[PATCH v7 0/3] block device interposer

Hi all.

I'm joyful to suggest the block device interposer (bdev_interposer) v7.
bdev_interposer allows to redirect bio requests to other block devices.

In this series of patches I suggest a different implementation of the bio
interception mechanism. Now the interposer is a different block device.
Instead of an additional hook, the function fops->submit_bio() of
the interposer device is used.
This implementation greatly simplifies the application of this
bdev_interposer in device-mapper. But there is one limitation - the size
of the interposer device must be greater than or equal to the size of
the original device.

The first patch adds the function blk_mq_is_queue_frozen(). It allows to
check a queue state.

The second patch is dedicated to bdev_interposer itself, which provides
the ability to redirect bio to the interposer device.

The third one adds the DM_INTERPOSED_FLAG flag. When this flag is
applied with the ioctl DM_TABLE_LOAD_CMD, the underlying devices are
opened without the FMODE_EXCL flag and connected via bdev_interposer.

Changes in this patchset v7:
  * the request interception mechanism. Now the interposer is
    a block device that receives requests instead of the original device;
  * code design fixes.

History:
v6 - https://patchwork.kernel.org/project/linux-block/cover/1614774618-22410-1-git-send-email-sergei.shtepa@veeam.com/
  * designed for 5.12;
  * thanks to the new design of the bio structure in v5.12, it is
    possible to perform interception not for the entire disk, but
    for each block device;
  * instead of the new ioctl DM_DEV_REMAP_CMD and the 'noexcl' option,
    the DM_INTERPOSED_FLAG flag for the ioctl DM_TABLE_LOAD_CMD is
    applied.

v5 - https://patchwork.kernel.org/project/linux-block/cover/1612881028-7878-1-git-send-email-sergei.shtepa@veeam.com/
 * rebase for v5.11-rc7;
 * patch set organization;
 * fix defects in documentation;
 * add some comments;
 * change mutex names for better code readability;
 * remove calling bd_unlink_disk_holder() for targets with non-exclusive
   flag;
 * change type for struct dm_remap_param from uint8_t to __u8.

v4 - https://patchwork.kernel.org/project/linux-block/cover/1612367638-3794-1-git-send-email-sergei.shtepa@veeam.com/
Mostly changes were made, due to Damien's comments:
 * on the design of the code;
 * by the patch set organization;
 * bug with passing a wrong parameter to dm_get_device();
 * description of the 'noexcl' parameter in the linear.rst.
Also added remap_and_filter.rst.

v3 - https://patchwork.kernel.org/project/linux-block/cover/1611853955-32167-1-git-send-email-sergei.shtepa@veeam.com/
In this version, I already suggested blk_interposer to apply to dm-linear.
Problems were solved:
 * Interception of bio requests from a specific device on the disk, not
   from the entire disk. To do this, we added the dm_interposed_dev
   structure and an interval tree to store these structures.
 * Implemented ioctl DM_DEV_REMAP_CMD. A patch with changes in the lvm2
   project was sent to the team lvm-devel@redhat.com.
 * Added the 'noexcl' option for dm-linear, which allows you to open
   the underlying block-device without FMODE_EXCL mode.

v2 - https://patchwork.kernel.org/project/linux-block/cover/1607518911-30692-1-git-send-email-sergei.shtepa@veeam.com/
I tried to suggest blk_interposer without using it in device mapper,
but with the addition of a sample of its use. It was then that I learned
about the maintainers' attitudes towards the samples directory :).

v1 - https://lwn.net/ml/linux-block/20201119164924.74401-1-hare@suse.de/
This Hannes's patch can be considered as a starting point, since this is
where the interception mechanism and the term blk_interposer itself
appeared. It became clear that blk_interposer can be useful for
device mapper.

before v1 - https://patchwork.kernel.org/project/linux-block/cover/1603271049-20681-1-git-send-email-sergei.shtepa@veeam.com/
I tried to offer a rather cumbersome blk-filter and a monster-like
blk-snap module for creating snapshots.

Thank you to everyone who was able to take the time to review
the previous versions.
I hope that this time I achieved the required quality.

Thanks,
Sergei.

Sergei Shtepa (3):
  block: add blk_mq_is_queue_frozen()
  block: add bdev_interposer
  dm: add DM_INTERPOSED_FLAG

 block/bio.c                   |  2 ++
 block/blk-core.c              | 57 ++++++++++++++++++++++++++++++++
 block/blk-mq.c                | 13 ++++++++
 block/genhd.c                 | 54 +++++++++++++++++++++++++++++++
 drivers/md/dm-core.h          |  3 ++
 drivers/md/dm-ioctl.c         | 13 ++++++++
 drivers/md/dm-table.c         | 61 +++++++++++++++++++++++++++++------
 drivers/md/dm.c               | 38 +++++++++++++++-------
 include/linux/blk-mq.h        |  1 +
 include/linux/blk_types.h     |  3 ++
 include/linux/blkdev.h        |  9 ++++++
 include/linux/device-mapper.h |  1 +
 include/uapi/linux/dm-ioctl.h |  6 ++++
 13 files changed, 240 insertions(+), 21 deletions(-)

-- 
2.20.1

bdev_interposer allows to redirect bio requests to another devices.

Signed-off-by: Sergei Shtepa <redacted>
---
 block/bio.c               |  2 ++
 block/blk-core.c          | 57 +++++++++++++++++++++++++++++++++++++++
 block/genhd.c             | 54 +++++++++++++++++++++++++++++++++++++
 include/linux/blk_types.h |  3 +++
 include/linux/blkdev.h    |  9 +++++++
 5 files changed, 125 insertions(+)

diff --git a/block/bio.c b/block/bio.c
index a1c4d2900c7a..0bfbf06475ee 100644
--- a/block/bio.c
+++ b/block/bio.c

@@ -640,6 +640,8 @@ void __bio_clone_fast(struct bio *bio, struct bio *bio_src)
 		bio_set_flag(bio, BIO_THROTTLED);
 	if (bio_flagged(bio_src, BIO_REMAPPED))
 		bio_set_flag(bio, BIO_REMAPPED);
+	if (bio_flagged(bio_src, BIO_INTERPOSED))
+		bio_set_flag(bio, BIO_INTERPOSED);
 	bio->bi_opf = bio_src->bi_opf;
 	bio->bi_ioprio = bio_src->bi_ioprio;
 	bio->bi_write_hint = bio_src->bi_write_hint;

diff --git a/block/blk-core.c b/block/blk-core.c
index fc60ff208497..da1abc4c27a9 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c

@@ -1018,6 +1018,55 @@ static blk_qc_t __submit_bio_noacct_mq(struct bio *bio)
 	return ret;
 }
 
+static noinline blk_qc_t submit_bio_interposed(struct bio *bio)
+{
+	blk_qc_t ret = BLK_QC_T_NONE;
+	struct bio_list bio_list[2] = { };
+	struct gendisk *orig_disk;
+
+	if (current->bio_list) {
+		bio_list_add(&current->bio_list[0], bio);
+		return BLK_QC_T_NONE;
+	}
+
+	orig_disk = bio->bi_bdev->bd_disk;
+	if (unlikely(bio_queue_enter(bio)))
+		return BLK_QC_T_NONE;
+
+	current->bio_list = bio_list;
+
+	do {
+		struct block_device *interposer = bio->bi_bdev->bd_interposer;
+
+		if (unlikely(!interposer)) {
+			/* interposer was removed */
+			bio_list_add(&current->bio_list[0], bio);
+			break;
+		}
+		/* assign bio to interposer device */
+		bio_set_dev(bio, interposer);
+		bio_set_flag(bio, BIO_INTERPOSED);
+
+		if (!submit_bio_checks(bio))
+			break;
+		/*
+		 * Because the current->bio_list is initialized,
+		 * the submit_bio callback will always return BLK_QC_T_NONE.
+		 */
+		interposer->bd_disk->fops->submit_bio(bio);
+	} while (false);
+
+	current->bio_list = NULL;
+
+	blk_queue_exit(orig_disk->queue);
+
+	/* Resubmit remaining bios */
+	while ((bio = bio_list_pop(&bio_list[0])))
+		ret = submit_bio_noacct(bio);
+
+	return ret;
+}
+
 /**
  * submit_bio_noacct - re-submit a bio to the block device layer for I/O
  * @bio:  The bio describing the location in memory and on the device.

