Re: [Xen-devel] [PATCH v7 for-xen-4.5 2/2] dpci: Replace tasklet with an softirq (v7)

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 27/09/14 02:33, Konrad Rzeszutek Wilk wrote:
> The existing tasklet mechanism has a single global
> spinlock that is taken every-time the global list
> is touched. And we use this lock quite a lot - when
> we call do_tasklet_work which is called via an softirq
> and from the idle loop. We take the lock on any
> operation on the tasklet_list.
>
> The problem we are facing is that there are quite a lot of
> tasklets scheduled. The most common one that is invoked is
> the one injecting the VIRQ_TIMER in the guest. Guests
> are not insane and don't set the one-shot or periodic
> clocks to be in sub 1ms intervals (causing said tasklet
> to be scheduled for such small intervalls).
>
> The problem appears when PCI passthrough devices are used
> over many sockets and we have an mix of heavy-interrupt
> guests and idle guests. The idle guests end up seeing
> 1/10 of its RUNNING timeslice eaten by the hypervisor
> (and 40% steal time).
>
> The mechanism by which we inject PCI interrupts is by
> hvm_do_IRQ_dpci which schedules the hvm_dirq_assist
> tasklet every time an interrupt is received.
> The callchain is:
>
> _asm_vmexit_handler
>  -> vmx_vmexit_handler
>     ->vmx_do_extint
>         -> do_IRQ
>             -> __do_IRQ_guest
>                 -> hvm_do_IRQ_dpci
>                    tasklet_schedule(&dpci->dirq_tasklet);
>                    [takes lock to put the tasklet on]
>
> [later on the schedule_tail is invoked which is 'vmx_do_resume']
>
> vmx_do_resume
>  -> vmx_asm_do_vmentry
>         -> call vmx_intr_assist
>           -> vmx_process_softirqs
>             -> do_softirq
>               [executes the tasklet function, takes the
>                lock again]
>
> While on other CPUs they might be sitting in a idle loop
> and invoked to deliver an VIRQ_TIMER, which also ends
> up taking the lock twice: first to schedule the
> v->arch.hvm_vcpu.assert_evtchn_irq_tasklet (accounted to
> the guests' BLOCKED_state); then to execute it - which is
> accounted for in the guest's RUNTIME_state.
>
> The end result is that on a 8 socket machine with
> PCI passthrough, where four sockets are busy with interrupts,
> and the other sockets have idle guests - we end up with
> the idle guests having around 40% steal time and 1/10
> of its timeslice (3ms out of 30 ms) being tied up
> taking the lock. The latency of the PCI interrupts delieved
> to guest is also hindered.
>
> With this patch the problem disappears completly.
> That is removing the lock for the PCI passthrough use-case
> (the 'hvm_dirq_assist' case) by not using tasklets at all.
>
> The patch is simple - instead of scheduling an tasklet
> we schedule our own softirq - HVM_DPCI_SOFTIRQ, which will
> take care of running 'hvm_dirq_assist'. The information we need
> on each CPU is which 'struct hvm_pirq_dpci' structure the
> 'hvm_dirq_assist' needs to run on. That is simple solved by
> threading the 'struct hvm_pirq_dpci' through a linked list.
> The rule of only running one 'hvm_dirq_assist' for only
> one 'hvm_pirq_dpci' is also preserved by having
> 'schedule_dpci_for' ignore any subsequent calls for an domain
> which has already been scheduled.
>
> Since we are using 'hvm_pirq_dpci' structure it is setup in
> 'pt_irq_create_bind' - we check if via 'pt_pirq_softirq_active'
> whether the 'hvm_dirq_assist' is still outstanding on the
> 'pirq' and if so spin up to a second calling softirq
> to flush it out. This replaces the 'tasklet_kill' which
> had been previously done in 'pt_irq_destroy_bind' via
> the 'pt_pirq_cleanup_check'.
>
> The mechanism to tear it down is more complex as there
> are three ways it can be executed. To make it simpler
> everything revolves around 'pt_pirq_softirq_active'. If it
