On Fri, 17 May 2013, Karthi wrote:
> Hi everyone,
>
> Our group works on a device mapper target called dmDedup that performs
> inline data deduplication. The dmDedup target internally uses the
> Persistent Metadata API (located in linux/drivers/md/persistent-data)to
> manage its meta-data.
>
> In our system, there is a main writer thread (or several of them) and a
> separate flusher thread. When a write request comes in, the writer thread
> checks for the duplicates in the data and updates corresponding meta-data.
> The flusher thread periodically (10 sec) flushes the meta-data to disk by
> calling the commit function in the transaction manager of Persistant
> Metadata API.
>
> When we benchmark dmDedup target with a sequential unique data write
> workload, we observed that write performance periodically drops to 0
> ops/sec for 8 to 15 seconds. It happens because the flusher thread blocks
> the writer thread. Our further analysis showed that, the writer thread is
> blocked on the dm_bufio mutex lock, which is acquired by the flusher thread
> on the dm_bufio_client when the dirty buffers are being flushed as a part
> of transaction. Though these dirty buffers are flushed asynchronously, the
> function enters a loop with the lock acquired and exits only when it places
> the request for all the dirty buffers. (drivers/md/dm-bufio.c -
> dm_bufio_write_dirty_buffers()). The writer thread is blocked as it needs
> this lock to lookup the metadata. The underlying I/O queue size is 128. The
> flusher thread waits to place IO request for all dirty buffers (which is
> much greater than 128) onto the queue. This way the writes becomes
> synchronous during this entire period, thus blocking the other writer
> thread.
>
> We want a metadata flush which is not affected by the writer thread so
> drastically. Specifically, we would like the flush to write out dirty
> buffers, while the writer still can update metadata (using COW to preserve
> consistency). Is it something that is possible with a current design of
> persistent metadata subsystem? Our current understanding is that during the
> commit all updates to the meta-data are blocked until all of the meta-data
> is written out.
>
> We appreciate any suggestions!
>
> Thank you!
>
Hi
Try this patch, it moves the code to submit buffers outside of the lock.
Tell us if it improves performance for you.
Mikulas
---
dm-bufio: drop lock when submitting writes
This patch changes dm-bufio so that it submits write I/Os outside of the
lock. If the number of submitted buffers is greater than the number of
requests on the target queue, submit_bio blocks. We want to block outside
of the lock to improve latency of other threads that may need the lock.
Signed-off-by: Mikulas Patocka <mpatocka@xxxxxxxxxx>
---
drivers/md/dm-bufio.c | 73 +++++++++++++++++++++++++++++++++++++++-----------
1 file changed, 58 insertions(+), 15 deletions(-)
Index: linux-3.9.3-fast/drivers/md/dm-bufio.c
===================================================================
--- linux-3.9.3-fast.orig/drivers/md/dm-bufio.c 2013-05-21 01:24:15.000000000 +0200
+++ linux-3.9.3-fast/drivers/md/dm-bufio.c 2013-05-21 02:03:26.000000000 +0200
@@ -145,6 +145,7 @@ struct dm_buffer {
unsigned long state;
unsigned long last_accessed;
struct dm_bufio_client *c;
+ struct list_head write_list;
struct bio bio;
struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
};
@@ -630,7 +631,8 @@ static int do_io_schedule(void *word)
* - Submit our write and don't wait on it. We set B_WRITING indicating
* that there is a write in progress.
*/
-static void __write_dirty_buffer(struct dm_buffer *b)
+static void __write_dirty_buffer(struct dm_buffer *b,
+ struct list_head *write_list)
{
if (!test_bit(B_DIRTY, &b->state))
return;
@@ -639,7 +641,24 @@ static void __write_dirty_buffer(struct
wait_on_bit_lock(&b->state, B_WRITING,
do_io_schedule, TASK_UNINTERRUPTIBLE);
- submit_io(b, WRITE, b->block, write_endio);
+ if (!write_list)
+ submit_io(b, WRITE, b->block, write_endio);
+ else
+ list_add_tail(&b->write_list, write_list);
+}
+
+static void __flush_write_list(struct list_head *write_list)
+{
+ struct blk_plug plug;
+ blk_start_plug(&plug);
+ while (!list_empty(write_list)) {
+ struct dm_buffer *b =
+ list_entry(write_list->next, struct dm_buffer, write_list);
+ list_del(&b->write_list);
+ submit_io(b, WRITE, b->block, write_endio);
+ dm_bufio_cond_resched();
+ }
+ blk_finish_plug(&plug);
}
/*
@@ -655,7 +674,7 @@ static void __make_buffer_clean(struct d
return;
wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
- __write_dirty_buffer(b);
+ __write_dirty_buffer(b, NULL);
wait_on_bit(&b->state, B_WRITING, do_io_schedule, TASK_UNINTERRUPTIBLE);
}
@@ -802,7 +821,8 @@ static void __free_buffer_wake(struct dm
wake_up(&c->free_buffer_wait);
}
-static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait)
+static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
+ struct list_head *write_list)
{
struct dm_buffer *b, *tmp;
@@ -818,7 +838,7 @@ static void __write_dirty_buffers_async(
if (no_wait && test_bit(B_WRITING, &b->state))
return;
- __write_dirty_buffer(b);
+ __write_dirty_buffer(b, write_list);
dm_bufio_cond_resched();
}
}
@@ -853,7 +873,8 @@ static void __get_memory_limit(struct dm
* If we are over threshold_buffers, start freeing buffers.
