[RESEND][PATCH v2 0/2] Fix misuse of blk_rq_stats in blk-iolatency

[RESEND][PATCH v2 0/2] Fix misuse of blk_rq_stats in blk-iolatency

[Pavel Begunkov (Silence)]

From: Pavel Begunkov <asml.silence@gmail.com>

There are implicit assumptions about struct blk_rq_stats, which make
it's very easy to misuse. The first patch fixes a bug caused by that.
The second employs type-system to prevent recurrences.

v2: rebase + reformulate commit messages (no code changes)

Acked-by: Josef Bacik <josef@toxicpanda.com>

Pavel Begunkov (2):
  blk-iolatency: Fix zero mean in previous stats
  blk-stats: Introduce explicit stat staging buffers

 block/blk-iolatency.c     | 60 ++++++++++++++++++++++++++++++---------
 block/blk-stat.c          | 48 +++++++++++++++++++++++--------
 block/blk-stat.h          |  9 ++++--
 include/linux/blk_types.h |  6 ++++
 4 files changed, 94 insertions(+), 29 deletions(-)

-- 
2.22.0

[PATCH v2 1/2] blk-iolatency: Fix zero mean in previous stats

[Pavel Begunkov (Silence)]

From: Pavel Begunkov <asml.silence@gmail.com>

struct blk_rq_stat has two implicit states in which it can be:
(1) per-cpu intermediate stats (i.e. staging, intermediate)
(2) final stats / aggregation of (1) (see blk_rq_stat_collect)

The states use different sets of fields. E.g. (1) uses @batch but not
@mean, and vise versa for (2). Functions operating on struct blk_rq_stat
have implicit assumptions about its state.

blk_rq_stat_sum() require @src argument to be in (1) and @dst in (2).
iolatency_check_latencies() violates that, and as a result,
iolat->cur_stat.rqs.mean is always 0 for non-ssd devices.

Use 2 distinct functions instead:
blk_rq_stat_collect() to collect intermediate stats (1)
blk_rq_stat_merge() to merge accumulated stats (2)

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
 block/blk-iolatency.c | 21 ++++++++++++++++-----
 block/blk-stat.c      | 20 ++++++++++++++++++--
 block/blk-stat.h      |  3 ++-
 3 files changed, 36 insertions(+), 8 deletions(-)

diff --git a/block/blk-iolatency.c b/block/blk-iolatency.c
index c128d50cb410..895c6e955f97 100644
--- a/block/blk-iolatency.c
+++ b/block/blk-iolatency.c
@@ -199,7 +199,7 @@ static inline void latency_stat_init(struct iolatency_grp *iolat,
 		blk_rq_stat_init(&stat->rqs);
 }
 
-static inline void latency_stat_sum(struct iolatency_grp *iolat,
+static inline void latency_stat_merge(struct iolatency_grp *iolat,
 				    struct latency_stat *sum,
 				    struct latency_stat *stat)
 {
@@ -207,7 +207,18 @@ static inline void latency_stat_sum(struct iolatency_grp *iolat,
 		sum->ps.total += stat->ps.total;
 		sum->ps.missed += stat->ps.missed;
 	} else
-		blk_rq_stat_sum(&sum->rqs, &stat->rqs);
+		blk_rq_stat_merge(&sum->rqs, &stat->rqs);
+}
+
+static inline void latency_stat_collect(struct iolatency_grp *iolat,
+					struct latency_stat *sum,
+					struct latency_stat *stat)
+{
+	if (iolat->ssd) {
+		sum->ps.total += stat->ps.total;
+		sum->ps.missed += stat->ps.missed;
+	} else
+		blk_rq_stat_collect(&sum->rqs, &stat->rqs);
 }
 
 static inline void latency_stat_record_time(struct iolatency_grp *iolat,
@@ -531,7 +542,7 @@ static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
 	for_each_online_cpu(cpu) {
 		struct latency_stat *s;
 		s = per_cpu_ptr(iolat->stats, cpu);
-		latency_stat_sum(iolat, &stat, s);
+		latency_stat_collect(iolat, &stat, s);
 		latency_stat_init(iolat, s);
 	}
 	preempt_enable();
@@ -552,7 +563,7 @@ static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
 	/* Somebody beat us to the punch, just bail. */
 	spin_lock_irqsave(&lat_info->lock, flags);
 