> returns -EGAIN that means there is an outstanding softirq
> that needs to finish running before we can continue tearing
> down. With that in mind:
>
> a) pci_clean_dpci_irq. This gets called when the guest is
>    being destroyed. We end up calling 'pt_pirq_softirq_active'
>    to see if it is OK to continue the destruction.
>
>    The scenarios in which the 'struct pirq' (and subsequently
>    the 'hvm_pirq_dpci') gets destroyed is when:
>
>    - guest did not use the pirq at all after setup.
>    - guest did use pirq, but decided to mask and left it in that
>      state.
>    - guest did use pirq, but crashed.
>
>    In all of those scenarios we end up calling
>    'pt_pirq_softirq_active' to check if the softirq is still
>    active. Read below on the 'pt_pirq_softirq_active' loop.
>
>    Note that before it would spin on an tasklet_kill waiting
>    the 'hvm_dirq_assist' to flush out. We now do it via
>    -EAGAIN to allow continuation.
>
> b) pt_irq_destroy_bind (guest disables the MSI). We double-check
>    that the softirq has run by piggy-backing on the existing
>    'pirq_cleanup_check' mechanism which calls 'pt_pirq_cleanup_check'.
>    We add the extra call to 'pt_pirq_softirq_active' in
>    'pt_pirq_cleanup_check'.
>
>    NOTE: Guests that use event channels unbind first the
>    event channel from PIRQs, so the 'pt_pirq_cleanup_check'
>    won't be called as eventch is set to zero. In that case
>    we either clean it up via the a) or c) mechanism.
>
>    There is an extra scenario regardless of eventch being
>    set or not: the guest did 'pt_irq_destroy_bind' while an
>    interrupt was triggered and softirq was scheduled (but had not
>    been run). It is OK to still run the softirq as
>    hvm_dirq_assist won't do anything (as the flags are
>    set to zero).
>
>    Note that before this patch we would spin forever in
>    tasklet_kill waiting for the 'hvm_dirq_assist' to flush out.
>    That 'spin' has been moved to 'pt_irq_create_bind' with
>    an -EAGAIN mechanism to not spin forever.
>
> c) pt_irq_create_bind (not a typo). The scenarios are:
>
>      - guest disables the MSI and then enables it
>        (rmmod and modprobe in a loop). We check if the
>        softirq has been scheduled.
>        Imagine the 'b)' with interrupts in flight and c) getting
>        called in a loop.
>
>      - we hit once the error paths in 'pt_irq_create_bind' while
>        an interrupt was triggered and softirq was scheduled.
>
> In all of those cases we will spin up to a second on
> 'pt_pirq_softirq_active' (at the start of the 'pt_irq_create_bind')
> with the event_lock spinlock dropped and calling
> 'process_pending_softirqs'. hvm_dirq_assist() will be executed and
> then the softirq will clear 'mapping' which signals that that we
> can re-use the 'hvm_pirq_dpci' structure.
>
> Note that the 'flags' variable is cleared at that
> point, so hvm_dirq_assist won't actually do anything..
>
> Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@xxxxxxxxxx>
> Suggested-by: Jan Beulich <JBeulich@xxxxxxxx>
>
> ---
> v2: On top of ref-cnts also have wait loop for the outstanding
>     'struct domain' that need to be processed.
> v3: Add -EAGAIN, fix up StyleGuide issues
> v4: Clean it up mode, redo per_cpu, this_cpu logic
> v5: Instead of threading struct domain, use hvm_pirq_dpci.
> v6: Ditch the 'state' bit, expand description, simplify
>     softirq and teardown sequence.
> v7: Flesh out the comments. Drop the use of domain refcounts
> ---
>  xen/arch/x86/domain.c         |   4 +-
>  xen/drivers/passthrough/io.c  | 152 
> ++++++++++++++++++++++++++++++++++++++----
>  xen/drivers/passthrough/pci.c |  31 ++++++---
>  xen/include/xen/hvm/irq.h     |   3 +-
>  xen/include/xen/pci.h         |   2 +-
>  xen/include/xen/softirq.h     |   3 +
>  6 files changed, 170 insertions(+), 25 deletions(-)
>
> diff --git a/xen/arch/x86/domain.c b/xen/arch/x86/domain.c
