diff mbox series

[RFC,1/3] block: add copy offload support

Message ID alpine.LRH.2.02.2202011331570.22481@file01.intranet.prod.int.rdu2.redhat.com
State New
Headers show
Series NVMe copy offload patches | expand

Commit Message

Mikulas Patocka Feb. 1, 2022, 6:32 p.m. UTC
Add generic copy offload support to the block layer.

We add two new bio types: REQ_OP_COPY_READ_TOKEN and
REQ_OP_COPY_WRITE_TOKEN. Their bio vector has one entry - a page
containing the token.

When we need to copy data, we send REQ_OP_COPY_READ_TOKEN to the source
device and then we send REQ_OP_COPY_WRITE_TOKEN to the destination device.

This patch introduces a new ioctl BLKCOPY that submits the copy operation.
BLKCOPY argument has four 64-bit numbers - source offset, destination
offset and length. The last number is returned by the ioctl and it is the
number of bytes that were actually copied.

For in-kernel users, we introduce a function blkdev_issue_copy.

Copying may fail anytime, the caller is required to fallback to explicit
copy.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>

---
 block/blk-core.c          |    7 +++
 block/blk-lib.c           |   89 ++++++++++++++++++++++++++++++++++++++++++++++
 block/blk-settings.c      |   12 ++++++
 block/blk-sysfs.c         |    7 +++
 block/blk.h               |    3 +
 block/ioctl.c             |   56 ++++++++++++++++++++++++++++
 include/linux/blk_types.h |    4 ++
 include/linux/blkdev.h    |   18 +++++++++
 include/uapi/linux/fs.h   |    1 
 9 files changed, 197 insertions(+)

Comments

Bart Van Assche Feb. 1, 2022, 7:18 p.m. UTC | #1
On 2/1/22 10:32, Mikulas Patocka wrote:
>   /**
> + * blk_queue_max_copy_sectors - set maximum copy offload sectors for the queue
> + * @q:  the request queue for the device
> + * @size:  the maximum copy offload sectors
> + */
> +void blk_queue_max_copy_sectors(struct request_queue *q, unsigned int size)
> +{
> +	q->limits.max_copy_sectors = size;
> +}
> +EXPORT_SYMBOL_GPL(blk_queue_max_copy_sectors);

Please either change the unit of 'size' into bytes or change its type 
into sector_t.

> +extern int blkdev_issue_copy(struct block_device *bdev1, sector_t sector1,
> +		      struct block_device *bdev2, sector_t sector2,
> +		      sector_t nr_sects, sector_t *copied, gfp_t gfp_mask);
> +

Only supporting copying between contiguous LBA ranges seems restrictive 
to me. I expect garbage collection by filesystems for UFS devices to 
perform better if multiple LBA ranges are submitted as a single SCSI 
XCOPY command.

A general comment about the approach: encoding the LBA range information 
in a bio payload is not compatible with bio splitting. How can the dm 
driver implement copy offloading without the ability to split copy 
offload bio's?

> +int blkdev_issue_copy(struct block_device *bdev1, sector_t sector1,
> +		      struct block_device *bdev2, sector_t sector2,
> +		      sector_t nr_sects, sector_t *copied, gfp_t gfp_mask)
> +{
> +	struct page *token;
> +	sector_t m;
> +	int r = 0;
> +	struct completion comp;

Consider using DECLARE_COMPLETION_ONSTACK() instead of a separate 
declaration and init_completion() call.

Thanks,

Bart.
Keith Busch Feb. 2, 2022, 4:21 p.m. UTC | #2
On Tue, Feb 01, 2022 at 01:32:29PM -0500, Mikulas Patocka wrote:
> +int blkdev_issue_copy(struct block_device *bdev1, sector_t sector1,
> +		      struct block_device *bdev2, sector_t sector2,
> +		      sector_t nr_sects, sector_t *copied, gfp_t gfp_mask)
> +{
> +	struct page *token;
> +	sector_t m;
> +	int r = 0;
> +	struct completion comp;
> +
> +	*copied = 0;
> +
> +	m = min(bdev_max_copy_sectors(bdev1), bdev_max_copy_sectors(bdev2));
> +	if (!m)
> +		return -EOPNOTSUPP;
> +	m = min(m, (sector_t)round_down(UINT_MAX, PAGE_SIZE) >> 9);
> +
> +	if (unlikely(bdev_read_only(bdev2)))
> +		return -EPERM;
> +
> +	token = alloc_page(gfp_mask);
> +	if (unlikely(!token))
> +		return -ENOMEM;
> +
> +	while (nr_sects) {
> +		struct bio *read_bio, *write_bio;
> +		sector_t this_step = min(nr_sects, m);
> +
> +		read_bio = bio_alloc(gfp_mask, 1);
> +		if (unlikely(!read_bio)) {
> +			r = -ENOMEM;
> +			break;
> +		}
> +		bio_set_op_attrs(read_bio, REQ_OP_COPY_READ_TOKEN, REQ_NOMERGE);
> +		bio_set_dev(read_bio, bdev1);
> +		__bio_add_page(read_bio, token, PAGE_SIZE, 0);

You have this "token" payload as driver specific data, but there's no
check that bdev1 and bdev2 subscribe to the same driver specific format.

