drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager_cik.c

   1 // SPDX-License-Identifier: GPL-2.0 OR MIT
   2 /*
   3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be included in
  13  * all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  21  * OTHER DEALINGS IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include "kfd_device_queue_manager.h"
  26 #include "cik_regs.h"
  27 #include "oss/oss_2_4_sh_mask.h"
  28 #include "gca/gfx_7_2_sh_mask.h"
  29
  30 static bool set_cache_memory_policy_cik(struct device_queue_manager *dqm,
  31                                    struct qcm_process_device *qpd,
  32                                    enum cache_policy default_policy,
  33                                    enum cache_policy alternate_policy,
  34                                    void __user *alternate_aperture_base,
  35                                    uint64_t alternate_aperture_size);
  36 static int update_qpd_cik(struct device_queue_manager *dqm,
  37                           struct qcm_process_device *qpd);
  38 static void init_sdma_vm(struct device_queue_manager *dqm,
  39                          struct queue *q,
  40                          struct qcm_process_device *qpd);
  41
  42 void device_queue_manager_init_cik(
  43         struct device_queue_manager_asic_ops *asic_ops)
  44 {
  45         asic_ops->set_cache_memory_policy = set_cache_memory_policy_cik;
  46         asic_ops->update_qpd = update_qpd_cik;
  47         asic_ops->init_sdma_vm = init_sdma_vm;
  48         asic_ops->mqd_manager_init = mqd_manager_init_cik;
  49 }
  50
  51 static uint32_t compute_sh_mem_bases_64bit(unsigned int top_address_nybble)
  52 {
  53         /* In 64-bit mode, we can only control the top 3 bits of the LDS,
  54          * scratch and GPUVM apertures.
  55          * The hardware fills in the remaining 59 bits according to the
  56          * following pattern:
  57          * LDS:         X0000000'00000000 - X0000001'00000000 (4GB)
  58          * Scratch:     X0000001'00000000 - X0000002'00000000 (4GB)
  59          * GPUVM:       Y0010000'00000000 - Y0020000'00000000 (1TB)
  60          *
  61          * (where X/Y is the configurable nybble with the low-bit 0)
  62          *
  63          * LDS and scratch will have the same top nybble programmed in the
  64          * top 3 bits of SH_MEM_BASES.PRIVATE_BASE.
  65          * GPUVM can have a different top nybble programmed in the
  66          * top 3 bits of SH_MEM_BASES.SHARED_BASE.
  67          * We don't bother to support different top nybbles
  68          * for LDS/Scratch and GPUVM.
  69          */
  70
  71         WARN_ON((top_address_nybble & 1) || top_address_nybble > 0xE ||
  72                 top_address_nybble == 0);
  73
  74         return PRIVATE_BASE(top_address_nybble << 12) |
  75                         SHARED_BASE(top_address_nybble << 12);
  76 }
  77
  78 static bool set_cache_memory_policy_cik(struct device_queue_manager *dqm,
  79                                    struct qcm_process_device *qpd,
  80                                    enum cache_policy default_policy,
  81                                    enum cache_policy alternate_policy,
  82                                    void __user *alternate_aperture_base,
  83                                    uint64_t alternate_aperture_size)
  84 {
  85         uint32_t default_mtype;
  86         uint32_t ape1_mtype;
  87
  88         default_mtype = (default_policy == cache_policy_coherent) ?
  89                         MTYPE_NONCACHED :
  90                         MTYPE_CACHED;
  91
  92         ape1_mtype = (alternate_policy == cache_policy_coherent) ?
  93                         MTYPE_NONCACHED :
  94                         MTYPE_CACHED;
  95
  96         qpd->sh_mem_config = (qpd->sh_mem_config & PTR32)
  97                         | ALIGNMENT_MODE(SH_MEM_ALIGNMENT_MODE_UNALIGNED)
  98                         | DEFAULT_MTYPE(default_mtype)
  99                         | APE1_MTYPE(ape1_mtype);
 100
 101         return true;
 102 }
 103
 104 static int update_qpd_cik(struct device_queue_manager *dqm,
 105                           struct qcm_process_device *qpd)
 106 {
 107         struct kfd_process_device *pdd;
 108         unsigned int temp;
 109
 110         pdd = qpd_to_pdd(qpd);
 111
 112         /* check if sh_mem_config register already configured */
 113         if (qpd->sh_mem_config == 0) {
 114                 qpd->sh_mem_config =
 115                         ALIGNMENT_MODE(SH_MEM_ALIGNMENT_MODE_UNALIGNED) |
 116                         DEFAULT_MTYPE(MTYPE_NONCACHED) |
 117                         APE1_MTYPE(MTYPE_NONCACHED);
 118                 qpd->sh_mem_ape1_limit = 0;
 119                 qpd->sh_mem_ape1_base = 0;
 120         }
 121
 122         /* On dGPU we're always in GPUVM64 addressing mode with 64-bit
 123          * aperture addresses.
 124          */
 125         temp = get_sh_mem_bases_nybble_64(pdd);
 126         qpd->sh_mem_bases = compute_sh_mem_bases_64bit(temp);
 127
 128         pr_debug("is32bit process: %d sh_mem_bases nybble: 0x%X and register 0x%X\n",
 129                 qpd->pqm->process->is_32bit_user_mode, temp, qpd->sh_mem_bases);
 130
 131         return 0;
 132 }
 133
 134 static void init_sdma_vm(struct device_queue_manager *dqm,
 135                          struct queue *q,
 136                          struct qcm_process_device *qpd)
 137 {
 138         /* On dGPU we're always in GPUVM64 addressing mode with 64-bit
 139          * aperture addresses.
 140          */
 141         q->properties.sdma_vm_addr =
 142                 ((get_sh_mem_bases_nybble_64(qpd_to_pdd(qpd))) <<
 143                  SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE__SHIFT) &
 144                 SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE_MASK;
 145 }