1. 03 Dec, 2015 13 commits
    • Nick Hoath's avatar
      drm/i915: Extend LRC pinning to cover GPU context writeback · 6d65ba94
      Nick Hoath authored
      Use the first retired request on a new context to unpin
      the old context. This ensures that the hw context remains
      bound until it has been written back to by the GPU.
      Now that the context is pinned until later in the request/context
      lifecycle, it no longer needs to be pinned from context_queue to
      retire_requests.
      This fixes an issue with GuC submission where the GPU might not
      have finished writing back the context before it is unpinned. This
      results in a GPU hang.
      
      v2: Moved the new pin to cover GuC submission (Alex Dai)
          Moved the new unpin to request_retire to fix coverage leak
      v3: Added switch to default context if freeing a still pinned
          context just in case the hw was actually still using it
      v4: Unwrapped context unpin to allow calling without a request
      v5: Only create a switch to idle context if the ring doesn't
          already have a request pending on it (Alex Dai)
          Rename unsaved to dirty to avoid double negatives (Dave Gordon)
          Changed _no_req postfix to __ prefix for consistency (Dave Gordon)
          Split out per engine cleanup from context_free as it
          was getting unwieldy
          Corrected locking (Dave Gordon)
      v6: Removed some bikeshedding (Mika Kuoppala)
          Added explanation of the GuC hang that this fixes (Daniel Vetter)
      v7: Removed extra per request pinning from ring reset code (Alex Dai)
          Added forced ring unpin/clean in error case in context free (Alex Dai)
      Signed-off-by: default avatarNick Hoath <nicholas.hoath@intel.com>
      Issue: VIZ-4277
      Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
      Cc: David Gordon <david.s.gordon@intel.com>
      Cc: Chris Wilson <chris@chris-wilson.co.uk>
      Cc: Alex Dai <yu.dai@intel.com>
      Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
      Reviewed-by: default avatarAlex Dai <yu.dai@intel.com>
      Signed-off-by: default avatarDaniel Vetter <daniel.vetter@ffwll.ch>
      6d65ba94
    • Alex Dai's avatar
      drm/i915/guc: Clean up locks in GuC · 5a843307
      Alex Dai authored
      For now, remove the spinlocks that protected the GuC's
      statistics block and work queue; they are only accessed
      by code that already holds the global struct_mutex, and
      so are redundant (until the big struct_mutex rewrite!).
      
      The specific problem that the spinlocks caused was that
      if the work queue was full, the driver would try to
      spinwait for one jiffy, but with interrupts disabled the
      jiffy count would not advance, leading to a system hang.
      The issue was found using test case igt/gem_close_race.
      
      The new version will usleep() instead, still holding
      the struct_mutex but without any spinlocks.
      
      v4: Reorganize commit message (Dave Gordon)
      v3: Remove unnecessary whitespace churn
      v2: Clean up wq_lock too
      v1: Clean up host2guc lock as well
      Signed-off-by: default avatarAlex Dai <yu.dai@intel.com>
      Reviewed-by: default avatarDave Gordon <david.s.gordon@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/1449104189-27591-1-git-send-email-yu.dai@intel.comSigned-off-by: default avatarDaniel Vetter <daniel.vetter@ffwll.ch>
      5a843307
    • Paulo Zanoni's avatar
      drm/i915: only recompress FBC after flushing a drawing operation · ee7d6cfa
      Paulo Zanoni authored
      There's no need to stop and restart FBC, which is quite expensive as
      we have to revalidate the CRTC state. After flushing a drawing
      operation we know the CRTC state hasn't changed, so a nuke
      (recompress) should be fine.
      
      v2: Make it simpler (Chris).
      v3: Rewrite the patch again due to patch order changes.
      v4: Rewrite commit message (Chris).
      Reviewed-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
      Signed-off-by: default avatarPaulo Zanoni <paulo.r.zanoni@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/
      ee7d6cfa
    • Paulo Zanoni's avatar
      drm/i915: get rid of FBC {,de}activation messages · 820bcabb
      Paulo Zanoni authored
      When running Cinnamon I see way too many pairs of these messages: many
      per second. Get rid of them as they're just telling us FBC is working
      as expected. We already have the messages for enable/disable, so we
      don't really need messages for activation/deactivation.
      
