drivers/gpu/drm/radeon/si_dma.c

   1 /*
   2  * Copyright 2013 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 "radeon_trace.h"
  28 #include "sid.h"
  29
  30 u32 si_gpu_check_soft_reset(struct radeon_device *rdev);
  31
  32 /**
  33  * si_dma_is_lockup - Check if the DMA engine is locked up
  34  *
  35  * @rdev: radeon_device pointer
  36  * @ring: radeon_ring structure holding ring information
  37  *
  38  * Check if the async DMA engine is locked up.
  39  * Returns true if the engine appears to be locked up, false if not.
  40  */
  41 bool si_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
  42 {
  43         u32 reset_mask = si_gpu_check_soft_reset(rdev);
  44         u32 mask;
  45
  46         if (ring->idx == R600_RING_TYPE_DMA_INDEX)
  47                 mask = RADEON_RESET_DMA;
  48         else
  49                 mask = RADEON_RESET_DMA1;
  50
  51         if (!(reset_mask & mask)) {
  52                 radeon_ring_lockup_update(ring);
  53                 return false;
  54         }
  55         /* force ring activities */
  56         radeon_ring_force_activity(rdev, ring);
  57         return radeon_ring_test_lockup(rdev, ring);
  58 }
  59
  60 /**
  61  * si_dma_vm_set_page - update the page tables using the DMA
  62  *
  63  * @rdev: radeon_device pointer
  64  * @ib: indirect buffer to fill with commands
  65  * @pe: addr of the page entry
  66  * @addr: dst addr to write into pe
  67  * @count: number of page entries to update
  68  * @incr: increase next addr by incr bytes
  69  * @flags: access flags
  70  *
  71  * Update the page tables using the DMA (SI).
  72  */
  73 void si_dma_vm_set_page(struct radeon_device *rdev,
  74                         struct radeon_ib *ib,
  75                         uint64_t pe,
  76                         uint64_t addr, unsigned count,
  77                         uint32_t incr, uint32_t flags)
  78 {
  79         uint64_t value;
  80         unsigned ndw;
  81
  82         trace_radeon_vm_set_page(pe, addr, count, incr, flags);
  83
  84         if (flags & R600_PTE_SYSTEM) {
  85                 while (count) {
  86                         ndw = count * 2;
  87                         if (ndw > 0xFFFFE)
  88                                 ndw = 0xFFFFE;
  89
  90                         /* for non-physically contiguous pages (system) */
  91                         ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw);
  92                         ib->ptr[ib->length_dw++] = pe;
  93                         ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
  94                         for (; ndw > 0; ndw -= 2, --count, pe += 8) {
  95                                 value = radeon_vm_map_gart(rdev, addr);
  96                                 value &= 0xFFFFFFFFFFFFF000ULL;
  97                                 addr += incr;
  98                                 value |= flags;
  99                                 ib->ptr[ib->length_dw++] = value;
 100                                 ib->ptr[ib->length_dw++] = upper_32_bits(value);
 101                         }
 102                 }
 103         } else {
 104                 while (count) {
 105                         ndw = count * 2;
 106                         if (ndw > 0xFFFFE)
 107                                 ndw = 0xFFFFE;
 108
 109                         if (flags & R600_PTE_VALID)
 110                                 value = addr;
 111                         else
 112                                 value = 0;
 113                         /* for physically contiguous pages (vram) */
 114                         ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
 115                         ib->ptr[ib->length_dw++] = pe; /* dst addr */
 116                         ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
 117                         ib->ptr[ib->length_dw++] = flags; /* mask */
 118                         ib->ptr[ib->length_dw++] = 0;
 119                         ib->ptr[ib->length_dw++] = value; /* value */
 120                         ib->ptr[ib->length_dw++] = upper_32_bits(value);
 121                         ib->ptr[ib->length_dw++] = incr; /* increment size */
 122                         ib->ptr[ib->length_dw++] = 0;
 123                         pe += ndw * 4;
