]> Git Repo - J-linux.git/blob - drivers/gpu/drm/panthor/panthor_heap.c
Merge tag 'fbdev-for-6.10-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/delle...
[J-linux.git] / drivers / gpu / drm / panthor / panthor_heap.c
1 // SPDX-License-Identifier: GPL-2.0 or MIT
2 /* Copyright 2023 Collabora ltd. */
3
4 #include <linux/iosys-map.h>
5 #include <linux/rwsem.h>
6
7 #include <drm/panthor_drm.h>
8
9 #include "panthor_device.h"
10 #include "panthor_gem.h"
11 #include "panthor_heap.h"
12 #include "panthor_mmu.h"
13 #include "panthor_regs.h"
14
15 /*
16  * The GPU heap context is an opaque structure used by the GPU to track the
17  * heap allocations. The driver should only touch it to initialize it (zero all
18  * fields). Because the CPU and GPU can both access this structure it is
19  * required to be GPU cache line aligned.
20  */
21 #define HEAP_CONTEXT_SIZE       32
22
23 /**
24  * struct panthor_heap_chunk_header - Heap chunk header
25  */
26 struct panthor_heap_chunk_header {
27         /**
28          * @next: Next heap chunk in the list.
29          *
30          * This is a GPU VA.
31          */
32         u64 next;
33
34         /** @unknown: MBZ. */
35         u32 unknown[14];
36 };
37
38 /**
39  * struct panthor_heap_chunk - Structure used to keep track of allocated heap chunks.
40  */
41 struct panthor_heap_chunk {
42         /** @node: Used to insert the heap chunk in panthor_heap::chunks. */
43         struct list_head node;
44
45         /** @bo: Buffer object backing the heap chunk. */
46         struct panthor_kernel_bo *bo;
47 };
48
49 /**
50  * struct panthor_heap - Structure used to manage tiler heap contexts.
51  */
52 struct panthor_heap {
53         /** @chunks: List containing all heap chunks allocated so far. */
54         struct list_head chunks;
55
56         /** @lock: Lock protecting insertion in the chunks list. */
57         struct mutex lock;
58
59         /** @chunk_size: Size of each chunk. */
60         u32 chunk_size;
61
62         /** @max_chunks: Maximum number of chunks. */
63         u32 max_chunks;
64
65         /**
66          * @target_in_flight: Number of in-flight render passes after which
67          * we'd let the FW wait for fragment job to finish instead of allocating new chunks.
68          */
69         u32 target_in_flight;
70
71         /** @chunk_count: Number of heap chunks currently allocated. */
72         u32 chunk_count;
73 };
74
75 #define MAX_HEAPS_PER_POOL    128
76
77 /**
78  * struct panthor_heap_pool - Pool of heap contexts
79  *
80  * The pool is attached to a panthor_file and can't be shared across processes.
81  */
82 struct panthor_heap_pool {
83         /** @refcount: Reference count. */
84         struct kref refcount;
85
86         /** @ptdev: Device. */
87         struct panthor_device *ptdev;
88
89         /** @vm: VM this pool is bound to. */
90         struct panthor_vm *vm;
91
92         /** @lock: Lock protecting access to @xa. */
93         struct rw_semaphore lock;
94
95         /** @xa: Array storing panthor_heap objects. */
96         struct xarray xa;
97
98         /** @gpu_contexts: Buffer object containing the GPU heap contexts. */
99         struct panthor_kernel_bo *gpu_contexts;
100 };
101
102 static int panthor_heap_ctx_stride(struct panthor_device *ptdev)
103 {
104         u32 l2_features = ptdev->gpu_info.l2_features;
105         u32 gpu_cache_line_size = GPU_L2_FEATURES_LINE_SIZE(l2_features);
106
107         return ALIGN(HEAP_CONTEXT_SIZE, gpu_cache_line_size);
108 }
109
110 static int panthor_get_heap_ctx_offset(struct panthor_heap_pool *pool, int id)
111 {
112         return panthor_heap_ctx_stride(pool->ptdev) * id;
113 }
114
115 static void *panthor_get_heap_ctx(struct panthor_heap_pool *pool, int id)
116 {
117         return pool->gpu_contexts->kmap +
118                panthor_get_heap_ctx_offset(pool, id);
119 }
120
121 static void panthor_free_heap_chunk(struct panthor_vm *vm,
122                                     struct panthor_heap *heap,
123                                     struct panthor_heap_chunk *chunk)
124 {
125         mutex_lock(&heap->lock);
126         list_del(&chunk->node);
