GitLab

Detail:
  ShareFS (using UDP) could get in a race condition when pushing files > 4kB
  to a remote share. The transmit buffer was large enough to send multiple
  frames without checking for any receives, so the acknowledge from the remote
  share was lost. Retry logic within ShareFS concluded the frames should be
  resent, with the same result - no ack - followed by resending. The computer
  would consume 100% of CPU time retrying until the cable was unplugged.
  Protocols using TCP were not affected due to window size negotiation inherent
  in that connection type.
  Reduce the transmit buffer size to 2048 bytes (+1 for a workaround for
  DeviceFS 0.69 and earlier) to match 3 of the other EtherUSB backends. This
  gives chance to check the receive queue between transmitted frames.

  Use symbolic names for arguments to usb_control().

  Increase the module version number reported to InetSetup by the AutoSense
  utility. This ensures that if a module with CDC support is softloaded and
  configured then on restart the RMEnsure in Choices:Internet.Startup will not
  leave the user without an interface if an older copy is present in ROM.
Admin:
  Submission for USB bounty.

Version 0.43. Tagged as 'EtherUSB-0_43'

Detail:
  Some communication devices have several configurations, including a CDC one
  which is not the default selected. Add these to the the products match
  mechanism (using table entries, or system variables) so that EtherUSB can
  automatically select the correct configuration when encountered.

  The configuration change is made on a callback when trying to open the CDC
  device, which on first pass returns a failure error. This is because at the
  time USBDriver sends the 'Attached' service call we can't issue star
  commands which affect the USB stack without the risk of a reentrancy problem.
  Once idle the callback fires and sets the configuration and the new
  descriptors can be read safely for a second pass at opening.

  If EtherUSB is subsequently killed, restore the default device configuration
  so the system is back in the state in which it was found - this avoids a hang
  with devices which don't start up in the CDC configuration by default.
Admin:
  Add a rule for debug to account for makefile changes in BuildSys 7.91.

Detail:
  Allow class CDC-ECM and those whose top level class is 0.0:0 too.
  Open an interrupt pipe in order to get status notifications, use this to
  populate the device abilities/status more accurately.
  When querying the MAC address, parse the language table before requesting
  the string.
  MTU correctly calculated.
  Log improvements to help debugging.

BuildSys 7.91 changes how top-level phony rules are overridden.
In this component, the `install` and `standalone` targets are redefined
after `CModule` is included, so they will be affected.

Version 0.42. Not tagged

Detail:
  Fix capitalisation of UpCall.h for case-sensitive filesystems

Version 0.42. Not tagged

EtherUSB stores the MAC internally in network byte order, except when the address is obtained via OS_ReadSysInfo 4, when it was stored reversed. This problem lay hidden because the only ROM in which EtherUSB is able to get an address that way is the Pi, whose HAL returned the address in the wrong order too.
Requires HAL_BCM2835-0_84.

Version 0.42. Tagged as 'EtherUSB-0_42'

Detail:
  c/smsc78xx, c/module, c/products, h/backends, Makefile - Add SMSC78xx driver, for the chip used in the Raspberry Pi 3B+
  c/config, c/mii, h/mii - Add support for gigabit MIIs & associated (auto) configuration
  c/net, h/net - Handle gigabit link types. Pass verbose flag through to backend info function.
  c/smsc95xx - Make configure_link const-correct
  Resources/UK/CmdHelp - Update *EJConfig syntaxt text to include 1000 option
  Resources/UK/Messages - Add 78xx backend name & device-specific info tokens
Admin:
  Tested on Raspberry Pi 3B & 3B+
  Manually configured gigabit links aren't supported yet
  Gigabit half-duplex isn't supported (78xx MAC limitation)
  For the 78xx, manually-configured full-duplex links appear to suffer terrible RX performance (as seen on 95xx). All supported auto modes appear to work correctly.


