xref: /linux/drivers/gpu/drm/panthor/panthor_gem.c (revision bf4afc53b77aeaa48b5409da5c8da6bb4eff7f43)
1 // SPDX-License-Identifier: GPL-2.0 or MIT
2 /* Copyright 2019 Linaro, Ltd, Rob Herring <robh@kernel.org> */
3 /* Copyright 2023 Collabora ltd. */
4 /* Copyright 2025 Amazon.com, Inc. or its affiliates */
5 
6 #include <linux/cleanup.h>
7 #include <linux/dma-buf.h>
8 #include <linux/dma-mapping.h>
9 #include <linux/err.h>
10 #include <linux/slab.h>
11 
12 #include <drm/drm_print.h>
13 #include <drm/panthor_drm.h>
14 
15 #include "panthor_device.h"
16 #include "panthor_drv.h"
17 #include "panthor_fw.h"
18 #include "panthor_gem.h"
19 #include "panthor_mmu.h"
20 
panthor_gem_init(struct panthor_device * ptdev)21 void panthor_gem_init(struct panthor_device *ptdev)
22 {
23 	int err;
24 
25 	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
26 	    !panthor_transparent_hugepage)
27 		return;
28 
29 	err = drm_gem_huge_mnt_create(&ptdev->base, "within_size");
30 	if (drm_gem_get_huge_mnt(&ptdev->base))
31 		drm_info(&ptdev->base, "Using Transparent Hugepage\n");
32 	else if (err)
33 		drm_warn(&ptdev->base, "Can't use Transparent Hugepage (%d)\n",
34 			 err);
35 }
36 
37 #ifdef CONFIG_DEBUG_FS
panthor_gem_debugfs_bo_init(struct panthor_gem_object * bo)38 static void panthor_gem_debugfs_bo_init(struct panthor_gem_object *bo)
39 {
40 	INIT_LIST_HEAD(&bo->debugfs.node);
41 }
42 
panthor_gem_debugfs_bo_add(struct panthor_gem_object * bo)43 static void panthor_gem_debugfs_bo_add(struct panthor_gem_object *bo)
44 {
45 	struct panthor_device *ptdev = container_of(bo->base.base.dev,
46 						    struct panthor_device, base);
47 
48 	bo->debugfs.creator.tgid = current->tgid;
49 	get_task_comm(bo->debugfs.creator.process_name, current->group_leader);
50 
51 	mutex_lock(&ptdev->gems.lock);
52 	list_add_tail(&bo->debugfs.node, &ptdev->gems.node);
53 	mutex_unlock(&ptdev->gems.lock);
54 }
55 
panthor_gem_debugfs_bo_rm(struct panthor_gem_object * bo)56 static void panthor_gem_debugfs_bo_rm(struct panthor_gem_object *bo)
57 {
58 	struct panthor_device *ptdev = container_of(bo->base.base.dev,
59 						    struct panthor_device, base);
60 
61 	if (list_empty(&bo->debugfs.node))
62 		return;
63 
64 	mutex_lock(&ptdev->gems.lock);
65 	list_del_init(&bo->debugfs.node);
66 	mutex_unlock(&ptdev->gems.lock);
67 }
68 
panthor_gem_debugfs_set_usage_flags(struct panthor_gem_object * bo,u32 usage_flags)69 static void panthor_gem_debugfs_set_usage_flags(struct panthor_gem_object *bo, u32 usage_flags)
70 {
71 	bo->debugfs.flags = usage_flags;
72 	panthor_gem_debugfs_bo_add(bo);
73 }
74 #else
panthor_gem_debugfs_bo_rm(struct panthor_gem_object * bo)75 static void panthor_gem_debugfs_bo_rm(struct panthor_gem_object *bo) {}
panthor_gem_debugfs_set_usage_flags(struct panthor_gem_object * bo,u32 usage_flags)76 static void panthor_gem_debugfs_set_usage_flags(struct panthor_gem_object *bo, u32 usage_flags) {}
panthor_gem_debugfs_bo_init(struct panthor_gem_object * bo)77 static void panthor_gem_debugfs_bo_init(struct panthor_gem_object *bo) {}
78 #endif
79 
80 static bool
should_map_wc(struct panthor_gem_object * bo,struct panthor_vm * exclusive_vm)81 should_map_wc(struct panthor_gem_object *bo, struct panthor_vm *exclusive_vm)
82 {
83 	struct panthor_device *ptdev = container_of(bo->base.base.dev, struct panthor_device, base);
84 
85 	/* We can't do uncached mappings if the device is coherent,
86 	 * because the zeroing done by the shmem layer at page allocation
87 	 * time happens on a cached mapping which isn't CPU-flushed (at least
88 	 * not on Arm64 where the flush is deferred to PTE setup time, and
89 	 * only done conditionally based on the mapping permissions). We can't
90 	 * rely on dma_map_sgtable()/dma_sync_sgtable_for_xxx() either to flush
91 	 * those, because they are NOPed if dma_dev_coherent() returns true.
