Commit | Line | Data |
---|---|---|
fe5db27d BG |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | ||
02c00b3a BG |
3 | #include "mmu.h" |
4 | #include "mmu_internal.h" | |
bb18842e | 5 | #include "mmutrace.h" |
2f2fad08 | 6 | #include "tdp_iter.h" |
fe5db27d | 7 | #include "tdp_mmu.h" |
02c00b3a | 8 | #include "spte.h" |
fe5db27d | 9 | |
9a77daac | 10 | #include <asm/cmpxchg.h> |
33dd3574 BG |
11 | #include <trace/events/kvm.h> |
12 | ||
71ba3f31 | 13 | static bool __read_mostly tdp_mmu_enabled = true; |
95fb5b02 | 14 | module_param_named(tdp_mmu, tdp_mmu_enabled, bool, 0644); |
fe5db27d BG |
15 | |
16 | /* Initializes the TDP MMU for the VM, if enabled. */ | |
d501f747 | 17 | bool kvm_mmu_init_tdp_mmu(struct kvm *kvm) |
fe5db27d | 18 | { |
897218ff | 19 | if (!tdp_enabled || !READ_ONCE(tdp_mmu_enabled)) |
d501f747 | 20 | return false; |
fe5db27d BG |
21 | |
22 | /* This should not be changed for the lifetime of the VM. */ | |
23 | kvm->arch.tdp_mmu_enabled = true; | |
02c00b3a BG |
24 | |
25 | INIT_LIST_HEAD(&kvm->arch.tdp_mmu_roots); | |
9a77daac | 26 | spin_lock_init(&kvm->arch.tdp_mmu_pages_lock); |
89c0fd49 | 27 | INIT_LIST_HEAD(&kvm->arch.tdp_mmu_pages); |
d501f747 BG |
28 | |
29 | return true; | |
fe5db27d BG |
30 | } |
31 | ||
226b8c8f SC |
32 | /* Arbitrarily returns true so that this may be used in if statements. */ |
33 | static __always_inline bool kvm_lockdep_assert_mmu_lock_held(struct kvm *kvm, | |
6103bc07 BG |
34 | bool shared) |
35 | { | |
36 | if (shared) | |
37 | lockdep_assert_held_read(&kvm->mmu_lock); | |
38 | else | |
39 | lockdep_assert_held_write(&kvm->mmu_lock); | |
226b8c8f SC |
40 | |
41 | return true; | |
6103bc07 BG |
42 | } |
43 | ||
fe5db27d BG |
44 | void kvm_mmu_uninit_tdp_mmu(struct kvm *kvm) |
45 | { | |
46 | if (!kvm->arch.tdp_mmu_enabled) | |
47 | return; | |
02c00b3a | 48 | |
524a1e4e | 49 | WARN_ON(!list_empty(&kvm->arch.tdp_mmu_pages)); |
02c00b3a | 50 | WARN_ON(!list_empty(&kvm->arch.tdp_mmu_roots)); |
7cca2d0b BG |
51 | |
52 | /* | |
53 | * Ensure that all the outstanding RCU callbacks to free shadow pages | |
54 | * can run before the VM is torn down. | |
55 | */ | |
56 | rcu_barrier(); | |
02c00b3a BG |
57 | } |
58 | ||
2bdb3d84 | 59 | static bool zap_gfn_range(struct kvm *kvm, struct kvm_mmu_page *root, |
6103bc07 BG |
60 | gfn_t start, gfn_t end, bool can_yield, bool flush, |
61 | bool shared); | |
2bdb3d84 BG |
62 | |
63 | static void tdp_mmu_free_sp(struct kvm_mmu_page *sp) | |
a889ea54 | 64 | { |
2bdb3d84 BG |
65 | free_page((unsigned long)sp->spt); |
66 | kmem_cache_free(mmu_page_header_cache, sp); | |
a889ea54 BG |
67 | } |
68 | ||
c0e64238 BG |
69 | /* |
70 | * This is called through call_rcu in order to free TDP page table memory | |
71 | * safely with respect to other kernel threads that may be operating on | |
72 | * the memory. | |
73 | * By only accessing TDP MMU page table memory in an RCU read critical | |
74 | * section, and freeing it after a grace period, lockless access to that | |
75 | * memory won't use it after it is freed. | |
76 | */ | |
77 | static void tdp_mmu_free_sp_rcu_callback(struct rcu_head *head) | |
a889ea54 | 78 | { |
c0e64238 BG |
79 | struct kvm_mmu_page *sp = container_of(head, struct kvm_mmu_page, |
80 | rcu_head); | |
a889ea54 | 81 | |
c0e64238 BG |
82 | tdp_mmu_free_sp(sp); |
83 | } | |
a889ea54 | 84 | |
6103bc07 BG |
85 | void kvm_tdp_mmu_put_root(struct kvm *kvm, struct kvm_mmu_page *root, |
86 | bool shared) | |
2bdb3d84 | 87 | { |
6103bc07 | 88 | kvm_lockdep_assert_mmu_lock_held(kvm, shared); |
a889ea54 | 89 | |
11cccf5c | 90 | if (!refcount_dec_and_test(&root->tdp_mmu_root_count)) |
2bdb3d84 BG |
91 | return; |
92 | ||
93 | WARN_ON(!root->tdp_mmu_page); | |
94 | ||
c0e64238 BG |
95 | spin_lock(&kvm->arch.tdp_mmu_pages_lock); |
96 | list_del_rcu(&root->link); | |
97 | spin_unlock(&kvm->arch.tdp_mmu_pages_lock); | |
2bdb3d84 | 98 | |
db01416b SC |
99 | /* |
100 | * A TLB flush is not necessary as KVM performs a local TLB flush when | |
101 | * allocating a new root (see kvm_mmu_load()), and when migrating vCPU | |
102 | * to a different pCPU. Note, the local TLB flush on reuse also | |
103 | * invalidates any paging-structure-cache entries, i.e. TLB entries for | |
104 | * intermediate paging structures, that may be zapped, as such entries | |
105 | * are associated with the ASID on both VMX and SVM. | |
106 | */ | |
107 | (void)zap_gfn_range(kvm, root, 0, -1ull, false, false, shared); | |
2bdb3d84 | 108 | |
c0e64238 | 109 | call_rcu(&root->rcu_head, tdp_mmu_free_sp_rcu_callback); |
a889ea54 BG |
110 | } |
111 | ||
cfc10997 | 112 | /* |
d62007ed SC |
113 | * Returns the next root after @prev_root (or the first root if @prev_root is |
114 | * NULL). A reference to the returned root is acquired, and the reference to | |
115 | * @prev_root is released (the caller obviously must hold a reference to | |
116 | * @prev_root if it's non-NULL). | |
117 | * | |
118 | * If @only_valid is true, invalid roots are skipped. | |
119 | * | |
120 | * Returns NULL if the end of tdp_mmu_roots was reached. | |
cfc10997 BG |
121 | */ |
122 | static struct kvm_mmu_page *tdp_mmu_next_root(struct kvm *kvm, | |
6103bc07 | 123 | struct kvm_mmu_page *prev_root, |
d62007ed | 124 | bool shared, bool only_valid) |
a889ea54 BG |
125 | { |
126 | struct kvm_mmu_page *next_root; | |
127 | ||
c0e64238 BG |
128 | rcu_read_lock(); |
129 | ||
cfc10997 | 130 | if (prev_root) |
c0e64238 BG |
131 | next_root = list_next_or_null_rcu(&kvm->arch.tdp_mmu_roots, |
132 | &prev_root->link, | |
133 | typeof(*prev_root), link); | |
cfc10997 | 134 | else |
c0e64238 BG |
135 | next_root = list_first_or_null_rcu(&kvm->arch.tdp_mmu_roots, |
136 | typeof(*next_root), link); | |
a889ea54 | 137 | |
04dc4e6c | 138 | while (next_root) { |
d62007ed | 139 | if ((!only_valid || !next_root->role.invalid) && |
ad6d6b94 | 140 | kvm_tdp_mmu_get_root(next_root)) |
04dc4e6c SC |
141 | break; |
142 | ||
c0e64238 BG |
143 | next_root = list_next_or_null_rcu(&kvm->arch.tdp_mmu_roots, |
144 | &next_root->link, typeof(*next_root), link); | |
04dc4e6c | 145 | } |
fb101293 | 146 | |
c0e64238 | 147 | rcu_read_unlock(); |
a889ea54 | 148 | |
cfc10997 | 149 | if (prev_root) |
6103bc07 | 150 | kvm_tdp_mmu_put_root(kvm, prev_root, shared); |
a889ea54 | 151 | |
a889ea54 BG |
152 | return next_root; |
153 | } | |
154 | ||
155 | /* | |
156 | * Note: this iterator gets and puts references to the roots it iterates over. | |
157 | * This makes it safe to release the MMU lock and yield within the loop, but | |
158 | * if exiting the loop early, the caller must drop the reference to the most | |
159 | * recent root. (Unless keeping a live reference is desirable.) | |
6103bc07 BG |
160 | * |
161 | * If shared is set, this function is operating under the MMU lock in read | |
162 | * mode. In the unlikely event that this thread must free a root, the lock | |
163 | * will be temporarily dropped and reacquired in write mode. | |
a889ea54 | 164 | */ |
d62007ed SC |
165 | #define __for_each_tdp_mmu_root_yield_safe(_kvm, _root, _as_id, _shared, _only_valid)\ |
166 | for (_root = tdp_mmu_next_root(_kvm, NULL, _shared, _only_valid); \ | |
167 | _root; \ | |
168 | _root = tdp_mmu_next_root(_kvm, _root, _shared, _only_valid)) \ | |
614f6970 PB |
169 | if (kvm_lockdep_assert_mmu_lock_held(_kvm, _shared) && \ |
170 | kvm_mmu_page_as_id(_root) != _as_id) { \ | |
a3f15bda | 171 | } else |
a889ea54 | 172 | |
d62007ed SC |
173 | #define for_each_valid_tdp_mmu_root_yield_safe(_kvm, _root, _as_id, _shared) \ |
174 | __for_each_tdp_mmu_root_yield_safe(_kvm, _root, _as_id, _shared, true) | |
175 | ||
614f6970 PB |
176 | #define for_each_tdp_mmu_root_yield_safe(_kvm, _root, _as_id) \ |
177 | __for_each_tdp_mmu_root_yield_safe(_kvm, _root, _as_id, false, false) | |
d62007ed | 178 | |
226b8c8f SC |
179 | /* |
180 | * Iterate over all TDP MMU roots. Requires that mmu_lock be held for write, | |
181 | * the implication being that any flow that holds mmu_lock for read is | |
182 | * inherently yield-friendly and should use the yield-safe variant above. | |
183 | * Holding mmu_lock for write obviates the need for RCU protection as the list | |
184 | * is guaranteed to be stable. | |
185 | */ | |
186 | #define for_each_tdp_mmu_root(_kvm, _root, _as_id) \ | |
187 | list_for_each_entry(_root, &_kvm->arch.tdp_mmu_roots, link) \ | |
188 | if (kvm_lockdep_assert_mmu_lock_held(_kvm, false) && \ | |
189 | kvm_mmu_page_as_id(_root) != _as_id) { \ | |
a3f15bda | 190 | } else |
02c00b3a | 191 | |
a82070b6 | 192 | static struct kvm_mmu_page *tdp_mmu_alloc_sp(struct kvm_vcpu *vcpu) |
02c00b3a BG |
193 | { |
194 | struct kvm_mmu_page *sp; | |
195 | ||
196 | sp = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_page_header_cache); | |
197 | sp->spt = kvm_mmu_memory_cache_alloc(&vcpu->arch.mmu_shadow_page_cache); | |
a82070b6 DM |
198 | |
199 | return sp; | |
200 | } | |
201 | ||
202 | static void tdp_mmu_init_sp(struct kvm_mmu_page *sp, gfn_t gfn, | |
203 | union kvm_mmu_page_role role) | |
204 | { | |
02c00b3a BG |
205 | set_page_private(virt_to_page(sp->spt), (unsigned long)sp); |
206 | ||
a3aca4de | 207 | sp->role = role; |
02c00b3a BG |
208 | sp->gfn = gfn; |
209 | sp->tdp_mmu_page = true; | |
210 | ||
33dd3574 | 211 | trace_kvm_mmu_get_page(sp, true); |
02c00b3a BG |
212 | } |
213 | ||
a82070b6 DM |
214 | static void tdp_mmu_init_child_sp(struct kvm_mmu_page *child_sp, |
215 | struct tdp_iter *iter) | |