-	latency_stat_sum(iolat, &iolat->cur_stat, &stat);
+	latency_stat_merge(iolat, &iolat->cur_stat, &stat);
 	lat_info->nr_samples -= iolat->nr_samples;
 	lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
 	iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
@@ -896,7 +907,7 @@ static size_t iolatency_ssd_stat(struct iolatency_grp *iolat, char *buf,
 	for_each_online_cpu(cpu) {
 		struct latency_stat *s;
 		s = per_cpu_ptr(iolat->stats, cpu);
-		latency_stat_sum(iolat, &stat, s);
+		latency_stat_collect(iolat, &stat, s);
 	}
 	preempt_enable();
 
diff --git a/block/blk-stat.c b/block/blk-stat.c
index 940f15d600f8..78389182b5d0 100644
--- a/block/blk-stat.c
+++ b/block/blk-stat.c
@@ -26,7 +26,7 @@ void blk_rq_stat_init(struct blk_rq_stat *stat)
 }
 
 /* src is a per-cpu stat, mean isn't initialized */
-void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
+void blk_rq_stat_collect(struct blk_rq_stat *dst, struct blk_rq_stat *src)
 {
 	if (!src->nr_samples)
 		return;
@@ -40,6 +40,21 @@ void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
 	dst->nr_samples += src->nr_samples;
 }
 
+void blk_rq_stat_merge(struct blk_rq_stat *dst, struct blk_rq_stat *src)
+{
+	if (!src->nr_samples)
+		return;
+
+	dst->min = min(dst->min, src->min);
+	dst->max = max(dst->max, src->max);
+
+	dst->mean = div_u64(src->mean * src->nr_samples +
+				dst->mean * dst->nr_samples,
+				dst->nr_samples + src->nr_samples);
+
+	dst->nr_samples += src->nr_samples;
+}
+
 void blk_rq_stat_add(struct blk_rq_stat *stat, u64 value)
 {
 	stat->min = min(stat->min, value);
@@ -90,7 +105,8 @@ static void blk_stat_timer_fn(struct timer_list *t)
 
 		cpu_stat = per_cpu_ptr(cb->cpu_stat, cpu);
 		for (bucket = 0; bucket < cb->buckets; bucket++) {
-			blk_rq_stat_sum(&cb->stat[bucket], &cpu_stat[bucket]);
+			blk_rq_stat_collect(&cb->stat[bucket],
+					    &cpu_stat[bucket]);
 			blk_rq_stat_init(&cpu_stat[bucket]);
 		}
 	}
diff --git a/block/blk-stat.h b/block/blk-stat.h
index 17b47a86eefb..5597ecc34ef5 100644
--- a/block/blk-stat.h
+++ b/block/blk-stat.h
@@ -165,7 +165,8 @@ static inline void blk_stat_activate_msecs(struct blk_stat_callback *cb,
 }
 
 void blk_rq_stat_add(struct blk_rq_stat *, u64);
-void blk_rq_stat_sum(struct blk_rq_stat *, struct blk_rq_stat *);
+void blk_rq_stat_collect(struct blk_rq_stat *dst, struct blk_rq_stat *src);
+void blk_rq_stat_merge(struct blk_rq_stat *dst, struct blk_rq_stat *src);
 void blk_rq_stat_init(struct blk_rq_stat *);
 
 #endif
-- 
2.22.0

[PATCH v2 2/2] blk-stats: Introduce explicit stat staging buffers

[Pavel Begunkov (Silence)]

From: Pavel Begunkov <asml.silence@gmail.com>

Split struct blk_rq_stat into 2:
1. struct blk_rq_stat_staging is for intermediate/staging stats.
These are usually per-cpu and in a hot-path
2. struct blk_rq_stat for accumulating staging stats