 124                         addr += (ndw / 2) * incr;
 125                         count -= ndw / 2;
 126                 }
 127         }
 128         while (ib->length_dw & 0x7)
 129                 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0);
 130 }
 131
 132 void si_dma_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
 133 {
 134         struct radeon_ring *ring = &rdev->ring[ridx];
 135
 136         if (vm == NULL)
 137                 return;
 138
 139         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
 140         if (vm->id < 8) {
 141                 radeon_ring_write(ring, (0xf << 16) | ((VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2)) >> 2));
 142         } else {
 143                 radeon_ring_write(ring, (0xf << 16) | ((VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm->id - 8) << 2)) >> 2));
 144         }
 145         radeon_ring_write(ring, vm->pd_gpu_addr >> 12);
 146
 147         /* flush hdp cache */
 148         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
 149         radeon_ring_write(ring, (0xf << 16) | (HDP_MEM_COHERENCY_FLUSH_CNTL >> 2));
 150         radeon_ring_write(ring, 1);
 151
 152         /* bits 0-7 are the VM contexts0-7 */
 153         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
 154         radeon_ring_write(ring, (0xf << 16) | (VM_INVALIDATE_REQUEST >> 2));
 155         radeon_ring_write(ring, 1 << vm->id);
 156 }
 157
 158 /**
 159  * si_copy_dma - copy pages using the DMA engine
 160  *
 161  * @rdev: radeon_device pointer
 162  * @src_offset: src GPU address
 163  * @dst_offset: dst GPU address
 164  * @num_gpu_pages: number of GPU pages to xfer
 165  * @fence: radeon fence object
 166  *
 167  * Copy GPU paging using the DMA engine (SI).
 168  * Used by the radeon ttm implementation to move pages if
 169  * registered as the asic copy callback.
 170  */
 171 int si_copy_dma(struct radeon_device *rdev,
 172                 uint64_t src_offset, uint64_t dst_offset,
 173                 unsigned num_gpu_pages,
 174                 struct radeon_fence **fence)
 175 {
 176         struct radeon_semaphore *sem = NULL;
 177         int ring_index = rdev->asic->copy.dma_ring_index;
 178         struct radeon_ring *ring = &rdev->ring[ring_index];
 179         u32 size_in_bytes, cur_size_in_bytes;
 180         int i, num_loops;
 181         int r = 0;
 182
 183         r = radeon_semaphore_create(rdev, &sem);
 184         if (r) {
 185                 DRM_ERROR("radeon: moving bo (%d).\n", r);
 186                 return r;
 187         }
 188
 189         size_in_bytes = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT);
 190         num_loops = DIV_ROUND_UP(size_in_bytes, 0xfffff);
 191         r = radeon_ring_lock(rdev, ring, num_loops * 5 + 11);
 192         if (r) {
 193                 DRM_ERROR("radeon: moving bo (%d).\n", r);
 194                 radeon_semaphore_free(rdev, &sem, NULL);
 195                 return r;
 196         }
 197
 198         radeon_semaphore_sync_to(sem, *fence);
 199         radeon_semaphore_sync_rings(rdev, sem, ring->idx);
 200
 201         for (i = 0; i < num_loops; i++) {
 202                 cur_size_in_bytes = size_in_bytes;
 203                 if (cur_size_in_bytes > 0xFFFFF)
 204                         cur_size_in_bytes = 0xFFFFF;
 205                 size_in_bytes -= cur_size_in_bytes;
 206                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_COPY, 1, 0, 0, cur_size_in_bytes));
 207                 radeon_ring_write(ring, dst_offset & 0xffffffff);
 208                 radeon_ring_write(ring, src_offset & 0xffffffff);
 209                 radeon_ring_write(ring, upper_32_bits(dst_offset) & 0xff);
 210                 radeon_ring_write(ring, upper_32_bits(src_offset) & 0xff);
 211                 src_offset += cur_size_in_bytes;
 212                 dst_offset += cur_size_in_bytes;
 213         }
 214
 215         r = radeon_fence_emit(rdev, fence, ring->idx);
 216         if (r) {
 217                 radeon_ring_unlock_undo(rdev, ring);
 218                 return r;
 219         }
 220
 221         radeon_ring_unlock_commit(rdev, ring);
 222         radeon_semaphore_free(rdev, &sem, *fence);
 223
 224         return r;
 225 }
 226