Popuplate the shadow for the shared portion of the CPU entry area, i.e.
the read-only IDT mapping, during KASAN initialization. A recent change
modified KASAN to map the per-CPU areas on-demand, but forgot to keep a
shadow for the common area that is shared amongst all CPUs.
Map the common area in KASAN init instead of letting idt_map_in_cea() do
the dirty work so that it Just Works in the unlikely event more shared
data is shoved into the CPU entry area.
The bug manifests as a not-present #PF when software attempts to lookup
an IDT entry, e.g. when KVM is handling IRQs on Intel CPUs (KVM performs
direct CALL to the IRQ handler to avoid the overhead of INTn):
BUG: unable to handle page fault for address: fffffbc0000001d8
#PF: supervisor read access in kernel mode
#PF: error_code(0x0000) - not-present page
PGD 16c03a067 P4D 16c03a067 PUD 0
Oops: 0000 [#1] PREEMPT SMP KASAN
CPU: 5 PID: 901 Comm: repro Tainted: G W 6.1.0-rc3+ #410
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
RIP: 0010:kasan_check_range+0xdf/0x190
vmx_handle_exit_irqoff+0x152/0x290 [kvm_intel]
vcpu_run+0x1d89/0x2bd0 [kvm]
kvm_arch_vcpu_ioctl_run+0x3ce/0xa70 [kvm]
kvm_vcpu_ioctl+0x349/0x900 [kvm]
__x64_sys_ioctl+0xb8/0xf0
do_syscall_64+0x2b/0x50
entry_SYSCALL_64_after_hwframe+0x46/0xb0
Fixes: 9fd429c28073 ("x86/kasan: Map shadow for percpu pages on demand")
Reported-by: [email protected]
Cc: Andrey Ryabinin <[email protected]>
Signed-off-by: Sean Christopherson <[email protected]>
---
arch/x86/mm/kasan_init_64.c | 12 +++++++++++-
1 file changed, 11 insertions(+), 1 deletion(-)
diff --git a/arch/x86/mm/kasan_init_64.c b/arch/x86/mm/kasan_init_64.c
index afc5e129ca7b..0302491d799d 100644
--- a/arch/x86/mm/kasan_init_64.c
+++ b/arch/x86/mm/kasan_init_64.c
@@ -341,7 +341,7 @@ void __init kasan_populate_shadow_for_vaddr(void *va, size_t size, int nid)
void __init kasan_init(void)
{
- unsigned long shadow_cea_begin, shadow_cea_end;
+ unsigned long shadow_cea_begin, shadow_cea_per_cpu_begin, shadow_cea_end;
int i;
memcpy(early_top_pgt, init_top_pgt, sizeof(early_top_pgt));
@@ -384,6 +384,7 @@ void __init kasan_init(void)
}
shadow_cea_begin = kasan_mem_to_shadow_align_down(CPU_ENTRY_AREA_BASE);
+ shadow_cea_per_cpu_begin = kasan_mem_to_shadow_align_up(CPU_ENTRY_AREA_PER_CPU);
shadow_cea_end = kasan_mem_to_shadow_align_up(CPU_ENTRY_AREA_BASE +
CPU_ENTRY_AREA_MAP_SIZE);
@@ -409,6 +410,15 @@ void __init kasan_init(void)
kasan_mem_to_shadow((void *)VMALLOC_END + 1),
(void *)shadow_cea_begin);
+ /*
+ * Populate the shadow for the shared portion of the CPU entry area.
+ * Shadows for the per-CPU areas are mapped on-demand, as each CPU's
+ * area is randomly placed somewhere in the 512GiB range and mapping
+ * the entire 512GiB range is prohibitively expensive.
+ */
+ kasan_populate_shadow(shadow_cea_begin,
+ shadow_cea_per_cpu_begin, 0);
+
kasan_populate_early_shadow((void *)shadow_cea_end,
kasan_mem_to_shadow((void *)__START_KERNEL_map));
--
2.38.1.431.g37b22c650d-goog
On 11/4/22 21:32, Sean Christopherson wrote:
> Popuplate the shadow for the shared portion of the CPU entry area, i.e.
> the read-only IDT mapping, during KASAN initialization. A recent change
> modified KASAN to map the per-CPU areas on-demand, but forgot to keep a
> shadow for the common area that is shared amongst all CPUs.
