Merge tag 'drm-next-2019-07-16' of git://anongit.freedesktop.org/drm/drm

Pull drm updates from Dave Airlie:
 "The biggest thing in this is the AMD Navi GPU support, this again
  contains a bunch of header files that are large. These are the new AMD
  RX5700 GPUs that just recently became available.

  New drivers:
   - ST-Ericsson MCDE driver
   - Ingenic JZ47xx SoC

  UAPI change:
   - HDR source metadata property

  Core:
   - HDR inforframes and EDID parsing
   - drm hdmi infoframe unpacking
   - remove prime sg_table caching into dma-buf
   - New gem vram helpers to reduce driver code
   - Lots of drmP.h removal
   - reservation fencing fix
   - documentation updates
   - drm_fb_helper_connector removed
   - mode name command handler rewrite

  fbcon:
   - Remove the fbcon notifiers

  ttm:
   - forward progress fixes

  dma-buf:
   - make mmap call optional
   - debugfs refcount fixes
   - dma-fence free with pending signals fix
   - each dma-buf gets an inode

  Panels:
   - Lots of additional panel bindings

  amdgpu:
   - initial navi10 support
   - avoid hw reset
   - HDR metadata support
   - new thermal sensors for vega asics
   - RAS fixes
   - use HMM rather than MMU notifier
   - xgmi topology via kfd
   - SR-IOV fixes
   - driver reload fixes
   - DC use a core bpc attribute
   - Aux fixes for DC
   - Bandwidth calc updates for DC
   - Clock handling refactor
   - kfd VEGAM support

  vmwgfx:
   - Coherent memory support changes

  i915:
   - HDR Support
   - HDMI i2c link
   - Icelake multi-segmented gamma support
   - GuC firmware update
   - Mule Creek Canyon PCH support for EHL
   - EHL platform updtes
   - move i915.alpha_support to i915.force_probe
   - runtime PM refactoring
   - VBT parsing refactoring
   - DSI fixes
   - struct mutex dependency reduction
   - GEM code reorg

  mali-dp:
   - Komeda driver features

  msm:
   - dsi vs EPROBE_DEFER fixes
   - msm8998 snapdragon 835 support
   - a540 gpu support
   - mdp5 and dpu interconnect support

  exynos:
   - drmP.h removal

  tegra:
   - misc fixes

  tda998x:
   - audio support improvements
   - pixel repeated mode support
   - quantisation range handling corrections
   - HDMI vendor info fix

  armada:
   - interlace support fix
   - overlay/video plane register handling refactor
   - add gamma support

  rockchip:
   - RX3328 support

  panfrost:
   - expose perf counters via hidden ioctls

  vkms:
   - enumerate CRC sources list

  ast:
   - rework BO handling

  mgag200:
   - rework BO handling

  dw-hdmi:
   - suspend/resume support

  rcar-du:
   - R8A774A1 Soc Support
   - LVDS dual-link mode support
   - Additional formats
   - Misc fixes

  omapdrm:
   - DSI command mode display support

  stm
   - fb modifier support
   - runtime PM support

  sun4i:
   - use vmap ops

  vc4:
   - binner bo binding rework

  v3d:
   - compute shader support
   - resync/sync fixes
   - job management refactoring

  lima:
   - NULL pointer in irq handler fix
   - scheduler default timeout

  virtio:
   - fence seqno support
   - trace events

  bochs:
   - misc fixes

  tc458767:
   - IRQ/HDP handling

  sii902x:
   - HDMI audio support

  atmel-hlcdc:
   - misc fixes

  meson:
   - zpos support"