127         heap->chunk_count--;
128         mutex_unlock(&heap->lock);
129
130         panthor_kernel_bo_destroy(vm, chunk->bo);
131         kfree(chunk);
132 }
133
134 static int panthor_alloc_heap_chunk(struct panthor_device *ptdev,
135                                     struct panthor_vm *vm,
136                                     struct panthor_heap *heap,
137                                     bool initial_chunk)
138 {
139         struct panthor_heap_chunk *chunk;
140         struct panthor_heap_chunk_header *hdr;
141         int ret;
142
143         chunk = kmalloc(sizeof(*chunk), GFP_KERNEL);
144         if (!chunk)
145                 return -ENOMEM;
146
147         chunk->bo = panthor_kernel_bo_create(ptdev, vm, heap->chunk_size,
148                                              DRM_PANTHOR_BO_NO_MMAP,
149                                              DRM_PANTHOR_VM_BIND_OP_MAP_NOEXEC,
150                                              PANTHOR_VM_KERNEL_AUTO_VA);
151         if (IS_ERR(chunk->bo)) {
152                 ret = PTR_ERR(chunk->bo);
153                 goto err_free_chunk;
154         }
155
156         ret = panthor_kernel_bo_vmap(chunk->bo);
157         if (ret)
158                 goto err_destroy_bo;
159
160         hdr = chunk->bo->kmap;
161         memset(hdr, 0, sizeof(*hdr));
162
163         if (initial_chunk && !list_empty(&heap->chunks)) {
164                 struct panthor_heap_chunk *prev_chunk;
165                 u64 prev_gpuva;
166
167                 prev_chunk = list_first_entry(&heap->chunks,
168                                               struct panthor_heap_chunk,
169                                               node);
170
171                 prev_gpuva = panthor_kernel_bo_gpuva(prev_chunk->bo);
172                 hdr->next = (prev_gpuva & GENMASK_ULL(63, 12)) |
173                             (heap->chunk_size >> 12);
174         }
175
176         panthor_kernel_bo_vunmap(chunk->bo);
177
178         mutex_lock(&heap->lock);
179         list_add(&chunk->node, &heap->chunks);
180         heap->chunk_count++;
181         mutex_unlock(&heap->lock);
182
183         return 0;
184
185 err_destroy_bo:
186         panthor_kernel_bo_destroy(vm, chunk->bo);
187
188 err_free_chunk:
189         kfree(chunk);
190
191         return ret;
192 }
193
194 static void panthor_free_heap_chunks(struct panthor_vm *vm,
195                                      struct panthor_heap *heap)
196 {
197         struct panthor_heap_chunk *chunk, *tmp;
198
199         list_for_each_entry_safe(chunk, tmp, &heap->chunks, node)
200                 panthor_free_heap_chunk(vm, heap, chunk);
201 }
202
203 static int panthor_alloc_heap_chunks(struct panthor_device *ptdev,
204                                      struct panthor_vm *vm,
205                                      struct panthor_heap *heap,
206                                      u32 chunk_count)
207 {
208         int ret;
209         u32 i;
210
211         for (i = 0; i < chunk_count; i++) {
212                 ret = panthor_alloc_heap_chunk(ptdev, vm, heap, true);
213                 if (ret)
214                         return ret;
215         }
216
217         return 0;
218 }
219
220 static int
221 panthor_heap_destroy_locked(struct panthor_heap_pool *pool, u32 handle)
222 {
223         struct panthor_heap *heap;
224
225         heap = xa_erase(&pool->xa, handle);
226         if (!heap)
227                 return -EINVAL;
228
229         panthor_free_heap_chunks(pool->vm, heap);
230         mutex_destroy(&heap->lock);
231         kfree(heap);
232         return 0;
233 }
234
235 /**
236  * panthor_heap_destroy() - Destroy a heap context
237  * @pool: Pool this context belongs to.
238  * @handle: Handle returned by panthor_heap_create().
239  */
240 int panthor_heap_destroy(struct panthor_heap_pool *pool, u32 handle)
241 {
242         int ret;
243
244         down_write(&pool->lock);
245         ret = panthor_heap_destroy_locked(pool, handle);
246         up_write(&pool->lock);
247
248         return ret;
249 }
250
251 /**
252  * panthor_heap_create() - Create a heap context
253  * @pool: Pool to instantiate the heap context from.