I thought we discussed defining something like a "copy domain" that
establishes which block devices can offload copy operations to/from each
other, and that should be checked before proceeding with the copy
operation.
Mikulas Patocka Feb. 2, 2022, 4:40 p.m. UTC | #3
On Wed, 2 Feb 2022, Keith Busch wrote:

> On Tue, Feb 01, 2022 at 01:32:29PM -0500, Mikulas Patocka wrote:
> > +int blkdev_issue_copy(struct block_device *bdev1, sector_t sector1,
> > +		      struct block_device *bdev2, sector_t sector2,
> > +		      sector_t nr_sects, sector_t *copied, gfp_t gfp_mask)
> > +{
> > +	struct page *token;
> > +	sector_t m;
> > +	int r = 0;
> > +	struct completion comp;
> > +
> > +	*copied = 0;
> > +
> > +	m = min(bdev_max_copy_sectors(bdev1), bdev_max_copy_sectors(bdev2));
> > +	if (!m)
> > +		return -EOPNOTSUPP;
> > +	m = min(m, (sector_t)round_down(UINT_MAX, PAGE_SIZE) >> 9);
> > +
> > +	if (unlikely(bdev_read_only(bdev2)))
> > +		return -EPERM;
> > +
> > +	token = alloc_page(gfp_mask);
> > +	if (unlikely(!token))
> > +		return -ENOMEM;
> > +
> > +	while (nr_sects) {
> > +		struct bio *read_bio, *write_bio;
> > +		sector_t this_step = min(nr_sects, m);
> > +
> > +		read_bio = bio_alloc(gfp_mask, 1);
> > +		if (unlikely(!read_bio)) {
> > +			r = -ENOMEM;
> > +			break;
> > +		}
> > +		bio_set_op_attrs(read_bio, REQ_OP_COPY_READ_TOKEN, REQ_NOMERGE);
> > +		bio_set_dev(read_bio, bdev1);
> > +		__bio_add_page(read_bio, token, PAGE_SIZE, 0);
> 
> You have this "token" payload as driver specific data, but there's no
> check that bdev1 and bdev2 subscribe to the same driver specific format.
> 
> I thought we discussed defining something like a "copy domain" that
> establishes which block devices can offload copy operations to/from each
> other, and that should be checked before proceeding with the copy
> operation.

There is nvme_setup_read_token that fills in the token:
	memcpy(token->subsys, "nvme", 4);
	token->ns = ns;
	token->src_sector = bio->bi_iter.bi_sector;
	token->sectors = bio->bi_iter.bi_size >> 9;

There is nvme_setup_write_token that checks these values:
	if (unlikely(memcmp(token->subsys, "nvme", 4)))
		return BLK_STS_NOTSUPP;
	if (unlikely(token->ns != ns))
		return BLK_STS_NOTSUPP;

So, if we attempt to copy data between the nvme subsystem and the scsi 
subsystem, the "subsys" check will fail. If we attempt to copy data 
between different nvme namespaces, the "ns" check will fail.

If the nvme standard gets extended with cross-namespace copies, we can 
check in nvme_setup_write_token if we can copy between the source and 
destination namespace.

Mikulas
Knight, Frederick Feb. 2, 2022, 6:40 p.m. UTC | #4
Just FYI about copy domains.

SCSI already has copy domains defined (it is optional in the standard).  Third-party Copy Descriptor - Copy Group Identifier. Here's the SCSI text:

A Copy Group is a set of logical units that have a high probability of using high performance methods (e.g.,
copy on write snapshot) for third-party copy operations involving logical units that are in the same Copy
Group. Each logical unit may be a member of zero or one Copy Group. A logical unit indicates membership in
a Copy Group using the Copy Group Identifier third-party copy descriptor (see 7.7.18.13).

If a third-party copy operation involves logical units that are in different Copy Groups, then that third-party copy
operation has a high probability of using low performance methods (e.g., copy manager read operations from
the source CSCD or copy manager write operations to the destination CSCD).

NVMe today can only copy between LBAs on the SAME Namespace.  There is a new project just getting started to allow copy across namespaces in the same NVM subsystem (included as part of that project is to define how the copy domains will work).  So for NVMe, copy domains are still a work in progress.

