1 /*
2 * SPDX-License-Identifier: MIT
3 *
4 * Copyright © 2011-2012 Intel Corporation
5 */
6
7 /*
8 * This file implements HW context support. On gen5+ a HW context consists of an
9 * opaque GPU object which is referenced at times of context saves and restores.
10 * With RC6 enabled, the context is also referenced as the GPU enters and exists
11 * from RC6 (GPU has it's own internal power context, except on gen5). Though
12 * something like a context does exist for the media ring, the code only
13 * supports contexts for the render ring.
14 *
15 * In software, there is a distinction between contexts created by the user,
16 * and the default HW context. The default HW context is used by GPU clients
17 * that do not request setup of their own hardware context. The default
18 * context's state is never restored to help prevent programming errors. This
19 * would happen if a client ran and piggy-backed off another clients GPU state.
20 * The default context only exists to give the GPU some offset to load as the
21 * current to invoke a save of the context we actually care about. In fact, the
22 * code could likely be constructed, albeit in a more complicated fashion, to
23 * never use the default context, though that limits the driver's ability to
24 * swap out, and/or destroy other contexts.
25 *
26 * All other contexts are created as a request by the GPU client. These contexts
27 * store GPU state, and thus allow GPU clients to not re-emit state (and
28 * potentially query certain state) at any time. The kernel driver makes
29 * certain that the appropriate commands are inserted.
30 *
31 * The context life cycle is semi-complicated in that context BOs may live
32 * longer than the context itself because of the way the hardware, and object
33 * tracking works. Below is a very crude representation of the state machine
34 * describing the context life.
35 * refcount pincount active
36 * S0: initial state 0 0 0
37 * S1: context created 1 0 0
38 * S2: context is currently running 2 1 X
39 * S3: GPU referenced, but not current 2 0 1
40 * S4: context is current, but destroyed 1 1 0
41 * S5: like S3, but destroyed 1 0 1
42 *
43 * The most common (but not all) transitions:
44 * S0->S1: client creates a context
45 * S1->S2: client submits execbuf with context
46 * S2->S3: other clients submits execbuf with context
47 * S3->S1: context object was retired
48 * S3->S2: clients submits another execbuf
49 * S2->S4: context destroy called with current context
50 * S3->S5->S0: destroy path
51 * S4->S5->S0: destroy path on current context
52 *
53 * There are two confusing terms used above:
54 * The "current context" means the context which is currently running on the
55 * GPU. The GPU has loaded its state already and has stored away the gtt
56 * offset of the BO. The GPU is not actively referencing the data at this
57 * offset, but it will on the next context switch. The only way to avoid this
58 * is to do a GPU reset.
59 *
60 * An "active context' is one which was previously the "current context" and is
61 * on the active list waiting for the next context switch to occur. Until this
62 * happens, the object must remain at the same gtt offset. It is therefore
63 * possible to destroy a context, but it is still active.
64 *
65 */
66
67 #include <linux/highmem.h>
68 #include <linux/log2.h>
69 #include <linux/nospec.h>
70
71 #include <drm/drm_cache.h>
72 #include <drm/drm_syncobj.h>
73
74 #include "gt/gen6_ppgtt.h"
75 #include "gt/intel_context.h"
76 #include "gt/intel_context_param.h"
77 #include "gt/intel_engine_heartbeat.h"
78 #include "gt/intel_engine_user.h"
79 #include "gt/intel_gpu_commands.h"
80 #include "gt/intel_ring.h"
81
82 #include "pxp/intel_pxp.h"
83
84 #include "i915_file_private.h"
85 #include "i915_gem_context.h"
86 #include "i915_trace.h"
87 #include "i915_user_extensions.h"
88
89 #define ALL_L3_SLICES(dev) (1 << NUM_L3_SLICES(dev)) - 1
90
91 static struct kmem_cache *slab_luts;
92
i915_lut_handle_alloc(void)93 struct i915_lut_handle *i915_lut_handle_alloc(void)
94 {
95 return kmem_cache_alloc(slab_luts, GFP_KERNEL);
96 }
97
i915_lut_handle_free(struct i915_lut_handle * lut)98 void i915_lut_handle_free(struct i915_lut_handle *lut)
99 {
100 return kmem_cache_free(slab_luts, lut);
101 }
102
lut_close(struct i915_gem_context * ctx)103 static void lut_close(struct i915_gem_context *ctx)
104 {
105 struct radix_tree_iter iter;
106 void __rcu **slot;
107
108 mutex_lock(&ctx->lut_mutex);
109 rcu_read_lock();
110 radix_tree_for_each_slot(slot, &ctx->handles_vma, &iter, 0) {
111 struct i915_vma *vma = rcu_dereference_raw(*slot);
112 struct drm_i915_gem_object *obj = vma->obj;
113 struct i915_lut_handle *lut;
114
115 if (!kref_get_unless_zero(&obj->base.refcount))
116 continue;
117
118 spin_lock(&obj->lut_lock);
119 list_for_each_entry(lut, &obj->lut_list, obj_link) {
120 if (lut->ctx != ctx)
121 continue;
122
123 if (lut->handle != iter.index)
124 continue;
125
126 list_del(&lut->obj_link);
127 break;
128 }
129 spin_unlock(&obj->lut_lock);
130
131 if (&lut->obj_link != &obj->lut_list) {
132 i915_lut_handle_free(lut);
133 radix_tree_iter_delete(&ctx->handles_vma, &iter, slot);
134 i915_vma_close(vma);
135 i915_gem_object_put(obj);
136 }
137
138 i915_gem_object_put(obj);
139 }
140 rcu_read_unlock();
141 mutex_unlock(&ctx->lut_mutex);
142 }
143
144 static struct intel_context *
lookup_user_engine(struct i915_gem_context * ctx,unsigned long flags,const struct i915_engine_class_instance * ci)145 lookup_user_engine(struct i915_gem_context *ctx,
146 unsigned long flags,
147 const struct i915_engine_class_instance *ci)
148 #define LOOKUP_USER_INDEX BIT(0)
149 {
150 int idx;
151
152 if (!!(flags & LOOKUP_USER_INDEX) != i915_gem_context_user_engines(ctx))
153 return ERR_PTR(-EINVAL);
154
155 if (!i915_gem_context_user_engines(ctx)) {
156 struct intel_engine_cs *engine;
157
158 engine = intel_engine_lookup_user(ctx->i915,
159 ci->engine_class,
160 ci->engine_instance);
161 if (!engine)
162 return ERR_PTR(-EINVAL);
163
164 idx = engine->legacy_idx;
165 } else {
166 idx = ci->engine_instance;
167 }
168
169 return i915_gem_context_get_engine(ctx, idx);
170 }
171
validate_priority(struct drm_i915_private * i915,const struct drm_i915_gem_context_param * args)172 static int validate_priority(struct drm_i915_private *i915,
173 const struct drm_i915_gem_context_param *args)
174 {
175 s64 priority = args->value;
176
177 if (args->size)
178 return -EINVAL;
179
180 if (!(i915->caps.scheduler & I915_SCHEDULER_CAP_PRIORITY))
181 return -ENODEV;
182
183 if (priority > I915_CONTEXT_MAX_USER_PRIORITY ||
184 priority < I915_CONTEXT_MIN_USER_PRIORITY)
185 return -EINVAL;
186
187 if (priority > I915_CONTEXT_DEFAULT_PRIORITY &&
188 !capable(CAP_SYS_NICE))
189 return -EPERM;
190
191 return 0;
192 }
193
proto_context_close(struct drm_i915_private * i915,struct i915_gem_proto_context * pc)194 static void proto_context_close(struct drm_i915_private *i915,
195 struct i915_gem_proto_context *pc)
196 {
197 int i;
198
199 if (pc->pxp_wakeref)
200 intel_runtime_pm_put(&i915->runtime_pm, pc->pxp_wakeref);
201 if (pc->vm)
202 i915_vm_put(pc->vm);
203 if (pc->user_engines) {
204 for (i = 0; i < pc->num_user_engines; i++)
205 kfree(pc->user_engines[i].siblings);
206 kfree(pc->user_engines);
207 }
208 kfree(pc);
209 }
210
proto_context_set_persistence(struct drm_i915_private * i915,struct i915_gem_proto_context * pc,bool persist)211 static int proto_context_set_persistence(struct drm_i915_private *i915,
212 struct i915_gem_proto_context *pc,
213 bool persist)
214 {
215 if (persist) {
216 /*
217 * Only contexts that are short-lived [that will expire or be
218 * reset] are allowed to survive past termination. We require
219 * hangcheck to ensure that the persistent requests are healthy.
220 */
221 if (!i915->params.enable_hangcheck)
222 return -EINVAL;
223
224 pc->user_flags |= BIT(UCONTEXT_PERSISTENCE);
225 } else {
226 /* To cancel a context we use "preempt-to-idle" */
227 if (!(i915->caps.scheduler & I915_SCHEDULER_CAP_PREEMPTION))
228 return -ENODEV;
229
230 /*
231 * If the cancel fails, we then need to reset, cleanly!
232 *
233 * If the per-engine reset fails, all hope is lost! We resort
234 * to a full GPU reset in that unlikely case, but realistically
235 * if the engine could not reset, the full reset does not fare
236 * much better. The damage has been done.
237 *
238 * However, if we cannot reset an engine by itself, we cannot
239 * cleanup a hanging persistent context without causing
240 * colateral damage, and we should not pretend we can by
241 * exposing the interface.
242 */
243 if (!intel_has_reset_engine(to_gt(i915)))
244 return -ENODEV;
245
246 pc->user_flags &= ~BIT(UCONTEXT_PERSISTENCE);
247 }
248
249 return 0;
250 }
251
proto_context_set_protected(struct drm_i915_private * i915,struct i915_gem_proto_context * pc,bool protected)252 static int proto_context_set_protected(struct drm_i915_private *i915,
253 struct i915_gem_proto_context *pc,
254 bool protected)
255 {
256 int ret = 0;
257
258 if (!protected) {
259 pc->uses_protected_content = false;
260 } else if (!intel_pxp_is_enabled(&to_gt(i915)->pxp)) {
261 ret = -ENODEV;
262 } else if ((pc->user_flags & BIT(UCONTEXT_RECOVERABLE)) ||
263 !(pc->user_flags & BIT(UCONTEXT_BANNABLE))) {
264 ret = -EPERM;
265 } else {
266 pc->uses_protected_content = true;
267
268 /*
269 * protected context usage requires the PXP session to be up,
270 * which in turn requires the device to be active.
