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 <linux/of.h>
94 #include <linux/memremap.h>
95 #include <asm/ultravisor.h>
96 #include <asm/mman.h>
97 #include <asm/kvm_ppc.h>
98 #include <asm/kvm_book3s_uvmem.h>
99 
100 static struct dev_pagemap kvmppc_uvmem_pgmap;
101 static unsigned long *kvmppc_uvmem_bitmap;
102 static DEFINE_SPINLOCK(kvmppc_uvmem_bitmap_lock);
103 
104 /*
105  * States of a GFN
106  * ---------------
107  * The GFN can be in one of the following states.
108  *
109  * (a) Secure - The GFN is secure. The GFN is associated with
110  *	a Secure VM, the contents of the GFN is not accessible
111  *	to the Hypervisor.  This GFN can be backed by a secure-PFN,
112  *	or can be backed by a normal-PFN with contents encrypted.
113  *	The former is true when the GFN is paged-in into the
114  *	ultravisor. The latter is true when the GFN is paged-out
115  *	of the ultravisor.
116  *
117  * (b) Shared - The GFN is shared. The GFN is associated with a
118  *	a secure VM. The contents of the GFN is accessible to
119  *	Hypervisor. This GFN is backed by a normal-PFN and its
120  *	content is un-encrypted.
121  *
122  * (c) Normal - The GFN is a normal. The GFN is associated with
123  *	a normal VM. The contents of the GFN is accessible to
124  *	the Hypervisor. Its content is never encrypted.
125  *
126  * States of a VM.
127  * ---------------
128  *
129  * Normal VM:  A VM whose contents are always accessible to
130  *	the hypervisor.  All its GFNs are normal-GFNs.
131  *
132  * Secure VM: A VM whose contents are not accessible to the
133  *	hypervisor without the VM's consent.  Its GFNs are
134  *	either Shared-GFN or Secure-GFNs.
135  *
136  * Transient VM: A Normal VM that is transitioning to secure VM.
137  *	The transition starts on successful return of
138  *	H_SVM_INIT_START, and ends on successful return
139  *	of H_SVM_INIT_DONE. This transient VM, can have GFNs
140  *	in any of the three states; i.e Secure-GFN, Shared-GFN,
141  *	and Normal-GFN.	The VM never executes in this state
142  *	in supervisor-mode.
143  *
144  * Memory slot State.
145  * -----------------------------
146  *	The state of a memory slot mirrors the state of the
147  *	VM the memory slot is associated with.
148  *
149  * VM State transition.
150  * --------------------
151  *
152  *  A VM always starts in Normal Mode.
153  *
154  *  H_SVM_INIT_START moves the VM into transient state. During this
155  *  time the Ultravisor may request some of its GFNs to be shared or
156  *  secured. So its GFNs can be in one of the three GFN states.
157  *
158  *  H_SVM_INIT_DONE moves the VM entirely from transient state to
159  *  secure-state. At this point any left-over normal-GFNs are
160  *  transitioned to Secure-GFN.
161  *
162  *  H_SVM_INIT_ABORT moves the transient VM back to normal VM.
163  *  All its GFNs are moved to Normal-GFNs.
164  *
165  *  UV_TERMINATE transitions the secure-VM back to normal-VM. All
166  *  the secure-GFN and shared-GFNs are tranistioned to normal-GFN
167  *  Note: The contents of the normal-GFN is undefined at this point.
168  *
169  * GFN state implementation:
170  * -------------------------
171  *
172  * Secure GFN is associated with a secure-PFN; also called uvmem_pfn,
173  * when the GFN is paged-in. Its pfn[] has KVMPPC_GFN_UVMEM_PFN flag
174  * set, and contains the value of the secure-PFN.
175  * It is associated with a normal-PFN; also called mem_pfn, when
176  * the GFN is pagedout. Its pfn[] has KVMPPC_GFN_MEM_PFN flag set.
177  * The value of the normal-PFN is not tracked.
178  *
179  * Shared GFN is associated with a normal-PFN. Its pfn[] has
180  * KVMPPC_UVMEM_SHARED_PFN flag set. The value of the normal-PFN
181  * is not tracked.
182  *
183  * Normal GFN is associated with normal-PFN. Its pfn[] has
184  * no flag set. The value of the normal-PFN is not tracked.
185  *
186  * Life cycle of a GFN
187  * --------------------
188  *
189  * --------------------------------------------------------------
190  * |        |     Share  |  Unshare | SVM       |H_SVM_INIT_DONE|
191  * |        |operation   |operation | abort/    |               |
192  * |        |            |          | terminate |               |
193  * -------------------------------------------------------------
194  * |        |            |          |           |               |
195  * | Secure |     Shared | Secure   |Normal     |Secure         |
196  * |        |            |          |           |               |
197  * | Shared |     Shared | Secure   |Normal     |Shared         |
198  * |        |            |          |           |               |
199  * | Normal |     Shared | Secure   |Normal     |Secure         |
200  * --------------------------------------------------------------
201  *
202  * Life cycle of a VM
203  * --------------------
204  *
205  * --------------------------------------------------------------------
206  * |         |  start    |  H_SVM_  |H_SVM_   |H_SVM_     |UV_SVM_    |
207  * |         |  VM       |INIT_START|INIT_DONE|INIT_ABORT |TERMINATE  |
208  * |         |           |          |         |           |           |
209  * --------- ----------------------------------------------------------
210  * |         |           |          |         |           |           |
211  * | Normal  | Normal    | Transient|Error    |Error      |Normal     |
212  * |         |           |          |         |           |           |
213  * | Secure  |   Error   | Error    |Error    |Error      |Normal     |
214  * |         |           |          |         |           |           |
215  * |Transient|   N/A     | Error    |Secure   |Normal     |Normal     |
216  * --------------------------------------------------------------------
217  */
218 
219 #define KVMPPC_GFN_UVMEM_PFN	(1UL << 63)
220 #define KVMPPC_GFN_MEM_PFN	(1UL << 62)
221 #define KVMPPC_GFN_SHARED	(1UL << 61)
222 #define KVMPPC_GFN_SECURE	(KVMPPC_GFN_UVMEM_PFN | KVMPPC_GFN_MEM_PFN)
223 #define KVMPPC_GFN_FLAG_MASK	(KVMPPC_GFN_SECURE | KVMPPC_GFN_SHARED)
224 #define KVMPPC_GFN_PFN_MASK	(~KVMPPC_GFN_FLAG_MASK)
225 
226 struct kvmppc_uvmem_slot {
227 	struct list_head list;
228 	unsigned long nr_pfns;
229 	unsigned long base_pfn;
230 	unsigned long *pfns;
231 };
232 struct kvmppc_uvmem_page_pvt {
233 	struct kvm *kvm;
234 	unsigned long gpa;
235 	bool skip_page_out;
236 	bool remove_gfn;
237 };
238 
239 bool kvmppc_uvmem_available(void)
240 {
241 	/*
242 	 * If kvmppc_uvmem_bitmap != NULL, then there is an ultravisor
243 	 * and our data structures have been initialized successfully.
