RFC: [PATCH] Add a sysfs entry queue/ignore_flushes.

RFC: [PATCH] Add a sysfs entry queue/ignore_flushes.

[Anthony DeRobertis]

I have an Oracle workload that ran quite fast under 2.6.32 due to mdraid
not supporting barriers. 2.6.33 slowed it down very substantially, and
its remains slow in 3.1. (Oracle is running on raw LVM logical volumes,
not filesystems, so ext4 barrier=0 doesn't help).

This patch adds a sysfs attribute to allow flush/fua to to turned off
(ignored) on a per-block-device basis. I have tested it on both a VM and
real hardware.

I'm not sure that queue_flags is the best place to put the bit, seems
like it may more logically go in flush_flags, but everything else is in
queue_flags.

I'm not subscribed, but will be looking for replies via list archives.
Feel free to CC me.

diff --git a/block/blk-core.c b/block/blk-core.c
index 90e1ffd..fd4f1c1 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -1515,9 +1515,11 @@ static inline void __generic_make_request(struct bio *bio)
 		/*
 		 * Filter flush bio's early so that make_request based
 		 * drivers without flush support don't have to worry
-		 * about them.
+		 * about them. Also filter if configured to ignore flushes.
 		 */
-		if ((bio->bi_rw & (REQ_FLUSH | REQ_FUA)) && !q->flush_flags) {
+		if ((bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
+			(!q->flush_flags ||
+			 test_bit(QUEUE_FLAG_IGNORE_FLUSH, &q->queue_flags))) {
 			bio->bi_rw &= ~(REQ_FLUSH | REQ_FUA);
 			if (!nr_sectors) {
 				err = 0;
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 0ee17b5..8f19827 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -214,6 +214,7 @@ queue_store_##name(struct request_queue *q, const char *page, size_t count) \
 
 QUEUE_SYSFS_BIT_FNS(nonrot, NONROT, 1);
 QUEUE_SYSFS_BIT_FNS(random, ADD_RANDOM, 0);
+QUEUE_SYSFS_BIT_FNS(ignore_flushes, IGNORE_FLUSH, 0);
 QUEUE_SYSFS_BIT_FNS(iostats, IO_STAT, 0);
 #undef QUEUE_SYSFS_BIT_FNS
 
@@ -385,6 +386,12 @@ static struct queue_sysfs_entry queue_random_entry = {
 	.store = queue_store_random,
 };
 
+static struct queue_sysfs_entry queue_ignore_flushes_entry = {
+	.attr = {.name = "ignore_flushes", .mode = S_IRUGO | S_IWUSR },
+	.show = queue_show_ignore_flushes,
+	.store = queue_store_ignore_flushes,
+};
+
 static struct attribute *default_attrs[] = {
 	&queue_requests_entry.attr,
 	&queue_ra_entry.attr,
@@ -407,6 +414,7 @@ static struct attribute *default_attrs[] = {
 	&queue_rq_affinity_entry.attr,
 	&queue_iostats_entry.attr,
 	&queue_random_entry.attr,
+	&queue_ignore_flushes_entry.attr,
 	NULL,
 };
 
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 84b15d5..61b39e0 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -408,6 +408,7 @@ struct request_queue {
 #define QUEUE_FLAG_ADD_RANDOM  16	/* Contributes to random pool */
 #define QUEUE_FLAG_SECDISCARD  17	/* supports SECDISCARD */
 #define QUEUE_FLAG_SAME_FORCE  18	/* force complete on same CPU */
+#define QUEUE_FLAG_IGNORE_FLUSH 19	/* ignore REQ_FLUSH and REQ_FUA */
 
 #define QUEUE_FLAG_DEFAULT	((1 << QUEUE_FLAG_IO_STAT) |		\
 				 (1 << QUEUE_FLAG_STACKABLE)	|	\

Re: RFC: [PATCH] Add a sysfs entry queue/ignore_flushes.

[Greg KH]

On Fri, Sep 09, 2011 at 04:34:05PM -0400, Anthony DeRobertis wrote:
> I have an Oracle workload that ran quite fast under 2.6.32 due to mdraid
> not supporting barriers. 2.6.33 slowed it down very substantially, and
> its remains slow in 3.1. (Oracle is running on raw LVM logical volumes,
> not filesystems, so ext4 barrier=0 doesn't help).
> 
> This patch adds a sysfs attribute to allow flush/fua to to turned off
> (ignored) on a per-block-device basis. I have tested it on both a VM and
> real hardware.
> 
> I'm not sure that queue_flags is the best place to put the bit, seems
> like it may more logically go in flush_flags, but everything else is in
> queue_flags.
> 
> I'm not subscribed, but will be looking for replies via list archives.
> Feel free to CC me.

If you add/remove/change sysfs attributes, you also need to do the same
to Documentation/ABI.

thanks,

greg k-h

Re: RFC: [PATCH] Add a sysfs entry queue/ignore_flushes.

