drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h

   1 /*
   2  * Copyright 2016 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: Christian König
  23  */
  24 #ifndef __AMDGPU_RING_H__
  25 #define __AMDGPU_RING_H__
  26
  27 #include <drm/amdgpu_drm.h>
  28 #include <drm/gpu_scheduler.h>
  29
  30 /* max number of rings */
  31 #define AMDGPU_MAX_RINGS                18
  32 #define AMDGPU_MAX_GFX_RINGS            1
  33 #define AMDGPU_MAX_COMPUTE_RINGS        8
  34 #define AMDGPU_MAX_VCE_RINGS            3
  35 #define AMDGPU_MAX_UVD_ENC_RINGS        2
  36
  37 /* some special values for the owner field */
  38 #define AMDGPU_FENCE_OWNER_UNDEFINED    ((void*)0ul)
  39 #define AMDGPU_FENCE_OWNER_VM           ((void*)1ul)
  40
  41 #define AMDGPU_FENCE_FLAG_64BIT         (1 << 0)
  42 #define AMDGPU_FENCE_FLAG_INT           (1 << 1)
  43
  44 enum amdgpu_ring_type {
  45         AMDGPU_RING_TYPE_GFX,
  46         AMDGPU_RING_TYPE_COMPUTE,
  47         AMDGPU_RING_TYPE_SDMA,
  48         AMDGPU_RING_TYPE_UVD,
  49         AMDGPU_RING_TYPE_VCE,
  50         AMDGPU_RING_TYPE_KIQ,
  51         AMDGPU_RING_TYPE_UVD_ENC,
  52         AMDGPU_RING_TYPE_VCN_DEC,
  53         AMDGPU_RING_TYPE_VCN_ENC
  54 };
  55
  56 struct amdgpu_device;
  57 struct amdgpu_ring;
  58 struct amdgpu_ib;
  59 struct amdgpu_cs_parser;
  60 struct amdgpu_job;
  61
  62 /*
  63  * Fences.
  64  */
  65 struct amdgpu_fence_driver {
  66         uint64_t                        gpu_addr;
  67         volatile uint32_t               *cpu_addr;
  68         /* sync_seq is protected by ring emission lock */
  69         uint32_t                        sync_seq;
  70         atomic_t                        last_seq;
  71         bool                            initialized;
  72         struct amdgpu_irq_src           *irq_src;
  73         unsigned                        irq_type;
  74         struct timer_list               fallback_timer;
  75         unsigned                        num_fences_mask;
  76         spinlock_t                      lock;
  77         struct dma_fence                **fences;
  78 };
  79
  80 int amdgpu_fence_driver_init(struct amdgpu_device *adev);
  81 void amdgpu_fence_driver_fini(struct amdgpu_device *adev);
  82 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring);
  83
  84 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
  85                                   unsigned num_hw_submission);
  86 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
  87                                    struct amdgpu_irq_src *irq_src,
  88                                    unsigned irq_type);
  89 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev);
  90 void amdgpu_fence_driver_resume(struct amdgpu_device *adev);
  91 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **fence);
  92 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s);
  93 void amdgpu_fence_process(struct amdgpu_ring *ring);
  94 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring);
  95 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
  96                                       uint32_t wait_seq,
  97                                       signed long timeout);
  98 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring);
  99
 100 /*
 101  * Rings.
 102  */
 103
 104 /* provided by hw blocks that expose a ring buffer for commands */
 105 struct amdgpu_ring_funcs {
 106         enum amdgpu_ring_type   type;
 107         uint32_t                align_mask;
 108         u32                     nop;
 109         bool                    support_64bit_ptrs;
 110         unsigned                vmhub;
 111
 112         /* ring read/write ptr handling */
 113         u64 (*get_rptr)(struct amdgpu_ring *ring);
 114         u64 (*get_wptr)(struct amdgpu_ring *ring);
 115         void (*set_wptr)(struct amdgpu_ring *ring);
 116         /* validating and patching of IBs */
 117         int (*parse_cs)(struct amdgpu_cs_parser *p, uint32_t ib_idx);
 118         /* constants to calculate how many DW are needed for an emit */
 119         unsigned emit_frame_size;
 120         unsigned emit_ib_size;
 121         /* command emit functions */
 122         void (*emit_ib)(struct amdgpu_ring *ring,
 123                         struct amdgpu_ib *ib,
 124                         unsigned vmid, bool ctx_switch);
 125         void (*emit_fence)(struct amdgpu_ring *ring, uint64_t addr,
 126                            uint64_t seq, unsigned flags);
 127         void (*emit_pipeline_sync)(struct amdgpu_ring *ring);
 128         void (*emit_vm_flush)(struct amdgpu_ring *ring, unsigned vmid,
 129                               unsigned pasid, uint64_t pd_addr);
 130         void (*emit_hdp_flush)(struct amdgpu_ring *ring);
 131         void (*emit_gds_switch)(struct amdgpu_ring *ring, uint32_t vmid,
 132                                 uint32_t gds_base, uint32_t gds_size,
 133                                 uint32_t gws_base, uint32_t gws_size,
 134                                 uint32_t oa_base, uint32_t oa_size);
 135         /* testing functions */
 136         int (*test_ring)(struct amdgpu_ring *ring);
 137         int (*test_ib)(struct amdgpu_ring *ring, long timeout);
 138         /* insert NOP packets */
 139         void (*insert_nop)(struct amdgpu_ring *ring, uint32_t count);
 140         void (*insert_start)(struct amdgpu_ring *ring);
 141         void (*insert_end)(struct amdgpu_ring *ring);
 142         /* pad the indirect buffer to the necessary number of dw */
 143         void (*pad_ib)(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
 144         unsigned (*init_cond_exec)(struct amdgpu_ring *ring);
 145         void (*patch_cond_exec)(struct amdgpu_ring *ring, unsigned offset);