271 */
272 pc->pxp_wakeref = intel_runtime_pm_get(&i915->runtime_pm);
273
274 if (!intel_pxp_is_active(&to_gt(i915)->pxp))
275 ret = intel_pxp_start(&to_gt(i915)->pxp);
276 }
277
278 return ret;
279 }
280
281 static struct i915_gem_proto_context *
proto_context_create(struct drm_i915_private * i915,unsigned int flags)282 proto_context_create(struct drm_i915_private *i915, unsigned int flags)
283 {
284 struct i915_gem_proto_context *pc, *err;
285
286 pc = kzalloc(sizeof(*pc), GFP_KERNEL);
287 if (!pc)
288 return ERR_PTR(-ENOMEM);
289
290 pc->num_user_engines = -1;
291 pc->user_engines = NULL;
292 pc->user_flags = BIT(UCONTEXT_BANNABLE) |
293 BIT(UCONTEXT_RECOVERABLE);
294 if (i915->params.enable_hangcheck)
295 pc->user_flags |= BIT(UCONTEXT_PERSISTENCE);
296 pc->sched.priority = I915_PRIORITY_NORMAL;
297
298 if (flags & I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE) {
299 if (!HAS_EXECLISTS(i915)) {
300 err = ERR_PTR(-EINVAL);
301 goto proto_close;
302 }
303 pc->single_timeline = true;
304 }
305
306 return pc;
307
308 proto_close:
309 proto_context_close(i915, pc);
310 return err;
311 }
312
proto_context_register_locked(struct drm_i915_file_private * fpriv,struct i915_gem_proto_context * pc,u32 * id)313 static int proto_context_register_locked(struct drm_i915_file_private *fpriv,
314 struct i915_gem_proto_context *pc,
315 u32 *id)
316 {
317 int ret;
318 void *old;
319
320 lockdep_assert_held(&fpriv->proto_context_lock);
321
322 ret = xa_alloc(&fpriv->context_xa, id, NULL, xa_limit_32b, GFP_KERNEL);
323 if (ret)
324 return ret;
325
326 old = xa_store(&fpriv->proto_context_xa, *id, pc, GFP_KERNEL);
327 if (xa_is_err(old)) {
328 xa_erase(&fpriv->context_xa, *id);
329 return xa_err(old);
330 }
331 WARN_ON(old);
332
333 return 0;
334 }
335
proto_context_register(struct drm_i915_file_private * fpriv,struct i915_gem_proto_context * pc,u32 * id)336 static int proto_context_register(struct drm_i915_file_private *fpriv,
337 struct i915_gem_proto_context *pc,
338 u32 *id)
339 {
340 int ret;
341
342 mutex_lock(&fpriv->proto_context_lock);
343 ret = proto_context_register_locked(fpriv, pc, id);
344 mutex_unlock(&fpriv->proto_context_lock);
345
346 return ret;
347 }
348
349 static struct i915_address_space *
i915_gem_vm_lookup(struct drm_i915_file_private * file_priv,u32 id)350 i915_gem_vm_lookup(struct drm_i915_file_private *file_priv, u32 id)
351 {
352 struct i915_address_space *vm;
353
354 xa_lock(&file_priv->vm_xa);
355 vm = xa_load(&file_priv->vm_xa, id);
356 if (vm)
357 kref_get(&vm->ref);
358 xa_unlock(&file_priv->vm_xa);
359
360 return vm;
361 }
362
set_proto_ctx_vm(struct drm_i915_file_private * fpriv,struct i915_gem_proto_context * pc,const struct drm_i915_gem_context_param * args)363 static int set_proto_ctx_vm(struct drm_i915_file_private *fpriv,
364 struct i915_gem_proto_context *pc,
365 const struct drm_i915_gem_context_param *args)
366 {
367 struct drm_i915_private *i915 = fpriv->dev_priv;
368 struct i915_address_space *vm;
369
370 if (args->size)
371 return -EINVAL;
372
373 if (!HAS_FULL_PPGTT(i915))
374 return -ENODEV;
375
376 if (upper_32_bits(args->value))
377 return -ENOENT;
378
379 vm = i915_gem_vm_lookup(fpriv, args->value);
380 if (!vm)
381 return -ENOENT;
382
383 if (pc->vm)
384 i915_vm_put(pc->vm);
385 pc->vm = vm;
386
387 return 0;
388 }
389
390 struct set_proto_ctx_engines {
391 struct drm_i915_private *i915;
392 unsigned num_engines;
393 struct i915_gem_proto_engine *engines;
394 };
395
396 static int
set_proto_ctx_engines_balance(struct i915_user_extension __user * base,void * data)397 set_proto_ctx_engines_balance(struct i915_user_extension __user *base,
398 void *data)
399 {
400 struct i915_context_engines_load_balance __user *ext =
401 container_of_user(base, typeof(*ext), base);
402 const struct set_proto_ctx_engines *set = data;
403 struct drm_i915_private *i915 = set->i915;
404 struct intel_engine_cs **siblings;
405 u16 num_siblings, idx;
406 unsigned int n;
407 int err;
408
409 if (!HAS_EXECLISTS(i915))
410 return -ENODEV;
411
412 if (get_user(idx, &ext->engine_index))
413 return -EFAULT;
414
415 if (idx >= set->num_engines) {
416 drm_dbg(&i915->drm, "Invalid placement value, %d >= %d\n",
417 idx, set->num_engines);
418 return -EINVAL;
419 }
420
421 idx = array_index_nospec(idx, set->num_engines);
422 if (set->engines[idx].type != I915_GEM_ENGINE_TYPE_INVALID) {
423 drm_dbg(&i915->drm,
424 "Invalid placement[%d], already occupied\n", idx);
425 return -EEXIST;
426 }
427
428 if (get_user(num_siblings, &ext->num_siblings))
429 return -EFAULT;
430
431 err = check_user_mbz(&ext->flags);
432 if (err)
433 return err;
434
435 err = check_user_mbz(&ext->mbz64);
436 if (err)
437 return err;
438
439 if (num_siblings == 0)
440 return 0;
441
442 siblings = kmalloc_array(num_siblings, sizeof(*siblings), GFP_KERNEL);
443 if (!siblings)
444 return -ENOMEM;
445
446 for (n = 0; n < num_siblings; n++) {
447 struct i915_engine_class_instance ci;
448
449 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) {
450 err = -EFAULT;
451 goto err_siblings;
452 }
453
454 siblings[n] = intel_engine_lookup_user(i915,
455 ci.engine_class,
456 ci.engine_instance);
457 if (!siblings[n]) {
458 drm_dbg(&i915->drm,
459 "Invalid sibling[%d]: { class:%d, inst:%d }\n",
460 n, ci.engine_class, ci.engine_instance);
461 err = -EINVAL;
462 goto err_siblings;
463 }
464 }
465
466 if (num_siblings == 1) {
467 set->engines[idx].type = I915_GEM_ENGINE_TYPE_PHYSICAL;
468 set->engines[idx].engine = siblings[0];
469 kfree(siblings);
470 } else {
471 set->engines[idx].type = I915_GEM_ENGINE_TYPE_BALANCED;
472 set->engines[idx].num_siblings = num_siblings;
473 set->engines[idx].siblings = siblings;
474 }
475
476 return 0;
477
478 err_siblings:
479 kfree(siblings);
480
481 return err;
482 }
483
484 static int
set_proto_ctx_engines_bond(struct i915_user_extension __user * base,void * data)485 set_proto_ctx_engines_bond(struct i915_user_extension __user *base, void *data)
486 {
487 struct i915_context_engines_bond __user *ext =
488 container_of_user(base, typeof(*ext), base);
489 const struct set_proto_ctx_engines *set = data;
490 struct drm_i915_private *i915 = set->i915;
491 struct i915_engine_class_instance ci;
492 struct intel_engine_cs *master;
493 u16 idx, num_bonds;
494 int err, n;
495
496 if (GRAPHICS_VER(i915) >= 12 && !IS_TIGERLAKE(i915) &&
497 !IS_ROCKETLAKE(i915) && !IS_ALDERLAKE_S(i915)) {
498 drm_dbg(&i915->drm,
499 "Bonding not supported on this platform\n");
500 return -ENODEV;
501 }
502
503 if (get_user(idx, &ext->virtual_index))
504 return -EFAULT;
505
506 if (idx >= set->num_engines) {
507 drm_dbg(&i915->drm,
508 "Invalid index for virtual engine: %d >= %d\n",
509 idx, set->num_engines);
510 return -EINVAL;
511 }
512
513 idx = array_index_nospec(idx, set->num_engines);
514 if (set->engines[idx].type == I915_GEM_ENGINE_TYPE_INVALID) {
515 drm_dbg(&i915->drm, "Invalid engine at %d\n", idx);
516 return -EINVAL;
517 }
518
519 if (set->engines[idx].type != I915_GEM_ENGINE_TYPE_PHYSICAL) {
520 drm_dbg(&i915->drm,
521 "Bonding with virtual engines not allowed\n");
522 return -EINVAL;
523 }
524
525 err = check_user_mbz(&ext->flags);
526 if (err)
527 return err;
528
529 for (n = 0; n < ARRAY_SIZE(ext->mbz64); n++) {
530 err = check_user_mbz(&ext->mbz64[n]);
531 if (err)
532 return err;
533 }
534
535 if (copy_from_user(&ci, &ext->master, sizeof(ci)))
536 return -EFAULT;
537
538 master = intel_engine_lookup_user(i915,
539 ci.engine_class,
540 ci.engine_instance);
541 if (!master) {
542 drm_dbg(&i915->drm,
543 "Unrecognised master engine: { class:%u, instance:%u }\n",
544 ci.engine_class, ci.engine_instance);
545 return -EINVAL;
546 }
547
548 if (intel_engine_uses_guc(master)) {
549 DRM_DEBUG("bonding extension not supported with GuC submission");
550 return -ENODEV;
551 }
552
553 if (get_user(num_bonds, &ext->num_bonds))
554 return -EFAULT;
555
556 for (n = 0; n < num_bonds; n++) {
557 struct intel_engine_cs *bond;
558
559 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci)))
560 return -EFAULT;
561
562 bond = intel_engine_lookup_user(i915,
563 ci.engine_class,
564 ci.engine_instance);
565 if (!bond) {
566 drm_dbg(&i915->drm,
567 "Unrecognised engine[%d] for bonding: { class:%d, instance: %d }\n",
568 n, ci.engine_class, ci.engine_instance);
569 return -EINVAL;
570 }
571 }
572
573 return 0;
574 }
575
576 static int
set_proto_ctx_engines_parallel_submit(struct i915_user_extension __user * base,void * data)577 set_proto_ctx_engines_parallel_submit(struct i915_user_extension __user *base,
578 void *data)
579 {
580 struct i915_context_engines_parallel_submit __user *ext =
581 container_of_user(base, typeof(*ext), base);
582 const struct set_proto_ctx_engines *set = data;
583 struct drm_i915_private *i915 = set->i915;
584 struct i915_engine_class_instance prev_engine;
585 u64 flags;
586 int err = 0, n, i, j;
587 u16 slot, width, num_siblings;
588 struct intel_engine_cs **siblings = NULL;
589 intel_engine_mask_t prev_mask;
590
591 if (get_user(slot, &ext->engine_index))
592 return -EFAULT;
593
594 if (get_user(width, &ext->width))
595 return -EFAULT;
596
597 if (get_user(num_siblings, &ext->num_siblings))
598 return -EFAULT;
599
600 if (!intel_uc_uses_guc_submission(&to_gt(i915)->uc) &&
601 num_siblings != 1) {
602 drm_dbg(&i915->drm, "Only 1 sibling (%d) supported in non-GuC mode\n",
603 num_siblings);
604 return -EINVAL;
605 }
606
607 if (slot >= set->num_engines) {
608 drm_dbg(&i915->drm, "Invalid placement value, %d >= %d\n",
609 slot, set->num_engines);
610 return -EINVAL;
611 }
612
613 if (set->engines[slot].type != I915_GEM_ENGINE_TYPE_INVALID) {
614 drm_dbg(&i915->drm,
615 "Invalid placement[%d], already occupied\n", slot);
616 return -EINVAL;
617 }
618
619 if (get_user(flags, &ext->flags))
620 return -EFAULT;
621
622 if (flags) {
623 drm_dbg(&i915->drm, "Unknown flags 0x%02llx", flags);
