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

source code of linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c