02c00b3a | 216 | { |
a3aca4de | 217 | struct kvm_mmu_page *parent_sp; |
02c00b3a | 218 | union kvm_mmu_page_role role; |
a3aca4de DM |
219 | |
220 | parent_sp = sptep_to_sp(rcu_dereference(iter->sptep)); | |
221 | ||
222 | role = parent_sp->role; | |
223 | role.level--; | |
224 | ||
a82070b6 | 225 | tdp_mmu_init_sp(child_sp, iter->gfn, role); |
a3aca4de DM |
226 | } |
227 | ||
228 | hpa_t kvm_tdp_mmu_get_vcpu_root_hpa(struct kvm_vcpu *vcpu) | |
229 | { | |
230 | union kvm_mmu_page_role role = vcpu->arch.mmu->mmu_role.base; | |
02c00b3a BG |
231 | struct kvm *kvm = vcpu->kvm; |
232 | struct kvm_mmu_page *root; | |
233 | ||
6e6ec584 | 234 | lockdep_assert_held_write(&kvm->mmu_lock); |
02c00b3a | 235 | |
04dc4e6c SC |
236 | /* |
237 | * Check for an existing root before allocating a new one. Note, the | |
238 | * role check prevents consuming an invalid root. | |
239 | */ | |
a3f15bda | 240 | for_each_tdp_mmu_root(kvm, root, kvm_mmu_role_as_id(role)) { |
fb101293 | 241 | if (root->role.word == role.word && |
ad6d6b94 | 242 | kvm_tdp_mmu_get_root(root)) |
6e6ec584 | 243 | goto out; |
02c00b3a BG |
244 | } |
245 | ||
a82070b6 DM |
246 | root = tdp_mmu_alloc_sp(vcpu); |
247 | tdp_mmu_init_sp(root, 0, role); | |
248 | ||
11cccf5c | 249 | refcount_set(&root->tdp_mmu_root_count, 1); |
02c00b3a | 250 | |
c0e64238 BG |
251 | spin_lock(&kvm->arch.tdp_mmu_pages_lock); |
252 | list_add_rcu(&root->link, &kvm->arch.tdp_mmu_roots); | |
253 | spin_unlock(&kvm->arch.tdp_mmu_pages_lock); | |
02c00b3a | 254 | |
6e6ec584 | 255 | out: |
02c00b3a | 256 | return __pa(root->spt); |
fe5db27d | 257 | } |
2f2fad08 BG |
258 | |
259 | static void handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn, | |
9a77daac BG |
260 | u64 old_spte, u64 new_spte, int level, |
261 | bool shared); | |
2f2fad08 | 262 | |
f8e14497 BG |
263 | static void handle_changed_spte_acc_track(u64 old_spte, u64 new_spte, int level) |
264 | { | |
f8e14497 BG |
265 | if (!is_shadow_present_pte(old_spte) || !is_last_spte(old_spte, level)) |
266 | return; | |
267 | ||
268 | if (is_accessed_spte(old_spte) && | |
64bb2769 SC |
269 | (!is_shadow_present_pte(new_spte) || !is_accessed_spte(new_spte) || |
270 | spte_to_pfn(old_spte) != spte_to_pfn(new_spte))) | |
f8e14497 BG |
271 | kvm_set_pfn_accessed(spte_to_pfn(old_spte)); |
272 | } | |
273 | ||
a6a0b05d BG |
274 | static void handle_changed_spte_dirty_log(struct kvm *kvm, int as_id, gfn_t gfn, |
275 | u64 old_spte, u64 new_spte, int level) | |
276 | { | |
277 | bool pfn_changed; | |
278 | struct kvm_memory_slot *slot; | |
279 | ||
280 | if (level > PG_LEVEL_4K) | |
281 | return; | |
282 | ||
283 | pfn_changed = spte_to_pfn(old_spte) != spte_to_pfn(new_spte); | |
284 | ||
285 | if ((!is_writable_pte(old_spte) || pfn_changed) && | |
286 | is_writable_pte(new_spte)) { | |
287 | slot = __gfn_to_memslot(__kvm_memslots(kvm, as_id), gfn); | |
fb04a1ed | 288 | mark_page_dirty_in_slot(kvm, slot, gfn); |
a6a0b05d BG |
289 | } |
290 | } | |
291 | ||
a9442f59 | 292 | /** |
c298a30c | 293 | * tdp_mmu_unlink_sp() - Remove a shadow page from the list of used pages |
a9442f59 BG |
294 | * |
295 | * @kvm: kvm instance | |
296 | * @sp: the page to be removed | |
9a77daac BG |
297 | * @shared: This operation may not be running under the exclusive use of |
298 | * the MMU lock and the operation must synchronize with other | |
299 | * threads that might be adding or removing pages. | |
a9442f59 | 300 | */ |
c298a30c DM |
301 | static void tdp_mmu_unlink_sp(struct kvm *kvm, struct kvm_mmu_page *sp, |
302 | bool shared) | |
a9442f59 | 303 | { |
9a77daac BG |
304 | if (shared) |
305 | spin_lock(&kvm->arch.tdp_mmu_pages_lock); | |
306 | else | |
307 | lockdep_assert_held_write(&kvm->mmu_lock); | |
a9442f59 BG |
308 | |
309 | list_del(&sp->link); | |
310 | if (sp->lpage_disallowed) | |
311 | unaccount_huge_nx_page(kvm, sp); | |
9a77daac BG |
312 | |
313 | if (shared) | |
314 | spin_unlock(&kvm->arch.tdp_mmu_pages_lock); | |
a9442f59 BG |
315 | } |
316 | ||
a066e61f | 317 | /** |
0f53dfa3 | 318 | * handle_removed_pt() - handle a page table removed from the TDP structure |
a066e61f BG |
319 | * |
320 | * @kvm: kvm instance | |
321 | * @pt: the page removed from the paging structure | |
9a77daac BG |
322 | * @shared: This operation may not be running under the exclusive use |
323 | * of the MMU lock and the operation must synchronize with other | |
324 | * threads that might be modifying SPTEs. | |
a066e61f BG |
325 | * |
326 | * Given a page table that has been removed from the TDP paging structure, | |
327 | * iterates through the page table to clear SPTEs and free child page tables. | |
70fb3e41 BG |
328 | * |
329 | * Note that pt is passed in as a tdp_ptep_t, but it does not need RCU | |
330 | * protection. Since this thread removed it from the paging structure, | |
331 | * this thread will be responsible for ensuring the page is freed. Hence the | |
332 | * early rcu_dereferences in the function. | |
a066e61f | 333 | */ |
0f53dfa3 | 334 | static void handle_removed_pt(struct kvm *kvm, tdp_ptep_t pt, bool shared) |
a066e61f | 335 | { |
70fb3e41 | 336 | struct kvm_mmu_page *sp = sptep_to_sp(rcu_dereference(pt)); |
a066e61f | 337 | int level = sp->role.level; |
e25f0e0c | 338 | gfn_t base_gfn = sp->gfn; |
a066e61f BG |
339 | int i; |
340 | ||
341 | trace_kvm_mmu_prepare_zap_page(sp); | |
342 | ||
c298a30c | 343 | tdp_mmu_unlink_sp(kvm, sp, shared); |
a066e61f BG |
344 | |
345 | for (i = 0; i < PT64_ENT_PER_PAGE; i++) { | |
574c3c55 BG |
346 | u64 *sptep = rcu_dereference(pt) + i; |
347 | gfn_t gfn = base_gfn + i * KVM_PAGES_PER_HPAGE(level); | |
348 | u64 old_child_spte; | |
9a77daac BG |
349 | |
350 | if (shared) { | |
e25f0e0c BG |
351 | /* |
352 | * Set the SPTE to a nonpresent value that other | |
353 | * threads will not overwrite. If the SPTE was | |
354 | * already marked as removed then another thread | |
355 | * handling a page fault could overwrite it, so | |
356 | * set the SPTE until it is set from some other | |
357 | * value to the removed SPTE value. | |
358 | */ | |
359 | for (;;) { | |
360 | old_child_spte = xchg(sptep, REMOVED_SPTE); | |
361 | if (!is_removed_spte(old_child_spte)) | |
362 | break; | |
363 | cpu_relax(); | |
364 | } | |
9a77daac | 365 | } else { |
8df9f1af SC |
366 | /* |
367 | * If the SPTE is not MMU-present, there is no backing | |
368 | * page associated with the SPTE and so no side effects | |
369 | * that need to be recorded, and exclusive ownership of | |
370 | * mmu_lock ensures the SPTE can't be made present. | |
371 | * Note, zapping MMIO SPTEs is also unnecessary as they | |
372 | * are guarded by the memslots generation, not by being | |
373 | * unreachable. | |
374 | */ | |
9a77daac | 375 | old_child_spte = READ_ONCE(*sptep); |
8df9f1af SC |
376 | if (!is_shadow_present_pte(old_child_spte)) |
377 | continue; | |
e25f0e0c BG |
378 | |
379 | /* | |
380 | * Marking the SPTE as a removed SPTE is not | |
381 | * strictly necessary here as the MMU lock will | |
382 | * stop other threads from concurrently modifying | |
383 | * this SPTE. Using the removed SPTE value keeps | |
384 | * the two branches consistent and simplifies | |
385 | * the function. | |
386 | */ | |
387 | WRITE_ONCE(*sptep, REMOVED_SPTE); | |
9a77daac | 388 | } |
e25f0e0c | 389 | handle_changed_spte(kvm, kvm_mmu_page_as_id(sp), gfn, |
f1b83255 | 390 | old_child_spte, REMOVED_SPTE, level, |
e25f0e0c | 391 | shared); |
a066e61f BG |
392 | } |
393 | ||
574c3c55 | 394 | kvm_flush_remote_tlbs_with_address(kvm, base_gfn, |
f1b83255 | 395 | KVM_PAGES_PER_HPAGE(level + 1)); |
a066e61f | 396 | |
7cca2d0b | 397 | call_rcu(&sp->rcu_head, tdp_mmu_free_sp_rcu_callback); |
a066e61f BG |
398 | } |
399 | ||
2f2fad08 | 400 | /** |
7f6231a3 | 401 | * __handle_changed_spte - handle bookkeeping associated with an SPTE change |
2f2fad08 BG |
402 | * @kvm: kvm instance |
403 | * @as_id: the address space of the paging structure the SPTE was a part of | |
404 | * @gfn: the base GFN that was mapped by the SPTE | |
405 | * @old_spte: The value of the SPTE before the change | |
406 | * @new_spte: The value of the SPTE after the change | |
407 | * @level: the level of the PT the SPTE is part of in the paging structure | |
9a77daac BG |
408 | * @shared: This operation may not be running under the exclusive use of |
409 | * the MMU lock and the operation must synchronize with other | |
410 | * threads that might be modifying SPTEs. | |
2f2fad08 BG |
411 | * |
412 | * Handle bookkeeping that might result from the modification of a SPTE. | |
413 | * This function must be called for all TDP SPTE modifications. | |
414 | */ | |
415 | static void __handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn, | |
9a77daac BG |
416 | u64 old_spte, u64 new_spte, int level, |
417 | bool shared) | |
2f2fad08 BG |
418 | { |
419 | bool was_present = is_shadow_present_pte(old_spte); | |
420 | bool is_present = is_shadow_present_pte(new_spte); | |
421 | bool was_leaf = was_present && is_last_spte(old_spte, level); | |
422 | bool is_leaf = is_present && is_last_spte(new_spte, level); | |
423 | bool pfn_changed = spte_to_pfn(old_spte) != spte_to_pfn(new_spte); | |
2f2fad08 BG |
424 | |
425 | WARN_ON(level > PT64_ROOT_MAX_LEVEL); | |
426 | WARN_ON(level < PG_LEVEL_4K); | |
764388ce | 427 | WARN_ON(gfn & (KVM_PAGES_PER_HPAGE(level) - 1)); |
2f2fad08 BG |
428 | |
429 | /* | |
430 | * If this warning were to trigger it would indicate that there was a | |
