                              Release Notes
                        Broadcom tg3 Linux Driver
                              Version 3.116j
                                01/21/2011

                          Broadcom Corporation
                         5300 California Avenue,
                            Irvine, CA 92617

              Copyright (c) 2010 - 2011 Broadcom Corporation
                           All rights reserved


v3.116j (January 21, 2011)
============================
    Fixes
    -----
        1) Problem: Driver doesn't compile if the NetQueue code is
                    disabled.
           Cause  : A NetQueue only variable is referenced outside
                    of the NetQueue preprocessor guards.
           Change : Surround the offending code with NetQueue
                    preprocessor guards.
           Impact : None.


v3.116i (January 19, 2011)
============================
    Fixes
    -----
        1) Problem: (CQ52356) Phy loopback test fails for 5718.
           Cause  : An EEE workaround causes the problem.
           Change : Disable EEE during the test.
           Impact : This bug could affect 5717, 5719, and 5720
                    ASIC revs.

        2) Problem: SNMP daemons can interfere with management
                    firmware when the device is down.
           Cause  : PHY access collisions between the SNMP daemon and
                    firmware cause unpredictable problems.
           Change : Forbid phy access if the device is down and
                    management firmware is running.
           Impact : This bug affects any device with management
                    firmware.


v3.116h (January 11, 2011)
============================
    Fixes
    -----
        1) Problem: (CQ52203) Unable to set MTU.
           Cause  : The vmware kernel needs the driver to explicitly
                    take the rtnl_lock while changing the MTU size.
                    This code was accidentally enabled for the
                    standard tg3 driver, which will cause a deadlock.
           Change : Fix the preprocessor condition that enables the
                    rtnl_locking code.
           Impact : This bug will affect all asic revs.

    Enhancements
    ------------
        1) Change : Added EEE interoperability fix.


v3.116g (January 10, 2011)
============================
    Enhancements
    ------------
        1) Change : Added NetQueue support for VMWare driver.


v3.116f (January 6, 2011)
============================
    Fixes
    -----
        1) Problem: (CQ51944) 5718 serdes loopback tests fail.
           Cause  : The MAC loopback test has been deprecated.  The
                    internal phy loopback test fails because link has
                    not yet come up when the test is executed.
           Change : Abort MAC loopback tests.  Wait for link for
                    phy loopback tests.
           Impact : The MAC loopback test has been disabled for all
                    CPMU enabled devices.  The phy loopback test fix
                    has been applied to all asic revs.


v3.116e (January 4, 2011)
============================
    Fixes
    -----
        1) Problem: (CQ51944) Unable to pass traffic on 5719.
           Cause  : The fix for CQ51280 accidentally prevented the
                    proper configuration of the rx producer ring
                    RCBs.
           Change : Allow the 5719 to execute the proper code paths to
                    setup the RCBs.
           Impact : This bug only affects the 5719 A0 revision.

    Enhancements
    ------------
        1) Change : Updated copyright strings.


v3.116d (December 23, 2010)
============================
    Fixes
    -----
        1) Problem: (CQ51280) Kernel panics with 5719 on bi-directional
                    Chariot test.
           Cause  : HW bug.
           Change : The best known way to workaround the bug is to
                    disable TSO and jumbo frames.
           Impact : This bug only affects the 5719 asic rev and will be
                    fixed in the A1 revision.


v3.116c (December 16, 2010)
============================
    Fixes
    -----
        1) Problem: (CQ46520) 5718 serdes loopback tests fail.
           Cause  : The hardware needs some time to adjust to the
                    new configuration before running the test.
           Change : For MAC loopback, the hardware needs about 30
                    milliseconds before the test passes.  For internal
                    phy loopback, the driver needs to poll register
                    0x460 for link before proceeding with the test.
           Impact : While the problems are only seen on the 5718S,
                    these changes should benefit all asic revs.

        2) Problem: TSO defaults to off for HW TSO devices on RH5.X
                    distros.
           Cause  : The driver was recently changed to show that TSO
                    implies checksum offload is enabled.  RH5.X kernels
                    use a different netdev feature bit for HW TSO
                    devices though.
           Change : Modify the code so that it checks for the old and
                    current IP checksum flags.
           Impact : This bug affects all HW TSO devices.