> index 8cfd1ca..8f6e17a 100644
> --- a/xen/arch/x86/domain.c
> +++ b/xen/arch/x86/domain.c
> @@ -1938,7 +1938,9 @@ int domain_relinquish_resources(struct domain *d)
>      switch ( d->arch.relmem )
>      {
>      case RELMEM_not_started:
> -        pci_release_devices(d);
> +        ret = pci_release_devices(d);
> +        if ( ret )
> +            return ret;
>  
>          /* Tear down paging-assistance stuff. */
>          paging_teardown(d);
> diff --git a/xen/drivers/passthrough/io.c b/xen/drivers/passthrough/io.c
> index 8534d63..40ad1dc 100644
> --- a/xen/drivers/passthrough/io.c
> +++ b/xen/drivers/passthrough/io.c
> @@ -20,14 +20,60 @@
>  
>  #include <xen/event.h>
>  #include <xen/iommu.h>
> +#include <xen/cpu.h>
>  #include <xen/irq.h>
>  #include <asm/hvm/irq.h>
>  #include <asm/hvm/iommu.h>
>  #include <asm/hvm/support.h>
>  #include <xen/hvm/irq.h>
> -#include <xen/tasklet.h>
>  
> -static void hvm_dirq_assist(unsigned long arg);
> +static DEFINE_PER_CPU(struct list_head, dpci_list);
> +
> +/*
> + * Should only be called from hvm_do_IRQ_dpci. We use the
> + * 'masked' as an gate to thwart multiple interrupts.
> + *
> + * The 'masked' is cleared by 'softirq_dpci' when it has
> + * completed executing 'hvm_dirq_assist'.
> + *
> + */
> +static void schedule_softirq_for(struct hvm_pirq_dpci *pirq_dpci)
> +{
> +    unsigned long flags;
> +
> +    if ( pirq_dpci->masked )
> +        return;
> +
> +    pirq_dpci->masked = 1;
> +
> +    local_irq_save(flags);
> +    list_add_tail(&pirq_dpci->softirq_list, &this_cpu(dpci_list));
> +    local_irq_restore(flags);
> +
> +    raise_softirq(HVM_DPCI_SOFTIRQ);
> +}
> +
> +/*
> + * If we are racing with softirq_dpci (masked is still set) we return
> + * -EAGAIN. Otherwise we return 0.
> + *
> + *  If it is -EAGAIN, it is the callers responsibility to make sure
> + *  that the softirq (with the event_lock dropped) has ran. We need
> + *  to flush out the outstanding 'dpci_softirq' (no more of them
> + *  will be added for this pirq as the IRQ action handler has been
> + *  reset in pt_irq_destroy_bind).
> + */
> +int pt_pirq_softirq_active(struct hvm_pirq_dpci *pirq_dpci)
> +{
> +    if ( pirq_dpci->masked )
> +        return -EAGAIN;
> +    /*
> +     * If in the future we would call 'schedule_softirq_for' right away
> +     * after 'pt_pirq_softirq_active' we MUST reset the list (otherwise it
> +     * might have stale data).
> +     */
> +    return 0;
> +}
>  
>  bool_t pt_irq_need_timer(uint32_t flags)
>  {
> @@ -97,10 +143,12 @@ int pt_irq_create_bind(
>      struct hvm_pirq_dpci *pirq_dpci;
>      struct pirq *info;
>      int rc, pirq = pt_irq_bind->machine_irq;
> +    s_time_t start = NOW();
>  
>      if ( pirq < 0 || pirq >= d->nr_pirqs )
>          return -EINVAL;
>  
> + restart:
>      spin_lock(&d->event_lock);
>  
>      hvm_irq_dpci = domain_get_irq_dpci(d);
> @@ -128,6 +176,25 @@ int pt_irq_create_bind(
>      }
>      pirq_dpci = pirq_dpci(info);
>      /*
> +     * A crude 'while' loop with us dropping the spinlock and giving
> +     * the softirq_dpci a chance to run.
> +     * We MUST check for this condition as the softirq could be scheduled
> +     * and hasn't run yet. We do this up to one second at which point we
> +     * give up. Note that this code replaced tasklet_kill which would have
> +     * spun forever and would do the same thing (wait to flush out
> +     * outstanding hvm_dirq_assist calls) - said tasklet_kill had been
> +     * in 'pt_pirq_cleanup_check' (called during pt_irq_destroy_bind)
> +     * and 'pci_clean_dpci_irq' (domain destroyed).
> +     */
> +    if ( pt_pirq_softirq_active(pirq_dpci) )
> +    {
> +        spin_unlock(&d->event_lock);
> +        process_pending_softirqs();
> +        if ( ( NOW() - start ) > SECONDS(1) )
> +            return -EAGAIN;
> +        goto restart;
> +    }

