KVM: arm64: Reload PTE after invoking walker callback on preorder traversal
PATCH 1/1 KVM: arm64: Reload PTE after invoking walker callback on preorder traversal
The preorder callback on the kvm_pgtable_stage2_map() path can replace a table with a block, then recursively free the detached table. The higher-level walking logic stashes the old page table entry and then walks the freed table, invoking the leaf callback and potentially freeing pgtable pages prematurely.
In normal operation, the call to tear down the detached stage-2 is indirected and uses an RCU callback to trigger the freeing. RCU is not available to pKVM, which is where this bug is triggered.
Change the behavior of the walker to reload the page table entry after invoking the walker callback on preorder traversal, as it does for leaf entries.
Tested on Pixel 6.
---
Based on: f1fcbaa18b28 (6.4-rc2)
The bug can be triggered by applying Will's FFA series [1] to android mainline [2] and booting a Pixel 6 in protected mode (pKVM).
[1] [email protected]
[2] https://android.googlesource.com/kernel/common/+/refs/tags/android-mainline-6.3
--- arch/arm64/include/asm/kvm_pgtable.h | 6 +++--- arch/arm64/kvm/hyp/pgtable.c | 14 +++++++++++++- 2 files changed, 16 insertions(+), 4 deletions(-)
arch/arm64/include/asm/kvm_pgtable.h +3 -3
@@ -631,9 +631,9 @@ int kvm_pgtable_stage2_flush(struct kvm_pgtable *pgt, u64 addr, u64 size);* * The walker will walk the page-table entries corresponding to the input * address range specified, visiting entries according to the walker flags. -* Invalid entries are treated as leaf entries. Leaf entries are reloaded -* after invoking the walker callback, allowing the walker to descend into -* a newly installed table. +* Invalid entries are treated as leaf entries. The visited page table entry is +* reloaded after invoking the walker callback, allowing the walker to descend +* into a newly installed table. * * Returning a negative error code from the walker callback function will * terminate the walk immediately with the same error code.
arch/arm64/kvm/hyp/pgtable.c +13 -1
@@ -207,14 +207,26 @@ static inline int __kvm_pgtable_visit(struct kvm_pgtable_walk_data *data, .flags = flags, };int ret =0 ; +bool reload = false;kvm_pteref_t childp;bool table =kvm_pte_table (ctx.old, level); -if (table && (ctx.flags &KVM_PGTABLE_WALK_TABLE_PRE )) +if (table && (ctx.flags &KVM_PGTABLE_WALK_TABLE_PRE )) { ret =kvm_pgtable_visitor_cb (data, &ctx,KVM_PGTABLE_WALK_TABLE_PRE ); + reload = true; + }if (!table && (ctx.flags &KVM_PGTABLE_WALK_LEAF )) { ret =kvm_pgtable_visitor_cb (data, &ctx,KVM_PGTABLE_WALK_LEAF ); + reload = true; + } + +/* +* Reload the page table after invoking the walker callback for leaf +* entries or after pre-order traversal, to allow the walker to descend +* into a newly installed or replaced table. +*/ +if (reload) { ctx.old =READ_ONCE (*ptep); table =kvm_pte_table (ctx.old, level); }
base-commit: f1fcbaa18b28dec10281551dfe6ed3a3ed80e3d6
Signature · 1 line
2.40.1.698.g37aff9b760-goog
Marc Zyngier
maintainerHi Fuad,
Quoted 7 lines · Fuad Tabba
The preorder callback on the
kvm_pgtable_stage2_map()path can replace a table with a block, then recursively free the detached table. The higher-level walking logic stashes the old page table entry and then walks the freed table, invoking the leaf callback and potentially freeing pgtable pages prematurely.
Quoted 8 lines
In normal operation, the call to tear down the detached stage-2 is indirected and uses an RCU callback to trigger the freeing. RCU is not available to pKVM, which is where this bug is triggered.
Change the behavior of the walker to reload the page table entry after invoking the walker callback on preorder traversal, as it does for leaf entries.
