diff options
Diffstat (limited to 'src/intel/perf')
| -rw-r--r-- | src/intel/perf/gen_perf.c | 120 | ||||
| -rw-r--r-- | src/intel/perf/gen_perf.h | 119 |
2 files changed, 120 insertions, 119 deletions
diff --git a/src/intel/perf/gen_perf.c b/src/intel/perf/gen_perf.c index 78e4b4c22b6..9b17b350363 100644 --- a/src/intel/perf/gen_perf.c +++ b/src/intel/perf/gen_perf.c @@ -69,6 +69,126 @@ #define MAP_READ (1 << 0) #define MAP_WRITE (1 << 1) +/** + * Periodic OA samples are read() into these buffer structures via the + * i915 perf kernel interface and appended to the + * perf_ctx->sample_buffers linked list. When we process the + * results of an OA metrics query we need to consider all the periodic + * samples between the Begin and End MI_REPORT_PERF_COUNT command + * markers. + * + * 'Periodic' is a simplification as there are other automatic reports + * written by the hardware also buffered here. + * + * Considering three queries, A, B and C: + * + * Time ----> + * ________________A_________________ + * | | + * | ________B_________ _____C___________ + * | | | | | | + * + * And an illustration of sample buffers read over this time frame: + * [HEAD ][ ][ ][ ][ ][ ][ ][ ][TAIL ] + * + * These nodes may hold samples for query A: + * [ ][ ][ A ][ A ][ A ][ A ][ A ][ ][ ] + * + * These nodes may hold samples for query B: + * [ ][ ][ B ][ B ][ B ][ ][ ][ ][ ] + * + * These nodes may hold samples for query C: + * [ ][ ][ ][ ][ ][ C ][ C ][ C ][ ] + * + * The illustration assumes we have an even distribution of periodic + * samples so all nodes have the same size plotted against time: + * + * Note, to simplify code, the list is never empty. + * + * With overlapping queries we can see that periodic OA reports may + * relate to multiple queries and care needs to be take to keep + * track of sample buffers until there are no queries that might + * depend on their contents. + * + * We use a node ref counting system where a reference ensures that a + * node and all following nodes can't be freed/recycled until the + * reference drops to zero. + * + * E.g. with a ref of one here: + * [ 0 ][ 0 ][ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ] + * + * These nodes could be freed or recycled ("reaped"): + * [ 0 ][ 0 ] + * + * These must be preserved until the leading ref drops to zero: + * [ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ] + * + * When a query starts we take a reference on the current tail of + * the list, knowing that no already-buffered samples can possibly + * relate to the newly-started query. A pointer to this node is + * also saved in the query object's ->oa.samples_head. + * + * E.g. starting query A while there are two nodes in .sample_buffers: + * ________________A________ + * | + * + * [ 0 ][ 1 ] + * ^_______ Add a reference and store pointer to node in + * A->oa.samples_head + * + * Moving forward to when the B query starts with no new buffer nodes: + * (for reference, i915 perf reads() are only done when queries finish) + * ________________A_______ + * | ________B___ + * | | + * + * [ 0 ][ 2 ] + * ^_______ Add a reference and store pointer to + * node in B->oa.samples_head + * + * Once a query is finished, after an OA query has become 'Ready', + * once the End OA report has landed and after we we have processed + * all the intermediate periodic samples then we drop the + * ->oa.samples_head reference we took at the start. + * + * So when the B query has finished we have: + * ________________A________ + * | ______B___________ + * | | | + * [ 0 ][ 1 ][ 0 ][ 0 ][ 0 ] + * ^_______ Drop B->oa.samples_head reference + * + * We still can't free these due to the A->oa.samples_head ref: + * [ 1 ][ 0 ][ 0 ][ 0 ] + * + * When the A query finishes: (note there's a new ref for C's samples_head) + * ________________A_________________ + * | | + * | _____C_________ + * | | | + * [ 0 ][ 0 ][ 0 ][ 0 ][ 1 ][ 0 ][ 0 ] + * ^_______ Drop A->oa.samples_head reference + * + * And we can now reap these nodes up to the C->oa.samples_head: + * [ X ][ X ][ X ][ X ] + * keeping -> [ 1 ][ 0 ][ 0 ] + * + * We reap old sample buffers each time we finish processing an OA + * query by iterating the sample_buffers list from the head until we + * find a referenced node and stop. + * + * Reaped buffers move to a perfquery.free_sample_buffers list and + * when we come to read() we first look to recycle a buffer from the + * free_sample_buffers list before allocating a new buffer. + */ +struct oa_sample_buf { + struct exec_node link; + int refcount; + int len; + uint8_t buf[I915_PERF_OA_SAMPLE_SIZE * 10]; + uint32_t