xref: /linux/lib/test_hmm.c (revision 662fa3d6099374c4615bf64d06895e3573b935b2)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * This is a module to test the HMM (Heterogeneous Memory Management)
4  * mirror and zone device private memory migration APIs of the kernel.
5  * Userspace programs can register with the driver to mirror their own address
6  * space and can use the device to read/write any valid virtual address.
7  */
8 #include <linux/init.h>
9 #include <linux/fs.h>
10 #include <linux/mm.h>
11 #include <linux/module.h>
12 #include <linux/kernel.h>
13 #include <linux/cdev.h>
14 #include <linux/device.h>
15 #include <linux/mutex.h>
16 #include <linux/rwsem.h>
17 #include <linux/sched.h>
18 #include <linux/slab.h>
19 #include <linux/highmem.h>
20 #include <linux/delay.h>
21 #include <linux/pagemap.h>
22 #include <linux/hmm.h>
23 #include <linux/vmalloc.h>
24 #include <linux/swap.h>
25 #include <linux/swapops.h>
26 #include <linux/sched/mm.h>
27 #include <linux/platform_device.h>
28 #include <linux/rmap.h>
29 
30 #include "test_hmm_uapi.h"
31 
32 #define DMIRROR_NDEVICES		2
33 #define DMIRROR_RANGE_FAULT_TIMEOUT	1000
34 #define DEVMEM_CHUNK_SIZE		(256 * 1024 * 1024U)
35 #define DEVMEM_CHUNKS_RESERVE		16
36 
37 static const struct dev_pagemap_ops dmirror_devmem_ops;
38 static const struct mmu_interval_notifier_ops dmirror_min_ops;
39 static dev_t dmirror_dev;
40 
41 struct dmirror_device;
42 
43 struct dmirror_bounce {
44 	void			*ptr;
45 	unsigned long		size;
46 	unsigned long		addr;
47 	unsigned long		cpages;
48 };
49 
50 #define DPT_XA_TAG_ATOMIC 1UL
51 #define DPT_XA_TAG_WRITE 3UL
52 
53 /*
54  * Data structure to track address ranges and register for mmu interval
55  * notifier updates.
56  */
57 struct dmirror_interval {
58 	struct mmu_interval_notifier	notifier;
59 	struct dmirror			*dmirror;
60 };
61 
62 /*
63  * Data attached to the open device file.
64  * Note that it might be shared after a fork().
65  */
66 struct dmirror {
67 	struct dmirror_device		*mdevice;
68 	struct xarray			pt;
69 	struct mmu_interval_notifier	notifier;
70 	struct mutex			mutex;
71 };
72 
73 /*
74  * ZONE_DEVICE pages for migration and simulating device memory.
75  */
76 struct dmirror_chunk {
77 	struct dev_pagemap	pagemap;
78 	struct dmirror_device	*mdevice;
79 };
80 
81 /*
82  * Per device data.
83  */
84 struct dmirror_device {
85 	struct cdev		cdevice;
86 	struct hmm_devmem	*devmem;
87 
88 	unsigned int		devmem_capacity;
89 	unsigned int		devmem_count;
90 	struct dmirror_chunk	**devmem_chunks;
91 	struct mutex		devmem_lock;	/* protects the above */
92 
93 	unsigned long		calloc;
94 	unsigned long		cfree;
95 	struct page		*free_pages;
96 	spinlock_t		lock;		/* protects the above */
97 };
98 
99 static struct dmirror_device dmirror_devices[DMIRROR_NDEVICES];
100 
101 static int dmirror_bounce_init(struct dmirror_bounce *bounce,
102 			       unsigned long addr,
103 			       unsigned long size)
104 {
105 	bounce->addr = addr;
106 	bounce->size = size;
107 	bounce->cpages = 0;
108 	bounce->ptr = vmalloc(size);
109 	if (!bounce->ptr)
110 		return -ENOMEM;
111 	return 0;
112 }
113 
114 static void dmirror_bounce_fini(struct dmirror_bounce *bounce)
115 {
116 	vfree(bounce->ptr);
117 }
118 
119 static int dmirror_fops_open(struct inode *inode, struct file *filp)
120 {
121 	struct cdev *cdev = inode->i_cdev;
122 	struct dmirror *dmirror;
123 	int ret;
124 
125 	/* Mirror this process address space */
126 	dmirror = kzalloc(sizeof(*dmirror), GFP_KERNEL);
127 	if (dmirror == NULL)
128 		return -ENOMEM;
129 
130 	dmirror->mdevice = container_of(cdev, struct dmirror_device, cdevice);
131 	mutex_init(&dmirror->mutex);
132 	xa_init(&dmirror->pt);
133 
134 	ret = mmu_interval_notifier_insert(&dmirror->notifier, current->mm,
135 				0, ULONG_MAX & PAGE_MASK, &dmirror_min_ops);
136 	if (ret) {
137 		kfree(dmirror);
138 		return ret;
139 	}
140 
141 	filp->private_data = dmirror;
142 	return 0;
143 }
144 
145 static int dmirror_fops_release(struct inode *inode, struct file *filp)
146 {
147 	struct dmirror *dmirror = filp->private_data;
148 
149 	mmu_interval_notifier_remove(&dmirror->notifier);
150 	xa_destroy(&dmirror->pt);
151 	kfree(dmirror);
152 	return 0;
153 }
154 
155 static struct dmirror_device *dmirror_page_to_device(struct page *page)
156 
157 {
158 	return container_of(page->pgmap, struct dmirror_chunk,
159 			    pagemap)->mdevice;
160 }
161 
162 static int dmirror_do_fault(struct dmirror *dmirror, struct hmm_range *range)
163 {
164 	unsigned long *pfns = range->hmm_pfns;
165 	unsigned long pfn;
166 
167 	for (pfn = (range->start >> PAGE_SHIFT);
168 	     pfn < (range->end >> PAGE_SHIFT);
169 	     pfn++, pfns++) {
170 		struct page *page;
171 		void *entry;
172 
173 		/*
174 		 * Since we asked for hmm_range_fault() to populate pages,
175 		 * it shouldn't return an error entry on success.
