1 // SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note
2 /*
3 *
4 * (C) COPYRIGHT 2019-2021 ARM Limited. All rights reserved.
5 *
6 * This program is free software and is provided to you under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation, and any use by you of this program is subject to the terms
9 * of such GNU license.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, you can access it online at
18 * http://www.gnu.org/licenses/gpl-2.0.html.
19 *
20 */
21
22 #include <tl/mali_kbase_tracepoints.h>
23 #include <tl/mali_kbase_timeline.h>
24 #include <tl/mali_kbase_timeline_priv.h>
25
26 #include <mali_kbase.h>
27
28 #define GPU_FEATURES_CROSS_STREAM_SYNC_MASK (1ull << 3ull)
29
kbase_create_timeline_objects(struct kbase_device * kbdev)30 void kbase_create_timeline_objects(struct kbase_device *kbdev)
31 {
32 unsigned int as_nr;
33 unsigned int slot_i;
34 struct kbase_context *kctx;
35 struct kbase_timeline *timeline = kbdev->timeline;
36 struct kbase_tlstream *summary =
37 &kbdev->timeline->streams[TL_STREAM_TYPE_OBJ_SUMMARY];
38 u32 const kbdev_has_cross_stream_sync =
39 (kbdev->gpu_props.props.raw_props.gpu_features &
40 GPU_FEATURES_CROSS_STREAM_SYNC_MASK) ?
41 1 :
42 0;
43 u32 const arch_maj = (kbdev->gpu_props.props.raw_props.gpu_id &
44 GPU_ID2_ARCH_MAJOR) >>
45 GPU_ID2_ARCH_MAJOR_SHIFT;
46 u32 const num_sb_entries = arch_maj >= 11 ? 16 : 8;
47 u32 const supports_gpu_sleep =
48 #ifdef KBASE_PM_RUNTIME
49 kbdev->pm.backend.gpu_sleep_supported;
50 #else
51 false;
52 #endif /* KBASE_PM_RUNTIME */
53
54 /* Summarize the Address Space objects. */
55 for (as_nr = 0; as_nr < kbdev->nr_hw_address_spaces; as_nr++)
56 __kbase_tlstream_tl_new_as(summary, &kbdev->as[as_nr], as_nr);
57
58 /* Create Legacy GPU object to track in AOM for dumping */
59 __kbase_tlstream_tl_new_gpu(summary,
60 kbdev,
61 kbdev->gpu_props.props.raw_props.gpu_id,
62 kbdev->gpu_props.num_cores);
63
64
65 for (as_nr = 0; as_nr < kbdev->nr_hw_address_spaces; as_nr++)
66 __kbase_tlstream_tl_lifelink_as_gpu(summary,
67 &kbdev->as[as_nr],
68 kbdev);
69
70 /* Trace the creation of a new kbase device and set its properties. */
71 __kbase_tlstream_tl_kbase_new_device(summary, kbdev->gpu_props.props.raw_props.gpu_id,
72 kbdev->gpu_props.num_cores,
73 kbdev->csf.global_iface.group_num,
74 kbdev->nr_hw_address_spaces, num_sb_entries,
75 kbdev_has_cross_stream_sync, supports_gpu_sleep);
76
77 /* Lock the context list, to ensure no changes to the list are made
78 * while we're summarizing the contexts and their contents.
79 */
80 mutex_lock(&timeline->tl_kctx_list_lock);
81
82 /* Hold the scheduler lock while we emit the current state
83 * We also need to continue holding the lock until after the first body
84 * stream tracepoints are emitted to ensure we don't change the
85 * scheduler until after then
86 */
87 mutex_lock(&kbdev->csf.scheduler.lock);
88
89 for (slot_i = 0; slot_i < kbdev->csf.global_iface.group_num; slot_i++) {
90
91 struct kbase_queue_group *group =
92 kbdev->csf.scheduler.csg_slots[slot_i].resident_group;
93
94 if (group)
95 __kbase_tlstream_tl_kbase_device_program_csg(
96 summary,
97 kbdev->gpu_props.props.raw_props.gpu_id,
98 group->kctx->id, group->handle, slot_i, 0);
99 }
100
101 /* Reset body stream buffers while holding the kctx lock.