624 return -EINVAL;
625 }
626
627 for (n = 0; n < ARRAY_SIZE(ext->mbz64); n++) {
628 err = check_user_mbz(&ext->mbz64[n]);
629 if (err)
630 return err;
631 }
632
633 if (width < 2) {
634 drm_dbg(&i915->drm, "Width (%d) < 2\n", width);
635 return -EINVAL;
636 }
637
638 if (num_siblings < 1) {
639 drm_dbg(&i915->drm, "Number siblings (%d) < 1\n",
640 num_siblings);
641 return -EINVAL;
642 }
643
644 siblings = kmalloc_array(num_siblings * width,
645 sizeof(*siblings),
646 GFP_KERNEL);
647 if (!siblings)
648 return -ENOMEM;
649
650 /* Create contexts / engines */
651 for (i = 0; i < width; ++i) {
652 intel_engine_mask_t current_mask = 0;
653
654 for (j = 0; j < num_siblings; ++j) {
655 struct i915_engine_class_instance ci;
656
657 n = i * num_siblings + j;
658 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) {
659 err = -EFAULT;
660 goto out_err;
661 }
662
663 siblings[n] =
664 intel_engine_lookup_user(i915, ci.engine_class,
665 ci.engine_instance);
666 if (!siblings[n]) {
667 drm_dbg(&i915->drm,
668 "Invalid sibling[%d]: { class:%d, inst:%d }\n",
669 n, ci.engine_class, ci.engine_instance);
670 err = -EINVAL;
671 goto out_err;
672 }
673
674 /*
675 * We don't support breadcrumb handshake on these
676 * classes
677 */
678 if (siblings[n]->class == RENDER_CLASS ||
679 siblings[n]->class == COMPUTE_CLASS) {
680 err = -EINVAL;
681 goto out_err;
682 }
683
684 if (n) {
685 if (prev_engine.engine_class !=
686 ci.engine_class) {
687 drm_dbg(&i915->drm,
688 "Mismatched class %d, %d\n",
689 prev_engine.engine_class,
690 ci.engine_class);
691 err = -EINVAL;
692 goto out_err;
693 }
694 }
695
696 prev_engine = ci;
697 current_mask |= siblings[n]->logical_mask;
698 }
699
700 if (i > 0) {
701 if (current_mask != prev_mask << 1) {
702 drm_dbg(&i915->drm,
703 "Non contiguous logical mask 0x%x, 0x%x\n",
704 prev_mask, current_mask);
705 err = -EINVAL;
706 goto out_err;
707 }
708 }
709 prev_mask = current_mask;
710 }
711
712 set->engines[slot].type = I915_GEM_ENGINE_TYPE_PARALLEL;
713 set->engines[slot].num_siblings = num_siblings;
714 set->engines[slot].width = width;
715 set->engines[slot].siblings = siblings;
716
717 return 0;
718
719 out_err:
720 kfree(siblings);
721
722 return err;
723 }
724
725 static const i915_user_extension_fn set_proto_ctx_engines_extensions[] = {
726 [I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE] = set_proto_ctx_engines_balance,
727 [I915_CONTEXT_ENGINES_EXT_BOND] = set_proto_ctx_engines_bond,
728 [I915_CONTEXT_ENGINES_EXT_PARALLEL_SUBMIT] =
729 set_proto_ctx_engines_parallel_submit,
730 };
731
set_proto_ctx_engines(struct drm_i915_file_private * fpriv,struct i915_gem_proto_context * pc,const struct drm_i915_gem_context_param * args)732 static int set_proto_ctx_engines(struct drm_i915_file_private *fpriv,
733 struct i915_gem_proto_context *pc,
734 const struct drm_i915_gem_context_param *args)
735 {
736 struct drm_i915_private *i915 = fpriv->dev_priv;
737 struct set_proto_ctx_engines set = { .i915 = i915 };
738 struct i915_context_param_engines __user *user =
739 u64_to_user_ptr(args->value);
740 unsigned int n;
741 u64 extensions;
742 int err;
743
744 if (pc->num_user_engines >= 0) {
745 drm_dbg(&i915->drm, "Cannot set engines twice");
746 return -EINVAL;
747 }
748
749 if (args->size < sizeof(*user) ||
750 !IS_ALIGNED(args->size - sizeof(*user), sizeof(*user->engines))) {
751 drm_dbg(&i915->drm, "Invalid size for engine array: %d\n",
752 args->size);
753 return -EINVAL;
754 }
755
756 set.num_engines = (args->size - sizeof(*user)) / sizeof(*user->engines);
757 /* RING_MASK has no shift so we can use it directly here */
758 if (set.num_engines > I915_EXEC_RING_MASK + 1)
759 return -EINVAL;
760
761 set.engines = kmalloc_array(set.num_engines, sizeof(*set.engines), GFP_KERNEL);
762 if (!set.engines)
763 return -ENOMEM;
764
765 for (n = 0; n < set.num_engines; n++) {
766 struct i915_engine_class_instance ci;
767 struct intel_engine_cs *engine;
768
769 if (copy_from_user(&ci, &user->engines[n], sizeof(ci))) {
770 kfree(set.engines);
771 return -EFAULT;
772 }
773
774 memset(&set.engines[n], 0, sizeof(set.engines[n]));
775
776 if (ci.engine_class == (u16)I915_ENGINE_CLASS_INVALID &&
777 ci.engine_instance == (u16)I915_ENGINE_CLASS_INVALID_NONE)
778 continue;
779
780 engine = intel_engine_lookup_user(i915,
781 ci.engine_class,
782 ci.engine_instance);
783 if (!engine) {
784 drm_dbg(&i915->drm,
785 "Invalid engine[%d]: { class:%d, instance:%d }\n",
786 n, ci.engine_class, ci.engine_instance);
787 kfree(set.engines);
788 return -ENOENT;
789 }
790
791 set.engines[n].type = I915_GEM_ENGINE_TYPE_PHYSICAL;
792 set.engines[n].engine = engine;
793 }
794
795 err = -EFAULT;
796 if (!get_user(extensions, &user->extensions))
797 err = i915_user_extensions(u64_to_user_ptr(extensions),
798 set_proto_ctx_engines_extensions,
799 ARRAY_SIZE(set_proto_ctx_engines_extensions),
800 &set);
801 if (err) {
802 kfree(set.engines);
803 return err;
804 }
805
806 pc->num_user_engines = set.num_engines;
807 pc->user_engines = set.engines;
808
809 return 0;
810 }
811
set_proto_ctx_sseu(struct drm_i915_file_private * fpriv,struct i915_gem_proto_context * pc,struct drm_i915_gem_context_param * args)812 static int set_proto_ctx_sseu(struct drm_i915_file_private *fpriv,
813 struct i915_gem_proto_context *pc,
814 struct drm_i915_gem_context_param *args)
815 {
816 struct drm_i915_private *i915 = fpriv->dev_priv;
817 struct drm_i915_gem_context_param_sseu user_sseu;
818 struct intel_sseu *sseu;
819 int ret;
820
821 if (args->size < sizeof(user_sseu))
822 return -EINVAL;
823
824 if (GRAPHICS_VER(i915) != 11)
825 return -ENODEV;
826
827 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
828 sizeof(user_sseu)))
829 return -EFAULT;
830
831 if (user_sseu.rsvd)
832 return -EINVAL;
833
834 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
835 return -EINVAL;
836
837 if (!!(user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX) != (pc->num_user_engines >= 0))
838 return -EINVAL;
839
840 if (pc->num_user_engines >= 0) {
841 int idx = user_sseu.engine.engine_instance;
842 struct i915_gem_proto_engine *pe;
843
844 if (idx >= pc->num_user_engines)
845 return -EINVAL;
846
847 pe = &pc->user_engines[idx];
848
849 /* Only render engine supports RPCS configuration. */
850 if (pe->engine->class != RENDER_CLASS)
851 return -EINVAL;
852
853 sseu = &pe->sseu;
854 } else {
855 /* Only render engine supports RPCS configuration. */
856 if (user_sseu.engine.engine_class != I915_ENGINE_CLASS_RENDER)
857 return -EINVAL;
858
859 /* There is only one render engine */
860 if (user_sseu.engine.engine_instance != 0)
861 return -EINVAL;
862
863 sseu = &pc->legacy_rcs_sseu;
864 }
865
866 ret = i915_gem_user_to_context_sseu(to_gt(i915), &user_sseu, sseu);
867 if (ret)
868 return ret;
869
870 args->size = sizeof(user_sseu);
871
872 return 0;
873 }
874
set_proto_ctx_param(struct drm_i915_file_private * fpriv,struct i915_gem_proto_context * pc,struct drm_i915_gem_context_param * args)875 static int set_proto_ctx_param(struct drm_i915_file_private *fpriv,
876 struct i915_gem_proto_context *pc,
877 struct drm_i915_gem_context_param *args)
878 {
879 int ret = 0;
880
881 switch (args->param) {
882 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
883 if (args->size)
884 ret = -EINVAL;
885 else if (args->value)
886 pc->user_flags |= BIT(UCONTEXT_NO_ERROR_CAPTURE);
887 else
888 pc->user_flags &= ~BIT(UCONTEXT_NO_ERROR_CAPTURE);
889 break;
890
891 case I915_CONTEXT_PARAM_BANNABLE:
892 if (args->size)
893 ret = -EINVAL;
894 else if (!capable(CAP_SYS_ADMIN) && !args->value)
895 ret = -EPERM;
896 else if (args->value)
897 pc->user_flags |= BIT(UCONTEXT_BANNABLE);
898 else if (pc->uses_protected_content)
899 ret = -EPERM;
900 else
901 pc->user_flags &= ~BIT(UCONTEXT_BANNABLE);
902 break;
903
904 case I915_CONTEXT_PARAM_RECOVERABLE:
905 if (args->size)
906 ret = -EINVAL;
907 else if (!args->value)
908 pc->user_flags &= ~BIT(UCONTEXT_RECOVERABLE);
909 else if (pc->uses_protected_content)
910 ret = -EPERM;
911 else
912 pc->user_flags |= BIT(UCONTEXT_RECOVERABLE);
913 break;
914
915 case I915_CONTEXT_PARAM_PRIORITY:
916 ret = validate_priority(fpriv->dev_priv, args);
917 if (!ret)
918 pc->sched.priority = args->value;
919 break;
920
921 case I915_CONTEXT_PARAM_SSEU:
922 ret = set_proto_ctx_sseu(fpriv, pc, args);
923 break;
924
925 case I915_CONTEXT_PARAM_VM:
926 ret = set_proto_ctx_vm(fpriv, pc, args);
927 break;
928
929 case I915_CONTEXT_PARAM_ENGINES:
930 ret = set_proto_ctx_engines(fpriv, pc, args);
931 break;
932
933 case I915_CONTEXT_PARAM_PERSISTENCE:
934 if (args->size)
935 ret = -EINVAL;
936 else
937 ret = proto_context_set_persistence(fpriv->dev_priv, pc,
938 args->value);
939 break;
940
941 case I915_CONTEXT_PARAM_PROTECTED_CONTENT:
942 ret = proto_context_set_protected(fpriv->dev_priv, pc,
943 args->value);
944 break;
945
946 case I915_CONTEXT_PARAM_NO_ZEROMAP:
947 case I915_CONTEXT_PARAM_BAN_PERIOD:
948 case I915_CONTEXT_PARAM_RINGSIZE:
949 default:
950 ret = -EINVAL;
951 break;
952 }
953
954 return ret;
955 }
956
intel_context_set_gem(struct intel_context * ce,struct i915_gem_context * ctx,struct intel_sseu sseu)957 static int intel_context_set_gem(struct intel_context *ce,
958 struct i915_gem_context *ctx,
959 struct intel_sseu sseu)
960 {
961 int ret = 0;
962
963 GEM_BUG_ON(rcu_access_pointer(ce->gem_context));
964 RCU_INIT_POINTER(ce->gem_context, ctx);
965
966 GEM_BUG_ON(intel_context_is_pinned(ce));
967 ce->ring_size = SZ_16K;
968
969 i915_vm_put(ce->vm);
970 ce->vm = i915_gem_context_get_eb_vm(ctx);
971
972 if (ctx->sched.priority >= I915_PRIORITY_NORMAL &&
973 intel_engine_has_timeslices(ce->engine) &&
974 intel_engine_has_semaphores(ce->engine))
975 __set_bit(CONTEXT_USE_SEMAPHORES, &ce->flags);
976
977 if (CONFIG_DRM_I915_REQUEST_TIMEOUT &&
978 ctx->i915->params.request_timeout_ms) {
979 unsigned int timeout_ms = ctx->i915->params.request_timeout_ms;