244 	 */
245 	return !!kvmppc_uvmem_bitmap;
246 }
247 
248 int kvmppc_uvmem_slot_init(struct kvm *kvm, const struct kvm_memory_slot *slot)
249 {
250 	struct kvmppc_uvmem_slot *p;
251 
252 	p = kzalloc(sizeof(*p), GFP_KERNEL);
253 	if (!p)
254 		return -ENOMEM;
255 	p->pfns = vcalloc(slot->npages, sizeof(*p->pfns));
256 	if (!p->pfns) {
257 		kfree(p);
258 		return -ENOMEM;
259 	}
260 	p->nr_pfns = slot->npages;
261 	p->base_pfn = slot->base_gfn;
262 
263 	mutex_lock(&kvm->arch.uvmem_lock);
264 	list_add(&p->list, &kvm->arch.uvmem_pfns);
265 	mutex_unlock(&kvm->arch.uvmem_lock);
266 
267 	return 0;
268 }
269 
270 /*
271  * All device PFNs are already released by the time we come here.
272  */
273 void kvmppc_uvmem_slot_free(struct kvm *kvm, const struct kvm_memory_slot *slot)
274 {
275 	struct kvmppc_uvmem_slot *p, *next;
276 
277 	mutex_lock(&kvm->arch.uvmem_lock);
278 	list_for_each_entry_safe(p, next, &kvm->arch.uvmem_pfns, list) {
279 		if (p->base_pfn == slot->base_gfn) {
280 			vfree(p->pfns);
281 			list_del(&p->list);
282 			kfree(p);
283 			break;
284 		}
285 	}
286 	mutex_unlock(&kvm->arch.uvmem_lock);
287 }
288 
289 static void kvmppc_mark_gfn(unsigned long gfn, struct kvm *kvm,
290 			unsigned long flag, unsigned long uvmem_pfn)
291 {
292 	struct kvmppc_uvmem_slot *p;
293 
294 	list_for_each_entry(p, &kvm->arch.uvmem_pfns, list) {
295 		if (gfn >= p->base_pfn && gfn < p->base_pfn + p->nr_pfns) {
296 			unsigned long index = gfn - p->base_pfn;
297 
298 			if (flag == KVMPPC_GFN_UVMEM_PFN)
299 				p->pfns[index] = uvmem_pfn | flag;
300 			else
301 				p->pfns[index] = flag;
302 			return;
303 		}
304 	}
305 }
306 
307 /* mark the GFN as secure-GFN associated with @uvmem pfn device-PFN. */
308 static void kvmppc_gfn_secure_uvmem_pfn(unsigned long gfn,
309 			unsigned long uvmem_pfn, struct kvm *kvm)
310 {
311 	kvmppc_mark_gfn(gfn, kvm, KVMPPC_GFN_UVMEM_PFN, uvmem_pfn);
312 }
313 
314 /* mark the GFN as secure-GFN associated with a memory-PFN. */
315 static void kvmppc_gfn_secure_mem_pfn(unsigned long gfn, struct kvm *kvm)
316 {
317 	kvmppc_mark_gfn(gfn, kvm, KVMPPC_GFN_MEM_PFN, 0);
318 }
319 
320 /* mark the GFN as a shared GFN. */
321 static void kvmppc_gfn_shared(unsigned long gfn, struct kvm *kvm)
322 {
323 	kvmppc_mark_gfn(gfn, kvm, KVMPPC_GFN_SHARED, 0);
324 }
325 
326 /* mark the GFN as a non-existent GFN. */
327 static void kvmppc_gfn_remove(unsigned long gfn, struct kvm *kvm)
328 {
329 	kvmppc_mark_gfn(gfn, kvm, 0, 0);
330 }
331 
332 /* return true, if the GFN is a secure-GFN backed by a secure-PFN */
333 static bool kvmppc_gfn_is_uvmem_pfn(unsigned long gfn, struct kvm *kvm,
334 				    unsigned long *uvmem_pfn)
335 {
336 	struct kvmppc_uvmem_slot *p;
337 
338 	list_for_each_entry(p, &kvm->arch.uvmem_pfns, list) {
339 		if (gfn >= p->base_pfn && gfn < p->base_pfn + p->nr_pfns) {
340 			unsigned long index = gfn - p->base_pfn;
341 
342 			if (p->pfns[index] & KVMPPC_GFN_UVMEM_PFN) {
343 				if (uvmem_pfn)
344 					*uvmem_pfn = p->pfns[index] &
345 						     KVMPPC_GFN_PFN_MASK;
346 				return true;
347 			} else
348 				return false;
349 		}
350 	}
351 	return false;
352 }
353 
354 /*
355  * starting from *gfn search for the next available GFN that is not yet
356  * transitioned to a secure GFN.  return the value of that GFN in *gfn.  If a
357  * GFN is found, return true, else return false
358  *
359  * Must be called with kvm->arch.uvmem_lock  held.