@@ -1029,6 +1078,14 @@ static blk_qc_t __submit_bio_noacct_mq(struct bio *bio)
  */
 blk_qc_t submit_bio_noacct(struct bio *bio)
 {
+	/*
+	 * Checking the BIO_INTERPOSED flag is necessary so that the bio
+	 * created by the bdev_interposer do not get to it for processing.
+	 */
+	if (bdev_has_interposer(bio->bi_bdev) &&
+	    !bio_flagged(bio, BIO_INTERPOSED))
+		return submit_bio_interposed(bio);
+
 	if (!submit_bio_checks(bio))
 		return BLK_QC_T_NONE;
 

diff --git a/block/genhd.c b/block/genhd.c
index c55e8f0fced1..c840ecffea68 100644
--- a/block/genhd.c
+++ b/block/genhd.c

@@ -30,6 +30,11 @@
 static struct kobject *block_depr;
 
 DECLARE_RWSEM(bdev_lookup_sem);
+/*
+ * Prevents different block-layer interposers from attaching or detaching
+ * to the block device at the same time.
+ */
+static DEFINE_MUTEX(bdev_interposer_attach_lock);
 
 /* for extended dynamic devt allocation, currently only one major is used */
 #define NR_EXT_DEVT		(1 << MINORBITS)

@@ -1940,3 +1945,52 @@ static void disk_release_events(struct gendisk *disk)
 	WARN_ON_ONCE(disk->ev && disk->ev->block != 1);
 	kfree(disk->ev);
 }
+
+int bdev_interposer_attach(struct block_device *original,
+			   struct block_device *interposer)
+{
+	int ret = 0;
+
+	if (WARN_ON(((!original) || (!interposer))))
+		return -EINVAL;
+	/*
+	 * interposer should be simple, no a multi-queue device
+	 */
+	if (!interposer->bd_disk->fops->submit_bio)
+		return -EINVAL;
+
+	if (WARN_ON(!blk_mq_is_queue_frozen(original->bd_disk->queue)))
+		return -EPERM;
+
+	mutex_lock(&bdev_interposer_attach_lock);
+
+	if (bdev_has_interposer(original))
+		ret = -EBUSY;
+	else {
+		original->bd_interposer = bdgrab(interposer);
+		if (!original->bd_interposer)
+			ret = -ENODEV;
+	}
+
+	mutex_unlock(&bdev_interposer_attach_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(bdev_interposer_attach);
+
+void bdev_interposer_detach(struct block_device *original)
+{
+	if (WARN_ON(!original))
+		return;
+
+	if (WARN_ON(!blk_mq_is_queue_frozen(original->bd_disk->queue)))
+		return;
+
+	mutex_lock(&bdev_interposer_attach_lock);
+	if (bdev_has_interposer(original)) {
+		bdput(original->bd_interposer);
+		original->bd_interposer = NULL;
+	}
+	mutex_unlock(&bdev_interposer_attach_lock);
+}
+EXPORT_SYMBOL_GPL(bdev_interposer_detach);

diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
index db026b6ec15a..13bda4732cf5 100644
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h

@@ -19,6 +19,7 @@ struct io_context;
 struct cgroup_subsys_state;
 typedef void (bio_end_io_t) (struct bio *);
 struct bio_crypt_ctx;
+struct bdev_interposer;
 
 struct block_device {
 	sector_t		bd_start_sect;

@@ -46,6 +47,7 @@ struct block_device {
 	spinlock_t		bd_size_lock; /* for bd_inode->i_size updates */
 	struct gendisk *	bd_disk;
 	struct backing_dev_info *bd_bdi;
+	struct block_device     *bd_interposer;
 
 	/* The counter of freeze processes */
 	int			bd_fsfreeze_count;

@@ -304,6 +306,7 @@ enum {
 	BIO_CGROUP_ACCT,	/* has been accounted to a cgroup */
 	BIO_TRACKED,		/* set if bio goes through the rq_qos path */
 	BIO_REMAPPED,
+	BIO_INTERPOSED,		/* bio was reassigned to another block device */
 	BIO_FLAG_LAST
 };
 

diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index bc6bc8383b43..90f62b4197da 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h

@@ -2031,4 +2031,13 @@ int fsync_bdev(struct block_device *bdev);
 int freeze_bdev(struct block_device *bdev);
 int thaw_bdev(struct block_device *bdev);
 
+static inline bool bdev_has_interposer(struct block_device *bdev)
+{
+	return (bdev->bd_interposer != NULL);
+};
+
+int bdev_interposer_attach(struct block_device *original,
+			   struct block_device *interposer);
+void bdev_interposer_detach(struct block_device *original);
+
 #endif /* _LINUX_BLKDEV_H */

-- 
2.20.1

blk_mq_is_queue_frozen() allow to assert that the queue is frozen.

Signed-off-by: Sergei Shtepa <redacted>
---
 block/blk-mq.c         | 13 +++++++++++++
 include/linux/blk-mq.h |  1 +
 2 files changed, 14 insertions(+)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index d4d7c1caa439..2f188a865024 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c

@@ -161,6 +161,19 @@ int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 }
 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 
+bool blk_mq_is_queue_frozen(struct request_queue *q)
+{
+	bool frozen;
+
+	mutex_lock(&q->mq_freeze_lock);
+	frozen = percpu_ref_is_dying(&q->q_usage_counter) &&
+		 percpu_ref_is_zero(&q->q_usage_counter);
+	mutex_unlock(&q->mq_freeze_lock);
+
+	return frozen;
+}
+EXPORT_SYMBOL_GPL(blk_mq_is_queue_frozen);
+
 /*
  * Guarantee no request is in use, so we can change any data structure of
  * the queue afterward.

diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 2c473c9b8990..6f01971abf7b 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h

@@ -533,6 +533,7 @@ void blk_freeze_queue_start(struct request_queue *q);
 void blk_mq_freeze_queue_wait(struct request_queue *q);
 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 				     unsigned long timeout);
+bool blk_mq_is_queue_frozen(struct request_queue *q);
 
 int blk_mq_map_queues(struct blk_mq_queue_map *qmap);
 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues);

-- 
2.20.1

DM_INTERPOSED_FLAG allow to create DM targets on "the fly".
Underlying block device opens without a flag FMODE_EXCL.
DM target receives bio from the original device via
bdev_interposer.

