drm/i915: Enable i915 perf stream for Haswell OA unit
authorRobert Bragg <robert@sixbynine.org>
Mon, 7 Nov 2016 19:49:52 +0000 (19:49 +0000)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Tue, 22 Nov 2016 13:38:13 +0000 (14:38 +0100)
Gen graphics hardware can be set up to periodically write snapshots of
performance counters into a circular buffer via its Observation
Architecture and this patch exposes that capability to userspace via the
i915 perf interface.

v2:
   Make sure to initialize ->specific_ctx_id when opening, without
   relying on _pin_notify hook, in case ctx already pinned.
v3:
   Revert back to pinning ctx upfront when opening stream, removing
   need to hook in to pinning and to update OACONTROL on the fly.

Signed-off-by: Robert Bragg <robert@sixbynine.org>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/20161107194957.3385-7-robert@sixbynine.org
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_perf.c
drivers/gpu/drm/i915/i915_reg.h
include/uapi/drm/i915_drm.h

index b11506f..3a69e21 100644 (file)
@@ -1797,6 +1797,11 @@ struct intel_wm_config {
        bool sprites_scaled;
 };
 
+struct i915_oa_format {
+       u32 format;
+       int size;
+};
+
 struct i915_oa_reg {
        i915_reg_t addr;
        u32 value;
@@ -1817,11 +1822,6 @@ struct i915_perf_stream_ops {
         */
        void (*disable)(struct i915_perf_stream *stream);
 
-       /* Return: true if any i915 perf records are ready to read()
-        * for this stream.
-        */
-       bool (*can_read)(struct i915_perf_stream *stream);
-
        /* Call poll_wait, passing a wait queue that will be woken
         * once there is something ready to read() for the stream
         */
@@ -1831,9 +1831,7 @@ struct i915_perf_stream_ops {
 
        /* For handling a blocking read, wait until there is something
         * to ready to read() for the stream. E.g. wait on the same
-        * wait queue that would be passed to poll_wait() until
-        * ->can_read() returns true (if its safe to call ->can_read()
-        * without the i915 perf lock held).
+        * wait queue that would be passed to poll_wait().
         */
        int (*wait_unlocked)(struct i915_perf_stream *stream);
 
@@ -1873,11 +1871,28 @@ struct i915_perf_stream {
        struct list_head link;
 
        u32 sample_flags;
+       int sample_size;
 
        struct i915_gem_context *ctx;
        bool enabled;
 
-       struct i915_perf_stream_ops *ops;
+       const struct i915_perf_stream_ops *ops;
+};
+
+struct i915_oa_ops {
+       void (*init_oa_buffer)(struct drm_i915_private *dev_priv);
+       int (*enable_metric_set)(struct drm_i915_private *dev_priv);
+       void (*disable_metric_set)(struct drm_i915_private *dev_priv);
+       void (*oa_enable)(struct drm_i915_private *dev_priv);
+       void (*oa_disable)(struct drm_i915_private *dev_priv);
+       void (*update_oacontrol)(struct drm_i915_private *dev_priv);
+       void (*update_hw_ctx_id_locked)(struct drm_i915_private *dev_priv,
+                                       u32 ctx_id);
+       int (*read)(struct i915_perf_stream *stream,
+                   char __user *buf,
+                   size_t count,
+                   size_t *offset);
+       bool (*oa_buffer_is_empty)(struct drm_i915_private *dev_priv);
 };
 
 struct drm_i915_private {
@@ -2177,16 +2192,47 @@ struct drm_i915_private {
 
        struct {
                bool initialized;
+
                struct mutex lock;
                struct list_head streams;
 
+               spinlock_t hook_lock;
+
                struct {
-                       u32 metrics_set;
+                       struct i915_perf_stream *exclusive_stream;
+
+                       u32 specific_ctx_id;
+                       struct i915_vma *pinned_rcs_vma;
+
+                       struct hrtimer poll_check_timer;
+                       wait_queue_head_t poll_wq;
+                       bool pollin;
+
+                       bool periodic;
+                       int period_exponent;
+                       int timestamp_frequency;
+
+                       int tail_margin;
+
+                       int metrics_set;
 
                        const struct i915_oa_reg *mux_regs;
                        int mux_regs_len;
                        const struct i915_oa_reg *b_counter_regs;
                        int b_counter_regs_len;
+
+                       struct {
+                               struct i915_vma *vma;
+                               u8 *vaddr;
+                               int format;
+                               int format_size;
+                       } oa_buffer;
+
+                       u32 gen7_latched_oastatus1;
+
+                       struct i915_oa_ops ops;
+                       const struct i915_oa_format *oa_formats;
+                       int n_builtin_sets;
                } oa;
        } perf;
 
index 777ce65..54653bc 100644 (file)
  */
 
 #include <linux/anon_inodes.h>
+#include <linux/sizes.h>
 
 #include "i915_drv.h"
+#include "i915_oa_hsw.h"
+
+/* HW requires this to be a power of two, between 128k and 16M, though driver
+ * is currently generally designed assuming the largest 16M size is used such
+ * that the overflow cases are unlikely in normal operation.
+ */
+#define OA_BUFFER_SIZE         SZ_16M
+
+#define OA_TAKEN(tail, head)   ((tail - head) & (OA_BUFFER_SIZE - 1))
+
+/* There's a HW race condition between OA unit tail pointer register updates and
+ * writes to memory whereby the tail pointer can sometimes get ahead of what's
+ * been written out to the OA buffer so far.
+ *
+ * Although this can be observed explicitly by checking for a zeroed report-id
+ * field in tail reports, it seems preferable to account for this earlier e.g.
+ * as part of the _oa_buffer_is_empty checks to minimize -EAGAIN polling cycles
+ * in this situation.
+ *
+ * To give time for the most recent reports to land before they may be copied to
+ * userspace, the driver operates as if the tail pointer effectively lags behind
+ * the HW tail pointer by 'tail_margin' bytes. The margin in bytes is calculated
+ * based on this constant in nanoseconds, the current OA sampling exponent
+ * and current report size.
+ *
+ * There is also a fallback check while reading to simply skip over reports with
+ * a zeroed report-id.
+ */
+#define OA_TAIL_MARGIN_NSEC    100000ULL
+
+/* frequency for checking whether the OA unit has written new reports to the
+ * circular OA buffer...
+ */
+#define POLL_FREQUENCY 200
+#define POLL_PERIOD (NSEC_PER_SEC / POLL_FREQUENCY)
+
+/* The maximum exponent the hardware accepts is 63 (essentially it selects one
+ * of the 64bit timestamp bits to trigger reports from) but there's currently
+ * no known use case for sampling as infrequently as once per 47 thousand years.
+ *
+ * Since the timestamps included in OA reports are only 32bits it seems
+ * reasonable to limit the OA exponent where it's still possible to account for
+ * overflow in OA report timestamps.
+ */
+#define OA_EXPONENT_MAX 31
+
+#define INVALID_CTX_ID 0xffffffff
+
+
+/* XXX: beware if future OA HW adds new report formats that the current
+ * code assumes all reports have a power-of-two size and ~(size - 1) can
+ * be used as a mask to align the OA tail pointer.
+ */
+static struct i915_oa_format hsw_oa_formats[I915_OA_FORMAT_MAX] = {
+       [I915_OA_FORMAT_A13]        = { 0, 64 },
+       [I915_OA_FORMAT_A29]        = { 1, 128 },
+       [I915_OA_FORMAT_A13_B8_C8]  = { 2, 128 },
+       /* A29_B8_C8 Disallowed as 192 bytes doesn't factor into buffer size */
+       [I915_OA_FORMAT_B4_C8]      = { 4, 64 },
+       [I915_OA_FORMAT_A45_B8_C8]  = { 5, 256 },
+       [I915_OA_FORMAT_B4_C8_A16]  = { 6, 128 },
+       [I915_OA_FORMAT_C4_B8]      = { 7, 64 },
+};
+
+#define SAMPLE_OA_REPORT      (1<<0)
 
 struct perf_open_properties {
        u32 sample_flags;
 