176 		 */
177 		WARN_ON(*pfns & HMM_PFN_ERROR);
178 		WARN_ON(!(*pfns & HMM_PFN_VALID));
179 
180 		page = hmm_pfn_to_page(*pfns);
181 		WARN_ON(!page);
182 
183 		entry = page;
184 		if (*pfns & HMM_PFN_WRITE)
185 			entry = xa_tag_pointer(entry, DPT_XA_TAG_WRITE);
186 		else if (WARN_ON(range->default_flags & HMM_PFN_WRITE))
187 			return -EFAULT;
188 		entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
189 		if (xa_is_err(entry))
190 			return xa_err(entry);
191 	}
192 
193 	return 0;
194 }
195 
196 static void dmirror_do_update(struct dmirror *dmirror, unsigned long start,
197 			      unsigned long end)
198 {
199 	unsigned long pfn;
200 	void *entry;
201 
202 	/*
203 	 * The XArray doesn't hold references to pages since it relies on
204 	 * the mmu notifier to clear page pointers when they become stale.
205 	 * Therefore, it is OK to just clear the entry.
206 	 */
207 	xa_for_each_range(&dmirror->pt, pfn, entry, start >> PAGE_SHIFT,
208 			  end >> PAGE_SHIFT)
209 		xa_erase(&dmirror->pt, pfn);
210 }
211 
212 static bool dmirror_interval_invalidate(struct mmu_interval_notifier *mni,
213 				const struct mmu_notifier_range *range,
214 				unsigned long cur_seq)
215 {
216 	struct dmirror *dmirror = container_of(mni, struct dmirror, notifier);
217 
218 	/*
219 	 * Ignore invalidation callbacks for device private pages since
220 	 * the invalidation is handled as part of the migration process.
221 	 */
222 	if (range->event == MMU_NOTIFY_MIGRATE &&
223 	    range->owner == dmirror->mdevice)
224 		return true;
225 
226 	if (mmu_notifier_range_blockable(range))
227 		mutex_lock(&dmirror->mutex);
228 	else if (!mutex_trylock(&dmirror->mutex))
229 		return false;
230 
231 	mmu_interval_set_seq(mni, cur_seq);
232 	dmirror_do_update(dmirror, range->start, range->end);
233 
234 	mutex_unlock(&dmirror->mutex);
235 	return true;
236 }
237 
238 static const struct mmu_interval_notifier_ops dmirror_min_ops = {
239 	.invalidate = dmirror_interval_invalidate,
240 };
241 
242 static int dmirror_range_fault(struct dmirror *dmirror,
243 				struct hmm_range *range)
244 {
245 	struct mm_struct *mm = dmirror->notifier.mm;
246 	unsigned long timeout =
247 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
248 	int ret;
249 
250 	while (true) {
251 		if (time_after(jiffies, timeout)) {
252 			ret = -EBUSY;
253 			goto out;
254 		}
255 
256 		range->notifier_seq = mmu_interval_read_begin(range->notifier);
257 		mmap_read_lock(mm);
258 		ret = hmm_range_fault(range);
259 		mmap_read_unlock(mm);
260 		if (ret) {
261 			if (ret == -EBUSY)
262 				continue;
263 			goto out;
264 		}
265 
266 		mutex_lock(&dmirror->mutex);
267 		if (mmu_interval_read_retry(range->notifier,
268 					    range->notifier_seq)) {
269 			mutex_unlock(&dmirror->mutex);
270 			continue;
271 		}
272 		break;
273 	}
274 
275 	ret = dmirror_do_fault(dmirror, range);
276 
277 	mutex_unlock(&dmirror->mutex);
278 out:
279 	return ret;
280 }
281 
282 static int dmirror_fault(struct dmirror *dmirror, unsigned long start,
283 			 unsigned long end, bool write)
284 {
285 	struct mm_struct *mm = dmirror->notifier.mm;
286 	unsigned long addr;
287 	unsigned long pfns[64];
288 	struct hmm_range range = {
289 		.notifier = &dmirror->notifier,
290 		.hmm_pfns = pfns,
291 		.pfn_flags_mask = 0,
292 		.default_flags =
293 			HMM_PFN_REQ_FAULT | (write ? HMM_PFN_REQ_WRITE : 0),
294 		.dev_private_owner = dmirror->mdevice,
295 	};
296 	int ret = 0;
297 
298 	/* Since the mm is for the mirrored process, get a reference first. */
299 	if (!mmget_not_zero(mm))
300 		return 0;
301 
302 	for (addr = start; addr < end; addr = range.end) {
303 		range.start = addr;
304 		range.end = min(addr + (ARRAY_SIZE(pfns) << PAGE_SHIFT), end);
305 
306 		ret = dmirror_range_fault(dmirror, &range);
307 		if (ret)
308 			break;
309 	}
310 
311 	mmput(mm);
312 	return ret;
313 }
314 
315 static int dmirror_do_read(struct dmirror *dmirror, unsigned long start,
316 			   unsigned long end, struct dmirror_bounce *bounce)
317 {
318 	unsigned long pfn;
319 	void *ptr;
320 
321 	ptr = bounce->ptr + ((start - bounce->addr) & PAGE_MASK);
322 
323 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
324 		void *entry;
325 		struct page *page;
326 		void *tmp;
327 
328 		entry = xa_load(&dmirror->pt, pfn);
329 		page = xa_untag_pointer(entry);
330 		if (!page)
331 			return -ENOENT;
332 
333 		tmp = kmap(page);
334 		memcpy(ptr, tmp, PAGE_SIZE);
335 		kunmap(page);
336 
337 		ptr += PAGE_SIZE;
338 		bounce->cpages++;
339 	}
340 
341 	return 0;
342 }
343 
344 static int dmirror_read(struct dmirror *dmirror, struct hmm_dmirror_cmd *cmd)
345 {
346 	struct dmirror_bounce bounce;
347 	unsigned long start, end;
348 	unsigned long size = cmd->npages << PAGE_SHIFT;
349 	int ret;
350 
