xref: /openbmc/linux/lib/test_hmm.c (revision 65b96377)
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/memremap.h>
16 #include <linux/mutex.h>
17 #include <linux/rwsem.h>
18 #include <linux/sched.h>
19 #include <linux/slab.h>
20 #include <linux/highmem.h>
21 #include <linux/delay.h>
22 #include <linux/pagemap.h>
23 #include <linux/hmm.h>
24 #include <linux/vmalloc.h>
25 #include <linux/swap.h>
26 #include <linux/swapops.h>
27 #include <linux/sched/mm.h>
28 #include <linux/platform_device.h>
29 #include <linux/rmap.h>
30 #include <linux/mmu_notifier.h>
31 #include <linux/migrate.h>
32 
33 #include "test_hmm_uapi.h"
34 
35 #define DMIRROR_NDEVICES		2
36 #define DMIRROR_RANGE_FAULT_TIMEOUT	1000
37 #define DEVMEM_CHUNK_SIZE		(256 * 1024 * 1024U)
38 #define DEVMEM_CHUNKS_RESERVE		16
39 
40 static const struct dev_pagemap_ops dmirror_devmem_ops;
41 static const struct mmu_interval_notifier_ops dmirror_min_ops;
42 static dev_t dmirror_dev;
43 
44 struct dmirror_device;
45 
46 struct dmirror_bounce {
47 	void			*ptr;
48 	unsigned long		size;
49 	unsigned long		addr;
50 	unsigned long		cpages;
51 };
52 
53 #define DPT_XA_TAG_ATOMIC 1UL
54 #define DPT_XA_TAG_WRITE 3UL
55 
56 /*
57  * Data structure to track address ranges and register for mmu interval
58  * notifier updates.
59  */
60 struct dmirror_interval {
61 	struct mmu_interval_notifier	notifier;
62 	struct dmirror			*dmirror;
63 };
64 
65 /*
66  * Data attached to the open device file.
67  * Note that it might be shared after a fork().
68  */
69 struct dmirror {
70 	struct dmirror_device		*mdevice;
71 	struct xarray			pt;
72 	struct mmu_interval_notifier	notifier;
73 	struct mutex			mutex;
74 };
75 
76 /*
77  * ZONE_DEVICE pages for migration and simulating device memory.
78  */
79 struct dmirror_chunk {
80 	struct dev_pagemap	pagemap;
81 	struct dmirror_device	*mdevice;
82 };
83 
84 /*
85  * Per device data.
86  */
87 struct dmirror_device {
88 	struct cdev		cdevice;
89 	struct hmm_devmem	*devmem;
90 
91 	unsigned int		devmem_capacity;
92 	unsigned int		devmem_count;
93 	struct dmirror_chunk	**devmem_chunks;
94 	struct mutex		devmem_lock;	/* protects the above */
95 
96 	unsigned long		calloc;
97 	unsigned long		cfree;
98 	struct page		*free_pages;
99 	spinlock_t		lock;		/* protects the above */
100 };
101 
102 static struct dmirror_device dmirror_devices[DMIRROR_NDEVICES];
103 
104 static int dmirror_bounce_init(struct dmirror_bounce *bounce,
105 			       unsigned long addr,
106 			       unsigned long size)
107 {
108 	bounce->addr = addr;
109 	bounce->size = size;
110 	bounce->cpages = 0;
111 	bounce->ptr = vmalloc(size);
112 	if (!bounce->ptr)
113 		return -ENOMEM;
114 	return 0;
115 }
116 
117 static void dmirror_bounce_fini(struct dmirror_bounce *bounce)
118 {
119 	vfree(bounce->ptr);
120 }
121 
122 static int dmirror_fops_open(struct inode *inode, struct file *filp)
123 {
124 	struct cdev *cdev = inode->i_cdev;
125 	struct dmirror *dmirror;
126 	int ret;
127 
128 	/* Mirror this process address space */
129 	dmirror = kzalloc(sizeof(*dmirror), GFP_KERNEL);
130 	if (dmirror == NULL)
131 		return -ENOMEM;
132 
133 	dmirror->mdevice = container_of(cdev, struct dmirror_device, cdevice);
134 	mutex_init(&dmirror->mutex);
135 	xa_init(&dmirror->pt);
136 
137 	ret = mmu_interval_notifier_insert(&dmirror->notifier, current->mm,
138 				0, ULONG_MAX & PAGE_MASK, &dmirror_min_ops);
139 	if (ret) {
140 		kfree(dmirror);
141 		return ret;
142 	}
143 
144 	filp->private_data = dmirror;
145 	return 0;
146 }
147 
148 static int dmirror_fops_release(struct inode *inode, struct file *filp)
149 {
150 	struct dmirror *dmirror = filp->private_data;
151 
152 	mmu_interval_notifier_remove(&dmirror->notifier);
153 	xa_destroy(&dmirror->pt);
154 	kfree(dmirror);
155 	return 0;
156 }
157 
158 static struct dmirror_device *dmirror_page_to_device(struct page *page)
159 
160 {
161 	return container_of(page->pgmap, struct dmirror_chunk,
162 			    pagemap)->mdevice;
163 }
164 
165 static int dmirror_do_fault(struct dmirror *dmirror, struct hmm_range *range)
166 {
167 	unsigned long *pfns = range->hmm_pfns;
168 	unsigned long pfn;
169 
170 	for (pfn = (range->start >> PAGE_SHIFT);
171 	     pfn < (range->end >> PAGE_SHIFT);
172 	     pfn++, pfns++) {
173 		struct page *page;
174 		void *entry;
175 
176 		/*
177 		 * Since we asked for hmm_range_fault() to populate pages,
178 		 * it shouldn't return an error entry on success.