92 	 *
93 	 * FIXME: Note that this problem is going to pop up again when we
94 	 * decide to support mapping buffers with the NO_MMAP flag as
95 	 * non-shareable (AKA buffers accessed only by the GPU), because we
96 	 * need the same CPU flush to happen after page allocation, otherwise
97 	 * there's a risk of data leak or late corruption caused by a dirty
98 	 * cacheline being evicted. At this point we'll need a way to force
99 	 * CPU cache maintenance regardless of whether the device is coherent
100 	 * or not.
101 	 */
102 	if (ptdev->coherent)
103 		return false;
104 
105 	/* Cached mappings are explicitly requested, so no write-combine. */
106 	if (bo->flags & DRM_PANTHOR_BO_WB_MMAP)
107 		return false;
108 
109 	/* The default is write-combine. */
110 	return true;
111 }
112 
panthor_gem_free_object(struct drm_gem_object * obj)113 static void panthor_gem_free_object(struct drm_gem_object *obj)
114 {
115 	struct panthor_gem_object *bo = to_panthor_bo(obj);
116 	struct drm_gem_object *vm_root_gem = bo->exclusive_vm_root_gem;
117 
118 	panthor_gem_debugfs_bo_rm(bo);
119 
120 	/*
121 	 * Label might have been allocated with kstrdup_const(),
122 	 * we need to take that into account when freeing the memory
123 	 */
124 	kfree_const(bo->label.str);
125 
126 	mutex_destroy(&bo->label.lock);
127 
128 	drm_gem_free_mmap_offset(&bo->base.base);
129 	drm_gem_shmem_free(&bo->base);
130 	drm_gem_object_put(vm_root_gem);
131 }
132 
133 /**
134  * panthor_kernel_bo_destroy() - Destroy a kernel buffer object
135  * @bo: Kernel buffer object to destroy. If NULL or an ERR_PTR(), the destruction
136  * is skipped.
137  */
panthor_kernel_bo_destroy(struct panthor_kernel_bo * bo)138 void panthor_kernel_bo_destroy(struct panthor_kernel_bo *bo)
139 {
140 	struct panthor_vm *vm;
141 
142 	if (IS_ERR_OR_NULL(bo))
143 		return;
144 
145 	vm = bo->vm;
146 	panthor_kernel_bo_vunmap(bo);
147 
148 	drm_WARN_ON(bo->obj->dev,
149 		    to_panthor_bo(bo->obj)->exclusive_vm_root_gem != panthor_vm_root_gem(vm));
150 	panthor_vm_unmap_range(vm, bo->va_node.start, bo->va_node.size);
151 	panthor_vm_free_va(vm, &bo->va_node);
152 	drm_gem_object_put(bo->obj);
153 	panthor_vm_put(vm);
154 	kfree(bo);
155 }
156 
157 /**
158  * panthor_kernel_bo_create() - Create and map a GEM object to a VM
159  * @ptdev: Device.
160  * @vm: VM to map the GEM to. If NULL, the kernel object is not GPU mapped.
161  * @size: Size of the buffer object.
162  * @bo_flags: Combination of drm_panthor_bo_flags flags.
