Re: [RFC KVM 24/27] kvm/isolation: KVM page fault handler
> On May 14, 2019, at 8:36 AM, Alexandre Chartre > wrote: > > >> On 5/14/19 9:21 AM, Peter Zijlstra wrote: >>> On Mon, May 13, 2019 at 07:02:30PM -0700, Andy Lutomirski wrote: >>> This sounds like a great use case for static_call(). PeterZ, do you >>> suppose we could wire up static_call() with the module infrastructure >>> to make it easy to do "static_call to such-and-such GPL module symbol >>> if that symbol is in a loaded module, else nop"? >> You're basically asking it to do dynamic linking. And I suppose that is >> technically possible. >> However, I'm really starting to think kvm (or at least these parts of it >> that want to play these games) had better not be a module anymore. > > Maybe we can use an atomic notifier (e.g. page_fault_notifier)? > > IMO that’s worse. I want to be able to read do_page_fault() and understand what happens and in what order. Having do_page_fault run with the wrong CR3 is so fundamental to its operation that it needs to be very obvious what’s happening.
Re: [RFC KVM 24/27] kvm/isolation: KVM page fault handler
On 5/14/19 9:21 AM, Peter Zijlstra wrote: On Mon, May 13, 2019 at 07:02:30PM -0700, Andy Lutomirski wrote: This sounds like a great use case for static_call(). PeterZ, do you suppose we could wire up static_call() with the module infrastructure to make it easy to do "static_call to such-and-such GPL module symbol if that symbol is in a loaded module, else nop"? You're basically asking it to do dynamic linking. And I suppose that is technically possible. However, I'm really starting to think kvm (or at least these parts of it that want to play these games) had better not be a module anymore. Maybe we can use an atomic notifier (e.g. page_fault_notifier)? alex.
Re: [RFC KVM 24/27] kvm/isolation: KVM page fault handler
On Mon, May 13, 2019 at 07:02:30PM -0700, Andy Lutomirski wrote: > This sounds like a great use case for static_call(). PeterZ, do you > suppose we could wire up static_call() with the module infrastructure > to make it easy to do "static_call to such-and-such GPL module symbol > if that symbol is in a loaded module, else nop"? You're basically asking it to do dynamic linking. And I suppose that is technically possible. However, I'm really starting to think kvm (or at least these parts of it that want to play these games) had better not be a module anymore.
Re: [RFC KVM 24/27] kvm/isolation: KVM page fault handler
On Mon, May 13, 2019 at 2:26 PM Liran Alon wrote: > > > > > On 13 May 2019, at 18:15, Peter Zijlstra wrote: > > > > On Mon, May 13, 2019 at 04:38:32PM +0200, Alexandre Chartre wrote: > >> diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c > >> index 46df4c6..317e105 100644 > >> --- a/arch/x86/mm/fault.c > >> +++ b/arch/x86/mm/fault.c > >> @@ -33,6 +33,10 @@ > >> #define CREATE_TRACE_POINTS > >> #include > >> > >> +bool (*kvm_page_fault_handler)(struct pt_regs *regs, unsigned long > >> error_code, > >> + unsigned long address); > >> +EXPORT_SYMBOL(kvm_page_fault_handler); > > > > NAK NAK NAK NAK > > > > This is one of the biggest anti-patterns around. > > I agree. > I think that mm should expose a mm_set_kvm_page_fault_handler() or something > (give it a better name). > Similar to how arch/x86/kernel/irq.c have > kvm_set_posted_intr_wakeup_handler(). > > -Liran > This sounds like a great use case for static_call(). PeterZ, do you suppose we could wire up static_call() with the module infrastructure to make it easy to do "static_call to such-and-such GPL module symbol if that symbol is in a loaded module, else nop"?