351 	start = cmd->addr;
352 	end = start + size;
353 	if (end < start)
354 		return -EINVAL;
355 
356 	ret = dmirror_bounce_init(&bounce, start, size);
357 	if (ret)
358 		return ret;
359 
360 	while (1) {
361 		mutex_lock(&dmirror->mutex);
362 		ret = dmirror_do_read(dmirror, start, end, &bounce);
363 		mutex_unlock(&dmirror->mutex);
364 		if (ret != -ENOENT)
365 			break;
366 
367 		start = cmd->addr + (bounce.cpages << PAGE_SHIFT);
368 		ret = dmirror_fault(dmirror, start, end, false);
369 		if (ret)
370 			break;
371 		cmd->faults++;
372 	}
373 
374 	if (ret == 0) {
375 		if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
376 				 bounce.size))
377 			ret = -EFAULT;
378 	}
379 	cmd->cpages = bounce.cpages;
380 	dmirror_bounce_fini(&bounce);
381 	return ret;
382 }
383 
384 static int dmirror_do_write(struct dmirror *dmirror, unsigned long start,
385 			    unsigned long end, struct dmirror_bounce *bounce)
386 {
387 	unsigned long pfn;
388 	void *ptr;
389 
390 	ptr = bounce->ptr + ((start - bounce->addr) & PAGE_MASK);
391 
392 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
393 		void *entry;
394 		struct page *page;
395 		void *tmp;
396 
397 		entry = xa_load(&dmirror->pt, pfn);
398 		page = xa_untag_pointer(entry);
399 		if (!page || xa_pointer_tag(entry) != DPT_XA_TAG_WRITE)
400 			return -ENOENT;
401 
402 		tmp = kmap(page);
403 		memcpy(tmp, ptr, PAGE_SIZE);
404 		kunmap(page);
405 
406 		ptr += PAGE_SIZE;
407 		bounce->cpages++;
408 	}
409 
410 	return 0;
411 }
412 
413 static int dmirror_write(struct dmirror *dmirror, struct hmm_dmirror_cmd *cmd)
414 {
415 	struct dmirror_bounce bounce;
416 	unsigned long start, end;
417 	unsigned long size = cmd->npages << PAGE_SHIFT;
418 	int ret;
419 
420 	start = cmd->addr;
421 	end = start + size;
422 	if (end < start)
423 		return -EINVAL;
424 
425 	ret = dmirror_bounce_init(&bounce, start, size);
426 	if (ret)
427 		return ret;
428 	if (copy_from_user(bounce.ptr, u64_to_user_ptr(cmd->ptr),
429 			   bounce.size)) {
430 		ret = -EFAULT;
431 		goto fini;
432 	}
433 
434 	while (1) {
435 		mutex_lock(&dmirror->mutex);
436 		ret = dmirror_do_write(dmirror, start, end, &bounce);
437 		mutex_unlock(&dmirror->mutex);
438 		if (ret != -ENOENT)
439 			break;
440 
441 		start = cmd->addr + (bounce.cpages << PAGE_SHIFT);
442 		ret = dmirror_fault(dmirror, start, end, true);
443 		if (ret)
444 			break;
445 		cmd->faults++;
446 	}
447 
448 fini:
449 	cmd->cpages = bounce.cpages;
450 	dmirror_bounce_fini(&bounce);
451 	return ret;
452 }
453 
454 static bool dmirror_allocate_chunk(struct dmirror_device *mdevice,
455 				   struct page **ppage)
456 {
457 	struct dmirror_chunk *devmem;
458 	struct resource *res;
459 	unsigned long pfn;
460 	unsigned long pfn_first;
461 	unsigned long pfn_last;
462 	void *ptr;
463 
464 	devmem = kzalloc(sizeof(*devmem), GFP_KERNEL);
465 	if (!devmem)
466 		return false;
467 
468 	res = request_free_mem_region(&iomem_resource, DEVMEM_CHUNK_SIZE,
469 				      "hmm_dmirror");
470 	if (IS_ERR(res))
471 		goto err_devmem;
472 
473 	devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
474 	devmem->pagemap.range.start = res->start;
475 	devmem->pagemap.range.end = res->end;
476 	devmem->pagemap.nr_range = 1;
477 	devmem->pagemap.ops = &dmirror_devmem_ops;
478 	devmem->pagemap.owner = mdevice;
479 
480 	mutex_lock(&mdevice->devmem_lock);
481 
482 	if (mdevice->devmem_count == mdevice->devmem_capacity) {
483 		struct dmirror_chunk **new_chunks;
484 		unsigned int new_capacity;
485 
486 		new_capacity = mdevice->devmem_capacity +
487 				DEVMEM_CHUNKS_RESERVE;
488 		new_chunks = krealloc(mdevice->devmem_chunks,
489 				sizeof(new_chunks[0]) * new_capacity,
490 				GFP_KERNEL);
491 		if (!new_chunks)
492 			goto err_release;
493 		mdevice->devmem_capacity = new_capacity;
494 		mdevice->devmem_chunks = new_chunks;
495 	}
496 
497 	ptr = memremap_pages(&devmem->pagemap, numa_node_id());
498 	if (IS_ERR(ptr))
499 		goto err_release;
500 
501 	devmem->mdevice = mdevice;
502 	pfn_first = devmem->pagemap.range.start >> PAGE_SHIFT;
503 	pfn_last = pfn_first + (range_len(&devmem->pagemap.range) >> PAGE_SHIFT);
504 	mdevice->devmem_chunks[mdevice->devmem_count++] = devmem;
505 
506 	mutex_unlock(&mdevice->devmem_lock);
507 
508 	pr_info("added new %u MB chunk (total %u chunks, %u MB) PFNs [0x%lx 0x%lx)\n",
509 		DEVMEM_CHUNK_SIZE / (1024 * 1024),
510 		mdevice->devmem_count,
511 		mdevice->devmem_count * (DEVMEM_CHUNK_SIZE / (1024 * 1024)),
512 		pfn_first, pfn_last);
513 
514 	spin_lock(&mdevice->lock);
515 	for (pfn = pfn_first; pfn < pfn_last; pfn++) {
516 		struct page *page = pfn_to_page(pfn);
517 
518 		page->zone_device_data = mdevice->free_pages;
519 		mdevice->free_pages = page;
520 	}
521 	if (ppage) {
522 		*ppage = mdevice->free_pages;