179 		 */
180 		WARN_ON(*pfns & HMM_PFN_ERROR);
181 		WARN_ON(!(*pfns & HMM_PFN_VALID));
182 
183 		page = hmm_pfn_to_page(*pfns);
184 		WARN_ON(!page);
185 
186 		entry = page;
187 		if (*pfns & HMM_PFN_WRITE)
188 			entry = xa_tag_pointer(entry, DPT_XA_TAG_WRITE);
189 		else if (WARN_ON(range->default_flags & HMM_PFN_WRITE))
190 			return -EFAULT;
191 		entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
192 		if (xa_is_err(entry))
193 			return xa_err(entry);
194 	}
195 
196 	return 0;
197 }
198 
199 static void dmirror_do_update(struct dmirror *dmirror, unsigned long start,
200 			      unsigned long end)
201 {
202 	unsigned long pfn;
203 	void *entry;
204 
205 	/*
206 	 * The XArray doesn't hold references to pages since it relies on
207 	 * the mmu notifier to clear page pointers when they become stale.
208 	 * Therefore, it is OK to just clear the entry.
209 	 */
210 	xa_for_each_range(&dmirror->pt, pfn, entry, start >> PAGE_SHIFT,
211 			  end >> PAGE_SHIFT)
212 		xa_erase(&dmirror->pt, pfn);
213 }
214 
215 static bool dmirror_interval_invalidate(struct mmu_interval_notifier *mni,
216 				const struct mmu_notifier_range *range,
217 				unsigned long cur_seq)
218 {
219 	struct dmirror *dmirror = container_of(mni, struct dmirror, notifier);
220 
221 	/*
222 	 * Ignore invalidation callbacks for device private pages since
223 	 * the invalidation is handled as part of the migration process.
224 	 */
225 	if (range->event == MMU_NOTIFY_MIGRATE &&
226 	    range->owner == dmirror->mdevice)
227 		return true;
228 
229 	if (mmu_notifier_range_blockable(range))
230 		mutex_lock(&dmirror->mutex);
231 	else if (!mutex_trylock(&dmirror->mutex))
232 		return false;
233 
234 	mmu_interval_set_seq(mni, cur_seq);
235 	dmirror_do_update(dmirror, range->start, range->end);
236 
237 	mutex_unlock(&dmirror->mutex);
238 	return true;
239 }
240 
241 static const struct mmu_interval_notifier_ops dmirror_min_ops = {
242 	.invalidate = dmirror_interval_invalidate,
243 };
244 
245 static int dmirror_range_fault(struct dmirror *dmirror,
246 				struct hmm_range *range)
247 {
248 	struct mm_struct *mm = dmirror->notifier.mm;
249 	unsigned long timeout =
250 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
251 	int ret;
252 
253 	while (true) {
254 		if (time_after(jiffies, timeout)) {
255 			ret = -EBUSY;
256 			goto out;
257 		}
258 
259 		range->notifier_seq = mmu_interval_read_begin(range->notifier);
260 		mmap_read_lock(mm);
261 		ret = hmm_range_fault(range);
262 		mmap_read_unlock(mm);
263 		if (ret) {
264 			if (ret == -EBUSY)
265 				continue;
266 			goto out;
267 		}
268 
269 		mutex_lock(&dmirror->mutex);
270 		if (mmu_interval_read_retry(range->notifier,
271 					    range->notifier_seq)) {
272 			mutex_unlock(&dmirror->mutex);
273 			continue;
274 		}
275 		break;
276 	}
277 
278 	ret = dmirror_do_fault(dmirror, range);
279 
280 	mutex_unlock(&dmirror->mutex);
281 out:
282 	return ret;
283 }
284 
285 static int dmirror_fault(struct dmirror *dmirror, unsigned long start,
286 			 unsigned long end, bool write)
287 {
288 	struct mm_struct *mm = dmirror->notifier.mm;
289 	unsigned long addr;
290 	unsigned long pfns[64];
291 	struct hmm_range range = {
292 		.notifier = &dmirror->notifier,
293 		.hmm_pfns = pfns,
294 		.pfn_flags_mask = 0,
295 		.default_flags =
296 			HMM_PFN_REQ_FAULT | (write ? HMM_PFN_REQ_WRITE : 0),
297 		.dev_private_owner = dmirror->mdevice,
298 	};
299 	int ret = 0;
300 
301 	/* Since the mm is for the mirrored process, get a reference first. */
302 	if (!mmget_not_zero(mm))
303 		return 0;
304 
305 	for (addr = start; addr < end; addr = range.end) {
306 		range.start = addr;
307 		range.end = min(addr + (ARRAY_SIZE(pfns) << PAGE_SHIFT), end);
308 
309 		ret = dmirror_range_fault(dmirror, &range);
310 		if (ret)
311 			break;
312 	}
313 
314 	mmput(mm);
315 	return ret;
316 }
317 
318 static int dmirror_do_read(struct dmirror *dmirror, unsigned long start,
319 			   unsigned long end, struct dmirror_bounce *bounce)
320 {
321 	unsigned long pfn;
322 	void *ptr;
323 
324 	ptr = bounce->ptr + ((start - bounce->addr) & PAGE_MASK);
325 
326 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
327 		void *entry;
328 		struct page *page;
329 		void *tmp;
330 
331 		entry = xa_load(&dmirror->pt, pfn);
332 		page = xa_untag_pointer(entry);
333 		if (!page)
334 			return -ENOENT;
335 
336 		tmp = kmap(page);
337 		memcpy(ptr, tmp, PAGE_SIZE);
338 		kunmap(page);
339 
340 		ptr += PAGE_SIZE;
341 		bounce->cpages++;
342 	}
343 
344 	return 0;
345 }
346 
347 static int dmirror_read(struct dmirror *dmirror, struct hmm_dmirror_cmd *cmd)
348 {
349 	struct dmirror_bounce bounce;
350 	unsigned long start, end;
351 	unsigned long size = cmd->npages << PAGE_SHIFT;
352 	int ret;
353 
354 	start = cmd->addr;
355 	end = start + size;
356 	if (end < start)
357 		return -EINVAL;
358 
359 	ret = dmirror_bounce_init(&bounce, start, size);
360 	if (ret)
361 		return ret;
362 
363 	while (1) {
364 		mutex_lock(&dmirror->mutex);
365 		ret = dmirror_do_read(dmirror, start, end, &bounce);
366 		mutex_unlock(&dmirror->mutex);
