1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2012-2014 Intel Corporation
5  *
6   * Based on amdgpu_mn, which bears the following notice:
7  *
8  * Copyright 2014 Advanced Micro Devices, Inc.
9  * All Rights Reserved.
10  *
11  * Permission is hereby granted, free of charge, to any person obtaining a
12  * copy of this software and associated documentation files (the
13  * "Software"), to deal in the Software without restriction, including
14  * without limitation the rights to use, copy, modify, merge, publish,
15  * distribute, sub license, and/or sell copies of the Software, and to
16  * permit persons to whom the Software is furnished to do so, subject to
17  * the following conditions:
18  *
19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25  * USE OR OTHER DEALINGS IN THE SOFTWARE.
26  *
27  * The above copyright notice and this permission notice (including the
28  * next paragraph) shall be included in all copies or substantial portions
29  * of the Software.
30  *
31  */
32 /*
33  * Authors:
34  *    Christian König <christian.koenig@amd.com>
35  */
36 
37 #include <linux/mmu_context.h>
38 #include <linux/mempolicy.h>
39 #include <linux/swap.h>
40 #include <linux/sched/mm.h>
41 
42 #include "i915_drv.h"
43 #include "i915_gem_ioctls.h"
44 #include "i915_gem_object.h"
45 #include "i915_gem_userptr.h"
46 #include "i915_scatterlist.h"
47 
48 #ifdef CONFIG_MMU_NOTIFIER
49 
50 /**
51  * i915_gem_userptr_invalidate - callback to notify about mm change
52  *
53  * @mni: the range (mm) is about to update
54  * @range: details on the invalidation
55  * @cur_seq: Value to pass to mmu_interval_set_seq()
56  *
57  * Block for operations on BOs to finish and mark pages as accessed and
58  * potentially dirty.
59  */
60 static bool i915_gem_userptr_invalidate(struct mmu_interval_notifier *mni,
61 					const struct mmu_notifier_range *range,
62 					unsigned long cur_seq)
63 {
64 	struct drm_i915_gem_object *obj = container_of(mni, struct drm_i915_gem_object, userptr.notifier);
65 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
66 	long r;
67 
68 	if (!mmu_notifier_range_blockable(range))
69 		return false;
70 
71 	write_lock(&i915->mm.notifier_lock);
72 
73 	mmu_interval_set_seq(mni, cur_seq);
74 
75 	write_unlock(&i915->mm.notifier_lock);
76 
77 	/*
78 	 * We don't wait when the process is exiting. This is valid
79 	 * because the object will be cleaned up anyway.
80 	 *
81 	 * This is also temporarily required as a hack, because we
82 	 * cannot currently force non-consistent batch buffers to preempt
83 	 * and reschedule by waiting on it, hanging processes on exit.
84 	 */
85 	if (current->flags & PF_EXITING)
86 		return true;
87 
88 	/* we will unbind on next submission, still have userptr pins */
89 	r = dma_resv_wait_timeout(obj->base.resv, true, false,
90 				  MAX_SCHEDULE_TIMEOUT);
91 	if (r <= 0)
92 		drm_err(&i915->drm, "(%ld) failed to wait for idle\n", r);
93 
94 	return true;
95 }
96 
97 static const struct mmu_interval_notifier_ops i915_gem_userptr_notifier_ops = {
98 	.invalidate = i915_gem_userptr_invalidate,
99 };
100 
101 static int
102 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj)
103 {
104 	return mmu_interval_notifier_insert(&obj->userptr.notifier, current->mm,
105 					    obj->userptr.ptr, obj->base.size,