That duplicates code, but
1. prevents misuses (compile-time check by type-system)
2. reduces memory needed (inc. per-cpu)
3. makes it easier to extend stats

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
---
 block/blk-iolatency.c     | 41 +++++++++++++++++++++++++++++----------
 block/blk-stat.c          | 30 +++++++++++++++++-----------
 block/blk-stat.h          |  8 +++++---
 include/linux/blk_types.h |  6 ++++++
 4 files changed, 61 insertions(+), 24 deletions(-)

diff --git a/block/blk-iolatency.c b/block/blk-iolatency.c
index 895c6e955f97..08a32dc4a7d6 100644
--- a/block/blk-iolatency.c
+++ b/block/blk-iolatency.c
@@ -130,9 +130,16 @@ struct latency_stat {
 	};
 };
 
+struct latency_stat_staging {
+	union {
+		struct percentile_stats ps;
+		struct blk_rq_stat_staging rqs;
+	};
+};
+
 struct iolatency_grp {
 	struct blkg_policy_data pd;
-	struct latency_stat __percpu *stats;
+	struct latency_stat_staging __percpu *stats;
 	struct latency_stat cur_stat;
 	struct blk_iolatency *blkiolat;
 	struct rq_depth rq_depth;
@@ -199,6 +206,16 @@ static inline void latency_stat_init(struct iolatency_grp *iolat,
 		blk_rq_stat_init(&stat->rqs);
 }
 
+static inline void latency_stat_init_staging(struct iolatency_grp *iolat,
+					     struct latency_stat_staging *stat)
+{
+	if (iolat->ssd) {
+		stat->ps.total = 0;
+		stat->ps.missed = 0;
+	} else
+		blk_rq_stat_init_staging(&stat->rqs);
+}
+
 static inline void latency_stat_merge(struct iolatency_grp *iolat,
 				    struct latency_stat *sum,
 				    struct latency_stat *stat)
@@ -212,7 +229,7 @@ static inline void latency_stat_merge(struct iolatency_grp *iolat,
 
 static inline void latency_stat_collect(struct iolatency_grp *iolat,
 					struct latency_stat *sum,
-					struct latency_stat *stat)
+					struct latency_stat_staging *stat)
 {
 	if (iolat->ssd) {
 		sum->ps.total += stat->ps.total;
@@ -224,7 +241,8 @@ static inline void latency_stat_collect(struct iolatency_grp *iolat,
 static inline void latency_stat_record_time(struct iolatency_grp *iolat,
 					    u64 req_time)
 {
-	struct latency_stat *stat = get_cpu_ptr(iolat->stats);
+	struct latency_stat_staging *stat = get_cpu_ptr(iolat->stats);
+
 	if (iolat->ssd) {
 		if (req_time >= iolat->min_lat_nsec)
 			stat->ps.missed++;
@@ -540,10 +558,11 @@ static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
 	latency_stat_init(iolat, &stat);
 	preempt_disable();
 	for_each_online_cpu(cpu) {
-		struct latency_stat *s;
+		struct latency_stat_staging *s;
+
 		s = per_cpu_ptr(iolat->stats, cpu);
 		latency_stat_collect(iolat, &stat, s);
-		latency_stat_init(iolat, s);
+		latency_stat_init_staging(iolat, s);
 	}
 	preempt_enable();
 
@@ -905,7 +924,8 @@ static size_t iolatency_ssd_stat(struct iolatency_grp *iolat, char *buf,
 	latency_stat_init(iolat, &stat);
 	preempt_disable();
 	for_each_online_cpu(cpu) {
-		struct latency_stat *s;
+		struct latency_stat_staging *s;
+
 		s = per_cpu_ptr(iolat->stats, cpu);
 		latency_stat_collect(iolat, &stat, s);
 	}
@@ -954,8 +974,8 @@ static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp,
 	iolat = kzalloc_node(sizeof(*iolat), gfp, q->node);
 	if (!iolat)
 		return NULL;
-	iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
-				       __alignof__(struct latency_stat), gfp);
+	iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat_staging),
+				__alignof__(struct latency_stat_staging), gfp);
 	if (!iolat->stats) {
 		kfree(iolat);
 		return NULL;
@@ -978,9 +998,10 @@ static void iolatency_pd_init(struct blkg_policy_data *pd)
 		iolat->ssd = false;
 