Thanks for the fix and the detailed explanation. A couple of nits, none of which deserve a respin on their own (I can fix up things when applying the patch).
Quoted 5 lines
Tested on Pixel 6.
Fixes 5c359cca1faf ("KVM: arm64: Tear down unlinked stage-2 subtree after break-before-make")
Spurious empty line. In general, please keep the trailers grouped together, as it otherwise tends to confuse git-interpret-trailers.
Quoted 47 lines
Suggested-by Oliver Upton <[email protected]>Signed-off-by Fuad Tabba <[email protected]>---
Based on:
f1fcbaa18b28(6.4-rc2)The bug can be triggered by applying Will's FFA series [1] to android mainline [2] and booting a Pixel 6 in protected mode (pKVM).
[1] [email protected]
[2] https://android.googlesource.com/kernel/common/+/refs/tags/android-mainline-6.3--- arch/arm64/include/asm/kvm_pgtable.h | 6 +++--- arch/arm64/kvm/hyp/pgtable.c | 14 +++++++++++++- 2 files changed, 16 insertions(+), 4 deletions(-)arch/arm64/include/asm/kvm_pgtable.h +3
-3@@ -631,9 +631,9 @@ int kvm_pgtable_stage2_flush(struct kvm_pgtable *pgt, u64 addr, u64 size);* * The walker will walk the page-table entries corresponding to the input * address range specified, visiting entries according to the walker flags. -* Invalid entries are treated as leaf entries. Leaf entries are reloaded -* after invoking the walker callback, allowing the walker to descend into -* a newly installed table. +* Invalid entries are treated as leaf entries. The visited page table entry is +* reloaded after invoking the walker callback, allowing the walker to descend +* into a newly installed table. * * Returning a negative error code from the walker callback function will * terminate the walk immediately with the same error code. arch/arm64/kvm/hyp/pgtable.c +1
-0@@ -207,14 +207,26 @@ static inline int __kvm_pgtable_visit(struct kvm_pgtable_walk_data *data, .flags = flags, };int ret =0 ; +bool reload = false;kvm_pteref_t childp;bool table =kvm_pte_table (ctx.old, level);
arch/arm64/kvm/hyp/pgtable.c static inline int __kvm_pgtable_visit(struct kvm_pgtable_walk_data *data, · line 222
Quoted 10 lines
arch/arm64/kvm/hyp/pgtable.c +5
-1-if (table && (ctx.flags &KVM_PGTABLE_WALK_TABLE_PRE )) +if (table && (ctx.flags &KVM_PGTABLE_WALK_TABLE_PRE )) { ret =kvm_pgtable_visitor_cb (data, &ctx,KVM_PGTABLE_WALK_TABLE_PRE ); + reload = true; + }if (!table && (ctx.flags &KVM_PGTABLE_WALK_LEAF )) { ret =kvm_pgtable_visitor_cb (data, &ctx,KVM_PGTABLE_WALK_LEAF ); + reload = true; + }
From these two clauses, it is clear that reload is always the value of (ctx.flags & KVM_PGTABLE_WALK_LEAF). That'd simplify the patch a bit.
