xref: /linux/lib/test_hmm.c (revision 43102a2012c2e2f8424d7eef52aede8e73cf2fed)
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		4
36 #define DMIRROR_RANGE_FAULT_TIMEOUT	1000
37 #define DEVMEM_CHUNK_SIZE		(256 * 1024 * 1024U)
38 #define DEVMEM_CHUNKS_RESERVE		16
39 
40 /*
41  * For device_private pages, dpage is just a dummy struct page
42  * representing a piece of device memory. dmirror_devmem_alloc_page
43  * allocates a real system memory page as backing storage to fake a
44  * real device. zone_device_data points to that backing page. But
45  * for device_coherent memory, the struct page represents real
46  * physical CPU-accessible memory that we can use directly.
47  */
48 #define BACKING_PAGE(page) (is_device_private_page((page)) ? \
49 			   (page)->zone_device_data : (page))
50 
51 static unsigned long spm_addr_dev0;
52 module_param(spm_addr_dev0, long, 0644);
53 MODULE_PARM_DESC(spm_addr_dev0,
54 		"Specify start address for SPM (special purpose memory) used for device 0. By setting this Coherent device type will be used. Make sure spm_addr_dev1 is set too. Minimum SPM size should be DEVMEM_CHUNK_SIZE.");
55 
56 static unsigned long spm_addr_dev1;
57 module_param(spm_addr_dev1, long, 0644);
58 MODULE_PARM_DESC(spm_addr_dev1,
59 		"Specify start address for SPM (special purpose memory) used for device 1. By setting this Coherent device type will be used. Make sure spm_addr_dev0 is set too. Minimum SPM size should be DEVMEM_CHUNK_SIZE.");
60 
61 static const struct dev_pagemap_ops dmirror_devmem_ops;
62 static const struct mmu_interval_notifier_ops dmirror_min_ops;
63 static dev_t dmirror_dev;
64 
65 struct dmirror_device;
66 
67 struct dmirror_bounce {
68 	void			*ptr;
69 	unsigned long		size;
70 	unsigned long		addr;
71 	unsigned long		cpages;
72 };
73 
74 #define DPT_XA_TAG_ATOMIC 1UL
75 #define DPT_XA_TAG_WRITE 3UL
76 
77 /*
78  * Data structure to track address ranges and register for mmu interval
79  * notifier updates.
80  */
81 struct dmirror_interval {
82 	struct mmu_interval_notifier	notifier;
83 	struct dmirror			*dmirror;
84 };
85 
86 /*
87  * Data attached to the open device file.
88  * Note that it might be shared after a fork().
89  */
90 struct dmirror {
91 	struct dmirror_device		*mdevice;
92 	struct xarray			pt;
93 	struct mmu_interval_notifier	notifier;
94 	struct mutex			mutex;
95 };
96 
97 /*
98  * ZONE_DEVICE pages for migration and simulating device memory.
99  */
100 struct dmirror_chunk {
101 	struct dev_pagemap	pagemap;
102 	struct dmirror_device	*mdevice;
103 	bool remove;
104 };
105 
106 /*
107  * Per device data.
108  */
109 struct dmirror_device {
110 	struct cdev		cdevice;
111 	unsigned int            zone_device_type;
112 	struct device		device;
113 
114 	unsigned int		devmem_capacity;
115 	unsigned int		devmem_count;
116 	struct dmirror_chunk	**devmem_chunks;
117 	struct mutex		devmem_lock;	/* protects the above */
118 
119 	unsigned long		calloc;
120 	unsigned long		cfree;
121 	struct page		*free_pages;
122 	spinlock_t		lock;		/* protects the above */
123 };
124 
125 static struct dmirror_device dmirror_devices[DMIRROR_NDEVICES];
126 
127 static int dmirror_bounce_init(struct dmirror_bounce *bounce,
128 			       unsigned long addr,
129 			       unsigned long size)
130 {
131 	bounce->addr = addr;
132 	bounce->size = size;
133 	bounce->cpages = 0;
134 	bounce->ptr = vmalloc(size);
135 	if (!bounce->ptr)
136 		return -ENOMEM;
137 	return 0;
138 }
139 
140 static bool dmirror_is_private_zone(struct dmirror_device *mdevice)
141 {
142 	return (mdevice->zone_device_type ==
143 		HMM_DMIRROR_MEMORY_DEVICE_PRIVATE) ? true : false;
144 }
145 
146 static enum migrate_vma_direction
147 dmirror_select_device(struct dmirror *dmirror)
148 {
149 	return (dmirror->mdevice->zone_device_type ==
150 		HMM_DMIRROR_MEMORY_DEVICE_PRIVATE) ?
151 		MIGRATE_VMA_SELECT_DEVICE_PRIVATE :
152 		MIGRATE_VMA_SELECT_DEVICE_COHERENT;
153 }
154 
155 static void dmirror_bounce_fini(struct dmirror_bounce *bounce)
156 {
157 	vfree(bounce->ptr);
158 }
159 
160 static int dmirror_fops_open(struct inode *inode, struct file *filp)
161 {
162 	struct cdev *cdev = inode->i_cdev;
163 	struct dmirror *dmirror;
164 	int ret;
165 
166 	/* Mirror this process address space */
167 	dmirror = kzalloc(sizeof(*dmirror), GFP_KERNEL);
168 	if (dmirror == NULL)
169 		return -ENOMEM;
170 
171 	dmirror->mdevice = container_of(cdev, struct dmirror_device, cdevice);
172 	mutex_init(&dmirror->mutex);
173 	xa_init(&dmirror->pt);
174 
175 	ret = mmu_interval_notifier_insert(&dmirror->notifier, current->mm,
176 				0, ULONG_MAX & PAGE_MASK, &dmirror_min_ops);
177 	if (ret) {
178 		kfree(dmirror);
179 		return ret;
180 	}
181 
182 	filp->private_data = dmirror;
183 	return 0;
184 }
185 
186 static int dmirror_fops_release(struct inode *inode, struct file *filp)
187 {
188 	struct dmirror *dmirror = filp->private_data;
189 
190 	mmu_interval_notifier_remove(&dmirror->notifier);
191 	xa_destroy(&dmirror->pt);
192 	kfree(dmirror);
193 	return 0;
194 }
195 
196 static struct dmirror_chunk *dmirror_page_to_chunk(struct page *page)
197 {
198 	return container_of(page->pgmap, struct dmirror_chunk, pagemap);
199 }
200 
201 static struct dmirror_device *dmirror_page_to_device(struct page *page)
202 
203 {
204 	return dmirror_page_to_chunk(page)->mdevice;
205 }
206 
207 static int dmirror_do_fault(struct dmirror *dmirror, struct hmm_range *range)
208 {
209 	unsigned long *pfns = range->hmm_pfns;
210 	unsigned long pfn;
211 
212 	for (pfn = (range->start >> PAGE_SHIFT);
213 	     pfn < (range->end >> PAGE_SHIFT);
214 	     pfn++, pfns++) {
215 		struct page *page;
216 		void *entry;
217 
218 		/*
219 		 * Since we asked for hmm_range_fault() to populate pages,
220 		 * it shouldn't return an error entry on success.