163  * @vm_map_flags: Combination of drm_panthor_vm_bind_op_flags (only those
164  * that are related to map operations).
165  * @gpu_va: GPU address assigned when mapping to the VM.
166  * If gpu_va == PANTHOR_VM_KERNEL_AUTO_VA, the virtual address will be
167  * automatically allocated.
168  * @name: Descriptive label of the BO's contents
169  *
170  * Return: A valid pointer in case of success, an ERR_PTR() otherwise.
171  */
172 struct panthor_kernel_bo *
panthor_kernel_bo_create(struct panthor_device * ptdev,struct panthor_vm * vm,size_t size,u32 bo_flags,u32 vm_map_flags,u64 gpu_va,const char * name)173 panthor_kernel_bo_create(struct panthor_device *ptdev, struct panthor_vm *vm,
174 			 size_t size, u32 bo_flags, u32 vm_map_flags,
175 			 u64 gpu_va, const char *name)
176 {
177 	struct drm_gem_shmem_object *obj;
178 	struct panthor_kernel_bo *kbo;
179 	struct panthor_gem_object *bo;
180 	u32 debug_flags = PANTHOR_DEBUGFS_GEM_USAGE_FLAG_KERNEL;
181 	int ret;
182 
183 	if (drm_WARN_ON(&ptdev->base, !vm))
184 		return ERR_PTR(-EINVAL);
185 
186 	kbo = kzalloc_obj(*kbo);
187 	if (!kbo)
188 		return ERR_PTR(-ENOMEM);
189 
190 	obj = drm_gem_shmem_create(&ptdev->base, size);
191 	if (IS_ERR(obj)) {
192 		ret = PTR_ERR(obj);
193 		goto err_free_bo;
194 	}
195 
196 	bo = to_panthor_bo(&obj->base);
197 	kbo->obj = &obj->base;
198 	bo->flags = bo_flags;
199 	bo->base.map_wc = should_map_wc(bo, vm);
200 	bo->exclusive_vm_root_gem = panthor_vm_root_gem(vm);
201 	drm_gem_object_get(bo->exclusive_vm_root_gem);
202 	bo->base.base.resv = bo->exclusive_vm_root_gem->resv;
203 
204 	if (vm == panthor_fw_vm(ptdev))
205 		debug_flags |= PANTHOR_DEBUGFS_GEM_USAGE_FLAG_FW_MAPPED;
206 
207 	panthor_gem_kernel_bo_set_label(kbo, name);
208 	panthor_gem_debugfs_set_usage_flags(to_panthor_bo(kbo->obj), debug_flags);
209 
210 	/* The system and GPU MMU page size might differ, which becomes a
211 	 * problem for FW sections that need to be mapped at explicit address
212 	 * since our PAGE_SIZE alignment might cover a VA range that's
213 	 * expected to be used for another section.
214 	 * Make sure we never map more than we need.
