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

   1 /*
   2  * Copyright 2018 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  */
  23
  24 /*
  25  * GPU doorbell structures, functions & helpers
  26  */
  27 struct amdgpu_doorbell {
  28         /* doorbell mmio */
  29         resource_size_t         base;
  30         resource_size_t         size;
  31         u32 __iomem             *ptr;
  32         u32                     num_doorbells;  /* Number of doorbells actually reserved for amdgpu. */
  33 };
  34
  35 /* Reserved doorbells for amdgpu (including multimedia).
  36  * KFD can use all the rest in the 2M doorbell bar.
  37  * For asic before vega10, doorbell is 32-bit, so the
  38  * index/offset is in dword. For vega10 and after, doorbell
  39  * can be 64-bit, so the index defined is in qword.
  40  */
  41 struct amdgpu_doorbell_index {
  42         uint32_t kiq;
  43         uint32_t mec_ring0;
  44         uint32_t mec_ring1;
  45         uint32_t mec_ring2;
  46         uint32_t mec_ring3;
  47         uint32_t mec_ring4;
  48         uint32_t mec_ring5;
  49         uint32_t mec_ring6;
  50         uint32_t mec_ring7;
  51         uint32_t userqueue_start;
  52         uint32_t userqueue_end;
  53         uint32_t gfx_ring0;
  54         uint32_t sdma_engine[8];
  55         uint32_t ih;
  56         union {
  57                 struct {
  58                         uint32_t vcn_ring0_1;
  59                         uint32_t vcn_ring2_3;
  60                         uint32_t vcn_ring4_5;
  61                         uint32_t vcn_ring6_7;
  62                 } vcn;
  63                 struct {
  64                         uint32_t uvd_ring0_1;
  65                         uint32_t uvd_ring2_3;
  66                         uint32_t uvd_ring4_5;
  67                         uint32_t uvd_ring6_7;
  68                         uint32_t vce_ring0_1;
  69                         uint32_t vce_ring2_3;
  70                         uint32_t vce_ring4_5;
  71                         uint32_t vce_ring6_7;
  72                 } uvd_vce;
  73         };
  74         uint32_t max_assignment;
  75         /* Per engine SDMA doorbell size in dword */
  76         uint32_t sdma_doorbell_range;
  77 };
  78
  79 typedef enum _AMDGPU_DOORBELL_ASSIGNMENT
  80 {
  81         AMDGPU_DOORBELL_KIQ                     = 0x000,
  82         AMDGPU_DOORBELL_HIQ                     = 0x001,
  83         AMDGPU_DOORBELL_DIQ                     = 0x002,
  84         AMDGPU_DOORBELL_MEC_RING0               = 0x010,
  85         AMDGPU_DOORBELL_MEC_RING1               = 0x011,
  86         AMDGPU_DOORBELL_MEC_RING2               = 0x012,
  87         AMDGPU_DOORBELL_MEC_RING3               = 0x013,
  88         AMDGPU_DOORBELL_MEC_RING4               = 0x014,
  89         AMDGPU_DOORBELL_MEC_RING5               = 0x015,
  90         AMDGPU_DOORBELL_MEC_RING6               = 0x016,
  91         AMDGPU_DOORBELL_MEC_RING7               = 0x017,
  92         AMDGPU_DOORBELL_GFX_RING0               = 0x020,
  93         AMDGPU_DOORBELL_sDMA_ENGINE0            = 0x1E0,
  94         AMDGPU_DOORBELL_sDMA_ENGINE1            = 0x1E1,
  95         AMDGPU_DOORBELL_IH                      = 0x1E8,
  96         AMDGPU_DOORBELL_MAX_ASSIGNMENT          = 0x3FF,
  97         AMDGPU_DOORBELL_INVALID                 = 0xFFFF
  98 } AMDGPU_DOORBELL_ASSIGNMENT;
  99