arch/arm64/kvm/hyp/pgtable.c static inline int __kvm_pgtable_visit(struct kvm_pgtable_walk_data *data, · line 232
Quoted 11 lines
arch/arm64/kvm/hyp/pgtable.c +7
-0+ +/* +* Reload the page table after invoking the walker callback for leaf +* entries or after pre-order traversal, to allow the walker to descend +* into a newly installed or replaced table. +*/ +if (reload) { ctx.old =READ_ONCE (*ptep); table =kvm_pte_table (ctx.old, level); }
Thanks,
M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
Marc Zyngier
maintainerQuoted 19 lines · Marc Zyngier
Quoted 19 lines
arch/arm64/kvm/hyp/pgtable.c +4
-1@@ -207,14 +207,26 @@ static inline int __kvm_pgtable_visit(struct kvm_pgtable_walk_data *data, .flags = flags, };int ret =0 ; +bool reload = false;kvm_pteref_t childp;bool table =kvm_pte_table (ctx.old, level); -if (table && (ctx.flags &KVM_PGTABLE_WALK_TABLE_PRE )) +if (table && (ctx.flags &KVM_PGTABLE_WALK_TABLE_PRE )) { ret =kvm_pgtable_visitor_cb (data, &ctx,KVM_PGTABLE_WALK_TABLE_PRE ); + reload = true; + }
Quoted 7 lines
arch/arm64/kvm/hyp/pgtable.c static inline int __kvm_pgtable_visit(struct kvm_pgtable_walk_data *data, · line 222
Quoted 4 lines
arch/arm64/kvm/hyp/pgtable.c +2
-0if (!table && (ctx.flags &KVM_PGTABLE_WALK_LEAF )) { ret =kvm_pgtable_visitor_cb (data, &ctx,KVM_PGTABLE_WALK_LEAF ); + reload = true; + }From these two clauses, it is clear that reload is always the value of (ctx.flags &
KVM_PGTABLE_WALK_LEAF). That'd simplify the patch a bit.
OK, it should be obvious by now that I cannot read, because TABLE_PRE and LEAF are very different things.
Sorry about the noise. I'll fix the trailer thing and stay quiet for the rest of the day...
M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
Quoted 30 lines · Marc Zyngier
Quoted 19 lines · Marc Zyngier
Quoted 19 lines
arch/arm64/kvm/hyp/pgtable.c +4
-1@@ -207,14 +207,26 @@ static inline int __kvm_pgtable_visit(struct kvm_pgtable_walk_data *data, .flags = flags, };int ret =0 ; +bool reload = false;kvm_pteref_t childp;bool table =kvm_pte_table (ctx.old, level); -if (table && (ctx.flags &KVM_PGTABLE_WALK_TABLE_PRE )) +if (table && (ctx.flags &KVM_PGTABLE_WALK_TABLE_PRE )) { ret =kvm_pgtable_visitor_cb (data, &ctx,KVM_PGTABLE_WALK_TABLE_PRE ); + reload = true; + }Quoted 7 lines
arch/arm64/kvm/hyp/pgtable.c static inline int __kvm_pgtable_visit(struct kvm_pgtable_walk_data *data, · line 222
Quoted 4 lines
arch/arm64/kvm/hyp/pgtable.c +2
-0if (!table && (ctx.flags &KVM_PGTABLE_WALK_LEAF )) { ret =kvm_pgtable_visitor_cb (data, &ctx,KVM_PGTABLE_WALK_LEAF ); + reload = true; + }From these two clauses, it is clear that reload is always the value of (ctx.flags &
KVM_PGTABLE_WALK_LEAF). That'd simplify the patch a bit.
Quoted 5 lines
OK, it should be obvious by now that I cannot read, because
TABLE_PREand LEAF are very different things.Sorry about the noise. I'll fix the trailer thing and stay quiet for the rest of the day...
:)
Thanks Marc!
/fuad
Quoted 5 lines
M.--
Without deviation from the norm, progress is not possible.
Quoted 6 lines · Fuad Tabba
The preorder callback on the
kvm_pgtable_stage2_map()path can replace a table with a block, then recursively free the detached table. The higher-level walking logic stashes the old page table entry and then walks the freed table, invoking the leaf callback and potentially freeing pgtable pages prematurely.
Quoted 6 lines
In normal operation, the call to tear down the detached stage-2 is indirected and uses an RCU callback to trigger the freeing. RCU is not available to pKVM, which is where this bug is triggered.
[...]
Applied to fixes, thanks!
[1/1] KVM: arm64: Reload PTE after invoking walker callback on preorder traversal
commit: a9f0e3d5a089d0844abb679a5e99f15010d53e25
Cheers,
M.
Signature · 1 line
Without deviation from the norm, progress is not possible.