Signed-off-by: Sergei Shtepa <redacted>
---
 drivers/md/dm-core.h          |  3 ++
 drivers/md/dm-ioctl.c         | 13 ++++++++
 drivers/md/dm-table.c         | 61 +++++++++++++++++++++++++++++------
 drivers/md/dm.c               | 38 +++++++++++++++-------
 include/linux/device-mapper.h |  1 +
 include/uapi/linux/dm-ioctl.h |  6 ++++
 6 files changed, 101 insertions(+), 21 deletions(-)

diff --git a/drivers/md/dm-core.h b/drivers/md/dm-core.h
index 5953ff2bd260..9eae419c7b18 100644
--- a/drivers/md/dm-core.h
+++ b/drivers/md/dm-core.h

@@ -114,6 +114,9 @@ struct mapped_device {
 	bool init_tio_pdu:1;
 
 	struct srcu_struct io_barrier;
+
+	/* attach target via block-layer interposer */
+	bool is_interposed;
 };
 
 void disable_discard(struct mapped_device *md);

diff --git a/drivers/md/dm-ioctl.c b/drivers/md/dm-ioctl.c
index 5e306bba4375..2b4c9557c283 100644
--- a/drivers/md/dm-ioctl.c
+++ b/drivers/md/dm-ioctl.c

@@ -1267,6 +1267,11 @@ static inline fmode_t get_mode(struct dm_ioctl *param)
 	return mode;
 }
 
+static inline bool get_interposer_flag(struct dm_ioctl *param)
+{
+	return (param->flags & DM_INTERPOSED_FLAG);
+}
+
 static int next_target(struct dm_target_spec *last, uint32_t next, void *end,
 		       struct dm_target_spec **spec, char **target_params)
 {

@@ -1293,6 +1298,10 @@ static int populate_table(struct dm_table *table,
 		DMWARN("populate_table: no targets specified");
 		return -EINVAL;
 	}
+	if (table->md->is_interposed && (param->target_count != 1)) {
+		DMWARN("%s: with interposer should be specified only one target", __func__);
+		return -EINVAL;
+	}
 
 	for (i = 0; i < param->target_count; i++) {
 

@@ -1338,6 +1347,8 @@ static int table_load(struct file *filp, struct dm_ioctl *param, size_t param_si
 	if (!md)
 		return -ENXIO;
 
+	md->is_interposed = get_interposer_flag(param);
+
 	r = dm_table_create(&t, get_mode(param), param->target_count, md);
 	if (r)
 		goto err;

@@ -2098,6 +2109,8 @@ int __init dm_early_create(struct dm_ioctl *dmi,
 	if (r)
 		goto err_hash_remove;
 
+	md->is_interposed = get_interposer_flag(dmi);
+
 	/* add targets */
 	for (i = 0; i < dmi->target_count; i++) {
 		r = dm_table_add_target(t, spec_array[i]->target_type,

diff --git a/drivers/md/dm-table.c b/drivers/md/dm-table.c
index 95391f78b8d5..f6e2eb3f8949 100644
--- a/drivers/md/dm-table.c
+++ b/drivers/md/dm-table.c

@@ -225,12 +225,13 @@ void dm_table_destroy(struct dm_table *t)
 /*
  * See if we've already got a device in the list.
  */
-static struct dm_dev_internal *find_device(struct list_head *l, dev_t dev)
+static struct dm_dev_internal *find_device(struct list_head *l, dev_t dev, bool is_interposed)
 {
 	struct dm_dev_internal *dd;
 
 	list_for_each_entry (dd, l, list)
-		if (dd->dm_dev->bdev->bd_dev == dev)
+		if ((dd->dm_dev->bdev->bd_dev == dev) &&
+		    (dd->dm_dev->is_interposed == is_interposed))
 			return dd;
 
 	return NULL;

@@ -358,6 +359,18 @@ dev_t dm_get_dev_t(const char *path)
 }
 EXPORT_SYMBOL_GPL(dm_get_dev_t);
 
+static inline void dm_disk_freeze(struct gendisk *disk)
+{
+	blk_mq_freeze_queue(disk->queue);
+	blk_mq_quiesce_queue(disk->queue);
+}
+
+static inline void dm_disk_unfreeze(struct gendisk *disk)
+{
+	blk_mq_unquiesce_queue(disk->queue);
+	blk_mq_unfreeze_queue(disk->queue);
+}
+
 /*
  * Add a device to the list, or just increment the usage count if
  * it's already present.

@@ -385,7 +398,7 @@ int dm_get_device(struct dm_target *ti, const char *path, fmode_t mode,
 			return -ENODEV;
 	}
 
-	dd = find_device(&t->devices, dev);
+	dd = find_device(&t->devices, dev, t->md->is_interposed);
 	if (!dd) {
 		dd = kmalloc(sizeof(*dd), GFP_KERNEL);
 		if (!dd)

@@ -398,15 +411,38 @@ int dm_get_device(struct dm_target *ti, const char *path, fmode_t mode,
 
 		refcount_set(&dd->count, 1);
 		list_add(&dd->list, &t->devices);
-		goto out;
-
 	} else if (dd->dm_dev->mode != (mode | dd->dm_dev->mode)) {
 		r = upgrade_mode(dd, mode, t->md);
 		if (r)
 			return r;
+		refcount_inc(&dd->count);
 	}
-	refcount_inc(&dd->count);
-out:
+
+	if (t->md->is_interposed) {
+		struct block_device *original = dd->dm_dev->bdev;
+		struct block_device *interposer = t->md->disk->part0;
+
+		if ((ti->begin != 0) || (ti->len < bdev_nr_sectors(original))) {
+			dm_put_device(ti, dd->dm_dev);
+			DMERR("The interposer device should not be less than the original.");
+			return -EINVAL;
+		}
+
+		/*
+		 * Attach mapped interposer device to original.
+		 * It is quite convenient that device mapper creates
+		 * one disk for one block device.
+		 */
+		dm_disk_freeze(original->bd_disk);
+		r = bdev_interposer_attach(original, interposer);
+		dm_disk_unfreeze(original->bd_disk);
+		if (r) {
+			dm_put_device(ti, dd->dm_dev);
+			DMERR("Failed to attach dm interposer.");
+			return r;
+		}
+	}
+
 	*result = dd->dm_dev;
 	return 0;
 }

@@ -446,6 +482,7 @@ void dm_put_device(struct dm_target *ti, struct dm_dev *d)
 {
 	int found = 0;
 	struct list_head *devices = &ti->table->devices;
+	struct mapped_device *md = ti->table->md;
 	struct dm_dev_internal *dd;
 
 	list_for_each_entry(dd, devices, list) {

@@ -456,11 +493,17 @@ void dm_put_device(struct dm_target *ti, struct dm_dev *d)
 	}
 	if (!found) {
 		DMWARN("%s: device %s not in table devices list",
-		       dm_device_name(ti->table->md), d->name);
+		       dm_device_name(md), d->name);
 		return;
 	}
+	if (md->is_interposed) {
+		dm_disk_freeze(d->bdev->bd_disk);
+		bdev_interposer_detach(d->bdev);
+		dm_disk_unfreeze(d->bdev->bd_disk);
+	}
+
 	if (refcount_dec_and_test(&dd->count)) {
-		dm_put_table_device(ti->table->md, d);
+		dm_put_table_device(md, d);
 		list_del(&dd->list);
 		kfree(dd);
 	}

diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index 50b693d776d6..466bf70a66b0 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c

@@ -762,16 +762,24 @@ static int open_table_device(struct table_device *td, dev_t dev,
 
 	BUG_ON(td->dm_dev.bdev);
 
-	bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _dm_claim_ptr);
-	if (IS_ERR(bdev))
-		return PTR_ERR(bdev);
+	if (md->is_interposed) {
 
-	r = bd_link_disk_holder(bdev, dm_disk(md));
-	if (r) {
-		blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
-		return r;
+		bdev = blkdev_get_by_dev(dev, td->dm_dev.mode, NULL);
+		if (IS_ERR(bdev))
+			return PTR_ERR(bdev);
+	} else {
+		bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _dm_claim_ptr);
+		if (IS_ERR(bdev))
+			return PTR_ERR(bdev);
+
+		r = bd_link_disk_holder(bdev, dm_disk(md));
+		if (r) {
+			blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
+			return r;
+		}
 	}
 