Version 0.41. Tagged as 'EtherUSB-0_41'

Detail:
  c/pegasus - Add mii_abilities() call to fix backend not working since extra sanity checks were added to config_parse_arguments() in EtherUSB 0.39. Also switch over to using the common status callback code, and provide a configure function.
  Resources/UK/CmdHelp - Update *ejconfig syntax message, since it was missing some things like 'noauto'.
Admin:
  Tested on beagleboard
  Fixes pegasus backend not working since EtherUSB 0.39:
  https://www.riscosopen.org/forum/forums/4/topics/10119


Version 0.40. Tagged as 'EtherUSB-0_40'

Detail:
  c/smsc95xx - Convert code to use the common MII backend. Reduce verbosity of debug output. Fix two sets of status callbacks running (callback_status_adapter + device_status). Tweak/refactor a couple of other bits as necessary in order to get 10M links working reliably.
  c/mii, h/mii, h/net - Allow backends to indicate their level of support for PAUSE functionality, so that we can advertise it correctly when performing auto-negotiation (previously no pause ability would have been advertised)
  c/config - Ensure config_parse_arguments passes through the PAUSE capabilities to the output net_config_t. Use the (currently unused) ERR_BAD_CONFIG for if autonegotiation is disabled but either speed or duplex setting not specified (avoids MII code or backends getting confused as to what options they should use)
  Resources/UK/Messages - Tweak wording of ERR_BAD_CONFIG so that it's more sensible for the new usage
  doc/!ReadMe - Remove the note about 10M links not being supported by the SMSC 95xx backend. Update the changelog to point people towards CVS for recent changes.
Admin:
  Tested on PandaBoard + Raspberry Pi 3, using SMSC 95xx and AX88772 devices
  SMSC testing results:
  * Previously, 10M links were non-functional
  * Previously, manually-configured 100M full duplex links would suffer terrible performance
  * Previously, any attempt to reconfigure a link on Pi 3 would result in a "Device broken or misbehaving" error? (or at least, that's how 0.38 is behaving for me now)
  * Now, all combinations of configuration (speed + link + auto/noauto) appear to work correctly, except for manually-configured full-duplex links, which still suffer from terrible performance
  AX88772 was used to verify MII changes haven't broken anything for other backends; testing showed no significant differences to previous version (all speed + link + auto/noauto modes working)
  Resolves ticket #342:
  https://www.riscosopen.org/tracker/tickets/342
  Note that currently the SMSC 95xx backend is the only backend to advertise PAUSE functionality via the common MII code; a future improvement would be to set up the other backends appropriately (set correct flags in net_abilities_t)


Version 0.39. Tagged as 'EtherUSB-0_39'

Detail:
  net.c: added support for 1000Mb speed
  smsc75xx.c: reduced MAX_RX_FIFOSIZE to 16k (avoids problems with Kernel keyboard scanning);
              corrected determination of ethernet speed (read from SMC75XX_MAC_CR register)
  net.h: added net_speed_1000Mb (in net_speed_t) and speed_1000Mb (in struct net_abilities_t)
  Messages: added Sp1000 and Dit21 for supporting 1000Mb speed
Admin:
  Submission from Colin Granville.

Version 0.38. Tagged as 'EtherUSB-0_38'

Detail:
  c/module - Don't pass silly error numbers to MessageTrans_ErrorLookup - the kernel will now fault them.
Admin:
  Tested on Raspberry Pi 3


Version 0.37. Tagged as 'EtherUSB-0_37'

Detail:
  This module previously used both SWP and SWPB instructions, which have
  been withdrawn in ARMv8. Now use SyncLib to provide the equivalent
  functionality, which will use LDREX/STREX or LDREXB/STREXB instead when
  appropriate.
Admin:
  Tested on Raspberry Pi 3.

Version 0.36. Tagged as 'EtherUSB-0_36'

Detail:
  More careful manipulation of the control register, and extra entries in the product lookup table, adding support for AX88772B controllers.
  Added EUI48 check function, blank MAC addresses are now replaced by locally generated ones in dongles so cheap they don't have one of their own.
Admin:
  Submission from Colin Granville.

Version 0.35. Tagged as 'EtherUSB-0_35'

Detail:
  Listen out for different service calls so EtherUSB keeps connected even if DeviceFS/USBDriver/HC backends get restarted.
Admin:
  Submission from Colin Granville.

Version 0.34. Tagged as 'EtherUSB-0_34'

Cut and paste mistake from smsc95xx during review.

Version 0.33. Tagged as 'EtherUSB-0_33'

Support for USB product ids 0x7500 and 0x7505.
Submission from Colin Granville.

Version 0.32. Tagged as 'EtherUSB-0_32'

ax88772.c:
  Modified to work with changes made to net.c and usb.c.
  Increased buffer sizes, allow more than one packet receipt.
ax88172.c/mcs7830.c/pegasus.c:
  Modified to work with changes made to net.c and usb.c.
smsc95xx.c:
  Copyright header simplified by using pronoun form "the copyright holders" rather than repeating this list.
  Modified to work with changes made to net.c and usb.c.
  Reading of MAC address from EEPROM fixed.
products.c:
  Added SMSC9730 device id (uses smsc95xx backend).
net.c/net.h:
  tx_queue removed, all transmit buffering is now done in the USB code.
usb.c:
  Function usb_write() blocks until room is available, or times out.
  Use USBDriver header rather than defining local SWI numbers.
Messages:
  Renamed DescSMC token to DescSMC9.
  Removed redundant Stat7 statistic.

