2 * Copyright 2016-2018 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
23 #include <linux/dma-fence.h>
24 #include <linux/spinlock.h>
25 #include <linux/atomic.h>
26 #include <linux/stacktrace.h>
27 #include <linux/sched.h>
28 #include <linux/slab.h>
29 #include <linux/sched/mm.h>
30 #include "amdgpu_amdkfd.h"
33 static const struct dma_fence_ops amdkfd_fence_ops;
34 static atomic_t fence_seq = ATOMIC_INIT(0);
37 * Fence helper functions to deal with KFD memory eviction.
38 * Big Idea - Since KFD submissions are done by user queues, a BO cannot be
39 * evicted unless all the user queues for that process are evicted.
41 * All the BOs in a process share an eviction fence. When process X wants
42 * to map VRAM memory but TTM can't find enough space, TTM will attempt to
43 * evict BOs from its LRU list. TTM checks if the BO is valuable to evict
44 * by calling ttm_device_funcs->eviction_valuable().
46 * ttm_device_funcs->eviction_valuable() - will return false if the BO belongs
47 * to process X. Otherwise, it will return true to indicate BO can be
50 * If ttm_device_funcs->eviction_valuable returns true, then TTM will continue
51 * the evcition process for that BO by calling ttm_bo_evict --> amdgpu_bo_move
52 * --> amdgpu_copy_buffer(). This sets up job in GPU scheduler.
54 * GPU Scheduler (amd_sched_main) - sets up a cb (fence_add_callback) to
55 * nofity when the BO is free to move. fence_add_callback --> enable_signaling
56 * --> amdgpu_amdkfd_fence.enable_signaling
58 * amdgpu_amdkfd_fence.enable_signaling - Start a work item that will quiesce
59 * user queues and signal fence. The work item will also start another delayed
60 * work item to restore BOs
63 struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
65 struct svm_range_bo *svm_bo)
67 struct amdgpu_amdkfd_fence *fence;
69 fence = kzalloc(sizeof(*fence), GFP_KERNEL);
73 /* This reference gets released in amdkfd_fence_release */
76 get_task_comm(fence->timeline_name, current);
77 spin_lock_init(&fence->lock);
78 fence->svm_bo = svm_bo;
79 dma_fence_init(&fence->base, &amdkfd_fence_ops, &fence->lock,
80 context, atomic_inc_return(&fence_seq));
85 struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f)
87 struct amdgpu_amdkfd_fence *fence;
92 fence = container_of(f, struct amdgpu_amdkfd_fence, base);
93 if (fence && f->ops == &amdkfd_fence_ops)
99 static const char *amdkfd_fence_get_driver_name(struct dma_fence *f)
101 return "amdgpu_amdkfd_fence";
104 static const char *amdkfd_fence_get_timeline_name(struct dma_fence *f)
106 struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
108 return fence->timeline_name;
112 * amdkfd_fence_enable_signaling - This gets called when TTM wants to evict
113 * a KFD BO and schedules a job to move the BO.
114 * If fence is already signaled return true.
115 * If fence is not signaled schedule a evict KFD process work item.
119 static bool amdkfd_fence_enable_signaling(struct dma_fence *f)
121 struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
126 if (dma_fence_is_signaled(f))
129 if (!fence->svm_bo) {
130 if (!kgd2kfd_schedule_evict_and_restore_process(fence->mm, f))
133 if (!svm_range_schedule_evict_svm_bo(fence))
140 * amdkfd_fence_release - callback that fence can be freed
144 * This function is called when the reference count becomes zero.
145 * Drops the mm_struct reference and RCU schedules freeing up the fence.
147 static void amdkfd_fence_release(struct dma_fence *f)
149 struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
151 /* Unconditionally signal the fence. The process is getting
155 return; /* Not an amdgpu_amdkfd_fence */
162 * amdkfd_fence_check_mm - Check whether to prevent eviction of @f by @mm
165 * @mm: [IN] mm that needs to be verified
167 * Check if @mm is same as that of the fence @f, if same return TRUE else
169 * For svm bo, which support vram overcommitment, always return FALSE.
171 bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
173 struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
177 else if (fence->mm == mm && !fence->svm_bo)
183 static const struct dma_fence_ops amdkfd_fence_ops = {
184 .get_driver_name = amdkfd_fence_get_driver_name,
185 .get_timeline_name = amdkfd_fence_get_timeline_name,
186 .enable_signaling = amdkfd_fence_enable_signaling,
187 .release = amdkfd_fence_release,