        3) Problem: TSS enabled devices can panic on Linux distros that
                    support it.
           Cause  : The driver's compatibility code incorrectly enables
                    itself, causing all tx packets to go to the first
                    tx ring.  When two packets (intended for different
                    rings get transmitted at the same time, the tx BD
                    ring and indexes get corrupted.  Ultimately this
                    causes a tx_recovery, from which the device never
                    recovers.
           Change : Change the compatibility code so that it enables
                    only at the proper times.
           Impact : This bug affects all TSS enabled devices.

    Enhancements
    ------------
        1) Change : Reduce the performance impact of the 5719 FIFO
                    overflow fix.
                    


v3.116b (December 9, 2010)
============================
    Enhancements
    ------------
        1) Change : Sync-up with upstream tg3 driver.


v3.115j (December 9, 2010)
============================
    Fixes
    -----
        1) Problem: (CQ46520) Loopback test fails when link is set to
                    10/100Mbps half duplex.
           Cause  : The loopback test is a full duplex test.  The half
                    duplex bit was set in register 0x404.
           Change : Clear the half duplex bit for the loopback test.
           Impact : This bug only seems to affect the 5785.

    Enhancements
    ------------
        1) Change : EEE debounce counters were accidentally set to very
                    relaxed values.  This has been fixed.
        2) Change : Fix 5701 packet loss by 8-byte aligning the rx_offset
                    needed for possible VLAN tag reinsertion.


v3.115i (December 1, 2010)
============================
    Fixes
    -----
        1) Problem: Tx data corruption on 5719.
           Cause  : Hardware bug requiring workaround.
           Change : Coalesce all buffers that contain 8-bytes or less.
           Impact : 5719 only.

        2) Problem: 57765, 5717, and 5719 devices can perform
                    suboptimally.  With packets sized around the MTU
                    size.
           Cause  : Driver sets the jumbo frame flag for packets that
                    were greater than 1500 bytes.
           Change : Only set the jumbo frame flag for frames that are
                    larger than 1518 in length.
           Impact : This problem only affects the 57765, 5717, and 5719
                    ASIC revs.

    Enhancements
    ------------
        1) Change : Added a compile time flag (TG3_NO_EEE) and a module
                    parameter (tg3_disable_eee) to selectively disable
                    EEE support.  See README.TXT for details.


v3.115h (November 16, 2010)
============================
    Fixes
    -----
        1) Problem: RDMA FIFO overrun can still happen with a Gen2
                    x4 PCIe link.
           Cause  : Hardware bug requiring workaround.
           Change : Adjust MRRS PCIE register to 512 bytes.
           Impact : 5719 only.

    Enhancements
    ------------
        1) Change : Sync EEE code with Windows driver.


v3.115g (November 15, 2010)
============================
    Fixes
    -----
        1) Problem: RDMA FIFO overrun can still happen with workaround
                    in place.
           Cause  : Hardware bug requiring workaround.
           Change : Adjust MRRS PCIE register based on PCIE link
                    speed and width.
           Impact : 5719 only.

        2) Problem: Driver crashes on load on RedHat 5.6.
           Cause  : alloc_etherdev_mq() should only be used if
                    the device supports multiple queues.
           Change : Change the code to use alloc_etherdev() instead.
           Impact : All multiqueue devices (5717, 5719, 57765 asic revs).


v3.115f (November 12, 2010)
============================
    Fixes
    -----
        1) Problem: RDMA FIFO overrun on 5719.
           Cause  : Hardware bug requiring workaround.
           Change : Adjust MRRS PCIE register based on PCIE link
                    speed and width.
           Impact : 5719 only.

v3.115e (November 8, 2010)
============================
    Fixes
    -----
        1) Problem: (CQ50579) 57765 EEE does not work until the
                    driver is unloaded and reloaded.
           Cause  : The phy did not enable EEE capabilities after
                    a phy reset.
           Change : Program the CPMU EEE registers before issuing
                    a phy reset.
           Impact : This bug only affects the 57765 B0+ asic rev.

        2) Problem: (CQ50823) Cannot build driver against latest
                    RedHat 5.6 release.
           Cause  : Compatibility code problems.
           Change : Fixed compatibility code.
           Impact : This only affects RedHat 5.6.