221 		 */
222 		WARN_ON(*pfns & HMM_PFN_ERROR);
223 		WARN_ON(!(*pfns & HMM_PFN_VALID));
224 
225 		page = hmm_pfn_to_page(*pfns);
226 		WARN_ON(!page);
227 
228 		entry = page;
229 		if (*pfns & HMM_PFN_WRITE)
230 			entry = xa_tag_pointer(entry, DPT_XA_TAG_WRITE);
231 		else if (WARN_ON(range->default_flags & HMM_PFN_WRITE))
232 			return -EFAULT;
233 		entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
234 		if (xa_is_err(entry))
235 			return xa_err(entry);
236 	}
237 
238 	return 0;
239 }
240 
241 static void dmirror_do_update(struct dmirror *dmirror, unsigned long start,
242 			      unsigned long end)
243 {
244 	unsigned long pfn;
245 	void *entry;
246 
247 	/*
248 	 * The XArray doesn't hold references to pages since it relies on
249 	 * the mmu notifier to clear page pointers when they become stale.
250 	 * Therefore, it is OK to just clear the entry.
251 	 */
252 	xa_for_each_range(&dmirror->pt, pfn, entry, start >> PAGE_SHIFT,
253 			  end >> PAGE_SHIFT)
254 		xa_erase(&dmirror->pt, pfn);
255 }
256 
257 static bool dmirror_interval_invalidate(struct mmu_interval_notifier *mni,
258 				const struct mmu_notifier_range *range,
259 				unsigned long cur_seq)
260 {
261 	struct dmirror *dmirror = container_of(mni, struct dmirror, notifier);
262 
263 	/*
264 	 * Ignore invalidation callbacks for device private pages since
265 	 * the invalidation is handled as part of the migration process.
266 	 */
267 	if (range->event == MMU_NOTIFY_MIGRATE &&
268 	    range->owner == dmirror->mdevice)
269 		return true;
270 
271 	if (mmu_notifier_range_blockable(range))
272 		mutex_lock(&dmirror->mutex);
273 	else if (!mutex_trylock(&dmirror->mutex))
274 		return false;
275 
276 	mmu_interval_set_seq(mni, cur_seq);
277 	dmirror_do_update(dmirror, range->start, range->end);
278 
279 	mutex_unlock(&dmirror->mutex);
280 	return true;
281 }
282 
283 static const struct mmu_interval_notifier_ops dmirror_min_ops = {
284 	.invalidate = dmirror_interval_invalidate,
285 };
286 
287 static int dmirror_range_fault(struct dmirror *dmirror,
288 				struct hmm_range *range)
289 {
290 	struct mm_struct *mm = dmirror->notifier.mm;
291 	unsigned long timeout =
292 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
293 	int ret;
294 
295 	while (true) {
296 		if (time_after(jiffies, timeout)) {
297 			ret = -EBUSY;
298 			goto out;
299 		}
300 
301 		range->notifier_seq = mmu_interval_read_begin(range->notifier);
302 		mmap_read_lock(mm);
303 		ret = hmm_range_fault(range);
304 		mmap_read_unlock(mm);
305 		if (ret) {
306 			if (ret == -EBUSY)
307 				continue;
308 			goto out;
309 		}
310 
311 		mutex_lock(&dmirror->mutex);
312 		if (mmu_interval_read_retry(range->notifier,
313 					    range->notifier_seq)) {
314 			mutex_unlock(&dmirror->mutex);
315 			continue;
316 		}
317 		break;
318 	}
319 
320 	ret = dmirror_do_fault(dmirror, range);
321 
322 	mutex_unlock(&dmirror->mutex);
323 out:
324 	return ret;
325 }
326 
327 static int dmirror_fault(struct dmirror *dmirror, unsigned long start,
328 			 unsigned long end, bool write)
329 {
330 	struct mm_struct *mm = dmirror->notifier.mm;
331 	unsigned long addr;
332 	unsigned long pfns[64];
333 	struct hmm_range range = {
334 		.notifier = &dmirror->notifier,
335 		.hmm_pfns = pfns,
336 		.pfn_flags_mask = 0,
337 		.default_flags =
338 			HMM_PFN_REQ_FAULT | (write ? HMM_PFN_REQ_WRITE : 0),
339 		.dev_private_owner = dmirror->mdevice,
340 	};
341 	int ret = 0;
342 
343 	/* Since the mm is for the mirrored process, get a reference first. */
344 	if (!mmget_not_zero(mm))
345 		return 0;
346 
347 	for (addr = start; addr < end; addr = range.end) {
348 		range.start = addr;
349 		range.end = min(addr + (ARRAY_SIZE(pfns) << PAGE_SHIFT), end);
350 
351 		ret = dmirror_range_fault(dmirror, &range);
352 		if (ret)
353 			break;
354 	}
355 
356 	mmput(mm);
357 	return ret;
358 }
359 
360 static int dmirror_do_read(struct dmirror *dmirror, unsigned long start,
361 			   unsigned long end, struct dmirror_bounce *bounce)
362 {
363 	unsigned long pfn;
364 	void *ptr;
365 
366 	ptr = bounce->ptr + ((start - bounce->addr) & PAGE_MASK);
367 
368 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
369 		void *entry;
370 		struct page *page;
371 
372 		entry = xa_load(&dmirror->pt, pfn);
373 		page = xa_untag_pointer(entry);
374 		if (!page)
375 			return -ENOENT;
376 
377 		memcpy_from_page(ptr, page, 0, PAGE_SIZE);
378 
379 		ptr += PAGE_SIZE;
380 		bounce->cpages++;
381 	}
382 
383 	return 0;
384 }
385 
386 static int dmirror_read(struct dmirror *dmirror, struct hmm_dmirror_cmd *cmd)
387 {
388 	struct dmirror_bounce bounce;
389 	unsigned long start, end;
390 	unsigned long size = cmd->npages << PAGE_SHIFT;
391 	int ret;
392 
393 	start = cmd->addr;
394 	end = start + size;
395 	if (end < start)
396 		return -EINVAL;
397 
398 	ret = dmirror_bounce_init(&bounce, start, size);
399 	if (ret)
400 		return ret;
401 
402 	while (1) {
403 		mutex_lock(&dmirror->mutex);
404 		ret = dmirror_do_read(dmirror, start, end, &bounce);
405 		mutex_unlock(&dmirror->mutex);
406 		if (ret != -ENOENT)
407 			break;
408 
409 		start = cmd->addr + (bounce.cpages << PAGE_SHIFT);
410 		ret = dmirror_fault(dmirror, start, end, false);
411 		if (ret)
412 			break;
413 		cmd->faults++;
414 	}
415 
416 	if (ret == 0) {
417 		if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
418 				 bounce.size))
419 			ret = -EFAULT;
420 	}
421 	cmd->cpages = bounce.cpages;
422 	dmirror_bounce_fini(&bounce);
