1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Secure pages management: Migration of pages between normal and secure
4  * memory of KVM guests.
5  *
6  * Copyright 2018 Bharata B Rao, IBM Corp. <bharata@linux.ibm.com>
7  */
8 
9 /*
10  * A pseries guest can be run as secure guest on Ultravisor-enabled
11  * POWER platforms. On such platforms, this driver will be used to manage
12  * the movement of guest pages between the normal memory managed by
13  * hypervisor (HV) and secure memory managed by Ultravisor (UV).
14  *
15  * The page-in or page-out requests from UV will come to HV as hcalls and
16  * HV will call back into UV via ultracalls to satisfy these page requests.
17  *
18  * Private ZONE_DEVICE memory equal to the amount of secure memory
19  * available in the platform for running secure guests is hotplugged.
20  * Whenever a page belonging to the guest becomes secure, a page from this
21  * private device memory is used to represent and track that secure page
22  * on the HV side. Some pages (like virtio buffers, VPA pages etc) are
23  * shared between UV and HV. However such pages aren't represented by
24  * device private memory and mappings to shared memory exist in both
25  * UV and HV page tables.
26  */
27 
28 /*
29  * Notes on locking
30  *
31  * kvm->arch.uvmem_lock is a per-guest lock that prevents concurrent
32  * page-in and page-out requests for the same GPA. Concurrent accesses
33  * can either come via UV (guest vCPUs requesting for same page)
34  * or when HV and guest simultaneously access the same page.
35  * This mutex serializes the migration of page from HV(normal) to
36  * UV(secure) and vice versa. So the serialization points are around
37  * migrate_vma routines and page-in/out routines.
38  *
39  * Per-guest mutex comes with a cost though. Mainly it serializes the
40  * fault path as page-out can occur when HV faults on accessing secure
41  * guest pages. Currently UV issues page-in requests for all the guest
42  * PFNs one at a time during early boot (UV_ESM uvcall), so this is
43  * not a cause for concern. Also currently the number of page-outs caused
44  * by HV touching secure pages is very very low. If an when UV supports
45  * overcommitting, then we might see concurrent guest driven page-outs.
46  *
47  * Locking order
48  *
49  * 1. kvm->srcu - Protects KVM memslots
50  * 2. kvm->mm->mmap_lock - find_vma, migrate_vma_pages and helpers, ksm_madvise
51  * 3. kvm->arch.uvmem_lock - protects read/writes to uvmem slots thus acting
52  *			     as sync-points for page-in/out
53  */
54 
55 /*
56  * Notes on page size
57  *
58  * Currently UV uses 2MB mappings internally, but will issue H_SVM_PAGE_IN
59  * and H_SVM_PAGE_OUT hcalls in PAGE_SIZE(64K) granularity. HV tracks
60  * secure GPAs at 64K page size and maintains one device PFN for each
61  * 64K secure GPA. UV_PAGE_IN and UV_PAGE_OUT calls by HV are also issued
62  * for 64K page at a time.
63  *
64  * HV faulting on secure pages: When HV touches any secure page, it
65  * faults and issues a UV_PAGE_OUT request with 64K page size. Currently
66  * UV splits and remaps the 2MB page if necessary and copies out the
67  * required 64K page contents.
68  *
69  * Shared pages: Whenever guest shares a secure page, UV will split and
70  * remap the 2MB page if required and issue H_SVM_PAGE_IN with 64K page size.
71  *
72  * HV invalidating a page: When a regular page belonging to secure
73  * guest gets unmapped, HV informs UV with UV_PAGE_INVAL of 64K
74  * page size. Using 64K page size is correct here because any non-secure
75  * page will essentially be of 64K page size. Splitting by UV during sharing
76  * and page-out ensures this.
77  *
78  * Page fault handling: When HV handles page fault of a page belonging
79  * to secure guest, it sends that to UV with a 64K UV_PAGE_IN request.
80  * Using 64K size is correct here too as UV would have split the 2MB page
81  * into 64k mappings and would have done page-outs earlier.
82  *
83  * In summary, the current secure pages handling code in HV assumes
84  * 64K page size and in fact fails any page-in/page-out requests of
85  * non-64K size upfront. If and when UV starts supporting multiple
86  * page-sizes, we need to break this assumption.
87  */
88 
89 #include <linux/pagemap.h>
90 #include <linux/migrate.h>
91 #include <linux/kvm_host.h>
92 #include <linux/ksm.h>
93 #include <asm/ultravisor.h>
94 #include <asm/mman.h>
95 #include <asm/kvm_ppc.h>
96 #include <asm/kvm_book3s_uvmem.h>
97 
98 static struct dev_pagemap kvmppc_uvmem_pgmap;
99 static unsigned long *kvmppc_uvmem_bitmap;
100 static DEFINE_SPINLOCK(kvmppc_uvmem_bitmap_lock);
101 
102 /*
103  * States of a GFN
104  * ---------------
105  * The GFN can be in one of the following states.
106  *
107  * (a) Secure - The GFN is secure. The GFN is associated with
108  *	a Secure VM, the contents of the GFN is not accessible
109  *	to the Hypervisor.  This GFN can be backed by a secure-PFN,
110  *	or can be backed by a normal-PFN with contents encrypted.
111  *	The former is true when the GFN is paged-in into the
112  *	ultravisor. The latter is true when the GFN is paged-out
113  *	of the ultravisor.
114  *
115  * (b) Shared - The GFN is shared. The GFN is associated with a
116  *	a secure VM. The contents of the GFN is accessible to
117  *	Hypervisor. This GFN is backed by a normal-PFN and its
118  *	content is un-encrypted.
119  *
120  * (c) Normal - The GFN is a normal. The GFN is associated with
121  *	a normal VM. The contents of the GFN is accesible to
122  *	the Hypervisor. Its content is never encrypted.
123  *
124  * States of a VM.
125  * ---------------
126  *
127  * Normal VM:  A VM whose contents are always accessible to
128  *	the hypervisor.  All its GFNs are normal-GFNs.
129  *
130  * Secure VM: A VM whose contents are not accessible to the
131  *	hypervisor without the VM's consent.  Its GFNs are
132  *	either Shared-GFN or Secure-GFNs.
133  *
134  * Transient VM: A Normal VM that is transitioning to secure VM.