        u64 single_context:1;
        u64 ctx_handle;
+
+       /* OA sampling state */
+       int metrics_set;
+       int oa_format;
+       bool oa_periodic;
+       int oa_period_exponent;
+};
+
+/* NB: This is either called via fops or the poll check hrtimer (atomic ctx)
+ *
+ * It's safe to read OA config state here unlocked, assuming that this is only
+ * called while the stream is enabled, while the global OA configuration can't
+ * be modified.
+ *
+ * Note: we don't lock around the head/tail reads even though there's the slim
+ * possibility of read() fop errors forcing a re-init of the OA buffer
+ * pointers.  A race here could result in a false positive !empty status which
+ * is acceptable.
+ */
+static bool gen7_oa_buffer_is_empty_fop_unlocked(struct drm_i915_private *dev_priv)
+{
+       int report_size = dev_priv->perf.oa.oa_buffer.format_size;
+       u32 oastatus2 = I915_READ(GEN7_OASTATUS2);
+       u32 oastatus1 = I915_READ(GEN7_OASTATUS1);
+       u32 head = oastatus2 & GEN7_OASTATUS2_HEAD_MASK;
+       u32 tail = oastatus1 & GEN7_OASTATUS1_TAIL_MASK;
+
+       return OA_TAKEN(tail, head) <
+               dev_priv->perf.oa.tail_margin + report_size;
+}
+
+/**
+ * Appends a status record to a userspace read() buffer.
+ */
+static int append_oa_status(struct i915_perf_stream *stream,
+                           char __user *buf,
+                           size_t count,
+                           size_t *offset,
+                           enum drm_i915_perf_record_type type)
+{
+       struct drm_i915_perf_record_header header = { type, 0, sizeof(header) };
+
+       if ((count - *offset) < header.size)
+               return -ENOSPC;
+
+       if (copy_to_user(buf + *offset, &header, sizeof(header)))
+               return -EFAULT;
+
+       (*offset) += header.size;
+
+       return 0;
+}
+
+/**
+ * Copies single OA report into userspace read() buffer.
+ */
+static int append_oa_sample(struct i915_perf_stream *stream,
+                           char __user *buf,
+                           size_t count,
+                           size_t *offset,
+                           const u8 *report)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+       int report_size = dev_priv->perf.oa.oa_buffer.format_size;
+       struct drm_i915_perf_record_header header;
+       u32 sample_flags = stream->sample_flags;
+
+       header.type = DRM_I915_PERF_RECORD_SAMPLE;
+       header.pad = 0;
+       header.size = stream->sample_size;
+
+       if ((count - *offset) < header.size)
+               return -ENOSPC;
+
+       buf += *offset;
+       if (copy_to_user(buf, &header, sizeof(header)))
+               return -EFAULT;
+       buf += sizeof(header);
+
+       if (sample_flags & SAMPLE_OA_REPORT) {
+               if (copy_to_user(buf, report, report_size))
+                       return -EFAULT;
+       }
+
+       (*offset) += header.size;
+
+       return 0;
+}
+
+/**
+ * Copies all buffered OA reports into userspace read() buffer.
+ * @stream: An i915-perf stream opened for OA metrics
+ * @buf: destination buffer given by userspace
+ * @count: the number of bytes userspace wants to read
+ * @offset: (inout): the current position for writing into @buf
+ * @head_ptr: (inout): the current oa buffer cpu read position
+ * @tail: the current oa buffer gpu write position
+ *
+ * Returns 0 on success, negative error code on failure.
+ *
+ * Notably any error condition resulting in a short read (-ENOSPC or
+ * -EFAULT) will be returned even though one or more records may
+ * have been successfully copied. In this case it's up to the caller
+ * to decide if the error should be squashed before returning to
+ * userspace.
+ *
+ * Note: reports are consumed from the head, and appended to the
+ * tail, so the head chases the tail?... If you think that's mad
+ * and back-to-front you're not alone, but this follows the
+ * Gen PRM naming convention.
+ */
+static int gen7_append_oa_reports(struct i915_perf_stream *stream,
+                                 char __user *buf,
+                                 size_t count,
+                                 size_t *offset,
+                                 u32 *head_ptr,
+                                 u32 tail)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+       int report_size = dev_priv->perf.oa.oa_buffer.format_size;
+       u8 *oa_buf_base = dev_priv->perf.oa.oa_buffer.vaddr;
+       int tail_margin = dev_priv->perf.oa.tail_margin;
+       u32 gtt_offset = i915_ggtt_offset(dev_priv->perf.oa.oa_buffer.vma);
+       u32 mask = (OA_BUFFER_SIZE - 1);
+       u32 head;
+       u32 taken;
+       int ret = 0;
+
+       if (WARN_ON(!stream->enabled))
+               return -EIO;
+
+       head = *head_ptr - gtt_offset;
+       tail -= gtt_offset;
+
+       /* The OA unit is expected to wrap the tail pointer according to the OA
+        * buffer size and since we should never write a misaligned head
+        * pointer we don't expect to read one back either...
+        */
+       if (tail > OA_BUFFER_SIZE || head > OA_BUFFER_SIZE ||
+           head % report_size) {
+               DRM_ERROR("Inconsistent OA buffer pointer (head = %u, tail = %u): force restart\n",
+                         head, tail);
+               dev_priv->perf.oa.ops.oa_disable(dev_priv);
+               dev_priv->perf.oa.ops.oa_enable(dev_priv);
+               *head_ptr = I915_READ(GEN7_OASTATUS2) &
+                       GEN7_OASTATUS2_HEAD_MASK;
+               return -EIO;
+       }
+
+
+       /* The tail pointer increases in 64 byte increments, not in report_size
+        * steps...
+        */
+       tail &= ~(report_size - 1);
+
+       /* Move the tail pointer back by the current tail_margin to account for
+        * the possibility that the latest reports may not have really landed
+        * in memory yet...
+        */
+
+       if (OA_TAKEN(tail, head) < report_size + tail_margin)
+               return -EAGAIN;
+
+       tail -= tail_margin;
+       tail &= mask;
+
+       for (/* none */;
+            (taken = OA_TAKEN(tail, head));
+            head = (head + report_size) & mask) {
+               u8 *report = oa_buf_base + head;
+               u32 *report32 = (void *)report;
+
+               /* All the report sizes factor neatly into the buffer
+                * size so we never expect to see a report split
+                * between the beginning and end of the buffer.
+                *
+                * Given the initial alignment check a misalignment
+                * here would imply a driver bug that would result
+                * in an overrun.
+                */
+               if (WARN_ON((OA_BUFFER_SIZE - head) < report_size)) {
+                       DRM_ERROR("Spurious OA head ptr: non-integral report offset\n");
+                       break;
+               }
+
+               /* The report-ID field for periodic samples includes
+                * some undocumented flags related to what triggered
+                * the report and is never expected to be zero so we
+                * can check that the report isn't invalid before
+                * copying it to userspace...
+                */
+               if (report32[0] == 0) {
+                       DRM_ERROR("Skipping spurious, invalid OA report\n");
+                       continue;
+               }
+
+               ret = append_oa_sample(stream, buf, count, offset, report);
+               if (ret)
+                       break;
+
+               /* The above report-id field sanity check is based on
+                * the assumption that the OA buffer is initially
+                * zeroed and we reset the field after copying so the
+                * check is still meaningful once old reports start
+                * being overwritten.
+                */
+               report32[0] = 0;
+       }
+
+       *head_ptr = gtt_offset + head;
+
+       return ret;
+}
+
+static int gen7_oa_read(struct i915_perf_stream *stream,
+                       char __user *buf,
+                       size_t count,
+                       size_t *offset)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+       int report_size = dev_priv->perf.oa.oa_buffer.format_size;
+       u32 oastatus2;
+       u32 oastatus1;
+       u32 head;
+       u32 tail;
+       int ret;
+
+       if (WARN_ON(!dev_priv->perf.oa.oa_buffer.vaddr))
+               return -EIO;
+
+       oastatus2 = I915_READ(GEN7_OASTATUS2);
+       oastatus1 = I915_READ(GEN7_OASTATUS1);
+
+       head = oastatus2 & GEN7_OASTATUS2_HEAD_MASK;
+       tail = oastatus1 & GEN7_OASTATUS1_TAIL_MASK;
+
+       /* XXX: On Haswell we don't have a safe way to clear oastatus1
+        * bits while the OA unit is enabled (while the tail pointer
+        * may be updated asynchronously) so we ignore status bits
+        * that have already been reported to userspace.
+        */
+       oastatus1 &= ~dev_priv->perf.oa.gen7_latched_oastatus1;
+
+       /* We treat OABUFFER_OVERFLOW as a significant error:
+        *
+        * - The status can be interpreted to mean that the buffer is
+        *   currently full (with a higher precedence than OA_TAKEN()
+        *   which will start to report a near-empty buffer after an
+        *   overflow) but it's awkward that we can't clear the status
+        *   on Haswell, so without a reset we won't be able to catch
+        *   the state again.
+        *
+        * - Since it also implies the HW has started overwriting old
+        *   reports it may also affect our sanity checks for invalid
+        *   reports when copying to userspace that assume new reports
+        *   are being written to cleared memory.
+        *
+        * - In the future we may want to introduce a flight recorder
+        *   mode where the driver will automatically maintain a safe
+        *   guard band between head/tail, avoiding this overflow
+        *   condition, but we avoid the added driver complexity for
+        *   now.
+        */
+       if (unlikely(oastatus1 & GEN7_OASTATUS1_OABUFFER_OVERFLOW)) {
+               ret = append_oa_status(stream, buf, count, offset,
+                                      DRM_I915_PERF_RECORD_OA_BUFFER_LOST);
+               if (ret)
+                       return ret;
+
+               DRM_ERROR("OA buffer overflow: force restart\n");
+
+               dev_priv->perf.oa.ops.oa_disable(dev_priv);
+               dev_priv->perf.oa.ops.oa_enable(dev_priv);
+
+               oastatus2 = I915_READ(GEN7_OASTATUS2);
+               oastatus1 = I915_READ(GEN7_OASTATUS1);
+
+               head = oastatus2 & GEN7_OASTATUS2_HEAD_MASK;
+               tail = oastatus1 & GEN7_OASTATUS1_TAIL_MASK;
+       }
+
+       if (unlikely(oastatus1 & GEN7_OASTATUS1_REPORT_LOST)) {
+               ret = append_oa_status(stream, buf, count, offset,
+                                      DRM_I915_PERF_RECORD_OA_REPORT_LOST);
+               if (ret)
+                       return ret;
+               dev_priv->perf.oa.gen7_latched_oastatus1 |=
+                       GEN7_OASTATUS1_REPORT_LOST;
+       }
+
+       ret = gen7_append_oa_reports(stream, buf, count, offset,
+                                    &head, tail);
+
+       /* All the report sizes are a power of two and the
+        * head should always be incremented by some multiple
+        * of the report size.
+        *
+        * A warning here, but notably if we later read back a
+        * misaligned pointer we will treat that as a bug since
+        * it could lead to a buffer overrun.
+        */
+       WARN_ONCE(head & (report_size - 1),
+                 "i915: Writing misaligned OA head pointer");
+
+       /* Note: we update the head pointer here even if an error
+        * was returned since the error may represent a short read
+        * where some some reports were successfully copied.
+        */
+       I915_WRITE(GEN7_OASTATUS2,
+                  ((head & GEN7_OASTATUS2_HEAD_MASK) |
+                   OA_MEM_SELECT_GGTT));
+
+       return ret;
+}
+
+static int i915_oa_wait_unlocked(struct i915_perf_stream *stream)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+
+       /* We would wait indefinitely if periodic sampling is not enabled */
+       if (!dev_priv->perf.oa.periodic)
+               return -EIO;
+
+       /* Note: the oa_buffer_is_empty() condition is ok to run unlocked as it
+        * just performs mmio reads of the OA buffer head + tail pointers and
+        * it's assumed we're handling some operation that implies the stream
+        * can't be destroyed until completion (such as a read()) that ensures
+        * the device + OA buffer can't disappear
+        */
+       return wait_event_interruptible(dev_priv->perf.oa.poll_wq,
+                                       !dev_priv->perf.oa.ops.oa_buffer_is_empty(dev_priv));
+}
+
+static void i915_oa_poll_wait(struct i915_perf_stream *stream,
+                             struct file *file,
