On Sat, Mar 07, 2020 at 12:42:06AM +0100, Thomas Gleixner wrote:
> While working on the entry consolidation I stumbled over the KVM async page
> fault handler and kvm_async_pf_task_wait() in particular. It took me a
> while to realize that the randomly sprinkled around rcu_irq_enter()/exit()
> invocations are just cargo cult programming. Several patches "fixed" RCU
> splats by curing the symptoms without noticing that the code is flawed
> from a design perspective.
>
> The main problem is that this async injection is not based on a proper
> handshake mechanism and only respects the minimal requirement, i.e. the
> guest is not in a state where it has interrupts disabled.
>
> Aside of that the actual code is a convoluted one fits it all swiss army
> knife. It is invoked from different places with different RCU constraints:
>
> 1) Host side:
>
> vcpu_enter_guest()
> kvm_x86_ops->handle_exit()
> kvm_handle_page_fault()
> kvm_async_pf_task_wait()
>
> The invocation happens from fully preemptible context.
>
> 2) Guest side:
>
> The async page fault interrupted:
>
> a) user space
>
> b) preemptible kernel code which is not in a RCU read side
> critical section
>
> c) non-preemtible kernel code or a RCU read side critical section
> or kernel code with CONFIG_PREEMPTION=n which allows not to
> differentiate between #2b and #2c.
>
> RCU is watching for:
>
> #1 The vCPU exited and current is definitely not the idle task
>
> #2a The #PF entry code on the guest went through enter_from_user_mode()
> which reactivates RCU
>
> #2b There is no preemptible, interrupts enabled code in the kernel
> which can run with RCU looking away. (The idle task is always
> non preemptible).
>
> I.e. all schedulable states (#1, #2a, #2b) do not need any of this RCU
> voodoo at all.
>
> In #2c RCU is eventually not watching, but as that state cannot schedule
> anyway there is no point to worry about it so it has to invoke
> rcu_irq_enter() before running that code. This can be optimized, but this
> will be done as an extra step in course of the entry code consolidation
> work.
In other words, any needed rcu_irq_enter() and rcu_irq_exit() are added
in one of the entry-code consolidation patches, and patch below depends
on that patch, correct?
Thanx, Paul
> So the proper solution for this is to:
>
> - Split kvm_async_pf_task_wait() into schedule and halt based waiting
> interfaces which share the enqueueing code.
>
> - Add comments (condensed form of this changelog) to spare others the
> time waste and pain of reverse engineering all of this with the help of
> uncomprehensible changelogs and code history.
>
> - Invoke kvm_async_pf_task_wait_schedule() from kvm_handle_page_fault(),
> user mode and schedulable kernel side async page faults (#1, #2a, #2b)
>
> - Invoke kvm_async_pf_task_wait_halt() for the non schedulable kernel
> case (#2c).
>
> For this case also remove the rcu_irq_exit()/enter() pair around the
> halt as it is just a pointless exercise:
>
> - vCPUs can VMEXIT at any any random point and can be scheduled out
> for an arbitrary amount of time by the host and this is not any
> different except that it voluntary triggers the exit via halt.
>
> - The interrupted context could have RCU watching already. So the
> rcu_irq_exit() before the halt is not gaining anything aside of
> confusing the reader. Claiming that this might prevent RCU stalls
> is just an illusion.
