xref: /linux/drivers/gpu/drm/xe/xe_vm_types.h (revision e9ef810dfee7a2227da9d423aecb0ced35faddbe)
1 /* SPDX-License-Identifier: MIT */
2 /*
3  * Copyright © 2022 Intel Corporation
4  */
5 
6 #ifndef _XE_VM_TYPES_H_
7 #define _XE_VM_TYPES_H_
8 
9 #include <drm/drm_gpusvm.h>
10 #include <drm/drm_gpuvm.h>
11 
12 #include <linux/dma-resv.h>
13 #include <linux/kref.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/scatterlist.h>
16 
17 #include "xe_device_types.h"
18 #include "xe_pt_types.h"
19 #include "xe_range_fence.h"
20 
21 struct xe_bo;
22 struct xe_svm_range;
23 struct xe_sync_entry;
24 struct xe_user_fence;
25 struct xe_vm;
26 struct xe_vm_pgtable_update_op;
27 
28 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG)
29 #define TEST_VM_OPS_ERROR
30 #define FORCE_OP_ERROR	BIT(31)
31 
32 #define FORCE_OP_ERROR_LOCK	0
33 #define FORCE_OP_ERROR_PREPARE	1
34 #define FORCE_OP_ERROR_RUN	2
35 #define FORCE_OP_ERROR_COUNT	3
36 #endif
37 
38 #define XE_VMA_READ_ONLY	DRM_GPUVA_USERBITS
39 #define XE_VMA_DESTROYED	(DRM_GPUVA_USERBITS << 1)
40 #define XE_VMA_ATOMIC_PTE_BIT	(DRM_GPUVA_USERBITS << 2)
41 #define XE_VMA_PTE_4K		(DRM_GPUVA_USERBITS << 3)
42 #define XE_VMA_PTE_2M		(DRM_GPUVA_USERBITS << 4)
43 #define XE_VMA_PTE_1G		(DRM_GPUVA_USERBITS << 5)
44 #define XE_VMA_PTE_64K		(DRM_GPUVA_USERBITS << 6)
45 #define XE_VMA_PTE_COMPACT	(DRM_GPUVA_USERBITS << 7)
46 #define XE_VMA_DUMPABLE		(DRM_GPUVA_USERBITS << 8)
47 #define XE_VMA_SYSTEM_ALLOCATOR	(DRM_GPUVA_USERBITS << 9)
48 
49 /** struct xe_userptr - User pointer */
50 struct xe_userptr {
51 	/** @invalidate_link: Link for the vm::userptr.invalidated list */
52 	struct list_head invalidate_link;
53 	/** @userptr: link into VM repin list if userptr. */
54 	struct list_head repin_link;
55 	/**
56 	 * @notifier: MMU notifier for user pointer (invalidation call back)
57 	 */
58 	struct mmu_interval_notifier notifier;
59 	/** @sgt: storage for a scatter gather table */
60 	struct sg_table sgt;
61 	/** @sg: allocated scatter gather table */
62 	struct sg_table *sg;
63 	/** @notifier_seq: notifier sequence number */
64 	unsigned long notifier_seq;
65 	/** @unmap_mutex: Mutex protecting dma-unmapping */
66 	struct mutex unmap_mutex;
67 	/**
68 	 * @initial_bind: user pointer has been bound at least once.
69 	 * write: vm->userptr.notifier_lock in read mode and vm->resv held.
70 	 * read: vm->userptr.notifier_lock in write mode or vm->resv held.
71 	 */
72 	bool initial_bind;
73 	/** @mapped: Whether the @sgt sg-table is dma-mapped. Protected by @unmap_mutex. */
74 	bool mapped;
75 #if IS_ENABLED(CONFIG_DRM_XE_USERPTR_INVAL_INJECT)
76 	u32 divisor;
77 #endif
78 };
79 
80 struct xe_vma {
81 	/** @gpuva: Base GPUVA object */
82 	struct drm_gpuva gpuva;
83 
84 	/**
85 	 * @combined_links: links into lists which are mutually exclusive.