106 					    &i915_gem_userptr_notifier_ops);
107 }
108 
109 static void i915_gem_object_userptr_drop_ref(struct drm_i915_gem_object *obj)
110 {
111 	struct page **pvec = NULL;
112 
113 	assert_object_held_shared(obj);
114 
115 	if (!--obj->userptr.page_ref) {
116 		pvec = obj->userptr.pvec;
117 		obj->userptr.pvec = NULL;
118 	}
119 	GEM_BUG_ON(obj->userptr.page_ref < 0);
120 
121 	if (pvec) {
122 		const unsigned long num_pages = obj->base.size >> PAGE_SHIFT;
123 
124 		unpin_user_pages(pvec, num_pages);
125 		kvfree(pvec);
126 	}
127 }
128 
129 static int i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
130 {
131 	const unsigned long num_pages = obj->base.size >> PAGE_SHIFT;
132 	unsigned int max_segment = i915_sg_segment_size();
133 	struct sg_table *st;
134 	unsigned int sg_page_sizes;
135 	struct page **pvec;
136 	int ret;
137 
138 	st = kmalloc(sizeof(*st), GFP_KERNEL);
139 	if (!st)
140 		return -ENOMEM;
141 
142 	if (!obj->userptr.page_ref) {
143 		ret = -EAGAIN;
144 		goto err_free;
145 	}
146 
147 	obj->userptr.page_ref++;
148 	pvec = obj->userptr.pvec;
149 
150 alloc_table:
151 	ret = sg_alloc_table_from_pages_segment(st, pvec, num_pages, 0,
152 						num_pages << PAGE_SHIFT,
153 						max_segment, GFP_KERNEL);
154 	if (ret)
155 		goto err;
156 
157 	ret = i915_gem_gtt_prepare_pages(obj, st);
158 	if (ret) {
159 		sg_free_table(st);
160 
161 		if (max_segment > PAGE_SIZE) {
162 			max_segment = PAGE_SIZE;
163 			goto alloc_table;
164 		}
165 
166 		goto err;
167 	}
168 
169 	WARN_ON_ONCE(!(obj->cache_coherent & I915_BO_CACHE_COHERENT_FOR_WRITE));
170 	if (i915_gem_object_can_bypass_llc(obj))
171 		obj->cache_dirty = true;
172 
173 	sg_page_sizes = i915_sg_dma_sizes(st->sgl);
174 	__i915_gem_object_set_pages(obj, st, sg_page_sizes);
175 
176 	return 0;
177 
178 err:
179 	i915_gem_object_userptr_drop_ref(obj);
180 err_free:
181 	kfree(st);
182 	return ret;
183 }
184 
185 static void
186 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
187 			   struct sg_table *pages)
188 {
189 	struct sgt_iter sgt_iter;
190 	struct page *page;
191 
192 	if (!pages)
193 		return;
194 
195 	__i915_gem_object_release_shmem(obj, pages, true);
196 	i915_gem_gtt_finish_pages(obj, pages);
197 
198 	/*
199 	 * We always mark objects as dirty when they are used by the GPU,
200 	 * just in case. However, if we set the vma as being read-only we know
201 	 * that the object will never have been written to.
202 	 */
203 	if (i915_gem_object_is_readonly(obj))
204 		obj->mm.dirty = false;
205 
206 	for_each_sgt_page(page, sgt_iter, pages) {
207 		if (obj->mm.dirty && trylock_page(page)) {
208 			/*
209 			 * As this may not be anonymous memory (e.g. shmem)
210 			 * but exist on a real mapping, we have to lock
211 			 * the page in order to dirty it -- holding
212 			 * the page reference is not sufficient to
213 			 * prevent the inode from being truncated.
214 			 * Play safe and take the lock.
215 			 *
216 			 * However...!
217 			 *
218 			 * The mmu-notifier can be invalidated for a
219 			 * migrate_page, that is alreadying holding the lock
220 			 * on the page. Such a try_to_unmap() will result
221 			 * in us calling put_pages() and so recursively try
222 			 * to lock the page. We avoid that deadlock with
223 			 * a trylock_page() and in exchange we risk missing
224 			 * some page dirtying.