431 | * missing MMU notifier or a race with some notifier handler. | |
432 | * A present, leaf SPTE should never be directly replaced with another | |
d9f6e12f | 433 | * present leaf SPTE pointing to a different PFN. A notifier handler |
2f2fad08 BG |
434 | * should be zapping the SPTE before the main MM's page table is |
435 | * changed, or the SPTE should be zeroed, and the TLBs flushed by the | |
436 | * thread before replacement. | |
437 | */ | |
438 | if (was_leaf && is_leaf && pfn_changed) { | |
439 | pr_err("Invalid SPTE change: cannot replace a present leaf\n" | |
440 | "SPTE with another present leaf SPTE mapping a\n" | |
441 | "different PFN!\n" | |
442 | "as_id: %d gfn: %llx old_spte: %llx new_spte: %llx level: %d", | |
443 | as_id, gfn, old_spte, new_spte, level); | |
444 | ||
445 | /* | |
446 | * Crash the host to prevent error propagation and guest data | |
d9f6e12f | 447 | * corruption. |
2f2fad08 BG |
448 | */ |
449 | BUG(); | |
450 | } | |
451 | ||
452 | if (old_spte == new_spte) | |
453 | return; | |
454 | ||
b9a98c34 BG |
455 | trace_kvm_tdp_mmu_spte_changed(as_id, gfn, level, old_spte, new_spte); |
456 | ||
115111ef DM |
457 | if (is_leaf) |
458 | check_spte_writable_invariants(new_spte); | |
459 | ||
2f2fad08 BG |
460 | /* |
461 | * The only times a SPTE should be changed from a non-present to | |
462 | * non-present state is when an MMIO entry is installed/modified/ | |
463 | * removed. In that case, there is nothing to do here. | |
464 | */ | |
465 | if (!was_present && !is_present) { | |
466 | /* | |
08f07c80 BG |
467 | * If this change does not involve a MMIO SPTE or removed SPTE, |
468 | * it is unexpected. Log the change, though it should not | |
469 | * impact the guest since both the former and current SPTEs | |
470 | * are nonpresent. | |
2f2fad08 | 471 | */ |
08f07c80 BG |
472 | if (WARN_ON(!is_mmio_spte(old_spte) && |
473 | !is_mmio_spte(new_spte) && | |
474 | !is_removed_spte(new_spte))) | |
2f2fad08 BG |
475 | pr_err("Unexpected SPTE change! Nonpresent SPTEs\n" |
476 | "should not be replaced with another,\n" | |
477 | "different nonpresent SPTE, unless one or both\n" | |
08f07c80 BG |
478 | "are MMIO SPTEs, or the new SPTE is\n" |
479 | "a temporary removed SPTE.\n" | |
2f2fad08 BG |
480 | "as_id: %d gfn: %llx old_spte: %llx new_spte: %llx level: %d", |
481 | as_id, gfn, old_spte, new_spte, level); | |
482 | return; | |
483 | } | |
484 | ||
71f51d2c MZ |
485 | if (is_leaf != was_leaf) |
486 | kvm_update_page_stats(kvm, level, is_leaf ? 1 : -1); | |
2f2fad08 BG |
487 | |
488 | if (was_leaf && is_dirty_spte(old_spte) && | |
64bb2769 | 489 | (!is_present || !is_dirty_spte(new_spte) || pfn_changed)) |
2f2fad08 BG |
490 | kvm_set_pfn_dirty(spte_to_pfn(old_spte)); |
491 | ||
492 | /* | |
493 | * Recursively handle child PTs if the change removed a subtree from | |
c8e5a0d0 SC |
494 | * the paging structure. Note the WARN on the PFN changing without the |
495 | * SPTE being converted to a hugepage (leaf) or being zapped. Shadow | |
496 | * pages are kernel allocations and should never be migrated. | |
2f2fad08 | 497 | */ |
c8e5a0d0 SC |
498 | if (was_present && !was_leaf && |
499 | (is_leaf || !is_present || WARN_ON_ONCE(pfn_changed))) | |
0f53dfa3 | 500 | handle_removed_pt(kvm, spte_to_child_pt(old_spte, level), shared); |
2f2fad08 BG |
501 | } |
502 | ||
503 | static void handle_changed_spte(struct kvm *kvm, int as_id, gfn_t gfn, | |
9a77daac BG |
504 | u64 old_spte, u64 new_spte, int level, |
505 | bool shared) | |
2f2fad08 | 506 | { |
9a77daac BG |
507 | __handle_changed_spte(kvm, as_id, gfn, old_spte, new_spte, level, |
508 | shared); | |
f8e14497 | 509 | handle_changed_spte_acc_track(old_spte, new_spte, level); |
a6a0b05d BG |
510 | handle_changed_spte_dirty_log(kvm, as_id, gfn, old_spte, |
511 | new_spte, level); | |
2f2fad08 | 512 | } |
faaf05b0 | 513 | |
9a77daac | 514 | /* |
6ccf4438 PB |
515 | * tdp_mmu_set_spte_atomic - Set a TDP MMU SPTE atomically |
516 | * and handle the associated bookkeeping. Do not mark the page dirty | |
24ae4cfa | 517 | * in KVM's dirty bitmaps. |
9a77daac | 518 | * |
3255530a DM |
519 | * If setting the SPTE fails because it has changed, iter->old_spte will be |
520 | * refreshed to the current value of the spte. | |
521 | * | |
9a77daac BG |
522 | * @kvm: kvm instance |
523 | * @iter: a tdp_iter instance currently on the SPTE that should be set | |
524 | * @new_spte: The value the SPTE should be set to | |
3e72c791 DM |
525 | * Return: |
526 | * * 0 - If the SPTE was set. | |
527 | * * -EBUSY - If the SPTE cannot be set. In this case this function will have | |
528 | * no side-effects other than setting iter->old_spte to the last | |
529 | * known value of the spte. | |
9a77daac | 530 | */ |
3e72c791 DM |
531 | static inline int tdp_mmu_set_spte_atomic(struct kvm *kvm, |
532 | struct tdp_iter *iter, | |
533 | u64 new_spte) | |
9a77daac | 534 | { |
3255530a DM |
535 | u64 *sptep = rcu_dereference(iter->sptep); |
536 | u64 old_spte; | |
537 | ||
3a0f64de SC |
538 | WARN_ON_ONCE(iter->yielded); |
539 | ||
9a77daac BG |
540 | lockdep_assert_held_read(&kvm->mmu_lock); |
541 | ||
08f07c80 BG |
542 | /* |
543 | * Do not change removed SPTEs. Only the thread that froze the SPTE | |
544 | * may modify it. | |
545 | */ | |
7a51393a | 546 | if (is_removed_spte(iter->old_spte)) |
3e72c791 | 547 | return -EBUSY; |
08f07c80 | 548 | |
6e8eb206 DM |
549 | /* |
550 | * Note, fast_pf_fix_direct_spte() can also modify TDP MMU SPTEs and | |
551 | * does not hold the mmu_lock. | |
552 | */ | |
3255530a DM |
553 | old_spte = cmpxchg64(sptep, iter->old_spte, new_spte); |
554 | if (old_spte != iter->old_spte) { | |
555 | /* | |
556 | * The page table entry was modified by a different logical | |
557 | * CPU. Refresh iter->old_spte with the current value so the | |
558 | * caller operates on fresh data, e.g. if it retries | |
559 | * tdp_mmu_set_spte_atomic(). | |
560 | */ | |
561 | iter->old_spte = old_spte; | |
3e72c791 | 562 | return -EBUSY; |
3255530a | 563 | } |
9a77daac | 564 | |
24ae4cfa BG |
565 | __handle_changed_spte(kvm, iter->as_id, iter->gfn, iter->old_spte, |
566 | new_spte, iter->level, true); | |
567 | handle_changed_spte_acc_track(iter->old_spte, new_spte, iter->level); | |
9a77daac | 568 | |
3e72c791 | 569 | return 0; |
9a77daac BG |
570 | } |
571 | ||
3e72c791 DM |
572 | static inline int tdp_mmu_zap_spte_atomic(struct kvm *kvm, |
573 | struct tdp_iter *iter) | |
08f07c80 | 574 | { |
3e72c791 DM |
575 | int ret; |
576 | ||
08f07c80 BG |
577 | /* |
578 | * Freeze the SPTE by setting it to a special, | |
579 | * non-present value. This will stop other threads from | |
580 | * immediately installing a present entry in its place | |
581 | * before the TLBs are flushed. | |
582 | */ | |
3e72c791 DM |
583 | ret = tdp_mmu_set_spte_atomic(kvm, iter, REMOVED_SPTE); |
584 | if (ret) | |
585 | return ret; | |
08f07c80 BG |
586 | |
587 | kvm_flush_remote_tlbs_with_address(kvm, iter->gfn, | |
588 | KVM_PAGES_PER_HPAGE(iter->level)); | |
589 | ||
590 | /* | |
591 | * No other thread can overwrite the removed SPTE as they | |
592 | * must either wait on the MMU lock or use | |
d9f6e12f | 593 | * tdp_mmu_set_spte_atomic which will not overwrite the |
08f07c80 BG |
594 | * special removed SPTE value. No bookkeeping is needed |
595 | * here since the SPTE is going from non-present | |
596 | * to non-present. | |
597 | */ | |
0e587aa7 | 598 | kvm_tdp_mmu_write_spte(iter->sptep, 0); |
08f07c80 | 599 | |
3e72c791 | 600 | return 0; |
08f07c80 BG |
601 | } |
602 | ||
9a77daac | 603 | |
fe43fa2f BG |
604 | /* |
605 | * __tdp_mmu_set_spte - Set a TDP MMU SPTE and handle the associated bookkeeping | |
626808d1 SC |
606 | * @kvm: KVM instance |
607 | * @as_id: Address space ID, i.e. regular vs. SMM | |
608 | * @sptep: Pointer to the SPTE | |
609 | * @old_spte: The current value of the SPTE | |
610 | * @new_spte: The new value that will be set for the SPTE | |
611 | * @gfn: The base GFN that was (or will be) mapped by the SPTE | |
612 | * @level: The level _containing_ the SPTE (its parent PT's level) | |
fe43fa2f BG |
613 | * @record_acc_track: Notify the MM subsystem of changes to the accessed state |
614 | * of the page. Should be set unless handling an MMU | |
615 | * notifier for access tracking. Leaving record_acc_track | |
616 | * unset in that case prevents page accesses from being | |
617 | * double counted. | |
618 | * @record_dirty_log: Record the page as dirty in the dirty bitmap if | |
619 | * appropriate for the change being made. Should be set | |
620 | * unless performing certain dirty logging operations. | |
621 | * Leaving record_dirty_log unset in that case prevents page | |
622 | * writes from being double counted. | |
623 | */ | |
626808d1 SC |
624 | static void __tdp_mmu_set_spte(struct kvm *kvm, int as_id, tdp_ptep_t sptep, |
625 | u64 old_spte, u64 new_spte, gfn_t gfn, int level, | |
626 | bool record_acc_track, bool record_dirty_log) | |
faaf05b0 | 627 | { |
531810ca | 628 | lockdep_assert_held_write(&kvm->mmu_lock); |
3a9a4aa5 | 629 | |
08f07c80 | 630 | /* |
966da62a | 631 | * No thread should be using this function to set SPTEs to or from the |
08f07c80 BG |
632 | * temporary removed SPTE value. |
633 | * If operating under the MMU lock in read mode, tdp_mmu_set_spte_atomic | |
634 | * should be used. If operating under the MMU lock in write mode, the | |
635 | * use of the removed SPTE should not be necessary. | |