215 	 */
216 	size = ALIGN(size, panthor_vm_page_size(vm));
217 	ret = panthor_vm_alloc_va(vm, gpu_va, size, &kbo->va_node);
218 	if (ret)
219 		goto err_put_obj;
220 
221 	ret = panthor_vm_map_bo_range(vm, bo, 0, size, kbo->va_node.start, vm_map_flags);
222 	if (ret)
223 		goto err_free_va;
224 
225 	kbo->vm = panthor_vm_get(vm);
226 	return kbo;
227 
228 err_free_va:
229 	panthor_vm_free_va(vm, &kbo->va_node);
230 
231 err_put_obj:
232 	drm_gem_object_put(&obj->base);
233 
234 err_free_bo:
235 	kfree(kbo);
236 	return ERR_PTR(ret);
237 }
238 
239 static struct sg_table *
panthor_gem_prime_map_dma_buf(struct dma_buf_attachment * attach,enum dma_data_direction dir)240 panthor_gem_prime_map_dma_buf(struct dma_buf_attachment *attach,
241 			      enum dma_data_direction dir)
242 {
243 	struct sg_table *sgt = drm_gem_map_dma_buf(attach, dir);
244 
245 	if (!IS_ERR(sgt))
246 		attach->priv = sgt;
247 
248 	return sgt;
249 }
250 
251 static void
panthor_gem_prime_unmap_dma_buf(struct dma_buf_attachment * attach,struct sg_table * sgt,enum dma_data_direction dir)252 panthor_gem_prime_unmap_dma_buf(struct dma_buf_attachment *attach,
253 				struct sg_table *sgt,
254 				enum dma_data_direction dir)
255 {
256 	attach->priv = NULL;
257 	drm_gem_unmap_dma_buf(attach, sgt, dir);
258 }
259 
260 static int
panthor_gem_prime_begin_cpu_access(struct dma_buf * dma_buf,enum dma_data_direction dir)261 panthor_gem_prime_begin_cpu_access(struct dma_buf *dma_buf,
262 				   enum dma_data_direction dir)
263 {
264 	struct drm_gem_object *obj = dma_buf->priv;
265 	struct drm_device *dev = obj->dev;
266 	struct drm_gem_shmem_object *shmem = to_drm_gem_shmem_obj(obj);
267 	struct dma_buf_attachment *attach;
268 
269 	dma_resv_lock(obj->resv, NULL);
270 	if (shmem->sgt)
271 		dma_sync_sgtable_for_cpu(dev->dev, shmem->sgt, dir);
272 
273 	if (shmem->vaddr)
274 		invalidate_kernel_vmap_range(shmem->vaddr, shmem->base.size);
275 
276 	list_for_each_entry(attach, &dma_buf->attachments, node) {
277 		struct sg_table *sgt = attach->priv;
278 
279 		if (sgt)
280 			dma_sync_sgtable_for_cpu(attach->dev, sgt, dir);
281 	}
282 	dma_resv_unlock(obj->resv);
283 
284 	return 0;
285 }
286 
287 static int
panthor_gem_prime_end_cpu_access(struct dma_buf * dma_buf,enum dma_data_direction dir)288 panthor_gem_prime_end_cpu_access(struct dma_buf *dma_buf,
289 				 enum dma_data_direction dir)
290 {
291 	struct drm_gem_object *obj = dma_buf->priv;
292 	struct drm_device *dev = obj->dev;
293 	struct drm_gem_shmem_object *shmem = to_drm_gem_shmem_obj(obj);
294 	struct dma_buf_attachment *attach;
295 
296 	dma_resv_lock(obj->resv, NULL);
297 	list_for_each_entry(attach, &dma_buf->attachments, node) {
298 		struct sg_table *sgt = attach->priv;
299 
300 		if (sgt)
301 			dma_sync_sgtable_for_device(attach->dev, sgt, dir);
302 	}
303 
304 	if (shmem->vaddr)
305 		flush_kernel_vmap_range(shmem->vaddr, shmem->base.size);
306 
307 	if (shmem->sgt)
308 		dma_sync_sgtable_for_device(dev->dev, shmem->sgt, dir);
309 
310 	dma_resv_unlock(obj->resv);
311 	return 0;
312 }
313 
314 static const struct dma_buf_ops panthor_dma_buf_ops = {
315 	.attach = drm_gem_map_attach,
316 	.detach = drm_gem_map_detach,
317 	.map_dma_buf = panthor_gem_prime_map_dma_buf,
318 	.unmap_dma_buf = panthor_gem_prime_unmap_dma_buf,
319 	.release = drm_gem_dmabuf_release,
320 	.mmap = drm_gem_dmabuf_mmap,
321 	.vmap = drm_gem_dmabuf_vmap,
322 	.vunmap = drm_gem_dmabuf_vunmap,
323 	.begin_cpu_access = panthor_gem_prime_begin_cpu_access,
324 	.end_cpu_access = panthor_gem_prime_end_cpu_access,
325 };
326 
327 static struct dma_buf *
panthor_gem_prime_export(struct drm_gem_object * obj,int flags)328 panthor_gem_prime_export(struct drm_gem_object *obj, int flags)
329 {
330 	struct drm_device *dev = obj->dev;
331 	struct dma_buf_export_info exp_info = {
332 		.exp_name = KBUILD_MODNAME,
333 		.owner = THIS_MODULE,
334 		.ops = &panthor_dma_buf_ops,
335 		.size = obj->size,
336 		.flags = flags,
337 		.priv = obj,
338 		.resv = obj->resv,
339 	};
340 
341 	/* We can't export GEMs that have an exclusive VM. */
342 	if (to_panthor_bo(obj)->exclusive_vm_root_gem)
343 		return ERR_PTR(-EINVAL);
344 
345 	return drm_gem_dmabuf_export(dev, &exp_info);
346 }
347 
348 struct drm_gem_object *
panthor_gem_prime_import(struct drm_device * dev,struct dma_buf * dma_buf)349 panthor_gem_prime_import(struct drm_device *dev,
350 			 struct dma_buf *dma_buf)
351 {
352 	struct drm_gem_object *obj = dma_buf->priv;
353 
354 	if (dma_buf->ops == &panthor_dma_buf_ops && obj->dev == dev) {
355 		/* Importing dmabuf exported from our own gem increases
356 		 * refcount on gem itself instead of f_count of dmabuf.