Just FYI.

	Fred

-----Original Message-----
From: Keith Busch <kbusch@kernel.org> 
Sent: Wednesday, February 2, 2022 11:22 AM
To: Mikulas Patocka <mpatocka@redhat.com>
Cc: Javier González <javier@javigon.com>; Chaitanya Kulkarni <chaitanyak@nvidia.com>; linux-block@vger.kernel.org; linux-scsi@vger.kernel.org; dm-devel@redhat.com; linux-nvme@lists.infradead.org; linux-fsdevel <linux-fsdevel@vger.kernel.org>; Jens Axboe <axboe@kernel.dk>; msnitzer@redhat.com >> msnitzer@redhat.com <msnitzer@redhat.com>; Bart Van Assche <bvanassche@acm.org>; martin.petersen@oracle.com >> Martin K. Petersen <martin.petersen@oracle.com>; roland@purestorage.com; Hannes Reinecke <hare@suse.de>; Christoph Hellwig <hch@lst.de>; Knight, Frederick <Frederick.Knight@netapp.com>; zach.brown@ni.com; osandov@fb.com; lsf-pc@lists.linux-foundation.org; djwong@kernel.org; josef@toxicpanda.com; clm@fb.com; dsterba@suse.com; tytso@mit.edu; jack@suse.com; Kanchan Joshi <joshi.k@samsung.com>
Subject: Re: [RFC PATCH 1/3] block: add copy offload support

NetApp Security WARNING: This is an external email. Do not click links or open attachments unless you recognize the sender and know the content is safe.




On Tue, Feb 01, 2022 at 01:32:29PM -0500, Mikulas Patocka wrote:
> +int blkdev_issue_copy(struct block_device *bdev1, sector_t sector1,
> +                   struct block_device *bdev2, sector_t sector2,
> +                   sector_t nr_sects, sector_t *copied, gfp_t 
> +gfp_mask) {
> +     struct page *token;
> +     sector_t m;
> +     int r = 0;
> +     struct completion comp;
> +
> +     *copied = 0;
> +
> +     m = min(bdev_max_copy_sectors(bdev1), bdev_max_copy_sectors(bdev2));
> +     if (!m)
> +             return -EOPNOTSUPP;
> +     m = min(m, (sector_t)round_down(UINT_MAX, PAGE_SIZE) >> 9);
> +
> +     if (unlikely(bdev_read_only(bdev2)))
> +             return -EPERM;
> +
> +     token = alloc_page(gfp_mask);
> +     if (unlikely(!token))
> +             return -ENOMEM;
> +
> +     while (nr_sects) {
> +             struct bio *read_bio, *write_bio;
> +             sector_t this_step = min(nr_sects, m);
> +
> +             read_bio = bio_alloc(gfp_mask, 1);
> +             if (unlikely(!read_bio)) {
> +                     r = -ENOMEM;
> +                     break;
> +             }
> +             bio_set_op_attrs(read_bio, REQ_OP_COPY_READ_TOKEN, REQ_NOMERGE);
> +             bio_set_dev(read_bio, bdev1);
> +             __bio_add_page(read_bio, token, PAGE_SIZE, 0);

You have this "token" payload as driver specific data, but there's no check that bdev1 and bdev2 subscribe to the same driver specific format.

I thought we discussed defining something like a "copy domain" that establishes which block devices can offload copy operations to/from each other, and that should be checked before proceeding with the copy operation.
Mikulas Patocka Feb. 3, 2022, 6:50 p.m. UTC | #5
On Tue, 1 Feb 2022, Bart Van Assche wrote:

> On 2/1/22 10:32, Mikulas Patocka wrote:
> >   /**
> > + * blk_queue_max_copy_sectors - set maximum copy offload sectors for the
> > queue
> > + * @q:  the request queue for the device
> > + * @size:  the maximum copy offload sectors
> > + */
> > +void blk_queue_max_copy_sectors(struct request_queue *q, unsigned int size)
> > +{
> > +	q->limits.max_copy_sectors = size;
> > +}
> > +EXPORT_SYMBOL_GPL(blk_queue_max_copy_sectors);
> 
> Please either change the unit of 'size' into bytes or change its type into
> sector_t.

blk_queue_chunk_sectors, blk_queue_max_discard_sectors, 
blk_queue_max_write_same_sectors, blk_queue_max_write_zeroes_sectors, 
blk_queue_max_zone_append_sectors also have the unit of sectors and the 
argument is "unsigned int". Should blk_queue_max_copy_sectors be 
different?