 	for_each_possible_cpu(cpu) {
-		struct latency_stat *stat;
+		struct latency_stat_staging *stat;
+
 		stat = per_cpu_ptr(iolat->stats, cpu);
-		latency_stat_init(iolat, stat);
+		latency_stat_init_staging(iolat, stat);
 	}
 
 	latency_stat_init(iolat, &iolat->cur_stat);
diff --git a/block/blk-stat.c b/block/blk-stat.c
index 78389182b5d0..d892ad2cb938 100644
--- a/block/blk-stat.c
+++ b/block/blk-stat.c
@@ -18,15 +18,22 @@ struct blk_queue_stats {
 	bool enable_accounting;
 };
 
+void blk_rq_stat_init_staging(struct blk_rq_stat_staging *stat)
+{
+	stat->min = -1ULL;
+	stat->max = 0;
+	stat->batch = 0;
+	stat->nr_samples = 0;
+}
+
 void blk_rq_stat_init(struct blk_rq_stat *stat)
 {
 	stat->min = -1ULL;
 	stat->max = stat->nr_samples = stat->mean = 0;
-	stat->batch = 0;
 }
 
-/* src is a per-cpu stat, mean isn't initialized */
-void blk_rq_stat_collect(struct blk_rq_stat *dst, struct blk_rq_stat *src)
+void blk_rq_stat_collect(struct blk_rq_stat *dst,
+			 struct blk_rq_stat_staging *src)
 {
 	if (!src->nr_samples)
 		return;
@@ -55,7 +62,7 @@ void blk_rq_stat_merge(struct blk_rq_stat *dst, struct blk_rq_stat *src)
 	dst->nr_samples += src->nr_samples;
 }
 
-void blk_rq_stat_add(struct blk_rq_stat *stat, u64 value)
+void blk_rq_stat_add(struct blk_rq_stat_staging *stat, u64 value)
 {
 	stat->min = min(stat->min, value);
 	stat->max = max(stat->max, value);
@@ -67,7 +74,7 @@ void blk_stat_add(struct request *rq, u64 now)
 {
 	struct request_queue *q = rq->q;
 	struct blk_stat_callback *cb;
-	struct blk_rq_stat *stat;
+	struct blk_rq_stat_staging *stat;
 	int bucket;
 	u64 value;
 
@@ -101,13 +108,13 @@ static void blk_stat_timer_fn(struct timer_list *t)
 		blk_rq_stat_init(&cb->stat[bucket]);
 
 	for_each_online_cpu(cpu) {
-		struct blk_rq_stat *cpu_stat;
+		struct blk_rq_stat_staging *cpu_stat;
 
 		cpu_stat = per_cpu_ptr(cb->cpu_stat, cpu);
 		for (bucket = 0; bucket < cb->buckets; bucket++) {
 			blk_rq_stat_collect(&cb->stat[bucket],
 					    &cpu_stat[bucket]);
-			blk_rq_stat_init(&cpu_stat[bucket]);
+			blk_rq_stat_init_staging(&cpu_stat[bucket]);
 		}
 	}
 
@@ -131,8 +138,9 @@ blk_stat_alloc_callback(void (*timer_fn)(struct blk_stat_callback *),
 		kfree(cb);
 		return NULL;
 	}
-	cb->cpu_stat = __alloc_percpu(buckets * sizeof(struct blk_rq_stat),
-				      __alignof__(struct blk_rq_stat));
+	cb->cpu_stat = __alloc_percpu(
+				buckets * sizeof(struct blk_rq_stat_staging),
+				__alignof__(struct blk_rq_stat_staging));
 	if (!cb->cpu_stat) {
 		kfree(cb->stat);
 		kfree(cb);
@@ -155,11 +163,11 @@ void blk_stat_add_callback(struct request_queue *q,
 	int cpu;
 