>
> Map the common area in KASAN init instead of letting idt_map_in_cea() do
> the dirty work so that it Just Works in the unlikely event more shared
> data is shoved into the CPU entry area.
>
> The bug manifests as a not-present #PF when software attempts to lookup
> an IDT entry, e.g. when KVM is handling IRQs on Intel CPUs (KVM performs
> direct CALL to the IRQ handler to avoid the overhead of INTn):
>
> BUG: unable to handle page fault for address: fffffbc0000001d8
> #PF: supervisor read access in kernel mode
> #PF: error_code(0x0000) - not-present page
> PGD 16c03a067 P4D 16c03a067 PUD 0
> Oops: 0000 [#1] PREEMPT SMP KASAN
> CPU: 5 PID: 901 Comm: repro Tainted: G W 6.1.0-rc3+ #410
> Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
> RIP: 0010:kasan_check_range+0xdf/0x190
> vmx_handle_exit_irqoff+0x152/0x290 [kvm_intel]
> vcpu_run+0x1d89/0x2bd0 [kvm]
> kvm_arch_vcpu_ioctl_run+0x3ce/0xa70 [kvm]
> kvm_vcpu_ioctl+0x349/0x900 [kvm]
> __x64_sys_ioctl+0xb8/0xf0
> do_syscall_64+0x2b/0x50
> entry_SYSCALL_64_after_hwframe+0x46/0xb0
>
> Fixes: 9fd429c28073 ("x86/kasan: Map shadow for percpu pages on demand")
> Reported-by: [email protected]
> Cc: Andrey Ryabinin <[email protected]>
> Signed-off-by: Sean Christopherson <[email protected]>
> ---
> arch/x86/mm/kasan_init_64.c | 12 +++++++++++-
> 1 file changed, 11 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/mm/kasan_init_64.c b/arch/x86/mm/kasan_init_64.c
> index afc5e129ca7b..0302491d799d 100644
> --- a/arch/x86/mm/kasan_init_64.c
> +++ b/arch/x86/mm/kasan_init_64.c
> @@ -341,7 +341,7 @@ void __init kasan_populate_shadow_for_vaddr(void *va, size_t size, int nid)
>
> void __init kasan_init(void)
> {
> - unsigned long shadow_cea_begin, shadow_cea_end;
> + unsigned long shadow_cea_begin, shadow_cea_per_cpu_begin, shadow_cea_end;
> int i;
>
> memcpy(early_top_pgt, init_top_pgt, sizeof(early_top_pgt));
> @@ -384,6 +384,7 @@ void __init kasan_init(void)
> }
>
> shadow_cea_begin = kasan_mem_to_shadow_align_down(CPU_ENTRY_AREA_BASE);
> + shadow_cea_per_cpu_begin = kasan_mem_to_shadow_align_up(CPU_ENTRY_AREA_PER_CPU);
> shadow_cea_end = kasan_mem_to_shadow_align_up(CPU_ENTRY_AREA_BASE +
> CPU_ENTRY_AREA_MAP_SIZE);
>
> @@ -409,6 +410,15 @@ void __init kasan_init(void)
> kasan_mem_to_shadow((void *)VMALLOC_END + 1),
> (void *)shadow_cea_begin);
>
> + /*
> + * Populate the shadow for the shared portion of the CPU entry area.
> + * Shadows for the per-CPU areas are mapped on-demand, as each CPU's
> + * area is randomly placed somewhere in the 512GiB range and mapping
> + * the entire 512GiB range is prohibitively expensive.
> + */
> + kasan_populate_shadow(shadow_cea_begin,
> + shadow_cea_per_cpu_begin, 0);
> +
I think we can extend the kasan_populate_early_shadow() call above up to
shadow_cea_per_cpu_begin point, instead of this.
populate_early_shadow() maps single RO zeroed page. No one should write to the shadow for IDT.
KASAN only needs writable shadow for linear mapping/stacks/vmalloc/global variables.