135  *	The transition starts on successful return of
136  *	H_SVM_INIT_START, and ends on successful return
137  *	of H_SVM_INIT_DONE. This transient VM, can have GFNs
138  *	in any of the three states; i.e Secure-GFN, Shared-GFN,
139  *	and Normal-GFN.	The VM never executes in this state
140  *	in supervisor-mode.
141  *
142  * Memory slot State.
143  * -----------------------------
144  *	The state of a memory slot mirrors the state of the
145  *	VM the memory slot is associated with.
146  *
147  * VM State transition.
148  * --------------------
149  *
150  *  A VM always starts in Normal Mode.
151  *
152  *  H_SVM_INIT_START moves the VM into transient state. During this
153  *  time the Ultravisor may request some of its GFNs to be shared or
154  *  secured. So its GFNs can be in one of the three GFN states.
155  *
156  *  H_SVM_INIT_DONE moves the VM entirely from transient state to
157  *  secure-state. At this point any left-over normal-GFNs are
158  *  transitioned to Secure-GFN.
159  *
160  *  H_SVM_INIT_ABORT moves the transient VM back to normal VM.
161  *  All its GFNs are moved to Normal-GFNs.
162  *
163  *  UV_TERMINATE transitions the secure-VM back to normal-VM. All
164  *  the secure-GFN and shared-GFNs are tranistioned to normal-GFN
165  *  Note: The contents of the normal-GFN is undefined at this point.
166  *
167  * GFN state implementation:
168  * -------------------------
169  *
170  * Secure GFN is associated with a secure-PFN; also called uvmem_pfn,
171  * when the GFN is paged-in. Its pfn[] has KVMPPC_GFN_UVMEM_PFN flag
172  * set, and contains the value of the secure-PFN.
173  * It is associated with a normal-PFN; also called mem_pfn, when
174  * the GFN is pagedout. Its pfn[] has KVMPPC_GFN_MEM_PFN flag set.
175  * The value of the normal-PFN is not tracked.
176  *
177  * Shared GFN is associated with a normal-PFN. Its pfn[] has
178  * KVMPPC_UVMEM_SHARED_PFN flag set. The value of the normal-PFN
179  * is not tracked.
180  *
181  * Normal GFN is associated with normal-PFN. Its pfn[] has
182  * no flag set. The value of the normal-PFN is not tracked.
183  *
184  * Life cycle of a GFN
185  * --------------------
186  *
187  * --------------------------------------------------------------
188  * |        |     Share  |  Unshare | SVM       |H_SVM_INIT_DONE|
189  * |        |operation   |operation | abort/    |               |
190  * |        |            |          | terminate |               |
191  * -------------------------------------------------------------
192  * |        |            |          |           |               |
193  * | Secure |     Shared | Secure   |Normal     |Secure         |
194  * |        |            |          |           |               |
195  * | Shared |     Shared | Secure   |Normal     |Shared         |
196  * |        |            |          |           |               |
197  * | Normal |     Shared | Secure   |Normal     |Secure         |
198  * --------------------------------------------------------------
199  *
200  * Life cycle of a VM
201  * --------------------
202  *
203  * --------------------------------------------------------------------
204  * |         |  start    |  H_SVM_  |H_SVM_   |H_SVM_     |UV_SVM_    |
205  * |         |  VM       |INIT_START|INIT_DONE|INIT_ABORT |TERMINATE  |
206  * |         |           |          |         |           |           |
207  * --------- ----------------------------------------------------------
208  * |         |           |          |         |           |           |
209  * | Normal  | Normal    | Transient|Error    |Error      |Normal     |
210  * |         |           |          |         |           |           |
211  * | Secure  |   Error   | Error    |Error    |Error      |Normal     |
212  * |         |           |          |         |           |           |
213  * |Transient|   N/A     | Error    |Secure   |Normal     |Normal     |
214  * --------------------------------------------------------------------
215  */
216 
217 #define KVMPPC_GFN_UVMEM_PFN	(1UL << 63)
218 #define KVMPPC_GFN_MEM_PFN	(1UL << 62)
219 #define KVMPPC_GFN_SHARED	(1UL << 61)
220 #define KVMPPC_GFN_SECURE	(KVMPPC_GFN_UVMEM_PFN | KVMPPC_GFN_MEM_PFN)
221 #define KVMPPC_GFN_FLAG_MASK	(KVMPPC_GFN_SECURE | KVMPPC_GFN_SHARED)
222 #define KVMPPC_GFN_PFN_MASK	(~KVMPPC_GFN_FLAG_MASK)
223 
224 struct kvmppc_uvmem_slot {
225 	struct list_head list;
226 	unsigned long nr_pfns;
227 	unsigned long base_pfn;
228 	unsigned long *pfns;
229 };
230 struct kvmppc_uvmem_page_pvt {
231 	struct kvm *kvm;
232 	unsigned long gpa;
233 	bool skip_page_out;
234 	bool remove_gfn;
235 };
236 
237 bool kvmppc_uvmem_available(void)
238 {
239 	/*
240 	 * If kvmppc_uvmem_bitmap != NULL, then there is an ultravisor
241 	 * and our data structures have been initialized successfully.
242 	 */
243 	return !!kvmppc_uvmem_bitmap;
244 }
245 
246 int kvmppc_uvmem_slot_init(struct kvm *kvm, const struct kvm_memory_slot *slot)
247 {
248 	struct kvmppc_uvmem_slot *p;
249 
250 	p = kzalloc(sizeof(*p), GFP_KERNEL);
251 	if (!p)
252 		return -ENOMEM;
253 	p->pfns = vzalloc(array_size(slot->npages, sizeof(*p->pfns)));
254 	if (!p->pfns) {
255 		kfree(p);
256 		return -ENOMEM;
257 	}
258 	p->nr_pfns = slot->npages;
259 	p->base_pfn = slot->base_gfn;
260 
261 	mutex_lock(&kvm->arch.uvmem_lock);
262 	list_add(&p->list, &kvm->arch.uvmem_pfns);
263 	mutex_unlock(&kvm->arch.uvmem_lock);
264 
265 	return 0;
266 }
267 
268 /*
269  * All device PFNs are already released by the time we come here.