636 | */ | |
626808d1 | 637 | WARN_ON(is_removed_spte(old_spte) || is_removed_spte(new_spte)); |
08f07c80 | 638 | |
626808d1 SC |
639 | kvm_tdp_mmu_write_spte(sptep, new_spte); |
640 | ||
641 | __handle_changed_spte(kvm, as_id, gfn, old_spte, new_spte, level, false); | |
f8e14497 | 642 | |
f8e14497 | 643 | if (record_acc_track) |
626808d1 | 644 | handle_changed_spte_acc_track(old_spte, new_spte, level); |
a6a0b05d | 645 | if (record_dirty_log) |
626808d1 SC |
646 | handle_changed_spte_dirty_log(kvm, as_id, gfn, old_spte, |
647 | new_spte, level); | |
648 | } | |
649 | ||
650 | static inline void _tdp_mmu_set_spte(struct kvm *kvm, struct tdp_iter *iter, | |
651 | u64 new_spte, bool record_acc_track, | |
652 | bool record_dirty_log) | |
653 | { | |
654 | WARN_ON_ONCE(iter->yielded); | |
655 | ||
656 | __tdp_mmu_set_spte(kvm, iter->as_id, iter->sptep, iter->old_spte, | |
657 | new_spte, iter->gfn, iter->level, | |
658 | record_acc_track, record_dirty_log); | |
f8e14497 BG |
659 | } |
660 | ||
661 | static inline void tdp_mmu_set_spte(struct kvm *kvm, struct tdp_iter *iter, | |
662 | u64 new_spte) | |
663 | { | |
626808d1 | 664 | _tdp_mmu_set_spte(kvm, iter, new_spte, true, true); |
f8e14497 | 665 | } |
faaf05b0 | 666 | |
f8e14497 BG |
667 | static inline void tdp_mmu_set_spte_no_acc_track(struct kvm *kvm, |
668 | struct tdp_iter *iter, | |
669 | u64 new_spte) | |
670 | { | |
626808d1 | 671 | _tdp_mmu_set_spte(kvm, iter, new_spte, false, true); |
a6a0b05d BG |
672 | } |
673 | ||
674 | static inline void tdp_mmu_set_spte_no_dirty_log(struct kvm *kvm, | |
675 | struct tdp_iter *iter, | |
676 | u64 new_spte) | |
677 | { | |
626808d1 | 678 | _tdp_mmu_set_spte(kvm, iter, new_spte, true, false); |
faaf05b0 BG |
679 | } |
680 | ||
681 | #define tdp_root_for_each_pte(_iter, _root, _start, _end) \ | |
77aa6075 | 682 | for_each_tdp_pte(_iter, _root, _start, _end) |
faaf05b0 | 683 | |
f8e14497 BG |
684 | #define tdp_root_for_each_leaf_pte(_iter, _root, _start, _end) \ |
685 | tdp_root_for_each_pte(_iter, _root, _start, _end) \ | |
686 | if (!is_shadow_present_pte(_iter.old_spte) || \ | |
687 | !is_last_spte(_iter.old_spte, _iter.level)) \ | |
688 | continue; \ | |
689 | else | |
690 | ||
bb18842e | 691 | #define tdp_mmu_for_each_pte(_iter, _mmu, _start, _end) \ |
b9e5603c | 692 | for_each_tdp_pte(_iter, to_shadow_page(_mmu->root.hpa), _start, _end) |
bb18842e | 693 | |
e28a436c BG |
694 | /* |
695 | * Yield if the MMU lock is contended or this thread needs to return control | |
696 | * to the scheduler. | |
697 | * | |
e139a34e BG |
698 | * If this function should yield and flush is set, it will perform a remote |
699 | * TLB flush before yielding. | |
700 | * | |
3a0f64de SC |
701 | * If this function yields, iter->yielded is set and the caller must skip to |
702 | * the next iteration, where tdp_iter_next() will reset the tdp_iter's walk | |
703 | * over the paging structures to allow the iterator to continue its traversal | |
704 | * from the paging structure root. | |
e28a436c | 705 | * |
3a0f64de | 706 | * Returns true if this function yielded. |
e28a436c | 707 | */ |
3a0f64de SC |
708 | static inline bool __must_check tdp_mmu_iter_cond_resched(struct kvm *kvm, |
709 | struct tdp_iter *iter, | |
710 | bool flush, bool shared) | |
a6a0b05d | 711 | { |
3a0f64de SC |
712 | WARN_ON(iter->yielded); |
713 | ||
ed5e484b BG |
714 | /* Ensure forward progress has been made before yielding. */ |
715 | if (iter->next_last_level_gfn == iter->yielded_gfn) | |
716 | return false; | |
717 | ||
531810ca | 718 | if (need_resched() || rwlock_needbreak(&kvm->mmu_lock)) { |
7cca2d0b BG |
719 | rcu_read_unlock(); |
720 | ||
e139a34e BG |
721 | if (flush) |
722 | kvm_flush_remote_tlbs(kvm); | |
723 | ||
6103bc07 BG |
724 | if (shared) |
725 | cond_resched_rwlock_read(&kvm->mmu_lock); | |
726 | else | |
727 | cond_resched_rwlock_write(&kvm->mmu_lock); | |
728 | ||
7cca2d0b | 729 | rcu_read_lock(); |
ed5e484b BG |
730 | |
731 | WARN_ON(iter->gfn > iter->next_last_level_gfn); | |
732 | ||
3a0f64de | 733 | iter->yielded = true; |
a6a0b05d | 734 | } |
e28a436c | 735 | |
3a0f64de | 736 | return iter->yielded; |
a6a0b05d BG |
737 | } |
738 | ||
faaf05b0 BG |
739 | /* |
740 | * Tears down the mappings for the range of gfns, [start, end), and frees the | |
741 | * non-root pages mapping GFNs strictly within that range. Returns true if | |
742 | * SPTEs have been cleared and a TLB flush is needed before releasing the | |
743 | * MMU lock. | |
6103bc07 | 744 | * |
063afacd BG |
745 | * If can_yield is true, will release the MMU lock and reschedule if the |
746 | * scheduler needs the CPU or there is contention on the MMU lock. If this | |
747 | * function cannot yield, it will not release the MMU lock or reschedule and | |
748 | * the caller must ensure it does not supply too large a GFN range, or the | |
6103bc07 BG |
749 | * operation can cause a soft lockup. |
750 | * | |
751 | * If shared is true, this thread holds the MMU lock in read mode and must | |
752 | * account for the possibility that other threads are modifying the paging | |
753 | * structures concurrently. If shared is false, this thread should hold the | |
754 | * MMU lock in write mode. | |
faaf05b0 BG |
755 | */ |
756 | static bool zap_gfn_range(struct kvm *kvm, struct kvm_mmu_page *root, | |
6103bc07 BG |
757 | gfn_t start, gfn_t end, bool can_yield, bool flush, |
758 | bool shared) | |
faaf05b0 | 759 | { |
524a1e4e SC |
760 | gfn_t max_gfn_host = 1ULL << (shadow_phys_bits - PAGE_SHIFT); |
761 | bool zap_all = (start == 0 && end >= max_gfn_host); | |
faaf05b0 | 762 | struct tdp_iter iter; |
faaf05b0 | 763 | |
0103098f SC |
764 | /* |
765 | * No need to try to step down in the iterator when zapping all SPTEs, | |
766 | * zapping the top-level non-leaf SPTEs will recurse on their children. | |
767 | */ | |
768 | int min_level = zap_all ? root->role.level : PG_LEVEL_4K; | |
769 | ||
524a1e4e SC |
770 | /* |
771 | * Bound the walk at host.MAXPHYADDR, guest accesses beyond that will | |
772 | * hit a #PF(RSVD) and never get to an EPT Violation/Misconfig / #NPF, | |
773 | * and so KVM will never install a SPTE for such addresses. | |
774 | */ | |
775 | end = min(end, max_gfn_host); | |
776 | ||
6103bc07 BG |
777 | kvm_lockdep_assert_mmu_lock_held(kvm, shared); |
778 | ||
7cca2d0b BG |
779 | rcu_read_lock(); |
780 | ||
77aa6075 | 781 | for_each_tdp_pte_min_level(iter, root, min_level, start, end) { |
6103bc07 | 782 | retry: |
1af4a960 | 783 | if (can_yield && |
6103bc07 | 784 | tdp_mmu_iter_cond_resched(kvm, &iter, flush, shared)) { |
a835429c | 785 | flush = false; |
1af4a960 BG |
786 | continue; |
787 | } | |
788 | ||
faaf05b0 BG |
789 | if (!is_shadow_present_pte(iter.old_spte)) |
790 | continue; | |
791 | ||
792 | /* | |
793 | * If this is a non-last-level SPTE that covers a larger range | |
794 | * than should be zapped, continue, and zap the mappings at a | |
524a1e4e | 795 | * lower level, except when zapping all SPTEs. |
faaf05b0 | 796 | */ |
524a1e4e SC |
797 | if (!zap_all && |
798 | (iter.gfn < start || | |
faaf05b0 BG |
799 | iter.gfn + KVM_PAGES_PER_HPAGE(iter.level) > end) && |
800 | !is_last_spte(iter.old_spte, iter.level)) | |
801 | continue; | |
802 | ||
6103bc07 BG |
803 | if (!shared) { |
804 | tdp_mmu_set_spte(kvm, &iter, 0); | |
805 | flush = true; | |
3e72c791 | 806 | } else if (tdp_mmu_zap_spte_atomic(kvm, &iter)) { |
6103bc07 BG |
807 | goto retry; |
808 | } | |
faaf05b0 | 809 | } |
7cca2d0b BG |
810 | |
811 | rcu_read_unlock(); | |
a835429c | 812 | return flush; |
faaf05b0 BG |
813 | } |
814 | ||
815 | /* | |
816 | * Tears down the mappings for the range of gfns, [start, end), and frees the | |
817 | * non-root pages mapping GFNs strictly within that range. Returns true if | |
818 | * SPTEs have been cleared and a TLB flush is needed before releasing the | |
819 | * MMU lock. | |
820 | */ | |
2b9663d8 | 821 | bool __kvm_tdp_mmu_zap_gfn_range(struct kvm *kvm, int as_id, gfn_t start, |
5a324c24 | 822 | gfn_t end, bool can_yield, bool flush) |
faaf05b0 BG |
823 | { |
824 | struct kvm_mmu_page *root; | |
faaf05b0 | 825 | |
614f6970 | 826 | for_each_tdp_mmu_root_yield_safe(kvm, root, as_id) |
6103bc07 | 827 | flush = zap_gfn_range(kvm, root, start, end, can_yield, flush, |
5a324c24 | 828 | false); |
faaf05b0 | 829 | |
faaf05b0 BG |
830 | return flush; |
831 | } | |
832 | ||
833 | void kvm_tdp_mmu_zap_all(struct kvm *kvm) | |
834 | { | |
2b9663d8 SC |
835 | bool flush = false; |
836 | int i; | |
837 | ||
838 | for (i = 0; i < KVM_ADDRESS_SPACE_NUM; i++) | |
5a324c24 | 839 | flush = kvm_tdp_mmu_zap_gfn_range(kvm, i, 0, -1ull, flush); |
faaf05b0 | 840 | |
faaf05b0 BG |
841 | if (flush) |
842 | kvm_flush_remote_tlbs(kvm); | |
843 | } | |
bb18842e | 844 | |
4c6654bd BG |
845 | static struct kvm_mmu_page *next_invalidated_root(struct kvm *kvm, |
846 | struct kvm_mmu_page *prev_root) | |
847 | { | |
848 | struct kvm_mmu_page *next_root; | |
849 | ||
850 | if (prev_root) | |
851 | next_root = list_next_or_null_rcu(&kvm->arch.tdp_mmu_roots, | |
852 | &prev_root->link, | |
853 | typeof(*prev_root), link); | |
854 | else | |
855 | next_root = list_first_or_null_rcu(&kvm->arch.tdp_mmu_roots, | |
856 | typeof(*next_root), link); | |
857 | ||
858 | while (next_root && !(next_root->role.invalid && | |
859 | refcount_read(&next_root->tdp_mmu_root_count))) | |
860 | next_root = list_next_or_null_rcu(&kvm->arch.tdp_mmu_roots, | |
