On legacy request_queues, request allocation through request_list
associates each request with its blkcg_gq. However, blk-mq doesn't
use request_list and its requests aren't associated with the matching
blkcg_gqs, preventing cgroup-aware tracking blk-mq IOs.
This patch adds blk_rq_[dis]associate_blkg() and use them in blk-mq
request alloc/free paths so that every request is associated with its
blkcg_gq while in-flight. The added overhead is minimal and per-cpu
in hot paths. In the future, we also should be able to remove the
more frequent per-bio blkcg_gq operations in blkcg_bio_issue_check()
for blk-mq.
* ->blkcg is added to blk_mq_alloc_data to carry the target blkcg in
the allocation path. I didn't #ifdef it for code simplicity. The
only overhead when cgroup is disabled is the extra pointer in the
data structure, which shouldn't matter.
* Moved set_start_time_ns() out of CONFIG_BLK_CGROUP together with
blk_rq_associate_blkg(). Both functions provide dummy
implementations when !CONFIG_BLK_CGROUP.
Signed-off-by: Tejun Heo <tj@xxxxxxxxxx>
---
block/blk-mq.c | 9 ++++++---
block/blk-mq.h | 1 +
include/linux/blk-cgroup.h | 42 ++++++++++++++++++++++++++++++++++++++++++
3 files changed, 49 insertions(+), 3 deletions(-)
diff --git a/block/blk-mq.c b/block/blk-mq.c
index af958c4..7cc64de 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -305,9 +305,10 @@ static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
rq->rq_disk = NULL;
rq->part = NULL;
rq->start_time = jiffies;
-#ifdef CONFIG_BLK_CGROUP
- rq->blkg = NULL;
+
+ blk_rq_associate_blkg(rq, data->blkcg);
set_start_time_ns(rq);
+#ifdef CONFIG_BLK_CGROUP
rq->io_start_time_ns = 0;
#endif
rq->nr_phys_segments = 0;
@@ -472,6 +473,8 @@ void blk_mq_free_request(struct request *rq)
}
}
+ blk_rq_disassociate_blkg(rq);
+
ctx->rq_completed[rq_is_sync(rq)]++;
if (rq->rq_flags & RQF_MQ_INFLIGHT)
atomic_dec(&hctx->nr_active);
@@ -1599,7 +1602,7 @@ static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
{
const int is_sync = op_is_sync(bio->bi_opf);
const int is_flush_fua = op_is_flush(bio->bi_opf);
- struct blk_mq_alloc_data data = { .flags = 0 };
+ struct blk_mq_alloc_data data = { .flags = 0, .blkcg = bio_blkcg(bio) };
struct request *rq;
unsigned int request_count = 0;
struct blk_plug *plug;
diff --git a/block/blk-mq.h b/block/blk-mq.h
index 4933af9..dfb1e1d 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -111,6 +111,7 @@ struct blk_mq_alloc_data {
struct request_queue *q;
unsigned int flags;
unsigned int shallow_depth;
+ struct blkcg *blkcg;
/* input & output parameter */
struct blk_mq_ctx *ctx;
diff --git a/include/linux/blk-cgroup.h b/include/linux/blk-cgroup.h
index 1de5158..96eed0f 100644
--- a/include/linux/blk-cgroup.h
+++ b/include/linux/blk-cgroup.h
@@ -725,6 +725,45 @@ static inline bool blkcg_bio_issue_check(struct request_queue *q,
return !throtl;
}
+/**
+ * blk_rq_associate_blkg - lookup and associate a request with its blkcg_gq
+ * @rq: request of interest
+ * @blkcg: target blkcg
+ *
+ * Associate @rq with the matching blkcg_cq of @blkcg. This is used by
+ * blk-mq to associate requests directly with blkgs without going through
+ * request_list.
+ */
+static inline void blk_rq_associate_blkg(struct request *rq, struct blkcg *blkcg)
+{
+ struct request_queue *q = rq->q;
+ struct blkcg_gq *blkg;
+
+ if (!blkcg) {
+ rq->blkg = q->root_blkg;
+ return;
+ }
+
+ rcu_read_lock();
+ blkg = blkg_lookup_create(blkcg, q);
+ if (likely(blkg == q->root_blkg || percpu_ref_tryget(&blkg->refcnt)))
+ rq->blkg = blkg;
+ else
+ rq->blkg = q->root_blkg;
+ rcu_read_unlock();
+}
+
+/**
+ * blk_rq_disassociate_blkg - undo blk_rq_associate_blkg()
+ * @rq: request of interest
+ */
+static inline void blk_rq_disassociate_blkg(struct request *rq)
+{
+ if (rq->blkg && rq->blkg != rq->q->root_blkg)
+ blkg_put(rq->blkg);
+ rq->blkg = NULL;
+}
+
#else /* CONFIG_BLK_CGROUP */
struct blkcg {
@@ -781,6 +820,9 @@ static inline struct request_list *blk_rq_rl(struct request *rq) { return &rq->q
static inline bool blkcg_bio_issue_check(struct request_queue *q,
struct bio *bio) { return true; }
+static inline void blk_rq_associate_blkg(struct request *rq, struct blkcg *blkcg) { }
+static inline void blk_rq_disassociate_blkg(struct request *rq) { }
+
#define blk_queue_for_each_rl(rl, q) \
for ((rl) = &(q)->root_rl; (rl); (rl) = NULL)
--
2.9.5
--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
