2 * Copyright 2014 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
28 * Christian König <christian.koenig@amd.com>
33 #include "amdgpu_trace.h"
34 #include "amdgpu_amdkfd.h"
36 struct amdgpu_sync_entry
{
37 struct hlist_node node
;
38 struct dma_fence
*fence
;
42 static struct kmem_cache
*amdgpu_sync_slab
;
45 * amdgpu_sync_create - zero init sync object
47 * @sync: sync object to initialize
49 * Just clear the sync object for now.
51 void amdgpu_sync_create(struct amdgpu_sync
*sync
)
53 hash_init(sync
->fences
);
54 sync
->last_vm_update
= NULL
;
58 * amdgpu_sync_same_dev - test if fence belong to us
60 * @adev: amdgpu device to use for the test
63 * Test if the fence was issued by us.
65 static bool amdgpu_sync_same_dev(struct amdgpu_device
*adev
,
68 struct drm_sched_fence
*s_fence
= to_drm_sched_fence(f
);
71 struct amdgpu_ring
*ring
;
73 ring
= container_of(s_fence
->sched
, struct amdgpu_ring
, sched
);
74 return ring
->adev
== adev
;
81 * amdgpu_sync_get_owner - extract the owner of a fence
83 * @fence: fence get the owner from
85 * Extract who originally created the fence.
87 static void *amdgpu_sync_get_owner(struct dma_fence
*f
)
89 struct drm_sched_fence
*s_fence
;
90 struct amdgpu_amdkfd_fence
*kfd_fence
;
93 return AMDGPU_FENCE_OWNER_UNDEFINED
;
95 s_fence
= to_drm_sched_fence(f
);
97 return s_fence
->owner
;
99 kfd_fence
= to_amdgpu_amdkfd_fence(f
);
101 return AMDGPU_FENCE_OWNER_KFD
;
103 return AMDGPU_FENCE_OWNER_UNDEFINED
;
107 * amdgpu_sync_keep_later - Keep the later fence
109 * @keep: existing fence to test
112 * Either keep the existing fence or the new one, depending which one is later.
114 static void amdgpu_sync_keep_later(struct dma_fence
**keep
,
115 struct dma_fence
*fence
)
117 if (*keep
&& dma_fence_is_later(*keep
, fence
))
120 dma_fence_put(*keep
);
121 *keep
= dma_fence_get(fence
);
125 * amdgpu_sync_add_later - add the fence to the hash
127 * @sync: sync object to add the fence to
130 * Tries to add the fence to an existing hash entry. Returns true when an entry
131 * was found, false otherwise.
133 static bool amdgpu_sync_add_later(struct amdgpu_sync
*sync
, struct dma_fence
*f
, bool explicit)
135 struct amdgpu_sync_entry
*e
;
137 hash_for_each_possible(sync
->fences
, e
, node
, f
->context
) {
138 if (unlikely(e
->fence
->context
!= f
->context
))
141 amdgpu_sync_keep_later(&e
->fence
, f
);
143 /* Preserve eplicit flag to not loose pipe line sync */
144 e
->explicit |= explicit;
152 * amdgpu_sync_fence - remember to sync to this fence
154 * @sync: sync object to add fence to
155 * @fence: fence to sync to
158 int amdgpu_sync_fence(struct amdgpu_device
*adev
, struct amdgpu_sync
*sync
,
159 struct dma_fence
*f
, bool explicit)
161 struct amdgpu_sync_entry
*e
;
165 if (amdgpu_sync_same_dev(adev
, f
) &&
166 amdgpu_sync_get_owner(f
) == AMDGPU_FENCE_OWNER_VM
)
167 amdgpu_sync_keep_later(&sync
->last_vm_update
, f
);
169 if (amdgpu_sync_add_later(sync
, f
, explicit))
172 e
= kmem_cache_alloc(amdgpu_sync_slab
, GFP_KERNEL
);
176 e
->explicit = explicit;
178 hash_add(sync
->fences
, &e
->node
, f
->context
);
179 e
->fence
= dma_fence_get(f
);
184 * amdgpu_sync_resv - sync to a reservation object
186 * @sync: sync object to add fences from reservation object to
187 * @resv: reservation object with embedded fence
188 * @explicit_sync: true if we should only sync to the exclusive fence
192 int amdgpu_sync_resv(struct amdgpu_device
*adev
,
193 struct amdgpu_sync
*sync
,
194 struct reservation_object
*resv
,
195 void *owner
, bool explicit_sync
)
197 struct reservation_object_list
*flist
;
206 /* always sync to the exclusive fence */
207 f
= reservation_object_get_excl(resv
);
208 r
= amdgpu_sync_fence(adev
, sync
, f
, false);
210 flist
= reservation_object_get_list(resv
);
214 for (i
= 0; i
< flist
->shared_count
; ++i
) {
215 f
= rcu_dereference_protected(flist
->shared
[i
],
216 reservation_object_held(resv
));
217 /* We only want to trigger KFD eviction fences on
218 * evict or move jobs. Skip KFD fences otherwise.