861 | &next_root->link, | |
862 | typeof(*next_root), link); | |
863 | ||
864 | return next_root; | |
865 | } | |
866 | ||
867 | /* | |
f28e9c7f SC |
868 | * Zap all invalidated roots to ensure all SPTEs are dropped before the "fast |
869 | * zap" completes. Since kvm_tdp_mmu_invalidate_all_roots() has acquired a | |
870 | * reference to each invalidated root, roots will not be freed until after this | |
871 | * function drops the gifted reference, e.g. so that vCPUs don't get stuck with | |
872 | * tearing down paging structures. | |
4c6654bd BG |
873 | */ |
874 | void kvm_tdp_mmu_zap_invalidated_roots(struct kvm *kvm) | |
875 | { | |
4c6654bd BG |
876 | struct kvm_mmu_page *next_root; |
877 | struct kvm_mmu_page *root; | |
4c6654bd BG |
878 | |
879 | lockdep_assert_held_read(&kvm->mmu_lock); | |
880 | ||
881 | rcu_read_lock(); | |
882 | ||
883 | root = next_invalidated_root(kvm, NULL); | |
884 | ||
885 | while (root) { | |
886 | next_root = next_invalidated_root(kvm, root); | |
887 | ||
888 | rcu_read_unlock(); | |
889 | ||
7ae5840e SC |
890 | /* |
891 | * A TLB flush is unnecessary, invalidated roots are guaranteed | |
892 | * to be unreachable by the guest (see kvm_tdp_mmu_put_root() | |
893 | * for more details), and unlike the legacy MMU, no vCPU kick | |
894 | * is needed to play nice with lockless shadow walks as the TDP | |
895 | * MMU protects its paging structures via RCU. Note, zapping | |
896 | * will still flush on yield, but that's a minor performance | |
897 | * blip and not a functional issue. | |
898 | */ | |
899 | (void)zap_gfn_range(kvm, root, 0, -1ull, true, false, true); | |
4c6654bd BG |
900 | |
901 | /* | |
902 | * Put the reference acquired in | |
903 | * kvm_tdp_mmu_invalidate_roots | |
904 | */ | |
905 | kvm_tdp_mmu_put_root(kvm, root, true); | |
906 | ||
907 | root = next_root; | |
908 | ||
909 | rcu_read_lock(); | |
910 | } | |
911 | ||
912 | rcu_read_unlock(); | |
faaf05b0 | 913 | } |
bb18842e | 914 | |
b7cccd39 | 915 | /* |
f28e9c7f SC |
916 | * Mark each TDP MMU root as invalid to prevent vCPUs from reusing a root that |
917 | * is about to be zapped, e.g. in response to a memslots update. The caller is | |
918 | * responsible for invoking kvm_tdp_mmu_zap_invalidated_roots() to do the actual | |
919 | * zapping. | |
b7cccd39 | 920 | * |
f28e9c7f SC |
921 | * Take a reference on all roots to prevent the root from being freed before it |
922 | * is zapped by this thread. Freeing a root is not a correctness issue, but if | |
923 | * a vCPU drops the last reference to a root prior to the root being zapped, it | |
924 | * will get stuck with tearing down the entire paging structure. | |
4c6654bd | 925 | * |
f28e9c7f SC |
926 | * Get a reference even if the root is already invalid, |
927 | * kvm_tdp_mmu_zap_invalidated_roots() assumes it was gifted a reference to all | |
928 | * invalid roots, e.g. there's no epoch to identify roots that were invalidated | |
929 | * by a previous call. Roots stay on the list until the last reference is | |
930 | * dropped, so even though all invalid roots are zapped, a root may not go away | |
931 | * for quite some time, e.g. if a vCPU blocks across multiple memslot updates. | |
932 | * | |
933 | * Because mmu_lock is held for write, it should be impossible to observe a | |
934 | * root with zero refcount, i.e. the list of roots cannot be stale. | |
4c6654bd | 935 | * |
b7cccd39 BG |
936 | * This has essentially the same effect for the TDP MMU |
937 | * as updating mmu_valid_gen does for the shadow MMU. | |
938 | */ | |
939 | void kvm_tdp_mmu_invalidate_all_roots(struct kvm *kvm) | |
940 | { | |
941 | struct kvm_mmu_page *root; | |
942 | ||
943 | lockdep_assert_held_write(&kvm->mmu_lock); | |
f28e9c7f SC |
944 | list_for_each_entry(root, &kvm->arch.tdp_mmu_roots, link) { |
945 | if (!WARN_ON_ONCE(!kvm_tdp_mmu_get_root(root))) | |
4c6654bd | 946 | root->role.invalid = true; |
f28e9c7f | 947 | } |
b7cccd39 BG |
948 | } |
949 | ||
bb18842e BG |
950 | /* |
951 | * Installs a last-level SPTE to handle a TDP page fault. | |
952 | * (NPT/EPT violation/misconfiguration) | |
953 | */ | |
cdc47767 PB |
954 | static int tdp_mmu_map_handle_target_level(struct kvm_vcpu *vcpu, |
955 | struct kvm_page_fault *fault, | |
956 | struct tdp_iter *iter) | |
bb18842e | 957 | { |
c435d4b7 | 958 | struct kvm_mmu_page *sp = sptep_to_sp(rcu_dereference(iter->sptep)); |
bb18842e | 959 | u64 new_spte; |
57a3e96d | 960 | int ret = RET_PF_FIXED; |
ad67e480 | 961 | bool wrprot = false; |
bb18842e | 962 | |
7158bee4 | 963 | WARN_ON(sp->role.level != fault->goal_level); |
e710c5f6 | 964 | if (unlikely(!fault->slot)) |
bb18842e | 965 | new_spte = make_mmio_spte(vcpu, iter->gfn, ACC_ALL); |
9a77daac | 966 | else |
53597858 | 967 | wrprot = make_spte(vcpu, sp, fault->slot, ACC_ALL, iter->gfn, |
2839180c | 968 | fault->pfn, iter->old_spte, fault->prefetch, true, |
7158bee4 | 969 | fault->map_writable, &new_spte); |
bb18842e BG |
970 | |
971 | if (new_spte == iter->old_spte) | |
972 | ret = RET_PF_SPURIOUS; | |
3e72c791 | 973 | else if (tdp_mmu_set_spte_atomic(vcpu->kvm, iter, new_spte)) |
9a77daac | 974 | return RET_PF_RETRY; |
bb18842e BG |
975 | |
976 | /* | |
977 | * If the page fault was caused by a write but the page is write | |
978 | * protected, emulation is needed. If the emulation was skipped, | |
979 | * the vCPU would have the same fault again. | |
980 | */ | |
ad67e480 | 981 | if (wrprot) { |
cdc47767 | 982 | if (fault->write) |
bb18842e | 983 | ret = RET_PF_EMULATE; |
bb18842e BG |
984 | } |
985 | ||
986 | /* If a MMIO SPTE is installed, the MMIO will need to be emulated. */ | |
9a77daac BG |
987 | if (unlikely(is_mmio_spte(new_spte))) { |
988 | trace_mark_mmio_spte(rcu_dereference(iter->sptep), iter->gfn, | |
989 | new_spte); | |
bb18842e | 990 | ret = RET_PF_EMULATE; |
3849e092 | 991 | } else { |
9a77daac BG |
992 | trace_kvm_mmu_set_spte(iter->level, iter->gfn, |
993 | rcu_dereference(iter->sptep)); | |
3849e092 | 994 | } |
bb18842e | 995 | |
857f8474 KH |
996 | /* |
997 | * Increase pf_fixed in both RET_PF_EMULATE and RET_PF_FIXED to be | |
998 | * consistent with legacy MMU behavior. | |
999 | */ | |
1000 | if (ret != RET_PF_SPURIOUS) | |
bb18842e BG |
1001 | vcpu->stat.pf_fixed++; |
1002 | ||
1003 | return ret; | |
1004 | } | |
1005 | ||
7b7e1ab6 | 1006 | /* |
cb00a70b DM |
1007 | * tdp_mmu_link_sp - Replace the given spte with an spte pointing to the |
1008 | * provided page table. | |
7b7e1ab6 DM |
1009 | * |
1010 | * @kvm: kvm instance | |
1011 | * @iter: a tdp_iter instance currently on the SPTE that should be set | |
1012 | * @sp: The new TDP page table to install. | |
1013 | * @account_nx: True if this page table is being installed to split a | |
1014 | * non-executable huge page. | |
cb00a70b | 1015 | * @shared: This operation is running under the MMU lock in read mode. |
7b7e1ab6 DM |
1016 | * |
1017 | * Returns: 0 if the new page table was installed. Non-0 if the page table | |
1018 | * could not be installed (e.g. the atomic compare-exchange failed). | |
1019 | */ | |
cb00a70b DM |
1020 | static int tdp_mmu_link_sp(struct kvm *kvm, struct tdp_iter *iter, |
1021 | struct kvm_mmu_page *sp, bool account_nx, | |
1022 | bool shared) | |
7b7e1ab6 DM |
1023 | { |
1024 | u64 spte = make_nonleaf_spte(sp->spt, !shadow_accessed_mask); | |
cb00a70b | 1025 | int ret = 0; |
7b7e1ab6 | 1026 | |
cb00a70b DM |
1027 | if (shared) { |
1028 | ret = tdp_mmu_set_spte_atomic(kvm, iter, spte); | |
1029 | if (ret) | |
1030 | return ret; | |
1031 | } else { | |
1032 | tdp_mmu_set_spte(kvm, iter, spte); | |
1033 | } | |
7b7e1ab6 DM |
1034 | |
1035 | spin_lock(&kvm->arch.tdp_mmu_pages_lock); | |
1036 | list_add(&sp->link, &kvm->arch.tdp_mmu_pages); | |
1037 | if (account_nx) | |
1038 | account_huge_nx_page(kvm, sp); | |
1039 | spin_unlock(&kvm->arch.tdp_mmu_pages_lock); | |
1040 | ||
1041 | return 0; | |
1042 | } | |
1043 | ||
bb18842e BG |
1044 | /* |
1045 | * Handle a TDP page fault (NPT/EPT violation/misconfiguration) by installing | |
1046 | * page tables and SPTEs to translate the faulting guest physical address. | |
1047 | */ | |
2f6305dd | 1048 | int kvm_tdp_mmu_map(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault) |
bb18842e | 1049 | { |
bb18842e BG |
1050 | struct kvm_mmu *mmu = vcpu->arch.mmu; |
1051 | struct tdp_iter iter; | |
89c0fd49 | 1052 | struct kvm_mmu_page *sp; |
bb18842e | 1053 | int ret; |
bb18842e | 1054 | |
73a3c659 | 1055 | kvm_mmu_hugepage_adjust(vcpu, fault); |
bb18842e | 1056 | |
f0066d94 | 1057 | trace_kvm_mmu_spte_requested(fault); |
7cca2d0b BG |
1058 | |
1059 | rcu_read_lock(); | |
1060 | ||
2f6305dd | 1061 | tdp_mmu_for_each_pte(iter, mmu, fault->gfn, fault->gfn + 1) { |
73a3c659 | 1062 | if (fault->nx_huge_page_workaround_enabled) |
536f0e6a | 1063 | disallowed_hugepage_adjust(fault, iter.old_spte, iter.level); |
bb18842e | 1064 | |
73a3c659 | 1065 | if (iter.level == fault->goal_level) |
bb18842e BG |
1066 | break; |
1067 | ||
1068 | /* | |
1069 | * If there is an SPTE mapping a large page at a higher level | |
1070 | * than the target, that SPTE must be cleared and replaced | |
1071 | * with a non-leaf SPTE. | |
1072 | */ | |
1073 | if (is_shadow_present_pte(iter.old_spte) && | |
1074 | is_large_pte(iter.old_spte)) { | |
3e72c791 | 1075 | if (tdp_mmu_zap_spte_atomic(vcpu->kvm, &iter)) |
9a77daac | 1076 | break; |
bb18842e | 1077 | |
bb18842e BG |
1078 | /* |
1079 | * The iter must explicitly re-read the spte here | |