+                             poll_table *wait)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+
+       poll_wait(file, &dev_priv->perf.oa.poll_wq, wait);
+}
+
+static int i915_oa_read(struct i915_perf_stream *stream,
+                       char __user *buf,
+                       size_t count,
+                       size_t *offset)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+
+       return dev_priv->perf.oa.ops.read(stream, buf, count, offset);
+}
+
+/* Determine the render context hw id, and ensure it remains fixed for the
+ * lifetime of the stream. This ensures that we don't have to worry about
+ * updating the context ID in OACONTROL on the fly.
+ */
+static int oa_get_render_ctx_id(struct i915_perf_stream *stream)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+       struct i915_vma *vma;
+       int ret;
+
+       ret = i915_mutex_lock_interruptible(&dev_priv->drm);
+       if (ret)
+               return ret;
+
+       /* As the ID is the gtt offset of the context's vma we pin
+        * the vma to ensure the ID remains fixed.
+        *
+        * NB: implied RCS engine...
+        */
+       vma = i915_gem_context_pin_legacy(stream->ctx, 0);
+       if (IS_ERR(vma)) {
+               ret = PTR_ERR(vma);
+               goto unlock;
+       }
+
+       dev_priv->perf.oa.pinned_rcs_vma = vma;
+
+       /* Explicitly track the ID (instead of calling i915_ggtt_offset()
+        * on the fly) considering the difference with gen8+ and
+        * execlists
+        */
+       dev_priv->perf.oa.specific_ctx_id = i915_ggtt_offset(vma);
+
+unlock:
+       mutex_unlock(&dev_priv->drm.struct_mutex);
+
+       return ret;
+}
+
+static void oa_put_render_ctx_id(struct i915_perf_stream *stream)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+
+       mutex_lock(&dev_priv->drm.struct_mutex);
+
+       i915_vma_unpin(dev_priv->perf.oa.pinned_rcs_vma);
+       dev_priv->perf.oa.pinned_rcs_vma = NULL;
+
+       dev_priv->perf.oa.specific_ctx_id = INVALID_CTX_ID;
+
+       mutex_unlock(&dev_priv->drm.struct_mutex);
+}
+
+static void
+free_oa_buffer(struct drm_i915_private *i915)
+{
+       mutex_lock(&i915->drm.struct_mutex);
+
+       i915_gem_object_unpin_map(i915->perf.oa.oa_buffer.vma->obj);
+       i915_vma_unpin(i915->perf.oa.oa_buffer.vma);
+       i915_gem_object_put(i915->perf.oa.oa_buffer.vma->obj);
+
+       i915->perf.oa.oa_buffer.vma = NULL;
+       i915->perf.oa.oa_buffer.vaddr = NULL;
+
+       mutex_unlock(&i915->drm.struct_mutex);
+}
+
+static void i915_oa_stream_destroy(struct i915_perf_stream *stream)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+
+       BUG_ON(stream != dev_priv->perf.oa.exclusive_stream);
+
+       dev_priv->perf.oa.ops.disable_metric_set(dev_priv);
+
+       free_oa_buffer(dev_priv);
+
+       intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
+       intel_runtime_pm_put(dev_priv);
+
+       if (stream->ctx)
+               oa_put_render_ctx_id(stream);
+
+       dev_priv->perf.oa.exclusive_stream = NULL;
+}
+
+static void gen7_init_oa_buffer(struct drm_i915_private *dev_priv)
+{
+       u32 gtt_offset = i915_ggtt_offset(dev_priv->perf.oa.oa_buffer.vma);
+
+       /* Pre-DevBDW: OABUFFER must be set with counters off,
+        * before OASTATUS1, but after OASTATUS2
+        */
+       I915_WRITE(GEN7_OASTATUS2, gtt_offset | OA_MEM_SELECT_GGTT); /* head */
+       I915_WRITE(GEN7_OABUFFER, gtt_offset);
+       I915_WRITE(GEN7_OASTATUS1, gtt_offset | OABUFFER_SIZE_16M); /* tail */
+
+       /* On Haswell we have to track which OASTATUS1 flags we've
+        * already seen since they can't be cleared while periodic
+        * sampling is enabled.
+        */
+       dev_priv->perf.oa.gen7_latched_oastatus1 = 0;
+
+       /* NB: although the OA buffer will initially be allocated
+        * zeroed via shmfs (and so this memset is redundant when
+        * first allocating), we may re-init the OA buffer, either
+        * when re-enabling a stream or in error/reset paths.
+        *
+        * The reason we clear the buffer for each re-init is for the
+        * sanity check in gen7_append_oa_reports() that looks at the
+        * report-id field to make sure it's non-zero which relies on
+        * the assumption that new reports are being written to zeroed
+        * memory...
+        */
+       memset(dev_priv->perf.oa.oa_buffer.vaddr, 0, OA_BUFFER_SIZE);
+
+       /* Maybe make ->pollin per-stream state if we support multiple
+        * concurrent streams in the future.
+        */
+       dev_priv->perf.oa.pollin = false;
+}
+
+static int alloc_oa_buffer(struct drm_i915_private *dev_priv)
+{
+       struct drm_i915_gem_object *bo;
+       struct i915_vma *vma;
+       int ret;
+
+       if (WARN_ON(dev_priv->perf.oa.oa_buffer.vma))
+               return -ENODEV;
+
+       ret = i915_mutex_lock_interruptible(&dev_priv->drm);
+       if (ret)
+               return ret;
+
+       BUILD_BUG_ON_NOT_POWER_OF_2(OA_BUFFER_SIZE);
+       BUILD_BUG_ON(OA_BUFFER_SIZE < SZ_128K || OA_BUFFER_SIZE > SZ_16M);
+
+       bo = i915_gem_object_create(&dev_priv->drm, OA_BUFFER_SIZE);
+       if (IS_ERR(bo)) {
+               DRM_ERROR("Failed to allocate OA buffer\n");
+               ret = PTR_ERR(bo);
+               goto unlock;
+       }
+
+       ret = i915_gem_object_set_cache_level(bo, I915_CACHE_LLC);
+       if (ret)
+               goto err_unref;
+
+       /* PreHSW required 512K alignment, HSW requires 16M */
+       vma = i915_gem_object_ggtt_pin(bo, NULL, 0, SZ_16M, 0);
+       if (IS_ERR(vma)) {
+               ret = PTR_ERR(vma);
+               goto err_unref;
+       }
+       dev_priv->perf.oa.oa_buffer.vma = vma;
+
+       dev_priv->perf.oa.oa_buffer.vaddr =
+               i915_gem_object_pin_map(bo, I915_MAP_WB);
+       if (IS_ERR(dev_priv->perf.oa.oa_buffer.vaddr)) {
+               ret = PTR_ERR(dev_priv->perf.oa.oa_buffer.vaddr);
+               goto err_unpin;
+       }
+
+       dev_priv->perf.oa.ops.init_oa_buffer(dev_priv);
+
+       DRM_DEBUG_DRIVER("OA Buffer initialized, gtt offset = 0x%x, vaddr = %p\n",
+                        i915_ggtt_offset(dev_priv->perf.oa.oa_buffer.vma),
+                        dev_priv->perf.oa.oa_buffer.vaddr);
+
+       goto unlock;
+
+err_unpin:
+       __i915_vma_unpin(vma);
+
+err_unref:
+       i915_gem_object_put(bo);
+
+       dev_priv->perf.oa.oa_buffer.vaddr = NULL;
+       dev_priv->perf.oa.oa_buffer.vma = NULL;
+
+unlock:
+       mutex_unlock(&dev_priv->drm.struct_mutex);
+       return ret;
+}
+
+static void config_oa_regs(struct drm_i915_private *dev_priv,
+                          const struct i915_oa_reg *regs,
+                          int n_regs)
+{
+       int i;
+
+       for (i = 0; i < n_regs; i++) {
+               const struct i915_oa_reg *reg = regs + i;
+
+               I915_WRITE(reg->addr, reg->value);
+       }
+}
+
+static int hsw_enable_metric_set(struct drm_i915_private *dev_priv)
+{
+       int ret = i915_oa_select_metric_set_hsw(dev_priv);
+
+       if (ret)
+               return ret;
+
+       I915_WRITE(GDT_CHICKEN_BITS, (I915_READ(GDT_CHICKEN_BITS) |
+                                     GT_NOA_ENABLE));
+
+       /* PRM:
+        *
+        * OA unit is using “crclk” for its functionality. When trunk
+        * level clock gating takes place, OA clock would be gated,
+        * unable to count the events from non-render clock domain.
+        * Render clock gating must be disabled when OA is enabled to
+        * count the events from non-render domain. Unit level clock
+        * gating for RCS should also be disabled.
+        */
+       I915_WRITE(GEN7_MISCCPCTL, (I915_READ(GEN7_MISCCPCTL) &
+                                   ~GEN7_DOP_CLOCK_GATE_ENABLE));
+       I915_WRITE(GEN6_UCGCTL1, (I915_READ(GEN6_UCGCTL1) |
+                                 GEN6_CSUNIT_CLOCK_GATE_DISABLE));
+
+       config_oa_regs(dev_priv, dev_priv->perf.oa.mux_regs,
+                      dev_priv->perf.oa.mux_regs_len);
+
+       /* It apparently takes a fairly long time for a new MUX
+        * configuration to be be applied after these register writes.
+        * This delay duration was derived empirically based on the
+        * render_basic config but hopefully it covers the maximum
+        * configuration latency.
+        *
+        * As a fallback, the checks in _append_oa_reports() to skip
+        * invalid OA reports do also seem to work to discard reports
+        * generated before this config has completed - albeit not
+        * silently.
+        *
+        * Unfortunately this is essentially a magic number, since we
+        * don't currently know of a reliable mechanism for predicting
+        * how long the MUX config will take to apply and besides
+        * seeing invalid reports we don't know of a reliable way to
+        * explicitly check that the MUX config has landed.
+        *
+        * It's even possible we've miss characterized the underlying
+        * problem - it just seems like the simplest explanation why
+        * a delay at this location would mitigate any invalid reports.
+        */
+       usleep_range(15000, 20000);
+
+       config_oa_regs(dev_priv, dev_priv->perf.oa.b_counter_regs,
+                      dev_priv->perf.oa.b_counter_regs_len);
+
+       return 0;
+}
+
+static void hsw_disable_metric_set(struct drm_i915_private *dev_priv)
+{
+       I915_WRITE(GEN6_UCGCTL1, (I915_READ(GEN6_UCGCTL1) &
+                                 ~GEN6_CSUNIT_CLOCK_GATE_DISABLE));
+       I915_WRITE(GEN7_MISCCPCTL, (I915_READ(GEN7_MISCCPCTL) |
+                                   GEN7_DOP_CLOCK_GATE_ENABLE));
+
+       I915_WRITE(GDT_CHICKEN_BITS, (I915_READ(GDT_CHICKEN_BITS) &
+                                     ~GT_NOA_ENABLE));
+}
+
+static void gen7_update_oacontrol_locked(struct drm_i915_private *dev_priv)
+{
+       assert_spin_locked(&dev_priv->perf.hook_lock);
+
+       if (dev_priv->perf.oa.exclusive_stream->enabled) {
+               struct i915_gem_context *ctx =
+                       dev_priv->perf.oa.exclusive_stream->ctx;
+               u32 ctx_id = dev_priv->perf.oa.specific_ctx_id;
+
+               bool periodic = dev_priv->perf.oa.periodic;
+               u32 period_exponent = dev_priv->perf.oa.period_exponent;
+               u32 report_format = dev_priv->perf.oa.oa_buffer.format;
+
+               I915_WRITE(GEN7_OACONTROL,
+                          (ctx_id & GEN7_OACONTROL_CTX_MASK) |
+                          (period_exponent <<
+                           GEN7_OACONTROL_TIMER_PERIOD_SHIFT) |
+                          (periodic ? GEN7_OACONTROL_TIMER_ENABLE : 0) |
+                          (report_format << GEN7_OACONTROL_FORMAT_SHIFT) |
+                          (ctx ? GEN7_OACONTROL_PER_CTX_ENABLE : 0) |
+                          GEN7_OACONTROL_ENABLE);
+       } else
+               I915_WRITE(GEN7_OACONTROL, 0);
+}
+
+static void gen7_oa_enable(struct drm_i915_private *dev_priv)
+{
+       unsigned long flags;
+
+       /* Reset buf pointers so we don't forward reports from before now.
+        *
+        * Think carefully if considering trying to avoid this, since it
+        * also ensures status flags and the buffer itself are cleared
+        * in error paths, and we have checks for invalid reports based
+        * on the assumption that certain fields are written to zeroed
+        * memory which this helps maintains.
+        */
+       gen7_init_oa_buffer(dev_priv);
+
+       spin_lock_irqsave(&dev_priv->perf.hook_lock, flags);
+       gen7_update_oacontrol_locked(dev_priv);
+       spin_unlock_irqrestore(&dev_priv->perf.hook_lock, flags);
+}
+
+static void i915_oa_stream_enable(struct i915_perf_stream *stream)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+
+       dev_priv->perf.oa.ops.oa_enable(dev_priv);
+
+       if (dev_priv->perf.oa.periodic)
+               hrtimer_start(&dev_priv->perf.oa.poll_check_timer,
+                             ns_to_ktime(POLL_PERIOD),
+                             HRTIMER_MODE_REL_PINNED);
+}
+
+static void gen7_oa_disable(struct drm_i915_private *dev_priv)
+{
+       I915_WRITE(GEN7_OACONTROL, 0);
+}
+
+static void i915_oa_stream_disable(struct i915_perf_stream *stream)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+
+       dev_priv->perf.oa.ops.oa_disable(dev_priv);
+
+       if (dev_priv->perf.oa.periodic)
+               hrtimer_cancel(&dev_priv->perf.oa.poll_check_timer);
+}
+
+static u64 oa_exponent_to_ns(struct drm_i915_private *dev_priv, int exponent)
+{
+       return 1000000000ULL * (2ULL << exponent) /
+               dev_priv->perf.oa.timestamp_frequency;
+}
+
+static const struct i915_perf_stream_ops i915_oa_stream_ops = {
+       .destroy = i915_oa_stream_destroy,
+       .enable = i915_oa_stream_enable,
+       .disable = i915_oa_stream_disable,
+       .wait_unlocked = i915_oa_wait_unlocked,
+       .poll_wait = i915_oa_poll_wait,
+       .read = i915_oa_read,
 };
 