86 	 * Locking: vm lock in write mode OR vm lock in read mode and the vm's
87 	 * resv.
88 	 */
89 	union {
90 		/** @rebind: link into VM if this VMA needs rebinding. */
91 		struct list_head rebind;
92 		/** @destroy: link to contested list when VM is being closed. */
93 		struct list_head destroy;
94 	} combined_links;
95 
96 	union {
97 		/** @destroy_cb: callback to destroy VMA when unbind job is done */
98 		struct dma_fence_cb destroy_cb;
99 		/** @destroy_work: worker to destroy this BO */
100 		struct work_struct destroy_work;
101 	};
102 
103 	/**
104 	 * @tile_invalidated: Tile mask of binding are invalidated for this VMA.
105 	 * protected by BO's resv and for userptrs, vm->userptr.notifier_lock in
106 	 * write mode for writing or vm->userptr.notifier_lock in read mode and
107 	 * the vm->resv. For stable reading, BO's resv or userptr
108 	 * vm->userptr.notifier_lock in read mode is required. Can be
109 	 * opportunistically read with READ_ONCE outside of locks.
110 	 */
111 	u8 tile_invalidated;
112 
113 	/** @tile_mask: Tile mask of where to create binding for this VMA */
114 	u8 tile_mask;
115 
116 	/**
117 	 * @tile_present: Tile mask of binding are present for this VMA.
118 	 * protected by vm->lock, vm->resv and for userptrs,
119 	 * vm->userptr.notifier_lock for writing. Needs either for reading,
120 	 * but if reading is done under the vm->lock only, it needs to be held
121 	 * in write mode.
122 	 */
123 	u8 tile_present;
124 
125 	/** @tile_staged: bind is staged for this VMA */
126 	u8 tile_staged;
127 
128 	/**
129 	 * @pat_index: The pat index to use when encoding the PTEs for this vma.
130 	 */
131 	u16 pat_index;
132 
133 	/**
134 	 * @ufence: The user fence that was provided with MAP.
135 	 * Needs to be signalled before UNMAP can be processed.
136 	 */
137 	struct xe_user_fence *ufence;
138 };
139 
140 /**
141  * struct xe_userptr_vma - A userptr vma subclass
142  * @vma: The vma.
143  * @userptr: Additional userptr information.
144  */
145 struct xe_userptr_vma {
146 	struct xe_vma vma;
147 	struct xe_userptr userptr;
148 };
149 
150 struct xe_device;
151 
152 struct xe_vm {
153 	/** @gpuvm: base GPUVM used to track VMAs */
154 	struct drm_gpuvm gpuvm;
155 
156 	/** @svm: Shared virtual memory state */
157 	struct {
158 		/** @svm.gpusvm: base GPUSVM used to track fault allocations */
159 		struct drm_gpusvm gpusvm;
160 		/**
161 		 * @svm.garbage_collector: Garbage collector which is used unmap
162 		 * SVM range's GPU bindings and destroy the ranges.
163 		 */
164 		struct {
165 			/** @svm.garbage_collector.lock: Protect's range list */
166 			spinlock_t lock;
167 			/**
168 			 * @svm.garbage_collector.range_list: List of SVM ranges
169 			 * in the garbage collector.
170 			 */
171 			struct list_head range_list;
172 			/**
173 			 * @svm.garbage_collector.work: Worker which the
174 			 * garbage collector runs on.
