]> Git Repo - linux.git/blob - drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
Merge tag 'turbostat-2025.02.02' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux.git] / drivers / gpu / drm / amd / amdgpu / amdgpu_ids.c
1 /*
2  * Copyright 2017 Advanced Micro Devices, Inc.
3  *
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:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
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.
21  *
22  */
23 #include "amdgpu_ids.h"
24
25 #include <linux/idr.h>
26 #include <linux/dma-fence-array.h>
27
28
29 #include "amdgpu.h"
30 #include "amdgpu_trace.h"
31
32 /*
33  * PASID manager
34  *
35  * PASIDs are global address space identifiers that can be shared
36  * between the GPU, an IOMMU and the driver. VMs on different devices
37  * may use the same PASID if they share the same address
38  * space. Therefore PASIDs are allocated using a global IDA. VMs are
39  * looked up from the PASID per amdgpu_device.
40  */
41 static DEFINE_IDA(amdgpu_pasid_ida);
42
43 /* Helper to free pasid from a fence callback */
44 struct amdgpu_pasid_cb {
45         struct dma_fence_cb cb;
46         u32 pasid;
47 };
48
49 /**
50  * amdgpu_pasid_alloc - Allocate a PASID
51  * @bits: Maximum width of the PASID in bits, must be at least 1
52  *
53  * Allocates a PASID of the given width while keeping smaller PASIDs
54  * available if possible.
55  *
56  * Returns a positive integer on success. Returns %-EINVAL if bits==0.
57  * Returns %-ENOSPC if no PASID was available. Returns %-ENOMEM on
58  * memory allocation failure.
59  */
60 int amdgpu_pasid_alloc(unsigned int bits)
61 {
62         int pasid = -EINVAL;
63
64         for (bits = min(bits, 31U); bits > 0; bits--) {
65                 pasid = ida_alloc_range(&amdgpu_pasid_ida, 1U << (bits - 1),
66                                         (1U << bits) - 1, GFP_KERNEL);
67                 if (pasid != -ENOSPC)
68                         break;
69         }
70
71         if (pasid >= 0)
72                 trace_amdgpu_pasid_allocated(pasid);
73
74         return pasid;
75 }
76
77 /**
78  * amdgpu_pasid_free - Free a PASID
79  * @pasid: PASID to free
80  */
81 void amdgpu_pasid_free(u32 pasid)
82 {
83         trace_amdgpu_pasid_freed(pasid);
84         ida_free(&amdgpu_pasid_ida, pasid);
85 }
86
87 static void amdgpu_pasid_free_cb(struct dma_fence *fence,
88                                  struct dma_fence_cb *_cb)
89 {
90         struct amdgpu_pasid_cb *cb =
91                 container_of(_cb, struct amdgpu_pasid_cb, cb);
92
93         amdgpu_pasid_free(cb->pasid);
94         dma_fence_put(fence);
95         kfree(cb);
96 }
97
98 /**
99  * amdgpu_pasid_free_delayed - free pasid when fences signal
100  *
101  * @resv: reservation object with the fences to wait for
102  * @pasid: pasid to free
103  *
104  * Free the pasid only after all the fences in resv are signaled.
105  */
106 void amdgpu_pasid_free_delayed(struct dma_resv *resv,
107                                u32 pasid)
108 {
109         struct amdgpu_pasid_cb *cb;
110         struct dma_fence *fence;
111         int r;
112
113         r = dma_resv_get_singleton(resv, DMA_RESV_USAGE_BOOKKEEP, &fence);
114         if (r)
115                 goto fallback;
116
117         if (!fence) {
118                 amdgpu_pasid_free(pasid);
119                 return;
120         }
121
122         cb = kmalloc(sizeof(*cb), GFP_KERNEL);
123         if (!cb) {
124                 /* Last resort when we are OOM */
125                 dma_fence_wait(fence, false);
126                 dma_fence_put(fence);
127                 amdgpu_pasid_free(pasid);
128         } else {
129                 cb->pasid = pasid;
130                 if (dma_fence_add_callback(fence, &cb->cb,
131                                            amdgpu_pasid_free_cb))
132                         amdgpu_pasid_free_cb(fence, &cb->cb);
133         }
134
135         return;
136
137 fallback:
138         /* Not enough memory for the delayed delete, as last resort
139          * block for all the fences to complete.