1080 | * because the new value informs the !present | |
1081 | * path below. | |
1082 | */ | |
0e587aa7 | 1083 | iter.old_spte = kvm_tdp_mmu_read_spte(iter.sptep); |
bb18842e BG |
1084 | } |
1085 | ||
1086 | if (!is_shadow_present_pte(iter.old_spte)) { | |
7b7e1ab6 DM |
1087 | bool account_nx = fault->huge_page_disallowed && |
1088 | fault->req_level >= iter.level; | |
1089 | ||
ff76d506 | 1090 | /* |
c4342633 | 1091 | * If SPTE has been frozen by another thread, just |
ff76d506 KH |
1092 | * give up and retry, avoiding unnecessary page table |
1093 | * allocation and free. | |
1094 | */ | |
1095 | if (is_removed_spte(iter.old_spte)) | |
1096 | break; | |
1097 | ||
a82070b6 DM |
1098 | sp = tdp_mmu_alloc_sp(vcpu); |
1099 | tdp_mmu_init_child_sp(sp, &iter); | |
1100 | ||
cb00a70b | 1101 | if (tdp_mmu_link_sp(vcpu->kvm, &iter, sp, account_nx, true)) { |
9a77daac BG |
1102 | tdp_mmu_free_sp(sp); |
1103 | break; | |
1104 | } | |
bb18842e BG |
1105 | } |
1106 | } | |
1107 | ||
73a3c659 | 1108 | if (iter.level != fault->goal_level) { |
7cca2d0b | 1109 | rcu_read_unlock(); |
bb18842e | 1110 | return RET_PF_RETRY; |
7cca2d0b | 1111 | } |
bb18842e | 1112 | |
cdc47767 | 1113 | ret = tdp_mmu_map_handle_target_level(vcpu, fault, &iter); |
7cca2d0b | 1114 | rcu_read_unlock(); |
bb18842e BG |
1115 | |
1116 | return ret; | |
1117 | } | |
063afacd | 1118 | |
3039bcc7 SC |
1119 | bool kvm_tdp_mmu_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range, |
1120 | bool flush) | |
063afacd | 1121 | { |
83b83a02 SC |
1122 | return __kvm_tdp_mmu_zap_gfn_range(kvm, range->slot->as_id, range->start, |
1123 | range->end, range->may_block, flush); | |
063afacd BG |
1124 | } |
1125 | ||
3039bcc7 SC |
1126 | typedef bool (*tdp_handler_t)(struct kvm *kvm, struct tdp_iter *iter, |
1127 | struct kvm_gfn_range *range); | |
063afacd | 1128 | |
3039bcc7 SC |
1129 | static __always_inline bool kvm_tdp_mmu_handle_gfn(struct kvm *kvm, |
1130 | struct kvm_gfn_range *range, | |
1131 | tdp_handler_t handler) | |
063afacd | 1132 | { |
3039bcc7 SC |
1133 | struct kvm_mmu_page *root; |
1134 | struct tdp_iter iter; | |
1135 | bool ret = false; | |
1136 | ||
e1eed584 SC |
1137 | /* |
1138 | * Don't support rescheduling, none of the MMU notifiers that funnel | |
1139 | * into this helper allow blocking; it'd be dead, wasteful code. | |
1140 | */ | |
3039bcc7 | 1141 | for_each_tdp_mmu_root(kvm, root, range->slot->as_id) { |
a151acec SC |
1142 | rcu_read_lock(); |
1143 | ||
3039bcc7 SC |
1144 | tdp_root_for_each_leaf_pte(iter, root, range->start, range->end) |
1145 | ret |= handler(kvm, &iter, range); | |
3039bcc7 | 1146 | |
a151acec SC |
1147 | rcu_read_unlock(); |
1148 | } | |
3039bcc7 SC |
1149 | |
1150 | return ret; | |
063afacd | 1151 | } |
f8e14497 BG |
1152 | |
1153 | /* | |
1154 | * Mark the SPTEs range of GFNs [start, end) unaccessed and return non-zero | |
1155 | * if any of the GFNs in the range have been accessed. | |
1156 | */ | |
3039bcc7 SC |
1157 | static bool age_gfn_range(struct kvm *kvm, struct tdp_iter *iter, |
1158 | struct kvm_gfn_range *range) | |
f8e14497 | 1159 | { |
f8e14497 BG |
1160 | u64 new_spte = 0; |
1161 | ||
3039bcc7 SC |
1162 | /* If we have a non-accessed entry we don't need to change the pte. */ |
1163 | if (!is_accessed_spte(iter->old_spte)) | |
1164 | return false; | |
7cca2d0b | 1165 | |
3039bcc7 SC |
1166 | new_spte = iter->old_spte; |
1167 | ||
1168 | if (spte_ad_enabled(new_spte)) { | |
1169 | new_spte &= ~shadow_accessed_mask; | |
1170 | } else { | |
f8e14497 | 1171 | /* |
3039bcc7 SC |
1172 | * Capture the dirty status of the page, so that it doesn't get |
1173 | * lost when the SPTE is marked for access tracking. | |
f8e14497 | 1174 | */ |
3039bcc7 SC |
1175 | if (is_writable_pte(new_spte)) |
1176 | kvm_set_pfn_dirty(spte_to_pfn(new_spte)); | |
f8e14497 | 1177 | |
3039bcc7 | 1178 | new_spte = mark_spte_for_access_track(new_spte); |
f8e14497 BG |
1179 | } |
1180 | ||
3039bcc7 | 1181 | tdp_mmu_set_spte_no_acc_track(kvm, iter, new_spte); |
7cca2d0b | 1182 | |
3039bcc7 | 1183 | return true; |
f8e14497 BG |
1184 | } |
1185 | ||
3039bcc7 | 1186 | bool kvm_tdp_mmu_age_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range) |
f8e14497 | 1187 | { |
3039bcc7 | 1188 | return kvm_tdp_mmu_handle_gfn(kvm, range, age_gfn_range); |
f8e14497 BG |
1189 | } |
1190 | ||
3039bcc7 SC |
1191 | static bool test_age_gfn(struct kvm *kvm, struct tdp_iter *iter, |
1192 | struct kvm_gfn_range *range) | |
f8e14497 | 1193 | { |
3039bcc7 | 1194 | return is_accessed_spte(iter->old_spte); |
f8e14497 BG |
1195 | } |
1196 | ||
3039bcc7 | 1197 | bool kvm_tdp_mmu_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range) |
f8e14497 | 1198 | { |
3039bcc7 | 1199 | return kvm_tdp_mmu_handle_gfn(kvm, range, test_age_gfn); |
f8e14497 | 1200 | } |
1d8dd6b3 | 1201 | |
3039bcc7 SC |
1202 | static bool set_spte_gfn(struct kvm *kvm, struct tdp_iter *iter, |
1203 | struct kvm_gfn_range *range) | |
1d8dd6b3 | 1204 | { |
1d8dd6b3 | 1205 | u64 new_spte; |
7cca2d0b | 1206 | |
3039bcc7 SC |
1207 | /* Huge pages aren't expected to be modified without first being zapped. */ |
1208 | WARN_ON(pte_huge(range->pte) || range->start + 1 != range->end); | |
1d8dd6b3 | 1209 | |
3039bcc7 SC |
1210 | if (iter->level != PG_LEVEL_4K || |
1211 | !is_shadow_present_pte(iter->old_spte)) | |
1212 | return false; | |
1d8dd6b3 | 1213 | |
3039bcc7 SC |
1214 | /* |
1215 | * Note, when changing a read-only SPTE, it's not strictly necessary to | |
1216 | * zero the SPTE before setting the new PFN, but doing so preserves the | |
1217 | * invariant that the PFN of a present * leaf SPTE can never change. | |
1218 | * See __handle_changed_spte(). | |
1219 | */ | |
1220 | tdp_mmu_set_spte(kvm, iter, 0); | |
1d8dd6b3 | 1221 | |
3039bcc7 SC |
1222 | if (!pte_write(range->pte)) { |
1223 | new_spte = kvm_mmu_changed_pte_notifier_make_spte(iter->old_spte, | |
1224 | pte_pfn(range->pte)); | |
1d8dd6b3 | 1225 | |
3039bcc7 | 1226 | tdp_mmu_set_spte(kvm, iter, new_spte); |
1d8dd6b3 BG |
1227 | } |
1228 | ||
3039bcc7 | 1229 | return true; |
1d8dd6b3 BG |
1230 | } |
1231 | ||
3039bcc7 SC |
1232 | /* |
1233 | * Handle the changed_pte MMU notifier for the TDP MMU. | |
1234 | * data is a pointer to the new pte_t mapping the HVA specified by the MMU | |
1235 | * notifier. | |
1236 | * Returns non-zero if a flush is needed before releasing the MMU lock. | |
1237 | */ | |
1238 | bool kvm_tdp_mmu_set_spte_gfn(struct kvm *kvm, struct kvm_gfn_range *range) | |
1d8dd6b3 | 1239 | { |
93fa50f6 SC |
1240 | /* |
1241 | * No need to handle the remote TLB flush under RCU protection, the | |
1242 | * target SPTE _must_ be a leaf SPTE, i.e. cannot result in freeing a | |
1243 | * shadow page. See the WARN on pfn_changed in __handle_changed_spte(). | |
1244 | */ | |
1245 | return kvm_tdp_mmu_handle_gfn(kvm, range, set_spte_gfn); | |
1d8dd6b3 BG |
1246 | } |
1247 | ||
a6a0b05d | 1248 | /* |
bedd9195 DM |
1249 | * Remove write access from all SPTEs at or above min_level that map GFNs |
1250 | * [start, end). Returns true if an SPTE has been changed and the TLBs need to | |
1251 | * be flushed. | |
a6a0b05d BG |
1252 | */ |
1253 | static bool wrprot_gfn_range(struct kvm *kvm, struct kvm_mmu_page *root, | |
1254 | gfn_t start, gfn_t end, int min_level) | |
1255 | { | |
1256 | struct tdp_iter iter; | |
1257 | u64 new_spte; | |
1258 | bool spte_set = false; | |
1259 | ||
7cca2d0b BG |
1260 | rcu_read_lock(); |
1261 | ||
a6a0b05d BG |
1262 | BUG_ON(min_level > KVM_MAX_HUGEPAGE_LEVEL); |
1263 | ||
77aa6075 | 1264 | for_each_tdp_pte_min_level(iter, root, min_level, start, end) { |
24ae4cfa BG |
1265 | retry: |
1266 | if (tdp_mmu_iter_cond_resched(kvm, &iter, false, true)) | |
1af4a960 BG |
1267 | continue; |
1268 | ||
a6a0b05d | 1269 | if (!is_shadow_present_pte(iter.old_spte) || |
0f99ee2c BG |
1270 | !is_last_spte(iter.old_spte, iter.level) || |
1271 | !(iter.old_spte & PT_WRITABLE_MASK)) | |
a6a0b05d BG |
1272 | continue; |
1273 | ||
1274 | new_spte = iter.old_spte & ~PT_WRITABLE_MASK; | |
1275 | ||
3e72c791 | 1276 | if (tdp_mmu_set_spte_atomic(kvm, &iter, new_spte)) |
24ae4cfa | 1277 | goto retry; |
3255530a | 1278 | |
a6a0b05d | 1279 | spte_set = true; |
a6a0b05d | 1280 | } |
7cca2d0b BG |
1281 | |
1282 | rcu_read_unlock(); | |
a6a0b05d BG |
1283 | return spte_set; |
1284 | } | |
1285 | ||
1286 | /* | |
1287 | * Remove write access from all the SPTEs mapping GFNs in the memslot. Will | |
1288 | * only affect leaf SPTEs down to min_level. | |
1289 | * Returns true if an SPTE has been changed and the TLBs need to be flushed. | |
1290 | */ | |
269e9552 HM |
1291 | bool kvm_tdp_mmu_wrprot_slot(struct kvm *kvm, |
1292 | const struct kvm_memory_slot *slot, int min_level) | |
a6a0b05d BG |
1293 | { |
1294 | struct kvm_mmu_page *root; | |
a6a0b05d BG |
1295 | bool spte_set = false; |
1296 | ||
24ae4cfa | 1297 | lockdep_assert_held_read(&kvm->mmu_lock); |
a6a0b05d | 1298 | |
d62007ed | 1299 | for_each_valid_tdp_mmu_root_yield_safe(kvm, root, slot->as_id, true) |
a6a0b05d BG |
1300 | spte_set |= wrprot_gfn_range(kvm, root, slot->base_gfn, |
1301 | slot->base_gfn + slot->npages, min_level); | |
a6a0b05d BG |
1302 | |
1303 | return spte_set; | |
1304 | } | |
1305 | ||
a3fe5dbd DM |
1306 | static struct kvm_mmu_page *__tdp_mmu_alloc_sp_for_split(gfp_t gfp) |
1307 | { | |
1308 | struct kvm_mmu_page *sp; | |
1309 | ||
1310 | gfp |= __GFP_ZERO; | |
1311 | ||
