drivers/gpu/drm/radeon/ni_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 "radeon_trace.h"
  28 #include "nid.h"
  29
  30 u32 cayman_gpu_check_soft_reset(struct radeon_device *rdev);
  31
  32 /*
  33  * DMA
  34  * Starting with R600, the GPU has an asynchronous
  35  * DMA engine.  The programming model is very similar
  36  * to the 3D engine (ring buffer, IBs, etc.), but the
  37  * DMA controller has it's own packet format that is
  38  * different form the PM4 format used by the 3D engine.
  39  * It supports copying data, writing embedded data,
  40  * solid fills, and a number of other things.  It also
  41  * has support for tiling/detiling of buffers.
  42  * Cayman and newer support two asynchronous DMA engines.
  43  */
  44
  45 /**
  46  * cayman_dma_ring_ib_execute - Schedule an IB on the DMA engine
  47  *
  48  * @rdev: radeon_device pointer
  49  * @ib: IB object to schedule
  50  *
  51  * Schedule an IB in the DMA ring (cayman-SI).
  52  */
  53 void cayman_dma_ring_ib_execute(struct radeon_device *rdev,
  54                                 struct radeon_ib *ib)
  55 {
  56         struct radeon_ring *ring = &rdev->ring[ib->ring];
  57
  58         if (rdev->wb.enabled) {
  59                 u32 next_rptr = ring->wptr + 4;
  60                 while ((next_rptr & 7) != 5)
  61                         next_rptr++;
  62                 next_rptr += 3;
  63                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1));
  64                 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
  65                 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xff);
  66                 radeon_ring_write(ring, next_rptr);
  67         }
  68
  69         /* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
  70          * Pad as necessary with NOPs.
  71          */
  72         while ((ring->wptr & 7) != 5)
  73                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0));
  74         radeon_ring_write(ring, DMA_IB_PACKET(DMA_PACKET_INDIRECT_BUFFER, ib->vm ? ib->vm->id : 0, 0));
  75         radeon_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
  76         radeon_ring_write(ring, (ib->length_dw << 12) | (upper_32_bits(ib->gpu_addr) & 0xFF));
  77
  78 }
  79
  80 /**
  81  * cayman_dma_stop - stop the async dma engines
  82  *
  83  * @rdev: radeon_device pointer
  84  *
  85  * Stop the async dma engines (cayman-SI).
  86  */
  87 void cayman_dma_stop(struct radeon_device *rdev)
  88 {
  89         u32 rb_cntl;
  90
  91         radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
  92
  93         /* dma0 */
  94         rb_cntl = RREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET);
  95         rb_cntl &= ~DMA_RB_ENABLE;
  96         WREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET, rb_cntl);
  97
  98         /* dma1 */
  99         rb_cntl = RREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET);
 100         rb_cntl &= ~DMA_RB_ENABLE;
 101         WREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET, rb_cntl);
 102
 103         rdev->ring[R600_RING_TYPE_DMA_INDEX].ready = false;
 104         rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX].ready = false;
 105 }
 106
 107 /**
 108  * cayman_dma_resume - setup and start the async dma engines
 109  *
 110  * @rdev: radeon_device pointer
 111  *
 112  * Set up the DMA ring buffers and enable them. (cayman-SI).
 113  * Returns 0 for success, error for failure.
