block: kyber: make kyber more friendly with merging

Currently, kyber is very unfriendly with merging. kyber depends
on ctx rq_list to do merging, however, most of time, it will not
leave any requests in ctx rq_list. This is because even if tokens
of one domain is used up, kyber will try to dispatch requests
from other domain and flush the rq_list there.

To improve this, we setup kyber_ctx_queue (kcq) which is similar
with ctx, but it has rq_lists for different domain and build same
mapping between kcq and khd as the ctx & hctx. Then we could merge,
insert and dispatch for different domains separately. At the same
time, only flush the rq_list of kcq when get domain token successfully.
Then if one domain token is used up, the requests could be left in
the rq_list of that domain and maybe merged with following io.

Following is my test result on machine with 8 cores and NVMe card
INTEL SSDPEKKR128G7

fio size=256m ioengine=libaio iodepth=64 direct=1 numjobs=8
seq/random
+------+---------------------------------------------------------------+
|patch?| bw(MB/s) |   iops    | slat(usec) |    clat(usec)   |  merge  |
+----------------------------------------------------------------------+
| w/o  |  606/612 | 151k/153k |  6.89/7.03 | 3349.21/3305.40 |   0/0   |
+----------------------------------------------------------------------+
| w/   | 1083/616 | 277k/154k |  4.93/6.95 | 1830.62/3279.95 | 223k/3k |
+----------------------------------------------------------------------+
When set numjobs to 16, the bw and iops could reach 1662MB/s and 425k
on my platform.

Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This commit is contained in:
Jianchao Wang 2018-05-30 10:47:40 -06:00 committed by Jens Axboe
parent 9c55873464
commit a6088845c2