 	for_each_possible_cpu(cpu) {
-		struct blk_rq_stat *cpu_stat;
+		struct blk_rq_stat_staging *cpu_stat;
 
 		cpu_stat = per_cpu_ptr(cb->cpu_stat, cpu);
 		for (bucket = 0; bucket < cb->buckets; bucket++)
-			blk_rq_stat_init(&cpu_stat[bucket]);
+			blk_rq_stat_init_staging(&cpu_stat[bucket]);
 	}
 
 	spin_lock(&q->stats->lock);
diff --git a/block/blk-stat.h b/block/blk-stat.h
index 5597ecc34ef5..e5c753fbd6e6 100644
--- a/block/blk-stat.h
+++ b/block/blk-stat.h
@@ -30,7 +30,7 @@ struct blk_stat_callback {
 	/**
 	 * @cpu_stat: Per-cpu statistics buckets.
 	 */
-	struct blk_rq_stat __percpu *cpu_stat;
+	struct blk_rq_stat_staging __percpu *cpu_stat;
 
 	/**
 	 * @bucket_fn: Given a request, returns which statistics bucket it
@@ -164,9 +164,11 @@ static inline void blk_stat_activate_msecs(struct blk_stat_callback *cb,
 	mod_timer(&cb->timer, jiffies + msecs_to_jiffies(msecs));
 }
 
-void blk_rq_stat_add(struct blk_rq_stat *, u64);
-void blk_rq_stat_collect(struct blk_rq_stat *dst, struct blk_rq_stat *src);
+void blk_rq_stat_add(struct blk_rq_stat_staging *stat, u64);
+void blk_rq_stat_collect(struct blk_rq_stat *dst,
+			 struct blk_rq_stat_staging *src);
 void blk_rq_stat_merge(struct blk_rq_stat *dst, struct blk_rq_stat *src);
 void blk_rq_stat_init(struct blk_rq_stat *);
+void blk_rq_stat_init_staging(struct blk_rq_stat_staging *stat);
 
 #endif
diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
index 6f7a8647cefb..5a7d8b70fcf0 100644
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h
@@ -452,7 +452,13 @@ struct blk_rq_stat {
 	u64 min;
 	u64 max;
 	u32 nr_samples;
+};
+
+struct blk_rq_stat_staging {
+	u64 min;
+	u64 max;
 	u64 batch;
+	u32 nr_samples;
 };
 
 #endif /* __LINUX_BLK_TYPES_H */
-- 
2.22.0

Re: [RESEND][PATCH v2 0/2] Fix misuse of blk_rq_stats in blk-iolatency

[Pavel Begunkov]

[-- Attachment #1.1: Type: text/plain, Size: 1096 bytes --]

Jens, could you please at last take a look?

The patches have been stalling here for almost 2 months
(see
https://lkml.kernel.org/lkml/ac0700a1-0984-417b-d5d8-35c4ba56f6f6@gmail.com/T/)

Thanks

On 06/09/2019 17:42, Pavel Begunkov (Silence) wrote:
> From: Pavel Begunkov <asml.silence@gmail.com>
> 
> There are implicit assumptions about struct blk_rq_stats, which make
> it's very easy to misuse. The first patch fixes a bug caused by that.
> The second employs type-system to prevent recurrences.
> 
> v2: rebase + reformulate commit messages (no code changes)
> 
> Acked-by: Josef Bacik <josef@toxicpanda.com>
> 
> Pavel Begunkov (2):
>   blk-iolatency: Fix zero mean in previous stats
>   blk-stats: Introduce explicit stat staging buffers
> 
>  block/blk-iolatency.c     | 60 ++++++++++++++++++++++++++++++---------
>  block/blk-stat.c          | 48 +++++++++++++++++++++++--------
>  block/blk-stat.h          |  9 ++++--
>  include/linux/blk_types.h |  6 ++++
>  4 files changed, 94 insertions(+), 29 deletions(-)
> 

-- 
Yours sincerely,
Pavel Begunkov


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]