| Message ID | 20200303221905.25866-7-umesh.nerlige.ramappa@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | drm/i915/perf: add OA interrupt support | expand |
On Tue, 03 Mar 2020 14:19:04 -0800, Umesh Nerlige Ramappa wrote: > > From: Lionel Landwerlin <lionel.g.landwerlin@intel.com> > > This let's the application choose to be driven by the interrupt > mechanism of the HW. In conjuction 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) > > Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com> > Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com> [snip] > + /** > + * Specifying this property sets up the interrupt mechanism for the OA > + * buffer in i915. This option in conjuction with a long polling delay > + * 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 5. > + */ > + DRM_I915_PERF_PROP_OA_ENABLE_INTERRUPT, What if we do not expose this parameter in the uapi at all and internally decide in i915 whether to leave the interrupt either always enabled or always disabled (and in that case always use the hrtimer)? This way we retain flexibility in i915 if hardware evolves in the future e.g. to use watermarks for the interrupt, without yielding control to userspace. Overall I feel we should avoid exposing unnecessary details of the internal implemenation to userspace, they would be neither interested in knowing internal details nor know how to properly use these parameters. Shouldn't the driver be able to make these kinds of decisions internally? At this point the only parameter which implicitly exposed to userspace is the hrtimer poll period, so perhaps all we need to do is to expose that in the uapi? Thoughts?
On 04/03/2020 07:47, Dixit, Ashutosh wrote: > On Tue, 03 Mar 2020 14:19:04 -0800, Umesh Nerlige Ramappa wrote: >> From: Lionel Landwerlin <lionel.g.landwerlin@intel.com> >> >> This let's the application choose to be driven by the interrupt >> mechanism of the HW. In conjuction 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) >> >> Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com> >> Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com> > [snip] > >> + /** >> + * Specifying this property sets up the interrupt mechanism for the OA >> + * buffer in i915. This option in conjuction with a long polling delay >> + * 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 5. >> + */ >> + DRM_I915_PERF_PROP_OA_ENABLE_INTERRUPT, > What if we do not expose this parameter in the uapi at all and internally > decide in i915 whether to leave the interrupt either always enabled or > always disabled (and in that case always use the hrtimer)? This way we > retain flexibility in i915 if hardware evolves in the future e.g. to use > watermarks for the interrupt, without yielding control to userspace. > > Overall I feel we should avoid exposing unnecessary details of the internal > implemenation to userspace, they would be neither interested in knowing > internal details nor know how to properly use these parameters. Shouldn't > the driver be able to make these kinds of decisions internally? > > At this point the only parameter which implicitly exposed to userspace is > the hrtimer poll period, so perhaps all we need to do is to expose that in > the uapi? Thoughts? I guess I agree with you. I can't remember why I exposed it to userspace. There might be one test that checks the stream reports LOST_BUFFER with no poll() wakeup, but I guess we could update it. -Lionel
On Tue, Mar 03, 2020 at 02:19:04PM -0800, Umesh Nerlige Ramappa wrote: >From: Lionel Landwerlin <lionel.g.landwerlin@intel.com> > >This let's the application choose to be driven by the interrupt >mechanism of the HW. In conjuction 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) > >Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com> >Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com> >--- > drivers/gpu/drm/i915/i915_perf.c | 58 +++++++++++++++++++++++--------- > include/uapi/drm/i915_drm.h | 10 ++++++ > 2 files changed, 53 insertions(+), 15 deletions(-) > >diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c >index 502961da840d..ab41cba85b40 100644 >--- a/drivers/gpu/drm/i915/i915_perf.c >+++ b/drivers/gpu/drm/i915/i915_perf.c >@@ -252,7 +252,7 @@ > * oa_buffer_check(). > * > * Most of the implementation details for this workaround are in >- * oa_buffer_check_unlocked() and _append_oa_reports() >+ * oa_buffer_check() and _append_oa_reports() > * > * Note for posterity: previously the driver used