+	td->dm_dev.is_interposed = md->is_interposed;
 	td->dm_dev.bdev = bdev;
 	td->dm_dev.dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
 	return 0;

@@ -785,20 +793,26 @@ static void close_table_device(struct table_device *td, struct mapped_device *md
 	if (!td->dm_dev.bdev)
 		return;
 
-	bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
-	blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
+	if (td->dm_dev.is_interposed)
+		blkdev_put(td->dm_dev.bdev, td->dm_dev.mode);
+	else {
+		bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
+		blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
+	}
 	put_dax(td->dm_dev.dax_dev);
 	td->dm_dev.bdev = NULL;
 	td->dm_dev.dax_dev = NULL;
 }
 
 static struct table_device *find_table_device(struct list_head *l, dev_t dev,
-					      fmode_t mode)
+					      fmode_t mode, bool is_interposed)
 {
 	struct table_device *td;
 
 	list_for_each_entry(td, l, list)
-		if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
+		if (td->dm_dev.bdev->bd_dev == dev &&
+		    td->dm_dev.mode == mode &&
+		    td->dm_dev.is_interposed == is_interposed)
 			return td;
 
 	return NULL;

@@ -811,7 +825,7 @@ int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
 	struct table_device *td;
 
 	mutex_lock(&md->table_devices_lock);
-	td = find_table_device(&md->table_devices, dev, mode);
+	td = find_table_device(&md->table_devices, dev, mode, md->is_interposed);
 	if (!td) {
 		td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
 		if (!td) {

diff --git a/include/linux/device-mapper.h b/include/linux/device-mapper.h
index 7f4ac87c0b32..76a6dfb1cb29 100644
--- a/include/linux/device-mapper.h
+++ b/include/linux/device-mapper.h

@@ -159,6 +159,7 @@ struct dm_dev {
 	struct block_device *bdev;
 	struct dax_device *dax_dev;
 	fmode_t mode;
+	bool is_interposed;
 	char name[16];
 };
 

diff --git a/include/uapi/linux/dm-ioctl.h b/include/uapi/linux/dm-ioctl.h
index fcff6669137b..fc4d06bb3dbb 100644
--- a/include/uapi/linux/dm-ioctl.h
+++ b/include/uapi/linux/dm-ioctl.h

@@ -362,4 +362,10 @@ enum {
  */
 #define DM_INTERNAL_SUSPEND_FLAG	(1 << 18) /* Out */
 
+/*
+ * If set, the underlying device should open without FMODE_EXCL
+ * and attach mapped device via bdev_interposer.
+ */
+#define DM_INTERPOSED_FLAG		(1 << 19) /* In */
+
 #endif				/* _LINUX_DM_IOCTL_H */

-- 
2.20.1

On Fri, Mar 12 2021 at 10:44am -0500,
Sergei Shtepa [off-list ref] wrote:

This flag is a mix of uses.  First it is used to store that
DM_INTERPOSED_FLAG was provided as input param during load.

And the same 'is_interposed' name is used in 'struct dm_dev'.

To me this state should be elevated to block core -- awkward for every
driver that might want to use blk-interposer to be sprinkling state
around its core structures.

So I'd prefer you:
1) rename 'struct mapped_device' to 'interpose' _and_ add it just after
   "bool init_tio_pdu:1;" with "bool interpose:1;" -- this reflects
   interpose was requested during load.
2) bdev_interposer_attach() should be made to set some block core state
   that is able to be tested to check if a device is_interposed.
3) Don't add an 'is_interposed' flag to 'struct dm_dev'

As I mention at the end: rename to DM_INTERPOSE_FLAG

This error/warning reads very awkwardly. Maybe?:
"%s: interposer requires only a single target be specified"

Think in make more sense to internalize the need to consider
md->interpose here.

So:

static struct dm_dev_internal *find_device(struct dm_table *t, dev_t dev)
{
        struct list_head *l = &t->devices;
        bool is_interposed = t->md->interpose;
        ...

 {
 	struct dm_dev_internal *dd;
 
 	list_for_each_entry (dd, l, list)
-		if (dd->dm_dev->bdev->bd_dev == dev)
+		if ((dd->dm_dev->bdev->bd_dev == dev) &&
+		    (dd->dm_dev->is_interposed == is_interposed))
 			return dd;

But why must this extra state be used/tested?  Seems like quite a deep
embedding of interposer into dm_dev_internal.. feels unnecessary.

These interfaces don't account for bio-based at all (pretty sure we've
been over this and you pointed out that they'll just return early), but
they also don't take steps to properly flush outstanding DM io.
Shouldn't you require DM devices do an internal suspend/resume?  And if
original device isn't DM then fallback to blk_mq calls?

This looks bogus... you cannot only increment refcount with the mode
check/upgrade branch (IIRC: I've made this same mistake in the past)

-out:
+
+	if (t->md->is_interposed) {
+		struct block_device *original = dd->dm_dev->bdev;
+		struct block_device *interposer = t->md->disk->part0;
+
+		if ((ti->begin != 0) || (ti->len < bdev_nr_sectors(original))) {
+			dm_put_device(ti, dd->dm_dev);
+			DMERR("The interposer device should not be less than the original.");
+			return -EINVAL;

Can you explain why allowing the device to be larger is meaningful?  Not
saying it isn't I'd just like to understand use-cases you forsee.

This _should_ hopefully get cleaned up by pushing such state into block
core's interposer interfaces.

But again, not seeing what utility/safety this extra flag is providing
to begin with.  Is this state _actually_ needed at all?

Again, I'd like this state to be part of 'struct block_device'

Please rename to DM_INTERPOSE_FLAG

+
 #endif				/* _LINUX_DM_IOCTL_H */
-- 
2.20.1

On Fri, Mar 12 2021 at 10:44am -0500,
Sergei Shtepa [off-list ref] wrote:

This is returning a frozen state that is immediately stale.  I don't
think any code calling this is providing the guarantees you think it
does due to the racey nature of this state once the mutex is dropped.

On Fri, Mar 12, 2021 at 02:06:41PM -0500, Mike Snitzer wrote:

This is returning a frozen state that is immediately stale.  I don't
think any code calling this is providing the guarantees you think it
does due to the racey nature of this state once the mutex is dropped.

The code only uses it for asserts in the form of WARN_ONs.

On Fri, Mar 12, 2021 at 06:44:54PM +0300, Sergei Shtepa wrote:

bdev_interposer allows to redirect bio requests to another devices.

I think this warrants a somewhat more detailed description.

The code itself looks pretty good to me now, a bunch of nitpicks and
a question below:

+static noinline blk_qc_t submit_bio_interposed(struct bio *bio)
+{
+	blk_qc_t ret = BLK_QC_T_NONE;
+	struct bio_list bio_list[2] = { };
+	struct gendisk *orig_disk;
+
+	if (current->bio_list) {
+		bio_list_add(&current->bio_list[0], bio);
+		return BLK_QC_T_NONE;
+	}

I don't think this case can ever happen:

 - current->bio_list != NULL means a ->submit_bio or blk_mq_submit_bio
   is active.  But if this device is being interposed this means the
   interposer recurses into itself, which should never happen.  So
   I think we'll want a WARN_ON_ONCE here as a debug check instead.

+
+	orig_disk = bio->bi_bdev->bd_disk;
+	if (unlikely(bio_queue_enter(bio)))
+		return BLK_QC_T_NONE;
+
+	current->bio_list = bio_list;
+
+	do {
+		struct block_device *interposer = bio->bi_bdev->bd_interposer;
+
+		if (unlikely(!interposer)) {
+			/* interposer was removed */
+			bio_list_add(&current->bio_list[0], bio);
+			break;
+		}
+		/* assign bio to interposer device */
+		bio_set_dev(bio, interposer);
+		bio_set_flag(bio, BIO_INTERPOSED);

Reassigning the bi_bdev here means the original source is lost by the
time we reach the interposer.  This initially seemed a little limiting,
but I guess the interposer driver can just record that information
locally, so we should be fine.  The big upside of this is that no
extra argument to submit_bio_checks, which means less changes to the
normal fast path, so if this works for everyone that is a nice
improvement over my draft.