367 		if (ret != -ENOENT)
368 			break;
369 
370 		start = cmd->addr + (bounce.cpages << PAGE_SHIFT);
371 		ret = dmirror_fault(dmirror, start, end, false);
372 		if (ret)
373 			break;
374 		cmd->faults++;
375 	}
376 
377 	if (ret == 0) {
378 		if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
379 				 bounce.size))
380 			ret = -EFAULT;
381 	}
382 	cmd->cpages = bounce.cpages;
383 	dmirror_bounce_fini(&bounce);
384 	return ret;
385 }
386 
387 static int dmirror_do_write(struct dmirror *dmirror, unsigned long start,
388 			    unsigned long end, struct dmirror_bounce *bounce)
389 {
390 	unsigned long pfn;
391 	void *ptr;
392 
393 	ptr = bounce->ptr + ((start - bounce->addr) & PAGE_MASK);
394 
395 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
396 		void *entry;
397 		struct page *page;
398 		void *tmp;
399 
400 		entry = xa_load(&dmirror->pt, pfn);
401 		page = xa_untag_pointer(entry);
402 		if (!page || xa_pointer_tag(entry) != DPT_XA_TAG_WRITE)
403 			return -ENOENT;
404 
405 		tmp = kmap(page);
406 		memcpy(tmp, ptr, PAGE_SIZE);
407 		kunmap(page);
408 
409 		ptr += PAGE_SIZE;
410 		bounce->cpages++;
411 	}
412 
413 	return 0;
414 }
415 
416 static int dmirror_write(struct dmirror *dmirror, struct hmm_dmirror_cmd *cmd)
417 {
418 	struct dmirror_bounce bounce;
419 	unsigned long start, end;
420 	unsigned long size = cmd->npages << PAGE_SHIFT;
421 	int ret;
422 
423 	start = cmd->addr;
424 	end = start + size;
425 	if (end < start)
426 		return -EINVAL;
427 
428 	ret = dmirror_bounce_init(&bounce, start, size);
429 	if (ret)
430 		return ret;
431 	if (copy_from_user(bounce.ptr, u64_to_user_ptr(cmd->ptr),
432 			   bounce.size)) {
433 		ret = -EFAULT;
434 		goto fini;
435 	}
436 
437 	while (1) {
438 		mutex_lock(&dmirror->mutex);
439 		ret = dmirror_do_write(dmirror, start, end, &bounce);
440 		mutex_unlock(&dmirror->mutex);
441 		if (ret != -ENOENT)
442 			break;
443 
444 		start = cmd->addr + (bounce.cpages << PAGE_SHIFT);
445 		ret = dmirror_fault(dmirror, start, end, true);
446 		if (ret)
447 			break;
448 		cmd->faults++;
449 	}
450 
451 fini:
452 	cmd->cpages = bounce.cpages;
453 	dmirror_bounce_fini(&bounce);
454 	return ret;
455 }
456 
457 static bool dmirror_allocate_chunk(struct dmirror_device *mdevice,
458 				   struct page **ppage)
459 {
460 	struct dmirror_chunk *devmem;
461 	struct resource *res;
462 	unsigned long pfn;
463 	unsigned long pfn_first;
464 	unsigned long pfn_last;
465 	void *ptr;
466 
467 	devmem = kzalloc(sizeof(*devmem), GFP_KERNEL);
468 	if (!devmem)
469 		return false;
470 
471 	res = request_free_mem_region(&iomem_resource, DEVMEM_CHUNK_SIZE,
472 				      "hmm_dmirror");
473 	if (IS_ERR(res))
474 		goto err_devmem;
475 
476 	devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
477 	devmem->pagemap.range.start = res->start;
478 	devmem->pagemap.range.end = res->end;
479 	devmem->pagemap.nr_range = 1;
480 	devmem->pagemap.ops = &dmirror_devmem_ops;
481 	devmem->pagemap.owner = mdevice;
482 
483 	mutex_lock(&mdevice->devmem_lock);
484 
485 	if (mdevice->devmem_count == mdevice->devmem_capacity) {
486 		struct dmirror_chunk **new_chunks;
487 		unsigned int new_capacity;
488 
489 		new_capacity = mdevice->devmem_capacity +
490 				DEVMEM_CHUNKS_RESERVE;
491 		new_chunks = krealloc(mdevice->devmem_chunks,
492 				sizeof(new_chunks[0]) * new_capacity,
493 				GFP_KERNEL);
494 		if (!new_chunks)
495 			goto err_release;
496 		mdevice->devmem_capacity = new_capacity;
497 		mdevice->devmem_chunks = new_chunks;
498 	}
499 
500 	ptr = memremap_pages(&devmem->pagemap, numa_node_id());
501 	if (IS_ERR(ptr))
502 		goto err_release;
503 
504 	devmem->mdevice = mdevice;
505 	pfn_first = devmem->pagemap.range.start >> PAGE_SHIFT;
506 	pfn_last = pfn_first + (range_len(&devmem->pagemap.range) >> PAGE_SHIFT);
507 	mdevice->devmem_chunks[mdevice->devmem_count++] = devmem;
508 
509 	mutex_unlock(&mdevice->devmem_lock);
510 
511 	pr_info("added new %u MB chunk (total %u chunks, %u MB) PFNs [0x%lx 0x%lx)\n",
512 		DEVMEM_CHUNK_SIZE / (1024 * 1024),
513 		mdevice->devmem_count,
514 		mdevice->devmem_count * (DEVMEM_CHUNK_SIZE / (1024 * 1024)),
515 		pfn_first, pfn_last);
516 
517 	spin_lock(&mdevice->lock);
518 	for (pfn = pfn_first; pfn < pfn_last; pfn++) {
519 		struct page *page = pfn_to_page(pfn);
520 
521 		page->zone_device_data = mdevice->free_pages;
522 		mdevice->free_pages = page;
523 	}
524 	if (ppage) {
525 		*ppage = mdevice->free_pages;
526 		mdevice->free_pages = (*ppage)->zone_device_data;
527 		mdevice->calloc++;
528 	}
529 	spin_unlock(&mdevice->lock);
530 
531 	return true;
532 
533 err_release:
534 	mutex_unlock(&mdevice->devmem_lock);
535 	release_mem_region(devmem->pagemap.range.start, range_len(&devmem->pagemap.range));
536 err_devmem:
537 	kfree(devmem);
538 
539 	return false;
540 }
541 
542 static struct page *dmirror_devmem_alloc_page(struct dmirror_device *mdevice)
543 {
544 	struct page *dpage = NULL;
545 	struct page *rpage;
546 
547 	/*
548 	 * This is a fake device so we alloc real system memory to store
549 	 * our device memory.