360  */
361 static bool kvmppc_next_nontransitioned_gfn(const struct kvm_memory_slot *memslot,
362 		struct kvm *kvm, unsigned long *gfn)
363 {
364 	struct kvmppc_uvmem_slot *p = NULL, *iter;
365 	bool ret = false;
366 	unsigned long i;
367 
368 	list_for_each_entry(iter, &kvm->arch.uvmem_pfns, list)
369 		if (*gfn >= iter->base_pfn && *gfn < iter->base_pfn + iter->nr_pfns) {
370 			p = iter;
371 			break;
372 		}
373 	if (!p)
374 		return ret;
375 	/*
376 	 * The code below assumes, one to one correspondence between
377 	 * kvmppc_uvmem_slot and memslot.
378 	 */
379 	for (i = *gfn; i < p->base_pfn + p->nr_pfns; i++) {
380 		unsigned long index = i - p->base_pfn;
381 
382 		if (!(p->pfns[index] & KVMPPC_GFN_FLAG_MASK)) {
383 			*gfn = i;
384 			ret = true;
385 			break;
386 		}
387 	}
388 	return ret;
389 }
390 
391 static int kvmppc_memslot_page_merge(struct kvm *kvm,
392 		const struct kvm_memory_slot *memslot, bool merge)
393 {
394 	unsigned long gfn = memslot->base_gfn;
395 	unsigned long end, start = gfn_to_hva(kvm, gfn);
396 	int ret = 0;
397 	struct vm_area_struct *vma;
398 	int merge_flag = (merge) ? MADV_MERGEABLE : MADV_UNMERGEABLE;
399 
400 	if (kvm_is_error_hva(start))
401 		return H_STATE;
402 
403 	end = start + (memslot->npages << PAGE_SHIFT);
404 
405 	mmap_write_lock(kvm->mm);
406 	do {
407 		vma = find_vma_intersection(kvm->mm, start, end);
408 		if (!vma) {
409 			ret = H_STATE;
410 			break;
411 		}
412 		ret = ksm_madvise(vma, vma->vm_start, vma->vm_end,
413 			  merge_flag, &vma->vm_flags);
414 		if (ret) {
415 			ret = H_STATE;
416 			break;
417 		}
418 		start = vma->vm_end;
419 	} while (end > vma->vm_end);
420 
421 	mmap_write_unlock(kvm->mm);
422 	return ret;
423 }
424 
425 static void __kvmppc_uvmem_memslot_delete(struct kvm *kvm,
426 		const struct kvm_memory_slot *memslot)
427 {
428 	uv_unregister_mem_slot(kvm->arch.lpid, memslot->id);
429 	kvmppc_uvmem_slot_free(kvm, memslot);
430 	kvmppc_memslot_page_merge(kvm, memslot, true);
431 }
432 
433 static int __kvmppc_uvmem_memslot_create(struct kvm *kvm,
434 		const struct kvm_memory_slot *memslot)
435 {
436 	int ret = H_PARAMETER;
437 
438 	if (kvmppc_memslot_page_merge(kvm, memslot, false))
439 		return ret;
440 
441 	if (kvmppc_uvmem_slot_init(kvm, memslot))
442 		goto out1;
443 
444 	ret = uv_register_mem_slot(kvm->arch.lpid,
445 				   memslot->base_gfn << PAGE_SHIFT,
446 				   memslot->npages * PAGE_SIZE,
447 				   0, memslot->id);
448 	if (ret < 0) {
449 		ret = H_PARAMETER;
450 		goto out;
451 	}
452 	return 0;
453 out:
454 	kvmppc_uvmem_slot_free(kvm, memslot);
455 out1:
456 	kvmppc_memslot_page_merge(kvm, memslot, true);
457 	return ret;
458 }
459 
460 unsigned long kvmppc_h_svm_init_start(struct kvm *kvm)
461 {
462 	struct kvm_memslots *slots;
463 	struct kvm_memory_slot *memslot, *m;
464 	int ret = H_SUCCESS;
465 	int srcu_idx, bkt;
466 
467 	kvm->arch.secure_guest = KVMPPC_SECURE_INIT_START;
468 
469 	if (!kvmppc_uvmem_bitmap)
470 		return H_UNSUPPORTED;
471 
472 	/* Only radix guests can be secure guests */
473 	if (!kvm_is_radix(kvm))
474 		return H_UNSUPPORTED;
475 
476 	/* NAK the transition to secure if not enabled */
477 	if (!kvm->arch.svm_enabled)
478 		return H_AUTHORITY;
479 
480 	srcu_idx = srcu_read_lock(&kvm->srcu);
481 
482 	/* register the memslot */
483 	slots = kvm_memslots(kvm);
484 	kvm_for_each_memslot(memslot, bkt, slots) {
485 		ret = __kvmppc_uvmem_memslot_create(kvm, memslot);
486 		if (ret)
487 			break;
488 	}
489 
490 	if (ret) {
491 		slots = kvm_memslots(kvm);
492 		kvm_for_each_memslot(m, bkt, slots) {
493 			if (m == memslot)
494 				break;
495 			__kvmppc_uvmem_memslot_delete(kvm, memslot);
496 		}
497 	}
498 
499 	srcu_read_unlock(&kvm->srcu, srcu_idx);
500 	return ret;
501 }
502 
503 /*
504  * Provision a new page on HV side and copy over the contents
505  * from secure memory using UV_PAGE_OUT uvcall.