 100 typedef enum _AMDGPU_VEGA20_DOORBELL_ASSIGNMENT
 101 {
 102         /* Compute + GFX: 0~255 */
 103         AMDGPU_VEGA20_DOORBELL_KIQ                     = 0x000,
 104         AMDGPU_VEGA20_DOORBELL_HIQ                     = 0x001,
 105         AMDGPU_VEGA20_DOORBELL_DIQ                     = 0x002,
 106         AMDGPU_VEGA20_DOORBELL_MEC_RING0               = 0x003,
 107         AMDGPU_VEGA20_DOORBELL_MEC_RING1               = 0x004,
 108         AMDGPU_VEGA20_DOORBELL_MEC_RING2               = 0x005,
 109         AMDGPU_VEGA20_DOORBELL_MEC_RING3               = 0x006,
 110         AMDGPU_VEGA20_DOORBELL_MEC_RING4               = 0x007,
 111         AMDGPU_VEGA20_DOORBELL_MEC_RING5               = 0x008,
 112         AMDGPU_VEGA20_DOORBELL_MEC_RING6               = 0x009,
 113         AMDGPU_VEGA20_DOORBELL_MEC_RING7               = 0x00A,
 114         AMDGPU_VEGA20_DOORBELL_USERQUEUE_START         = 0x00B,
 115         AMDGPU_VEGA20_DOORBELL_USERQUEUE_END           = 0x08A,
 116         AMDGPU_VEGA20_DOORBELL_GFX_RING0               = 0x08B,
 117         /* SDMA:256~335*/
 118         AMDGPU_VEGA20_DOORBELL_sDMA_ENGINE0            = 0x100,
 119         AMDGPU_VEGA20_DOORBELL_sDMA_ENGINE1            = 0x10A,
 120         AMDGPU_VEGA20_DOORBELL_sDMA_ENGINE2            = 0x114,
 121         AMDGPU_VEGA20_DOORBELL_sDMA_ENGINE3            = 0x11E,
 122         AMDGPU_VEGA20_DOORBELL_sDMA_ENGINE4            = 0x128,
 123         AMDGPU_VEGA20_DOORBELL_sDMA_ENGINE5            = 0x132,
 124         AMDGPU_VEGA20_DOORBELL_sDMA_ENGINE6            = 0x13C,
 125         AMDGPU_VEGA20_DOORBELL_sDMA_ENGINE7            = 0x146,
 126         /* IH: 376~391 */
 127         AMDGPU_VEGA20_DOORBELL_IH                      = 0x178,
 128         /* MMSCH: 392~407
 129          * overlap the doorbell assignment with VCN as they are  mutually exclusive
 130          * VCE engine's doorbell is 32 bit and two VCE ring share one QWORD
 131          */
 132         AMDGPU_VEGA20_DOORBELL64_VCN0_1                  = 0x188, /* lower 32 bits for VNC0 and upper 32 bits for VNC1 */
 133         AMDGPU_VEGA20_DOORBELL64_VCN2_3                  = 0x189,
 134         AMDGPU_VEGA20_DOORBELL64_VCN4_5                  = 0x18A,
 135         AMDGPU_VEGA20_DOORBELL64_VCN6_7                  = 0x18B,
 136
 137         AMDGPU_VEGA20_DOORBELL64_UVD_RING0_1             = 0x188,
 138         AMDGPU_VEGA20_DOORBELL64_UVD_RING2_3             = 0x189,
 139         AMDGPU_VEGA20_DOORBELL64_UVD_RING4_5             = 0x18A,
 140         AMDGPU_VEGA20_DOORBELL64_UVD_RING6_7             = 0x18B,
 141
 142         AMDGPU_VEGA20_DOORBELL64_VCE_RING0_1             = 0x18C,
 143         AMDGPU_VEGA20_DOORBELL64_VCE_RING2_3             = 0x18D,
 144         AMDGPU_VEGA20_DOORBELL64_VCE_RING4_5             = 0x18E,
 145         AMDGPU_VEGA20_DOORBELL64_VCE_RING6_7             = 0x18F,
 146         AMDGPU_VEGA20_DOORBELL_MAX_ASSIGNMENT            = 0x18F,
 147         AMDGPU_VEGA20_DOORBELL_INVALID                   = 0xFFFF
 148 } AMDGPU_VEGA20_DOORBELL_ASSIGNMENT;
 149
 150 /*
 151  * 64bit doorbell, offset are in QWORD, occupy 2KB doorbell space
 152  */
 153 typedef enum _AMDGPU_DOORBELL64_ASSIGNMENT
 154 {
 155         /*
 156          * All compute related doorbells: kiq, hiq, diq, traditional compute queue, user queue, should locate in
 157          * a continues range so that programming CP_MEC_DOORBELL_RANGE_LOWER/UPPER can cover this range.