last_timestamp; +}; + struct gen_perf_query_object * gen_perf_new_query(struct gen_perf_context *perf_ctx, unsigned query_index) { diff --git a/src/intel/perf/gen_perf.h b/src/intel/perf/gen_perf.h index 1dab55a85e6..aed10d0d61b 100644 --- a/src/intel/perf/gen_perf.h +++ b/src/intel/perf/gen_perf.h @@ -248,125 +248,6 @@ struct gen_perf_config { } vtbl; }; -/** - * Periodic OA samples are read() into these buffer structures via the - * i915 perf kernel interface and appended to the - * brw->perfquery.sample_buffers linked list. When we process the - * results of an OA metrics query we need to consider all the periodic - * samples between the Begin and End MI_REPORT_PERF_COUNT command - * markers. - * - * 'Periodic' is a simplification as there are other automatic reports - * written by the hardware also buffered here. - * - * Considering three queries, A, B and C: - * - * Time ----> - * ________________A_________________ - * | | - * | ________B_________ _____C___________ - * | | | | | | - * - * And an illustration of sample buffers read over this time frame: - * [HEAD ][ ][ ][ ][ ][ ][ ][ ][TAIL ] - * - * These nodes may hold samples for query A: - * [ ][ ][ A ][ A ][ A ][ A ][ A ][ ][ ] - * - * These nodes may hold samples for query B: - * [ ][ ][ B ][ B ][ B ][ ][ ][ ][ ] - * - * These nodes may hold samples for query C: - * [ ][ ][ ][ ][ ][ C ][ C ][ C ][ ] - * - * The illustration assumes we have an even distribution of periodic - * samples so all nodes have the same size plotted against time: - * - * Note, to simplify code, the list is never empty. - * - * With overlapping queries we can see that periodic OA reports may - * relate to multiple queries and care needs to be take to keep - * track of sample buffers until there are no queries that might - * depend on their contents. - * - * We use a node ref counting system where a reference ensures that a - * node and all following nodes can't be freed/recycled until the - * reference drops to zero. - * - * E.g. with a ref of one here: - * [ 0 ][ 0 ][ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ] - * - * These nodes could be freed or recycled ("reaped"): - * [ 0 ][ 0 ] - * - * These must be preserved until the leading ref drops to zero: - * [ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ] - * - * When a query starts we take a reference on the current tail of - * the list, knowing that no already-buffered samples can possibly - * relate to the newly-started query. A pointer to this node is - * also saved in the query object's ->oa.samples_head. - * - * E.g. starting query A while there are two nodes in .sample_buffers: - * ________________A________ - * | - * - * [ 0 ][ 1 ] - * ^_______ Add a reference and store pointer to node in - * A->oa.samples_head - * - * Moving forward to when the B query starts with no new buffer nodes: - * (for reference, i915 perf reads() are only done when queries finish) - * ________________A_______ - * | ________B___ - * | | - * - * [ 0 ][ 2 ] - * ^_______ Add a reference and store pointer to - * node in B->oa.samples_head - * - * Once a query is finished, after an OA query has become 'Ready', - * once the End OA report has landed and after we we have processed - * all the intermediate periodic samples then we drop the - * ->oa.samples_head reference we took at the start. - * - * So when the B query has finished we have: - * ________________A________ - * | ______B___________ - * | | | - * [ 0 ][ 1 ][ 0 ][ 0 ][ 0 ] - * ^_______ Drop B->oa.samples_head reference - * - * We still can't free these due to the A->oa.samples_head ref: - * [ 1 ][ 0 ][ 0 ][ 0 ] - * - * When the A query finishes: (note there's a new ref for C's samples_head) - * ________________A_________________ - * | | - * | _____C_________ - * | | | - * [ 0 ][ 0 ][ 0 ][ 0 ][ 1 ][ 0 ][ 0 ] - * ^_______ Drop A->oa.samples_head reference - * - * And we can now reap these nodes up to the C->oa.samples_head: - * [ X ][ X ][ X ][ X ] - * keeping -> [ 1 ][ 0 ][ 0 ] - * - * We reap old sample buffers each time we finish processing an OA - * query by iterating the sample_buffers list from the head until we - * find a referenced node and stop. - * - * Reaped buffers move to a perfquery.free_sample_buffers list and - * when we come to read() we first look to recycle a buffer from the - * free_sample_buffers list before allocating a new buffer. - */ -struct oa_sample_buf { - struct exec_node link; - int refcount; - int len; - uint8_t buf[I915_PERF_OA_SAMPLE_SIZE * 10]; - uint32_t last_timestamp; -}; /** * gen representation of a performance query object. |