140          */
141         dma_resv_wait_timeout(resv, DMA_RESV_USAGE_BOOKKEEP,
142                               false, MAX_SCHEDULE_TIMEOUT);
143         amdgpu_pasid_free(pasid);
144 }
145
146 /*
147  * VMID manager
148  *
149  * VMIDs are a per VMHUB identifier for page tables handling.
150  */
151
152 /**
153  * amdgpu_vmid_had_gpu_reset - check if reset occured since last use
154  *
155  * @adev: amdgpu_device pointer
156  * @id: VMID structure
157  *
158  * Check if GPU reset occured since last use of the VMID.
159  */
160 bool amdgpu_vmid_had_gpu_reset(struct amdgpu_device *adev,
161                                struct amdgpu_vmid *id)
162 {
163         return id->current_gpu_reset_count !=
164                 atomic_read(&adev->gpu_reset_counter);
165 }
166
167 /* Check if we need to switch to another set of resources */
168 static bool amdgpu_vmid_gds_switch_needed(struct amdgpu_vmid *id,
169                                           struct amdgpu_job *job)
170 {
171         return id->gds_base != job->gds_base ||
172                 id->gds_size != job->gds_size ||
173                 id->gws_base != job->gws_base ||
174                 id->gws_size != job->gws_size ||
175                 id->oa_base != job->oa_base ||
176                 id->oa_size != job->oa_size;
177 }
178
179 /* Check if the id is compatible with the job */
180 static bool amdgpu_vmid_compatible(struct amdgpu_vmid *id,
181                                    struct amdgpu_job *job)
182 {
183         return  id->pd_gpu_addr == job->vm_pd_addr &&
184                 !amdgpu_vmid_gds_switch_needed(id, job);
185 }
186
187 /**
188  * amdgpu_vmid_grab_idle - grab idle VMID
189  *
190  * @ring: ring we want to submit job to
191  * @idle: resulting idle VMID
192  * @fence: fence to wait for if no id could be grabbed
193  *
194  * Try to find an idle VMID, if none is idle add a fence to wait to the sync
195  * object. Returns -ENOMEM when we are out of memory.
196  */
197 static int amdgpu_vmid_grab_idle(struct amdgpu_ring *ring,
198                                  struct amdgpu_vmid **idle,
199                                  struct dma_fence **fence)
200 {
201         struct amdgpu_device *adev = ring->adev;
202         unsigned vmhub = ring->vm_hub;
203         struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
204         struct dma_fence **fences;
205         unsigned i;
206
207         if (!dma_fence_is_signaled(ring->vmid_wait)) {
208                 *fence = dma_fence_get(ring->vmid_wait);
209                 return 0;
210         }
211
212         fences = kmalloc_array(id_mgr->num_ids, sizeof(void *), GFP_KERNEL);
213         if (!fences)
214                 return -ENOMEM;
215
216         /* Check if we have an idle VMID */
217         i = 0;
218         list_for_each_entry((*idle), &id_mgr->ids_lru, list) {
219                 /* Don't use per engine and per process VMID at the same time */
220                 struct amdgpu_ring *r = adev->vm_manager.concurrent_flush ?
221                         NULL : ring;
222
223                 fences[i] = amdgpu_sync_peek_fence(&(*idle)->active, r);
224                 if (!fences[i])
225                         break;
226                 ++i;
227         }
228
229         /* If we can't find a idle VMID to use, wait till one becomes available */
230         if (&(*idle)->list == &id_mgr->ids_lru) {
231                 u64 fence_context = adev->vm_manager.fence_context + ring->idx;
232                 unsigned seqno = ++adev->vm_manager.seqno[ring->idx];
233                 struct dma_fence_array *array;
234                 unsigned j;
235
236                 *idle = NULL;
237                 for (j = 0; j < i; ++j)
238                         dma_fence_get(fences[j]);
239
240                 array = dma_fence_array_create(i, fences, fence_context,
241                                                seqno, true);
242                 if (!array) {
243                         for (j = 0; j < i; ++j)
244                                 dma_fence_put(fences[j]);
245                         kfree(fences);
246                         return -ENOMEM;
247                 }
248
249                 *fence = dma_fence_get(&array->base);
250                 dma_fence_put(ring->vmid_wait);
251                 ring->vmid_wait = &array->base;
252                 return 0;
253         }
254         kfree(fences);
255
256         return 0;
257 }
258
259 /**
260  * amdgpu_vmid_grab_reserved - try to assign reserved VMID
261  *
262  * @vm: vm to allocate id for
263  * @ring: ring we want to submit job to
264  * @job: job who wants to use the VMID
265  * @id: resulting VMID
266  * @fence: fence to wait for if no id could be grabbed
267  *
268  * Try to assign a reserved VMID.
