xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c (revision 3e93d5bbcbfc3808f83712c0701f9d4c148cc8ed)
1 /*
2  * Copyright 2008 Advanced Micro Devices, Inc.
3  * Copyright 2008 Red Hat Inc.
4  * Copyright 2009 Jerome Glisse.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  *
24  * Authors: Dave Airlie
25  *          Alex Deucher
26  *          Jerome Glisse
27  */
28 
29 #include <linux/dma-fence-array.h>
30 #include <linux/interval_tree_generic.h>
31 #include <linux/idr.h>
32 #include <linux/dma-buf.h>
33 
34 #include <drm/amdgpu_drm.h>
35 #include <drm/drm_drv.h>
36 #include <drm/ttm/ttm_tt.h>
37 #include <drm/drm_exec.h>
38 #include "amdgpu.h"
39 #include "amdgpu_vm.h"
40 #include "amdgpu_trace.h"
41 #include "amdgpu_amdkfd.h"
42 #include "amdgpu_gmc.h"
43 #include "amdgpu_xgmi.h"
44 #include "amdgpu_dma_buf.h"
45 #include "amdgpu_res_cursor.h"
46 #include "kfd_svm.h"
47 
48 /**
49  * DOC: GPUVM
50  *
51  * GPUVM is the MMU functionality provided on the GPU.
52  * GPUVM is similar to the legacy GART on older asics, however
53  * rather than there being a single global GART table
54  * for the entire GPU, there can be multiple GPUVM page tables active
55  * at any given time.  The GPUVM page tables can contain a mix
56  * VRAM pages and system pages (both memory and MMIO) and system pages
57  * can be mapped as snooped (cached system pages) or unsnooped
58  * (uncached system pages).
59  *
60  * Each active GPUVM has an ID associated with it and there is a page table
61  * linked with each VMID.  When executing a command buffer,
62  * the kernel tells the engine what VMID to use for that command
63  * buffer.  VMIDs are allocated dynamically as commands are submitted.
64  * The userspace drivers maintain their own address space and the kernel
65  * sets up their pages tables accordingly when they submit their
66  * command buffers and a VMID is assigned.
67  * The hardware supports up to 16 active GPUVMs at any given time.
68  *
69  * Each GPUVM is represented by a 1-2 or 1-5 level page table, depending
70  * on the ASIC family.  GPUVM supports RWX attributes on each page as well
71  * as other features such as encryption and caching attributes.
72  *
73  * VMID 0 is special.  It is the GPUVM used for the kernel driver.  In
74  * addition to an aperture managed by a page table, VMID 0 also has
75  * several other apertures.  There is an aperture for direct access to VRAM
76  * and there is a legacy AGP aperture which just forwards accesses directly
77  * to the matching system physical addresses (or IOVAs when an IOMMU is
78  * present).  These apertures provide direct access to these memories without
79  * incurring the overhead of a page table.  VMID 0 is used by the kernel
80  * driver for tasks like memory management.
81  *
82  * GPU clients (i.e., engines on the GPU) use GPUVM VMIDs to access memory.
83  * For user applications, each application can have their own unique GPUVM
84  * address space.  The application manages the address space and the kernel
85  * driver manages the GPUVM page tables for each process.  If an GPU client
86  * accesses an invalid page, it will generate a GPU page fault, similar to
87  * accessing an invalid page on a CPU.
88  */
89 
90 #define START(node) ((node)->start)
91 #define LAST(node) ((node)->last)
92 
93 INTERVAL_TREE_DEFINE(struct amdgpu_bo_va_mapping, rb, uint64_t, __subtree_last,
94 		     START, LAST, static, amdgpu_vm_it)
95 
96 #undef START
97 #undef LAST
98 
99 /**
100  * struct amdgpu_prt_cb - Helper to disable partial resident texture feature from a fence callback
101  */
102 struct amdgpu_prt_cb {
103 
104 	/**
105 	 * @adev: amdgpu device
106 	 */
107 	struct amdgpu_device *adev;
108 
109 	/**
110 	 * @cb: callback
111 	 */
112 	struct dma_fence_cb cb;
113 };
114 
115 /**
116  * struct amdgpu_vm_tlb_seq_struct - Helper to increment the TLB flush sequence
117  */
118 struct amdgpu_vm_tlb_seq_struct {
119 	/**
120 	 * @vm: pointer to the amdgpu_vm structure to set the fence sequence on
121 	 */
122 	struct amdgpu_vm *vm;
123 
124 	/**
125 	 * @cb: callback
126 	 */
127 	struct dma_fence_cb cb;
128 };
129 
130 /**
131  * amdgpu_vm_set_pasid - manage pasid and vm ptr mapping
132  *
133  * @adev: amdgpu_device pointer
134  * @vm: amdgpu_vm pointer
135  * @pasid: the pasid the VM is using on this GPU
136  *
137  * Set the pasid this VM is using on this GPU, can also be used to remove the
138  * pasid by passing in zero.
139  *
140  */
amdgpu_vm_set_pasid(struct amdgpu_device * adev,struct amdgpu_vm * vm,u32 pasid)141 int amdgpu_vm_set_pasid(struct amdgpu_device *adev, struct amdgpu_vm *vm,
142 			u32 pasid)
143 {
144 	int r;
145 
146 	if (vm->pasid == pasid)
147 		return 0;
148 
149 	if (vm->pasid) {
150 		r = xa_err(xa_erase_irq(&adev->vm_manager.pasids, vm->pasid));
151 		if (r < 0)
152 			return r;
153 
154 		vm->pasid = 0;
155 	}
156 
157 	if (pasid) {
158 		r = xa_err(xa_store_irq(&adev->vm_manager.pasids, pasid, vm,
159 					GFP_KERNEL));
160 		if (r < 0)
161 			return r;
162 
163 		vm->pasid = pasid;
164 	}
165 
166 
167 	return 0;
168 }
169 
170 /**
171  * amdgpu_vm_bo_evicted - vm_bo is evicted
172  *
173  * @vm_bo: vm_bo which is evicted
174  *
175  * State for PDs/PTs and per VM BOs which are not at the location they should
176  * be.
177  */
amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base * vm_bo)178 static void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)
179 {
180 	struct amdgpu_vm *vm = vm_bo->vm;
181 	struct amdgpu_bo *bo = vm_bo->bo;
182 
183 	vm_bo->moved = true;
184 	spin_lock(&vm_bo->vm->status_lock);
185 	if (bo->tbo.type == ttm_bo_type_kernel)
186 		list_move(&vm_bo->vm_status, &vm->evicted);
187 	else
188 		list_move_tail(&vm_bo->vm_status, &vm->evicted);
189 	spin_unlock(&vm_bo->vm->status_lock);
190 }
191 /**
192  * amdgpu_vm_bo_moved - vm_bo is moved
193  *
194  * @vm_bo: vm_bo which is moved
195  *
196  * State for per VM BOs which are moved, but that change is not yet reflected
197  * in the page tables.
198  */
amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base * vm_bo)199 static void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)
200 {
201 	spin_lock(&vm_bo->vm->status_lock);
202 	list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
203 	spin_unlock(&vm_bo->vm->status_lock);
204 }
205 
206 /**
207  * amdgpu_vm_bo_idle - vm_bo is idle
208  *
209  * @vm_bo: vm_bo which is now idle
210  *
211  * State for PDs/PTs and per VM BOs which have gone through the state machine
212  * and are now idle.
213  */
amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base * vm_bo)214 static void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)
215 {
216 	spin_lock(&vm_bo->vm->status_lock);
217 	list_move(&vm_bo->vm_status, &vm_bo->vm->idle);
218 	spin_unlock(&vm_bo->vm->status_lock);
219 	vm_bo->moved = false;
220 }
221 
222 /**
223  * amdgpu_vm_bo_invalidated - vm_bo is invalidated
224  *
225  * @vm_bo: vm_bo which is now invalidated
226  *
227  * State for normal BOs which are invalidated and that change not yet reflected
228  * in the PTs.
229  */
amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base * vm_bo)230 static void amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base *vm_bo)
231 {
232 	spin_lock(&vm_bo->vm->status_lock);
233 	list_move(&vm_bo->vm_status, &vm_bo->vm->invalidated);
234 	spin_unlock(&vm_bo->vm->status_lock);
235 }
236 
237 /**
238  * amdgpu_vm_bo_evicted_user - vm_bo is evicted
239  *
240  * @vm_bo: vm_bo which is evicted
241  *
242  * State for BOs used by user mode queues which are not at the location they
243  * should be.
244  */
amdgpu_vm_bo_evicted_user(struct amdgpu_vm_bo_base * vm_bo)245 static void amdgpu_vm_bo_evicted_user(struct amdgpu_vm_bo_base *vm_bo)
246 {
247 	vm_bo->moved = true;
248 	spin_lock(&vm_bo->vm->status_lock);
249 	list_move(&vm_bo->vm_status, &vm_bo->vm->evicted_user);
250 	spin_unlock(&vm_bo->vm->status_lock);
251 }
252 
253 /**
254  * amdgpu_vm_bo_relocated - vm_bo is reloacted
255  *
256  * @vm_bo: vm_bo which is relocated
257  *
258  * State for PDs/PTs which needs to update their parent PD.
259  * For the root PD, just move to idle state.
260  */
amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base * vm_bo)261 static void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)
262 {
263 	if (vm_bo->bo->parent) {
264 		spin_lock(&vm_bo->vm->status_lock);
265 		list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
266 		spin_unlock(&vm_bo->vm->status_lock);
267 	} else {
268 		amdgpu_vm_bo_idle(vm_bo);
269 	}
270 }
271 
272 /**
273  * amdgpu_vm_bo_done - vm_bo is done
274  *
275  * @vm_bo: vm_bo which is now done
276  *
277  * State for normal BOs which are invalidated and that change has been updated
278  * in the PTs.
279  */
amdgpu_vm_bo_done(struct amdgpu_vm_bo_base * vm_bo)280 static void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)
281 {
282 	spin_lock(&vm_bo->vm->status_lock);
283 	list_move(&vm_bo->vm_status, &vm_bo->vm->done);
284 	spin_unlock(&vm_bo->vm->status_lock);
285 }
286 
287 /**
288  * amdgpu_vm_bo_reset_state_machine - reset the vm_bo state machine
289  * @vm: the VM which state machine to reset
290  *
291  * Move all vm_bo object in the VM into a state where they will be updated
292  * again during validation.
293  */
amdgpu_vm_bo_reset_state_machine(struct amdgpu_vm * vm)294 static void amdgpu_vm_bo_reset_state_machine(struct amdgpu_vm *vm)
295 {
296 	struct amdgpu_vm_bo_base *vm_bo, *tmp;
297 
298 	spin_lock(&vm->status_lock);
299 	list_splice_init(&vm->done, &vm->invalidated);
300 	list_for_each_entry(vm_bo, &vm->invalidated, vm_status)
301 		vm_bo->moved = true;
302 	list_for_each_entry_safe(vm_bo, tmp, &vm->idle, vm_status) {
303 		struct amdgpu_bo *bo = vm_bo->bo;
304 
305 		vm_bo->moved = true;
306 		if (!bo || bo->tbo.type != ttm_bo_type_kernel)
307 			list_move(&vm_bo->vm_status, &vm_bo->vm->moved);
308 		else if (bo->parent)
309 			list_move(&vm_bo->vm_status, &vm_bo->vm->relocated);
310 	}
311 	spin_unlock(&vm->status_lock);
312 }
313 
314 /**
315  * amdgpu_vm_update_shared - helper to update shared memory stat
316  * @base: base structure for tracking BO usage in a VM
317  *
318  * Takes the vm status_lock and updates the shared memory stat. If the basic
319  * stat changed (e.g. buffer was moved) amdgpu_vm_update_stats need to be called
320  * as well.
321  */
amdgpu_vm_update_shared(struct amdgpu_vm_bo_base * base)322 static void amdgpu_vm_update_shared(struct amdgpu_vm_bo_base *base)
323 {
324 	struct amdgpu_vm *vm = base->vm;
325 	struct amdgpu_bo *bo = base->bo;
326 	uint64_t size = amdgpu_bo_size(bo);
327 	uint32_t bo_memtype = amdgpu_bo_mem_stats_placement(bo);
328 	bool shared;
329 
330 	spin_lock(&vm->status_lock);
331 	shared = drm_gem_object_is_shared_for_memory_stats(&bo->tbo.base);
332 	if (base->shared != shared) {
333 		base->shared = shared;
334 		if (shared) {
335 			vm->stats[bo_memtype].drm.shared += size;
336 			vm->stats[bo_memtype].drm.private -= size;
337 		} else {
338 			vm->stats[bo_memtype].drm.shared -= size;
339 			vm->stats[bo_memtype].drm.private += size;
340 		}
341 	}
342 	spin_unlock(&vm->status_lock);
343 }
344 
345 /**
346  * amdgpu_vm_bo_update_shared - callback when bo gets shared/unshared
347  * @bo: amdgpu buffer object
348  *
349  * Update the per VM stats for all the vm if needed from private to shared or
350  * vice versa.
351  */
amdgpu_vm_bo_update_shared(struct amdgpu_bo * bo)352 void amdgpu_vm_bo_update_shared(struct amdgpu_bo *bo)
353 {
354 	struct amdgpu_vm_bo_base *base;
355 
356 	for (base = bo->vm_bo; base; base = base->next)
357 		amdgpu_vm_update_shared(base);
358 }
359 
360 /**
361  * amdgpu_vm_update_stats_locked - helper to update normal memory stat
362  * @base: base structure for tracking BO usage in a VM
363  * @res:  the ttm_resource to use for the purpose of accounting, may or may not
364  *        be bo->tbo.resource
365  * @sign: if we should add (+1) or subtract (-1) from the stat
366  *
367  * Caller need to have the vm status_lock held. Useful for when multiple update
368  * need to happen at the same time.
369  */
amdgpu_vm_update_stats_locked(struct amdgpu_vm_bo_base * base,struct ttm_resource * res,int sign)370 static void amdgpu_vm_update_stats_locked(struct amdgpu_vm_bo_base *base,
371 			    struct ttm_resource *res, int sign)
372 {
373 	struct amdgpu_vm *vm = base->vm;
374 	struct amdgpu_bo *bo = base->bo;
375 	int64_t size = sign * amdgpu_bo_size(bo);
376 	uint32_t bo_memtype = amdgpu_bo_mem_stats_placement(bo);
377 
378 	/* For drm-total- and drm-shared-, BO are accounted by their preferred
379 	 * placement, see also amdgpu_bo_mem_stats_placement.
380 	 */
381 	if (base->shared)
382 		vm->stats[bo_memtype].drm.shared += size;
383 	else
384 		vm->stats[bo_memtype].drm.private += size;
385 
386 	if (res && res->mem_type < __AMDGPU_PL_NUM) {
387 		uint32_t res_memtype = res->mem_type;
388 
389 		vm->stats[res_memtype].drm.resident += size;
390 		/* BO only count as purgeable if it is resident,
391 		 * since otherwise there's nothing to purge.
392 		 */
393 		if (bo->flags & AMDGPU_GEM_CREATE_DISCARDABLE)
394 			vm->stats[res_memtype].drm.purgeable += size;
395 		if (!(bo->preferred_domains & amdgpu_mem_type_to_domain(res_memtype)))
396 			vm->stats[bo_memtype].evicted += size;
397 	}
398 }
399 
400 /**
401  * amdgpu_vm_update_stats - helper to update normal memory stat
402  * @base: base structure for tracking BO usage in a VM
403  * @res:  the ttm_resource to use for the purpose of accounting, may or may not
404  *        be bo->tbo.resource
405  * @sign: if we should add (+1) or subtract (-1) from the stat
406  *
407  * Updates the basic memory stat when bo is added/deleted/moved.