357 		 */
358 		drm_gem_object_get(obj);
359 		return obj;
360 	}
361 
362 	return drm_gem_prime_import(dev, dma_buf);
363 }
364 
panthor_gem_status(struct drm_gem_object * obj)365 static enum drm_gem_object_status panthor_gem_status(struct drm_gem_object *obj)
366 {
367 	struct panthor_gem_object *bo = to_panthor_bo(obj);
368 	enum drm_gem_object_status res = 0;
369 
370 	if (drm_gem_is_imported(&bo->base.base) || bo->base.pages)
371 		res |= DRM_GEM_OBJECT_RESIDENT;
372 
373 	return res;
374 }
375 
376 static const struct drm_gem_object_funcs panthor_gem_funcs = {
377 	.free = panthor_gem_free_object,
378 	.print_info = drm_gem_shmem_object_print_info,
379 	.pin = drm_gem_shmem_object_pin,
380 	.unpin = drm_gem_shmem_object_unpin,
381 	.get_sg_table = drm_gem_shmem_object_get_sg_table,
382 	.vmap = drm_gem_shmem_object_vmap,
383 	.vunmap = drm_gem_shmem_object_vunmap,
384 	.mmap = drm_gem_shmem_object_mmap,
385 	.status = panthor_gem_status,
386 	.export = panthor_gem_prime_export,
387 	.vm_ops = &drm_gem_shmem_vm_ops,
388 };
389 
390 /**
391  * panthor_gem_create_object - Implementation of driver->gem_create_object.
392  * @ddev: DRM device
393  * @size: Size in bytes of the memory the object will reference
394  *
395  * This lets the GEM helpers allocate object structs for us, and keep
396  * our BO stats correct.
397  */
panthor_gem_create_object(struct drm_device * ddev,size_t size)398 struct drm_gem_object *panthor_gem_create_object(struct drm_device *ddev, size_t size)
399 {
400 	struct panthor_gem_object *obj;
401 
402 	obj = kzalloc_obj(*obj);
403 	if (!obj)
404 		return ERR_PTR(-ENOMEM);
405 
406 	obj->base.base.funcs = &panthor_gem_funcs;
407 	mutex_init(&obj->label.lock);
408 
409 	panthor_gem_debugfs_bo_init(obj);
410 
411 	return &obj->base.base;
412 }
413 
414 /**
415  * panthor_gem_create_with_handle() - Create a GEM object and attach it to a handle.
416  * @file: DRM file.
417  * @ddev: DRM device.
418  * @exclusive_vm: Exclusive VM. Not NULL if the GEM object can't be shared.
419  * @size: Size of the GEM object to allocate.
420  * @flags: Combination of drm_panthor_bo_flags flags.
421  * @handle: Pointer holding the handle pointing to the new GEM object.