254  * @initial_chunk_count: Number of chunk allocated at initialization time.
255  * Must be at least 1.
256  * @chunk_size: The size of each chunk. Must be a power of two between 256k
257  * and 2M.
258  * @max_chunks: Maximum number of chunks that can be allocated.
259  * @target_in_flight: Maximum number of in-flight render passes.
260  * @heap_ctx_gpu_va: Pointer holding the GPU address of the allocated heap
261  * context.
262  * @first_chunk_gpu_va: Pointer holding the GPU address of the first chunk
263  * assigned to the heap context.
264  *
265  * Return: a positive handle on success, a negative error otherwise.
266  */
267 int panthor_heap_create(struct panthor_heap_pool *pool,
268                         u32 initial_chunk_count,
269                         u32 chunk_size,
270                         u32 max_chunks,
271                         u32 target_in_flight,
272                         u64 *heap_ctx_gpu_va,
273                         u64 *first_chunk_gpu_va)
274 {
275         struct panthor_heap *heap;
276         struct panthor_heap_chunk *first_chunk;
277         struct panthor_vm *vm;
278         int ret = 0;
279         u32 id;
280
281         if (initial_chunk_count == 0)
282                 return -EINVAL;
283
284         if (hweight32(chunk_size) != 1 ||
285             chunk_size < SZ_256K || chunk_size > SZ_2M)
286                 return -EINVAL;
287
288         down_read(&pool->lock);
289         vm = panthor_vm_get(pool->vm);
290         up_read(&pool->lock);
291
292         /* The pool has been destroyed, we can't create a new heap. */
293         if (!vm)
294                 return -EINVAL;
295
296         heap = kzalloc(sizeof(*heap), GFP_KERNEL);
297         if (!heap) {
298                 ret = -ENOMEM;
299                 goto err_put_vm;
300         }
301
302         mutex_init(&heap->lock);
303         INIT_LIST_HEAD(&heap->chunks);
304         heap->chunk_size = chunk_size;
305         heap->max_chunks = max_chunks;
306         heap->target_in_flight = target_in_flight;
307
308         ret = panthor_alloc_heap_chunks(pool->ptdev, vm, heap,
309                                         initial_chunk_count);
310         if (ret)
311                 goto err_free_heap;
312
313         first_chunk = list_first_entry(&heap->chunks,
314                                        struct panthor_heap_chunk,
315                                        node);
316         *first_chunk_gpu_va = panthor_kernel_bo_gpuva(first_chunk->bo);
317
318         down_write(&pool->lock);
319         /* The pool has been destroyed, we can't create a new heap. */
320         if (!pool->vm) {
321                 ret = -EINVAL;
322         } else {
323                 ret = xa_alloc(&pool->xa, &id, heap, XA_LIMIT(1, MAX_HEAPS_PER_POOL), GFP_KERNEL);
324                 if (!ret) {
325                         void *gpu_ctx = panthor_get_heap_ctx(pool, id);
326
327                         memset(gpu_ctx, 0, panthor_heap_ctx_stride(pool->ptdev));
328                         *heap_ctx_gpu_va = panthor_kernel_bo_gpuva(pool->gpu_contexts) +
329                                            panthor_get_heap_ctx_offset(pool, id);
330                 }
331         }
332         up_write(&pool->lock);
333
334         if (ret)
335                 goto err_free_heap;
336
337         panthor_vm_put(vm);
338         return id;
339
340 err_free_heap:
341         panthor_free_heap_chunks(pool->vm, heap);
342         mutex_destroy(&heap->lock);
343         kfree(heap);
344
345 err_put_vm:
346         panthor_vm_put(vm);
347         return ret;
348 }
349
350 /**
351  * panthor_heap_return_chunk() - Return an unused heap chunk
352  * @pool: The pool this heap belongs to.
353  * @heap_gpu_va: The GPU address of the heap context.
354  * @chunk_gpu_va: The chunk VA to return.
355  *
356  * This function is used when a chunk allocated with panthor_heap_grow()
357  * couldn't be linked to the heap context through the FW interface because
358  * the group requesting the allocation was scheduled out in the meantime.
