RE: [PATCH 5/6] ACPI/EC: Cleanup QR_SC command processing by adding a kernel thread to poll EC events.

[Zheng, Lv]

Hi, Rafael

> From: Rafael J. Wysocki [mailto:rjw@xxxxxxxxxxxxx]
> Sent: Wednesday, November 12, 2014 9:17 AM
> 
> On Monday, November 03, 2014 01:16:37 PM Lv Zheng wrote:
> > There is wait code in the QR_SC command processing, which makes it not
> > suitable to be put into a work queue item (see bug 82611). And there is
> > case that the SCI_EVT cannot trigger GPE, though all commands have polling
> > mode implemented, the event cannot be polled (see bug 77431).
> >
> > So if the QR_SC command can be put into a seperate IRQ thread, then the
> > work queue will not be blocked by the QR_SC command processing and we can
> > also trigger polling using the thread. Using IRQ thread also allows us to
> > change the EC GPE handler into the threaded IRQ model when possible.
> >
> > This patch thus adds the IRQ polling thread for SCI_EVT polling and removes
> > QR_SC processing work item.
> >
> > The reasons why we do not put a loop in the event poller to poll event
> > until the returned query value is 0:
> >   Some platforms return non 0 query value even when SCI_EVT=0, if we put a
> >   loop in the poller, our command flush mechanism could never execute to
> >   an end thus the system suspending process could be blocked. One SCI_EVT
> >   triggering one QR_EC is current logic and has been proven to be working
> >   for so long time.
> >
> > The reasons why it is not implemented directly using threaded IRQ are:
> > 1. ACPICA doesn't support threaded IRQ as not all of the OSPMs support
> >    threaded IRQ while GPE handler registerations are done through ACPICA.
> > 2. The IRQ processing code need to be identical for both the IRQ handler
> >    and the thread callback, while currently, though the command GPE
> >    handling is ready for both IRQ and polling mode, only the event GPE is
> >    is polled in the event polling thread and the command is polled in the
> >    user threads.
> > So we use a standalone kernel thread, if the above situations are changed
> > in the future, we can easily convert the code into the threaded IRQ style.
> >
> > The old EC_FLAGS_QUERY_PENDING is converted to EC_FLAGS_EVENT_ENABLED in
> > this patch, so that its naming is consistent with EC_FLAGS_EVENT_PENDING.
> > The original flag doesn't co-work with SCI_EVT well, this patch refines
> > its usage by enforcing a event polling wakeup indication as:
> >   EC_FLAGS_EVENT_ENABLED && EC_FLAGS_EVENT_PENDING
> > So unless the both of the flags are set, the threaded event poller will
> > not be woken up.
> >
> > This patch invokes acpi_ec_submit_request() after having detected SCI_EVT
> > and invokes acpi_ec_complete_request() before having the QR_EC command
> > processed. This is useful for implementing GPE storm prevention for
> > malicous "level triggered" SCI_EVT. But the storm prevention is not
> > implemented in this patch.
> >
> > Since the acpi_ec_submit_request() invoked after detecting the SCI_EVT is
> > paired with acpi_ec_complete_request() invoked after completing QR_EC
> > command, acpi_ec_submit_flushable_request() then need to be modified to
> > allow QR_EC command to be submitted during this period to revert the
> > increased reference count. This period can be determined by the event
> > polling indication:
> >   EC_FLAGS_EVENT_ENABLED && EC_FLAGS_EVENT_PENDING
> > Without enhancing this check in acpi_ec_submit_flushable_request(), QR_EC
> > command will not be executed to decrease the reference count added after
> > detecting the SCI_EVT, thus the system suspending will be blocked because
> > the reference count equals to 2. Such check is common for flushing code.