523 		mdevice->free_pages = (*ppage)->zone_device_data;
524 		mdevice->calloc++;
525 	}
526 	spin_unlock(&mdevice->lock);
527 
528 	return true;
529 
530 err_release:
531 	mutex_unlock(&mdevice->devmem_lock);
532 	release_mem_region(devmem->pagemap.range.start, range_len(&devmem->pagemap.range));
533 err_devmem:
534 	kfree(devmem);
535 
536 	return false;
537 }
538 
539 static struct page *dmirror_devmem_alloc_page(struct dmirror_device *mdevice)
540 {
541 	struct page *dpage = NULL;
542 	struct page *rpage;
543 
544 	/*
545 	 * This is a fake device so we alloc real system memory to store
546 	 * our device memory.
547 	 */
548 	rpage = alloc_page(GFP_HIGHUSER);
549 	if (!rpage)
550 		return NULL;
551 
552 	spin_lock(&mdevice->lock);
553 
554 	if (mdevice->free_pages) {
555 		dpage = mdevice->free_pages;
556 		mdevice->free_pages = dpage->zone_device_data;
557 		mdevice->calloc++;
558 		spin_unlock(&mdevice->lock);
559 	} else {
560 		spin_unlock(&mdevice->lock);
561 		if (!dmirror_allocate_chunk(mdevice, &dpage))
562 			goto error;
563 	}
564 
565 	dpage->zone_device_data = rpage;
566 	get_page(dpage);
567 	lock_page(dpage);
568 	return dpage;
569 
570 error:
571 	__free_page(rpage);
572 	return NULL;
573 }
574 
575 static void dmirror_migrate_alloc_and_copy(struct migrate_vma *args,
576 					   struct dmirror *dmirror)
577 {
578 	struct dmirror_device *mdevice = dmirror->mdevice;
579 	const unsigned long *src = args->src;
580 	unsigned long *dst = args->dst;
581 	unsigned long addr;
582 
583 	for (addr = args->start; addr < args->end; addr += PAGE_SIZE,
584 						   src++, dst++) {
585 		struct page *spage;
586 		struct page *dpage;
587 		struct page *rpage;
588 
589 		if (!(*src & MIGRATE_PFN_MIGRATE))
590 			continue;
591 
592 		/*
593 		 * Note that spage might be NULL which is OK since it is an
594 		 * unallocated pte_none() or read-only zero page.
595 		 */
596 		spage = migrate_pfn_to_page(*src);
597 
598 		dpage = dmirror_devmem_alloc_page(mdevice);
599 		if (!dpage)
600 			continue;
601 
602 		rpage = dpage->zone_device_data;
603 		if (spage)
604 			copy_highpage(rpage, spage);
605 		else
606 			clear_highpage(rpage);
607 
608 		/*
609 		 * Normally, a device would use the page->zone_device_data to
610 		 * point to the mirror but here we use it to hold the page for
611 		 * the simulated device memory and that page holds the pointer
612 		 * to the mirror.
613 		 */
614 		rpage->zone_device_data = dmirror;
615 
616 		*dst = migrate_pfn(page_to_pfn(dpage)) |
617 			    MIGRATE_PFN_LOCKED;
618 		if ((*src & MIGRATE_PFN_WRITE) ||
619 		    (!spage && args->vma->vm_flags & VM_WRITE))
620 			*dst |= MIGRATE_PFN_WRITE;
621 	}
622 }
623 
624 static int dmirror_check_atomic(struct dmirror *dmirror, unsigned long start,
625 			     unsigned long end)
626 {
627 	unsigned long pfn;
628 
629 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
630 		void *entry;
631 		struct page *page;
632 
633 		entry = xa_load(&dmirror->pt, pfn);
634 		page = xa_untag_pointer(entry);
635 		if (xa_pointer_tag(entry) == DPT_XA_TAG_ATOMIC)
636 			return -EPERM;
637 	}
638 
639 	return 0;
640 }
641 
642 static int dmirror_atomic_map(unsigned long start, unsigned long end,
643 			      struct page **pages, struct dmirror *dmirror)
644 {
645 	unsigned long pfn, mapped = 0;
646 	int i;
647 
648 	/* Map the migrated pages into the device's page tables. */
649 	mutex_lock(&dmirror->mutex);
650 
651 	for (i = 0, pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++, i++) {
652 		void *entry;
653 
654 		if (!pages[i])
655 			continue;
656 
657 		entry = pages[i];
658 		entry = xa_tag_pointer(entry, DPT_XA_TAG_ATOMIC);
659 		entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
660 		if (xa_is_err(entry)) {
661 			mutex_unlock(&dmirror->mutex);
662 			return xa_err(entry);
663 		}
664 
665 		mapped++;
666 	}
667 
668 	mutex_unlock(&dmirror->mutex);
669 	return mapped;
670 }
671 
672 static int dmirror_migrate_finalize_and_map(struct migrate_vma *args,
673 					    struct dmirror *dmirror)
674 {
675 	unsigned long start = args->start;
676 	unsigned long end = args->end;
677 	const unsigned long *src = args->src;
678 	const unsigned long *dst = args->dst;
679 	unsigned long pfn;
680 
681 	/* Map the migrated pages into the device's page tables. */
682 	mutex_lock(&dmirror->mutex);
683 
684 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++,
685 								src++, dst++) {
686 		struct page *dpage;
687 		void *entry;
688 
689 		if (!(*src & MIGRATE_PFN_MIGRATE))
690 			continue;
691 
692 		dpage = migrate_pfn_to_page(*dst);
693 		if (!dpage)
694 			continue;
695 
696 		/*
697 		 * Store the page that holds the data so the page table
698 		 * doesn't have to deal with ZONE_DEVICE private pages.