269  */
270 static int amdgpu_vmid_grab_reserved(struct amdgpu_vm *vm,
271                                      struct amdgpu_ring *ring,
272                                      struct amdgpu_job *job,
273                                      struct amdgpu_vmid **id,
274                                      struct dma_fence **fence)
275 {
276         struct amdgpu_device *adev = ring->adev;
277         unsigned vmhub = ring->vm_hub;
278         struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
279         uint64_t fence_context = adev->fence_context + ring->idx;
280         bool needs_flush = vm->use_cpu_for_update;
281         uint64_t updates = amdgpu_vm_tlb_seq(vm);
282         int r;
283
284         *id = id_mgr->reserved;
285         if ((*id)->owner != vm->immediate.fence_context ||
286             !amdgpu_vmid_compatible(*id, job) ||
287             (*id)->flushed_updates < updates ||
288             !(*id)->last_flush ||
289             ((*id)->last_flush->context != fence_context &&
290              !dma_fence_is_signaled((*id)->last_flush))) {
291                 struct dma_fence *tmp;
292
293                 /* Wait for the gang to be assembled before using a
294                  * reserved VMID or otherwise the gang could deadlock.
295                  */
296                 tmp = amdgpu_device_get_gang(adev);
297                 if (!dma_fence_is_signaled(tmp) && tmp != job->gang_submit) {
298                         *id = NULL;
299                         *fence = tmp;
300                         return 0;
301                 }
302                 dma_fence_put(tmp);
303
304                 /* Make sure the id is owned by the gang before proceeding */
305                 if (!job->gang_submit ||
306                     (*id)->owner != vm->immediate.fence_context) {
307
308                         /* Don't use per engine and per process VMID at the
309                          * same time
310                          */
311                         if (adev->vm_manager.concurrent_flush)
312                                 ring = NULL;
313
314                         /* to prevent one context starved by another context */
315                         (*id)->pd_gpu_addr = 0;
316                         tmp = amdgpu_sync_peek_fence(&(*id)->active, ring);
317                         if (tmp) {
318                                 *id = NULL;
319                                 *fence = dma_fence_get(tmp);
320                                 return 0;
321                         }
322                 }
323                 needs_flush = true;
324         }
325
326         /* Good we can use this VMID. Remember this submission as
327         * user of the VMID.
328         */
329         r = amdgpu_sync_fence(&(*id)->active, &job->base.s_fence->finished);
330         if (r)
331                 return r;
332
333         job->vm_needs_flush = needs_flush;
334         job->spm_update_needed = true;
335         return 0;
336 }
337
338 /**
339  * amdgpu_vmid_grab_used - try to reuse a VMID
340  *
341  * @vm: vm to allocate id for
342  * @ring: ring we want to submit job to
343  * @job: job who wants to use the VMID
344  * @id: resulting VMID
345  *
346  * Try to reuse a VMID for this submission.
347  */
348 static int amdgpu_vmid_grab_used(struct amdgpu_vm *vm,
349                                  struct amdgpu_ring *ring,
350                                  struct amdgpu_job *job,
351                                  struct amdgpu_vmid **id)
352 {
353         struct amdgpu_device *adev = ring->adev;
354         unsigned vmhub = ring->vm_hub;
355         struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
356         uint64_t fence_context = adev->fence_context + ring->idx;
357         uint64_t updates = amdgpu_vm_tlb_seq(vm);
358         int r;
359
360         job->vm_needs_flush = vm->use_cpu_for_update;
361
362         /* Check if we can use a VMID already assigned to this VM */
363         list_for_each_entry_reverse((*id), &id_mgr->ids_lru, list) {
364                 bool needs_flush = vm->use_cpu_for_update;
365
366                 /* Check all the prerequisites to using this VMID */
367                 if ((*id)->owner != vm->immediate.fence_context)
368                         continue;
369
370                 if (!amdgpu_vmid_compatible(*id, job))
371                         continue;
372
373                 if (!(*id)->last_flush ||
374                     ((*id)->last_flush->context != fence_context &&
375                      !dma_fence_is_signaled((*id)->last_flush)))
376                         needs_flush = true;
377
378                 if ((*id)->flushed_updates < updates)
379                         needs_flush = true;
380
381                 if (needs_flush && !adev->vm_manager.concurrent_flush)
382                         continue;
383
384                 /* Good, we can use this VMID. Remember this submission as
385                  * user of the VMID.