> >
> > Reference: https://bugzilla.kernel.org/show_bug.cgi?id=82611
> > Reference: https://bugzilla.kernel.org/show_bug.cgi?id=77431
> > Signed-off-by: Lv Zheng <lv.zheng@xxxxxxxxx>
> > Tested-by: Ortwin GlÃck <odi@xxxxxx>
> > ---
> >  drivers/acpi/ec.c       |  194 ++++++++++++++++++++++++++++++++++-------------
> >  drivers/acpi/internal.h |    1 +
> >  2 files changed, 144 insertions(+), 51 deletions(-)
> >
> > diff --git a/drivers/acpi/ec.c b/drivers/acpi/ec.c
> > index a76794a..7089081 100644
> > --- a/drivers/acpi/ec.c
> > +++ b/drivers/acpi/ec.c
> > @@ -44,6 +44,7 @@
> >  #include <linux/slab.h>
> >  #include <linux/acpi.h>
> >  #include <linux/dmi.h>
> > +#include <linux/kthread.h>
> >  #include <asm/io.h>
> >
> >  #include "internal.h"
> > @@ -75,7 +76,8 @@ enum ec_command {
> >  					 * when trying to clear the EC */
> >
> >  enum {
> > -	EC_FLAGS_QUERY_PENDING,		/* Query is pending */
> > +	EC_FLAGS_EVENT_ENABLED,		/* Event is enabled */
> > +	EC_FLAGS_EVENT_PENDING,		/* Event is pending */
> >  	EC_FLAGS_GPE_STORM,		/* GPE storm detected */
> >  	EC_FLAGS_HANDLERS_INSTALLED,	/* Handlers for GPE and
> >  					 * OpReg are installed */
> > @@ -124,6 +126,7 @@ struct transaction {
> >  static struct acpi_ec_query_handler *
> >  acpi_ec_get_query_handler(struct acpi_ec_query_handler *handler);
> >  static void acpi_ec_put_query_handler(struct acpi_ec_query_handler *handler);
> > +static int acpi_ec_sync_query(struct acpi_ec *ec, u8 *data);
> >
> >  struct acpi_ec *boot_ec, *first_ec;
> >  EXPORT_SYMBOL(first_ec);
> > @@ -149,6 +152,12 @@ static bool acpi_ec_flushed(struct acpi_ec *ec)
> >  	return ec->reference_count == 1;
> >  }
> >
> > +static bool acpi_ec_has_pending_event(struct acpi_ec *ec)
> > +{
> > +	return test_bit(EC_FLAGS_EVENT_ENABLED, &ec->flags) &&
> > +	       test_bit(EC_FLAGS_EVENT_PENDING, &ec->flags);
> > +}
> > +
> >  /* --------------------------------------------------------------------------
> >   *                           GPE Enhancement
> >   * -------------------------------------------------------------------------- */
> > @@ -177,20 +186,93 @@ static void acpi_ec_complete_request(struct acpi_ec *ec)
> >   *                                      the flush operation is not in
> >   *                                      progress
> >   * @ec: the EC device
> > + * @check_event: check whether event is pending
> >   *
> >   * This function must be used before taking a new action that should hold
> >   * the reference count.  If this function returns false, then the action
> >   * must be discarded or it will prevent the flush operation from being
> >   * completed.
> > + *
> > + * During flushing, QR_EC command need to pass this check when there is a
> > + * pending event, so that the reference count held for the pending event
> > + * can be decreased by the completion of the QR_EC command.
> >   */
> > -static bool acpi_ec_submit_flushable_request(struct acpi_ec *ec)
> > +static bool acpi_ec_submit_flushable_request(struct acpi_ec *ec,
> > +					     bool check_event)
> >  {
> > -	if (!acpi_ec_started(ec))
> > -		return false;
> > +	if (!acpi_ec_started(ec)) {
> > +		if (!check_event || !acpi_ec_has_pending_event(ec))
> > +			return false;
> > +	}
> >  	acpi_ec_submit_request(ec);
> >  	return true;
> >  }
> >
> > +static void acpi_ec_enable_event(struct acpi_ec *ec)
> > +{
> > +	unsigned long flags;
> > +
> > +	spin_lock_irqsave(&ec->lock, flags);
> > +	/* Hold reference for pending event */