270  */
271 void kvmppc_uvmem_slot_free(struct kvm *kvm, const struct kvm_memory_slot *slot)
272 {
273 	struct kvmppc_uvmem_slot *p, *next;
274 
275 	mutex_lock(&kvm->arch.uvmem_lock);
276 	list_for_each_entry_safe(p, next, &kvm->arch.uvmem_pfns, list) {
277 		if (p->base_pfn == slot->base_gfn) {
278 			vfree(p->pfns);
279 			list_del(&p->list);
280 			kfree(p);
281 			break;
282 		}
283 	}
284 	mutex_unlock(&kvm->arch.uvmem_lock);
285 }
286 
287 static void kvmppc_mark_gfn(unsigned long gfn, struct kvm *kvm,
288 			unsigned long flag, unsigned long uvmem_pfn)
289 {
290 	struct kvmppc_uvmem_slot *p;
291 
292 	list_for_each_entry(p, &kvm->arch.uvmem_pfns, list) {
293 		if (gfn >= p->base_pfn && gfn < p->base_pfn + p->nr_pfns) {
294 			unsigned long index = gfn - p->base_pfn;
295 
296 			if (flag == KVMPPC_GFN_UVMEM_PFN)
297 				p->pfns[index] = uvmem_pfn | flag;
298 			else
299 				p->pfns[index] = flag;
300 			return;
301 		}
302 	}
303 }
304 
305 /* mark the GFN as secure-GFN associated with @uvmem pfn device-PFN. */
306 static void kvmppc_gfn_secure_uvmem_pfn(unsigned long gfn,
307 			unsigned long uvmem_pfn, struct kvm *kvm)
308 {
309 	kvmppc_mark_gfn(gfn, kvm, KVMPPC_GFN_UVMEM_PFN, uvmem_pfn);
310 }
311 
312 /* mark the GFN as secure-GFN associated with a memory-PFN. */
313 static void kvmppc_gfn_secure_mem_pfn(unsigned long gfn, struct kvm *kvm)
314 {
315 	kvmppc_mark_gfn(gfn, kvm, KVMPPC_GFN_MEM_PFN, 0);
316 }
317 
318 /* mark the GFN as a shared GFN. */
319 static void kvmppc_gfn_shared(unsigned long gfn, struct kvm *kvm)
320 {
321 	kvmppc_mark_gfn(gfn, kvm, KVMPPC_GFN_SHARED, 0);
322 }
323 
324 /* mark the GFN as a non-existent GFN. */
325 static void kvmppc_gfn_remove(unsigned long gfn, struct kvm *kvm)
326 {
327 	kvmppc_mark_gfn(gfn, kvm, 0, 0);
328 }
329 
330 /* return true, if the GFN is a secure-GFN backed by a secure-PFN */
331 static bool kvmppc_gfn_is_uvmem_pfn(unsigned long gfn, struct kvm *kvm,
332 				    unsigned long *uvmem_pfn)
333 {
334 	struct kvmppc_uvmem_slot *p;
335 
336 	list_for_each_entry(p, &kvm->arch.uvmem_pfns, list) {
337 		if (gfn >= p->base_pfn && gfn < p->base_pfn + p->nr_pfns) {
338 			unsigned long index = gfn - p->base_pfn;
339 
340 			if (p->pfns[index] & KVMPPC_GFN_UVMEM_PFN) {
341 				if (uvmem_pfn)
342 					*uvmem_pfn = p->pfns[index] &
343 						     KVMPPC_GFN_PFN_MASK;
344 				return true;
345 			} else
346 				return false;
347 		}
348 	}
349 	return false;
350 }
351 
352 /*
353  * starting from *gfn search for the next available GFN that is not yet
354  * transitioned to a secure GFN.  return the value of that GFN in *gfn.  If a
355  * GFN is found, return true, else return false
356  *
357  * Must be called with kvm->arch.uvmem_lock  held.
358  */
359 static bool kvmppc_next_nontransitioned_gfn(const struct kvm_memory_slot *memslot,
360 		struct kvm *kvm, unsigned long *gfn)
361 {
362 	struct kvmppc_uvmem_slot *p;
363 	bool ret = false;
364 	unsigned long i;
365 
366 	list_for_each_entry(p, &kvm->arch.uvmem_pfns, list)
367 		if (*gfn >= p->base_pfn && *gfn < p->base_pfn + p->nr_pfns)
368 			break;
369 	if (!p)
370 		return ret;
371 	/*
372 	 * The code below assumes, one to one correspondence between
373 	 * kvmppc_uvmem_slot and memslot.
374 	 */
375 	for (i = *gfn; i < p->base_pfn + p->nr_pfns; i++) {
376 		unsigned long index = i - p->base_pfn;
377 
378 		if (!(p->pfns[index] & KVMPPC_GFN_FLAG_MASK)) {
379 			*gfn = i;
380 			ret = true;
381 			break;
382 		}
383 	}
384 	return ret;
385 }
386 
387 static int kvmppc_memslot_page_merge(struct kvm *kvm,
388 		const struct kvm_memory_slot *memslot, bool merge)
389 {
390 	unsigned long gfn = memslot->base_gfn;
391 	unsigned long end, start = gfn_to_hva(kvm, gfn);
392 	int ret = 0;
393 	struct vm_area_struct *vma;
394 	int merge_flag = (merge) ? MADV_MERGEABLE : MADV_UNMERGEABLE;
395 
396 	if (kvm_is_error_hva(start))
397 		return H_STATE;
398 
399 	end = start + (memslot->npages << PAGE_SHIFT);
400 
401 	mmap_write_lock(kvm->mm);
402 	do {
403 		vma = find_vma_intersection(kvm->mm, start, end);
404 		if (!vma) {
405 			ret = H_STATE;
406 			break;
407 		}
408 		ret = ksm_madvise(vma, vma->vm_start, vma->vm_end,
409 			  merge_flag, &vma->vm_flags);
410 		if (ret) {
411 			ret = H_STATE;
412 			break;
413 		}
414 		start = vma->vm_end;
415 	} while (end > vma->vm_end);
416 
417 	mmap_write_unlock(kvm->mm);
418 	return ret;
419 }
420 
421 static void __kvmppc_uvmem_memslot_delete(struct kvm *kvm,
422 		const struct kvm_memory_slot *memslot)
423 {
424 	uv_unregister_mem_slot(kvm->arch.lpid, memslot->id);
425 	kvmppc_uvmem_slot_free(kvm, memslot);
426 	kvmppc_memslot_page_merge(kvm, memslot, true);
427 }
428 