408  */
amdgpu_vm_update_stats(struct amdgpu_vm_bo_base * base,struct ttm_resource * res,int sign)409 void amdgpu_vm_update_stats(struct amdgpu_vm_bo_base *base,
410 			    struct ttm_resource *res, int sign)
411 {
412 	struct amdgpu_vm *vm = base->vm;
413 
414 	spin_lock(&vm->status_lock);
415 	amdgpu_vm_update_stats_locked(base, res, sign);
416 	spin_unlock(&vm->status_lock);
417 }
418 
419 /**
420  * amdgpu_vm_bo_base_init - Adds bo to the list of bos associated with the vm
421  *
422  * @base: base structure for tracking BO usage in a VM
423  * @vm: vm to which bo is to be added
424  * @bo: amdgpu buffer object
425  *
426  * Initialize a bo_va_base structure and add it to the appropriate lists
427  *
428  */
amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base * base,struct amdgpu_vm * vm,struct amdgpu_bo * bo)429 void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base,
430 			    struct amdgpu_vm *vm, struct amdgpu_bo *bo)
431 {
432 	base->vm = vm;
433 	base->bo = bo;
434 	base->next = NULL;
435 	INIT_LIST_HEAD(&base->vm_status);
436 
437 	if (!bo)
438 		return;
439 	base->next = bo->vm_bo;
440 	bo->vm_bo = base;
441 
442 	spin_lock(&vm->status_lock);
443 	base->shared = drm_gem_object_is_shared_for_memory_stats(&bo->tbo.base);
444 	amdgpu_vm_update_stats_locked(base, bo->tbo.resource, +1);
445 	spin_unlock(&vm->status_lock);
446 
447 	if (!amdgpu_vm_is_bo_always_valid(vm, bo))
448 		return;
449 
450 	dma_resv_assert_held(vm->root.bo->tbo.base.resv);
451 
452 	ttm_bo_set_bulk_move(&bo->tbo, &vm->lru_bulk_move);
453 	if (bo->tbo.type == ttm_bo_type_kernel && bo->parent)
454 		amdgpu_vm_bo_relocated(base);
455 	else
456 		amdgpu_vm_bo_idle(base);
457 
458 	if (bo->preferred_domains &
459 	    amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type))
460 		return;
461 
462 	/*
463 	 * we checked all the prerequisites, but it looks like this per vm bo
464 	 * is currently evicted. add the bo to the evicted list to make sure it
465 	 * is validated on next vm use to avoid fault.
466 	 * */
467 	amdgpu_vm_bo_evicted(base);
468 }
469 
470 /**
471  * amdgpu_vm_lock_pd - lock PD in drm_exec
472  *
473  * @vm: vm providing the BOs
474  * @exec: drm execution context
475  * @num_fences: number of extra fences to reserve
476  *
477  * Lock the VM root PD in the DRM execution context.
478  */
amdgpu_vm_lock_pd(struct amdgpu_vm * vm,struct drm_exec * exec,unsigned int num_fences)479 int amdgpu_vm_lock_pd(struct amdgpu_vm *vm, struct drm_exec *exec,
480 		      unsigned int num_fences)
481 {
482 	/* We need at least two fences for the VM PD/PT updates */
483 	return drm_exec_prepare_obj(exec, &vm->root.bo->tbo.base,
484 				    2 + num_fences);
485 }
486 
487 /**
488  * amdgpu_vm_move_to_lru_tail - move all BOs to the end of LRU
489  *
490  * @adev: amdgpu device pointer
491  * @vm: vm providing the BOs
492  *
493  * Move all BOs to the end of LRU and remember their positions to put them
494  * together.
495  */
amdgpu_vm_move_to_lru_tail(struct amdgpu_device * adev,struct amdgpu_vm * vm)496 void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
497 				struct amdgpu_vm *vm)
498 {
499 	spin_lock(&adev->mman.bdev.lru_lock);
500 	ttm_lru_bulk_move_tail(&vm->lru_bulk_move);
501 	spin_unlock(&adev->mman.bdev.lru_lock);
502 }
503 
504 /* Create scheduler entities for page table updates */
amdgpu_vm_init_entities(struct amdgpu_device * adev,struct amdgpu_vm * vm)505 static int amdgpu_vm_init_entities(struct amdgpu_device *adev,
506 				   struct amdgpu_vm *vm)
507 {
508 	int r;
509 
510 	r = drm_sched_entity_init(&vm->immediate, DRM_SCHED_PRIORITY_NORMAL,
511 				  adev->vm_manager.vm_pte_scheds,
512 				  adev->vm_manager.vm_pte_num_scheds, NULL);
513 	if (r)
514 		goto error;
515 
516 	return drm_sched_entity_init(&vm->delayed, DRM_SCHED_PRIORITY_NORMAL,
517 				     adev->vm_manager.vm_pte_scheds,
518 				     adev->vm_manager.vm_pte_num_scheds, NULL);
519 
520 error:
521 	drm_sched_entity_destroy(&vm->immediate);
522 	return r;
523 }
524 
525 /* Destroy the entities for page table updates again */
amdgpu_vm_fini_entities(struct amdgpu_vm * vm)526 static void amdgpu_vm_fini_entities(struct amdgpu_vm *vm)
527 {
528 	drm_sched_entity_destroy(&vm->immediate);
529 	drm_sched_entity_destroy(&vm->delayed);
530 }
531 
532 /**
533  * amdgpu_vm_generation - return the page table re-generation counter
534  * @adev: the amdgpu_device
535  * @vm: optional VM to check, might be NULL
536  *
537  * Returns a page table re-generation token to allow checking if submissions
538  * are still valid to use this VM. The VM parameter might be NULL in which case
539  * just the VRAM lost counter will be used.
540  */
amdgpu_vm_generation(struct amdgpu_device * adev,struct amdgpu_vm * vm)541 uint64_t amdgpu_vm_generation(struct amdgpu_device *adev, struct amdgpu_vm *vm)
542 {
543 	uint64_t result = (u64)atomic_read(&adev->vram_lost_counter) << 32;
544 
545 	if (!vm)
546 		return result;
547 
548 	result += lower_32_bits(vm->generation);
549 	/* Add one if the page tables will be re-generated on next CS */
550 	if (drm_sched_entity_error(&vm->delayed))
551 		++result;
552 
553 	return result;
554 }
555 
556 /**
557  * amdgpu_vm_validate - validate evicted BOs tracked in the VM
558  *
559  * @adev: amdgpu device pointer
560  * @vm: vm providing the BOs
561  * @ticket: optional reservation ticket used to reserve the VM
562  * @validate: callback to do the validation
563  * @param: parameter for the validation callback
564  *
565  * Validate the page table BOs and per-VM BOs on command submission if
566  * necessary. If a ticket is given, also try to validate evicted user queue
567  * BOs. They must already be reserved with the given ticket.
568  *
569  * Returns:
570  * Validation result.
571  */
amdgpu_vm_validate(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct ww_acquire_ctx * ticket,int (* validate)(void * p,struct amdgpu_bo * bo),void * param)572 int amdgpu_vm_validate(struct amdgpu_device *adev, struct amdgpu_vm *vm,
573 		       struct ww_acquire_ctx *ticket,
574 		       int (*validate)(void *p, struct amdgpu_bo *bo),
575 		       void *param)
576 {
577 	uint64_t new_vm_generation = amdgpu_vm_generation(adev, vm);
578 	struct amdgpu_vm_bo_base *bo_base;
579 	struct amdgpu_bo *bo;
580 	int r;
581 
582 	if (vm->generation != new_vm_generation) {
583 		vm->generation = new_vm_generation;
584 		amdgpu_vm_bo_reset_state_machine(vm);
585 		amdgpu_vm_fini_entities(vm);
586 		r = amdgpu_vm_init_entities(adev, vm);
587 		if (r)
588 			return r;
589 	}
590 
591 	spin_lock(&vm->status_lock);
592 	while (!list_empty(&vm->evicted)) {
593 		bo_base = list_first_entry(&vm->evicted,
594 					   struct amdgpu_vm_bo_base,
595 					   vm_status);
596 		spin_unlock(&vm->status_lock);
597 
598 		bo = bo_base->bo;
599 
600 		r = validate(param, bo);
601 		if (r)
602 			return r;
603 
604 		if (bo->tbo.type != ttm_bo_type_kernel) {
605 			amdgpu_vm_bo_moved(bo_base);
606 		} else {
607 			vm->update_funcs->map_table(to_amdgpu_bo_vm(bo));
608 			amdgpu_vm_bo_relocated(bo_base);
609 		}
610 		spin_lock(&vm->status_lock);
611 	}
612 	while (ticket && !list_empty(&vm->evicted_user)) {
613 		bo_base = list_first_entry(&vm->evicted_user,
614 					   struct amdgpu_vm_bo_base,
615 					   vm_status);
616 		spin_unlock(&vm->status_lock);
617 
618 		bo = bo_base->bo;
619 
620 		if (dma_resv_locking_ctx(bo->tbo.base.resv) != ticket) {
621 			struct amdgpu_task_info *ti = amdgpu_vm_get_task_info_vm(vm);
622 
623 			pr_warn_ratelimited("Evicted user BO is not reserved\n");
624 			if (ti) {
625 				pr_warn_ratelimited("pid %d\n", ti->task.pid);
626 				amdgpu_vm_put_task_info(ti);
627 			}
628 
629 			return -EINVAL;
630 		}
631 
632 		r = validate(param, bo);
633 		if (r)
634 			return r;
635 
636 		amdgpu_vm_bo_invalidated(bo_base);
637 
638 		spin_lock(&vm->status_lock);
639 	}
640 	spin_unlock(&vm->status_lock);
641 
642 	amdgpu_vm_eviction_lock(vm);
643 	vm->evicting = false;
644 	amdgpu_vm_eviction_unlock(vm);
645 
646 	return 0;
647 }
648 
649 /**
650  * amdgpu_vm_ready - check VM is ready for updates
651  *
652  * @vm: VM to check
653  *
654  * Check if all VM PDs/PTs are ready for updates
655  *
656  * Returns:
657  * True if VM is not evicting and all VM entities are not stopped
658  */
amdgpu_vm_ready(struct amdgpu_vm * vm)659 bool amdgpu_vm_ready(struct amdgpu_vm *vm)
660 {
661 	bool ret;
662 
663 	amdgpu_vm_eviction_lock(vm);
664 	ret = !vm->evicting;
665 	amdgpu_vm_eviction_unlock(vm);
666 
667 	spin_lock(&vm->status_lock);
668 	ret &= list_empty(&vm->evicted);
669 	spin_unlock(&vm->status_lock);
670 
671 	spin_lock(&vm->immediate.lock);
672 	ret &= !vm->immediate.stopped;
673 	spin_unlock(&vm->immediate.lock);
674 
675 	spin_lock(&vm->delayed.lock);
676 	ret &= !vm->delayed.stopped;
677 	spin_unlock(&vm->delayed.lock);
678 
679 	return ret;
680 }
681 
682 /**
683  * amdgpu_vm_check_compute_bug - check whether asic has compute vm bug
684  *
685  * @adev: amdgpu_device pointer
686  */
amdgpu_vm_check_compute_bug(struct amdgpu_device * adev)687 void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)
688 {
689 	const struct amdgpu_ip_block *ip_block;
690 	bool has_compute_vm_bug;
691 	struct amdgpu_ring *ring;
692 	int i;
693 
694 	has_compute_vm_bug = false;
695 
696 	ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
697 	if (ip_block) {
698 		/* Compute has a VM bug for GFX version < 7.
699 		   Compute has a VM bug for GFX 8 MEC firmware version < 673.*/
700 		if (ip_block->version->major <= 7)
701 			has_compute_vm_bug = true;
702 		else if (ip_block->version->major == 8)
703 			if (adev->gfx.mec_fw_version < 673)
704 				has_compute_vm_bug = true;
705 	}
706 
707 	for (i = 0; i < adev->num_rings; i++) {
708 		ring = adev->rings[i];
709 		if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
710 			/* only compute rings */
711 			ring->has_compute_vm_bug = has_compute_vm_bug;
712 		else
713 			ring->has_compute_vm_bug = false;
714 	}
715 }
716 
717 /**
718  * amdgpu_vm_need_pipeline_sync - Check if pipe sync is needed for job.
719  *
720  * @ring: ring on which the job will be submitted
721  * @job: job to submit
722  *
723  * Returns:
724  * True if sync is needed.
725  */
amdgpu_vm_need_pipeline_sync(struct amdgpu_ring * ring,struct amdgpu_job * job)726 bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring,
727 				  struct amdgpu_job *job)
728 {
729 	struct amdgpu_device *adev = ring->adev;
730 	unsigned vmhub = ring->vm_hub;
731 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
732 
733 	if (job->vmid == 0)
734 		return false;
735 
736 	if (job->vm_needs_flush || ring->has_compute_vm_bug)
737 		return true;
738 
739 	if (ring->funcs->emit_gds_switch && job->gds_switch_needed)
740 		return true;
741 
742 	if (amdgpu_vmid_had_gpu_reset(adev, &id_mgr->ids[job->vmid]))
743 		return true;
744 
745 	return false;
746 }
747 
748 /**
749  * amdgpu_vm_flush - hardware flush the vm
750  *
751  * @ring: ring to use for flush
752  * @job:  related job
753  * @need_pipe_sync: is pipe sync needed
754  *
755  * Emit a VM flush when it is necessary.
756  *
757  * Returns:
758  * 0 on success, errno otherwise.
759  */
amdgpu_vm_flush(struct amdgpu_ring * ring,struct amdgpu_job * job,bool need_pipe_sync)760 int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
761 		    bool need_pipe_sync)
762 {
763 	struct amdgpu_device *adev = ring->adev;
764 	struct amdgpu_isolation *isolation = &adev->isolation[ring->xcp_id];
765 	unsigned vmhub = ring->vm_hub;
766 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
767 	struct amdgpu_vmid *id = &id_mgr->ids[job->vmid];
768 	bool spm_update_needed = job->spm_update_needed;
769 	bool gds_switch_needed = ring->funcs->emit_gds_switch &&
770 		job->gds_switch_needed;
771 	bool vm_flush_needed = job->vm_needs_flush;
772 	bool cleaner_shader_needed = false;
773 	bool pasid_mapping_needed = false;
774 	struct dma_fence *fence = NULL;
775 	struct amdgpu_fence *af;
776 	unsigned int patch;
777 	int r;
778 
779 	if (amdgpu_vmid_had_gpu_reset(adev, id)) {
780 		gds_switch_needed = true;
781 		vm_flush_needed = true;
782 		pasid_mapping_needed = true;
783 		spm_update_needed = true;
784 	}
785 
786 	mutex_lock(&id_mgr->lock);
787 	if (id->pasid != job->pasid || !id->pasid_mapping ||
788 	    !dma_fence_is_signaled(id->pasid_mapping))
789 		pasid_mapping_needed = true;
790 	mutex_unlock(&id_mgr->lock);
791 
792 	gds_switch_needed &= !!ring->funcs->emit_gds_switch;
793 	vm_flush_needed &= !!ring->funcs->emit_vm_flush  &&
794 			job->vm_pd_addr != AMDGPU_BO_INVALID_OFFSET;
795 	pasid_mapping_needed &= adev->gmc.gmc_funcs->emit_pasid_mapping &&
796 		ring->funcs->emit_wreg;
797 
798 	cleaner_shader_needed = job->run_cleaner_shader &&
799 		adev->gfx.enable_cleaner_shader &&
800 		ring->funcs->emit_cleaner_shader && job->base.s_fence &&
801 		&job->base.s_fence->scheduled == isolation->spearhead;
802 
803 	if (!vm_flush_needed && !gds_switch_needed && !need_pipe_sync &&
804 	    !cleaner_shader_needed)
805 		return 0;
806 
807 	amdgpu_ring_ib_begin(ring);
808 	if (ring->funcs->init_cond_exec)
809 		patch = amdgpu_ring_init_cond_exec(ring,
810 						   ring->cond_exe_gpu_addr);
811 
812 	if (need_pipe_sync)
813 		amdgpu_ring_emit_pipeline_sync(ring);
814 
815 	if (cleaner_shader_needed)
816 		ring->funcs->emit_cleaner_shader(ring);
817 
818 	if (vm_flush_needed) {
819 		trace_amdgpu_vm_flush(ring, job->vmid, job->vm_pd_addr);
820 		amdgpu_ring_emit_vm_flush(ring, job->vmid, job->vm_pd_addr);
821 	}
822 
823 	if (pasid_mapping_needed)
824 		amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
825 
826 	if (spm_update_needed && adev->gfx.rlc.funcs->update_spm_vmid)
827 		adev->gfx.rlc.funcs->update_spm_vmid(adev, ring, job->vmid);
828 
829 	if (ring->funcs->emit_gds_switch &&
830 	    gds_switch_needed) {
831 		amdgpu_ring_emit_gds_switch(ring, job->vmid, job->gds_base,
832 					    job->gds_size, job->gws_base,
833 					    job->gws_size, job->oa_base,
834 					    job->oa_size);
835 	}
836 
837 	if (vm_flush_needed || pasid_mapping_needed || cleaner_shader_needed) {
838 		r = amdgpu_fence_emit(ring, &fence, NULL, 0);
839 		if (r)
840 			return r;
841 		/* this is part of the job's context */
842 		af = container_of(fence, struct amdgpu_fence, base);
843 		af->context = job->base.s_fence ? job->base.s_fence->finished.context : 0;
844 	}
845 
846 	if (vm_flush_needed) {
847 		mutex_lock(&id_mgr->lock);
848 		dma_fence_put(id->last_flush);
849 		id->last_flush = dma_fence_get(fence);
850 		id->current_gpu_reset_count =
851 			atomic_read(&adev->gpu_reset_counter);
852 		mutex_unlock(&id_mgr->lock);
853 	}
854 
855 	if (pasid_mapping_needed) {
856 		mutex_lock(&id_mgr->lock);
857 		id->pasid = job->pasid;
858 		dma_fence_put(id->pasid_mapping);
859 		id->pasid_mapping = dma_fence_get(fence);
860 		mutex_unlock(&id_mgr->lock);
861 	}
862 
863 	/*
864 	 * Make sure that all other submissions wait for the cleaner shader to
865 	 * finish before we push them to the HW.