550 	 */
551 	rpage = alloc_page(GFP_HIGHUSER);
552 	if (!rpage)
553 		return NULL;
554 
555 	spin_lock(&mdevice->lock);
556 
557 	if (mdevice->free_pages) {
558 		dpage = mdevice->free_pages;
559 		mdevice->free_pages = dpage->zone_device_data;
560 		mdevice->calloc++;
561 		spin_unlock(&mdevice->lock);
562 	} else {
563 		spin_unlock(&mdevice->lock);
564 		if (!dmirror_allocate_chunk(mdevice, &dpage))
565 			goto error;
566 	}
567 
568 	dpage->zone_device_data = rpage;
569 	lock_page(dpage);
570 	return dpage;
571 
572 error:
573 	__free_page(rpage);
574 	return NULL;
575 }
576 
577 static void dmirror_migrate_alloc_and_copy(struct migrate_vma *args,
578 					   struct dmirror *dmirror)
579 {
580 	struct dmirror_device *mdevice = dmirror->mdevice;
581 	const unsigned long *src = args->src;
582 	unsigned long *dst = args->dst;
583 	unsigned long addr;
584 
585 	for (addr = args->start; addr < args->end; addr += PAGE_SIZE,
586 						   src++, dst++) {
587 		struct page *spage;
588 		struct page *dpage;
589 		struct page *rpage;
590 
591 		if (!(*src & MIGRATE_PFN_MIGRATE))
592 			continue;
593 
594 		/*
595 		 * Note that spage might be NULL which is OK since it is an
596 		 * unallocated pte_none() or read-only zero page.
597 		 */
598 		spage = migrate_pfn_to_page(*src);
599 
600 		dpage = dmirror_devmem_alloc_page(mdevice);
601 		if (!dpage)
602 			continue;
603 
604 		rpage = dpage->zone_device_data;
605 		if (spage)
606 			copy_highpage(rpage, spage);
607 		else
608 			clear_highpage(rpage);
609 
610 		/*
611 		 * Normally, a device would use the page->zone_device_data to
612 		 * point to the mirror but here we use it to hold the page for
613 		 * the simulated device memory and that page holds the pointer
614 		 * to the mirror.
615 		 */
616 		rpage->zone_device_data = dmirror;
617 
618 		*dst = migrate_pfn(page_to_pfn(dpage));
619 		if ((*src & MIGRATE_PFN_WRITE) ||
620 		    (!spage && args->vma->vm_flags & VM_WRITE))
621 			*dst |= MIGRATE_PFN_WRITE;
622 	}
623 }
624 
625 static int dmirror_check_atomic(struct dmirror *dmirror, unsigned long start,
626 			     unsigned long end)
627 {
628 	unsigned long pfn;
629 
630 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
631 		void *entry;
632 
633 		entry = xa_load(&dmirror->pt, pfn);
634 		if (xa_pointer_tag(entry) == DPT_XA_TAG_ATOMIC)
635 			return -EPERM;
636 	}
637 
638 	return 0;
639 }
640 
641 static int dmirror_atomic_map(unsigned long start, unsigned long end,
642 			      struct page **pages, struct dmirror *dmirror)
643 {
644 	unsigned long pfn, mapped = 0;
645 	int i;
646 
647 	/* Map the migrated pages into the device's page tables. */
648 	mutex_lock(&dmirror->mutex);
649 
650 	for (i = 0, pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++, i++) {
651 		void *entry;
652 
653 		if (!pages[i])
654 			continue;
655 
656 		entry = pages[i];
657 		entry = xa_tag_pointer(entry, DPT_XA_TAG_ATOMIC);
658 		entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
659 		if (xa_is_err(entry)) {
660 			mutex_unlock(&dmirror->mutex);
661 			return xa_err(entry);
662 		}
663 
664 		mapped++;
665 	}
666 
667 	mutex_unlock(&dmirror->mutex);
668 	return mapped;
669 }
670 
671 static int dmirror_migrate_finalize_and_map(struct migrate_vma *args,
672 					    struct dmirror *dmirror)
673 {
674 	unsigned long start = args->start;
675 	unsigned long end = args->end;
676 	const unsigned long *src = args->src;
677 	const unsigned long *dst = args->dst;
678 	unsigned long pfn;
679 
680 	/* Map the migrated pages into the device's page tables. */
681 	mutex_lock(&dmirror->mutex);
682 
683 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++,
684 								src++, dst++) {
685 		struct page *dpage;
686 		void *entry;
687 
688 		if (!(*src & MIGRATE_PFN_MIGRATE))
689 			continue;
690 
691 		dpage = migrate_pfn_to_page(*dst);
692 		if (!dpage)
693 			continue;
694 
695 		/*
696 		 * Store the page that holds the data so the page table
697 		 * doesn't have to deal with ZONE_DEVICE private pages.