429 static int __kvmppc_uvmem_memslot_create(struct kvm *kvm,
430 		const struct kvm_memory_slot *memslot)
431 {
432 	int ret = H_PARAMETER;
433 
434 	if (kvmppc_memslot_page_merge(kvm, memslot, false))
435 		return ret;
436 
437 	if (kvmppc_uvmem_slot_init(kvm, memslot))
438 		goto out1;
439 
440 	ret = uv_register_mem_slot(kvm->arch.lpid,
441 				   memslot->base_gfn << PAGE_SHIFT,
442 				   memslot->npages * PAGE_SIZE,
443 				   0, memslot->id);
444 	if (ret < 0) {
445 		ret = H_PARAMETER;
446 		goto out;
447 	}
448 	return 0;
449 out:
450 	kvmppc_uvmem_slot_free(kvm, memslot);
451 out1:
452 	kvmppc_memslot_page_merge(kvm, memslot, true);
453 	return ret;
454 }
455 
456 unsigned long kvmppc_h_svm_init_start(struct kvm *kvm)
457 {
458 	struct kvm_memslots *slots;
459 	struct kvm_memory_slot *memslot, *m;
460 	int ret = H_SUCCESS;
461 	int srcu_idx;
462 
463 	kvm->arch.secure_guest = KVMPPC_SECURE_INIT_START;
464 
465 	if (!kvmppc_uvmem_bitmap)
466 		return H_UNSUPPORTED;
467 
468 	/* Only radix guests can be secure guests */
469 	if (!kvm_is_radix(kvm))
470 		return H_UNSUPPORTED;
471 
472 	/* NAK the transition to secure if not enabled */
473 	if (!kvm->arch.svm_enabled)
474 		return H_AUTHORITY;
475 
476 	srcu_idx = srcu_read_lock(&kvm->srcu);
477 
478 	/* register the memslot */
479 	slots = kvm_memslots(kvm);
480 	kvm_for_each_memslot(memslot, slots) {
481 		ret = __kvmppc_uvmem_memslot_create(kvm, memslot);
482 		if (ret)
483 			break;
484 	}
485 
486 	if (ret) {
487 		slots = kvm_memslots(kvm);
488 		kvm_for_each_memslot(m, slots) {
489 			if (m == memslot)
490 				break;
491 			__kvmppc_uvmem_memslot_delete(kvm, memslot);
492 		}
493 	}
494 
495 	srcu_read_unlock(&kvm->srcu, srcu_idx);
496 	return ret;
497 }
498 
499 /*
500  * Provision a new page on HV side and copy over the contents
501  * from secure memory using UV_PAGE_OUT uvcall.
502  * Caller must held kvm->arch.uvmem_lock.
503  */
504 static int __kvmppc_svm_page_out(struct vm_area_struct *vma,
505 		unsigned long start,
506 		unsigned long end, unsigned long page_shift,
507 		struct kvm *kvm, unsigned long gpa)
508 {
509 	unsigned long src_pfn, dst_pfn = 0;
510 	struct migrate_vma mig;
511 	struct page *dpage, *spage;
512 	struct kvmppc_uvmem_page_pvt *pvt;
513 	unsigned long pfn;
514 	int ret = U_SUCCESS;
515 
516 	memset(&mig, 0, sizeof(mig));
517 	mig.vma = vma;
518 	mig.start = start;
519 	mig.end = end;
520 	mig.src = &src_pfn;
521 	mig.dst = &dst_pfn;
522 	mig.pgmap_owner = &kvmppc_uvmem_pgmap;
523 	mig.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE;
524 
525 	/* The requested page is already paged-out, nothing to do */
526 	if (!kvmppc_gfn_is_uvmem_pfn(gpa >> page_shift, kvm, NULL))
527 		return ret;
528 
529 	ret = migrate_vma_setup(&mig);
530 	if (ret)
531 		return -1;
532 
533 	spage = migrate_pfn_to_page(*mig.src);
534 	if (!spage || !(*mig.src & MIGRATE_PFN_MIGRATE))
535 		goto out_finalize;
536 
537 	if (!is_zone_device_page(spage))
538 		goto out_finalize;
539 
540 	dpage = alloc_page_vma(GFP_HIGHUSER, vma, start);
541 	if (!dpage) {
542 		ret = -1;
543 		goto out_finalize;
544 	}
545 
546 	lock_page(dpage);
547 	pvt = spage->zone_device_data;
548 	pfn = page_to_pfn(dpage);
549 
550 	/*
551 	 * This function is used in two cases:
552 	 * - When HV touches a secure page, for which we do UV_PAGE_OUT
553 	 * - When a secure page is converted to shared page, we *get*
554 	 *   the page to essentially unmap the device page. In this
555 	 *   case we skip page-out.
556 	 */
557 	if (!pvt->skip_page_out)
558 		ret = uv_page_out(kvm->arch.lpid, pfn << page_shift,
559 				  gpa, 0, page_shift);
560 
561 	if (ret == U_SUCCESS)
562 		*mig.dst = migrate_pfn(pfn) | MIGRATE_PFN_LOCKED;
563 	else {
564 		unlock_page(dpage);
565 		__free_page(dpage);
566 		goto out_finalize;
567 	}
568 
569 	migrate_vma_pages(&mig);
570 
571 out_finalize:
572 	migrate_vma_finalize(&mig);
573 	return ret;
574 }
575 
576 static inline int kvmppc_svm_page_out(struct vm_area_struct *vma,
577 				      unsigned long start, unsigned long end,
578 				      unsigned long page_shift,
579 				      struct kvm *kvm, unsigned long gpa)
580 {
581 	int ret;
582 
583 	mutex_lock(&kvm->arch.uvmem_lock);
584 	ret = __kvmppc_svm_page_out(vma, start, end, page_shift, kvm, gpa);
585 	mutex_unlock(&kvm->arch.uvmem_lock);
586 
587 	return ret;
588 }
589 
590 /*
591  * Drop device pages that we maintain for the secure guest
592  *
593  * We first mark the pages to be skipped from UV_PAGE_OUT when there
594  * is HV side fault on these pages. Next we *get* these pages, forcing
595  * fault on them, do fault time migration to replace the device PTEs in
596  * QEMU page table with normal PTEs from newly allocated pages.
