Release Notes

1.0 September 25, 1999.

Highlights

• The following VIPL routines have been implemented:
  • VipRecvNotify, VipSendNotify, and VipCQNotify
  • VipResizeCQ
  • VipQueryMem and VipSetMemAttributes
  This leaves only the peer-to-peer functions missing from the list of functions required for Functional Conformance and for Full Conformance.

• Updated the M-VIA tulip driver to use the 0.90q version as its base. This adds support for NICs based on 21143 chips.

• Added a driver for Intel EtherExpress Pro/100 cards.

• Numerous bugs have been fixed.

• There is a first cut at a 3c59x driver that almost works.

Details

• Updated the M-VIA tulip driver to use the 0.90q version as its base. This adds support for NICs based on 21143 chips.

• Added a driver for Intel EtherExpress Pro/100 cards.

• SMP configuration has moved back into Makefile.config.

• Eliminated race which resulted in reuse of NicInstance numbers.

• Eliminated kernel panic when a packet-sent interrupt arrived after the sending VI had been destroyed.

• Eliminated a kernel hang which occured under certain circumstances when an RDMA transfer failed w/ a protection violation.

• No longer get kernel headers from /usr/src/linux, /usr/include/linux or /usr/include/asm. This is important because those directories may not be for the same kernel version as we are compiling for. We now get all kernel headers from /usr/src/linux-$(KERNEL-VERSION)/include.

• The Hamachi driver now deregisters the device on close as it should.

• Eliminated some memory leaks from the ERing-class drivers.

• Fixed a kernel deadlock which would occur if the CChan retransmit timer occured while the ERing device was busy.

• The MTU for the loopback device is no longer the unrealistic 4Gb value used before. It is now based on the maximum amount of registered memory .

• VipRecvNotify, VipSendNotify, and VipCQNotify have been implemented. Be warned, however, that you should not invoke VipDestroyVi from a call-back on the same VI or strange things might happen.

• Implemented VipResizeCQ.

• Implemented VipQueryMem and VipSetMemAttributes.

• Two new M-VIA specific calls have been added for caching of application level data associated with a VI. This is in the spirit of secion 9.2 of the Intel VIA Developer's Guide, but with calling conventions more like the rest of VIA.

          VIP_RETURN
          mviaSetViAppData(IN     VIP_VI_HANDLE   ViHandle,
                           IN     VIP_PVOID64     AppData);
          VIP_RETURN
          mviaGetViAppData(IN     VIP_VI_HANDLE   ViHandle,
                           OUT    VIP_PVOID64     *AppData);
  

• Debugging wrappers for kmalloc/kfree and vmalloc/vfree have been added to the drivers and kernel agent. These may be enabled by adding "-DVIPK_DEBUG_MALLOC" to the MOD_CFLAGS in Makefile.config

1.0b3 (formerly 0.9.3) April 9, 1999.

Highlights

With the release of version 1.0b3, M-VIA is on the verge of providing complete Functional Conformance to the VIPL API. M-VIA has been Early Adopter conformant since its initial release and version 1.0b3 includes key functionality and infrastructure to move to the next level of conformance.

The major components of this release are support for asynchronous remote disconnect notification, pthreads compatibility, and a new connection manager. These components provide increased functionality to the application developer as well as necessary functionality and infrastructure for achieving Functional Conformance.

Remote disconnect notification allows user applications to detect and gracefully recover from broken connections. Moreover, it provides one of the missing components of Functional Conformance from previous version of M-VIA, allowing M-VIA to pass many more Intel conformance suite tests.

Pthreads compatibility provides application developers the ability to write multi-threaded applications and is the first step in providing a completely thread safe VIPL. Operations performed within a multi-threaded application on different VIs are inherently thread safe, but an application must currently provide its own explicit locks if multiple threads access a single VI. The use of pthreads instead of the Linux specific clone system call used in previous versions is the first step in the development of a fully thread safe VIPL. Additionally, the use of pthreads will enable the development of the VIPL asynchronous Notify functions.