175 			 */
176 			struct work_struct work;
177 		} garbage_collector;
178 	} svm;
179 
180 	struct xe_device *xe;
181 
182 	/* exec queue used for (un)binding vma's */
183 	struct xe_exec_queue *q[XE_MAX_TILES_PER_DEVICE];
184 
185 	/** @lru_bulk_move: Bulk LRU move list for this VM's BOs */
186 	struct ttm_lru_bulk_move lru_bulk_move;
187 
188 	u64 size;
189 
190 	struct xe_pt *pt_root[XE_MAX_TILES_PER_DEVICE];
191 	struct xe_pt *scratch_pt[XE_MAX_TILES_PER_DEVICE][XE_VM_MAX_LEVEL];
192 
193 	/**
194 	 * @flags: flags for this VM, statically setup a creation time aside
195 	 * from XE_VM_FLAG_BANNED which requires vm->lock to set / read safely
196 	 */
197 #define XE_VM_FLAG_64K			BIT(0)
198 #define XE_VM_FLAG_LR_MODE		BIT(1)
199 #define XE_VM_FLAG_MIGRATION		BIT(2)
200 #define XE_VM_FLAG_SCRATCH_PAGE		BIT(3)
201 #define XE_VM_FLAG_FAULT_MODE		BIT(4)
202 #define XE_VM_FLAG_BANNED		BIT(5)
203 #define XE_VM_FLAG_TILE_ID(flags)	FIELD_GET(GENMASK(7, 6), flags)
204 #define XE_VM_FLAG_SET_TILE_ID(tile)	FIELD_PREP(GENMASK(7, 6), (tile)->id)
205 #define XE_VM_FLAG_GSC			BIT(8)
206 	unsigned long flags;
207 
208 	/** @composite_fence_ctx: context composite fence */
209 	u64 composite_fence_ctx;
210 	/** @composite_fence_seqno: seqno for composite fence */
211 	u32 composite_fence_seqno;
212 
213 	/**
214 	 * @lock: outer most lock, protects objects of anything attached to this
215 	 * VM
216 	 */
217 	struct rw_semaphore lock;
218 	/**
219 	 * @snap_mutex: Mutex used to guard insertions and removals from gpuva,
220 	 * so we can take a snapshot safely from devcoredump.
221 	 */
222 	struct mutex snap_mutex;
223 
224 	/**
225 	 * @rebind_list: list of VMAs that need rebinding. Protected by the
226 	 * vm->lock in write mode, OR (the vm->lock in read mode and the
227 	 * vm resv).
228 	 */
229 	struct list_head rebind_list;
230 
231 	/**
232 	 * @destroy_work: worker to destroy VM, needed as a dma_fence signaling
233 	 * from an irq context can be last put and the destroy needs to be able
234 	 * to sleep.
235 	 */
236 	struct work_struct destroy_work;
237 
238 	/**
239 	 * @rftree: range fence tree to track updates to page table structure.
240 	 * Used to implement conflict tracking between independent bind engines.
241 	 */
242 	struct xe_range_fence_tree rftree[XE_MAX_TILES_PER_DEVICE];
243 
244 	const struct xe_pt_ops *pt_ops;
245 
246 	/** @userptr: user pointer state */
247 	struct {
248 		/**
249 		 * @userptr.repin_list: list of VMAs which are user pointers,
250 		 * and needs repinning. Protected by @lock.
251 		 */
252 		struct list_head repin_list;
253 		/**
254 		 * @notifier_lock: protects notifier in write mode and
255 		 * submission in read mode.
256 		 */
257 		struct rw_semaphore notifier_lock;
258 		/**
259 		 * @userptr.invalidated_lock: Protects the
260 		 * @userptr.invalidated list.
261 		 */
262 		spinlock_t invalidated_lock;
263 		/**
264 		 * @userptr.invalidated: List of invalidated userptrs, not yet
265 		 * picked
266 		 * up for revalidation. Protected from access with the
267 		 * @invalidated_lock. Removing items from the list
268 		 * additionally requires @lock in write mode, and adding
269 		 * items to the list requires either the @userptr.notifer_lock in
270 		 * write mode, OR @lock in write mode.
271 		 */
272 		struct list_head invalidated;
273 	} userptr;
274 
275 	/** @preempt: preempt state */
276 	struct {
277 		/**
278 		 * @min_run_period_ms: The minimum run period before preempting
279 		 * an engine again
280 		 */
281 		s64 min_run_period_ms;
282 		/** @exec_queues: list of exec queues attached to this VM */
283 		struct list_head exec_queues;
284 		/** @num_exec_queues: number exec queues attached to this VM */
285 		int num_exec_queues;
286 		/**
287 		 * @rebind_deactivated: Whether rebind has been temporarily deactivated
288 		 * due to no work available. Protected by the vm resv.