506  * Caller must held kvm->arch.uvmem_lock.
507  */
508 static int __kvmppc_svm_page_out(struct vm_area_struct *vma,
509 		unsigned long start,
510 		unsigned long end, unsigned long page_shift,
511 		struct kvm *kvm, unsigned long gpa)
512 {
513 	unsigned long src_pfn, dst_pfn = 0;
514 	struct migrate_vma mig;
515 	struct page *dpage, *spage;
516 	struct kvmppc_uvmem_page_pvt *pvt;
517 	unsigned long pfn;
518 	int ret = U_SUCCESS;
519 
520 	memset(&mig, 0, sizeof(mig));
521 	mig.vma = vma;
522 	mig.start = start;
523 	mig.end = end;
524 	mig.src = &src_pfn;
525 	mig.dst = &dst_pfn;
526 	mig.pgmap_owner = &kvmppc_uvmem_pgmap;
527 	mig.flags = MIGRATE_VMA_SELECT_DEVICE_PRIVATE;
528 
529 	/* The requested page is already paged-out, nothing to do */
530 	if (!kvmppc_gfn_is_uvmem_pfn(gpa >> page_shift, kvm, NULL))
531 		return ret;
532 
533 	ret = migrate_vma_setup(&mig);
534 	if (ret)
535 		return -1;
536 
537 	spage = migrate_pfn_to_page(*mig.src);
538 	if (!spage || !(*mig.src & MIGRATE_PFN_MIGRATE))
539 		goto out_finalize;
540 
541 	if (!is_zone_device_page(spage))
542 		goto out_finalize;
543 
544 	dpage = alloc_page_vma(GFP_HIGHUSER, vma, start);
545 	if (!dpage) {
546 		ret = -1;
547 		goto out_finalize;
548 	}
549 
550 	lock_page(dpage);
551 	pvt = spage->zone_device_data;
552 	pfn = page_to_pfn(dpage);
553 
554 	/*
555 	 * This function is used in two cases:
556 	 * - When HV touches a secure page, for which we do UV_PAGE_OUT
557 	 * - When a secure page is converted to shared page, we *get*
558 	 *   the page to essentially unmap the device page. In this
559 	 *   case we skip page-out.
560 	 */
561 	if (!pvt->skip_page_out)
562 		ret = uv_page_out(kvm->arch.lpid, pfn << page_shift,
563 				  gpa, 0, page_shift);
564 
565 	if (ret == U_SUCCESS)
566 		*mig.dst = migrate_pfn(pfn);
567 	else {
568 		unlock_page(dpage);
569 		__free_page(dpage);
570 		goto out_finalize;
571 	}
572 
573 	migrate_vma_pages(&mig);
574 
575 out_finalize:
576 	migrate_vma_finalize(&mig);
577 	return ret;
578 }
579 
580 static inline int kvmppc_svm_page_out(struct vm_area_struct *vma,
581 				      unsigned long start, unsigned long end,
582 				      unsigned long page_shift,
583 				      struct kvm *kvm, unsigned long gpa)
584 {
585 	int ret;
586 
587 	mutex_lock(&kvm->arch.uvmem_lock);
588 	ret = __kvmppc_svm_page_out(vma, start, end, page_shift, kvm, gpa);
589 	mutex_unlock(&kvm->arch.uvmem_lock);
590 
591 	return ret;
592 }
593 
594 /*
595  * Drop device pages that we maintain for the secure guest
596  *
597  * We first mark the pages to be skipped from UV_PAGE_OUT when there
598  * is HV side fault on these pages. Next we *get* these pages, forcing
599  * fault on them, do fault time migration to replace the device PTEs in
600  * QEMU page table with normal PTEs from newly allocated pages.
601  */
602 void kvmppc_uvmem_drop_pages(const struct kvm_memory_slot *slot,
603 			     struct kvm *kvm, bool skip_page_out)
604 {
605 	int i;
606 	struct kvmppc_uvmem_page_pvt *pvt;
607 	struct page *uvmem_page;
608 	struct vm_area_struct *vma = NULL;
609 	unsigned long uvmem_pfn, gfn;
610 	unsigned long addr;
611 
612 	mmap_read_lock(kvm->mm);
613 
614 	addr = slot->userspace_addr;
615 
616 	gfn = slot->base_gfn;
617 	for (i = slot->npages; i; --i, ++gfn, addr += PAGE_SIZE) {
618 
619 		/* Fetch the VMA if addr is not in the latest fetched one */
620 		if (!vma || addr >= vma->vm_end) {
621 			vma = vma_lookup(kvm->mm, addr);
622 			if (!vma) {
623 				pr_err("Can't find VMA for gfn:0x%lx\n", gfn);
624 				break;
625 			}
626 		}
627 
628 		mutex_lock(&kvm->arch.uvmem_lock);
629 
630 		if (kvmppc_gfn_is_uvmem_pfn(gfn, kvm, &uvmem_pfn)) {
631 			uvmem_page = pfn_to_page(uvmem_pfn);
632 			pvt = uvmem_page->zone_device_data;
633 			pvt->skip_page_out = skip_page_out;
634 			pvt->remove_gfn = true;
635 
636 			if (__kvmppc_svm_page_out(vma, addr, addr + PAGE_SIZE,
637 						  PAGE_SHIFT, kvm, pvt->gpa))
638 				pr_err("Can't page out gpa:0x%lx addr:0x%lx\n",
639 				       pvt->gpa, addr);
640 		} else {
641 			/* Remove the shared flag if any */
642 			kvmppc_gfn_remove(gfn, kvm);
643 		}
644 
645 		mutex_unlock(&kvm->arch.uvmem_lock);
646 	}
647 
648 	mmap_read_unlock(kvm->mm);
649 }
650 
651 unsigned long kvmppc_h_svm_init_abort(struct kvm *kvm)
652 {
653 	int srcu_idx, bkt;
654 	struct kvm_memory_slot *memslot;
655 
656 	/*
657 	 * Expect to be called only after INIT_START and before INIT_DONE.