The new connection manager is more robust than in the previous versions of M-VIA. The previous connection manager was a prototype to provide the required user level connection management features, however, under certain circumstances it would fail, resulting in a machine lockup. The new connection manager also makes it easier to develop the peer-to-peer connection routines required for Functional Conformance.

Details

• Asynchronous remote disconnect notification is delivered in the event of an error on a remote VI, allowing user applications to detect and gracefully recover from broken connections.

• Single-processor loopback performance has been significantly improved by dynamically configuring the spin wait based on the number of CPUs and weather the connection is local or remote.

• M-VIA is now pthreads compatible, allowing users to write multi-threaded applications.
  
  M-VIA applications must now be compiled with -D_REENTRANT and linked with the -lpthread flag. In addition, glibc 2 is required (we have not tested M-VIA with previous versions.)

  Note, this change does not cause extra overhead to be incurred by the user application. Previous releases used a thread to handle asynchronous errors, but did so in a non-portable way (using the clone system call.) The error handler thread is blocked until an error occurs and does not consume CPU cycles.

  The use of signals to trigger the error handler was ruled out because the Linux pthreads library uses SIGUSR1 and SIGUSR2 internally and pthreads compatibility is an M-VIA goal. However, a future version of M-VIA will provide a compile time option as to weather to use signals or pthreads for asynchronous error handling.

• Connection management has been made more robust. The previous connection manager was a prototype to provide the required user level connection management features, however, under certain circumstances it would fail, resulting in a machine lockup. Additionally, the new connection management paves the road for the development of peer-to-peer connection management services.

• Linux 2.2.X support.

• Updated M-VIA tulip driver to use the Linux 2.2.x tulip driver as its base.

• Fixed a machine lockup which would occur after disconnecting a VI with circular work queues, i.e. reposting a descriptor before it has completed.

• Changed the the MaxReliabilityLevel attribute of the VIP_NIC_ATTRIBUTES structure to ReliabilityLevelSupport to conform with the 1.0 specification.

• Changed the the RDMARead attribute of the VIP_NIC_ATTRIBUTES structure to RDMAReadSupport to conform with the 1.0 specification.

• Changed memory deregistration to use free_page instead of atomic_dec of the page count. This allows Linux to reclaim the memory more quickly. (Thanks to Matt Welsh for pointing this out.)

• Removed SMP configuration from Makefile.config. This is now handled automatically in the linux kernel headers.

• The license has been changed so that it does not repeat what is already required by U.S. export control law (e.g. restriction on export to rogue states).

1.0b2 (formerly 0.9.2) November 17, 1998. Changes include:

• Support for the Packet Engines GNIC-II (Hamachi) gigabit Ethernet card. This was accidentally released in 1.0b1, but a) it didn't work and b) performance was terrible. It now works better, and latency is now about 20 us. The performance changes will also help TCP performance, which improves from more than 200us latency to about 60us latency.

  Hamachi cards seem to be finicky, and if you have problems bringing up the network after changing drivers, you may have to power cycle the machine (not just reboot).

• Hostfile parsing was changed. This fixed a bug in which extra lines in the host file caused a failure in VipNSInit. The format of loopback address in the file is now changed. To specify a name for loopback, use lines of the form:

             loopback     localhost      # Note that support for comments
             loopback     myhostname     # is also more robust now. 
  

• The scheduling timeout interface changed with 2.1.127 and further updates were made with 2.1.128 (a rather significant set of changes, we think, for feature-frozen code that is trying to converge to 2.2). 1.0b1 does not work with 2.1.127 or later. 1.0b2 does not work with 2.1.127 or earlier. 1.0b2 has been tested with 2.1.128. It may work with later kernel versions.

• Fixed a few bugs relating to waiting on send queues and internal VIP_NET_ADDRESS handling reported by Raymond Namyst.

• Fixed warnings reported by egcs. M-VIA now compiles under egcs as well as gcc.

1.0b1 (formerly 0.9.1) Initial Release (beta). November 5, 1998

---

M-VIA 1.0. Fri, 17 Sep 1999 09:13:43 -0700.
Copyright (C) 1998,1999 Berkeley Lab.