[Intel-gfx] [PATCH] drm/i915/perf: Do not clear pollin for small user read buffers

[Lionel Landwerlin]

On 26/03/2020 06:43, Ashutosh Dixit wrote:
> It is wrong to block the user thread in the next poll when OA data is
> already available which could not fit in the user buffer provided in
> the previous read. In several cases the exact user buffer size is not
> known. Blocking user space in poll can lead to data loss when the
> buffer size used is smaller than the available data.
>
> This change fixes this issue and allows user space to read all OA data
> even when using a buffer size smaller than the available data using
> multiple non-blocking reads rather than staying blocked in poll till
> the next timer interrupt.
>
> v2: Fix ret value for blocking reads (Umesh)
>
> Cc: Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com>
> Cc: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
> Signed-off-by: Ashutosh Dixit <ashutosh.dixit at intel.com>
> ---
>   drivers/gpu/drm/i915/i915_perf.c | 63 ++++++--------------------------
>   1 file changed, 12 insertions(+), 51 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
> index 3222f6cd8255..e2d083efba6d 100644
> --- a/drivers/gpu/drm/i915/i915_perf.c
> +++ b/drivers/gpu/drm/i915/i915_perf.c
> @@ -2957,49 +2957,6 @@ void i915_oa_init_reg_state(const struct intel_context *ce,
>   		gen8_update_reg_state_unlocked(ce, stream);
>   }
>   
> -/**
> - * i915_perf_read_locked - &i915_perf_stream_ops->read with error normalisation
> - * @stream: An i915 perf stream
> - * @file: An i915 perf stream file
> - * @buf: destination buffer given by userspace
> - * @count: the number of bytes userspace wants to read
> - * @ppos: (inout) file seek position (unused)
> - *
> - * Besides wrapping &i915_perf_stream_ops->read this provides a common place to
> - * ensure that if we've successfully copied any data then reporting that takes
> - * precedence over any internal error status, so the data isn't lost.
> - *
> - * For example ret will be -ENOSPC whenever there is more buffered data than
> - * can be copied to userspace, but that's only interesting if we weren't able
> - * to copy some data because it implies the userspace buffer is too small to
> - * receive a single record (and we never split records).
> - *
> - * Another case with ret == -EFAULT is more of a grey area since it would seem
> - * like bad form for userspace to ask us to overrun its buffer, but the user
> - * knows best:
> - *
> - *   http://yarchive.net/comp/linux/partial_reads_writes.html
> - *
> - * Returns: The number of bytes copied or a negative error code on failure.
> - */
> -static ssize_t i915_perf_read_locked(struct i915_perf_stream *stream,
> -				     struct file *file,
> -				     char __user *buf,
> -				     size_t count,
> -				     loff_t *ppos)
> -{
> -	/* Note we keep the offset (aka bytes read) separate from any
> -	 * error status so that the final check for whether we return
> -	 * the bytes read with a higher precedence than any error (see
> -	 * comment below) doesn't need to be handled/duplicated in
> -	 * stream->ops->read() implementations.
> -	 */
> -	size_t offset = 0;
> -	int ret = stream->ops->read(stream, buf, count, &offset);
> -
> -	return offset ?: (ret ?: -EAGAIN);
> -}
> -
>   /**
>    * i915_perf_read - handles read() FOP for i915 perf stream FDs
>    * @file: An i915 perf stream file
> @@ -3025,6 +2982,8 @@ static ssize_t i915_perf_read(struct file *file,
>   {
>   	struct i915_perf_stream *stream = file->private_data;
>   	struct i915_perf *perf = stream->perf;
> +	size_t offset = 0;
> +	int __ret;
>   	ssize_t ret;
>   
>   	/* To ensure it's handled consistently we simply treat all reads of a
> @@ -3048,16 +3007,19 @@ static ssize_t i915_perf_read(struct file *file,
>   				return ret;
>   
>   			mutex_lock(&perf->lock);
> -			ret = i915_perf_read_locked(stream, file,
> -						    buf, count, ppos);
> +			__ret = stream->ops->read(stream, buf, count, &offset);
> +			ret = offset ?: (__ret ?: -EAGAIN);
>   			mutex_unlock(&perf->lock);
>   		} while (ret == -EAGAIN);
>   	} else {
>   		mutex_lock(&perf->lock);
> -		ret = i915_perf_read_locked(stream, file, buf, count, ppos);
> +		__ret = stream->ops->read(stream, buf, count, &offset);
> +		ret = offset ?: (__ret ?: -EAGAIN);
>   		mutex_unlock(&perf->lock);
>   	}
>   
> +	/* Possible values for __ret are 0, -EFAULT, -ENOSPC, -EAGAIN, ... */
> +
>   	/* We allow the poll checking to sometimes report false positive EPOLLIN
>   	 * events where we might actually report EAGAIN on read() if there's
>   	 * not really any data available. In this situation though we don't
> @@ -3065,13 +3027,12 @@ static ssize_t i915_perf_read(struct file *file,
>   	 * and read() returning -EAGAIN. Clearing the oa.pollin state here
>   	 * effectively ensures we back off until the next hrtimer callback
>   	 * before reporting another EPOLLIN event.
> +	 * The exception to this is if ops->read() returned -ENOSPC which means
> +	 * that more OA data is available than could fit in the user provided
> +	 * buffer. In this case we want the next poll() call to not block.
>   	 */
> -	if (ret >= 0 || ret == -EAGAIN) {
> -		/* Maybe make ->pollin per-stream state if we support multiple
> -		 * concurrent streams in the future.
> -		 */
> +	if ((ret > 0 || ret == -EAGAIN) && __ret != -ENOSPC)
>   		stream->pollin = false;
> -	}
>   
>   	return ret;
>   }

I think this reset of the pollin field is in the wrong place in the driver.

The decision of whether pollin is true/false should be based off the 
difference between head/tail pointers.


In my opinion the best place to do this in at the end of 
gen7/8_append_oa_reports functions, under the stream->oa_buffer.ptr_lock.

If everything has been read up to the tail pointer, then there is 
nothing to wake up userspace for, otherwise leave pollin untouched.


-Lionel