      v2: Rebase after changing the patch order.
      Reviewed-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
      Signed-off-by: default avatarPaulo Zanoni <paulo.r.zanoni@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/
      820bcabb
    • Paulo Zanoni's avatar
      drm/i915: kill fbc.uncompressed_size · 559d9135
      Paulo Zanoni authored
      Directly call intel_fbc_calculate_cfb_size() in the only place that
      actually needs it, and use the proper check before removing the stolen
      node. IMHO, this change makes our code easier to understand.
      
      v2: Use drm_mm_node_allocated() (Chris).
      Reviewed-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
      Signed-off-by: default avatarPaulo Zanoni <paulo.r.zanoni@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/
      559d9135
    • Paulo Zanoni's avatar
      drm/i915: use a single intel_fbc_work struct · 128d7356
      Paulo Zanoni authored
      This was already on my TODO list, and was requested both by Chris and
      Ville, for different reasons. The advantages are avoiding a frequent
      malloc/free pair, and the locality of having the work structure
      embedded in dev_priv. The maximum used memory is also smaller since
      previously we could have multiple allocated intel_fbc_work structs at
      the same time, and now we'll always have a single one - the one
      embedded on dev_priv. Of course, we're now using a little more memory
      on the cases where there's nothing scheduled.
      
      The biggest challenge here is to keep everything synchronized the way
      it was before.
      
      Currently, when we try to activate FBC, we allocate a new
      intel_fbc_work structure. Then later when we conclude we must delay
      the FBC activation a little more, we allocate a new intel_fbc_work
      struct, and then adjust dev_priv->fbc.fbc_work to point to the new
      struct. So when the old work runs - at intel_fbc_work_fn() - it will
      check that dev_priv->fbc.fbc_work points to something else, so it does
      nothing. Everything is also protected by fbc.lock.
      
      Just cancelling the old delayed work doesn't work because we might
      just cancel it after the work function already started to run, but
      while it is still waiting to grab fbc.lock. That's why we use the
      "dev_priv->fbc.fbc_work == work" check described in the paragraph
      above.
      
      So now that we have a single work struct we have to introduce a new
      way to synchronize everything. So we're making the work function a
      normal work instead of a delayed work, and it will be responsible for
      sleeping the appropriate amount of time itself. This way, after it
      wakes up it can grab the lock, ask "were we delayed or cancelled?" and
      then go back to sleep, enable FBC or give up.
      
      v2:
        - Spelling fixes.
        - Rebase after changing the patch order.
        - Fix ms/jiffies confusion.
      
      Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1)
      Signed-off-by: default avatarPaulo Zanoni <paulo.r.zanoni@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/
      128d7356
    • Paulo Zanoni's avatar
      drm/i915: check for FBC planes in the same place as the pipes · e6cd6dc1
      Paulo Zanoni authored
      This moves the pre-gen4 check from update() to enable(). The HAS_DDI
      in the original code is not needed since only gen 2/3 have the plane
      swapping code.
      
      v2: Rebase.
      v3: Extract fbc_on_plane_a_only() (Chris).
      Reviewed-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
      Signed-off-by: default avatarPaulo Zanoni <paulo.r.zanoni@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/
      e6cd6dc1
    • Paulo Zanoni's avatar
      drm/i915: alloc/free the FBC CFB during enable/disable · c5ecd469
      Paulo Zanoni authored
      One of the problems with the current code is that it frees the CFB and
      releases its drm_mm node as soon as we flip FBC's enable bit. This is
      bad because after we disable FBC the hardware may still use the CFB
      for the rest of the frame, so in theory we should only release the
      drm_mm node one frame after we disable FBC. Otherwise, a stolen memory
      allocation done right after an FBC disable may result in either
      corrupted memory for the new owner of that memory region or corrupted
      screen/underruns in case the new owner changes it while the hardware
      is still reading it. This case is not exactly easy to reproduce since
      we currently don't do a lot of stolen memory allocations, but I see
      patches on the mailing list trying to expose stolen memory to user
      space, so races will be possible.
      
      I thought about three different approaches to solve this, and they all
      have downsides.
      