1s seems like overkill here, and far too long to loop in Xen context
even if we are processing softirqs.

Why is it so long?  Surely the softirq will be processed after the first
call to process_pending_softirqs(), or are we possibly waiting on
another pcpu to run its softirqs?

> +    /*
>       * The 'pt_irq_create_bind' can be called right after 
> 'pt_irq_destroy_bind'
>       * was called. The 'pirq_cleanup_check' which would free the structure
>       * is only called if the event channel for the PIRQ is active. However
> @@ -425,14 +492,12 @@ void pt_pirq_init(struct domain *d, struct 
> hvm_pirq_dpci *dpci)
>  {
>      INIT_LIST_HEAD(&dpci->digl_list);
>      dpci->gmsi.dest_vcpu_id = -1;
> -    softirq_tasklet_init(&dpci->tasklet, hvm_dirq_assist, (unsigned 
> long)dpci);
>  }
>  
>  bool_t pt_pirq_cleanup_check(struct hvm_pirq_dpci *dpci)
>  {
> -    if ( !dpci->flags )
> +    if ( !dpci->flags && !pt_pirq_softirq_active(dpci) )
>      {
> -        tasklet_kill(&dpci->tasklet);
>          dpci->dom = NULL;
>          return 1;
>      }
> @@ -475,8 +540,7 @@ int hvm_do_IRQ_dpci(struct domain *d, struct pirq *pirq)
>           !(pirq_dpci->flags & HVM_IRQ_DPCI_MAPPED) )
>          return 0;
>  
> -    pirq_dpci->masked = 1;
> -    tasklet_schedule(&pirq_dpci->tasklet);
> +    schedule_softirq_for(pirq_dpci);
>      return 1;
>  }
>  
> @@ -530,9 +594,8 @@ void hvm_dpci_msi_eoi(struct domain *d, int vector)
>      spin_unlock(&d->event_lock);
>  }
>  
> -static void hvm_dirq_assist(unsigned long arg)
> +static void hvm_dirq_assist(struct hvm_pirq_dpci *pirq_dpci)
>  {
> -    struct hvm_pirq_dpci *pirq_dpci = (struct hvm_pirq_dpci *)arg;
>      struct domain *d = pirq_dpci->dom;
>  
>      /*
> @@ -544,14 +607,12 @@ static void hvm_dirq_assist(unsigned long arg)
>       * do nothing except clear the ->masked field anyhow.
>       */
>      if ( !d )
> -    {
> -        pirq_dpci->masked = 0;
>          return;
> -    }
> +
>      ASSERT(d->arch.hvm_domain.irq.dpci);
>  
>      spin_lock(&d->event_lock);
> -    if ( test_and_clear_bool(pirq_dpci->masked) )
> +    if ( pirq_dpci->masked )
>      {
>          struct pirq *pirq = dpci_pirq(pirq_dpci);
>          const struct dev_intx_gsi_link *digl;
> @@ -653,3 +714,68 @@ void hvm_dpci_eoi(struct domain *d, unsigned int 
> guest_gsi,
>  unlock:
>      spin_unlock(&d->event_lock);
>  }
> +
> +static void dpci_softirq(void)
> +{
> +    struct hvm_pirq_dpci *pirq_dpci;
> +    unsigned int cpu = smp_processor_id();
> +    LIST_HEAD(our_list);
> +
> +    local_irq_disable();
> +    list_splice_init(&per_cpu(dpci_list, cpu), &our_list);
> +    local_irq_enable();
> +
> +    while ( !list_empty(&our_list) )
> +    {
> +        pirq_dpci = list_entry(our_list.next, struct hvm_pirq_dpci, 
> softirq_list);
> +        list_del(&pirq_dpci->softirq_list);
> +
> +        hvm_dirq_assist(pirq_dpci);
> +        pirq_dpci->masked = 0;
> +    }
> +}
> +
> +static int cpu_callback(
> +    struct notifier_block *nfb, unsigned long action, void *hcpu)
> +{
> +    unsigned int cpu = (unsigned long)hcpu;
> +
> +    switch ( action )
> +    {
> +    case CPU_UP_PREPARE:
> +        INIT_LIST_HEAD(&per_cpu(dpci_list, cpu));
> +        break;
> +    case CPU_UP_CANCELED:
> +    case CPU_DEAD:
> +        /*
> +         * On CPU_DYING this callback is called (on the CPU that is dying)
> +         * with an possible HVM_DPIC_SOFTIRQ pending - at which point we can
> +         * clear out any outstanding domains (by the virtue of the idle loop
> +         * calling the softirq later). In CPU_DEAD case the CPU is deaf and
> +         * there are no pending softirqs for us to handle so we can chill.
> +         */
> +        ASSERT(list_empty(&per_cpu(dpci_list, cpu)));
> +        break;
> +    default:
> +        break;
> +    }
> +
> +    return NOTIFY_DONE;
> +}
> +
> +static struct notifier_block cpu_nfb = {
> +    .notifier_call = cpu_callback,
> +};
> +
> +static int __init setup_dpci_softirq(void)
> +{
> +    unsigned int cpu;
> +
> +    for_each_online_cpu(cpu)
> +        INIT_LIST_HEAD(&per_cpu(dpci_list, cpu));

