kernel/drivers/gpu/drm/radeon/evergreen_dma.c

   1 /*
   2  * Copyright 2010 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * Authors: Alex Deucher
  23  */
  24 #include <drm/drmP.h>
  25 #include "radeon.h"
  26 #include "radeon_asic.h"
  27 #include "evergreend.h"
  28
  29 u32 evergreen_gpu_check_soft_reset(struct radeon_device *rdev);
  30
  31 /**
  32  * evergreen_dma_fence_ring_emit - emit a fence on the DMA ring
  33  *
  34  * @rdev: radeon_device pointer
  35  * @fence: radeon fence object
  36  *
  37  * Add a DMA fence packet to the ring to write
  38  * the fence seq number and DMA trap packet to generate
  39  * an interrupt if needed (evergreen-SI).
  40  */
  41 void evergreen_dma_fence_ring_emit(struct radeon_device *rdev,
  42                                    struct radeon_fence *fence)
  43 {
  44         struct radeon_ring *ring = &rdev->ring[fence->ring];
  45         u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
  46         /* write the fence */
  47         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0));
  48         radeon_ring_write(ring, addr & 0xfffffffc);
  49         radeon_ring_write(ring, (upper_32_bits(addr) & 0xff));
  50         radeon_ring_write(ring, fence->seq);
  51         /* generate an interrupt */
  52         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_TRAP, 0, 0));
  53         /* flush HDP */
  54         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0));
  55         radeon_ring_write(ring, (0xf << 16) | (HDP_MEM_COHERENCY_FLUSH_CNTL >> 2));
  56         radeon_ring_write(ring, 1);
  57 }
  58
  59 /**
  60  * evergreen_dma_ring_ib_execute - schedule an IB on the DMA engine
  61  *
  62  * @rdev: radeon_device pointer
  63  * @ib: IB object to schedule
  64  *
  65  * Schedule an IB in the DMA ring (evergreen).
  66  */
  67 void evergreen_dma_ring_ib_execute(struct radeon_device *rdev,
  68                                    struct radeon_ib *ib)
  69 {
  70         struct radeon_ring *ring = &rdev->ring[ib->ring];
  71
  72         if (rdev->wb.enabled) {
  73                 u32 next_rptr = ring->wptr + 4;
  74                 while ((next_rptr & 7) != 5)
  75                         next_rptr++;
  76                 next_rptr += 3;
  77                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 1));
  78                 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
  79                 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xff);
  80                 radeon_ring_write(ring, next_rptr);
  81         }
  82
  83         /* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
  84          * Pad as necessary with NOPs.
  85          */
  86         while ((ring->wptr & 7) != 5)
  87                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0));
  88         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_INDIRECT_BUFFER, 0, 0));
  89         radeon_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
  90         radeon_ring_write(ring, (ib->length_dw << 12) | (upper_32_bits(ib->gpu_addr) & 0xFF));
  91
  92 }
  93
  94 /**
  95  * evergreen_copy_dma - copy pages using the DMA engine
  96  *
  97  * @rdev: radeon_device pointer
  98  * @src_offset: src GPU address
  99  * @dst_offset: dst GPU address
 100  * @num_gpu_pages: number of GPU pages to xfer
 101  * @fence: radeon fence object
 102  *
 103  * Copy GPU paging using the DMA engine (evergreen-cayman).
 104  * Used by the radeon ttm implementation to move pages if
 105  * registered as the asic copy callback.
 106  */
 107 struct radeon_fence *evergreen_copy_dma(struct radeon_device *rdev,
 108                                         uint64_t src_offset,
 109                                         uint64_t dst_offset,
 110                                         unsigned num_gpu_pages,
 111                                         struct reservation_object *resv)
 112 {
 113         struct radeon_fence *fence;
 114         struct radeon_sync sync;
 115         int ring_index = rdev->asic->copy.dma_ring_index;
 116         struct radeon_ring *ring = &rdev->ring[ring_index];
 117         u32 size_in_dw, cur_size_in_dw;
 118         int i, num_loops;
 119         int r = 0;
 120
 121         radeon_sync_create(&sync);
 122
 123         size_in_dw = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT) / 4;
 124         num_loops = DIV_ROUND_UP(size_in_dw, 0xfffff);
 125         r = radeon_ring_lock(rdev, ring, num_loops * 5 + 11);
 126         if (r) {
 127                 DRM_ERROR("radeon: moving bo (%d).\n", r);
 128                 radeon_sync_free(rdev, &sync, NULL);
 129                 return ERR_PTR(r);
 130         }
 131
 132         radeon_sync_resv(rdev, &sync, resv, false);
 133         radeon_sync_rings(rdev, &sync, ring->idx);
 134
 135         for (i = 0; i < num_loops; i++) {
 136                 cur_size_in_dw = size_in_dw;
 137                 if (cur_size_in_dw > 0xFFFFF)
 138                         cur_size_in_dw = 0xFFFFF;
 139                 size_in_dw -= cur_size_in_dw;
 140                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_COPY, 0, cur_size_in_dw));
 141                 radeon_ring_write(ring, dst_offset & 0xfffffffc);
 142                 radeon_ring_write(ring, src_offset & 0xfffffffc);
 143                 radeon_ring_write(ring, upper_32_bits(dst_offset) & 0xff);
 144                 radeon_ring_write(ring, upper_32_bits(src_offset) & 0xff);
 145                 src_offset += cur_size_in_dw * 4;
 146                 dst_offset += cur_size_in_dw * 4;
 147         }
 148
 149         r = radeon_fence_emit(rdev, &fence, ring->idx);
 150         if (r) {
 151                 radeon_ring_unlock_undo(rdev, ring);
 152                 radeon_sync_free(rdev, &sync, NULL);
 153                 return ERR_PTR(r);
 154         }
 155
 156         radeon_ring_unlock_commit(rdev, ring, false);
 157         radeon_sync_free(rdev, &sync, fence);
 158
 159         return fence;
 160 }
 161
 162 /**
 163  * evergreen_dma_is_lockup - Check if the DMA engine is locked up
 164  *
 165  * @rdev: radeon_device pointer
 166  * @ring: radeon_ring structure holding ring information
 167  *
 168  * Check if the async DMA engine is locked up.
 169  * Returns true if the engine appears to be locked up, false if not.
 170  */
 171 bool evergreen_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
 172 {
 173         u32 reset_mask = evergreen_gpu_check_soft_reset(rdev);
 174
 175         if (!(reset_mask & RADEON_RESET_DMA)) {
 176                 radeon_ring_lockup_update(rdev, ring);
 177                 return false;
 178         }
 179         return radeon_ring_test_lockup(rdev, ring);
 180 }
 181
 182