> > +extern int blkdev_issue_copy(struct block_device *bdev1, sector_t sector1,
> > +		      struct block_device *bdev2, sector_t sector2,
> > +		      sector_t nr_sects, sector_t *copied, gfp_t gfp_mask);
> > +
> 
> Only supporting copying between contiguous LBA ranges seems restrictive to me.
> I expect garbage collection by filesystems for UFS devices to perform better
> if multiple LBA ranges are submitted as a single SCSI XCOPY command.

NVMe has a possibility to copy multiple source ranges into one destination 
range. But I think that leveraging this capability would just make the 
code excessively complex.

> A general comment about the approach: encoding the LBA range information in a
> bio payload is not compatible with bio splitting. How can the dm driver
> implement copy offloading without the ability to split copy offload bio's?

I don't expect the copy bios to be split.

One possibility is to just return -EOPNOTSUPP and fall back to explicit 
copy if the bio crosses dm target boundary (that's what my previous patch 
for SCSI XCOPY did).

Another possibility is to return the split boundary in the token and retry 
both bios will smaller length. But this approach would prevent us from 
submitting the REQ_OP_COPY_WRITE_TOKEN bios asynchronously.

I'm not sure which of these two approaches is better.

> > +int blkdev_issue_copy(struct block_device *bdev1, sector_t sector1,
> > +		      struct block_device *bdev2, sector_t sector2,
> > +		      sector_t nr_sects, sector_t *copied, gfp_t gfp_mask)
> > +{
> > +	struct page *token;
> > +	sector_t m;
> > +	int r = 0;
> > +	struct completion comp;
> 
> Consider using DECLARE_COMPLETION_ONSTACK() instead of a separate declaration
> and init_completion() call.

OK.

> Thanks,
> 
> Bart.

Mikulas
Keith Busch Feb. 3, 2022, 8:11 p.m. UTC | #6
On Thu, Feb 03, 2022 at 01:50:06PM -0500, Mikulas Patocka wrote:
> On Tue, 1 Feb 2022, Bart Van Assche wrote:
> > Only supporting copying between contiguous LBA ranges seems restrictive to me.
> > I expect garbage collection by filesystems for UFS devices to perform better
> > if multiple LBA ranges are submitted as a single SCSI XCOPY command.
> 
> NVMe has a possibility to copy multiple source ranges into one destination 
> range. But I think that leveraging this capability would just make the 
> code excessively complex.

The point is to defrag discontiguous blocks into a single range. The
capability loses a major value proposition without multiple sources.
Bart Van Assche Feb. 3, 2022, 10:49 p.m. UTC | #7
On 2/3/22 10:50, Mikulas Patocka wrote:
> On Tue, 1 Feb 2022, Bart Van Assche wrote:
>> On 2/1/22 10:32, Mikulas Patocka wrote:
>>>    /**
>>> + * blk_queue_max_copy_sectors - set maximum copy offload sectors for the
>>> queue
>>> + * @q:  the request queue for the device
>>> + * @size:  the maximum copy offload sectors
>>> + */
>>> +void blk_queue_max_copy_sectors(struct request_queue *q, unsigned int size)
>>> +{
>>> +	q->limits.max_copy_sectors = size;
>>> +}
>>> +EXPORT_SYMBOL_GPL(blk_queue_max_copy_sectors);
>>
>> Please either change the unit of 'size' into bytes or change its type into
>> sector_t.
> 
> blk_queue_chunk_sectors, blk_queue_max_discard_sectors,
> blk_queue_max_write_same_sectors, blk_queue_max_write_zeroes_sectors,
> blk_queue_max_zone_append_sectors also have the unit of sectors and the
> argument is "unsigned int". Should blk_queue_max_copy_sectors be
> different?

As far as I know using the type sector_t for variables that represent a 
number of sectors is a widely followed convention:

$ git grep -w sector_t | wc -l
2575

I would appreciate it if that convention would be used consistently, 
even if that means modifying existing code.

Thanks,

Bart.
Mikulas Patocka Feb. 4, 2022, 12:09 p.m. UTC | #8
On Thu, 3 Feb 2022, Bart Van Assche wrote:

> On 2/3/22 10:50, Mikulas Patocka wrote:
> > On Tue, 1 Feb 2022, Bart Van Assche wrote:
> > > On 2/1/22 10:32, Mikulas Patocka wrote:
> > > >    /**
> > > > + * blk_queue_max_copy_sectors - set maximum copy offload sectors for
> > > > the
> > > > queue
> > > > + * @q:  the request queue for the device
> > > > + * @size:  the maximum copy offload sectors
> > > > + */
> > > > +void blk_queue_max_copy_sectors(struct request_queue *q, unsigned int
> > > > size)
> > > > +{
> > > > +	q->limits.max_copy_sectors = size;
> > > > +}
> > > > +EXPORT_SYMBOL_GPL(blk_queue_max_copy_sectors);
> > > 
> > > Please either change the unit of 'size' into bytes or change its type into
> > > sector_t.
> > 
> > blk_queue_chunk_sectors, blk_queue_max_discard_sectors,
> > blk_queue_max_write_same_sectors, blk_queue_max_write_zeroes_sectors,
> > blk_queue_max_zone_append_sectors also have the unit of sectors and the
> > argument is "unsigned int". Should blk_queue_max_copy_sectors be
> > different?
> 
> As far as I know using the type sector_t for variables that represent a number
> of sectors is a widely followed convention:
> 
> $ git grep -w sector_t | wc -l
> 2575
> 
> I would appreciate it if that convention would be used consistently, even if
> that means modifying existing code.
> 
> Thanks,
> 
> Bart.