980
981 intel_context_set_watchdog_us(ce, (u64)timeout_ms * 1000);
982 }
983
984 /* A valid SSEU has no zero fields */
985 if (sseu.slice_mask && !WARN_ON(ce->engine->class != RENDER_CLASS))
986 ret = intel_context_reconfigure_sseu(ce, sseu);
987
988 return ret;
989 }
990
__unpin_engines(struct i915_gem_engines * e,unsigned int count)991 static void __unpin_engines(struct i915_gem_engines *e, unsigned int count)
992 {
993 while (count--) {
994 struct intel_context *ce = e->engines[count], *child;
995
996 if (!ce || !test_bit(CONTEXT_PERMA_PIN, &ce->flags))
997 continue;
998
999 for_each_child(ce, child)
1000 intel_context_unpin(child);
1001 intel_context_unpin(ce);
1002 }
1003 }
1004
unpin_engines(struct i915_gem_engines * e)1005 static void unpin_engines(struct i915_gem_engines *e)
1006 {
1007 __unpin_engines(e, e->num_engines);
1008 }
1009
__free_engines(struct i915_gem_engines * e,unsigned int count)1010 static void __free_engines(struct i915_gem_engines *e, unsigned int count)
1011 {
1012 while (count--) {
1013 if (!e->engines[count])
1014 continue;
1015
1016 intel_context_put(e->engines[count]);
1017 }
1018 kfree(e);
1019 }
1020
free_engines(struct i915_gem_engines * e)1021 static void free_engines(struct i915_gem_engines *e)
1022 {
1023 __free_engines(e, e->num_engines);
1024 }
1025
free_engines_rcu(struct rcu_head * rcu)1026 static void free_engines_rcu(struct rcu_head *rcu)
1027 {
1028 struct i915_gem_engines *engines =
1029 container_of(rcu, struct i915_gem_engines, rcu);
1030
1031 i915_sw_fence_fini(&engines->fence);
1032 free_engines(engines);
1033 }
1034
accumulate_runtime(struct i915_drm_client * client,struct i915_gem_engines * engines)1035 static void accumulate_runtime(struct i915_drm_client *client,
1036 struct i915_gem_engines *engines)
1037 {
1038 struct i915_gem_engines_iter it;
1039 struct intel_context *ce;
1040
1041 if (!client)
1042 return;
1043
1044 /* Transfer accumulated runtime to the parent GEM context. */
1045 for_each_gem_engine(ce, engines, it) {
1046 unsigned int class = ce->engine->uabi_class;
1047
1048 GEM_BUG_ON(class >= ARRAY_SIZE(client->past_runtime));
1049 atomic64_add(intel_context_get_total_runtime_ns(ce),
1050 &client->past_runtime[class]);
1051 }
1052 }
1053
1054 static int
engines_notify(struct i915_sw_fence * fence,enum i915_sw_fence_notify state)1055 engines_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
1056 {
1057 struct i915_gem_engines *engines =
1058 container_of(fence, typeof(*engines), fence);
1059 struct i915_gem_context *ctx = engines->ctx;
1060
1061 switch (state) {
1062 case FENCE_COMPLETE:
1063 if (!list_empty(&engines->link)) {
1064 unsigned long flags;
1065
1066 spin_lock_irqsave(&ctx->stale.lock, flags);
1067 list_del(&engines->link);
1068 spin_unlock_irqrestore(&ctx->stale.lock, flags);
1069 }
1070 accumulate_runtime(ctx->client, engines);
1071 i915_gem_context_put(ctx);
1072
1073 break;
1074
1075 case FENCE_FREE:
1076 init_rcu_head(&engines->rcu);
1077 call_rcu(&engines->rcu, free_engines_rcu);
1078 break;
1079 }
1080
1081 return NOTIFY_DONE;
1082 }
1083
alloc_engines(unsigned int count)1084 static struct i915_gem_engines *alloc_engines(unsigned int count)
1085 {
1086 struct i915_gem_engines *e;
1087
1088 e = kzalloc(struct_size(e, engines, count), GFP_KERNEL);
1089 if (!e)
1090 return NULL;
1091
1092 i915_sw_fence_init(&e->fence, engines_notify);
1093 return e;
1094 }
1095
default_engines(struct i915_gem_context * ctx,struct intel_sseu rcs_sseu)1096 static struct i915_gem_engines *default_engines(struct i915_gem_context *ctx,
1097 struct intel_sseu rcs_sseu)
1098 {
1099 const struct intel_gt *gt = to_gt(ctx->i915);
1100 struct intel_engine_cs *engine;
1101 struct i915_gem_engines *e, *err;
1102 enum intel_engine_id id;
1103
1104 e = alloc_engines(I915_NUM_ENGINES);
1105 if (!e)
1106 return ERR_PTR(-ENOMEM);
1107
1108 for_each_engine(engine, gt, id) {
1109 struct intel_context *ce;
1110 struct intel_sseu sseu = {};
1111 int ret;
1112
1113 if (engine->legacy_idx == INVALID_ENGINE)
1114 continue;
1115
1116 GEM_BUG_ON(engine->legacy_idx >= I915_NUM_ENGINES);
1117 GEM_BUG_ON(e->engines[engine->legacy_idx]);
1118
1119 ce = intel_context_create(engine);
1120 if (IS_ERR(ce)) {
1121 err = ERR_CAST(ce);
1122 goto free_engines;
1123 }
1124
1125 e->engines[engine->legacy_idx] = ce;
1126 e->num_engines = max(e->num_engines, engine->legacy_idx + 1);
1127
1128 if (engine->class == RENDER_CLASS)
1129 sseu = rcs_sseu;
1130
1131 ret = intel_context_set_gem(ce, ctx, sseu);
1132 if (ret) {
1133 err = ERR_PTR(ret);
1134 goto free_engines;
1135 }
1136
1137 }
1138
1139 return e;
1140
1141 free_engines:
1142 free_engines(e);
1143 return err;
1144 }
1145
perma_pin_contexts(struct intel_context * ce)1146 static int perma_pin_contexts(struct intel_context *ce)
1147 {
1148 struct intel_context *child;
1149 int i = 0, j = 0, ret;
1150
1151 GEM_BUG_ON(!intel_context_is_parent(ce));
1152
1153 ret = intel_context_pin(ce);
1154 if (unlikely(ret))
1155 return ret;
1156
1157 for_each_child(ce, child) {
1158 ret = intel_context_pin(child);
1159 if (unlikely(ret))
1160 goto unwind;
1161 ++i;
1162 }
1163
1164 set_bit(CONTEXT_PERMA_PIN, &ce->flags);
1165
1166 return 0;
1167
1168 unwind:
1169 intel_context_unpin(ce);
1170 for_each_child(ce, child) {
1171 if (j++ < i)
1172 intel_context_unpin(child);
1173 else
1174 break;
1175 }
1176
1177 return ret;
1178 }
1179
user_engines(struct i915_gem_context * ctx,unsigned int num_engines,struct i915_gem_proto_engine * pe)1180 static struct i915_gem_engines *user_engines(struct i915_gem_context *ctx,
1181 unsigned int num_engines,
1182 struct i915_gem_proto_engine *pe)
1183 {
1184 struct i915_gem_engines *e, *err;
1185 unsigned int n;
1186
1187 e = alloc_engines(num_engines);
1188 if (!e)
1189 return ERR_PTR(-ENOMEM);
1190 e->num_engines = num_engines;
1191
1192 for (n = 0; n < num_engines; n++) {
1193 struct intel_context *ce, *child;
1194 int ret;
1195
1196 switch (pe[n].type) {
1197 case I915_GEM_ENGINE_TYPE_PHYSICAL:
1198 ce = intel_context_create(pe[n].engine);
1199 break;
1200
1201 case I915_GEM_ENGINE_TYPE_BALANCED:
1202 ce = intel_engine_create_virtual(pe[n].siblings,
1203 pe[n].num_siblings, 0);
1204 break;
1205
1206 case I915_GEM_ENGINE_TYPE_PARALLEL:
1207 ce = intel_engine_create_parallel(pe[n].siblings,
1208 pe[n].num_siblings,
1209 pe[n].width);
1210 break;
1211
1212 case I915_GEM_ENGINE_TYPE_INVALID:
1213 default:
1214 GEM_WARN_ON(pe[n].type != I915_GEM_ENGINE_TYPE_INVALID);
1215 continue;
1216 }
1217
1218 if (IS_ERR(ce)) {
1219 err = ERR_CAST(ce);
1220 goto free_engines;
1221 }
1222
1223 e->engines[n] = ce;
1224
1225 ret = intel_context_set_gem(ce, ctx, pe->sseu);
1226 if (ret) {
1227 err = ERR_PTR(ret);
1228 goto free_engines;
1229 }
1230 for_each_child(ce, child) {
1231 ret = intel_context_set_gem(child, ctx, pe->sseu);
1232 if (ret) {
1233 err = ERR_PTR(ret);
1234 goto free_engines;
1235 }
1236 }
1237
1238 /*
1239 * XXX: Must be done after calling intel_context_set_gem as that
1240 * function changes the ring size. The ring is allocated when
1241 * the context is pinned. If the ring size is changed after
1242 * allocation we have a mismatch of the ring size and will cause
1243 * the context to hang. Presumably with a bit of reordering we
1244 * could move the perma-pin step to the backend function
1245 * intel_engine_create_parallel.
1246 */
1247 if (pe[n].type == I915_GEM_ENGINE_TYPE_PARALLEL) {
1248 ret = perma_pin_contexts(ce);
1249 if (ret) {
1250 err = ERR_PTR(ret);
1251 goto free_engines;
1252 }
1253 }
1254 }
1255
1256 return e;
1257
1258 free_engines:
1259 free_engines(e);
1260 return err;
1261 }
1262
i915_gem_context_release_work(struct work_struct * work)1263 static void i915_gem_context_release_work(struct work_struct *work)
1264 {
1265 struct i915_gem_context *ctx = container_of(work, typeof(*ctx),
1266 release_work);
1267 struct i915_address_space *vm;
1268
1269 trace_i915_context_free(ctx);
1270 GEM_BUG_ON(!i915_gem_context_is_closed(ctx));
1271
1272 spin_lock(&ctx->i915->gem.contexts.lock);
1273 list_del(&ctx->link);
1274 spin_unlock(&ctx->i915->gem.contexts.lock);
1275
1276 if (ctx->syncobj)
1277 drm_syncobj_put(ctx->syncobj);
1278
1279 vm = ctx->vm;
1280 if (vm)
1281 i915_vm_put(vm);
1282
1283 if (ctx->pxp_wakeref)
1284 intel_runtime_pm_put(&ctx->i915->runtime_pm, ctx->pxp_wakeref);
1285
1286 if (ctx->client)
1287 i915_drm_client_put(ctx->client);
1288
1289 mutex_destroy(&ctx->engines_mutex);
1290 mutex_destroy(&ctx->lut_mutex);
1291
1292 put_pid(ctx->pid);
1293 mutex_destroy(&ctx->mutex);
1294
1295 kfree_rcu(ctx, rcu);
1296 }
1297
i915_gem_context_release(struct kref * ref)1298 void i915_gem_context_release(struct kref *ref)
1299 {
1300 struct i915_gem_context *ctx = container_of(ref, typeof(*ctx), ref);
1301
1302 queue_work(ctx->i915->wq, &ctx->release_work);
1303 }
1304
1305 static inline struct i915_gem_engines *
__context_engines_static(const struct i915_gem_context * ctx)1306 __context_engines_static(const struct i915_gem_context *ctx)
1307 {
1308 return rcu_dereference_protected(ctx->engines, true);
1309 }
1310
__reset_context(struct i915_gem_context * ctx,struct intel_engine_cs * engine)1311 static void __reset_context(struct i915_gem_context *ctx,
1312 struct intel_engine_cs *engine)
1313 {
1314 intel_gt_handle_error(engine->gt, engine->mask, 0,
1315 "context closure in %s", ctx->name);
1316 }
1317
__cancel_engine(struct intel_engine_cs * engine)1318 static bool __cancel_engine(struct intel_engine_cs *engine)
1319 {
1320 /*
1321 * Send a "high priority pulse" down the engine to cause the
1322 * current request to be momentarily preempted. (If it fails to
1323 * be preempted, it will be reset). As we have marked our context
1324 * as banned, any incomplete request, including any running, will
1325 * be skipped following the preemption.
