Quantcast

[xen master] x86/vm_event: fix race between __context_switch() and vm_event_resume()

classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

[xen master] x86/vm_event: fix race between __context_switch() and vm_event_resume()

patchbot
commit 461f0482033b09ade2f3e38c8c33f1c15bc8defc
Author:     Razvan Cojocaru <[hidden email]>
AuthorDate: Mon May 8 14:54:00 2017 +0200
Commit:     Jan Beulich <[hidden email]>
CommitDate: Mon May 8 14:54:00 2017 +0200

    x86/vm_event: fix race between __context_switch() and vm_event_resume()
   
    The introspection agent can reply to a vm_event faster than
    vmx_vmexit_handler() can complete in some cases, where it is then
    not safe for vm_event_set_registers() to modify v->arch.user_regs.
    In the test scenario, we were stepping over an INT3 breakpoint by
    setting RIP += 1. The quick reply tended to complete before the VCPU
    triggering the introspection event had properly paused and been
    descheduled. If the reply occurs before __context_switch() happens,
    __context_switch() clobbers the reply by overwriting
    v->arch.user_regs from the stack. If we don't pass through
    __context_switch() (due to switching to the idle vCPU), reply data
    wouldn't be picked up when switching back straight to the original
    vCPU.
   
    This patch ensures that vm_event_resume() code only sets per-VCPU
    data to be used for the actual setting of registers later in
    hvm_do_resume() (similar to the model used to control setting of CRs
    and MSRs).
   
    The patch additionally removes the sync_vcpu_execstate(v) call from
    vm_event_resume(), which is no longer necessary, which removes the
    associated broadcast TLB flush (read: performance improvement).
   
    Signed-off-by: Razvan Cojocaru <[hidden email]>
    Signed-off-by: Andrew Cooper <[hidden email]>
    Reviewed-by: Jan Beulich <[hidden email]>
    Acked-by: Tamas K Lengyel <[hidden email]>
    Release-acked-by: Julien Grall <[hidden email]>
---
 xen/arch/x86/hvm/vm_event.c    | 35 +++++++++++++++++++++++++++++++++++
 xen/arch/x86/vm_event.c        | 22 ++--------------------
 xen/common/vm_event.c          | 17 ++++++++++-------
 xen/include/asm-x86/vm_event.h |  2 ++
 4 files changed, 49 insertions(+), 27 deletions(-)

diff --git a/xen/arch/x86/hvm/vm_event.c b/xen/arch/x86/hvm/vm_event.c
index 821236e..28d08a6 100644
--- a/xen/arch/x86/hvm/vm_event.c
+++ b/xen/arch/x86/hvm/vm_event.c
@@ -25,12 +25,47 @@
 #include <asm/hvm/support.h>
 #include <asm/vm_event.h>
 
+static void hvm_vm_event_set_registers(const struct vcpu *v)
+{
+    ASSERT(v == current);
+
+    if ( unlikely(v->arch.vm_event->set_gprs) )
+    {
+        struct cpu_user_regs *regs = guest_cpu_user_regs();
+
+        regs->rax = v->arch.vm_event->gprs.rax;
+        regs->rbx = v->arch.vm_event->gprs.rbx;
+        regs->rcx = v->arch.vm_event->gprs.rcx;
+        regs->rdx = v->arch.vm_event->gprs.rdx;
+        regs->rsp = v->arch.vm_event->gprs.rsp;
+        regs->rbp = v->arch.vm_event->gprs.rbp;
+        regs->rsi = v->arch.vm_event->gprs.rsi;
+        regs->rdi = v->arch.vm_event->gprs.rdi;
+
+        regs->r8 = v->arch.vm_event->gprs.r8;
+        regs->r9 = v->arch.vm_event->gprs.r9;
+        regs->r10 = v->arch.vm_event->gprs.r10;
+        regs->r11 = v->arch.vm_event->gprs.r11;
+        regs->r12 = v->arch.vm_event->gprs.r12;
+        regs->r13 = v->arch.vm_event->gprs.r13;
+        regs->r14 = v->arch.vm_event->gprs.r14;
+        regs->r15 = v->arch.vm_event->gprs.r15;
+
+        regs->rflags = v->arch.vm_event->gprs.rflags;
+        regs->rip = v->arch.vm_event->gprs.rip;
+
+        v->arch.vm_event->set_gprs = false;
+    }
+}
+
 void hvm_vm_event_do_resume(struct vcpu *v)
 {
     struct monitor_write_data *w;
 
     ASSERT(v->arch.vm_event);
 