225 			 */
226 			set_page_dirty(page);
227 			unlock_page(page);
228 		}
229 
230 		mark_page_accessed(page);
231 	}
232 	obj->mm.dirty = false;
233 
234 	sg_free_table(pages);
235 	kfree(pages);
236 
237 	i915_gem_object_userptr_drop_ref(obj);
238 }
239 
240 static int i915_gem_object_userptr_unbind(struct drm_i915_gem_object *obj)
241 {
242 	struct sg_table *pages;
243 	int err;
244 
245 	err = i915_gem_object_unbind(obj, I915_GEM_OBJECT_UNBIND_ACTIVE);
246 	if (err)
247 		return err;
248 
249 	if (GEM_WARN_ON(i915_gem_object_has_pinned_pages(obj)))
250 		return -EBUSY;
251 
252 	assert_object_held(obj);
253 
254 	pages = __i915_gem_object_unset_pages(obj);
255 	if (!IS_ERR_OR_NULL(pages))
256 		i915_gem_userptr_put_pages(obj, pages);
257 
258 	return err;
259 }
260 
261 int i915_gem_object_userptr_submit_init(struct drm_i915_gem_object *obj)
262 {
263 	const unsigned long num_pages = obj->base.size >> PAGE_SHIFT;
264 	struct page **pvec;
265 	unsigned int gup_flags = 0;
266 	unsigned long notifier_seq;
267 	int pinned, ret;
268 
269 	if (obj->userptr.notifier.mm != current->mm)
270 		return -EFAULT;
271 
272 	notifier_seq = mmu_interval_read_begin(&obj->userptr.notifier);
273 
274 	ret = i915_gem_object_lock_interruptible(obj, NULL);
275 	if (ret)
276 		return ret;
277 
278 	if (notifier_seq == obj->userptr.notifier_seq && obj->userptr.pvec) {
279 		i915_gem_object_unlock(obj);
280 		return 0;
281 	}
282 
283 	ret = i915_gem_object_userptr_unbind(obj);
284 	i915_gem_object_unlock(obj);
285 	if (ret)
286 		return ret;
287 
288 	pvec = kvmalloc_array(num_pages, sizeof(struct page *), GFP_KERNEL);
289 	if (!pvec)
290 		return -ENOMEM;
291 
292 	if (!i915_gem_object_is_readonly(obj))
293 		gup_flags |= FOLL_WRITE;
294 
295 	pinned = ret = 0;
296 	while (pinned < num_pages) {
297 		ret = pin_user_pages_fast(obj->userptr.ptr + pinned * PAGE_SIZE,
298 					  num_pages - pinned, gup_flags,
299 					  &pvec[pinned]);
300 		if (ret < 0)
301 			goto out;
302 
303 		pinned += ret;
304 	}
305 	ret = 0;
306 
307 	ret = i915_gem_object_lock_interruptible(obj, NULL);
308 	if (ret)
309 		goto out;
310 
311 	if (mmu_interval_read_retry(&obj->userptr.notifier,
312 		!obj->userptr.page_ref ? notifier_seq :
313 		obj->userptr.notifier_seq)) {
314 		ret = -EAGAIN;
315 		goto out_unlock;
316 	}
317 
318 	if (!obj->userptr.page_ref++) {
319 		obj->userptr.pvec = pvec;
320 		obj->userptr.notifier_seq = notifier_seq;
321 		pvec = NULL;
322 		ret = ____i915_gem_object_get_pages(obj);
323 	}
324 
325 	obj->userptr.page_ref--;
326 
327 out_unlock:
328 	i915_gem_object_unlock(obj);
329 
330 out:
331 	if (pvec) {
332 		unpin_user_pages(pvec, pinned);
333 		kvfree(pvec);
334 	}
335 
336 	return ret;
337 }
338 
339 int i915_gem_object_userptr_submit_done(struct drm_i915_gem_object *obj)
340 {
341 	if (mmu_interval_read_retry(&obj->userptr.notifier,
342 				    obj->userptr.notifier_seq)) {
343 		/* We collided with the mmu notifier, need to retry */
344 
345 		return -EAGAIN;
346 	}
347 
348 	return 0;
349 }
350 
351 int i915_gem_object_userptr_validate(struct drm_i915_gem_object *obj)
352 {
353 	int err;
354 
355 	err = i915_gem_object_userptr_submit_init(obj);
356 	if (err)
357 		return err;
358 
359 	err = i915_gem_object_lock_interruptible(obj, NULL);
360 	if (!err) {
361 		/*
362 		 * Since we only check validity, not use the pages,
363 		 * it doesn't matter if we collide with the mmu notifier,
364 		 * and -EAGAIN handling is not required.