> > +	if (!acpi_ec_submit_flushable_request(ec, false)) {
> > +		spin_unlock_irqrestore(&ec->lock, flags);
> > +		return;
> > +	}
> > +	set_bit(EC_FLAGS_EVENT_ENABLED, &ec->flags);
> > +	if (test_bit(EC_FLAGS_EVENT_PENDING, &ec->flags)) {
> > +		pr_debug("***** Event pending *****\n");
> > +		wake_up_process(ec->thread);
> > +		spin_unlock_irqrestore(&ec->lock, flags);
> > +		return;
> > +	}
> > +	acpi_ec_complete_request(ec);
> > +	spin_unlock_irqrestore(&ec->lock, flags);
> > +}
> > +
> > +static void __acpi_ec_set_event(struct acpi_ec *ec)
> > +{
> > +	/* Hold reference for pending event */
> > +	if (!acpi_ec_submit_flushable_request(ec, false))
> > +		return;
> > +	if (!test_and_set_bit(EC_FLAGS_EVENT_PENDING, &ec->flags)) {
> > +		pr_debug("***** Event pending *****\n");
> > +		if (test_bit(EC_FLAGS_EVENT_ENABLED, &ec->flags)) {
> > +			wake_up_process(ec->thread);
> > +			return;
> > +		}
> > +	}
> > +	acpi_ec_complete_request(ec);
> > +}
> > +
> > +static void __acpi_ec_complete_event(struct acpi_ec *ec)
> > +{
> > +	if (test_and_clear_bit(EC_FLAGS_EVENT_PENDING, &ec->flags)) {
> > +		/* Unhold reference for pending event */
> > +		acpi_ec_complete_request(ec);
> > +		pr_debug("***** Event running *****\n");
> > +	}
> > +}
> > +
> > +int acpi_ec_wait_for_event(struct acpi_ec *ec)
> > +{
> > +	unsigned long flags;
> > +
> > +	set_current_state(TASK_INTERRUPTIBLE);
> > +	while (!kthread_should_stop()) {
> > +		spin_lock_irqsave(&ec->lock, flags);
> > +		if (acpi_ec_has_pending_event(ec)) {
> > +			spin_unlock_irqrestore(&ec->lock, flags);
> > +			__set_current_state(TASK_RUNNING);
> > +			return 0;
> > +		}
> > +		spin_unlock_irqrestore(&ec->lock, flags);
> > +		schedule();
> > +		set_current_state(TASK_INTERRUPTIBLE);
> > +	}
> > +	__set_current_state(TASK_RUNNING);
> > +	return -1;
> > +}
> > +
> >  /* --------------------------------------------------------------------------
> >   *                           Transaction Management
> >   * -------------------------------------------------------------------------- */
> > @@ -298,7 +380,7 @@ static bool advance_transaction(struct acpi_ec *ec)
> >  			t->flags |= ACPI_EC_COMMAND_COMPLETE;
> >  			wakeup = true;
> >  		}
> > -		return wakeup;
> > +		goto out;
> >  	} else {
> >  		if (EC_FLAGS_QUERY_HANDSHAKE &&
> >  		    !(status & ACPI_EC_FLAG_SCI) &&
> > @@ -312,9 +394,11 @@ static bool advance_transaction(struct acpi_ec *ec)
> >  		} else if ((status & ACPI_EC_FLAG_IBF) == 0) {
> >  			acpi_ec_write_cmd(ec, t->command);
> >  			t->flags |= ACPI_EC_COMMAND_POLL;
> > +			if (t->command == ACPI_EC_COMMAND_QUERY)
> > +				__acpi_ec_complete_event(ec);
> >  		} else
> >  			goto err;
> > -		return wakeup;
> > +		goto out;
> >  	}
> >  err:
> >  	/*
> > @@ -325,6 +409,10 @@ err:
> >  		if (in_interrupt() && t)
> >  			++t->irq_count;
> >  	}
> > +out:
> > +	if (status & ACPI_EC_FLAG_SCI &&
> > +	    (!t || t->flags & ACPI_EC_COMMAND_COMPLETE))
> > +		__acpi_ec_set_event(ec);
> >  	return wakeup;
> >  }
> >
> > @@ -335,17 +423,6 @@ static void start_transaction(struct acpi_ec *ec)
> >  	(void)advance_transaction(ec);
> >  }
> >
> > -static int acpi_ec_sync_query(struct acpi_ec *ec, u8 *data);
> > -
> > -static int ec_check_sci_sync(struct acpi_ec *ec, u8 state)
> > -{
> > -	if (state & ACPI_EC_FLAG_SCI) {
> > -		if (!test_and_set_bit(EC_FLAGS_QUERY_PENDING, &ec->flags))
> > -			return acpi_ec_sync_query(ec, NULL);
> > -	}
> > -	return 0;
> > -}
> > -
> >  static int ec_poll(struct acpi_ec *ec)
> >  {