359  */
360 int panthor_heap_return_chunk(struct panthor_heap_pool *pool,
361                               u64 heap_gpu_va,
362                               u64 chunk_gpu_va)
363 {
364         u64 offset = heap_gpu_va - panthor_kernel_bo_gpuva(pool->gpu_contexts);
365         u32 heap_id = (u32)offset / panthor_heap_ctx_stride(pool->ptdev);
366         struct panthor_heap_chunk *chunk, *tmp, *removed = NULL;
367         struct panthor_heap *heap;
368         int ret;
369
370         if (offset > U32_MAX || heap_id >= MAX_HEAPS_PER_POOL)
371                 return -EINVAL;
372
373         down_read(&pool->lock);
374         heap = xa_load(&pool->xa, heap_id);
375         if (!heap) {
376                 ret = -EINVAL;
377                 goto out_unlock;
378         }
379
380         chunk_gpu_va &= GENMASK_ULL(63, 12);
381
382         mutex_lock(&heap->lock);
383         list_for_each_entry_safe(chunk, tmp, &heap->chunks, node) {
384                 if (panthor_kernel_bo_gpuva(chunk->bo) == chunk_gpu_va) {
385                         removed = chunk;
386                         list_del(&chunk->node);
387                         heap->chunk_count--;
388                         break;
389                 }
390         }
391         mutex_unlock(&heap->lock);
392
393         if (removed) {
394                 panthor_kernel_bo_destroy(pool->vm, chunk->bo);
395                 kfree(chunk);
396                 ret = 0;
397         } else {
398                 ret = -EINVAL;
399         }
400
401 out_unlock:
402         up_read(&pool->lock);
403         return ret;
404 }
405
406 /**
407  * panthor_heap_grow() - Make a heap context grow.
408  * @pool: The pool this heap belongs to.
409  * @heap_gpu_va: The GPU address of the heap context.
410  * @renderpasses_in_flight: Number of render passes currently in-flight.
411  * @pending_frag_count: Number of fragment jobs waiting for execution/completion.
412  * @new_chunk_gpu_va: Pointer used to return the chunk VA.
413  */
414 int panthor_heap_grow(struct panthor_heap_pool *pool,
415                       u64 heap_gpu_va,
416                       u32 renderpasses_in_flight,
417                       u32 pending_frag_count,
418                       u64 *new_chunk_gpu_va)
419 {
420         u64 offset = heap_gpu_va - panthor_kernel_bo_gpuva(pool->gpu_contexts);
421         u32 heap_id = (u32)offset / panthor_heap_ctx_stride(pool->ptdev);
422         struct panthor_heap_chunk *chunk;
423         struct panthor_heap *heap;
424         int ret;
425
426         if (offset > U32_MAX || heap_id >= MAX_HEAPS_PER_POOL)
427                 return -EINVAL;
428
429         down_read(&pool->lock);
430         heap = xa_load(&pool->xa, heap_id);
431         if (!heap) {
432                 ret = -EINVAL;
433                 goto out_unlock;
434         }
435
436         /* If we reached the target in-flight render passes, or if we
437          * reached the maximum number of chunks, let the FW figure another way to
438          * find some memory (wait for render passes to finish, or call the exception
439          * handler provided by the userspace driver, if any).
440          */
441         if (renderpasses_in_flight > heap->target_in_flight ||
442             (pending_frag_count > 0 && heap->chunk_count >= heap->max_chunks)) {
443                 ret = -EBUSY;
444                 goto out_unlock;
445         } else if (heap->chunk_count >= heap->max_chunks) {
446                 ret = -ENOMEM;
447                 goto out_unlock;
448         }
449
450         /* FIXME: panthor_alloc_heap_chunk() triggers a kernel BO creation,
451          * which goes through the blocking allocation path. Ultimately, we
452          * want a non-blocking allocation, so we can immediately report to the
453          * FW when the system is running out of memory. In that case, the FW
454          * can call a user-provided exception handler, which might try to free
455          * some tiler memory by issuing an intermediate fragment job. If the
456          * exception handler can't do anything, it will flag the queue as
457          * faulty so the job that triggered this tiler chunk allocation and all
458          * further jobs in this queue fail immediately instead of having to
459          * wait for the job timeout.
460          */
461         ret = panthor_alloc_heap_chunk(pool->ptdev, pool->vm, heap, false);
462         if (ret)
463                 goto out_unlock;
464
465         chunk = list_first_entry(&heap->chunks,
466                                  struct panthor_heap_chunk,
467                                  node);
468         *new_chunk_gpu_va = (panthor_kernel_bo_gpuva(chunk->bo) & GENMASK_ULL(63, 12)) |
469                             (heap->chunk_size >> 12);
470         ret = 0;
471
472 out_unlock:
473         up_read(&pool->lock);
474         return ret;
475 }
476
477 static void panthor_heap_pool_release(struct kref *refcount)
478 {
479         struct panthor_heap_pool *pool =
480                 container_of(refcount, struct panthor_heap_pool, refcount);
481
482         xa_destroy(&pool->xa);
483         kfree(pool);
484 }
485
486 /**
487  * panthor_heap_pool_put() - Release a heap pool reference
488  * @pool: Pool to release the reference on. Can be NULL.