289 		 */
290 		bool rebind_deactivated;
291 		/**
292 		 * @rebind_work: worker to rebind invalidated userptrs / evicted
293 		 * BOs
294 		 */
295 		struct work_struct rebind_work;
296 	} preempt;
297 
298 	/** @um: unified memory state */
299 	struct {
300 		/** @asid: address space ID, unique to each VM */
301 		u32 asid;
302 		/**
303 		 * @last_fault_vma: Last fault VMA, used for fast lookup when we
304 		 * get a flood of faults to the same VMA
305 		 */
306 		struct xe_vma *last_fault_vma;
307 	} usm;
308 
309 	/** @error_capture: allow to track errors */
310 	struct {
311 		/** @capture_once: capture only one error per VM */
312 		bool capture_once;
313 	} error_capture;
314 
315 	/**
316 	 * @tlb_flush_seqno: Required TLB flush seqno for the next exec.
317 	 * protected by the vm resv.
318 	 */
319 	u64 tlb_flush_seqno;
320 	/**
321 	 * @validating: The task that is currently making bos resident for this vm.
322 	 * Protected by the VM's resv for writing. Opportunistic reading can be done
323 	 * using READ_ONCE. Note: This is a workaround for the
324 	 * TTM eviction_valuable() callback not being passed a struct
325 	 * ttm_operation_context(). Future work might want to address this.
326 	 */
327 	struct task_struct *validating;
328 	/** @batch_invalidate_tlb: Always invalidate TLB before batch start */
329 	bool batch_invalidate_tlb;
330 	/** @xef: XE file handle for tracking this VM's drm client */
331 	struct xe_file *xef;
332 };
333 
334 /** struct xe_vma_op_map - VMA map operation */
335 struct xe_vma_op_map {
336 	/** @vma: VMA to map */
337 	struct xe_vma *vma;
338 	/** @immediate: Immediate bind */
339 	bool immediate;
340 	/** @read_only: Read only */
341 	bool read_only;
342 	/** @is_null: is NULL binding */
343 	bool is_null;
344 	/** @is_cpu_addr_mirror: is CPU address mirror binding */
345 	bool is_cpu_addr_mirror;
346 	/** @dumpable: whether BO is dumped on GPU hang */
347 	bool dumpable;
348 	/** @invalidate: invalidate the VMA before bind */
349 	bool invalidate_on_bind;
350 	/** @pat_index: The pat index to use for this operation. */
351 	u16 pat_index;
352 };
353 
354 /** struct xe_vma_op_remap - VMA remap operation */
355 struct xe_vma_op_remap {
356 	/** @prev: VMA preceding part of a split mapping */
357 	struct xe_vma *prev;
358 	/** @next: VMA subsequent part of a split mapping */
359 	struct xe_vma *next;
360 	/** @start: start of the VMA unmap */
361 	u64 start;
362 	/** @range: range of the VMA unmap */
363 	u64 range;
364 	/** @skip_prev: skip prev rebind */
365 	bool skip_prev;
366 	/** @skip_next: skip next rebind */
367 	bool skip_next;
368 	/** @unmap_done: unmap operation in done */
369 	bool unmap_done;
370 };
371 
372 /** struct xe_vma_op_prefetch - VMA prefetch operation */
373 struct xe_vma_op_prefetch {
374 	/** @region: memory region to prefetch to */
375 	u32 region;
376 };
377 
378 /** struct xe_vma_op_map_range - VMA map range operation */
379 struct xe_vma_op_map_range {
380 	/** @vma: VMA to map (system allocator VMA) */
381 	struct xe_vma *vma;
382 	/** @range: SVM range to map */