Changing the sector limit variables in struct queue_limits from unsigned 
int to sector_t would increase the size of the structure and its cache 
footprint.

And we can't send bios larger than 4GiB anyway because bi_size is 32-bit.

Jens, what do you think about it? Should the sectors limits be sector_t?

Mikulas
Jens Axboe Feb. 4, 2022, 1:34 p.m. UTC | #9
On 2/4/22 5:09 AM, Mikulas Patocka wrote:
> 
> 
> On Thu, 3 Feb 2022, Bart Van Assche wrote:
> 
>> On 2/3/22 10:50, Mikulas Patocka wrote:
>>> On Tue, 1 Feb 2022, Bart Van Assche wrote:
>>>> On 2/1/22 10:32, Mikulas Patocka wrote:
>>>>>    /**
>>>>> + * blk_queue_max_copy_sectors - set maximum copy offload sectors for
>>>>> the
>>>>> queue
>>>>> + * @q:  the request queue for the device
>>>>> + * @size:  the maximum copy offload sectors
>>>>> + */
>>>>> +void blk_queue_max_copy_sectors(struct request_queue *q, unsigned int
>>>>> size)
>>>>> +{
>>>>> +	q->limits.max_copy_sectors = size;
>>>>> +}
>>>>> +EXPORT_SYMBOL_GPL(blk_queue_max_copy_sectors);
>>>>
>>>> Please either change the unit of 'size' into bytes or change its type into
>>>> sector_t.
>>>
>>> blk_queue_chunk_sectors, blk_queue_max_discard_sectors,
>>> blk_queue_max_write_same_sectors, blk_queue_max_write_zeroes_sectors,
>>> blk_queue_max_zone_append_sectors also have the unit of sectors and the
>>> argument is "unsigned int". Should blk_queue_max_copy_sectors be
>>> different?
>>
>> As far as I know using the type sector_t for variables that represent a number
>> of sectors is a widely followed convention:
>>
>> $ git grep -w sector_t | wc -l
>> 2575
>>
>> I would appreciate it if that convention would be used consistently, even if
>> that means modifying existing code.
>>
>> Thanks,
>>
>> Bart.
> 
> Changing the sector limit variables in struct queue_limits from
> unsigned int to sector_t would increase the size of the structure and
> its cache footprint.
> 
> And we can't send bios larger than 4GiB anyway because bi_size is
> 32-bit.
> 
> Jens, what do you think about it? Should the sectors limits be
> sector_t?

Why make it larger than it needs to?
diff mbox series

Patch

Index: linux-2.6/block/blk-settings.c
===================================================================
--- linux-2.6.orig/block/blk-settings.c	2022-01-26 19:12:30.000000000 +0100
+++ linux-2.6/block/blk-settings.c	2022-01-27 20:43:27.000000000 +0100
@@ -57,6 +57,7 @@  void blk_set_default_limits(struct queue
 	lim->misaligned = 0;
 	lim->zoned = BLK_ZONED_NONE;
 	lim->zone_write_granularity = 0;
+	lim->max_copy_sectors = 0;
 }
 EXPORT_SYMBOL(blk_set_default_limits);
 
@@ -365,6 +366,17 @@  void blk_queue_zone_write_granularity(st
 EXPORT_SYMBOL_GPL(blk_queue_zone_write_granularity);
 