> >  	unsigned long flags;
> > @@ -384,12 +461,13 @@ static int acpi_ec_transaction_unlocked(struct acpi_ec *ec,
> >  	unsigned long tmp;
> >  	int ret = 0;
> >
> > +	bool is_query = !!(t->command == ACPI_EC_COMMAND_QUERY);
> >  	if (EC_FLAGS_MSI)
> >  		udelay(ACPI_EC_MSI_UDELAY);
> >  	/* start transaction */
> >  	spin_lock_irqsave(&ec->lock, tmp);
> >  	/* Enable GPE for command processing (IBF=0/OBF=1) */
> > -	if (!acpi_ec_submit_flushable_request(ec)) {
> > +	if (!acpi_ec_submit_flushable_request(ec, is_query)) {
> >  		ret = -EINVAL;
> >  		goto unlock;
> >  	}
> > @@ -398,10 +476,6 @@ static int acpi_ec_transaction_unlocked(struct acpi_ec *ec,
> >  	pr_debug("***** Command(%s) started *****\n",
> >  		 acpi_ec_cmd_string(t->command));
> >  	start_transaction(ec);
> > -	if (ec->curr->command == ACPI_EC_COMMAND_QUERY) {
> > -		clear_bit(EC_FLAGS_QUERY_PENDING, &ec->flags);
> > -		pr_debug("***** Event stopped *****\n");
> > -	}
> >  	spin_unlock_irqrestore(&ec->lock, tmp);
> >  	ret = ec_poll(ec);
> >  	spin_lock_irqsave(&ec->lock, tmp);
> > @@ -440,8 +514,6 @@ static int acpi_ec_transaction(struct acpi_ec *ec, struct transaction *t)
> >
> >  	status = acpi_ec_transaction_unlocked(ec, t);
> >
> > -	/* check if we received SCI during transaction */
> > -	ec_check_sci_sync(ec, acpi_ec_read_status(ec));
> >  	if (test_bit(EC_FLAGS_GPE_STORM, &ec->flags)) {
> >  		msleep(1);
> >  		/* It is safe to enable the GPE outside of the transaction. */
> > @@ -792,29 +864,6 @@ static int acpi_ec_sync_query(struct acpi_ec *ec, u8 *data)
> >  	return 0;
> >  }
> >
> > -static void acpi_ec_gpe_query(void *ec_cxt)
> > -{
> > -	struct acpi_ec *ec = ec_cxt;
> > -
> > -	if (!ec)
> > -		return;
> > -	mutex_lock(&ec->mutex);
> > -	acpi_ec_sync_query(ec, NULL);
> > -	mutex_unlock(&ec->mutex);
> > -}
> > -
> > -static int ec_check_sci(struct acpi_ec *ec, u8 state)
> > -{
> > -	if (state & ACPI_EC_FLAG_SCI) {
> > -		if (!test_and_set_bit(EC_FLAGS_QUERY_PENDING, &ec->flags)) {
> > -			pr_debug("***** Event started *****\n");
> > -			return acpi_os_execute(OSL_NOTIFY_HANDLER,
> > -				acpi_ec_gpe_query, ec);
> > -		}
> > -	}
> > -	return 0;
> > -}
> > -
> >  static u32 acpi_ec_gpe_handler(acpi_handle gpe_device,
> >  	u32 gpe_number, void *data)
> >  {
> > @@ -825,10 +874,46 @@ static u32 acpi_ec_gpe_handler(acpi_handle gpe_device,
> >  	if (advance_transaction(ec))
> >  		wake_up(&ec->wait);
> >  	spin_unlock_irqrestore(&ec->lock, flags);
> > -	ec_check_sci(ec, acpi_ec_read_status(ec));
> >  	return ACPI_INTERRUPT_HANDLED | ACPI_REENABLE_GPE;
> >  }
> >
> > +static int acpi_ec_event_poller(void *context)
> > +{
> > +	struct acpi_ec *ec = context;
> > +
> > +	while (!acpi_ec_wait_for_event(ec)) {
> > +		pr_debug("***** Event poller started *****\n");
> > +		mutex_lock(&ec->mutex);
> > +		(void)acpi_ec_sync_query(ec, NULL);
> > +		mutex_unlock(&ec->mutex);
> > +		pr_debug("***** Event poller stopped *****\n");
> > +	}
> > +	return 0;
> > +}
> > +
> > +static int ec_create_event_poller(struct acpi_ec *ec)
> > +{
> > +	struct task_struct *t;
> > +
> > +	t = kthread_run(acpi_ec_event_poller, ec, "ec/gpe-%lu", ec->gpe);
> 
> Does it have to be a kernel thread?
> 
> What about using a workqueue instead?

Actually I just want to use threaded IRQ here in response to Andi Kleen's comment.
If acpi_irq is registered as threaded IRQ, then acpi_ec_event_poller() will be the callback from it.
Since ACPICA is not ready for threaded IRQ currently, we cannot proceed at this point.