423 	return ret;
424 }
425 
426 static int dmirror_do_write(struct dmirror *dmirror, unsigned long start,
427 			    unsigned long end, struct dmirror_bounce *bounce)
428 {
429 	unsigned long pfn;
430 	void *ptr;
431 
432 	ptr = bounce->ptr + ((start - bounce->addr) & PAGE_MASK);
433 
434 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
435 		void *entry;
436 		struct page *page;
437 
438 		entry = xa_load(&dmirror->pt, pfn);
439 		page = xa_untag_pointer(entry);
440 		if (!page || xa_pointer_tag(entry) != DPT_XA_TAG_WRITE)
441 			return -ENOENT;
442 
443 		memcpy_to_page(page, 0, ptr, PAGE_SIZE);
444 
445 		ptr += PAGE_SIZE;
446 		bounce->cpages++;
447 	}
448 
449 	return 0;
450 }
451 
452 static int dmirror_write(struct dmirror *dmirror, struct hmm_dmirror_cmd *cmd)
453 {
454 	struct dmirror_bounce bounce;
455 	unsigned long start, end;
456 	unsigned long size = cmd->npages << PAGE_SHIFT;
457 	int ret;
458 
459 	start = cmd->addr;
460 	end = start + size;
461 	if (end < start)
462 		return -EINVAL;
463 
464 	ret = dmirror_bounce_init(&bounce, start, size);
465 	if (ret)
466 		return ret;
467 	if (copy_from_user(bounce.ptr, u64_to_user_ptr(cmd->ptr),
468 			   bounce.size)) {
469 		ret = -EFAULT;
470 		goto fini;
471 	}
472 
473 	while (1) {
474 		mutex_lock(&dmirror->mutex);
475 		ret = dmirror_do_write(dmirror, start, end, &bounce);
476 		mutex_unlock(&dmirror->mutex);
477 		if (ret != -ENOENT)
478 			break;
479 
480 		start = cmd->addr + (bounce.cpages << PAGE_SHIFT);
481 		ret = dmirror_fault(dmirror, start, end, true);
482 		if (ret)
483 			break;
484 		cmd->faults++;
485 	}
486 
487 fini:
488 	cmd->cpages = bounce.cpages;
489 	dmirror_bounce_fini(&bounce);
490 	return ret;
491 }
492 
493 static int dmirror_allocate_chunk(struct dmirror_device *mdevice,
494 				   struct page **ppage)
495 {
496 	struct dmirror_chunk *devmem;
497 	struct resource *res = NULL;
498 	unsigned long pfn;
499 	unsigned long pfn_first;
500 	unsigned long pfn_last;
501 	void *ptr;
502 	int ret = -ENOMEM;
503 
504 	devmem = kzalloc(sizeof(*devmem), GFP_KERNEL);
505 	if (!devmem)
506 		return ret;
507 
508 	switch (mdevice->zone_device_type) {
509 	case HMM_DMIRROR_MEMORY_DEVICE_PRIVATE:
510 		res = request_free_mem_region(&iomem_resource, DEVMEM_CHUNK_SIZE,
511 					      "hmm_dmirror");
512 		if (IS_ERR_OR_NULL(res))
513 			goto err_devmem;
514 		devmem->pagemap.range.start = res->start;
515 		devmem->pagemap.range.end = res->end;
516 		devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
517 		break;
518 	case HMM_DMIRROR_MEMORY_DEVICE_COHERENT:
519 		devmem->pagemap.range.start = (MINOR(mdevice->cdevice.dev) - 2) ?
520 							spm_addr_dev0 :
521 							spm_addr_dev1;
522 		devmem->pagemap.range.end = devmem->pagemap.range.start +
523 					    DEVMEM_CHUNK_SIZE - 1;
524 		devmem->pagemap.type = MEMORY_DEVICE_COHERENT;
525 		break;
526 	default:
527 		ret = -EINVAL;
528 		goto err_devmem;
529 	}
530 
531 	devmem->pagemap.nr_range = 1;
532 	devmem->pagemap.ops = &dmirror_devmem_ops;
533 	devmem->pagemap.owner = mdevice;
534 
535 	mutex_lock(&mdevice->devmem_lock);
536 
537 	if (mdevice->devmem_count == mdevice->devmem_capacity) {
538 		struct dmirror_chunk **new_chunks;
539 		unsigned int new_capacity;
540 
541 		new_capacity = mdevice->devmem_capacity +
542 				DEVMEM_CHUNKS_RESERVE;
543 		new_chunks = krealloc(mdevice->devmem_chunks,
544 				sizeof(new_chunks[0]) * new_capacity,
545 				GFP_KERNEL);
546 		if (!new_chunks)
547 			goto err_release;
548 		mdevice->devmem_capacity = new_capacity;
549 		mdevice->devmem_chunks = new_chunks;
550 	}
551 	ptr = memremap_pages(&devmem->pagemap, numa_node_id());
552 	if (IS_ERR_OR_NULL(ptr)) {
553 		if (ptr)
554 			ret = PTR_ERR(ptr);
555 		else
556 			ret = -EFAULT;
557 		goto err_release;
558 	}
559 
560 	devmem->mdevice = mdevice;
561 	pfn_first = devmem->pagemap.range.start >> PAGE_SHIFT;
562 	pfn_last = pfn_first + (range_len(&devmem->pagemap.range) >> PAGE_SHIFT);
563 	mdevice->devmem_chunks[mdevice->devmem_count++] = devmem;
564 
565 	mutex_unlock(&mdevice->devmem_lock);
566 
567 	pr_info("added new %u MB chunk (total %u chunks, %u MB) PFNs [0x%lx 0x%lx)\n",
568 		DEVMEM_CHUNK_SIZE / (1024 * 1024),
569 		mdevice->devmem_count,
570 		mdevice->devmem_count * (DEVMEM_CHUNK_SIZE / (1024 * 1024)),
571 		pfn_first, pfn_last);
572 
573 	spin_lock(&mdevice->lock);
574 	for (pfn = pfn_first; pfn < pfn_last; pfn++) {
575 		struct page *page = pfn_to_page(pfn);
576 
577 		page->zone_device_data = mdevice->free_pages;
578 		mdevice->free_pages = page;
579 	}
580 	if (ppage) {
581 		*ppage = mdevice->free_pages;
582 		mdevice->free_pages = (*ppage)->zone_device_data;
583 		mdevice->calloc++;
584 	}
585 	spin_unlock(&mdevice->lock);
586 
587 	return 0;
588 
589 err_release:
590 	mutex_unlock(&mdevice->devmem_lock);
591 	if (res && devmem->pagemap.type == MEMORY_DEVICE_PRIVATE)
592 		release_mem_region(devmem->pagemap.range.start,
593 				   range_len(&devmem->pagemap.range));
594 err_devmem:
595 	kfree(devmem);
596 
597 	return ret;
598 }
599 
600 static struct page *dmirror_devmem_alloc_page(struct dmirror_device *mdevice)
601 {
602 	struct page *dpage = NULL;
603 	struct page *rpage = NULL;
604 
605 	/*
606 	 * For ZONE_DEVICE private type, this is a fake device so we allocate
607 	 * real system memory to store our device memory.
608 	 * For ZONE_DEVICE coherent type we use the actual dpage to store the
609 	 * data and ignore rpage.
