These changes are the raw update to linux-4.4.6-rt14. Kernel sources
[kvmfornfv.git] / kernel / drivers / gpu / drm / amd / amdkfd / kfd_device_queue_manager_vi.c
1 /*
2  * Copyright 2014 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 #include "kfd_device_queue_manager.h"
25 #include "gca/gfx_8_0_enum.h"
26 #include "gca/gfx_8_0_sh_mask.h"
27 #include "gca/gfx_8_0_enum.h"
28 #include "oss/oss_3_0_sh_mask.h"
29
30 static bool set_cache_memory_policy_vi(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 register_process_vi(struct device_queue_manager *dqm,
37                                         struct qcm_process_device *qpd);
38 static int initialize_cpsch_vi(struct device_queue_manager *dqm);
39 static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
40                                 struct qcm_process_device *qpd);
41
42 void device_queue_manager_init_vi(struct device_queue_manager_asic_ops *ops)
43 {
44         ops->set_cache_memory_policy = set_cache_memory_policy_vi;
45         ops->register_process = register_process_vi;
46         ops->initialize = initialize_cpsch_vi;
47         ops->init_sdma_vm = init_sdma_vm;
48 }
49
50 static uint32_t compute_sh_mem_bases_64bit(unsigned int top_address_nybble)
51 {
52         /* In 64-bit mode, we can only control the top 3 bits of the LDS,
53          * scratch and GPUVM apertures.
54          * The hardware fills in the remaining 59 bits according to the
55          * following pattern:
56          * LDS:         X0000000'00000000 - X0000001'00000000 (4GB)
57          * Scratch:     X0000001'00000000 - X0000002'00000000 (4GB)
58          * GPUVM:       Y0010000'00000000 - Y0020000'00000000 (1TB)
59          *
60          * (where X/Y is the configurable nybble with the low-bit 0)
61          *
62          * LDS and scratch will have the same top nybble programmed in the
63          * top 3 bits of SH_MEM_BASES.PRIVATE_BASE.
64          * GPUVM can have a different top nybble programmed in the
65          * top 3 bits of SH_MEM_BASES.SHARED_BASE.
66          * We don't bother to support different top nybbles
67          * for LDS/Scratch and GPUVM.
68          */
69
70         BUG_ON((top_address_nybble & 1) || top_address_nybble > 0xE ||
71                 top_address_nybble == 0);
72
73         return top_address_nybble << 12 |
74                         (top_address_nybble << 12) <<
75                         SH_MEM_BASES__SHARED_BASE__SHIFT;
76 }
77
78 static bool set_cache_memory_policy_vi(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_CC :
90                         MTYPE_NC;
91
92         ape1_mtype = (alternate_policy == cache_policy_coherent) ?
93                         MTYPE_CC :
94                         MTYPE_NC;
95
96         qpd->sh_mem_config = (qpd->sh_mem_config &
97                         SH_MEM_CONFIG__ADDRESS_MODE_MASK) |
98                 SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
99                                 SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
100                 default_mtype << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
101                 ape1_mtype << SH_MEM_CONFIG__APE1_MTYPE__SHIFT |
102                 SH_MEM_CONFIG__PRIVATE_ATC_MASK;
103
104         return true;
105 }
106
107 static int register_process_vi(struct device_queue_manager *dqm,
108                                         struct qcm_process_device *qpd)
109 {
110         struct kfd_process_device *pdd;
111         unsigned int temp;
112
113         BUG_ON(!dqm || !qpd);
114
115         pdd = qpd_to_pdd(qpd);
116
117         /* check if sh_mem_config register already configured */
118         if (qpd->sh_mem_config == 0) {
119                 qpd->sh_mem_config =
120                         SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
121                                 SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
122                         MTYPE_CC << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
123                         MTYPE_CC << SH_MEM_CONFIG__APE1_MTYPE__SHIFT |
124                         SH_MEM_CONFIG__PRIVATE_ATC_MASK;
125
126                 qpd->sh_mem_ape1_limit = 0;
127                 qpd->sh_mem_ape1_base = 0;
128         }
129
130         if (qpd->pqm->process->is_32bit_user_mode) {
131                 temp = get_sh_mem_bases_32(pdd);
132                 qpd->sh_mem_bases = temp << SH_MEM_BASES__SHARED_BASE__SHIFT;
133                 qpd->sh_mem_config |= SH_MEM_ADDRESS_MODE_HSA32 <<
134                                         SH_MEM_CONFIG__ADDRESS_MODE__SHIFT;
135         } else {
136                 temp = get_sh_mem_bases_nybble_64(pdd);
137                 qpd->sh_mem_bases = compute_sh_mem_bases_64bit(temp);
138                 qpd->sh_mem_config |= SH_MEM_ADDRESS_MODE_HSA64 <<
139                         SH_MEM_CONFIG__ADDRESS_MODE__SHIFT;
140         }
141
142         pr_debug("kfd: is32bit process: %d sh_mem_bases nybble: 0x%X and register 0x%X\n",
143                 qpd->pqm->process->is_32bit_user_mode, temp, qpd->sh_mem_bases);
144
145         return 0;
146 }
147
148 static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
149                                 struct qcm_process_device *qpd)
150 {
151         uint32_t value = (1 << SDMA0_RLC0_VIRTUAL_ADDR__ATC__SHIFT);
152
153         if (q->process->is_32bit_user_mode)
154                 value |= (1 << SDMA0_RLC0_VIRTUAL_ADDR__PTR32__SHIFT) |
155                                 get_sh_mem_bases_32(qpd_to_pdd(qpd));
156         else
157                 value |= ((get_sh_mem_bases_nybble_64(qpd_to_pdd(qpd))) <<
158                                 SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE__SHIFT) &
159                                 SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE_MASK;
160
161         q->properties.sdma_vm_addr = value;
162 }
163
164 static int initialize_cpsch_vi(struct device_queue_manager *dqm)
165 {
166         return 0;
167 }