866 	 */
867 	if (cleaner_shader_needed) {
868 		trace_amdgpu_cleaner_shader(ring, fence);
869 		mutex_lock(&adev->enforce_isolation_mutex);
870 		dma_fence_put(isolation->spearhead);
871 		isolation->spearhead = dma_fence_get(fence);
872 		mutex_unlock(&adev->enforce_isolation_mutex);
873 	}
874 	dma_fence_put(fence);
875 
876 	amdgpu_ring_patch_cond_exec(ring, patch);
877 
878 	/* the double SWITCH_BUFFER here *cannot* be skipped by COND_EXEC */
879 	if (ring->funcs->emit_switch_buffer) {
880 		amdgpu_ring_emit_switch_buffer(ring);
881 		amdgpu_ring_emit_switch_buffer(ring);
882 	}
883 
884 	amdgpu_ring_ib_end(ring);
885 	return 0;
886 }
887 
888 /**
889  * amdgpu_vm_bo_find - find the bo_va for a specific vm & bo
890  *
891  * @vm: requested vm
892  * @bo: requested buffer object
893  *
894  * Find @bo inside the requested vm.
895  * Search inside the @bos vm list for the requested vm
896  * Returns the found bo_va or NULL if none is found
897  *
898  * Object has to be reserved!
899  *
900  * Returns:
901  * Found bo_va or NULL.
902  */
amdgpu_vm_bo_find(struct amdgpu_vm * vm,struct amdgpu_bo * bo)903 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
904 				       struct amdgpu_bo *bo)
905 {
906 	struct amdgpu_vm_bo_base *base;
907 
908 	for (base = bo->vm_bo; base; base = base->next) {
909 		if (base->vm != vm)
910 			continue;
911 
912 		return container_of(base, struct amdgpu_bo_va, base);
913 	}
914 	return NULL;
915 }
916 
917 /**
918  * amdgpu_vm_map_gart - Resolve gart mapping of addr
919  *
920  * @pages_addr: optional DMA address to use for lookup
921  * @addr: the unmapped addr
922  *
923  * Look up the physical address of the page that the pte resolves
924  * to.
925  *
926  * Returns:
927  * The pointer for the page table entry.
928  */
amdgpu_vm_map_gart(const dma_addr_t * pages_addr,uint64_t addr)929 uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
930 {
931 	uint64_t result;
932 
933 	/* page table offset */
934 	result = pages_addr[addr >> PAGE_SHIFT];
935 
936 	/* in case cpu page size != gpu page size*/
937 	result |= addr & (~PAGE_MASK);
938 
939 	result &= 0xFFFFFFFFFFFFF000ULL;
940 
941 	return result;
942 }
943 
944 /**
945  * amdgpu_vm_update_pdes - make sure that all directories are valid
946  *
947  * @adev: amdgpu_device pointer
948  * @vm: requested vm
949  * @immediate: submit immediately to the paging queue
950  *
951  * Makes sure all directories are up to date.
952  *
953  * Returns:
954  * 0 for success, error for failure.
955  */
amdgpu_vm_update_pdes(struct amdgpu_device * adev,struct amdgpu_vm * vm,bool immediate)956 int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
957 			  struct amdgpu_vm *vm, bool immediate)
958 {
959 	struct amdgpu_vm_update_params params;
960 	struct amdgpu_vm_bo_base *entry;
961 	bool flush_tlb_needed = false;
962 	LIST_HEAD(relocated);
963 	int r, idx;
964 
965 	spin_lock(&vm->status_lock);
966 	list_splice_init(&vm->relocated, &relocated);
967 	spin_unlock(&vm->status_lock);
968 
969 	if (list_empty(&relocated))
970 		return 0;
971 
972 	if (!drm_dev_enter(adev_to_drm(adev), &idx))
973 		return -ENODEV;
974 
975 	memset(&params, 0, sizeof(params));
976 	params.adev = adev;
977 	params.vm = vm;
978 	params.immediate = immediate;
979 
980 	r = vm->update_funcs->prepare(&params, NULL);
981 	if (r)
982 		goto error;
983 
984 	list_for_each_entry(entry, &relocated, vm_status) {
985 		/* vm_flush_needed after updating moved PDEs */
986 		flush_tlb_needed |= entry->moved;
987 
988 		r = amdgpu_vm_pde_update(&params, entry);
989 		if (r)
990 			goto error;
991 	}
992 
993 	r = vm->update_funcs->commit(&params, &vm->last_update);
994 	if (r)
995 		goto error;
996 
997 	if (flush_tlb_needed)
998 		atomic64_inc(&vm->tlb_seq);
999 
1000 	while (!list_empty(&relocated)) {
1001 		entry = list_first_entry(&relocated, struct amdgpu_vm_bo_base,
1002 					 vm_status);
1003 		amdgpu_vm_bo_idle(entry);
1004 	}
1005 
1006 error:
1007 	drm_dev_exit(idx);
1008 	return r;
1009 }
1010 
1011 /**
1012  * amdgpu_vm_tlb_seq_cb - make sure to increment tlb sequence
1013  * @fence: unused
1014  * @cb: the callback structure
1015  *
1016  * Increments the tlb sequence to make sure that future CS execute a VM flush.
1017  */
amdgpu_vm_tlb_seq_cb(struct dma_fence * fence,struct dma_fence_cb * cb)1018 static void amdgpu_vm_tlb_seq_cb(struct dma_fence *fence,
1019 				 struct dma_fence_cb *cb)
1020 {
1021 	struct amdgpu_vm_tlb_seq_struct *tlb_cb;
1022 
1023 	tlb_cb = container_of(cb, typeof(*tlb_cb), cb);
1024 	atomic64_inc(&tlb_cb->vm->tlb_seq);
1025 	kfree(tlb_cb);
1026 }
1027 
1028 /**
1029  * amdgpu_vm_tlb_flush - prepare TLB flush
1030  *
1031  * @params: parameters for update
1032  * @fence: input fence to sync TLB flush with
1033  * @tlb_cb: the callback structure
1034  *
1035  * Increments the tlb sequence to make sure that future CS execute a VM flush.
1036  */
1037 static void
amdgpu_vm_tlb_flush(struct amdgpu_vm_update_params * params,struct dma_fence ** fence,struct amdgpu_vm_tlb_seq_struct * tlb_cb)1038 amdgpu_vm_tlb_flush(struct amdgpu_vm_update_params *params,
1039 		    struct dma_fence **fence,
1040 		    struct amdgpu_vm_tlb_seq_struct *tlb_cb)
1041 {
1042 	struct amdgpu_vm *vm = params->vm;
1043 
1044 	tlb_cb->vm = vm;
1045 	if (!fence || !*fence) {
1046 		amdgpu_vm_tlb_seq_cb(NULL, &tlb_cb->cb);
1047 		return;
1048 	}
1049 
1050 	if (!dma_fence_add_callback(*fence, &tlb_cb->cb,
1051 				    amdgpu_vm_tlb_seq_cb)) {
1052 		dma_fence_put(vm->last_tlb_flush);
1053 		vm->last_tlb_flush = dma_fence_get(*fence);
1054 	} else {
1055 		amdgpu_vm_tlb_seq_cb(NULL, &tlb_cb->cb);
1056 	}
1057 
1058 	/* Prepare a TLB flush fence to be attached to PTs */
1059 	if (!params->unlocked && vm->is_compute_context) {
1060 		amdgpu_vm_tlb_fence_create(params->adev, vm, fence);
1061 
1062 		/* Makes sure no PD/PT is freed before the flush */
1063 		dma_resv_add_fence(vm->root.bo->tbo.base.resv, *fence,
1064 				   DMA_RESV_USAGE_BOOKKEEP);
1065 	}
1066 }
1067 
1068 /**
1069  * amdgpu_vm_update_range - update a range in the vm page table
1070  *
1071  * @adev: amdgpu_device pointer to use for commands
1072  * @vm: the VM to update the range
1073  * @immediate: immediate submission in a page fault
1074  * @unlocked: unlocked invalidation during MM callback
1075  * @flush_tlb: trigger tlb invalidation after update completed
1076  * @allow_override: change MTYPE for local NUMA nodes
1077  * @sync: fences we need to sync to
1078  * @start: start of mapped range
1079  * @last: last mapped entry
1080  * @flags: flags for the entries
1081  * @offset: offset into nodes and pages_addr
1082  * @vram_base: base for vram mappings
1083  * @res: ttm_resource to map
1084  * @pages_addr: DMA addresses to use for mapping
1085  * @fence: optional resulting fence
1086  *
1087  * Fill in the page table entries between @start and @last.
1088  *
1089  * Returns:
1090  * 0 for success, negative erro code for failure.
1091  */
amdgpu_vm_update_range(struct amdgpu_device * adev,struct amdgpu_vm * vm,bool immediate,bool unlocked,bool flush_tlb,bool allow_override,struct amdgpu_sync * sync,uint64_t start,uint64_t last,uint64_t flags,uint64_t offset,uint64_t vram_base,struct ttm_resource * res,dma_addr_t * pages_addr,struct dma_fence ** fence)1092 int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
1093 			   bool immediate, bool unlocked, bool flush_tlb,
1094 			   bool allow_override, struct amdgpu_sync *sync,
1095 			   uint64_t start, uint64_t last, uint64_t flags,
1096 			   uint64_t offset, uint64_t vram_base,
1097 			   struct ttm_resource *res, dma_addr_t *pages_addr,
1098 			   struct dma_fence **fence)
1099 {
1100 	struct amdgpu_vm_tlb_seq_struct *tlb_cb;
1101 	struct amdgpu_vm_update_params params;
1102 	struct amdgpu_res_cursor cursor;
1103 	int r, idx;
1104 
1105 	if (!drm_dev_enter(adev_to_drm(adev), &idx))
1106 		return -ENODEV;
1107 
1108 	tlb_cb = kmalloc(sizeof(*tlb_cb), GFP_KERNEL);
1109 	if (!tlb_cb) {
1110 		drm_dev_exit(idx);
1111 		return -ENOMEM;
1112 	}
1113 
1114 	/* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
1115 	 * heavy-weight flush TLB unconditionally.
1116 	 */
1117 	flush_tlb |= adev->gmc.xgmi.num_physical_nodes &&
1118 		     amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 0);
1119 
1120 	/*
1121 	 * On GFX8 and older any 8 PTE block with a valid bit set enters the TLB
1122 	 */
1123 	flush_tlb |= amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 0, 0);
1124 
1125 	memset(&params, 0, sizeof(params));
1126 	params.adev = adev;
1127 	params.vm = vm;
1128 	params.immediate = immediate;
1129 	params.pages_addr = pages_addr;
1130 	params.unlocked = unlocked;
1131 	params.needs_flush = flush_tlb;
1132 	params.allow_override = allow_override;
1133 	INIT_LIST_HEAD(&params.tlb_flush_waitlist);
1134 
1135 	amdgpu_vm_eviction_lock(vm);
1136 	if (vm->evicting) {
1137 		r = -EBUSY;
1138 		goto error_free;
1139 	}
1140 
1141 	if (!unlocked && !dma_fence_is_signaled(vm->last_unlocked)) {
1142 		struct dma_fence *tmp = dma_fence_get_stub();
1143 
1144 		amdgpu_bo_fence(vm->root.bo, vm->last_unlocked, true);
1145 		swap(vm->last_unlocked, tmp);
1146 		dma_fence_put(tmp);
1147 	}
1148 
1149 	r = vm->update_funcs->prepare(&params, sync);
1150 	if (r)
1151 		goto error_free;
1152 
1153 	amdgpu_res_first(pages_addr ? NULL : res, offset,
1154 			 (last - start + 1) * AMDGPU_GPU_PAGE_SIZE, &cursor);
1155 	while (cursor.remaining) {
1156 		uint64_t tmp, num_entries, addr;
1157 
1158 		num_entries = cursor.size >> AMDGPU_GPU_PAGE_SHIFT;
1159 		if (pages_addr) {
1160 			bool contiguous = true;
1161 
1162 			if (num_entries > AMDGPU_GPU_PAGES_IN_CPU_PAGE) {
1163 				uint64_t pfn = cursor.start >> PAGE_SHIFT;
1164 				uint64_t count;
1165 
1166 				contiguous = pages_addr[pfn + 1] ==
1167 					pages_addr[pfn] + PAGE_SIZE;
1168 
1169 				tmp = num_entries /
1170 					AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1171 				for (count = 2; count < tmp; ++count) {
1172 					uint64_t idx = pfn + count;
1173 
1174 					if (contiguous != (pages_addr[idx] ==
1175 					    pages_addr[idx - 1] + PAGE_SIZE))
1176 						break;
1177 				}
1178 				if (!contiguous)
1179 					count--;
1180 				num_entries = count *
1181 					AMDGPU_GPU_PAGES_IN_CPU_PAGE;
1182 			}
1183 
1184 			if (!contiguous) {
1185 				addr = cursor.start;
1186 				params.pages_addr = pages_addr;
1187 			} else {
1188 				addr = pages_addr[cursor.start >> PAGE_SHIFT];
1189 				params.pages_addr = NULL;
1190 			}
1191 
1192 		} else if (flags & (AMDGPU_PTE_VALID | AMDGPU_PTE_PRT_FLAG(adev))) {
1193 			addr = vram_base + cursor.start;
1194 		} else {
1195 			addr = 0;
1196 		}
1197 
1198 		tmp = start + num_entries;
1199 		r = amdgpu_vm_ptes_update(&params, start, tmp, addr, flags);
1200 		if (r)
1201 			goto error_free;
1202 
1203 		amdgpu_res_next(&cursor, num_entries * AMDGPU_GPU_PAGE_SIZE);
1204 		start = tmp;
1205 	}
1206 
1207 	r = vm->update_funcs->commit(&params, fence);
1208 	if (r)
1209 		goto error_free;
1210 
1211 	if (params.needs_flush) {
1212 		amdgpu_vm_tlb_flush(&params, fence, tlb_cb);
1213 		tlb_cb = NULL;
1214 	}
1215 
1216 	amdgpu_vm_pt_free_list(adev, &params);
1217 
1218 error_free:
1219 	kfree(tlb_cb);
1220 	amdgpu_vm_eviction_unlock(vm);
1221 	drm_dev_exit(idx);
1222 	return r;
1223 }
1224 
amdgpu_vm_get_memory(struct amdgpu_vm * vm,struct amdgpu_mem_stats stats[__AMDGPU_PL_NUM])1225 void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
1226 			  struct amdgpu_mem_stats stats[__AMDGPU_PL_NUM])
1227 {
1228 	spin_lock(&vm->status_lock);
1229 	memcpy(stats, vm->stats, sizeof(*stats) * __AMDGPU_PL_NUM);
1230 	spin_unlock(&vm->status_lock);
1231 }
1232 
1233 /**
1234  * amdgpu_vm_bo_update - update all BO mappings in the vm page table
1235  *
1236  * @adev: amdgpu_device pointer
1237  * @bo_va: requested BO and VM object
1238  * @clear: if true clear the entries
1239  *
1240  * Fill in the page table entries for @bo_va.