422  *
423  * Return: Zero on success
424  */
425 int
panthor_gem_create_with_handle(struct drm_file * file,struct drm_device * ddev,struct panthor_vm * exclusive_vm,u64 * size,u32 flags,u32 * handle)426 panthor_gem_create_with_handle(struct drm_file *file,
427 			       struct drm_device *ddev,
428 			       struct panthor_vm *exclusive_vm,
429 			       u64 *size, u32 flags, u32 *handle)
430 {
431 	int ret;
432 	struct drm_gem_shmem_object *shmem;
433 	struct panthor_gem_object *bo;
434 
435 	shmem = drm_gem_shmem_create(ddev, *size);
436 	if (IS_ERR(shmem))
437 		return PTR_ERR(shmem);
438 
439 	bo = to_panthor_bo(&shmem->base);
440 	bo->flags = flags;
441 	bo->base.map_wc = should_map_wc(bo, exclusive_vm);
442 
443 	if (exclusive_vm) {
444 		bo->exclusive_vm_root_gem = panthor_vm_root_gem(exclusive_vm);
445 		drm_gem_object_get(bo->exclusive_vm_root_gem);
446 		bo->base.base.resv = bo->exclusive_vm_root_gem->resv;
447 	}
448 
449 	panthor_gem_debugfs_set_usage_flags(bo, 0);
450 
451 	/* If this is a write-combine mapping, we query the sgt to force a CPU
452 	 * cache flush (dma_map_sgtable() is called when the sgt is created).
453 	 * This ensures the zero-ing is visible to any uncached mapping created
454 	 * by vmap/mmap.
455 	 * FIXME: Ideally this should be done when pages are allocated, not at
456 	 * BO creation time.
457 	 */
458 	if (shmem->map_wc) {
459 		struct sg_table *sgt;
460 
461 		sgt = drm_gem_shmem_get_pages_sgt(shmem);
462 		if (IS_ERR(sgt)) {
463 			ret = PTR_ERR(sgt);
464 			goto out_put_gem;
465 		}
466 	}
467 
468 	/*
469 	 * Allocate an id of idr table where the obj is registered
470 	 * and handle has the id what user can see.
471 	 */
472 	ret = drm_gem_handle_create(file, &shmem->base, handle);
473 	if (!ret)
474 		*size = bo->base.base.size;
475 
476 out_put_gem:
477 	/* drop reference from allocate - handle holds it now. */
478 	drm_gem_object_put(&shmem->base);
479 
480 	return ret;
481 }
482 
483 void
panthor_gem_bo_set_label(struct drm_gem_object * obj,const char * label)484 panthor_gem_bo_set_label(struct drm_gem_object *obj, const char *label)
485 {
486 	struct panthor_gem_object *bo = to_panthor_bo(obj);
487 	const char *old_label;
488 
489 	scoped_guard(mutex, &bo->label.lock) {
490 		old_label = bo->label.str;
491 		bo->label.str = label;
492 	}
493 
494 	kfree_const(old_label);
495 }
496 
497 void
panthor_gem_kernel_bo_set_label(struct panthor_kernel_bo * bo,const char * label)498 panthor_gem_kernel_bo_set_label(struct panthor_kernel_bo *bo, const char *label)
499 {
500 	const char *str;
501 
502 	/* We should never attempt labelling a UM-exposed GEM object */
503 	if (drm_WARN_ON(bo->obj->dev, bo->obj->handle_count > 0))
504 		return;
505 
506 	if (!label)
507 		return;
508 
509 	str = kstrdup_const(label, GFP_KERNEL);
510 	if (!str) {
511 		/* Failing to allocate memory for a label isn't a fatal condition */
512 		drm_warn(bo->obj->dev, "Not enough memory to allocate BO label");
513 		return;
514 	}
515 
516 	panthor_gem_bo_set_label(bo->obj, str);
517 }
518 
519 int
panthor_gem_sync(struct drm_gem_object * obj,u32 type,u64 offset,u64 size)520 panthor_gem_sync(struct drm_gem_object *obj, u32 type,
521 		 u64 offset, u64 size)
522 {