1326 *
1327 * If there is no hangchecking (one of the reasons why we try to
1328 * cancel the context) and no forced preemption, there may be no
1329 * means by which we reset the GPU and evict the persistent hog.
1330 * Ergo if we are unable to inject a preemptive pulse that can
1331 * kill the banned context, we fallback to doing a local reset
1332 * instead.
1333 */
1334 return intel_engine_pulse(engine) == 0;
1335 }
1336
active_engine(struct intel_context * ce)1337 static struct intel_engine_cs *active_engine(struct intel_context *ce)
1338 {
1339 struct intel_engine_cs *engine = NULL;
1340 struct i915_request *rq;
1341
1342 if (intel_context_has_inflight(ce))
1343 return intel_context_inflight(ce);
1344
1345 if (!ce->timeline)
1346 return NULL;
1347
1348 /*
1349 * rq->link is only SLAB_TYPESAFE_BY_RCU, we need to hold a reference
1350 * to the request to prevent it being transferred to a new timeline
1351 * (and onto a new timeline->requests list).
1352 */
1353 rcu_read_lock();
1354 list_for_each_entry_reverse(rq, &ce->timeline->requests, link) {
1355 bool found;
1356
1357 /* timeline is already completed upto this point? */
1358 if (!i915_request_get_rcu(rq))
1359 break;
1360
1361 /* Check with the backend if the request is inflight */
1362 found = true;
1363 if (likely(rcu_access_pointer(rq->timeline) == ce->timeline))
1364 found = i915_request_active_engine(rq, &engine);
1365
1366 i915_request_put(rq);
1367 if (found)
1368 break;
1369 }
1370 rcu_read_unlock();
1371
1372 return engine;
1373 }
1374
1375 static void
kill_engines(struct i915_gem_engines * engines,bool exit,bool persistent)1376 kill_engines(struct i915_gem_engines *engines, bool exit, bool persistent)
1377 {
1378 struct i915_gem_engines_iter it;
1379 struct intel_context *ce;
1380
1381 /*
1382 * Map the user's engine back to the actual engines; one virtual
1383 * engine will be mapped to multiple engines, and using ctx->engine[]
1384 * the same engine may be have multiple instances in the user's map.
1385 * However, we only care about pending requests, so only include
1386 * engines on which there are incomplete requests.
1387 */
1388 for_each_gem_engine(ce, engines, it) {
1389 struct intel_engine_cs *engine;
1390
1391 if ((exit || !persistent) && intel_context_revoke(ce))
1392 continue; /* Already marked. */
1393
1394 /*
1395 * Check the current active state of this context; if we
1396 * are currently executing on the GPU we need to evict
1397 * ourselves. On the other hand, if we haven't yet been
1398 * submitted to the GPU or if everything is complete,
1399 * we have nothing to do.
1400 */
1401 engine = active_engine(ce);
1402
1403 /* First attempt to gracefully cancel the context */
1404 if (engine && !__cancel_engine(engine) && (exit || !persistent))
1405 /*
1406 * If we are unable to send a preemptive pulse to bump
1407 * the context from the GPU, we have to resort to a full
1408 * reset. We hope the collateral damage is worth it.
1409 */
1410 __reset_context(engines->ctx, engine);
1411 }
1412 }
1413
kill_context(struct i915_gem_context * ctx)1414 static void kill_context(struct i915_gem_context *ctx)
1415 {
1416 struct i915_gem_engines *pos, *next;
1417
1418 spin_lock_irq(&ctx->stale.lock);
1419 GEM_BUG_ON(!i915_gem_context_is_closed(ctx));
1420 list_for_each_entry_safe(pos, next, &ctx->stale.engines, link) {
1421 if (!i915_sw_fence_await(&pos->fence)) {
1422 list_del_init(&pos->link);
1423 continue;
1424 }
1425
1426 spin_unlock_irq(&ctx->stale.lock);
1427
1428 kill_engines(pos, !ctx->i915->params.enable_hangcheck,
1429 i915_gem_context_is_persistent(ctx));
1430
1431 spin_lock_irq(&ctx->stale.lock);
1432 GEM_BUG_ON(i915_sw_fence_signaled(&pos->fence));
1433 list_safe_reset_next(pos, next, link);
1434 list_del_init(&pos->link); /* decouple from FENCE_COMPLETE */
1435
1436 i915_sw_fence_complete(&pos->fence);
1437 }
1438 spin_unlock_irq(&ctx->stale.lock);
1439 }
1440
engines_idle_release(struct i915_gem_context * ctx,struct i915_gem_engines * engines)1441 static void engines_idle_release(struct i915_gem_context *ctx,
1442 struct i915_gem_engines *engines)
1443 {
1444 struct i915_gem_engines_iter it;
1445 struct intel_context *ce;
1446
1447 INIT_LIST_HEAD(&engines->link);
1448
1449 engines->ctx = i915_gem_context_get(ctx);
1450
1451 for_each_gem_engine(ce, engines, it) {
1452 int err;
1453
1454 /* serialises with execbuf */
1455 set_bit(CONTEXT_CLOSED_BIT, &ce->flags);
1456 if (!intel_context_pin_if_active(ce))
1457 continue;
1458
1459 /* Wait until context is finally scheduled out and retired */
1460 err = i915_sw_fence_await_active(&engines->fence,
1461 &ce->active,
1462 I915_ACTIVE_AWAIT_BARRIER);
1463 intel_context_unpin(ce);
1464 if (err)
1465 goto kill;
1466 }
1467
1468 spin_lock_irq(&ctx->stale.lock);
1469 if (!i915_gem_context_is_closed(ctx))
1470 list_add_tail(&engines->link, &ctx->stale.engines);
1471 spin_unlock_irq(&ctx->stale.lock);
1472
1473 kill:
1474 if (list_empty(&engines->link)) /* raced, already closed */
1475 kill_engines(engines, true,
1476 i915_gem_context_is_persistent(ctx));
1477
1478 i915_sw_fence_commit(&engines->fence);
1479 }
1480
set_closed_name(struct i915_gem_context * ctx)1481 static void set_closed_name(struct i915_gem_context *ctx)
1482 {
1483 char *s;
1484
1485 /* Replace '[]' with '<>' to indicate closed in debug prints */
1486
1487 s = strrchr(ctx->name, '[');
1488 if (!s)
1489 return;
1490
1491 *s = '<';
1492
1493 s = strchr(s + 1, ']');
1494 if (s)
1495 *s = '>';
1496 }
1497
context_close(struct i915_gem_context * ctx)1498 static void context_close(struct i915_gem_context *ctx)
1499 {
1500 struct i915_drm_client *client;
1501
1502 /* Flush any concurrent set_engines() */
1503 mutex_lock(&ctx->engines_mutex);
1504 unpin_engines(__context_engines_static(ctx));
1505 engines_idle_release(ctx, rcu_replace_pointer(ctx->engines, NULL, 1));
1506 i915_gem_context_set_closed(ctx);
1507 mutex_unlock(&ctx->engines_mutex);
1508
1509 mutex_lock(&ctx->mutex);
1510
1511 set_closed_name(ctx);
1512
1513 /*
1514 * The LUT uses the VMA as a backpointer to unref the object,
1515 * so we need to clear the LUT before we close all the VMA (inside
1516 * the ppgtt).
1517 */
1518 lut_close(ctx);
1519
1520 ctx->file_priv = ERR_PTR(-EBADF);
1521
1522 client = ctx->client;
1523 if (client) {
1524 spin_lock(&client->ctx_lock);
1525 list_del_rcu(&ctx->client_link);
1526 spin_unlock(&client->ctx_lock);
1527 }
1528
1529 mutex_unlock(&ctx->mutex);
1530
1531 /*
1532 * If the user has disabled hangchecking, we can not be sure that
1533 * the batches will ever complete after the context is closed,
1534 * keeping the context and all resources pinned forever. So in this
1535 * case we opt to forcibly kill off all remaining requests on
1536 * context close.
1537 */
1538 kill_context(ctx);
1539
1540 i915_gem_context_put(ctx);
1541 }
1542
__context_set_persistence(struct i915_gem_context * ctx,bool state)1543 static int __context_set_persistence(struct i915_gem_context *ctx, bool state)
1544 {
1545 if (i915_gem_context_is_persistent(ctx) == state)
1546 return 0;
1547
1548 if (state) {
1549 /*
1550 * Only contexts that are short-lived [that will expire or be
1551 * reset] are allowed to survive past termination. We require
1552 * hangcheck to ensure that the persistent requests are healthy.
1553 */
1554 if (!ctx->i915->params.enable_hangcheck)
1555 return -EINVAL;
1556
1557 i915_gem_context_set_persistence(ctx);
1558 } else {
1559 /* To cancel a context we use "preempt-to-idle" */
1560 if (!(ctx->i915->caps.scheduler & I915_SCHEDULER_CAP_PREEMPTION))
1561 return -ENODEV;
1562
1563 /*
1564 * If the cancel fails, we then need to reset, cleanly!
1565 *
1566 * If the per-engine reset fails, all hope is lost! We resort
1567 * to a full GPU reset in that unlikely case, but realistically
1568 * if the engine could not reset, the full reset does not fare
1569 * much better. The damage has been done.
1570 *
1571 * However, if we cannot reset an engine by itself, we cannot
1572 * cleanup a hanging persistent context without causing
1573 * colateral damage, and we should not pretend we can by
1574 * exposing the interface.
1575 */
1576 if (!intel_has_reset_engine(to_gt(ctx->i915)))
1577 return -ENODEV;
1578
1579 i915_gem_context_clear_persistence(ctx);
1580 }
1581
1582 return 0;
1583 }
1584
1585 static struct i915_gem_context *
i915_gem_create_context(struct drm_i915_private * i915,const struct i915_gem_proto_context * pc)1586 i915_gem_create_context(struct drm_i915_private *i915,
1587 const struct i915_gem_proto_context *pc)
1588 {
1589 struct i915_gem_context *ctx;
1590 struct i915_address_space *vm = NULL;
1591 struct i915_gem_engines *e;
1592 int err;
1593 int i;
1594
1595 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1596 if (!ctx)
1597 return ERR_PTR(-ENOMEM);
1598
1599 kref_init(&ctx->ref);
1600 ctx->i915 = i915;
1601 ctx->sched = pc->sched;
1602 mutex_init(&ctx->mutex);
1603 INIT_LIST_HEAD(&ctx->link);
1604 INIT_WORK(&ctx->release_work, i915_gem_context_release_work);
1605
1606 spin_lock_init(&ctx->stale.lock);
1607 INIT_LIST_HEAD(&ctx->stale.engines);
1608
1609 if (pc->vm) {
1610 vm = i915_vm_get(pc->vm);
1611 } else if (HAS_FULL_PPGTT(i915)) {
1612 struct i915_ppgtt *ppgtt;
1613
1614 ppgtt = i915_ppgtt_create(to_gt(i915), 0);
1615 if (IS_ERR(ppgtt)) {
1616 drm_dbg(&i915->drm, "PPGTT setup failed (%ld)\n",
1617 PTR_ERR(ppgtt));
1618 err = PTR_ERR(ppgtt);
1619 goto err_ctx;
1620 }
1621 vm = &ppgtt->vm;
1622 }
1623 if (vm)
1624 ctx->vm = vm;
1625
1626 mutex_init(&ctx->engines_mutex);
1627 if (pc->num_user_engines >= 0) {
1628 i915_gem_context_set_user_engines(ctx);
1629 e = user_engines(ctx, pc->num_user_engines, pc->user_engines);
1630 } else {
1631 i915_gem_context_clear_user_engines(ctx);
1632 e = default_engines(ctx, pc->legacy_rcs_sseu);
1633 }
1634 if (IS_ERR(e)) {
1635 err = PTR_ERR(e);
1636 goto err_vm;
1637 }
1638 RCU_INIT_POINTER(ctx->engines, e);
1639
1640 INIT_RADIX_TREE(&ctx->handles_vma, GFP_KERNEL);
1641 mutex_init(&ctx->lut_mutex);
1642
1643 /* NB: Mark all slices as needing a remap so that when the context first
1644 * loads it will restore whatever remap state already exists. If there
1645 * is no remap info, it will be a NOP. */
1646 ctx->remap_slice = ALL_L3_SLICES(i915);
1647
1648 ctx->user_flags = pc->user_flags;
1649
1650 for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp); i++)
1651 ctx->hang_timestamp[i] = jiffies - CONTEXT_FAST_HANG_JIFFIES;
1652
1653 if (pc->single_timeline) {
1654 err = drm_syncobj_create(&ctx->syncobj,
1655 DRM_SYNCOBJ_CREATE_SIGNALED,
1656 NULL);
1657 if (err)
1658 goto err_engines;
1659 }
1660
1661 if (pc->uses_protected_content) {
1662 ctx->pxp_wakeref = intel_runtime_pm_get(&i915->runtime_pm);
1663 ctx->uses_protected_content = true;
1664 }
1665
1666 trace_i915_context_create(ctx);
1667
1668 return ctx;
1669
1670 err_engines:
1671 free_engines(e);
1672 err_vm:
1673 if (ctx->vm)
1674 i915_vm_put(ctx->vm);
1675 err_ctx:
1676 kfree(ctx);
1677 return ERR_PTR(err);
1678 }
1679
init_contexts(struct i915_gem_contexts * gc)1680 static void init_contexts(struct i915_gem_contexts *gc)
1681 {
1682 spin_lock_init(&gc->lock);
1683 INIT_LIST_HEAD(&gc->list);
1684 }
1685
i915_gem_init__contexts(struct drm_i915_private * i915)1686 void i915_gem_init__contexts(struct drm_i915_private *i915)
1687 {
1688 init_contexts(&i915->gem.contexts);
1689 }
1690
1691 /*
1692 * Note that this implicitly consumes the ctx reference, by placing
1693 * the ctx in the context_xa.
