Re: [Intel-gfx] [PATCH 3/3] drm/i915/perf: add interrupt enabling parameter
On Mon, 13 Apr 2020 08:48:22 -0700, Umesh Nerlige Ramappa wrote:
>
> @@ -556,16 +559,23 @@ static bool oa_buffer_check_unlocked(struct
> i915_perf_stream *stream)
> * waiting on an event to occur. These checks are redundant when hrtimer
> events
> * will call oa_buffer_check_unlocked to update the oa_buffer tail pointers.
> The
> * redundant checks add cpu overhead. We simplify the check to reduce cpu
> - * overhead.
> + * overhead. For interrupt events, we still need to make sure that
> + * oa_buffer_check_unlocked is called when an interrupt occurs.
> */
> static bool oa_buffer_check_reports(struct i915_perf_stream *stream)
> {
> unsigned long flags;
> bool available;
>
> - spin_lock_irqsave(>oa_buffer.ptr_lock, flags);
> - available = stream->oa_buffer.tail != stream->oa_buffer.head;
> - spin_unlock_irqrestore(>oa_buffer.ptr_lock, flags);
> + if (!stream->oa_interrupt_monitor) {
> + spin_lock_irqsave(>oa_buffer.ptr_lock, flags);
> + available = stream->oa_buffer.tail != stream->oa_buffer.head;
> + spin_unlock_irqrestore(>oa_buffer.ptr_lock, flags);
> + } else {
> + if (stream->half_full_count_last !=
> + atomic64_read(>half_full_count))
> + available = oa_buffer_check_unlocked(stream);
> + }
This logic is broken if we have both the interrupt and the timer enabled?
Anyway as I said for Patch 1, oa_buffer_check_reports() should not be
needed (and hopefully neither the half_full_count's as per the comment in
Patch 2).
> @@ -3710,13 +3736,16 @@ static int read_properties_unlocked(struct i915_perf
> *perf,
> break;
> }
> case DRM_I915_PERF_PROP_POLL_OA_PERIOD:
> - if (value < 10 /* 100us */) {
> + if (value > 0 && value < 10 /* 100us */) {
> DRM_DEBUG("OA availability timer too small
> (%lluns < 100us)\n",
> value);
> return -EINVAL;
> }
> props->poll_oa_period = value;
> break;
> + case DRM_I915_PERF_PROP_OA_ENABLE_INTERRUPT:
> + props->oa_interrupt_monitor = value != 0;
At one point it was fashionable to write this as '= !!value' but I believe
even that is not needed now and this can be written as '= value'.
> @@ -3725,6 +3754,19 @@ static int read_properties_unlocked(struct i915_perf
> *perf,
> uprop += 2;
> }
>
> + /*
> + * Blocking read need to be waken up by some mechanism. If no polling
> + * of the HEAD/TAIL register is done by the kernel and no interrupt is
> + * enabled, we'll never be able to wake up.
> + */
> + if ((open_flags & I915_PERF_FLAG_FD_NONBLOCK) == 0 &&
> + !props->poll_oa_period &&
> + !props->oa_interrupt_monitor) {
> + DRM_DEBUG("Requesting a blocking stream with no polling period "
> + "& no interrupt.\n");
> + return -EINVAL;
> + }
Shouldn't this be true for non-blocking reads too since the poll() can
block indefinitely if both timer and interrupt are disabled? I think we
should disallow disabling both in either case.
> diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
> index 14b67cd6b54b..947e65b937eb 100644
> --- a/include/uapi/drm/i915_drm.h
> +++ b/include/uapi/drm/i915_drm.h
> @@ -1987,12 +1987,23 @@ enum drm_i915_perf_property_id {
>* the driver if this parameter is not specified. Note that larger timer
>* values will reduce cpu consumption during OA perf captures. However,
>* excessively large values would potentially result in OA buffer
> - * overwrites as captures reach end of the OA buffer.
> + * overwrites as captures reach end of the OA buffer. A value of 0 means
> + * no hrtimer will be started.
>*
>* This property is available in perf revision 5.
>*/
> DRM_I915_PERF_PROP_POLL_OA_PERIOD,
>
> + /**
> + * Specifying this property sets up the interrupt mechanism for the OA
> + * buffer in i915. This option in conjunction with a long polling period
> + * for avaibility of OA data can reduce CPU load significantly if you
> + * do not care about OA data being read as soon as it's available.
> + *
> + * This property is available in perf revision 6.
> + */
> + DRM_I915_PERF_PROP_OA_ENABLE_INTERRUPT,
I am still torn about exposing this new control to userspace. If we don't
we can have the interrupt always enabled and just document that the timer
can be disabled and when the timer is disabled the the OA sampling will be
driven off the interrupt. Anyway, if we decide that exposing this to user
space is better or more explicit, I'll go with it.