1241  *
1242  * Returns:
1243  * 0 for success, -EINVAL for failure.
1244  */
amdgpu_vm_bo_update(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va,bool clear)1245 int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va,
1246 			bool clear)
1247 {
1248 	struct amdgpu_bo *bo = bo_va->base.bo;
1249 	struct amdgpu_vm *vm = bo_va->base.vm;
1250 	struct amdgpu_bo_va_mapping *mapping;
1251 	struct dma_fence **last_update;
1252 	dma_addr_t *pages_addr = NULL;
1253 	struct ttm_resource *mem;
1254 	struct amdgpu_sync sync;
1255 	bool flush_tlb = clear;
1256 	uint64_t vram_base;
1257 	uint64_t flags;
1258 	bool uncached;
1259 	int r;
1260 
1261 	amdgpu_sync_create(&sync);
1262 	if (clear) {
1263 		mem = NULL;
1264 
1265 		/* Implicitly sync to command submissions in the same VM before
1266 		 * unmapping.
1267 		 */
1268 		r = amdgpu_sync_resv(adev, &sync, vm->root.bo->tbo.base.resv,
1269 				     AMDGPU_SYNC_EQ_OWNER, vm);
1270 		if (r)
1271 			goto error_free;
1272 		if (bo) {
1273 			r = amdgpu_sync_kfd(&sync, bo->tbo.base.resv);
1274 			if (r)
1275 				goto error_free;
1276 		}
1277 	} else if (!bo) {
1278 		mem = NULL;
1279 
1280 		/* PRT map operations don't need to sync to anything. */
1281 
1282 	} else {
1283 		struct drm_gem_object *obj = &bo->tbo.base;
1284 
1285 		if (drm_gem_is_imported(obj) && bo_va->is_xgmi) {
1286 			struct dma_buf *dma_buf = obj->import_attach->dmabuf;
1287 			struct drm_gem_object *gobj = dma_buf->priv;
1288 			struct amdgpu_bo *abo = gem_to_amdgpu_bo(gobj);
1289 
1290 			if (abo->tbo.resource &&
1291 			    abo->tbo.resource->mem_type == TTM_PL_VRAM)
1292 				bo = gem_to_amdgpu_bo(gobj);
1293 		}
1294 		mem = bo->tbo.resource;
1295 		if (mem && (mem->mem_type == TTM_PL_TT ||
1296 			    mem->mem_type == AMDGPU_PL_PREEMPT))
1297 			pages_addr = bo->tbo.ttm->dma_address;
1298 
1299 		/* Implicitly sync to moving fences before mapping anything */
1300 		r = amdgpu_sync_resv(adev, &sync, bo->tbo.base.resv,
1301 				     AMDGPU_SYNC_EXPLICIT, vm);
1302 		if (r)
1303 			goto error_free;
1304 	}
1305 
1306 	if (bo) {
1307 		struct amdgpu_device *bo_adev;
1308 
1309 		flags = amdgpu_ttm_tt_pte_flags(adev, bo->tbo.ttm, mem);
1310 
1311 		if (amdgpu_bo_encrypted(bo))
1312 			flags |= AMDGPU_PTE_TMZ;
1313 
1314 		bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
1315 		vram_base = bo_adev->vm_manager.vram_base_offset;
1316 		uncached = (bo->flags & AMDGPU_GEM_CREATE_UNCACHED) != 0;
1317 	} else {
1318 		flags = 0x0;
1319 		vram_base = 0;
1320 		uncached = false;
1321 	}
1322 
1323 	if (clear || amdgpu_vm_is_bo_always_valid(vm, bo))
1324 		last_update = &vm->last_update;
1325 	else
1326 		last_update = &bo_va->last_pt_update;
1327 
1328 	if (!clear && bo_va->base.moved) {
1329 		flush_tlb = true;
1330 		list_splice_init(&bo_va->valids, &bo_va->invalids);
1331 
1332 	} else if (bo_va->cleared != clear) {
1333 		list_splice_init(&bo_va->valids, &bo_va->invalids);
1334 	}
1335 
1336 	list_for_each_entry(mapping, &bo_va->invalids, list) {
1337 		uint64_t update_flags = flags;
1338 
1339 		/* normally,bo_va->flags only contians READABLE and WIRTEABLE bit go here
1340 		 * but in case of something, we filter the flags in first place
1341 		 */
1342 		if (!(mapping->flags & AMDGPU_PTE_READABLE))
1343 			update_flags &= ~AMDGPU_PTE_READABLE;
1344 		if (!(mapping->flags & AMDGPU_PTE_WRITEABLE))
1345 			update_flags &= ~AMDGPU_PTE_WRITEABLE;
1346 
1347 		/* Apply ASIC specific mapping flags */
1348 		amdgpu_gmc_get_vm_pte(adev, mapping, &update_flags);
1349 
1350 		trace_amdgpu_vm_bo_update(mapping);
1351 
1352 		r = amdgpu_vm_update_range(adev, vm, false, false, flush_tlb,
1353 					   !uncached, &sync, mapping->start,
1354 					   mapping->last, update_flags,
1355 					   mapping->offset, vram_base, mem,
1356 					   pages_addr, last_update);
1357 		if (r)
1358 			goto error_free;
1359 	}
1360 
1361 	/* If the BO is not in its preferred location add it back to
1362 	 * the evicted list so that it gets validated again on the
1363 	 * next command submission.
1364 	 */
1365 	if (amdgpu_vm_is_bo_always_valid(vm, bo)) {
1366 		if (bo->tbo.resource &&
1367 		    !(bo->preferred_domains &
1368 		      amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type)))
1369 			amdgpu_vm_bo_evicted(&bo_va->base);
1370 		else
1371 			amdgpu_vm_bo_idle(&bo_va->base);
1372 	} else {
1373 		amdgpu_vm_bo_done(&bo_va->base);
1374 	}
1375 
1376 	list_splice_init(&bo_va->invalids, &bo_va->valids);
1377 	bo_va->cleared = clear;
1378 	bo_va->base.moved = false;
1379 
1380 	if (trace_amdgpu_vm_bo_mapping_enabled()) {
1381 		list_for_each_entry(mapping, &bo_va->valids, list)
1382 			trace_amdgpu_vm_bo_mapping(mapping);
1383 	}
1384 
1385 error_free:
1386 	amdgpu_sync_free(&sync);
1387 	return r;
1388 }
1389 
1390 /**
1391  * amdgpu_vm_update_prt_state - update the global PRT state
1392  *
1393  * @adev: amdgpu_device pointer
1394  */
amdgpu_vm_update_prt_state(struct amdgpu_device * adev)1395 static void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)
1396 {
1397 	unsigned long flags;
1398 	bool enable;
1399 
1400 	spin_lock_irqsave(&adev->vm_manager.prt_lock, flags);
1401 	enable = !!atomic_read(&adev->vm_manager.num_prt_users);
1402 	adev->gmc.gmc_funcs->set_prt(adev, enable);
1403 	spin_unlock_irqrestore(&adev->vm_manager.prt_lock, flags);
1404 }
1405 
1406 /**
1407  * amdgpu_vm_prt_get - add a PRT user
1408  *
1409  * @adev: amdgpu_device pointer
1410  */
amdgpu_vm_prt_get(struct amdgpu_device * adev)1411 static void amdgpu_vm_prt_get(struct amdgpu_device *adev)
1412 {
1413 	if (!adev->gmc.gmc_funcs->set_prt)
1414 		return;
1415 
1416 	if (atomic_inc_return(&adev->vm_manager.num_prt_users) == 1)
1417 		amdgpu_vm_update_prt_state(adev);
1418 }
1419 
1420 /**
1421  * amdgpu_vm_prt_put - drop a PRT user
1422  *
1423  * @adev: amdgpu_device pointer
1424  */
amdgpu_vm_prt_put(struct amdgpu_device * adev)1425 static void amdgpu_vm_prt_put(struct amdgpu_device *adev)
1426 {
1427 	if (atomic_dec_return(&adev->vm_manager.num_prt_users) == 0)
1428 		amdgpu_vm_update_prt_state(adev);
1429 }
1430 
1431 /**
1432  * amdgpu_vm_prt_cb - callback for updating the PRT status
1433  *
1434  * @fence: fence for the callback
1435  * @_cb: the callback function
1436  */
amdgpu_vm_prt_cb(struct dma_fence * fence,struct dma_fence_cb * _cb)1437 static void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)
1438 {
1439 	struct amdgpu_prt_cb *cb = container_of(_cb, struct amdgpu_prt_cb, cb);
1440 
1441 	amdgpu_vm_prt_put(cb->adev);
1442 	kfree(cb);
1443 }
1444 
1445 /**
1446  * amdgpu_vm_add_prt_cb - add callback for updating the PRT status
1447  *
1448  * @adev: amdgpu_device pointer
1449  * @fence: fence for the callback
1450  */
amdgpu_vm_add_prt_cb(struct amdgpu_device * adev,struct dma_fence * fence)1451 static void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev,
1452 				 struct dma_fence *fence)
1453 {
1454 	struct amdgpu_prt_cb *cb;
1455 
1456 	if (!adev->gmc.gmc_funcs->set_prt)
1457 		return;
1458 
1459 	cb = kmalloc(sizeof(struct amdgpu_prt_cb), GFP_KERNEL);
1460 	if (!cb) {
1461 		/* Last resort when we are OOM */
1462 		if (fence)
1463 			dma_fence_wait(fence, false);
1464 
1465 		amdgpu_vm_prt_put(adev);
1466 	} else {
1467 		cb->adev = adev;
1468 		if (!fence || dma_fence_add_callback(fence, &cb->cb,
1469 						     amdgpu_vm_prt_cb))
1470 			amdgpu_vm_prt_cb(fence, &cb->cb);
1471 	}
1472 }
1473 
1474 /**
1475  * amdgpu_vm_free_mapping - free a mapping
1476  *
1477  * @adev: amdgpu_device pointer
1478  * @vm: requested vm
1479  * @mapping: mapping to be freed
1480  * @fence: fence of the unmap operation
1481  *
1482  * Free a mapping and make sure we decrease the PRT usage count if applicable.
1483  */
amdgpu_vm_free_mapping(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct amdgpu_bo_va_mapping * mapping,struct dma_fence * fence)1484 static void amdgpu_vm_free_mapping(struct amdgpu_device *adev,
1485 				   struct amdgpu_vm *vm,
1486 				   struct amdgpu_bo_va_mapping *mapping,
1487 				   struct dma_fence *fence)
1488 {
1489 	if (mapping->flags & AMDGPU_PTE_PRT_FLAG(adev))
1490 		amdgpu_vm_add_prt_cb(adev, fence);
1491 	kfree(mapping);
1492 }
1493 
1494 /**
1495  * amdgpu_vm_prt_fini - finish all prt mappings
1496  *
1497  * @adev: amdgpu_device pointer
1498  * @vm: requested vm
1499  *
1500  * Register a cleanup callback to disable PRT support after VM dies.
1501  */
amdgpu_vm_prt_fini(struct amdgpu_device * adev,struct amdgpu_vm * vm)1502 static void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
1503 {
1504 	struct dma_resv *resv = vm->root.bo->tbo.base.resv;
1505 	struct dma_resv_iter cursor;
1506 	struct dma_fence *fence;
1507 
1508 	dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP, fence) {
1509 		/* Add a callback for each fence in the reservation object */
1510 		amdgpu_vm_prt_get(adev);
1511 		amdgpu_vm_add_prt_cb(adev, fence);
1512 	}
1513 }
1514 
1515 /**
1516  * amdgpu_vm_clear_freed - clear freed BOs in the PT
1517  *
1518  * @adev: amdgpu_device pointer
1519  * @vm: requested vm
1520  * @fence: optional resulting fence (unchanged if no work needed to be done
1521  * or if an error occurred)
1522  *
1523  * Make sure all freed BOs are cleared in the PT.
1524  * PTs have to be reserved and mutex must be locked!
1525  *
1526  * Returns:
1527  * 0 for success.
1528  *
1529  */
amdgpu_vm_clear_freed(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct dma_fence ** fence)1530 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
1531 			  struct amdgpu_vm *vm,
1532 			  struct dma_fence **fence)
1533 {
1534 	struct amdgpu_bo_va_mapping *mapping;
1535 	struct dma_fence *f = NULL;
1536 	struct amdgpu_sync sync;
1537 	int r;
1538 
1539 
1540 	/*
1541 	 * Implicitly sync to command submissions in the same VM before
1542 	 * unmapping.
1543 	 */
1544 	amdgpu_sync_create(&sync);
1545 	r = amdgpu_sync_resv(adev, &sync, vm->root.bo->tbo.base.resv,
1546 			     AMDGPU_SYNC_EQ_OWNER, vm);
1547 	if (r)
1548 		goto error_free;
1549 
1550 	while (!list_empty(&vm->freed)) {
1551 		mapping = list_first_entry(&vm->freed,
1552 			struct amdgpu_bo_va_mapping, list);
1553 		list_del(&mapping->list);
1554 
1555 		r = amdgpu_vm_update_range(adev, vm, false, false, true, false,
1556 					   &sync, mapping->start, mapping->last,
1557 					   0, 0, 0, NULL, NULL, &f);
1558 		amdgpu_vm_free_mapping(adev, vm, mapping, f);
1559 		if (r) {
1560 			dma_fence_put(f);
1561 			goto error_free;
1562 		}
1563 	}
1564 
1565 	if (fence && f) {
1566 		dma_fence_put(*fence);
1567 		*fence = f;
1568 	} else {
1569 		dma_fence_put(f);
1570 	}
1571 
1572 error_free:
1573 	amdgpu_sync_free(&sync);
1574 	return r;
1575 
1576 }
1577 
1578 /**
1579  * amdgpu_vm_handle_moved - handle moved BOs in the PT
1580  *
1581  * @adev: amdgpu_device pointer
1582  * @vm: requested vm
1583  * @ticket: optional reservation ticket used to reserve the VM
1584  *
1585  * Make sure all BOs which are moved are updated in the PTs.
1586  *
1587  * Returns:
1588  * 0 for success.
1589  *
1590  * PTs have to be reserved!