698 		 */
699 		entry = dpage->zone_device_data;
700 		if (*dst & MIGRATE_PFN_WRITE)
701 			entry = xa_tag_pointer(entry, DPT_XA_TAG_WRITE);
702 		entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
703 		if (xa_is_err(entry)) {
704 			mutex_unlock(&dmirror->mutex);
705 			return xa_err(entry);
706 		}
707 	}
708 
709 	mutex_unlock(&dmirror->mutex);
710 	return 0;
711 }
712 
713 static int dmirror_exclusive(struct dmirror *dmirror,
714 			     struct hmm_dmirror_cmd *cmd)
715 {
716 	unsigned long start, end, addr;
717 	unsigned long size = cmd->npages << PAGE_SHIFT;
718 	struct mm_struct *mm = dmirror->notifier.mm;
719 	struct page *pages[64];
720 	struct dmirror_bounce bounce;
721 	unsigned long next;
722 	int ret;
723 
724 	start = cmd->addr;
725 	end = start + size;
726 	if (end < start)
727 		return -EINVAL;
728 
729 	/* Since the mm is for the mirrored process, get a reference first. */
730 	if (!mmget_not_zero(mm))
731 		return -EINVAL;
732 
733 	mmap_read_lock(mm);
734 	for (addr = start; addr < end; addr = next) {
735 		unsigned long mapped;
736 		int i;
737 
738 		if (end < addr + (ARRAY_SIZE(pages) << PAGE_SHIFT))
739 			next = end;
740 		else
741 			next = addr + (ARRAY_SIZE(pages) << PAGE_SHIFT);
742 
743 		ret = make_device_exclusive_range(mm, addr, next, pages, NULL);
744 		mapped = dmirror_atomic_map(addr, next, pages, dmirror);
745 		for (i = 0; i < ret; i++) {
746 			if (pages[i]) {
747 				unlock_page(pages[i]);
748 				put_page(pages[i]);
749 			}
750 		}
751 
752 		if (addr + (mapped << PAGE_SHIFT) < next) {
753 			mmap_read_unlock(mm);
754 			mmput(mm);
755 			return -EBUSY;
756 		}
757 	}
758 	mmap_read_unlock(mm);
759 	mmput(mm);
760 
761 	/* Return the migrated data for verification. */
762 	ret = dmirror_bounce_init(&bounce, start, size);
763 	if (ret)
764 		return ret;
765 	mutex_lock(&dmirror->mutex);
766 	ret = dmirror_do_read(dmirror, start, end, &bounce);
767 	mutex_unlock(&dmirror->mutex);
768 	if (ret == 0) {
769 		if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
770 				 bounce.size))
771 			ret = -EFAULT;
772 	}
773 
774 	cmd->cpages = bounce.cpages;
775 	dmirror_bounce_fini(&bounce);
776 	return ret;
777 }
778 
779 static int dmirror_migrate(struct dmirror *dmirror,
780 			   struct hmm_dmirror_cmd *cmd)
781 {
782 	unsigned long start, end, addr;
783 	unsigned long size = cmd->npages << PAGE_SHIFT;
784 	struct mm_struct *mm = dmirror->notifier.mm;
785 	struct vm_area_struct *vma;
786 	unsigned long src_pfns[64];
787 	unsigned long dst_pfns[64];
788 	struct dmirror_bounce bounce;
789 	struct migrate_vma args;
790 	unsigned long next;
791 	int ret;
792 
793 	start = cmd->addr;
794 	end = start + size;
795 	if (end < start)
796 		return -EINVAL;
797 
798 	/* Since the mm is for the mirrored process, get a reference first. */
799 	if (!mmget_not_zero(mm))
800 		return -EINVAL;
801 
802 	mmap_read_lock(mm);
803 	for (addr = start; addr < end; addr = next) {
804 		vma = vma_lookup(mm, addr);
805 		if (!vma || !(vma->vm_flags & VM_READ)) {
806 			ret = -EINVAL;
807 			goto out;
808 		}
809 		next = min(end, addr + (ARRAY_SIZE(src_pfns) << PAGE_SHIFT));
810 		if (next > vma->vm_end)
811 			next = vma->vm_end;
812 
813 		args.vma = vma;
814 		args.src = src_pfns;
815 		args.dst = dst_pfns;
816 		args.start = addr;
817 		args.end = next;
818 		args.pgmap_owner = dmirror->mdevice;
819 		args.flags = MIGRATE_VMA_SELECT_SYSTEM;
820 		ret = migrate_vma_setup(&args);
821 		if (ret)
822 			goto out;
823 
824 		dmirror_migrate_alloc_and_copy(&args, dmirror);
825 		migrate_vma_pages(&args);
826 		dmirror_migrate_finalize_and_map(&args, dmirror);
827 		migrate_vma_finalize(&args);
828 	}
829 	mmap_read_unlock(mm);
830 	mmput(mm);
831 
832 	/* Return the migrated data for verification. */
833 	ret = dmirror_bounce_init(&bounce, start, size);
834 	if (ret)
835 		return ret;
836 	mutex_lock(&dmirror->mutex);
837 	ret = dmirror_do_read(dmirror, start, end, &bounce);
838 	mutex_unlock(&dmirror->mutex);
839 	if (ret == 0) {
840 		if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
841 				 bounce.size))
842 			ret = -EFAULT;
843 	}
844 	cmd->cpages = bounce.cpages;
845 	dmirror_bounce_fini(&bounce);
846 	return ret;
847 
848 out:
849 	mmap_read_unlock(mm);
850 	mmput(mm);
851 	return ret;
852 }
853 