In either case, we need to add to the documentation above
[Intel-gfx] [PATCH 3/3] drm/i915/perf: add interrupt enabling parameter
From: Lionel Landwerlin
This let's the application choose to be driven by the interrupt
mechanism of the HW. In conjunction with long periods for checks for
the availability of data on the CPU, this can reduce the CPU load when
doing capture of OA data.
v2: Version the new parameter (Joonas)
v3: Rebase (Umesh)
v4: Disable hrtimer if poll oa period is zero
Signed-off-by: Lionel Landwerlin
Signed-off-by: Umesh Nerlige Ramappa
---
drivers/gpu/drm/i915/i915_perf.c | 64 +++-
include/uapi/drm/i915_drm.h | 13 ++-
2 files changed, 67 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
index a9423af67bb8..d3396129d828 100644
--- a/drivers/gpu/drm/i915/i915_perf.c
+++ b/drivers/gpu/drm/i915/i915_perf.c
@@ -463,6 +463,7 @@ static u32 gen7_oa_hw_tail_read(struct i915_perf_stream
*stream)
*/
static bool oa_buffer_check_unlocked(struct i915_perf_stream *stream)
{
+ u64 half_full_count = atomic64_read(>half_full_count);
u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma);
int report_size = stream->oa_buffer.format_size;
unsigned long flags;
@@ -539,6 +540,8 @@ static bool oa_buffer_check_unlocked(struct
i915_perf_stream *stream)
pollin = OA_TAKEN(stream->oa_buffer.tail - gtt_offset,
stream->oa_buffer.head - gtt_offset) >= report_size;
+ stream->half_full_count_last = half_full_count;
+
spin_unlock_irqrestore(>oa_buffer.ptr_lock, flags);
return pollin;
@@ -556,16 +559,23 @@ static bool oa_buffer_check_unlocked(struct
i915_perf_stream *stream)
* waiting on an event to occur. These checks are redundant when hrtimer events
* will call oa_buffer_check_unlocked to update the oa_buffer tail pointers.
The
* redundant checks add cpu overhead. We simplify the check to reduce cpu
- * overhead.
+ * overhead. For interrupt events, we still need to make sure that
+ * oa_buffer_check_unlocked is called when an interrupt occurs.
*/
static bool oa_buffer_check_reports(struct i915_perf_stream *stream)
{
unsigned long flags;
bool available;
- spin_lock_irqsave(>oa_buffer.ptr_lock, flags);
- available = stream->oa_buffer.tail != stream->oa_buffer.head;
- spin_unlock_irqrestore(>oa_buffer.ptr_lock, flags);
+ if (!stream->oa_interrupt_monitor) {
+ spin_lock_irqsave(>oa_buffer.ptr_lock, flags);
+ available = stream->oa_buffer.tail != stream->oa_buffer.head;
+ spin_unlock_irqrestore(>oa_buffer.ptr_lock, flags);
+ } else {
+ if (stream->half_full_count_last !=
+ atomic64_read(>half_full_count))
+ available = oa_buffer_check_unlocked(stream);
+ }
return available;
}
@@ -1163,8 +1173,8 @@ static int gen7_oa_read(struct i915_perf_stream *stream,
* @stream: An i915-perf stream opened for OA metrics
*
* Called when userspace tries to read() from a blocking stream FD opened
- * for OA metrics. It waits until the hrtimer callback finds a non-empty
- * OA buffer and wakes us.
+ * for OA metrics. It waits until either the hrtimer callback finds a non-empty
+ * OA buffer or the OA interrupt kicks in and wakes us.
*
* Note: it's acceptable to have this return with some false positives
* since any subsequent read handling will return -EAGAIN if there isn't
@@ -2696,7 +2706,7 @@ static void i915_oa_stream_enable(struct i915_perf_stream
*stream)
stream->perf->ops.oa_enable(stream);
- if (stream->periodic)
+ if (stream->periodic && stream->poll_oa_period)
hrtimer_start(>poll_check_timer,
ns_to_ktime(stream->poll_oa_period),
HRTIMER_MODE_REL_PINNED);
@@ -2770,6 +2780,10 @@ static void i915_oa_stream_disable(struct
i915_perf_stream *stream)
{
stream->perf->ops.oa_disable(stream);
+ stream->half_full_count_last = 0;
+ atomic64_set(>half_full_count,
+stream->half_full_count_last);
+
if (stream->periodic)
hrtimer_cancel(>poll_check_timer);
}
@@ -3137,6 +3151,16 @@ static __poll_t i915_perf_poll_locked(struct
i915_perf_stream *stream,
stream->ops->poll_wait(stream, file, wait);
+ /*
+* Only check the half buffer full notifications if requested by the
+* user.
+*/
+ if (stream->oa_interrupt_monitor &&
+ (stream->half_full_count_last !=
+atomic64_read(>half_full_count))) {
+ stream->pollin = oa_buffer_check_unlocked(stream);
+ }
+
/* Note: we don't explicitly check whether there's something to read
* here since this path may be very hot depending on what else
* userspace is polling, or on the timeout in use. We rely solely on
@@ -3554,6 +3578,7 @@ static u64 oa_exponent_to_ns(struct i915_perf *perf,
