GitLab

Detail:
  DMA.s, in DMAForceActivate: Support for the NoInitIRQ flag for interrupt-driven buffer-type channels (API 1+)
  DMA.s, 'testloop': Fix handling of interrupt-driven buffer-type channels so DMAManager doesn't immediately crash or malfunction
  DMA.s, 'overrun', 'earlyoverrun', 'nooverrun': Fix handling of interrupt-driven buffer-type channels so IRQ state is cleared before the next stage of the transfer is programmed (to avoid missing IRQs)
  DMA.s, 'finished': Fix handling of interrupt-driven buffer-type channels so R8 is loaded with the logical channel block ptr, as expected by subsequent function calls.
  Front.s: Make dmaq_Trampoline larger to allow it to contain the HAL device ptr
  GetAll.s: Add Hdr:HALEntries to GET list, as required by 'ClaimVectors' fix in ModHead.s
  ModHead.s: Update controller handling code to cope with API versions 0.1 (TestIRQ2) & 1 ('Deallocate' passed channel device ptr)
  ModHead.s: Fix 'ClaimVectors' to call HAL_IRQEnable to enable IRQs for the detected channels/controllers.
  ModHead.s: Fix IRQ handler trampolines so that they call HALDevice_TestIRQ on the channel instead of branching to a random address in memory
  SWI.s: Support for controller API 1 (pass channel device ptr to 'Deallocate' in R2)
  hdr/DMADevice: Extra flags & entries for new DMA controller & buffer API versions
Admin:
  Tested on rev C2 beagleboard with the incoming HAL DMA driver. API revisions are documented in more detail on the ROOL wiki.


Version 0.18. Tagged as 'DMA-0_18'

Detail:
  Uses suffixed file extensions for compatiblity with both objasm and asasm.
Admin:
  Supplied by Peter Naulls, tested at ROOL

Version 0.17. Not tagged

Detail:
  * Service_PagesSafe handler would abort if the restarting transfer had
    more than one entry in its scatter list
  * Logical channels remained linked into the chain during mass
    deregistration in module finalisation, allowing new transfers to be
    queued
  * Logical channel (RMA) block wasn't freed in DMA_DeregisterChannel
Admin:
  Tested on Tungsten with UDMA ADFS, pscramble and the Task Manager.

Version 0.17. Not tagged

No longer complains if last scatter list entry isn't a multiple of the transfer unit size, as long as we're using a unit-multiple subpart of the entry. Needed by ADFS.

Version 0.17. Not tagged

Detail:
  Features:
  * Supports physical channels using a new type of HAL device (&303) where
    the hardware takes a pointer to a scatter list up front to effect the
    entire transfer, and which may report hardware errors at any time.
    For full details, see Doc.HAL_DMAAPI.
  * Page tables are no longer generated when using a bounce-buffer, since
    they aren't actually any use except in determining if unsafe pages
    intersect with the transfer. Instead, bounce-buffer transfers are halted
    whenever any page becomes unsafe, and no new bounce-buffer transfers can
    be started while any pages are unsafe.
  * For both bounce-buffer and list-type devices, the Activate entry is
    now called after the transfer is programmed (with SetCurrentTransfer
    or SetListTransfer), since this is likely to be more useful for these
    types of devices. Note that for interrupt-driven buffer devices, it
    is still assumed that the default state of the device is to be in an
    underrun interrupt state, so the transfer is not programmed until the
    first time the interrupt routine is entered.
  * Now ensures that scatter list start addresses and lengths are all
    multiples of the transfer unit size (this check is done during page
    table construction, so will be returned via the Completed callback,
    rather than directly from DMA_QueueTransfer like the total length check).
  * When there is no interrupt associated with a physical channel (both for
    buffer- or list-type devices), if the polling of the hardware within SWI
    DMA_ExamineTransfer detects that the transfer has completed, then
    the transfer is shut down in the same was as though an interrupt had
    happened (ie the Completed callback is called without an error).
  * Added SWI DMA_AllocateLogicalChannels to act as a centralised allocator
    for session-unique logical DMA channels. Primarily a service to
    third-party authors of "HAL" devices.
  Bugfixes:
  * Now handles Service_MessageFileClosed to allow safe MessageTrans lookups
    in the background.
  * If a controller device was removed before its constituent channel
    devices, removing the channel device would abort.
  * Interrupt trampoline construction was wrong, the address of the
    interrupt handler code wasn't relocated. Also, non-vectored interrupts
    weren't correctly handled.
  * Transfers suspended by Service_PagesUnsafe weren't blocking the logical
    channel, so you could queue another transfer on the same channel and it
    would start immediately (although transfers already queued before the
    service call wouldn't have been started).
  * Transfers which had completed within Service_PagesUnsafe handler weren't
    being delinked from the queue.
  * Service_PagesSafe wasn't handled properly, meaning that none of the
    transfers that were halted by Service_PagesUnsafe were restarted! This
    was due to the use of routines in common with DMA_SuspendTransfer and
    DMA_ResumeTransfer which started with the first HALised version - these
    suspend a logical channel, unlike the original IOMD code which
    effectively suspends a physical channel. Thus the code to search for
    transfers to restart needs to be quite different, since the physical
    channel may be idle, or occupied with another logical channel, and
    each physical channel may have multiple associated logical channels
    that were on unsafe transfers!
  * When you deregistered a logical channel (including during module death)
    the Completed callback was called for all queued transfers, incorrectly
    including those that had not yet had the matching Start callback called.
  * There was a stack imbalance in the buffer device interrupt handler
    because the retained part of the old code assumed Hdr:Proc stack
    management, which the new code doesn't use (it's entered with the return
    address on the stack).
Admin:
  Tested with UDMA test harnesses on Tungsten (ADFS not yet complete).

Version 0.17. Tagged as 'DMA-0_17'

Version 0.16. Tagged as 'DMA-0_16'

Detail:
  Also
  * added Ursula service call table (not even the Phoebe version had one!)
  * fixed a bug: the Service_PagesUnsafe handler was returning with IRQs
    still disabled
Admin:
  Not tested.

Version 0.15. Tagged as 'DMA-0_15'

  Updated to build using objasm instead of aasm.
  Sources changed to be objasm-compatible.
Admin:
  Requires Library 0.71 or later.
  Requires BuildSys 3.06 or later.
  Requires Env 0.65 or later.

Version 0.14. Tagged as 'DMA-0_14'

Detail:
	A change in Version 0.52 of HdrSrc meant that this component would
not build. This has now been fixed.
	Makefile now uses strdepend in the clean phase to remove all the
dynamic dependencies.
	Placed under srccommit control.

Admin:
	Built.

Version 0.13. Tagged as 'DMA-0_13'