+static int i915_oa_stream_init(struct i915_perf_stream *stream,
+                              struct drm_i915_perf_open_param *param,
+                              struct perf_open_properties *props)
+{
+       struct drm_i915_private *dev_priv = stream->dev_priv;
+       int format_size;
+       int ret;
+
+       if (!(props->sample_flags & SAMPLE_OA_REPORT)) {
+               DRM_ERROR("Only OA report sampling supported\n");
+               return -EINVAL;
+       }
+
+       if (!dev_priv->perf.oa.ops.init_oa_buffer) {
+               DRM_ERROR("OA unit not supported\n");
+               return -ENODEV;
+       }
+
+       /* To avoid the complexity of having to accurately filter
+        * counter reports and marshal to the appropriate client
+        * we currently only allow exclusive access
+        */
+       if (dev_priv->perf.oa.exclusive_stream) {
+               DRM_ERROR("OA unit already in use\n");
+               return -EBUSY;
+       }
+
+       if (!props->metrics_set) {
+               DRM_ERROR("OA metric set not specified\n");
+               return -EINVAL;
+       }
+
+       if (!props->oa_format) {
+               DRM_ERROR("OA report format not specified\n");
+               return -EINVAL;
+       }
+
+       stream->sample_size = sizeof(struct drm_i915_perf_record_header);
+
+       format_size = dev_priv->perf.oa.oa_formats[props->oa_format].size;
+
+       stream->sample_flags |= SAMPLE_OA_REPORT;
+       stream->sample_size += format_size;
+
+       dev_priv->perf.oa.oa_buffer.format_size = format_size;
+       if (WARN_ON(dev_priv->perf.oa.oa_buffer.format_size == 0))
+               return -EINVAL;
+
+       dev_priv->perf.oa.oa_buffer.format =
+               dev_priv->perf.oa.oa_formats[props->oa_format].format;
+
+       dev_priv->perf.oa.metrics_set = props->metrics_set;
+
+       dev_priv->perf.oa.periodic = props->oa_periodic;
+       if (dev_priv->perf.oa.periodic) {
+               u64 period_ns = oa_exponent_to_ns(dev_priv,
+                                                 props->oa_period_exponent);
+
+               dev_priv->perf.oa.period_exponent = props->oa_period_exponent;
+
+               /* See comment for OA_TAIL_MARGIN_NSEC for details
+                * about this tail_margin...
+                */
+               dev_priv->perf.oa.tail_margin =
+                       ((OA_TAIL_MARGIN_NSEC / period_ns) + 1) * format_size;
+       }
+
+       if (stream->ctx) {
+               ret = oa_get_render_ctx_id(stream);
+               if (ret)
+                       return ret;
+       }
+
+       ret = alloc_oa_buffer(dev_priv);
+       if (ret)
+               goto err_oa_buf_alloc;
+
+       /* PRM - observability performance counters:
+        *
+        *   OACONTROL, performance counter enable, note:
+        *
+        *   "When this bit is set, in order to have coherent counts,
+        *   RC6 power state and trunk clock gating must be disabled.
+        *   This can be achieved by programming MMIO registers as
+        *   0xA094=0 and 0xA090[31]=1"
+        *
+        *   In our case we are expecting that taking pm + FORCEWAKE
+        *   references will effectively disable RC6.
+        */
+       intel_runtime_pm_get(dev_priv);
+       intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
+
+       ret = dev_priv->perf.oa.ops.enable_metric_set(dev_priv);
+       if (ret)
+               goto err_enable;
+
+       stream->ops = &i915_oa_stream_ops;
+
+       dev_priv->perf.oa.exclusive_stream = stream;
+
+       return 0;
+
+err_enable:
+       intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
+       intel_runtime_pm_put(dev_priv);
+       free_oa_buffer(dev_priv);
+
+err_oa_buf_alloc:
+       if (stream->ctx)
+               oa_put_render_ctx_id(stream);
+
+       return ret;
+}
+
 static ssize_t i915_perf_read_locked(struct i915_perf_stream *stream,
                                     struct file *file,
                                     char __user *buf,
@@ -78,8 +961,20 @@ static ssize_t i915_perf_read(struct file *file,
        struct drm_i915_private *dev_priv = stream->dev_priv;
        ssize_t ret;
 