 158          *  Compute related doorbells are allocated from 0x00 to 0x8a
 159          */
 160
 161
 162         /* kernel scheduling */
 163         AMDGPU_DOORBELL64_KIQ                     = 0x00,
 164
 165         /* HSA interface queue and debug queue */
 166         AMDGPU_DOORBELL64_HIQ                     = 0x01,
 167         AMDGPU_DOORBELL64_DIQ                     = 0x02,
 168
 169         /* Compute engines */
 170         AMDGPU_DOORBELL64_MEC_RING0               = 0x03,
 171         AMDGPU_DOORBELL64_MEC_RING1               = 0x04,
 172         AMDGPU_DOORBELL64_MEC_RING2               = 0x05,
 173         AMDGPU_DOORBELL64_MEC_RING3               = 0x06,
 174         AMDGPU_DOORBELL64_MEC_RING4               = 0x07,
 175         AMDGPU_DOORBELL64_MEC_RING5               = 0x08,
 176         AMDGPU_DOORBELL64_MEC_RING6               = 0x09,
 177         AMDGPU_DOORBELL64_MEC_RING7               = 0x0a,
 178
 179         /* User queue doorbell range (128 doorbells) */
 180         AMDGPU_DOORBELL64_USERQUEUE_START         = 0x0b,
 181         AMDGPU_DOORBELL64_USERQUEUE_END           = 0x8a,
 182
 183         /* Graphics engine */
 184         AMDGPU_DOORBELL64_GFX_RING0               = 0x8b,
 185
 186         /*
 187          * Other graphics doorbells can be allocated here: from 0x8c to 0xdf
 188          * Graphics voltage island aperture 1
 189          * default non-graphics QWORD index is 0xe0 - 0xFF inclusive
 190          */
 191
 192         /* For vega10 sriov, the sdma doorbell must be fixed as follow
 193          * to keep the same setting with host driver, or it will
 194          * happen conflicts
 195          */
 196         AMDGPU_DOORBELL64_sDMA_ENGINE0            = 0xF0,
 197         AMDGPU_DOORBELL64_sDMA_HI_PRI_ENGINE0     = 0xF1,
 198         AMDGPU_DOORBELL64_sDMA_ENGINE1            = 0xF2,
 199         AMDGPU_DOORBELL64_sDMA_HI_PRI_ENGINE1     = 0xF3,
 200
 201         /* Interrupt handler */
 202         AMDGPU_DOORBELL64_IH                      = 0xF4,  /* For legacy interrupt ring buffer */
 203         AMDGPU_DOORBELL64_IH_RING1                = 0xF5,  /* For page migration request log */
 204         AMDGPU_DOORBELL64_IH_RING2                = 0xF6,  /* For page migration translation/invalidation log */
 205
 206         /* VCN engine use 32 bits doorbell  */
 207         AMDGPU_DOORBELL64_VCN0_1                  = 0xF8, /* lower 32 bits for VNC0 and upper 32 bits for VNC1 */
 208         AMDGPU_DOORBELL64_VCN2_3                  = 0xF9,
 209         AMDGPU_DOORBELL64_VCN4_5                  = 0xFA,
 210         AMDGPU_DOORBELL64_VCN6_7                  = 0xFB,
 211
 212         /* overlap the doorbell assignment with VCN as they are  mutually exclusive
 213          * VCE engine's doorbell is 32 bit and two VCE ring share one QWORD
 214          */
 215         AMDGPU_DOORBELL64_UVD_RING0_1             = 0xF8,
 216         AMDGPU_DOORBELL64_UVD_RING2_3             = 0xF9,
 217         AMDGPU_DOORBELL64_UVD_RING4_5             = 0xFA,
 218         AMDGPU_DOORBELL64_UVD_RING6_7             = 0xFB,
 219
 220         AMDGPU_DOORBELL64_VCE_RING0_1             = 0xFC,
 221         AMDGPU_DOORBELL64_VCE_RING2_3             = 0xFD,
 222         AMDGPU_DOORBELL64_VCE_RING4_5             = 0xFE,
 223         AMDGPU_DOORBELL64_VCE_RING6_7             = 0xFF,
 224
 225         AMDGPU_DOORBELL64_MAX_ASSIGNMENT          = 0xFF,
 226         AMDGPU_DOORBELL64_INVALID                 = 0xFFFF
 227 } AMDGPU_DOORBELL64_ASSIGNMENT;
 228
 229 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index);
 230 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v);
 231 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index);
 232 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v);
 233
 234 #define RDOORBELL32(index) amdgpu_mm_rdoorbell(adev, (index))
 235 #define WDOORBELL32(index, v) amdgpu_mm_wdoorbell(adev, (index), (v))
 236 #define RDOORBELL64(index) amdgpu_mm_rdoorbell64(adev, (index))
 237 #define WDOORBELL64(index, v) amdgpu_mm_wdoorbell64(adev, (index), (v))
 238