1591  */
amdgpu_vm_handle_moved(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct ww_acquire_ctx * ticket)1592 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
1593 			   struct amdgpu_vm *vm,
1594 			   struct ww_acquire_ctx *ticket)
1595 {
1596 	struct amdgpu_bo_va *bo_va;
1597 	struct dma_resv *resv;
1598 	bool clear, unlock;
1599 	int r;
1600 
1601 	spin_lock(&vm->status_lock);
1602 	while (!list_empty(&vm->moved)) {
1603 		bo_va = list_first_entry(&vm->moved, struct amdgpu_bo_va,
1604 					 base.vm_status);
1605 		spin_unlock(&vm->status_lock);
1606 
1607 		/* Per VM BOs never need to bo cleared in the page tables */
1608 		r = amdgpu_vm_bo_update(adev, bo_va, false);
1609 		if (r)
1610 			return r;
1611 		spin_lock(&vm->status_lock);
1612 	}
1613 
1614 	while (!list_empty(&vm->invalidated)) {
1615 		bo_va = list_first_entry(&vm->invalidated, struct amdgpu_bo_va,
1616 					 base.vm_status);
1617 		resv = bo_va->base.bo->tbo.base.resv;
1618 		spin_unlock(&vm->status_lock);
1619 
1620 		/* Try to reserve the BO to avoid clearing its ptes */
1621 		if (!adev->debug_vm && dma_resv_trylock(resv)) {
1622 			clear = false;
1623 			unlock = true;
1624 		/* The caller is already holding the reservation lock */
1625 		} else if (ticket && dma_resv_locking_ctx(resv) == ticket) {
1626 			clear = false;
1627 			unlock = false;
1628 		/* Somebody else is using the BO right now */
1629 		} else {
1630 			clear = true;
1631 			unlock = false;
1632 		}
1633 
1634 		r = amdgpu_vm_bo_update(adev, bo_va, clear);
1635 
1636 		if (unlock)
1637 			dma_resv_unlock(resv);
1638 		if (r)
1639 			return r;
1640 
1641 		/* Remember evicted DMABuf imports in compute VMs for later
1642 		 * validation
1643 		 */
1644 		if (vm->is_compute_context &&
1645 		    drm_gem_is_imported(&bo_va->base.bo->tbo.base) &&
1646 		    (!bo_va->base.bo->tbo.resource ||
1647 		     bo_va->base.bo->tbo.resource->mem_type == TTM_PL_SYSTEM))
1648 			amdgpu_vm_bo_evicted_user(&bo_va->base);
1649 
1650 		spin_lock(&vm->status_lock);
1651 	}
1652 	spin_unlock(&vm->status_lock);
1653 
1654 	return 0;
1655 }
1656 
1657 /**
1658  * amdgpu_vm_flush_compute_tlb - Flush TLB on compute VM
1659  *
1660  * @adev: amdgpu_device pointer
1661  * @vm: requested vm
1662  * @flush_type: flush type
1663  * @xcc_mask: mask of XCCs that belong to the compute partition in need of a TLB flush.
1664  *
1665  * Flush TLB if needed for a compute VM.
1666  *
1667  * Returns:
1668  * 0 for success.
1669  */
amdgpu_vm_flush_compute_tlb(struct amdgpu_device * adev,struct amdgpu_vm * vm,uint32_t flush_type,uint32_t xcc_mask)1670 int amdgpu_vm_flush_compute_tlb(struct amdgpu_device *adev,
1671 				struct amdgpu_vm *vm,
1672 				uint32_t flush_type,
1673 				uint32_t xcc_mask)
1674 {
1675 	uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm);
1676 	bool all_hub = false;
1677 	int xcc = 0, r = 0;
1678 
1679 	WARN_ON_ONCE(!vm->is_compute_context);
1680 
1681 	/*
1682 	 * It can be that we race and lose here, but that is extremely unlikely
1683 	 * and the worst thing which could happen is that we flush the changes
1684 	 * into the TLB once more which is harmless.
1685 	 */
1686 	if (atomic64_xchg(&vm->kfd_last_flushed_seq, tlb_seq) == tlb_seq)
1687 		return 0;
1688 
1689 	if (adev->family == AMDGPU_FAMILY_AI ||
1690 	    adev->family == AMDGPU_FAMILY_RV)
1691 		all_hub = true;
1692 
1693 	for_each_inst(xcc, xcc_mask) {
1694 		r = amdgpu_gmc_flush_gpu_tlb_pasid(adev, vm->pasid, flush_type,
1695 						   all_hub, xcc);
1696 		if (r)
1697 			break;
1698 	}
1699 	return r;
1700 }
1701 
1702 /**
1703  * amdgpu_vm_bo_add - add a bo to a specific vm
1704  *
1705  * @adev: amdgpu_device pointer
1706  * @vm: requested vm
1707  * @bo: amdgpu buffer object
1708  *
1709  * Add @bo into the requested vm.
1710  * Add @bo to the list of bos associated with the vm
1711  *
1712  * Returns:
1713  * Newly added bo_va or NULL for failure
1714  *
1715  * Object has to be reserved!
1716  */
amdgpu_vm_bo_add(struct amdgpu_device * adev,struct amdgpu_vm * vm,struct amdgpu_bo * bo)1717 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
1718 				      struct amdgpu_vm *vm,
1719 				      struct amdgpu_bo *bo)
1720 {
1721 	struct amdgpu_bo_va *bo_va;
1722 
1723 	bo_va = kzalloc(sizeof(struct amdgpu_bo_va), GFP_KERNEL);
1724 	if (bo_va == NULL) {
1725 		return NULL;
1726 	}
1727 	amdgpu_vm_bo_base_init(&bo_va->base, vm, bo);
1728 
1729 	bo_va->ref_count = 1;
1730 	bo_va->last_pt_update = dma_fence_get_stub();
1731 	INIT_LIST_HEAD(&bo_va->valids);
1732 	INIT_LIST_HEAD(&bo_va->invalids);
1733 
1734 	if (!bo)
1735 		return bo_va;
1736 
1737 	dma_resv_assert_held(bo->tbo.base.resv);
1738 	if (amdgpu_dmabuf_is_xgmi_accessible(adev, bo)) {
1739 		bo_va->is_xgmi = true;
1740 		/* Power up XGMI if it can be potentially used */
1741 		amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MAX_VEGA20);
1742 	}
1743 
1744 	return bo_va;
1745 }
1746 
1747 
1748 /**
1749  * amdgpu_vm_bo_insert_map - insert a new mapping
1750  *
1751  * @adev: amdgpu_device pointer
1752  * @bo_va: bo_va to store the address
1753  * @mapping: the mapping to insert
1754  *
1755  * Insert a new mapping into all structures.
1756  */
amdgpu_vm_bo_insert_map(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va,struct amdgpu_bo_va_mapping * mapping)1757 static void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev,
1758 				    struct amdgpu_bo_va *bo_va,
1759 				    struct amdgpu_bo_va_mapping *mapping)
1760 {
1761 	struct amdgpu_vm *vm = bo_va->base.vm;
1762 	struct amdgpu_bo *bo = bo_va->base.bo;
1763 
1764 	mapping->bo_va = bo_va;
1765 	list_add(&mapping->list, &bo_va->invalids);
1766 	amdgpu_vm_it_insert(mapping, &vm->va);
1767 
1768 	if (mapping->flags & AMDGPU_PTE_PRT_FLAG(adev))
1769 		amdgpu_vm_prt_get(adev);
1770 
1771 	if (amdgpu_vm_is_bo_always_valid(vm, bo) && !bo_va->base.moved)
1772 		amdgpu_vm_bo_moved(&bo_va->base);
1773 
1774 	trace_amdgpu_vm_bo_map(bo_va, mapping);
1775 }
1776 
1777 /* Validate operation parameters to prevent potential abuse */
amdgpu_vm_verify_parameters(struct amdgpu_device * adev,struct amdgpu_bo * bo,uint64_t saddr,uint64_t offset,uint64_t size)1778 static int amdgpu_vm_verify_parameters(struct amdgpu_device *adev,
1779 					  struct amdgpu_bo *bo,
1780 					  uint64_t saddr,
1781 					  uint64_t offset,
1782 					  uint64_t size)
1783 {
1784 	uint64_t tmp, lpfn;
1785 
1786 	if (saddr & AMDGPU_GPU_PAGE_MASK
1787 	    || offset & AMDGPU_GPU_PAGE_MASK
1788 	    || size & AMDGPU_GPU_PAGE_MASK)
1789 		return -EINVAL;
1790 
1791 	if (check_add_overflow(saddr, size, &tmp)
1792 	    || check_add_overflow(offset, size, &tmp)
1793 	    || size == 0 /* which also leads to end < begin */)
1794 		return -EINVAL;
1795 
1796 	/* make sure object fit at this offset */
1797 	if (bo && offset + size > amdgpu_bo_size(bo))
1798 		return -EINVAL;
1799 
1800 	/* Ensure last pfn not exceed max_pfn */
1801 	lpfn = (saddr + size - 1) >> AMDGPU_GPU_PAGE_SHIFT;
1802 	if (lpfn >= adev->vm_manager.max_pfn)
1803 		return -EINVAL;
1804 
1805 	return 0;
1806 }
1807 
1808 /**
1809  * amdgpu_vm_bo_map - map bo inside a vm
1810  *
1811  * @adev: amdgpu_device pointer
1812  * @bo_va: bo_va to store the address
1813  * @saddr: where to map the BO
1814  * @offset: requested offset in the BO
1815  * @size: BO size in bytes
1816  * @flags: attributes of pages (read/write/valid/etc.)
1817  *
1818  * Add a mapping of the BO at the specefied addr into the VM.
1819  *
1820  * Returns:
1821  * 0 for success, error for failure.
1822  *
1823  * Object has to be reserved and unreserved outside!
1824  */
amdgpu_vm_bo_map(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va,uint64_t saddr,uint64_t offset,uint64_t size,uint64_t flags)1825 int amdgpu_vm_bo_map(struct amdgpu_device *adev,
1826 		     struct amdgpu_bo_va *bo_va,
1827 		     uint64_t saddr, uint64_t offset,
1828 		     uint64_t size, uint64_t flags)
1829 {
1830 	struct amdgpu_bo_va_mapping *mapping, *tmp;
1831 	struct amdgpu_bo *bo = bo_va->base.bo;
1832 	struct amdgpu_vm *vm = bo_va->base.vm;
1833 	uint64_t eaddr;
1834 	int r;
1835 
1836 	r = amdgpu_vm_verify_parameters(adev, bo, saddr, offset, size);
1837 	if (r)
1838 		return r;
1839 
1840 	saddr /= AMDGPU_GPU_PAGE_SIZE;
1841 	eaddr = saddr + (size - 1) / AMDGPU_GPU_PAGE_SIZE;
1842 
1843 	tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
1844 	if (tmp) {
1845 		/* bo and tmp overlap, invalid addr */
1846 		dev_err(adev->dev, "bo %p va 0x%010Lx-0x%010Lx conflict with "
1847 			"0x%010Lx-0x%010Lx\n", bo, saddr, eaddr,
1848 			tmp->start, tmp->last + 1);
1849 		return -EINVAL;
1850 	}
1851 
1852 	mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1853 	if (!mapping)
1854 		return -ENOMEM;
1855 
1856 	mapping->start = saddr;
1857 	mapping->last = eaddr;
1858 	mapping->offset = offset;
1859 	mapping->flags = flags;
1860 
1861 	amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
1862 
1863 	return 0;
1864 }
1865 
1866 /**
1867  * amdgpu_vm_bo_replace_map - map bo inside a vm, replacing existing mappings
1868  *
1869  * @adev: amdgpu_device pointer
1870  * @bo_va: bo_va to store the address
1871  * @saddr: where to map the BO
1872  * @offset: requested offset in the BO
1873  * @size: BO size in bytes
1874  * @flags: attributes of pages (read/write/valid/etc.)
1875  *
1876  * Add a mapping of the BO at the specefied addr into the VM. Replace existing
1877  * mappings as we do so.
1878  *
1879  * Returns:
1880  * 0 for success, error for failure.
1881  *
1882  * Object has to be reserved and unreserved outside!
1883  */
amdgpu_vm_bo_replace_map(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va,uint64_t saddr,uint64_t offset,uint64_t size,uint64_t flags)1884 int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev,
1885 			     struct amdgpu_bo_va *bo_va,
1886 			     uint64_t saddr, uint64_t offset,
1887 			     uint64_t size, uint64_t flags)
1888 {
1889 	struct amdgpu_bo_va_mapping *mapping;
1890 	struct amdgpu_bo *bo = bo_va->base.bo;
1891 	uint64_t eaddr;
1892 	int r;
1893 
1894 	r = amdgpu_vm_verify_parameters(adev, bo, saddr, offset, size);
1895 	if (r)
1896 		return r;
1897 
1898 	/* Allocate all the needed memory */
1899 	mapping = kmalloc(sizeof(*mapping), GFP_KERNEL);
1900 	if (!mapping)
1901 		return -ENOMEM;
1902 
1903 	r = amdgpu_vm_bo_clear_mappings(adev, bo_va->base.vm, saddr, size);
1904 	if (r) {
1905 		kfree(mapping);
1906 		return r;
1907 	}
1908 
1909 	saddr /= AMDGPU_GPU_PAGE_SIZE;
1910 	eaddr = saddr + (size - 1) / AMDGPU_GPU_PAGE_SIZE;
1911 
1912 	mapping->start = saddr;
1913 	mapping->last = eaddr;
1914 	mapping->offset = offset;
1915 	mapping->flags = flags;
1916 
1917 	amdgpu_vm_bo_insert_map(adev, bo_va, mapping);
1918 
1919 	return 0;
1920 }
1921 
1922 /**
1923  * amdgpu_vm_bo_unmap - remove bo mapping from vm
1924  *
1925  * @adev: amdgpu_device pointer
1926  * @bo_va: bo_va to remove the address from
1927  * @saddr: where to the BO is mapped
1928  *
1929  * Remove a mapping of the BO at the specefied addr from the VM.
1930  *
1931  * Returns:
1932  * 0 for success, error for failure.
1933  *
1934  * Object has to be reserved and unreserved outside!
1935  */
amdgpu_vm_bo_unmap(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va,uint64_t saddr)1936 int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
1937 		       struct amdgpu_bo_va *bo_va,
1938 		       uint64_t saddr)
1939 {
1940 	struct amdgpu_bo_va_mapping *mapping;
1941 	struct amdgpu_vm *vm = bo_va->base.vm;
1942 	bool valid = true;
1943 
1944 	saddr /= AMDGPU_GPU_PAGE_SIZE;
1945 
1946 	list_for_each_entry(mapping, &bo_va->valids, list) {
1947 		if (mapping->start == saddr)
1948 			break;
1949 	}
1950 
1951 	if (&mapping->list == &bo_va->valids) {
1952 		valid = false;
1953 
1954 		list_for_each_entry(mapping, &bo_va->invalids, list) {
1955 			if (mapping->start == saddr)
1956 				break;
1957 		}
1958 
1959 		if (&mapping->list == &bo_va->invalids)
1960 			return -ENOENT;
1961 	}
1962 
1963 	list_del(&mapping->list);
1964 	amdgpu_vm_it_remove(mapping, &vm->va);
1965 	mapping->bo_va = NULL;
1966 	trace_amdgpu_vm_bo_unmap(bo_va, mapping);
1967 
1968 	if (valid)
1969 		list_add(&mapping->list, &vm->freed);
1970 	else
1971 		amdgpu_vm_free_mapping(adev, vm, mapping,
1972 				       bo_va->last_pt_update);
1973 
1974 	return 0;
1975 }
1976 
1977 /**
1978  * amdgpu_vm_bo_clear_mappings - remove all mappings in a specific range
1979  *
1980  * @adev: amdgpu_device pointer
1981  * @vm: VM structure to use
1982  * @saddr: start of the range
1983  * @size: size of the range
1984  *
1985  * Remove all mappings in a range, split them as appropriate.
1986  *
1987  * Returns:
1988  * 0 for success, error for failure.