854 static void dmirror_mkentry(struct dmirror *dmirror, struct hmm_range *range,
855 			    unsigned char *perm, unsigned long entry)
856 {
857 	struct page *page;
858 
859 	if (entry & HMM_PFN_ERROR) {
860 		*perm = HMM_DMIRROR_PROT_ERROR;
861 		return;
862 	}
863 	if (!(entry & HMM_PFN_VALID)) {
864 		*perm = HMM_DMIRROR_PROT_NONE;
865 		return;
866 	}
867 
868 	page = hmm_pfn_to_page(entry);
869 	if (is_device_private_page(page)) {
870 		/* Is the page migrated to this device or some other? */
871 		if (dmirror->mdevice == dmirror_page_to_device(page))
872 			*perm = HMM_DMIRROR_PROT_DEV_PRIVATE_LOCAL;
873 		else
874 			*perm = HMM_DMIRROR_PROT_DEV_PRIVATE_REMOTE;
875 	} else if (is_zero_pfn(page_to_pfn(page)))
876 		*perm = HMM_DMIRROR_PROT_ZERO;
877 	else
878 		*perm = HMM_DMIRROR_PROT_NONE;
879 	if (entry & HMM_PFN_WRITE)
880 		*perm |= HMM_DMIRROR_PROT_WRITE;
881 	else
882 		*perm |= HMM_DMIRROR_PROT_READ;
883 	if (hmm_pfn_to_map_order(entry) + PAGE_SHIFT == PMD_SHIFT)
884 		*perm |= HMM_DMIRROR_PROT_PMD;
885 	else if (hmm_pfn_to_map_order(entry) + PAGE_SHIFT == PUD_SHIFT)
886 		*perm |= HMM_DMIRROR_PROT_PUD;
887 }
888 
889 static bool dmirror_snapshot_invalidate(struct mmu_interval_notifier *mni,
890 				const struct mmu_notifier_range *range,
891 				unsigned long cur_seq)
892 {
893 	struct dmirror_interval *dmi =
894 		container_of(mni, struct dmirror_interval, notifier);
895 	struct dmirror *dmirror = dmi->dmirror;
896 
897 	if (mmu_notifier_range_blockable(range))
898 		mutex_lock(&dmirror->mutex);
899 	else if (!mutex_trylock(&dmirror->mutex))
900 		return false;
901 
902 	/*
903 	 * Snapshots only need to set the sequence number since any
904 	 * invalidation in the interval invalidates the whole snapshot.
905 	 */
906 	mmu_interval_set_seq(mni, cur_seq);
907 
908 	mutex_unlock(&dmirror->mutex);
909 	return true;
910 }
911 
912 static const struct mmu_interval_notifier_ops dmirror_mrn_ops = {
913 	.invalidate = dmirror_snapshot_invalidate,
914 };
915 
916 static int dmirror_range_snapshot(struct dmirror *dmirror,
917 				  struct hmm_range *range,
918 				  unsigned char *perm)
919 {
920 	struct mm_struct *mm = dmirror->notifier.mm;
921 	struct dmirror_interval notifier;
922 	unsigned long timeout =
923 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
924 	unsigned long i;
925 	unsigned long n;
926 	int ret = 0;
927 
928 	notifier.dmirror = dmirror;
929 	range->notifier = &notifier.notifier;
930 
931 	ret = mmu_interval_notifier_insert(range->notifier, mm,
932 			range->start, range->end - range->start,
933 			&dmirror_mrn_ops);
934 	if (ret)
935 		return ret;
936 
937 	while (true) {
938 		if (time_after(jiffies, timeout)) {
939 			ret = -EBUSY;
940 			goto out;
941 		}
942 
943 		range->notifier_seq = mmu_interval_read_begin(range->notifier);
944 
945 		mmap_read_lock(mm);
946 		ret = hmm_range_fault(range);
947 		mmap_read_unlock(mm);
948 		if (ret) {
949 			if (ret == -EBUSY)
950 				continue;
951 			goto out;
952 		}
953 
954 		mutex_lock(&dmirror->mutex);
955 		if (mmu_interval_read_retry(range->notifier,
956 					    range->notifier_seq)) {
957 			mutex_unlock(&dmirror->mutex);
958 			continue;
959 		}
960 		break;
961 	}
962 
963 	n = (range->end - range->start) >> PAGE_SHIFT;
964 	for (i = 0; i < n; i++)
965 		dmirror_mkentry(dmirror, range, perm + i, range->hmm_pfns[i]);
966 
967 	mutex_unlock(&dmirror->mutex);
968 out:
969 	mmu_interval_notifier_remove(range->notifier);
970 	return ret;
971 }
972 
973 static int dmirror_snapshot(struct dmirror *dmirror,
974 			    struct hmm_dmirror_cmd *cmd)
975 {
976 	struct mm_struct *mm = dmirror->notifier.mm;
977 	unsigned long start, end;
978 	unsigned long size = cmd->npages << PAGE_SHIFT;
979 	unsigned long addr;
980 	unsigned long next;
981 	unsigned long pfns[64];
982 	unsigned char perm[64];
983 	char __user *uptr;
984 	struct hmm_range range = {
985 		.hmm_pfns = pfns,
986 		.dev_private_owner = dmirror->mdevice,
987 	};
988 	int ret = 0;
989 
990 	start = cmd->addr;
991 	end = start + size;
992 	if (end < start)
993 		return -EINVAL;
994 
995 	/* Since the mm is for the mirrored process, get a reference first. */
996 	if (!mmget_not_zero(mm))
997 		return -EINVAL;
998 
999 	/*
1000 	 * Register a temporary notifier to detect invalidations even if it
1001 	 * overlaps with other mmu_interval_notifiers.