+
+		if (!submit_bio_checks(bio))
+			break;
+		/*
+		 * Because the current->bio_list is initialized,
+		 * the submit_bio callback will always return BLK_QC_T_NONE.
+		 */
+		interposer->bd_disk->fops->submit_bio(bio);
+	} while (false);

I find the do { ... } while (false) idiom here a little strange.  Normal
kernel style would be a goto done instead of the breaks.

+int bdev_interposer_attach(struct block_device *original,
+			   struct block_device *interposer)

A kerneldoc comment for bdev_interposer_attach (and
bdev_interposer_detach) would be nice to explain the API a little more.

+{
+	int ret = 0;
+
+	if (WARN_ON(((!original) || (!interposer))))
+		return -EINVAL;

No need for the inner two levels of braces.

+	 * interposer should be simple, no a multi-queue device
+	 */
+	if (!interposer->bd_disk->fops->submit_bio)

Please use queue_is_mq() instead.

+	if (bdev_has_interposer(original))
+		ret = -EBUSY;
+	else {
+		original->bd_interposer = bdgrab(interposer);

Just thinking out a loud:  what happens if the interposed device
goes away?  Shouldn't we at very least also make sure this
gabs another refererence on bdev as well?

+struct bdev_interposer;

Not needed any more.

+static inline bool bdev_has_interposer(struct block_device *bdev)
+{
+	return (bdev->bd_interposer != NULL);
+};

No need for the braces.

On Fri, Mar 12, 2021 at 06:44:55PM +0300, Sergei Shtepa wrote:

DM_INTERPOSED_FLAG allow to create DM targets on "the fly".
Underlying block device opens without a flag FMODE_EXCL.
DM target receives bio from the original device via
bdev_interposer.

This is more of a philopical comment, but the idea of just letting the
interposed reopen the device by itself seems like a bad idea.  I think
that is probably better hidden in the block layer interposer attachment
function, which could do the extra blkdev_get_by_dev for the caller.

The 03/14/2021 12:14, Christoph Hellwig wrote:

On Fri, Mar 12, 2021 at 02:06:41PM -0500, Mike Snitzer wrote:

quoted

This is returning a frozen state that is immediately stale.  I don't
think any code calling this is providing the guarantees you think it
does due to the racey nature of this state once the mutex is dropped.

The code only uses it for asserts in the form of WARN_ONs.

But perhaps it is possible to come up with a more elegant solution?
I'll think about it.
-- 
Sergei Shtepa
Veeam Software developer.

The 03/12/2021 22:00, Mike Snitzer wrote:

On Fri, Mar 12 2021 at 10:44am -0500,
Sergei Shtepa [off-list ref] wrote:

quoted

DM_INTERPOSED_FLAG allow to create DM targets on "the fly".
Underlying block device opens without a flag FMODE_EXCL.
DM target receives bio from the original device via
bdev_interposer.

Signed-off-by: Sergei Shtepa <redacted>
---
 drivers/md/dm-core.h          |  3 ++
 drivers/md/dm-ioctl.c         | 13 ++++++++
 drivers/md/dm-table.c         | 61 +++++++++++++++++++++++++++++------
 drivers/md/dm.c               | 38 +++++++++++++++-------
 include/linux/device-mapper.h |  1 +
 include/uapi/linux/dm-ioctl.h |  6 ++++
 6 files changed, 101 insertions(+), 21 deletions(-)

diff --git a/drivers/md/dm-core.h b/drivers/md/dm-core.h
index 5953ff2bd260..9eae419c7b18 100644
--- a/drivers/md/dm-core.h
+++ b/drivers/md/dm-core.h

@@ -114,6 +114,9 @@ struct mapped_device {
 	bool init_tio_pdu:1;
 
 	struct srcu_struct io_barrier;
+
+	/* attach target via block-layer interposer */
+	bool is_interposed;
 };

This flag is a mix of uses.  First it is used to store that
DM_INTERPOSED_FLAG was provided as input param during load.

And the same 'is_interposed' name is used in 'struct dm_dev'.

To me this state should be elevated to block core -- awkward for every
driver that might want to use blk-interposer to be sprinkling state
around its core structures.

So I'd prefer you:
1) rename 'struct mapped_device' to 'interpose' _and_ add it just after
   "bool init_tio_pdu:1;" with "bool interpose:1;" -- this reflects
   interpose was requested during load.
2) bdev_interposer_attach() should be made to set some block core state
   that is able to be tested to check if a device is_interposed.
3) Don't add an 'is_interposed' flag to 'struct dm_dev'

Ok, but little fix in "rename 'struct mapped_device' to 'interpose'".
Maybe "rename 'bool is_interposed' to 'bool interpose:1'"?
I think I understand your idea, I'll try to implement it.

quoted

 
 void disable_discard(struct mapped_device *md);

diff --git a/drivers/md/dm-ioctl.c b/drivers/md/dm-ioctl.c
index 5e306bba4375..2b4c9557c283 100644
--- a/drivers/md/dm-ioctl.c
+++ b/drivers/md/dm-ioctl.c

@@ -1267,6 +1267,11 @@ static inline fmode_t get_mode(struct dm_ioctl *param)
 	return mode;
 }
 
+static inline bool get_interposer_flag(struct dm_ioctl *param)
+{
+	return (param->flags & DM_INTERPOSED_FLAG);
+}
+

As I mention at the end: rename to DM_INTERPOSE_FLAG

Ok.

quoted

 static int next_target(struct dm_target_spec *last, uint32_t next, void *end,
 		       struct dm_target_spec **spec, char **target_params)
 {

@@ -1293,6 +1298,10 @@ static int populate_table(struct dm_table *table,
 		DMWARN("populate_table: no targets specified");
 		return -EINVAL;
 	}
+	if (table->md->is_interposed && (param->target_count != 1)) {
+		DMWARN("%s: with interposer should be specified only one target", __func__);

This error/warning reads very awkwardly. Maybe?:
"%s: interposer requires only a single target be specified"

Ok.

quoted

+		return -EINVAL;
+	}
 
 	for (i = 0; i < param->target_count; i++) {
 

@@ -1338,6 +1347,8 @@ static int table_load(struct file *filp, struct dm_ioctl *param, size_t param_si
 	if (!md)
 		return -ENXIO;
 
+	md->is_interposed = get_interposer_flag(param);
+
 	r = dm_table_create(&t, get_mode(param), param->target_count, md);
 	if (r)
 		goto err;

@@ -2098,6 +2109,8 @@ int __init dm_early_create(struct dm_ioctl *dmi,
 	if (r)
 		goto err_hash_remove;
 
+	md->is_interposed = get_interposer_flag(dmi);
+
 	/* add targets */
 	for (i = 0; i < dmi->target_count; i++) {
 		r = dm_table_add_target(t, spec_array[i]->target_type,

diff --git a/drivers/md/dm-table.c b/drivers/md/dm-table.c
index 95391f78b8d5..f6e2eb3f8949 100644
--- a/drivers/md/dm-table.c
+++ b/drivers/md/dm-table.c

@@ -225,12 +225,13 @@ void dm_table_destroy(struct dm_table *t)
 /*
  * See if we've already got a device in the list.
  */
-static struct dm_dev_internal *find_device(struct list_head *l, dev_t dev)
+static struct dm_dev_internal *find_device(struct list_head *l, dev_t dev, bool is_interposed)

Think in make more sense to internalize the need to consider
md->interpose here.

