Sean Hefty [Mon, 25 Mar 2013 21:20:33 +0000 (14:20 -0700)]
Call ucma_ib_init() during startup
Commit 272c3cc024d0e5854cbafa6c2f1e8560398a68d7 removed
automatic calls to ucma_ib_init(). It was added back for
rsocket calls, but not for the older rdma_cm calls. The
result is that ibacm support ends up being disabled. Fix
this by calling ucma_ib_init() where needed.
This problem was reported by Jens Domke
<domke.j.aa@m.titech.ac.jp>
Sean Hefty [Mon, 25 Mar 2013 21:20:33 +0000 (14:20 -0700)]
Call ucma_ib_init() during startup
ucma_ib_init() was removed from ucma_init() to avoid
recursive callback issues when using the librspreload
library over rsockets. Although rsockets calls ucma_ib_init,
other pieces of the code do not. The result is that
ibacm ends up being disabled unless rsockets has been
used. Fix this by calling ucma_ib_init() where needed.
This problem was reported by Jens Domke
<domke.j.aa@m.titech.ac.jp>
Sean Hefty [Mon, 25 Mar 2013 21:20:33 +0000 (14:20 -0700)]
Call ucma_ib_init() during startup
ucma_ib_init() was removed from ucma_init() to avoid
recursive callback issues when using the librspreload
library over rsockets. Although rsockets calls ucma_ib_init,
other pieces of the code do not. The result is that
ibacm ends up being disabled unless rsockets has been
used. Fix this by calling ucma_ib_init() where needed.
Sean Hefty [Tue, 5 Mar 2013 21:39:00 +0000 (13:39 -0800)]
librdmacm/rstream: Adjust transfer counts
A customer has reported that 100 transfers for large sizes
is insufficient to achieve the top bandwidth over a high
latency link. Contrary, 100,000 transfers is too high for
small latency ping-pong over a slow connection.
Sean Hefty [Tue, 5 Mar 2013 21:39:00 +0000 (13:39 -0800)]
librdmacm/rstream: Adjust transfer counts
A customer has reported that 100 transfers for large sizes
is insufficient to achieve the top bandwidth over a high
latency link. Contrary, 100,000 transfers is too high for
small latency ping-pong over a slow connection.
Sean Hefty [Tue, 5 Mar 2013 21:39:00 +0000 (13:39 -0800)]
librdmacm/rstream: Adjust transfer counts
A customer has reported that 100 transfers for large sizes
is insufficient to achieve the top bandwidth over a high
latency link. Contrary, 100,000 transfers is too high for
small latency ping-pong over a slow connection.
Sean Hefty [Tue, 5 Mar 2013 21:39:00 +0000 (13:39 -0800)]
librdmacm/rstream: Adjust transfer counts
A customer has reported that 100 transfers for large sizes
is insufficient to achieve the top bandwidth over a high
latency link. Contrary, 100,000 transfers is too high for
small latency ping-pong over a slow connection.
Sean Hefty [Wed, 20 Feb 2013 04:03:58 +0000 (20:03 -0800)]
librdmacm/rsocket: Fix resetting O_NONBLOCK after calling shutdown
Shutdown switches an rsocket from nonblocking to blocking to
ensure that all data has been sent. After completing all
transfers, it should switch back to nonblocking; this handles
partial shutdown situations, where only half the connection
is shut down. However, the code uses the value of '1' to
set the nonblocking flag, rather than O_NONBLOCK. Fix this.
Sean Hefty [Wed, 20 Feb 2013 04:03:58 +0000 (20:03 -0800)]
rsocket: Fix resetting O_NONBLOCK after calling shutdown
Shutdown switches an rsocket from nonblocking to blocking to
ensure that all data has been sent. After completing all
transfers, it should switch back to nonblocking; this handles
partial shutdown situations, where only half the connection
is shut down. However, the code uses the value of '1' to
set the nonblocking flag, rather than O_NONBLOCK. Fix this.
Sean Hefty [Wed, 20 Feb 2013 04:03:58 +0000 (20:03 -0800)]
rsocket: Fix resetting O_NONBLOCK after calling shutdown
Shutdown switches an rsocket from nonblocking to blocking to
ensure that all data has been sent. After completing all
transfers, it should switch back to nonblocking; this handles
partial shutdown situations, where only half the connection
is shut down. However, the code uses the value of '1' to
set the nonblocking flag, rather than O_NONBLOCK. Fix this.