to define an effective tail > * pointer that lagged the real pointer by a 'tail margin' measured in bytes >@@ -447,8 +447,9 @@ static u32 gen7_oa_hw_tail_read(struct i915_perf_stream *stream) > } > > /** >- * oa_buffer_check_unlocked - check for data and update tail ptr state >+ * oa_buffer_check - check for data and update tail ptr state > * @stream: i915 stream instance >+ * @lock: whether to take the oa_buffer spin lock > * > * This is either called via fops (for blocking reads in user ctx) or the poll > * check hrtimer (atomic ctx) to check the OA buffer tail pointer and check >@@ -470,8 +471,9 @@ static u32 gen7_oa_hw_tail_read(struct i915_perf_stream *stream) > * > * Returns: %true if the OA buffer contains data, else %false > */ >-static bool oa_buffer_check_unlocked(struct i915_perf_stream *stream) >+static bool oa_buffer_check(struct i915_perf_stream *stream, bool lock) Hi Lionel, All callers seem to set the lock to true when calling oa_buffer_check(). Do you recall why the parameter was introduced? If not, we probably want to remove this change. Thanks, Umesh
On 10/03/2020 22:08, Umesh Nerlige Ramappa wrote: > On Tue, Mar 03, 2020 at 02:19:04PM -0800, Umesh Nerlige Ramappa wrote: >> From: Lionel Landwerlin <lionel.g.landwerlin@intel.com> >> >> This let's the application choose to be driven by the interrupt >> mechanism of the HW. In conjuction 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) >> >> Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com> >> Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com> >> --- >> drivers/gpu/drm/i915/i915_perf.c | 58 +++++++++++++++++++++++--------- >> include/uapi/drm/i915_drm.h | 10 ++++++ >> 2 files changed, 53 insertions(+), 15 deletions(-) >> >> diff --git a/drivers/gpu/drm/i915/i915_perf.c >> b/drivers/gpu/drm/i915/i915_perf.c >> index 502961da840d..ab41cba85b40 100644 >> --- a/drivers/gpu/drm/i915/i915_perf.c >> +++ b/drivers/gpu/drm/i915/i915_perf.c >> @@ -252,7 +252,7 @@ >> * oa_buffer_check(). >> * >> * Most of the implementation details for this workaround are in >> - * oa_buffer_check_unlocked() and _append_oa_reports() >> + * oa_buffer_check() and _append_oa_reports() >> * >> * Note for posterity: previously the driver used to define an >> effective tail >> * pointer that lagged the real pointer by a 'tail margin' measured >> in bytes >> @@ -447,8 +447,9 @@ static u32 gen7_oa_hw_tail_read(struct >> i915_perf_stream *stream) >> } >> >> /** >> - * oa_buffer_check_unlocked - check for data and update tail ptr state >> + * oa_buffer_check - check for data and update tail ptr state >> * @stream: i915 stream instance >> + * @lock: whether to take the oa_buffer spin lock >> * >> * This is either called via fops (for blocking reads in user ctx) or >> the poll >> * check hrtimer (atomic ctx) to check the OA buffer tail pointer and >> check >> @@ -470,8 +471,9 @@ static u32 gen7_oa_hw_tail_read(struct >> i915_perf_stream *stream) >> * >> * Returns: %true if the OA buffer contains data, else %false >> */ >> -static bool oa_buffer_check_unlocked(struct i915_perf_stream *stream) >> +static bool oa_buffer_check(struct i915_perf_stream *stream, bool lock) > > Hi Lionel, > > All callers seem to set the lock to true when calling > oa_buffer_check(). Do you recall why the parameter was introduced? > If not, we probably want to remove this change. > > Thanks, > Umesh Err... Sorry, I don't remember. It's probably a leftover the initial iteration where I was trying to get the OA head/tail register from the interrupt. I guess you can drop that param and leave the function with the _unlocked prefix. Thanks, -Lionel
diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c index 502961da840d..ab41cba85b40 100644 --- a/drivers/gpu/drm/i915/i915_perf.c +++ b/drivers/gpu/drm/i915/i915_perf.c @@ -252,7 +252,7 @@ * oa_buffer_check(). * * Most of the implementation details for this workaround are in - * oa_buffer_check_unlocked() and _append_oa_reports() + * oa_buffer_check() and _append_oa_reports() * * Note for posterity: previously the driver used to define an effective tail * pointer that lagged the real pointer by a 'tail margin' measured in bytes @@ -447,8 +447,9 @@ static u32 gen7_oa_hw_tail_read(struct i915_perf_stream *stream) } /** - * oa_buffer_check_unlocked - check for data and update tail ptr state + * oa_buffer_check - check for data and update tail ptr state * @stream: i915 stream instance + * @lock: whether to take the oa_buffer spin lock * * This