383 	struct xe_svm_range *range;
384 };
385 
386 /** struct xe_vma_op_unmap_range - VMA unmap range operation */
387 struct xe_vma_op_unmap_range {
388 	/** @range: SVM range to unmap */
389 	struct xe_svm_range *range;
390 };
391 
392 /** struct xe_vma_op_prefetch_range - VMA prefetch range operation */
393 struct xe_vma_op_prefetch_range {
394 	/** @range: xarray for SVM ranges data */
395 	struct xarray range;
396 	/** @ranges_count: number of svm ranges to map */
397 	u32 ranges_count;
398 	/** @region: memory region to prefetch to */
399 	u32 region;
400 };
401 
402 /** enum xe_vma_op_flags - flags for VMA operation */
403 enum xe_vma_op_flags {
404 	/** @XE_VMA_OP_COMMITTED: VMA operation committed */
405 	XE_VMA_OP_COMMITTED		= BIT(0),
406 	/** @XE_VMA_OP_PREV_COMMITTED: Previous VMA operation committed */
407 	XE_VMA_OP_PREV_COMMITTED	= BIT(1),
408 	/** @XE_VMA_OP_NEXT_COMMITTED: Next VMA operation committed */
409 	XE_VMA_OP_NEXT_COMMITTED	= BIT(2),
410 };
411 
412 /** enum xe_vma_subop - VMA sub-operation */
413 enum xe_vma_subop {
414 	/** @XE_VMA_SUBOP_MAP_RANGE: Map range */
415 	XE_VMA_SUBOP_MAP_RANGE,
416 	/** @XE_VMA_SUBOP_UNMAP_RANGE: Unmap range */
417 	XE_VMA_SUBOP_UNMAP_RANGE,
418 };
419 
420 /** struct xe_vma_op - VMA operation */
421 struct xe_vma_op {
422 	/** @base: GPUVA base operation */
423 	struct drm_gpuva_op base;
424 	/** @link: async operation link */
425 	struct list_head link;
426 	/** @flags: operation flags */
427 	enum xe_vma_op_flags flags;
428 	/** @subop: user defined sub-operation */
429 	enum xe_vma_subop subop;
430 	/** @tile_mask: Tile mask for operation */
431 	u8 tile_mask;
432 
433 	union {
434 		/** @map: VMA map operation specific data */
435 		struct xe_vma_op_map map;
436 		/** @remap: VMA remap operation specific data */
437 		struct xe_vma_op_remap remap;
438 		/** @prefetch: VMA prefetch operation specific data */
439 		struct xe_vma_op_prefetch prefetch;
440 		/** @map_range: VMA map range operation specific data */
441 		struct xe_vma_op_map_range map_range;
442 		/** @unmap_range: VMA unmap range operation specific data */
443 		struct xe_vma_op_unmap_range unmap_range;
444 		/** @prefetch_range: VMA prefetch range operation specific data */
445 		struct xe_vma_op_prefetch_range prefetch_range;
446 	};
447 };
448 
449 /** struct xe_vma_ops - VMA operations */
450 struct xe_vma_ops {
451 	/** @list: list of VMA operations */
452 	struct list_head list;
453 	/** @vm: VM */
454 	struct xe_vm *vm;
455 	/** @q: exec queue for VMA operations */
456 	struct xe_exec_queue *q;
457 	/** @syncs: syncs these operation */
458 	struct xe_sync_entry *syncs;
459 	/** @num_syncs: number of syncs */
460 	u32 num_syncs;
461 	/** @pt_update_ops: page table update operations */
462 	struct xe_vm_pgtable_update_ops pt_update_ops[XE_MAX_TILES_PER_DEVICE];
463 	/** @flag: signify the properties within xe_vma_ops*/
464 #define XE_VMA_OPS_FLAG_HAS_SVM_PREFETCH BIT(0)
465 	u32 flags;
466 #ifdef TEST_VM_OPS_ERROR
467 	/** @inject_error: inject error to test error handling */
468 	bool inject_error;
469 #endif
470 };
471 
472 #endif
473