658 	 * If INIT_DONE was completed, use normal VM termination sequence.
659 	 */
660 	if (!(kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START))
661 		return H_UNSUPPORTED;
662 
663 	if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
664 		return H_STATE;
665 
666 	srcu_idx = srcu_read_lock(&kvm->srcu);
667 
668 	kvm_for_each_memslot(memslot, bkt, kvm_memslots(kvm))
669 		kvmppc_uvmem_drop_pages(memslot, kvm, false);
670 
671 	srcu_read_unlock(&kvm->srcu, srcu_idx);
672 
673 	kvm->arch.secure_guest = 0;
674 	uv_svm_terminate(kvm->arch.lpid);
675 
676 	return H_PARAMETER;
677 }
678 
679 /*
680  * Get a free device PFN from the pool
681  *
682  * Called when a normal page is moved to secure memory (UV_PAGE_IN). Device
683  * PFN will be used to keep track of the secure page on HV side.
684  *
685  * Called with kvm->arch.uvmem_lock held
686  */
687 static struct page *kvmppc_uvmem_get_page(unsigned long gpa, struct kvm *kvm)
688 {
689 	struct page *dpage = NULL;
690 	unsigned long bit, uvmem_pfn;
691 	struct kvmppc_uvmem_page_pvt *pvt;
692 	unsigned long pfn_last, pfn_first;
693 
694 	pfn_first = kvmppc_uvmem_pgmap.range.start >> PAGE_SHIFT;
695 	pfn_last = pfn_first +
696 		   (range_len(&kvmppc_uvmem_pgmap.range) >> PAGE_SHIFT);
697 
698 	spin_lock(&kvmppc_uvmem_bitmap_lock);
699 	bit = find_first_zero_bit(kvmppc_uvmem_bitmap,
700 				  pfn_last - pfn_first);
701 	if (bit >= (pfn_last - pfn_first))
702 		goto out;
703 	bitmap_set(kvmppc_uvmem_bitmap, bit, 1);
704 	spin_unlock(&kvmppc_uvmem_bitmap_lock);
705 
706 	pvt = kzalloc(sizeof(*pvt), GFP_KERNEL);
707 	if (!pvt)
708 		goto out_clear;
709 
710 	uvmem_pfn = bit + pfn_first;
711 	kvmppc_gfn_secure_uvmem_pfn(gpa >> PAGE_SHIFT, uvmem_pfn, kvm);
712 
713 	pvt->gpa = gpa;
714 	pvt->kvm = kvm;
715 
716 	dpage = pfn_to_page(uvmem_pfn);
717 	dpage->zone_device_data = pvt;
718 	lock_page(dpage);
719 	return dpage;
720 out_clear:
721 	spin_lock(&kvmppc_uvmem_bitmap_lock);
722 	bitmap_clear(kvmppc_uvmem_bitmap, bit, 1);
723 out:
724 	spin_unlock(&kvmppc_uvmem_bitmap_lock);
725 	return NULL;
726 }
727 
728 /*
729  * Alloc a PFN from private device memory pool. If @pagein is true,
730  * copy page from normal memory to secure memory using UV_PAGE_IN uvcall.
731  */
732 static int kvmppc_svm_page_in(struct vm_area_struct *vma,
733 		unsigned long start,
734 		unsigned long end, unsigned long gpa, struct kvm *kvm,
735 		unsigned long page_shift,
736 		bool pagein)
737 {
738 	unsigned long src_pfn, dst_pfn = 0;
739 	struct migrate_vma mig;
740 	struct page *spage;
741 	unsigned long pfn;
742 	struct page *dpage;
743 	int ret = 0;
744 
745 	memset(&mig, 0, sizeof(mig));
746 	mig.vma = vma;
747 	mig.start = start;
748 	mig.end = end;
749 	mig.src = &src_pfn;
750 	mig.dst = &dst_pfn;
751 	mig.flags = MIGRATE_VMA_SELECT_SYSTEM;
752 
753 	ret = migrate_vma_setup(&mig);
754 	if (ret)
755 		return ret;
756 
757 	if (!(*mig.src & MIGRATE_PFN_MIGRATE)) {
758 		ret = -1;
759 		goto out_finalize;
760 	}
761 
762 	dpage = kvmppc_uvmem_get_page(gpa, kvm);
763 	if (!dpage) {
764 		ret = -1;
765 		goto out_finalize;
766 	}
767 
768 	if (pagein) {
769 		pfn = *mig.src >> MIGRATE_PFN_SHIFT;
770 		spage = migrate_pfn_to_page(*mig.src);
771 		if (spage) {
772 			ret = uv_page_in(kvm->arch.lpid, pfn << page_shift,
773 					gpa, 0, page_shift);
774 			if (ret)
775 				goto out_finalize;
776 		}
777 	}
778 
779 	*mig.dst = migrate_pfn(page_to_pfn(dpage));
780 	migrate_vma_pages(&mig);
781 out_finalize:
782 	migrate_vma_finalize(&mig);
783 	return ret;
784 }
785 
786 static int kvmppc_uv_migrate_mem_slot(struct kvm *kvm,
787 		const struct kvm_memory_slot *memslot)
788 {
789 	unsigned long gfn = memslot->base_gfn;
790 	struct vm_area_struct *vma;
791 	unsigned long start, end;
792 	int ret = 0;
793 
794 	mmap_read_lock(kvm->mm);
795 	mutex_lock(&kvm->arch.uvmem_lock);
796 	while (kvmppc_next_nontransitioned_gfn(memslot, kvm, &gfn)) {
797 		ret = H_STATE;
798 		start = gfn_to_hva(kvm, gfn);
799 		if (kvm_is_error_hva(start))
800 			break;
801 
802 		end = start + (1UL << PAGE_SHIFT);
803 		vma = find_vma_intersection(kvm->mm, start, end);
804 		if (!vma || vma->vm_start > start || vma->vm_end < end)
805 			break;
806 
807 		ret = kvmppc_svm_page_in(vma, start, end,
808 				(gfn << PAGE_SHIFT), kvm, PAGE_SHIFT, false);
809 		if (ret) {
810 			ret = H_STATE;
811 			break;
812 		}
813 
814 		/* relinquish the cpu if needed */
815 		cond_resched();
816 	}
817 	mutex_unlock(&kvm->arch.uvmem_lock);
818 	mmap_read_unlock(kvm->mm);
819 	return ret;
820 }
821 
822 unsigned long kvmppc_h_svm_init_done(struct kvm *kvm)
823 {
824 	struct kvm_memslots *slots;
825 	struct kvm_memory_slot *memslot;
826 	int srcu_idx, bkt;
827 	long ret = H_SUCCESS;
828 
829 	if (!(kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START))
830 		return H_UNSUPPORTED;
831 
832 	/* migrate any unmoved normal pfn to device pfns*/
833 	srcu_idx = srcu_read_lock(&kvm->srcu);
834 	slots = kvm_memslots(kvm);
835 	kvm_for_each_memslot(memslot, bkt, slots) {
836 		ret = kvmppc_uv_migrate_mem_slot(kvm, memslot);
837 		if (ret) {
838 			/*
839 			 * The pages will remain transitioned.
