1 // SPDX-License-Identifier: GPL-2.0
2
3 #ifndef __KVM_X86_MMU_TDP_ITER_H
4 #define __KVM_X86_MMU_TDP_ITER_H
5
6 #include <linux/kvm_host.h>
7
8 #include "mmu.h"
9 #include "spte.h"
10
11 /*
12 * TDP MMU SPTEs are RCU protected to allow paging structures (non-leaf SPTEs)
13 * to be zapped while holding mmu_lock for read, and to allow TLB flushes to be
14 * batched without having to collect the list of zapped SPs. Flows that can
15 * remove SPs must service pending TLB flushes prior to dropping RCU protection.
16 */
kvm_tdp_mmu_read_spte(tdp_ptep_t sptep)17 static inline u64 kvm_tdp_mmu_read_spte(tdp_ptep_t sptep)
18 {
19 return READ_ONCE(*rcu_dereference(sptep));
20 }
21
kvm_tdp_mmu_write_spte_atomic(tdp_ptep_t sptep,u64 new_spte)22 static inline u64 kvm_tdp_mmu_write_spte_atomic(tdp_ptep_t sptep, u64 new_spte)
23 {
24 return xchg(rcu_dereference(sptep), new_spte);
25 }
26
__kvm_tdp_mmu_write_spte(tdp_ptep_t sptep,u64 new_spte)27 static inline void __kvm_tdp_mmu_write_spte(tdp_ptep_t sptep, u64 new_spte)
28 {
29 WRITE_ONCE(*rcu_dereference(sptep), new_spte);
30 }
31
32 /*
33 * SPTEs must be modified atomically if they are shadow-present, leaf
34 * SPTEs, and have volatile bits, i.e. has bits that can be set outside
35 * of mmu_lock. The Writable bit can be set by KVM's fast page fault
36 * handler, and Accessed and Dirty bits can be set by the CPU.
37 *
38 * Note, non-leaf SPTEs do have Accessed bits and those bits are
39 * technically volatile, but KVM doesn't consume the Accessed bit of
40 * non-leaf SPTEs, i.e. KVM doesn't care if it clobbers the bit. This
41 * logic needs to be reassessed if KVM were to use non-leaf Accessed
42 * bits, e.g. to skip stepping down into child SPTEs when aging SPTEs.
43 */
kvm_tdp_mmu_spte_need_atomic_write(u64 old_spte,int level)44 static inline bool kvm_tdp_mmu_spte_need_atomic_write(u64 old_spte, int level)
45 {
46 return is_shadow_present_pte(old_spte) &&
47 is_last_spte(old_spte, level) &&
48 spte_has_volatile_bits(old_spte);
49 }
50
kvm_tdp_mmu_write_spte(tdp_ptep_t sptep,u64 old_spte,u64 new_spte,int level)51 static inline u64 kvm_tdp_mmu_write_spte(tdp_ptep_t sptep, u64 old_spte,
52 u64 new_spte, int level)
53 {
54 if (kvm_tdp_mmu_spte_need_atomic_write(old_spte, level))
55 return kvm_tdp_mmu_write_spte_atomic(sptep, new_spte);
56
57 __kvm_tdp_mmu_write_spte(sptep, new_spte);
58 return old_spte;
59 }
60
tdp_mmu_clear_spte_bits(tdp_ptep_t sptep,u64 old_spte,u64 mask,int level)61 static inline u64 tdp_mmu_clear_spte_bits(tdp_ptep_t sptep, u64 old_spte,
62 u64 mask, int level)
63 {
64 atomic64_t *sptep_atomic;
65
66 if (kvm_tdp_mmu_spte_need_atomic_write(old_spte, level)) {
67 sptep_atomic = (atomic64_t *)rcu_dereference(sptep);
68 return (u64)atomic64_fetch_and(~mask, sptep_atomic);
69 }
70
71 __kvm_tdp_mmu_write_spte(sptep, old_spte & ~mask);
72 return old_spte;
73 }
74
75 /*
76 * A TDP iterator performs a pre-order walk over a TDP paging structure.
77 */
78 struct tdp_iter {
79 /*
80 * The iterator will traverse the paging structure towards the mapping
81 * for this GFN.
82 */
83 gfn_t next_last_level_gfn;
84 /*
85 * The next_last_level_gfn at the time when the thread last
86 * yielded. Only yielding when the next_last_level_gfn !=
87 * yielded_gfn helps ensure forward progress.
88 */
89 gfn_t yielded_gfn;
90 /* Pointers to the page tables traversed to reach the current SPTE */
91 tdp_ptep_t pt_path[PT64_ROOT_MAX_LEVEL];
92 /* A pointer to the current SPTE */
93 tdp_ptep_t sptep;
94 /* The lowest GFN mapped by the current SPTE */
95 gfn_t gfn;
96 /* The level of the root page given to the iterator */
97 int root_level;
98 /* The lowest level the iterator should traverse to */
99 int min_level;
100 /* The iterator's current level within the paging structure */
101 int level;
102 /* The address space ID, i.e. SMM vs. regular. */
103 int as_id;
104 /* A snapshot of the value at sptep */
105 u64 old_spte;
106 /*
107 * Whether the iterator has a valid state. This will be false if the
108 * iterator walks off the end of the paging structure.
109 */
110 bool valid;
111 /*
112 * True if KVM dropped mmu_lock and yielded in the middle of a walk, in
113 * which case tdp_iter_next() needs to restart the walk at the root
114 * level instead of advancing to the next entry.
115 */
116 bool yielded;
117 };
118
119 /*
120 * Iterates over every SPTE mapping the GFN range [start, end) in a
121 * preorder traversal.
122 */
123 #define for_each_tdp_pte_min_level(iter, root, min_level, start, end) \
124 for (tdp_iter_start(&iter, root, min_level, start); \
125 iter.valid && iter.gfn < end; \
126 tdp_iter_next(&iter))
127
128 #define for_each_tdp_pte(iter, root, start, end) \
129 for_each_tdp_pte_min_level(iter, root, PG_LEVEL_4K, start, end)
130
131 tdp_ptep_t spte_to_child_pt(u64 pte, int level);
132
133 void tdp_iter_start(struct tdp_iter *iter, struct kvm_mmu_page *root,
134 int min_level, gfn_t next_last_level_gfn);
135 void tdp_iter_next(struct tdp_iter *iter);
136 void tdp_iter_restart(struct tdp_iter *iter);
137
138 #endif /* __KVM_X86_MMU_TDP_ITER_H */
139