 /**
+ * blk_queue_max_copy_sectors - set maximum copy offload sectors for the queue
+ * @q:  the request queue for the device
+ * @size:  the maximum copy offload sectors
+ */
+void blk_queue_max_copy_sectors(struct request_queue *q, unsigned int size)
+{
+	q->limits.max_copy_sectors = size;
+}
+EXPORT_SYMBOL_GPL(blk_queue_max_copy_sectors);
+
+/**
  * blk_queue_alignment_offset - set physical block alignment offset
  * @q:	the request queue for the device
  * @offset: alignment offset in bytes
Index: linux-2.6/include/linux/blkdev.h
===================================================================
--- linux-2.6.orig/include/linux/blkdev.h	2022-01-26 19:12:30.000000000 +0100
+++ linux-2.6/include/linux/blkdev.h	2022-01-29 17:46:03.000000000 +0100
@@ -103,6 +103,7 @@  struct queue_limits {
 	unsigned int		discard_granularity;
 	unsigned int		discard_alignment;
 	unsigned int		zone_write_granularity;
+	unsigned int		max_copy_sectors;
 
 	unsigned short		max_segments;
 	unsigned short		max_integrity_segments;
@@ -706,6 +707,7 @@  extern void blk_queue_max_zone_append_se
 extern void blk_queue_physical_block_size(struct request_queue *, unsigned int);
 void blk_queue_zone_write_granularity(struct request_queue *q,
 				      unsigned int size);
+void blk_queue_max_copy_sectors(struct request_queue *q, unsigned int size);
 extern void blk_queue_alignment_offset(struct request_queue *q,
 				       unsigned int alignment);
 void disk_update_readahead(struct gendisk *disk);
@@ -862,6 +864,10 @@  extern int __blkdev_issue_zeroout(struct
 extern int blkdev_issue_zeroout(struct block_device *bdev, sector_t sector,
 		sector_t nr_sects, gfp_t gfp_mask, unsigned flags);
 
+extern int blkdev_issue_copy(struct block_device *bdev1, sector_t sector1,
+		      struct block_device *bdev2, sector_t sector2,
+		      sector_t nr_sects, sector_t *copied, gfp_t gfp_mask);
+
 static inline int sb_issue_discard(struct super_block *sb, sector_t block,
 		sector_t nr_blocks, gfp_t gfp_mask, unsigned long flags)
 {
@@ -1001,6 +1007,18 @@  bdev_zone_write_granularity(struct block
 	return queue_zone_write_granularity(bdev_get_queue(bdev));
 }
 
+static inline unsigned int
+queue_max_copy_sectors(const struct request_queue *q)
+{
+	return q->limits.max_copy_sectors;
+}
+
+static inline unsigned int
+bdev_max_copy_sectors(struct block_device *bdev)
+{
+	return queue_max_copy_sectors(bdev_get_queue(bdev));
+}
+
 static inline int queue_alignment_offset(const struct request_queue *q)
 {
 	if (q->limits.misaligned)
Index: linux-2.6/block/blk-sysfs.c
===================================================================
--- linux-2.6.orig/block/blk-sysfs.c	2022-01-26 19:12:30.000000000 +0100
+++ linux-2.6/block/blk-sysfs.c	2022-01-26 19:12:30.000000000 +0100
@@ -230,6 +230,11 @@  static ssize_t queue_zone_write_granular
 	return queue_var_show(queue_zone_write_granularity(q), page);
 }
 
+static ssize_t queue_max_copy_sectors_show(struct request_queue *q, char *page)
+{
+	return queue_var_show(queue_max_copy_sectors(q), page);
+}
+
 static ssize_t queue_zone_append_max_show(struct request_queue *q, char *page)
 {
 	unsigned long long max_sectors = q->limits.max_zone_append_sectors;
@@ -591,6 +596,7 @@  QUEUE_RO_ENTRY(queue_write_same_max, "wr
 QUEUE_RO_ENTRY(queue_write_zeroes_max, "write_zeroes_max_bytes");
 QUEUE_RO_ENTRY(queue_zone_append_max, "zone_append_max_bytes");
 QUEUE_RO_ENTRY(queue_zone_write_granularity, "zone_write_granularity");
+QUEUE_RO_ENTRY(queue_max_copy_sectors, "max_copy_sectors");
 
 QUEUE_RO_ENTRY(queue_zoned, "zoned");
 QUEUE_RO_ENTRY(queue_nr_zones, "nr_zones");
@@ -647,6 +653,7 @@  static struct attribute *queue_attrs[] =
 	&queue_write_zeroes_max_entry.attr,
 	&queue_zone_append_max_entry.attr,
 	&queue_zone_write_granularity_entry.attr,
+	&queue_max_copy_sectors_entry.attr,
 	&queue_nonrot_entry.attr,
 	&queue_zoned_entry.attr,
 	&queue_nr_zones_entry.attr,
Index: linux-2.6/include/linux/blk_types.h
===================================================================
--- linux-2.6.orig/include/linux/blk_types.h	2022-01-06 18:55:01.000000000 +0100
+++ linux-2.6/include/linux/blk_types.h	2022-01-29 17:47:44.000000000 +0100
@@ -371,6 +371,10 @@  enum req_opf {
 	/* reset all the zone present on the device */
 	REQ_OP_ZONE_RESET_ALL	= 17,
 