* tag 'drm-next-2019-07-16' of git://anongit.freedesktop.org/drm/drm: (1815 commits)
  Revert "Merge branch 'vmwgfx-next' of git://people.freedesktop.org/~thomash/linux into drm-next"
  Revert "mm: adjust apply_to_pfn_range interface for dropped token."
  mm: adjust apply_to_pfn_range interface for dropped token.
  drm/amdgpu/navi10: add uclk activity sensor
  drm/amdgpu: properly guard the generic discovery code
  drm/amdgpu: add missing documentation on new module parameters
  drm/amdgpu: don't invalidate caches in RELEASE_MEM, only do the writeback
  drm/amd/display: avoid 64-bit division
  drm/amdgpu/psp11: simplify the ucode register logic
  drm/amdgpu: properly guard DC support in navi code
  drm/amd/powerplay: vega20: fix uninitialized variable use
  drm/amd/display: dcn20: include linux/delay.h
  amdgpu: make pmu support optional
  drm/amd/powerplay: Zero initialize current_rpm in vega20_get_fan_speed_percent
  drm/amd/powerplay: Zero initialize freq in smu_v11_0_get_current_clk_freq
  drm/amd/powerplay: Use memset to initialize metrics structs
  drm/amdgpu/mes10.1: Fix header guard
  drm/amd/powerplay: add temperature sensor support for navi10
  drm/amdgpu: fix scheduler timeout calc
  drm/amdgpu: Prepare for hmm_range_register API change (v2)
  ...

1 files changed, 183 insertions, 2 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
index 8930d66f2204..20ce158490db 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
@@ -24,8 +24,11 @@
  */
 
 #include <linux/kthread.h>
-#include <drm/drmP.h>
-#include <linux/debugfs.h>
+#include <linux/pci.h>
+#include <linux/uaccess.h>
+
+#include <drm/drm_debugfs.h>
+
 #include "amdgpu.h"
 
 /**
@@ -920,17 +923,195 @@ static const struct drm_info_list amdgpu_debugfs_list[] = {
 	{"amdgpu_evict_gtt", &amdgpu_debugfs_evict_gtt},
 };
 
