• Daniel Vetter's avatar
    dma-buf: mmap support · 4c78513e
    Daniel Vetter authored
    Compared to Rob Clark's RFC I've ditched the prepare/finish hooks
    and corresponding ioctls on the dma_buf file. The major reason for
    that is that many people seem to be under the impression that this is
    also for synchronization with outstanding asynchronous processsing.
    I'm pretty massively opposed to this because:
    
    - It boils down reinventing a new rather general-purpose userspace
      synchronization interface. If we look at things like futexes, this
      is hard to get right.
    - Furthermore a lot of kernel code has to interact with this
      synchronization primitive. This smells a look like the dri1 hw_lock,
      a horror show I prefer not to reinvent.
    - Even more fun is that multiple different subsystems would interact
      here, so we have plenty of opportunities to create funny deadlock
      scenarios.
    
    I think synchronization is a wholesale different problem from data
    sharing and should be tackled as an orthogonal problem.
    
    Now we could demand that prepare/finish may only ensure cache
    coherency (as Rob intended), but that runs up into the next problem:
    We not only need mmap support to facilitate sw-only processing nodes
    in a pipeline (without jumping through hoops by importing the dma_buf
    into some sw-access only importer), which allows for a nicer
    ION->dma-buf upgrade path for existing Android userspace. We also need
    mmap support for existing importing subsystems to support existing
    userspace libraries. And a loot of these subsystems are expected to
    export coherent userspace mappings.
    
    So prepare/finish can only ever be optional and the exporter /needs/
    to support coherent mappings. Given that mmap access is always
    somewhat fallback-y in nature I've decided to drop this optimization,
    instead of just making it optional. If we demonstrate a clear need for
    this, supported by benchmark results, we can always add it in again
    later as an optional extension.
    
    Other differences compared to Rob's RFC is the above mentioned support
    for mapping a dma-buf through facilities provided by the importer.
    Which results in mmap support no longer being optional.
    
    Note that this dma-buf mmap patch does _not_ support every possible
    insanity an existing subsystem could pull of with mmap: Because it
    does not allow to intercept pagefaults and shoot down ptes importing
    subsystems can't add some magic of their own at these points (e.g. to
    automatically synchronize with outstanding rendering or set up some
    special resources). I've done a cursory read through a few mmap
    implementions of various subsytems and I'm hopeful that we can avoid
    this (and the complexity it'd bring with it).
    
    Additonally I've extended the documentation a bit to explain the hows
    and whys of this mmap extension.
    
    In case we ever want to add support for explicitly cache maneged
    userspace mmap with a prepare/finish ioctl pair, we could specify that
    userspace needs to mmap a different part of the dma_buf, e.g. the
    range starting at dma_buf->size up to dma_buf->size*2. This works
    because the size of a dma_buf is invariant over it's lifetime. The
    exporter would obviously need to fall back to coherent mappings for
    both ranges if a legacy clients maps the coherent range and the
    architecture cannot suppor conflicting caching policies. Also, this
    would obviously be optional and userspace needs to be able to fall
    back to coherent mappings.
    
    v2:
    - Spelling fixes from Rob Clark.
    - Compile fix for !DMA_BUF from Rob Clark.
    - Extend commit message to explain how explicitly cache managed mmap
      support could be added later.
    - Extend the documentation with implementations notes for exporters
      that need to manually fake coherency.
    
    v3:
    - dma_buf pointer initialization goof-up noticed by Rebecca Schultz
      Zavin.
    
    Cc: Rob Clark <rob.clark@linaro.org>
    Cc: Rebecca Schultz Zavin <rebecca@android.com>
    Acked-by: default avatarRob Clark <rob.clark@linaro.org>
    Signed-Off-by: default avatarDaniel Vetter <daniel.vetter@ffwll.ch>
    Signed-off-by: default avatarSumit Semwal <sumit.semwal@linaro.org>
    4c78513e
dma-buf-sharing.txt 19 KB