Re: [RFC KVM 24/27] kvm/isolation: KVM page fault handler
> On 13 May 2019, at 18:15, Peter Zijlstra wrote: > > On Mon, May 13, 2019 at 04:38:32PM +0200, Alexandre Chartre wrote: >> diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c >> index 46df4c6..317e105 100644 >> --- a/arch/x86/mm/fault.c >> +++ b/arch/x86/mm/fault.c >> @@ -33,6 +33,10 @@ >> #define CREATE_TRACE_POINTS >> #include >> >> +bool (*kvm_page_fault_handler)(struct pt_regs *regs, unsigned long >> error_code, >> + unsigned long address); >> +EXPORT_SYMBOL(kvm_page_fault_handler); > > NAK NAK NAK NAK > > This is one of the biggest anti-patterns around. I agree. I think that mm should expose a mm_set_kvm_page_fault_handler() or something (give it a better name). Similar to how arch/x86/kernel/irq.c have kvm_set_posted_intr_wakeup_handler(). -Liran
Re: [RFC KVM 24/27] kvm/isolation: KVM page fault handler
On 5/13/19 6:02 PM, Andy Lutomirski wrote: On Mon, May 13, 2019 at 7:39 AM Alexandre Chartre wrote: The KVM page fault handler handles page fault occurring while using the KVM address space by switching to the kernel address space and retrying the access (except if the fault occurs while switching to the kernel address space). Processing of page faults occurring while using the kernel address space is unchanged. Page fault log is cleared when creating a vm so that page fault information doesn't persist when qemu is stopped and restarted. Are you saying that a page fault will just exit isolation? This completely defeats most of the security, right? Sure, it still helps with side channels, but not with actual software bugs. Yes, page fault exit isolation so that the faulty instruction can be retried with the full kernel address space. When exiting isolation, we also want to kick the sibling hyperthread and pinned it so that it can't steal secret while we use the kernel address page, but that's not implemented in this serie (see TODO comment in kvm_isolation_exit() in patch 25 "kvm/isolation: implement actual KVM isolation enter/exit"). alex.
Re: [RFC KVM 24/27] kvm/isolation: KVM page fault handler
On Mon, May 13, 2019 at 7:39 AM Alexandre Chartre wrote: > > The KVM page fault handler handles page fault occurring while using > the KVM address space by switching to the kernel address space and > retrying the access (except if the fault occurs while switching > to the kernel address space). Processing of page faults occurring > while using the kernel address space is unchanged. > > Page fault log is cleared when creating a vm so that page fault > information doesn't persist when qemu is stopped and restarted. Are you saying that a page fault will just exit isolation? This completely defeats most of the security, right? Sure, it still helps with side channels, but not with actual software bugs.
Re: [RFC KVM 24/27] kvm/isolation: KVM page fault handler
On Mon, May 13, 2019 at 04:38:32PM +0200, Alexandre Chartre wrote: > diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c > index 46df4c6..317e105 100644 > --- a/arch/x86/mm/fault.c > +++ b/arch/x86/mm/fault.c > @@ -33,6 +33,10 @@ > #define CREATE_TRACE_POINTS > #include > > +bool (*kvm_page_fault_handler)(struct pt_regs *regs, unsigned long > error_code, > +unsigned long address); > +EXPORT_SYMBOL(kvm_page_fault_handler); NAK NAK NAK NAK This is one of the biggest anti-patterns around.