+    hvm_vm_event_set_registers(v);
+
     w = &v->arch.vm_event->write_data;
 
     if ( unlikely(v->arch.vm_event->emulate_flags) )
diff --git a/xen/arch/x86/vm_event.c b/xen/arch/x86/vm_event.c
index 502ccc2..a6ea42c 100644
--- a/xen/arch/x86/vm_event.c
+++ b/xen/arch/x86/vm_event.c
@@ -113,26 +113,8 @@ void vm_event_set_registers(struct vcpu *v, vm_event_response_t *rsp)
 {
     ASSERT(atomic_read(&v->vm_event_pause_count));
 
-    v->arch.user_regs.rax = rsp->data.regs.x86.rax;
-    v->arch.user_regs.rbx = rsp->data.regs.x86.rbx;
-    v->arch.user_regs.rcx = rsp->data.regs.x86.rcx;
-    v->arch.user_regs.rdx = rsp->data.regs.x86.rdx;
-    v->arch.user_regs.rsp = rsp->data.regs.x86.rsp;
-    v->arch.user_regs.rbp = rsp->data.regs.x86.rbp;
-    v->arch.user_regs.rsi = rsp->data.regs.x86.rsi;
-    v->arch.user_regs.rdi = rsp->data.regs.x86.rdi;
-
-    v->arch.user_regs.r8 = rsp->data.regs.x86.r8;
-    v->arch.user_regs.r9 = rsp->data.regs.x86.r9;
-    v->arch.user_regs.r10 = rsp->data.regs.x86.r10;
-    v->arch.user_regs.r11 = rsp->data.regs.x86.r11;
-    v->arch.user_regs.r12 = rsp->data.regs.x86.r12;
-    v->arch.user_regs.r13 = rsp->data.regs.x86.r13;
-    v->arch.user_regs.r14 = rsp->data.regs.x86.r14;
-    v->arch.user_regs.r15 = rsp->data.regs.x86.r15;
-
-    v->arch.user_regs.rflags = rsp->data.regs.x86.rflags;
-    v->arch.user_regs.rip = rsp->data.regs.x86.rip;
+    v->arch.vm_event->gprs = rsp->data.regs.x86;
+    v->arch.vm_event->set_gprs = true;
 }
 
 void vm_event_monitor_next_interrupt(struct vcpu *v)
diff --git a/xen/common/vm_event.c b/xen/common/vm_event.c
index 0fe9a53..9291db6 100644
--- a/xen/common/vm_event.c
+++ b/xen/common/vm_event.c
@@ -357,6 +357,16 @@ void vm_event_resume(struct domain *d, struct vm_event_domain *ved)
 {
     vm_event_response_t rsp;
 
+    /*
+     * vm_event_resume() runs in either XEN_DOMCTL_VM_EVENT_OP_*, or
+     * EVTCHN_send context from the introspection consumer. Both contexts
+     * are guaranteed not to be the subject of vm_event responses.
+     * While we could ASSERT(v != current) for each VCPU in d in the loop
+     * below, this covers the case where we would need to iterate over all
+     * of them more succintly.
+     */
+    ASSERT(d != current->domain);
+
     /* Pull all responses off the ring. */
     while ( vm_event_get_response(d, ved, &rsp) )
     {
@@ -375,13 +385,6 @@ void vm_event_resume(struct domain *d, struct vm_event_domain *ved)
         v = d->vcpu[rsp.vcpu_id];
 
         /*
-         * Make sure the vCPU state has been synchronized for the custom
-         * handlers.
-         */
-        if ( atomic_read(&v->vm_event_pause_count) )
-            sync_vcpu_execstate(v);
-
-        /*
          * In some cases the response type needs extra handling, so here
          * we call the appropriate handlers.
          */
diff --git a/xen/include/asm-x86/vm_event.h b/xen/include/asm-x86/vm_event.h
index ca73f99..39e73c8 100644
--- a/xen/include/asm-x86/vm_event.h
+++ b/xen/include/asm-x86/vm_event.h
@@ -32,6 +32,8 @@ struct arch_vm_event {
         struct vm_event_emul_insn_data insn;
     } emul;
     struct monitor_write_data write_data;
+    struct vm_event_regs_x86 gprs;
+    bool set_gprs;
 };
 
 int vm_event_init_domain(struct domain *d);
--
generated by git-patchbot for /home/xen/git/xen.git#master

_______________________________________________
Xen-changelog mailing list
[hidden email]
https://lists.xenproject.org/xen-changelog
Loading...