386                  */
387                 r = amdgpu_sync_fence(&(*id)->active,
388                                       &job->base.s_fence->finished);
389                 if (r)
390                         return r;
391
392                 job->vm_needs_flush |= needs_flush;
393                 return 0;
394         }
395
396         *id = NULL;
397         return 0;
398 }
399
400 /**
401  * amdgpu_vmid_grab - allocate the next free VMID
402  *
403  * @vm: vm to allocate id for
404  * @ring: ring we want to submit job to
405  * @job: job who wants to use the VMID
406  * @fence: fence to wait for if no id could be grabbed
407  *
408  * Allocate an id for the vm, adding fences to the sync obj as necessary.
409  */
410 int amdgpu_vmid_grab(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
411                      struct amdgpu_job *job, struct dma_fence **fence)
412 {
413         struct amdgpu_device *adev = ring->adev;
414         unsigned vmhub = ring->vm_hub;
415         struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
416         struct amdgpu_vmid *idle = NULL;
417         struct amdgpu_vmid *id = NULL;
418         int r = 0;
419
420         mutex_lock(&id_mgr->lock);
421         r = amdgpu_vmid_grab_idle(ring, &idle, fence);
422         if (r || !idle)
423                 goto error;
424
425         if (amdgpu_vmid_uses_reserved(adev, vm, vmhub)) {
426                 r = amdgpu_vmid_grab_reserved(vm, ring, job, &id, fence);
427                 if (r || !id)
428                         goto error;
429         } else {
430                 r = amdgpu_vmid_grab_used(vm, ring, job, &id);
431                 if (r)
432                         goto error;
433
434                 if (!id) {
435                         /* Still no ID to use? Then use the idle one found earlier */
436                         id = idle;
437
438                         /* Remember this submission as user of the VMID */
439                         r = amdgpu_sync_fence(&id->active,
440                                               &job->base.s_fence->finished);
441                         if (r)
442                                 goto error;
443
444                         job->vm_needs_flush = true;
445                 }
446
447                 list_move_tail(&id->list, &id_mgr->ids_lru);
448         }
449
450         job->gds_switch_needed = amdgpu_vmid_gds_switch_needed(id, job);
451         if (job->vm_needs_flush) {
452                 id->flushed_updates = amdgpu_vm_tlb_seq(vm);
453                 dma_fence_put(id->last_flush);
454                 id->last_flush = NULL;
455         }
456         job->vmid = id - id_mgr->ids;
457         job->pasid = vm->pasid;
458
459         id->gds_base = job->gds_base;
460         id->gds_size = job->gds_size;
461         id->gws_base = job->gws_base;
462         id->gws_size = job->gws_size;
463         id->oa_base = job->oa_base;
464         id->oa_size = job->oa_size;
465         id->pd_gpu_addr = job->vm_pd_addr;
466         id->owner = vm->immediate.fence_context;
467
468         trace_amdgpu_vm_grab_id(vm, ring, job);
469
470 error:
471         mutex_unlock(&id_mgr->lock);
472         return r;
473 }
474
475 /*
476  * amdgpu_vmid_uses_reserved - check if a VM will use a reserved VMID
477  * @adev: amdgpu_device pointer
478  * @vm: the VM to check
479  * @vmhub: the VMHUB which will be used
480  *
481  * Returns: True if the VM will use a reserved VMID.
482  */
483 bool amdgpu_vmid_uses_reserved(struct amdgpu_device *adev,
484                                struct amdgpu_vm *vm, unsigned int vmhub)
485 {
486         return vm->reserved_vmid[vmhub] ||
487                 (adev->enforce_isolation[(vm->root.bo->xcp_id != AMDGPU_XCP_NO_PARTITION) ?