Submission from Colin Granville.
Tested on Pi B+ and OMAP3, not tested on OMAP4, but believed to markedly improve its performance over version 0.30.

Version 0.31. Tagged as 'EtherUSB-0_31'

The FSCV, FindV, UpcallV handlers were marked as generic veneers rather than vector handlers.
Spotted by Colin Granville.

Version 0.30. Tagged as 'EtherUSB-0_30'

Retagged as EtherUSB-0_29.

Step change improvement in transmit speed 'usb_upcall_hook_handler' removed filling of tx pipe. This fixes problems of the tx pipe being very slow. It was never needed anyway.

When transmitting, the tx pipe is driven by inserts from the socket driver and transfering that data to USB kicks USB writes into action.
Underflow isn't a problem USB will send as data arrives in the buffer.
Submission from Colin Granville.

Testing with a Raspberry Pi connected via LanManFS to a WinXP server *COPYing a 10MB file from SDFS to Windows, and *COPYing it back to null:
  EtherUSB 0.26  uplink   0.07 MB/s
                 downlink 2.96 MB/s
  This version   uplink   1.31 MB/s
                 downlink 3.05 MB/s
inspection in Wireshark shows no dropped TCP segments or ACKs during the test, previously duplicated ACKs and missing segments would cause most of the uplink slowness while the TCP/IP stack backed off for ~1.5s.
Note: other filing systems and use cases may vary.
In the case of the SMSC91xx backend, ticket #342 remains - sending to a 10Mbps server (even if through a 100Mbps switch) causes dropped segments/duplicated ACKs as before.

usb.c: upcall handler change as above.
smsc91xx.c: typos in comments and debug ammended.


Version 0.29. Tagged as 'EtherUSB-0_29'

Errors from net_attempt_transmit() are now passed back to the Internet module.
Fix from Colin Granville.

Version 0.28. Tagged as 'EtherUSB-0_28'

There's an element of duplication between the definitions of net.h and those in TCPIPLibs:net/ethernet.h which would be better to resolve, but for now just mirror its definition locally.
Not tagged.

Fixes to reads in the AX88772 backend 'read_packet' changed to read more than 1 ethernet packet. The device actually operates in burst mode so reads can contain more than 1 ethernet packet, the previous version assumed only ever 1 so dropped others.

Size of usb buffer passed to usb_open was reduced as it didn't help being bigger so may as well save memory.
Submission from Colin Granville.

Version 0.27. Tagged as 'EtherUSB-0_27'

AutoSense script updated to be able to parse both the hardwired english output (0.08-0.25 inclusive) and internationalised text (0.26+).
Tested RAM loaded on a Raspberry Pi.

Version 0.26. Tagged as 'EtherUSB-0_26'

The polarity information was being stuffed into the wrong stats byte.
Fixes problem reported
  http://www.riscosopen.org/forum/forums/4/topics/2121

Version 0.25. Tagged as 'EtherUSB-0_25'

Add more ways of specifying device configuration. Adjust unit number allocation & SMSC MAC handling.

Detail:
  c/smsc95xx - Adjust MAC handling some more:
    - Unit number check removed from EtherUSB$MAC_Configured code in fill_arbitraryMAC(), to increase compatibility with previous version
    - EtherUSB$MAC_Configured check is now performed before the devices currently-programmed MAC is checked; this should ensure it always takes priority over the other methods (especially when reiniting EtherUSB after setting the variable)
    - OS_ReadSysInfo 4 & machine ID hash checks remain in their previous place, after the currently-programmed MAC check
  c/config, h/config:
    - Add support for using EtherUSB$VP_* and EtherUSB$Loc_* system variables to configure devices, by specifying USB vendor/product IDs or USB device tree locations, respectively.
    - Allow config_reserved_units() to select a devices unit number by looking for VP_ or Loc_ system variables.
  c/net, h/net:
    - Add USB device vendor & product ID, and device tree location, to net_device_t struct.
    - Display vendor/product ID & device tree location in verbose ejinfo output
    - Adjust unit number selection: Attempt to select a unit number before calling backend->start, so that the unit number can be used in MAC allocation if required (i.e. by net_default_mac()).
    - Change net_default_mac() to use OS_ReadSysInfo 5 instead of OS_ReadSysInfo 2; OS_ReadSysInfo 2 strips out the first and last byte of the ID, making the psuedo-unique MAC somewhat less unique.
  c/usb, h/usb - Add utility function for reading the location of a device within the USB device tree.
Admin:
  Tested on BB-xM, Raspberry Pi
  Overriding MAC via EtherUSB$MAC_Configured now works as expected (was previously failing due to unit number being MODULE_MAX_UNITS)
  Unit number reservation/selection & device configuration via all three variable types (MAC, VP, Loc) verified