523 	struct panthor_gem_object *bo = to_panthor_bo(obj);
524 	struct drm_gem_shmem_object *shmem = &bo->base;
525 	const struct drm_device *dev = shmem->base.dev;
526 	struct sg_table *sgt;
527 	struct scatterlist *sgl;
528 	unsigned int count;
529 
530 	/* Make sure the range is in bounds. */
531 	if (offset + size < offset || offset + size > shmem->base.size)
532 		return -EINVAL;
533 
534 	/* Disallow CPU-cache maintenance on imported buffers. */
535 	if (drm_gem_is_imported(&shmem->base))
536 		return -EINVAL;
537 
538 	switch (type) {
539 	case DRM_PANTHOR_BO_SYNC_CPU_CACHE_FLUSH:
540 	case DRM_PANTHOR_BO_SYNC_CPU_CACHE_FLUSH_AND_INVALIDATE:
541 		break;
542 
543 	default:
544 		return -EINVAL;
545 	}
546 
547 	/* Don't bother if it's WC-mapped */
548 	if (shmem->map_wc)
549 		return 0;
550 
551 	/* Nothing to do if the size is zero. */
552 	if (size == 0)
553 		return 0;
554 
555 	sgt = drm_gem_shmem_get_pages_sgt(shmem);
556 	if (IS_ERR(sgt))
557 		return PTR_ERR(sgt);
558 
559 	for_each_sgtable_dma_sg(sgt, sgl, count) {
560 		if (size == 0)
561 			break;
562 
563 		dma_addr_t paddr = sg_dma_address(sgl);
564 		size_t len = sg_dma_len(sgl);
565 
566 		if (len <= offset) {
567 			offset -= len;
568 			continue;
569 		}
570 
571 		paddr += offset;
572 		len -= offset;
573 		len = min_t(size_t, len, size);
574 		size -= len;
575 		offset = 0;
576 
577 		/* It's unclear whether dma_sync_xxx() is the right API to do CPU
578 		 * cache maintenance given an IOMMU can register their own
579 		 * implementation doing more than just CPU cache flushes/invalidation,
580 		 * and what we really care about here is CPU caches only, but that's
581 		 * the best we have that is both arch-agnostic and does at least the
582 		 * CPU cache maintenance on a <page,offset,size> tuple.
583 		 *
584 		 * Also, I wish we could do a single
585 		 *
586 		 *      dma_sync_single_for_device(BIDIR)
587 		 *
588 		 * and get a flush+invalidate, but that's not how it's implemented
589 		 * in practice (at least on arm64), so we have to make it
590 		 *
591 		 *      dma_sync_single_for_device(TO_DEVICE)
592 		 *      dma_sync_single_for_cpu(FROM_DEVICE)
593 		 *
594 		 * for the flush+invalidate case.
595 		 */
596 		dma_sync_single_for_device(dev->dev, paddr, len, DMA_TO_DEVICE);
597 		if (type == DRM_PANTHOR_BO_SYNC_CPU_CACHE_FLUSH_AND_INVALIDATE)
598 			dma_sync_single_for_cpu(dev->dev, paddr, len, DMA_FROM_DEVICE);
599 	}
600 
601 	return 0;
602 }
603 
604 #ifdef CONFIG_DEBUG_FS
605 struct gem_size_totals {
606 	size_t size;
607 	size_t resident;
608 	size_t reclaimable;
609 };
610 
panthor_gem_debugfs_print_flag_names(struct seq_file * m)611 static void panthor_gem_debugfs_print_flag_names(struct seq_file *m)
612 {
613 	int len;
614 	int i;
615 
616 	static const char * const gem_state_flags_names[] = {
617 		[PANTHOR_DEBUGFS_GEM_STATE_IMPORTED_BIT] = "imported",
618 		[PANTHOR_DEBUGFS_GEM_STATE_EXPORTED_BIT] = "exported",
619 	};
620 
621 	static const char * const gem_usage_flags_names[] = {
622 		[PANTHOR_DEBUGFS_GEM_USAGE_KERNEL_BIT] = "kernel",
623 		[PANTHOR_DEBUGFS_GEM_USAGE_FW_MAPPED_BIT] = "fw-mapped",
624 	};
625 
626 	seq_puts(m, "GEM state flags: ");
