sys/dev/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 /**
  30  * evergreen_dma_fence_ring_emit - emit a fence on the DMA ring
  31  *
  32  * @rdev: radeon_device pointer
  33  * @fence: radeon fence object
  34  *
  35  * Add a DMA fence packet to the ring to write
  36  * the fence seq number and DMA trap packet to generate
  37  * an interrupt if needed (evergreen-SI).
  38  */
  39 void evergreen_dma_fence_ring_emit(struct radeon_device *rdev,
  40                                    struct radeon_fence *fence)
  41 {
  42         struct radeon_ring *ring = &rdev->ring[fence->ring];
  43         u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
  44         /* write the fence */
  45         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0));
  46         radeon_ring_write(ring, addr & 0xfffffffc);
  47         radeon_ring_write(ring, (upper_32_bits(addr) & 0xff));
  48         radeon_ring_write(ring, fence->seq);
  49         /* generate an interrupt */
  50         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_TRAP, 0, 0));
  51         /* flush HDP */
  52         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0));
  53         radeon_ring_write(ring, (0xf << 16) | (HDP_MEM_COHERENCY_FLUSH_CNTL >> 2));
  54         radeon_ring_write(ring, 1);
  55 }
  56
  57 /**
  58  * evergreen_dma_ring_ib_execute - schedule an IB on the DMA engine
  59  *
  60  * @rdev: radeon_device pointer
  61  * @ib: IB object to schedule
  62  *
  63  * Schedule an IB in the DMA ring (evergreen).
  64  */
  65 void evergreen_dma_ring_ib_execute(struct radeon_device *rdev,
  66                                    struct radeon_ib *ib)
  67 {
  68         struct radeon_ring *ring = &rdev->ring[ib->ring];
  69
  70         if (rdev->wb.enabled) {
  71                 u32 next_rptr = ring->wptr + 4;
  72                 while ((next_rptr & 7) != 5)
  73                         next_rptr++;
  74                 next_rptr += 3;
  75                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 1));
  76                 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
  77                 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xff);
  78                 radeon_ring_write(ring, next_rptr);
  79         }
  80
  81         /* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
  82          * Pad as necessary with NOPs.
  83          */
  84         while ((ring->wptr & 7) != 5)
  85                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0));
  86         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_INDIRECT_BUFFER, 0, 0));
  87         radeon_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
  88         radeon_ring_write(ring, (ib->length_dw << 12) | (upper_32_bits(ib->gpu_addr) & 0xFF));
  89
  90 }
  91
  92 /**
  93  * evergreen_copy_dma - copy pages using the DMA engine
  94  *
  95  * @rdev: radeon_device pointer
  96  * @src_offset: src GPU address
  97  * @dst_offset: dst GPU address
  98  * @num_gpu_pages: number of GPU pages to xfer
  99  * @fence: radeon fence object
 100  *
 101  * Copy GPU paging using the DMA engine (evergreen-cayman).
 102  * Used by the radeon ttm implementation to move pages if
 103  * registered as the asic copy callback.
 104  */
 105 int evergreen_copy_dma(struct radeon_device *rdev,
 106                        uint64_t src_offset, uint64_t dst_offset,
 107                        unsigned num_gpu_pages,
 108                        struct radeon_fence **fence)
 109 {
 110         struct radeon_semaphore *sem = NULL;
 111         int ring_index = rdev->asic->copy.dma_ring_index;
 112         struct radeon_ring *ring = &rdev->ring[ring_index];
 113         u32 size_in_dw, cur_size_in_dw;
 114         int i, num_loops;
 115         int r = 0;
 116
 117         r = radeon_semaphore_create(rdev, &sem);
 118         if (r) {
 119                 DRM_ERROR("radeon: moving bo (%d).\n", r);
 120                 return r;
 121         }
 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_semaphore_free(rdev, &sem, NULL);
 129                 return r;
 130         }
 131
 132         if (radeon_fence_need_sync(*fence, ring->idx)) {
 133                 radeon_semaphore_sync_rings(rdev, sem, (*fence)->ring,
 134                                             ring->idx);
 135                 radeon_fence_note_sync(*fence, ring->idx);
 136         } else {
 137                 radeon_semaphore_free(rdev, &sem, NULL);
 138         }
 139
 140         for (i = 0; i < num_loops; i++) {
 141                 cur_size_in_dw = size_in_dw;
 142                 if (cur_size_in_dw > 0xFFFFF)
 143                         cur_size_in_dw = 0xFFFFF;
 144                 size_in_dw -= cur_size_in_dw;
 145                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_COPY, 0, cur_size_in_dw));
 146                 radeon_ring_write(ring, dst_offset & 0xfffffffc);
 147                 radeon_ring_write(ring, src_offset & 0xfffffffc);
 148                 radeon_ring_write(ring, upper_32_bits(dst_offset) & 0xff);
 149                 radeon_ring_write(ring, upper_32_bits(src_offset) & 0xff);
 150                 src_offset += cur_size_in_dw * 4;
 151                 dst_offset += cur_size_in_dw * 4;
 152         }
 153
 154         r = radeon_fence_emit(rdev, fence, ring->idx);
 155         if (r) {
 156                 radeon_ring_unlock_undo(rdev, ring);
 157                 return r;
 158         }
 159
 160         radeon_ring_unlock_commit(rdev, ring);
 161         radeon_semaphore_free(rdev, &sem, *fence);
 162
 163         return r;
 164 }
 165
 166 /**
 167  * evergreen_dma_is_lockup - Check if the DMA engine is locked up
 168  *
 169  * @rdev: radeon_device pointer
 170  * @ring: radeon_ring structure holding ring information
 171  *
 172  * Check if the async DMA engine is locked up.
 173  * Returns true if the engine appears to be locked up, false if not.
 174  */
 175 bool evergreen_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
 176 {
 177         u32 reset_mask = evergreen_gpu_check_soft_reset(rdev);
 178
 179         if (!(reset_mask & RADEON_RESET_DMA)) {
 180                 radeon_ring_lockup_update(ring);
 181                 return false;
 182         }
 183         /* force ring activities */
 184         radeon_ring_force_activity(rdev, ring);
 185         return radeon_ring_test_lockup(rdev, ring);
 186 }