[Anthony DeRobertis]

On 09/11/2011 01:08 AM, Greg KH wrote:
> If you add/remove/change sysfs attributes, you also need to do the same
> to Documentation/ABI.

I've added in documentation. Not sure who to set Contact: to, if that
should be the maintainer of the block layer or me, or...

Here is the updated patch, with documentation:

commit bd69b2eea10cfa217e91f82bac62b32800cb26bb
Author: Anthony DeRobertis <anthony@derobert.net>
Date:   Fri Sep 9 10:54:25 2011 -0400

    Add a sysfs entry queue/ignore_flushes.
    
    Set to ignore_flushes to 1 to have the block layer silently discard
    REQ_FLUSH and REQ_FUA. Great speed increase at the risk of great data
    loss.
    
    Defaults to 0, of course.
    
    Signed-off-by: Anthony DeRobertis <aderobertis@metrics.net>

diff --git a/Documentation/ABI/testing/sysfs-block b/Documentation/ABI/testing/sysfs-block
index c1eb41c..b0da935 100644
--- a/Documentation/ABI/testing/sysfs-block
+++ b/Documentation/ABI/testing/sysfs-block
@@ -206,3 +206,14 @@ Description:
 		when a discarded area is read the discard_zeroes_data
 		parameter will be set to one. Otherwise it will be 0 and
 		the result of reading a discarded area is undefined.
+
+What:		/sys/block/<disk>/queue/ignore_flushes
+Date:		September 2011
+Contact:	
+Description:
+		When this is set to 0 (the default), the device will
+		honor flush and force unit access (FUA) requests. When
+		set to 1, flush and FUA requests will be silently
+		ignored. Ignoring these requests potentially bosts
+		performance at the cost of data integrity if the machine
+		crashes or loses power.
diff --git a/block/blk-core.c b/block/blk-core.c
index 90e1ffd..fd4f1c1 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -1515,9 +1515,11 @@ static inline void __generic_make_request(struct bio *bio)
 		/*
 		 * Filter flush bio's early so that make_request based
 		 * drivers without flush support don't have to worry
-		 * about them.
+		 * about them. Also filter if configured to ignore flushes.
 		 */
-		if ((bio->bi_rw & (REQ_FLUSH | REQ_FUA)) && !q->flush_flags) {
+		if ((bio->bi_rw & (REQ_FLUSH | REQ_FUA)) &&
+			(!q->flush_flags ||
+			 test_bit(QUEUE_FLAG_IGNORE_FLUSH, &q->queue_flags))) {
 			bio->bi_rw &= ~(REQ_FLUSH | REQ_FUA);
 			if (!nr_sectors) {
 				err = 0;
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 0ee17b5..8f19827 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -214,6 +214,7 @@ queue_store_##name(struct request_queue *q, const char *page, size_t count) \
 
 QUEUE_SYSFS_BIT_FNS(nonrot, NONROT, 1);
 QUEUE_SYSFS_BIT_FNS(random, ADD_RANDOM, 0);
+QUEUE_SYSFS_BIT_FNS(ignore_flushes, IGNORE_FLUSH, 0);
 QUEUE_SYSFS_BIT_FNS(iostats, IO_STAT, 0);
 #undef QUEUE_SYSFS_BIT_FNS
 
@@ -385,6 +386,12 @@ static struct queue_sysfs_entry queue_random_entry = {
 	.store = queue_store_random,
 };
 
+static struct queue_sysfs_entry queue_ignore_flushes_entry = {
+	.attr = {.name = "ignore_flushes", .mode = S_IRUGO | S_IWUSR },
+	.show = queue_show_ignore_flushes,
+	.store = queue_store_ignore_flushes,
+};
+
 static struct attribute *default_attrs[] = {
 	&queue_requests_entry.attr,
 	&queue_ra_entry.attr,
@@ -407,6 +414,7 @@ static struct attribute *default_attrs[] = {
 	&queue_rq_affinity_entry.attr,
 	&queue_iostats_entry.attr,
 	&queue_random_entry.attr,
+	&queue_ignore_flushes_entry.attr,
 	NULL,
 };
 
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 84b15d5..61b39e0 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -408,6 +408,7 @@ struct request_queue {
 #define QUEUE_FLAG_ADD_RANDOM  16	/* Contributes to random pool */
 #define QUEUE_FLAG_SECDISCARD  17	/* supports SECDISCARD */
 #define QUEUE_FLAG_SAME_FORCE  18	/* force complete on same CPU */
+#define QUEUE_FLAG_IGNORE_FLUSH 19	/* ignore REQ_FLUSH and REQ_FUA */
 
 #define QUEUE_FLAG_DEFAULT	((1 << QUEUE_FLAG_IO_STAT) |		\
 				 (1 << QUEUE_FLAG_STACKABLE)	|	\