102 * As we are holding the lock, we can guarantee that no kctx creation or
103 * deletion tracepoints can be fired from outside of this function by
104 * some other thread.
105 */
106 kbase_timeline_streams_body_reset(timeline);
107
108 mutex_unlock(&kbdev->csf.scheduler.lock);
109
110 /* For each context in the device... */
111 list_for_each_entry(kctx, &timeline->tl_kctx_list, tl_kctx_list_node) {
112 size_t i;
113 struct kbase_tlstream *body =
114 &timeline->streams[TL_STREAM_TYPE_OBJ];
115
116 /* Lock the context's KCPU queues, to ensure no KCPU-queue
117 * related actions can occur in this context from now on.
118 */
119 mutex_lock(&kctx->csf.kcpu_queues.lock);
120
121 /* Acquire the MMU lock, to ensure we don't get a concurrent
122 * address space assignment while summarizing this context's
123 * address space.
124 */
125 mutex_lock(&kbdev->mmu_hw_mutex);
126
127 /* Trace the context itself into the body stream, not the
128 * summary stream.
129 * We place this in the body to ensure it is ordered after any
130 * other tracepoints related to the contents of the context that
131 * might have been fired before acquiring all of the per-context
132 * locks.
133 * This ensures that those tracepoints will not actually affect
134 * the object model state, as they reference a context that
135 * hasn't been traced yet. They may, however, cause benign
136 * errors to be emitted.
137 */
138 __kbase_tlstream_tl_kbase_new_ctx(body, kctx->id,
139 kbdev->gpu_props.props.raw_props.gpu_id);
140
141 /* Also trace with the legacy AOM tracepoint for dumping */
142 __kbase_tlstream_tl_new_ctx(body,
143 kctx,
144 kctx->id,
145 (u32)(kctx->tgid));
146
147 /* Trace the currently assigned address space */
148 if (kctx->as_nr != KBASEP_AS_NR_INVALID)
149 __kbase_tlstream_tl_kbase_ctx_assign_as(body, kctx->id,
150 kctx->as_nr);
151
152
153 /* Trace all KCPU queues in the context into the body stream.
154 * As we acquired the KCPU lock after resetting the body stream,
155 * it's possible that some KCPU-related events for this context
156 * occurred between that reset and now.
157 * These will cause errors to be emitted when parsing the
158 * timeline, but they will not affect the correctness of the
159 * object model.
160 */
161 for (i = 0; i < KBASEP_MAX_KCPU_QUEUES; i++) {
162 const struct kbase_kcpu_command_queue *kcpu_queue =
163 kctx->csf.kcpu_queues.array[i];
164
165 if (kcpu_queue)
166 __kbase_tlstream_tl_kbase_new_kcpuqueue(
167 body, kcpu_queue, kcpu_queue->id, kcpu_queue->kctx->id,
168 kcpu_queue->num_pending_cmds);
169 }
170
171 mutex_unlock(&kbdev->mmu_hw_mutex);
172 mutex_unlock(&kctx->csf.kcpu_queues.lock);
173
174 /* Now that all per-context locks for this context have been
175 * released, any per-context tracepoints that are fired from
176 * any other threads will go into the body stream after
177 * everything that was just summarised into the body stream in
178 * this iteration of the loop, so will start to correctly update
179 * the object model state.
180 */
181 }
182
183 mutex_unlock(&timeline->tl_kctx_list_lock);
184
185 /* Static object are placed into summary packet that needs to be
186 * transmitted first. Flush all streams to make it available to
187 * user space.
188 */
189 kbase_timeline_streams_flush(timeline);
190 }
191