840 			 * Its the callers responsibility to
841 			 * terminate the VM, which will undo
842 			 * all state of the VM. Till then
843 			 * this VM is in a erroneous state.
844 			 * Its KVMPPC_SECURE_INIT_DONE will
845 			 * remain unset.
846 			 */
847 			ret = H_STATE;
848 			goto out;
849 		}
850 	}
851 
852 	kvm->arch.secure_guest |= KVMPPC_SECURE_INIT_DONE;
853 	pr_info("LPID %d went secure\n", kvm->arch.lpid);
854 
855 out:
856 	srcu_read_unlock(&kvm->srcu, srcu_idx);
857 	return ret;
858 }
859 
860 /*
861  * Shares the page with HV, thus making it a normal page.
862  *
863  * - If the page is already secure, then provision a new page and share
864  * - If the page is a normal page, share the existing page
865  *
866  * In the former case, uses dev_pagemap_ops.migrate_to_ram handler
867  * to unmap the device page from QEMU's page tables.
868  */
869 static unsigned long kvmppc_share_page(struct kvm *kvm, unsigned long gpa,
870 		unsigned long page_shift)
871 {
872 
873 	int ret = H_PARAMETER;
874 	struct page *uvmem_page;
875 	struct kvmppc_uvmem_page_pvt *pvt;
876 	unsigned long pfn;
877 	unsigned long gfn = gpa >> page_shift;
878 	int srcu_idx;
879 	unsigned long uvmem_pfn;
880 
881 	srcu_idx = srcu_read_lock(&kvm->srcu);
882 	mutex_lock(&kvm->arch.uvmem_lock);
883 	if (kvmppc_gfn_is_uvmem_pfn(gfn, kvm, &uvmem_pfn)) {
884 		uvmem_page = pfn_to_page(uvmem_pfn);
885 		pvt = uvmem_page->zone_device_data;
886 		pvt->skip_page_out = true;
887 		/*
888 		 * do not drop the GFN. It is a valid GFN
889 		 * that is transitioned to a shared GFN.
890 		 */
891 		pvt->remove_gfn = false;
892 	}
893 
894 retry:
895 	mutex_unlock(&kvm->arch.uvmem_lock);
896 	pfn = gfn_to_pfn(kvm, gfn);
897 	if (is_error_noslot_pfn(pfn))
898 		goto out;
899 
900 	mutex_lock(&kvm->arch.uvmem_lock);
901 	if (kvmppc_gfn_is_uvmem_pfn(gfn, kvm, &uvmem_pfn)) {
902 		uvmem_page = pfn_to_page(uvmem_pfn);
903 		pvt = uvmem_page->zone_device_data;
904 		pvt->skip_page_out = true;
905 		pvt->remove_gfn = false; /* it continues to be a valid GFN */
906 		kvm_release_pfn_clean(pfn);
907 		goto retry;
908 	}
909 
910 	if (!uv_page_in(kvm->arch.lpid, pfn << page_shift, gpa, 0,
911 				page_shift)) {
912 		kvmppc_gfn_shared(gfn, kvm);
913 		ret = H_SUCCESS;
914 	}
915 	kvm_release_pfn_clean(pfn);
916 	mutex_unlock(&kvm->arch.uvmem_lock);
917 out:
918 	srcu_read_unlock(&kvm->srcu, srcu_idx);
919 	return ret;
920 }
921 
922 /*
923  * H_SVM_PAGE_IN: Move page from normal memory to secure memory.
924  *
925  * H_PAGE_IN_SHARED flag makes the page shared which means that the same
926  * memory in is visible from both UV and HV.