597  */
598 void kvmppc_uvmem_drop_pages(const struct kvm_memory_slot *slot,
599 			     struct kvm *kvm, bool skip_page_out)
600 {
601 	int i;
602 	struct kvmppc_uvmem_page_pvt *pvt;
603 	struct page *uvmem_page;
604 	struct vm_area_struct *vma = NULL;
605 	unsigned long uvmem_pfn, gfn;
606 	unsigned long addr;
607 
608 	mmap_read_lock(kvm->mm);
609 
610 	addr = slot->userspace_addr;
611 
612 	gfn = slot->base_gfn;
613 	for (i = slot->npages; i; --i, ++gfn, addr += PAGE_SIZE) {
614 
615 		/* Fetch the VMA if addr is not in the latest fetched one */
616 		if (!vma || addr >= vma->vm_end) {
617 			vma = find_vma_intersection(kvm->mm, addr, addr+1);
618 			if (!vma) {
619 				pr_err("Can't find VMA for gfn:0x%lx\n", gfn);
620 				break;
621 			}
622 		}
623 
624 		mutex_lock(&kvm->arch.uvmem_lock);
625 
626 		if (kvmppc_gfn_is_uvmem_pfn(gfn, kvm, &uvmem_pfn)) {
627 			uvmem_page = pfn_to_page(uvmem_pfn);
628 			pvt = uvmem_page->zone_device_data;
629 			pvt->skip_page_out = skip_page_out;
630 			pvt->remove_gfn = true;
631 
632 			if (__kvmppc_svm_page_out(vma, addr, addr + PAGE_SIZE,
633 						  PAGE_SHIFT, kvm, pvt->gpa))
634 				pr_err("Can't page out gpa:0x%lx addr:0x%lx\n",
635 				       pvt->gpa, addr);
636 		} else {
637 			/* Remove the shared flag if any */
638 			kvmppc_gfn_remove(gfn, kvm);
639 		}
640 
641 		mutex_unlock(&kvm->arch.uvmem_lock);
642 	}
643 
644 	mmap_read_unlock(kvm->mm);
645 }
646 
647 unsigned long kvmppc_h_svm_init_abort(struct kvm *kvm)
648 {
649 	int srcu_idx;
650 	struct kvm_memory_slot *memslot;
651 
652 	/*
653 	 * Expect to be called only after INIT_START and before INIT_DONE.
654 	 * If INIT_DONE was completed, use normal VM termination sequence.
655 	 */
656 	if (!(kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START))
657 		return H_UNSUPPORTED;
658 
659 	if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
660 		return H_STATE;
661 
662 	srcu_idx = srcu_read_lock(&kvm->srcu);
663 
664 	kvm_for_each_memslot(memslot, kvm_memslots(kvm))
665 		kvmppc_uvmem_drop_pages(memslot, kvm, false);
666 
667 	srcu_read_unlock(&kvm->srcu, srcu_idx);
668 
669 	kvm->arch.secure_guest = 0;
670 	uv_svm_terminate(kvm->arch.lpid);
671 
672 	return H_PARAMETER;
673 }
674 
675 /*
676  * Get a free device PFN from the pool
677  *
678  * Called when a normal page is moved to secure memory (UV_PAGE_IN). Device
679  * PFN will be used to keep track of the secure page on HV side.
680  *
681  * Called with kvm->arch.uvmem_lock held
682  */
683 static struct page *kvmppc_uvmem_get_page(unsigned long gpa, struct kvm *kvm)
684 {
685 	struct page *dpage = NULL;
686 	unsigned long bit, uvmem_pfn;
687 	struct kvmppc_uvmem_page_pvt *pvt;
688 	unsigned long pfn_last, pfn_first;
689 
690 	pfn_first = kvmppc_uvmem_pgmap.res.start >> PAGE_SHIFT;
691 	pfn_last = pfn_first +
692 		   (resource_size(&kvmppc_uvmem_pgmap.res) >> PAGE_SHIFT);
693 
694 	spin_lock(&kvmppc_uvmem_bitmap_lock);
695 	bit = find_first_zero_bit(kvmppc_uvmem_bitmap,
696 				  pfn_last - pfn_first);
697 	if (bit >= (pfn_last - pfn_first))
698 		goto out;
699 	bitmap_set(kvmppc_uvmem_bitmap, bit, 1);
700 	spin_unlock(&kvmppc_uvmem_bitmap_lock);
701 
702 	pvt = kzalloc(sizeof(*pvt), GFP_KERNEL);
703 	if (!pvt)
704 		goto out_clear;
705 
706 	uvmem_pfn = bit + pfn_first;
707 	kvmppc_gfn_secure_uvmem_pfn(gpa >> PAGE_SHIFT, uvmem_pfn, kvm);
708 
709 	pvt->gpa = gpa;
710 	pvt->kvm = kvm;
711 
712 	dpage = pfn_to_page(uvmem_pfn);
713 	dpage->zone_device_data = pvt;
714 	get_page(dpage);
715 	lock_page(dpage);
716 	return dpage;
717 out_clear:
718 	spin_lock(&kvmppc_uvmem_bitmap_lock);
719 	bitmap_clear(kvmppc_uvmem_bitmap, bit, 1);
720 out:
721 	spin_unlock(&kvmppc_uvmem_bitmap_lock);
722 	return NULL;
723 }
724 
725 /*
726  * Alloc a PFN from private device memory pool. If @pagein is true,
727  * copy page from normal memory to secure memory using UV_PAGE_IN uvcall.