View File

@ -72,6 +72,19 @@ static const unsigned int kyber_batch_size[] = {
[KYBER_OTHER] = 8, [KYBER_OTHER] = 8,
}; };
/*
* There is a same mapping between ctx & hctx and kcq & khd,
* we use request->mq_ctx->index_hw to index the kcq in khd.
*/
struct kyber_ctx_queue {
/*
* Used to ensure operations on rq_list and kcq_map to be an atmoic one.
* Also protect the rqs on rq_list when merge.
*/
spinlock_t lock;
struct list_head rq_list[KYBER_NUM_DOMAINS];
} ____cacheline_aligned_in_smp;
struct kyber_queue_data { struct kyber_queue_data {
struct request_queue *q; struct request_queue *q;
@ -99,6 +112,8 @@ struct kyber_hctx_data {
struct list_head rqs[KYBER_NUM_DOMAINS]; struct list_head rqs[KYBER_NUM_DOMAINS];
unsigned int cur_domain; unsigned int cur_domain;
unsigned int batching; unsigned int batching;
struct kyber_ctx_queue *kcqs;
struct sbitmap kcq_map[KYBER_NUM_DOMAINS];
wait_queue_entry_t domain_wait[KYBER_NUM_DOMAINS]; wait_queue_entry_t domain_wait[KYBER_NUM_DOMAINS];
struct sbq_wait_state *domain_ws[KYBER_NUM_DOMAINS]; struct sbq_wait_state *domain_ws[KYBER_NUM_DOMAINS];
atomic_t wait_index[KYBER_NUM_DOMAINS]; atomic_t wait_index[KYBER_NUM_DOMAINS];
@ -107,10 +122,8 @@ struct kyber_hctx_data {
static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags, static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
void *key); void *key);
static int rq_sched_domain(const struct request *rq) static unsigned int kyber_sched_domain(unsigned int op)
{ {
unsigned int op = rq->cmd_flags;
if ((op & REQ_OP_MASK) == REQ_OP_READ) if ((op & REQ_OP_MASK) == REQ_OP_READ)
return KYBER_READ; return KYBER_READ;
else if ((op & REQ_OP_MASK) == REQ_OP_WRITE && op_is_sync(op)) else if ((op & REQ_OP_MASK) == REQ_OP_WRITE && op_is_sync(op))
@ -284,6 +297,11 @@ static unsigned int kyber_sched_tags_shift(struct kyber_queue_data *kqd)
return kqd->q->queue_hw_ctx[0]->sched_tags->bitmap_tags.sb.shift; return kqd->q->queue_hw_ctx[0]->sched_tags->bitmap_tags.sb.shift;
} }
static int kyber_bucket_fn(const struct request *rq)
{
return kyber_sched_domain(rq->cmd_flags);
}
static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q) static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
{ {
struct kyber_queue_data *kqd; struct kyber_queue_data *kqd;
@ -297,7 +315,7 @@ static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
goto err; goto err;
kqd->q = q; kqd->q = q;
kqd->cb = blk_stat_alloc_callback(kyber_stat_timer_fn, rq_sched_domain, kqd->cb = blk_stat_alloc_callback(kyber_stat_timer_fn, kyber_bucket_fn,
KYBER_NUM_DOMAINS, kqd); KYBER_NUM_DOMAINS, kqd);
if (!kqd->cb) if (!kqd->cb)
goto err_kqd; goto err_kqd;
@ -376,6 +394,15 @@ static void kyber_exit_sched(struct elevator_queue *e)
kfree(kqd); kfree(kqd);
} }
static void kyber_ctx_queue_init(struct kyber_ctx_queue *kcq)
{
unsigned int i;
spin_lock_init(&kcq->lock);
for (i = 0; i < KYBER_NUM_DOMAINS; i++)
INIT_LIST_HEAD(&kcq->rq_list[i]);
}
static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
{ {
struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data; struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
@ -386,6 +413,24 @@ static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
if (!khd) if (!khd)
return -ENOMEM; return -ENOMEM;
khd->kcqs = kmalloc_array_node(hctx->nr_ctx,
sizeof(struct kyber_ctx_queue),
GFP_KERNEL, hctx->numa_node);
if (!khd->kcqs)
goto err_khd;
for (i = 0; i < hctx->nr_ctx; i++)
kyber_ctx_queue_init(&khd->kcqs[i]);
for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
if (sbitmap_init_node(&khd->kcq_map[i], hctx->nr_ctx,
ilog2(8), GFP_KERNEL, hctx->numa_node)) {
while (--i >= 0)
sbitmap_free(&khd->kcq_map[i]);
goto err_kcqs;
}
}
spin_lock_init(&khd->lock); spin_lock_init(&khd->lock);
for (i = 0; i < KYBER_NUM_DOMAINS; i++) { for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
@ -405,10 +450,22 @@ static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
kqd->async_depth); kqd->async_depth);
return 0; return 0;
err_kcqs:
kfree(khd->kcqs);
err_khd:
kfree(khd);
return -ENOMEM;
} }
static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
{ {
struct kyber_hctx_data *khd = hctx->sched_data;
int i;
for (i = 0; i < KYBER_NUM_DOMAINS; i++)
