1 // SPDX-License-Identifier: MIT
2 /*
3 * Copyright © 2022 Intel Corporation
4 */
5
6 #include "xe_bb.h"
7
8 #include "instructions/xe_mi_commands.h"
9 #include "xe_assert.h"
10 #include "xe_device.h"
11 #include "xe_exec_queue_types.h"
12 #include "xe_gt.h"
13 #include "xe_hw_fence.h"
14 #include "xe_sa.h"
15 #include "xe_sched_job.h"
16 #include "xe_vm_types.h"
17
bb_prefetch(struct xe_gt * gt)18 static int bb_prefetch(struct xe_gt *gt)
19 {
20 struct xe_device *xe = gt_to_xe(gt);
21
22 if (GRAPHICS_VERx100(xe) >= 1250 && !xe_gt_is_media_type(gt))
23 /*
24 * RCS and CCS require 1K, although other engines would be
25 * okay with 512.
26 */
27 return SZ_1K;
28 else
29 return SZ_512;
30 }
31
xe_bb_new(struct xe_gt * gt,u32 dwords,bool usm)32 struct xe_bb *xe_bb_new(struct xe_gt *gt, u32 dwords, bool usm)
33 {
34 struct xe_tile *tile = gt_to_tile(gt);
35 struct xe_bb *bb = kmalloc(sizeof(*bb), GFP_KERNEL);
36 int err;
37
38 if (!bb)
39 return ERR_PTR(-ENOMEM);
40
41 /*
42 * We need to allocate space for the requested number of dwords,
43 * one additional MI_BATCH_BUFFER_END dword, and additional buffer
44 * space to accomodate the platform-specific hardware prefetch
45 * requirements.
46 */
47 bb->bo = xe_sa_bo_new(!usm ? tile->mem.kernel_bb_pool : gt->usm.bb_pool,
48 4 * (dwords + 1) + bb_prefetch(gt));
49 if (IS_ERR(bb->bo)) {
50 err = PTR_ERR(bb->bo);
51 goto err;
52 }
53
54 bb->cs = xe_sa_bo_cpu_addr(bb->bo);
55 bb->len = 0;
56
57 return bb;
58 err:
59 kfree(bb);
60 return ERR_PTR(err);
61 }
62
63 static struct xe_sched_job *
__xe_bb_create_job(struct xe_exec_queue * q,struct xe_bb * bb,u64 * addr)64 __xe_bb_create_job(struct xe_exec_queue *q, struct xe_bb *bb, u64 *addr)
65 {
66 u32 size = drm_suballoc_size(bb->bo);
67
68 if (bb->len == 0 || bb->cs[bb->len - 1] != MI_BATCH_BUFFER_END)
69 bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
70
71 xe_gt_assert(q->gt, bb->len * 4 + bb_prefetch(q->gt) <= size);
72
73 xe_sa_bo_flush_write(bb->bo);
74
75 return xe_sched_job_create(q, addr);
76 }
77
xe_bb_create_migration_job(struct xe_exec_queue * q,struct xe_bb * bb,u64 batch_base_ofs,u32 second_idx)78 struct xe_sched_job *xe_bb_create_migration_job(struct xe_exec_queue *q,
79 struct xe_bb *bb,
80 u64 batch_base_ofs,
81 u32 second_idx)
82 {
83 u64 addr[2] = {
84 batch_base_ofs + drm_suballoc_soffset(bb->bo),
85 batch_base_ofs + drm_suballoc_soffset(bb->bo) +
86 4 * second_idx,
87 };
88
89 xe_gt_assert(q->gt, second_idx <= bb->len);
90 xe_gt_assert(q->gt, xe_sched_job_is_migration(q));
91 xe_gt_assert(q->gt, q->width == 1);
92
93 return __xe_bb_create_job(q, bb, addr);
94 }
95
xe_bb_create_job(struct xe_exec_queue * q,struct xe_bb * bb)96 struct xe_sched_job *xe_bb_create_job(struct xe_exec_queue *q,
97 struct xe_bb *bb)
98 {
99 u64 addr = xe_sa_bo_gpu_addr(bb->bo);
100
101 xe_gt_assert(q->gt, !xe_sched_job_is_migration(q));
102 xe_gt_assert(q->gt, q->width == 1);
103 return __xe_bb_create_job(q, bb, &addr);
104 }
105
xe_bb_free(struct xe_bb * bb,struct dma_fence * fence)106 void xe_bb_free(struct xe_bb *bb, struct dma_fence *fence)
107 {
108 if (!bb)
109 return;
110
111 xe_sa_bo_free(bb->bo, fence);
112 kfree(bb);
113 }
114