610 	 */
611 	if (dmirror_is_private_zone(mdevice)) {
612 		rpage = alloc_page(GFP_HIGHUSER);
613 		if (!rpage)
614 			return NULL;
615 	}
616 	spin_lock(&mdevice->lock);
617 
618 	if (mdevice->free_pages) {
619 		dpage = mdevice->free_pages;
620 		mdevice->free_pages = dpage->zone_device_data;
621 		mdevice->calloc++;
622 		spin_unlock(&mdevice->lock);
623 	} else {
624 		spin_unlock(&mdevice->lock);
625 		if (dmirror_allocate_chunk(mdevice, &dpage))
626 			goto error;
627 	}
628 
629 	zone_device_page_init(dpage);
630 	dpage->zone_device_data = rpage;
631 	return dpage;
632 
633 error:
634 	if (rpage)
635 		__free_page(rpage);
636 	return NULL;
637 }
638 
639 static void dmirror_migrate_alloc_and_copy(struct migrate_vma *args,
640 					   struct dmirror *dmirror)
641 {
642 	struct dmirror_device *mdevice = dmirror->mdevice;
643 	const unsigned long *src = args->src;
644 	unsigned long *dst = args->dst;
645 	unsigned long addr;
646 
647 	for (addr = args->start; addr < args->end; addr += PAGE_SIZE,
648 						   src++, dst++) {
649 		struct page *spage;
650 		struct page *dpage;
651 		struct page *rpage;
652 
653 		if (!(*src & MIGRATE_PFN_MIGRATE))
654 			continue;
655 
656 		/*
657 		 * Note that spage might be NULL which is OK since it is an
658 		 * unallocated pte_none() or read-only zero page.
659 		 */
660 		spage = migrate_pfn_to_page(*src);
661 		if (WARN(spage && is_zone_device_page(spage),
662 		     "page already in device spage pfn: 0x%lx\n",
663 		     page_to_pfn(spage)))
664 			continue;
665 
666 		dpage = dmirror_devmem_alloc_page(mdevice);
667 		if (!dpage)
668 			continue;
669 
670 		rpage = BACKING_PAGE(dpage);
671 		if (spage)
672 			copy_highpage(rpage, spage);
673 		else
674 			clear_highpage(rpage);
675 
676 		/*
677 		 * Normally, a device would use the page->zone_device_data to
678 		 * point to the mirror but here we use it to hold the page for
679 		 * the simulated device memory and that page holds the pointer
680 		 * to the mirror.
681 		 */
682 		rpage->zone_device_data = dmirror;
683 
684 		pr_debug("migrating from sys to dev pfn src: 0x%lx pfn dst: 0x%lx\n",
685 			 page_to_pfn(spage), page_to_pfn(dpage));
686 		*dst = migrate_pfn(page_to_pfn(dpage));
687 		if ((*src & MIGRATE_PFN_WRITE) ||
688 		    (!spage && args->vma->vm_flags & VM_WRITE))
689 			*dst |= MIGRATE_PFN_WRITE;
690 	}
691 }
692 
693 static int dmirror_check_atomic(struct dmirror *dmirror, unsigned long start,
694 			     unsigned long end)
695 {
696 	unsigned long pfn;
697 
698 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
699 		void *entry;
700 
701 		entry = xa_load(&dmirror->pt, pfn);
702 		if (xa_pointer_tag(entry) == DPT_XA_TAG_ATOMIC)
703 			return -EPERM;
704 	}
705 
706 	return 0;
707 }
708 
709 static int dmirror_atomic_map(unsigned long start, unsigned long end,
710 			      struct page **pages, struct dmirror *dmirror)
711 {
712 	unsigned long pfn, mapped = 0;
713 	int i;
714 
715 	/* Map the migrated pages into the device's page tables. */
716 	mutex_lock(&dmirror->mutex);
717 
718 	for (i = 0, pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++, i++) {
719 		void *entry;
720 
721 		if (!pages[i])
722 			continue;
723 
724 		entry = pages[i];
725 		entry = xa_tag_pointer(entry, DPT_XA_TAG_ATOMIC);
726 		entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
727 		if (xa_is_err(entry)) {
728 			mutex_unlock(&dmirror->mutex);
729 			return xa_err(entry);
730 		}
731 
732 		mapped++;
733 	}
734 
735 	mutex_unlock(&dmirror->mutex);
736 	return mapped;
737 }
738 
739 static int dmirror_migrate_finalize_and_map(struct migrate_vma *args,
740 					    struct dmirror *dmirror)
741 {
742 	unsigned long start = args->start;
743 	unsigned long end = args->end;
744 	const unsigned long *src = args->src;
745 	const unsigned long *dst = args->dst;
746 	unsigned long pfn;
747 
748 	/* Map the migrated pages into the device's page tables. */
749 	mutex_lock(&dmirror->mutex);
750 
751 	for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++,
752 								src++, dst++) {
753 		struct page *dpage;
754 		void *entry;
755 
756 		if (!(*src & MIGRATE_PFN_MIGRATE))
757 			continue;
758 
759 		dpage = migrate_pfn_to_page(*dst);
760 		if (!dpage)
761 			continue;
762 
763 		entry = BACKING_PAGE(dpage);
764 		if (*dst & MIGRATE_PFN_WRITE)
765 			entry = xa_tag_pointer(entry, DPT_XA_TAG_WRITE);
766 		entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
767 		if (xa_is_err(entry)) {
768 			mutex_unlock(&dmirror->mutex);
769 			return xa_err(entry);
770 		}
771 	}
772 
773 	mutex_unlock(&dmirror->mutex);
774 	return 0;
775 }
776 
777 static int dmirror_exclusive(struct dmirror *dmirror,
778 			     struct hmm_dmirror_cmd *cmd)
779 {
780 	unsigned long start, end, addr;
781 	unsigned long size = cmd->npages << PAGE_SHIFT;
782 	struct mm_struct *mm = dmirror->notifier.mm;
783 	struct page *pages[64];
784 	struct dmirror_bounce bounce;
785 	unsigned long next;
786 	int ret;
787 
788 	start = cmd->addr;
789 	end = start + size;
790 	if (end < start)
791 		return -EINVAL;
792 
793 	/* Since the mm is for the mirrored process, get a reference first. */
794 	if (!mmget_not_zero(mm))
795 		return -EINVAL;
796 
797 	mmap_read_lock(mm);
798 	for (addr = start; addr < end; addr = next) {
799 		unsigned long mapped = 0;
800 		int i;
801 
802 		next = min(end, addr + (ARRAY_SIZE(pages) << PAGE_SHIFT));
803 
804 		ret = make_device_exclusive_range(mm, addr, next, pages, NULL);
805 		/*
806 		 * Do dmirror_atomic_map() iff all pages are marked for
807 		 * exclusive access to avoid accessing uninitialized
808 		 * fields of pages.
