summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorJens Axboe <axboe@kernel.dk>2019-12-29 22:15:21 -0700
committerJens Axboe <axboe@kernel.dk>2020-01-17 08:27:16 -0700
commit619c842f5985b0a237252cc631ee3508841dc849 (patch)
treeaffdbfcc3e7c63124b1da24b370a11252735a904
parentbb2d58c57d22e4fc290f3eb6a072af09c1d7e4b6 (diff)
blk-mq: allocate tags in batchesfor-5.6/block-test
Instead of grabbing tags one by one, grab a batch and store the local cache in the software queue. Then subsequent tag allocations can just grab free tags from there, without having to hit the shared tag map. We flush these batches out if we run out of tags on the hardware queue. The intent here is this should rarely happen. This works very well in practice, with anywhere from 40-60 batch counts seen regularly in testing. Signed-off-by: Jens Axboe <axboe@kernel.dk>
-rw-r--r--block/blk-mq-debugfs.c18
-rw-r--r--block/blk-mq-tag.c110
-rw-r--r--block/blk-mq-tag.h3
-rw-r--r--block/blk-mq.c16
-rw-r--r--block/blk-mq.h5
-rw-r--r--include/linux/blk-mq.h2
6 files changed, 149 insertions, 5 deletions
diff --git a/block/blk-mq-debugfs.c b/block/blk-mq-debugfs.c
index e789f830ff59..914be72d080e 100644
--- a/block/blk-mq-debugfs.c
+++ b/block/blk-mq-debugfs.c
@@ -659,6 +659,23 @@ CTX_RQ_SEQ_OPS(default, HCTX_TYPE_DEFAULT);
CTX_RQ_SEQ_OPS(read, HCTX_TYPE_READ);
CTX_RQ_SEQ_OPS(poll, HCTX_TYPE_POLL);
+static ssize_t ctx_tag_hit_write(void *data, const char __user *buf,
+ size_t count, loff_t *ppos)
+{
+ struct blk_mq_ctx *ctx = data;
+
+ ctx->tag_hit = ctx->tag_refill = 0;
+ return count;
+}
+
+static int ctx_tag_hit_show(void *data, struct seq_file *m)
+{
+ struct blk_mq_ctx *ctx = data;
+
+ seq_printf(m, "hit=%lu refills=%lu\n", ctx->tag_hit, ctx->tag_refill);
+ return 0;
+}
+
static int ctx_dispatched_show(void *data, struct seq_file *m)
{
struct blk_mq_ctx *ctx = data;
@@ -800,6 +817,7 @@ static const struct blk_mq_debugfs_attr blk_mq_debugfs_ctx_attrs[] = {
{"read_rq_list", 0400, .seq_ops = &ctx_read_rq_list_seq_ops},
{"poll_rq_list", 0400, .seq_ops = &ctx_poll_rq_list_seq_ops},
{"dispatched", 0600, ctx_dispatched_show, ctx_dispatched_write},
+ {"tag_hit", 0600, ctx_tag_hit_show, ctx_tag_hit_write},
{"merged", 0600, ctx_merged_show, ctx_merged_write},
{"completed", 0600, ctx_completed_show, ctx_completed_write},
{},
diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index fbacde454718..fe2273eec312 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -8,9 +8,10 @@
*/
#include <linux/kernel.h>
#include <linux/module.h>
-
#include <linux/blk-mq.h>
#include <linux/delay.h>
+#include <linux/cpu.h>
+
#include "blk.h"
#include "blk-mq.h"
#include "blk-mq-tag.h"
@@ -99,6 +100,103 @@ static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
return __sbitmap_queue_get(bt);
}
+void blk_mq_tag_ctx_flush_batch(struct blk_mq_hw_ctx *hctx,
+ struct blk_mq_ctx *ctx)
+{
+ struct sbitmap_queue *bt = &hctx->tags->bitmap_tags;
+ unsigned int i;
+
+ for (i = 0; i < hctx->queue->tag_set->nr_maps; i++) {