+	/* copy offload bios */
+	REQ_OP_COPY_READ_TOKEN	= 18,
+	REQ_OP_COPY_WRITE_TOKEN	= 19,
+
 	/* Driver private requests */
 	REQ_OP_DRV_IN		= 34,
 	REQ_OP_DRV_OUT		= 35,
Index: linux-2.6/block/blk-lib.c
===================================================================
--- linux-2.6.orig/block/blk-lib.c	2021-08-18 13:59:55.000000000 +0200
+++ linux-2.6/block/blk-lib.c	2022-01-30 17:33:04.000000000 +0100
@@ -440,3 +440,92 @@  retry:
 	return ret;
 }
 EXPORT_SYMBOL(blkdev_issue_zeroout);
+
+static void bio_wake_completion(struct bio *bio)
+{
+	struct completion *comp = bio->bi_private;
+	complete(comp);
+}
+
+int blkdev_issue_copy(struct block_device *bdev1, sector_t sector1,
+		      struct block_device *bdev2, sector_t sector2,
+		      sector_t nr_sects, sector_t *copied, gfp_t gfp_mask)
+{
+	struct page *token;
+	sector_t m;
+	int r = 0;
+	struct completion comp;
+
+	*copied = 0;
+
+	m = min(bdev_max_copy_sectors(bdev1), bdev_max_copy_sectors(bdev2));
+	if (!m)
+		return -EOPNOTSUPP;
+	m = min(m, (sector_t)round_down(UINT_MAX, PAGE_SIZE) >> 9);
+
+	if (unlikely(bdev_read_only(bdev2)))
+		return -EPERM;
+
+	token = alloc_page(gfp_mask);
+	if (unlikely(!token))
+		return -ENOMEM;
+
+	while (nr_sects) {
+		struct bio *read_bio, *write_bio;
+		sector_t this_step = min(nr_sects, m);
+
+		read_bio = bio_alloc(gfp_mask, 1);
+		if (unlikely(!read_bio)) {
+			r = -ENOMEM;
+			break;
+		}
+		bio_set_op_attrs(read_bio, REQ_OP_COPY_READ_TOKEN, REQ_NOMERGE);
+		bio_set_dev(read_bio, bdev1);
+		__bio_add_page(read_bio, token, PAGE_SIZE, 0);
+		read_bio->bi_iter.bi_sector = sector1;
+		read_bio->bi_iter.bi_size = this_step << 9;
+		read_bio->bi_private = &comp;
+		read_bio->bi_end_io = bio_wake_completion;
+		init_completion(&comp);
+		submit_bio(read_bio);
+		wait_for_completion(&comp);
+		if (unlikely(read_bio->bi_status != BLK_STS_OK)) {
+			r = blk_status_to_errno(read_bio->bi_status);
+			bio_put(read_bio);
+			break;
+		}
+		bio_put(read_bio);
+
+		write_bio = bio_alloc(gfp_mask, 1);
+		if (unlikely(!write_bio)) {
+			r = -ENOMEM;
+			break;
+		}
+		bio_set_op_attrs(write_bio, REQ_OP_COPY_WRITE_TOKEN, REQ_NOMERGE);
+		bio_set_dev(write_bio, bdev2);
+		__bio_add_page(write_bio, token, PAGE_SIZE, 0);
+		write_bio->bi_iter.bi_sector = sector2;
+		write_bio->bi_iter.bi_size = this_step << 9;
+		write_bio->bi_private = &comp;
+		write_bio->bi_end_io = bio_wake_completion;
+		reinit_completion(&comp);
+		submit_bio(write_bio);
+		wait_for_completion(&comp);
+		if (unlikely(write_bio->bi_status != BLK_STS_OK)) {
+			r = blk_status_to_errno(write_bio->bi_status);
+			bio_put(write_bio);
+			break;
+		}
+		bio_put(write_bio);
+
+		sector1 += this_step;
+		sector2 += this_step;
+		nr_sects -= this_step;
+		*copied += this_step;
+	}
+
+	__free_page(token);
+
+	return r;
+}
+EXPORT_SYMBOL(blkdev_issue_copy);
Index: linux-2.6/block/ioctl.c
===================================================================
--- linux-2.6.orig/block/ioctl.c	2022-01-24 15:10:40.000000000 +0100
+++ linux-2.6/block/ioctl.c	2022-01-30 13:43:35.000000000 +0100
@@ -165,6 +165,60 @@  fail:
 	return err;
 }
 