365 		 */
366 		err = i915_gem_object_pin_pages(obj);
367 		if (!err)
368 			i915_gem_object_unpin_pages(obj);
369 
370 		i915_gem_object_unlock(obj);
371 	}
372 
373 	return err;
374 }
375 
376 static void
377 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
378 {
379 	GEM_WARN_ON(obj->userptr.page_ref);
380 
381 	mmu_interval_notifier_remove(&obj->userptr.notifier);
382 	obj->userptr.notifier.mm = NULL;
383 }
384 
385 static int
386 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
387 {
388 	drm_dbg(obj->base.dev, "Exporting userptr no longer allowed\n");
389 
390 	return -EINVAL;
391 }
392 
393 static int
394 i915_gem_userptr_pwrite(struct drm_i915_gem_object *obj,
395 			const struct drm_i915_gem_pwrite *args)
396 {
397 	drm_dbg(obj->base.dev, "pwrite to userptr no longer allowed\n");
398 
399 	return -EINVAL;
400 }
401 
402 static int
403 i915_gem_userptr_pread(struct drm_i915_gem_object *obj,
404 		       const struct drm_i915_gem_pread *args)
405 {
406 	drm_dbg(obj->base.dev, "pread from userptr no longer allowed\n");
407 
408 	return -EINVAL;
409 }
410 
411 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
412 	.name = "i915_gem_object_userptr",
413 	.flags = I915_GEM_OBJECT_IS_SHRINKABLE |
414 		 I915_GEM_OBJECT_NO_MMAP |
415 		 I915_GEM_OBJECT_IS_PROXY,
416 	.get_pages = i915_gem_userptr_get_pages,
417 	.put_pages = i915_gem_userptr_put_pages,
418 	.dmabuf_export = i915_gem_userptr_dmabuf_export,
419 	.pwrite = i915_gem_userptr_pwrite,
420 	.pread = i915_gem_userptr_pread,
421 	.release = i915_gem_userptr_release,
422 };
423 
424 #endif
425 
426 static int
427 probe_range(struct mm_struct *mm, unsigned long addr, unsigned long len)
428 {
429 	const unsigned long end = addr + len;
430 	struct vm_area_struct *vma;
431 	int ret = -EFAULT;
432 
433 	mmap_read_lock(mm);
434 	for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
435 		/* Check for holes, note that we also update the addr below */
436 		if (vma->vm_start > addr)
437 			break;
438 
439 		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
440 			break;
441 
442 		if (vma->vm_end >= end) {
443 			ret = 0;
444 			break;
445 		}
446 
447 		addr = vma->vm_end;
448 	}
449 	mmap_read_unlock(mm);
450 
451 	return ret;
452 }
453 
454 /*
455  * Creates a new mm object that wraps some normal memory from the process
456  * context - user memory.
457  *
458  * We impose several restrictions upon the memory being mapped
459  * into the GPU.
460  * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
461  * 2. It must be normal system memory, not a pointer into another map of IO
462  *    space (e.g. it must not be a GTT mmapping of another object).
463  * 3. We only allow a bo as large as we could in theory map into the GTT,
464  *    that is we limit the size to the total size of the GTT.
465  * 4. The bo is marked as being snoopable. The backing pages are left
466  *    accessible directly by the CPU, but reads and writes by the GPU may
467  *    incur the cost of a snoop (unless you have an LLC architecture).
468  *
469  * Synchronisation between multiple users and the GPU is left to userspace
470  * through the normal set-domain-ioctl. The kernel will enforce that the
471  * GPU relinquishes the VMA before it is returned back to the system
472  * i.e. upon free(), munmap() or process termination. However, the userspace
473  * malloc() library may not immediately relinquish the VMA after free() and
474  * instead reuse it whilst the GPU is still reading and writing to the VMA.
475  * Caveat emptor.
476  *
477  * Also note, that the object created here is not currently a "first class"
478  * object, in that several ioctls are banned. These are the CPU access
479  * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
480  * direct access via your pointer rather than use those ioctls. Another
481  * restriction is that we do not allow userptr surfaces to be pinned to the
482  * hardware and so we reject any attempt to create a framebuffer out of a
483  * userptr.
484  *
485  * If you think this is a good interface to use to pass GPU memory between
486  * drivers, please use dma-buf instead. In fact, wherever possible use
487  * dma-buf instead.