+ struct blk_mq_ctx_type *type = &ctx->type[i];
+
+ if (!type->tags)
+ continue;
+
+ __sbitmap_queue_clear_batch(bt, type->tag_offset, type->tags);
+ type->tags = 0;
+ }
+}
+
+static void ctx_flush_ipi(void *data)
+{
+ struct blk_mq_hw_ctx *hctx = data;
+ struct blk_mq_ctx *ctx;
+
+ ctx = __blk_mq_get_ctx(hctx->queue, smp_processor_id());
+ blk_mq_tag_ctx_flush_batch(hctx, ctx);
+ atomic_dec(&hctx->flush_pending);
+}
+
+static void blk_mq_tag_flush_batches(struct blk_mq_hw_ctx *hctx)
+{
+ int cpu, err;
+
+ if (atomic_cmpxchg(&hctx->flush_pending, 0, hctx->nr_ctx))
+ return;
+ cpus_read_lock();
+ for_each_cpu(cpu, hctx->cpumask) {
+ err = smp_call_function_single(cpu, ctx_flush_ipi, hctx, 0);
+ if (err)
+ atomic_dec(&hctx->flush_pending);
+ }
+ cpus_read_unlock();
+}
+
+void blk_mq_tag_queue_flush_batches(struct request_queue *q)
+{
+ struct blk_mq_hw_ctx *hctx;
+ unsigned int i;
+
+ queue_for_each_hw_ctx(q, hctx, i)
+ blk_mq_tag_flush_batches(hctx);
+}
+
+static int blk_mq_get_tag_batch(struct blk_mq_alloc_data *data)
+{
+ struct blk_mq_hw_ctx *hctx = data->hctx;
+ struct blk_mq_ctx_type *type;
+ struct blk_mq_ctx *ctx = data->ctx;
+ struct blk_mq_tags *tags;
+ struct sbitmap_queue *bt;
+ int tag = -1;
+
+ if (!ctx || (data->flags & BLK_MQ_REQ_INTERNAL))
+ return -1;
+
+ tags = hctx->tags;
+ bt = &tags->bitmap_tags;
+ /* don't do batches for round-robin or (very) sparse maps */
+ if (bt->round_robin || bt->sb.shift < ilog2(BITS_PER_LONG) - 1)
+ return -1;
+
+ /* we could make do with preempt disable, but we need to block flush */
+ local_irq_disable();
+ if (unlikely(ctx->cpu != smp_processor_id()))
+ goto out;
+
+ type = &ctx->type[hctx->type];
+
+ if (type->tags) {
+get_tag:
+ ctx->tag_hit++;
+
+ tag = __ffs(type->tags);
+ type->tags &= ~(1UL << tag);
+ tag += type->tag_offset;
+out:
+ local_irq_enable();
+ return tag;
+ }
+
+ /* no current tag cache, attempt to refill a batch */
+ if (!__sbitmap_queue_get_batch(bt, &type->tag_offset, &type->tags)) {
+ ctx->tag_refill++;
+ goto get_tag;
+ }
+
+ goto out;
+}
+
unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
{
struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
@@ -116,8 +214,13 @@ unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
bt = &tags->breserved_tags;
tag_offset = 0;
} else {
- bt = &tags->bitmap_tags;
tag_offset = tags->nr_reserved_tags;
+
+ tag = blk_mq_get_tag_batch(data);
+ if (tag != -1)
+ goto found_tag;
+
+ bt = &tags->bitmap_tags;
}
tag = __blk_mq_get_tag(data, bt);
@@ -146,6 +249,9 @@ unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
if (tag != -1)
break;
+ if (!(data->flags & BLK_MQ_REQ_RESERVED))
+ blk_mq_tag_flush_batches(data->hctx);
+
sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
tag = __blk_mq_get_tag(data, bt);
diff --git a/block/blk-mq-tag.h b/block/blk-mq-tag.h
index 15bc74acb57e..b5964fff1630 100644
--- a/block/blk-mq-tag.h
+++ b/block/blk-mq-tag.h
@@ -34,6 +34,9 @@ extern int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