1002 	 */
1003 	uptr = u64_to_user_ptr(cmd->ptr);
1004 	for (addr = start; addr < end; addr = next) {
1005 		unsigned long n;
1006 
1007 		next = min(addr + (ARRAY_SIZE(pfns) << PAGE_SHIFT), end);
1008 		range.start = addr;
1009 		range.end = next;
1010 
1011 		ret = dmirror_range_snapshot(dmirror, &range, perm);
1012 		if (ret)
1013 			break;
1014 
1015 		n = (range.end - range.start) >> PAGE_SHIFT;
1016 		if (copy_to_user(uptr, perm, n)) {
1017 			ret = -EFAULT;
1018 			break;
1019 		}
1020 
1021 		cmd->cpages += n;
1022 		uptr += n;
1023 	}
1024 	mmput(mm);
1025 
1026 	return ret;
1027 }
1028 
1029 static long dmirror_fops_unlocked_ioctl(struct file *filp,
1030 					unsigned int command,
1031 					unsigned long arg)
1032 {
1033 	void __user *uarg = (void __user *)arg;
1034 	struct hmm_dmirror_cmd cmd;
1035 	struct dmirror *dmirror;
1036 	int ret;
1037 
1038 	dmirror = filp->private_data;
1039 	if (!dmirror)
1040 		return -EINVAL;
1041 
1042 	if (copy_from_user(&cmd, uarg, sizeof(cmd)))
1043 		return -EFAULT;
1044 
1045 	if (cmd.addr & ~PAGE_MASK)
1046 		return -EINVAL;
1047 	if (cmd.addr >= (cmd.addr + (cmd.npages << PAGE_SHIFT)))
1048 		return -EINVAL;
1049 
1050 	cmd.cpages = 0;
1051 	cmd.faults = 0;
1052 
1053 	switch (command) {
1054 	case HMM_DMIRROR_READ:
1055 		ret = dmirror_read(dmirror, &cmd);
1056 		break;
1057 
1058 	case HMM_DMIRROR_WRITE:
1059 		ret = dmirror_write(dmirror, &cmd);
1060 		break;
1061 
1062 	case HMM_DMIRROR_MIGRATE:
1063 		ret = dmirror_migrate(dmirror, &cmd);
1064 		break;
1065 
1066 	case HMM_DMIRROR_EXCLUSIVE:
1067 		ret = dmirror_exclusive(dmirror, &cmd);
1068 		break;
1069 
1070 	case HMM_DMIRROR_CHECK_EXCLUSIVE:
1071 		ret = dmirror_check_atomic(dmirror, cmd.addr,
1072 					cmd.addr + (cmd.npages << PAGE_SHIFT));
1073 		break;
1074 
1075 	case HMM_DMIRROR_SNAPSHOT:
1076 		ret = dmirror_snapshot(dmirror, &cmd);
1077 		break;
1078 
1079 	default:
1080 		return -EINVAL;
1081 	}
1082 	if (ret)
1083 		return ret;
1084 
1085 	if (copy_to_user(uarg, &cmd, sizeof(cmd)))
1086 		return -EFAULT;
1087 
1088 	return 0;
1089 }
1090 
1091 static int dmirror_fops_mmap(struct file *file, struct vm_area_struct *vma)
1092 {
1093 	unsigned long addr;
1094 
1095 	for (addr = vma->vm_start; addr < vma->vm_end; addr += PAGE_SIZE) {
1096 		struct page *page;
1097 		int ret;
1098 
1099 		page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1100 		if (!page)
1101 			return -ENOMEM;
1102 
1103 		ret = vm_insert_page(vma, addr, page);
1104 		if (ret) {
1105 			__free_page(page);
1106 			return ret;
1107 		}
1108 		put_page(page);
1109 	}
1110 
1111 	return 0;
1112 }
1113 
1114 static const struct file_operations dmirror_fops = {
1115 	.open		= dmirror_fops_open,
1116 	.release	= dmirror_fops_release,
1117 	.mmap		= dmirror_fops_mmap,
1118 	.unlocked_ioctl = dmirror_fops_unlocked_ioctl,
1119 	.llseek		= default_llseek,
1120 	.owner		= THIS_MODULE,
1121 };
1122 
1123 static void dmirror_devmem_free(struct page *page)
1124 {
1125 	struct page *rpage = page->zone_device_data;
1126 	struct dmirror_device *mdevice;
1127 
1128 	if (rpage)
1129 		__free_page(rpage);
1130 
1131 	mdevice = dmirror_page_to_device(page);
1132 
1133 	spin_lock(&mdevice->lock);
1134 	mdevice->cfree++;
1135 	page->zone_device_data = mdevice->free_pages;
1136 	mdevice->free_pages = page;
1137 	spin_unlock(&mdevice->lock);
1138 }
1139 
1140 static vm_fault_t dmirror_devmem_fault_alloc_and_copy(struct migrate_vma *args,
1141 						      struct dmirror *dmirror)
1142 {
1143 	const unsigned long *src = args->src;
1144 	unsigned long *dst = args->dst;
1145 	unsigned long start = args->start;
1146 	unsigned long end = args->end;
1147 	unsigned long addr;
1148 
1149 	for (addr = start; addr < end; addr += PAGE_SIZE,
1150 				       src++, dst++) {
1151 		struct page *dpage, *spage;
1152 
1153 		spage = migrate_pfn_to_page(*src);
1154 		if (!spage || !(*src & MIGRATE_PFN_MIGRATE))
1155 			continue;
1156 		spage = spage->zone_device_data;
1157 
1158 		dpage = alloc_page_vma(GFP_HIGHUSER_MOVABLE, args->vma, addr);
1159 		if (!dpage)
1160 			continue;
1161 
1162 		lock_page(dpage);
1163 		xa_erase(&dmirror->pt, addr >> PAGE_SHIFT);
1164 		copy_highpage(dpage, spage);
1165 		*dst = migrate_pfn(page_to_pfn(dpage));
1166 		if (*src & MIGRATE_PFN_WRITE)
1167 			*dst |= MIGRATE_PFN_WRITE;
1168 	}
1169 	return 0;
1170 }
1171 
1172 static vm_fault_t dmirror_devmem_fault(struct vm_fault *vmf)
1173 {
1174 	struct migrate_vma args;
1175 	unsigned long src_pfns;
1176 	unsigned long dst_pfns;
1177 	struct page *rpage;
1178 	struct dmirror *dmirror;
1179 	vm_fault_t ret;
1180 
1181 	/*
1182 	 * Normally, a device would use the page->zone_device_data to point to
1183 	 * the mirror but here we use it to hold the page for the simulated
1184 	 * device memory and that page holds the pointer to the mirror.