1694 */
gem_context_register(struct i915_gem_context * ctx,struct drm_i915_file_private * fpriv,u32 id)1695 static void gem_context_register(struct i915_gem_context *ctx,
1696 struct drm_i915_file_private *fpriv,
1697 u32 id)
1698 {
1699 struct drm_i915_private *i915 = ctx->i915;
1700 void *old;
1701
1702 ctx->file_priv = fpriv;
1703
1704 ctx->pid = get_task_pid(current, PIDTYPE_PID);
1705 ctx->client = i915_drm_client_get(fpriv->client);
1706
1707 snprintf(ctx->name, sizeof(ctx->name), "%s[%d]",
1708 current->comm, pid_nr(ctx->pid));
1709
1710 spin_lock(&ctx->client->ctx_lock);
1711 list_add_tail_rcu(&ctx->client_link, &ctx->client->ctx_list);
1712 spin_unlock(&ctx->client->ctx_lock);
1713
1714 spin_lock(&i915->gem.contexts.lock);
1715 list_add_tail(&ctx->link, &i915->gem.contexts.list);
1716 spin_unlock(&i915->gem.contexts.lock);
1717
1718 /* And finally expose ourselves to userspace via the idr */
1719 old = xa_store(&fpriv->context_xa, id, ctx, GFP_KERNEL);
1720 WARN_ON(old);
1721 }
1722
i915_gem_context_open(struct drm_i915_private * i915,struct drm_file * file)1723 int i915_gem_context_open(struct drm_i915_private *i915,
1724 struct drm_file *file)
1725 {
1726 struct drm_i915_file_private *file_priv = file->driver_priv;
1727 struct i915_gem_proto_context *pc;
1728 struct i915_gem_context *ctx;
1729 int err;
1730
1731 mutex_init(&file_priv->proto_context_lock);
1732 xa_init_flags(&file_priv->proto_context_xa, XA_FLAGS_ALLOC);
1733
1734 /* 0 reserved for the default context */
1735 xa_init_flags(&file_priv->context_xa, XA_FLAGS_ALLOC1);
1736
1737 /* 0 reserved for invalid/unassigned ppgtt */
1738 xa_init_flags(&file_priv->vm_xa, XA_FLAGS_ALLOC1);
1739
1740 pc = proto_context_create(i915, 0);
1741 if (IS_ERR(pc)) {
1742 err = PTR_ERR(pc);
1743 goto err;
1744 }
1745
1746 ctx = i915_gem_create_context(i915, pc);
1747 proto_context_close(i915, pc);
1748 if (IS_ERR(ctx)) {
1749 err = PTR_ERR(ctx);
1750 goto err;
1751 }
1752
1753 gem_context_register(ctx, file_priv, 0);
1754
1755 return 0;
1756
1757 err:
1758 xa_destroy(&file_priv->vm_xa);
1759 xa_destroy(&file_priv->context_xa);
1760 xa_destroy(&file_priv->proto_context_xa);
1761 mutex_destroy(&file_priv->proto_context_lock);
1762 return err;
1763 }
1764
i915_gem_context_close(struct drm_file * file)1765 void i915_gem_context_close(struct drm_file *file)
1766 {
1767 struct drm_i915_file_private *file_priv = file->driver_priv;
1768 struct i915_gem_proto_context *pc;
1769 struct i915_address_space *vm;
1770 struct i915_gem_context *ctx;
1771 unsigned long idx;
1772
1773 xa_for_each(&file_priv->proto_context_xa, idx, pc)
1774 proto_context_close(file_priv->dev_priv, pc);
1775 xa_destroy(&file_priv->proto_context_xa);
1776 mutex_destroy(&file_priv->proto_context_lock);
1777
1778 xa_for_each(&file_priv->context_xa, idx, ctx)
1779 context_close(ctx);
1780 xa_destroy(&file_priv->context_xa);
1781
1782 xa_for_each(&file_priv->vm_xa, idx, vm)
1783 i915_vm_put(vm);
1784 xa_destroy(&file_priv->vm_xa);
1785 }
1786
i915_gem_vm_create_ioctl(struct drm_device * dev,void * data,struct drm_file * file)1787 int i915_gem_vm_create_ioctl(struct drm_device *dev, void *data,
1788 struct drm_file *file)
1789 {
1790 struct drm_i915_private *i915 = to_i915(dev);
1791 struct drm_i915_gem_vm_control *args = data;
1792 struct drm_i915_file_private *file_priv = file->driver_priv;
1793 struct i915_ppgtt *ppgtt;
1794 u32 id;
1795 int err;
1796
1797 if (!HAS_FULL_PPGTT(i915))
1798 return -ENODEV;
1799
1800 if (args->flags)
1801 return -EINVAL;
1802
1803 ppgtt = i915_ppgtt_create(to_gt(i915), 0);
1804 if (IS_ERR(ppgtt))
1805 return PTR_ERR(ppgtt);
1806
1807 if (args->extensions) {
1808 err = i915_user_extensions(u64_to_user_ptr(args->extensions),
1809 NULL, 0,
1810 ppgtt);
1811 if (err)
1812 goto err_put;
1813 }
1814
1815 err = xa_alloc(&file_priv->vm_xa, &id, &ppgtt->vm,
1816 xa_limit_32b, GFP_KERNEL);
1817 if (err)
1818 goto err_put;
1819
1820 GEM_BUG_ON(id == 0); /* reserved for invalid/unassigned ppgtt */
1821 args->vm_id = id;
1822 return 0;
1823
1824 err_put:
1825 i915_vm_put(&ppgtt->vm);
1826 return err;
1827 }
1828
i915_gem_vm_destroy_ioctl(struct drm_device * dev,void * data,struct drm_file * file)1829 int i915_gem_vm_destroy_ioctl(struct drm_device *dev, void *data,
1830 struct drm_file *file)
1831 {
1832 struct drm_i915_file_private *file_priv = file->driver_priv;
1833 struct drm_i915_gem_vm_control *args = data;
1834 struct i915_address_space *vm;
1835
1836 if (args->flags)
1837 return -EINVAL;
1838
1839 if (args->extensions)
1840 return -EINVAL;
1841
1842 vm = xa_erase(&file_priv->vm_xa, args->vm_id);
1843 if (!vm)
1844 return -ENOENT;
1845
1846 i915_vm_put(vm);
1847 return 0;
1848 }
1849
get_ppgtt(struct drm_i915_file_private * file_priv,struct i915_gem_context * ctx,struct drm_i915_gem_context_param * args)1850 static int get_ppgtt(struct drm_i915_file_private *file_priv,
1851 struct i915_gem_context *ctx,
1852 struct drm_i915_gem_context_param *args)
1853 {
1854 struct i915_address_space *vm;
1855 int err;
1856 u32 id;
1857
1858 if (!i915_gem_context_has_full_ppgtt(ctx))
1859 return -ENODEV;
1860
1861 vm = ctx->vm;
1862 GEM_BUG_ON(!vm);
1863
1864 err = xa_alloc(&file_priv->vm_xa, &id, vm, xa_limit_32b, GFP_KERNEL);
1865 if (err)
1866 return err;
1867
1868 i915_vm_get(vm);
1869
1870 GEM_BUG_ON(id == 0); /* reserved for invalid/unassigned ppgtt */
1871 args->value = id;
1872 args->size = 0;
1873
1874 return err;
1875 }
1876
1877 int
i915_gem_user_to_context_sseu(struct intel_gt * gt,const struct drm_i915_gem_context_param_sseu * user,struct intel_sseu * context)1878 i915_gem_user_to_context_sseu(struct intel_gt *gt,
1879 const struct drm_i915_gem_context_param_sseu *user,
1880 struct intel_sseu *context)
1881 {
1882 const struct sseu_dev_info *device = >->info.sseu;
1883 struct drm_i915_private *i915 = gt->i915;
1884 unsigned int dev_subslice_mask = intel_sseu_get_hsw_subslices(device, 0);
1885
1886 /* No zeros in any field. */
1887 if (!user->slice_mask || !user->subslice_mask ||
1888 !user->min_eus_per_subslice || !user->max_eus_per_subslice)
1889 return -EINVAL;
1890
1891 /* Max > min. */
1892 if (user->max_eus_per_subslice < user->min_eus_per_subslice)
1893 return -EINVAL;
1894
1895 /*
1896 * Some future proofing on the types since the uAPI is wider than the
1897 * current internal implementation.
1898 */
1899 if (overflows_type(user->slice_mask, context->slice_mask) ||
1900 overflows_type(user->subslice_mask, context->subslice_mask) ||
1901 overflows_type(user->min_eus_per_subslice,
1902 context->min_eus_per_subslice) ||
1903 overflows_type(user->max_eus_per_subslice,
1904 context->max_eus_per_subslice))
1905 return -EINVAL;
1906
1907 /* Check validity against hardware. */
1908 if (user->slice_mask & ~device->slice_mask)
1909 return -EINVAL;
1910
1911 if (user->subslice_mask & ~dev_subslice_mask)
1912 return -EINVAL;
1913
1914 if (user->max_eus_per_subslice > device->max_eus_per_subslice)
1915 return -EINVAL;
1916
1917 context->slice_mask = user->slice_mask;
1918 context->subslice_mask = user->subslice_mask;
1919 context->min_eus_per_subslice = user->min_eus_per_subslice;
1920 context->max_eus_per_subslice = user->max_eus_per_subslice;
1921
1922 /* Part specific restrictions. */
1923 if (GRAPHICS_VER(i915) == 11) {
1924 unsigned int hw_s = hweight8(device->slice_mask);
1925 unsigned int hw_ss_per_s = hweight8(dev_subslice_mask);
1926 unsigned int req_s = hweight8(context->slice_mask);
1927 unsigned int req_ss = hweight8(context->subslice_mask);
1928
1929 /*
1930 * Only full subslice enablement is possible if more than one
1931 * slice is turned on.