927  */
928 unsigned long kvmppc_h_svm_page_in(struct kvm *kvm, unsigned long gpa,
929 		unsigned long flags,
930 		unsigned long page_shift)
931 {
932 	unsigned long start, end;
933 	struct vm_area_struct *vma;
934 	int srcu_idx;
935 	unsigned long gfn = gpa >> page_shift;
936 	int ret;
937 
938 	if (!(kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START))
939 		return H_UNSUPPORTED;
940 
941 	if (page_shift != PAGE_SHIFT)
942 		return H_P3;
943 
944 	if (flags & ~H_PAGE_IN_SHARED)
945 		return H_P2;
946 
947 	if (flags & H_PAGE_IN_SHARED)
948 		return kvmppc_share_page(kvm, gpa, page_shift);
949 
950 	ret = H_PARAMETER;
951 	srcu_idx = srcu_read_lock(&kvm->srcu);
952 	mmap_read_lock(kvm->mm);
953 
954 	start = gfn_to_hva(kvm, gfn);
955 	if (kvm_is_error_hva(start))
956 		goto out;
957 
958 	mutex_lock(&kvm->arch.uvmem_lock);
959 	/* Fail the page-in request of an already paged-in page */
960 	if (kvmppc_gfn_is_uvmem_pfn(gfn, kvm, NULL))
961 		goto out_unlock;
962 
963 	end = start + (1UL << page_shift);
964 	vma = find_vma_intersection(kvm->mm, start, end);
965 	if (!vma || vma->vm_start > start || vma->vm_end < end)
966 		goto out_unlock;
967 
968 	if (kvmppc_svm_page_in(vma, start, end, gpa, kvm, page_shift,
969 				true))
970 		goto out_unlock;
971 
972 	ret = H_SUCCESS;
973 
974 out_unlock:
975 	mutex_unlock(&kvm->arch.uvmem_lock);
976 out:
977 	mmap_read_unlock(kvm->mm);
978 	srcu_read_unlock(&kvm->srcu, srcu_idx);
979 	return ret;
980 }
981 
982 
983 /*
984  * Fault handler callback that gets called when HV touches any page that
985  * has been moved to secure memory, we ask UV to give back the page by
986  * issuing UV_PAGE_OUT uvcall.
987  *
988  * This eventually results in dropping of device PFN and the newly
989  * provisioned page/PFN gets populated in QEMU page tables.
990  */
991 static vm_fault_t kvmppc_uvmem_migrate_to_ram(struct vm_fault *vmf)
992 {
993 	struct kvmppc_uvmem_page_pvt *pvt = vmf->page->zone_device_data;
994 
995 	if (kvmppc_svm_page_out(vmf->vma, vmf->address,
996 				vmf->address + PAGE_SIZE, PAGE_SHIFT,
997 				pvt->kvm, pvt->gpa))
998 		return VM_FAULT_SIGBUS;
999 	else
1000 		return 0;
1001 }
1002 
1003 /*
1004  * Release the device PFN back to the pool
1005  *
1006  * Gets called when secure GFN tranistions from a secure-PFN
1007  * to a normal PFN during H_SVM_PAGE_OUT.
1008  * Gets called with kvm->arch.uvmem_lock held.
1009  */
1010 static void kvmppc_uvmem_page_free(struct page *page)
1011 {
1012 	unsigned long pfn = page_to_pfn(page) -
1013 			(kvmppc_uvmem_pgmap.range.start >> PAGE_SHIFT);
1014 	struct kvmppc_uvmem_page_pvt *pvt;
1015 
1016 	spin_lock(&kvmppc_uvmem_bitmap_lock);
1017 	bitmap_clear(kvmppc_uvmem_bitmap, pfn, 1);
1018 	spin_unlock(&kvmppc_uvmem_bitmap_lock);
1019 
1020 	pvt = page->zone_device_data;
1021 	page->zone_device_data = NULL;
1022 	if (pvt->remove_gfn)
1023 		kvmppc_gfn_remove(pvt->gpa >> PAGE_SHIFT, pvt->kvm);
1024 	else
1025 		kvmppc_gfn_secure_mem_pfn(pvt->gpa >> PAGE_SHIFT, pvt->kvm);
1026 	kfree(pvt);
1027 }
1028 
1029 static const struct dev_pagemap_ops kvmppc_uvmem_ops = {
1030 	.page_free = kvmppc_uvmem_page_free,
1031 	.migrate_to_ram	= kvmppc_uvmem_migrate_to_ram,
1032 };
1033 
1034 /*
1035  * H_SVM_PAGE_OUT: Move page from secure memory to normal memory.
1036  */
1037 unsigned long
1038 kvmppc_h_svm_page_out(struct kvm *kvm, unsigned long gpa,
1039 		      unsigned long flags, unsigned long page_shift)
1040 {
1041 	unsigned long gfn = gpa >> page_shift;
1042 	unsigned long start, end;
1043 	struct vm_area_struct *vma;
1044 	int srcu_idx;
1045 	int ret;
1046 
1047 	if (!(kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START))
1048 		return H_UNSUPPORTED;
1049 
1050 	if (page_shift != PAGE_SHIFT)
1051 		return H_P3;
1052 
1053 	if (flags)
1054 		return H_P2;
1055 
1056 	ret = H_PARAMETER;
1057 	srcu_idx = srcu_read_lock(&kvm->srcu);
1058 	mmap_read_lock(kvm->mm);
1059 	start = gfn_to_hva(kvm, gfn);
1060 	if (kvm_is_error_hva(start))
1061 		goto out;
1062 
1063 	end = start + (1UL << page_shift);
1064 	vma = find_vma_intersection(kvm->mm, start, end);
1065 	if (!vma || vma->vm_start > start || vma->vm_end < end)
1066 		goto out;
1067 
1068 	if (!kvmppc_svm_page_out(vma, start, end, page_shift, kvm, gpa))