sbitmap_free(&khd->kcq_map[i]);
kfree(khd->kcqs);
kfree(hctx->sched_data); kfree(hctx->sched_data);
} }
@ -430,7 +487,7 @@ static void rq_clear_domain_token(struct kyber_queue_data *kqd,
nr = rq_get_domain_token(rq); nr = rq_get_domain_token(rq);
if (nr != -1) { if (nr != -1) {
sched_domain = rq_sched_domain(rq); sched_domain = kyber_sched_domain(rq->cmd_flags);
sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr, sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr,
rq->mq_ctx->cpu); rq->mq_ctx->cpu);
} }
@ -449,11 +506,51 @@ static void kyber_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
} }
} }
static bool kyber_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
{
struct kyber_hctx_data *khd = hctx->sched_data;
struct blk_mq_ctx *ctx = blk_mq_get_ctx(hctx->queue);
struct kyber_ctx_queue *kcq = &khd->kcqs[ctx->index_hw];
unsigned int sched_domain = kyber_sched_domain(bio->bi_opf);
struct list_head *rq_list = &kcq->rq_list[sched_domain];
bool merged;
spin_lock(&kcq->lock);
merged = blk_mq_bio_list_merge(hctx->queue, rq_list, bio);
spin_unlock(&kcq->lock);
blk_mq_put_ctx(ctx);
return merged;
}
static void kyber_prepare_request(struct request *rq, struct bio *bio) static void kyber_prepare_request(struct request *rq, struct bio *bio)
{ {
rq_set_domain_token(rq, -1); rq_set_domain_token(rq, -1);
} }
static void kyber_insert_requests(struct blk_mq_hw_ctx *hctx,
struct list_head *rq_list, bool at_head)
{
struct kyber_hctx_data *khd = hctx->sched_data;
struct request *rq, *next;
list_for_each_entry_safe(rq, next, rq_list, queuelist) {
unsigned int sched_domain = kyber_sched_domain(rq->cmd_flags);
struct kyber_ctx_queue *kcq = &khd->kcqs[rq->mq_ctx->index_hw];
struct list_head *head = &kcq->rq_list[sched_domain];
spin_lock(&kcq->lock);
if (at_head)
list_move(&rq->queuelist, head);
else
list_move_tail(&rq->queuelist, head);
sbitmap_set_bit(&khd->kcq_map[sched_domain],
rq->mq_ctx->index_hw);
blk_mq_sched_request_inserted(rq);
spin_unlock(&kcq->lock);
}
}
static void kyber_finish_request(struct request *rq) static void kyber_finish_request(struct request *rq)
{ {
struct kyber_queue_data *kqd = rq->q->elevator->elevator_data; struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
@ -472,7 +569,7 @@ static void kyber_completed_request(struct request *rq)
* Check if this request met our latency goal. If not, quickly gather * Check if this request met our latency goal. If not, quickly gather
* some statistics and start throttling. * some statistics and start throttling.
*/ */
sched_domain = rq_sched_domain(rq); sched_domain = kyber_sched_domain(rq->cmd_flags);
switch (sched_domain) { switch (sched_domain) {
case KYBER_READ: case KYBER_READ:
target = kqd->read_lat_nsec; target = kqd->read_lat_nsec;
@ -498,19 +595,38 @@ static void kyber_completed_request(struct request *rq)
blk_stat_activate_msecs(kqd->cb, 10); blk_stat_activate_msecs(kqd->cb, 10);
} }
static void kyber_flush_busy_ctxs(struct kyber_hctx_data *khd, struct flush_kcq_data {
struct blk_mq_hw_ctx *hctx) struct kyber_hctx_data *khd;
unsigned int sched_domain;
struct list_head *list;
};
static bool flush_busy_kcq(struct sbitmap *sb, unsigned int bitnr, void *data)
{ {
LIST_HEAD(rq_list); struct flush_kcq_data *flush_data = data;
struct request *rq, *next; struct kyber_ctx_queue *kcq = &flush_data->khd->kcqs[bitnr];
blk_mq_flush_busy_ctxs(hctx, &rq_list); spin_lock(&kcq->lock);
list_for_each_entry_safe(rq, next, &rq_list, queuelist) { list_splice_tail_init(&kcq->rq_list[flush_data->sched_domain],
unsigned int sched_domain; flush_data->list);
sbitmap_clear_bit(sb, bitnr);
spin_unlock(&kcq->lock);
sched_domain = rq_sched_domain(rq); return true;
list_move_tail(&rq->queuelist, &khd->rqs[sched_domain]); }
}
static void kyber_flush_busy_kcqs(struct kyber_hctx_data *khd,
unsigned int sched_domain,
struct list_head *list)
{
struct flush_kcq_data data = {
.khd = khd,
.sched_domain = sched_domain,
.list = list,
};
sbitmap_for_each_set(&khd->kcq_map[sched_domain],
flush_busy_kcq, &data);
} }
static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags, static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
@ -573,26 +689,23 @@ static int kyber_get_domain_token(struct kyber_queue_data *kqd,