489  */
490 void panthor_heap_pool_put(struct panthor_heap_pool *pool)
491 {
492         if (pool)
493                 kref_put(&pool->refcount, panthor_heap_pool_release);
494 }
495
496 /**
497  * panthor_heap_pool_get() - Get a heap pool reference
498  * @pool: Pool to get the reference on. Can be NULL.
499  *
500  * Return: @pool.
501  */
502 struct panthor_heap_pool *
503 panthor_heap_pool_get(struct panthor_heap_pool *pool)
504 {
505         if (pool)
506                 kref_get(&pool->refcount);
507
508         return pool;
509 }
510
511 /**
512  * panthor_heap_pool_create() - Create a heap pool
513  * @ptdev: Device.
514  * @vm: The VM this heap pool will be attached to.
515  *
516  * Heap pools might contain up to 128 heap contexts, and are per-VM.
517  *
518  * Return: A valid pointer on success, a negative error code otherwise.
519  */
520 struct panthor_heap_pool *
521 panthor_heap_pool_create(struct panthor_device *ptdev, struct panthor_vm *vm)
522 {
523         size_t bosize = ALIGN(MAX_HEAPS_PER_POOL *
524                               panthor_heap_ctx_stride(ptdev),
525                               4096);
526         struct panthor_heap_pool *pool;
527         int ret = 0;
528
529         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
530         if (!pool)
531                 return ERR_PTR(-ENOMEM);
532
533         /* We want a weak ref here: the heap pool belongs to the VM, so we're
534          * sure that, as long as the heap pool exists, the VM exists too.
535          */
536         pool->vm = vm;
537         pool->ptdev = ptdev;
538         init_rwsem(&pool->lock);
539         xa_init_flags(&pool->xa, XA_FLAGS_ALLOC1);
540         kref_init(&pool->refcount);
541
542         pool->gpu_contexts = panthor_kernel_bo_create(ptdev, vm, bosize,
543                                                       DRM_PANTHOR_BO_NO_MMAP,
544                                                       DRM_PANTHOR_VM_BIND_OP_MAP_NOEXEC,
545                                                       PANTHOR_VM_KERNEL_AUTO_VA);
546         if (IS_ERR(pool->gpu_contexts)) {
547                 ret = PTR_ERR(pool->gpu_contexts);
548                 goto err_destroy_pool;
549         }
550
551         ret = panthor_kernel_bo_vmap(pool->gpu_contexts);
552         if (ret)
553                 goto err_destroy_pool;
554
555         return pool;
556
557 err_destroy_pool:
558         panthor_heap_pool_destroy(pool);
559         return ERR_PTR(ret);
560 }
561
562 /**
563  * panthor_heap_pool_destroy() - Destroy a heap pool.
564  * @pool: Pool to destroy.
565  *
566  * This function destroys all heap contexts and their resources. Thus
567  * preventing any use of the heap context or the chunk attached to them
568  * after that point.
569  *
570  * If the GPU still has access to some heap contexts, a fault should be
571  * triggered, which should flag the command stream groups using these
572  * context as faulty.
573  *
574  * The heap pool object is only released when all references to this pool
575  * are released.
576  */
577 void panthor_heap_pool_destroy(struct panthor_heap_pool *pool)
578 {
579         struct panthor_heap *heap;
580         unsigned long i;
581
582         if (!pool)
583                 return;
584
585         down_write(&pool->lock);
586         xa_for_each(&pool->xa, i, heap)
587                 drm_WARN_ON(&pool->ptdev->base, panthor_heap_destroy_locked(pool, i));
588
589         if (!IS_ERR_OR_NULL(pool->gpu_contexts))
590                 panthor_kernel_bo_destroy(pool->vm, pool->gpu_contexts);
591
592         /* Reflects the fact the pool has been destroyed. */
593         pool->vm = NULL;
594         up_write(&pool->lock);
595
596         panthor_heap_pool_put(pool);
597 }
This page took 0.069192 seconds and 4 git commands to generate.