809 		 */
810 		if (ret == (next - addr) >> PAGE_SHIFT)
811 			mapped = dmirror_atomic_map(addr, next, pages, dmirror);
812 		for (i = 0; i < ret; i++) {
813 			if (pages[i]) {
814 				unlock_page(pages[i]);
815 				put_page(pages[i]);
816 			}
817 		}
818 
819 		if (addr + (mapped << PAGE_SHIFT) < next) {
820 			mmap_read_unlock(mm);
821 			mmput(mm);
822 			return -EBUSY;
823 		}
824 	}
825 	mmap_read_unlock(mm);
826 	mmput(mm);
827 
828 	/* Return the migrated data for verification. */
829 	ret = dmirror_bounce_init(&bounce, start, size);
830 	if (ret)
831 		return ret;
832 	mutex_lock(&dmirror->mutex);
833 	ret = dmirror_do_read(dmirror, start, end, &bounce);
834 	mutex_unlock(&dmirror->mutex);
835 	if (ret == 0) {
836 		if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
837 				 bounce.size))
838 			ret = -EFAULT;
839 	}
840 
841 	cmd->cpages = bounce.cpages;
842 	dmirror_bounce_fini(&bounce);
843 	return ret;
844 }
845 
846 static vm_fault_t dmirror_devmem_fault_alloc_and_copy(struct migrate_vma *args,
847 						      struct dmirror *dmirror)
848 {
849 	const unsigned long *src = args->src;
850 	unsigned long *dst = args->dst;
851 	unsigned long start = args->start;
852 	unsigned long end = args->end;
853 	unsigned long addr;
854 
855 	for (addr = start; addr < end; addr += PAGE_SIZE,
856 				       src++, dst++) {
857 		struct page *dpage, *spage;
858 
859 		spage = migrate_pfn_to_page(*src);
860 		if (!spage || !(*src & MIGRATE_PFN_MIGRATE))
861 			continue;
862 
863 		if (WARN_ON(!is_device_private_page(spage) &&
864 			    !is_device_coherent_page(spage)))
865 			continue;
866 		spage = BACKING_PAGE(spage);
867 		dpage = alloc_page_vma(GFP_HIGHUSER_MOVABLE, args->vma, addr);
868 		if (!dpage)
869 			continue;
870 		pr_debug("migrating from dev to sys pfn src: 0x%lx pfn dst: 0x%lx\n",
871 			 page_to_pfn(spage), page_to_pfn(dpage));
872 
873 		lock_page(dpage);
874 		xa_erase(&dmirror->pt, addr >> PAGE_SHIFT);
875 		copy_highpage(dpage, spage);
876 		*dst = migrate_pfn(page_to_pfn(dpage));
877 		if (*src & MIGRATE_PFN_WRITE)
878 			*dst |= MIGRATE_PFN_WRITE;
879 	}
880 	return 0;
881 }
882 
883 static unsigned long
884 dmirror_successful_migrated_pages(struct migrate_vma *migrate)
885 {
886 	unsigned long cpages = 0;
887 	unsigned long i;
888 
889 	for (i = 0; i < migrate->npages; i++) {
890 		if (migrate->src[i] & MIGRATE_PFN_VALID &&
891 		    migrate->src[i] & MIGRATE_PFN_MIGRATE)
892 			cpages++;
893 	}
894 	return cpages;
895 }
896 
897 static int dmirror_migrate_to_system(struct dmirror *dmirror,
898 				     struct hmm_dmirror_cmd *cmd)
899 {
900 	unsigned long start, end, addr;
901 	unsigned long size = cmd->npages << PAGE_SHIFT;
902 	struct mm_struct *mm = dmirror->notifier.mm;
903 	struct vm_area_struct *vma;
904 	unsigned long src_pfns[64] = { 0 };
905 	unsigned long dst_pfns[64] = { 0 };
906 	struct migrate_vma args = { 0 };
907 	unsigned long next;
908 	int ret;
909 
910 	start = cmd->addr;
911 	end = start + size;
912 	if (end < start)
913 		return -EINVAL;
914 
915 	/* Since the mm is for the mirrored process, get a reference first. */
916 	if (!mmget_not_zero(mm))
917 		return -EINVAL;
918 
919 	cmd->cpages = 0;
920 	mmap_read_lock(mm);
921 	for (addr = start; addr < end; addr = next) {
922 		vma = vma_lookup(mm, addr);
923 		if (!vma || !(vma->vm_flags & VM_READ)) {
924 			ret = -EINVAL;
925 			goto out;
926 		}
927 		next = min(end, addr + (ARRAY_SIZE(src_pfns) << PAGE_SHIFT));
928 		if (next > vma->vm_end)
929 			next = vma->vm_end;
930 
931 		args.vma = vma;
932 		args.src = src_pfns;
933 		args.dst = dst_pfns;
934 		args.start = addr;
935 		args.end = next;
936 		args.pgmap_owner = dmirror->mdevice;
937 		args.flags = dmirror_select_device(dmirror);
938 
939 		ret = migrate_vma_setup(&args);
940 		if (ret)
941 			goto out;
942 
943 		pr_debug("Migrating from device mem to sys mem\n");
944 		dmirror_devmem_fault_alloc_and_copy(&args, dmirror);
945 
946 		migrate_vma_pages(&args);
947 		cmd->cpages += dmirror_successful_migrated_pages(&args);
948 		migrate_vma_finalize(&args);
949 	}
950 out:
951 	mmap_read_unlock(mm);
952 	mmput(mm);
953 
954 	return ret;
955 }
956 
957 static int dmirror_migrate_to_device(struct dmirror *dmirror,
958 				struct hmm_dmirror_cmd *cmd)
959 {
960 	unsigned long start, end, addr;
961 	unsigned long size = cmd->npages << PAGE_SHIFT;
962 	struct mm_struct *mm = dmirror->notifier.mm;
963 	struct vm_area_struct *vma;
964 	unsigned long src_pfns[64] = { 0 };
965 	unsigned long dst_pfns[64] = { 0 };
966 	struct dmirror_bounce bounce;
967 	struct migrate_vma args = { 0 };
968 	unsigned long next;
969 	int ret;
970 
971 	start = cmd->addr;
972 	end = start + size;
973 	if (end < start)
974 		return -EINVAL;
975 
976 	/* Since the mm is for the mirrored process, get a reference first. */
977 	if (!mmget_not_zero(mm))
978 		return -EINVAL;
979 
980 	mmap_read_lock(mm);
981 	for (addr = start; addr < end; addr = next) {
982 		vma = vma_lookup(mm, addr);
983 		if (!vma || !(vma->vm_flags & VM_READ)) {
984 			ret = -EINVAL;
985 			goto out;
986 		}
987 		next = min(end, addr + (ARRAY_SIZE(src_pfns) << PAGE_SHIFT));
988 		if (next > vma->vm_end)
989 			next = vma->vm_end;
990 
991 		args.vma = vma;
992 		args.src = src_pfns;
993 		args.dst = dst_pfns;
994 		args.start = addr;
995 		args.end = next;
996 		args.pgmap_owner = dmirror->mdevice;
997 		args.flags = MIGRATE_VMA_SELECT_SYSTEM;
998 		ret = migrate_vma_setup(&args);
999 		if (ret)
1000 			goto out;
1001 
1002 		pr_debug("Migrating from sys mem to device mem\n");
1003 		dmirror_migrate_alloc_and_copy(&args, dmirror);
1004 		migrate_vma_pages(&args);
1005 		dmirror_migrate_finalize_and_map(&args, dmirror);
1006 		migrate_vma_finalize(&args);
1007 	}
1008 	mmap_read_unlock(mm);
1009 	mmput(mm);
1010 
1011 	/*
1012 	 * Return the migrated data for verification.