>
> Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> ---
> arch/x86/include/asm/kvm_para.h | 2
> arch/x86/kernel/kvm.c | 156 ++++++++++++++++++++++++++++------------
> arch/x86/kvm/mmu/mmu.c | 2
> 3 files changed, 115 insertions(+), 45 deletions(-)
>
> --- a/arch/x86/include/asm/kvm_para.h
> +++ b/arch/x86/include/asm/kvm_para.h
> @@ -88,7 +88,7 @@ static inline long kvm_hypercall4(unsign
> bool kvm_para_available(void);
> unsigned int kvm_arch_para_features(void);
> unsigned int kvm_arch_para_hints(void);
> -void kvm_async_pf_task_wait(u32 token, int interrupt_kernel);
> +void kvm_async_pf_task_wait_schedule(u32 token);
> void kvm_async_pf_task_wake(u32 token);
> u32 kvm_read_and_reset_pf_reason(void);
> void kvm_disable_steal_time(void);
> --- a/arch/x86/kernel/kvm.c
> +++ b/arch/x86/kernel/kvm.c
> @@ -75,7 +75,7 @@ struct kvm_task_sleep_node {
> struct swait_queue_head wq;
> u32 token;
> int cpu;
> - bool halted;
> + bool use_halt;
> };
>
> static struct kvm_task_sleep_head {
> @@ -98,75 +98,145 @@ static struct kvm_task_sleep_node *_find
> return NULL;
> }
>
> -/*
> - * @interrupt_kernel: Is this called from a routine which interrupts the kernel
> - * (other than user space)?
> - */
> -void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
> +static bool kvm_async_pf_queue_task(u32 token, bool use_halt,
> + struct kvm_task_sleep_node *n)
> {
> u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
> struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
> - struct kvm_task_sleep_node n, *e;
> - DECLARE_SWAITQUEUE(wait);
> -
> - rcu_irq_enter();
> + struct kvm_task_sleep_node *e;
>
> raw_spin_lock(&b->lock);
> e = _find_apf_task(b, token);
> if (e) {
> /* dummy entry exist -> wake up was delivered ahead of PF */
> hlist_del(&e->link);
> - kfree(e);
> raw_spin_unlock(&b->lock);
> + kfree(e);
> + return false;
> + }
>
> - rcu_irq_exit();
> + n->token = token;
> + n->cpu = smp_processor_id();
> + n->use_halt = use_halt;
> + init_swait_queue_head(&n->wq);
> + hlist_add_head(&n->link, &b->list);
> + raw_spin_unlock(&b->lock);
> + return true;
> +}
> +
> +/*
> + * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
> + * @token: Token to identify the sleep node entry
> + *
> + * Invoked from the async pagefault handling code or from the VM exit page
> + * fault handler. In both cases RCU is watching.
> + */
> +void kvm_async_pf_task_wait_schedule(u32 token)
> +{
> + struct kvm_task_sleep_node n;
> + DECLARE_SWAITQUEUE(wait);
> +
> + lockdep_assert_irqs_disabled();
> +
> + if (!kvm_async_pf_queue_task(token, false, &n))
> return;
> +
> + for (;;) {
> + prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
> + if (hlist_unhashed(&n.link))
> + break;
> +
> + local_irq_enable();
> + schedule();
> + local_irq_disable();
> }
> + finish_swait(&n.wq, &wait);
> +}
> +EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
>
> - n.token = token;
> - n.cpu = smp_processor_id();
> - n.halted = is_idle_task(current) ||
> - (IS_ENABLED(CONFIG_PREEMPT_COUNT)
> - ? preempt_count() > 1 || rcu_preempt_depth()
> - : interrupt_kernel);
> - init_swait_queue_head(&n.wq);
> - hlist_add_head(&n.link, &b->list);
> - raw_spin_unlock(&b->lock);
> +/*
> + * Invoked from the async page fault handler.
> + */
> +static void kvm_async_pf_task_wait_halt(u32 token)
> +{
> + struct kvm_task_sleep_node n;
> +
> + if (!kvm_async_pf_queue_task(token, true, &n))
> + return;
>
> for (;;) {
> - if (!n.halted)
> - prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
> if (hlist_unhashed(&n.link))
> break;
> + /*
> + * No point in doing anything about RCU here. Any RCU read
> + * side critical section or RCU watching section can be
> + * interrupted by VMEXITs and the host is free to keep the
> + * vCPU scheduled out as long as it sees fit. This is not
> + * any different just because of the halt induced voluntary
> + * VMEXIT.