+       /* To ensure it's handled consistently we simply treat all reads of a
+        * disabled stream as an error. In particular it might otherwise lead
+        * to a deadlock for blocking file descriptors...
+        */
+       if (!stream->enabled)
+               return -EIO;
+
        if (!(file->f_flags & O_NONBLOCK)) {
-               /* Allow false positives from stream->ops->wait_unlocked.
+               /* There's the small chance of false positives from
+                * stream->ops->wait_unlocked.
+                *
+                * E.g. with single context filtering since we only wait until
+                * oabuffer has >= 1 report we don't immediately know whether
+                * any reports really belong to the current context
                 */
                do {
                        ret = stream->ops->wait_unlocked(stream);
@@ -97,21 +992,51 @@ static ssize_t i915_perf_read(struct file *file,
                mutex_unlock(&dev_priv->perf.lock);
        }
 
+       if (ret >= 0) {
+               /* Maybe make ->pollin per-stream state if we support multiple
+                * concurrent streams in the future.
+                */
+               dev_priv->perf.oa.pollin = false;
+       }
+
        return ret;
 }
 
-static unsigned int i915_perf_poll_locked(struct i915_perf_stream *stream,
+static enum hrtimer_restart oa_poll_check_timer_cb(struct hrtimer *hrtimer)
+{
+       struct drm_i915_private *dev_priv =
+               container_of(hrtimer, typeof(*dev_priv),
+                            perf.oa.poll_check_timer);
+
+       if (!dev_priv->perf.oa.ops.oa_buffer_is_empty(dev_priv)) {
+               dev_priv->perf.oa.pollin = true;
+               wake_up(&dev_priv->perf.oa.poll_wq);
+       }
+
+       hrtimer_forward_now(hrtimer, ns_to_ktime(POLL_PERIOD));
+
+       return HRTIMER_RESTART;
+}
+
+static unsigned int i915_perf_poll_locked(struct drm_i915_private *dev_priv,
+                                         struct i915_perf_stream *stream,
                                          struct file *file,
                                          poll_table *wait)
 {
-       unsigned int streams = 0;
+       unsigned int events = 0;
 
        stream->ops->poll_wait(stream, file, wait);
 
-       if (stream->ops->can_read(stream))
-               streams |= POLLIN;
+       /* 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
+        * the hrtimer/oa_poll_check_timer_cb to notify us when there are
+        * samples to read.
+        */
+       if (dev_priv->perf.oa.pollin)
+               events |= POLLIN;
 
-       return streams;
+       return events;
 }
 
 static unsigned int i915_perf_poll(struct file *file, poll_table *wait)
@@ -121,7 +1046,7 @@ static unsigned int i915_perf_poll(struct file *file, poll_table *wait)
        int ret;
 
        mutex_lock(&dev_priv->perf.lock);
-       ret = i915_perf_poll_locked(stream, file, wait);
+       ret = i915_perf_poll_locked(dev_priv, stream, file, wait);
        mutex_unlock(&dev_priv->perf.lock);
 
        return ret;
@@ -285,18 +1210,21 @@ i915_perf_open_ioctl_locked(struct drm_i915_private *dev_priv,
                goto err_ctx;
        }
 
-       stream->sample_flags = props->sample_flags;
        stream->dev_priv = dev_priv;
        stream->ctx = specific_ctx;
 
-       /*
-        * TODO: support sampling something
-        *
-        * For now this is as far as we can go.
+       ret = i915_oa_stream_init(stream, param, props);
+       if (ret)
+               goto err_alloc;
+
+       /* we avoid simply assigning stream->sample_flags = props->sample_flags
+        * to have _stream_init check the combination of sample flags more
+        * thoroughly, but still this is the expected result at this point.
         */
-       DRM_ERROR("Unsupported i915 perf stream configuration\n");
-       ret = -EINVAL;
-       goto err_alloc;
+       if (WARN_ON(stream->sample_flags != props->sample_flags)) {
+               ret = -ENODEV;
+               goto err_alloc;
+       }
 
        list_add(&stream->link, &dev_priv->perf.streams);
 
@@ -382,6 +1310,56 @@ static int read_properties_unlocked(struct drm_i915_private *dev_priv,
                        props->single_context = 1;
                        props->ctx_handle = value;
                        break;
+               case DRM_I915_PERF_PROP_SAMPLE_OA:
+                       props->sample_flags |= SAMPLE_OA_REPORT;
+                       break;
+               case DRM_I915_PERF_PROP_OA_METRICS_SET:
+                       if (value == 0 ||
+                           value > dev_priv->perf.oa.n_builtin_sets) {
+                               DRM_ERROR("Unknown OA metric set ID");
+                               return -EINVAL;
+                       }
+                       props->metrics_set = value;
+                       break;
+               case DRM_I915_PERF_PROP_OA_FORMAT:
+                       if (value == 0 || value >= I915_OA_FORMAT_MAX) {
+                               DRM_ERROR("Invalid OA report format\n");
+                               return -EINVAL;
+                       }
+                       if (!dev_priv->perf.oa.oa_formats[value].size) {
+                               DRM_ERROR("Invalid OA report format\n");
+                               return -EINVAL;
+                       }
+                       props->oa_format = value;
+                       break;
+               case DRM_I915_PERF_PROP_OA_EXPONENT:
+                       if (value > OA_EXPONENT_MAX) {
+                               DRM_ERROR("OA timer exponent too high (> %u)\n",
+                                         OA_EXPONENT_MAX);
+                               return -EINVAL;
+                       }
+
+                       /* NB: The exponent represents a period as follows:
+                        *
+                        *   80ns * 2^(period_exponent + 1)
+                        *
+                        * Theoretically we can program the OA unit to sample
+                        * every 160ns but don't allow that by default unless
+                        * root.
+                        *
+                        * Referring to perf's
+                        * kernel.perf_event_max_sample_rate for a precedent
+                        * (100000 by default); with an OA exponent of 6 we get
+                        * a period of 10.240 microseconds -just under 100000Hz
+                        */
+                       if (value < 6 && !capable(CAP_SYS_ADMIN)) {
+                               DRM_ERROR("Minimum OA sampling exponent is 6 without root privileges\n");
+                               return -EACCES;
+                       }
+
+                       props->oa_periodic = true;
+                       props->oa_period_exponent = value;
+                       break;
                default:
                        MISSING_CASE(id);
                        DRM_ERROR("Unknown i915 perf property ID");
@@ -432,8 +1410,33 @@ int i915_perf_open_ioctl(struct drm_device *dev, void *data,
 
 void i915_perf_init(struct drm_i915_private *dev_priv)
 {
+       if (!IS_HASWELL(dev_priv))
+               return;
+
+       hrtimer_init(&dev_priv->perf.oa.poll_check_timer,
+                    CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+       dev_priv->perf.oa.poll_check_timer.function = oa_poll_check_timer_cb;
+       init_waitqueue_head(&dev_priv->perf.oa.poll_wq);
+
        INIT_LIST_HEAD(&dev_priv->perf.streams);
        mutex_init(&dev_priv->perf.lock);
+       spin_lock_init(&dev_priv->perf.hook_lock);
+
+       dev_priv->perf.oa.ops.init_oa_buffer = gen7_init_oa_buffer;
+       dev_priv->perf.oa.ops.enable_metric_set = hsw_enable_metric_set;
+       dev_priv->perf.oa.ops.disable_metric_set = hsw_disable_metric_set;
+       dev_priv->perf.oa.ops.oa_enable = gen7_oa_enable;
+       dev_priv->perf.oa.ops.oa_disable = gen7_oa_disable;
+       dev_priv->perf.oa.ops.read = gen7_oa_read;
+       dev_priv->perf.oa.ops.oa_buffer_is_empty =
+               gen7_oa_buffer_is_empty_fop_unlocked;
+
+       dev_priv->perf.oa.timestamp_frequency = 12500000;
+
+       dev_priv->perf.oa.oa_formats = hsw_oa_formats;
+
+       dev_priv->perf.oa.n_builtin_sets =
+               i915_oa_n_builtin_metric_sets_hsw;
 
        dev_priv->perf.initialized = true;
 }
@@ -443,7 +1446,6 @@ void i915_perf_fini(struct drm_i915_private *dev_priv)
        if (!dev_priv->perf.initialized)
                return;
 
