These two patches enable a driver to control if tags should be initialized or
use an existing tags structure.
This is for example needed for the NVMe driver, that have a shared number of
queues for each block device it expose.
Matias Bjorling (2):
blk-mq: allow request queues to share tags map
blk-mq: add maps_tags fn and add usage
block/blk-mq-tag.c | 22 +++++++++++++++++++++-
block/blk-mq.c | 11 +++++++----
include/linux/blk-mq.h | 12 ++++++++++++
3 files changed, 40 insertions(+), 5 deletions(-)
--
1.8.3.2
Now that the tags mapping allows shared mapping, update blk-mq to allow
the driver to control initialization of the tags structure.
Signed-off-by: Matias Bjorling <[email protected]>
---
block/blk-mq.c | 11 +++++++----
include/linux/blk-mq.h | 12 ++++++++++++
2 files changed, 19 insertions(+), 4 deletions(-)
diff --git a/block/blk-mq.c b/block/blk-mq.c
index f21ec96..9e25555 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -1093,7 +1093,7 @@ static size_t order_to_size(unsigned int order)
}
static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
- unsigned int reserved_tags, int node)
+ struct blk_mq_reg *reg, int node)
{
unsigned int i, j, entries_per_page, max_order = 4;
size_t rq_size, left;
@@ -1150,7 +1150,7 @@ static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
}
}
- if (i < (reserved_tags + BLK_MQ_TAG_MIN))
+ if (i < (reg->reserved_tags + BLK_MQ_TAG_MIN))
goto err_rq_map;
else if (i != hctx->queue_depth) {
hctx->queue_depth = i;
@@ -1158,7 +1158,10 @@ static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
__func__, i);
}
- hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
+ if (reg->ops->map_tags)
+ hctx->tags = reg->ops->map_tags(hctx, reg, node);
+ else
+ hctx->tags = blk_mq_init_tags(hctx->queue_depth, reg->reserved_tags, node);
if (!hctx->tags) {
err_rq_map:
blk_mq_free_rq_map(hctx);
@@ -1198,7 +1201,7 @@ static int blk_mq_init_hw_queues(struct request_queue *q,
blk_mq_hctx_notify, hctx);
blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
- if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
+ if (blk_mq_init_rq_map(hctx, reg, node))
break;
/*
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 746042ff..eb6df6a 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -66,6 +66,8 @@ typedef struct blk_mq_hw_ctx *(alloc_hctx_fn)(struct blk_mq_reg *,unsigned int);
typedef void (free_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int);
typedef int (init_hctx_fn)(struct blk_mq_hw_ctx *, void *, unsigned int);
typedef void (exit_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int);
+typedef struct blk_mq_tags *(map_tags_fn)(struct blk_mq_hw_ctx *,
+ struct blk_mq_reg *, unsigned int node);
struct blk_mq_ops {
/*
@@ -96,6 +98,12 @@ struct blk_mq_ops {
*/
init_hctx_fn *init_hctx;
exit_hctx_fn *exit_hctx;
+
+ /*
+ * If a tags structure is to be shared between multiple request queues,
+ * allow the driver to initialize or reuse a tags structure.
+ */
+ map_tags_fn *map_tags;
};
enum {
@@ -138,6 +146,10 @@ void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx);
void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx);
void blk_mq_start_stopped_hw_queues(struct request_queue *q);
+struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
+ unsigned int reserved_tags, int node);
+void blk_mq_init_tags_shared(struct blk_mq_tags *tags);
+
/*
* Driver command data is immediately after the request. So subtract request
* size to get back to the original request.
--
1.8.3.2
Some devices, such as NVMe, initializes a limited number of queues. These are
shared by all the block devices. Allow a driver to tap into the tags
structure and decide if an existing tag structure should be used.
Signed-off-by: Matias Bjorling <[email protected]>
---
block/blk-mq-tag.c | 22 +++++++++++++++++++++-
1 file changed, 21 insertions(+), 1 deletion(-)
diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index d64a02f..c7b49d5 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -1,4 +1,5 @@
#include <linux/kernel.h>
+#include <linux/kref.h>
#include <linux/module.h>
#include <linux/percpu_ida.h>
@@ -18,6 +19,8 @@ struct blk_mq_tags {
struct percpu_ida free_tags;
struct percpu_ida reserved_tags;
+
+ struct kref ref_count;
};
void blk_mq_wait_for_tags(struct blk_mq_tags *tags)
@@ -116,6 +119,12 @@ void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
kfree(tag_map);
}
+void blk_mq_init_tags_shared(struct blk_mq_tags *tags)
+{
+ kref_get(&tags->ref_count);
+}
+EXPORT_SYMBOL_GPL(blk_mq_init_tags_shared);
+
struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
unsigned int reserved_tags, int node)
{
@@ -145,6 +154,8 @@ struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
tags->nr_max_cache = nr_cache;
tags->nr_batch_move = max(1u, nr_cache / 2);
+ kref_init(&tags->ref_count);
+
ret = __percpu_ida_init(&tags->free_tags, tags->nr_tags -
tags->nr_reserved_tags,
tags->nr_max_cache,
@@ -171,14 +182,23 @@ err_free_tags:
kfree(tags);
return NULL;
}
+EXPORT_SYMBOL_GPL(blk_mq_init_tags);
-void blk_mq_free_tags(struct blk_mq_tags *tags)
+static void __blk_mq_free_tags(struct kref *kref)
{
+ struct blk_mq_tags *tags = container_of(kref, struct blk_mq_tags,
+ ref_count);
+
percpu_ida_destroy(&tags->free_tags);
percpu_ida_destroy(&tags->reserved_tags);
kfree(tags);
}
+void blk_mq_free_tags(struct blk_mq_tags *tags)
+{
+ kref_put(&tags->ref_count, __blk_mq_free_tags);
+}
+
ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
{
char *orig_page = page;
--
1.8.3.2