    Enhancements
    ------------
        1) Change : Reenabled EEE support.


v3.115d (November 1, 2010)
============================
    Enhancements
    ------------
        1) Change : Reordered tx members in the tg3_napi structure
                    and cacheline-aligned the tx and rx member variable
                    sections.  Improves small packet performance
                    by ~3-4%.
        2) Change : Removed EEE support.


v3.115c (October 27, 2010)
============================
    Fixes
    -----
        1) Problem: (CQ50635) Pass-through does not work with 5718S.
           Cause  : The driver was not preserving the APE rx and tx
                    settings in the mac mode register for serdes
                    devices.
           Change : Preserve the APE rx and tx settings.
           Impact : This bug should only affect 5718S and 5719S.

    Enhancements
    ------------
        1) Change : Enable gphy APD for 5717 and newer devices.
        2) Change : Remove 8-byte DMA fragment length workaround for
                    5719 devices.
        3) Change : Apply 10Mbps CPMU clock workaround for all 57765
                    devices, not just the A0 revision.
        4) Change : Reenable TSS for 5719 devices.
        5) Change : Enable Multiple DMA Read Request featuer for 5719.
        6) Change : Apply tx margin fix for 5719 devices.


v3.115b (October 18, 2010)
============================
    Fixes
    -----
        1) Problem: (CQ50405) Unable to show 5718 bootcode version.
           Cause  : A new sb format had been introduced.
           Change : Add support for sb format 1 revision 6.
           Impact : This bug should only affect 5717 and 5719 asic
                    revs.

    Enhancements
    ------------
        1) Change : Sync-up with upstream tg3 driver.


v3.114e (October 12, 2010)
============================
    Enhancements
    ------------
        1) Change : Enable EEE support for 57765 B0 and later devices.


v3.114d (October 5, 2010)
============================
    Fixes
    -----
        1) Problem: (CQ50234) Driver crashes when opening non-jumbo
                    capable devices.
           Cause  : Recently the driver was changed to abstain from
                    allocating resources for the jumbo producer ring(s)
                    if the device was not jumbo capable.  One of the
                    tests mistakenly allowed non-jumbo capable devices
                    to attempt to access unallocated resources.  The
                    result was a kernel panic.
           Change : Fix the test so that non-jumbo capable devices are
                    properly blocked.
           Impact : This affects all non-jumbo capable devices.


v3.114c (October 4, 2010)
============================
    Fixes
    -----
        1) Problem: Driver fails to compile for VMWare.
           Cause  : Code errors missed due to build system
                    misconfiguration.
           Change : Fixed the build system configuration and
                    then fixed the syntax errors.
           Impact : This problem only affects the VMWare driver.


v3.114b (October 1, 2010)
============================
    Enhancements
    ------------
        1) Change : Sync-up with upstream tg3 driver.


v3.113i (September 30, 2010)
============================
    Fixes
    -----
        1) Problem: (CQ50133) When device is closed, ethtool can
                    report link when it is not available.
           Cause  : The driver only fills in speed / duplex information
                    when the device has been opened.  When the device
                    is closed, the fields are left alone.  When ethtool
                    interprets the information returned from a closed
                    device, it will use whatever value is in that
                    field and interpret it incorrectly.
           Change : Set the speed and duplex fields to the same invalid
                    values that would be used if the link were down.
           Impact : Link always shows down when device is down.  Speed
                    and duplex values are obviously invalid.


v3.113h (September 29, 2010)
============================
    Enhancements
    ------------
        1) Change : Enlarged the standard producer ring size from 512 to 2k.
                    Enlarged the jumbo producer ring size from 256 to 1k.
                    Enlarged the rx return ring size from 1K to 4K.
        2) Change : Removed 5724 device ID.


v3.113g (September 28, 2010)
============================
    Fixes
    -----
        1) Problem: (CQ49992) 5718 partno shows as "none".
           Cause  : The code correctly obtained the partno, but
                    incorrectly branched to the "none" section
                    afterwards.
           Change : Fix the code that caused the branch.
           Impact : All devices.

        2) Problem: (CQ50001) Cannot compile on SuSE 9.X or SuSE 10.X
           Cause  : Compatibility code problems.
           Change : Fixed compatibility code.
           Impact : This bug only affects these two platforms.