So:

static struct dm_dev_internal *find_device(struct dm_table *t, dev_t dev)
{
        struct list_head *l = &t->devices;
        bool is_interposed = t->md->interpose;
        ...

Yes.

quoted

 {
 	struct dm_dev_internal *dd;
 
 	list_for_each_entry (dd, l, list)
-		if (dd->dm_dev->bdev->bd_dev == dev)
+		if ((dd->dm_dev->bdev->bd_dev == dev) &&
+		    (dd->dm_dev->is_interposed == is_interposed))
 			return dd;

But why must this extra state be used/tested?  Seems like quite a deep
embedding of interposer into dm_dev_internal.. feels unnecessary.

Yeah, I guess so.
Since dm_table is unique for each dm_target and in our case will generally
contain only one device.

quoted

 
 	return NULL;

@@ -358,6 +359,18 @@ dev_t dm_get_dev_t(const char *path)
 }
 EXPORT_SYMBOL_GPL(dm_get_dev_t);
 
+static inline void dm_disk_freeze(struct gendisk *disk)
+{
+	blk_mq_freeze_queue(disk->queue);
+	blk_mq_quiesce_queue(disk->queue);
+}
+
+static inline void dm_disk_unfreeze(struct gendisk *disk)
+{
+	blk_mq_unquiesce_queue(disk->queue);
+	blk_mq_unfreeze_queue(disk->queue);
+}
+

These interfaces don't account for bio-based at all (pretty sure we've
been over this and you pointed out that they'll just return early), but
they also don't take steps to properly flush outstanding DM io.
Shouldn't you require DM devices do an internal suspend/resume?  And if
original device isn't DM then fallback to blk_mq calls?

I thought that was enough.
The function dm_disk_freeze() guarantees that no bio will be processed,
which means that we can safely attach or detach the interposer.
All bios will continue their processing after dm_disk_unfreeze().
Sorry, but I don't understand why it is required for DM devices to do
internal suspend/resume.
If this is not too difficult, can you explain what problem can cause
a simple queue freeze?

quoted

 /*
  * Add a device to the list, or just increment the usage count if
  * it's already present.

@@ -385,7 +398,7 @@ int dm_get_device(struct dm_target *ti, const char *path, fmode_t mode,
 			return -ENODEV;
 	}
 
-	dd = find_device(&t->devices, dev);
+	dd = find_device(&t->devices, dev, t->md->is_interposed);
 	if (!dd) {
 		dd = kmalloc(sizeof(*dd), GFP_KERNEL);
 		if (!dd)

@@ -398,15 +411,38 @@ int dm_get_device(struct dm_target *ti, const char *path, fmode_t mode,
 
 		refcount_set(&dd->count, 1);
 		list_add(&dd->list, &t->devices);
-		goto out;
-
 	} else if (dd->dm_dev->mode != (mode | dd->dm_dev->mode)) {
 		r = upgrade_mode(dd, mode, t->md);
 		if (r)
 			return r;
+		refcount_inc(&dd->count);
 	}
-	refcount_inc(&dd->count);

This looks bogus... you cannot only increment refcount with the mode
check/upgrade branch (IIRC: I've made this same mistake in the past)

I realized my mistake. I wanted to remove the unnecessary goto,
but I change the logic.

quoted

-out:
+
+	if (t->md->is_interposed) {
+		struct block_device *original = dd->dm_dev->bdev;
+		struct block_device *interposer = t->md->disk->part0;
+
+		if ((ti->begin != 0) || (ti->len < bdev_nr_sectors(original))) {
+			dm_put_device(ti, dd->dm_dev);
+			DMERR("The interposer device should not be less than the original.");
+			return -EINVAL;

Can you explain why allowing the device to be larger is meaningful?  Not
saying it isn't I'd just like to understand use-cases you forsee.

Hmm... Maybe strict equality would be better.

quoted

+		}
+
+		/*
+		 * Attach mapped interposer device to original.
+		 * It is quite convenient that device mapper creates
+		 * one disk for one block device.
+		 */
+		dm_disk_freeze(original->bd_disk);
+		r = bdev_interposer_attach(original, interposer);
+		dm_disk_unfreeze(original->bd_disk);
+		if (r) {
+			dm_put_device(ti, dd->dm_dev);
+			DMERR("Failed to attach dm interposer.");
+			return r;
+		}
+	}
+
 	*result = dd->dm_dev;
 	return 0;
 }

@@ -446,6 +482,7 @@ void dm_put_device(struct dm_target *ti, struct dm_dev *d)
 {
 	int found = 0;
 	struct list_head *devices = &ti->table->devices;
+	struct mapped_device *md = ti->table->md;
 	struct dm_dev_internal *dd;
 
 	list_for_each_entry(dd, devices, list) {

@@ -456,11 +493,17 @@ void dm_put_device(struct dm_target *ti, struct dm_dev *d)
 	}
 	if (!found) {
 		DMWARN("%s: device %s not in table devices list",
-		       dm_device_name(ti->table->md), d->name);
+		       dm_device_name(md), d->name);
 		return;
 	}
+	if (md->is_interposed) {
+		dm_disk_freeze(d->bdev->bd_disk);
+		bdev_interposer_detach(d->bdev);
+		dm_disk_unfreeze(d->bdev->bd_disk);
+	}
+
 	if (refcount_dec_and_test(&dd->count)) {
-		dm_put_table_device(ti->table->md, d);
+		dm_put_table_device(md, d);
 		list_del(&dd->list);
 		kfree(dd);
 	}

diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index 50b693d776d6..466bf70a66b0 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c

@@ -762,16 +762,24 @@ static int open_table_device(struct table_device *td, dev_t dev,
 
 	BUG_ON(td->dm_dev.bdev);
 
-	bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _dm_claim_ptr);
-	if (IS_ERR(bdev))
-		return PTR_ERR(bdev);
+	if (md->is_interposed) {
 
-	r = bd_link_disk_holder(bdev, dm_disk(md));
-	if (r) {
-		blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
-		return r;
+		bdev = blkdev_get_by_dev(dev, td->dm_dev.mode, NULL);
+		if (IS_ERR(bdev))
+			return PTR_ERR(bdev);
+	} else {
+		bdev = blkdev_get_by_dev(dev, td->dm_dev.mode | FMODE_EXCL, _dm_claim_ptr);
+		if (IS_ERR(bdev))
+			return PTR_ERR(bdev);
+
+		r = bd_link_disk_holder(bdev, dm_disk(md));
+		if (r) {
+			blkdev_put(bdev, td->dm_dev.mode | FMODE_EXCL);
+			return r;
+		}
 	}
 
+	td->dm_dev.is_interposed = md->is_interposed;

This _should_ hopefully get cleaned up by pushing such state into block
core's interposer interfaces.

But again, not seeing what utility/safety this extra flag is providing
to begin with.  Is this state _actually_ needed at all?