1312 | sp = kmem_cache_alloc(mmu_page_header_cache, gfp); | |
1313 | if (!sp) | |
1314 | return NULL; | |
1315 | ||
1316 | sp->spt = (void *)__get_free_page(gfp); | |
1317 | if (!sp->spt) { | |
1318 | kmem_cache_free(mmu_page_header_cache, sp); | |
1319 | return NULL; | |
1320 | } | |
1321 | ||
1322 | return sp; | |
1323 | } | |
1324 | ||
1325 | static struct kvm_mmu_page *tdp_mmu_alloc_sp_for_split(struct kvm *kvm, | |
cb00a70b DM |
1326 | struct tdp_iter *iter, |
1327 | bool shared) | |
a3fe5dbd DM |
1328 | { |
1329 | struct kvm_mmu_page *sp; | |
1330 | ||
a3fe5dbd DM |
1331 | /* |
1332 | * Since we are allocating while under the MMU lock we have to be | |
1333 | * careful about GFP flags. Use GFP_NOWAIT to avoid blocking on direct | |
1334 | * reclaim and to avoid making any filesystem callbacks (which can end | |
1335 | * up invoking KVM MMU notifiers, resulting in a deadlock). | |
1336 | * | |
1337 | * If this allocation fails we drop the lock and retry with reclaim | |
1338 | * allowed. | |
1339 | */ | |
1340 | sp = __tdp_mmu_alloc_sp_for_split(GFP_NOWAIT | __GFP_ACCOUNT); | |
1341 | if (sp) | |
1342 | return sp; | |
1343 | ||
1344 | rcu_read_unlock(); | |
cb00a70b DM |
1345 | |
1346 | if (shared) | |
1347 | read_unlock(&kvm->mmu_lock); | |
1348 | else | |
1349 | write_unlock(&kvm->mmu_lock); | |
a3fe5dbd DM |
1350 | |
1351 | iter->yielded = true; | |
1352 | sp = __tdp_mmu_alloc_sp_for_split(GFP_KERNEL_ACCOUNT); | |
1353 | ||
cb00a70b DM |
1354 | if (shared) |
1355 | read_lock(&kvm->mmu_lock); | |
1356 | else | |
1357 | write_lock(&kvm->mmu_lock); | |
1358 | ||
a3fe5dbd DM |
1359 | rcu_read_lock(); |
1360 | ||
1361 | return sp; | |
1362 | } | |
1363 | ||
cb00a70b DM |
1364 | static int tdp_mmu_split_huge_page(struct kvm *kvm, struct tdp_iter *iter, |
1365 | struct kvm_mmu_page *sp, bool shared) | |
a3fe5dbd DM |
1366 | { |
1367 | const u64 huge_spte = iter->old_spte; | |
1368 | const int level = iter->level; | |
1369 | int ret, i; | |
1370 | ||
1371 | tdp_mmu_init_child_sp(sp, iter); | |
1372 | ||
1373 | /* | |
1374 | * No need for atomics when writing to sp->spt since the page table has | |
1375 | * not been linked in yet and thus is not reachable from any other CPU. | |
1376 | */ | |
1377 | for (i = 0; i < PT64_ENT_PER_PAGE; i++) | |
1378 | sp->spt[i] = make_huge_page_split_spte(huge_spte, level, i); | |
1379 | ||
1380 | /* | |
1381 | * Replace the huge spte with a pointer to the populated lower level | |
1382 | * page table. Since we are making this change without a TLB flush vCPUs | |
1383 | * will see a mix of the split mappings and the original huge mapping, | |
1384 | * depending on what's currently in their TLB. This is fine from a | |
1385 | * correctness standpoint since the translation will be the same either | |
1386 | * way. | |
1387 | */ | |
cb00a70b | 1388 | ret = tdp_mmu_link_sp(kvm, iter, sp, false, shared); |
a3fe5dbd | 1389 | if (ret) |
e0b728b1 | 1390 | goto out; |
a3fe5dbd DM |
1391 | |
1392 | /* | |
1393 | * tdp_mmu_link_sp_atomic() will handle subtracting the huge page we | |
1394 | * are overwriting from the page stats. But we have to manually update | |
1395 | * the page stats with the new present child pages. | |
1396 | */ | |
1397 | kvm_update_page_stats(kvm, level - 1, PT64_ENT_PER_PAGE); | |
1398 | ||
e0b728b1 DM |
1399 | out: |
1400 | trace_kvm_mmu_split_huge_page(iter->gfn, huge_spte, level, ret); | |
1401 | return ret; | |
a3fe5dbd DM |
1402 | } |
1403 | ||
1404 | static int tdp_mmu_split_huge_pages_root(struct kvm *kvm, | |
1405 | struct kvm_mmu_page *root, | |
1406 | gfn_t start, gfn_t end, | |
cb00a70b | 1407 | int target_level, bool shared) |
a3fe5dbd DM |
1408 | { |
1409 | struct kvm_mmu_page *sp = NULL; | |
1410 | struct tdp_iter iter; | |
1411 | int ret = 0; | |
1412 | ||
1413 | rcu_read_lock(); | |
1414 | ||
1415 | /* | |
1416 | * Traverse the page table splitting all huge pages above the target | |
1417 | * level into one lower level. For example, if we encounter a 1GB page | |
1418 | * we split it into 512 2MB pages. | |
1419 | * | |
1420 | * Since the TDP iterator uses a pre-order traversal, we are guaranteed | |
1421 | * to visit an SPTE before ever visiting its children, which means we | |
1422 | * will correctly recursively split huge pages that are more than one | |
1423 | * level above the target level (e.g. splitting a 1GB to 512 2MB pages, | |
1424 | * and then splitting each of those to 512 4KB pages). | |
1425 | */ | |
1426 | for_each_tdp_pte_min_level(iter, root, target_level + 1, start, end) { | |
1427 | retry: | |
cb00a70b | 1428 | if (tdp_mmu_iter_cond_resched(kvm, &iter, false, shared)) |
a3fe5dbd DM |
1429 | continue; |
1430 | ||
1431 | if (!is_shadow_present_pte(iter.old_spte) || !is_large_pte(iter.old_spte)) | |
1432 | continue; | |
1433 | ||
1434 | if (!sp) { | |
cb00a70b | 1435 | sp = tdp_mmu_alloc_sp_for_split(kvm, &iter, shared); |
a3fe5dbd DM |
1436 | if (!sp) { |
1437 | ret = -ENOMEM; | |
e0b728b1 DM |
1438 | trace_kvm_mmu_split_huge_page(iter.gfn, |
1439 | iter.old_spte, | |
1440 | iter.level, ret); | |
a3fe5dbd DM |
1441 | break; |
1442 | } | |
1443 | ||
1444 | if (iter.yielded) | |
1445 | continue; | |
1446 | } | |
1447 | ||
cb00a70b | 1448 | if (tdp_mmu_split_huge_page(kvm, &iter, sp, shared)) |
a3fe5dbd DM |
1449 | goto retry; |
1450 | ||
1451 | sp = NULL; | |
1452 | } | |
1453 | ||
1454 | rcu_read_unlock(); | |
1455 | ||
1456 | /* | |
1457 | * It's possible to exit the loop having never used the last sp if, for | |
1458 | * example, a vCPU doing HugePage NX splitting wins the race and | |
1459 | * installs its own sp in place of the last sp we tried to split. | |
1460 | */ | |
1461 | if (sp) | |
1462 | tdp_mmu_free_sp(sp); | |
1463 | ||
a3fe5dbd DM |
1464 | return ret; |
1465 | } | |
1466 | ||
cb00a70b | 1467 | |
a3fe5dbd DM |
1468 | /* |
1469 | * Try to split all huge pages mapped by the TDP MMU down to the target level. | |
1470 | */ | |
1471 | void kvm_tdp_mmu_try_split_huge_pages(struct kvm *kvm, | |
1472 | const struct kvm_memory_slot *slot, | |
1473 | gfn_t start, gfn_t end, | |
cb00a70b | 1474 | int target_level, bool shared) |
a3fe5dbd DM |
1475 | { |
1476 | struct kvm_mmu_page *root; | |
1477 | int r = 0; | |
1478 | ||
cb00a70b | 1479 | kvm_lockdep_assert_mmu_lock_held(kvm, shared); |
a3fe5dbd | 1480 | |
7c554d8e | 1481 | for_each_valid_tdp_mmu_root_yield_safe(kvm, root, slot->as_id, shared) { |
cb00a70b | 1482 | r = tdp_mmu_split_huge_pages_root(kvm, root, start, end, target_level, shared); |
a3fe5dbd | 1483 | if (r) { |
cb00a70b | 1484 | kvm_tdp_mmu_put_root(kvm, root, shared); |
a3fe5dbd DM |
1485 | break; |
1486 | } | |
1487 | } | |
1488 | } | |
1489 | ||
a6a0b05d BG |
1490 | /* |
1491 | * Clear the dirty status of all the SPTEs mapping GFNs in the memslot. If | |
1492 | * AD bits are enabled, this will involve clearing the dirty bit on each SPTE. | |
1493 | * If AD bits are not enabled, this will require clearing the writable bit on | |
1494 | * each SPTE. Returns true if an SPTE has been changed and the TLBs need to | |
1495 | * be flushed. | |
1496 | */ | |
1497 | static bool clear_dirty_gfn_range(struct kvm *kvm, struct kvm_mmu_page *root, | |
1498 | gfn_t start, gfn_t end) | |
1499 | { | |
1500 | struct tdp_iter iter; | |
1501 | u64 new_spte; | |
1502 | bool spte_set = false; | |
1503 | ||
7cca2d0b BG |
1504 | rcu_read_lock(); |
1505 | ||
a6a0b05d | 1506 | tdp_root_for_each_leaf_pte(iter, root, start, end) { |
24ae4cfa BG |
1507 | retry: |
1508 | if (tdp_mmu_iter_cond_resched(kvm, &iter, false, true)) | |
1af4a960 BG |
1509 | continue; |
1510 | ||
3354ef5a SC |
1511 | if (!is_shadow_present_pte(iter.old_spte)) |
1512 | continue; | |
1513 | ||
a6a0b05d BG |
1514 | if (spte_ad_need_write_protect(iter.old_spte)) { |
1515 | if (is_writable_pte(iter.old_spte)) | |
1516 | new_spte = iter.old_spte & ~PT_WRITABLE_MASK; | |
1517 | else | |
1518 | continue; | |
1519 | } else { | |
1520 | if (iter.old_spte & shadow_dirty_mask) | |
1521 | new_spte = iter.old_spte & ~shadow_dirty_mask; | |
1522 | else | |
1523 | continue; | |
1524 | } | |
1525 | ||
3e72c791 | 1526 | if (tdp_mmu_set_spte_atomic(kvm, &iter, new_spte)) |
24ae4cfa | 1527 | goto retry; |
3255530a | 1528 | |
a6a0b05d | 1529 | spte_set = true; |
a6a0b05d | 1530 | } |
7cca2d0b BG |
1531 | |
1532 | rcu_read_unlock(); | |
a6a0b05d BG |
1533 | return spte_set; |
1534 | } | |
1535 | ||
1536 | /* | |
1537 | * Clear the dirty status of all the SPTEs mapping GFNs in the memslot. If | |
1538 | * AD bits are enabled, this will involve clearing the dirty bit on each SPTE. | |
1539 | * If AD bits are not enabled, this will require clearing the writable bit on | |
1540 | * each SPTE. Returns true if an SPTE has been changed and the TLBs need to | |
1541 | * be flushed. | |
1542 | */ | |
269e9552 HM |
1543 | bool kvm_tdp_mmu_clear_dirty_slot(struct kvm *kvm, |
1544 | const struct kvm_memory_slot *slot) | |
a6a0b05d BG |
1545 | { |
1546 | struct kvm_mmu_page *root; | |
a6a0b05d BG |
1547 | bool spte_set = false; |
1548 | ||
24ae4cfa | 1549 | lockdep_assert_held_read(&kvm->mmu_lock); |
a6a0b05d | 1550 | |
d62007ed | 1551 | for_each_valid_tdp_mmu_root_yield_safe(kvm, root, slot->as_id, true) |
a6a0b05d BG |
1552 | spte_set |= clear_dirty_gfn_range(kvm, root, slot->base_gfn, |
1553 | slot->base_gfn + slot->npages); | |
a6a0b05d BG |
1554 | |
1555 | return spte_set; | |
1556 | } | |
1557 | ||
1558 | /* | |
1559 | * Clears the dirty status of all the 4k SPTEs mapping GFNs for which a bit is | |