699 		 */
700 		entry = dpage->zone_device_data;
701 		if (*dst & MIGRATE_PFN_WRITE)
702 			entry = xa_tag_pointer(entry, DPT_XA_TAG_WRITE);
703 		entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
704 		if (xa_is_err(entry)) {
705 			mutex_unlock(&dmirror->mutex);
706 			return xa_err(entry);
707 		}
708 	}
709 
710 	mutex_unlock(&dmirror->mutex);
711 	return 0;
712 }
713 
714 static int dmirror_exclusive(struct dmirror *dmirror,
715 			     struct hmm_dmirror_cmd *cmd)
716 {
717 	unsigned long start, end, addr;
718 	unsigned long size = cmd->npages << PAGE_SHIFT;
719 	struct mm_struct *mm = dmirror->notifier.mm;
720 	struct page *pages[64];
721 	struct dmirror_bounce bounce;
722 	unsigned long next;
723 	int ret;
724 
725 	start = cmd->addr;
726 	end = start + size;
727 	if (end < start)
728 		return -EINVAL;
729 
730 	/* Since the mm is for the mirrored process, get a reference first. */
731 	if (!mmget_not_zero(mm))
732 		return -EINVAL;
733 
734 	mmap_read_lock(mm);
735 	for (addr = start; addr < end; addr = next) {
736 		unsigned long mapped;
737 		int i;
738 
739 		if (end < addr + (ARRAY_SIZE(pages) << PAGE_SHIFT))
740 			next = end;
741 		else
742 			next = addr + (ARRAY_SIZE(pages) << PAGE_SHIFT);
743 
744 		ret = make_device_exclusive_range(mm, addr, next, pages, NULL);
745 		mapped = dmirror_atomic_map(addr, next, pages, dmirror);
746 		for (i = 0; i < ret; i++) {
747 			if (pages[i]) {
748 				unlock_page(pages[i]);
749 				put_page(pages[i]);
750 			}
751 		}
752 
753 		if (addr + (mapped << PAGE_SHIFT) < next) {
754 			mmap_read_unlock(mm);
755 			mmput(mm);
756 			return -EBUSY;
757 		}
758 	}
759 	mmap_read_unlock(mm);
760 	mmput(mm);
761 
762 	/* Return the migrated data for verification. */
763 	ret = dmirror_bounce_init(&bounce, start, size);
764 	if (ret)
765 		return ret;
766 	mutex_lock(&dmirror->mutex);
767 	ret = dmirror_do_read(dmirror, start, end, &bounce);
768 	mutex_unlock(&dmirror->mutex);
769 	if (ret == 0) {
770 		if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
771 				 bounce.size))
772 			ret = -EFAULT;
773 	}
774 
775 	cmd->cpages = bounce.cpages;
776 	dmirror_bounce_fini(&bounce);
777 	return ret;
778 }
779 
780 static int dmirror_migrate(struct dmirror *dmirror,
781 			   struct hmm_dmirror_cmd *cmd)
782 {
783 	unsigned long start, end, addr;
784 	unsigned long size = cmd->npages << PAGE_SHIFT;
785 	struct mm_struct *mm = dmirror->notifier.mm;
786 	struct vm_area_struct *vma;
787 	unsigned long src_pfns[64];
788 	unsigned long dst_pfns[64];
789 	struct dmirror_bounce bounce;
790 	struct migrate_vma args;
791 	unsigned long next;
792 	int ret;
793 
794 	start = cmd->addr;
795 	end = start + size;
796 	if (end < start)
797 		return -EINVAL;
798 
799 	/* Since the mm is for the mirrored process, get a reference first. */
800 	if (!mmget_not_zero(mm))
801 		return -EINVAL;
802 
803 	mmap_read_lock(mm);
804 	for (addr = start; addr < end; addr = next) {
805 		vma = vma_lookup(mm, addr);
806 		if (!vma || !(vma->vm_flags & VM_READ)) {
807 			ret = -EINVAL;
808 			goto out;
809 		}
810 		next = min(end, addr + (ARRAY_SIZE(src_pfns) << PAGE_SHIFT));
811 		if (next > vma->vm_end)
812 			next = vma->vm_end;
813 
814 		args.vma = vma;
815 		args.src = src_pfns;
816 		args.dst = dst_pfns;
817 		args.start = addr;
818 		args.end = next;
819 		args.pgmap_owner = dmirror->mdevice;
820 		args.flags = MIGRATE_VMA_SELECT_SYSTEM;
821 		ret = migrate_vma_setup(&args);
822 		if (ret)
823 			goto out;
824 
825 		dmirror_migrate_alloc_and_copy(&args, dmirror);
826 		migrate_vma_pages(&args);
827 		dmirror_migrate_finalize_and_map(&args, dmirror);
828 		migrate_vma_finalize(&args);
829 	}
830 	mmap_read_unlock(mm);
831 	mmput(mm);
832 
833 	/* Return the migrated data for verification. */
834 	ret = dmirror_bounce_init(&bounce, start, size);
835 	if (ret)
836 		return ret;
837 	mutex_lock(&dmirror->mutex);
838 	ret = dmirror_do_read(dmirror, start, end, &bounce);
839 	mutex_unlock(&dmirror->mutex);
840 	if (ret == 0) {
841 		if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
842 				 bounce.size))
843 			ret = -EFAULT;
844 	}
845 	cmd->cpages = bounce.cpages;
846 	dmirror_bounce_fini(&bounce);
847 	return ret;
848 
849 out:
850 	mmap_read_unlock(mm);
851 	mmput(mm);
852 	return ret;
853 }
854 
