Do not allow runtime pm autosuspend to remove userspace GGTT mmaps too
quickly. For example, igt sets the autosuspend delay to 0, and so we
immediately attempt to perform runtime suspend upon releasing the
wakeref. Unfortunately, that involves tearing down GGTT mmaps as they
require an active device.
Override the autosuspend for GGTT mmaps, by keeping the wakeref around
for 250ms after populating the PTE for a fresh mmap.
v2: Prefer refcount_t for its under/overflow error detection
v3: Flush the user runtime autosuspend prior to system system.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190527115114.13448-1-chris@chris-wilson.co.uk
+config DRM_I915_USERFAULT_AUTOSUSPEND
+ int "Runtime autosuspend delay for userspace GGTT mmaps (ms)"
+ default 250 # milliseconds
+ help
+ On runtime suspend, as we suspend the device, we have to revoke
+ userspace GGTT mmaps and force userspace to take a pagefault on
+ their next access. The revocation and subsequent recreation of
+ the GGTT mmap can be very slow and so we impose a small hysteris
+ that complements the runtime-pm autosuspend and provides a lower
+ floor on the autosuspend delay.
+
+ May be 0 to disable the extra delay and solely use the device level
+ runtime pm autosuspend delay tunable.
+
config DRM_I915_SPIN_REQUEST
int
default 5 # microseconds
*/
struct list_head userfault_list;
+ /* Manual runtime pm autosuspend delay for user GGTT mmaps */
+ struct intel_wakeref_auto userfault_wakeref;
+
/**
* List of objects which are pending destruction.
*/
assert_rpm_wakelock_held(dev_priv);
if (!i915_vma_set_userfault(vma) && !obj->userfault_count++)
list_add(&obj->userfault_link, &dev_priv->mm.userfault_list);
+ if (CONFIG_DRM_I915_USERFAULT_AUTOSUSPEND)
+ intel_wakeref_auto(&dev_priv->mm.userfault_wakeref,
+ msecs_to_jiffies_timeout(CONFIG_DRM_I915_USERFAULT_AUTOSUSPEND));
GEM_BUG_ON(!obj->userfault_count);
i915_vma_set_ggtt_write(vma);
{
GEM_BUG_ON(dev_priv->gt.awake);
+ intel_wakeref_auto_fini(&dev_priv->mm.userfault_wakeref);
+
i915_gem_suspend_late(dev_priv);
intel_disable_gt_powersave(dev_priv);
INIT_LIST_HEAD(&i915->mm.unbound_list);
INIT_LIST_HEAD(&i915->mm.bound_list);
INIT_LIST_HEAD(&i915->mm.fence_list);
+
INIT_LIST_HEAD(&i915->mm.userfault_list);
+ intel_wakeref_auto_init(&i915->mm.userfault_wakeref, i915);
INIT_WORK(&i915->mm.free_work, __i915_gem_free_work);
}
{
GEM_TRACE("\n");
+ intel_wakeref_auto(&i915->mm.userfault_wakeref, 0);
flush_workqueue(i915->wq);
mutex_lock(&i915->drm.struct_mutex);
atomic_set(&wf->count, 0);
wf->wakeref = 0;
}
+
+static void wakeref_auto_timeout(struct timer_list *t)
+{
+ struct intel_wakeref_auto *wf = from_timer(wf, t, timer);
+ intel_wakeref_t wakeref;
+ unsigned long flags;
+
+ if (!refcount_dec_and_lock_irqsave(&wf->count, &wf->lock, &flags))
+ return;
+
+ wakeref = fetch_and_zero(&wf->wakeref);
+ spin_unlock_irqrestore(&wf->lock, flags);
+
+ intel_runtime_pm_put(wf->i915, wakeref);
+}
+
+void intel_wakeref_auto_init(struct intel_wakeref_auto *wf,
+ struct drm_i915_private *i915)
+{
+ spin_lock_init(&wf->lock);
+ timer_setup(&wf->timer, wakeref_auto_timeout, 0);
+ refcount_set(&wf->count, 0);
+ wf->wakeref = 0;
+ wf->i915 = i915;
+}
+
+void intel_wakeref_auto(struct intel_wakeref_auto *wf, unsigned long timeout)
+{
+ unsigned long flags;
+
+ if (!timeout) {
+ if (del_timer_sync(&wf->timer))
+ wakeref_auto_timeout(&wf->timer);
+ return;
+ }
+
+ /* Our mission is that we only extend an already active wakeref */
+ assert_rpm_wakelock_held(wf->i915);
+
+ if (!refcount_inc_not_zero(&wf->count)) {
+ spin_lock_irqsave(&wf->lock, flags);
+ if (!refcount_read(&wf->count)) {
+ GEM_BUG_ON(wf->wakeref);
+ wf->wakeref = intel_runtime_pm_get_if_in_use(wf->i915);
+ }
+ refcount_inc(&wf->count);
+ spin_unlock_irqrestore(&wf->lock, flags);
+ }
+
+ /*
+ * If we extend a pending timer, we will only get a single timer
+ * callback and so need to cancel the local inc by running the
+ * elided callback to keep the wf->count balanced.
+ */
+ if (mod_timer(&wf->timer, jiffies + timeout))
+ wakeref_auto_timeout(&wf->timer);
+}
+
+void intel_wakeref_auto_fini(struct intel_wakeref_auto *wf)
+{
+ intel_wakeref_auto(wf, 0);
+ GEM_BUG_ON(wf->wakeref);
+}
#include <linux/atomic.h>
#include <linux/mutex.h>
+#include <linux/refcount.h>
#include <linux/stackdepot.h>
+#include <linux/timer.h>
struct drm_i915_private;
return READ_ONCE(wf->wakeref);
}
+struct intel_wakeref_auto {
+ struct drm_i915_private *i915;
+ struct timer_list timer;
+ intel_wakeref_t wakeref;
+ spinlock_t lock;
+ refcount_t count;
+};
+
+/**
+ * intel_wakeref_auto: Delay the runtime-pm autosuspend
+ * @wf: the wakeref
+ * @timeout: relative timeout in jiffies
+ *
+ * The runtime-pm core uses a suspend delay after the last wakeref
+ * is released before triggering runtime suspend of the device. That
+ * delay is configurable via sysfs with little regard to the device
+ * characteristics. Instead, we want to tune the autosuspend based on our
+ * HW knowledge. intel_wakeref_auto() delays the sleep by the supplied
+ * timeout.
+ *
+ * Pass @timeout = 0 to cancel a previous autosuspend by executing the
+ * suspend immediately.
+ */
+void intel_wakeref_auto(struct intel_wakeref_auto *wf, unsigned long timeout);
+
+void intel_wakeref_auto_init(struct intel_wakeref_auto *wf,
+ struct drm_i915_private *i915);
+void intel_wakeref_auto_fini(struct intel_wakeref_auto *wf);
+
#endif /* INTEL_WAKEREF_H */
OSDN Git Service
