv6, unless
IPV6_V6ONLY is set.
Is that right? What will the API be for this? Maybe create_cm_id should get an
Rdma_create_id() already takes a struct sockaddr *, which has an address
family selector (sa_family) to define the contained address format. Why is
that one not sufficient?
Rdma_bind() and rdma_resolve_addr() take struct sockaddr *. Rdma_create_id()
only has an event handler, context, and port
Looking at rdma_listen(), the code I see checks for bound state before
proceeding to listen:
int rdma_listen(struct rdma_cm_id *id, int backlog)
{
struct rdma_id_private *id_priv;
int ret;
id_priv = container_of(id, struct rdma_id_private, id);
if
701 - 800 of 2277 matches
Mail list logo