1185 	 */
1186 	rpage = vmf->page->zone_device_data;
1187 	dmirror = rpage->zone_device_data;
1188 
1189 	/* FIXME demonstrate how we can adjust migrate range */
1190 	args.vma = vmf->vma;
1191 	args.start = vmf->address;
1192 	args.end = args.start + PAGE_SIZE;
1193 	args.src = &src_pfns;
1194 	args.dst = &dst_pfns;
1195 	args.pgmap_owner = dmirror->mdevice;
1196 	args.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE;
1197 
1198 	if (migrate_vma_setup(&args))
1199 		return VM_FAULT_SIGBUS;
1200 
1201 	ret = dmirror_devmem_fault_alloc_and_copy(&args, dmirror);
1202 	if (ret)
1203 		return ret;
1204 	migrate_vma_pages(&args);
1205 	/*
1206 	 * No device finalize step is needed since
1207 	 * dmirror_devmem_fault_alloc_and_copy() will have already
1208 	 * invalidated the device page table.
1209 	 */
1210 	migrate_vma_finalize(&args);
1211 	return 0;
1212 }
1213 
1214 static const struct dev_pagemap_ops dmirror_devmem_ops = {
1215 	.page_free	= dmirror_devmem_free,
1216 	.migrate_to_ram	= dmirror_devmem_fault,
1217 };
1218 
1219 static int dmirror_device_init(struct dmirror_device *mdevice, int id)
1220 {
1221 	dev_t dev;
1222 	int ret;
1223 
1224 	dev = MKDEV(MAJOR(dmirror_dev), id);
1225 	mutex_init(&mdevice->devmem_lock);
1226 	spin_lock_init(&mdevice->lock);
1227 
1228 	cdev_init(&mdevice->cdevice, &dmirror_fops);
1229 	mdevice->cdevice.owner = THIS_MODULE;
1230 	ret = cdev_add(&mdevice->cdevice, dev, 1);
1231 	if (ret)
1232 		return ret;
1233 
1234 	/* Build a list of free ZONE_DEVICE private struct pages */
1235 	dmirror_allocate_chunk(mdevice, NULL);
1236 
1237 	return 0;
1238 }
1239 
1240 static void dmirror_device_remove(struct dmirror_device *mdevice)
1241 {
1242 	unsigned int i;
1243 
1244 	if (mdevice->devmem_chunks) {
1245 		for (i = 0; i < mdevice->devmem_count; i++) {
1246 			struct dmirror_chunk *devmem =
1247 				mdevice->devmem_chunks[i];
1248 
1249 			memunmap_pages(&devmem->pagemap);
1250 			release_mem_region(devmem->pagemap.range.start,
1251 					   range_len(&devmem->pagemap.range));
1252 			kfree(devmem);
1253 		}
1254 		kfree(mdevice->devmem_chunks);
1255 	}
1256 
1257 	cdev_del(&mdevice->cdevice);
1258 }
1259 
1260 static int __init hmm_dmirror_init(void)
1261 {
1262 	int ret;
1263 	int id;
1264 
1265 	ret = alloc_chrdev_region(&dmirror_dev, 0, DMIRROR_NDEVICES,
1266 				  "HMM_DMIRROR");
1267 	if (ret)
1268 		goto err_unreg;
1269 
1270 	for (id = 0; id < DMIRROR_NDEVICES; id++) {
1271 		ret = dmirror_device_init(dmirror_devices + id, id);
1272 		if (ret)
1273 			goto err_chrdev;
1274 	}
1275 
1276 	pr_info("HMM test module loaded. This is only for testing HMM.\n");
1277 	return 0;
1278 
1279 err_chrdev:
1280 	while (--id >= 0)
1281 		dmirror_device_remove(dmirror_devices + id);
1282 	unregister_chrdev_region(dmirror_dev, DMIRROR_NDEVICES);
1283 err_unreg:
1284 	return ret;
1285 }
1286 
1287 static void __exit hmm_dmirror_exit(void)
1288 {
1289 	int id;
1290 
1291 	for (id = 0; id < DMIRROR_NDEVICES; id++)
1292 		dmirror_device_remove(dmirror_devices + id);
1293 	unregister_chrdev_region(dmirror_dev, DMIRROR_NDEVICES);
1294 }
1295 
1296 module_init(hmm_dmirror_init);
1297 module_exit(hmm_dmirror_exit);
1298 MODULE_LICENSE("GPL");
1299