-       /* Currently nothing to clean up */
-
+       memset(&dev_priv->perf.oa.ops, 0, sizeof(dev_priv->perf.oa.ops));
        dev_priv->perf.initialized = false;
 }
index ee87073..fcad8fa 100644 (file)
@@ -616,6 +616,343 @@ static inline bool i915_mmio_reg_valid(i915_reg_t reg)
 #define HSW_CS_GPR_UDW(n)               _MMIO(0x2600 + (n) * 8 + 4)
 
 #define GEN7_OACONTROL _MMIO(0x2360)
+#define  GEN7_OACONTROL_CTX_MASK           0xFFFFF000
+#define  GEN7_OACONTROL_TIMER_PERIOD_MASK   0x3F
+#define  GEN7_OACONTROL_TIMER_PERIOD_SHIFT  6
+#define  GEN7_OACONTROL_TIMER_ENABLE       (1<<5)
+#define  GEN7_OACONTROL_FORMAT_A13         (0<<2)
+#define  GEN7_OACONTROL_FORMAT_A29         (1<<2)
+#define  GEN7_OACONTROL_FORMAT_A13_B8_C8    (2<<2)
+#define  GEN7_OACONTROL_FORMAT_A29_B8_C8    (3<<2)
+#define  GEN7_OACONTROL_FORMAT_B4_C8       (4<<2)
+#define  GEN7_OACONTROL_FORMAT_A45_B8_C8    (5<<2)
+#define  GEN7_OACONTROL_FORMAT_B4_C8_A16    (6<<2)
+#define  GEN7_OACONTROL_FORMAT_C4_B8       (7<<2)
+#define  GEN7_OACONTROL_FORMAT_SHIFT       2
+#define  GEN7_OACONTROL_PER_CTX_ENABLE     (1<<1)
+#define  GEN7_OACONTROL_ENABLE             (1<<0)
+
+#define GEN8_OACTXID _MMIO(0x2364)
+
+#define GEN8_OACONTROL _MMIO(0x2B00)
+#define  GEN8_OA_REPORT_FORMAT_A12         (0<<2)
+#define  GEN8_OA_REPORT_FORMAT_A12_B8_C8    (2<<2)
+#define  GEN8_OA_REPORT_FORMAT_A36_B8_C8    (5<<2)
+#define  GEN8_OA_REPORT_FORMAT_C4_B8       (7<<2)
+#define  GEN8_OA_REPORT_FORMAT_SHIFT       2
+#define  GEN8_OA_SPECIFIC_CONTEXT_ENABLE    (1<<1)
+#define  GEN8_OA_COUNTER_ENABLE             (1<<0)
+
+#define GEN8_OACTXCONTROL _MMIO(0x2360)
+#define  GEN8_OA_TIMER_PERIOD_MASK         0x3F
+#define  GEN8_OA_TIMER_PERIOD_SHIFT        2
+#define  GEN8_OA_TIMER_ENABLE              (1<<1)
+#define  GEN8_OA_COUNTER_RESUME                    (1<<0)
+
+#define GEN7_OABUFFER _MMIO(0x23B0) /* R/W */
+#define  GEN7_OABUFFER_OVERRUN_DISABLE     (1<<3)
+#define  GEN7_OABUFFER_EDGE_TRIGGER        (1<<2)
+#define  GEN7_OABUFFER_STOP_RESUME_ENABLE   (1<<1)
+#define  GEN7_OABUFFER_RESUME              (1<<0)
+
+#define GEN8_OABUFFER _MMIO(0x2b14)
+
+#define GEN7_OASTATUS1 _MMIO(0x2364)
+#define  GEN7_OASTATUS1_TAIL_MASK          0xffffffc0
+#define  GEN7_OASTATUS1_COUNTER_OVERFLOW    (1<<2)
+#define  GEN7_OASTATUS1_OABUFFER_OVERFLOW   (1<<1)
+#define  GEN7_OASTATUS1_REPORT_LOST        (1<<0)
+
+#define GEN7_OASTATUS2 _MMIO(0x2368)
+#define GEN7_OASTATUS2_HEAD_MASK    0xffffffc0
+
+#define GEN8_OASTATUS _MMIO(0x2b08)
+#define  GEN8_OASTATUS_OVERRUN_STATUS      (1<<3)
+#define  GEN8_OASTATUS_COUNTER_OVERFLOW     (1<<2)
+#define  GEN8_OASTATUS_OABUFFER_OVERFLOW    (1<<1)
+#define  GEN8_OASTATUS_REPORT_LOST         (1<<0)
+
+#define GEN8_OAHEADPTR _MMIO(0x2B0C)
+#define GEN8_OATAILPTR _MMIO(0x2B10)
+
+#define OABUFFER_SIZE_128K  (0<<3)
+#define OABUFFER_SIZE_256K  (1<<3)
+#define OABUFFER_SIZE_512K  (2<<3)
+#define OABUFFER_SIZE_1M    (3<<3)
+#define OABUFFER_SIZE_2M    (4<<3)
+#define OABUFFER_SIZE_4M    (5<<3)
+#define OABUFFER_SIZE_8M    (6<<3)
+#define OABUFFER_SIZE_16M   (7<<3)
+
+#define OA_MEM_SELECT_GGTT  (1<<0)
+
+#define EU_PERF_CNTL0      _MMIO(0xe458)
+
+#define GDT_CHICKEN_BITS    _MMIO(0x9840)
+#define GT_NOA_ENABLE      0x00000080
+
+/*
+ * OA Boolean state
+ */
+
+#define OAREPORTTRIG1 _MMIO(0x2740)
+#define OAREPORTTRIG1_THRESHOLD_MASK 0xffff
+#define OAREPORTTRIG1_EDGE_LEVEL_TRIGER_SELECT_MASK 0xffff0000 /* 0=level */
+
+#define OAREPORTTRIG2 _MMIO(0x2744)
+#define OAREPORTTRIG2_INVERT_A_0  (1<<0)
+#define OAREPORTTRIG2_INVERT_A_1  (1<<1)
+#define OAREPORTTRIG2_INVERT_A_2  (1<<2)
+#define OAREPORTTRIG2_INVERT_A_3  (1<<3)
+#define OAREPORTTRIG2_INVERT_A_4  (1<<4)
+#define OAREPORTTRIG2_INVERT_A_5  (1<<5)
+#define OAREPORTTRIG2_INVERT_A_6  (1<<6)
+#define OAREPORTTRIG2_INVERT_A_7  (1<<7)
+#define OAREPORTTRIG2_INVERT_A_8  (1<<8)
+#define OAREPORTTRIG2_INVERT_A_9  (1<<9)
+#define OAREPORTTRIG2_INVERT_A_10 (1<<10)
+#define OAREPORTTRIG2_INVERT_A_11 (1<<11)
+#define OAREPORTTRIG2_INVERT_A_12 (1<<12)
+#define OAREPORTTRIG2_INVERT_A_13 (1<<13)
+#define OAREPORTTRIG2_INVERT_A_14 (1<<14)
+#define OAREPORTTRIG2_INVERT_A_15 (1<<15)
+#define OAREPORTTRIG2_INVERT_B_0  (1<<16)
+#define OAREPORTTRIG2_INVERT_B_1  (1<<17)
+#define OAREPORTTRIG2_INVERT_B_2  (1<<18)
+#define OAREPORTTRIG2_INVERT_B_3  (1<<19)
+#define OAREPORTTRIG2_INVERT_C_0  (1<<20)
+#define OAREPORTTRIG2_INVERT_C_1  (1<<21)
+#define OAREPORTTRIG2_INVERT_D_0  (1<<22)
+#define OAREPORTTRIG2_THRESHOLD_ENABLE     (1<<23)
+#define OAREPORTTRIG2_REPORT_TRIGGER_ENABLE (1<<31)
+
+#define OAREPORTTRIG3 _MMIO(0x2748)
+#define OAREPORTTRIG3_NOA_SELECT_MASK      0xf
+#define OAREPORTTRIG3_NOA_SELECT_8_SHIFT    0
+#define OAREPORTTRIG3_NOA_SELECT_9_SHIFT    4
+#define OAREPORTTRIG3_NOA_SELECT_10_SHIFT   8
+#define OAREPORTTRIG3_NOA_SELECT_11_SHIFT   12
+#define OAREPORTTRIG3_NOA_SELECT_12_SHIFT   16
+#define OAREPORTTRIG3_NOA_SELECT_13_SHIFT   20
+#define OAREPORTTRIG3_NOA_SELECT_14_SHIFT   24
+#define OAREPORTTRIG3_NOA_SELECT_15_SHIFT   28
+
+#define OAREPORTTRIG4 _MMIO(0x274c)
+#define OAREPORTTRIG4_NOA_SELECT_MASK      0xf
+#define OAREPORTTRIG4_NOA_SELECT_0_SHIFT    0
+#define OAREPORTTRIG4_NOA_SELECT_1_SHIFT    4
+#define OAREPORTTRIG4_NOA_SELECT_2_SHIFT    8
+#define OAREPORTTRIG4_NOA_SELECT_3_SHIFT    12
+#define OAREPORTTRIG4_NOA_SELECT_4_SHIFT    16
+#define OAREPORTTRIG4_NOA_SELECT_5_SHIFT    20
+#define OAREPORTTRIG4_NOA_SELECT_6_SHIFT    24
+#define OAREPORTTRIG4_NOA_SELECT_7_SHIFT    28
+
+#define OAREPORTTRIG5 _MMIO(0x2750)
+#define OAREPORTTRIG5_THRESHOLD_MASK 0xffff
+#define OAREPORTTRIG5_EDGE_LEVEL_TRIGER_SELECT_MASK 0xffff0000 /* 0=level */
+
+#define OAREPORTTRIG6 _MMIO(0x2754)
+#define OAREPORTTRIG6_INVERT_A_0  (1<<0)
+#define OAREPORTTRIG6_INVERT_A_1  (1<<1)
+#define OAREPORTTRIG6_INVERT_A_2  (1<<2)
+#define OAREPORTTRIG6_INVERT_A_3  (1<<3)
+#define OAREPORTTRIG6_INVERT_A_4  (1<<4)
+#define OAREPORTTRIG6_INVERT_A_5  (1<<5)
+#define OAREPORTTRIG6_INVERT_A_6  (1<<6)
+#define OAREPORTTRIG6_INVERT_A_7  (1<<7)
+#define OAREPORTTRIG6_INVERT_A_8  (1<<8)
+#define OAREPORTTRIG6_INVERT_A_9  (1<<9)
+#define OAREPORTTRIG6_INVERT_A_10 (1<<10)
+#define OAREPORTTRIG6_INVERT_A_11 (1<<11)
+#define OAREPORTTRIG6_INVERT_A_12 (1<<12)
+#define OAREPORTTRIG6_INVERT_A_13 (1<<13)
+#define OAREPORTTRIG6_INVERT_A_14 (1<<14)
+#define OAREPORTTRIG6_INVERT_A_15 (1<<15)
+#define OAREPORTTRIG6_INVERT_B_0  (1<<16)
+#define OAREPORTTRIG6_INVERT_B_1  (1<<17)
+#define OAREPORTTRIG6_INVERT_B_2  (1<<18)
+#define OAREPORTTRIG6_INVERT_B_3  (1<<19)
+#define OAREPORTTRIG6_INVERT_C_0  (1<<20)
+#define OAREPORTTRIG6_INVERT_C_1  (1<<21)
+#define OAREPORTTRIG6_INVERT_D_0  (1<<22)
+#define OAREPORTTRIG6_THRESHOLD_ENABLE     (1<<23)
+#define OAREPORTTRIG6_REPORT_TRIGGER_ENABLE (1<<31)
+