      The first approach would be to simply use multiple drm_mm nodes and
      freeing the unused ones only after a frame has passed. The problem
      with this approach is that since stolen memory is rather small,
      there's a risk we just won't be able to allocate a new CFB from stolen
      if the previous one was not freed yet. This could happen in case we
      quickly disable FBC from pipe A and decide to enable it on pipe B, or
      just if we change pipe A's fb stride while FBC is enabled.
      
      The second approach would be similar to the first one, but maintaining
      a single drm_mm node and keeping track of when it can be reused. This
      would remove the disadvantage of not having enough space for two
      nodes, but would create the new problem where we may not be able to
      enable FBC at the point intel_fbc_update() is called, so we would have
      to add more code to retry updating FBC after the time has passed. And
      that can quickly get too complex since we can get invalidate, flush,
      disable and other calls in the middle of the wait.
      
      Both solutions above - and also the current code - have the problem
      that we unnecessarily free+realloc FBC during invalidate+flush
      operations even if the CFB size doesn't change.
      
      The third option would be to move the allocation/deallocation to
      enable/disable. This makes sure that the pipe is always disabled when
      we allocate/deallocate the CFB, so there's no risk that the FBC
      hardware may read or write to the memory right after it is freed from
      drm_mm. The downside is that it is possible for user space to change
      the buffer stride without triggering a disable/enable - only
      deactivate/activate -, so we'll have to handle this case somehow - see
      igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It
      could be possible to implement a way to free+alloc the CFB during said
      stride change, but it would involve a lot of book-keeping - exactly as
      mentioned above - just for on case, so for now I'll keep it simple and
      just deactivate FBC. Besides, we may not even need to disable FBC
      since we do CFB over-allocation.
      
      Note from Chris: "Starting a fullscreen client that covers a single
      monitor in a multi-monitor setup will trigger a change in stride on
      one of the CRTCs (the monitors will be flipped independently).". It
      shouldn't be a huge problem if we lose FBC on multi-monitor setups
      since these setups already have problems reaching deep PC states
      anyway.
      
      v2: Rebase after changing the patch order.
      v3:
        - Remove references to the stride change case being "uncommon" and
          paste Chris' example.
        - Rebase after a change in a previous patch.
      Reviewed-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
      Signed-off-by: default avatarPaulo Zanoni <paulo.r.zanoni@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/
      c5ecd469
    • Paulo Zanoni's avatar
      drm/i915: introduce intel_fbc_{enable,disable} · d029bcad
      Paulo Zanoni authored
      The goal is to call FBC enable/disable only once per modeset, while
      activate/deactivate/update will be called multiple times.
      
      The enable() function will be responsible for deciding if a CRTC will
      have FBC on it and then it will "lock" FBC on this CRTC: it won't be
      possible to change FBC's CRTC until disable(). With this, all checks
      and resource acquisition that only need to be done once per modeset
      can be moved from update() to enable(). And then the update(),
      activate() and deactivate() code will also get simpler since they
      won't need to worry about the CRTC being changed.
      
      The disable() function will do the reverse operation of enable(). One
      of its features is that it should only be called while the pipe is
      already off. This guarantees that FBC is stopped and nothing is
      using the CFB.
      
      With this, the activate() and deactivate() functions just start and
      temporarily stop FBC. They are the ones touching the hardware enable
      bit, so HW state reflects dev_priv->crtc.active.
      
      The last function remaining is update(). A lot of times I thought
      about renaming update() to activate() or try_to_activate() since it's
      called when we want to activate FBC. The thing is that update() may
      not only decide to activate FBC, but also deactivate or keep it on the
      same state, so I'll leave this name for now.
      
      Moving code to enable() and disable() will also help in case we decide
      to move FBC to pipe_config or something else later.
      
      The current patch only puts the very basic code on enable() and
      disable(). The next commits will take care of moving more stuff from
      update() to the new functions.
      
      v2:
        - Rebase.
        - Improve commit message (Chris).
      v3: Rebase after changing the patch order.
      v4: Rebase again after upstream changes.
      Reviewed-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
      Signed-off-by: default avatarPaulo Zanoni <paulo.r.zanoni@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/
      d029bcad
    • Paulo Zanoni's avatar
      drm/i915: introduce is_active/activate/deactivate to the FBC terminology · 0e631adc
      Paulo Zanoni authored
      The long term goal is to have enable/disable as the higher level
      functions and activate/deactivate as the lower level functions, just
      like we do for PSR and for the CRTC. This way, we'll run enable and
      disable once per modeset, while update, activate and deactivate will
      be run many times. With this, we can move the checks and code that
      need to run only once per modeset to enable(), making the code simpler
      and possibly a little faster.
      