1989  */
amdgpu_vm_bo_clear_mappings(struct amdgpu_device * adev,struct amdgpu_vm * vm,uint64_t saddr,uint64_t size)1990 int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev,
1991 				struct amdgpu_vm *vm,
1992 				uint64_t saddr, uint64_t size)
1993 {
1994 	struct amdgpu_bo_va_mapping *before, *after, *tmp, *next;
1995 	LIST_HEAD(removed);
1996 	uint64_t eaddr;
1997 	int r;
1998 
1999 	r = amdgpu_vm_verify_parameters(adev, NULL, saddr, 0, size);
2000 	if (r)
2001 		return r;
2002 
2003 	saddr /= AMDGPU_GPU_PAGE_SIZE;
2004 	eaddr = saddr + (size - 1) / AMDGPU_GPU_PAGE_SIZE;
2005 
2006 	/* Allocate all the needed memory */
2007 	before = kzalloc(sizeof(*before), GFP_KERNEL);
2008 	if (!before)
2009 		return -ENOMEM;
2010 	INIT_LIST_HEAD(&before->list);
2011 
2012 	after = kzalloc(sizeof(*after), GFP_KERNEL);
2013 	if (!after) {
2014 		kfree(before);
2015 		return -ENOMEM;
2016 	}
2017 	INIT_LIST_HEAD(&after->list);
2018 
2019 	/* Now gather all removed mappings */
2020 	tmp = amdgpu_vm_it_iter_first(&vm->va, saddr, eaddr);
2021 	while (tmp) {
2022 		/* Remember mapping split at the start */
2023 		if (tmp->start < saddr) {
2024 			before->start = tmp->start;
2025 			before->last = saddr - 1;
2026 			before->offset = tmp->offset;
2027 			before->flags = tmp->flags;
2028 			before->bo_va = tmp->bo_va;
2029 			list_add(&before->list, &tmp->bo_va->invalids);
2030 		}
2031 
2032 		/* Remember mapping split at the end */
2033 		if (tmp->last > eaddr) {
2034 			after->start = eaddr + 1;
2035 			after->last = tmp->last;
2036 			after->offset = tmp->offset;
2037 			after->offset += (after->start - tmp->start) << PAGE_SHIFT;
2038 			after->flags = tmp->flags;
2039 			after->bo_va = tmp->bo_va;
2040 			list_add(&after->list, &tmp->bo_va->invalids);
2041 		}
2042 
2043 		list_del(&tmp->list);
2044 		list_add(&tmp->list, &removed);
2045 
2046 		tmp = amdgpu_vm_it_iter_next(tmp, saddr, eaddr);
2047 	}
2048 
2049 	/* And free them up */
2050 	list_for_each_entry_safe(tmp, next, &removed, list) {
2051 		amdgpu_vm_it_remove(tmp, &vm->va);
2052 		list_del(&tmp->list);
2053 
2054 		if (tmp->start < saddr)
2055 		    tmp->start = saddr;
2056 		if (tmp->last > eaddr)
2057 		    tmp->last = eaddr;
2058 
2059 		tmp->bo_va = NULL;
2060 		list_add(&tmp->list, &vm->freed);
2061 		trace_amdgpu_vm_bo_unmap(NULL, tmp);
2062 	}
2063 
2064 	/* Insert partial mapping before the range */
2065 	if (!list_empty(&before->list)) {
2066 		struct amdgpu_bo *bo = before->bo_va->base.bo;
2067 
2068 		amdgpu_vm_it_insert(before, &vm->va);
2069 		if (before->flags & AMDGPU_PTE_PRT_FLAG(adev))
2070 			amdgpu_vm_prt_get(adev);
2071 
2072 		if (amdgpu_vm_is_bo_always_valid(vm, bo) &&
2073 		    !before->bo_va->base.moved)
2074 			amdgpu_vm_bo_moved(&before->bo_va->base);
2075 	} else {
2076 		kfree(before);
2077 	}
2078 
2079 	/* Insert partial mapping after the range */
2080 	if (!list_empty(&after->list)) {
2081 		struct amdgpu_bo *bo = after->bo_va->base.bo;
2082 
2083 		amdgpu_vm_it_insert(after, &vm->va);
2084 		if (after->flags & AMDGPU_PTE_PRT_FLAG(adev))
2085 			amdgpu_vm_prt_get(adev);
2086 
2087 		if (amdgpu_vm_is_bo_always_valid(vm, bo) &&
2088 		    !after->bo_va->base.moved)
2089 			amdgpu_vm_bo_moved(&after->bo_va->base);
2090 	} else {
2091 		kfree(after);
2092 	}
2093 
2094 	return 0;
2095 }
2096 
2097 /**
2098  * amdgpu_vm_bo_lookup_mapping - find mapping by address
2099  *
2100  * @vm: the requested VM
2101  * @addr: the address
2102  *
2103  * Find a mapping by it's address.
2104  *
2105  * Returns:
2106  * The amdgpu_bo_va_mapping matching for addr or NULL
2107  *
2108  */
amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm * vm,uint64_t addr)2109 struct amdgpu_bo_va_mapping *amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm,
2110 							 uint64_t addr)
2111 {
2112 	return amdgpu_vm_it_iter_first(&vm->va, addr, addr);
2113 }
2114 
2115 /**
2116  * amdgpu_vm_bo_trace_cs - trace all reserved mappings
2117  *
2118  * @vm: the requested vm
2119  * @ticket: CS ticket
2120  *
2121  * Trace all mappings of BOs reserved during a command submission.
2122  */
amdgpu_vm_bo_trace_cs(struct amdgpu_vm * vm,struct ww_acquire_ctx * ticket)2123 void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)
2124 {
2125 	struct amdgpu_bo_va_mapping *mapping;
2126 
2127 	if (!trace_amdgpu_vm_bo_cs_enabled())
2128 		return;
2129 
2130 	for (mapping = amdgpu_vm_it_iter_first(&vm->va, 0, U64_MAX); mapping;
2131 	     mapping = amdgpu_vm_it_iter_next(mapping, 0, U64_MAX)) {
2132 		if (mapping->bo_va && mapping->bo_va->base.bo) {
2133 			struct amdgpu_bo *bo;
2134 
2135 			bo = mapping->bo_va->base.bo;
2136 			if (dma_resv_locking_ctx(bo->tbo.base.resv) !=
2137 			    ticket)
2138 				continue;
2139 		}
2140 
2141 		trace_amdgpu_vm_bo_cs(mapping);
2142 	}
2143 }
2144 
2145 /**
2146  * amdgpu_vm_bo_del - remove a bo from a specific vm
2147  *
2148  * @adev: amdgpu_device pointer
2149  * @bo_va: requested bo_va
2150  *
2151  * Remove @bo_va->bo from the requested vm.
2152  *
2153  * Object have to be reserved!
2154  */
amdgpu_vm_bo_del(struct amdgpu_device * adev,struct amdgpu_bo_va * bo_va)2155 void amdgpu_vm_bo_del(struct amdgpu_device *adev,
2156 		      struct amdgpu_bo_va *bo_va)
2157 {
2158 	struct amdgpu_bo_va_mapping *mapping, *next;
2159 	struct amdgpu_bo *bo = bo_va->base.bo;
2160 	struct amdgpu_vm *vm = bo_va->base.vm;
2161 	struct amdgpu_vm_bo_base **base;
2162 
2163 	dma_resv_assert_held(vm->root.bo->tbo.base.resv);
2164 
2165 	if (bo) {
2166 		dma_resv_assert_held(bo->tbo.base.resv);
2167 		if (amdgpu_vm_is_bo_always_valid(vm, bo))
2168 			ttm_bo_set_bulk_move(&bo->tbo, NULL);
2169 
2170 		for (base = &bo_va->base.bo->vm_bo; *base;
2171 		     base = &(*base)->next) {
2172 			if (*base != &bo_va->base)
2173 				continue;
2174 
2175 			amdgpu_vm_update_stats(*base, bo->tbo.resource, -1);
2176 			*base = bo_va->base.next;
2177 			break;
2178 		}
2179 	}
2180 
2181 	spin_lock(&vm->status_lock);
2182 	list_del(&bo_va->base.vm_status);
2183 	spin_unlock(&vm->status_lock);
2184 
2185 	list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
2186 		list_del(&mapping->list);
2187 		amdgpu_vm_it_remove(mapping, &vm->va);
2188 		mapping->bo_va = NULL;
2189 		trace_amdgpu_vm_bo_unmap(bo_va, mapping);
2190 		list_add(&mapping->list, &vm->freed);
2191 	}
2192 	list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
2193 		list_del(&mapping->list);
2194 		amdgpu_vm_it_remove(mapping, &vm->va);
2195 		amdgpu_vm_free_mapping(adev, vm, mapping,
2196 				       bo_va->last_pt_update);
2197 	}
2198 
2199 	dma_fence_put(bo_va->last_pt_update);
2200 
2201 	if (bo && bo_va->is_xgmi)
2202 		amdgpu_xgmi_set_pstate(adev, AMDGPU_XGMI_PSTATE_MIN);
2203 
2204 	kfree(bo_va);
2205 }
2206 
2207 /**
2208  * amdgpu_vm_evictable - check if we can evict a VM
2209  *
2210  * @bo: A page table of the VM.
2211  *
2212  * Check if it is possible to evict a VM.
2213  */
amdgpu_vm_evictable(struct amdgpu_bo * bo)2214 bool amdgpu_vm_evictable(struct amdgpu_bo *bo)
2215 {
2216 	struct amdgpu_vm_bo_base *bo_base = bo->vm_bo;
2217 
2218 	/* Page tables of a destroyed VM can go away immediately */
2219 	if (!bo_base || !bo_base->vm)
2220 		return true;
2221 
2222 	/* Don't evict VM page tables while they are busy */
2223 	if (!dma_resv_test_signaled(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP))
2224 		return false;
2225 
2226 	/* Try to block ongoing updates */
2227 	if (!amdgpu_vm_eviction_trylock(bo_base->vm))
2228 		return false;
2229 
2230 	/* Don't evict VM page tables while they are updated */
2231 	if (!dma_fence_is_signaled(bo_base->vm->last_unlocked)) {
2232 		amdgpu_vm_eviction_unlock(bo_base->vm);
2233 		return false;
2234 	}
2235 
2236 	bo_base->vm->evicting = true;
2237 	amdgpu_vm_eviction_unlock(bo_base->vm);
2238 	return true;
2239 }
2240 
2241 /**
2242  * amdgpu_vm_bo_invalidate - mark the bo as invalid
2243  *
2244  * @bo: amdgpu buffer object
2245  * @evicted: is the BO evicted
2246  *
2247  * Mark @bo as invalid.
2248  */
amdgpu_vm_bo_invalidate(struct amdgpu_bo * bo,bool evicted)2249 void amdgpu_vm_bo_invalidate(struct amdgpu_bo *bo, bool evicted)
2250 {
2251 	struct amdgpu_vm_bo_base *bo_base;
2252 
2253 	for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
2254 		struct amdgpu_vm *vm = bo_base->vm;
2255 
2256 		if (evicted && amdgpu_vm_is_bo_always_valid(vm, bo)) {
2257 			amdgpu_vm_bo_evicted(bo_base);
2258 			continue;
2259 		}
2260 
2261 		if (bo_base->moved)
2262 			continue;
2263 		bo_base->moved = true;
2264 
2265 		if (bo->tbo.type == ttm_bo_type_kernel)
2266 			amdgpu_vm_bo_relocated(bo_base);
2267 		else if (amdgpu_vm_is_bo_always_valid(vm, bo))
2268 			amdgpu_vm_bo_moved(bo_base);
2269 		else
2270 			amdgpu_vm_bo_invalidated(bo_base);
2271 	}
2272 }
2273 
2274 /**
2275  * amdgpu_vm_bo_move - handle BO move
2276  *
2277  * @bo: amdgpu buffer object
2278  * @new_mem: the new placement of the BO move
2279  * @evicted: is the BO evicted
2280  *
2281  * Update the memory stats for the new placement and mark @bo as invalid.
2282  */
amdgpu_vm_bo_move(struct amdgpu_bo * bo,struct ttm_resource * new_mem,bool evicted)2283 void amdgpu_vm_bo_move(struct amdgpu_bo *bo, struct ttm_resource *new_mem,
2284 		       bool evicted)
2285 {
2286 	struct amdgpu_vm_bo_base *bo_base;
2287 
2288 	for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
2289 		struct amdgpu_vm *vm = bo_base->vm;
2290 
2291 		spin_lock(&vm->status_lock);
2292 		amdgpu_vm_update_stats_locked(bo_base, bo->tbo.resource, -1);
2293 		amdgpu_vm_update_stats_locked(bo_base, new_mem, +1);
2294 		spin_unlock(&vm->status_lock);
2295 	}
2296 
2297 	amdgpu_vm_bo_invalidate(bo, evicted);
2298 }
2299 
2300 /**
2301  * amdgpu_vm_get_block_size - calculate VM page table size as power of two
2302  *
2303  * @vm_size: VM size
2304  *
2305  * Returns:
2306  * VM page table as power of two
2307  */
amdgpu_vm_get_block_size(uint64_t vm_size)2308 static uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)
2309 {
2310 	/* Total bits covered by PD + PTs */
2311 	unsigned bits = ilog2(vm_size) + 18;
2312 
2313 	/* Make sure the PD is 4K in size up to 8GB address space.
2314 	   Above that split equal between PD and PTs */
2315 	if (vm_size <= 8)
2316 		return (bits - 9);
2317 	else
2318 		return ((bits + 3) / 2);
2319 }
2320 
2321 /**
2322  * amdgpu_vm_adjust_size - adjust vm size, block size and fragment size
2323  *
2324  * @adev: amdgpu_device pointer
2325  * @min_vm_size: the minimum vm size in GB if it's set auto
2326  * @fragment_size_default: Default PTE fragment size
2327  * @max_level: max VMPT level
2328  * @max_bits: max address space size in bits
2329  *
2330  */
amdgpu_vm_adjust_size(struct amdgpu_device * adev,uint32_t min_vm_size,uint32_t fragment_size_default,unsigned max_level,unsigned max_bits)2331 void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size,
2332 			   uint32_t fragment_size_default, unsigned max_level,
2333 			   unsigned max_bits)
2334 {
2335 	unsigned int max_size = 1 << (max_bits - 30);
2336 	unsigned int vm_size;
2337 	uint64_t tmp;
2338 
2339 	/* adjust vm size first */
2340 	if (amdgpu_vm_size != -1) {
2341 		vm_size = amdgpu_vm_size;
2342 		if (vm_size > max_size) {
2343 			dev_warn(adev->dev, "VM size (%d) too large, max is %u GB\n",
2344 				 amdgpu_vm_size, max_size);
2345 			vm_size = max_size;
2346 		}
2347 	} else {
2348 		struct sysinfo si;
2349 		unsigned int phys_ram_gb;
2350 
2351 		/* Optimal VM size depends on the amount of physical
2352 		 * RAM available. Underlying requirements and
2353 		 * assumptions:
2354 		 *
2355 		 *  - Need to map system memory and VRAM from all GPUs
2356 		 *     - VRAM from other GPUs not known here
2357 		 *     - Assume VRAM <= system memory
2358 		 *  - On GFX8 and older, VM space can be segmented for
2359 		 *    different MTYPEs
2360 		 *  - Need to allow room for fragmentation, guard pages etc.
2361 		 *
2362 		 * This adds up to a rough guess of system memory x3.
2363 		 * Round up to power of two to maximize the available
2364 		 * VM size with the given page table size.