> + *
> + * Also the async page fault could have interrupted any RCU
> + * watching context, so invoking rcu_irq_exit()/enter()
> + * around this is not gaining anything.
> + */
> + native_safe_halt();
> + local_irq_disable();
> + }
> +}
>
> - rcu_irq_exit();
> +/* Invoked from the async page fault handler */
> +static void kvm_async_pf_task_wait(u32 token, bool usermode)
> +{
> + bool can_schedule;
>
> - if (!n.halted) {
> - local_irq_enable();
> - schedule();
> - local_irq_disable();
> - } else {
> - /*
> - * We cannot reschedule. So halt.
> - */
> - native_safe_halt();
> - local_irq_disable();
> - }
> + /*
> + * No need to check whether interrupts were disabled because the
> + * host will (hopefully) only inject an async page fault into
> + * interrupt enabled regions.
> + *
> + * If CONFIG_PREEMPTION is enabled then check whether the code
> + * which triggered the page fault is preemptible. This covers user
> + * mode as well because preempt_count() is obviously 0 there.
> + *
> + * The check for rcu_preempt_depth() is also required because
> + * voluntary scheduling inside a rcu read locked section is not
> + * allowed.
> + *
> + * The idle task is already covered by this because idle always
> + * has a preempt count > 0.
> + *
> + * If CONFIG_PREEMPTION is disabled only allow scheduling when
> + * coming from user mode as there is no indication whether the
> + * context which triggered the page fault could schedule or not.
> + */
> + if (IS_ENABLED(CONFIG_PREEMPTION))
> + can_schedule = preempt_count() + rcu_preempt_depth() == 0;
> + else
> + can_schedule = usermode;
>
> + /*
> + * If the kernel context is allowed to schedule then RCU is
> + * watching because no preemptible code in the kernel is inside RCU
> + * idle state. So it can be treated like user mode. User mode is
> + * safe because the #PF entry invoked enter_from_user_mode().
> + *
> + * For the non schedulable case invoke rcu_irq_enter() for
> + * now. This will be moved out to the pagefault entry code later
> + * and only invoked when really needed.
> + */
> + if (can_schedule) {
> + kvm_async_pf_task_wait_schedule(token);
> + } else {
> rcu_irq_enter();
> + kvm_async_pf_task_wait_halt(token);
> + rcu_irq_exit();
> }
> - if (!n.halted)
> - finish_swait(&n.wq, &wait);
> -
> - rcu_irq_exit();
> - return;
> }
> -EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
>
> static void apf_task_wake_one(struct kvm_task_sleep_node *n)
> {
> hlist_del_init(&n->link);
> - if (n->halted)
> + if (n->use_halt)
> smp_send_reschedule(n->cpu);
> else if (swq_has_sleeper(&n->wq))
> swake_up_one(&n->wq);
> @@ -255,7 +325,7 @@ bool __kvm_handle_async_pf(struct pt_reg
> return false;
> case KVM_PV_REASON_PAGE_NOT_PRESENT:
> /* page is swapped out by the host. */
> - kvm_async_pf_task_wait(token, !user_mode(regs));
> + kvm_async_pf_task_wait(token, user_mode(regs));
> return true;
> case KVM_PV_REASON_PAGE_READY:
> rcu_irq_enter();
> --- a/arch/x86/kvm/mmu/mmu.c
> +++ b/arch/x86/kvm/mmu/mmu.c
> @@ -4205,7 +4205,7 @@ int kvm_handle_page_fault(struct kvm_vcp
> case KVM_PV_REASON_PAGE_NOT_PRESENT:
> vcpu->arch.apf.host_apf_reason = 0;
> local_irq_disable();
> - kvm_async_pf_task_wait(fault_address, 0);
> + kvm_async_pf_task_wait_schedule(fault_address);
> local_irq_enable();
> break;
> case KVM_PV_REASON_PAGE_READY:
>