        3) Problem: NAPI code will crash if netpoll is enabled.
           Cause  : Netpoll visits each NAPI instance of a device during
                    a poll, whether it is enabled or not.  If an
                    inactive NAPI instance has any uninitialized
                    pointers, the driver will crash.
           Change : Change the code to only add as many NAPI instances
                    as there are allocated interrupt vectors.  Free the
                    NAPI instances when the device is closed.
           Impact : This bug affects all devices.

    Enhancements
    ------------
        1) Change : Replaced hardcoded constants with preprocessor
                    macros.
        2) Change : Added 5717 asic rev partno strings, in case the
                    NVRAM does not contain a VPD section.
        3) Change : Modified the TXMBUF margin to avoid data corruption
                    on 5719 devices.
        4) Change : Set the request size to 4K to improve performance on
                    5719 devices.


v3.113f (September 24, 2010)
============================
    Enhancements
    ------------
        1) Change : Added EEE support for B0 revisions of the
                    5717 asic rev.
        2) Change : Added TSS netq support.  All current hardware
                    implementations are broken, so this code is
                    effectively disabled.


v3.113e (September 24, 2010)
============================
    Enhancements
    ------------
        1) Change : Sync-up with upstream tg3 driver.


v3.113d (September 1, 2010)
===========================
    Enhancements
    ------------
        1) Change : Added baseline netqueue support.  Multiple queues
                    are enabled in the hardware but not exposed to
                    the system.
        2) Change : Added fix for CQ25155 (FIFO overflow fix).
        3) Change : Fixed jumbo frames capability report for VMWare
                    driver.


v3.113c (August 18, 2010)
=========================
    Enhancements
    ------------
        1) Change : Removed APE events for NCSI firmware.  They are
                    not used.
        2) Change : Modified code to respect the semantics of
                    TG3_IRQ_MAX_VECS and tp->irq_max.
        3) Change : Moved the producer ring data structures into
                    the per-napi instance structure.
        4) Change : Modified makeflags.sh to report when the
                    mandatory kernel source directory argument is
                    missing.
        5) Change : Moved global rx_mode tracking variable for
                    vmware to a per-device member.


v3.113b (August 03, 2010)
=========================
    Enhancements
    ------------
        1) Change : Sync-up with upstream tg3 driver.


v3.112e (July 29, 2010)
========================
    Fixes
    -----
        1) Problem: (CQ49020) Driver does not compile on SLES 10.3 or
                    SLES 9.x.
           Cause  : The preprocessor guard in the previous release was
                    misapplied.
           Change : Correct the preprocessor guard.
           Impact : The driver compiles on SLES 10.3 and SLES 9.x.


v3.112d (July 28, 2010)
========================
    Fixes
    -----
        1) Problem: (CQ48897) Driver fails to show NCSI firmware
                    version, or driver shows wrong firmware
                    description for NCSI firmware.
           Cause  : The firmware wasn't setting the "status ready"
                    bit, which caused the driver to abort the fw
                    version retrieval routine.  Once that was fixed
                    the driver mistakenly assumed DASH was the only
                    type of firmware to be running on the APE.
           Change : Inspect the "firmware features" APE shared
                    memory location to determine the type of firmware.
           Impact : This bug only affects the 5717 and perhaps the
                    upcoming 5719 asic revs.

        2) Problem: (CQ49020) Driver does not compile on SLES 10.3 or
                    SLES 9.x.
           Cause  : The compile fails because the driver uses
                    unrecognised IPv6 TSO preprocessor constants.
                    These platforms do not support IPv6 TSO.
           Change : Protect the unrecognised constants with a
                    preprocessor guard.
           Impact : The driver compiles on SLES 10.3 and SLES 9.x.

        3) Problem: (CQ49066) 5723 device issues an NMI on driver load.
           Cause  : The 5723 received a PCIe packet whose payload was
                    256 bytes in length and interpreted it as a
                    malformed PCIe packet.
           Change : The driver had code which attempted to "correct" the
                    negotiated PCIe payload size to 128 bytes.  This code
                    was introduced for the 5750 ASIC rev and was
                    conservatively applied to all devices older than and
                    including the 5784 ASIC rev.  The change was to
                    remove the 5784 ASIC rev from the list of devices
                    this code should be applied towards.
           Impact : This bug only affects the 5784 ASIC rev.
                    