2365 		 */
2366 		si_meminfo(&si);
2367 		phys_ram_gb = ((uint64_t)si.totalram * si.mem_unit +
2368 			       (1 << 30) - 1) >> 30;
2369 		vm_size = roundup_pow_of_two(
2370 			clamp(phys_ram_gb * 3, min_vm_size, max_size));
2371 	}
2372 
2373 	adev->vm_manager.max_pfn = (uint64_t)vm_size << 18;
2374 
2375 	tmp = roundup_pow_of_two(adev->vm_manager.max_pfn);
2376 	if (amdgpu_vm_block_size != -1)
2377 		tmp >>= amdgpu_vm_block_size - 9;
2378 	tmp = DIV_ROUND_UP(fls64(tmp) - 1, 9) - 1;
2379 	adev->vm_manager.num_level = min_t(unsigned int, max_level, tmp);
2380 	switch (adev->vm_manager.num_level) {
2381 	case 3:
2382 		adev->vm_manager.root_level = AMDGPU_VM_PDB2;
2383 		break;
2384 	case 2:
2385 		adev->vm_manager.root_level = AMDGPU_VM_PDB1;
2386 		break;
2387 	case 1:
2388 		adev->vm_manager.root_level = AMDGPU_VM_PDB0;
2389 		break;
2390 	default:
2391 		dev_err(adev->dev, "VMPT only supports 2~4+1 levels\n");
2392 	}
2393 	/* block size depends on vm size and hw setup*/
2394 	if (amdgpu_vm_block_size != -1)
2395 		adev->vm_manager.block_size =
2396 			min((unsigned)amdgpu_vm_block_size, max_bits
2397 			    - AMDGPU_GPU_PAGE_SHIFT
2398 			    - 9 * adev->vm_manager.num_level);
2399 	else if (adev->vm_manager.num_level > 1)
2400 		adev->vm_manager.block_size = 9;
2401 	else
2402 		adev->vm_manager.block_size = amdgpu_vm_get_block_size(tmp);
2403 
2404 	if (amdgpu_vm_fragment_size == -1)
2405 		adev->vm_manager.fragment_size = fragment_size_default;
2406 	else
2407 		adev->vm_manager.fragment_size = amdgpu_vm_fragment_size;
2408 
2409 	dev_info(
2410 		adev->dev,
2411 		"vm size is %u GB, %u levels, block size is %u-bit, fragment size is %u-bit\n",
2412 		vm_size, adev->vm_manager.num_level + 1,
2413 		adev->vm_manager.block_size, adev->vm_manager.fragment_size);
2414 }
2415 
2416 /**
2417  * amdgpu_vm_wait_idle - wait for the VM to become idle
2418  *
2419  * @vm: VM object to wait for
2420  * @timeout: timeout to wait for VM to become idle
2421  */
amdgpu_vm_wait_idle(struct amdgpu_vm * vm,long timeout)2422 long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)
2423 {
2424 	timeout = drm_sched_entity_flush(&vm->immediate, timeout);
2425 	if (timeout <= 0)
2426 		return timeout;
2427 
2428 	return drm_sched_entity_flush(&vm->delayed, timeout);
2429 }
2430 
amdgpu_vm_destroy_task_info(struct kref * kref)2431 static void amdgpu_vm_destroy_task_info(struct kref *kref)
2432 {
2433 	struct amdgpu_task_info *ti = container_of(kref, struct amdgpu_task_info, refcount);
2434 
2435 	kfree(ti);
2436 }
2437 
2438 static inline struct amdgpu_vm *
amdgpu_vm_get_vm_from_pasid(struct amdgpu_device * adev,u32 pasid)2439 amdgpu_vm_get_vm_from_pasid(struct amdgpu_device *adev, u32 pasid)
2440 {
2441 	struct amdgpu_vm *vm;
2442 	unsigned long flags;
2443 
2444 	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
2445 	vm = xa_load(&adev->vm_manager.pasids, pasid);
2446 	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
2447 
2448 	return vm;
2449 }
2450 
2451 /**
2452  * amdgpu_vm_put_task_info - reference down the vm task_info ptr
2453  *
2454  * @task_info: task_info struct under discussion.
2455  *
2456  * frees the vm task_info ptr at the last put
2457  */
amdgpu_vm_put_task_info(struct amdgpu_task_info * task_info)2458 void amdgpu_vm_put_task_info(struct amdgpu_task_info *task_info)
2459 {
2460 	if (task_info)
2461 		kref_put(&task_info->refcount, amdgpu_vm_destroy_task_info);
2462 }
2463 
2464 /**
2465  * amdgpu_vm_get_task_info_vm - Extracts task info for a vm.
2466  *
2467  * @vm: VM to get info from
2468  *
2469  * Returns the reference counted task_info structure, which must be
2470  * referenced down with amdgpu_vm_put_task_info.
2471  */
2472 struct amdgpu_task_info *
amdgpu_vm_get_task_info_vm(struct amdgpu_vm * vm)2473 amdgpu_vm_get_task_info_vm(struct amdgpu_vm *vm)
2474 {
2475 	struct amdgpu_task_info *ti = NULL;
2476 
2477 	if (vm) {
2478 		ti = vm->task_info;
2479 		kref_get(&vm->task_info->refcount);
2480 	}
2481 
2482 	return ti;
2483 }
2484 
2485 /**
2486  * amdgpu_vm_get_task_info_pasid - Extracts task info for a PASID.
2487  *
2488  * @adev: drm device pointer
2489  * @pasid: PASID identifier for VM
2490  *
2491  * Returns the reference counted task_info structure, which must be
2492  * referenced down with amdgpu_vm_put_task_info.
2493  */
2494 struct amdgpu_task_info *
amdgpu_vm_get_task_info_pasid(struct amdgpu_device * adev,u32 pasid)2495 amdgpu_vm_get_task_info_pasid(struct amdgpu_device *adev, u32 pasid)
2496 {
2497 	return amdgpu_vm_get_task_info_vm(
2498 			amdgpu_vm_get_vm_from_pasid(adev, pasid));
2499 }
2500 
amdgpu_vm_create_task_info(struct amdgpu_vm * vm)2501 static int amdgpu_vm_create_task_info(struct amdgpu_vm *vm)
2502 {
2503 	vm->task_info = kzalloc(sizeof(struct amdgpu_task_info), GFP_KERNEL);
2504 	if (!vm->task_info)
2505 		return -ENOMEM;
2506 
2507 	kref_init(&vm->task_info->refcount);
2508 	return 0;
2509 }
2510 
2511 /**
2512  * amdgpu_vm_set_task_info - Sets VMs task info.
2513  *
2514  * @vm: vm for which to set the info
2515  */
amdgpu_vm_set_task_info(struct amdgpu_vm * vm)2516 void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)
2517 {
2518 	if (!vm->task_info)
2519 		return;
2520 
2521 	if (vm->task_info->task.pid == current->pid)
2522 		return;
2523 
2524 	vm->task_info->task.pid = current->pid;
2525 	get_task_comm(vm->task_info->task.comm, current);
2526 
2527 	if (current->group_leader->mm != current->mm)
2528 		return;
2529 
2530 	vm->task_info->tgid = current->group_leader->pid;
2531 	get_task_comm(vm->task_info->process_name, current->group_leader);
2532 }
2533 
2534 /**
2535  * amdgpu_vm_init - initialize a vm instance
2536  *
2537  * @adev: amdgpu_device pointer
2538  * @vm: requested vm
2539  * @xcp_id: GPU partition selection id
2540  *
2541  * Init @vm fields.
2542  *
2543  * Returns:
2544  * 0 for success, error for failure.
2545  */
amdgpu_vm_init(struct amdgpu_device * adev,struct amdgpu_vm * vm,int32_t xcp_id)2546 int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
2547 		   int32_t xcp_id)
2548 {
2549 	struct amdgpu_bo *root_bo;
2550 	struct amdgpu_bo_vm *root;
2551 	int r, i;
2552 
2553 	vm->va = RB_ROOT_CACHED;
2554 	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
2555 		vm->reserved_vmid[i] = NULL;
2556 	INIT_LIST_HEAD(&vm->evicted);
2557 	INIT_LIST_HEAD(&vm->evicted_user);
2558 	INIT_LIST_HEAD(&vm->relocated);
2559 	INIT_LIST_HEAD(&vm->moved);
2560 	INIT_LIST_HEAD(&vm->idle);
2561 	INIT_LIST_HEAD(&vm->invalidated);
2562 	spin_lock_init(&vm->status_lock);
2563 	INIT_LIST_HEAD(&vm->freed);
2564 	INIT_LIST_HEAD(&vm->done);
2565 	INIT_KFIFO(vm->faults);
2566 
2567 	r = amdgpu_vm_init_entities(adev, vm);
2568 	if (r)
2569 		return r;
2570 
2571 	ttm_lru_bulk_move_init(&vm->lru_bulk_move);
2572 
2573 	vm->is_compute_context = false;
2574 
2575 	vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2576 				    AMDGPU_VM_USE_CPU_FOR_GFX);
2577 
2578 	dev_dbg(adev->dev, "VM update mode is %s\n",
2579 		vm->use_cpu_for_update ? "CPU" : "SDMA");
2580 	WARN_ONCE((vm->use_cpu_for_update &&
2581 		   !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2582 		  "CPU update of VM recommended only for large BAR system\n");
2583 
2584 	if (vm->use_cpu_for_update)
2585 		vm->update_funcs = &amdgpu_vm_cpu_funcs;
2586 	else
2587 		vm->update_funcs = &amdgpu_vm_sdma_funcs;
2588 
2589 	vm->last_update = dma_fence_get_stub();
2590 	vm->last_unlocked = dma_fence_get_stub();
2591 	vm->last_tlb_flush = dma_fence_get_stub();
2592 	vm->generation = amdgpu_vm_generation(adev, NULL);
2593 
2594 	mutex_init(&vm->eviction_lock);
2595 	vm->evicting = false;
2596 	vm->tlb_fence_context = dma_fence_context_alloc(1);
2597 
2598 	r = amdgpu_vm_pt_create(adev, vm, adev->vm_manager.root_level,
2599 				false, &root, xcp_id);
2600 	if (r)
2601 		goto error_free_delayed;
2602 
2603 	root_bo = amdgpu_bo_ref(&root->bo);
2604 	r = amdgpu_bo_reserve(root_bo, true);
2605 	if (r) {
2606 		amdgpu_bo_unref(&root_bo);
2607 		goto error_free_delayed;
2608 	}
2609 
2610 	amdgpu_vm_bo_base_init(&vm->root, vm, root_bo);
2611 	r = dma_resv_reserve_fences(root_bo->tbo.base.resv, 1);
2612 	if (r)
2613 		goto error_free_root;
2614 
2615 	r = amdgpu_vm_pt_clear(adev, vm, root, false);
2616 	if (r)
2617 		goto error_free_root;
2618 
2619 	r = amdgpu_vm_create_task_info(vm);
2620 	if (r)
2621 		dev_dbg(adev->dev, "Failed to create task info for VM\n");
2622 
2623 	amdgpu_bo_unreserve(vm->root.bo);
2624 	amdgpu_bo_unref(&root_bo);
2625 
2626 	return 0;
2627 
2628 error_free_root:
2629 	amdgpu_vm_pt_free_root(adev, vm);
2630 	amdgpu_bo_unreserve(vm->root.bo);
2631 	amdgpu_bo_unref(&root_bo);
2632 
2633 error_free_delayed:
2634 	dma_fence_put(vm->last_tlb_flush);
2635 	dma_fence_put(vm->last_unlocked);
2636 	ttm_lru_bulk_move_fini(&adev->mman.bdev, &vm->lru_bulk_move);
2637 	amdgpu_vm_fini_entities(vm);
2638 
2639 	return r;
2640 }
2641 
2642 /**
2643  * amdgpu_vm_make_compute - Turn a GFX VM into a compute VM
2644  *
2645  * @adev: amdgpu_device pointer
2646  * @vm: requested vm
2647  *
2648  * This only works on GFX VMs that don't have any BOs added and no
2649  * page tables allocated yet.
2650  *
2651  * Changes the following VM parameters:
2652  * - use_cpu_for_update
2653  * - pte_supports_ats
2654  *
2655  * Reinitializes the page directory to reflect the changed ATS
2656  * setting.
2657  *
2658  * Returns:
2659  * 0 for success, -errno for errors.
2660  */
amdgpu_vm_make_compute(struct amdgpu_device * adev,struct amdgpu_vm * vm)2661 int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2662 {
2663 	int r;
2664 
2665 	r = amdgpu_bo_reserve(vm->root.bo, true);
2666 	if (r)
2667 		return r;
2668 
2669 	/* Update VM state */
2670 	vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
2671 				    AMDGPU_VM_USE_CPU_FOR_COMPUTE);
2672 	dev_dbg(adev->dev, "VM update mode is %s\n",
2673 		vm->use_cpu_for_update ? "CPU" : "SDMA");
2674 	WARN_ONCE((vm->use_cpu_for_update &&
2675 		   !amdgpu_gmc_vram_full_visible(&adev->gmc)),
2676 		  "CPU update of VM recommended only for large BAR system\n");
2677 
2678 	if (vm->use_cpu_for_update) {
2679 		/* Sync with last SDMA update/clear before switching to CPU */
2680 		r = amdgpu_bo_sync_wait(vm->root.bo,
2681 					AMDGPU_FENCE_OWNER_UNDEFINED, true);
2682 		if (r)
2683 			goto unreserve_bo;
2684 
2685 		vm->update_funcs = &amdgpu_vm_cpu_funcs;
2686 		r = amdgpu_vm_pt_map_tables(adev, vm);
2687 		if (r)
2688 			goto unreserve_bo;
2689 
2690 	} else {
2691 		vm->update_funcs = &amdgpu_vm_sdma_funcs;
2692 	}
2693 
2694 	dma_fence_put(vm->last_update);
2695 	vm->last_update = dma_fence_get_stub();
2696 	vm->is_compute_context = true;
2697 
2698 unreserve_bo:
2699 	amdgpu_bo_unreserve(vm->root.bo);
2700 	return r;
2701 }
2702 
amdgpu_vm_stats_is_zero(struct amdgpu_vm * vm)2703 static int amdgpu_vm_stats_is_zero(struct amdgpu_vm *vm)
2704 {
2705 	for (int i = 0; i < __AMDGPU_PL_NUM; ++i) {
2706 		if (!(drm_memory_stats_is_zero(&vm->stats[i].drm) &&
2707 		      vm->stats[i].evicted == 0))
2708 			return false;
2709 	}
2710 	return true;
2711 }
2712 
2713 /**
2714  * amdgpu_vm_fini - tear down a vm instance
2715  *
2716  * @adev: amdgpu_device pointer
2717  * @vm: requested vm
2718  *
2719  * Tear down @vm.
2720  * Unbind the VM and remove all bos from the vm bo list
2721  */
amdgpu_vm_fini(struct amdgpu_device * adev,struct amdgpu_vm * vm)2722 void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
2723 {
2724 	struct amdgpu_bo_va_mapping *mapping, *tmp;
2725 	bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
2726 	struct amdgpu_bo *root;
2727 	unsigned long flags;
2728 	int i;
2729 
2730 	amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
2731 
2732 	root = amdgpu_bo_ref(vm->root.bo);
2733 	amdgpu_bo_reserve(root, true);
2734 	amdgpu_vm_set_pasid(adev, vm, 0);
2735 	dma_fence_wait(vm->last_unlocked, false);
2736 	dma_fence_put(vm->last_unlocked);
2737 	dma_fence_wait(vm->last_tlb_flush, false);
2738 	/* Make sure that all fence callbacks have completed */
2739 	spin_lock_irqsave(vm->last_tlb_flush->lock, flags);
2740 	spin_unlock_irqrestore(vm->last_tlb_flush->lock, flags);
2741 	dma_fence_put(vm->last_tlb_flush);
2742 
2743 	list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
2744 		if (mapping->flags & AMDGPU_PTE_PRT_FLAG(adev) && prt_fini_needed) {
2745 			amdgpu_vm_prt_fini(adev, vm);
2746 			prt_fini_needed = false;
2747 		}
2748 
2749 		list_del(&mapping->list);
2750 		amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
2751 	}
2752 
2753 	amdgpu_vm_pt_free_root(adev, vm);
2754 	amdgpu_bo_unreserve(root);
2755 	amdgpu_bo_unref(&root);
2756 	WARN_ON(vm->root.bo);
2757 
2758 	amdgpu_vm_fini_entities(vm);
2759 
2760 	if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
2761 		dev_err(adev->dev, "still active bo inside vm\n");
2762 	}
2763 	rbtree_postorder_for_each_entry_safe(mapping, tmp,
2764 					     &vm->va.rb_root, rb) {
2765 		/* Don't remove the mapping here, we don't want to trigger a
2766 		 * rebalance and the tree is about to be destroyed anyway.
2767 		 */
2768 		list_del(&mapping->list);
2769 		kfree(mapping);
2770 	}
2771 
2772 	dma_fence_put(vm->last_update);
2773 
2774 	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++) {
2775 		if (vm->reserved_vmid[i]) {
2776 			amdgpu_vmid_free_reserved(adev, i);
2777 			vm->reserved_vmid[i] = false;
2778 		}
2779 	}
2780 
2781 	ttm_lru_bulk_move_fini(&adev->mman.bdev, &vm->lru_bulk_move);
2782 
2783 	if (!amdgpu_vm_stats_is_zero(vm)) {
2784 		struct amdgpu_task_info *ti = vm->task_info;
2785 
2786 		dev_warn(adev->dev,
2787 			 "VM memory stats for proc %s(%d) task %s(%d) is non-zero when fini\n",
2788 			 ti->process_name, ti->task.pid, ti->task.comm, ti->tgid);
2789 	}
2790 
2791 	amdgpu_vm_put_task_info(vm->task_info);
2792 }
2793 
2794 /**
2795  * amdgpu_vm_manager_init - init the VM manager
2796  *
2797  * @adev: amdgpu_device pointer
2798  *
2799  * Initialize the VM manager structures
2800  */
amdgpu_vm_manager_init(struct amdgpu_device * adev)2801 void amdgpu_vm_manager_init(struct amdgpu_device *adev)
2802 {
2803 	unsigned i;
2804 
2805 	/* Concurrent flushes are only possible starting with Vega10 and
2806 	 * are broken on Navi10 and Navi14.