Yes, it is not necessary.

quoted

 	td->dm_dev.bdev = bdev;
 	td->dm_dev.dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
 	return 0;

@@ -785,20 +793,26 @@ static void close_table_device(struct table_device *td, struct mapped_device *md
 	if (!td->dm_dev.bdev)
 		return;
 
-	bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
-	blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
+	if (td->dm_dev.is_interposed)
+		blkdev_put(td->dm_dev.bdev, td->dm_dev.mode);
+	else {
+		bd_unlink_disk_holder(td->dm_dev.bdev, dm_disk(md));
+		blkdev_put(td->dm_dev.bdev, td->dm_dev.mode | FMODE_EXCL);
+	}
 	put_dax(td->dm_dev.dax_dev);
 	td->dm_dev.bdev = NULL;
 	td->dm_dev.dax_dev = NULL;
 }
 
 static struct table_device *find_table_device(struct list_head *l, dev_t dev,
-					      fmode_t mode)
+					      fmode_t mode, bool is_interposed)
 {
 	struct table_device *td;
 
 	list_for_each_entry(td, l, list)
-		if (td->dm_dev.bdev->bd_dev == dev && td->dm_dev.mode == mode)
+		if (td->dm_dev.bdev->bd_dev == dev &&
+		    td->dm_dev.mode == mode &&
+		    td->dm_dev.is_interposed == is_interposed)
 			return td;
 
 	return NULL;

@@ -811,7 +825,7 @@ int dm_get_table_device(struct mapped_device *md, dev_t dev, fmode_t mode,
 	struct table_device *td;
 
 	mutex_lock(&md->table_devices_lock);
-	td = find_table_device(&md->table_devices, dev, mode);
+	td = find_table_device(&md->table_devices, dev, mode, md->is_interposed);
 	if (!td) {
 		td = kmalloc_node(sizeof(*td), GFP_KERNEL, md->numa_node_id);
 		if (!td) {

diff --git a/include/linux/device-mapper.h b/include/linux/device-mapper.h
index 7f4ac87c0b32..76a6dfb1cb29 100644
--- a/include/linux/device-mapper.h
+++ b/include/linux/device-mapper.h

@@ -159,6 +159,7 @@ struct dm_dev {
 	struct block_device *bdev;
 	struct dax_device *dax_dev;
 	fmode_t mode;
+	bool is_interposed;

Again, I'd like this state to be part of 'struct block_device'

Ok.

quoted

 	char name[16];
 };
 

diff --git a/include/uapi/linux/dm-ioctl.h b/include/uapi/linux/dm-ioctl.h
index fcff6669137b..fc4d06bb3dbb 100644
--- a/include/uapi/linux/dm-ioctl.h
+++ b/include/uapi/linux/dm-ioctl.h

@@ -362,4 +362,10 @@ enum {
  */
 #define DM_INTERNAL_SUSPEND_FLAG	(1 << 18) /* Out */
 
+/*
+ * If set, the underlying device should open without FMODE_EXCL
+ * and attach mapped device via bdev_interposer.
+ */
+#define DM_INTERPOSED_FLAG		(1 << 19) /* In */

Please rename to DM_INTERPOSE_FLAG

Ok.

quoted

+
 #endif				/* _LINUX_DM_IOCTL_H */
-- 
2.20.1

-- 
Sergei Shtepa
Veeam Software developer.

The 03/14/2021 12:28, Christoph Hellwig wrote:

On Fri, Mar 12, 2021 at 06:44:54PM +0300, Sergei Shtepa wrote:

quoted

bdev_interposer allows to redirect bio requests to another devices.

I think this warrants a somewhat more detailed description.

The code itself looks pretty good to me now, a bunch of nitpicks and
a question below:

quoted

+static noinline blk_qc_t submit_bio_interposed(struct bio *bio)
+{
+	blk_qc_t ret = BLK_QC_T_NONE;
+	struct bio_list bio_list[2] = { };
+	struct gendisk *orig_disk;
+
+	if (current->bio_list) {
+		bio_list_add(&current->bio_list[0], bio);
+		return BLK_QC_T_NONE;
+	}

I don't think this case can ever happen:

 - current->bio_list != NULL means a ->submit_bio or blk_mq_submit_bio
   is active.  But if this device is being interposed this means the
   interposer recurses into itself, which should never happen.  So
   I think we'll want a WARN_ON_ONCE here as a debug check instead.

Yes, it is.
Completely remove this check or add "BUG_ON(current->bio_list);" for
an emergency?

quoted

+
+	orig_disk = bio->bi_bdev->bd_disk;
+	if (unlikely(bio_queue_enter(bio)))
+		return BLK_QC_T_NONE;
+
+	current->bio_list = bio_list;
+
+	do {
+		struct block_device *interposer = bio->bi_bdev->bd_interposer;
+
+		if (unlikely(!interposer)) {
+			/* interposer was removed */
+			bio_list_add(&current->bio_list[0], bio);
+			break;
+		}
+		/* assign bio to interposer device */
+		bio_set_dev(bio, interposer);
+		bio_set_flag(bio, BIO_INTERPOSED);

Reassigning the bi_bdev here means the original source is lost by the
time we reach the interposer.  This initially seemed a little limiting,
but I guess the interposer driver can just record that information
locally, so we should be fine.  The big upside of this is that no
extra argument to submit_bio_checks, which means less changes to the
normal fast path, so if this works for everyone that is a nice
improvement over my draft.

quoted

+
+		if (!submit_bio_checks(bio))
+			break;
+		/*
+		 * Because the current->bio_list is initialized,
+		 * the submit_bio callback will always return BLK_QC_T_NONE.
+		 */
+		interposer->bd_disk->fops->submit_bio(bio);
+	} while (false);

I find the do { ... } while (false) idiom here a little strange.  Normal
kernel style would be a goto done instead of the breaks.

Ok. I'll use the normal kernel style.

quoted

+int bdev_interposer_attach(struct block_device *original,
+			   struct block_device *interposer)

A kerneldoc comment for bdev_interposer_attach (and
bdev_interposer_detach) would be nice to explain the API a little more.

Yes, I should add kerneldoc comments.

quoted

+{
+	int ret = 0;
+
+	if (WARN_ON(((!original) || (!interposer))))
+		return -EINVAL;

No need for the inner two levels of braces.

Ok.

quoted

+	 * interposer should be simple, no a multi-queue device
+	 */
+	if (!interposer->bd_disk->fops->submit_bio)

Please use queue_is_mq() instead.

Ok.

quoted

+	if (bdev_has_interposer(original))
+		ret = -EBUSY;
+	else {
+		original->bd_interposer = bdgrab(interposer);

Just thinking out a loud:  what happens if the interposed device
goes away?  Shouldn't we at very least also make sure this
gabs another refererence on bdev as well?

If the original device is removed from the system, the interposer device
will be permanently occupied. I need to add an interposer release when
deleting a block device.

quoted

+struct bdev_interposer;

Not needed any more.

Yes.

quoted

+static inline bool bdev_has_interposer(struct block_device *bdev)
+{
+	return (bdev->bd_interposer != NULL);
+};

No need for the braces.

Ok.

-- 
Sergei Shtepa
Veeam Software developer.

The 03/14/2021 12:30, Christoph Hellwig wrote:

On Fri, Mar 12, 2021 at 06:44:55PM +0300, Sergei Shtepa wrote:

quoted

DM_INTERPOSED_FLAG allow to create DM targets on "the fly".
Underlying block device opens without a flag FMODE_EXCL.
DM target receives bio from the original device via
bdev_interposer.

This is more of a philopical comment, but the idea of just letting the
interposed reopen the device by itself seems like a bad idea.  I think
that is probably better hidden in the block layer interposer attachment
function, which could do the extra blkdev_get_by_dev for the caller.

I suppose this cannot be implemented, since we need to change the behavior
for block devices that already have been opened.

-- 
Sergei Shtepa
Veeam Software developer.

On Fri, Mar 12, 2021 at 06:44:54PM +0300, Sergei Shtepa wrote:

Given original request queue may become live when calling attach() and
detach(), see below comment. bdev_interposer_detach() may be run
when running ->submit_bio(), meantime the interposer device is
gone during the period, then kernel oops.

The original request queue may become live now...