Hmm - its loops like this that make me think we should have a function
to explicitly initialise the percpu area of cpus at the point at which
they come up.  Still, this is fine for now, and another item for the
todo list.

> +
> +    open_softirq(HVM_DPCI_SOFTIRQ, dpci_softirq);
> +    register_cpu_notifier(&cpu_nfb);
> +    return 0;
> +}
> +__initcall(setup_dpci_softirq);
> diff --git a/xen/drivers/passthrough/pci.c b/xen/drivers/passthrough/pci.c
> index 81e8a3a..d147189 100644
> --- a/xen/drivers/passthrough/pci.c
> +++ b/xen/drivers/passthrough/pci.c
> @@ -767,40 +767,51 @@ static int pci_clean_dpci_irq(struct domain *d,
>          xfree(digl);
>      }
>  
> -    tasklet_kill(&pirq_dpci->tasklet);
> -
> -    return 0;
> +    return pt_pirq_softirq_active(pirq_dpci);
>  }
>  
> -static void pci_clean_dpci_irqs(struct domain *d)
> +static int pci_clean_dpci_irqs(struct domain *d)
>  {
>      struct hvm_irq_dpci *hvm_irq_dpci = NULL;
>  
>      if ( !iommu_enabled )
> -        return;
> +        return -ESRCH;

I realise that our return codes are inconsistent, but can we avoid
further overloading of return values.  -ESRCH is generally "no such domain"

Certain IOMMU related failures currently use -ENODEV, which is perhaps
more appropriate.