+static void amdgpu_ib_preempt_fences_swap(struct amdgpu_ring *ring,
+					  struct dma_fence **fences)
+{
+	struct amdgpu_fence_driver *drv = &ring->fence_drv;
+	uint32_t sync_seq, last_seq;
+
+	last_seq = atomic_read(&ring->fence_drv.last_seq);
+	sync_seq = ring->fence_drv.sync_seq;
+
+	last_seq &= drv->num_fences_mask;
+	sync_seq &= drv->num_fences_mask;
+
+	do {
+		struct dma_fence *fence, **ptr;
+
+		++last_seq;
+		last_seq &= drv->num_fences_mask;
+		ptr = &drv->fences[last_seq];
+
+		fence = rcu_dereference_protected(*ptr, 1);
+		RCU_INIT_POINTER(*ptr, NULL);
+
+		if (!fence)
+			continue;
+
+		fences[last_seq] = fence;
+
+	} while (last_seq != sync_seq);
+}
+
+static void amdgpu_ib_preempt_signal_fences(struct dma_fence **fences,
+					    int length)
+{
+	int i;
+	struct dma_fence *fence;
+
+	for (i = 0; i < length; i++) {
+		fence = fences[i];
+		if (!fence)
+			continue;
+		dma_fence_signal(fence);
+		dma_fence_put(fence);
+	}
+}
+
+static void amdgpu_ib_preempt_job_recovery(struct drm_gpu_scheduler *sched)
+{
+	struct drm_sched_job *s_job;
+	struct dma_fence *fence;
+
+	spin_lock(&sched->job_list_lock);
+	list_for_each_entry(s_job, &sched->ring_mirror_list, node) {
+		fence = sched->ops->run_job(s_job);
+		dma_fence_put(fence);
+	}
+	spin_unlock(&sched->job_list_lock);
+}
+
+static void amdgpu_ib_preempt_mark_partial_job(struct amdgpu_ring *ring)
+{
+	struct amdgpu_job *job;
+	struct drm_sched_job *s_job;
+	uint32_t preempt_seq;
+	struct dma_fence *fence, **ptr;
+	struct amdgpu_fence_driver *drv = &ring->fence_drv;
+	struct drm_gpu_scheduler *sched = &ring->sched;
+
+	if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
+		return;
+
+	preempt_seq = le32_to_cpu(*(drv->cpu_addr + 2));
+	if (preempt_seq <= atomic_read(&drv->last_seq))
+		return;
+
+	preempt_seq &= drv->num_fences_mask;
+	ptr = &drv->fences[preempt_seq];
+	fence = rcu_dereference_protected(*ptr, 1);
+
+	spin_lock(&sched->job_list_lock);
+	list_for_each_entry(s_job, &sched->ring_mirror_list, node) {
+		job = to_amdgpu_job(s_job);
+		if (job->fence == fence)
+			/* mark the job as preempted */
+			job->preemption_status |= AMDGPU_IB_PREEMPTED;
+	}
+	spin_unlock(&sched->job_list_lock);
+}
+
+static int amdgpu_debugfs_ib_preempt(void *data, u64 val)
+{
+	int r, resched, length;
+	struct amdgpu_ring *ring;
+	struct dma_fence **fences = NULL;
+	struct amdgpu_device *adev = (struct amdgpu_device *)data;
+
+	if (val >= AMDGPU_MAX_RINGS)
+		return -EINVAL;
+
+	ring = adev->rings[val];
+
+	if (!ring || !ring->funcs->preempt_ib || !ring->sched.thread)
+		return -EINVAL;
+
+	/* the last preemption failed */
+	if (ring->trail_seq != le32_to_cpu(*ring->trail_fence_cpu_addr))
+		return -EBUSY;
+
+	length = ring->fence_drv.num_fences_mask + 1;
+	fences = kcalloc(length, sizeof(void *), GFP_KERNEL);
+	if (!fences)
+		return -ENOMEM;
+
+	/* stop the scheduler */
+	kthread_park(ring->sched.thread);
+
+	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
+
+	/* preempt the IB */
+	r = amdgpu_ring_preempt_ib(ring);
+	if (r) {
+		DRM_WARN("failed to preempt ring %d\n", ring->idx);
+		goto failure;
+	}
+
+	amdgpu_fence_process(ring);
+
+	if (atomic_read(&ring->fence_drv.last_seq) !=
+	    ring->fence_drv.sync_seq) {
+		DRM_INFO("ring %d was preempted\n", ring->idx);
+
+		amdgpu_ib_preempt_mark_partial_job(ring);
+
+		/* swap out the old fences */
+		amdgpu_ib_preempt_fences_swap(ring, fences);
+
+		amdgpu_fence_driver_force_completion(ring);
+
+		/* resubmit unfinished jobs */
+		amdgpu_ib_preempt_job_recovery(&ring->sched);
+
+		/* wait for jobs finished */
+		amdgpu_fence_wait_empty(ring);
+
+		/* signal the old fences */
+		amdgpu_ib_preempt_signal_fences(fences, length);
+	}
+
+failure:
+	/* restart the scheduler */
+	kthread_unpark(ring->sched.thread);
+
+	ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
+
+	if (fences)
+		kfree(fences);
+
+	return 0;
+}
+
+DEFINE_SIMPLE_ATTRIBUTE(fops_ib_preempt, NULL,
+			amdgpu_debugfs_ib_preempt, "%llu\n");
+
 int amdgpu_debugfs_init(struct amdgpu_device *adev)
 {
+	adev->debugfs_preempt =
+		debugfs_create_file("amdgpu_preempt_ib", 0600,
+				    adev->ddev->primary->debugfs_root,
+				    (void *)adev, &fops_ib_preempt);
+	if (!(adev->debugfs_preempt)) {
+		DRM_ERROR("unable to create amdgpu_preempt_ib debugsfs file\n");
+		return -EIO;
+	}
+
 	return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_list,
 					ARRAY_SIZE(amdgpu_debugfs_list));
 }
 
+void amdgpu_debugfs_preempt_cleanup(struct amdgpu_device *adev)
+{
+	if (adev->debugfs_preempt)
+		debugfs_remove(adev->debugfs_preempt);
+}
+
 #else
 int amdgpu_debugfs_init(struct amdgpu_device *adev)
 {
 	return 0;
 }
+void amdgpu_debugfs_preempt_cleanup(struct amdgpu_device *adev) { }
 int amdgpu_debugfs_regs_init(struct amdgpu_device *adev)
 {
 	return 0;