* If we're over "limit_buffers", block until we get under the limit.
*/
-static void __check_watermark(struct dm_bufio_client *c)
+static void __check_watermark(struct dm_bufio_client *c,
+ struct list_head *write_list)
{
unsigned long threshold_buffers, limit_buffers;
@@ -872,7 +893,7 @@ static void __check_watermark(struct dm_
}
if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
- __write_dirty_buffers_async(c, 1);
+ __write_dirty_buffers_async(c, 1, write_list);
}
/*
@@ -897,7 +918,8 @@ static struct dm_buffer *__find(struct d
*--------------------------------------------------------------*/
static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
- enum new_flag nf, int *need_submit)
+ enum new_flag nf, int *need_submit,
+ struct list_head *write_list)
{
struct dm_buffer *b, *new_b = NULL;
@@ -924,7 +946,7 @@ static struct dm_buffer *__bufio_new(str
goto found_buffer;
}
- __check_watermark(c);
+ __check_watermark(c, write_list);
b = new_b;
b->hold_count = 1;
@@ -992,10 +1014,14 @@ static void *new_read(struct dm_bufio_cl
int need_submit;
struct dm_buffer *b;
+ LIST_HEAD(write_list);
+
dm_bufio_lock(c);
- b = __bufio_new(c, block, nf, &need_submit);
+ b = __bufio_new(c, block, nf, &need_submit, &write_list);
dm_bufio_unlock(c);
+ __flush_write_list(&write_list);
+
if (!b)
return b;
@@ -1047,6 +1073,8 @@ void dm_bufio_prefetch(struct dm_bufio_c
{
struct blk_plug plug;
+ LIST_HEAD(write_list);
+
BUG_ON(dm_bufio_in_request());
blk_start_plug(&plug);
@@ -1055,7 +1083,15 @@ void dm_bufio_prefetch(struct dm_bufio_c
for (; n_blocks--; block++) {
int need_submit;
struct dm_buffer *b;
- b = __bufio_new(c, block, NF_PREFETCH, &need_submit);
+ b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
+ &write_list);
+ if (unlikely(!list_empty(&write_list))) {
+ dm_bufio_unlock(c);
+ blk_finish_plug(&plug);
+ __flush_write_list(&write_list);
+ blk_start_plug(&plug);
+ dm_bufio_lock(c);
+ }
if (unlikely(b != NULL)) {
dm_bufio_unlock(c);
@@ -1069,7 +1105,6 @@ void dm_bufio_prefetch(struct dm_bufio_c
goto flush_plug;
dm_bufio_lock(c);
}
-
}
dm_bufio_unlock(c);
@@ -1126,11 +1161,14 @@ EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_d
void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
{
+ LIST_HEAD(write_list);
+
BUG_ON(dm_bufio_in_request());
dm_bufio_lock(c);
- __write_dirty_buffers_async(c, 0);
+ __write_dirty_buffers_async(c, 0, &write_list);
dm_bufio_unlock(c);
+ __flush_write_list(&write_list);
}
EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
@@ -1147,8 +1185,13 @@ int dm_bufio_write_dirty_buffers(struct
unsigned long buffers_processed = 0;
struct dm_buffer *b, *tmp;
+ LIST_HEAD(write_list);
+
+ dm_bufio_lock(c);
+ __write_dirty_buffers_async(c, 0, &write_list);
+ dm_bufio_unlock(c);
+ __flush_write_list(&write_list);
dm_bufio_lock(c);
- __write_dirty_buffers_async(c, 0);
again:
list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
@@ -1274,7 +1317,7 @@ retry:
BUG_ON(!b->hold_count);
BUG_ON(test_bit(B_READING, &b->state));
- __write_dirty_buffer(b);
+ __write_dirty_buffer(b, NULL);
if (b->hold_count == 1) {
wait_on_bit(&b->state, B_WRITING,
do_io_schedule, TASK_UNINTERRUPTIBLE);
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