2807 	 */
2808 	adev->vm_manager.concurrent_flush = !(adev->asic_type < CHIP_VEGA10 ||
2809 					      adev->asic_type == CHIP_NAVI10 ||
2810 					      adev->asic_type == CHIP_NAVI14);
2811 	amdgpu_vmid_mgr_init(adev);
2812 
2813 	adev->vm_manager.fence_context =
2814 		dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2815 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
2816 		adev->vm_manager.seqno[i] = 0;
2817 
2818 	spin_lock_init(&adev->vm_manager.prt_lock);
2819 	atomic_set(&adev->vm_manager.num_prt_users, 0);
2820 
2821 	/* If not overridden by the user, by default, only in large BAR systems
2822 	 * Compute VM tables will be updated by CPU
2823 	 */
2824 #ifdef CONFIG_X86_64
2825 	if (amdgpu_vm_update_mode == -1) {
2826 		/* For asic with VF MMIO access protection
2827 		 * avoid using CPU for VM table updates
2828 		 */
2829 		if (amdgpu_gmc_vram_full_visible(&adev->gmc) &&
2830 		    !amdgpu_sriov_vf_mmio_access_protection(adev))
2831 			adev->vm_manager.vm_update_mode =
2832 				AMDGPU_VM_USE_CPU_FOR_COMPUTE;
2833 		else
2834 			adev->vm_manager.vm_update_mode = 0;
2835 	} else
2836 		adev->vm_manager.vm_update_mode = amdgpu_vm_update_mode;
2837 #else
2838 	adev->vm_manager.vm_update_mode = 0;
2839 #endif
2840 
2841 	xa_init_flags(&adev->vm_manager.pasids, XA_FLAGS_LOCK_IRQ);
2842 }
2843 
2844 /**
2845  * amdgpu_vm_manager_fini - cleanup VM manager
2846  *
2847  * @adev: amdgpu_device pointer
2848  *
2849  * Cleanup the VM manager and free resources.
2850  */
amdgpu_vm_manager_fini(struct amdgpu_device * adev)2851 void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
2852 {
2853 	WARN_ON(!xa_empty(&adev->vm_manager.pasids));
2854 	xa_destroy(&adev->vm_manager.pasids);
2855 
2856 	amdgpu_vmid_mgr_fini(adev);
2857 }
2858 
2859 /**
2860  * amdgpu_vm_ioctl - Manages VMID reservation for vm hubs.
2861  *
2862  * @dev: drm device pointer
2863  * @data: drm_amdgpu_vm
2864  * @filp: drm file pointer
2865  *
2866  * Returns:
2867  * 0 for success, -errno for errors.
2868  */
amdgpu_vm_ioctl(struct drm_device * dev,void * data,struct drm_file * filp)2869 int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
2870 {
2871 	union drm_amdgpu_vm *args = data;
2872 	struct amdgpu_device *adev = drm_to_adev(dev);
2873 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
2874 
2875 	/* No valid flags defined yet */
2876 	if (args->in.flags)
2877 		return -EINVAL;
2878 
2879 	switch (args->in.op) {
2880 	case AMDGPU_VM_OP_RESERVE_VMID:
2881 		/* We only have requirement to reserve vmid from gfxhub */
2882 		if (!fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)]) {
2883 			amdgpu_vmid_alloc_reserved(adev, AMDGPU_GFXHUB(0));
2884 			fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)] = true;
2885 		}
2886 
2887 		break;
2888 	case AMDGPU_VM_OP_UNRESERVE_VMID:
2889 		if (fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)]) {
2890 			amdgpu_vmid_free_reserved(adev, AMDGPU_GFXHUB(0));
2891 			fpriv->vm.reserved_vmid[AMDGPU_GFXHUB(0)] = false;
2892 		}
2893 		break;
2894 	default:
2895 		return -EINVAL;
2896 	}
2897 
2898 	return 0;
2899 }
2900 
2901 /**
2902  * amdgpu_vm_handle_fault - graceful handling of VM faults.
2903  * @adev: amdgpu device pointer
2904  * @pasid: PASID of the VM
2905  * @ts: Timestamp of the fault
2906  * @vmid: VMID, only used for GFX 9.4.3.
2907  * @node_id: Node_id received in IH cookie. Only applicable for
2908  *           GFX 9.4.3.
2909  * @addr: Address of the fault
2910  * @write_fault: true is write fault, false is read fault
2911  *
2912  * Try to gracefully handle a VM fault. Return true if the fault was handled and
2913  * shouldn't be reported any more.
2914  */
amdgpu_vm_handle_fault(struct amdgpu_device * adev,u32 pasid,u32 vmid,u32 node_id,uint64_t addr,uint64_t ts,bool write_fault)2915 bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
2916 			    u32 vmid, u32 node_id, uint64_t addr, uint64_t ts,
2917 			    bool write_fault)
2918 {
2919 	bool is_compute_context = false;
2920 	struct amdgpu_bo *root;
2921 	unsigned long irqflags;
2922 	uint64_t value, flags;
2923 	struct amdgpu_vm *vm;
2924 	int r;
2925 
2926 	xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
2927 	vm = xa_load(&adev->vm_manager.pasids, pasid);
2928 	if (vm) {
2929 		root = amdgpu_bo_ref(vm->root.bo);
2930 		is_compute_context = vm->is_compute_context;
2931 	} else {
2932 		root = NULL;
2933 	}
2934 	xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
2935 
2936 	if (!root)
2937 		return false;
2938 
2939 	addr /= AMDGPU_GPU_PAGE_SIZE;
2940 
2941 	if (is_compute_context && !svm_range_restore_pages(adev, pasid, vmid,
2942 	    node_id, addr, ts, write_fault)) {
2943 		amdgpu_bo_unref(&root);
2944 		return true;
2945 	}
2946 
2947 	r = amdgpu_bo_reserve(root, true);
2948 	if (r)
2949 		goto error_unref;
2950 
2951 	/* Double check that the VM still exists */
2952 	xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
2953 	vm = xa_load(&adev->vm_manager.pasids, pasid);
2954 	if (vm && vm->root.bo != root)
2955 		vm = NULL;
2956 	xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
2957 	if (!vm)
2958 		goto error_unlock;
2959 
2960 	flags = AMDGPU_PTE_VALID | AMDGPU_PTE_SNOOPED |
2961 		AMDGPU_PTE_SYSTEM;
2962 
2963 	if (is_compute_context) {
2964 		/* Intentionally setting invalid PTE flag
2965 		 * combination to force a no-retry-fault
2966 		 */
2967 		flags = AMDGPU_VM_NORETRY_FLAGS;
2968 		value = 0;
2969 	} else if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_NEVER) {
2970 		/* Redirect the access to the dummy page */
2971 		value = adev->dummy_page_addr;
2972 		flags |= AMDGPU_PTE_EXECUTABLE | AMDGPU_PTE_READABLE |
2973 			AMDGPU_PTE_WRITEABLE;
2974 
2975 	} else {
2976 		/* Let the hw retry silently on the PTE */
2977 		value = 0;
2978 	}
2979 
2980 	r = dma_resv_reserve_fences(root->tbo.base.resv, 1);
2981 	if (r) {
2982 		pr_debug("failed %d to reserve fence slot\n", r);
2983 		goto error_unlock;
2984 	}
2985 
2986 	r = amdgpu_vm_update_range(adev, vm, true, false, false, false,
2987 				   NULL, addr, addr, flags, value, 0, NULL, NULL, NULL);
2988 	if (r)
2989 		goto error_unlock;
2990 
2991 	r = amdgpu_vm_update_pdes(adev, vm, true);
2992 
2993 error_unlock:
2994 	amdgpu_bo_unreserve(root);
2995 	if (r < 0)
2996 		dev_err(adev->dev, "Can't handle page fault (%d)\n", r);
2997 
2998 error_unref:
2999 	amdgpu_bo_unref(&root);
3000 
3001 	return false;
3002 }
3003 
3004 #if defined(CONFIG_DEBUG_FS)
3005 /**
3006  * amdgpu_debugfs_vm_bo_info  - print BO info for the VM
3007  *
3008  * @vm: Requested VM for printing BO info
3009  * @m: debugfs file
3010  *
3011  * Print BO information in debugfs file for the VM
3012  */
amdgpu_debugfs_vm_bo_info(struct amdgpu_vm * vm,struct seq_file * m)3013 void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m)
3014 {
3015 	struct amdgpu_bo_va *bo_va, *tmp;
3016 	u64 total_idle = 0;
3017 	u64 total_evicted = 0;
3018 	u64 total_relocated = 0;
3019 	u64 total_moved = 0;
3020 	u64 total_invalidated = 0;
3021 	u64 total_done = 0;
3022 	unsigned int total_idle_objs = 0;
3023 	unsigned int total_evicted_objs = 0;
3024 	unsigned int total_relocated_objs = 0;
3025 	unsigned int total_moved_objs = 0;
3026 	unsigned int total_invalidated_objs = 0;
3027 	unsigned int total_done_objs = 0;
3028 	unsigned int id = 0;
3029 
3030 	spin_lock(&vm->status_lock);
3031 	seq_puts(m, "\tIdle BOs:\n");
3032 	list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
3033 		if (!bo_va->base.bo)
3034 			continue;
3035 		total_idle += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3036 	}
3037 	total_idle_objs = id;
3038 	id = 0;
3039 
3040 	seq_puts(m, "\tEvicted BOs:\n");
3041 	list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
3042 		if (!bo_va->base.bo)
3043 			continue;
3044 		total_evicted += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3045 	}
3046 	total_evicted_objs = id;
3047 	id = 0;
3048 
3049 	seq_puts(m, "\tRelocated BOs:\n");
3050 	list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
3051 		if (!bo_va->base.bo)
3052 			continue;
3053 		total_relocated += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3054 	}
3055 	total_relocated_objs = id;
3056 	id = 0;
3057 
3058 	seq_puts(m, "\tMoved BOs:\n");
3059 	list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
3060 		if (!bo_va->base.bo)
3061 			continue;
3062 		total_moved += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3063 	}
3064 	total_moved_objs = id;
3065 	id = 0;
3066 
3067 	seq_puts(m, "\tInvalidated BOs:\n");
3068 	list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
3069 		if (!bo_va->base.bo)
3070 			continue;
3071 		total_invalidated += amdgpu_bo_print_info(id++,	bo_va->base.bo, m);
3072 	}
3073 	total_invalidated_objs = id;
3074 	id = 0;
3075 
3076 	seq_puts(m, "\tDone BOs:\n");
3077 	list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
3078 		if (!bo_va->base.bo)
3079 			continue;
3080 		total_done += amdgpu_bo_print_info(id++, bo_va->base.bo, m);
3081 	}
3082 	spin_unlock(&vm->status_lock);
3083 	total_done_objs = id;
3084 
3085 	seq_printf(m, "\tTotal idle size:        %12lld\tobjs:\t%d\n", total_idle,
3086 		   total_idle_objs);
3087 	seq_printf(m, "\tTotal evicted size:     %12lld\tobjs:\t%d\n", total_evicted,
3088 		   total_evicted_objs);
3089 	seq_printf(m, "\tTotal relocated size:   %12lld\tobjs:\t%d\n", total_relocated,
3090 		   total_relocated_objs);
3091 	seq_printf(m, "\tTotal moved size:       %12lld\tobjs:\t%d\n", total_moved,
3092 		   total_moved_objs);
3093 	seq_printf(m, "\tTotal invalidated size: %12lld\tobjs:\t%d\n", total_invalidated,
3094 		   total_invalidated_objs);
3095 	seq_printf(m, "\tTotal done size:        %12lld\tobjs:\t%d\n", total_done,
3096 		   total_done_objs);
3097 }
3098 #endif
3099 
3100 /**
3101  * amdgpu_vm_update_fault_cache - update cached fault into.
3102  * @adev: amdgpu device pointer
3103  * @pasid: PASID of the VM
3104  * @addr: Address of the fault
3105  * @status: GPUVM fault status register
3106  * @vmhub: which vmhub got the fault
3107  *
3108  * Cache the fault info for later use by userspace in debugging.
3109  */
amdgpu_vm_update_fault_cache(struct amdgpu_device * adev,unsigned int pasid,uint64_t addr,uint32_t status,unsigned int vmhub)3110 void amdgpu_vm_update_fault_cache(struct amdgpu_device *adev,
3111 				  unsigned int pasid,
3112 				  uint64_t addr,
3113 				  uint32_t status,
3114 				  unsigned int vmhub)
3115 {
3116 	struct amdgpu_vm *vm;
3117 	unsigned long flags;
3118 
3119 	xa_lock_irqsave(&adev->vm_manager.pasids, flags);
3120 
3121 	vm = xa_load(&adev->vm_manager.pasids, pasid);
3122 	/* Don't update the fault cache if status is 0.  In the multiple
3123 	 * fault case, subsequent faults will return a 0 status which is
3124 	 * useless for userspace and replaces the useful fault status, so
3125 	 * only update if status is non-0.
3126 	 */
3127 	if (vm && status) {
3128 		vm->fault_info.addr = addr;
3129 		vm->fault_info.status = status;
3130 		/*
3131 		 * Update the fault information globally for later usage
3132 		 * when vm could be stale or freed.
3133 		 */
3134 		adev->vm_manager.fault_info.addr = addr;
3135 		adev->vm_manager.fault_info.vmhub = vmhub;
3136 		adev->vm_manager.fault_info.status = status;
3137 
3138 		if (AMDGPU_IS_GFXHUB(vmhub)) {
3139 			vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_GFX;
3140 			vm->fault_info.vmhub |=
3141 				(vmhub - AMDGPU_GFXHUB_START) << AMDGPU_VMHUB_IDX_SHIFT;
3142 		} else if (AMDGPU_IS_MMHUB0(vmhub)) {
3143 			vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_MM0;
3144 			vm->fault_info.vmhub |=
3145 				(vmhub - AMDGPU_MMHUB0_START) << AMDGPU_VMHUB_IDX_SHIFT;
3146 		} else if (AMDGPU_IS_MMHUB1(vmhub)) {
3147 			vm->fault_info.vmhub = AMDGPU_VMHUB_TYPE_MM1;
3148 			vm->fault_info.vmhub |=
3149 				(vmhub - AMDGPU_MMHUB1_START) << AMDGPU_VMHUB_IDX_SHIFT;
3150 		} else {
3151 			WARN_ONCE(1, "Invalid vmhub %u\n", vmhub);
3152 		}
3153 	}
3154 	xa_unlock_irqrestore(&adev->vm_manager.pasids, flags);
3155 }
3156 
3157 /**
3158  * amdgpu_vm_is_bo_always_valid - check if the BO is VM always valid
3159  *
3160  * @vm: VM to test against.
3161  * @bo: BO to be tested.
3162  *
3163  * Returns true if the BO shares the dma_resv object with the root PD and is
3164  * always guaranteed to be valid inside the VM.
3165  */
amdgpu_vm_is_bo_always_valid(struct amdgpu_vm * vm,struct amdgpu_bo * bo)3166 bool amdgpu_vm_is_bo_always_valid(struct amdgpu_vm *vm, struct amdgpu_bo *bo)
3167 {
3168 	return bo && bo->tbo.base.resv == vm->root.bo->tbo.base.resv;
3169 }
3170 
amdgpu_vm_print_task_info(struct amdgpu_device * adev,struct amdgpu_task_info * task_info)3171 void amdgpu_vm_print_task_info(struct amdgpu_device *adev,
3172 			       struct amdgpu_task_info *task_info)
3173 {
3174 	dev_err(adev->dev,
3175 		" Process %s pid %d thread %s pid %d\n",
3176 		task_info->process_name, task_info->tgid,
3177 		task_info->task.comm, task_info->task.pid);
3178 }
3179