static struct request * static struct request *
kyber_dispatch_cur_domain(struct kyber_queue_data *kqd, kyber_dispatch_cur_domain(struct kyber_queue_data *kqd,
struct kyber_hctx_data *khd, struct kyber_hctx_data *khd,
struct blk_mq_hw_ctx *hctx, struct blk_mq_hw_ctx *hctx)
bool *flushed)
{ {
struct list_head *rqs; struct list_head *rqs;
struct request *rq; struct request *rq;
int nr; int nr;
rqs = &khd->rqs[khd->cur_domain]; rqs = &khd->rqs[khd->cur_domain];
rq = list_first_entry_or_null(rqs, struct request, queuelist);
/* /*
* If there wasn't already a pending request and we haven't flushed the * If we already have a flushed request, then we just need to get a
* software queues yet, flush the software queues and check again. * token for it. Otherwise, if there are pending requests in the kcqs,
* flush the kcqs, but only if we can get a token. If not, we should
* leave the requests in the kcqs so that they can be merged. Note that
* khd->lock serializes the flushes, so if we observed any bit set in
* the kcq_map, we will always get a request.
*/ */
if (!rq && !*flushed) { rq = list_first_entry_or_null(rqs, struct request, queuelist);
kyber_flush_busy_ctxs(khd, hctx);
*flushed = true;
rq = list_first_entry_or_null(rqs, struct request, queuelist);
}
if (rq) { if (rq) {
nr = kyber_get_domain_token(kqd, khd, hctx); nr = kyber_get_domain_token(kqd, khd, hctx);
if (nr >= 0) { if (nr >= 0) {
@ -601,6 +714,16 @@ kyber_dispatch_cur_domain(struct kyber_queue_data *kqd,
list_del_init(&rq->queuelist); list_del_init(&rq->queuelist);
return rq; return rq;
} }
} else if (sbitmap_any_bit_set(&khd->kcq_map[khd->cur_domain])) {
nr = kyber_get_domain_token(kqd, khd, hctx);
if (nr >= 0) {
kyber_flush_busy_kcqs(khd, khd->cur_domain, rqs);
rq = list_first_entry(rqs, struct request, queuelist);
khd->batching++;
rq_set_domain_token(rq, nr);
list_del_init(&rq->queuelist);
return rq;
}
} }
/* There were either no pending requests or no tokens. */ /* There were either no pending requests or no tokens. */
@ -611,7 +734,6 @@ static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx)
{ {
struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data; struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
struct kyber_hctx_data *khd = hctx->sched_data; struct kyber_hctx_data *khd = hctx->sched_data;
bool flushed = false;
struct request *rq; struct request *rq;
int i; int i;
@ -622,7 +744,7 @@ static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx)
* from the batch. * from the batch.
*/ */
if (khd->batching < kyber_batch_size[khd->cur_domain]) { if (khd->batching < kyber_batch_size[khd->cur_domain]) {
rq = kyber_dispatch_cur_domain(kqd, khd, hctx, &flushed); rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
if (rq) if (rq)
goto out; goto out;
} }
@ -643,7 +765,7 @@ static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx)
else else
khd->cur_domain++; khd->cur_domain++;
rq = kyber_dispatch_cur_domain(kqd, khd, hctx, &flushed); rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
if (rq) if (rq)
goto out; goto out;
} }
@ -660,10 +782,12 @@ static bool kyber_has_work(struct blk_mq_hw_ctx *hctx)
int i; int i;
for (i = 0; i < KYBER_NUM_DOMAINS; i++) { for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
if (!list_empty_careful(&khd->rqs[i])) if (!list_empty_careful(&khd->rqs[i]) ||
sbitmap_any_bit_set(&khd->kcq_map[i]))
return true; return true;
} }
return sbitmap_any_bit_set(&hctx->ctx_map);
return false;
} }
#define KYBER_LAT_SHOW_STORE(op) \ #define KYBER_LAT_SHOW_STORE(op) \
@ -834,7 +958,9 @@ static struct elevator_type kyber_sched = {
.init_hctx = kyber_init_hctx, .init_hctx = kyber_init_hctx,
.exit_hctx = kyber_exit_hctx, .exit_hctx = kyber_exit_hctx,
.limit_depth = kyber_limit_depth, .limit_depth = kyber_limit_depth,
.bio_merge = kyber_bio_merge,
.prepare_request = kyber_prepare_request, .prepare_request = kyber_prepare_request,
.insert_requests = kyber_insert_requests,
.finish_request = kyber_finish_request, .finish_request = kyber_finish_request,
.requeue_request = kyber_finish_request, .requeue_request = kyber_finish_request,
.completed_request = kyber_completed_request, .completed_request = kyber_completed_request,