x86, kvm: use correct GFP flags for preemption disabled

Commit ddd7ed842627 ("x86/kvm: Alloc dummy async #PF token outside of
raw spinlock") leads to the following Smatch static checker warning:

	arch/x86/kernel/kvm.c:212 kvm_async_pf_task_wake()
	warn: sleeping in atomic context

arch/x86/kernel/kvm.c
    202         raw_spin_lock(&b->lock);
    203         n = _find_apf_task(b, token);
    204         if (!n) {
    205                 /*
    206                  * Async #PF not yet handled, add a dummy entry for the token.
    207                  * Allocating the token must be down outside of the raw lock
    208                  * as the allocator is preemptible on PREEMPT_RT kernels.
    209                  */
    210                 if (!dummy) {
    211                         raw_spin_unlock(&b->lock);
--> 212                         dummy = kzalloc(sizeof(*dummy), GFP_KERNEL);
                                                                ^^^^^^^^^^
Smatch thinks the caller has preempt disabled.  The `smdb.py preempt
kvm_async_pf_task_wake` output call tree is:

sysvec_kvm_asyncpf_interrupt() <- disables preempt
-> __sysvec_kvm_asyncpf_interrupt()
   -> kvm_async_pf_task_wake()

The caller is this:

arch/x86/kernel/kvm.c
   290        DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
   291        {
   292                struct pt_regs *old_regs = set_irq_regs(regs);
   293                u32 token;
   294
   295                ack_APIC_irq();
   296
   297                inc_irq_stat(irq_hv_callback_count);
   298
   299                if (__this_cpu_read(apf_reason.enabled)) {
   300                        token = __this_cpu_read(apf_reason.token);
   301                        kvm_async_pf_task_wake(token);
   302                        __this_cpu_write(apf_reason.token, 0);
   303                        wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1);
   304                }
   305
   306                set_irq_regs(old_regs);
   307        }

The DEFINE_IDTENTRY_SYSVEC() is a wrapper that calls this function
from the call_on_irqstack_cond().  It's inside the call_on_irqstack_cond()
where preempt is disabled (unless it's already disabled).  The
irq_enter/exit_rcu() functions disable/enable preempt.

Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
index 35b3c5836703bbf63d1c6683e7411a332c17cafa..1a3658f7e6d99c9c3ff7d1be057eeac2aa945c16 100644 (file)
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c
@@ -209,7 +209,7 @@ again:
                 */
                if (!dummy) {
                        raw_spin_unlock(&b->lock);
-                       dummy = kzalloc(sizeof(*dummy), GFP_KERNEL);
+                       dummy = kzalloc(sizeof(*dummy), GFP_ATOMIC);
 
                        /*
                         * Continue looping on allocation failure, eventually