 114  */
 115 int cayman_dma_resume(struct radeon_device *rdev)
 116 {
 117         struct radeon_ring *ring;
 118         u32 rb_cntl, dma_cntl, ib_cntl;
 119         u32 rb_bufsz;
 120         u32 reg_offset, wb_offset;
 121         int i, r;
 122
 123         /* Reset dma */
 124         WREG32(SRBM_SOFT_RESET, SOFT_RESET_DMA | SOFT_RESET_DMA1);
 125         RREG32(SRBM_SOFT_RESET);
 126         udelay(50);
 127         WREG32(SRBM_SOFT_RESET, 0);
 128
 129         for (i = 0; i < 2; i++) {
 130                 if (i == 0) {
 131                         ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
 132                         reg_offset = DMA0_REGISTER_OFFSET;
 133                         wb_offset = R600_WB_DMA_RPTR_OFFSET;
 134                 } else {
 135                         ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
 136                         reg_offset = DMA1_REGISTER_OFFSET;
 137                         wb_offset = CAYMAN_WB_DMA1_RPTR_OFFSET;
 138                 }
 139
 140                 WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL + reg_offset, 0);
 141                 WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL + reg_offset, 0);
 142
 143                 /* Set ring buffer size in dwords */
 144                 rb_bufsz = order_base_2(ring->ring_size / 4);
 145                 rb_cntl = rb_bufsz << 1;
 146 #ifdef __BIG_ENDIAN
 147                 rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
 148 #endif
 149                 WREG32(DMA_RB_CNTL + reg_offset, rb_cntl);
 150
 151                 /* Initialize the ring buffer's read and write pointers */
 152                 WREG32(DMA_RB_RPTR + reg_offset, 0);
 153                 WREG32(DMA_RB_WPTR + reg_offset, 0);
 154
 155                 /* set the wb address whether it's enabled or not */
 156                 WREG32(DMA_RB_RPTR_ADDR_HI + reg_offset,
 157                        upper_32_bits(rdev->wb.gpu_addr + wb_offset) & 0xFF);
 158                 WREG32(DMA_RB_RPTR_ADDR_LO + reg_offset,
 159                        ((rdev->wb.gpu_addr + wb_offset) & 0xFFFFFFFC));
 160
 161                 if (rdev->wb.enabled)
 162                         rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
 163
 164                 WREG32(DMA_RB_BASE + reg_offset, ring->gpu_addr >> 8);
 165
 166                 /* enable DMA IBs */
 167                 ib_cntl = DMA_IB_ENABLE | CMD_VMID_FORCE;
 168 #ifdef __BIG_ENDIAN
 169                 ib_cntl |= DMA_IB_SWAP_ENABLE;
 170 #endif
 171                 WREG32(DMA_IB_CNTL + reg_offset, ib_cntl);
 172
 173                 dma_cntl = RREG32(DMA_CNTL + reg_offset);
 174                 dma_cntl &= ~CTXEMPTY_INT_ENABLE;
 175                 WREG32(DMA_CNTL + reg_offset, dma_cntl);
 176
 177                 ring->wptr = 0;
 178                 WREG32(DMA_RB_WPTR + reg_offset, ring->wptr << 2);
 179
 180                 ring->rptr = RREG32(DMA_RB_RPTR + reg_offset) >> 2;
 181
 182                 WREG32(DMA_RB_CNTL + reg_offset, rb_cntl | DMA_RB_ENABLE);
 183
 184                 ring->ready = true;
 185
 186                 r = radeon_ring_test(rdev, ring->idx, ring);
 187                 if (r) {
 188                         ring->ready = false;
 189                         return r;
 190                 }
 191         }
 192
 193         radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
 194
 195         return 0;
 196 }
 197
 198 /**
 199  * cayman_dma_fini - tear down the async dma engines
 200  *
 201  * @rdev: radeon_device pointer
 202  *
 203  * Stop the async dma engines and free the rings (cayman-SI).
 204  */
 205 void cayman_dma_fini(struct radeon_device *rdev)
 206 {
 207         cayman_dma_stop(rdev);
 208         radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX]);
 209         radeon_ring_fini(rdev, &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX]);
 210 }
 211
 212 /**
 213  * cayman_dma_is_lockup - Check if the DMA engine is locked up
 214  *
 215  * @rdev: radeon_device pointer
 216  * @ring: radeon_ring structure holding ring information
 217  *
 218  * Check if the async DMA engine is locked up.
 219  * Returns true if the engine appears to be locked up, false if not.
 220  */
 221 bool cayman_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
 222 {
 223         u32 reset_mask = cayman_gpu_check_soft_reset(rdev);
 224         u32 mask;
 225
 226         if (ring->idx == R600_RING_TYPE_DMA_INDEX)
 227                 mask = RADEON_RESET_DMA;
 228         else
 229                 mask = RADEON_RESET_DMA1;
 230
 231         if (!(reset_mask & mask)) {
 232                 radeon_ring_lockup_update(ring);
 233                 return false;
 234         }
 235         /* force ring activities */
 236         radeon_ring_force_activity(rdev, ring);
 237         return radeon_ring_test_lockup(rdev, ring);
 238 }
 239
 240 /**
 241  * cayman_dma_vm_set_page - update the page tables using the DMA
 242  *
 243  * @rdev: radeon_device pointer
 244  * @ib: indirect buffer to fill with commands
 245  * @pe: addr of the page entry
 246  * @addr: dst addr to write into pe
 247  * @count: number of page entries to update
 248  * @incr: increase next addr by incr bytes
 249  * @flags: hw access flags
 250  *
 251  * Update the page tables using the DMA (cayman/TN).