728  */
729 static int kvmppc_svm_page_in(struct vm_area_struct *vma,
730 		unsigned long start,
731 		unsigned long end, unsigned long gpa, struct kvm *kvm,
732 		unsigned long page_shift,
733 		bool pagein)
734 {
735 	unsigned long src_pfn, dst_pfn = 0;
736 	struct migrate_vma mig;
737 	struct page *spage;
738 	unsigned long pfn;
739 	struct page *dpage;
740 	int ret = 0;
741 
742 	memset(&mig, 0, sizeof(mig));
743 	mig.vma = vma;
744 	mig.start = start;
745 	mig.end = end;
746 	mig.src = &src_pfn;
747 	mig.dst = &dst_pfn;
748 	mig.flags = MIGRATE_VMA_SELECT_SYSTEM;
749 
750 	ret = migrate_vma_setup(&mig);
751 	if (ret)
752 		return ret;
753 
754 	if (!(*mig.src & MIGRATE_PFN_MIGRATE)) {
755 		ret = -1;
756 		goto out_finalize;
757 	}
758 
759 	dpage = kvmppc_uvmem_get_page(gpa, kvm);
760 	if (!dpage) {
761 		ret = -1;
762 		goto out_finalize;
763 	}
764 
765 	if (pagein) {
766 		pfn = *mig.src >> MIGRATE_PFN_SHIFT;
767 		spage = migrate_pfn_to_page(*mig.src);
768 		if (spage) {
769 			ret = uv_page_in(kvm->arch.lpid, pfn << page_shift,
770 					gpa, 0, page_shift);
771 			if (ret)
772 				goto out_finalize;
773 		}
774 	}
775 
776 	*mig.dst = migrate_pfn(page_to_pfn(dpage)) | MIGRATE_PFN_LOCKED;
777 	migrate_vma_pages(&mig);
778 out_finalize:
779 	migrate_vma_finalize(&mig);
780 	return ret;
781 }
782 
783 static int kvmppc_uv_migrate_mem_slot(struct kvm *kvm,
784 		const struct kvm_memory_slot *memslot)
785 {
786 	unsigned long gfn = memslot->base_gfn;
787 	struct vm_area_struct *vma;
788 	unsigned long start, end;
789 	int ret = 0;
790 
791 	mmap_read_lock(kvm->mm);
792 	mutex_lock(&kvm->arch.uvmem_lock);
793 	while (kvmppc_next_nontransitioned_gfn(memslot, kvm, &gfn)) {
794 		ret = H_STATE;
795 		start = gfn_to_hva(kvm, gfn);
796 		if (kvm_is_error_hva(start))
797 			break;
798 
799 		end = start + (1UL << PAGE_SHIFT);
800 		vma = find_vma_intersection(kvm->mm, start, end);
801 		if (!vma || vma->vm_start > start || vma->vm_end < end)
802 			break;
803 
804 		ret = kvmppc_svm_page_in(vma, start, end,
805 				(gfn << PAGE_SHIFT), kvm, PAGE_SHIFT, false);
806 		if (ret) {
807 			ret = H_STATE;
808 			break;
809 		}
810 
811 		/* relinquish the cpu if needed */
812 		cond_resched();
813 	}
814 	mutex_unlock(&kvm->arch.uvmem_lock);
815 	mmap_read_unlock(kvm->mm);
816 	return ret;
817 }
818 
819 unsigned long kvmppc_h_svm_init_done(struct kvm *kvm)
820 {
821 	struct kvm_memslots *slots;
822 	struct kvm_memory_slot *memslot;
823 	int srcu_idx;
824 	long ret = H_SUCCESS;
825 
826 	if (!(kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START))
827 		return H_UNSUPPORTED;
828 
829 	/* migrate any unmoved normal pfn to device pfns*/
830 	srcu_idx = srcu_read_lock(&kvm->srcu);
831 	slots = kvm_memslots(kvm);
832 	kvm_for_each_memslot(memslot, slots) {
833 		ret = kvmppc_uv_migrate_mem_slot(kvm, memslot);
834 		if (ret) {
835 			/*
836 			 * The pages will remain transitioned.
837 			 * Its the callers responsibility to
838 			 * terminate the VM, which will undo
839 			 * all state of the VM. Till then
840 			 * this VM is in a erroneous state.
841 			 * Its KVMPPC_SECURE_INIT_DONE will
842 			 * remain unset.
843 			 */
844 			ret = H_STATE;
845 			goto out;
846 		}
847 	}
848 
849 	kvm->arch.secure_guest |= KVMPPC_SECURE_INIT_DONE;
850 	pr_info("LPID %d went secure\n", kvm->arch.lpid);
851 
852 out:
853 	srcu_read_unlock(&kvm->srcu, srcu_idx);
854 	return ret;
855 }
856 
857 /*
858  * Shares the page with HV, thus making it a normal page.
859  *
860  * - If the page is already secure, then provision a new page and share
861  * - If the page is a normal page, share the existing page
862  *
863  * In the former case, uses dev_pagemap_ops.migrate_to_ram handler
864  * to unmap the device page from QEMU's page tables.
865  */
866 static unsigned long kvmppc_share_page(struct kvm *kvm, unsigned long gpa,
867 		unsigned long page_shift)
868 {
869 
870 	int ret = H_PARAMETER;
871 	struct page *uvmem_page;
872 	struct kvmppc_uvmem_page_pvt *pvt;
873 	unsigned long pfn;
874 	unsigned long gfn = gpa >> page_shift;
875 	int srcu_idx;
876 	unsigned long uvmem_pfn;
877 
878 	srcu_idx = srcu_read_lock(&kvm->srcu);
879 	mutex_lock(&kvm->arch.uvmem_lock);
880 	if (kvmppc_gfn_is_uvmem_pfn(gfn, kvm, &uvmem_pfn)) {
881 		uvmem_page = pfn_to_page(uvmem_pfn);
882 		pvt = uvmem_page->zone_device_data;
883 		pvt->skip_page_out = true;
884 		/*
885 		 * do not drop the GFN. It is a valid GFN
886 		 * that is transitioned to a shared GFN.
887 		 */
888 		pvt->remove_gfn = false;
889 	}
890 
891 retry:
892 	mutex_unlock(&kvm->arch.uvmem_lock);
893 	pfn = gfn_to_pfn(kvm, gfn);
894 	if (is_error_noslot_pfn(pfn))
895 		goto out;
896 
897 	mutex_lock(&kvm->arch.uvmem_lock);
898 	if (kvmppc_gfn_is_uvmem_pfn(gfn, kvm, &uvmem_pfn)) {
899 		uvmem_page = pfn_to_page(uvmem_pfn);
900 		pvt = uvmem_page->zone_device_data;
901 		pvt->skip_page_out = true;
902 		pvt->remove_gfn = false; /* it continues to be a valid GFN */
903 		kvm_release_pfn_clean(pfn);
904 		goto retry;
905 	}
906 
907 	if (!uv_page_in(kvm->arch.lpid, pfn << page_shift, gpa, 0,
908 				page_shift)) {
909 		kvmppc_gfn_shared(gfn, kvm);
910 		ret = H_SUCCESS;
911 	}
912 	kvm_release_pfn_clean(pfn);
913 	mutex_unlock(&kvm->arch.uvmem_lock);
914 out:
915 	srcu_read_unlock(&kvm->srcu, srcu_idx);
916 	return ret;
917 }
918 
919 /*
920  * H_SVM_PAGE_IN: Move page from normal memory to secure memory.