    Enhancements
    ------------
        1) Change : Increased small packet performance by 3% by recoding
                    transmit related memory barriers.


v3.112c (July 22, 2010)
========================
    Fixes
    -----
        1) Problem: (CQ48173, CQ48172, CQ48171) Link failure (at 1Gbps)
                    when changing link from forced mode.  No link when
                    link is forced to 10Mbps.  Link flaps at 100Mbps
                    when forced.
           Cause  : Driver failed to clear bit 5 of register 0x18,
                    shadow reg 0x7.  This bit enables RGMII inband
                    status updates.
           Change : Set the above bit.
           Impact : Link comes up and passes traffic at all forced
                    and autonegotiated speeds.

    Enhancements
    ------------
        1) Change : Added PCI configuration space register read and
                    write ioctls for VMWare driver.


v3.112b (July 12, 2010)
========================
    Enhancements
    ------------
        1) Change : Sync-up with upstream tg3 driver.


v3.111b (July 01, 2010)
========================
    Enhancements
    ------------
        1) Change : Sync-up with upstream tg3 driver.


v3.110j (June 28, 2010)
========================
    Fixes
    -----
        1) Problem: Fix for CQ48242, CQ48186 did not work as expected.
           Cause  : Test compares RSS enable flag against the wrong
                    flags variable.
           Change : Use the correct flags variable.
           Impact : This problem affects 5717 and 57765 asic revs.


v3.110i (June 23, 2010)
========================
    Fixes
    -----
        1) Problem: (CQ48242, CQ48186) Traffic fails or poor performance.
           Cause  : A recent change allowed the driver to allocate a
                    single MSI-X vector, rather than falling back to MSIs.
                    The coalescing register setup function incorrectly
                    used a flag, that indicates the use of MSI-X vectors,
                    to determine whether or not to configure non-zero
                    rx coalescing parameters for vector 0.  Zeroed rx
                    coalescing parameters essentially disables rx
                    interrupts.
           Change : Use the RSS enable flag instead.
           Impact : Rx interrupts are restored.  This problem affects 5717
                    and 5719 ASIC revs.


v3.110h (June 15, 2010)
========================
    Enhancements
    ------------
        1) Change : Replace another VMWARE_ESX_40_DDK preprocessor check
                    with its __VMKLNX__ constant equivalent.
        2) Change : Enable TSO on VMWare async builds.
        3) Change : Add three missing devices to tg3.xml file.


v3.110g (June 8, 2010)
========================
    Fixes
    -----
        1) Problem: (CQ47824) 57765 devices will occasionally not be
                    able to ping.
           Cause  : The fix and reasoning behind the fix for CQ44458
                    below are incorrect.  The indirection table
                    addresses rx queues, not MSI-X vectors.
           Change : Revert the fix for CQ44458.
           Impact : Pings work as expected.  The root cause for CQ44458
                    is that the incoming rx packets just don't hash to
                    the last RSS queue.

    Enhancements
    ------------
        1) Change : Integrate changes based on code review.
        2) Change : Make MSI-X interrupt allocation failure checks
                    more robust.
        3) Change : Add support for 50612E phy for 5785.


v3.110f (May 28, 2010)
========================
    Fixes
    -----
        1) Problem: BD and / or packet corruption happens under high
                    traffic conditions for 5717 B0 and 5719 revs.
           Cause  : The multiple DMA read engine still has problems.
           Change : Turn off the multiple DMA read engine and coalesce
                    all packets that have fragments that are less than
                    8 bytes in length.
           Impact : This problem affects all 5717 and 5719 devices.

        2) Problem: Link up via parallel detect does not work.
           Cause  : The phy on 5717 and 5719 devices determines link up
                    via parallel detection internally.  When the ISR
                    processes the link change though, the remote
                    partner's link advertisement register will read 0x0.
                    This ultimately causes the driver to incorrectly
                    conclude that the link is down.
           Change : Trust the link status in phy register 0x1.  If
                    the remote partner's advertisement register reads
                    0x0, assume link is up via parallel detect.
           Impact : This problem affects all 5717 and 5719 serdes
                    devices.