1069 		ret = H_SUCCESS;
1070 out:
1071 	mmap_read_unlock(kvm->mm);
1072 	srcu_read_unlock(&kvm->srcu, srcu_idx);
1073 	return ret;
1074 }
1075 
1076 int kvmppc_send_page_to_uv(struct kvm *kvm, unsigned long gfn)
1077 {
1078 	unsigned long pfn;
1079 	int ret = U_SUCCESS;
1080 
1081 	pfn = gfn_to_pfn(kvm, gfn);
1082 	if (is_error_noslot_pfn(pfn))
1083 		return -EFAULT;
1084 
1085 	mutex_lock(&kvm->arch.uvmem_lock);
1086 	if (kvmppc_gfn_is_uvmem_pfn(gfn, kvm, NULL))
1087 		goto out;
1088 
1089 	ret = uv_page_in(kvm->arch.lpid, pfn << PAGE_SHIFT, gfn << PAGE_SHIFT,
1090 			 0, PAGE_SHIFT);
1091 out:
1092 	kvm_release_pfn_clean(pfn);
1093 	mutex_unlock(&kvm->arch.uvmem_lock);
1094 	return (ret == U_SUCCESS) ? RESUME_GUEST : -EFAULT;
1095 }
1096 
1097 int kvmppc_uvmem_memslot_create(struct kvm *kvm, const struct kvm_memory_slot *new)
1098 {
1099 	int ret = __kvmppc_uvmem_memslot_create(kvm, new);
1100 
1101 	if (!ret)
1102 		ret = kvmppc_uv_migrate_mem_slot(kvm, new);
1103 
1104 	return ret;
1105 }
1106 
1107 void kvmppc_uvmem_memslot_delete(struct kvm *kvm, const struct kvm_memory_slot *old)
1108 {
1109 	__kvmppc_uvmem_memslot_delete(kvm, old);
1110 }
1111 
1112 static u64 kvmppc_get_secmem_size(void)
1113 {
1114 	struct device_node *np;
1115 	int i, len;
1116 	const __be32 *prop;
1117 	u64 size = 0;
1118 
1119 	/*
1120 	 * First try the new ibm,secure-memory nodes which supersede the
1121 	 * secure-memory-ranges property.
1122 	 * If we found some, no need to read the deprecated ones.
1123 	 */
1124 	for_each_compatible_node(np, NULL, "ibm,secure-memory") {
1125 		prop = of_get_property(np, "reg", &len);
1126 		if (!prop)
1127 			continue;
1128 		size += of_read_number(prop + 2, 2);
1129 	}
1130 	if (size)
1131 		return size;
1132 
1133 	np = of_find_compatible_node(NULL, NULL, "ibm,uv-firmware");
1134 	if (!np)
1135 		goto out;
1136 
1137 	prop = of_get_property(np, "secure-memory-ranges", &len);
1138 	if (!prop)
1139 		goto out_put;
1140 
1141 	for (i = 0; i < len / (sizeof(*prop) * 4); i++)
1142 		size += of_read_number(prop + (i * 4) + 2, 2);
1143 
1144 out_put:
1145 	of_node_put(np);
1146 out:
1147 	return size;
1148 }
1149 
1150 int kvmppc_uvmem_init(void)
1151 {
1152 	int ret = 0;
1153 	unsigned long size;
1154 	struct resource *res;
1155 	void *addr;
1156 	unsigned long pfn_last, pfn_first;
1157 
1158 	size = kvmppc_get_secmem_size();
1159 	if (!size) {
1160 		/*
1161 		 * Don't fail the initialization of kvm-hv module if
1162 		 * the platform doesn't export ibm,uv-firmware node.
1163 		 * Let normal guests run on such PEF-disabled platform.
1164 		 */
1165 		pr_info("KVMPPC-UVMEM: No support for secure guests\n");
1166 		goto out;
1167 	}
1168 
1169 	res = request_free_mem_region(&iomem_resource, size, "kvmppc_uvmem");
1170 	if (IS_ERR(res)) {
1171 		ret = PTR_ERR(res);
1172 		goto out;
1173 	}
1174 
1175 	kvmppc_uvmem_pgmap.type = MEMORY_DEVICE_PRIVATE;
1176 	kvmppc_uvmem_pgmap.range.start = res->start;
1177 	kvmppc_uvmem_pgmap.range.end = res->end;
1178 	kvmppc_uvmem_pgmap.nr_range = 1;
1179 	kvmppc_uvmem_pgmap.ops = &kvmppc_uvmem_ops;
1180 	/* just one global instance: */
1181 	kvmppc_uvmem_pgmap.owner = &kvmppc_uvmem_pgmap;
1182 	addr = memremap_pages(&kvmppc_uvmem_pgmap, NUMA_NO_NODE);
1183 	if (IS_ERR(addr)) {
1184 		ret = PTR_ERR(addr);
1185 		goto out_free_region;
1186 	}
1187 
1188 	pfn_first = res->start >> PAGE_SHIFT;
1189 	pfn_last = pfn_first + (resource_size(res) >> PAGE_SHIFT);
1190 	kvmppc_uvmem_bitmap = kcalloc(BITS_TO_LONGS(pfn_last - pfn_first),
1191 				      sizeof(unsigned long), GFP_KERNEL);
1192 	if (!kvmppc_uvmem_bitmap) {
1193 		ret = -ENOMEM;
1194 		goto out_unmap;
1195 	}
1196 
1197 	pr_info("KVMPPC-UVMEM: Secure Memory size 0x%lx\n", size);
1198 	return ret;
1199 out_unmap:
1200 	memunmap_pages(&kvmppc_uvmem_pgmap);
1201 out_free_region:
1202 	release_mem_region(res->start, size);
1203 out:
1204 	return ret;
1205 }
1206 
1207 void kvmppc_uvmem_free(void)
1208 {
1209 	if (!kvmppc_uvmem_bitmap)
1210 		return;
1211 
1212 	memunmap_pages(&kvmppc_uvmem_pgmap);
1213 	release_mem_region(kvmppc_uvmem_pgmap.range.start,
1214 			   range_len(&kvmppc_uvmem_pgmap.range));
1215 	kfree(kvmppc_uvmem_bitmap);
1216 }
1217