      This patch is just the first step on the conversion: it starts by
      converting the current low level functions from enable/disable to
      activate/deactivate. This patch by itself has no benefits other than
      making review and rebase easier. Please see the next patches for more
      details on the conversion.
      
      v2:
        - Rebase.
        - Improve commit message (Chris).
      v3: Rebase after changing the patch order.
      Reviewed-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
      Signed-off-by: default avatarPaulo Zanoni <paulo.r.zanoni@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/
      0e631adc
    • Paulo Zanoni's avatar
      drm/i915: pass the crtc as an argument to intel_fbc_update() · 754d1133
      Paulo Zanoni authored
      There's no need to reevaluate the status of every single crtc when a
      single crtc changes its state.
      
      With this, we're cutting the case where due to a change in pipe B,
      intel_fbc_update() is called, then intel_fbc_find_crtc() concludes FBC
      should be enabled on pipe A, then it completely rechecks the state of
      pipe A only to conclude FBC should remain enabled on pipe A. If any
      change on pipe A triggers a need to recompute whether FBC is valid on
      pipe A, then at some point someone is going to call
      intel_fbc_update(PIPE_A).
      
      The addition of intel_fbc_deactivate() is necessary so we keep track
      of the previously selected CRTC when we do invalidate/flush. We're
      also going to continue the enable/disable/activate/deactivate concept
      in the next patches.
      
      v2: Rebase.
      v3: Rebase after changing the patch order.
      Reviewed-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
      Signed-off-by: default avatarPaulo Zanoni <paulo.r.zanoni@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/
      754d1133
    • Paulo Zanoni's avatar
      drm/i915: set dev_priv->fbc.crtc before scheduling the enable work · e9c5fd26
      Paulo Zanoni authored
      This thing where we need to get the crtc either from the work
      structure or the fbc structure itself is confusing and unnecessary.
      Set fbc.crtc right when scheduling the enable work so we can always
      use it.
      
      The problem is not what gets passed and how to retrieve it. The
      problem is that when we're in the other parts of the code we always
      have to keep in mind that if FBC is already enabled we have to get the
      CRTC from place A, if FBC is scheduled we have to get the CRTC from
      place B, and if it's disabled there's no CRTC. Having a single place
      to retrieve the CRTC from allows us to treat the "is enabled" and "is
      scheduled" cases as the same case, reducing the mistake surface. I
      guess I should add this to the commit message.
      
      Besides the immediate advantages, this is also going to make one of
      the next commits much simpler. And even later, when we introduce
      enable/disable + activate/deactivate, this will be even simpler as
      we'll set the CRTC at enable time. So all the
      activate/deactivate/update code can just look at the single CRTC
      variable regardless of the current state.
      
      v2: Improve commit message (Chris).
      v3: Rebase after changing the patch order.
      Reviewed-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
      Signed-off-by: default avatarPaulo Zanoni <paulo.r.zanoni@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/
      e9c5fd26
    • Paulo Zanoni's avatar
      drm/i915: fix the CFB size check · 90d5234f
      Paulo Zanoni authored
      In function find_compression_threshold() we try to over-allocate CFB
      space in order to reduce reallocations and fragmentation, and we're
      not considering that at the CFB size check. Consider it.
      
      There is also a longer-term plan to kill
      dev_priv->fbc.uncompressed_size, but this will come later.
      
      v2: Use drm_mm_node_allocated() (Chris).
      Reviewed-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
      Signed-off-by: default avatarPaulo Zanoni <paulo.r.zanoni@intel.com>
      Link: http://patchwork.freedesktop.org/patch/msgid/
      90d5234f
  2. 02 Dec, 2015 15 commits
  3. 01 Dec, 2015 4 commits
  4. 30 Nov, 2015 4 commits
  5. 26 Nov, 2015 4 commits