488                                          vm->root.bo->xcp_id : 0] &&
489                  AMDGPU_IS_GFXHUB(vmhub));
490 }
491
492 int amdgpu_vmid_alloc_reserved(struct amdgpu_device *adev,
493                                unsigned vmhub)
494 {
495         struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
496
497         mutex_lock(&id_mgr->lock);
498
499         ++id_mgr->reserved_use_count;
500         if (!id_mgr->reserved) {
501                 struct amdgpu_vmid *id;
502
503                 id = list_first_entry(&id_mgr->ids_lru, struct amdgpu_vmid,
504                                       list);
505                 /* Remove from normal round robin handling */
506                 list_del_init(&id->list);
507                 id_mgr->reserved = id;
508         }
509
510         mutex_unlock(&id_mgr->lock);
511         return 0;
512 }
513
514 void amdgpu_vmid_free_reserved(struct amdgpu_device *adev,
515                                unsigned vmhub)
516 {
517         struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
518
519         mutex_lock(&id_mgr->lock);
520         if (!--id_mgr->reserved_use_count) {
521                 /* give the reserved ID back to normal round robin */
522                 list_add(&id_mgr->reserved->list, &id_mgr->ids_lru);
523                 id_mgr->reserved = NULL;
524         }
525
526         mutex_unlock(&id_mgr->lock);
527 }
528
529 /**
530  * amdgpu_vmid_reset - reset VMID to zero
531  *
532  * @adev: amdgpu device structure
533  * @vmhub: vmhub type
534  * @vmid: vmid number to use
535  *
536  * Reset saved GDW, GWS and OA to force switch on next flush.
537  */
538 void amdgpu_vmid_reset(struct amdgpu_device *adev, unsigned vmhub,
539                        unsigned vmid)
540 {
541         struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
542         struct amdgpu_vmid *id = &id_mgr->ids[vmid];
543
544         mutex_lock(&id_mgr->lock);
545         id->owner = 0;
546         id->gds_base = 0;
547         id->gds_size = 0;
548         id->gws_base = 0;
549         id->gws_size = 0;
550         id->oa_base = 0;
551         id->oa_size = 0;
552         mutex_unlock(&id_mgr->lock);
553 }
554
555 /**
556  * amdgpu_vmid_reset_all - reset VMID to zero
557  *
558  * @adev: amdgpu device structure
559  *
560  * Reset VMID to force flush on next use
561  */
562 void amdgpu_vmid_reset_all(struct amdgpu_device *adev)
563 {
564         unsigned i, j;
565
566         for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
567                 struct amdgpu_vmid_mgr *id_mgr =
568                         &adev->vm_manager.id_mgr[i];
569
570                 for (j = 1; j < id_mgr->num_ids; ++j)
571                         amdgpu_vmid_reset(adev, i, j);
572         }
573 }
574
575 /**
576  * amdgpu_vmid_mgr_init - init the VMID manager
577  *
578  * @adev: amdgpu_device pointer
579  *
580  * Initialize the VM manager structures
581  */
582 void amdgpu_vmid_mgr_init(struct amdgpu_device *adev)
583 {
584         unsigned i, j;
585
586         for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
587                 struct amdgpu_vmid_mgr *id_mgr =
588                         &adev->vm_manager.id_mgr[i];
589
590                 mutex_init(&id_mgr->lock);
591                 INIT_LIST_HEAD(&id_mgr->ids_lru);
592                 id_mgr->reserved_use_count = 0;
593
594                 /* manage only VMIDs not used by KFD */
595                 id_mgr->num_ids = adev->vm_manager.first_kfd_vmid;
596
597                 /* skip over VMID 0, since it is the system VM */
598                 for (j = 1; j < id_mgr->num_ids; ++j) {
599                         amdgpu_vmid_reset(adev, i, j);
600                         amdgpu_sync_create(&id_mgr->ids[j].active);
601                         list_add_tail(&id_mgr->ids[j].list, &id_mgr->ids_lru);
602                 }
603         }
604         /* alloc a default reserved vmid to enforce isolation */
605         for (i = 0; i < (adev->xcp_mgr ? adev->xcp_mgr->num_xcps : 1); i++) {
606                 if (adev->enforce_isolation[i])
607                         amdgpu_vmid_alloc_reserved(adev, AMDGPU_GFXHUB(i));
608         }
609 }
610
611 /**
612  * amdgpu_vmid_mgr_fini - cleanup VM manager
613  *
614  * @adev: amdgpu_device pointer
615  *
616  * Cleanup the VM manager and free resources.
617  */
618 void amdgpu_vmid_mgr_fini(struct amdgpu_device *adev)
619 {
620         unsigned i, j;
621
622         for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
623                 struct amdgpu_vmid_mgr *id_mgr =
624                         &adev->vm_manager.id_mgr[i];
625
626                 mutex_destroy(&id_mgr->lock);
627                 for (j = 0; j < AMDGPU_NUM_VMID; ++j) {
628                         struct amdgpu_vmid *id = &id_mgr->ids[j];
629
630                         amdgpu_sync_free(&id->active);
631                         dma_fence_put(id->last_flush);
632                         dma_fence_put(id->pasid_mapping);
633                 }
634         }
635 }
This page took 0.062118 seconds and 4 git commands to generate.