diff options
-rw-r--r-- | drivers/gpu/nvgpu/gp10b/cde_gp10b.c | 84 |
1 files changed, 84 insertions, 0 deletions
diff --git a/drivers/gpu/nvgpu/gp10b/cde_gp10b.c b/drivers/gpu/nvgpu/gp10b/cde_gp10b.c index acb8aee3..dadec4d2 100644 --- a/drivers/gpu/nvgpu/gp10b/cde_gp10b.c +++ b/drivers/gpu/nvgpu/gp10b/cde_gp10b.c | |||
@@ -58,7 +58,91 @@ static void gp10b_cde_get_program_numbers(struct gk20a *g, | |||
58 | *vprog_out = vprog; | 58 | *vprog_out = vprog; |
59 | } | 59 | } |
60 | 60 | ||
61 | static bool gp10b_need_scatter_buffer(struct gk20a *g) | ||
62 | { | ||
63 | return g->mm.bypass_smmu; | ||
64 | } | ||
65 | |||
66 | static u8 parity(u32 a) | ||
67 | { | ||
68 | a ^= a>>16u; | ||
69 | a ^= a>>8u; | ||
70 | a ^= a>>4u; | ||
71 | a &= 0xfu; | ||
72 | return (0x6996u >> a) & 1u; | ||
73 | } | ||
74 | |||
75 | static int gp10b_populate_scatter_buffer(struct gk20a *g, | ||
76 | struct sg_table *sgt, | ||
77 | size_t surface_size, | ||
78 | void *scatter_buffer_ptr, | ||
79 | size_t scatter_buffer_size) | ||
80 | { | ||
81 | /* map scatter buffer to CPU VA and fill it */ | ||
82 | const u32 page_size_log2 = 12; | ||
83 | const u32 page_size = 1 << page_size_log2; | ||
84 | const u32 page_size_shift = page_size_log2 - 7u; | ||
85 | |||
86 | /* 0011 1111 1111 1111 1111 1110 0100 1000 */ | ||
87 | const u32 getSliceMaskGP10B = 0x3ffffe48; | ||
88 | u8 *scatter_buffer = scatter_buffer_ptr; | ||
89 | |||
90 | size_t i; | ||
91 | struct scatterlist *sg = NULL; | ||
92 | u8 d = 0; | ||
93 | size_t page = 0; | ||
94 | size_t pages_left; | ||
95 | |||
96 | surface_size = round_up(surface_size, page_size); | ||
97 | |||
98 | pages_left = surface_size >> page_size_log2; | ||
99 | if ((pages_left >> 3) > scatter_buffer_size) | ||
100 | return -ENOMEM; | ||
101 | |||
102 | for_each_sg(sgt->sgl, sg, sgt->nents, i) { | ||
103 | unsigned int j; | ||
104 | u64 surf_pa = sg_phys(sg); | ||
105 | unsigned int n = (int)(sg->length >> page_size_log2); | ||
106 | |||
107 | gk20a_dbg(gpu_dbg_cde, "surfPA=0x%llx + %d pages", surf_pa, n); | ||
108 | |||
109 | for (j=0; j < n && pages_left > 0; j++, surf_pa += page_size) { | ||
110 | u32 addr = (((u32)(surf_pa>>7)) & getSliceMaskGP10B) >> page_size_shift; | ||
111 | u8 scatter_bit = parity(addr); | ||
112 | u8 bit = page & 7; | ||
113 | |||
114 | d |= scatter_bit << bit; | ||
115 | if (bit == 7) { | ||
116 | scatter_buffer[page >> 3] = d; | ||
117 | d = 0; | ||
118 | } | ||
119 | |||
120 | ++page; | ||
121 | --pages_left; | ||
122 | } | ||
123 | |||
124 | if (pages_left == 0) | ||
125 | break; | ||
126 | } | ||
127 | |||
128 | /* write the last byte in case the number of pages is not divisible by 8 */ | ||
129 | if ((page & 7) != 0) | ||
130 | scatter_buffer[page >> 3] = d; | ||
131 | |||
132 | #if defined(GK20A_DEBUG) | ||
133 | if (unlikely(gpu_dbg_cde & gk20a_dbg_mask)) { | ||
134 | gk20a_dbg(gpu_dbg_cde, "scatterBuffer content:"); | ||
135 | for (i=0; i < page>>3; i++) { | ||
136 | gk20a_dbg(gpu_dbg_cde, " %x", scatter_buffer[i]); | ||
137 | } | ||
138 | } | ||
139 | #endif | ||
140 | return 0; | ||
141 | } | ||
142 | |||
61 | void gp10b_init_cde_ops(struct gpu_ops *gops) | 143 | void gp10b_init_cde_ops(struct gpu_ops *gops) |
62 | { | 144 | { |
63 | gops->cde.get_program_numbers = gp10b_cde_get_program_numbers; | 145 | gops->cde.get_program_numbers = gp10b_cde_get_program_numbers; |
146 | gops->cde.need_scatter_buffer = gp10b_need_scatter_buffer; | ||
147 | gops->cde.populate_scatter_buffer = gp10b_populate_scatter_buffer; | ||
64 | } | 148 | } |