Version 0.24. Tagged as 'EtherUSB-0_24'

Detail:
  c/smsc95xx - Removed old code that was setting the MAC address via OS_ReadSysInfo 4. Code now correctly uses the logic described previously.
Admin:
  Tested on Raspberry Pi


Version 0.23. Retagged as 'EtherUSB-0_23'

Detail:
  c/net, h/net - Make net_default_mac() non-static. Add net_machine_mac() function for looking up MAC by OS_ReadSysInfo 4.
  c/smsc95xx - Change MAC selection logic. Old order was OS_ReadSysInfo 4 -> EtherUSB$MAC_Configured -> 100% hardcoded default. New order is EtherUSB$MAC_Configured -> OS_ReadSysInfo 4 -> hashed machine ID + unit number -> hardcoded default (with unit number). This new logic means BB-xM's should now have unique MACs instead of all sharing the same MAC.
Admin:
  Tested on BB-xM, Raspberry Pi. Raspberry Pi retains the MAC that's specified by the firmware, while the BB-xM now uses the hashed machine ID instead of the hardcoded default.
  Based on changes supplied by Willi Theiss.


Version 0.23. Tagged as 'EtherUSB-0_23'

Detail:
  c/net, c/usb - Tidy up some debugging code so it's only active in debug builds.
Admin:
  Changes supplied by Willi Theiss, to fix hangs on restart (on OMAP4) when the SysLog module is active.
  Tested locally on BB-xM


Version 0.22. Tagged as 'EtherUSB-0_22'

Detail:
  c/cmodule, c/usb, cmhg/EtherUSBHdr, h/usb - Sit on FSControlV & FindV so we can watch for file closures and catch any attempt to close our devicefs streams. Also need to sit on Service_Reset to make sure FileSwitch doesn't get the vector calls before us.
  If the device streams get closed without our notice, then the next attempt to write to a buffer (or potentially read from one?) via the direct buffer manager calls will result in RMA corruption.
Admin:
  Tested on Raspberry Pi with high processor vectors
  Fixes occasional data aborts while sat on "shutdown complete" screen (since *shutdown will close the devicefs streams)


Version 0.21. Tagged as 'EtherUSB-0_21'

Admin:
	tested in RPi build


Version 0.20. Tagged as 'EtherUSB-0_20'

Detail:
  c/net - In set_inet_variables(), no check was being made for whether OS_ReadVarVal returned an error or not, resulting in a null pointer dereference if <Inet$EtherType> does exist
Admin:
  Tested on Raspberry Pi with high processor vectors


Version 0.19. Tagged as 'EtherUSB-0_19'

Changed *EJINFO output to refer to bus/device to match *USBDEVICES nomenclature.

Version 0.18. Tagged as 'EtherUSB-0_18'

Save 4.5kB by making the table of errors with objasm (so they're not all padded to 252 bytes by the compiler to fit into _kernel_oserror structs).

The 'moduleerr.s' source should be made redundant by internationalising the module.

Version 0.17. Tagged as 'EtherUSB-0_17'

Now uses AsmUtils for irqs_on/irqs_off
Squash a few unused variable warnings, not in the backends though as porting updates would be more complex for those that come from (for example) NetBSD.
Changed to use structure definitions from sys/dcistructs.h (mostly an exercise in prefixing member names).

Version 0.16. Tagged as 'EtherUSB-0_16'

Use ModHand/Variables/RISCOS/OsByte headers.
Note, this has meant some header files needed reordering to prevent circular dependencies.

Version 0.15. Tagged as 'EtherUSB-0_15'

Makefile recreated from fragments
CMHG file renamed to avoid needing special rule to tell where it is.

Version 0.14. Tagged as 'EtherUSB-0_14'

No code changes, not tagged.