855 static void dmirror_mkentry(struct dmirror *dmirror, struct hmm_range *range,
856 			    unsigned char *perm, unsigned long entry)
857 {
858 	struct page *page;
859 
860 	if (entry & HMM_PFN_ERROR) {
861 		*perm = HMM_DMIRROR_PROT_ERROR;
862 		return;
863 	}
864 	if (!(entry & HMM_PFN_VALID)) {
865 		*perm = HMM_DMIRROR_PROT_NONE;
866 		return;
867 	}
868 
869 	page = hmm_pfn_to_page(entry);
870 	if (is_device_private_page(page)) {
871 		/* Is the page migrated to this device or some other? */
872 		if (dmirror->mdevice == dmirror_page_to_device(page))
873 			*perm = HMM_DMIRROR_PROT_DEV_PRIVATE_LOCAL;
874 		else
875 			*perm = HMM_DMIRROR_PROT_DEV_PRIVATE_REMOTE;
876 	} else if (is_zero_pfn(page_to_pfn(page)))
877 		*perm = HMM_DMIRROR_PROT_ZERO;
878 	else
879 		*perm = HMM_DMIRROR_PROT_NONE;
880 	if (entry & HMM_PFN_WRITE)
881 		*perm |= HMM_DMIRROR_PROT_WRITE;
882 	else
883 		*perm |= HMM_DMIRROR_PROT_READ;
884 	if (hmm_pfn_to_map_order(entry) + PAGE_SHIFT == PMD_SHIFT)
885 		*perm |= HMM_DMIRROR_PROT_PMD;
886 	else if (hmm_pfn_to_map_order(entry) + PAGE_SHIFT == PUD_SHIFT)
887 		*perm |= HMM_DMIRROR_PROT_PUD;
888 }
889 
890 static bool dmirror_snapshot_invalidate(struct mmu_interval_notifier *mni,
891 				const struct mmu_notifier_range *range,
892 				unsigned long cur_seq)
893 {
894 	struct dmirror_interval *dmi =
895 		container_of(mni, struct dmirror_interval, notifier);
896 	struct dmirror *dmirror = dmi->dmirror;
897 
898 	if (mmu_notifier_range_blockable(range))
899 		mutex_lock(&dmirror->mutex);
900 	else if (!mutex_trylock(&dmirror->mutex))
901 		return false;
902 
903 	/*
904 	 * Snapshots only need to set the sequence number since any
905 	 * invalidation in the interval invalidates the whole snapshot.
906 	 */
907 	mmu_interval_set_seq(mni, cur_seq);
908 
909 	mutex_unlock(&dmirror->mutex);
910 	return true;
911 }
912 
913 static const struct mmu_interval_notifier_ops dmirror_mrn_ops = {
914 	.invalidate = dmirror_snapshot_invalidate,
915 };
916 
917 static int dmirror_range_snapshot(struct dmirror *dmirror,
918 				  struct hmm_range *range,
919 				  unsigned char *perm)
920 {
921 	struct mm_struct *mm = dmirror->notifier.mm;
922 	struct dmirror_interval notifier;
923 	unsigned long timeout =
924 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
925 	unsigned long i;
926 	unsigned long n;
927 	int ret = 0;
928 
929 	notifier.dmirror = dmirror;
930 	range->notifier = &notifier.notifier;
931 
932 	ret = mmu_interval_notifier_insert(range->notifier, mm,
933 			range->start, range->end - range->start,
934 			&dmirror_mrn_ops);
935 	if (ret)
936 		return ret;
937 
938 	while (true) {
939 		if (time_after(jiffies, timeout)) {
940 			ret = -EBUSY;
941 			goto out;
942 		}
943 
944 		range->notifier_seq = mmu_interval_read_begin(range->notifier);
945 
946 		mmap_read_lock(mm);
947 		ret = hmm_range_fault(range);
948 		mmap_read_unlock(mm);
949 		if (ret) {
950 			if (ret == -EBUSY)
951 				continue;
952 			goto out;
953 		}
954 
955 		mutex_lock(&dmirror->mutex);
956 		if (mmu_interval_read_retry(range->notifier,
957 					    range->notifier_seq)) {
958 			mutex_unlock(&dmirror->mutex);
959 			continue;
960 		}
961 		break;
962 	}
963 
964 	n = (range->end - range->start) >> PAGE_SHIFT;
965 	for (i = 0; i < n; i++)
966 		dmirror_mkentry(dmirror, range, perm + i, range->hmm_pfns[i]);
967 
968 	mutex_unlock(&dmirror->mutex);
969 out:
970 	mmu_interval_notifier_remove(range->notifier);
971 	return ret;
972 }
973 
974 static int dmirror_snapshot(struct dmirror *dmirror,
975 			    struct hmm_dmirror_cmd *cmd)
976 {
977 	struct mm_struct *mm = dmirror->notifier.mm;
978 	unsigned long start, end;
979 	unsigned long size = cmd->npages << PAGE_SHIFT;
980 	unsigned long addr;
981 	unsigned long next;
982 	unsigned long pfns[64];
983 	unsigned char perm[64];
984 	char __user *uptr;
985 	struct hmm_range range = {
986 		.hmm_pfns = pfns,
987 		.dev_private_owner = dmirror->mdevice,
988 	};
989 	int ret = 0;
990 
991 	start = cmd->addr;
992 	end = start + size;
993 	if (end < start)
994 		return -EINVAL;
995 
996 	/* Since the mm is for the mirrored process, get a reference first. */
997 	if (!mmget_not_zero(mm))
998 		return -EINVAL;
999 
1000 	/*
1001 	 * Register a temporary notifier to detect invalidations even if it
1002 	 * overlaps with other mmu_interval_notifiers.