+#define OAREPORTTRIG7 _MMIO(0x2758)
+#define OAREPORTTRIG7_NOA_SELECT_MASK      0xf
+#define OAREPORTTRIG7_NOA_SELECT_8_SHIFT    0
+#define OAREPORTTRIG7_NOA_SELECT_9_SHIFT    4
+#define OAREPORTTRIG7_NOA_SELECT_10_SHIFT   8
+#define OAREPORTTRIG7_NOA_SELECT_11_SHIFT   12
+#define OAREPORTTRIG7_NOA_SELECT_12_SHIFT   16
+#define OAREPORTTRIG7_NOA_SELECT_13_SHIFT   20
+#define OAREPORTTRIG7_NOA_SELECT_14_SHIFT   24
+#define OAREPORTTRIG7_NOA_SELECT_15_SHIFT   28
+
+#define OAREPORTTRIG8 _MMIO(0x275c)
+#define OAREPORTTRIG8_NOA_SELECT_MASK      0xf
+#define OAREPORTTRIG8_NOA_SELECT_0_SHIFT    0
+#define OAREPORTTRIG8_NOA_SELECT_1_SHIFT    4
+#define OAREPORTTRIG8_NOA_SELECT_2_SHIFT    8
+#define OAREPORTTRIG8_NOA_SELECT_3_SHIFT    12
+#define OAREPORTTRIG8_NOA_SELECT_4_SHIFT    16
+#define OAREPORTTRIG8_NOA_SELECT_5_SHIFT    20
+#define OAREPORTTRIG8_NOA_SELECT_6_SHIFT    24
+#define OAREPORTTRIG8_NOA_SELECT_7_SHIFT    28
+
+#define OASTARTTRIG1 _MMIO(0x2710)
+#define OASTARTTRIG1_THRESHOLD_COUNT_MASK_MBZ 0xffff0000
+#define OASTARTTRIG1_THRESHOLD_MASK          0xffff
+
+#define OASTARTTRIG2 _MMIO(0x2714)
+#define OASTARTTRIG2_INVERT_A_0 (1<<0)
+#define OASTARTTRIG2_INVERT_A_1 (1<<1)
+#define OASTARTTRIG2_INVERT_A_2 (1<<2)
+#define OASTARTTRIG2_INVERT_A_3 (1<<3)
+#define OASTARTTRIG2_INVERT_A_4 (1<<4)
+#define OASTARTTRIG2_INVERT_A_5 (1<<5)
+#define OASTARTTRIG2_INVERT_A_6 (1<<6)
+#define OASTARTTRIG2_INVERT_A_7 (1<<7)
+#define OASTARTTRIG2_INVERT_A_8 (1<<8)
+#define OASTARTTRIG2_INVERT_A_9 (1<<9)
+#define OASTARTTRIG2_INVERT_A_10 (1<<10)
+#define OASTARTTRIG2_INVERT_A_11 (1<<11)
+#define OASTARTTRIG2_INVERT_A_12 (1<<12)
+#define OASTARTTRIG2_INVERT_A_13 (1<<13)
+#define OASTARTTRIG2_INVERT_A_14 (1<<14)
+#define OASTARTTRIG2_INVERT_A_15 (1<<15)
+#define OASTARTTRIG2_INVERT_B_0 (1<<16)
+#define OASTARTTRIG2_INVERT_B_1 (1<<17)
+#define OASTARTTRIG2_INVERT_B_2 (1<<18)
+#define OASTARTTRIG2_INVERT_B_3 (1<<19)
+#define OASTARTTRIG2_INVERT_C_0 (1<<20)
+#define OASTARTTRIG2_INVERT_C_1 (1<<21)
+#define OASTARTTRIG2_INVERT_D_0 (1<<22)
+#define OASTARTTRIG2_THRESHOLD_ENABLE      (1<<23)
+#define OASTARTTRIG2_START_TRIG_FLAG_MBZ    (1<<24)
+#define OASTARTTRIG2_EVENT_SELECT_0  (1<<28)
+#define OASTARTTRIG2_EVENT_SELECT_1  (1<<29)
+#define OASTARTTRIG2_EVENT_SELECT_2  (1<<30)
+#define OASTARTTRIG2_EVENT_SELECT_3  (1<<31)
+
+#define OASTARTTRIG3 _MMIO(0x2718)
+#define OASTARTTRIG3_NOA_SELECT_MASK      0xf
+#define OASTARTTRIG3_NOA_SELECT_8_SHIFT    0
+#define OASTARTTRIG3_NOA_SELECT_9_SHIFT    4
+#define OASTARTTRIG3_NOA_SELECT_10_SHIFT   8
+#define OASTARTTRIG3_NOA_SELECT_11_SHIFT   12
+#define OASTARTTRIG3_NOA_SELECT_12_SHIFT   16
+#define OASTARTTRIG3_NOA_SELECT_13_SHIFT   20
+#define OASTARTTRIG3_NOA_SELECT_14_SHIFT   24
+#define OASTARTTRIG3_NOA_SELECT_15_SHIFT   28
+
+#define OASTARTTRIG4 _MMIO(0x271c)
+#define OASTARTTRIG4_NOA_SELECT_MASK       0xf
+#define OASTARTTRIG4_NOA_SELECT_0_SHIFT    0
+#define OASTARTTRIG4_NOA_SELECT_1_SHIFT    4
+#define OASTARTTRIG4_NOA_SELECT_2_SHIFT    8
+#define OASTARTTRIG4_NOA_SELECT_3_SHIFT    12
+#define OASTARTTRIG4_NOA_SELECT_4_SHIFT    16
+#define OASTARTTRIG4_NOA_SELECT_5_SHIFT    20
+#define OASTARTTRIG4_NOA_SELECT_6_SHIFT    24
+#define OASTARTTRIG4_NOA_SELECT_7_SHIFT    28
+
+#define OASTARTTRIG5 _MMIO(0x2720)
+#define OASTARTTRIG5_THRESHOLD_COUNT_MASK_MBZ 0xffff0000
+#define OASTARTTRIG5_THRESHOLD_MASK          0xffff
+
+#define OASTARTTRIG6 _MMIO(0x2724)
+#define OASTARTTRIG6_INVERT_A_0 (1<<0)
+#define OASTARTTRIG6_INVERT_A_1 (1<<1)
+#define OASTARTTRIG6_INVERT_A_2 (1<<2)
+#define OASTARTTRIG6_INVERT_A_3 (1<<3)
+#define OASTARTTRIG6_INVERT_A_4 (1<<4)
+#define OASTARTTRIG6_INVERT_A_5 (1<<5)
+#define OASTARTTRIG6_INVERT_A_6 (1<<6)
+#define OASTARTTRIG6_INVERT_A_7 (1<<7)
+#define OASTARTTRIG6_INVERT_A_8 (1<<8)
+#define OASTARTTRIG6_INVERT_A_9 (1<<9)
+#define OASTARTTRIG6_INVERT_A_10 (1<<10)
+#define OASTARTTRIG6_INVERT_A_11 (1<<11)
+#define OASTARTTRIG6_INVERT_A_12 (1<<12)
+#define OASTARTTRIG6_INVERT_A_13 (1<<13)
+#define OASTARTTRIG6_INVERT_A_14 (1<<14)
+#define OASTARTTRIG6_INVERT_A_15 (1<<15)
+#define OASTARTTRIG6_INVERT_B_0 (1<<16)
+#define OASTARTTRIG6_INVERT_B_1 (1<<17)
+#define OASTARTTRIG6_INVERT_B_2 (1<<18)
+#define OASTARTTRIG6_INVERT_B_3 (1<<19)
+#define OASTARTTRIG6_INVERT_C_0 (1<<20)
+#define OASTARTTRIG6_INVERT_C_1 (1<<21)
+#define OASTARTTRIG6_INVERT_D_0 (1<<22)
+#define OASTARTTRIG6_THRESHOLD_ENABLE      (1<<23)
+#define OASTARTTRIG6_START_TRIG_FLAG_MBZ    (1<<24)
+#define OASTARTTRIG6_EVENT_SELECT_4  (1<<28)
+#define OASTARTTRIG6_EVENT_SELECT_5  (1<<29)
+#define OASTARTTRIG6_EVENT_SELECT_6  (1<<30)
+#define OASTARTTRIG6_EVENT_SELECT_7  (1<<31)
+
+#define OASTARTTRIG7 _MMIO(0x2728)
+#define OASTARTTRIG7_NOA_SELECT_MASK      0xf
+#define OASTARTTRIG7_NOA_SELECT_8_SHIFT    0
+#define OASTARTTRIG7_NOA_SELECT_9_SHIFT    4
+#define OASTARTTRIG7_NOA_SELECT_10_SHIFT   8
+#define OASTARTTRIG7_NOA_SELECT_11_SHIFT   12
+#define OASTARTTRIG7_NOA_SELECT_12_SHIFT   16
+#define OASTARTTRIG7_NOA_SELECT_13_SHIFT   20
+#define OASTARTTRIG7_NOA_SELECT_14_SHIFT   24
+#define OASTARTTRIG7_NOA_SELECT_15_SHIFT   28
+
+#define OASTARTTRIG8 _MMIO(0x272c)
+#define OASTARTTRIG8_NOA_SELECT_MASK      0xf
+#define OASTARTTRIG8_NOA_SELECT_0_SHIFT    0
+#define OASTARTTRIG8_NOA_SELECT_1_SHIFT    4
+#define OASTARTTRIG8_NOA_SELECT_2_SHIFT    8
+#define OASTARTTRIG8_NOA_SELECT_3_SHIFT    12
+#define OASTARTTRIG8_NOA_SELECT_4_SHIFT    16
+#define OASTARTTRIG8_NOA_SELECT_5_SHIFT    20
+#define OASTARTTRIG8_NOA_SELECT_6_SHIFT    24
+#define OASTARTTRIG8_NOA_SELECT_7_SHIFT    28
+
+/* CECX_0 */
+#define OACEC_COMPARE_LESS_OR_EQUAL    6
+#define OACEC_COMPARE_NOT_EQUAL                5
+#define OACEC_COMPARE_LESS_THAN                4
+#define OACEC_COMPARE_GREATER_OR_EQUAL 3
+#define OACEC_COMPARE_EQUAL            2
+#define OACEC_COMPARE_GREATER_THAN     1
+#define OACEC_COMPARE_ANY_EQUAL                0
+
+#define OACEC_COMPARE_VALUE_MASK    0xffff
+#define OACEC_COMPARE_VALUE_SHIFT   3
+
+#define OACEC_SELECT_NOA       (0<<19)
+#define OACEC_SELECT_PREV      (1<<19)
+#define OACEC_SELECT_BOOLEAN   (2<<19)
+
+/* CECX_1 */
+#define OACEC_MASK_MASK                    0xffff
+#define OACEC_CONSIDERATIONS_MASK   0xffff
+#define OACEC_CONSIDERATIONS_SHIFT  16
+
+#define OACEC0_0 _MMIO(0x2770)
+#define OACEC0_1 _MMIO(0x2774)
+#define OACEC1_0 _MMIO(0x2778)
+#define OACEC1_1 _MMIO(0x277c)
+#define OACEC2_0 _MMIO(0x2780)
+#define OACEC2_1 _MMIO(0x2784)
+#define OACEC3_0 _MMIO(0x2788)
+#define OACEC3_1 _MMIO(0x278c)
+#define OACEC4_0 _MMIO(0x2790)
+#define OACEC4_1 _MMIO(0x2794)
+#define OACEC5_0 _MMIO(0x2798)
+#define OACEC5_1 _MMIO(0x279c)
+#define OACEC6_0 _MMIO(0x27a0)
+#define OACEC6_1 _MMIO(0x27a4)
+#define OACEC7_0 _MMIO(0x27a8)
+#define OACEC7_1 _MMIO(0x27ac)
+
 