627 	for (i = 0, len = ARRAY_SIZE(gem_state_flags_names); i < len; i++) {
628 		if (!gem_state_flags_names[i])
629 			continue;
630 		seq_printf(m, "%s (0x%x)%s", gem_state_flags_names[i],
631 			   (u32)BIT(i), (i < len - 1) ? ", " : "\n");
632 	}
633 
634 	seq_puts(m, "GEM usage flags: ");
635 	for (i = 0, len = ARRAY_SIZE(gem_usage_flags_names); i < len; i++) {
636 		if (!gem_usage_flags_names[i])
637 			continue;
638 		seq_printf(m, "%s (0x%x)%s", gem_usage_flags_names[i],
639 			   (u32)BIT(i), (i < len - 1) ? ", " : "\n\n");
640 	}
641 }
642 
panthor_gem_debugfs_bo_print(struct panthor_gem_object * bo,struct seq_file * m,struct gem_size_totals * totals)643 static void panthor_gem_debugfs_bo_print(struct panthor_gem_object *bo,
644 					 struct seq_file *m,
645 					 struct gem_size_totals *totals)
646 {
647 	unsigned int refcount = kref_read(&bo->base.base.refcount);
648 	char creator_info[32] = {};
649 	size_t resident_size;
650 	u32 gem_usage_flags = bo->debugfs.flags;
651 	u32 gem_state_flags = 0;
652 
653 	/* Skip BOs being destroyed. */
654 	if (!refcount)
655 		return;
656 
657 	resident_size = bo->base.pages ? bo->base.base.size : 0;
658 
659 	snprintf(creator_info, sizeof(creator_info),
660 		 "%s/%d", bo->debugfs.creator.process_name, bo->debugfs.creator.tgid);
661 	seq_printf(m, "%-32s%-16d%-16d%-16zd%-16zd0x%-16lx",
662 		   creator_info,
663 		   bo->base.base.name,
664 		   refcount,
665 		   bo->base.base.size,
666 		   resident_size,
667 		   drm_vma_node_start(&bo->base.base.vma_node));
668 
669 	if (bo->base.base.import_attach)
670 		gem_state_flags |= PANTHOR_DEBUGFS_GEM_STATE_FLAG_IMPORTED;
671 	if (bo->base.base.dma_buf)
672 		gem_state_flags |= PANTHOR_DEBUGFS_GEM_STATE_FLAG_EXPORTED;
673 
674 	seq_printf(m, "0x%-8x 0x%-10x", gem_state_flags, gem_usage_flags);
675 
676 	scoped_guard(mutex, &bo->label.lock) {
677 		seq_printf(m, "%s\n", bo->label.str ? : "");
678 	}
679 
680 	totals->size += bo->base.base.size;
681 	totals->resident += resident_size;
682 	if (bo->base.madv > 0)
683 		totals->reclaimable += resident_size;
684 }
685 
panthor_gem_debugfs_print_bos(struct panthor_device * ptdev,struct seq_file * m)686 void panthor_gem_debugfs_print_bos(struct panthor_device *ptdev,
687 				   struct seq_file *m)
688 {
689 	struct gem_size_totals totals = {0};
690 	struct panthor_gem_object *bo;
691 
692 	panthor_gem_debugfs_print_flag_names(m);
693 
694 	seq_puts(m, "created-by                      global-name     refcount        size            resident-size   file-offset       state      usage       label\n");
695 	seq_puts(m, "----------------------------------------------------------------------------------------------------------------------------------------------\n");
696 
697 	scoped_guard(mutex, &ptdev->gems.lock) {
698 		list_for_each_entry(bo, &ptdev->gems.node, debugfs.node) {
699 			panthor_gem_debugfs_bo_print(bo, m, &totals);
700 		}
701 	}
702 
703 	seq_puts(m, "==============================================================================================================================================\n");
704 	seq_printf(m, "Total size: %zd, Total resident: %zd, Total reclaimable: %zd\n",
705 		   totals.size, totals.resident, totals.reclaimable);
706 }
707 #endif
708