1003 	 */
1004 	uptr = u64_to_user_ptr(cmd->ptr);
1005 	for (addr = start; addr < end; addr = next) {
1006 		unsigned long n;
1007 
1008 		next = min(addr + (ARRAY_SIZE(pfns) << PAGE_SHIFT), end);
1009 		range.start = addr;
1010 		range.end = next;
1011 
1012 		ret = dmirror_range_snapshot(dmirror, &range, perm);
1013 		if (ret)
1014 			break;
1015 
1016 		n = (range.end - range.start) >> PAGE_SHIFT;
1017 		if (copy_to_user(uptr, perm, n)) {
1018 			ret = -EFAULT;
1019 			break;
1020 		}
1021 
1022 		cmd->cpages += n;
1023 		uptr += n;
1024 	}
1025 	mmput(mm);
1026 
1027 	return ret;
1028 }
1029 
1030 static long dmirror_fops_unlocked_ioctl(struct file *filp,
1031 					unsigned int command,
1032 					unsigned long arg)
1033 {
1034 	void __user *uarg = (void __user *)arg;
1035 	struct hmm_dmirror_cmd cmd;
1036 	struct dmirror *dmirror;
1037 	int ret;
1038 
1039 	dmirror = filp->private_data;
1040 	if (!dmirror)
1041 		return -EINVAL;
1042 
1043 	if (copy_from_user(&cmd, uarg, sizeof(cmd)))
1044 		return -EFAULT;
1045 
1046 	if (cmd.addr & ~PAGE_MASK)
1047 		return -EINVAL;
1048 	if (cmd.addr >= (cmd.addr + (cmd.npages << PAGE_SHIFT)))
1049 		return -EINVAL;
1050 
1051 	cmd.cpages = 0;
1052 	cmd.faults = 0;
1053 
1054 	switch (command) {
1055 	case HMM_DMIRROR_READ:
1056 		ret = dmirror_read(dmirror, &cmd);
1057 		break;
1058 
1059 	case HMM_DMIRROR_WRITE:
1060 		ret = dmirror_write(dmirror, &cmd);
1061 		break;
1062 
1063 	case HMM_DMIRROR_MIGRATE:
1064 		ret = dmirror_migrate(dmirror, &cmd);
1065 		break;
1066 
1067 	case HMM_DMIRROR_EXCLUSIVE:
1068 		ret = dmirror_exclusive(dmirror, &cmd);
1069 		break;
1070 
1071 	case HMM_DMIRROR_CHECK_EXCLUSIVE:
1072 		ret = dmirror_check_atomic(dmirror, cmd.addr,
1073 					cmd.addr + (cmd.npages << PAGE_SHIFT));
1074 		break;
1075 
1076 	case HMM_DMIRROR_SNAPSHOT:
1077 		ret = dmirror_snapshot(dmirror, &cmd);
1078 		break;
1079 
1080 	default:
1081 		return -EINVAL;
1082 	}
1083 	if (ret)
1084 		return ret;
1085 
1086 	if (copy_to_user(uarg, &cmd, sizeof(cmd)))
1087 		return -EFAULT;
1088 
1089 	return 0;
1090 }
1091 
1092 static const struct file_operations dmirror_fops = {
1093 	.open		= dmirror_fops_open,
1094 	.release	= dmirror_fops_release,
1095 	.unlocked_ioctl = dmirror_fops_unlocked_ioctl,
1096 	.llseek		= default_llseek,
1097 	.owner		= THIS_MODULE,
1098 };
1099 
1100 static void dmirror_devmem_free(struct page *page)
1101 {
1102 	struct page *rpage = page->zone_device_data;
1103 	struct dmirror_device *mdevice;
1104 
1105 	if (rpage)
1106 		__free_page(rpage);
1107 
1108 	mdevice = dmirror_page_to_device(page);
1109 
1110 	spin_lock(&mdevice->lock);
1111 	mdevice->cfree++;
1112 	page->zone_device_data = mdevice->free_pages;
1113 	mdevice->free_pages = page;
1114 	spin_unlock(&mdevice->lock);
1115 }
1116 
1117 static vm_fault_t dmirror_devmem_fault_alloc_and_copy(struct migrate_vma *args,
1118 						      struct dmirror *dmirror)
1119 {
1120 	const unsigned long *src = args->src;
1121 	unsigned long *dst = args->dst;
1122 	unsigned long start = args->start;
1123 	unsigned long end = args->end;
1124 	unsigned long addr;
1125 
1126 	for (addr = start; addr < end; addr += PAGE_SIZE,
1127 				       src++, dst++) {
1128 		struct page *dpage, *spage;
1129 
1130 		spage = migrate_pfn_to_page(*src);
1131 		if (!spage || !(*src & MIGRATE_PFN_MIGRATE))
1132 			continue;
1133 		spage = spage->zone_device_data;
1134 
1135 		dpage = alloc_page_vma(GFP_HIGHUSER_MOVABLE, args->vma, addr);
1136 		if (!dpage)
1137 			continue;
1138 
1139 		lock_page(dpage);
1140 		xa_erase(&dmirror->pt, addr >> PAGE_SHIFT);
1141 		copy_highpage(dpage, spage);
1142 		*dst = migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED;
1143 		if (*src & MIGRATE_PFN_WRITE)
1144 			*dst |= MIGRATE_PFN_WRITE;
1145 	}
1146 	return 0;
1147 }
1148 
1149 static vm_fault_t dmirror_devmem_fault(struct vm_fault *vmf)
1150 {
1151 	struct migrate_vma args;
1152 	unsigned long src_pfns;
1153 	unsigned long dst_pfns;
1154 	struct page *rpage;
1155 	struct dmirror *dmirror;
1156 	vm_fault_t ret;
1157 
1158 	/*
1159 	 * Normally, a device would use the page->zone_device_data to point to
1160 	 * the mirror but here we use it to hold the page for the simulated
1161 	 * device memory and that page holds the pointer to the mirror.