 #define _GEN7_PIPEA_DE_LOAD_SL 0x70068
 #define _GEN7_PIPEB_DE_LOAD_SL 0x71068
@@ -6914,6 +7251,7 @@ enum {
 # define GEN6_RCCUNIT_CLOCK_GATE_DISABLE               (1 << 11)
 
 #define GEN6_UCGCTL3                           _MMIO(0x9408)
+# define GEN6_OACSUNIT_CLOCK_GATE_DISABLE              (1 << 20)
 
 #define GEN7_UCGCTL4                           _MMIO(0x940c)
 #define  GEN7_L3BANK2X_CLOCK_GATE_DISABLE      (1<<25)
index 192dc1f..bdfc688 100644 (file)
@@ -1230,6 +1230,18 @@ struct drm_i915_gem_context_param {
        __u64 value;
 };
 
+enum drm_i915_oa_format {
+       I915_OA_FORMAT_A13 = 1,
+       I915_OA_FORMAT_A29,
+       I915_OA_FORMAT_A13_B8_C8,
+       I915_OA_FORMAT_B4_C8,
+       I915_OA_FORMAT_A45_B8_C8,
+       I915_OA_FORMAT_B4_C8_A16,
+       I915_OA_FORMAT_C4_B8,
+
+       I915_OA_FORMAT_MAX          /* non-ABI */
+};
+
 enum drm_i915_perf_property_id {
        /**
         * Open the stream for a specific context handle (as used with
@@ -1238,6 +1250,32 @@ enum drm_i915_perf_property_id {
         */
        DRM_I915_PERF_PROP_CTX_HANDLE = 1,
 
+       /**
+        * A value of 1 requests the inclusion of raw OA unit reports as
+        * part of stream samples.
+        */
+       DRM_I915_PERF_PROP_SAMPLE_OA,
+
+       /**
+        * The value specifies which set of OA unit metrics should be
+        * be configured, defining the contents of any OA unit reports.
+        */
+       DRM_I915_PERF_PROP_OA_METRICS_SET,
+
+       /**
+        * The value specifies the size and layout of OA unit reports.
+        */
+       DRM_I915_PERF_PROP_OA_FORMAT,
+
+       /**
+        * Specifying this property implicitly requests periodic OA unit
+        * sampling and (at least on Haswell) the sampling frequency is derived
+        * from this exponent as follows:
+        *
+        *   80ns * 2^(period_exponent + 1)
+        */
+       DRM_I915_PERF_PROP_OA_EXPONENT,
+
        DRM_I915_PERF_PROP_MAX /* non-ABI */
 };
 
@@ -1257,7 +1295,23 @@ struct drm_i915_perf_open_param {
        __u64 __user properties_ptr;
 };
 
+/**
+ * Enable data capture for a stream that was either opened in a disabled state
+ * via I915_PERF_FLAG_DISABLED or was later disabled via
+ * I915_PERF_IOCTL_DISABLE.
+ *
+ * It is intended to be cheaper to disable and enable a stream than it may be
+ * to close and re-open a stream with the same configuration.
+ *
+ * It's undefined whether any pending data for the stream will be lost.
+ */
 #define I915_PERF_IOCTL_ENABLE _IO('i', 0x0)
+
+/**
+ * Disable data capture for a stream.
+ *
+ * It is an error to try and read a stream that is disabled.
+ */
 #define I915_PERF_IOCTL_DISABLE        _IO('i', 0x1)
 
 /**
@@ -1281,17 +1335,30 @@ enum drm_i915_perf_record_type {
         * every sample.
         *
         * The order of these sample properties given by userspace has no
-        * affect on the ordering of data within a sample. The order will be
+        * affect on the ordering of data within a sample. The order is
         * documented here.
         *
         * struct {
         *     struct drm_i915_perf_record_header header;
         *
-        *     TODO: itemize extensible sample data here
+        *     { u32 oa_report[]; } && DRM_I915_PERF_PROP_SAMPLE_OA
         * };
         */
        DRM_I915_PERF_RECORD_SAMPLE = 1,
 
+       /*
+        * Indicates that one or more OA reports were not written by the
+        * hardware. This can happen for example if an MI_REPORT_PERF_COUNT
+        * command collides with periodic sampling - which would be more likely
+        * at higher sampling frequencies.
+        */
+       DRM_I915_PERF_RECORD_OA_REPORT_LOST = 2,
+
+       /**
+        * An error occurred that resulted in all pending OA reports being lost.
+        */
+       DRM_I915_PERF_RECORD_OA_BUFFER_LOST = 3,
+
        DRM_I915_PERF_RECORD_MAX /* non-ABI */
 };