> kasan_populate_early_shadow((void *)shadow_cea_end,
> kasan_mem_to_shadow((void *)__START_KERNEL_map));
>
On 11/8/22 23:03, Sean Christopherson wrote:
> On Tue, Nov 08, 2022, Andrey Ryabinin wrote:
>>
>> On 11/4/22 21:32, Sean Christopherson wrote:
>>> @@ -409,6 +410,15 @@ void __init kasan_init(void)
>>> kasan_mem_to_shadow((void *)VMALLOC_END + 1),
>>> (void *)shadow_cea_begin);
>>>
>>> + /*
>>> + * Populate the shadow for the shared portion of the CPU entry area.
>>> + * Shadows for the per-CPU areas are mapped on-demand, as each CPU's
>>> + * area is randomly placed somewhere in the 512GiB range and mapping
>>> + * the entire 512GiB range is prohibitively expensive.
>>> + */
>>> + kasan_populate_shadow(shadow_cea_begin,
>>> + shadow_cea_per_cpu_begin, 0);
>>> +
>>
>> I think we can extend the kasan_populate_early_shadow() call above up to
>> shadow_cea_per_cpu_begin point, instead of this.
>> populate_early_shadow() maps single RO zeroed page. No one should write to the shadow for IDT.
>> KASAN only needs writable shadow for linear mapping/stacks/vmalloc/global variables.
>
> Is that the only difference between the "early" and "normal" variants?
It is. kasan_populate_shadow() allocates new memory and maps it, while the "early" one maps
'kasan_early_shadow_page'
> If so, renaming them to kasan_populate_ro_shadow() vs. kasan_populate_rw_shadow() would
> make this code much more intuitive for non-KASAN folks.
>
Agreed.
On Tue, Nov 08, 2022, Andrey Ryabinin wrote:
>
> On 11/4/22 21:32, Sean Christopherson wrote:
> > @@ -409,6 +410,15 @@ void __init kasan_init(void)
> > kasan_mem_to_shadow((void *)VMALLOC_END + 1),
> > (void *)shadow_cea_begin);
> >
> > + /*
> > + * Populate the shadow for the shared portion of the CPU entry area.
> > + * Shadows for the per-CPU areas are mapped on-demand, as each CPU's
> > + * area is randomly placed somewhere in the 512GiB range and mapping
> > + * the entire 512GiB range is prohibitively expensive.
> > + */
> > + kasan_populate_shadow(shadow_cea_begin,
> > + shadow_cea_per_cpu_begin, 0);
> > +
>
> I think we can extend the kasan_populate_early_shadow() call above up to
> shadow_cea_per_cpu_begin point, instead of this.
> populate_early_shadow() maps single RO zeroed page. No one should write to the shadow for IDT.
> KASAN only needs writable shadow for linear mapping/stacks/vmalloc/global variables.
Is that the only difference between the "early" and "normal" variants? If so,
renaming them to kasan_populate_ro_shadow() vs. kasan_populate_rw_shadow() would
make this code much more intuitive for non-KASAN folks.
>
> > kasan_populate_early_shadow((void *)shadow_cea_end,
> > kasan_mem_to_shadow((void *)__START_KERNEL_map));
> >
On Tue, Nov 08, 2022, Andrey Ryabinin wrote:
>
> On 11/4/22 21:32, Sean Christopherson wrote:
> > @@ -409,6 +410,15 @@ void __init kasan_init(void)
> > kasan_mem_to_shadow((void *)VMALLOC_END + 1),
> > (void *)shadow_cea_begin);
> >
> > + /*
> > + * Populate the shadow for the shared portion of the CPU entry area.
> > + * Shadows for the per-CPU areas are mapped on-demand, as each CPU's
> > + * area is randomly placed somewhere in the 512GiB range and mapping
> > + * the entire 512GiB range is prohibitively expensive.
> > + */
> > + kasan_populate_shadow(shadow_cea_begin,
> > + shadow_cea_per_cpu_begin, 0);
> > +
>
> I think we can extend the kasan_populate_early_shadow() call above up to
> shadow_cea_per_cpu_begin point, instead of this.
> populate_early_shadow() maps single RO zeroed page. No one should write to the shadow for IDT.
> KASAN only needs writable shadow for linear mapping/stacks/vmalloc/global variables.
Any objection to simply converting this to use kasan_populate_early_shadow(),
i.e. to keeping a separate "populate" call for the CPU entry area? Purely so
that it's more obvious that a small portion of the overall CPU entry area is
mapped during init.