488  */
489 int
490 i915_gem_userptr_ioctl(struct drm_device *dev,
491 		       void *data,
492 		       struct drm_file *file)
493 {
494 	static struct lock_class_key __maybe_unused lock_class;
495 	struct drm_i915_private *dev_priv = to_i915(dev);
496 	struct drm_i915_gem_userptr *args = data;
497 	struct drm_i915_gem_object __maybe_unused *obj;
498 	int __maybe_unused ret;
499 	u32 __maybe_unused handle;
500 
501 	if (!HAS_LLC(dev_priv) && !HAS_SNOOP(dev_priv)) {
502 		/* We cannot support coherent userptr objects on hw without
503 		 * LLC and broken snooping.
504 		 */
505 		return -ENODEV;
506 	}
507 
508 	if (args->flags & ~(I915_USERPTR_READ_ONLY |
509 			    I915_USERPTR_UNSYNCHRONIZED |
510 			    I915_USERPTR_PROBE))
511 		return -EINVAL;
512 
513 	if (i915_gem_object_size_2big(args->user_size))
514 		return -E2BIG;
515 
516 	if (!args->user_size)
517 		return -EINVAL;
518 
519 	if (offset_in_page(args->user_ptr | args->user_size))
520 		return -EINVAL;
521 
522 	if (!access_ok((char __user *)(unsigned long)args->user_ptr, args->user_size))
523 		return -EFAULT;
524 
525 	if (args->flags & I915_USERPTR_UNSYNCHRONIZED)
526 		return -ENODEV;
527 
528 	if (args->flags & I915_USERPTR_READ_ONLY) {
529 		/*
530 		 * On almost all of the older hw, we cannot tell the GPU that
531 		 * a page is readonly.
532 		 */
533 		if (!to_gt(dev_priv)->vm->has_read_only)
534 			return -ENODEV;
535 	}
536 
537 	if (args->flags & I915_USERPTR_PROBE) {
538 		/*
539 		 * Check that the range pointed to represents real struct
540 		 * pages and not iomappings (at this moment in time!)
541 		 */
542 		ret = probe_range(current->mm, args->user_ptr, args->user_size);
543 		if (ret)
544 			return ret;
545 	}
546 
547 #ifdef CONFIG_MMU_NOTIFIER
548 	obj = i915_gem_object_alloc();
549 	if (obj == NULL)
550 		return -ENOMEM;
551 
552 	drm_gem_private_object_init(dev, &obj->base, args->user_size);
553 	i915_gem_object_init(obj, &i915_gem_userptr_ops, &lock_class,
554 			     I915_BO_ALLOC_USER);
555 	obj->mem_flags = I915_BO_FLAG_STRUCT_PAGE;
556 	obj->read_domains = I915_GEM_DOMAIN_CPU;
557 	obj->write_domain = I915_GEM_DOMAIN_CPU;
558 	i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
559 
560 	obj->userptr.ptr = args->user_ptr;
561 	obj->userptr.notifier_seq = ULONG_MAX;
562 	if (args->flags & I915_USERPTR_READ_ONLY)
563 		i915_gem_object_set_readonly(obj);
564 
565 	/* And keep a pointer to the current->mm for resolving the user pages
566 	 * at binding. This means that we need to hook into the mmu_notifier
567 	 * in order to detect if the mmu is destroyed.
568 	 */
569 	ret = i915_gem_userptr_init__mmu_notifier(obj);
570 	if (ret == 0)
571 		ret = drm_gem_handle_create(file, &obj->base, &handle);
572 
573 	/* drop reference from allocate - handle holds it now */
574 	i915_gem_object_put(obj);
575 	if (ret)
576 		return ret;
577 
578 	args->handle = handle;
579 	return 0;
580 #else
581 	return -ENODEV;
582 #endif
583 }
584 
585 int i915_gem_init_userptr(struct drm_i915_private *dev_priv)
586 {
587 #ifdef CONFIG_MMU_NOTIFIER
588 	rwlock_init(&dev_priv->mm.notifier_lock);
589 #endif
590 
591 	return 0;
592 }
593 
594 void i915_gem_cleanup_userptr(struct drm_i915_private *dev_priv)
595 {
596 }
597