extern void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool);
void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
void *priv);
+void blk_mq_tag_queue_flush_batches(struct request_queue *q);
+void blk_mq_tag_ctx_flush_batch(struct blk_mq_hw_ctx *hctx,
+ struct blk_mq_ctx *ctx);
static inline struct sbq_wait_state *bt_wait_ptr(struct sbitmap_queue *bt,
struct blk_mq_hw_ctx *hctx)
diff --git a/block/blk-mq.c b/block/blk-mq.c
index cc48a0ffa5ec..81140f61a7c9 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -2255,6 +2255,8 @@ static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
ctx = __blk_mq_get_ctx(hctx->queue, cpu);
type = hctx->type;
+ blk_mq_tag_ctx_flush_batch(hctx, ctx);
+
spin_lock(&ctx->lock);
if (!list_empty(&ctx->type[type].rq_list)) {
list_splice_init(&ctx->type[type].rq_list, &tmp);
@@ -2436,8 +2438,10 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
__ctx->cpu = i;
spin_lock_init(&__ctx->lock);
- for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
+ for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++) {
INIT_LIST_HEAD(&__ctx->type[k].rq_list);
+ __ctx->type[k].tags = 0;
+ }
/*
* Set local node, IFF we have more than one hw queue. If
@@ -2521,6 +2525,7 @@ static void blk_mq_map_swqueue(struct request_queue *q)
}
hctx = blk_mq_map_queue_type(q, j, i);
+ ctx->type[j].tags = 0;
ctx->hctxs[j] = hctx;
/*
* If the CPU is already set in the mask, then we've
@@ -2542,9 +2547,11 @@ static void blk_mq_map_swqueue(struct request_queue *q)
BUG_ON(!hctx->nr_ctx);
}
- for (; j < HCTX_MAX_TYPES; j++)
+ for (; j < HCTX_MAX_TYPES; j++) {
ctx->hctxs[j] = blk_mq_map_queue_type(q,
HCTX_TYPE_DEFAULT, i);
+ ctx->type[j].tags = 0;
+ }
}
queue_for_each_hw_ctx(q, hctx, i) {
@@ -3298,8 +3305,11 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
return;
- list_for_each_entry(q, &set->tag_list, tag_set_list)
+ list_for_each_entry(q, &set->tag_list, tag_set_list) {
+ blk_mq_tag_queue_flush_batches(q);
blk_mq_freeze_queue(q);
+ }
+
/*
* Switch IO scheduler to 'none', cleaning up the data associated
* with the previous scheduler. We will switch back once we are done
diff --git a/block/blk-mq.h b/block/blk-mq.h
index 271f16771499..b6095cc50921 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -14,6 +14,10 @@ struct blk_mq_ctxs {
struct blk_mq_ctx_type {
struct list_head rq_list;
+
+ /* tag batch cache */
+ unsigned long tags;
+ unsigned int tag_offset;
};
/**
@@ -23,6 +27,7 @@ struct blk_mq_ctx {
struct {
spinlock_t lock;
struct blk_mq_ctx_type type[HCTX_MAX_TYPES];
+ unsigned long tag_hit, tag_refill;
} ____cacheline_aligned_in_smp;
unsigned int cpu;
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 11cfd6470b1a..2c6a8657a72c 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -140,6 +140,8 @@ struct blk_mq_hw_ctx {
*/
atomic_t nr_active;
+ atomic_t flush_pending;
+
/** @cpuhp_dead: List to store request if some CPU die. */
struct hlist_node cpuhp_dead;
/** @kobj: Kernel object for sysfs. */