is either called via fops (for blocking reads in user ctx) or the poll * check hrtimer (atomic ctx) to check the OA buffer tail pointer and check @@ -470,8 +471,9 @@ static u32 gen7_oa_hw_tail_read(struct i915_perf_stream *stream) * * Returns: %true if the OA buffer contains data, else %false */ -static bool oa_buffer_check_unlocked(struct i915_perf_stream *stream) +static bool oa_buffer_check(struct i915_perf_stream *stream, bool lock) { + u64 half_full_count = atomic64_read(&stream->half_full_count); u32 gtt_offset = i915_ggtt_offset(stream->oa_buffer.vma); int report_size = stream->oa_buffer.format_size; unsigned long flags; @@ -482,7 +484,8 @@ static bool oa_buffer_check_unlocked(struct i915_perf_stream *stream) * could result in an OA buffer reset which might reset the head, * tails[] and aged_tail state. */ - spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags); + if (lock) + spin_lock_irqsave(&stream->oa_buffer.ptr_lock, flags); hw_tail = stream->perf->ops.oa_hw_tail_read(stream); @@ -558,7 +561,10 @@ static bool oa_buffer_check_unlocked(struct i915_perf_stream *stream) stream->oa_buffer.aging_timestamp = now; } - spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags); + stream->half_full_count_last = half_full_count; + + if (lock) + spin_unlock_irqrestore(&stream->oa_buffer.ptr_lock, flags); return OA_TAKEN(stream->oa_buffer.tail - gtt_offset, stream->oa_buffer.head - gtt_offset) >= report_size; @@ -1169,9 +1175,9 @@ static int gen7_oa_read(struct i915_perf_stream *stream, * i915_oa_wait_unlocked - handles blocking IO until OA data available * @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. + * Called when userspace tries to read() from a blocking stream FD opened 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 @@ -1186,7 +1192,7 @@ static int i915_oa_wait_unlocked(struct i915_perf_stream *stream) return -EIO; return wait_event_interruptible(stream->poll_wq, - oa_buffer_check_unlocked(stream)); + oa_buffer_check(stream, true)); } /** @@ -2733,6 +2739,10 @@ 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(&stream->half_full_count, + stream->half_full_count_last); + if (stream->periodic) hrtimer_cancel(&stream->poll_check_timer); } @@ -3075,7 +3085,7 @@ static enum hrtimer_restart oa_poll_check_timer_cb(struct hrtimer *hrtimer) struct i915_perf_stream *stream = container_of(hrtimer, typeof(*stream), poll_check_timer); - if (oa_buffer_check_unlocked(stream)) { + if (oa_buffer_check(stream, true)) { stream->pollin = true; wake_up(&stream->poll_wq); } @@ -3109,6 +3119,16 @@ 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(&stream->half_full_count))) { + stream->pollin = oa_buffer_check(stream, true); + } + /* 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 @@ -3667,6 +3687,9 @@ static int read_properties_unlocked(struct i915_perf *perf, } props->poll_oa_period = value; break; + case DRM_I915_PERF_PROP_OA_ENABLE_INTERRUPT: + props->oa_interrupt_monitor = value != 0; + break; case DRM_I915_PERF_PROP_MAX: MISSING_CASE(id); return -EINVAL; @@ -3677,12 +3700,14 @@ static int read_properties_unlocked(struct i915_perf *perf, /* * Blocking read need to be waken up by some mechanism. If no polling - * of the HEAD/TAIL register is done by the kernel, we'll never be - * able to wake up. + * 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) { - DRM_DEBUG("Requesting a blocking stream with no polling period.\n"); + !props->poll_oa_period && + !props->oa_interrupt_monitor) { + DRM_DEBUG("Requesting a blocking stream with no polling period " + "& no interrupt.\n"); return -EINVAL; } @@ -4523,8 +4548,11 @@ int i915_perf_ioctl_version(void) * 4: Add DRM_I915_PERF_PROP_POLL_OA_DELAY parameter that controls * enable/disable as well as the interval for the hrtimer used to * check for OA data. + * + * 5: Add DRM_I915_PERF_PROP_OA_ENABLE_INTERRUPT paramter to + * enable/disable interrupts in OA. */ - return 4; + return 5; } #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h index 131cb237d19c..f609ff4ceccb 100644 --- a/include/uapi/drm/i915_drm.h +++ b/include/uapi/drm/i915_drm.h @@ -1979,6 +1979,16 @@ enum drm_i915_perf_property_id { */ DRM_I915_PERF_PROP_POLL_OA_DELAY, + /** + * Specifying this property sets up the interrupt mechanism for the OA + * buffer in i915. This option in conjuction with a long polling delay + * 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 5. + */ + DRM_I915_PERF_PROP_OA_ENABLE_INTERRUPT, + DRM_I915_PERF_PROP_MAX /* non-ABI */ };