1013 	 * Only for pages in device zone
1014 	 */
1015 	ret = dmirror_bounce_init(&bounce, start, size);
1016 	if (ret)
1017 		return ret;
1018 	mutex_lock(&dmirror->mutex);
1019 	ret = dmirror_do_read(dmirror, start, end, &bounce);
1020 	mutex_unlock(&dmirror->mutex);
1021 	if (ret == 0) {
1022 		if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
1023 				 bounce.size))
1024 			ret = -EFAULT;
1025 	}
1026 	cmd->cpages = bounce.cpages;
1027 	dmirror_bounce_fini(&bounce);
1028 	return ret;
1029 
1030 out:
1031 	mmap_read_unlock(mm);
1032 	mmput(mm);
1033 	return ret;
1034 }
1035 
1036 static void dmirror_mkentry(struct dmirror *dmirror, struct hmm_range *range,
1037 			    unsigned char *perm, unsigned long entry)
1038 {
1039 	struct page *page;
1040 
1041 	if (entry & HMM_PFN_ERROR) {
1042 		*perm = HMM_DMIRROR_PROT_ERROR;
1043 		return;
1044 	}
1045 	if (!(entry & HMM_PFN_VALID)) {
1046 		*perm = HMM_DMIRROR_PROT_NONE;
1047 		return;
1048 	}
1049 
1050 	page = hmm_pfn_to_page(entry);
1051 	if (is_device_private_page(page)) {
1052 		/* Is the page migrated to this device or some other? */
1053 		if (dmirror->mdevice == dmirror_page_to_device(page))
1054 			*perm = HMM_DMIRROR_PROT_DEV_PRIVATE_LOCAL;
1055 		else
1056 			*perm = HMM_DMIRROR_PROT_DEV_PRIVATE_REMOTE;
1057 	} else if (is_device_coherent_page(page)) {
1058 		/* Is the page migrated to this device or some other? */
1059 		if (dmirror->mdevice == dmirror_page_to_device(page))
1060 			*perm = HMM_DMIRROR_PROT_DEV_COHERENT_LOCAL;
1061 		else
1062 			*perm = HMM_DMIRROR_PROT_DEV_COHERENT_REMOTE;
1063 	} else if (is_zero_pfn(page_to_pfn(page)))
1064 		*perm = HMM_DMIRROR_PROT_ZERO;
1065 	else
1066 		*perm = HMM_DMIRROR_PROT_NONE;
1067 	if (entry & HMM_PFN_WRITE)
1068 		*perm |= HMM_DMIRROR_PROT_WRITE;
1069 	else
1070 		*perm |= HMM_DMIRROR_PROT_READ;
1071 	if (hmm_pfn_to_map_order(entry) + PAGE_SHIFT == PMD_SHIFT)
1072 		*perm |= HMM_DMIRROR_PROT_PMD;
1073 	else if (hmm_pfn_to_map_order(entry) + PAGE_SHIFT == PUD_SHIFT)
1074 		*perm |= HMM_DMIRROR_PROT_PUD;
1075 }
1076 
1077 static bool dmirror_snapshot_invalidate(struct mmu_interval_notifier *mni,
1078 				const struct mmu_notifier_range *range,
1079 				unsigned long cur_seq)
1080 {
1081 	struct dmirror_interval *dmi =
1082 		container_of(mni, struct dmirror_interval, notifier);
1083 	struct dmirror *dmirror = dmi->dmirror;
1084 
1085 	if (mmu_notifier_range_blockable(range))
1086 		mutex_lock(&dmirror->mutex);
1087 	else if (!mutex_trylock(&dmirror->mutex))
1088 		return false;
1089 
1090 	/*
1091 	 * Snapshots only need to set the sequence number since any
1092 	 * invalidation in the interval invalidates the whole snapshot.
1093 	 */
1094 	mmu_interval_set_seq(mni, cur_seq);
1095 
1096 	mutex_unlock(&dmirror->mutex);
1097 	return true;
1098 }
1099 
1100 static const struct mmu_interval_notifier_ops dmirror_mrn_ops = {
1101 	.invalidate = dmirror_snapshot_invalidate,
1102 };
1103 
1104 static int dmirror_range_snapshot(struct dmirror *dmirror,
1105 				  struct hmm_range *range,
1106 				  unsigned char *perm)
1107 {
1108 	struct mm_struct *mm = dmirror->notifier.mm;
1109 	struct dmirror_interval notifier;
1110 	unsigned long timeout =
1111 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
1112 	unsigned long i;
1113 	unsigned long n;
1114 	int ret = 0;
1115 
1116 	notifier.dmirror = dmirror;
1117 	range->notifier = &notifier.notifier;
1118 
1119 	ret = mmu_interval_notifier_insert(range->notifier, mm,
1120 			range->start, range->end - range->start,
1121 			&dmirror_mrn_ops);
1122 	if (ret)
1123 		return ret;
1124 
1125 	while (true) {
1126 		if (time_after(jiffies, timeout)) {
1127 			ret = -EBUSY;
1128 			goto out;
1129 		}
1130 
1131 		range->notifier_seq = mmu_interval_read_begin(range->notifier);
1132 
1133 		mmap_read_lock(mm);
1134 		ret = hmm_range_fault(range);
1135 		mmap_read_unlock(mm);
1136 		if (ret) {
1137 			if (ret == -EBUSY)
1138 				continue;
1139 			goto out;
1140 		}
1141 
1142 		mutex_lock(&dmirror->mutex);
1143 		if (mmu_interval_read_retry(range->notifier,
1144 					    range->notifier_seq)) {
1145 			mutex_unlock(&dmirror->mutex);
1146 			continue;
1147 		}
1148 		break;
1149 	}
1150 
1151 	n = (range->end - range->start) >> PAGE_SHIFT;
1152 	for (i = 0; i < n; i++)
1153 		dmirror_mkentry(dmirror, range, perm + i, range->hmm_pfns[i]);
1154 
1155 	mutex_unlock(&dmirror->mutex);
1156 out:
1157 	mmu_interval_notifier_remove(range->notifier);
1158 	return ret;
1159 }
1160 
1161 static int dmirror_snapshot(struct dmirror *dmirror,
1162 			    struct hmm_dmirror_cmd *cmd)
1163 {
1164 	struct mm_struct *mm = dmirror->notifier.mm;
1165 	unsigned long start, end;
1166 	unsigned long size = cmd->npages << PAGE_SHIFT;
1167 	unsigned long addr;
1168 	unsigned long next;
1169 	unsigned long pfns[64];
1170 	unsigned char perm[64];
1171 	char __user *uptr;
1172 	struct hmm_range range = {
1173 		.hmm_pfns = pfns,
1174 		.dev_private_owner = dmirror->mdevice,
1175 	};
1176 	int ret = 0;
1177 
1178 	start = cmd->addr;
1179 	end = start + size;
1180 	if (end < start)
1181 		return -EINVAL;
1182 
1183 	/* Since the mm is for the mirrored process, get a reference first. */
1184 	if (!mmget_not_zero(mm))
1185 		return -EINVAL;
1186 
1187 	/*
1188 	 * Register a temporary notifier to detect invalidations even if it
1189 	 * overlaps with other mmu_interval_notifiers.