+static int blk_ioctl_copy(struct block_device *bdev, fmode_t mode,
+		unsigned long arg)
+{
+	uint64_t range[4];
+	uint64_t start1, start2, end1, end2, len;
+	sector_t copied = 0;
+	struct inode *inode = bdev->bd_inode;
+	int err;
+
+	if (!(mode & FMODE_WRITE)) {
+		err = -EBADF;
+		goto fail1;
+	}
+
+	if (copy_from_user(range, (void __user *)arg, 24)) {
+		err = -EFAULT;
+		goto fail1;
+	}
+
+	start1 = range[0];
+	start2 = range[1];
+	len = range[2];
+	end1 = start1 + len - 1;
+	end2 = start2 + len - 1;
+
+	if ((start1 | start2 | len) & 511)
+		return -EINVAL;
+	if (end1 >= (uint64_t)bdev_nr_bytes(bdev))
+		return -EINVAL;
+	if (end2 >= (uint64_t)bdev_nr_bytes(bdev))
+		return -EINVAL;
+	if (end1 < start1)
+		return -EINVAL;
+	if (end2 < start2)
+		return -EINVAL;
+
+	filemap_invalidate_lock(inode->i_mapping);
+	err = truncate_bdev_range(bdev, mode, start2, end2);
+	if (err)
+		goto fail2;
+
+	err = blkdev_issue_copy(bdev, start1 >> 9, bdev, start2 >> 9, len >> 9, &copied, GFP_KERNEL);
+
+fail2:
+	filemap_invalidate_unlock(inode->i_mapping);
+
+fail1:
+	range[3] = (uint64_t)copied << 9;
+	if (copy_to_user((void __user *)(arg + 24), &range[3], 8))
+		err = -EFAULT;
+
+	return err;
+}
+
 static int put_ushort(unsigned short __user *argp, unsigned short val)
 {
 	return put_user(val, argp);
@@ -459,6 +513,8 @@  static int blkdev_common_ioctl(struct bl
 		return blk_ioctl_zeroout(bdev, mode, arg);
 	case BLKGETDISKSEQ:
 		return put_u64(argp, bdev->bd_disk->diskseq);
+	case BLKCOPY:
+		return blk_ioctl_copy(bdev, mode, arg);
 	case BLKREPORTZONE:
 		return blkdev_report_zones_ioctl(bdev, mode, cmd, arg);
 	case BLKRESETZONE:
Index: linux-2.6/include/uapi/linux/fs.h
===================================================================
--- linux-2.6.orig/include/uapi/linux/fs.h	2021-09-23 17:07:02.000000000 +0200
+++ linux-2.6/include/uapi/linux/fs.h	2022-01-27 19:05:46.000000000 +0100
@@ -185,6 +185,7 @@  struct fsxattr {
 #define BLKROTATIONAL _IO(0x12,126)
 #define BLKZEROOUT _IO(0x12,127)
 #define BLKGETDISKSEQ _IOR(0x12,128,__u64)
+#define BLKCOPY _IO(0x12,129)
 /*
  * A jump here: 130-136 are reserved for zoned block devices
  * (see uapi/linux/blkzoned.h)
Index: linux-2.6/block/blk.h
===================================================================
--- linux-2.6.orig/block/blk.h	2022-01-24 15:10:40.000000000 +0100
+++ linux-2.6/block/blk.h	2022-01-29 18:10:28.000000000 +0100
@@ -288,6 +288,9 @@  static inline bool blk_may_split(struct
 	case REQ_OP_WRITE_ZEROES:
 	case REQ_OP_WRITE_SAME:
 		return true; /* non-trivial splitting decisions */
+	case REQ_OP_COPY_READ_TOKEN:
+	case REQ_OP_COPY_WRITE_TOKEN:
+		return false;
 	default:
 		break;
 	}
Index: linux-2.6/block/blk-core.c
===================================================================
--- linux-2.6.orig/block/blk-core.c	2022-01-24 15:10:40.000000000 +0100
+++ linux-2.6/block/blk-core.c	2022-02-01 15:53:39.000000000 +0100
@@ -124,6 +124,8 @@  static const char *const blk_op_name[] =
 	REQ_OP_NAME(ZONE_APPEND),
 	REQ_OP_NAME(WRITE_SAME),
 	REQ_OP_NAME(WRITE_ZEROES),
+	REQ_OP_NAME(COPY_READ_TOKEN),
+	REQ_OP_NAME(COPY_WRITE_TOKEN),
 	REQ_OP_NAME(DRV_IN),
 	REQ_OP_NAME(DRV_OUT),
 };
@@ -758,6 +760,11 @@  noinline_for_stack bool submit_bio_check
 		if (!q->limits.max_write_zeroes_sectors)
 			goto not_supported;
 		break;
+	case REQ_OP_COPY_READ_TOKEN:
+	case REQ_OP_COPY_WRITE_TOKEN:
+		if (!q->limits.max_copy_sectors)
+			goto not_supported;
+		break;
 	default:
 		break;
 	}