1162 	 */
1163 	rpage = vmf->page->zone_device_data;
1164 	dmirror = rpage->zone_device_data;
1165 
1166 	/* FIXME demonstrate how we can adjust migrate range */
1167 	args.vma = vmf->vma;
1168 	args.start = vmf->address;
1169 	args.end = args.start + PAGE_SIZE;
1170 	args.src = &src_pfns;
1171 	args.dst = &dst_pfns;
1172 	args.pgmap_owner = dmirror->mdevice;
1173 	args.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE;
1174 
1175 	if (migrate_vma_setup(&args))
1176 		return VM_FAULT_SIGBUS;
1177 
1178 	ret = dmirror_devmem_fault_alloc_and_copy(&args, dmirror);
1179 	if (ret)
1180 		return ret;
1181 	migrate_vma_pages(&args);
1182 	/*
1183 	 * No device finalize step is needed since
1184 	 * dmirror_devmem_fault_alloc_and_copy() will have already
1185 	 * invalidated the device page table.
1186 	 */
1187 	migrate_vma_finalize(&args);
1188 	return 0;
1189 }
1190 
1191 static const struct dev_pagemap_ops dmirror_devmem_ops = {
1192 	.page_free	= dmirror_devmem_free,
1193 	.migrate_to_ram	= dmirror_devmem_fault,
1194 };
1195 
1196 static int dmirror_device_init(struct dmirror_device *mdevice, int id)
1197 {
1198 	dev_t dev;
1199 	int ret;
1200 
1201 	dev = MKDEV(MAJOR(dmirror_dev), id);
1202 	mutex_init(&mdevice->devmem_lock);
1203 	spin_lock_init(&mdevice->lock);
1204 
1205 	cdev_init(&mdevice->cdevice, &dmirror_fops);
1206 	mdevice->cdevice.owner = THIS_MODULE;
1207 	ret = cdev_add(&mdevice->cdevice, dev, 1);
1208 	if (ret)
1209 		return ret;
1210 
1211 	/* Build a list of free ZONE_DEVICE private struct pages */
1212 	dmirror_allocate_chunk(mdevice, NULL);
1213 
1214 	return 0;
1215 }
1216 
1217 static void dmirror_device_remove(struct dmirror_device *mdevice)
1218 {
1219 	unsigned int i;
1220 
1221 	if (mdevice->devmem_chunks) {
1222 		for (i = 0; i < mdevice->devmem_count; i++) {
1223 			struct dmirror_chunk *devmem =
1224 				mdevice->devmem_chunks[i];
1225 
1226 			memunmap_pages(&devmem->pagemap);
1227 			release_mem_region(devmem->pagemap.range.start,
1228 					   range_len(&devmem->pagemap.range));
1229 			kfree(devmem);
1230 		}
1231 		kfree(mdevice->devmem_chunks);
1232 	}
1233 
1234 	cdev_del(&mdevice->cdevice);
1235 }
1236 
1237 static int __init hmm_dmirror_init(void)
1238 {
1239 	int ret;
1240 	int id;
1241 
1242 	ret = alloc_chrdev_region(&dmirror_dev, 0, DMIRROR_NDEVICES,
1243 				  "HMM_DMIRROR");
1244 	if (ret)
1245 		goto err_unreg;
1246 
1247 	for (id = 0; id < DMIRROR_NDEVICES; id++) {
1248 		ret = dmirror_device_init(dmirror_devices + id, id);
1249 		if (ret)
1250 			goto err_chrdev;
1251 	}
1252 
1253 	pr_info("HMM test module loaded. This is only for testing HMM.\n");
1254 	return 0;
1255 
1256 err_chrdev:
1257 	while (--id >= 0)
1258 		dmirror_device_remove(dmirror_devices + id);
1259 	unregister_chrdev_region(dmirror_dev, DMIRROR_NDEVICES);
1260 err_unreg:
1261 	return ret;
1262 }
1263 
1264 static void __exit hmm_dmirror_exit(void)
1265 {
1266 	int id;
1267 
1268 	for (id = 0; id < DMIRROR_NDEVICES; id++)
1269 		dmirror_device_remove(dmirror_devices + id);
1270 	unregister_chrdev_region(dmirror_dev, DMIRROR_NDEVICES);
1271 }
1272 
1273 module_init(hmm_dmirror_init);
1274 module_exit(hmm_dmirror_exit);
1275 MODULE_LICENSE("GPL");
1276