1190 	 */
1191 	uptr = u64_to_user_ptr(cmd->ptr);
1192 	for (addr = start; addr < end; addr = next) {
1193 		unsigned long n;
1194 
1195 		next = min(addr + (ARRAY_SIZE(pfns) << PAGE_SHIFT), end);
1196 		range.start = addr;
1197 		range.end = next;
1198 
1199 		ret = dmirror_range_snapshot(dmirror, &range, perm);
1200 		if (ret)
1201 			break;
1202 
1203 		n = (range.end - range.start) >> PAGE_SHIFT;
1204 		if (copy_to_user(uptr, perm, n)) {
1205 			ret = -EFAULT;
1206 			break;
1207 		}
1208 
1209 		cmd->cpages += n;
1210 		uptr += n;
1211 	}
1212 	mmput(mm);
1213 
1214 	return ret;
1215 }
1216 
1217 static void dmirror_device_evict_chunk(struct dmirror_chunk *chunk)
1218 {
1219 	unsigned long start_pfn = chunk->pagemap.range.start >> PAGE_SHIFT;
1220 	unsigned long end_pfn = chunk->pagemap.range.end >> PAGE_SHIFT;
1221 	unsigned long npages = end_pfn - start_pfn + 1;
1222 	unsigned long i;
1223 	unsigned long *src_pfns;
1224 	unsigned long *dst_pfns;
1225 
1226 	src_pfns = kvcalloc(npages, sizeof(*src_pfns), GFP_KERNEL | __GFP_NOFAIL);
1227 	dst_pfns = kvcalloc(npages, sizeof(*dst_pfns), GFP_KERNEL | __GFP_NOFAIL);
1228 
1229 	migrate_device_range(src_pfns, start_pfn, npages);
1230 	for (i = 0; i < npages; i++) {
1231 		struct page *dpage, *spage;
1232 
1233 		spage = migrate_pfn_to_page(src_pfns[i]);
1234 		if (!spage || !(src_pfns[i] & MIGRATE_PFN_MIGRATE))
1235 			continue;
1236 
1237 		if (WARN_ON(!is_device_private_page(spage) &&
1238 			    !is_device_coherent_page(spage)))
1239 			continue;
1240 		spage = BACKING_PAGE(spage);
1241 		dpage = alloc_page(GFP_HIGHUSER_MOVABLE | __GFP_NOFAIL);
1242 		lock_page(dpage);
1243 		copy_highpage(dpage, spage);
1244 		dst_pfns[i] = migrate_pfn(page_to_pfn(dpage));
1245 		if (src_pfns[i] & MIGRATE_PFN_WRITE)
1246 			dst_pfns[i] |= MIGRATE_PFN_WRITE;
1247 	}
1248 	migrate_device_pages(src_pfns, dst_pfns, npages);
1249 	migrate_device_finalize(src_pfns, dst_pfns, npages);
1250 	kvfree(src_pfns);
1251 	kvfree(dst_pfns);
1252 }
1253 
1254 /* Removes free pages from the free list so they can't be re-allocated */
1255 static void dmirror_remove_free_pages(struct dmirror_chunk *devmem)
1256 {
1257 	struct dmirror_device *mdevice = devmem->mdevice;
1258 	struct page *page;
1259 
1260 	for (page = mdevice->free_pages; page; page = page->zone_device_data)
1261 		if (dmirror_page_to_chunk(page) == devmem)
1262 			mdevice->free_pages = page->zone_device_data;
1263 }
1264 
1265 static void dmirror_device_remove_chunks(struct dmirror_device *mdevice)
1266 {
1267 	unsigned int i;
1268 
1269 	mutex_lock(&mdevice->devmem_lock);
1270 	if (mdevice->devmem_chunks) {
1271 		for (i = 0; i < mdevice->devmem_count; i++) {
1272 			struct dmirror_chunk *devmem =
1273 				mdevice->devmem_chunks[i];
1274 
1275 			spin_lock(&mdevice->lock);
1276 			devmem->remove = true;
1277 			dmirror_remove_free_pages(devmem);
1278 			spin_unlock(&mdevice->lock);
1279 
1280 			dmirror_device_evict_chunk(devmem);
1281 			memunmap_pages(&devmem->pagemap);
1282 			if (devmem->pagemap.type == MEMORY_DEVICE_PRIVATE)
1283 				release_mem_region(devmem->pagemap.range.start,
1284 						   range_len(&devmem->pagemap.range));
1285 			kfree(devmem);
1286 		}
1287 		mdevice->devmem_count = 0;
1288 		mdevice->devmem_capacity = 0;
1289 		mdevice->free_pages = NULL;
1290 		kfree(mdevice->devmem_chunks);
1291 		mdevice->devmem_chunks = NULL;
1292 	}
1293 	mutex_unlock(&mdevice->devmem_lock);
1294 }
1295 
1296 static long dmirror_fops_unlocked_ioctl(struct file *filp,
1297 					unsigned int command,
1298 					unsigned long arg)
1299 {
1300 	void __user *uarg = (void __user *)arg;
1301 	struct hmm_dmirror_cmd cmd;
1302 	struct dmirror *dmirror;
1303 	int ret;
1304 
1305 	dmirror = filp->private_data;
1306 	if (!dmirror)
1307 		return -EINVAL;
1308 
1309 	if (copy_from_user(&cmd, uarg, sizeof(cmd)))
1310 		return -EFAULT;
1311 
1312 	if (cmd.addr & ~PAGE_MASK)
1313 		return -EINVAL;
1314 	if (cmd.addr >= (cmd.addr + (cmd.npages << PAGE_SHIFT)))
1315 		return -EINVAL;
1316 
1317 	cmd.cpages = 0;
1318 	cmd.faults = 0;
1319 
1320 	switch (command) {
1321 	case HMM_DMIRROR_READ:
1322 		ret = dmirror_read(dmirror, &cmd);
1323 		break;
1324 
1325 	case HMM_DMIRROR_WRITE:
1326 		ret = dmirror_write(dmirror, &cmd);
1327 		break;
1328 
1329 	case HMM_DMIRROR_MIGRATE_TO_DEV:
1330 		ret = dmirror_migrate_to_device(dmirror, &cmd);
1331 		break;
1332 
1333 	case HMM_DMIRROR_MIGRATE_TO_SYS:
1334 		ret = dmirror_migrate_to_system(dmirror, &cmd);
1335 		break;
1336 
1337 	case HMM_DMIRROR_EXCLUSIVE:
1338 		ret = dmirror_exclusive(dmirror, &cmd);
1339 		break;
1340 
1341 	case HMM_DMIRROR_CHECK_EXCLUSIVE:
1342 		ret = dmirror_check_atomic(dmirror, cmd.addr,
1343 					cmd.addr + (cmd.npages << PAGE_SHIFT));
1344 		break;
1345 
1346 	case HMM_DMIRROR_SNAPSHOT:
1347 		ret = dmirror_snapshot(dmirror, &cmd);
1348 		break;
1349 
1350 	case HMM_DMIRROR_RELEASE:
1351 		dmirror_device_remove_chunks(dmirror->mdevice);
1352 		ret = 0;
1353 		break;
1354 
1355 	default:
1356 		return -EINVAL;
1357 	}
1358 	if (ret)
1359 		return ret;
1360 
1361 	if (copy_to_user(uarg, &cmd, sizeof(cmd)))
1362 		return -EFAULT;
1363 
1364 	return 0;
1365 }
1366 
1367 static int dmirror_fops_mmap(struct file *file, struct vm_area_struct *vma)
1368 {
1369 	unsigned long addr;
1370 
1371 	for (addr = vma->vm_start; addr < vma->vm_end; addr += PAGE_SIZE) {
1372 		struct page *page;
1373 		int ret;
1374 
1375 		page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1376 		if (!page)
1377 			return -ENOMEM;
1378 
1379 		ret = vm_insert_page(vma, addr, page);
1380 		if (ret) {
1381 			__free_page(page);
1382 			return ret;
1383 		}
1384 		put_page(page);
1385 	}
1386 
1387 	return 0;
1388 }
1389 
1390 static const struct file_operations dmirror_fops = {
1391 	.open		= dmirror_fops_open,
1392 	.release	= dmirror_fops_release,
1393 	.mmap		= dmirror_fops_mmap,
1394 	.unlocked_ioctl = dmirror_fops_unlocked_ioctl,
1395 	.llseek		= default_llseek,
1396 	.owner		= THIS_MODULE,
1397 };
1398 
1399 static void dmirror_devmem_free(struct page *page)
1400 {
1401 	struct page *rpage = BACKING_PAGE(page);
1402 	struct dmirror_device *mdevice;
1403 
1404 	if (rpage != page)
1405 		__free_page(rpage);
1406 
1407 	mdevice = dmirror_page_to_device(page);
1408 	spin_lock(&mdevice->lock);
1409 
1410 	/* Return page to our allocator if not freeing the chunk */
1411 	if (!dmirror_page_to_chunk(page)->remove) {
1412 		mdevice->cfree++;
1413 		page->zone_device_data = mdevice->free_pages;
1414 		mdevice->free_pages = page;
1415 	}
1416 	spin_unlock(&mdevice->lock);
1417 }
1418 
1419 static vm_fault_t dmirror_devmem_fault(struct vm_fault *vmf)
1420 {
1421 	struct migrate_vma args = { 0 };
1422 	unsigned long src_pfns = 0;
1423 	unsigned long dst_pfns = 0;
1424 	struct page *rpage;
1425 	struct dmirror *dmirror;
1426 	vm_fault_t ret;
1427 
1428 	/*
1429 	 * Normally, a device would use the page->zone_device_data to point to
1430 	 * the mirror but here we use it to hold the page for the simulated
1431 	 * device memory and that page holds the pointer to the mirror.