 146         /* note usage for clock and power gating */
 147         void (*begin_use)(struct amdgpu_ring *ring);
 148         void (*end_use)(struct amdgpu_ring *ring);
 149         void (*emit_switch_buffer) (struct amdgpu_ring *ring);
 150         void (*emit_cntxcntl) (struct amdgpu_ring *ring, uint32_t flags);
 151         void (*emit_rreg)(struct amdgpu_ring *ring, uint32_t reg);
 152         void (*emit_wreg)(struct amdgpu_ring *ring, uint32_t reg, uint32_t val);
 153         void (*emit_reg_wait)(struct amdgpu_ring *ring, uint32_t reg,
 154                               uint32_t val, uint32_t mask);
 155         void (*emit_tmz)(struct amdgpu_ring *ring, bool start);
 156         /* priority functions */
 157         void (*set_priority) (struct amdgpu_ring *ring,
 158                               enum drm_sched_priority priority);
 159 };
 160
 161 struct amdgpu_ring {
 162         struct amdgpu_device            *adev;
 163         const struct amdgpu_ring_funcs  *funcs;
 164         struct amdgpu_fence_driver      fence_drv;
 165         struct drm_gpu_scheduler        sched;
 166         struct list_head                lru_list;
 167
 168         struct amdgpu_bo        *ring_obj;
 169         volatile uint32_t       *ring;
 170         unsigned                rptr_offs;
 171         u64                     wptr;
 172         u64                     wptr_old;
 173         unsigned                ring_size;
 174         unsigned                max_dw;
 175         int                     count_dw;
 176         uint64_t                gpu_addr;
 177         uint64_t                ptr_mask;
 178         uint32_t                buf_mask;
 179         bool                    ready;
 180         u32                     idx;
 181         u32                     me;
 182         u32                     pipe;
 183         u32                     queue;
 184         struct amdgpu_bo        *mqd_obj;
 185         uint64_t                mqd_gpu_addr;
 186         void                    *mqd_ptr;
 187         uint64_t                eop_gpu_addr;
 188         u32                     doorbell_index;
 189         bool                    use_doorbell;
 190         bool                    use_pollmem;
 191         unsigned                wptr_offs;
 192         unsigned                fence_offs;
 193         uint64_t                current_ctx;
 194         char                    name[16];
 195         unsigned                cond_exe_offs;
 196         u64                     cond_exe_gpu_addr;
 197         volatile u32            *cond_exe_cpu_addr;
 198         unsigned                vm_inv_eng;
 199         bool                    has_compute_vm_bug;
 200
 201         atomic_t                num_jobs[DRM_SCHED_PRIORITY_MAX];
 202         struct mutex            priority_mutex;
 203         /* protected by priority_mutex */
 204         int                     priority;
 205
 206 #if defined(CONFIG_DEBUG_FS)
 207         struct dentry *ent;
 208 #endif
 209 };
 210
 211 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw);
 212 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count);
 213 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
 214 void amdgpu_ring_commit(struct amdgpu_ring *ring);
 215 void amdgpu_ring_undo(struct amdgpu_ring *ring);
 216 void amdgpu_ring_priority_get(struct amdgpu_ring *ring,
 217                               enum drm_sched_priority priority);
 218 void amdgpu_ring_priority_put(struct amdgpu_ring *ring,
 219                               enum drm_sched_priority priority);
 220 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
 221                      unsigned ring_size, struct amdgpu_irq_src *irq_src,
 222                      unsigned irq_type);
 223 void amdgpu_ring_fini(struct amdgpu_ring *ring);
 224 int amdgpu_ring_lru_get(struct amdgpu_device *adev, int type,
 225                         int *blacklist, int num_blacklist,
 226                         bool lru_pipe_order, struct amdgpu_ring **ring);
 227 void amdgpu_ring_lru_touch(struct amdgpu_device *adev, struct amdgpu_ring *ring);
 228 static inline void amdgpu_ring_clear_ring(struct amdgpu_ring *ring)
 229 {
 230         int i = 0;
 231         while (i <= ring->buf_mask)
 232                 ring->ring[i++] = ring->funcs->nop;
 233
 234 }
 235
 236 static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v)
 237 {
 238         if (ring->count_dw <= 0)
 239                 DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
 240         ring->ring[ring->wptr++ & ring->buf_mask] = v;
 241         ring->wptr &= ring->ptr_mask;
 242         ring->count_dw--;
 243 }
 244
 245 static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring,
 246                                               void *src, int count_dw)
 247 {
 248         unsigned occupied, chunk1, chunk2;
 249         void *dst;
 250
 251         if (unlikely(ring->count_dw < count_dw))
 252                 DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
 253
 254         occupied = ring->wptr & ring->buf_mask;
 255         dst = (void *)&ring->ring[occupied];
 256         chunk1 = ring->buf_mask + 1 - occupied;
 257         chunk1 = (chunk1 >= count_dw) ? count_dw: chunk1;
 258         chunk2 = count_dw - chunk1;
 259         chunk1 <<= 2;
 260         chunk2 <<= 2;
 261
 262         if (chunk1)
 263                 memcpy(dst, src, chunk1);
 264
 265         if (chunk2) {
 266                 src += chunk1;
 267                 dst = (void *)ring->ring;
 268                 memcpy(dst, src, chunk2);
 269         }
 270
 271         ring->wptr += count_dw;
 272         ring->wptr &= ring->ptr_mask;
 273         ring->count_dw -= count_dw;
 274 }
 275
 276 #endif