>  
>      if ( !is_hvm_domain(d) )
> -        return;
> +        return -EINVAL;
>  
>      spin_lock(&d->event_lock);
>      hvm_irq_dpci = domain_get_irq_dpci(d);
>      if ( hvm_irq_dpci != NULL )
>      {
> -        pt_pirq_iterate(d, pci_clean_dpci_irq, NULL);
> +        int ret = pt_pirq_iterate(d, pci_clean_dpci_irq, NULL);
> +
> +        if ( ret )
> +        {
> +            spin_unlock(&d->event_lock);
> +            return ret;
> +        }
>  
>          d->arch.hvm_domain.irq.dpci = NULL;
>          free_hvm_irq_dpci(hvm_irq_dpci);
>      }
>      spin_unlock(&d->event_lock);
> +    return 0;
>  }
>  
> -void pci_release_devices(struct domain *d)
> +int pci_release_devices(struct domain *d)
>  {
>      struct pci_dev *pdev;
>      u8 bus, devfn;
> +    int ret;
>  
>      spin_lock(&pcidevs_lock);
> -    pci_clean_dpci_irqs(d);
> +    ret = pci_clean_dpci_irqs(d);
> +    if ( ret == -EAGAIN )
> +    {
> +        spin_unlock(&pcidevs_lock);
> +        return ret;
> +    }
>      while ( (pdev = pci_get_pdev_by_domain(d, -1, -1, -1)) )
>      {
>          bus = pdev->bus;
> @@ -811,6 +822,8 @@ void pci_release_devices(struct domain *d)
>                     PCI_SLOT(devfn), PCI_FUNC(devfn));
>      }
>      spin_unlock(&pcidevs_lock);
> +
> +    return 0;
>  }
>  
>  #define PCI_CLASS_BRIDGE_HOST    0x0600
> diff --git a/xen/include/xen/hvm/irq.h b/xen/include/xen/hvm/irq.h
> index 94a550a..4fc6dcf 100644
> --- a/xen/include/xen/hvm/irq.h
> +++ b/xen/include/xen/hvm/irq.h
> @@ -99,7 +99,7 @@ struct hvm_pirq_dpci {
>      struct domain *dom;
>      struct hvm_gmsi_info gmsi;
>      struct timer timer;
> -    struct tasklet tasklet;
> +    struct list_head softirq_list;
>  };
>  
>  void pt_pirq_init(struct domain *, struct hvm_pirq_dpci *);
> @@ -109,6 +109,7 @@ int pt_pirq_iterate(struct domain *d,
>                                struct hvm_pirq_dpci *, void *arg),
>                      void *arg);
>  
> +int pt_pirq_softirq_active(struct hvm_pirq_dpci *);

"struct hvm_pirq_dpci *pirq_dpci" please.

~Andrew

>  /* Modify state of a PCI INTx wire. */
>  void hvm_pci_intx_assert(
>      struct domain *d, unsigned int device, unsigned int intx);
> diff --git a/xen/include/xen/pci.h b/xen/include/xen/pci.h
> index 91520bc..5f295f3 100644
> --- a/xen/include/xen/pci.h
> +++ b/xen/include/xen/pci.h
> @@ -99,7 +99,7 @@ struct pci_dev *pci_lock_domain_pdev(
>  
>  void setup_hwdom_pci_devices(struct domain *,
>                              int (*)(u8 devfn, struct pci_dev *));
> -void pci_release_devices(struct domain *d);
> +int pci_release_devices(struct domain *d);
>  int pci_add_segment(u16 seg);
>  const unsigned long *pci_get_ro_map(u16 seg);
>  int pci_add_device(u16 seg, u8 bus, u8 devfn, const struct pci_dev_info *);
> diff --git a/xen/include/xen/softirq.h b/xen/include/xen/softirq.h
> index 0895a16..f3da5df 100644
> --- a/xen/include/xen/softirq.h
> +++ b/xen/include/xen/softirq.h
> @@ -8,6 +8,9 @@ enum {
>      NEW_TLBFLUSH_CLOCK_PERIOD_SOFTIRQ,
>      RCU_SOFTIRQ,
>      TASKLET_SOFTIRQ,
> +#ifdef HAS_PASSTHROUGH
> +    HVM_DPCI_SOFTIRQ,
> +#endif
>      NR_COMMON_SOFTIRQS
>  };
>  



_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel