[RFC] Thin provisioning bits

[RFC] Thin provisioning bits

[Martin K. Petersen]

Just for discussion: This is what I have in my tree.  These two patches
correspond roughly to the ones Christoph posted earlier today.

I also support UNMAP and choose between UNMAP and WRITE SAME(16) depending
on values reported in the block limits VPD...

[PATCH 1/2] block: Expose discard granularity

[Martin K. Petersen]

While SSDs track block usage on a per-sector basis, RAID arrays often
have allocation blocks that are bigger.  Allow the discard granularity
to be set and expose it in sysfs.

Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
---
 block/blk-settings.c   |   16 +++++++++++++++-
 block/blk-sysfs.c      |   22 ++++++++++++++++++++++
 include/linux/blkdev.h |   13 +++++++++++++
 3 files changed, 50 insertions(+), 1 deletions(-)

diff --git a/block/blk-settings.c b/block/blk-settings.c
index 66d4aa8..60a0bec 100644
--- a/block/blk-settings.c
+++ b/block/blk-settings.c
@@ -96,7 +96,8 @@ void blk_set_default_limits(struct queue_limits *lim)
 	lim->max_segment_size = MAX_SEGMENT_SIZE;
 	lim->max_sectors = BLK_DEF_MAX_SECTORS;
 	lim->max_hw_sectors = INT_MAX;
-	lim->max_discard_sectors = SAFE_MAX_SECTORS;
+	lim->max_discard_sectors = 0;
+	lim->discard_granularity = 0;
 	lim->logical_block_size = lim->physical_block_size = lim->io_min = 512;
 	lim->bounce_pfn = (unsigned long)(BLK_BOUNCE_ANY >> PAGE_SHIFT);
 	lim->alignment_offset = 0;
@@ -252,6 +253,19 @@ void blk_queue_max_discard_sectors(struct request_queue *q,
 EXPORT_SYMBOL(blk_queue_max_discard_sectors);
 
 /**
+ * blk_queue_discard_granularity - set minimum size for discard requests
+ * @q:	the request queue for the device
+ * @gran:  the smallest unit the device can discard (in bytes)
+ *
+ * Description:
+ */
+void blk_queue_discard_granularity(struct request_queue *q, unsigned int gran)
+{
+	q->limits.discard_granularity = gran;
+}
+EXPORT_SYMBOL(blk_queue_discard_granularity);
+
+/**
  * blk_queue_max_phys_segments - set max phys segments for a request for this queue
  * @q:  the request queue for the device
  * @max_segments:  max number of segments
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 8a6d81a..3e9c947 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -126,6 +126,16 @@ static ssize_t queue_io_opt_show(struct request_queue *q, char *page)
 	return queue_var_show(queue_io_opt(q), page);
 }
 
+static ssize_t queue_discard_granularity_show(struct request_queue *q, char *page)
+{
+	return queue_var_show(queue_discard_granularity(q), page);
+}
+
+static ssize_t queue_discard_max_show(struct request_queue *q, char *page)
+{
+	return queue_var_show(queue_max_discard_sectors(q) << 9, page);
+}
+
 static ssize_t
 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
 {
@@ -293,6 +303,16 @@ static struct queue_sysfs_entry queue_io_opt_entry = {
 	.show = queue_io_opt_show,
 };
 
+static struct queue_sysfs_entry queue_discard_granularity_entry = {
+	.attr = {.name = "discard_granularity", .mode = S_IRUGO },
+	.show = queue_discard_granularity_show,
+};
+
+static struct queue_sysfs_entry queue_discard_max_entry = {
+	.attr = {.name = "discard_max_bytes", .mode = S_IRUGO },
+	.show = queue_discard_max_show,
+};
+
 static struct queue_sysfs_entry queue_nonrot_entry = {
 	.attr = {.name = "rotational", .mode = S_IRUGO | S_IWUSR },
 	.show = queue_nonrot_show,
@@ -328,6 +348,8 @@ static struct attribute *default_attrs[] = {
 	&queue_physical_block_size_entry.attr,
 	&queue_io_min_entry.attr,
 	&queue_io_opt_entry.attr,
+	&queue_discard_granularity_entry.attr,
+	&queue_discard_max_entry.attr,
 	&queue_nonrot_entry.attr,
 	&queue_nomerges_entry.attr,
 	&queue_rq_affinity_entry.attr,
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 221cecd..d5bb834 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -312,6 +312,7 @@ struct queue_limits {
 	unsigned int		io_min;
 	unsigned int		io_opt;
 	unsigned int		max_discard_sectors;
+	unsigned int		discard_granularity;
 
 	unsigned short		logical_block_size;
 	unsigned short		max_hw_segments;
@@ -939,6 +940,8 @@ extern void blk_limits_io_min(struct queue_limits *limits, unsigned int min);
 extern void blk_queue_io_min(struct request_queue *q, unsigned int min);
 extern void blk_limits_io_opt(struct queue_limits *limits, unsigned int opt);
 extern void blk_queue_io_opt(struct request_queue *q, unsigned int opt);
+extern void blk_queue_discard_granularity(struct request_queue *q,
+					  unsigned int gran);
 extern void blk_set_default_limits(struct queue_limits *lim);
 extern int blk_stack_limits(struct queue_limits *t, struct queue_limits *b,
 			    sector_t offset);
@@ -1134,6 +1137,16 @@ static inline int bdev_alignment_offset(struct block_device *bdev)
 	return q->limits.alignment_offset;
 }
 
+static inline unsigned int queue_discard_granularity(struct request_queue *q)
+{
+	return q->limits.discard_granularity;
+}
+
+static inline unsigned int queue_max_discard_sectors(struct request_queue *q)
+{
+	return q->limits.max_discard_sectors;
+}
+
 static inline int queue_dma_alignment(struct request_queue *q)
 {
 	return q ? q->dma_alignment : 511;
-- 
1.6.0.6

Re: [PATCH 1/2] block: Expose discard granularity

[Christoph Hellwig]

On Thu, Oct 29, 2009 at 11:30:08PM -0400, Martin K. Petersen wrote:
> + * blk_queue_discard_granularity - set minimum size for discard requests
> + * @q:	the request queue for the device
> + * @gran:  the smallest unit the device can discard (in bytes)
> + *
> + * Description:
> + */
> +void blk_queue_discard_granularity(struct request_queue *q, unsigned int gran)
> +{
> +	q->limits.discard_granularity = gran;
> +}
> +EXPORT_SYMBOL(blk_queue_discard_granularity);
> +

> +static inline unsigned int queue_discard_granularity(struct request_queue *q)
> +{
> +	return q->limits.discard_granularity;
> +}
> +
> +static inline unsigned int queue_max_discard_sectors(struct request_queue *q)
> +{
> +	return q->limits.max_discard_sectors;
> +}
> +


Btw, I really hate these wrapper, they make followign the code much
harder.  Already annoyed me for the I/O topology bits.

Re: [PATCH 1/2] block: Expose discard granularity

[Martin K. Petersen]

>>>>> "Christoph" == Christoph Hellwig <hch@infradead.org> writes:

>> +static inline unsigned int queue_discard_granularity(struct

Christoph> Btw, I really hate these wrapper, they make followign the
Christoph> code much harder.  Already annoyed me for the I/O topology
Christoph> bits.

They all follow the block layer calling conventions.  They are also
there to prevent the disaster that was the topology merge which exposed
code all over the kernel that was poking directly at fields in struct
request.

-- 
Martin K. Petersen	Oracle Linux Engineering

Re: [PATCH 1/2] block: Expose discard granularity

[Christoph Hellwig]

On Mon, Nov 02, 2009 at 08:31:08AM -0500, Martin K. Petersen wrote:
> They all follow the block layer calling conventions.  They are also
> there to prevent the disaster that was the topology merge which exposed
> code all over the kernel that was poking directly at fields in struct
> request.

It's pretty nasy historic conventions we don't follow anywhere else.
If Jens insist on keeping the obsfucation we of course have to, but
recently even the block layer moved away from all that wrapping.

Re: [PATCH 1/2] block: Expose discard granularity

[Jens Axboe]

On Tue, Nov 03 2009, Christoph Hellwig wrote:
> On Mon, Nov 02, 2009 at 08:31:08AM -0500, Martin K. Petersen wrote:
> > They all follow the block layer calling conventions.  They are also
> > there to prevent the disaster that was the topology merge which exposed
> > code all over the kernel that was poking directly at fields in struct
> > request.
> 
> It's pretty nasy historic conventions we don't follow anywhere else.
> If Jens insist on keeping the obsfucation we of course have to, but
> recently even the block layer moved away from all that wrapping.

I'm not too fond of it either, I have no problems removing the wrappers.
They are transitionally useful as Martin points out, but in the long run
we should simply get rid of them as they hinder code readability.

-- 
Jens Axboe

[PATCH 2/2] sd: WRITE SAME(16) / UNMAP support

[Martin K. Petersen]

Implement a prepare discard function that sends either WRITE SAME(16) or
UNMAP(10) depending on parameters indicated by the device in the block
limits VPD.

Extract unmap constraints and report them to the block layer.

Based in part by a patch by Christoph Hellwig <hch@lst.de>.

Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
---
 drivers/scsi/sd.c |   93 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 drivers/scsi/sd.h |    2 +
 2 files changed, 95 insertions(+), 0 deletions(-)

diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c
index 9093c72..72e0d78 100644
--- a/drivers/scsi/sd.c
+++ b/drivers/scsi/sd.c
@@ -264,6 +264,15 @@ sd_show_app_tag_own(struct device *dev, struct device_attribute *attr,
 	return snprintf(buf, 20, "%u\n", sdkp->ATO);
 }
 
+static ssize_t
+sd_show_thin_provisioning(struct device *dev, struct device_attribute *attr,
+			  char *buf)
+{
+	struct scsi_disk *sdkp = to_scsi_disk(dev);
+
+	return snprintf(buf, 20, "%u\n", sdkp->thin_provisioning);
+}
+
 static struct device_attribute sd_disk_attrs[] = {
 	__ATTR(cache_type, S_IRUGO|S_IWUSR, sd_show_cache_type,
 	       sd_store_cache_type),
@@ -274,6 +283,7 @@ static struct device_attribute sd_disk_attrs[] = {
 	       sd_store_manage_start_stop),
 	__ATTR(protection_type, S_IRUGO, sd_show_protection_type, NULL),
 	__ATTR(app_tag_own, S_IRUGO, sd_show_app_tag_own, NULL),
+	__ATTR(thin_provisioning, S_IRUGO, sd_show_thin_provisioning, NULL),
 	__ATTR_NULL,
 };
 
@@ -399,6 +409,57 @@ static void sd_prot_op(struct scsi_cmnd *scmd, unsigned int dif)
 }
 
 /**
+ * sd_prepare_discard - unmap blocks on thinly provisioned device
+ * @rq: Request to prepare
+ *
+ * Will issue either UNMAP or WRITE SAME(16) depending on preference
+ * indicated by target device.
+ **/
+static int sd_prepare_discard(struct request *rq)
+{
+	struct scsi_disk *sdkp = scsi_disk(rq->rq_disk);
+	struct bio *bio = rq->bio;
+	sector_t sector = bio->bi_sector;
+	unsigned int num = bio_sectors(bio);
+
+	if (sdkp->device->sector_size == 4096) {
+		sector >>= 3;
+		num >>= 3;
+	}
+
+	rq->cmd_type = REQ_TYPE_BLOCK_PC;
+	rq->timeout = SD_TIMEOUT;
+
+	memset(rq->cmd, 0, rq->cmd_len);
+
+	if (sdkp->unmap) {
+		char *buf = kmap_atomic(bio_page(bio), KM_USER0);
+
+		rq->cmd[0] = UNMAP;
+		rq->cmd[8] = 24;
+		rq->cmd_len = 10;
+
+		/* Ensure that data length matches payload */
+		rq->__data_len = bio->bi_size = bio->bi_io_vec->bv_len = 24;
+
+		put_unaligned_be16(6 + 16, &buf[0]);
+		put_unaligned_be16(16, &buf[2]);
+		put_unaligned_be64(sector, &buf[8]);
+		put_unaligned_be32(num, &buf[16]);
+
+		kunmap_atomic(buf, KM_USER0);
+	} else {
+		rq->cmd[0] = WRITE_SAME_16;
+		rq->cmd[1] = 0x8; /* UNMAP */
+		put_unaligned_be64(sector, &rq->cmd[2]);
+		put_unaligned_be32(num, &rq->cmd[10]);
+		rq->cmd_len = 16;
+	}
+
+	return BLKPREP_OK;
+}
+
+/**
  *	sd_init_command - build a scsi (read or write) command from
  *	information in the request structure.
  *	@SCpnt: pointer to mid-level's per scsi command structure that
@@ -418,6 +479,13 @@ static int sd_prep_fn(struct request_queue *q, struct request *rq)
 	int ret, host_dif;
 	unsigned char protect;
 
+	/*
+	 * Discard request come in as REQ_TYPE_FS but we turn them into
+	 * block PC requests to make life easier.
+	 */
+	if (blk_discard_rq(rq))
+		ret = sd_prepare_discard(rq);
+
 	if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
 		ret = scsi_setup_blk_pc_cmnd(sdp, rq);
 		goto out;
@@ -1432,6 +1500,9 @@ static int read_capacity_16(struct scsi_disk *sdkp, struct scsi_device *sdp,
 		sd_printk(KERN_NOTICE, sdkp,
 			  "physical block alignment offset: %u\n", alignment);
 
+	if (buffer[14] & 0x80)
+		sdkp->thin_provisioning = 1;
+
 	sdkp->capacity = lba + 1;
 	return sector_size;
 }
@@ -1863,6 +1934,7 @@ void sd_read_app_tag_own(struct scsi_disk *sdkp, unsigned char *buffer)
  */
 static void sd_read_block_limits(struct scsi_disk *sdkp)
 {
+	struct scsi_device *sdp = sdkp->device;
 	unsigned int sector_sz = sdkp->device->sector_size;
 	char *buffer;
 
@@ -1877,6 +1949,27 @@ static void sd_read_block_limits(struct scsi_disk *sdkp)
 	blk_queue_io_opt(sdkp->disk->queue,
 			 get_unaligned_be32(&buffer[12]) * sector_sz);
 
+	if (sdkp->thin_provisioning && buffer[3] == 0x3c) {
+		unsigned int lba_count, desc_count, granularity;
+		unsigned int max_sectors = UINT_MAX;
+
+		lba_count = get_unaligned_be32(&buffer[20]);
+		desc_count = get_unaligned_be32(&buffer[24]);
+		granularity = get_unaligned_be32(&buffer[28]) * sector_sz;
+
+		if (lba_count && desc_count) {
+			sdkp->unmap = 1;
+			max_sectors = lba_count * sector_sz >> 9;
+		}
+
+		if (granularity == 0)
+			granularity = sector_sz;
+
+		blk_queue_max_discard_sectors(sdp->request_queue, max_sectors);
+		blk_queue_discard_granularity(sdp->request_queue, granularity);
+		queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, sdkp->disk->queue);
+	}
+
 	kfree(buffer);
 }
 
diff --git a/drivers/scsi/sd.h b/drivers/scsi/sd.h
index e374804..43d3caf 100644
--- a/drivers/scsi/sd.h
+++ b/drivers/scsi/sd.h
@@ -60,6 +60,8 @@ struct scsi_disk {
 	unsigned	RCD : 1;	/* state of disk RCD bit, unused */
 	unsigned	DPOFUA : 1;	/* state of disk DPOFUA bit */
 	unsigned	first_scan : 1;
+	unsigned	thin_provisioning : 1;
+	unsigned	unmap : 1;
 };
 #define to_scsi_disk(obj) container_of(obj,struct scsi_disk,dev)
 
-- 
1.6.0.6

Re: [PATCH 2/2] sd: WRITE SAME(16) / UNMAP support

[Douglas Gilbert]

Martin K. Petersen wrote:
> Implement a prepare discard function that sends either WRITE SAME(16) or
> UNMAP(10) depending on parameters indicated by the device in the block
> limits VPD.

And if both are supported by the logical unit, the patch
prefers UNMAP?

Doug Gilbert

Re: [PATCH 2/2] sd: WRITE SAME(16) / UNMAP support

[Martin K. Petersen]

>>>>> "Doug" == Douglas Gilbert <dgilbert@interlog.com> writes:

Doug> And if both are supported by the logical unit, the patch prefers
Doug> UNMAP?

Yes.

SBC states that if the device reports MAXIMUM UNMAP LBA COUNT > 1 and
MAXIMUM UNMAP DESCRIPTOR COUNT > 1 then the device supports UNMAP.  And
in that case that's what I'll issue.  In all other cases I'll send out
WRITE SAME(16).  I believe that approach is what's currently considered
best practice.

-- 
Martin K. Petersen	Oracle Linux Engineering

Re: [PATCH 2/2] sd: WRITE SAME(16) / UNMAP support

[Ric Wheeler]

On 10/30/2009 12:53 AM, Martin K. Petersen wrote:
>>>>>> "Doug" == Douglas Gilbert<dgilbert@interlog.com>  writes:
>>>>>>              
> Doug>  And if both are supported by the logical unit, the patch prefers
> Doug>  UNMAP?
>
> Yes.
>
> SBC states that if the device reports MAXIMUM UNMAP LBA COUNT>  1 and
> MAXIMUM UNMAP DESCRIPTOR COUNT>  1 then the device supports UNMAP.  And
> in that case that's what I'll issue.  In all other cases I'll send out
> WRITE SAME(16).  I believe that approach is what's currently considered
> best practice.
>    

This sounds like the correct thing to do.

We should at the same time try to unify the file system mount options so 
we roll out the testing in a careful way.

Specifically, I would suggest that we default to "not" issuing discards 
by default and that we try to use the same mount option for any file 
system that supports barrier discards. My worry is that we will fry 
SSD's (like the reported issues with the Intel SSD's and Windows 7) or 
have horrific performance on arrays that are not tuned for fine grained 
discards :-)

Ric

Re: [PATCH 2/2] sd: WRITE SAME(16) / UNMAP support

[Christoph Hellwig]

On Thu, Oct 29, 2009 at 11:30:09PM -0400, Martin K. Petersen wrote:
> +sd_show_thin_provisioning(struct device *dev, struct device_attribute *attr,
> +			  char *buf)
> +{
> +	struct scsi_disk *sdkp = to_scsi_disk(dev);
> +
> +	return snprintf(buf, 20, "%u\n", sdkp->thin_provisioning);
> +}

If we want to expose this we'd better do it at the block layer so that
it also coverts non-scsi dicard support.  Besides the mtd one already in
tree I also plan to add a discard command to virtio-blk.

Re: [PATCH 2/2] sd: WRITE SAME(16) / UNMAP support

[Martin K. Petersen]

>>>>> "Christoph" == Christoph Hellwig <hch@infradead.org> writes:

Christoph> If we want to expose this we'd better do it at the block
Christoph> layer

I agree.  I've already had several requests for "is this blkdev thinly
provisioned?"...

-- 
Martin K. Petersen	Oracle Linux Engineering

Re: [PATCH 2/2] sd: WRITE SAME(16) / UNMAP support

[Christoph Hellwig]

On Mon, Nov 02, 2009 at 08:32:08AM -0500, Martin K. Petersen wrote:
> >>>>> "Christoph" == Christoph Hellwig <hch@infradead.org> writes:
> 
> Christoph> If we want to expose this we'd better do it at the block
> Christoph> layer
> 
> I agree.  I've already had several requests for "is this blkdev thinly
> provisioned?"...

Btw, I wonder if we should make a distinction between truely thing
provisioned and just discard capable devices like SSDs or other FTLs.

Thin provisioning update

[Martin K. Petersen]

Christoph,

Here are some more recent bits for you to tinker with.

The first patch exports unmap granularity and alignment up the stack.  I have
also implemented support for topology stacking of these parameters.  A reported
discard granularity of 0 means fully provisioned.

In the second patch I made a few changes to the TP detection and extraction of
VPD values.

I have been hunting for heuristics for the WRITE SAME case but haven't found any
good ones.  I'm thinking we probably want to distinguish between a real array
with NV-backed unmap queueing and a thinly provisioned disk on a virtualization
server where there might be a real impact from having to zero out partial blocks.
I guess that's easy to handle with virtio but I'm not so sure how to do it when
emulating a SCSI disk.

[PATCH 1/2] block: Expose discard granularity

[Martin K. Petersen]

While SSDs track block usage on a per-sector basis, RAID arrays often
have allocation blocks that are bigger.  Allow the discard granularity
and alignment to be set and teach the topology stacking logic how to
handle them.

Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
---
 block/blk-settings.c   |   46 ++++++++++++++++++++++++++++++++++++----------
 block/blk-sysfs.c      |   22 ++++++++++++++++++++++
 block/genhd.c          |   12 ++++++++++++
 fs/partitions/check.c  |   12 ++++++++++++
 include/linux/blkdev.h |   18 ++++++++++++++++++
 include/linux/genhd.h  |    1 +
 6 files changed, 101 insertions(+), 10 deletions(-)

diff --git a/block/blk-settings.c b/block/blk-settings.c
index 66d4aa8..7f986ca 100644
--- a/block/blk-settings.c
+++ b/block/blk-settings.c
@@ -96,7 +96,10 @@ void blk_set_default_limits(struct queue_limits *lim)
 	lim->max_segment_size = MAX_SEGMENT_SIZE;
 	lim->max_sectors = BLK_DEF_MAX_SECTORS;
 	lim->max_hw_sectors = INT_MAX;
-	lim->max_discard_sectors = SAFE_MAX_SECTORS;
+	lim->max_discard_sectors = 0;
+	lim->discard_granularity = 0;
+	lim->discard_alignment = 0;
+	lim->discard_misaligned = 0;
 	lim->logical_block_size = lim->physical_block_size = lim->io_min = 512;
 	lim->bounce_pfn = (unsigned long)(BLK_BOUNCE_ANY >> PAGE_SHIFT);
 	lim->alignment_offset = 0;
@@ -488,6 +491,16 @@ void blk_queue_stack_limits(struct request_queue *t, struct request_queue *b)
 }
 EXPORT_SYMBOL(blk_queue_stack_limits);
 
+static unsigned int lcm(unsigned int a, unsigned int b)
+{
+	if (a && b)
+		return (a * b) / gcd(a, b);
+	else if (b)
+		return b;
+
+	return a;
+}
+
 /**
  * blk_stack_limits - adjust queue_limits for stacked devices
  * @t:	the stacking driver limits (top)
@@ -502,6 +515,10 @@ EXPORT_SYMBOL(blk_queue_stack_limits);
 int blk_stack_limits(struct queue_limits *t, struct queue_limits *b,
 		     sector_t offset)
 {
+	int ret;
+
+	ret = 0;
+
 	t->max_sectors = min_not_zero(t->max_sectors, b->max_sectors);
 	t->max_hw_sectors = min_not_zero(t->max_hw_sectors, b->max_hw_sectors);
 	t->bounce_pfn = min_not_zero(t->bounce_pfn, b->bounce_pfn);
@@ -531,7 +548,13 @@ int blk_stack_limits(struct queue_limits *t, struct queue_limits *b,
 	if (offset &&
 	    (offset & (b->physical_block_size - 1)) != b->alignment_offset) {
 		t->misaligned = 1;
-		return -1;
+		ret = -1;
+	}
+
+	if (offset &&
+	    (offset & (b->discard_granularity - 1)) != b->discard_alignment) {
+		t->discard_misaligned = 1;
+		ret = -1;
 	}
 
 	/* If top has no alignment offset, inherit from bottom */
@@ -539,23 +562,26 @@ int blk_stack_limits(struct queue_limits *t, struct queue_limits *b,
 		t->alignment_offset =
 			b->alignment_offset & (b->physical_block_size - 1);
 
+	if (!t->discard_alignment)
+		t->discard_alignment =
+			b->discard_alignment & (b->discard_granularity - 1);
+
 	/* Top device aligned on logical block boundary? */
 	if (t->alignment_offset & (t->logical_block_size - 1)) {
 		t->misaligned = 1;
-		return -1;
+		ret = -1;
 	}
 
-	/* Find lcm() of optimal I/O size */
-	if (t->io_opt && b->io_opt)
-		t->io_opt = (t->io_opt * b->io_opt) / gcd(t->io_opt, b->io_opt);
-	else if (b->io_opt)
-		t->io_opt = b->io_opt;
+	/* Find lcm() of optimal I/O size and granularity */
+	t->io_opt = lcm(t->io_opt, b->io_opt);
+	t->discard_granularity = lcm(t->discard_granularity,
+				     b->discard_granularity);
 
 	/* Verify that optimal I/O size is a multiple of io_min */
 	if (t->io_min && t->io_opt % t->io_min)
-		return -1;
+		ret = -1;
 
-	return 0;
+	return ret;
 }
 EXPORT_SYMBOL(blk_stack_limits);
 
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 8a6d81a..3147145 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -126,6 +126,16 @@ static ssize_t queue_io_opt_show(struct request_queue *q, char *page)
 	return queue_var_show(queue_io_opt(q), page);
 }
 
+static ssize_t queue_discard_granularity_show(struct request_queue *q, char *page)
+{
+	return queue_var_show(q->limits.discard_granularity, page);
+}
+
+static ssize_t queue_discard_max_show(struct request_queue *q, char *page)
+{
+	return queue_var_show(q->limits.max_discard_sectors << 9, page);
+}
+
 static ssize_t
 queue_max_sectors_store(struct request_queue *q, const char *page, size_t count)
 {
@@ -293,6 +303,16 @@ static struct queue_sysfs_entry queue_io_opt_entry = {
 	.show = queue_io_opt_show,
 };
 
+static struct queue_sysfs_entry queue_discard_granularity_entry = {
+	.attr = {.name = "discard_granularity", .mode = S_IRUGO },
+	.show = queue_discard_granularity_show,
+};
+
+static struct queue_sysfs_entry queue_discard_max_entry = {
+	.attr = {.name = "discard_max_bytes", .mode = S_IRUGO },
+	.show = queue_discard_max_show,
+};
+
 static struct queue_sysfs_entry queue_nonrot_entry = {
 	.attr = {.name = "rotational", .mode = S_IRUGO | S_IWUSR },
 	.show = queue_nonrot_show,
@@ -328,6 +348,8 @@ static struct attribute *default_attrs[] = {
 	&queue_physical_block_size_entry.attr,
 	&queue_io_min_entry.attr,
 	&queue_io_opt_entry.attr,
+	&queue_discard_granularity_entry.attr,
+	&queue_discard_max_entry.attr,
 	&queue_nonrot_entry.attr,
 	&queue_nomerges_entry.attr,
 	&queue_rq_affinity_entry.attr,
diff --git a/block/genhd.c b/block/genhd.c
index 517e433..b11a4ad 100644
--- a/block/genhd.c
+++ b/block/genhd.c
@@ -861,12 +861,23 @@ static ssize_t disk_alignment_offset_show(struct device *dev,
 	return sprintf(buf, "%d\n", queue_alignment_offset(disk->queue));
 }
 
+static ssize_t disk_discard_alignment_show(struct device *dev,
+					   struct device_attribute *attr,
+					   char *buf)
+{
+	struct gendisk *disk = dev_to_disk(dev);
+
+	return sprintf(buf, "%u\n", queue_discard_alignment(disk->queue));
+}
+
 static DEVICE_ATTR(range, S_IRUGO, disk_range_show, NULL);
 static DEVICE_ATTR(ext_range, S_IRUGO, disk_ext_range_show, NULL);
 static DEVICE_ATTR(removable, S_IRUGO, disk_removable_show, NULL);
 static DEVICE_ATTR(ro, S_IRUGO, disk_ro_show, NULL);
 static DEVICE_ATTR(size, S_IRUGO, part_size_show, NULL);
 static DEVICE_ATTR(alignment_offset, S_IRUGO, disk_alignment_offset_show, NULL);
+static DEVICE_ATTR(discard_alignment, S_IRUGO, disk_discard_alignment_show,
+		   NULL);
 static DEVICE_ATTR(capability, S_IRUGO, disk_capability_show, NULL);
 static DEVICE_ATTR(stat, S_IRUGO, part_stat_show, NULL);
 static DEVICE_ATTR(inflight, S_IRUGO, part_inflight_show, NULL);
@@ -887,6 +898,7 @@ static struct attribute *disk_attrs[] = {
 	&dev_attr_ro.attr,
 	&dev_attr_size.attr,
 	&dev_attr_alignment_offset.attr,
+	&dev_attr_discard_alignment.attr,
 	&dev_attr_capability.attr,
 	&dev_attr_stat.attr,
 	&dev_attr_inflight.attr,
diff --git a/fs/partitions/check.c b/fs/partitions/check.c
index 7b685e1..64bc899 100644
--- a/fs/partitions/check.c
+++ b/fs/partitions/check.c
@@ -226,6 +226,13 @@ ssize_t part_alignment_offset_show(struct device *dev,
 	return sprintf(buf, "%llu\n", (unsigned long long)p->alignment_offset);
 }
 