1560 | * set in mask, starting at gfn. The given memslot is expected to contain all | |
1561 | * the GFNs represented by set bits in the mask. If AD bits are enabled, | |
1562 | * clearing the dirty status will involve clearing the dirty bit on each SPTE | |
1563 | * or, if AD bits are not enabled, clearing the writable bit on each SPTE. | |
1564 | */ | |
1565 | static void clear_dirty_pt_masked(struct kvm *kvm, struct kvm_mmu_page *root, | |
1566 | gfn_t gfn, unsigned long mask, bool wrprot) | |
1567 | { | |
1568 | struct tdp_iter iter; | |
1569 | u64 new_spte; | |
1570 | ||
7cca2d0b BG |
1571 | rcu_read_lock(); |
1572 | ||
a6a0b05d BG |
1573 | tdp_root_for_each_leaf_pte(iter, root, gfn + __ffs(mask), |
1574 | gfn + BITS_PER_LONG) { | |
1575 | if (!mask) | |
1576 | break; | |
1577 | ||
1578 | if (iter.level > PG_LEVEL_4K || | |
1579 | !(mask & (1UL << (iter.gfn - gfn)))) | |
1580 | continue; | |
1581 | ||
f1b3b06a BG |
1582 | mask &= ~(1UL << (iter.gfn - gfn)); |
1583 | ||
a6a0b05d BG |
1584 | if (wrprot || spte_ad_need_write_protect(iter.old_spte)) { |
1585 | if (is_writable_pte(iter.old_spte)) | |
1586 | new_spte = iter.old_spte & ~PT_WRITABLE_MASK; | |
1587 | else | |
1588 | continue; | |
1589 | } else { | |
1590 | if (iter.old_spte & shadow_dirty_mask) | |
1591 | new_spte = iter.old_spte & ~shadow_dirty_mask; | |
1592 | else | |
1593 | continue; | |
1594 | } | |
1595 | ||
1596 | tdp_mmu_set_spte_no_dirty_log(kvm, &iter, new_spte); | |
a6a0b05d | 1597 | } |
7cca2d0b BG |
1598 | |
1599 | rcu_read_unlock(); | |
a6a0b05d BG |
1600 | } |
1601 | ||
1602 | /* | |
1603 | * Clears the dirty status of all the 4k SPTEs mapping GFNs for which a bit is | |
1604 | * set in mask, starting at gfn. The given memslot is expected to contain all | |
1605 | * the GFNs represented by set bits in the mask. If AD bits are enabled, | |
1606 | * clearing the dirty status will involve clearing the dirty bit on each SPTE | |
1607 | * or, if AD bits are not enabled, clearing the writable bit on each SPTE. | |
1608 | */ | |
1609 | void kvm_tdp_mmu_clear_dirty_pt_masked(struct kvm *kvm, | |
1610 | struct kvm_memory_slot *slot, | |
1611 | gfn_t gfn, unsigned long mask, | |
1612 | bool wrprot) | |
1613 | { | |
1614 | struct kvm_mmu_page *root; | |
a6a0b05d | 1615 | |
531810ca | 1616 | lockdep_assert_held_write(&kvm->mmu_lock); |
a3f15bda | 1617 | for_each_tdp_mmu_root(kvm, root, slot->as_id) |
a6a0b05d | 1618 | clear_dirty_pt_masked(kvm, root, gfn, mask, wrprot); |
a6a0b05d BG |
1619 | } |
1620 | ||
14881998 | 1621 | /* |
87aa9ec9 BG |
1622 | * Clear leaf entries which could be replaced by large mappings, for |
1623 | * GFNs within the slot. | |
14881998 | 1624 | */ |
4b85c921 | 1625 | static void zap_collapsible_spte_range(struct kvm *kvm, |
14881998 | 1626 | struct kvm_mmu_page *root, |
4b85c921 | 1627 | const struct kvm_memory_slot *slot) |
14881998 | 1628 | { |
9eba50f8 SC |
1629 | gfn_t start = slot->base_gfn; |
1630 | gfn_t end = start + slot->npages; | |
14881998 BG |
1631 | struct tdp_iter iter; |
1632 | kvm_pfn_t pfn; | |
14881998 | 1633 | |
7cca2d0b BG |
1634 | rcu_read_lock(); |
1635 | ||
14881998 | 1636 | tdp_root_for_each_pte(iter, root, start, end) { |
2db6f772 | 1637 | retry: |
4b85c921 | 1638 | if (tdp_mmu_iter_cond_resched(kvm, &iter, false, true)) |
1af4a960 | 1639 | continue; |
1af4a960 | 1640 | |
14881998 | 1641 | if (!is_shadow_present_pte(iter.old_spte) || |
87aa9ec9 | 1642 | !is_last_spte(iter.old_spte, iter.level)) |
14881998 BG |
1643 | continue; |
1644 | ||
1645 | pfn = spte_to_pfn(iter.old_spte); | |
1646 | if (kvm_is_reserved_pfn(pfn) || | |
9eba50f8 SC |
1647 | iter.level >= kvm_mmu_max_mapping_level(kvm, slot, iter.gfn, |
1648 | pfn, PG_LEVEL_NUM)) | |
14881998 BG |
1649 | continue; |
1650 | ||
4b85c921 | 1651 | /* Note, a successful atomic zap also does a remote TLB flush. */ |
3e72c791 | 1652 | if (tdp_mmu_zap_spte_atomic(kvm, &iter)) |
2db6f772 | 1653 | goto retry; |
14881998 BG |
1654 | } |
1655 | ||
7cca2d0b | 1656 | rcu_read_unlock(); |
14881998 BG |
1657 | } |
1658 | ||
1659 | /* | |
1660 | * Clear non-leaf entries (and free associated page tables) which could | |
1661 | * be replaced by large mappings, for GFNs within the slot. | |
1662 | */ | |
4b85c921 SC |
1663 | void kvm_tdp_mmu_zap_collapsible_sptes(struct kvm *kvm, |
1664 | const struct kvm_memory_slot *slot) | |
14881998 BG |
1665 | { |
1666 | struct kvm_mmu_page *root; | |
14881998 | 1667 | |
2db6f772 | 1668 | lockdep_assert_held_read(&kvm->mmu_lock); |
14881998 | 1669 | |
d62007ed | 1670 | for_each_valid_tdp_mmu_root_yield_safe(kvm, root, slot->as_id, true) |
4b85c921 | 1671 | zap_collapsible_spte_range(kvm, root, slot); |
14881998 | 1672 | } |
46044f72 BG |
1673 | |
1674 | /* | |
1675 | * Removes write access on the last level SPTE mapping this GFN and unsets the | |
5fc3424f | 1676 | * MMU-writable bit to ensure future writes continue to be intercepted. |
46044f72 BG |
1677 | * Returns true if an SPTE was set and a TLB flush is needed. |
1678 | */ | |
1679 | static bool write_protect_gfn(struct kvm *kvm, struct kvm_mmu_page *root, | |
3ad93562 | 1680 | gfn_t gfn, int min_level) |
46044f72 BG |
1681 | { |
1682 | struct tdp_iter iter; | |
1683 | u64 new_spte; | |
1684 | bool spte_set = false; | |
1685 | ||
3ad93562 KZ |
1686 | BUG_ON(min_level > KVM_MAX_HUGEPAGE_LEVEL); |
1687 | ||
7cca2d0b BG |
1688 | rcu_read_lock(); |
1689 | ||
77aa6075 | 1690 | for_each_tdp_pte_min_level(iter, root, min_level, gfn, gfn + 1) { |
3ad93562 KZ |
1691 | if (!is_shadow_present_pte(iter.old_spte) || |
1692 | !is_last_spte(iter.old_spte, iter.level)) | |
1693 | continue; | |
1694 | ||
46044f72 | 1695 | new_spte = iter.old_spte & |
5fc3424f | 1696 | ~(PT_WRITABLE_MASK | shadow_mmu_writable_mask); |
46044f72 | 1697 | |
7c8a4742 DM |
1698 | if (new_spte == iter.old_spte) |
1699 | break; | |
1700 | ||
46044f72 BG |
1701 | tdp_mmu_set_spte(kvm, &iter, new_spte); |
1702 | spte_set = true; | |
1703 | } | |
1704 | ||
7cca2d0b BG |
1705 | rcu_read_unlock(); |
1706 | ||
46044f72 BG |
1707 | return spte_set; |
1708 | } | |
1709 | ||
1710 | /* | |
1711 | * Removes write access on the last level SPTE mapping this GFN and unsets the | |
5fc3424f | 1712 | * MMU-writable bit to ensure future writes continue to be intercepted. |
46044f72 BG |
1713 | * Returns true if an SPTE was set and a TLB flush is needed. |
1714 | */ | |
1715 | bool kvm_tdp_mmu_write_protect_gfn(struct kvm *kvm, | |
3ad93562 KZ |
1716 | struct kvm_memory_slot *slot, gfn_t gfn, |
1717 | int min_level) | |
46044f72 BG |
1718 | { |
1719 | struct kvm_mmu_page *root; | |
46044f72 BG |
1720 | bool spte_set = false; |
1721 | ||
531810ca | 1722 | lockdep_assert_held_write(&kvm->mmu_lock); |
a3f15bda | 1723 | for_each_tdp_mmu_root(kvm, root, slot->as_id) |
3ad93562 | 1724 | spte_set |= write_protect_gfn(kvm, root, gfn, min_level); |
a3f15bda | 1725 | |
46044f72 BG |
1726 | return spte_set; |
1727 | } | |
1728 | ||
95fb5b02 BG |
1729 | /* |
1730 | * Return the level of the lowest level SPTE added to sptes. | |
1731 | * That SPTE may be non-present. | |
c5c8c7c5 DM |
1732 | * |
1733 | * Must be called between kvm_tdp_mmu_walk_lockless_{begin,end}. | |
95fb5b02 | 1734 | */ |
39b4d43e SC |
1735 | int kvm_tdp_mmu_get_walk(struct kvm_vcpu *vcpu, u64 addr, u64 *sptes, |
1736 | int *root_level) | |
95fb5b02 BG |
1737 | { |
1738 | struct tdp_iter iter; | |
1739 | struct kvm_mmu *mmu = vcpu->arch.mmu; | |
95fb5b02 | 1740 | gfn_t gfn = addr >> PAGE_SHIFT; |
2aa07893 | 1741 | int leaf = -1; |
95fb5b02 | 1742 | |
39b4d43e | 1743 | *root_level = vcpu->arch.mmu->shadow_root_level; |
95fb5b02 BG |
1744 | |
1745 | tdp_mmu_for_each_pte(iter, mmu, gfn, gfn + 1) { | |
1746 | leaf = iter.level; | |
dde81f94 | 1747 | sptes[leaf] = iter.old_spte; |
95fb5b02 BG |
1748 | } |
1749 | ||
1750 | return leaf; | |
1751 | } | |
6e8eb206 DM |
1752 | |
1753 | /* | |
1754 | * Returns the last level spte pointer of the shadow page walk for the given | |
1755 | * gpa, and sets *spte to the spte value. This spte may be non-preset. If no | |
1756 | * walk could be performed, returns NULL and *spte does not contain valid data. | |
1757 | * | |
1758 | * Contract: | |
1759 | * - Must be called between kvm_tdp_mmu_walk_lockless_{begin,end}. | |
1760 | * - The returned sptep must not be used after kvm_tdp_mmu_walk_lockless_end. | |
1761 | * | |
1762 | * WARNING: This function is only intended to be called during fast_page_fault. | |
1763 | */ | |
1764 | u64 *kvm_tdp_mmu_fast_pf_get_last_sptep(struct kvm_vcpu *vcpu, u64 addr, | |
1765 | u64 *spte) | |
1766 | { | |
1767 | struct tdp_iter iter; | |
1768 | struct kvm_mmu *mmu = vcpu->arch.mmu; | |
1769 | gfn_t gfn = addr >> PAGE_SHIFT; | |
1770 | tdp_ptep_t sptep = NULL; | |
1771 | ||
1772 | tdp_mmu_for_each_pte(iter, mmu, gfn, gfn + 1) { | |
1773 | *spte = iter.old_spte; | |
1774 | sptep = iter.sptep; | |
1775 | } | |
1776 | ||
1777 | /* | |
1778 | * Perform the rcu_dereference to get the raw spte pointer value since | |
1779 | * we are passing it up to fast_page_fault, which is shared with the | |
1780 | * legacy MMU and thus does not retain the TDP MMU-specific __rcu | |
1781 | * annotation. | |
1782 | * | |
1783 | * This is safe since fast_page_fault obeys the contracts of this | |
1784 | * function as well as all TDP MMU contracts around modifying SPTEs | |
1785 | * outside of mmu_lock. | |
1786 | */ | |
1787 | return rcu_dereference(sptep); | |
1788 | } |