921  *
922  * H_PAGE_IN_SHARED flag makes the page shared which means that the same
923  * memory in is visible from both UV and HV.
924  */
925 unsigned long kvmppc_h_svm_page_in(struct kvm *kvm, unsigned long gpa,
926 		unsigned long flags,
927 		unsigned long page_shift)
928 {
929 	unsigned long start, end;
930 	struct vm_area_struct *vma;
931 	int srcu_idx;
932 	unsigned long gfn = gpa >> page_shift;
933 	int ret;
934 
935 	if (!(kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START))
936 		return H_UNSUPPORTED;
937 
938 	if (page_shift != PAGE_SHIFT)
939 		return H_P3;
940 
941 	if (flags & ~H_PAGE_IN_SHARED)
942 		return H_P2;
943 
944 	if (flags & H_PAGE_IN_SHARED)
945 		return kvmppc_share_page(kvm, gpa, page_shift);
946 
947 	ret = H_PARAMETER;
948 	srcu_idx = srcu_read_lock(&kvm->srcu);
949 	mmap_read_lock(kvm->mm);
950 
951 	start = gfn_to_hva(kvm, gfn);
952 	if (kvm_is_error_hva(start))
953 		goto out;
954 
955 	mutex_lock(&kvm->arch.uvmem_lock);
956 	/* Fail the page-in request of an already paged-in page */
957 	if (kvmppc_gfn_is_uvmem_pfn(gfn, kvm, NULL))
958 		goto out_unlock;
959 
960 	end = start + (1UL << page_shift);
961 	vma = find_vma_intersection(kvm->mm, start, end);
962 	if (!vma || vma->vm_start > start || vma->vm_end < end)
963 		goto out_unlock;
964 
965 	if (kvmppc_svm_page_in(vma, start, end, gpa, kvm, page_shift,
966 				true))
967 		goto out_unlock;
968 
969 	ret = H_SUCCESS;
970 
971 out_unlock:
972 	mutex_unlock(&kvm->arch.uvmem_lock);
973 out:
974 	mmap_read_unlock(kvm->mm);
975 	srcu_read_unlock(&kvm->srcu, srcu_idx);
976 	return ret;
977 }
978 
979 
980 /*
981  * Fault handler callback that gets called when HV touches any page that
982  * has been moved to secure memory, we ask UV to give back the page by
983  * issuing UV_PAGE_OUT uvcall.
984  *
985  * This eventually results in dropping of device PFN and the newly
986  * provisioned page/PFN gets populated in QEMU page tables.
987  */
988 static vm_fault_t kvmppc_uvmem_migrate_to_ram(struct vm_fault *vmf)
989 {
990 	struct kvmppc_uvmem_page_pvt *pvt = vmf->page->zone_device_data;
991 
992 	if (kvmppc_svm_page_out(vmf->vma, vmf->address,
993 				vmf->address + PAGE_SIZE, PAGE_SHIFT,
994 				pvt->kvm, pvt->gpa))
995 		return VM_FAULT_SIGBUS;
996 	else
997 		return 0;
998 }
999 
1000 /*
1001  * Release the device PFN back to the pool
1002  *
1003  * Gets called when secure GFN tranistions from a secure-PFN
1004  * to a normal PFN during H_SVM_PAGE_OUT.
1005  * Gets called with kvm->arch.uvmem_lock held.
1006  */
1007 static void kvmppc_uvmem_page_free(struct page *page)
1008 {
1009 	unsigned long pfn = page_to_pfn(page) -
1010 			(kvmppc_uvmem_pgmap.res.start >> PAGE_SHIFT);
1011 	struct kvmppc_uvmem_page_pvt *pvt;
1012 
1013 	spin_lock(&kvmppc_uvmem_bitmap_lock);
1014 	bitmap_clear(kvmppc_uvmem_bitmap, pfn, 1);
1015 	spin_unlock(&kvmppc_uvmem_bitmap_lock);
1016 
1017 	pvt = page->zone_device_data;
1018 	page->zone_device_data = NULL;
1019 	if (pvt->remove_gfn)
1020 		kvmppc_gfn_remove(pvt->gpa >> PAGE_SHIFT, pvt->kvm);
1021 	else
1022 		kvmppc_gfn_secure_mem_pfn(pvt->gpa >> PAGE_SHIFT, pvt->kvm);
1023 	kfree(pvt);
1024 }
1025 
1026 static const struct dev_pagemap_ops kvmppc_uvmem_ops = {
1027 	.page_free = kvmppc_uvmem_page_free,
1028 	.migrate_to_ram	= kvmppc_uvmem_migrate_to_ram,
1029 };
1030 
1031 /*
1032  * H_SVM_PAGE_OUT: Move page from secure memory to normal memory.