    Enhancements
    ------------
        1) Change : For the 5717 (and 5719), the driver uses the PCI
                    function number to determine which MDIO bus address
                    the driver should use to address the current
                    device's phy.  This release changes the code to
                    use the function number provided by the kernel
                    rather than decoding a device register.  The
                    5719 changes the layout of this register, so the
                    benefit of this change is more maintainable code.


v3.110e (May 13, 2010)
========================
    Fixes
    -----
        1) Problem: Driver issues ALIGN not defined warnings on 2.4
                    kernels.
           Cause  : Missing compatibility code.
           Change : The driver's compatibility code will define this
                    macro if it does not already exist.
           Impact : This problem was only seen on 2.4 kernels.

        2) Problem: Driver compile fails on newer kernels.
           Cause  : Recent makefile changes assumed `uname -p` would
                    return x86_64, i686, etc.  On some systems, the
                    command actually returns a processor specific
                    descriptive string.
           Change : Use `uname -m` instead, which seems to get it right
                    in all cases so far.
           Impact : This bug affects any system that does not meet the
                    old assumptions.

    Enhancements
    ------------
        1) Change : Allow the allocation of a single MSI-X interrupt if
                    the system supports more than one CPU.  The benefit
                    is that the system can migrate that interrupt to
                    any CPU if needed.


v3.110d (May 11, 2010)
========================
    Fixes
    -----
        1) Problem: (CQ45614) Driver fails to build if kernel variant
                    binaries are not installed.
           Cause  : The driver's makefile looks at the installed kernel
                    binary directory in /lib/modules for a symlink that
                    points to the kernel sources.  If that symlink is
                    not found, the build ultimately fails.
           Change : Change the makefile to look for kernel sources in
                    known locations if the kernel binary symlinks are
                    missing.
           Impact : This problem can happen anytime somebody runs make
                    with KVER set and with only the kernel devel packages
                    installed.

        2) Problem: (CQ44458) CPU3 does not get utilized when RSS mode is
                    enabled.
           Cause  : The RSS indirection table indexes into the MSI-X
                    vector table, not table of RX queues.
           Change : Add one to the indirection table index so that the table
                    always routes packets to valid vectors.
           Impact : This problem happens on 5717 and 57765.  At the time of
                    this writing, these are the only two devices that
                    implement RSS and MSI-X interrupts.

        3) Problem: (CQ44456) 72hour cc32 stress test fails.
           Cause  : The driver was leaking memory.  An out of memory
                    condition could cause the reboot.
           Change : Free all rx rings when deconstructing the device.
           Impact : This problem happens on 5717 and 57765.  At the time of
                    this writing, these are the only two devices that
                    implement RSS and MSI-X interrupts.


v3.110c (April 27, 2010)
========================
    Fixes
    -----
        1) Problem: (CQ47032) Driver version incorrect in Dash
                    Management console.
           Cause  : The driver posted a static 3.81 version to
                    APE shared memory.
           Change : Implement a driver version reporting scheme
                    that will always stay accurate.
           Impact : This problem only afflicts APE-enabled devices.
                    At the time of this writing, only the 5761 and
                    5717 devices are affected.

        2) Problem: Driver implements APE mutex locking incorrectly
                    for the 5717.
           Cause  : APE mutex registers changed location.
           Change : Recode the driver to use the new mutex register
                    location.
           Impact : This bug only affects the 5717 asic rev.

        3) Problem: On SLES 11.X systems, modules that are marked
                    externally supported prevent kernel updates.  This
                    is regarded as a security concern.
           Cause  : The system's update mechanisms simply won't update
                    kernels with such modules in place.
           Change : Remove the Module.supported target from the driver's
                    makefile.  This will render the module unloadable
                    in a default install.  Update the README.TXT file to
                    describe the problem and how to work around it.
           Impact : This problem will affect all SLES 11 and newer
                    installs.
        4) Problem: Jumbo frames are corrupted at 100Mbps.
           Cause  : Driver was accidentally disabling the tx mbuf lockup
                    fix in register 0x45c.
           Change : Set bit 8 of 0x45c to enable lockup fix.
           Impact : This problem affects all 5755 and newer devices.

    Enhancements
    ------------
        1) Change : Enable the multiple DMA read facilities for 5717
                    B0 devices.
        2) Change : Removed the Module.supported file as a build target.
                    This means the admin will have to toggle the
                    allow_unsupported_modules flag in
                    /etc/modprobe.d/unsupported-modules.  The README.TXT
                    has been updated to address this.

v3.110b (April 13, 2010)
========================
    Initial release notes.