[RFC KVM 24/27] kvm/isolation: KVM page fault handler
The KVM page fault handler handles page fault occurring while using the KVM address space by switching to the kernel address space and retrying the access (except if the fault occurs while switching to the kernel address space). Processing of page faults occurring while using the kernel address space is unchanged. Page fault log is cleared when creating a vm so that page fault information doesn't persist when qemu is stopped and restarted. The KVM module parameter page_fault_stack can be used to disable dumping stack trace when a page fault occurs while using the KVM address space. The fault will still be reported but without the stack trace. Signed-off-by: Alexandre Chartre --- arch/x86/kernel/dumpstack.c |1 + arch/x86/kvm/isolation.c| 202 +++ arch/x86/mm/fault.c | 12 +++ 3 files changed, 215 insertions(+), 0 deletions(-) diff --git a/arch/x86/kernel/dumpstack.c b/arch/x86/kernel/dumpstack.c index 2b58864..aa28763 100644 --- a/arch/x86/kernel/dumpstack.c +++ b/arch/x86/kernel/dumpstack.c @@ -292,6 +292,7 @@ void show_stack(struct task_struct *task, unsigned long *sp) show_trace_log_lvl(task, NULL, sp, KERN_DEFAULT); } +EXPORT_SYMBOL(show_stack); void show_stack_regs(struct pt_regs *regs) { diff --git a/arch/x86/kvm/isolation.c b/arch/x86/kvm/isolation.c index e7979b3..db0a7ce 100644 --- a/arch/x86/kvm/isolation.c +++ b/arch/x86/kvm/isolation.c @@ -8,6 +8,7 @@ #include #include #include +#include #include #include @@ -17,6 +18,9 @@ #include "isolation.h" +extern bool (*kvm_page_fault_handler)(struct pt_regs *regs, + unsigned long error_code, + unsigned long address); enum page_table_level { PGT_LEVEL_PTE, @@ -91,6 +95,25 @@ struct kvm_range_mapping { static LIST_HEAD(kvm_range_mapping_list); static DEFINE_MUTEX(kvm_range_mapping_lock); +/* + * When a page fault occurs, while running with the KVM address space, + * the KVM page fault handler prints information about the fault (in + * particular the stack trace), and it switches back to the kernel + * address space. + * + * Information printed by the KVM page fault handler can be used to find + * out data not mapped in the KVM address space. Then the KVM address + * space can be augmented to include the missing mapping so that we don't + * fault at that same place anymore. + * + * The following variables keep track of page faults occurring while running + * with the KVM address space to prevent displaying the same information. + */ + +#define KVM_LAST_FAULT_COUNT 128 + +static unsigned long kvm_last_fault[KVM_LAST_FAULT_COUNT]; + struct mm_struct kvm_mm = { .mm_rb = RB_ROOT, @@ -126,6 +149,14 @@ static void kvm_clear_mapping(void *ptr, size_t size, static bool __read_mostly address_space_isolation; module_param(address_space_isolation, bool, 0444); +/* + * When set to true, KVM dumps the stack when a page fault occurs while + * running with the KVM address space. Otherwise the page fault is still + * reported but without the stack trace. + */ +static bool __read_mostly page_fault_stack = true; +module_param(page_fault_stack, bool, 0444); + static struct kvm_range_mapping *kvm_get_range_mapping_locked(void *ptr, bool *subset) { @@ -1195,6 +1226,173 @@ static void kvm_reset_all_task_mapping(void) mutex_unlock(&kvm_task_mapping_lock); } +static int bad_address(void *p) +{ + unsigned long dummy; + + return probe_kernel_address((unsigned long *)p, dummy); +} + +static void kvm_dump_pagetable(pgd_t *base, unsigned long address) +{ + pgd_t *pgd = base + pgd_index(address); + p4d_t *p4d; + pud_t *pud; + pmd_t *pmd; + pte_t *pte; + + pr_info("BASE %px ", base); + + if (bad_address(pgd)) + goto bad; + + pr_cont("PGD %lx ", pgd_val(*pgd)); + + if (!pgd_present(*pgd)) + goto out; + + p4d = p4d_offset(pgd, address); + if (bad_address(p4d)) + goto bad; + + pr_cont("P4D %lx ", p4d_val(*p4d)); + if (!p4d_present(*p4d) || p4d_large(*p4d)) + goto out; + + pud = pud_offset(p4d, address); + if (bad_address(pud)) + goto bad; + + pr_cont("PUD %lx ", pud_val(*pud)); + if (!pud_present(*pud) || pud_large(*pud)) + goto out; + + pmd = pmd_offset(pud, address); + if (bad_address(pmd)) + goto bad; + + pr_cont("PMD %lx ", pmd_val(*pmd)); + if (!pmd_present(*pmd) || pmd_large(*pmd)) + goto out; + + pte = pte_offset_kernel(pmd, address); + if (bad_address(pte)) + goto bad; + + pr_cont("PTE %lx", pte_val(*pte)); +out: + pr_cont("\n"); + return; +bad: + pr_info("BAD\n"); +} + +static void kvm_clear_page_f