1432 	 */
1433 	rpage = vmf->page->zone_device_data;
1434 	dmirror = rpage->zone_device_data;
1435 
1436 	/* FIXME demonstrate how we can adjust migrate range */
1437 	args.vma = vmf->vma;
1438 	args.start = vmf->address;
1439 	args.end = args.start + PAGE_SIZE;
1440 	args.src = &src_pfns;
1441 	args.dst = &dst_pfns;
1442 	args.pgmap_owner = dmirror->mdevice;
1443 	args.flags = dmirror_select_device(dmirror);
1444 	args.fault_page = vmf->page;
1445 
1446 	if (migrate_vma_setup(&args))
1447 		return VM_FAULT_SIGBUS;
1448 
1449 	ret = dmirror_devmem_fault_alloc_and_copy(&args, dmirror);
1450 	if (ret)
1451 		return ret;
1452 	migrate_vma_pages(&args);
1453 	/*
1454 	 * No device finalize step is needed since
1455 	 * dmirror_devmem_fault_alloc_and_copy() will have already
1456 	 * invalidated the device page table.
1457 	 */
1458 	migrate_vma_finalize(&args);
1459 	return 0;
1460 }
1461 
1462 static const struct dev_pagemap_ops dmirror_devmem_ops = {
1463 	.page_free	= dmirror_devmem_free,
1464 	.migrate_to_ram	= dmirror_devmem_fault,
1465 };
1466 
1467 static int dmirror_device_init(struct dmirror_device *mdevice, int id)
1468 {
1469 	dev_t dev;
1470 	int ret;
1471 
1472 	dev = MKDEV(MAJOR(dmirror_dev), id);
1473 	mutex_init(&mdevice->devmem_lock);
1474 	spin_lock_init(&mdevice->lock);
1475 
1476 	cdev_init(&mdevice->cdevice, &dmirror_fops);
1477 	mdevice->cdevice.owner = THIS_MODULE;
1478 	device_initialize(&mdevice->device);
1479 	mdevice->device.devt = dev;
1480 
1481 	ret = dev_set_name(&mdevice->device, "hmm_dmirror%u", id);
1482 	if (ret)
1483 		return ret;
1484 
1485 	ret = cdev_device_add(&mdevice->cdevice, &mdevice->device);
1486 	if (ret)
1487 		return ret;
1488 
1489 	/* Build a list of free ZONE_DEVICE struct pages */
1490 	return dmirror_allocate_chunk(mdevice, NULL);
1491 }
1492 
1493 static void dmirror_device_remove(struct dmirror_device *mdevice)
1494 {
1495 	dmirror_device_remove_chunks(mdevice);
1496 	cdev_device_del(&mdevice->cdevice, &mdevice->device);
1497 }
1498 
1499 static int __init hmm_dmirror_init(void)
1500 {
1501 	int ret;
1502 	int id = 0;
1503 	int ndevices = 0;
1504 
1505 	ret = alloc_chrdev_region(&dmirror_dev, 0, DMIRROR_NDEVICES,
1506 				  "HMM_DMIRROR");
1507 	if (ret)
1508 		goto err_unreg;
1509 
1510 	memset(dmirror_devices, 0, DMIRROR_NDEVICES * sizeof(dmirror_devices[0]));
1511 	dmirror_devices[ndevices++].zone_device_type =
1512 				HMM_DMIRROR_MEMORY_DEVICE_PRIVATE;
1513 	dmirror_devices[ndevices++].zone_device_type =
1514 				HMM_DMIRROR_MEMORY_DEVICE_PRIVATE;
1515 	if (spm_addr_dev0 && spm_addr_dev1) {
1516 		dmirror_devices[ndevices++].zone_device_type =
1517 					HMM_DMIRROR_MEMORY_DEVICE_COHERENT;
1518 		dmirror_devices[ndevices++].zone_device_type =
1519 					HMM_DMIRROR_MEMORY_DEVICE_COHERENT;
1520 	}
1521 	for (id = 0; id < ndevices; id++) {
1522 		ret = dmirror_device_init(dmirror_devices + id, id);
1523 		if (ret)
1524 			goto err_chrdev;
1525 	}
1526 
1527 	pr_info("HMM test module loaded. This is only for testing HMM.\n");
1528 	return 0;
1529 
1530 err_chrdev:
1531 	while (--id >= 0)
1532 		dmirror_device_remove(dmirror_devices + id);
1533 	unregister_chrdev_region(dmirror_dev, DMIRROR_NDEVICES);
1534 err_unreg:
1535 	return ret;
1536 }
1537 
1538 static void __exit hmm_dmirror_exit(void)
1539 {
1540 	int id;
1541 
1542 	for (id = 0; id < DMIRROR_NDEVICES; id++)
1543 		if (dmirror_devices[id].zone_device_type)
1544 			dmirror_device_remove(dmirror_devices + id);
1545 	unregister_chrdev_region(dmirror_dev, DMIRROR_NDEVICES);
1546 }
1547 
1548 module_init(hmm_dmirror_init);
1549 module_exit(hmm_dmirror_exit);
1550 MODULE_DESCRIPTION("HMM (Heterogeneous Memory Management) test module");
1551 MODULE_LICENSE("GPL");
1552