 252  */
 253 void cayman_dma_vm_set_page(struct radeon_device *rdev,
 254                             struct radeon_ib *ib,
 255                             uint64_t pe,
 256                             uint64_t addr, unsigned count,
 257                             uint32_t incr, uint32_t flags)
 258 {
 259         uint64_t value;
 260         unsigned ndw;
 261
 262         trace_radeon_vm_set_page(pe, addr, count, incr, flags);
 263
 264         if ((flags & R600_PTE_SYSTEM) || (count == 1)) {
 265                 while (count) {
 266                         ndw = count * 2;
 267                         if (ndw > 0xFFFFE)
 268                                 ndw = 0xFFFFE;
 269
 270                         /* for non-physically contiguous pages (system) */
 271                         ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, ndw);
 272                         ib->ptr[ib->length_dw++] = pe;
 273                         ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
 274                         for (; ndw > 0; ndw -= 2, --count, pe += 8) {
 275                                 if (flags & R600_PTE_SYSTEM) {
 276                                         value = radeon_vm_map_gart(rdev, addr);
 277                                         value &= 0xFFFFFFFFFFFFF000ULL;
 278                                 } else if (flags & R600_PTE_VALID) {
 279                                         value = addr;
 280                                 } else {
 281                                         value = 0;
 282                                 }
 283                                 addr += incr;
 284                                 value |= flags;
 285                                 ib->ptr[ib->length_dw++] = value;
 286                                 ib->ptr[ib->length_dw++] = upper_32_bits(value);
 287                         }
 288                 }
 289         } else {
 290                 while (count) {
 291                         ndw = count * 2;
 292                         if (ndw > 0xFFFFE)
 293                                 ndw = 0xFFFFE;
 294
 295                         if (flags & R600_PTE_VALID)
 296                                 value = addr;
 297                         else
 298                                 value = 0;
 299                         /* for physically contiguous pages (vram) */
 300                         ib->ptr[ib->length_dw++] = DMA_PTE_PDE_PACKET(ndw);
 301                         ib->ptr[ib->length_dw++] = pe; /* dst addr */
 302                         ib->ptr[ib->length_dw++] = upper_32_bits(pe) & 0xff;
 303                         ib->ptr[ib->length_dw++] = flags; /* mask */
 304                         ib->ptr[ib->length_dw++] = 0;
 305                         ib->ptr[ib->length_dw++] = value; /* value */
 306                         ib->ptr[ib->length_dw++] = upper_32_bits(value);
 307                         ib->ptr[ib->length_dw++] = incr; /* increment size */
 308                         ib->ptr[ib->length_dw++] = 0;
 309                         pe += ndw * 4;
 310                         addr += (ndw / 2) * incr;
 311                         count -= ndw / 2;
 312                 }
 313         }
 314         while (ib->length_dw & 0x7)
 315                 ib->ptr[ib->length_dw++] = DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0);
 316 }
 317
 318 void cayman_dma_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
 319 {
 320         struct radeon_ring *ring = &rdev->ring[ridx];
 321
 322         if (vm == NULL)
 323                 return;
 324
 325         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0));
 326         radeon_ring_write(ring, (0xf << 16) | ((VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2)) >> 2));
 327         radeon_ring_write(ring, vm->pd_gpu_addr >> 12);
 328
 329         /* flush hdp cache */
 330         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0));
 331         radeon_ring_write(ring, (0xf << 16) | (HDP_MEM_COHERENCY_FLUSH_CNTL >> 2));
 332         radeon_ring_write(ring, 1);
 333
 334         /* bits 0-7 are the VM contexts0-7 */
 335         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0));
 336         radeon_ring_write(ring, (0xf << 16) | (VM_INVALIDATE_REQUEST >> 2));
 337         radeon_ring_write(ring, 1 << vm->id);
 338 }
 339