So I copied the threaded IRQ code from kernel/irq/manage.c here to prepare threaded IRQ logics.

Using a separate work queue, we didn't decrease the kernel thread count.
And the code written for the work item cannot be derived when things are switched to the threaded IRQ.
So I used kthread here.

Thanks and best regards
-Lv

> 
> > +	if (IS_ERR(t)) {
> > +		pr_err("failed to create event poller %lu\n", ec->gpe);
> > +		return PTR_ERR(t);
> > +	}
> > +	get_task_struct(t);
> > +	ec->thread = t;
> > +	return 0;
> > +}
> > +
> > +static void ec_delete_event_poller(struct acpi_ec *ec)
> > +{
> > +	struct task_struct *t = ec->thread;
> > +
> > +	ec->thread = NULL;
> > +	kthread_stop(t);
> > +	put_task_struct(t);
> > +}
> > +
> >  /* --------------------------------------------------------------------------
> >   *                           Address Space Management
> >   * -------------------------------------------------------------------------- */
> > @@ -884,7 +969,6 @@ static struct acpi_ec *make_acpi_ec(void)
> >
> >  	if (!ec)
> >  		return NULL;
> > -	ec->flags = 1 << EC_FLAGS_QUERY_PENDING;
> >  	mutex_init(&ec->mutex);
> >  	init_waitqueue_head(&ec->wait);
> >  	INIT_LIST_HEAD(&ec->list);
> > @@ -940,15 +1024,21 @@ ec_parse_device(acpi_handle handle, u32 Level, void *context, void **retval)
> >
> >  static int ec_install_handlers(struct acpi_ec *ec)
> >  {
> > +	int ret;
> >  	acpi_status status;
> >
> >  	if (test_bit(EC_FLAGS_HANDLERS_INSTALLED, &ec->flags))
> >  		return 0;
> > +	ret = ec_create_event_poller(ec);
> > +	if (ret)
> > +		return ret;
> >  	status = acpi_install_gpe_handler(NULL, ec->gpe,
> >  				  ACPI_GPE_EDGE_TRIGGERED,
> >  				  &acpi_ec_gpe_handler, ec);
> > -	if (ACPI_FAILURE(status))
> > +	if (ACPI_FAILURE(status)) {
> > +		ec_delete_event_poller(ec);
> >  		return -ENODEV;
> > +	}
> >
> >  	acpi_ec_start(ec);
> >  	status = acpi_install_address_space_handler(ec->handle,
> > @@ -968,6 +1058,7 @@ static int ec_install_handlers(struct acpi_ec *ec)
> >  			acpi_ec_stop(ec);
> >  			acpi_remove_gpe_handler(NULL, ec->gpe,
> >  				&acpi_ec_gpe_handler);
> > +			ec_delete_event_poller(ec);
> >  			return -ENODEV;
> >  		}
> >  	}
> > @@ -985,6 +1076,7 @@ static void ec_remove_handlers(struct acpi_ec *ec)
> >  	if (ACPI_FAILURE(acpi_remove_gpe_handler(NULL, ec->gpe,
> >  				&acpi_ec_gpe_handler)))
> >  		pr_err("failed to remove gpe handler\n");
> > +	ec_delete_event_poller(ec);
> >  	clear_bit(EC_FLAGS_HANDLERS_INSTALLED, &ec->flags);
> >  }
> >
> > @@ -1032,7 +1124,7 @@ static int acpi_ec_add(struct acpi_device *device)
> >  	ret = ec_install_handlers(ec);
> >
> >  	/* EC is fully operational, allow queries */
> > -	clear_bit(EC_FLAGS_QUERY_PENDING, &ec->flags);
> > +	acpi_ec_enable_event(ec);
> >
> >  	/* Clear stale _Q events if hardware might require that */
> >  	if (EC_FLAGS_CLEAR_ON_RESUME) {
> > diff --git a/drivers/acpi/internal.h b/drivers/acpi/internal.h
> > index bbcfe0b..20a569c 100644
> > --- a/drivers/acpi/internal.h
> > +++ b/drivers/acpi/internal.h
> > @@ -128,6 +128,7 @@ struct acpi_ec {
> >  	struct list_head list;
> >  	struct transaction *curr;
> >  	spinlock_t lock;
> > +	struct task_struct *thread;
> >  };
> >
> >  extern struct acpi_ec *first_ec;
> >
> 
> --
> I speak only for myself.
> Rafael J. Wysocki, Intel Open Source Technology Center.