1932 */
1933 if (req_s > 1 && req_ss != hw_ss_per_s)
1934 return -EINVAL;
1935
1936 /*
1937 * If more than four (SScount bitfield limit) subslices are
1938 * requested then the number has to be even.
1939 */
1940 if (req_ss > 4 && (req_ss & 1))
1941 return -EINVAL;
1942
1943 /*
1944 * If only one slice is enabled and subslice count is below the
1945 * device full enablement, it must be at most half of the all
1946 * available subslices.
1947 */
1948 if (req_s == 1 && req_ss < hw_ss_per_s &&
1949 req_ss > (hw_ss_per_s / 2))
1950 return -EINVAL;
1951
1952 /* ABI restriction - VME use case only. */
1953
1954 /* All slices or one slice only. */
1955 if (req_s != 1 && req_s != hw_s)
1956 return -EINVAL;
1957
1958 /*
1959 * Half subslices or full enablement only when one slice is
1960 * enabled.
1961 */
1962 if (req_s == 1 &&
1963 (req_ss != hw_ss_per_s && req_ss != (hw_ss_per_s / 2)))
1964 return -EINVAL;
1965
1966 /* No EU configuration changes. */
1967 if ((user->min_eus_per_subslice !=
1968 device->max_eus_per_subslice) ||
1969 (user->max_eus_per_subslice !=
1970 device->max_eus_per_subslice))
1971 return -EINVAL;
1972 }
1973
1974 return 0;
1975 }
1976
set_sseu(struct i915_gem_context * ctx,struct drm_i915_gem_context_param * args)1977 static int set_sseu(struct i915_gem_context *ctx,
1978 struct drm_i915_gem_context_param *args)
1979 {
1980 struct drm_i915_private *i915 = ctx->i915;
1981 struct drm_i915_gem_context_param_sseu user_sseu;
1982 struct intel_context *ce;
1983 struct intel_sseu sseu;
1984 unsigned long lookup;
1985 int ret;
1986
1987 if (args->size < sizeof(user_sseu))
1988 return -EINVAL;
1989
1990 if (GRAPHICS_VER(i915) != 11)
1991 return -ENODEV;
1992
1993 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
1994 sizeof(user_sseu)))
1995 return -EFAULT;
1996
1997 if (user_sseu.rsvd)
1998 return -EINVAL;
1999
2000 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
2001 return -EINVAL;
2002
2003 lookup = 0;
2004 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
2005 lookup |= LOOKUP_USER_INDEX;
2006
2007 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
2008 if (IS_ERR(ce))
2009 return PTR_ERR(ce);
2010
2011 /* Only render engine supports RPCS configuration. */
2012 if (ce->engine->class != RENDER_CLASS) {
2013 ret = -ENODEV;
2014 goto out_ce;
2015 }
2016
2017 ret = i915_gem_user_to_context_sseu(ce->engine->gt, &user_sseu, &sseu);
2018 if (ret)
2019 goto out_ce;
2020
2021 ret = intel_context_reconfigure_sseu(ce, sseu);
2022 if (ret)
2023 goto out_ce;
2024
2025 args->size = sizeof(user_sseu);
2026
2027 out_ce:
2028 intel_context_put(ce);
2029 return ret;
2030 }
2031
2032 static int
set_persistence(struct i915_gem_context * ctx,const struct drm_i915_gem_context_param * args)2033 set_persistence(struct i915_gem_context *ctx,
2034 const struct drm_i915_gem_context_param *args)
2035 {
2036 if (args->size)
2037 return -EINVAL;
2038
2039 return __context_set_persistence(ctx, args->value);
2040 }
2041
set_priority(struct i915_gem_context * ctx,const struct drm_i915_gem_context_param * args)2042 static int set_priority(struct i915_gem_context *ctx,
2043 const struct drm_i915_gem_context_param *args)
2044 {
2045 struct i915_gem_engines_iter it;
2046 struct intel_context *ce;
2047 int err;
2048
2049 err = validate_priority(ctx->i915, args);
2050 if (err)
2051 return err;
2052
2053 ctx->sched.priority = args->value;
2054
2055 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
2056 if (!intel_engine_has_timeslices(ce->engine))
2057 continue;
2058
2059 if (ctx->sched.priority >= I915_PRIORITY_NORMAL &&
2060 intel_engine_has_semaphores(ce->engine))
2061 intel_context_set_use_semaphores(ce);
2062 else
2063 intel_context_clear_use_semaphores(ce);
2064 }
2065 i915_gem_context_unlock_engines(ctx);
2066
2067 return 0;
2068 }
2069
get_protected(struct i915_gem_context * ctx,struct drm_i915_gem_context_param * args)2070 static int get_protected(struct i915_gem_context *ctx,
2071 struct drm_i915_gem_context_param *args)
2072 {
2073 args->size = 0;
2074 args->value = i915_gem_context_uses_protected_content(ctx);
2075
2076 return 0;
2077 }
2078
ctx_setparam(struct drm_i915_file_private * fpriv,struct i915_gem_context * ctx,struct drm_i915_gem_context_param * args)2079 static int ctx_setparam(struct drm_i915_file_private *fpriv,
2080 struct i915_gem_context *ctx,
2081 struct drm_i915_gem_context_param *args)
2082 {
2083 int ret = 0;
2084
2085 switch (args->param) {
2086 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
2087 if (args->size)
2088 ret = -EINVAL;
2089 else if (args->value)
2090 i915_gem_context_set_no_error_capture(ctx);
2091 else
2092 i915_gem_context_clear_no_error_capture(ctx);
2093 break;
2094
2095 case I915_CONTEXT_PARAM_BANNABLE:
2096 if (args->size)
2097 ret = -EINVAL;
2098 else if (!capable(CAP_SYS_ADMIN) && !args->value)
2099 ret = -EPERM;
2100 else if (args->value)
2101 i915_gem_context_set_bannable(ctx);
2102 else if (i915_gem_context_uses_protected_content(ctx))
2103 ret = -EPERM; /* can't clear this for protected contexts */
2104 else
2105 i915_gem_context_clear_bannable(ctx);
2106 break;
2107
2108 case I915_CONTEXT_PARAM_RECOVERABLE:
2109 if (args->size)
2110 ret = -EINVAL;
2111 else if (!args->value)
2112 i915_gem_context_clear_recoverable(ctx);
2113 else if (i915_gem_context_uses_protected_content(ctx))
2114 ret = -EPERM; /* can't set this for protected contexts */
2115 else
2116 i915_gem_context_set_recoverable(ctx);
2117 break;
2118
2119 case I915_CONTEXT_PARAM_PRIORITY:
2120 ret = set_priority(ctx, args);
2121 break;
2122
2123 case I915_CONTEXT_PARAM_SSEU:
2124 ret = set_sseu(ctx, args);
2125 break;
2126
2127 case I915_CONTEXT_PARAM_PERSISTENCE:
2128 ret = set_persistence(ctx, args);
2129 break;
2130
2131 case I915_CONTEXT_PARAM_PROTECTED_CONTENT:
2132 case I915_CONTEXT_PARAM_NO_ZEROMAP:
2133 case I915_CONTEXT_PARAM_BAN_PERIOD:
2134 case I915_CONTEXT_PARAM_RINGSIZE:
2135 case I915_CONTEXT_PARAM_VM:
2136 case I915_CONTEXT_PARAM_ENGINES:
2137 default:
2138 ret = -EINVAL;
2139 break;
2140 }
2141
2142 return ret;
2143 }
2144
2145 struct create_ext {
2146 struct i915_gem_proto_context *pc;
2147 struct drm_i915_file_private *fpriv;
2148 };
2149
create_setparam(struct i915_user_extension __user * ext,void * data)2150 static int create_setparam(struct i915_user_extension __user *ext, void *data)
2151 {
2152 struct drm_i915_gem_context_create_ext_setparam local;
2153 const struct create_ext *arg = data;
2154
2155 if (copy_from_user(&local, ext, sizeof(local)))
2156 return -EFAULT;
2157
2158 if (local.param.ctx_id)
2159 return -EINVAL;
2160
2161 return set_proto_ctx_param(arg->fpriv, arg->pc, &local.param);
2162 }
2163
invalid_ext(struct i915_user_extension __user * ext,void * data)2164 static int invalid_ext(struct i915_user_extension __user *ext, void *data)
2165 {
2166 return -EINVAL;
2167 }
2168
2169 static const i915_user_extension_fn create_extensions[] = {
2170 [I915_CONTEXT_CREATE_EXT_SETPARAM] = create_setparam,
2171 [I915_CONTEXT_CREATE_EXT_CLONE] = invalid_ext,
2172 };
2173
client_is_banned(struct drm_i915_file_private * file_priv)2174 static bool client_is_banned(struct drm_i915_file_private *file_priv)
2175 {
2176 return atomic_read(&file_priv->ban_score) >= I915_CLIENT_SCORE_BANNED;
2177 }
2178
2179 static inline struct i915_gem_context *
__context_lookup(struct drm_i915_file_private * file_priv,u32 id)2180 __context_lookup(struct drm_i915_file_private *file_priv, u32 id)
2181 {
2182 struct i915_gem_context *ctx;
2183
2184 rcu_read_lock();
2185 ctx = xa_load(&file_priv->context_xa, id);
2186 if (ctx && !kref_get_unless_zero(&ctx->ref))
2187 ctx = NULL;
2188 rcu_read_unlock();
2189
2190 return ctx;
2191 }
2192
2193 static struct i915_gem_context *
finalize_create_context_locked(struct drm_i915_file_private * file_priv,struct i915_gem_proto_context * pc,u32 id)2194 finalize_create_context_locked(struct drm_i915_file_private *file_priv,
2195 struct i915_gem_proto_context *pc, u32 id)
2196 {
2197 struct i915_gem_context *ctx;
2198 void *old;
2199
2200 lockdep_assert_held(&file_priv->proto_context_lock);
2201
2202 ctx = i915_gem_create_context(file_priv->dev_priv, pc);
2203 if (IS_ERR(ctx))
2204 return ctx;
2205
2206 /*
2207 * One for the xarray and one for the caller. We need to grab
2208 * the reference *prior* to making the ctx visble to userspace
2209 * in gem_context_register(), as at any point after that
2210 * userspace can try to race us with another thread destroying
2211 * the context under our feet.