+
+	mutex_lock(&bdev_interposer_attach_lock);
+
+	if (bdev_has_interposer(original))
+		ret = -EBUSY;
+	else {
+		original->bd_interposer = bdgrab(interposer);
+		if (!original->bd_interposer)
+			ret = -ENODEV;
+	}
+
+	mutex_unlock(&bdev_interposer_attach_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(bdev_interposer_attach);
+
+void bdev_interposer_detach(struct block_device *original)
+{
+	if (WARN_ON(!original))
+		return;
+
+	if (WARN_ON(!blk_mq_is_queue_frozen(original->bd_disk->queue)))
+		return;

The original request queue may become live now...


-- 
Ming

On Mon, Mar 15, 2021 at 04:25:09PM +0300, Sergei Shtepa wrote:

The 03/14/2021 12:30, Christoph Hellwig wrote:

quoted

On Fri, Mar 12, 2021 at 06:44:55PM +0300, Sergei Shtepa wrote:

quoted

DM_INTERPOSED_FLAG allow to create DM targets on "the fly".
Underlying block device opens without a flag FMODE_EXCL.
DM target receives bio from the original device via
bdev_interposer.

This is more of a philopical comment, but the idea of just letting the
interposed reopen the device by itself seems like a bad idea.  I think
that is probably better hidden in the block layer interposer attachment
function, which could do the extra blkdev_get_by_dev for the caller.

I suppose this cannot be implemented, since we need to change the behavior
for block devices that already have been opened.

That's not what I mean.  Take a look at the patch relative to your
series to let me know what you think.  The new blkdev_interposer_attach
now takes a dev_t + mode for the original device and opens it on
behalf of the interposer.  It also moves the queue freezing into the
API, which should address the concerns about the helper and adds a few
more sanity checks.

On Tue, Mar 16, 2021 at 03:23:14PM +0000, Christoph Hellwig wrote:

On Mon, Mar 15, 2021 at 04:25:09PM +0300, Sergei Shtepa wrote:

quoted

The 03/14/2021 12:30, Christoph Hellwig wrote:

quoted

On Fri, Mar 12, 2021 at 06:44:55PM +0300, Sergei Shtepa wrote:

quoted

DM_INTERPOSED_FLAG allow to create DM targets on "the fly".
Underlying block device opens without a flag FMODE_EXCL.
DM target receives bio from the original device via
bdev_interposer.

This is more of a philopical comment, but the idea of just letting the
interposed reopen the device by itself seems like a bad idea.  I think
that is probably better hidden in the block layer interposer attachment
function, which could do the extra blkdev_get_by_dev for the caller.

I suppose this cannot be implemented, since we need to change the behavior
for block devices that already have been opened.

That's not what I mean.  Take a look at the patch relative to your
series to let me know what you think.  The new blkdev_interposer_attach
now takes a dev_t + mode for the original device and opens it on
behalf of the interposer.  It also moves the queue freezing into the
API, which should address the concerns about the helper and adds a few
more sanity checks.

And now actually with the diff:

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 2f188a865024ac..d4d7c1caa43966 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c

@@ -161,19 +161,6 @@ int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 }
 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 
-bool blk_mq_is_queue_frozen(struct request_queue *q)
-{
-	bool frozen;
-
-	mutex_lock(&q->mq_freeze_lock);
-	frozen = percpu_ref_is_dying(&q->q_usage_counter) &&
-		 percpu_ref_is_zero(&q->q_usage_counter);
-	mutex_unlock(&q->mq_freeze_lock);
-
-	return frozen;
-}
-EXPORT_SYMBOL_GPL(blk_mq_is_queue_frozen);
-
 /*
  * Guarantee no request is in use, so we can change any data structure of
  * the queue afterward.

diff --git a/block/genhd.c b/block/genhd.c
index fa406b972371ae..64d6338b08cc87 100644
--- a/block/genhd.c
+++ b/block/genhd.c

@@ -1944,51 +1944,70 @@ static void disk_release_events(struct gendisk *disk)
 	kfree(disk->ev);
 }
 
-int bdev_interposer_attach(struct block_device *original,
+struct block_device *blkdev_interposer_attach(dev_t dev, fmode_t mode,
 			   struct block_device *interposer)
 {
+	struct block_device *bdev;
 	int ret = 0;
 
-	if (WARN_ON(((!original) || (!interposer))))
-		return -EINVAL;
-	/*
-	 * interposer should be simple, no a multi-queue device
-	 */
-	if (!interposer->bd_disk->fops->submit_bio)
-		return -EINVAL;
+	if (WARN_ON_ONCE(!bdev_is_partition(interposer)))
+		return ERR_PTR(-EINVAL);
+	if (WARN_ON_ONCE(!queue_is_mq(interposer->bd_disk->queue)))
+		return ERR_PTR(-EINVAL);
 
-	if (WARN_ON(!blk_mq_is_queue_frozen(original->bd_disk->queue)))
-		return -EPERM;
+	bdev = blkdev_get_by_dev(dev, mode, NULL);
+	if (IS_ERR(bdev))
+		return bdev;
 
-	mutex_lock(&bdev_interposer_attach_lock);
+	ret = -EINVAL;
+	if (WARN_ON_ONCE(bdev_nr_sectors(bdev) != bdev_nr_sectors(interposer)))
+		goto out;
 
-	if (bdev_has_interposer(original))
-		ret = -EBUSY;
-	else {
-		original->bd_interposer = bdgrab(interposer);
-		if (!original->bd_interposer)
-			ret = -ENODEV;
-	}
+	blk_mq_freeze_queue(bdev->bd_disk->queue);
+	blk_mq_quiesce_queue(bdev->bd_disk->queue);
 
+	mutex_lock(&bdev_interposer_attach_lock);
+	ret = -EBUSY;
+	if (bdev_has_interposer(bdev))
+		goto out_unlock;
+	ret = -ENODEV;
+	bdev->bd_interposer = bdgrab(interposer);
+	if (!bdev->bd_interposer)
+		goto out_unlock;
+	ret = 0;
+out_unlock:
 	mutex_unlock(&bdev_interposer_attach_lock);
 
-	return ret;
+	blk_mq_unquiesce_queue(bdev->bd_disk->queue);
+	blk_mq_unfreeze_queue(bdev->bd_disk->queue);
+out:
+	if (ret) {
+		blkdev_put(bdev, mode);
+		bdev = ERR_PTR(ret);
+	}
+
+	return bdev;
 }
-EXPORT_SYMBOL_GPL(bdev_interposer_attach);
+EXPORT_SYMBOL_GPL(blkdev_interposer_attach);
 
-void bdev_interposer_detach(struct block_device *original)
+void blkdev_interposer_detach(struct block_device *bdev, fmode_t mode)
 {
-	if (WARN_ON(!original))
-		return;
+	struct block_device *interposer;
 
-	if (WARN_ON(!blk_mq_is_queue_frozen(original->bd_disk->queue)))
+	if (WARN_ON_ONCE(!bdev_has_interposer(bdev)))
 		return;
 
+	blk_mq_freeze_queue(bdev->bd_disk->queue);
+	blk_mq_quiesce_queue(bdev->bd_disk->queue);
+
 	mutex_lock(&bdev_interposer_attach_lock);
-	if (bdev_has_interposer(original)) {
-		bdput(original->bd_interposer);
-		original->bd_interposer = NULL;
-	}
+	interposer = bdev->bd_interposer;
+	bdev->bd_interposer = NULL;
 	mutex_unlock(&bdev_interposer_attach_lock);
+
+	blk_mq_unquiesce_queue(bdev->bd_disk->queue);
+	blk_mq_unfreeze_queue(bdev->bd_disk->queue);
+
+	blkdev_put(interposer, mode);
 }
-EXPORT_SYMBOL_GPL(bdev_interposer_detach);
+EXPORT_SYMBOL_GPL(blkdev_interposer_detach);