+ssize_t part_discard_alignment_show(struct device *dev,
+				   struct device_attribute *attr, char *buf)
+{
+	struct hd_struct *p = dev_to_part(dev);
+	return sprintf(buf, "%u\n", p->discard_alignment);
+}
+
 ssize_t part_stat_show(struct device *dev,
 		       struct device_attribute *attr, char *buf)
 {
@@ -288,6 +295,8 @@ static DEVICE_ATTR(partition, S_IRUGO, part_partition_show, NULL);
 static DEVICE_ATTR(start, S_IRUGO, part_start_show, NULL);
 static DEVICE_ATTR(size, S_IRUGO, part_size_show, NULL);
 static DEVICE_ATTR(alignment_offset, S_IRUGO, part_alignment_offset_show, NULL);
+static DEVICE_ATTR(discard_alignment, S_IRUGO, part_discard_alignment_show,
+		   NULL);
 static DEVICE_ATTR(stat, S_IRUGO, part_stat_show, NULL);
 static DEVICE_ATTR(inflight, S_IRUGO, part_inflight_show, NULL);
 #ifdef CONFIG_FAIL_MAKE_REQUEST
@@ -300,6 +309,7 @@ static struct attribute *part_attrs[] = {
 	&dev_attr_start.attr,
 	&dev_attr_size.attr,
 	&dev_attr_alignment_offset.attr,
+	&dev_attr_discard_alignment.attr,
 	&dev_attr_stat.attr,
 	&dev_attr_inflight.attr,
 #ifdef CONFIG_FAIL_MAKE_REQUEST
@@ -403,6 +413,8 @@ struct hd_struct *add_partition(struct gendisk *disk, int partno,
 
 	p->start_sect = start;
 	p->alignment_offset = queue_sector_alignment_offset(disk->queue, start);
+	p->discard_alignment = queue_sector_discard_alignment(disk->queue,
+							      start);
 	p->nr_sects = len;
 	p->partno = partno;
 	p->policy = get_disk_ro(disk);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 221cecd..3b67221 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -312,12 +312,15 @@ struct queue_limits {
 	unsigned int		io_min;
 	unsigned int		io_opt;
 	unsigned int		max_discard_sectors;
+	unsigned int		discard_granularity;
+	unsigned int		discard_alignment;
 
 	unsigned short		logical_block_size;
 	unsigned short		max_hw_segments;
 	unsigned short		max_phys_segments;
 
 	unsigned char		misaligned;
+	unsigned char		discard_misaligned;
 	unsigned char		no_cluster;
 };
 
@@ -1134,6 +1137,21 @@ static inline int bdev_alignment_offset(struct block_device *bdev)
 	return q->limits.alignment_offset;
 }
 
+static inline int queue_discard_alignment(struct request_queue *q)
+{
+	if (q->limits.discard_misaligned)
+		return -1;
+
+	return q->limits.discard_alignment;
+}
+
+static inline int queue_sector_discard_alignment(struct request_queue *q,
+						 sector_t sector)
+{
+	return ((sector << 9) - q->limits.discard_alignment)
+		& (q->limits.discard_granularity - 1);
+}
+
 static inline int queue_dma_alignment(struct request_queue *q)
 {
 	return q ? q->dma_alignment : 511;
diff --git a/include/linux/genhd.h b/include/linux/genhd.h
index 297df45..c6c0c41 100644
--- a/include/linux/genhd.h
+++ b/include/linux/genhd.h
@@ -91,6 +91,7 @@ struct hd_struct {
 	sector_t start_sect;
 	sector_t nr_sects;
 	sector_t alignment_offset;
+	unsigned int discard_alignment;
 	struct device __dev;
 	struct kobject *holder_dir;
 	int policy, partno;
-- 
1.6.0.6