2212 */
2213 i915_gem_context_get(ctx);
2214
2215 gem_context_register(ctx, file_priv, id);
2216
2217 old = xa_erase(&file_priv->proto_context_xa, id);
2218 GEM_BUG_ON(old != pc);
2219 proto_context_close(file_priv->dev_priv, pc);
2220
2221 return ctx;
2222 }
2223
2224 struct i915_gem_context *
i915_gem_context_lookup(struct drm_i915_file_private * file_priv,u32 id)2225 i915_gem_context_lookup(struct drm_i915_file_private *file_priv, u32 id)
2226 {
2227 struct i915_gem_proto_context *pc;
2228 struct i915_gem_context *ctx;
2229
2230 ctx = __context_lookup(file_priv, id);
2231 if (ctx)
2232 return ctx;
2233
2234 mutex_lock(&file_priv->proto_context_lock);
2235 /* Try one more time under the lock */
2236 ctx = __context_lookup(file_priv, id);
2237 if (!ctx) {
2238 pc = xa_load(&file_priv->proto_context_xa, id);
2239 if (!pc)
2240 ctx = ERR_PTR(-ENOENT);
2241 else
2242 ctx = finalize_create_context_locked(file_priv, pc, id);
2243 }
2244 mutex_unlock(&file_priv->proto_context_lock);
2245
2246 return ctx;
2247 }
2248
i915_gem_context_create_ioctl(struct drm_device * dev,void * data,struct drm_file * file)2249 int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
2250 struct drm_file *file)
2251 {
2252 struct drm_i915_private *i915 = to_i915(dev);
2253 struct drm_i915_gem_context_create_ext *args = data;
2254 struct create_ext ext_data;
2255 int ret;
2256 u32 id;
2257
2258 if (!DRIVER_CAPS(i915)->has_logical_contexts)
2259 return -ENODEV;
2260
2261 if (args->flags & I915_CONTEXT_CREATE_FLAGS_UNKNOWN)
2262 return -EINVAL;
2263
2264 ret = intel_gt_terminally_wedged(to_gt(i915));
2265 if (ret)
2266 return ret;
2267
2268 ext_data.fpriv = file->driver_priv;
2269 if (client_is_banned(ext_data.fpriv)) {
2270 drm_dbg(&i915->drm,
2271 "client %s[%d] banned from creating ctx\n",
2272 current->comm, task_pid_nr(current));
2273 return -EIO;
2274 }
2275
2276 ext_data.pc = proto_context_create(i915, args->flags);
2277 if (IS_ERR(ext_data.pc))
2278 return PTR_ERR(ext_data.pc);
2279
2280 if (args->flags & I915_CONTEXT_CREATE_FLAGS_USE_EXTENSIONS) {
2281 ret = i915_user_extensions(u64_to_user_ptr(args->extensions),
2282 create_extensions,
2283 ARRAY_SIZE(create_extensions),
2284 &ext_data);
2285 if (ret)
2286 goto err_pc;
2287 }
2288
2289 if (GRAPHICS_VER(i915) > 12) {
2290 struct i915_gem_context *ctx;
2291
2292 /* Get ourselves a context ID */
2293 ret = xa_alloc(&ext_data.fpriv->context_xa, &id, NULL,
2294 xa_limit_32b, GFP_KERNEL);
2295 if (ret)
2296 goto err_pc;
2297
2298 ctx = i915_gem_create_context(i915, ext_data.pc);
2299 if (IS_ERR(ctx)) {
2300 ret = PTR_ERR(ctx);
2301 goto err_pc;
2302 }
2303
2304 proto_context_close(i915, ext_data.pc);
2305 gem_context_register(ctx, ext_data.fpriv, id);
2306 } else {
2307 ret = proto_context_register(ext_data.fpriv, ext_data.pc, &id);
2308 if (ret < 0)
2309 goto err_pc;
2310 }
2311
2312 args->ctx_id = id;
2313 drm_dbg(&i915->drm, "HW context %d created\n", args->ctx_id);
2314
2315 return 0;
2316
2317 err_pc:
2318 proto_context_close(i915, ext_data.pc);
2319 return ret;
2320 }
2321
i915_gem_context_destroy_ioctl(struct drm_device * dev,void * data,struct drm_file * file)2322 int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
2323 struct drm_file *file)
2324 {
2325 struct drm_i915_gem_context_destroy *args = data;
2326 struct drm_i915_file_private *file_priv = file->driver_priv;
2327 struct i915_gem_proto_context *pc;
2328 struct i915_gem_context *ctx;
2329
2330 if (args->pad != 0)
2331 return -EINVAL;
2332
2333 if (!args->ctx_id)
2334 return -ENOENT;
2335
2336 /* We need to hold the proto-context lock here to prevent races
2337 * with finalize_create_context_locked().
2338 */
2339 mutex_lock(&file_priv->proto_context_lock);
2340 ctx = xa_erase(&file_priv->context_xa, args->ctx_id);
2341 pc = xa_erase(&file_priv->proto_context_xa, args->ctx_id);
2342 mutex_unlock(&file_priv->proto_context_lock);
2343
2344 if (!ctx && !pc)
2345 return -ENOENT;
2346 GEM_WARN_ON(ctx && pc);
2347
2348 if (pc)
2349 proto_context_close(file_priv->dev_priv, pc);
2350
2351 if (ctx)
2352 context_close(ctx);
2353
2354 return 0;
2355 }
2356
get_sseu(struct i915_gem_context * ctx,struct drm_i915_gem_context_param * args)2357 static int get_sseu(struct i915_gem_context *ctx,
2358 struct drm_i915_gem_context_param *args)
2359 {
2360 struct drm_i915_gem_context_param_sseu user_sseu;
2361 struct intel_context *ce;
2362 unsigned long lookup;
2363 int err;
2364
2365 if (args->size == 0)
2366 goto out;
2367 else if (args->size < sizeof(user_sseu))
2368 return -EINVAL;
2369
2370 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
2371 sizeof(user_sseu)))
2372 return -EFAULT;
2373
2374 if (user_sseu.rsvd)
2375 return -EINVAL;
2376
2377 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
2378 return -EINVAL;
2379
2380 lookup = 0;
2381 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
2382 lookup |= LOOKUP_USER_INDEX;
2383
2384 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
2385 if (IS_ERR(ce))
2386 return PTR_ERR(ce);
2387
2388 err = intel_context_lock_pinned(ce); /* serialises with set_sseu */
2389 if (err) {
2390 intel_context_put(ce);
2391 return err;
2392 }
2393
2394 user_sseu.slice_mask = ce->sseu.slice_mask;
2395 user_sseu.subslice_mask = ce->sseu.subslice_mask;
2396 user_sseu.min_eus_per_subslice = ce->sseu.min_eus_per_subslice;
2397 user_sseu.max_eus_per_subslice = ce->sseu.max_eus_per_subslice;
2398
2399 intel_context_unlock_pinned(ce);
2400 intel_context_put(ce);
2401
2402 if (copy_to_user(u64_to_user_ptr(args->value), &user_sseu,
2403 sizeof(user_sseu)))
2404 return -EFAULT;
2405
2406 out:
2407 args->size = sizeof(user_sseu);
2408
2409 return 0;
2410 }
2411
i915_gem_context_getparam_ioctl(struct drm_device * dev,void * data,struct drm_file * file)2412 int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
2413 struct drm_file *file)
2414 {
2415 struct drm_i915_file_private *file_priv = file->driver_priv;
2416 struct drm_i915_gem_context_param *args = data;
2417 struct i915_gem_context *ctx;
2418 struct i915_address_space *vm;
2419 int ret = 0;
2420
2421 ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
2422 if (IS_ERR(ctx))
2423 return PTR_ERR(ctx);
2424
2425 switch (args->param) {
2426 case I915_CONTEXT_PARAM_GTT_SIZE:
2427 args->size = 0;
2428 vm = i915_gem_context_get_eb_vm(ctx);
2429 args->value = vm->total;
2430 i915_vm_put(vm);
2431
2432 break;
2433
2434 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
2435 args->size = 0;
2436 args->value = i915_gem_context_no_error_capture(ctx);
2437 break;
2438
2439 case I915_CONTEXT_PARAM_BANNABLE:
2440 args->size = 0;
2441 args->value = i915_gem_context_is_bannable(ctx);
2442 break;
2443
2444 case I915_CONTEXT_PARAM_RECOVERABLE:
2445 args->size = 0;
2446 args->value = i915_gem_context_is_recoverable(ctx);
2447 break;
2448
2449 case I915_CONTEXT_PARAM_PRIORITY:
2450 args->size = 0;
2451 args->value = ctx->sched.priority;
2452 break;
2453
2454 case I915_CONTEXT_PARAM_SSEU:
2455 ret = get_sseu(ctx, args);
2456 break;
2457
2458 case I915_CONTEXT_PARAM_VM:
2459 ret = get_ppgtt(file_priv, ctx, args);
2460 break;
2461
2462 case I915_CONTEXT_PARAM_PERSISTENCE:
2463 args->size = 0;
2464 args->value = i915_gem_context_is_persistent(ctx);
2465 break;
2466
2467 case I915_CONTEXT_PARAM_PROTECTED_CONTENT:
2468 ret = get_protected(ctx, args);
2469 break;
2470
2471 case I915_CONTEXT_PARAM_NO_ZEROMAP:
2472 case I915_CONTEXT_PARAM_BAN_PERIOD:
2473 case I915_CONTEXT_PARAM_ENGINES:
2474 case I915_CONTEXT_PARAM_RINGSIZE:
2475 default:
2476 ret = -EINVAL;
2477 break;
2478 }
2479
2480 i915_gem_context_put(ctx);
2481 return ret;
2482 }
2483
i915_gem_context_setparam_ioctl(struct drm_device * dev,void * data,struct drm_file * file)2484 int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
2485 struct drm_file *file)
2486 {
2487 struct drm_i915_file_private *file_priv = file->driver_priv;
2488 struct drm_i915_gem_context_param *args = data;
2489 struct i915_gem_proto_context *pc;
2490 struct i915_gem_context *ctx;
2491 int ret = 0;
2492
2493 mutex_lock(&file_priv->proto_context_lock);
2494 ctx = __context_lookup(file_priv, args->ctx_id);
2495 if (!ctx) {
2496 pc = xa_load(&file_priv->proto_context_xa, args->ctx_id);
2497 if (pc) {
2498 /* Contexts should be finalized inside
2499 * GEM_CONTEXT_CREATE starting with graphics
2500 * version 13.
2501 */
2502 WARN_ON(GRAPHICS_VER(file_priv->dev_priv) > 12);
2503 ret = set_proto_ctx_param(file_priv, pc, args);
2504 } else {
2505 ret = -ENOENT;
2506 }
2507 }
2508 mutex_unlock(&file_priv->proto_context_lock);
2509
2510 if (ctx) {
2511 ret = ctx_setparam(file_priv, ctx, args);
2512 i915_gem_context_put(ctx);
2513 }
2514
2515 return ret;
2516 }
2517
i915_gem_context_reset_stats_ioctl(struct drm_device * dev,void * data,struct drm_file * file)2518 int i915_gem_context_reset_stats_ioctl(struct drm_device *dev,
2519 void *data, struct drm_file *file)
2520 {
2521 struct drm_i915_private *i915 = to_i915(dev);
2522 struct drm_i915_reset_stats *args = data;
2523 struct i915_gem_context *ctx;
2524
2525 if (args->flags || args->pad)
2526 return -EINVAL;
2527
2528 ctx = i915_gem_context_lookup(file->driver_priv, args->ctx_id);
2529 if (IS_ERR(ctx))
2530 return PTR_ERR(ctx);
2531
2532 /*
2533 * We opt for unserialised reads here. This may result in tearing
2534 * in the extremely unlikely event of a GPU hang on this context
2535 * as we are querying them. If we need that extra layer of protection,
2536 * we should wrap the hangstats with a seqlock.
2537 */
2538
2539 if (capable(CAP_SYS_ADMIN))
2540 args->reset_count = i915_reset_count(&i915->gpu_error);
2541 else
2542 args->reset_count = 0;
2543
2544 args->batch_active = atomic_read(&ctx->guilty_count);
2545 args->batch_pending = atomic_read(&ctx->active_count);
2546
2547 i915_gem_context_put(ctx);
2548 return 0;
2549 }
2550
2551 /* GEM context-engines iterator: for_each_gem_engine() */
2552 struct intel_context *
i915_gem_engines_iter_next(struct i915_gem_engines_iter * it)2553 i915_gem_engines_iter_next(struct i915_gem_engines_iter *it)
2554 {
2555 const struct i915_gem_engines *e = it->engines;
2556 struct intel_context *ctx;
2557
2558 if (unlikely(!e))
2559 return NULL;
2560
2561 do {
2562 if (it->idx >= e->num_engines)
2563 return NULL;
2564
2565 ctx = e->engines[it->idx++];
2566 } while (!ctx);
2567
2568 return ctx;
2569 }
2570
2571 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2572 #include "selftests/mock_context.c"
2573 #include "selftests/i915_gem_context.c"
2574 #endif
2575
i915_gem_context_module_exit(void)2576 void i915_gem_context_module_exit(void)
2577 {
2578 kmem_cache_destroy(slab_luts);
2579 }
2580
i915_gem_context_module_init(void)2581 int __init i915_gem_context_module_init(void)
2582 {
2583 slab_luts = KMEM_CACHE(i915_lut_handle, 0);
2584 if (!slab_luts)
2585 return -ENOMEM;
2586
2587 return 0;
2588 }
2589