220 fence_owner
= amdgpu_sync_get_owner(f
);
221 if (fence_owner
== AMDGPU_FENCE_OWNER_KFD
&&
222 owner
!= AMDGPU_FENCE_OWNER_UNDEFINED
)
225 if (amdgpu_sync_same_dev(adev
, f
)) {
226 /* VM updates are only interesting
227 * for other VM updates and moves.
229 if ((owner
!= AMDGPU_FENCE_OWNER_UNDEFINED
) &&
230 (fence_owner
!= AMDGPU_FENCE_OWNER_UNDEFINED
) &&
231 ((owner
== AMDGPU_FENCE_OWNER_VM
) !=
232 (fence_owner
== AMDGPU_FENCE_OWNER_VM
)))
235 /* Ignore fence from the same owner and explicit one as
236 * long as it isn't undefined.
238 if (owner
!= AMDGPU_FENCE_OWNER_UNDEFINED
&&
239 (fence_owner
== owner
|| explicit_sync
))
243 r
= amdgpu_sync_fence(adev
, sync
, f
, false);
251 * amdgpu_sync_peek_fence - get the next fence not signaled yet
253 * @sync: the sync object
254 * @ring: optional ring to use for test
256 * Returns the next fence not signaled yet without removing it from the sync
259 struct dma_fence
*amdgpu_sync_peek_fence(struct amdgpu_sync
*sync
,
260 struct amdgpu_ring
*ring
)
262 struct amdgpu_sync_entry
*e
;
263 struct hlist_node
*tmp
;
266 hash_for_each_safe(sync
->fences
, i
, tmp
, e
, node
) {
267 struct dma_fence
*f
= e
->fence
;
268 struct drm_sched_fence
*s_fence
= to_drm_sched_fence(f
);
270 if (dma_fence_is_signaled(f
)) {
273 kmem_cache_free(amdgpu_sync_slab
, e
);
276 if (ring
&& s_fence
) {
277 /* For fences from the same ring it is sufficient
278 * when they are scheduled.
280 if (s_fence
->sched
== &ring
->sched
) {
281 if (dma_fence_is_signaled(&s_fence
->scheduled
))
284 return &s_fence
->scheduled
;
295 * amdgpu_sync_get_fence - get the next fence from the sync object
297 * @sync: sync object to use
298 * @explicit: true if the next fence is explicit
300 * Get and removes the next fence from the sync object not signaled yet.
302 struct dma_fence
*amdgpu_sync_get_fence(struct amdgpu_sync
*sync
, bool *explicit)
304 struct amdgpu_sync_entry
*e
;
305 struct hlist_node
*tmp
;
308 hash_for_each_safe(sync
->fences
, i
, tmp
, e
, node
) {
312 *explicit = e
->explicit;
315 kmem_cache_free(amdgpu_sync_slab
, e
);
317 if (!dma_fence_is_signaled(f
))
326 * amdgpu_sync_clone - clone a sync object
328 * @source: sync object to clone
329 * @clone: pointer to destination sync object
331 * Adds references to all unsignaled fences in @source to @clone. Also
332 * removes signaled fences from @source while at it.
334 int amdgpu_sync_clone(struct amdgpu_sync
*source
, struct amdgpu_sync
*clone
)
336 struct amdgpu_sync_entry
*e
;
337 struct hlist_node
*tmp
;
341 hash_for_each_safe(source
->fences
, i
, tmp
, e
, node
) {
343 if (!dma_fence_is_signaled(f
)) {
344 r
= amdgpu_sync_fence(NULL
, clone
, f
, e
->explicit);
350 kmem_cache_free(amdgpu_sync_slab
, e
);
354 dma_fence_put(clone
->last_vm_update
);
355 clone
->last_vm_update
= dma_fence_get(source
->last_vm_update
);
360 int amdgpu_sync_wait(struct amdgpu_sync
*sync
, bool intr
)
362 struct amdgpu_sync_entry
*e
;
363 struct hlist_node
*tmp
;
366 hash_for_each_safe(sync
->fences
, i
, tmp
, e
, node
) {
367 r
= dma_fence_wait(e
->fence
, intr
);
372 dma_fence_put(e
->fence
);
373 kmem_cache_free(amdgpu_sync_slab
, e
);
380 * amdgpu_sync_free - free the sync object
382 * @sync: sync object to use
384 * Free the sync object.
386 void amdgpu_sync_free(struct amdgpu_sync
*sync
)
388 struct amdgpu_sync_entry
*e
;
389 struct hlist_node
*tmp
;
392 hash_for_each_safe(sync
->fences
, i
, tmp
, e
, node
) {
394 dma_fence_put(e
->fence
);
395 kmem_cache_free(amdgpu_sync_slab
, e
);
398 dma_fence_put(sync
->last_vm_update
);
402 * amdgpu_sync_init - init sync object subsystem
404 * Allocate the slab allocator.
406 int amdgpu_sync_init(void)
408 amdgpu_sync_slab
= kmem_cache_create(
409 "amdgpu_sync", sizeof(struct amdgpu_sync_entry
), 0,
410 SLAB_HWCACHE_ALIGN
, NULL
);
411 if (!amdgpu_sync_slab
)
418 * amdgpu_sync_fini - fini sync object subsystem
420 * Free the slab allocator.
422 void amdgpu_sync_fini(void)
424 kmem_cache_destroy(amdgpu_sync_slab
);