1033  */
1034 unsigned long
1035 kvmppc_h_svm_page_out(struct kvm *kvm, unsigned long gpa,
1036 		      unsigned long flags, unsigned long page_shift)
1037 {
1038 	unsigned long gfn = gpa >> page_shift;
1039 	unsigned long start, end;
1040 	struct vm_area_struct *vma;
1041 	int srcu_idx;
1042 	int ret;
1043 
1044 	if (!(kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START))
1045 		return H_UNSUPPORTED;
1046 
1047 	if (page_shift != PAGE_SHIFT)
1048 		return H_P3;
1049 
1050 	if (flags)
1051 		return H_P2;
1052 
1053 	ret = H_PARAMETER;
1054 	srcu_idx = srcu_read_lock(&kvm->srcu);
1055 	mmap_read_lock(kvm->mm);
1056 	start = gfn_to_hva(kvm, gfn);
1057 	if (kvm_is_error_hva(start))
1058 		goto out;
1059 
1060 	end = start + (1UL << page_shift);
1061 	vma = find_vma_intersection(kvm->mm, start, end);
1062 	if (!vma || vma->vm_start > start || vma->vm_end < end)
1063 		goto out;
1064 
1065 	if (!kvmppc_svm_page_out(vma, start, end, page_shift, kvm, gpa))
1066 		ret = H_SUCCESS;
1067 out:
1068 	mmap_read_unlock(kvm->mm);
1069 	srcu_read_unlock(&kvm->srcu, srcu_idx);
1070 	return ret;
1071 }
1072 
1073 int kvmppc_send_page_to_uv(struct kvm *kvm, unsigned long gfn)
1074 {
1075 	unsigned long pfn;
1076 	int ret = U_SUCCESS;
1077 
1078 	pfn = gfn_to_pfn(kvm, gfn);
1079 	if (is_error_noslot_pfn(pfn))
1080 		return -EFAULT;
1081 
1082 	mutex_lock(&kvm->arch.uvmem_lock);
1083 	if (kvmppc_gfn_is_uvmem_pfn(gfn, kvm, NULL))
1084 		goto out;
1085 
1086 	ret = uv_page_in(kvm->arch.lpid, pfn << PAGE_SHIFT, gfn << PAGE_SHIFT,
1087 			 0, PAGE_SHIFT);
1088 out:
1089 	kvm_release_pfn_clean(pfn);
1090 	mutex_unlock(&kvm->arch.uvmem_lock);
1091 	return (ret == U_SUCCESS) ? RESUME_GUEST : -EFAULT;
1092 }
1093 
1094 int kvmppc_uvmem_memslot_create(struct kvm *kvm, const struct kvm_memory_slot *new)
1095 {
1096 	int ret = __kvmppc_uvmem_memslot_create(kvm, new);
1097 
1098 	if (!ret)
1099 		ret = kvmppc_uv_migrate_mem_slot(kvm, new);
1100 
1101 	return ret;
1102 }
1103 
1104 void kvmppc_uvmem_memslot_delete(struct kvm *kvm, const struct kvm_memory_slot *old)
1105 {
1106 	__kvmppc_uvmem_memslot_delete(kvm, old);
1107 }
1108 
1109 static u64 kvmppc_get_secmem_size(void)
1110 {
1111 	struct device_node *np;
1112 	int i, len;
1113 	const __be32 *prop;
1114 	u64 size = 0;
1115 
1116 	/*
1117 	 * First try the new ibm,secure-memory nodes which supersede the
1118 	 * secure-memory-ranges property.
1119 	 * If we found some, no need to read the deprecated ones.
1120 	 */
1121 	for_each_compatible_node(np, NULL, "ibm,secure-memory") {
1122 		prop = of_get_property(np, "reg", &len);
1123 		if (!prop)
1124 			continue;
1125 		size += of_read_number(prop + 2, 2);
1126 	}
1127 	if (size)
1128 		return size;
1129 
1130 	np = of_find_compatible_node(NULL, NULL, "ibm,uv-firmware");
1131 	if (!np)
1132 		goto out;
1133 
1134 	prop = of_get_property(np, "secure-memory-ranges", &len);
1135 	if (!prop)
1136 		goto out_put;
1137 
1138 	for (i = 0; i < len / (sizeof(*prop) * 4); i++)
1139 		size += of_read_number(prop + (i * 4) + 2, 2);
1140 
1141 out_put:
1142 	of_node_put(np);
1143 out:
1144 	return size;
1145 }
1146 
1147 int kvmppc_uvmem_init(void)
1148 {
1149 	int ret = 0;
1150 	unsigned long size;
1151 	struct resource *res;
1152 	void *addr;
1153 	unsigned long pfn_last, pfn_first;
1154 
1155 	size = kvmppc_get_secmem_size();
1156 	if (!size) {
1157 		/*
1158 		 * Don't fail the initialization of kvm-hv module if
1159 		 * the platform doesn't export ibm,uv-firmware node.
1160 		 * Let normal guests run on such PEF-disabled platform.
1161 		 */
1162 		pr_info("KVMPPC-UVMEM: No support for secure guests\n");
1163 		goto out;
1164 	}
1165 
1166 	res = request_free_mem_region(&iomem_resource, size, "kvmppc_uvmem");
1167 	if (IS_ERR(res)) {
1168 		ret = PTR_ERR(res);
1169 		goto out;
1170 	}
1171 
1172 	kvmppc_uvmem_pgmap.type = MEMORY_DEVICE_PRIVATE;
1173 	kvmppc_uvmem_pgmap.res = *res;
1174 	kvmppc_uvmem_pgmap.ops = &kvmppc_uvmem_ops;
1175 	/* just one global instance: */
1176 	kvmppc_uvmem_pgmap.owner = &kvmppc_uvmem_pgmap;
1177 	addr = memremap_pages(&kvmppc_uvmem_pgmap, NUMA_NO_NODE);
1178 	if (IS_ERR(addr)) {
1179 		ret = PTR_ERR(addr);
1180 		goto out_free_region;
1181 	}
1182 
1183 	pfn_first = res->start >> PAGE_SHIFT;
1184 	pfn_last = pfn_first + (resource_size(res) >> PAGE_SHIFT);
1185 	kvmppc_uvmem_bitmap = kcalloc(BITS_TO_LONGS(pfn_last - pfn_first),
1186 				      sizeof(unsigned long), GFP_KERNEL);
1187 	if (!kvmppc_uvmem_bitmap) {
1188 		ret = -ENOMEM;
1189 		goto out_unmap;
1190 	}
1191 
1192 	pr_info("KVMPPC-UVMEM: Secure Memory size 0x%lx\n", size);
1193 	return ret;
1194 out_unmap:
1195 	memunmap_pages(&kvmppc_uvmem_pgmap);
1196 out_free_region:
1197 	release_mem_region(res->start, size);
1198 out:
1199 	return ret;
1200 }
1201 
1202 void kvmppc_uvmem_free(void)
1203 {
1204 	if (!kvmppc_uvmem_bitmap)
1205 		return;
1206 
1207 	memunmap_pages(&kvmppc_uvmem_pgmap);
1208 	release_mem_region(kvmppc_uvmem_pgmap.res.start,
1209 			   resource_size(&kvmppc_uvmem_pgmap.res));
1210 	kfree(kvmppc_uvmem_bitmap);
1211 }
1212